Intern. J. of Research in Marketing 31 (2014) 127–140
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
From academic research to marketing practice: Exploring the marketing
science value chain
John H. Roberts a,⁎, Ujwal Kayande b, Stefan Stremersch c,d
a
London Business School and Australian National University, Canberra, Australia
Melbourne Business School, Melbourne, Australia
Erasmus School of Economics, Erasmus University Rotterdam, The Netherlands
d
IESE Business School, University of Navarra, Barcelona, Spain
b
c
a r t i c l e
i n f o
Article history:
First received in 28 September 2012 and was
under review for 4 months
Available online 1 October 2013
Area Editor: Dominique M. Hanssens
Guest Editor: Marnik G. Dekimpe
a b s t r a c t
We aim to investigate the impact of marketing science articles and tools on the practice of marketing. This impact
may be direct (e.g., an academic article may be adapted to solve a practical problem) or indirect (e.g., its contents
may be incorporated into practitioners' tools, which then influence marketing decision making). We use the term
“marketing science value chain” to describe these diffusion steps, and survey marketing managers, marketing science intermediaries (practicing marketing analysts), and marketing academics to calibrate the value chain.
In our sample, we find that (1) the impact of marketing science is perceived to be largest on decisions such as the
management of brands, pricing, new products, product portfolios, and customer/market selection, and (2) tools
such as segmentation, survey-based choice models, marketing mix models, and pre-test market models have the
largest impact on marketing decisions. Exemplary papers from 1982 to 2003 that achieved dual – academic and
practice – impact are Guadagni and Little (1983) and Green and Srinivasan (1990). Overall, our results are encouraging. First, we find that the impact of marketing science has been largest on marketing decision areas
that are important to practice. Second, we find moderate alignment between academic impact and practice impact. Third, we identify antecedents of practice impact among dual impact marketing science papers. Fourth, we
discover more recent trends and initiatives in the period 2004–2012, such as the increased importance of big data
and the rise of digital and mobile communication, using the marketing science value chain as an organizing
framework.
© 2013 The Authors. Published by Elsevier B.V. Open access under CC BY-NC-ND license.
1. Introduction
Does marketing science research affect marketing practice? Which
decisions have marketing science articles supported? To which tools
has marketing science contributed? Which marketing science articles
have had dual impact on both science and practice? These are key questions that we address in this paper. We define marketing science as the
development and use of quantifiable concepts and quantitative tools
to understand marketplace behavior and the effect of marketing activity
upon it. From this definition, one would consider it reasonable for
marketing scientists to seek impact on marketing practice, i.e., seek
relevance.
However, marketing scientists have recently rekindled the age-old
debate on rigor versus relevance. On the one hand, marketing science
has been very successful in attracting scholars from other fields such
⁎ Corresponding author.
E-mail addresses: jhroberts@london.edu (J.H. Roberts), U.Kayande@mbs.edu
(U. Kayande), stremersch@ese.eur.nl (S. Stremersch).
as economics, statistics, econometrics and psychology. This inflow of
talented scientists from other fields has clearly added to the rigor of
marketing science and has allowed the development of new techniques.
On the other hand, a number of academic scholars have recently called
for more emphasis to be placed on the application of marketing science
to industry problems, rather than rigor per se (e.g., Lehmann, McAlister,
& Staelin, 2011; Lilien, 2011; Reibstein, Day, & Wind, 2009). Such application may also show positive returns to firms. Germann, Lilien, and
Rangaswamy (2013) find that increasing analytics deployment by
firms leads to an improvement in their return on assets.
Despite the importance of this debate for our field and the strong interest in the drivers of academic impact (e.g., see Stremersch & Verhoef,
2005; Stremersch, Verniers, & Verhoef, 2007), empirical examination of
the impact of marketing science on practice is rare. Valuable exceptions
are Bucklin and Gupta (1999), Cattin and Wittink (1982), Wittink and
Cattin (1989), and Wittink, Vriens, and Burhenne (1994). However,
their application areas were narrow. Wittink and his colleagues studied
the commercial use of conjoint analysis in North America and Europe,
while Bucklin and Gupta studied the usage of scanner data and the
models that scholars have developed to analyze them. Other scholars
http://dx.doi.org/10.1016/j.ijresmar.2013.07.006
0167-8116 © 2013 The Authors. Published by Elsevier B.V. Open access under CC BY-NC-ND license.
128
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
have conceptually reviewed the impact of marketing science and prescribed areas in which marketing science might have an impact in the
future. In a special issue of the International Journal of Research in
Marketing, Leeflang and Wittink (2000) summarized the areas in
which marketing science has been used to inform management decisions. Roberts (2000) acknowledged the breadth of marketing science
applications, but lamented the depth of penetration of marketing science (i.e., the proportion of management decisions informed by marketing science models). Lilien, Roberts, and Shankar (2013) take an
applications-based approach to best practice. However, there has been
no broad systematic investigation of which marketing science articles
and tools have been applied, the decisions that these concepts and
tools have informed, and the perceptions of different stakeholders of
the usefulness of marketing science in informing decisions. We aim to
address this void.
We develop the concept of the marketing science value chain, which
captures the diffusion of insights from academic articles in a direct
(e.g., from article to practice) or indirect (e.g., from article to marketing
science tool to practice) manner. We survey the primary agents in this
value chain – marketing managers, marketing science intermediaries
(marketing analysts), and marketing academics – to calibrate the practice impact of marketing science in all its facets.
2. Methodology
2.1. The framework: The marketing science value chain
An important step in our methodology is a conceptualization of the
marketing science value chain. Our representation of this chain, illustrated in Fig. 1, depicts activities (full arrows) by which marketing science is translated from academic knowledge to practical tools, and
thence to marketing action, as well as the participants involved in the
chain.
First, new knowledge (marketing science articles) is developed,
often but not always, by marketing academics.1 Second, knowledge
conversion occurs when new knowledge in articles is adapted and
integrated into practical tools and approaches, often but again not always, by marketing intermediaries, such as market research agencies (e.g. ACNielsen or GfK), marketing and strategy consultancies
(e.g., McKinsey or Bain), specialist niche marketing consulting firms
(e.g. Advanis or Simon-Kucher Partners), or the marketing science division of a marketing organization (e.g. Novartis or General Mills). Third,
knowledge application occurs when marketing managers implement
marketing science knowledge via practical tools to make marketing
decisions.
While we contend in Fig. 1 that marketing intermediaries play a critical role in the diffusion process, we allow for a direct path as well (disintermediation). For example, marketing academics may work directly
with marketing managers to have their tools adopted (marketing science push) or a firm's internal analysts may actively seek out solutions
to address the firm's specific problems (marketing science pull). Alternatively, the locus of conceptual innovation may fall further down the
value chain (user innovation). Moreover, diffusion may occur through
routes other than through intermediaries (for example, via specialist
books such as Lilien, Kotler, & Moorthy, 1992; Wierenga & van
Bruggen, 2000, and Lilien, Rangaswamy, & De Bruyn, 2007 or general
texts such as Kotler & Keller, 2012). In other words, the “direct” influence in Fig. 1 may include a number of further sub-stages that we do
not explicitly identify or calibrate.
2.2. The elements: Decisions, tools and articles
In Fig. 1, we identify three core elements in the marketing science
value chain: decisions, tools, and articles. Selection of stimuli in each
of these elements is a critical part of our methodology, especially
considering the scope of our study. Not only have thousands of marketing articles been published across many journals, but marketing managers make decisions to solve marketing problems in a wide variety of
areas (pricing, promotions, sales force management, etc.), using a considerable range of marketing science tools (segmentation tools, choice
models, etc.) to assist in that decision making. To make our calibration
practically feasible, we decided to limit the three sets of stimuli to 12 decision areas, 12 marketing science tools, and 20 marketing science articles. We decided on these limits iteratively, by trading off the need for a
comprehensive classification of the decisions, tools, and articles against
the time required for respondents to react to the stimuli. In Section 3.5
we discuss the dynamics of these three elements.
2.2.1. Decisions
Marketing decisions refer to the choice of management actions regarding any part of the firm's marketing activity. To categorize marketing decisions, we followed a four-step procedure. First, we examined
subject areas used at the major marketing journals and in leading marketing management textbooks. Second, we integrated and synthesized
these lists to create an exhaustive inventory. Third, we aggregated the
different decision areas into higher order categories, to create a manageable number. Finally, we tested our list with practicing managers and
the Executive Committee of the Marketing Science Institute, and refined
it based on their feedback. Our final list of marketing decision areas is:
1. Brand management: Developing, positioning and managing existing
brands.
Fig. 1. The marketing science value chain.
1
For example, a study of Marketing Science over the period 1982–2003 shows that of
1072 article authors, 1001 of them were academics (93.4%) Authors with multiple articles
are counted as many times as they have (co-)authored an article.
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
2. New product/service management: New product development, management and diffusion.
3. Marketing strategy: Product line, multi-product and portfolio
strategies.
4. Advertising management: Advertising spending, planning and
design.
5. Promotion management: Promotion decisions.
6. Pricing management: Pricing decisions.
7. Sales force management: Sales force size, allocation, and compensation decisions.
8. Channel management: Channel strategy, design, and monitoring.
9. Customer/market selection: Targeting decisions.
10. Relationship management: Customer value assessment and maximization, acquisition, retention, and relationship management.
11. Managing marketing investments: Organizing for higher returns and
internal marketing.
12. Service/product quality management: Any aspect of quality
management.
2.2.2. Tools
Tools are approaches and methodologies that can be used to support
marketing decisions. To categorize marketing science tools, we followed
a procedure similar to the one used for decision areas (using marketing
research and marketing analysis texts). Our list of tools is:
1.
2.
3.
4.
5.
6.
7.
8.
9.
Segmentation tools: latent class segmentation, cluster analysis, etc.
Perceptual mapping: multidimensional scaling, factor analysis, etc.
Survey-based choice models: conjoint analysis, discrete choice, etc.
Panel-based choice models: choice models, stochastic models, etc.
Pre-test market models: ASSESSOR, durable pre-testing, etc.
New product models: diffusion models, dynamic models, etc.
Aggregate marketing response models: marketing mix models, etc.
Sales force allocation models: Call planning models, etc.
Customer satisfaction models: Models of service quality, satisfaction,
etc.
10. Game theory models: Models of competition, channel structure, etc.
11. Customer lifetime value models: Loyalty and direct marketing models,
etc.
12. Marketing metrics: Accounting models, internal rate of return, etc.
2.2.3. Articles
We selected candidate articles for the twenty marketing science articles by applying four filters. First, we filter the journals and time period
from which to sample. Second, we select 200 articles in the sampled
journals and time period, which have made the highest academic impact, measured by age-adjusted citations. Third, we reduce the list of
200 to 100, by weighing impact with the likelihood to which an article
represents marketing science. Fourth, we reduce the list of 100 highimpact marketing articles to the 20 articles that marketing intermediaries rated as most impactful on marketing practice. Next, we explain
this procedure in greater detail.
For the first filter, our aim was to achieve a good representation of
major marketing journals, which we based on prior scientometric
work in marketing (Stremersch et al., 2007). We excluded the Journal
of Consumer Research (JCR) as it is not an outlet that typically publishes
marketing science articles. We added Management Science, because it
consistently features in the Financial Times Top 45, for example, and
has a marketing section. This step thus led us to the following selection
of journals: International Journal of Research in Marketing (IJRM), Journal
of Marketing (JM), Journal of Marketing Research (JMR), Management Science (MGS) and Marketing Science (MKS).
Next, we assessed how long the journals were covered in the Social
Science Citation Index. Young journals need time to mature and become
academically and practically impactful, which may make them less suited for our goals, even if they are a top journal. IJRM is the youngest top
journal in the set and was not included in the Social Science Citation
129
index until 1997. Therefore, in 2006, it was very unlikely for IJRM articles
from the period 1997–2003 to have amassed enough citations to be
among the top 200 age-adjusted cited articles and be included in our
further analytical steps. Later analyses on an expanded sample that included IJRM showed this assessment to be accurate. The most highly
ranked IJRM article was Geyskens, Steenkamp, and Kumar (1998) at
rank 255. We selected the period 1982–2003 as observation window.
We chose the start year of our data to coincide with the launch of Marketing Science in 1982. We chose the end year of 2003 to allow articles at
least 2 full years for their impact to materialize (this is common in citation studies, see Stremersch et al., 2007).
Second, we rank-ordered the resultant 5556 articles on their
academic importance, as measured by age-adjusted citations (see
Stremersch et al., 2007 for a similar procedure). As citations show a
time trend, we first de-trended our measure by regressing the number
of citations of an article i (CITEi) on the number of quarters (Qi) that
have passed between publication and the quarter in which we gathered
the citations and its square (Qi2), including a constant (across all articles). We conducted this study in the 3rd quarter of 2006 and, thus,
we obtained the stock of citations that were in ISI databases, in that
quarter.
As CITEi shows over-dispersion, we specified a negative binomial
count model and optimized with quadratic hill climbing. As expected,
our results indicated an inverted U-shaped time trend (the estimated
coefficients for Qi and Qi2 were 0.07 and −4.76E−04 respectively,
both significant at p b 0.001; R2 = 0.035). We obtain standardized residuals from the model, denoted by CITERESIDi, which can be regarded
as a time-corrected citation measure of academic impact. We retained
the top 200 articles ranked on this academic impact measure.
Third, we examined the MGS articles in this top 200 and excluded
the 71 articles that did not consider a marketing subject, because they
could not possibly be “marketing” science. Next, we calculated the extent to which each of the 129 remaining marketing articles is a marketing “science” article. We found the task of defining marketing “science”
difficult. After many discussions with experts, we came to the following
working definition: “Marketing science is the development and use of
quantifiable concepts and quantitative tools to understand marketplace
behavior and the effect of marketing activity upon it.”2
To determine whether a specific article satisfied this definition, we
asked five pairs of two marketing science experts – members of the
Marketing Science and IJRM Editorial Boards, and leading marketing intermediaries – to individually code 100 articles published in the four
journals in a hold out sample published in 2004–2005,3 as marketing
science articles, or not. The proportion of agreement between the raters
was 0.77, which translated into a proportional reduction of loss (PRL)
inter-rater reliability measure of 0.72 (Rust & Cooil, 1994), satisfactory
for the exploratory nature of our research. We created a variable that
took the value 1 if both raters agreed it was a marketing science article,
0 otherwise.
Next, we inventoried the number of equations to measure an
article's mathematical sophistication (also used by Stremersch &
Verhoef, 2005), the methodologies an article uses, going from qualitative techniques to time series and analytical models, and the number
of referenced articles in econometrics, statistics and mathematics. Stepwise logistic regression revealed two significant predictors: the number
of equations and whether the methodology used factor and/or cluster
analysis or not. The more equations an article contains, the higher the
likelihood of it being considered a marketing science article. Articles
that use factor or cluster analyses are generally less perceived as a
2
This definition aligns closely to the definition of marketing analytics of Germann
et al.'s (2013).
3
We selected 25 articles from each of the four journals. Article selection was random for
JM, JMR, and Marketing Science. For Management Science, we inventoried 25 articles from
2004 to 2005 that were marketing-related. The list of 100 articles is provided in Web Appendix 1.1.
130
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
Table 1
The 100 academically most impactful papers in marketing science (ordered by practice impact, and then by academic impact; complete bibliography is available in the Web Appendix 3).
Rank
Authors, publication year
Cites total
CITERESID
PROBMKS
Academic impact: MKSIMPACT
Practice impact: INTIMPACT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
Green and Srinivasan (1990)
Louviere and Woodworth (1983)
Aaker and Keller (1990)
Cattin and Wittink (1982)
Guadagni and Little (1983)
Mahajan et al. (1990)
Rust et al. (1995)
Hauser and Shugan (1983)
Fornell, Johnson, Anderson, Cha, and Bryant (1996)
Griffin and Hauser (1993)
Day (1994)
Punj and Stewart (1983)
Fornell (1992)
Vanheerde et al. (2003)
Hunt and Morgan (1995)
Anderson, Fornell, and Lehmann (1994)
Simonson and Tversky (1992)
Boulding et al. (1993)
Parasuraman, Zeithaml, and Berry (1985)
Keller (1993)
Yu & Cooper (1983)
Urban, Carter, Gaskin & Mucha (1986)
Carpenter & Nakamoto (1989)
Zeithaml, Berry & Parasuraman (1996)
Dickson & Sawyer (1990)
Zeithaml, Parasuraman & Berry (1985)
Joreskog & Sorbom (1982)
Day & Wensley (1988)
Thaler (1985)
Kamakura & Russell (1989)
Zeithaml (1988)
Bolton (1998)
Tversky & Simonson (1993)
Churchill & Surprenant (1982)
Fornell & Bookstein (1982)
Mittal & Kamakura (2001)
Srivastava, Shervani & Fahey (1998)
Churchill, Ford, Hartley & Walker (1985)
Gupta (1988)
Teas (1993)
Anderson & Sullivan (1993)
Gutman (1982)
Jaworski and Kohli (1993)
Slater & Narver (1994)
Mcguire, TW & Staelin (1983)
Parasuraman, Zeithaml & Berry (1994)
Mackenzie & Lutz (1989)
Robinson & Fornell (1985)
Bitner, Booms & Tetreault (1990)
Bolton & Lemon (1999)
Henard & Szymanski (2001)
Bitner (1990)
Perreault & Leigh (1989)
Ruekert & Walker (1987)
Mackenzie, Lutz & Belch (1986)
Alba, Lynch, Weitz, Janiszewski, Lutz, Sawyer & Wood (1997)
Webster (1992)
Haubl & Trifts (2000)
Bearden, Sharma & Teel (1982)
Han, Kim & Srivastava (1998)
Dwyer & Schurr & Oh (1987)
Lynch & Ariely (2000)
Pollay (1986)
Bitner (1992)
Cronin & Taylor (1992)
Oliver (1999)
Garbarino & Johnson (1999)
Crosby, Evans & Cowles (1990)
Cronin & Taylor (1994)
Rindfleisch & Heide (1997)
Kalwani & Narayandas (1995)
Ganesan (1994)
Doney & Cannon (1997)
Morgan and Hunt (1994)
Bakos (1997)
292
195
170
152
431
268
146
152
136
166
321
263
159
25
149
217
213
250
765
250
192
162
157
191
160
225
133
233
532
242
390
85
121
262
210
56
92
161
206
135
200
157
411
238
140
178
208
174
183
62
42
294
168
185
167
182
273
60
125
97
632
83
157
281
399
81
114
259
153
92
117
311
218
690
156
4.34
2.99
2.21
2.48
7.44
3.82
2.88
2.21
3.29
2.62
6.63
4.40
2.55
2.01
2.95
4.18
3.37
4.23
12.08
4.23
3.13
2.07
1.97
2.65
2.09
5.41
2.04
3.14
8.31
3.37
5.64
2.31
1.90
4.58
3.20
2.62
2.55
2.14
2.72
2.18
3.34
2.65
7.64
5.22
2.08
3.14
2.82
2.33
2.76
2.00
2.08
4.28
2.13
2.39
2.11
5.15
4.57
2.26
1.89
3.12
9.53
3.35
1.99
4.03
6.82
2.64
4.15
3.74
2.62
2.43
2.12
6.07
6.05
14.52
4.52
0.47
0.78
0.45
0.45
0.80
0.87
0.77
0.93
0.45
0.47
0.45
0.47
0.80
0.72
0.45
0.65
0.53
0.42
0.45
0.45
0.47
0.57
0.55
0.45
0.45
0.14
0.84
0.45
0.53
0.84
0.45
0.69
0.78
0.13
0.49
0.59
0.45
0.45
0.74
0.65
0.65
0.45
0.53
0.45
0.95
0.57
0.47
0.55
0.45
0.61
0.49
0.45
0.65
0.45
0.45
0.45
0.45
0.49
0.45
0.16
0.45
0.47
0.45
0.45
0.21
0.45
0.13
0.13
0.45
0.45
0.45
0.13
0.13
0.45
0.89
2.04
2.35
1.00
1.12
5.94
3.31
2.22
2.04
1.48
1.23
2.98
2.07
2.04
1.45
1.33
2.73
1.80
1.79
5.44
1.90
1.47
1.19
1.09
1.19
0.94
0.76
1.71
1.41
4.43
2.81
2.54
1.59
1.49
0.60
1.57
1.56
1.15
0.96
2.01
1.42
2.18
1.19
4.07
2.35
1.98
1.80
1.33
1.29
1.24
1.23
1.02
1.93
1.39
1.08
0.95
2.32
2.06
1.11
0.85
0.51
4.29
1.58
0.89
1.82
1.45
1.19
0.54
0.49
1.18
1.09
0.96
0.80
0.79
6.54
4.04
4.22
3.56
3.50
3.25
3.22
3.11
3.00
3.00
3.00
2.89
2.67
2.67
2.67
2.63
2.63
2.44
2.38
2.38
2.25
2.25
2.25
2.25
2.22
2.13
2.13
2.13
2.11
2.11
2.00
2.00
2.00
2.00
2.00
2.00
1.89
1.89
1.88
1.88
1.75
1.75
1.67
1.67
1.63
1.63
1.63
1.63
1.63
1.63
1.63
1.63
1.63
1.50
1.50
1.50
1.50
1.38
1.38
1.38
1.38
1.33
1.25
1.25
1.25
1.22
1.22
1.22
1.22
1.22
1.13
1.13
1.13
1.13
1.13
1.00
1.00
131
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
Table 1 (continued)
Rank
Authors, publication year
Cites total
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
Narver & Slater (1990)
Anderson, Hakansson & Johanson (1994)
Deshpande, Farley & Webster (1993)
Kohli & Jaworski (1990)
Jeuland & Shugan (1983)
Gorn (1982)
Anderson & Coughlan (1987)
Phillips, Chang & Buzzell (1983)
Gaski (1984)
Novak, Hoffman & Yung (2000)
Lovelock (1983)
Solomon, Surprenant, Czepiel & Gutman (1985)
Anderson & Narus (1990)
Zirger & Maidique (1990)
Deshpande & Zaltman (1982)
Sinkula (1994)
Anderson & Weitz (1989)
Hirschman & Holbrook (1982)
Huber & McCann (1982)
Tse & Wilton (1988)
Anderson & Weitz (1992)
Hoffman & Novak (1996)
Slater & Narver (1995)
Ferrell & Gresham (1985)
Gerbing and Anderson (1988)
440
139
119
442
185
170
213
166
164
92
191
158
442
150
248
152
185
234
133
175
265
287
197
290
453
CITERESID
6.83
2.55
1.92
6.73
2.87
2.99
2.89
2.57
2.30
3.77
2.98
2.12
6.67
1.93
4.19
2.60
2.39
4.12
2.10
2.21
4.18
7.31
3.54
4.27
6.64
PROBMKS
Academic impact: MKSIMPACT
Practice impact: INTIMPACT
0.45
0.45
0.45
0.45
0.96
0.45
0.45
0.45
0.45
0.13
0.45
0.45
0.13
0.45
0.13
0.45
0.45
0.45
0.67
0.47
0.17
0.45
0.45
0.45
0.21
3.08
1.15
0.87
3.03
2.76
1.35
1.30
1.16
1.03
0.50
1.34
0.96
0.88
0.87
0.55
1.17
1.08
1.86
1.41
1.04
0.73
3.29
1.59
1.92
1.41
1.00
1.00
1.00
0.89
0.89
0.89
0.89
0.89
0.89
0.89
0.78
0.75
0.75
0.67
0.67
0.63
0.63
0.56
0.56
0.56
0.56
0.50
0.44
0.33
0.33
Notes:
1. In the case of ties in practice impact, we reverted to academic impact to determine which articles got into the top 20.
2. CITERESID is age-adjusted citation impact, measured by the residual from the negative binomial model with citations as the dependent variable and quarters since publication and its
square as the independent variables.
3. PROBMKS is the probability that the article is a marketing science article, see Section 2.2.3 for details.
4. MKSIMPACT = CITERESID × PROBMKS.
5. INTIMPACT is awareness-adjusted impact, which is the average impact across all respondents assuming that the impact is 0 for articles of which the respondent is not aware.
marketing science article, probably because these are exploratory techniques. The fit of this model is reasonable; the hit rate was 75%, which
compares favorably to chance (50.5%). We applied these model coefficients, calibrated on the out-of-sample 2004–2005 articles, to the 129
marketing articles identified earlier and retrieve an estimated probability that an article is marketing science, denoted as PROBMKSi.
We then weighted the age-adjusted citation impact (CITERESIDi) by
the likelihood of the article being marketing science (PROBMKSi) to obtain our final measure of marketing science academic impact for each
article (MKSIMPACTi). We rank-ordered the 129 articles on this latter
measure and selected the top 100 articles. We provide the full list of
100 articles and all metrics in Table 1. Complete references are included
in Web appendix 1.3. Table 1 shows that our methodology leads to credible results, with substantial face validity. For instance, Guadagni and
Little (1983) and Mahajan, Muller, and Bass (1990) are more likely to
be regarded as being marketing science articles than Morgan and
Hunt (1994) and Jaworski and Kohli (1993).
Because one of our goals is to survey academics and intermediaries
on the impact on practice of individual marketing science articles, we
needed to reduce the list of 100 articles to 20, to make the task manageable for our respondents. In the final reduction from 100 articles to 20,
we wanted to account for practice impact and asked 34 marketing intermediaries to rate the practical impact of four randomized blocks of 25
articles. The respondents were from a larger pool of 54 intermediaries
(63% response rate) who worked in marketing science intermediary
roles in firms such as AC Nielsen, Mercer, GfK, and McKinsey. These intermediaries were specifically selected because (i) they had previously
published papers in or were on the Editorial Board of Marketing Science,
and/or (ii) were past or current members of the Practice Committee of
INFORMS ISMS. We asked these 34 respondents if they were aware of
each article and, if so, the impact on practice that they believed that it
had had, using a 5-point verbally anchored scale (1 = no influence;
5 = extremely influential). We gave a score of 0 to those articles of
which the respondents were not aware, assuming that there could not
be a direct impact if the respondent was not even aware of the article
when prompted. We then calculated an average impact across all respondents for each article, calling it an awareness-adjusted practice
impact score (denoted as INTIMPACT). Rank-ordering all 100 articles
on INTIMPACT allowed us to select the 20 highest ranked articles,
which we used in our large-scale survey of academics and intermediaries. We found no significant differences in the average awarenessadjusted practice impact score across the four groups of intermediaries.
We acknowledge that starting with a citation screen (as well as a
screen in terms of journal outlet) may preclude consideration of some
papers with high impact on practice, but low impact on scholarship.
Our intention though was not to measure which were the marketing articles with the highest practice impact per se. Rather our intention was
to identify marketing papers with high dual impact, including both academic and practice impact.
2.3. The participants: Managers, intermediaries and academics
We use samples from each participant population (managers, intermediaries, and academics) to inventory the impact of marketing science
on marketing practice, along the marketing science value chain, described in Fig. 1. We do not expect marketing managers to be aware of
many, if any, academic articles, even where those articles have been incorporated into the marketing science tools that they routinely use.
Thus, marketing managers can inform us only on knowledge conversion
(tools) and knowledge application (decisions). However, we also calibrate managers' perceived importance of different areas of marketing
decision making.
2.3.1. Sample of managers
Our sample of senior marketing managers consisted of Marketing
Science Institute and Institute for the Study of Business Markets
132
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
(ISBM) members and company contacts. Both institutes graciously
emailed a request from us to their members. In total, we solicited survey
participation from 477 managers, of whom 94 (20%)4 provided usable
responses. While this group comes from a well-defined population, it almost certainly has a bias towards greater sophistication. This sophistication is likely to introduce an upward bias in the perceived impact of
tools and their influence in different areas (the absolute impact of marketing science). However, there is no reason to believe that this bias will
be very different for different tools and decision areas, meaning any bias
in the relative effects will likely be considerably less.
2.3.2. Sample of intermediaries
We used four sources to create the sample of intermediaries. First,
we examined all articles published by practitioner analysts in Marketing
Science and included those authors in our sample. Second, we examined
the editorial boards of our target journals and included any intermediaries on these boards. Third, the Marketing Science Institute contacted
the marketing intermediaries among their members on our behalf. Finally, we surveyed marketing intermediaries attending the 2007 ISMS
Marketing Science Practice Conference, held at the Wharton School. In
total, we solicited participation from 93 intermediaries, of whom 34
(37%) participated in the main survey. 21 of these respondents worked
at marketing and/or management consulting firms such as McKinsey,
AC Nielsen, and Millward Brown, while 13 respondents worked in
firms such as General Motors, IBM, and Campbell Soup.
and practitioners; (7) the stage of their career in which they wrote the
article; and (8) the reasons that may have made the article impactful.
We summarize our data collection approach in Fig. 2.
3. Results
Moving up the value chain illustrated in Fig. 1, we present the results
of our research in four stages: the relative impact of marketing science
on different decision making areas (Section 3.1), the impact that different marketing science tools and approaches have had on marketing
practice (Section 3.2), the impact of the twenty articles on marketing
decisions and tools (Section 3.3), and the antecedents of “dual” (academic and practice) impact from a survey of the authors of 20 top articles (Section 3.4). In Section 3.5, we identify trends since 2004 in the
application and use of marketing science.
3.1. Impact of marketing science on marketing decisions
To inventory the impact of marketing science on marketing decision
areas, we first present the self-stated importance of each decision area
by manager respondents. Next, we present the extent to which our respondents felt that marketing science had impacted each marketing decision area. We end with graphically presenting the alignment between
impact of marketing science on and the importance of the decision
areas.
2.3.3. Sample of academics
We defined the sampling frame of marketing academics to be academic marketing science members of the editorial boards of the target
journals. We excluded the authors of the current paper from this sampling frame. To identify the “marketing science” members of those editorial boards, we used a peer review process, in which we asked ten
marketing science experts to indicate whether they would classify
members of these editorial boards (223 in total) as marketing scientists
or not.5 Of the 223 editorial board members in total, 126 were classified
as marketing scientists, of whom 84 (67%) ultimately responded to our
survey.
3.1.1. Importance of decision areas
In Table 2, we report the self-stated importance of each of the decision areas to the company, classified by type of firm (B2B, B2C, both
B2B and B2C, and total). Overall, pricing management is rated the
most important (aggregated across types of firms), while promotion
management is rated the least important. However, there are notable
differences across B2B and B2C firms. Managers of B2B firms consider
pricing management to be the most important decision area, followed
by customer/market selection and product portfolio management.
Managers of B2C firms consider brand management and new product
management to be the most important decision areas.
2.4. The instruments: Surveys among participants
3.1.2. Impact of marketing science on decision areas
In Table 3, we present the perceived impact of marketing science on
specific marketing decision areas, as perceived by academics (A), intermediaries (I), and managers (M). According to managers, marketing science has had the biggest impact on brand management decisions and
pricing decisions (mean = 3.77 for both), and new product/service
management and customer/market selection (mean = 3.66 for both).
Academics feel that marketing science has made the biggest impact on
brand management, new product/service management and promotion
management. Intermediaries sense that marketing science has made
the biggest impact on pricing management, promotion management,
and new product/service management.
Interestingly, academics believe that marketing science had the biggest impact on promotion management among all decision areas
(mean = 3.76), while managers consider that it had the smallest influence among all areas (mean = 3.14). For other areas, such as new product/service management, both seem to agree much more as to the
relatively large extent to which marketing science has impacted such
decisions (means = 3.70 and 3.66 respectively for academics and managers). Overall, Table 3 shows that while there is consensus between the
academic and intermediary groups (ρAI = 0.62) and some moderate
level of consensus between the intermediary and manager groups
(ρIM = 0.39), there is much disagreement between academics and
managers (ρAM = 0.17), pointing to the bridging role of marketing
intermediaries.
In Table 3, we also present how managers perceived the impact of
marketing science on different decision areas, split by type of firm. As
expected, the results indicate some differences by type of firm. While
Our instruments are as follows (see Web Appendix 1.2 for details).
The survey to managers measured: (1) the overall influence of each of
the 12 tools on marketing practice; (2) the overall influence of marketing science on each of the 12 marketing decision areas; and (3) the importance of the 12 marketing decision areas to their company. The
survey to intermediaries and academics measured: (1) the overall influence of the 20 marketing science articles on marketing practice; (2) the
overall influence of the 12 tools on marketing practice; and (3) the overall influence of marketing science on the 12 marketing decision areas.
We also collected respondent background data for each sample.
Additionally, we surveyed the authors of the top 20 dual impact articles to probe: (1) other scholars who influenced the development and
execution of the article; (2) academic ideas underlying the article, including the important papers on which the article was built; (3) practitioner influence on the development and execution of the article; (4)
the practical ideas underlying the article; (5) whether there was cooperation with practitioners when developing the article; (6) any diffusion efforts the authors undertook to diffuse their work to academics
4
The response rate for the MSI sample was 53% and for the ISBM sample (where the
participant request was less personalized), it was 16%. Note that our email solicitation included a URL, which increases the likelihood of the email being classified by spam filters as
spam and thus not reaching many members of our sample. As a result, the response rate
we report is a lower bound. This comment applies to all three samples (managers, intermediaries, and academics).
5
The inter-rater reliability using a separate sub-sample was 0.90, sufficiently high to indicate that our classification procedure is reliable.
133
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
Stimuli
Article selection
Tool and Decision selection
What is marketing science?
Who is a marketing scientist?
Marketing Science
Editorial Board
Screening of 100 most cited
34 Marketing
marketing science articles to 20 Intermediaries
What are the key areas
of marketing decisions?
MSI
Executive
Committee
What are the key tools
and approaches used?
Main Questionnaire
Manager Survey (N = 94)
Intermediary Survey (N = 34)
Academic Survey (N = 84)
Importance of decision areas
Impact of marketing science on
marketing decision areas
Impact of marketing science on
marketing decision areas
Impact of marketing science tools
Impact of marketing science tools
Impact of articles
Impact of articles
Manager Survey (N = 4)
Intermediary Survey (N = 5)
Academic Survey (N = 4)
Impact of 12 marketing science
tools on 12 marketing decision
areas
Impact of 12 marketing science tools
on 12 marketing decision areas
Impact of 12 marketing science
tools on 12 marketing decision
areas
Impact of marketing science on
marketing decision areas
Impact of marketing science tools
Transition Matrices
Impact of 20 Articles on 12
marketing science tools
Impact of 20 Articles on 12
marketing decision areas
Impact of 20 Articles on 12
marketing science tools
Impact of 20 Articles on 12
marketing decision areas
Antecedents of impactful papers
Survey of authors of 20 marketing science articles with high academic and practice impact
• Influence (academic, industry, literature, problem • Industry co-operation
• Effort to diffuse findings
• Author background (experience)
Fig. 2. Overview of the primary data collection approach.
B2B managers perceive the biggest impact on pricing management, B2C
managers perceive the impact to be largest on customer insight management. However, there is moderate consistency (ρB2B, B2C =0.45).
3.1.3. Alignment between importance of decision areas and impact of
marketing science
To examine whether the impact of marketing science on decision
areas is aligned with the importance of the decision area to managers,
we plot the importance against (managerial perceptions of) impact in
Fig. 3. Considering the differences in importance as well as perceived
impact across managers from different types of firms, we present the
B2B and B2C plots separately. (We have not included the plots for
firms that do both since these largely lie between the two).
Table 2
Average importance of decision areas according to managers in different types of firms
(ordered per Table 3).
Decision areas
B2B
(N = 59)
B2C
(N = 10)
B2B & B2C
(N = 25)
Total
(N = 94)
Brand management
Pricing management
New product/service management
Customer/market selection
Product portfolio management
Customer insight management
Service/product quality
management
Channel management
Relationship management
Salesforce management
Advertising management
Promotion management
3.51
4.03
3.78
3.79
3.79
3.16
3.57
4.60
4.30
4.60
4.20
4.20
4.20
3.80
4.04
4.12
3.80
3.84
3.76
3.80
3.52
3.77
4.09
3.87
3.85
3.83
3.45
3.58
3.24
3.62
3.62
2.69
2.68
4.10
3.60
4.30
3.90
4.00
3.72
3.56
3.60
3.24
3.12
3.46
3.60
3.69
2.97
2.95
Scale: 1: Of no importance. 5: Extremely important.
Both plots indicate that, by and large, the impact of marketing science is aligned with the perceived importance of the decision area.
The most notable examples of under-performance are sales force management and service/product quality for both groups, relationship management for B2B, and advertising and channel management for B2C.
Table 3
Average impact of marketing science on decision areas (ordered by managers'
perceptions; numbers represent average impact given awareness).
Managers
Decision areas
Academics Intermediaries All
B2B
Brand management
Pricing management
New product/service
management
Customer/market selectionb
Product portfolio
managementb
Customer insight
managementb
Service/product quality
management
Channel managementb,c
Relationship management
Sales force managementa,c
Advertising management
Promotion managementb,c
Average perceived impact
3.75
3.53
3.70
3.56
3.85
3.68
3.77 3.80 4.10 3.54
3.77 3.82 3.80 3.63
3.66 3.68 3.90 3.50
3.24
2.94
3.58
3.26
3.66 3.70 3.60 3.58
3.55 3.55 3.60 3.54
2.95
3.31
3.42 3.29 4.20 3.38
3.37
3.13
3.41 3.36 3.30 3.58
2.72
3.29
3.43
3.22
3.76
3.32
2.71
3.25
2.80
3.47
3.71
3.36
3.40
3.37
3.26
3.15
3.14
3.46
3.40
3.40
3.29
2.93
3.04
3.44
B2C
3.44
3.56
3.44
3.40
3.60
3.66
B2B &
B2C
3.38
3.21
3.13
3.54
3.17
3.43
Scale: 1: No influence at all 5: Extremely influential.
a
Academics-intermediaries significantly different at p b 0.05.
b
Academics-managers significantly different at p b 0.05.
c
Intermediaries-managers significantly different at p b 0.05. Significance assessed with
the Welch–Satterthwaite t-test.
Degree to which Decision Area is Influenced By
Marketing Science
Degree to which Decision Area is Influenced By
Marketing Science
134
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
Table 4
Average impact of marketing science tools on marketing practice, according to academics,
intermediaries, and managers (ordered by intermediaries' perceptions, numbers represent
average impact given awareness).
B2B Firms only (N=58)
4.10
3.90
Brand mgmt
Customer/Mkt selection
New product mgmt
3.50
2.90
2.70
2.50
2.50
Tools/approaches
Academics Intermediaries All
B2B
B2C
Segmentation toolsc
Survey-based choice
modelsa,b,c
Aggregate marketing mix
modelsa,b,c
Pre-test market modelsb,c
Marketing metrics
New product modelsb
Customer life time value
modelsb,c
Panel-based choice modelsb,c
Perceptual mappinga,b
Customer satisfaction
modela
Sales force allocation
modelsb
Game theory models
Average Perceived Impact
4.29
3.71
4.44
4.15
4.02 4.00 4.30 3.96
3.25 3.06 3.50 3.58
3.36
4.06
2.99 2.88 3.40 3.00
3.93
3.54
3.78
3.84
3.94
3.77
3.74
3.63
3.38
3.73
3.37
3.07
3.76
3.99
3.83
3.58
3.53
3.39
2.82 2.73 3.11 2.87
3.19 3.14 3.80 3.04
3.59 3.66 3.33 3.52
3.62
3.23
3.07 3.02 3.25 3.13
2.18
3.65
2.12
3.63
2.41 2.51 2.44 2.19
3.24 3.18 3.46 3.27
Product portfolio mgmt
Channel mgmt
Relationship mgmt
Service/prod quality mgmt
3.30
3.10
Managers
Pricing mgmt
3.70
Customer insights mgmt
Salesforce mgmt
Promotion mgmt
Advtg mgmt
Degree of Influence = 1.49+ 0.56 Importance of Decision
3.00
3.50
4.50
4.00
Importance of Decision Area
B2C Firms only (N=10)
Brand mgmt
Customer insights mgmt
4.10
3.90
New product mgmt
Pricing mgmt
3.70
Promotion mgmt
Product portfolio mgmt
Customer/Mkt selection
Relationship mgmt
3.50
Channel mgmt
Salesforce mgmt
Advtg mgmt
3.30
Service/prod quality mgmt
2.98
3.72
3.27
3.18
4.30
3.67
3.67
2.70
B2B &
B2C
3.71
3.76
3.48
3.00
Scale: 1: No influence at all 5: Extremely influential.
a
Academics-intermediaries significantly different at p b 0.05.
b
Academics-managers significantly different at p b 0.05.
c
Intermediaries-managers significantly different at p b 0.05. Significance assessed with
the Welch–Satterthwaite t-test.
3.10
encouraging. But again, we note that our sample is likely biased toward
high levels of sophistication.
2.90
2.70
2.50
2.50
Degree of Influence = 1.16 + 0.60 Importance of Decision
3.00
3.50
4.00
4.50
Importance of Decision Area
Fig. 3. Impact of marketing science versus importance of decision area (both according to
managers).
3.2. Impact of marketing tools on marketing practice
Having gauged the decisions that are important to the firm and the
extent to which marketing science has influenced them, we examine
the tools that provide one route by which that influence is felt. In
Table 4, we present the average impact of marketing science tools on
marketing practice, as perceived by academics, intermediaries and
managers. We also provide a split of manager perceptions, according
to whether they are in a B2B, B2C, or both B2B and B2C firm.
According to managers, the top three marketing science tools and
approaches are: (1) marketing segmentation tools (mean = 4.02), (2)
marketing metrics (mean = 3.73), and (3) customer satisfaction
models (mean = 3.59). While segmentation tools are also the number
1 pick of academics and intermediaries, opinions diverge on the other
ones. Survey-based choice models (number 2 among intermediaries,
mean = 4.15) and perceptual mapping techniques (number 2 among
academics, mean = 3.99) had less of an impact on marketing practice,
according to the marketing managers (means = 3.25 and 3.19 respectively for survey-based choice models and perceptual mapping techniques). Other tools that were consistently found to significantly
impact practice are pre-test market models (number 3 or 4 in the
three groups) and new product models (number 5 or 6 in the three
groups). The different samples also consistently agree on the lack of
practical impact of game theory models. The agreement between
groups as to the impact of different tools is a lot stronger than the agreement we found on the impact of marketing science on the different
decision areas: ρAI = 0.80, ρIM = 0.70, and ρAM = 0.73. Managers' average awareness of marketing science tools was close to 90%, which is
3.3. Impact of articles on marketing tools and directly on marketing practice
We continue to calibrate practice impact up the value chain in Fig. 1
by examining select marketing science articles and the effect that they
have had both on marketing science tools and directly on marketing decision making. We first report results from our precalibration of the top
100 marketing science papers according to academic impact among
marketing intermediaries, after which we report on the results from
the complete survey of the authors of the top 20 marketing science papers with “dual” impact.
In Fig. 4, we plot the academic impact of the top 100 marketing science articles in Table 1 (MKSIMPACT) against the awareness-adjusted
impact on practice as perceived by the 34 marketing intermediaries
from the precalibration (INTIMPACT). Individual points may be identified by reference to Table 1. While there is a significant relationship between academic and practice impact, it is weak (ρ = 0.19). We find it
more insightful to divide the graph into four quadrants, through a median split on both dimensions. Articles in the bottom left quadrant of Fig. 4
have not had a major impact on practice (e.g., Gerbing & Anderson,
1988), and are also below the median for these 100 articles on academic
impact. (Note that all 100 candidates for inclusion fall in the top 5% of
age-adjusted citation in the profession's top four quantitative journals.)
The articles on the bottom right are primarily knowledge drivers — that
is, articles that have had above-median academic impact (relative to the
100 papers in this pool), but have had below-median practice impact
(e.g., Morgan & Hunt, 1994). The articles on the top left quadrant are
practice drivers — articles that have had below-median academic impact among the top 100 pool, but have had above-median practice impact (e.g., Aaker & Keller, 1990). The top right quadrant consists of
articles that have had dual impact, exceptional academic as well as practice impact (e.g., Guadagni & Little, 1983). The selection from top 100 on
academic impact to top 20 on dual impact represent articles from both
the top-left and the top-right quadrants in Fig. 4 (see Web Appendix
2.2 for articles by quadrant).
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
135
3.4. Antecedents of practice impact among dual impact marketing science
articles
Awareness-adjusted Practice Impact Score
Aaker & Keller (1990)
4.50
Guadagni & Little (1983)
4.00
3.50
3.00
2.50
2.00
1.50
Morgan & Hunt (1994)
1.00
0.50
0.00
0.00
Gerbing & Anderson (1988)
2.00
4.00
6.00
8.00
Academic Impact Score
Notes:
1. Awareness adjusted practice impact score is INTIMPACT from Table 1. It is the average impact
of the article assuming that the impact=0 for articles of which respondents are not aware.
2. Academic Impact Score is MKSIMPACT from Table 1, which is the age-adjusted citation score,
further adjusted by the probability of the paper being marketing science.
Fig. 4. Contrast of academic and practice impact of 100 selected articles. Notes: Awareness
adjusted practice impact score is INTIMPACT from Table 1. It is the average impact of the
article assuming that the impact = 0 for articles of which respondents are not aware. Academic Impact Score is MKSIMPACT from Table 1, which is the age-adjusted citation score,
further adjusted by the probability of the paper being marketing science.
In Table 5, we present the results of asking our sample of intermediaries (N = 34) and academics (N = 84) to evaluate the practice impact
of each of the 20 dual-impact articles we identified earlier. In this table,
we present the impact score given awareness for each article6 as well as
awareness-adjusted practice impact. Although we need to be careful in
drawing very strong conclusions (given quite large standard deviations), Guadagni and Little (1983) and Green and Srinivasan (1990)
show the highest impact on practice, both as perceived by academics
(mean = 4.28 and 4.17 respectively) and intermediaries (mean =
4.17 and 3.97 respectively). Overall, the ranking across the two samples
is quite consistent (ρAI = 0.63). Notable exceptions include Louviere
and Woodworth (1983), Vanheerde, Gupta, and Wittink (2003), and
Simonson and Tversky (1992), all of which intermediaries accredit a
significantly higher impact on practice than academics, while only
Fornell (1992) shows the opposite. Finally, there is a correlation of
0.65 between the practice impact of these 20 articles gauged from the
pre-calibration sample of intermediaries and the calibration sample of
intermediaries. (Respondents in the precalibration and calibration samples responded to different tasks, precluding any aggregation of data
across samples).
Table 3 describes the impact that marketing science has had on different marketing decisions, and Tables 4 and 5 show the influence of different tools and articles, respectively. We also solicited the more
detailed transition matrices of individual articles' impact on individual
tools and decisions, and individual tools on individual decisions, from
a sub-sample of our respondents. We include and discuss these transition matrices in the marketing science value chain in Web Appendix
2.1. Additionally, many respondents provided open ended comments
(included as Web Appendix 2.2). Perhaps the most interesting aspect
of those is the variety of “mental maps” with which managers, intermediaries and academics think about marketing science applications.
6
As before, although we also report conditional impact (impact given awareness), our
awareness adjusted impact assumes that for an article to have impact a respondent must
have awareness of it when prompted.
As described earlier in our methodology section, we surveyed the
authors of the twenty dual-impact articles, shown in Table 5, to learn
from their experiences that go beyond the obvious, or possibly deviate
from some norms in our field. Participation in our survey of these author
teams was 100% (by article). 17 out of the 20 papers had multiple authors. Of those 17, multiple authors in 9 cases responded to our survey.
Unsurprisingly, many expected themes emerged from these responses;
themes that have been previously identified in the academic and practitioner literature. They include advice from authors to look for gaps in
the literature, to ensure a strong grounding in prior theory, to find interesting, unsolved problems that are important to managers, and to fuel
the diffusion process, not relying on good ideas to automatically be
adopted. Below we focus on the three most interesting new themes
that emerged. In addition, Guadagni and Little (2008) share their recollection in a Marketing Science commentary, which they based on our
survey to them.
3.4.1. Symbiosis with consulting
Many of the authors referred to the symbiosis of their research with
consulting as a fertile ground for dual impact papers. Rick Staelin
describing Boulding, Kalra, Staelin, and Zeithaml (1993) stated “This
paper started with a “consulting” project for the School [Fuqua School of
Business, Duke University] trying to improve the service quality of our
teaching/delivery system.” Jordan Louviere speaking of Louviere and
Woodworth (1983) said “[The problem] came from a consulting project in Australia. I was asked by the Bureau of Transport Economics to
help them forecast demand for Qantas flights on transpacific routes.”
Many of the authors also (co-)founded professional services companies to commercialize their work. For example, Roland Rust mentioned
forming a company to commercialize the approach of Rust, Zahorik,
and Keiningham (1995). John Little attributes his logit model's practical
success largely to the commercialized products based on it. Louviere
worked with DRC to commercialize the method he had developed.
MDS started selling Hauser and Shugan's (1983) Defender model. Hauser
joined Bob Klein in founding Applied Marketing Science, Inc. to commercialize the “voice of the customer” methodology (Griffin & Hauser, 1993).
3.4.2. Going against the grain at the right time
A common topic in many responses was that they went against the
grain at the right point in time. Times were either ripe for the radical innovation the authors introduced or the authors rode on a new technology wave that came to transform industry. About the former, Roland Rust
nicely phrases it as follows: “We went against the grain, which meant
that acceptance of our ideas ensured minds were changed.” Peter
Guadagni and John Little attribute part of the success of Guadagni and
Little (1983) more to the latter, an impeccable sense of timing: “Much
of the impact was due to its early use of data from UPC scanners.”
This does not mean that dual impact author teams were not also
firmly grounded in basic theory, despite going against the grain. For example, Peter Guadagni and John Little say: “Consumers make choices to
maximize utility. This came from basic economic theory.” In the same
vein, John Hauser on Hauser and Shugan (1983) mentions: “There
was the Brandaid model by John Little in which he used a multiplicative
form for the effects of advertising and distribution. Coupled with
Lancaster's model, this gave us an empirically-relevant, but analytically
tractable model with which to study the problem.” Indeed, it is of interest that the 20 top papers by practice impact in Table 1 contained an average of 12 equations and 54 references (compared to 5 and 37
respectively for articles ranked 21–100, p b 0.05).
3.4.3. Working with experience
A long track record of some of the authors and influencers seems to
be an essential component of dual impact teams. All author teams have
136
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
Table 5
Average impact of marketing science articles on marketing practice (ranked by intermediaries' perceptions of impact).
Intermediaries (I) (N = 34).
Academics (A) (N = 84).
Article
Awareness Impact
Std.
Rank
AwarenessAwareness Impact
Std.
Rank
AwarenessDifference test
(%)
(Avg|Aware) Error (impact) adjusted impact (%)
(Avg|Aware) Error (impact) adjusted impact in A–I impact
Guadagni and Little (1983)
Green and Srinivasan (1990)
Louviere and Woodworth (1983)
Griffin and Hauser (1993)
Keller (1993)
Cattin and Wittink (1982)
Parasuraman et al. (1985)
Mahajan et al. (1990)
Fornell et al. (1996)
Aaker and Keller (1990)
Vanheerde et al. (2003)
Hauser and Shugan (1983)
Simonson and Tversky (1992)
Rust et al. (1995)
Anderson et al. (1994)
Boulding et al. (1993)
Punj and Stewart (1983)
Day (1994)
Fornell (1992)
Hunt and Morgan (1995)
Average across articles
85
85
76
74
85
85
65
91
76
79
74
74
71
71
59
68
65
65
62
47
73
4.17
3.97
3.92
3.64
3.48
3.41
3.41
3.35
3.27
2.96
2.96
2.92
2.88
2.83
2.75
2.74
2.73
2.68
2.48
2.44
3.15
0.19
0.18
0.21
0.22
0.20
0.20
0.25
0.20
0.20
0.20
0.19
0.24
0.22
0.19
0.22
0.16
0.24
0.27
0.21
0.27
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
3.56
3.38
3.00
2.68
2.97
2.91
2.21
3.06
2.5
2.35
2.18
2.15
2.03
2.00
1.62
1.85
1.76
1.74
1.53
1.15
98
96
81
94
94
92
94
98
90
92
89
93
87
92
92
82
79
86
90
73
90
4.28
4.17
2.76
3.32
3.78
3.23
3.87
3.13
3.63
3.30
2.43
3.03
2.27
3.12
3.10
2.74
2.73
3.19
3.29
2.02
3.17
0.11
0.10
0.14
0.12
0.11
0.13
0.11
0.13
0.12
0.12
0.11
0.12
0.13
0.12
0.11
0.14
0.14
0.12
0.12
0.13
1
2
15
6
4
9
3
11
5
7
18
14
19
12
13
16
17
10
8
20
4.18
4.02
2.24
3.12
3.56
2.96
3.64
3.06
3.29
3.02
2.17
2.81
1.98
2.86
2.85
2.25
2.14
2.74
2.98
1.46
0.50
1.02
−4.60a
−1.32
1.35
−0.75
1.69
−0.93
1.59
1.43
−2.46b
0.40
−2.38b
1.28
1.46
0.00
0.00
1.71
3.30a
−1.39
Notes:
1. Scale: 1: No influence at all to 5: Extremely influential.
2. ap b 0.01, bp b 0.05, using the Welch–Satterthwaite t-test to test for differences in impact given awareness across academic and intermediary samples.
3. Awareness-adjusted impact is equal to awareness proportion multiplied by impact given awareness. Awareness-adjusted impact assumes that the impact of an article is 0 if the respondent is not aware of the article. Correlation between the two measures is 0.94 for intermediaries and 0.99 for academics.
at least one scholar with an academic career of over 15 years before coauthoring the paper (with the exception of Keller, 1993 article). The
most senior author in 14 of the top 20 papers by practice impact in
Table 1 held a named chair, in contrast to 23 out of the remaining 80
high academic impact articles (p b 0.01). It appears that significant academic experience is close to a prerequisite to writing an article that has
large dual impact. In addition, industry experience may help. Authors
who responded to our survey also had an average 6.75 years of experience in industry.
Authors frequently mentioned close liaison with industry. Eight out
of 20 teams worked with practitioners on developing at least part of
their ideas. Many other sources are mentioned on the practitioner
side, both at intermediaries and marketing companies. Top sources are
the Marketing Science Institute (mentioned by 5 author teams out of
20) as a source of inspiration. As individual practitioners, these authors
mention people such as Bob Klein, Steve Gaskin, Richard M. Johnson,
and Steve Cohen (3 or more mentions).
Academic colleagues with an influence are mainly scholars' coauthors, colleagues from the same department, or scholars on whose
work authors built. Within the marketing profession, Glen Urban and
Al Silk received three or more mentions. John Hauser notes on Hauser
and Shugan (1983): “There were many influences. Chief was the Assessor model by Silk and Urban, which was a pre-test market model to predict the shares of new products. However, for every innovator, there
were many defenders. We wanted to know what was the best defensive
strategy.” Authors also cite inspiration from well-known scholars
outside their own field. Scholars mentioned in that category are Doug
Carroll, Dan McFadden, Albert Hirschman, Herman Wold, and Frank Andrews (2 or more mentions).
3.5. Trends since 2004
It is useful to examine changes in the environment in the past nine
years and to use our findings to consider likely trends in the impact of
marketing science. To do that, we return to the marketing science
value chain and examine separately changes to the decisions managers
make, the tools that they use, and the articles that have driven the development of those tools.
3.5.1. Trends in management decisions
Clearly, a number of environmental changes have affected the way
in which managers need to relate to their marketplaces. These include
a greater availability of addressable data (i.e. big data) and the rise of
digital and mobile communications, both in terms of access to markets
and communications between consumers (such as social networks).
To formalize our examination of these trends, we assessed the changing
content of marketing management textbooks. We examined marketing
management texts rather than cutting edge methodology books because, at this stage of the marketing science value chain, it is the overall
managerial decision making environment we wish to study. An examination of sales lists at amazon.com shows that Kotler/Kotler and Keller's
Marketing Management (in its various guises) dominates this market.
For example, on February 23, 2013 “A Framework for Marketing Management” (5th edition) was 6632 on the best seller list with the closest
non-Kotler competitor coming in at 56,620. Therefore, we looked at the
evolution of this text over time: before the beginning of our study
(1980), four years into our study (1988), at the end of our study
(2003), and most recently (2012). The results are included as Web Appendix 3.1. We note the rising importance of branding, customer management and integrated marketing over this time.
Because textbooks may be backward looking, we also examined
trends in the Marketing Science Institute's Research Priorities which
are, themselves, derived from surveys among academics and their
members, who are all senior managers (Web Appendix 3.2). As
expected, we see more recent topics in this list such as understanding
mobile marketing opportunities, the role of social networks, and the
harnessing of “big data.” The survey of our authors would suggest that
these environmental shifts in possibility and priority bring with them
the opportunity to go against the grain at the right time. An obvious
analogy is John Little's view that his adoption of logit modeling was a direct result of the availability of vast quantities of panel scanner data
which enabled a new, less aggregate way of modeling response to
changes in the marketing mix.
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
3.5.2. Trends in tools available
Clearly, many changes have occurred in the statistical tools available to the industry marketing analyst (and marketing intermediary) since 2004. Kluwer's Series in Quantitative Marketing, edited
by Josh Eliashberg, provides an excellent resource describing advances in many of the tools available. Many of these are driven by the
availability of vast amounts of customer data and with them, the rise
of data mining (see Humby, Hunt, & Phillips, 2008 for an example).
Much of this work is being conducted by information systems groups
rather than marketers. As well as models that account for observed heterogeneity, models that account for unobserved heterogeneity are also
gaining traction. Lilien (2011) speaks to the relative success of models
that may be implemented by automatic algorithm, rather than as a
managerial decision aid, which is an interesting distinction.
To gain a more systematic view of trends in the tools being used in
industry, we examined the programs of the American Marketing
Association's Advanced Research Techniques (ART) Forum from 2002
to 2013. The ART Forum is an annual meeting of academics, intermediaries, and practicing managers which discusses new and emerging
marketing science techniques, as well as conducting tutorials in newlyestablished ones. A summary of these programs is included as Web
Appendix 3.3. We observe that a number of 12 types of tool we identified
continue to be important over the following nine years (including
discrete choice conjoint analysis, customer lifetime value models, and
segmentation techniques). Second, we notice the introduction of new
sets of tools, of which the most important are social media and network
analysis methods from 2010 to 2013, including viral models, recommendation systems, and user generated content. Also of growing importance are text mining methods (2012) and agent-based modeling (2008
and 2012). Finally, many of the tools that we have described have undergone substantial development and enhancement. Primary among those
are the areas of survey based and panel based choice models. The Bayesian treatment of heterogeneity (from 2002 onwards), introduction of
new measurement bases such as MaxDiff, and data augmentation techniques stand out. In a rare study of the prevalence of marketing science
tool usage, Orme (2013) notes fourteen major trends over the past ten
years in the use of Sawtooth software (probably the leader in conjoint/
choice analysis software). Primary among those are the mainstreaming
of Hierarchical Bayes, the decline of ratings based conjoint, the emergence of MaxDiff scaling, and new applications/methods such as menu
based choice, optimization, and adaptive designs.
3.5.3. Trends in marketing science articles
We undertook an examination of the papers published in IJRM, JM,
JMR, MGS, and MKS for the period 2004 to 2010. We included IJRM
given the more recent time period of study and its recognized importance as a top academic journal (Pieters, Baumgartner, Vermunt, &
Bijmolt, 1999). We obtain a CITERESID (see Section 2.2.3) on each of
the journals separately (given that we search for recent trends, they
may pop up in one journal specifically). In this model, we used the number of quarters to December, 2010 as a measure of age of the article.
Next, we have ranked CITERESID per journal and provide the top 10
per journal in Table 6. Note that we validated that the inclusion of
IJRM was appropriate by estimating CITERESID also on the full sample
of all articles jointly and found IJRM had 2 representatives in the top
50 (3 in top 100), marking the gradual maturation of IJRM as the youngest member of top journals in marketing.
A content-analysis of the 50 papers in Table 6 indicates that the
topics of research that have been cited the most are word of mouth
and social networks and relationship marketing/management.
In the absence of a formal survey of the impact of marketing science
articles since 2004, one way to gain some feel for those that have affected the tools that intermediaries (academics and managers) use to address marketing decisions is to look at those articles that have been
mentioned in patents. Because such citations are likely to indicate an article providing the foundation of new tools, we undertook a search using
137
Google Patents for mentions of articles in our target journals in patents issued by the US Patents and Trade Office (USPTO). To allow comparability
with our sample period of 1983 to 2003, we also looked historically at that
period as well. The results are included as Web Appendix 3.4. Marketing
papers from the five target journals received a total of 1317 citations
from patents issued by the USPTO. The first paper to receive a patent citation was published in the Journal of Marketing in 1940. The data indicate a
significantly increasing trend of marketing papers being cited in patents.
Almost half of the citations (625) to historical marketing papers published
in the five target journals have come from patents issued since 2004.
Marketing papers published since 2004 have attracted 39 of those 625
citations. The 39 patent citations were obtained by a total of 27 papers
published since 2004 in IJRM (2 papers), JM (2), JMR (5), MGS (5), and
MKS (13). Papers on the following topics received more than one citation:
pricing and promotions (10), movies (4), online behavior models (4), retail assortment models (3), customer lifetime value models (2), conjoint
(2), forecasting (2), innovation (2), and social networks (2).
One interesting trend is the level of engagement of marketing intermediaries and managers in the knowledge generation process. In 1983
(the beginning of our sample period), approximately half of the participants at the ISMS Marketing Science Conference held at the University
of Southern California came from industry. By 2012, only 37 out of 930
attendees (4%) were from industry. However, general conferences have
been replaced by specialized conferences such the biennial ISMS Practice Conference. Similarly, the Gary Lilien ISMS-MSI Practice Prize has
maintained industry connections with our top journals in terms of authors. The proportion of industry authors of Marketing Science articles
fell from 7% in the period 1983 to 2003 to 5% between 2004 and 2012.
However, 35 of these 68 industry authors from 2004 to 2012 were a
part of Practice Prize Finalist papers, showing the important role special
events can have in stemming the disconnect between academic researchers in marketing and those who have to use their research.
3.5.4. Other marketing science trends
A number of other trends emerged in the development and application of marketing science over the past nine years. First, it has become
more international at all levels of the value chain. In terms of managerial
decision making, globalization has become a major driver of change. In
terms of tools, at the American Marketing Association Advanced Research Techniques Forum, the ratio of North American academic presenters to those from other continents went from 15/1 in 2003 to 22/6
in 2008 and 19/6 in 2013. At the other end of the value chain, the number of authors publishing from outside North America in the top marketing journals is increasing. Looking at the authorship profile of the
top 100 articles (by age-adjusted citation impact) published in the five
top journals from 2004 to 2010, we find that 22% of the authors of papers from 2004 to 2007 were from non-US locations, while this number
increased from 11% in 2004 to 33% in 2010. (See Stremersch & Verhoef,
2005 for evidence of globalization of authorship on the same sample of
journals, but including all articles between 1964 and 2002, not merely
the top cited articles). Also special fora that aim to bridge the gap between academics and practitioners can enable globalization. 11 of the
25 finalists of the Lilien ISMS-MSI Practice Prize Competition since its inception have come from outside North America (seven from Europe,
three from Australia, and one from the Asia Pacific region). Entries
from Europe have won the prize four out of the seven times.
4. Discussion
4.1. Summary
We have calibrated the relative impact of marketing science research on practice, using our marketing science value chain as a central
framework. It is reassuring to see that the impact of marketing science
on marketing decisions has been largely felt in areas that are of the
greatest importance to the firm (see Fig. 3). Moreover, the managers
138
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
Table 6
Top 10 Articles from 2004 to 2012, listed by journal in order of age-adjusted citations.
Articles by journal
Total citations
Age-adjusted impact
Topic
International Journal of Research in Marketing
Reinartz, Haenlein & Henseler (2009)
Peres, Muller & Mahajan (2010)
Dholakia, Bagozzi & Pearo (2004)
Burgess, Steenkamp (2006)
Bagozzi & Dholakia (2006)
Street & Burgess & Louviere (2005)
Goldenberg, Libai & Muller (2010)
Du, Bhattacharya & Sen (2007)
Verhoef, Neslin & Vroomen (2007)
De Bruyn & Lilien (2008)
40
26
169
71
83
83
18
47
51
30
7.31
5.70
5.53
3.91
3.61
3.31
3.19
3.15
3.14
2.80
Journal of Marketing
Vargo & Lusch (2004)
Schau, Muniz & Arnould (2009)
Palmatier, Dant, Grewal & Evans (2006)
Trusov, Bucklin & Pauwels (2009)
Kozinets, de Valck, Wojnicki & Wilner (2010)
Luo & Bhattacharya (2006)
Tuli, Kohli & Bharadwaj (2007)
Brakus, Schmitt & Zarantonello (2009)
Palmatier, Dant & Grewal (2007)
Rust, Lemon & Zeithaml (2004)
1029
82
215
72
40
146
110
55
94
310
10.47
5.37
4.61
4.59
2.82
2.78
2.73
2.70
2.49
2.39
Marketing theory
Customer communities
Relationship Mktg & Mgmt
WOM/networks
WOM/networks
CSR
Mass customization
Brand
Relationship Mktg & Mgmt
Customer equity
Journal of Marketing Research
Chevalier & Mayzlin (2006)
Bergkvist & Rossiter (2007)
Gupta, Lehmann & Stuart (2004)
Mazar, Amir & Ariely (2008)
Reinartz, Krafft & Hoyer (2004)
Srinivasan & Hanssens (2009)
Rindfleisch, Malter, Ganesan & Moorman (2008)
Trusov, Bodapati & Bucklin (2010)
Nair, Manchanda & Bhatia (2010)
Petrin & Train (2010)
284
200
187
91
190
64
81
21
18
25
10.05
8.14
6.03
6.02
5.78
5.24
4.13
2.95
2.77
2.52
WOM
Research methodology/survey research
Customer equity
Behavioral theory
Relationship Mktg & Mgmt
Metrics and firm value
Research methodology/survey research
Social networks
Social networks
Research methodology/choice
Marketing Science
Fiebig, Keane, Louviere & Wasi (2010)
Godes & Mayzlin (2009)
Keller & Lehmann (2006)
Hauser, Tellis & Griffin (2006)
Godes & Mayzlin (2004)
Gupta & Zeithaml (2006)
Rust & Chung (2006)
Zhang (2010)
Van den Bulte & Joshi (2007)
Eliashberg, Elberse & Leenders (2006)
46
54
128
122
224
90
79
23
52
61
8.13
7.36
7.18
6.80
5.81
4.76
4.06
3.39
2.97
2.91
Research methodology/choice
WOM
Brand
Diffusion/innovation
WOM/networks
Metrics and firm value
Relationship Mktg & Mgmt
Learning
Social networks/innovation
Movies
41
37
54
75
19
44
32
23
27
72
6.55
3.92
3.36
3.26
3.04
2.83
2.41
2.33
2.27
2.04
Search
Pricing
WOM/social networks
Pricing
Mass customization
Diffusion/innovation
Remanufacturing
Recommender systems
Online marketing
Social networks
Management Science
Ghose & Yang (2009)
Cachon & Swinney (2009)
Chen & Xie (2008)
Su (2007)
Franke, Schreier & Kaiser (2010)
Rahmandad & Sterman (2008)
Atasu, Sarvary & Van Wassenhove (2008)
Fleder & Hosanagar (2009)
Forman, Ghose & Goldfarb (2009)
Grewal, Lilien & Mallapragada (2006)
Research methodology/SEM
Diffusion/innovation
Social networks
Emerging markets
Social networks
Research methodology/choice
Network externalities
Corporate Social Responsibility (CSR)
Multichannel shoppers
Word of Mouth (WOM)
Note: Age-adjusted impact is estimated as the residual from a journal-specific negative binomial model relating number of citations to the age of the article (as measured by the number of
quarters to December, 2012). The model includes linear and squared age terms to capture the non-linear time trend of citations.
in our sample are aware of the marketing science tools available to
them, and there is a correlation between managers, academics, and intermediaries on the perception of the impact of those tools. Marketing
science articles that have influenced practice come in a wide range of
flavors. Some articles do not include empirical work (e.g., Hauser and
Shugan's Defender model), while others use only laboratory data
(e.g., Aaker and Keller's brand extension work). The survey among
authors of top dual impact articles provides excellent pointers as to
what it takes to write a top-journal article that achieves high academic and practice impact: symbiosis with consulting, going against
the grain at the right time, and working with experience. Examining
more recent developments in our field since 2004, we were able to
document the rise of digitization, mobile communications, and social networking, as well as further globalization of academia and
the important role of special fora. We now discuss implications of
our research for academia and practice, limitations of our research,
and ideas for future research in this area.
4.2. Implications for academia
Many marketing science academics may not see impacting practice
as their primary goal, letting the practice impact occur as a by-product
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
at best. A goal of practical impact might even be seen as counterproductive from the perspective of academic impact, distracting researchers
from their primary mission and potentially compromising the rigor
and integrity with which a problem is studied. Our study points to
several counterarguments as to why the two goals may not necessarily
be in conflict. First, practical problems may provide inspiration for new
breakthroughs as old tools are found inappropriate to solve them
(e.g., Louviere & Woodworth, 1983). Second, practical problems
lure academics away from the ivory tower, in which they may be
held captive by dominant paradigms.
Scholars who seek high practical impact may want to focus their research on decisions that are of greater importance to firms. In Table 2,
we identified such areas to be pricing management, new product management, customer and market selection, and product portfolio management. While scholars may very well choose their research area
using other inputs as well, we are able to offer scholars general advice
on the challenging road to practical impact, from surveying top 20
dual impact authors. Research in symbiosis with consulting may prove
to be a fertile ground for dual impact papers. The right timing in tackling
the problem and the willingness to go against the grain seem crucial as
well. Too early and radical a new idea may not find acceptance yet, too
late and a colleague may beat the researcher to the punch. That dual impact papers require a strong grounding both in marketing science and
practice, may explain why we find a disproportionate number of highly
experienced scholars in our 20 top dual impact papers.
139
response rate), we believe that it could introduce considerable bias.
We have attempted to address this by focusing largely on relative rather than absolute effects.
• Alternative knowledge diffusion routes. Textbooks, magazines and
newspapers represent important, alternate ways by which new marketing knowledge diffuses. Similarly, organizations such as ACNielsen,
Sawtooth, and Advanis are responsible for knowledge generation that
may not always begin in journal articles. Because we are not claiming
a complete catalog of the sources and transition nodes of marketing
science knowledge diffusion, this is less of a problem.
• We focus on success and that brings with it a number of benefits, as
well as being easier to observe. However, the lack of a control sample
of “failures” means that we cannot discriminate between that which
works and that which does not (though we can, to some extent, examine correlates of drivers of the degree of success).
Research in marketing science has relevance to many marketing decisions. At least that is what we find from the practitioners we surveyed.
Even though our samples may be biased towards the sophisticated end
of practice, our results are encouraging. Intermediaries consider segmentation tools and survey-based choice models to be most influential
relative to other tools. Intermediaries find individual articles, such as
Guadagni and Little (1983), Green and Srinivasan (1990), and Louviere
and Woodworth (1983) to be very influential on practice.
Our paper provides a good primer on marketing science for marketing
practitioners. It reviews an impressive body of top marketing science articles with dual impact. Therefore, it provides a guide to marketing science
research for (i) marketing practitioners with an interest in discovering
new areas or (ii) young market research professionals. This paper can
help them discover for which decisions or tools it is useful to turn to marketing science research, as well as which specific articles provide potentially useful insights and tools to which they should be exposed.
Having taken the first step in an effort to calibrate the effect of marketing science on marketing practice, we find ourselves faced with a number
of interesting but unanswered questions. These include the possibility of
a more comprehensive mapping and measures built up from marketing
practice, rather than down from journal articles. In terms of a more comprehensive mapping, it would be useful to consider other knowledge
vehicles (e.g., textbooks, magazines and newspapers), routes (e.g., user
knowledge generation and seminars), and participants (e.g., specialist
training educators). More representative samples would allow inferences
to be drawn about absolute impact rather than just relative impact. Finally, the unit of analysis we used is that of articles published in the period
1982–2003. Had it been scholars or over a longer timeframe, other researchers may have been more strongly represented.
The measure of relative rather than absolute impact raises another
issue; that of market penetration of marketing science knowledge and
tools (e.g., Roberts, 2000). Marketing science tools and the articles on
which they are based may be used in a wide variety of marketing decision making situations (i.e., the opportunity set is large). A more appropriate benchmark might perhaps be, “Of all the situations to which
these tools could have provided insight, in what per cent are the tools
actually being applied?” Our sense is that the number is low. If this is indeed the case, it is presumably hard for us to argue that the marketing
science tools currently in the market are in any way “standard” approaches to marketing and the measurement of its effect. We could contrast this penetration to that of approaches taught in other management
disciplines, such as accounting and finance, for example.
Overall, we hope that we have identified the basis for a continued
and richer study of the marketing science value chain.
4.4. Limitations and future research
Acknowledgments
In undertaking any research with as many dimensions as in our
study, researchers must make a number of choices and assumptions.
Our primary motivation in designing our research was to have a methodology that was objective and verifiable. To do so, we set up criteria upon
which to design our study, carefully evaluating those criteria and
obtaining input from a variety of knowledgeable sources at each stage
of the research. Yet, we understand that other scholars may have
approached the study differently and/or identified other study design
criteria. Some significant limitations of our research include the following:
We were inspired and supported by the Practice Prize Committee of
the INFORMS Society for Marketing Science, the Marketing Science Institute (MSI) and the Institute for the Study of Business Markets
(ISBM) in the execution of this work. Many individuals have also contributed to the ideas contained in the paper, particularly Gary Lilien
and Bruce Hardie. We would like to thank the Guest Editor, Area Editor
and two anonymous reviewers who provided constructive and insightful feedback to us. John Roberts acknowledges support from the London
Business School Centre for Marketing and Stefan Stremersch from the
Erasmus Center for Marketing and Innovation.
We have conducted many secondary analyses in the context of this
project, which we do not report in the interest of brevity. Please contact
the first author should you have an interest in obtaining supplementary
materials.
4.3. Implications for practice
• Citations as a screening mechanism. We are acutely aware of the irony
of starting to measure impact on practice with a list ranked by academic impact (i.e., citations). We tried to minimize this effect by including a pre-calibration stage. At worst, however, we can claim to
have gauged the practice impact of the population of highly cited marketing science articles (what we call dual impact).
• Biased sample. The use of MSI and ISBM led to practitioner samples
that were likely skewed towards greater sophistication. While this
likely skew might improve the reliability of responses (and the
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://dx.
doi.org/10.1016/j.ijresmar.2013.07.006.
140
J.H. Roberts et al. / Intern. J. of Research in Marketing 31 (2014) 127–140
References
Aaker, D. A., & Keller, K. L. (1990). Consumer evaluations of brand extensions. Journal of
Marketing, 54(1), 27–41.
Anderson, E. W., Fornell, C., & Lehmann, D. R. (1994). Customer satisfaction, market share,
and profitability: Findings from Sweden. Journal of Marketing, 58(3), 53–66.
Boulding, W., Kalra, A., Staelin, R., & Zeithaml, V. A. (1993). A dynamic process model of
service quality — From expectations to behavioral intentions. Journal of Marketing
Research, 30(1), 7–27.
Bucklin, R. E., & Gupta, S. (1999). Commercial use of UPC scanner data: Industry and academic perspectives. Marketing Science, 18(3), 247–273.
Cattin, P., & Wittink, D. R. (1982). Commercial use of conjoint-analysis — A survey. Journal
of Marketing, 46(3), 44–53.
Day, G. S. (1994). The capabilities of market-driven organizations. Journal of Marketing,
58(4), 37–52.
Fornell, C. (1992). A national customer satisfaction barometer — The Swedish experience.
Journal of Marketing, 56(1), 6–21.
Fornell, C., Johnson, M.D., Anderson, E. W., Cha, J. S., & Bryant, B. E. (1996). The American
customer satisfaction index: Nature, purpose, and findings. Journal of Marketing,
60(4), 7–18.
Gerbing, D. W., & Anderson, J. C. (1988). An updated paradigm for scale development incorporating unidimensionality and its assessment. Journal of Marketing Research,
25(2), 186–192.
Germann, F., Lilien, G. L., & Rangaswamy, A. (2013). Performance implications of
deploying marketing analytics. International Journal of Research in Marketing, 2,
114–128.
Geyskens, I., Steenkamp, J. E. B.M., & Kumar, N. (1998). Generalizations about trust in marketing channel relationships using meta-analysis. International Journal of Research in
Marketing, 15(3), 223–248.
Green, P. E., & Srinivasan, V. (1990). Conjoint-analysis in marketing — New developments
with implications for research. Journal of Marketing, 54(4), 3–19.
Griffin, A., & Hauser, J. R. (1993). The voice of the customer. Marketing Science, 12(1),
1–27.
Guadagni, P., & Little, J.D. C. (1983). A logit model of brand choice calibrated on scanner
data. Marketing Science, 2(3), 203–238.
Guadagni, P., & Little, J.D. C. (2008). A logit model of brand choice calibrated on scanner
data: A 25th anniversary perspective. Marketing Science, 27(1), 26–28.
Hauser, J. R., & Shugan, S. (1983). Defensive marketing strategies. Marketing Science, 2(4),
319–360.
Humby, C., Hunt, T., & Phillips, T. (2008). Scoring points: How Tesco continues to win
customer loyalty. London, UK: Kogan Page.
Hunt, S. D., & Morgan, R. M. (1995). The comparative advantage theory of competition.
Journal of Marketing, 59(2), 1–15.
Jaworski, B. J., & Kohli, A. K. (1993). Market orientation — Antecedents and consequences.
Journal of Marketing, 57(3), 53–70.
Keller, K. L. (1993). Conceptualizing, measuring, and managing customer-based brand equity. Journal of Marketing, 57(1), 1–22.
Kotler, P., & Keller, K. L. (2012). Marketing management 14th edition. Upper Saddle River,
NJ: Prentice Hall.
Leeflang, P.S. H., & Wittink, D. R. (2000). Building models for marketing decisions: Past,
present, future. International Journal of Research in Marketing, 17(2–3), 105–126.
Lehmann, D. R., McAlister, L., & Staelin, R. (2011). Sophistication in research in marketing.
Journal of Marketing, 75(July), 155–165.
Lilien, G. L. (2011). Bridging the academic-practitioner divide in marketing decision
models. Journal of Marketing, 75(2), 196–210.
Lilien, G. L., Kotler, P., & Moorthy, K. S. (1992). Marketing models. Englewood Cliffs: Prentice Hall.
Lilien, Gary L., Rangaswamy, Arvind, & De Bruyn, Arnaud (2007). Principles of marketing
engineering. : Trafford.
Lilien, G. L., Roberts, J. H., & Shankar, V. (2013). Effective marketing science applications:
Insights from the ISMS practice prize finalist papers and projects. Marketing Science,
32(2), 229–245.
Louviere, J. J., & Woodworth, G. (1983). Design and analysis of simulated consumer choice
or allocation experiments — An approach based on aggregate data. Journal of
Marketing Research, 20(4), 350–367.
Mahajan, V., Muller, F. M., & Bass, (1990). New product diffusion-models in marketing —
A review and directions for research. Journal of Marketing, 54(1), 1–26.
Morgan, R. M., & Hunt, S. D. (1994). The commitment–trust theory of relationship marketing. Journal of Marketing, 58(3), 20–38.
Orme, B. (2013). Advances and trends in marketing science from the Sawtooth Software perspective. Orem, UT: Sawtooth Software, Inc.
Parasuraman, A., Zeithaml, V. A., & Berry, L. L. (1985). A conceptual-model of service
quality and its implications for future-research. Journal of Marketing, 49(4),
41–50.
Pieters, R., Baumgartner, H., Vermunt, J., & Bijmolt, T. H. A. (1999). Importance and similarity in the evolving citation network of the International Journal of Research in Marketing. International Journal of Research in Marketing, 16(2), 113–127.
Punj, G., & Stewart, D. W. (1983). Cluster-analysis in marketing-research — Review and
suggestions for application. Journal of Marketing Research, 20(2), 134–148.
Reibstein, D. J., Day, G., & Wind, J. (2009). Guest editorial: Is marketing academia losing its
way? Journal of Marketing, 73(June), 1–3.
Roberts, J. H. (2000). The intersection of modelling potential and practice. International
Journal of Research in Marketing, 13(3), 127–134.
Rust, R. T., & Cooil, B. (1994). Reliability measures for qualitative data: Theory and implications. Journal of Marketing Research, 31(1), 1–14.
Rust, R. T., Zahorik, A. J., & Keiningham, T. L. (1995). Return on quality (ROQ) — Making
service quality financially accountable. Journal of Marketing, 59(2), 58–70.
Simonson, I., & Tversky, A. (1992). Choice in context — Trade-off contrast and extremeness aversion. Journal of Marketing Research, 29(3), 281–295.
Stremersch, S., & Verhoef, P. C. (2005). Globalization of authorship in the marketing discipline: Does it help or hinder the field? Marketing Science, 24(4), 585–594.
Stremersch, S., Verniers, I., & Verhoef, P. C. (2007). The quest for citations: Drivers of article impact. Journal of Marketing, 71(3), 171–193.
Vanheerde, H. J., Gupta, S., & Wittink, D. R. (2003). Is 75% of the sales promotion bump
due to brand switching? No, only 33% is. Journal of Marketing Research, 40(4),
481–491.
Wierenga, B., & van Bruggen, G. H. (2000). Marketing management support systems: Principles, tools and implementation. Boston: Kluwer Academic Publishers.
Wittink, D. R., & Cattin, P. (1989). Commercial use of conjoint analysis: An update. Journal
of Marketing, 53(3), 91–96.
Wittink, D. R., Vriens, M., & Burhenne, W. (1994). Commercial use of conjoint analysis in
Europe: Results and critical reflections. International Journal of Research in Marketing,
11(1), 41–52.
Intern. J. of Research in Marketing 31 (2014) 147–155
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
Full Length Article
Probabilistic selling vs. markdown selling: Price discrimination and
management of demand uncertainty in retailing☆
Dan Hamilton Rice a,⁎, Scott A. Fay b,1, Jinhong Xie c,2
a
b
c
Department of Marketing, E.J. Ourso College of Business, Louisiana State University, 2119 Building Education Complex, Baton Rouge, LA 70803-6314, United States
Department of Marketing, Martin J. Whitman School of Management, Syracuse University, 721 University Avenue, Syracuse, NY 13244, United States
Department of Marketing, University of Florida, P.O. Box 117155, Gainesville, FL 32611-7155, United States
a r t i c l e
i n f o
Article history:
First received in 26 September 2011
and was under review for 12 months
Available online 12 October 2013
Area Editor: Els Gijsbrechts
Keywords:
Probabilistic selling
Pricing
Demand uncertainty
Markdowns
Price discrimination
a b s t r a c t
Markdown selling (i.e., price reductions over the course of the selling season) is a strategy to implement price
discrimination and to manage market uncertainty that has been widely adopted by retailers. This paper explores
the potential advantage of introducing an additional tool to the arsenal of retailers, probabilistic selling
(i.e., offering consumers a choice to buy a product that can turn out to be any item from a predetermined set of
distinct items). We show that both probabilistic and markdown selling strategies serve as price discrimination
tools by offering buyers an option to purchase a “damaged” good (an uncertain product under the former and
delayed consumption of a product under the latter). However, the two strategies segment markets based on
different types of buyer heterogeneity: buyer preference strength under probabilistic selling and buyer patience
under markdown selling. Our analytical model reveals that, compared with markdown selling, probabilistic selling
can (1) improve margin management by increasing revenue from full-price sales and reducing the magnitude of
discounts; and (2) improve inventory utilization by reducing stockouts and the amount of excess inventory. We
identify the conditions required for probabilistic selling to be more profitable than markdown selling.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
In an effort to obtain the maximum profit across a diverse set of
customers, retailers often offer price reductions over the course of the
selling season. It is estimated that one-third of all goods are sold at
marked-down prices (Friend & Walker, 2001) and discounts due to
markdowns by US retailers amount to $200 B a year (Levy, Grewal,
Kopalle, & Hess, 2004). Although costly, markdowns can be a valuable
tool for improving profit margin management because they allow the
retailer to price discriminate across time, i.e., sell the product at a high
price early in the season to customers who value the product highly
and are unwilling to wait, and at a discounted price later in the season
to customers who are willing to delay their purchases (Besbes & Lobel,
☆ The paper has benefited from the helpful comments made by the seminar participants
at the University of California, The Ohio State University, Rice University, Baylor University,
University of Missouri–Kansas City, Syracuse University, Purdue University (Krannert),
University of North Carolina–Charlotte, University of Illinois at Urbana–Champaign, the
Marketing Science Conference (2009; Ann Arbor, Michigan) and the UTD Marketing
Conference (University of Texas–Dallas).
⁎ Corresponding author at: Department of Marketing, Louisiana State University,
Room 2100, Business Education Complex, Baton Rouge, LA 70808-6314, United States.
Tel.: +1 225 578 8788; fax: +1 225 578 8616.
E-mail addresses: danrice@lsu.edu (D.H. Rice), scfay@syr.edu (S.A. Fay),
jinhong.xie@warrington.ufl.edu (J. Xie).
1
Tel.: +1 315 443 3456; fax: +1 315 442 1461.
2
Tel.: +1 352 273 3270; fax: +1 352 846 0457.
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.08.006
2012; Nair, 2007; Su, 2007). The markdown strategy can also enhance
inventory management for retailers who are unable to accurately
predict consumers' demand for each particular product (Lazear, 1986),
e.g., by starting with a high price and reducing the price if units of the
item remain unsold. Retailers are continually searching for more
efficient ways to improve margin management and enhance inventory
utilization.3 In this paper, we consider one such alternate selling
mechanism, namely probabilistic selling (PS), and show that there are
situations in which this mechanism can be advantageous relative to
traditional markdowns both in enhancing price discrimination and in
overcoming the main problems associated with demand uncertainty,
namely stockouts and excess inventory.
A probabilistic product is an offer involving the probability of
obtaining any one of a set of multiple distinct items (Fay & Xie, 2008).
Probabilistic selling (PS) is a selling strategy under which the seller
creates probabilistic goods using the seller's distinct products or services
and offers such goods to potential buyers as additional purchase choices.
Notable examples of sellers of probabilistic products include priceline.
3
Previous research has focused on developing sophisticated dynamic markdown
algorithms (e.g., Bitran & Mondschein, 1997; Chung, Flynn, & Zhu, 2009; Mantrala &
Rao, 2001; Sullivan, 2005), implementing inventory management systems (Friend &
Walker, 2001; Khouja, 1995; Ross, 1997), and identifying alternate ways to dispose of
distressed goods, such as via off-price retailers and outlet stores (Coughlan & Soberman,
2005; Levy & Weitz, 2004, p. 56; Petruzzi & Monahan, 2003) or online auctions (Wang,
Gal-Or, & Chatterjee, 2009; Wood, Alford, Jackson, & Gilley, 2005).
148
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
Table 1
Related literature and distinguishing characteristics of current paper.
Research focus
Developing theory and applications of
probabilistic (opaque) selling strategy
Developing decision support system to
implement probabilistic (opaque)
selling strategy
Research paper
Current paper
Fay and Xie (2012)
Fay (2008)
Fay and Xie (2010)
Fay and Xie (2008)
Jiang (2007)
Shapiro and Zillante (2009)
Jerath et al. (2010)
Zouaoui and Rao (2009)
Granados et al. (2008)
Post (2010)
Anderson (2009)
Anderson and Xie (2012)
Gallego and Phillips (2004)
Mang et al. (2012)
Petrick, Steinhardt, Gonsch,
and Klein (2012)
Methodology a
AM, LE
AM
AM
AM
AM
AM
LE
AM
E
AM, E
AM
AM
E
AM
E
AM
Endogenous
variables
Price
Capacity
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
Yes
Yes
No
No
No
No
No
No
No
No
No
No
No
No
No
No
Consumers optimally
time purchase
Probabilistic good can
cannibalize full-price sales
Yes
No
No
Yes b
No
No
No
Yes
No
No
No
No
No
No
Yes
No
Yes
No
No
Yes
No
No
No
No
Yes
Yes
Yes
No
Yes
No
Yes
No
a
“AM” = Analytical MODELING; “E” = Empirical; “LE” = Lab experiments.
In Fay and Xie (2010), the two selling periods are the advanced period (prior to consumers learning their true valuations) and the spot period. In the current paper, consumers know
their valuations in both periods. Thus, the major difference between the two periods is the time delay rather than differences in the information available to consumers.
b
com, lastminutetravel.com, and hotwire.com, websites where consumers
can purchase travel services for which specific attributes of the service
(e.g., the itinerary of the flight, the location of the hotel, or the identity
of the car rental company) are not revealed until after payment. Recently,
the idea of offering probabilistic goods has also been adopted by several
online retailers (e.g., swimoutlet.com, agonswim.com, speedo.com, and
kidsurplus.com) who offer discounted “grab bag” apparel and shoes,
where patterns and styles are chosen randomly by the website.4 As
technological advances make it much more practical to implement PS
both in online and brick-and-mortar shopping environments, more
retailers can potentially benefit from adopting this novel selling strategy
(Fay & Xie, 2008). While the existing research on PS has significantly
advanced our understanding of the fundamental drivers of PS and
illustrates its general applicability, it is important to extend the research
to understand how this novel strategy may address some unique
problems in the retailing industry and to explore whether PS can be a
valuable alternative to offering late-season markdowns.
Most retailers strategically invest in inventory prior to the selling
season, control the prices of their products over the entire selling season,
and must account for how consumers time their purchases in response
to these chosen prices. We introduce a model that incorporates each of
these key characteristics. As shown in Table 1, among the current
research on PS,5 ours is the only model that incorporates all of the
following three key characteristics: (1) The seller optimally chooses its
prices for the probabilistic goods and the specified goods; (2) the seller
optimally adjusts its inventory orders when introducing probabilistic
goods; and (3) consumers strategically choose when to purchase in
order to maximize their expected surplus. By incorporating these
three critical factors, we are able to develop the theory and implications
of PS for the retailing industry. In particular, our model enables us to
compare discounting on the basis of time (high initial price and a
discounted price if the consumer delays her purchase) versus
discounting on the basis of product opacity (i.e., setting a high price
for each specified good and a discounted price if the consumer will
purchase the probabilistic good). Thus, the paper's primary contribution
4
See an example at http://www.swimoutlet.com/product_p/1623.htm
Several papers (Gallego & Phillips, 2004; Mang et al., 2012, and Petrick et al., 2012)
consider a seller who does not assign products to buyers of the probabilistic good until a
time that is substantially later than the day of purchase. They refer to this business model
as flexible selling rather than PS. However, consistent with Fay and Xie (2012), we
consider these papers as part of the PS literature since delaying product assignment can
be viewed as an alternative way of implementing the PS strategy.
5
is that it is the first to examine the profit advantage of the PS strategy
relative to the more commonly utilized strategy of marking down
merchandise over time, i.e., the markdown (MD) selling strategy. We
identify factors under which PS can be a more useful tool for retailers
as they attempt to price discriminate across consumers. We find that
PS and MD can be complementary strategies since, in some market
settings, PS is a profitable form of price discrimination whereas MD is
not, while, in other market settings, price discrimination is profitable
via MD but not profitable via PS.
A second contribution of the paper is that, by introducing a model
that allows a probabilistic good to cannibalize full-price sales, we can
examine the factors that affect the extent of cannibalization by the
probabilistic good and determine whether PS can remain advantageous
in its presence. Most extant analytical research on PS utilizes a Hotelling
model to account for consumer heterogeneity (Fay, 2008; Fay & Xie,
2008, 2012; Jerath, Netessine, & Veeraraghavan, 2010; Jiang, 2007). A
feature of the Hotelling model is that all consumers have the same
expected value for the probabilistic good. As a result, price can be set
at this common expected value, thus eliminating consumer surplus for
all buyers of the probabilistic good. Since the probabilistic good does
not generate positive surplus, the seller does not have to worry about
any consumers switching from a higher-priced specified good to the
lower-priced probabilistic good. However, cannibalization is a crucial
concern under MD in the retailing industry because retailers are
apprehensive that a discounted price at the end of the season may
entice both low- and high-valuation buyers to delay their purchases
(especially if the magnitude of the discount is very large). Thus, to
provide an adequate comparison of PS with MD, the model must be
capable of capturing the cannibalization effect under both strategies.
Note that several empirical studies incorporate the cannibalization
effect into their model estimations (Anderson & Xie, 2012; Granados,
Gupta, & Kauffman, 2008; Mang, Post, & Spann, 2012; Zouaoui and
Rao (2009)). However, since demand is modeled in reduced form in
these papers, i.e., cross-price effects exist between the probabilistic
good and the specified goods, these studies do not analyze the factors
which affect the magnitude of this cannibalization effect or how
cannibalization impacts the profitability of PS, as we do here.
The rest of this paper is organized as follows. In the next section, we
use a lab experiment to illustrate the potential advantages of the PS and
MD strategies relative to a No Discounting strategy. In Section 3, we
illustrate how both PS and MD can enable a retailer to price discriminate
and then compare the profitability of these two strategies. In Section 4,
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
we extend the analytical model to allow for demand uncertainty and
demonstrate that our key results continue to hold, and are even
strengthened, in such markets. We conclude the paper by summarizing
our results and suggesting areas for future research.
149
and MD may enable a firm to effectively segment its customers.
Specifically, we develop a model to explain the conditions under which
PS will be more profitable than MD, and vice-versa.
3. Price discrimination: Probabilistic selling versus markdowns
2. Two advantageous strategies: Probabilistic selling
and markdowns
2.1. Motivation
For MD and PS to be viable selling strategies, they must be more
advantageous than offering no discounts to consumers, i.e., the No
Discounting (ND) strategy. We performed an experimental study to
explore how consumers respond to these different types of discounts
versus a situation in which no discounts are offered. One hundred and
thirty-eight undergraduate business students participated in the webbased study in exchange for extra credit, purportedly to help
researchers better understand the online purchase behavior of college
students. The design was a 2 (Selling Strategy: MD, PS) × 2 (discount:
5% or 25%) between-subjects design with an ND control condition.
Participants were randomly assigned to one of the five conditions. Five
students were eliminated for failure to follow directions, which left
133 responses for analysis.
Participants were told that they would be shown offers for products
(T-shirts) that had recently been for sale online, and then would be
asked to answer the purchase questions. Next, the concept of
probabilistic goods was explained to the participants, and they were
told that the displayed offers may or may not include an option to
purchase a probabilistic good. Participants were asked to make a
purchase decision for T-shirt offers in Period 1 and then to rank their
strength of preference (SOP) between the distinct T-Shirts on a scale
of 0 to 50, where SOP is a measure of the difference between valuations
for one's preferred and one's less-preferred product. Heterogeneity in
the strength of preferences has been previously hypothesized as a key
factor in determining the profitability of PS (Fay & Xie, 2008). In the
MD condition, full-priced items were offered in Period 1 and, if no
purchase occurred in this period, marked-down items were offered in
Period 2. In the PS condition, full-priced items as well as the probabilistic
good were offered in Period 1, but there were no second-period
offerings. In the ND control condition, both full-priced items were
offered in Period 1 only. If participants chose to purchase a T-Shirt in
Period 1, the shopping experience ended in all conditions.
2.2. Results
A one-way ANOVA for revenue generated by the purchase decisions
of the participants was conducted across the five selling conditions. The
results indicate a significant effect of selling condition (F(4128) = 4.79,
p=.001). Specific contrasts indicate that the average revenue under ND
($5.86) was less than under the PS 5% condition (MPS5% = $13.31,
t(133) = 4.00, p b .001), the PS 25% condition (MPS25% = $12.16,
t(133) = 3.31, p = .001), the MD 5% condition (MMD5% = $9.53,
t(133) = 1.95, p = .05) and the MD 25% condition (MMD5% = $9.42,
t(133) = 1.95, p = .05). For the two PS conditions, SOP was significantly
greater for those buying full-priced goods (MDS = 34.05) than for
those who purchased a discounted good (MPS = 22.18; F(1,42) = 7.03,
p = .01). No significant preference difference was present in the MD
conditions between those who chose full price and those who chose a
discounted good (p N .8).
Together these findings suggest that, for our sample, both MD and PS
were effective at increasing revenue relative to ND, which raises the
question of when each strategy works best. The significant effect of
SOP is consistent with the arguments in the extant literature that
heterogeneity in the strength of consumers' preferences is a fundamental
profit driver for PS. In the following sections, we use an analytical model
to explore an environment, similar to our experiment, in which both PS
In this section, we introduce a formal mathematical model to
explore whether, and under what conditions PS is more profitable
than MD. We begin by focusing on how PS and MD enable price
discrimination. In Section 4, we extend the model to incorporate
demand uncertainty and capacity constraints. Our model allows for
asymmetric product preferences, costly inventory, endogenous capacity
constraints, heterogeneous discount rates, and a spectrum of different
product valuations by consumers. Overall, we find that, although PS is
not more profitable than MD in all scenarios, there exists a sufficiently
broad range of situations in which PS is advantageous to warrant
increased attention to this new tool.
3.1. Modeling assumptions
3.1.1. Seller behavior
Consider a retailer with two products, A and B, (e.g., a red T-shirt and
a white T-shirt) and two possible selling periods (Periods 1 and 2). Prior
to the first selling period, the retailer orders KA units of product A and KB
units of product B, where each unit costs c. The seller has three
alternative selling strategies:
(1) No Discounting (ND), under which the prices offered do not
change over time, i.e., the price of product A and the price of
product B is PND in both the first and second periods.
(2) Markdown selling (MD), under which prices vary over time.
Specifically, the price in the first period (of product A and of
product B) is P1MD and the price of each product is P2MD in the
second period.
(3) Probabilistic selling (PS), under which the seller offers each
product individually and also a probabilistic product that can
turn out to be either product A or product B. Specifically, the
price of product A and the price of product B is P1PS. In addition,
the seller offers a probabilistic good in the first period (at a
price of P0PS). After purchase, the firm immediately determines
which product the buyer will receive, each of which is equally
likely.6
3.1.2. Buyer behavior
Each consumer purchases at most one product, choosing the
purchase option that yields the highest net surplus. There are two
types of consumers: θ = {H,L}, where θ = H represents consumers
with high product valuations and θ = L represents consumers with
low product valuations. Each type makes up half of the total population
and we normalize the size of each segment to one. Let vθF be consumer
θ's value in Period 1 for her favored product and vθU be consumer θ's
value for her less favored product. Thus, by definition, we have vHF N vLF,
vHF N vHU and vLF N vLU. To reduce notation, we normalize the valuations
so that the maximum valuation is one: vHF = 1. Furthermore, products
are more highly valued if they are purchased in Period 1 rather than in
Period 2. Valuations are lower in the second period due to consumer
impatience, loss of perceived newness of the product, or loss of the
opportunity to use the product during the first period. Specifically, in
Period 2, a consumer's favored product is valued at dθvθF and her lessfavored product is valued at dθvθU, where dθ is naturally restricted to
6
Fay and Xie (2008) demonstrate that a seller typically finds an equal probability of
assignment optimal under various demand conditions. Thus, the assumption of equal
probability is commonly made (e.g., Fay & Xie, 2010; Jiang, 2007).
150
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
the parameter region 0 b dθ b 1 for θ = {H,L}. We allow the products to
differ in their aggregate popularity. In particular, α of consumers
"1 !
#
≤α ≤1 favor the “popular” product and 1 − α of customers favor
2
the “unpopular” product. Note that a specific consumer's favored
product may actually be the less popular one. Throughout this section,
we assume that product A is the popular product and that the seller
and the buyers both know this. In Section 4, we extend the model to
allow for demand uncertainty.
The seller chooses prices (P1MD, P2MD) to maximize its total profit
given the constraints in (2)9:
MD
¼ 1−Max½dH −dL vLF ; 0%
P1
MD
P2
These prices result in a profit of:
MD
Π
3.2. Three strategies
3.2.1. No Discounting (ND)
Under ND, the firm sells in both periods at a price PND. All sales will
occur in the first period. The seller chooses the inventory orders (KA,
KB) and price, PND, in order to maximize its profit. These optimal values
and the resulting profit are:
ND
KA ¼
P
ND
¼
f
f
c ≤ ^c
2ð1−α Þ
c ≤ ^c
^cbc ≤1; K ND
^cbc ≤1;
1−α
B ¼
c N1
0
c N1
vLF
c ≤ ^c
2ðvLF −cÞ c ≤ ^c
^cbc ≤1; Π ND ¼
^cbc ≤1
1
1−c
N=A
c N1
0
c N1
2α
α
0
where ^c ¼ 2νLF −1
f
f
ð1Þ
3.2.2. Markdown selling (MD)
In MD, the seller offers each specified product in both the first and
second periods (at prices of P1MD and P2MD, respectively). Under this
strategy, each H-type consumer buys her favorite product in the first
period and each L-type consumer buys her favorite product in the
second period.7,8 To meet demand, the seller needs KMD
= 2α and
A
MD
KND
and P2MD induce such a purchasing
B = 2(1 − α). The prices P1
pattern only if the following incentive compatibility and participation
constraints are met:
MD
½P1% : 1‐P 1 ≥0
MD
½P2% : dL vLF ‐P 2 ≥0
$
%
MD
MD
½IC1% : 1‐P 1 − dH −P 2 ≥0
$
%
MD
MD
½IC2% : dL vLF ‐P 2 − vLF −P 1 ≥0
ð2Þ
7
Alternative strategies in which H-types wait to purchase until the second period and
L-types either purchase in the first period or also wait to purchase until the second period
would yield strictly less profit.
8
One could envision a scenario in which L-type consumers also purchase in the first
period. However, our focus is on identifying scenarios in which MD is strictly more
profitable than ND. If L-types also purchase in the first period, no markdown sales occur
and thus MD and ND would be identical.
¼ 1−Max½dH −dL ν LF ; 0% þ dL ν LF −2c
ð4Þ
3.2.3. Probabilistic selling (PS)
Under PS, in addition to selling each specified product in the first
period (each at a price of P1PS), the seller also offers a probabilistic good
in the first period (at a price of PPS
O ). After purchase, the firm immediately
determines which product the buyer will consume (each of which is
equally likely). Under this strategy, each H-type consumer buys her
favored product in the first period and each L-type consumer buys the
probabilistic good (also in the first period).10 To meet demand, the seller
1
2αþ1
1
3−2α
selects K MD
and K MD
A ¼αþ2¼ 2
B ¼ ð1−α Þ þ 2 ¼ 2 . The expected
value of the probabilistic good to consumer θ is vθo = (vθF + vθU)/2. The
prices P1PS and PPS
0 induce such a purchasing pattern only if the following
incentive compatibility and participation constraints are met:
h
If costs are low ðc ≤ ^cÞ, the seller orders sufficient capacity to serve all
ND
consumers: KND
A + KB = 2. Price is set so that L-type consumers are
willing to purchase. For moderate costs ð^cbc ≤1Þ, the seller only orders
ND
enough capacity to serve the H-type consumers, KND
A + KB = 1, and
the price is set so that H-type consumers are just willing to purchase.
For higher costs, c N 1, it is impossible for the seller to earn a positive
profit.
For the remainder of the paper, we assume that c ≤ ^c , so that it is
optimal to serve both consumer types, enabling us to focus on the role
of price discrimination (since costs will be the same under all three
strategies). For c N ^c , total sales under either MD or PS will be higher
than under ND, and thus the role of price discrimination is confounded
with the role of market expansion. Furthermore, even at these higher
costs, the magnitude of c does not impact the focal comparison between
MD and PS, because these two strategies require the same amount of
inventory and thus incur the same costs.
ð3Þ
¼ dL vLF
h
′
P1
′
P2
i
PS
: 1−P 1 ≥0
i v þv
PS
LU
: LF
−P 0 ≥0
2
&
'
1 þ vHU
PS
PS
: 1−P 1 −
−P 0 ≥0
2
h
i v þv
$
%
′
PS
PS
LF
LU
−P 0 − vLF −P 1 ≥0
IC2 :
2
ð5Þ
h
IC1
′
i
The seller chooses prices (P1PS and P0PS) to maximize its total profit
given the constraints in Eq. (5)11:
(
)
(
)
1 þ vHU vLF þ vLU
1−vHU þ vLF þ vLU
PS
P 1 ¼ 1−Max
−
; 0 ¼ Min
;1
2
2
2
v þ vLU
PS
P o ¼ LF
2
ð6Þ
These prices result in a profit of:
PS
Π
(
)
1−ν HU þ vLF þ vLU
v þ vLU
; 1 þ LF
−2c
¼ Min
2
2
ð7Þ
3.3. Comparison of profit
Lemma 1 summarizes the conditions under which MD and PS,
respectively, are more profitable than ND. Proofs of the Lemmas,
Corollaries, and Propositions are given in the Web Appendix.
9
At the optimal solution, constraint [P2] must be binding. If P2MD was set such that [P2]
did not hold with equality, then it would be possible for the seller to raise this secondperiod price and still sell to L-type consumers (and possibly even sell to the H-types at a
higher first-period price), thus increasing its profit. Similarly, either [P1] or [IC1] must
bind. If P1MD was set such that neither of these constraints held with equality, then it
would be possible for the seller to raise this first-period price and still sell to H-type
consumers, thus increasing its profit.
10
An alternative strategy in which H-types also buy the probabilistic good would yield
strictly less profit since the firm can charge a higher price for a consumer's favored good
than it can for the probabilistic good.
11
At the optimal solution, constraint [P2’] must be binding. If P0PS was set such that [P2’]
did not hold with equality, then it would be possible for the seller to raise the price of the
probabilistic product and still sell it to L-type consumers, thus increasing its profit.
Similarly, either [P1’] or [IC1’] must bind. If P1PS was set such that neither of these
constraints held with equality, then it would be possible for the seller to raise the price
of the specified goods and still sell to H-type consumers, thus increasing its profit.
151
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
Lemma 1. (Conditions required for price discrimination)
Compared with the No Discounting strategy, which earns the seller
the same profit from both H- and L-type consumers,
a) Markdown selling allows the seller to benefit from price
discrimination if the increase in profit from H-type consumers
(who buy in the first period) is larger than the decrease in profit
from L-type consumers (who buy in the second period).
b) Probabilistic selling allows the seller to benefit from price
discrimination if the increase in profit from H-type consumers
(who buy the specified goods) is larger than the decrease in profit
from the L-type consumers (who buy probabilistic good).
Mathematically,
PS
Π NΠ
Π
MD
ND
NΠ
ND
PS
PS
PS
if ΔPrice to H‐Types −ΔPrice to L‐Types N0
MD
PS
ND
PS
¼
ND
PS
MD
P 1 −P 0 ; ΔPrice to H‐Types
PS
Π −Π
þ
MP
ND
¼ P 1 −P 1 ;
ND
MD
ΔMD
Price to L ‐ Types = P1 − P2 , and these prices are given in Eqs. (1),
(3), and (6), respectively.
Two important corollaries to Lemma 1 are:
Corollary 1. A necessary condition for the inequality in Lemma 1a) to be
satisfied is dL N dH, i.e., H-types are less patient than L-types.
Corollary 2. A necessary condition for the inequality in Lemma 1b) to be
satisfied is vLF-vLU b 1-vHU, i.e., H-types have stronger product preferences
than L-types.
These two corollaries indicate that “damaging” the good (either by
delaying consumption or pairing it with one's less-favored good)
must create greater disutility for the high- than for the low-value
consumers in order for each respective pricing strategy to be profitable.
Under MD, the discounted second-period product must be relatively
attractive to the low-value consumers (so that the second-period
price is not too low), but also relatively unattractive to the high-value
consumers (so that they are willing to pay a premium to purchase in
the earlier period). A necessary condition to achieve a net advantage
from MD is that high-value consumers must be less willing to wait
than low-value consumers. Su (2007) and Besbes and Lobel (2012)
also point out that inter-temporal price discrimination is predicated
on there being an inverse relationship between consumers' patience
and their valuations. Similarly, under PS, the probabilistic good must
be relatively more appealing to the low-value consumers so that the
discount for the probabilistic good does not have to be too large and,
thus, a relatively high price can be obtained for the specified goods.
Furthermore, although Lemma 1 indicates that both MD and PS can
be used to segment the market, the basis of this segmentation is very
different in each case: MD is based on buyer heterogeneity in their
discount for time (Corollary 1) and PS is based on buyer heterogeneity
in their product preference strength (Corollary 2). Such differences
create the potential for retailers to benefit from PS, as summarized by
the following proposition.
Proposition 1. (Profit advantage of probabilistic selling)
(a) PS expands the scenarios under which the seller can benefit from
price discrimination, i.e., there is a subset of parameters such that
PS is more profitable than ND, but MD is not.
(b) PS is more profitable than MD if it can create the following two
advantages (or if one of these advantages is sufficiently large):
1. A more valuable “damaged” good, i.e., L-type consumers are willing
to pay more for the probabilistic product in the first period than for
their preferred product in the second period.
&
MD
¼
$v þ v
%
LF
LU
−dL vLF
2 ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
Difference in how much L‐types value a
probabilistic product in the the first period
versus their favored product in the
second period:
(
)'
1 þ vHU vLF þ vLU
−
;0
Max½dH −dL vLF ; 0%‐Max
2
2
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
N0
Difference in surplus the seller must provide to H‐types to get them to forego
purchasing the discounted in the markdown selling strategy versus
in the probabilistic selling strategy
¼
MD
if ΔPrice to H‐Types −ΔPrice to L‐Types N0
where ΔPrice to H‐Types ¼ P 1 −P 1 ; ΔPrice to L‐Types
2. A lessened cannibalization threat, i.e., H-type consumers demand
less surplus to forgo purchasing the discounted product under PS
than under MD.
Formally, PS earns the retailer a higher profit if the following
condition holds:
ΔValue to L‐Types
þ
ΔValue to H‐Types
(c) The profit advantage of PS relative to MD (ΠPS–ΠMD) increases as
1. L-types' valuations for the probabilistic good increase (i.e., a larger
VLU.)
2. L-types become more impatient (i.e., a lower dL)
3. H-types' valuations for the probabilistic good decrease (i.e., a smaller
VHU).
4. H-types become more patient (i.e., a higher dH).
Fig. 1 illustrates the main results from Proposition 1. In Fig. 1a), we
show the impact of vLU and dL on the relative profit of PS, MD, and ND
(holding vLF, vHU, and dH constant). Fig. 1b) shows the impact of vHU
and dH on the relative profit, holding vLF, vLU, and dL constant. The
unshaded regions indicate parameter values for which neither MD nor
PS can outperform ND. In the polka-dotted regions, PS is more profitable
than both MD and ND. In shaded regions, MD is more profitable than
both PS and ND.
Consistent with Corollary 1, MD is not a useful price discrimination
tool if either L-type consumers are too impatient (i.e., dL b 2/3 in
Fig. 1a) or H-type consumers are too patient (i.e., dH N 5/8 in Fig. 1b).
However, even with such parameters, L-type consumers may have
sufficiently weak preferences (i.e., vLU N ½ in Fig. 1a) and/or H-types
may have sufficiently strong preferences (i.e., vHU b ½ in Fig. 1b) so
that PS is more profitable than ND. Thus, PS expands the range of market
settings under which the seller can price discriminate, i.e., in the regions
denoted “Only PS is advantageous to ND,” price discrimination on the
basis of time is not profitable, but price discrimination on the basis of
consumer preference strength is profitable. In contrast, in the regions
denoted “Only MD is advantageous to ND,” price discrimination is
profitable if it is done on the basis of time but not on the basis of
consumer preference strength. Thus, PS and MD can be viewed as
complementary strategies, with one strategy often being profitable in
markets where the other strategy would not be beneficial.
In markets where both MD and PS outperform ND, the relative
advantage of the two price discrimination mechanisms depends on
two effects: (1) Difference in Value to L-types, i.e., how valuations for
the low-value consumers differ across the two price discrimination
mechanisms, and (2) Difference in Surplus to H-types, i.e., how the
amount of surplus that must be allocated to the high-value consumers
differs between MD and PS.12 First consider the Value-to-L-types effect.
Under both MD and PS, low-value consumers purchase a product that is
12
In a Hotelling model, which is most often employed in the extant literature (e.g., Fay &
Xie, 2008, 2010; Jerath et al., 2010; Jiang, 2007), the Surplus to H-types effect would be
missing since, in that model, no consumer obtains a positive surplus from purchasing
the probabilistic good and, thus, no extra incentive is needed to induce strongpreferenced consumers to buy their preferred good. Therefore, the current model has
the advantage of allowing us to analyze the cannibalization effect of PS.
152
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
a)
vLF = ¾ ; v HU
b)
dH = ½
3
vLU
1
2
d L = ¾ ; v LU = ½
1
Only PS is
advantageous
to ND
4
v LF
B1
B2
vHU
C
1
2
B2
Only MD is
advantageous
to ND
Neither PS nor MD is
advantageous to ND
Neither PS nor
MD is
advantageous
to ND
Only MD is
advantageous
to ND
Only PS is
advantageous
to ND
B1
A
0
0
0
2
3
1
0
dL
5
8
1
dH
Fig. 1. The advantage of price discrimination. The patterns of the regions in Fig. 1a and b represent which strategy is most advantageous for the retailer. In the shaded regions, MD is the
most profitable strategy. In the polka-dotted regions, PS is the most profitable strategy. In the unshaded regions, ND is the most profitable strategy. In the regions labeled A, B1, B2, and C,
both PS and MD are more profitable than ND. In region A, PS generates more revenue than MD from both the L-types and the H-Types. In region C, PS generates less revenue from both
consumer types. In regions B1 and B2, PS generates higher revenue from the L-types but lower revenue from the H-types. The net effect is that PS is more profitable than MD in region B1,
but less profitable in region B2. Note: The maximum value of vLU in Fig. 1a is ¾ since, by definition, vLU b vLF = ¾.
less valuable to them than their preferred product in the first period and
the seller charges a price that is just low enough to induce them to make
this purchase. Thus, the difference in revenue per L-type customer
between MD and PS equals the difference between how much such a
customer values a probabilistic good in Period 1 and how much she
values her preferred good in Period 2. The relative advantage of PS is
greater the weaker the SOPs of low-value consumers and the less
patient they are (Proposition 1(c)). On the other hand, PS is unable to
create an advantage (relative to MD) in Value to L-Types if low-value
customers are patient but “picky” about which product they consume.
We now turn to the Surplus to H-types. Under both MD and PS, the
seller must set its first-period price for the specified goods such that the
high-value consumers will purchase their preferred products in Period
1. Specifically, the seller must reduce the price below these consumers'
willingness-to-pay so they receive at least as much surplus as they
would receive if they purchased the discounted product. This maximum
obtainable price depends critically upon how much the H-type
consumers value the alternate purchase offering and also on the price
of this alternate purchase option. Specifically, as indicated in
Proposition 1(c), the PS advantage increases as the SOP of low-value
consumers weakens (so that P0PS is higher), as low-value consumers
become less patient (so that P2MD is lower), as the SOP of high-value
consumers become stronger (so that purchasing the probabilistic good
would yield less surplus), and as the latter become more patient (so
that waiting to purchase until the second period would yield more
surplus). On the other hand, PS fails to generate an advantage in Surplus
to H-types if low-value customers are patient and picky, while highvalue customers are impatient and are not picky. In short, whether or
not PS generates an advantage in Surplus to H-types depends on
whether it is easier for the seller to prevent high-value consumers
from opting for the (discounted) probabilistic good or to prevent
them from waiting until the second period to purchase their preferred
products.
As shown in the equation in Proposition 1(b), PS is preferable to MD
only if the sum of these two effects is positive.13 If PS creates a more
valuable “damaged” product for low-value consumers and a lower (or
no change in) required surplus for high-value consumers, then PS is
clearly more profitable than MD. This occurs in the region labeled “A”
in Fig. 1b. On the other hand, neither advantage is present in region
13
Note, we assume that the low- and the high-value segments are of equal size. With
asymmetric segment sizes, these two effects would need to be weighted according to each
segment's size.
“C” of Fig. 1a. Here, MD is more profitable than PS. If the two effects go
in different directions, the larger effect determines which pricing
discrimination mechanism is more profitable. Thus, even if PS has one
disadvantage, this strategy can still be more profitable than MD if its
advantage is greater than its disadvantage. In the regions labeled “B1,”
PS creates a sufficiently large Value-to-L-types effect to offset the
negative Surplus-to-H-types effect. However, in the regions labeled
“B2,” the negative Surplus-to-H-types effect is of a greater magnitude,
and thus MD is more profitable than PS.
3.4. Summary
In sum, for a price discrimination mechanism (e.g., PS or MD) to
improve profit, it must create a new purchase option that is attractive
to low-value consumers (thus enabling the firm to earn significant
revenue from the new purchase option), but is not attractive to highvalue consumers (so that high margins can be maintained for the
original products). Whether varying prices over time (via the MD
strategy) or creating a probabilistic product (via the PS strategy) is
more advantageous, depends crucially upon the heterogeneity in
consumers' degree of patience and the strength of their preferences.
Thus, in some markets we would expect PS to be more profitable than
MD and in other markets for the reverse to be true (depending on
whether or not the condition from Proposition 1(b) holds).
4. Model extension: Demand uncertainty
Since retailers often cannot predict which products will be more
popular, inventory is depleted asymmetrically, with popular products
selling faster and unpopular products selling more slowly than
expected. Unpopular items that remain after the primary selling season
are often severely marked down. In this section, we incorporate demand
uncertainty into the base model from Section 3 to examine whether PS
remains a viable strategy (relative to MD) in such market settings and to
garner additional insights into how demand uncertainty impacts the
tradeoff between the two strategies. Specifically, at the time inventory
orders are made and when first-period prices are chosen, the seller
does not know whether product A or product B will be the more popular
good. We assume that the seller knows the value of α, i.e., that one good
will be more popular than the other, but not whether α will apply to
product A or product B. Instead, the seller believes each product is
equally likely to be the popular one. In period 2, the seller learns
which product is popular and can adjust its prices accordingly.
153
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
In this section, we derive the optimal inventory orders that occur
under MD and PS and make the simplifying assumptions that c b Min
[dL vLF, (vLF + vLU)/2], dH = dL/2, and vLU = vHU = 0. These conditions
guarantee that MD and PS are advantageous in the absence of demand
uncertainty, i.e., the conditions given in Corollary 1 and Corollary 2
both hold, and that the L-type's value of both the probabilistic good
under PS and the delayed product under MD exceed the cost of
acquiring a unit of a product in order to meet this demand. Lemma 2
gives the optimal solution under PS when demand uncertainty is
present:
Lemma 2. (PS with demand uncertainty)
If dH = dL/2, vLU = vHU = 0 and c b vLF/2, in the presence of demand
uncertainty
(a) The firm will purchase 2αþ1
2 units of both product A and product B.
(b) The resulting profit is ΠPS;DU ¼ vLF þ 12 −ð2α þ 1Þc:
Lemma 2 shows that, if costs are sufficiently low, the inventory order
under PS will enable the firm to sell to all customers (of both types).
Demand uncertainty does not impact the prices that the firm sets or
the amount of sales. However, the firm must order 2αþ1
units of both
2
product A and product B, whereas without demand uncertainty,
the#
" 2αþ1
3−2α
firm would order units of the 2αþ1
popular
good
and
b
2
2
2
units of the unpopular good. Thus, demand uncertainty leads to excess
inventory for the unpopular good and reduces profit.
Proposition 2 compares profit and inventory utilization under PS to
that which occurs under MD when demand uncertainty is present.
Proposition 2. (Demand uncertainty with costly inventory)
Probabilistic selling addresses demand uncertainty more efficiently than
does markdown selling. Specifically,
"
+
! ,
#
(a) For sufficiently low costs cbMax α; 2 3 dL vLF , demand uncertainty
reduces the profit under markdown selling more than under
probabilistic selling.
(b) Relative to markdown selling, probabilistic selling achieves a higher
level of inventory utilization and reduced wastage for all cb dL 2νLF .
(c) Total sales are (weakly) higher under probabilistic selling than under
markdown selling.
In order to understand Proposition 2, it is instructive to examine
how products with asymmetric demand are allocated in the absence
of demand uncertainty. Under MD, H-type consumers purchase their
preferred good (where α prefer the popular good) in the first period
and L-type consumers purchase their preferred good (also where α
prefer the popular good) in the second period. Thus, across the two
selling periods, 2α units of the popular good and 2(1 − α) units of the
unpopular good are sold. Under PS, in the first period, H-type consumers
purchase their preferred good (where α prefers the popular good) and
L-type consumers will purchase the probabilistic good (where a fraction
½ will be assigned to the popular good). Thus, in total, α þ 12 ¼ 1þ2α
2 units
of the popular good and 1−α þ 12 ¼ 3−2α
units
of
the
unpopular
good
are
2
consumed. In the absence of demand uncertainty, under both PS and
MD, the firm orders inventory to exactly meet the demand for each
product, which involves purchasing more units of the popular product.
It is critical to note that total consumption of the popular good and the
difference in consumption between the popular and the unpopular
good are both higher under MD than under PS.
When demand uncertainty exists, the firm does not know which
product will be more popular and thus inventory orders cannot be
tied to a product's popularity. Mathematically, this is equivalent to
imposing a constraint, KsA = KsB, s = {ND, MD, PS} on the maximization
problems in Section 3. It is intuitive that, as Proposition 2(a) states,
this constraint will have a larger detrimental impact on MD since,
without such a restriction, the inventory orders would be more
asymmetric under MD than under PS.
Under PS, as long as costs are not too large, Lemma 2 shows that the
firm's best response to demand uncertainty is to order K PS;DU
¼ K PS;DU
¼
B
A
2αþ1
units.
Some
units
of
the
unpopular
good
will
go
unsold.
Thus,
even
2
though the firm generates the same revenue as in the absence of
demand uncertainty, higher costs are incurred (due to a larger total
amount of inventory).
Under
MD,% one of two options will be optimal. First, if costs are very
$
low c≤ dL2νLF , the firm will order KMD,DU
= KMD,DU
= 2α units so that
A
B
sales are the same as under demand uncertainty. Since the firm sells
to all consumers and at the same prices as when there was no demand
uncertainty, the latter has no effect on the revenue earned under MD.
However, costs increase due to the higher inventory order, thus leading
to unsold units remaining and lower profit. Fig. 2a graphs the number of
unsold units under MD and shows the magnitude of the profit decrease
that demand uncertainty inflicts on the MD strategy (ΔMD,DU) as a
function of demand asymmetry (α). Notice that, as α increases, the
seller faces more severe demand uncertainty. Two results are apparent:
(1) MD results in more unsold units than does PS (since inventory
orders are higher under MD); and (2) demand uncertainty reduces
b) Moderate
a) Low
2.0
Sales (PS)
2.0
Unsold (MD)
1.5
1.5
Unsold (PS)
1.0
Sales (MD)
1.0
MD, DU
0.5
0.5
MD, DU
PS, DU
PS, DU
0.6
0.7
0.8
0.9
1.0
0.6
0.7
0.8
0.9
1.0
Fig. 2. Impact of demand uncertainty on PS and MD. The graphs are drawn assuming vLF = dL = 3/4. For Fig. 2a, which is drawn for c = .2, total sales equals 2 for both MD and PS. Fig. 2b is
drawn assuming c = .3.
154
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
profit more for MD than for PS. As demand uncertainty grows, the
magnitude of both of these effects
increases.
$
%
Second, if costs are higher cN dL2νLF , the retailer will order less than
2α units of each good. This scenario is illustrated in Fig. 2b. While
ordering less inventory reduces the number of unsold units, this also
creates stockouts, as some or all of the L-type consumers would be
unable to purchase in the second period. As stated by Proposition 2(c),
total sales would be higher under PS since, in this case, all consumers
receive one unit of one product, whereas under MD some consumers
do not purchase anything. This reduction in revenue makes the profit
decrease from demand uncertainty larger under MD than under PS.
It is important to recognize that, whichever of these two options is
adopted under MD, PS offers the seller a more effective tool for
addressing demand uncertainty. Relative to PS, MD either requires a
larger investment in inventory (as in option one) or fewer total sales
(as in option two). In sum, not only does PS remain a viable strategy
for retailers who face demand uncertainty, the relative advantage of PS
over MD becomes even stronger when demand uncertainty is present.
5. Concluding comments
5.1. The benefits of probabilistic selling
Markdowns are commonplace in retailing and can result in painful
repercussions, such as very low (or even negative) margins and a
consumer mentality of delaying purchases in anticipation of a big sale.
In this paper, we show that probabilistic selling (PS) can be preferred
over a typical markdown (MD) strategy. In particular, we illustrate that
introducing probabilistic products can be beneficial because they (1)
allow the seller to obtain higher prices from customers with strong
preferences for one product over alternatives (because probabilistic
products may present less of a cannibalization threat to full-price sales
than do traditional time-dependent markdowns); (2) reduce the size of
necessary discounts (because consumers may be willing to pay more
for a randomly selected product at the beginning of the season than for
their preferred product late in the season); and (3) more effectively
address the negative consequences of demand uncertainty and limited
capacity by helping a firm prevent stockouts and reduce excess inventory.
We find that, while both probabilistic products and markdowns can be
used to segment consumers, PS is most effective when low-value
consumers have weak preferences and high-value consumers have
strong preferences. In contrast, markdown pricing relies on specific
differences in consumers' willingness to delay purchase until a later
date. The markdown pricing strategy is most effective when low-value
consumers are relatively patient and high-value consumers are relatively
impatient. Because the MD and PS strategies rely on distinctly different
mechanisms to segment the market, PS is a valuable additional tool for
retailers, especially those who operate in markets where markdowns
are ineffective (e.g., non-fashion products or products unaffected by
seasonality so that consumer arrival times are not negatively correlated
with product valuations or price sensitivity).
5.2. Limitations and future research
While markdown pricing strategies are ubiquitous, the PS strategy is
still nascent. More research is needed to fully flesh out how to optimally
implement this selling strategy. For example, it is unclear how PS
impacts which products (and how many) a seller should stock. Other
research has considered a version of a probabilistic good (which they
call a “flexible good”) in which the seller does not assign products to
buyers until after demand uncertainty is resolved (e.g., Gallego &
Phillips, 2004; Mang et al., 2012; Post, 2010). As refinements to PS
are made, such as determining the optimal product mix and the
optimal timing of product assignments, the profit obtainable from
the strategy should increase and, thus, the market settings in which
PS is advantageous to traditional markdowns should be enlarged beyond
the regions identified in this paper, which utilizes a base case (and a
rather conservative) definition of PS.
The paper employs a stylized model and thus has several limitations.
First, in our model, we assume no inventory holding costs and that firstand second-period profits are equally weighted. If holding costs and
discounting of future cash flow were incorporated, PS would have the
additional benefit over MD of shifting sales forward in time. This
advantage would be especially important to sellers who desire rapid
turnover or incur significant inventory holding costs. Second, there
may be circumstances in practice in which it is beneficial to offer both
temporal-based discounts and discounted probabilistic products.
Third, our stylized model leads to the firm charging the same price for
each product (both in the first and second periods). In addition, we
assume that inventory orders can only be placed prior to the selling
season. We believe that our assumptions are quite realistic for many
retailers since there are many cases in retailing where markdowns are
symmetric and mid-period inventory acquisition is not possible.
However, in practice, some retailers do offer different-sized markdowns
as they better learn the demand for each particular product and others
are able to replenish inventory throughout the selling season. Relaxing
these (and other) assumptions of the model could be a valuable
direction for future research.
Additional research questions that remain unresolved include: What
are the ramifications of PS on suppliers (e.g., their market power and
effects on supply chain dynamics)? Can PS be effective when the seller
is uncertain about total category demand rather than the relative
popularity of each specific item? Do some consumers prefer the
probabilistic good because (rather than in spite of the fact that) it offers
a gamble? Does PS alter consumer preferences (e.g., dilute brand- or
store loyalty)?
Web Appendix
Supplementary data to this article can be found online at http://dx.
doi.org/10.1016/j.ijresmar.2013.08.006.
References
Anderson, C. K. (2009). Setting prices on priceline. Interfaces, 39(4), 307–315.
Anderson, C. K., & Xie, X. (2012). A choice-based dynamic programming approach for
setting opaque prices. Production and Operations Management, 21(3), 590–605.
Besbes, O., & Lobel, I. (2012). Intertemporal price discrimination: Structure and computation
of optimal policies. Columbia Business School Research Paper No. 12/46 (August 7,
2012, downloaded Dec. 4, 2012 from http://papers.ssrn.com/sol3/papers.cfm?
abstract_id=2126312)
Bitran, G. R., & Mondschein, S. V. (1997). Periodic pricing of seasonal products in retailing.
Management Science, 43(1), 64–79.
Chung, C. -S., Flynn, J., & Zhu, J. (2009). The newsvendor problem with an in-season price
adjustment. European Journal of Operational Research, 198(1), 148–156.
Coughlan, A., & Soberman, D. (2005). Strategic segmentation using outlet malls. International
Journal of Research in Marketing, 22(1), 61–86.
Fay, S. (2008). Selling an opaque product through an intermediary: The case of disguising
one's product. Journal of Retailing, 84(1), 59–75.
Fay, S., & Xie, J. (2008). Probabilistic goods: A creative way of selling products and
services. Marketing Science, 27(4), 674–690.
Fay, S., & Xie, J. (2010). The economics of buyer uncertainty: Advance selling vs.
probabilistic selling. Marketing Science, 29(6), 1040–1057.
Fay, S., & Xie, J. (2012). Timing of commitment as a strategic variable: Using probabilistic
selling to enhance inventory. Working Paper.
Friend, S.C., & Walker, P. H. (2001). Welcome to the new world of merchandising. Harvard
Business Review, 79(10), 133–141.
Gallego, Guillermo, & Phillips, Robert (2004). Revenue management of flexible products.
Manufacturing and service operations management, 6(4), 321–337.
Granados, N., Gupta, A., & Kauffman, R. J. (2008). Designing online selling mechanisms:
Transparency levels and prices. Decision Support Systems, 45(4), 729–745.
Jerath, K., Netessine, S., & Veeraraghavan, S. K. (2010). Revenue management with
strategic customers: Last-minute selling and opaque selling. Management Science,
56(3), 430–448.
Jiang, Y. (2007). Price discrimination with opaque products. Journal of Revenue and Pricing
Management, 6(2), 118–134.
Khouja, M. (1995). The newsboy problem under progressive multiple discounts. European
Journal of Operational Research, 84(2), 458–466.
Lazear, E. P. (1986). Retail Pricing and Clearance Sales. American Economic Review, 76(1),
14–32.
D.H. Rice et al. / Intern. J. of Research in Marketing 31 (2014) 147–155
Levy, M., Grewal, D., Kopalle, P. K., & Hess, J.D. (2004). Emerging trends in retail pricing
practice: Implications for research. Journal of Retailing, 80(3), xiii–xxi.
Levy, M., & Weitz, B.A. (2004). Retailing management (5th ed.)Boston: McGraw-Hill Irwin.
Mang, S., Post, D., & Spann, M. (2012). Pricing of flexible products. Review of Managerial
Science, 6(4), 361–374.
Mantrala, M. K., & Rao, S. (2001). A decision-support system that helps retailers decide
order quantities and markdowns for fashion goods. Interfaces, 31(3), S146–S165.
Nair, H. (2007). Intertemporal price discrimination with forward-looking consumers:
Application to the US market for console video-games. Quantitative Marketing and
Economics, 5(3), 239–292.
Petrick, A., Steinhardt, C., Gonsch, J., & Klein, R. (2012). Using flexible products to
cope with demand uncertainty in revenue management. OR Spectrum, 34(1),
215–242.
Petruzzi, N. C., & Monahan, G. E. (2003). Managing fashion goods inventories:
Dynamic recourse for retailers with outlet stores. IIE Transactions, 35(11),
1033–1047.
155
Post, D. (2010). Variable opaque products in the airline industry: A tool to fill the gaps and
increase revenues. Journal of Revenue and Pricing Management, 9(4), 292–299.
Ross, J. R. (1997). Inventory management systems cut costs while keeping store shelves
full. Stores, 79(7), 78.
Shapiro, D., & Zillante, A. (2009). Naming your own price mechanisms: Revenue gain or
drain? Journal of Economic Behavior and Organization, 72, 725–737.
Su, X. (2007). Intertemporal pricing with strategic customer behavior. Management
Science, 53(5), 726–741.
Sullivan, L. (2005). Fine-tuned pricing. Information Week, 1052, 36–41 (Aug 15-Aug 22).
Wang, T., Gal-Or, E., & Chatterjee, R. (2009). The name-your-own-price channel in
the travel industry: An analytical exploration. Management Science, 55(6), 968–979.
Wood, C. M., Alford, B.L., Jackson, R. W., & Gilley, O. W. (2005). Can retailers get higher
prices for “end-of-life” inventory through online auctions? Journal of Retailing, 81(3),
181–190.
Zouaoui, F., & Rao, B. V. (2009). Dynamic pricing of opaque airline tickets. Journal of
Revenue and Pricing Management, 8(2/3), 148–154.
Intern. J. of Research in Marketing 31 (2014) 156–167
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
Full Length Article
Does retailer CSR enhance behavioral loyalty? A case for
benefit segmentation
Kusum L. Ailawadi ⁎, Scott A. Neslin, Y. Jackie Luan, Gail Ayala Taylor
Tuck School of Business, Dartmouth College, 100 Tuck Hall, Hanover, NH 03755, United States
a r t i c l e
i n f o
Article history:
First received in 8 November 2011 and was
under review for 10 months
Available online 16 October 2013
Area Editor: Els Gijsbrechts
Keywords:
Corporate social responsibility (CSR)
Dimensions of CSR
Retail store patronage
Attitude towards store
Loyalty
Share of wallet
a b s t r a c t
We study the effects of consumer perceptions of four types of corporate social responsibility (CSR) activities on
their behavioral loyalty toward retailers. The four activities are environmental friendliness, community support,
selling locally produced products, and treating employees fairly. Behavioral loyalty is measured by share-ofwallet (SOW). We control for other retailer attributes that drive attitudes and SOW, and examine how the market
is segmented in terms of consumer response. We partition the total effect of CSR on SOW into a direct effect and
an indirect effect mediated through attitude towards the store. These effects differ by CSR activity and customer
segment. The effects on attitude are positive and positive attitude enhances SOW, so the indirect effects on SOW
are positive. While we generally find positive total effects, the total effect of one of the CSR activities,
environmental friendliness, is significantly negative for one group of consumers. The magnitude of CSR's total
impact on SOW is not only statistically significant but also managerially meaningful in an industry where
every share point carries a substantial dollar amount. We characterize the customer segments and conclude
with implications for how best a retailer can manage its CSR initiatives.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
Corporate social responsibility (CSR) refers to a firm's moral, ethical
and social obligations beyond its own economic interests (Brown &
Dacin, 1997; McWilliams & Siegel, 2001). As CSR gains strategic
importance in the eyes of senior management, companies are engaging
in a wide range of CSR programs including environmental sustainability,
community support, cause-related marketing, and employee enablement.
They are investing significantly in publicizing their CSR initiatives in
the hope of strengthening relationships with employees, customers,
investors, and the broader community. But, as noted by Luo and
Bhattacharya (2009) and others, CSR programs compete for resources
that can alternatively be channeled to other areas such as innovation or
service improvement. Not surprisingly, both academics and practitioners
want to determine the returns to CSR investments. The purpose of this
paper is to investigate CSR returns by examining the impact on behavioral
loyalty, focusing on the retail grocery industry.
Prior research has assessed returns to CSR efforts by examining
financial performance. Despite a large body of empirical research, the
jury is still out regarding this question. Most studies use the Kinder,
Lydenburg, Domini (KLD) index of corporate social performance to
quantify CSR efforts. The majority of these studies show a positive effect
and recent work suggests that CSR reduces firm-specific risk (Luo &
Bhattacharya, 2009). But some researchers report a substantial number
of insignificant and even negative effects, and methodological and
⁎ Corresponding author. Tel.: +1 603 646 2845; fax: +1 603 646 1308.
E-mail address: kusum.ailawadi@dartmouth.edu (K.L. Ailawadi).
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.09.003
theoretical criticisms of the studies abound (see Margolis & Walsh,
2003 and Orlitzky, Schmidt, & Rynes, 2003 for reviews). These mixed
results are attributable in part to the fact that CSR has multiple
dimensions whose impact varies across industries, stakeholder groups,
and individuals within a stakeholder group (Berman, Wicks, Kotha, &
Jones, 1999; Hillman & Keim, 2001; Sen & Bhattacharya, 2001).
Godfrey and Hatch (2007) and Raghubir, Roberts, Lemon, and Winer
(2010) note that there is a need to conduct industry-specific studies
and to distinguish between different dimensions of CSR.
One of the firm's most relevant stakeholders is its customers. Social
identity theory and consumer–company identification research suggest
that consumers should embrace the more positive and distinctive identity
of a company that engages in CSR (e.g., Bhattacharya & Sen, 2003; Sen &
Bhattacharya, 2001). Thus, customers should reward such companies
with greater loyalty, ultimately enhancing the firm's financial value. But,
research on how customers respond to CSR efforts is more limited.
Consumer polls paint a rosy picture for CSR initiatives, but they suffer
from social desirability bias and other validity concerns (see Auger,
Burke, Devinney, & Louviere, 2003 and Cotte & Trudel, 2009 for critiques
of these polls). Academic work shows that, by and large, consumers
exhibit more favorable attitudes towards socially responsible companies
(e.g., Du, Bhattacharya, & Sen, 2007; Klein & Dawar, 2004; Lichtenstein,
Drumwright, & Braig, 2004; Luo & Bhattacharya, 2006) but there is
considerable heterogeneity in response (e.g., Barone, Miyazaki, &
Taylor, 2000; Bhattacharya & Sen, 2004; Brown & Dacin, 1997; Sen &
Bhattacharya, 2001).
Importantly, it is not clear whether these positive effects translate
into behavioral loyalty, for example in the form of share of wallet
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
(SOW). Previous research is largely based on laboratory experiments
and measures attitudes and intentions rather than actual behavior.
Subjects are typically presented with a description of a company's CSR
record and then asked about their attitudes and/or purchase likelihood.
Given the salience of the CSR information in the experiment, its impact
may be overstated compared to the real-life purchase environment
in which several other factors – product quality, price, assortment,
convenience, etc. – influence choice. Bhattacharya and Sen (2004)
note that, while CSR initiatives produce positive company attitudes,
this may not translate into greater purchase behavior because
consumers are reluctant to trade off CSR for core attributes such as
price. This suggests that attitudes may mediate the impact of CSR
activities on behavioral loyalty, but CSR activities may have direct
effects as well. In addition, the limited external validity of this body of
work has led researchers like Sen and Bhattacharya (2001) and Du
et al. (2007) to call for more research based on data collected in actual
marketing environments and in the context of competitive offerings.
Thus, prior research reveals the need to: (1) distinguish between
different dimensions of CSR; (2) study the response of specific
stakeholder groups in individual industries; (3) link consumers' CSR
perceptions to their behavioral loyalty in addition to attitude; (4) control
for other core firm attributes from which consumers derive utility;
(5) examine heterogeneity in CSR response across individuals; and
(6) study real-world data.
To address this need, we study the effects of key CSR activities in the
grocery retail industry on behavioral loyalty. We survey consumers in a
geographical market to measure their perceptions of CSR and other
attributes, as well as overall attitude, with respect to all major grocery
retailers in that market, and measure their behavioral loyalty to these
retailers.
We use these data to specify and estimate a model of behavioral
loyalty that allows for attitudes to mediate the impact of CSR, and for
heterogeneity in consumer response. We examine four CSR activities:
environmental friendliness, community support, selling local products,
and treating employees fairly. In sum, we (1) measure the effects of
CSR on behavioral loyalty in a field setting while controlling for other
drivers of consumer preferences, (2) allow attitudes to mediate these
effects; (3) show how these effects differ across key CSR dimensions
in an industry that represents a major sector of the economy
(U.S. sales of $580billion in 2010); and (4) investigate how the response
to CSR dimensions varies across consumers.
The rest of our paper is organized as follows. We first develop our
conceptual framework and describe the data used for our analysis. This
is followed by a presentation of our results and we conclude our paper
with a discussion of the implications for researchers and managers.
2. Conceptual development
Fig. 1 depicts our conceptual model. It allows consumers' perceptions
of CSR to influence behavioral loyalty (measured as share of wallet)
through overall attitude as well as directly, while incorporating the
impact of other retailer attributes that the literature identifies as
important influencers of store patronage. We discuss each major
element of the framework below, moving from left to right in the figure.
2.1. The dimensions of CSR
The literature generally follows the KLD classification of CSR into six
dimensions – employee support, diversity, community support,
environment, products, and non-U.S. operations. Bhattacharya and
Sen (2004) propose that consumers may respond more positively to
CSR initiatives that directly affect their experience with the firm.
Bhattacharya, Sen, and Korschun (2008) also note that stakeholders'
response depends upon the benefits they themselves derive from the
CSR activities. Related to this is the notion that pro-social behavior is
motivated by both selfish and selfless altruism, where the ultimate goal
157
of the former is self-benefit with helping others being an instrumental
goal, while the ultimate goal of the latter is helping others with selfbenefit as an unintended consequence (Batson & Shaw, 1991; Krishna,
2011).
Consumer response is also expected to be more positive for
initiatives that are integrated into the core positioning of the firm/
brand (Du et al., 2007), as long as this does not generate negative
perceptions regarding the firm's motives (Barone, Norman, & Miyazaki,
2007). This suggests that dimensions of CSR that only contribute to
broad social good and that are less integrated with a retailer's core
offering (e.g., those related to the environment or community) should
have a less positive effect on consumer loyalty. In contrast, CSR
dimensions that provide both societal and personal benefit and are
integrated into a retailer's core offering (e.g., those related to the product
or service experience) should have a more positive effect.
We examine four CSR activities that are relevant in our empirical
context: environmental friendliness, community support, selling local
products, and treating employees fairly. The last two relate directly to
the customer's shopping experience, while the first two do not. While
the four CSR activities can be grouped into customer-experience versus
non-customer-experience, all four are quite different. We therefore
examine them separately; the results will reveal whether they exert
similar or different effects.
2.2. Other retailer attributes
Although our focus is on the effect of CSR, we must control for other
retailer attributes that affect loyalty and may be correlated with CSR,
especially given previous findings regarding consumers' unwillingness
to trade off other attributes for CSR (Barone et al., 2000; Luo &
Bhattacharya, 2006; Sen & Bhattacharya, 2001). A review of the
literature shows that the drivers of retail store image and patronage
can be categorized into a few key attributes – price, assortment, product
quality, deals, in-store service and social experience, and convenience of
location (Ailawadi & Keller, 2004; Baker, Parasuraman, Grewal, & Voss,
2002; Lindquist, 1974; Mazursky & Jacoby, 1986; Verhoef, Neslin, &
Vroomen, 2007). We include these in our model.
2.3. Mediation model
Consumer perceptions of CSR and other store attributes can affect
behavioral loyalty directly or indirectly through overall attitude towards
the store. The indirect route is supported by models of consumer
decision-making such as the theory of reasoned action (Fishbein &
Ajzen, 1975) later broadened into the theory of planned behavior
(Ajzen, 1991).
However, attitudes may not fully mediate the impact of perceptions
on behavior. Perceptions of a store's CSR activities may influence
behavior not just because of what CSR says about the store (as would
be measured by overall store attitude), but also because of what it says
about oneself (e.g., the social identity literature cited earlier). Also, social
scientists have identified an “automaticity” effect, whereby behavior
may be induced by cues in the environment without a conscious thought
process (Bargh, Chen, & Burrows, 1996). The social atmosphere in the
store or CSR activities might serve as such cues, evoking a
categorization/stereotype that compels the consumer to shop at or
avoid a store (Bitner, 1992). Similarly, a convenient location or special
deals may directly cause a consumer to shop at that store, without the
elaborate thought process assumed by the formation of overall attitude.
In summary, the total effect of CSR and other store attributes on
behavioral loyalty comprised an indirect effect (mediated by overall
attitude toward the store) and a direct effect. Since prior research
shows a positive effect of CSR on attitude and attitude is positively
correlated with behavior, albeit weakly, we expect a positive indirect
effect. However, the direct effect may well be negative if, as some
have suggested, the total effect of CSR on behavior is not positive. The
158
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
Consumer Heterogeneity
CSR-Corporate Ability Belief
CSR-Costs Belief
Demographics
Grocery Budget
Retailer CSR
Environmental Friendliness
Community Support
Local Products
Employee Fairness
Attitude Towards the Store
Behavioral Loyalty
(Share of Wallet)
Other Retailer Attributes
Assortment selection
Product quality
Price
Deals
In-store service
In-store social environment
Location convenience
Fig. 1. The impact of corporate social responsibility on behavioral loyalty.
magnitude of the total effect and its decomposition into the direct and
indirect components is an empirical question that we will investigate.
2.4. Heterogeneity in consumer response
As noted previously, researchers have found considerable variation in
consumers' response to CSR. This may be due to how much consumers
personally believe in the activity (Sen & Bhattacharya, 2001), whether
they believe that CSR impinges on a company's corporate abilities
(Brown & Dacin, 1997), and how much importance they place on other
aspects of the company's core offering, such as price and service
(Bhattacharya & Sen, 2004). Also, research has shown that consumers
vary in the value they place on other store attributes, e.g., how much
they are willing to engage in price search (e.g., Talukdar, Gauri, &
Grewal, 2010; Urbany, Dickson, & Kalapurakal, 1996). Thus, our
conceptual model allows for heterogeneity in the response to CSR as
well as the other store attributes.
3. Method
3.1. Sample
Our data come primarily from a survey administered to customers of
a retail grocery chain located in the northeastern U.S. This “focal”
retailer positions itself strongly as a socially responsible retailer. With
the retailer's cooperation, we mailed a letter to its approximately
16,000 active loyalty program members (i.e., those who made at least
one purchase at the retailer in the previous 6 months) inviting them
to participate in the survey that could be completed online or on
paper. Paper copies were made available and collected at all of the
retailer's stores. The purpose of the survey was introduced in general
terms (“to better understand and serve the needs of customers”)
without mentioning CSR or any other specific area. It was made clear
that the project was being conducted by a team of academic researchers
at a nearby university. A lottery of ten gift certificates worth $100 each,
redeemable at area businesses, was used to encourage participation.
In total, 2884 responses were obtained during the 1-month period
when the survey was live, representing a response rate of about 18%.
Note that the sampling frame consists of loyalty program members.
However, 77% of the total sales by the focal retailer are to members of
its program so this is a highly relevant sampling frame for studying
the focal firm's customers.
3.2. Measures
The survey comprised four main sections. The first section collected
information on the respondent's share of wallet (hereafter SOW),
measured as percentage of total grocery spending in the past 6 months
with the focal retailer as well as the seven major competing retailers in
the area. We also allowed the respondent to indicate “other stores” not
listed in the survey. The median (mean) SOW for these other stores is 0%
(9.3%), indicating that the eight retailers included in our study account
for most of the respondents' grocery spending.
The second section asked for respondents' perceptions of the focal
retailer on the key attributes identified in the retailing and store image
literature (such as product quality, price, and in-store service) and the
four CSR dimensions identified earlier, as well as their overall attitude
towards the retailer. Items for all constructs used a five-point scale, and
the ordering of the items was randomized across respondents.
The third section asked for respondents' perceptions of a second
store on the same items. In the online version of the survey, the second
store was randomly generated among the competing stores that
received at least 10% SOW from the respondent (this section was
skipped if no competing stores receive more than 10% SOW). This
ensured that the respondent had some familiarity with the second
store being evaluated. In the paper version of the survey, the identity
of the second store was randomized across multiple versions of the
questionnaire, and the respondent was instructed to skip this section
if he or she was unfamiliar with the particular store. The last section of
the survey gathered self-reported importance of various retailer
attributes and standard demographic and psychographic information.
In order to ensure variation not just across but within respondents
we retained only respondents who rated two stores. After responses
159
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
with missing data were discarded, our final sample consisted of 3492
observations from 1746 respondents. To assess how representative
our sample is of the sampling frame, we compared it to the focal
retailer's population of active loyalty program members. Table 1
summarizes this comparison and shows that our respondents are not
very different from the population of loyalty program members in
terms of duration of program membership and total spending and
number of trips during the 6 months preceding the study.
Table 2 presents measures of each variable, along with descriptive
statistics and, for multi-item variables, reliabilities.1 As noted previously,
we selected the store attributes based on the retailing and store image
literature. We fine-tuned our selection based on qualitative interviews
with four managers from the focal retailer and a convenience sample of
fifteen consumers who were familiar with most of the stores in our
study. The interviews led us to include not just the size of a retailer's
assortment but the extent to which the retailer offers unique items not
available elsewhere. They also led us to measure the social experience/
clientele aspects of the store through two separate attributes, i.e., how
much a consumer feels they have in common with the clientele and
how wealthy they perceive the clientele to be.
3.3. Data quality
SOW is of central interest so we first compare the self-reported SOW
with actual spending compiled from the focal retailer's customer
database.2 The correlation between respondents' self-reported SOW at
the focal retailer and their actual spending there over 6 months prior
to survey administration is 0.61. We also computed respondents' SOW
at the focal retailer from their actual spending and the weekly total
grocery budget they reported in the survey. The correlation between
computed and self-reported SOW is 0.71. These correlations are much
larger than typically reported between perceived and objective measures
(e.g., Bommer, Johnson, Rich, Podsakoff, & Mackenzie, 1995), and suggest
that self-reported SOW has strong convergent validity.
Table 3 provides descriptive statistics of attribute ratings across
retailers. The table shows substantial variation in mean ratings both
within and across retailers. As the focal store positions itself on CSR
and communicates its CSR activities via the quarterly newsletter to
program members, its website, and in-store signage, it is not surprising
that it rates highest on these dimensions. It also stands out in carrying
unique items, product quality, in-store service, and assortment.
However, it is rated poorly on price and promotions, and is perceived
as having a wealthy clientele. Thus, consumers don't uniformly rate it
positively or negatively on all attributes, alleviating concerns about
social desirability and halo effects. Ratings of other retailers also show
substantial variation across attributes and they generally have face
validity. For example, retailers F and G position themselves strongly
on price and consumers' mean price perceptions are in line with this.
Similarly, retailers D and F are discount/big box stores and consumers'
perceptions of these stores as offering attractive prices but being low
on assortment and in-store service are in accordance with this. Finally,
retailer H, who receives the poorest ratings, was struggling and closed
its store in the year following our study. Overall, the pattern suggests
that respondents rated the stores realistically and alleviates concerns
about halo effects.
1
We adapted existing measures where possible (e.g., Baker et al., 2002; the GfK
consumer surveys used by van Heerde, Gijsbrechts, & Pauwels, 2008; Verhoef et al.,
2007) and developed and pretested others. While multi-item measures may have been
desirable for all constructs (but see Bergkvist and Rossiter (2007) for findings to the
contrary), survey length was an issue given the number of constructs and the need for
respondents to rate two retailers. Therefore, we used single items for some variables.
2
The retailer estimates that well over 90% of members' purchases are captured in their
database.
Table 1
Comparison of sample with member population.
Variable
Sample mean
(std. deviation)
Population mean#
(std. deviation)
Total spending in last 6 months ($)
1549
(1597)
38
(42)
145
(118)
1631
(1976)
35
(44)
147
(121)
Number of trips in last 6 months
Number of months as member
#
The population is the full set of active loyalty program members in the focal retailer's
database.
As an additional check for halo effects, we examined the percentage
of observations that showed high (4 or 5 on the 5-point scale) ratings
for all attributes, and also what percentage of respondents who rated a
retailer high on CSR also rated the retailer high on other attributes. We
found that the observations with highly positive ratings on all attributes
comprise less than 1% of the sample. Further, in terms of CSR versus other
attributes, only a minority rated a given store high on both, depending on
the CSR and other attribute involved. For example, less than 50% rated a
retailer high on both environmental friendliness and quality. Only 5%
rated a retailer high on both environmental friendliness and price.
Finally, we interviewed five grocery retail experts in the area and
asked them to rate the stores (excluding their own) on the key
attributes in our study. Overall, their ratings are consistent with those
of our sample. For example, all of them rated the focal retailer highest
on CSR attributes, assortment, quality, unique items, and wealthy
clientele and worst on price and promotions, although the difference
in ratings of the focal retailer and the next best was very small on
assortment and quality.
In summary, the SOW measure exhibits strong convergent validity,
the mean attribute ratings have good face validity as well as discriminant
validity, and there is little evidence of halo effects. In addition, concerns
about common method bias are alleviated (Rindfleisch, Malter,
Ganesan, & Moorman, 2008) because (a) our key dependent variable,
SOW, precedes the independent variables in the survey; (b) the order
of items relating to CSR, overall attitude, and all other store attributes is
randomized across respondents; and (c) SOW is measured using a
different measurement scale than that used for other retailer attributes.
3.4. Model
The framework in Fig. 1 translates to a model with two equations, one
for attitude and one for SOW. Both equations include the CSR dimensions
and other store attributes as independent variables; the SOW equation
also includes attitude. All variables are mean-centered relative to their
grand means in the full sample. This does not affect estimates when
there are no interactions and makes it easy to interpret main effects
when interactions are subsequently included as a robustness check.
Unobserved heterogeneity in parameters across consumers can be
incorporated by either the continuous (e.g., Gönül & Srinivasan, 1993)
or the finite mixture method (e.g., Kamakura & Russell, 1989). We use
the latter while accounting for the dependence between observations
from the same respondent in a panel data model. Our choice of the finite
mixture (also called latent class) model is dictated by the fact that we
are interested not just in controlling for heterogeneity but in identifying
actionable consumer segments whose size and preferences provide
important managerial insights.
Since the latent segments are formed based on response to all model
variables, the segment-level parameter estimates characterize the
consumer segments not only in terms of how they respond to CSR
activities but also in terms of the value they place on other retailer
attributes. In addition, per Fig. 1, we use demographics and other
160
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
Table 2
Measurement of model variables.
Variables
Dependent variables
Attitude (ATT)
(α = 0.88)
Behavioral loyalty (SOW)⁎
CSR
Environmental Friendliness (CSREnv)
Community Support (CSRCom)
Local Products (CSRLoc)
Employee Fairness (CSREmp)
Mean
3.64
35.9
Survey itemsa,b
SD
1.02
26.4
I consider myself a loyal customer at Retailer A.
I would recommend Retailer A to my friends.
I would go out of my way to shop at Retailer A.
In the last 6 months, what percentage of your grocery spending
was in Retailer A? (0–100%)
3.62
3.70
3.32
3.47
1.19
1.26
1.47
0.95
I believe that Retailer A has environmentally friendly policies.
I believe that Retailer A cares about the local community.
I believe that Retailer A offers a large selection of local products.
I believe that Retailer A treats employees fairly.
2.92
0.94
3.93
1.01
3.64
0.79
3.73
0.95
3.89
3.50
3.35
3.10
3.61
1.01
1.36
0.99
1.29
1.32
I can get the same items at lower prices in other stores than Retailer A.
Prices at Retailer A are good compared to other stores. (reverse coded)
I am confident in the quality of products at Retailer A.
The quality of products sold at Retailer A is high.
There are special deals available on many products at Retailer A.
When items are on sale at Retailer A, the discounts are deep.
The atmosphere at Retailer A is pleasant.
Help is always available when I need it at Retailer A.
It is easy to find things at Retailer A.
Retailer A offers a big selection of items in many product categories.
I can find unique products at Retailer A that are not available elsewhere.
I have a lot in common with others who shop at Retailer A.
Shoppers at Retailer A tend to be wealthier than at other stores.
Retailer A's location is convenient for me.
Education
CSR-ability belief (CSR-ability)
0.24
0.31
0.23
0.46
2.37
0.43
0.47
0.42
0.50
1.04
CSR-cost belief (CSR-Cost)
Price importance
3.66
3.47
0.87
0.99
Quality importance
4.35
0.72
Service importance
3.30
1.02
CSR importance
3.44
1.12
Seek local
4.08
0.90
Other retailer attributes
Price (Price)
(α = 0.66)
Quality (Qual)
(α = 0.90)
Deals (Deal)
(α = 0.71)
In-store service (Instor)
(α = 0.79)
Assortment selection (Assort)
Unique items (Unique)
Similar shoppers (Similar)
Wealthy shoppers (Wlthy)
Location convenience (Convloc)
Consumer characteristics
Age
Income
a
b
Age-High = 1 if age greater than 65, 0 otherwise
Income-High = 1 if household income is greater than $100 K, 0 otherwise
Income-No Report = 1 if respondent prefer not to report income, 0 otherwise
Educ-High = 1 if more than college graduation, 0 otherwise
Environmental and social responsibility makes it difficult for companies to
best serve their customers.
Environmental and social responsibility programs increase a company's costs.
How important is everyday price when you decide where to do most of your
grocery shopping?
How important is product quality when you decide where to do most of your
grocery shopping?
How important is in-store service when you decide where to do most of your
grocery shopping?
How important is environmental and social responsibility when you decide
where to do most of your grocery shopping?
I seek out locally grown and locally produced foods.
All items except SOW are measured on a 5-point scale with 5 = “strongly agree” or “extremely important” and 1 = “strongly disagree” or “not at all important”.
In the survey, “Retailer A” is replaced by each retailer's actual name.
consumer characteristics to explain the probability of belonging to a
given segment. The equation for attitude is:
Attir ¼ δc0 þ δc1 CSREnvir þ δc2 CSRComir þ δc3 CSRLocir þ δc4 CSREmpir
þ δc5 Priceir þ δc6 Assortir þ δc7 Uniqueir þ δc8 Qualir þ δc9 Dealir þ δc10 Instorir
þδc11 Similarir þ δc12 Wlthyir þ δc13 Convlocir þ ηir
ð1Þ
makes it difficult for a firm to effectively serve its customers, and belief
that CSR raises a firm's costs. Complete definitions are listed in Table 2.
The equation for SOW is the same as the attitude equation, except
that Attir is included on the right-hand side of the equation:
SOWir ¼ βs0 þ βs1 CSREnvir þ βs2 CSRComir þ βs3 CSRLocir þ βs4 CSREmpir þ
þ βs5 Priceir þ βs6 Assortir þ βs7 Uniqueir þ βs8 Qualir þ βs9 Dealir þ βs10 Instorir
þβs11 Similarir þ βs12 Wlthyir þ βs13 Convlocir þ βs14 Attir þ εir
ð3Þ
where the variables are as defined in Table 2; the subscripts refer to
consumer i and retailer r, and c = {1,2,…C} indicates the latent class or
segment. The prior probability that consumer i belongs to segment c is
given by:
Here s = {1,2,…S} indicates the latent class or segment and the prior
probability that consumer i belongs to segment s is given by:
! 0 "
exp Z i :γc
Prob ði ¼ cÞ ¼ XC
# 0 $
exp Z i :γ c
c0 ¼1
! 0 "
exp Z i :γ s
Prob ði ¼ sÞ ¼ XS
# 0 $
exp Z i :γ s
s0 ¼1
ð2Þ
where Zi is a vector of consumer i's characteristics comprising income,
education, age, average, weekly grocery expenditure, belief that CSR
ð4Þ
We estimate the attitude and SOW models separately, using the
concomitant variable latent class approach (Greene, 2003; Wedel &
161
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
Table 3
Descriptive statistics across retailers.
Variables
Environmental friendliness
Community support
Local products
Employee fairness
Price
Assortment selection
Unique items
Product quality
Deals
In-store service
Similar shoppers
Wealthy shoppers
Location convenience
Attitude
SOW (among raters of store)
SOW (full sample n = 1746)
Mean (std. error) for retailer
A (n = 1746)
B (n = 416)
C (n = 575)
D (n = 273)
E (n = 176)
F (n = 167)
G (n = 98)
H (n = 41)
4.55
(.02)
4.68
(.02)
4.54
(.02)
4.01
(.02)
3.45
(.02)
4.11
(.02)
4.58
(.02)
4.67
(.01)
3.50
(.02)
4.31
(.02)
3.74
(.02)
4.10
(.02)
3.86
(.03)
4.20
(.02)
39.72
(.67)
39.72
(.67)
2.74
(.04)
2.74
(.04)
2.18
(.04)
2.99
(.03)
2.67
(.04)
3.87
(.04)
2.52
(.05)
3.17
(.04)
3.77
(.03)
3.27
(.04)
2.98
(.04)
2.17
(.04)
3.36
(.06)
2.84
(.04)
31.10
(1.12)
12.04
(.46)
2.76
(.03)
2.81
(.04)
2.36
(.04)
3.02
(.02)
2.37
(.03)
4.01
(.03)
2.64
(.04)
3.30
(.03)
3.95
(.03)
3.26
(.03)
3.05
(.04)
2.14
(.03)
3.36
(.05)
3.16
(.04)
36.64
(1.02)
18.92
(.57)
2.47
(.05)
2.39
(.05)
1.49
(.04)
2.88
(.04)
2.22
(.05)
2.84
(.07)
2.55
(.07)
3.16
(.05)
3.69
(.05)
2.73
(.05)
2.73
(.05)
2.06
(.05)
3.06
(.07)
3.32
(.05)
19.68
(.77)
6.32
(.27)
3.02
(.07)
3.06
(.08)
2.58
(.07)
3.09
(.05)
2.81
(.06)
3.66
(.07)
2.64
(.08)
3.42
(.07)
3.35
(.05)
3.46
(.07)
3.34
(.06)
2.60
(.06)
4.24
(.08)
3.01
(.07)
43.35
(2.17)
6.72
(.43)
2.34
(.07)
2.41
(.08)
1.46
(.05)
2.34
(.07)
2.02
(.06)
3.48
(.08)
2.19
(.08)
2.57
(.07)
3.76
(.07)
2.61
(.07)
2.58
(.07)
1.56
(.06)
2.98
(.10)
2.75
(.08)
20.92
(1.25)
4.49
(.24)
2.90
(.08)
3.19
(.08)
2.38
(.10)
3.17
(.07)
1.71
(.07)
4.06
(.09)
2.75
(.12)
3.54
(.09)
4.21
(.07)
3.64
(.08)
3.08
(.10)
1.84
(.07)
3.09
(.14)
3.88
(.09)
45.54
(3.05)
4.18
(.35)
2.37
(.13)
2.51
(.16)
1.90
(.15)
2.78
(.14)
2.74
(.11)
2.83
(.20)
1.81
(.14)
2.67
(.15)
3.42
(.12)
2.99
(.15)
2.63
(.16)
1.76
(.16)
3.78
(.23)
2.36
(.17)
25.39
(3.38)
1.20
(.15)
Kamakura, 2000), and select the number of latent segments using the
Bayesian Information Criterion (BIC).3
4. Results
Table 4 presents correlations between key variables in our model.
The CSR dimensions are highly correlated as would be expected. There
are also strong correlations between CSR dimensions and some other
retailer attributes. This underscores the importance of controlling for
the latter to ensure that the estimated effects of CSR are not biased
due to omitted variables. The strong correlations also suggest the
possibility of multicollinearity so we computed variance inflation
factors (VIFs) and condition indices for all the variables in our model
before proceeding further. The highest VIFs are for the CSR dimensions,
but even these are all less than 5, well below levels of 10 or higher that
are considered problematic. The condition number for the model
variables is only 6.2, well below the ad hoc standard of 30 that is often
used.
Next, we ensured that our mediation model is supported by the data
(Baron & Kenny, 1986; Zhao, Lynch, & Chen, 2010). First we estimated
equations for the effects of store attributes on attitude and SOW, and
found significant effects of most attributes in both equations. Then, we
added attitude to the SOW model and found that most attributes
3
Bucklin, Gupta, and Siddarth (1998) note that the trade-off between separate and
joint estimation is one of better fit for the former vs. parsimony (fewer segments) for
the latter. The better fit for the separate approach is due to its flexibility. A customer
may be in attitude segment 1 and then be in either SOW segment 1 or 2. This is in contrast
to the joint approach, which forces all customers who are in attitude segment 1 to be in
SOW segment 1. We opted for the separate approach because we are interested in
segmentation and therefore flexibility is important.
continued to have significant direct effects on SOW after controlling
for attitude, with a significant change in their magnitude. Thus, the
data support partial mediation of the effects of store attributes on
SOW by attitude.
4.1. Attitude model
According to the BIC criterion, two segments provided the best fit for
the attitude model, with Segment 2 being the majority segment (64.4%
versus 35.6% of the sample). Segment-level parameters for the store
attributes are provided in the first two columns of Table 5. The CSR
coefficients all have positive signs and are statistically significant in four
out of eight cases. This confirms that, by and large, CSR improves
customer attitudes toward the store. Segments 1 and 2 differ interestingly
in the emphasis they put on various CSR activities. Segment 1 places more
emphasis on environmental friendliness whereas Segment 2 places more
emphasis on employee fairness. The segments are similar in their attitude
response to community support and local products. Both value the former
and neither has an attitudinal response to the latter. Both segments
have the expected signs for other store attributes, differing only in the
magnitude of some effects. Perhaps the most important difference
between them is that Segment 2 places more emphasis on promotional
deals and on price and less emphasis on unique items and quality.
In addition, it is helpful to characterize segments in terms of
consumer characteristics like demographic variables and to see if their
CSR response is related to beliefs about how CSR affects corporate ability
(Brown & Dacin, 1997; Sen & Bhattacharya, 2001). The bottom panel of
Table 5 provides the effects of these concomitant variables on the
probability of belonging to Segment 1 versus Segment 2. We find that
higher age and more education increase the likelihood of being in
Segment 1 and belief that CSR limits a firm's ability to effectively serve
its customers decreases the likelihood of being in Segment 1.
162
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
Table 4
Correlations among model variables.
EnFr
Env. friendliness
Comm. support
Local products
Employ. fairness
Price
Assort. selection
Unique items
Prod. quality
Deals
In-store service
Similar shoppers
Wlthy. shoppers
Loc. conv.
Attitude
Share of wallet
CmSup
1
.82
.81
.65
.40
.31
.71
.78
−.00
.72
.49
.68
.24
.64
.21
LocP
EmFair
Price
Assor
Uniq
Qual
Deal
1
.05
.76
.53
.65
.24
.73
.23
1
.16
.14
−.19
.03
.23
.16
Serv
Sim
Wlth
LCon
Att
SOW
1
.44
1
1
.81
.63
.37
.34
.73
.79
.01
.71
.50
.66
.24
.65
.21
1
.61
.45
.33
.75
.78
−.05
.72
.48
.74
.23
.61
.22
1
.25
.27
.52
.61
.06
.62
.43
.54
.20
.54
.21
1
−.03
.40
.32
−.42
.22
.13
.56
.15
.03
−.10
1
.33
.39
.30
.42
.32
.18
.05
.37
.17
1
.72
−.02
.61
.43
.65
.15
.59
.11
1
.53
.55
.28
.69
.31
1
.38
.23
.54
.30
1
.21
.45
.10
1
.21
.30
Table 5
Attitude and share-of-wallet models with main effects.
Independent variable
Store attributes:
Envir. friendliness
Community support
Local products
Employee fairness
Price
Assortment selection
Unique items
Product quality
Deals
In-store service
Similar shoppers
Wealthy shoppers
Location conven.
Attitude
Concomitant variables#:
CSR-ability belief
CSR-costs belief
Education-high
Income-high
Income-no report
Age-high
Wkly grocery spend
Attitude
Share of wallet
Seg 1
(34.2%)
Std.
error
.173⁎⁎⁎
.098⁎⁎⁎
.017
.004
−.119⁎⁎⁎
.030
.026
.025
.029
.028
.022
.020
.036
.030
.032
.025
.023
.014
.015
.147⁎⁎⁎
.382⁎⁎⁎
−.028
.156⁎⁎⁎
.062⁎⁎
.030
.023
–
−.307⁎⁎⁎
−.019
.593⁎⁎⁎
.029
.317
.662⁎⁎⁎
.054
.099
.107
.194
.238
.243
.231
.172
Seg 2
(65.8%)
.005
.078⁎⁎⁎
.016
.078⁎⁎⁎
−.296⁎⁎⁎
−.016
.074⁎⁎⁎
.287⁎⁎⁎
.139⁎⁎⁎
.194⁎⁎⁎
.149⁎⁎⁎
−.025
.015
–
–
–
–
–
–
–
–
Std.
error
.029
.025
.022
.021
.021
.016
.018
.026
.023
.024
.017
.020
.012
Seg 1
(27.7%)
2.681⁎⁎⁎
1.864⁎
7.039⁎⁎⁎
.853
.266
−1.252⁎
2.288⁎⁎⁎
.446
−1.007
1.672
.601
3.633⁎⁎⁎
1.659⁎⁎⁎
Share of wallet mediated
Std.
error
Seg 2
(72.3%)
Std.
error
.985
.986
.744
.819
.803
.666
.705
1.060
.845
1.059
.720
.734
.432
−1.287
−1.069
1.218⁎
1.800⁎⁎
−5.616⁎⁎⁎
.822
.699
.681
.745
.662
.575
.560
.939
.754
.865
.577
.614
.376
–
.938
−3.564⁎⁎⁎
−.071
1.094
4.590⁎⁎⁎
4.007⁎⁎⁎
−1.865⁎⁎⁎
4.566⁎⁎⁎
–
−.358⁎⁎⁎
.033
.372⁎⁎
.079
.444⁎⁎⁎
.188
.478⁎⁎⁎
.078
.086
.149
.183
.186
.178
.131
–
–
–
–
–
–
–
Seg 1
(25.4%)
1.944⁎⁎
.767
7.246⁎⁎⁎
.796
1.570⁎
−.880
1.577⁎⁎
−3.578⁎⁎⁎
−2.074⁎⁎
.905
−.364
3.398⁎⁎⁎
1.253⁎⁎⁎
8.649⁎⁎⁎
−.321⁎⁎⁎
.031
.337⁎⁎
.044
.399⁎⁎
.215
.550⁎⁎⁎
Std.
error
Seg 2
(74.6%)
Std.
error
.988
.969
.766
.799
.806
.632
.695
1.122
.840
1.059
.750
.718
.432
.966
−1.961⁎⁎⁎
−1.774⁎⁎⁎
1.365⁎⁎
.765
.659
.654
.691
.655
.546
.524
.902
.708
.868
.546
.580
.358
.769
.083
.092
.160
.195
.198
.187
.140
.905
−2.584⁎⁎⁎
.962⁎
−4.584⁎⁎⁎
−3.119⁎⁎⁎
.100
2.182⁎⁎
2.552⁎⁎⁎
−1.524⁎⁎⁎
4.541⁎⁎⁎
11.096⁎⁎⁎
–
–
–
–
–
–
–
Standard errors are in parentheses. Effects of CSR variables are in bold.
#
Effect of concomitant variables on probability of membership in segment 1 versus 2.
⁎⁎⁎ p b 0.01.
⁎⁎ p b 0.05.
⁎ p b 0.10.
4.2. SOW model
The BIC criterion supports a two-segment solution for the SOW
model also. Table 5 shows the non-mediated SOW model (i.e., without
attitude) in columns 3 and 4, and the mediated SOW model (i.e., with
attitude) in columns 5 and 6. Contrasting the two models yields two
important conclusions: First, attitude is clearly important. Its coefficient
is positive and highly significant in both segments. Second, many
CSR and other store attributes remain statistically significant in the
mediated SOW model showing that attitude partially mediates the
relationship between attributes and SOW. Thus, we use the mediated
SOW model in the remainder of our discussion.
The two segments in this SOW model differ substantially in the
direct effects of CSR and some other attributes like price, unique
items, deals, and wealthy shoppers. The first segment shows more
positive direct effects of CSR, unique items and wealthy shoppers, a
more negative direct effect of deals, and an insignificant direct effect
of price. In terms of the concomitant variables, customers who are
highly educated, bigger spenders, and who refuse to report income
are more likely to be in Segment 1.4
It is important to point out that interpreting the direct CSR/attribute
coefficients of the mediated SOW model in isolation is not particularly
relevant. Managerially, what is of interest is the total effect, i.e., the
impact of a change in an attribute, e.g., a CSR dimension, on SOW. This
δSOW δAtt
total effect is dSOW
and is given by δSOW
δCSR þ δAtt % δCSR . The first term is
dCSR
4
It may seem odd for a “missing data” code to have strong explanatory power.
However, in light of the other features of Segment 1, “Income-no report” probably means
that the customer is higher income. The importance of missing data codes such as
represented by the income-no report variable is consistent with Blattberg et al.
(2008, p. 307), who recommend the use of missing-variable coding in database
marketing models.
163
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
the direct effect of the CSR dimension; the second is the indirect effect of
CSR on SOW through the mediating attitude variable. The
decomposition of the total effect shows it is possible for CSR to inspire
positive attitudes but little behavioral action. A positive effect of CSR
on attitude, together with a positive effect of attitude on behavior
δAtt
makes the δSOW
δAtt % δCSR term positive. Therefore, for the total effect on
behavior to be non-positive, the direct effect, δSOW
δCSR , has to be negative.
In the next section, we calculate the total effects and interpret these
in conjunction with the coefficients reported in Table 5. Since the total
# δAtt $
effects combine coefficients
the attitude model δCSR
and the
#δSOW from
$
mediated SOW model δAtt ; δSOW
δCSR , they depend on which segment
(1 or 2) the consumer belongs to in the attitude model and the
mediated SOW model.
4.3. Total effect of CSR activities on SOW
First, we classify each consumer in Attitude Segment 1 or 2 and in
SOW Segment 1 or 2, using the posterior probabilities of segment
membership.5 This generates four groups of consumers: Group 1/1
comprising consumers who are in Attitude Segment 1 and SOW
Segment 1, Group 2/1 comprising consumers who are Attitude Segment
2 and SOW Segment 1, and so on. Next, we compute the total effect of
each model variable (i.e., the total number of units by which SOW
changes for a one unit increase in the variable) in each of the four
groups, using parameter estimates from the corresponding Attitude
and SOW segments. Finally, we use bootstrapping to obtain standard
errors for these total effects (Efron & Tibshirani, 1993) in each of the
four groups. For each group, we draw 500 bootstrap samples with
replacement, estimate our models and compute the total effects of
each model variable for each sample. The standard deviation of a
given total effect across the bootstrapped samples is its standard error.
Table 6 provides the total effects and standard errors for all four
groups. The first take-away is that the two attitude segments overlap
only partially with the two SOW segments. The largest portion of the
sample is in Group 2/2 but a significant proportion of consumers is in
Groups 2/1 and 1/2. This underscores the advantage of not forcing a
single segmentation scheme for attitude and SOW. The second takeaway is that total SOW returns differ by group and CSR dimension,
underscoring the importance of segmentation. Third, of the 16 CSR
total effects, 10 are statistically significant – nine with a positive sign
and one with a negative sign. Clearly CSR exerts an important impact
on SOW.
Groups 1/1 and 2/1 are relatively small (11.6% and 13.8% of the
sample, respectively). However, CSR is very important to these
consumers – significant and positive in seven of eight cases. Both of
these groups place very high emphasis on local products: a one-point
increase in local products perceptions garners over seven SOW points.
They also respond similarly to community support: a one-point increase
in community support garners 1.62 SOW points for Group 1/1 and 1.44
SOW points for Group 2/1. With respect to the remaining two CSR
dimensions, Group 1/1 places more emphasis on environmental
friendliness, while Group 2/1 places more emphasis on employee
fairness. Referring back to Table 5, we see several significant CSR effects
both in the attitude and mediated SOW models for attitude Segments 1
and 2 and SOW Segment 1, so the strong total effects for Groups 1/1 and
1/2 are as expected.
CSR exerts different effects on Groups 1/2 and 2/2. This follows from
the weaker and even negative direct effects in the mediated SOW model
5
The most transparent way to assess the “quality” of classification is to examine the
posterior probability of each respondent being in the segment he/she is assigned to. The
closer these probabilities are to 1, the better the quality of classification. We find that
the average probability of being in the assigned segment is 0.89 for the SOW model and
0.77 for the attitude model. Thus, the quality of assignment is very high, especially for
the SOW model.
Table 6
Total effects on share of wallet.
Group defined by
Total effect of
Attitude Seg1
SOW Seg1
(size = 11.6%)
Attitude Seg2
SOW Seg1
(size = 13.8%)
Attitude Seg1
SOW Seg2
(size = 22.6%)
Attitude Seg2
SOW Seg2
(size = 52.0%)
Environmental
friendliness
Community
support
Local products
3.440⁎⁎⁎
(1.064)
1.615⁎
1.987⁎⁎⁎
(.782)
1.442⁎⁎
(1.856)
7.393⁎⁎⁎
(.723)
7.384⁎⁎⁎
−1.906⁎⁎
(.933)
−.909
(.961)
1.543⁎
(1.074)
.831
(.820)
.541
(.846)
−.750
(.704)
2.848⁎⁎⁎
(.639)
−.274
(1.023)
−2.316⁎⁎⁎
(.778)
1.471⁎⁎
(.734)
−.990
(.664)
−1.018⁎⁎
(.883)
2.254⁎⁎⁎
(.461)
2.217⁎⁎⁎
(.516)
−1.096
(.826)
−.872
(.722)
2.583⁎⁎⁎
−.041
(1.419)
−.687
(1.202)
1.554
(1.108)
.949
(1.021)
−3.904⁎⁎⁎
(1.056)
1.128
(.857)
−2.953⁎⁎⁎
(.932)
1.120
(1.492)
−.211
(1.173)
3.913⁎⁎⁎
(.754)
4.335⁎⁎⁎
(.928)
.172
(.819)
3.657⁎⁎⁎
(.782)
1.452⁎⁎⁎
(.847)
.925
(.726)
3.182⁎⁎⁎
(.614)
1.383⁎⁎⁎
(1.312)
3.240⁎⁎⁎
(1.016)
−1.191
(.939)
4.796⁎⁎⁎
(.989)
4.205⁎⁎⁎
(.591)
−1.801⁎⁎⁎
(.710)
4.707⁎⁎⁎
(.502)
(.453)
(.619)
(.409)
Employee
fairness
Price
Assortment
selection
Unique items
Product quality
Deals
In-store service
Similar
shoppers
Wealthy
shoppers
Location
convenience
(.860)
1.770⁎⁎
(.771)
−5.868⁎⁎⁎
(.684)
.784
(.601)
−3.763⁎⁎⁎
(.764)
.066
(1.007)
1.642⁎⁎
Note 1: Total effect = (Direct effect on SOW) + (Effect on attitude)⁎(Effect of attitude on
SOW).
Note 2: Bootstrapped standard errors are in parentheses. Effects of CSR variables are in bold.
⁎⁎⁎ p b 0.01.
⁎⁎ p b 0.05.
⁎ p b 0.10.
for segment 2. CSR does not exert a significant impact on SOW for Group
1/2. The local products effect is close but does not achieve statistical
significance at the 0.10 level. So for 22.6% of the consumer base, CSR is
a non-factor in terms of SOW.
Group 2/2, the largest portion of consumers (52.0%), presents a
mixed response to CSR. Local products and employee fairness exert
positive total effects. A one-point increase in employee fairness is
worth 1.77 SOW points and a one-point increase in local products yields
1.54 points of SOW. Note that these two CSR variables are related to a
consumer's shopping experience, so a stronger response is expected –
these CSR efforts provide not just a societal but also a personal benefit.
In contrast, the impact of community support is not significant, and
environmental friendliness has a negative total impact, significant at
the 0.05 level. This is only one negative total impact out of 16 total
CSR effects, but nevertheless it is intriguing and begs explanation.
We offer an ex post psychological explanation for the negative impact
of environmental friendliness in this group – attribution. In particular,
consumers who see a retailer devoting attention to a CSR cause that is
not related to their experience with the store might infer that the
retailer's attention is being diverted from serving customers. Consistent
with this hypothesis, we find that one of the concomitant variables, the
perception that “CSR makes it difficult for companies to serve their
customers effectively,” is associated with a higher probability of being
in attitude segment two (where environmental friendliness has an
insignificant effect), and a higher probability of being in SOW segment
two (where the direct effect of environmental friendliness is negative).
The net result is that consumers in Group 2/2 are concerned that the
retailer's attention to CSR activities for the broad societal good takes
their attention away from the customer. This shows up as a significantly
164
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
negative total response to environmental friendliness and an insignificant total response to community support, the two CSR dimensions
that have a societal but not necessarily a personal benefit.
Overall, Table 6 shows that some CSR efforts can have a strongly
positive impact on SOW. The positive effect is especially strong for 25%
of our sample. For 52% of the consumers, CSR related to the consumer's
experience in the store – local products and employee fairness – has a
significant positive effect, but broader societal good related CSR,
particularly environmental friendliness, detracts from SOW. There is a
third group – 23% of consumers – for whom CSR is a non-factor.
4.4. Total impact of other store attributes on SOW
Table 6 shows the total effects of other store attributes on SOW,
suggesting important contrasts among the groups. Groups 1/1 and 1/2
are not price sensitive, not deal prone, and value unique and special
items and a wealthy clientele. Recall that these two groups are the
most responsive to CSR. In contrast, Groups 1/2 and 2/2 are price
sensitive and value similar shoppers, not wealthy ones. Group 2/2 is
particularly price sensitive and deal prone. They respond to unique
items by reducing SOW. Perhaps these consumers think of such a
store as a place to shop for special items, not for everyday needs. The
survey includes a question as to whether the consumer shopped at a
chain mainly for special items not available elsewhere. Consistent
with our logic, this item is more positively correlated with the Unique
attribute in SOW segment 2 than in SOW segment 1 (correlation =
0.38 versus 0.17), and its mean is significantly higher for SOW segment
2 (3.95) than for SOW segment 1 (3.09).
Perhaps the most surprising result is the lack of importance of
quality. Table 5 shows that both Attitude segments have a positive effect
of quality, but the direct effect on SOW is negative for both SOW
segments. Table 6 shows that the total effect is not significant. Everyone
likes high quality “theoretically,” but when it comes to actual shopping,
quality may not mean much within the range of the data. The stores
carry more or less the same packaged goods brands so, as the retail
experts we interviewed noted, quality differs primarily in fresh produce,
which is not a large part of total grocery spending.
4.5. Additional characterization of the four groups
The concomitant variables discussed previously characterize the
Attitude and SOW segments. In Table 7, we summarize additional selfreported characteristics of the four groups. The contrast between Groups
1/1 and 2/2 (respectively the most and least CSR responsive groups
Table 7
Shopping characteristics of the four groups.
Mean (standard error) in group defined by
Variable
Attitude Seg1 Attitude Seg2 Attitude Seg1 Attitude Seg2
SOW Seg2
SOW Seg2
SOW Seg1
SOW Seg1
(size = 11.6%) (size = 13.8%) (size = 22.6%) (size = 52.0%)
SOW at focal
chain (%)
SOW at low price
chains (%)
(chains
C + D + F + G)
Importance of
price
Importance of
quality
Importance of
in-store service
Importance of CSR
73.82⁎
(.74)
12.61⁎
71.31⁎
(.78)
18.80⁎
35.01⁎
(1.15)
30.58⁎
(.77)
(.71)
(1.33)
3.03⁎
(.07)
4.58⁎
3.29⁎
(.06)
3.93⁎
(.04)
3.47⁎
(.05)
3.42⁎
(.07)
3.79⁎
(.07)
4.38⁎
(.07)
3.57⁎
(.08)
4.25⁎
3.31⁎
(.05)
4.35
(.04)
3.34
(.05)
3.57⁎
(.05)
4.23⁎
(.06)
(.05)
(.04)
Seeking local
products
⁎ Mean is significantly different from group 4 at p b 0.01.
25.71
(.74)
44.12
(1.02)
3.70
(.03)
4.27
(.03)
3.21
(.03)
3.27
(.04)
3.91
(.03)
based on our results in Table 6) is particularly interesting. First, Groups
1/1 and 2/1 have much higher SOW at the focal retailer than the other
two groups, and the self-reported importance of CSR in store choice,
especially in group 1/1, is significantly higher than for the other two
groups. This makes sense given the focal retailer's superior performance
on CSR and the high responsiveness of these groups to CSR. Groups 1/1
and 1/2 also have much lower SOW at the price oriented retailers
(Chains C, D, F, and G), and the self-reported importance of price in
their store choice is significantly lower. Again this is in line with our
model-based results showing that Groups 1/1 and 1/2 are much less
price sensitive and much less deal prone than the other two groups,
especially Group 2/2. Finally, Groups 1/1 and 1/2 report significantly
higher importance of in-store service and quality in their store choice,
and they report seeking local products more than Group 2/2. This,
too, is in line with our model-based findings. Thus, the differences in
these self-reported characteristics across the four groups provide high
convergent validity for our model-based results.
4.6. Robustness checks
We conducted several analyses to establish the robustness of our
results. The robustness checks relate to (i) use of self-reported SOW
versus SOW computed from purchases, (ii) multicollinearity, (iii) chain
differences in CSR response, and (iv) potential nonlinear effects of CSR.
We summarize our findings below but full details are available upon
request.
4.6.1. Self-reported versus computed SOW
The SOW measure used in our analyses is self-reported. We have
respondents' actual purchase data from the focal retailer but we are
unable to use actual SOW because we do not have the same information
about their purchases from other retailers. However, we tested the
robustness of our results with actual purchase data in two ways. First,
we computed SOW at the focal retailer from the respondents' actual
purchases at the focal retailer and their total weekly grocery budget as
reported in the survey. Then we re-estimated our model after replacing
the self-reported SOW by this computed SOW for observations relating
to the focal retailer. We found no substantive difference in results.
Further, we estimated the model using only observations for the focal
retailer. This allowed us to directly compare results between selfreported and computed SOW because both were available. It was
reassuring that the results were very similar for the two SOW measures.
4.6.2. Multicollinearity
Although the VIFs and condition indices in our model are within
acceptable limits, we conducted additional checks given the high
correlations between the CSR dimensions. Multicollinearity generally
increases standard errors, rendering the estimated coefficients unstable
across model variations. Sometimes it can result in “wrong signs”, so we
wished to make sure that the negative direct effects of some CSR
dimensions are not an artifact of multicollinearity. We re-estimated
the mediated SOW model by dropping one CSR dimension at a time
and found that the results were very robust – e.g., the negative effects
of environmental friendliness and community support in the second
segment remained. Noting the high correlations between the quality
attribute and in-store service as well as unique items attributes, we
ran three additional models, dropping in-store service in one, unique
items in the other, and both in the third, to see if the negative quality
coefficient held up. Indeed it did. The signs of the CSR variables held
up as well, although there were a few instances where significance
levels changed. Finally, we subjected the multi-item store attributes to
a principal components analysis and re-estimated the models with the
orthogonal component scores. The CSR results were unchanged.
We also examined the correlation matrices for SOW Segments 1 and
2 and found them to be similar. We examined correlations for the focal
chain versus other chains. These were somewhat smaller than in the
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
overall sample but, as we discuss in the next section, this is at the cost of
less variation. Overall, therefore, we find that multicollinearity is not a
problem and it is not responsible for negative or insignificant CSR
effects.
4.6.3. Focal versus other chains
Our analysis relies on between chain/within customer variation to
estimate the impact of CSR. However, it may be argued that our results,
especially the negative direct effects of environmental friendliness and
community support, are driven mainly by focal chain which is positioned
much higher on CSR dimensions. Although separating the chains reduces
variation in the CSR dimensions and hurts statistical power, we repeated
our analysis after deleting the focal chain. As we expected, this resulted
in fewer segments and reduced statistical significance. However, the
direct effects of environmental friendliness and community support do
not turn positive even when we exclude the focal chain. Indeed, the
direct effect of community support remains significantly negative in
one segment, just as in the full sample. We also found that the signs for
local products and employee fairness are all positive, and in two of four
cases, significant, despite the reduction in statistical power. In summary,
the basic pattern of results persists when we remove the focal chain from
the analysis, despite the weaker statistical significance resulting from
attenuated variation in the CSR variables.
4.6.4. Nonlinear effects of CSR
It is possible that the impact of CSR is nonlinear. For example, the
negative impact we find for environmental friendliness could be due
to “over-reaching” on the part of the focal chain, i.e., going too far in
its emphasis on environmental friendliness. This could lead to an
inverted U effect on SOW. We examined this by adding dummy
variables for high CSR values, i.e., equal to 1 if the CSR variable is rated
4 or 5 on our 1 to 5 scale (the variables are measured on an interval
rather than a ratio scale, so it is not appropriate to include a quadratic
term).
Although model fit improved, we found that the negative effects we
discussed earlier cannot be attributed to an inverted-U relationship.
Specifically, the dummy variables for environmental friendliness and
community support are not significant for the second SOW segment
where our model showed negative effects of environmental friendliness
and community support. And directionally, they support the opposite of
an inverted U effect. For example, the effect of community support
becomes less negative at high levels. The only dimension for which we
see an inverted U is one that did not have a negative effect in our
model – local products. We find that the effect of local products gets
less positive at high levels.
Adding these dummy variables aggravates multicollinearity (we now
obtain VIFs above 5 and the condition number is 15). Further, the
magnitude of the dummy variable for local products is implausibly
high in Segment 1. Overall, therefore, although there are some nonlinear
effects (as shown by the improved fit), they do not explain the negative
effects in our original model, nor do they change our results directionally.
We are also concerned that this is manipulating the 1–5 scale data too
much. Overall, we conclude that the exacerbation of multicollinearity is
not worth the additional insights and therefore we retain our original
model. However, we note that non-monotonic CSR effects present a
fruitful direction for future research.
A final nonlinear effect is potential interactions, especially between
CSR efforts and other store attributes. We used the “data-driven”
procedure proposed by Bijmolt, Van Heerde, and Pietres (2005) to
identify potential interactions. Specifically, we added one set of interactions at a time (between the CSR activities and another attribute),
determined the number of segments using BIC, and then determined
whether this set of interactions improved the fit of the model. We did
not find any interactions for the attitude model, while we found
interactions of CSR with quality and unique items for the SOW model.
165
We used the results to recalculate the total CSR effects on SOW and
found the pattern of results to be similar to those shown in Table 6.6
5. Conclusion
We have used field data to quantify the impact of competing grocery
retailers' CSR activities on consumers' behavioral loyalty towards these
firms. We measure behavioral loyalty as share of wallet and distinguish
among four types of CSR and among consumer segments. We
investigate the role of consumers' attitude toward the store, allowing
it to mediate the impact of CSR and other store attributes on SOW.
Our key findings are as follows:
(a) CSR perceptions have a direct effect on SOW as well as an indirect
effect through attitude toward the store. The effects on attitudes
are generally positive and attitudes enhance SOW. However,
some of the direct effects are negative reducing the total effect
on SOW, which is a combination of indirect and direct effects.
(b) The total effect on SOW varies substantially across segments and
CSR activities. Among the four types of CSR activities in our study,
selling locally produced products has strong universal appeal.
Employee fairness also has a positive, albeit weaker, impact
across segments. Environmental friendliness is a double-edged
sword with respect to SOW. A substantial segment of consumers
reacts negatively to it.
(c) For the largest group comprising over 50% of our sample, a oneunit improvement in perception (on a 1–5 scale) of local products
or employee fairness increases SOW by 1.5 and 1.8 points
respectively. However, this group is turned off by environmental
friendliness – a similar effort on this dimension loses 1.9 SOW
points. For two other groups, comprising approximately 25% of
our sample, the SOW gain from local products is much larger
and improvements in environmental friendliness and community
support also garner substantial increases in SOW. Thus, there is a
strong case for benefit segmentation in CSR efforts.
(d) These groups can be distinguished based on education, age,
income, and belief that CSR activities limits a firm's ability to
effectively serve its customers. They can also be distinguished
based on their response to other store attributes. Compared to
the smaller groups that respond positively to all CSR, members
of the group with negative SOW response to environmental
friendliness and community support are more price sensitive
and place greater value on assortment and location convenience.
They are turned off by perceptions of exclusivity such as unique
items and a wealthy clientele, have a smaller weekly grocery
budget, and are more likely to believe that CSR efforts hinders
the retailer's ability to serve its customers effectively.
These findings have several important implications. First, the potential
gains and losses of SOW due to improvements in CSR perceptions are
managerially meaningful. U.S. supermarket sales exceed $550 billion
annually and median store sales is over $25 million (Food Marketing
Institute, 2010), so every SOW point carries a substantial dollar amount.
For instance, consider Kroger, one of the largest U.S. grocery chains with
approximately $76 billion in annual sales. It recorded a market share
gain of 0.61 points in its major markets in 2008, and a total gain of 2.25
market share points over a four-year period. These gains were considered
very “impressive” in its press release announcing the fiscal year results
(The Kroger Company, 2009).
Second, not all CSR initiatives are equally important or meaningful.
The best CSR initiatives are closely integrated into the company's core
customer offering. CSR activities that are directly tied to the customer's
experience with the firm – the front-end employees and the products –
generate a higher return that is less contingent upon consumers'
6
Complete details are available from the authors.
166
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
idiosyncratic beliefs about the relationship between CSR and corporate
abilities.
Third, our results highlight the importance of targeting CSR
communications to consumers. For the largest group, communicating
environmental friendliness hurts SOW. This does not mean that firms
should act in ways that are environmentally unfriendly or that exploit
their employees. For one, consumers react much more strongly to
negative news about a firm's CSR than they might to positive news
(Bhattacharya & Sen, 2004). For another, CSR never detracts from
overall attitudes toward the store and therefore may lead to other
pro-firm behavior not captured in SOW, such as word of mouth referrals
and advocacy, and higher willingness to forgive occasional lapses
(Bhattacharya & Sen, 2004; Klein & Dawar, 2004). Also, the smaller
groups that value CSR for the broader social good are strategically
important. Their lifetime value to the retailer is likely to be very high
given their lower price sensitivity and high loyalty. In our sample, the
average SOW of these segments with the focal retailer is over 70%, and
other research shows that the “green consumer” spends more and is
more brand and retailer loyal (GMA-Deloitte, 2009). All this suggests
that while core-offering related CSR lends itself to a more uniform,
mass-market communication approach, non-core related CSR is more
nuanced and requires both careful messaging and careful targeting.
This is definitely both feasible and cost-effective for a retailer
that already has a loyalty program and communicates directly with its
consumers, e.g., through e-mail. For instance, all consumers should
receive information about a retailer's local product selection and related
consumption benefits such as freshness and lower pesticide levels. As
also noted in the GMA-Deloitte (2009) study, green products are most
effective when they represent a broader value proposition encompassing
multiple purchase drivers. Only the higher educated, higher income
consumers who use reusable bags or support environmental
organizations should receive information about the environmental
benefits of local products and about the retailer's environmental
programs. Even retailers who do not have a loyalty program can
use the rich geo-segmentation data available from tools such as
Nielsen's PRIZM system to target CSR communications by zip-code
(see Blattberg, Kim, & Neslin, 2008; pp. 197–206).
Beyond targeting based on consumer demographics and psychographics, our findings highlight the importance of appropriate
messaging. Firms must tie their CSR effort not just to the broader social
good but also communicate how those efforts translate into a better
customer experience. They need to combat the view that some CSR
activities do not directly benefit the customer and interfere with the
firm's ability to serve its customers. For example, communications
about a retailer's environmental programs (e.g., energy and water
conservation or waste reduction) should emphasize how they reduce
costs and allow the retailer to invest in products and services that the
consumer values and/or reduce prices. The British grocery retailer,
Tesco, even rewards customers with loyalty program points if they
take actions that benefit the environment, e.g., use a reusable shopping
bag. Indeed, Manget, Roche, and Munnich (2009) find that the most
popular environmentally friendly actions that consumers themselves
undertake involve saving money as well.
Fourth, it is dangerous for companies to charge higher prices
because they perform well on CSR. Although the segment that values
all types of CSR does tend to be less price sensitive, it is small. The largest
group, that values only some types of CSR, is significantly more price
sensitive. Further, we did not find any interactions between CSR efforts
and price response, i.e., CSR does not decrease price sensitivity. This
recommendation is also consistent with the GMA-Deloitte (2009)
finding that consumers don't see why a green product should cost
more if it is manufactured with less packaging/waste or if it is not
transported far.
Fifth, firms need to measure the costs of their CSR initiatives
realistically to calculate the ROI of their CSR investment. These costs
were not available to us so we cannot make these calculations. We do
wish to note that, just as not all CSR activities bring equal SOW benefits,
not all of them are equally costly. Indeed, offering local products, the
activity with the biggest benefit in our study, may not incur incremental
costs (Bustillo & Kesmodel, 2011). Sourcing local products may actually
be cheaper for a retailer due to lower transportation and spoilage costs
and more negotiating leverage over local, often smaller, suppliers.
Similarly, environmentally friendly practices such as reducing plastic
or water or energy use, or reducing waste, may lower costs, the savings
from which can be communicated and passed along to consumers. The
point is that costs, which vary substantially across CSR activities, are
quantifiable, and our research shows how to quantify economic benefits
from the revenue side.
Finally, our results underscore the importance of distinguishing
between attitudes and behavior in CSR research. Previous studies have
suggested that positive attitudes engendered by CSR may not translate
into higher purchase incidence, but, to the best of our knowledge, the
current research is the first to quantify the interrelationship between
attitudes and behavioral loyalty. The conclusion is that attitudes
partially mediate the relationship between CSR and SOW, so evaluation
of CSR must entail both attitudinal and behavioral measures. Our results
show that only measuring impact on attitudes paints a rosier than
warranted picture of CSR.
We note the limitations of our work and some important future
research opportunities. First, our sample comes from the loyalty program
of the focal retailer who is strongly positioned on CSR. Although much
other empirical work has also been done using loyalty program
members, we recognize that consumers in our sample may not be
representative of the population as a whole. In particular, they may be
more responsive to CSR, having chosen to enroll in the focal retailer's
program, so our results may reasonably be viewed as an upper bound
on the SOW returns of CSR. Even for this sample, we find considerable
heterogeneity in CSR response. We hope future research can validate
our findings with a broader sample.
Second, we have identified consumer segments and developed
profiles that can be used for targeted messaging. However, we have
not delved deeply into the underlying reasons for consumers' distinct
perspectives about CSR. For example, we posit the negative impact of
environmental friendliness on SOW in one group may be due to
attribution on the part of the consumer that these efforts detract from
the retailer's ability to serve the customer. Although we find only this
one negative effect out of sixteen effects examined, its significance
within a sizeable segment of consumers and the fact that environmental
efforts are the most commonly publicized CSR initiatives underscores
the importance of further research on this issue.
Third, we studied the major stakeholder group for grocery retailers,
but it is also important to study how other stakeholders such as
employees and investors respond to each CSR dimension. CSR dimensions
like environmental friendliness and community support may well have
significant effects on these stakeholders and therefore on financial returns
even though they have little direct impact on consumers' behavioral
loyalty. Fourth, our research is cross-sectional and there are always
questions of causality with cross-sectional research. Strictly speaking,
our research finds associations though theory suggests the associations
are causal. We note that for reasons reported earlier, we do not believe
halo effects were a problem. In addition, we note that the fact that we
found two CSR dimensions negatively related to SOW and two positively
related suggests that reverse causality is not at play here – respondents
didn't simply rate their favorite store positively on CSR. Still, a longitudinal
study over a period in which CSR policies are changed would be very
valuable, especially as some retailers like Wal-Mart are investing
significantly in environmentally friendly stores, products, and suppliers,
and taking steps to improve their reputation on treatment of employees.
It could also help to distinguish between the effects of positive versus
negative changes in CSR.
Finally, we examined the impact of CSR dimensions in the grocery
retail industry. As discussed at the outset of this paper, the impact of
K.L. Ailawadi et al. / Intern. J. of Research in Marketing 31 (2014) 156–167
CSR should be industry-specific, so industry focus is important. But we
hope future researchers will build on our work by conducting fieldbased analysis of the impact of CSR dimensions in other industries.
Acknowledgments
The authors thank an anonymous grocery retail chain in the
northeastern U.S. for access to their loyalty program members and
data. They thank Rong Guo, Lynn Foster-Johnson, and Paul Wolfson for
their invaluable research support. Finally, we thank the AE and the
reviewers for their constructive and valuable comments.
Appendix A
Supplementary data to this article can be found online at www.
runmycode.org.
References
Ailawadi, K. L., & Keller, K. L. (2004). Understanding retail branding: Conceptual insights
and research priorities. Journal of Retailing, 80(4), 331–342.
Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human
Decision Processes, 50, 179–211.
Auger, P., Burke, P., Devinney, T. M., & Louviere, J. J. (2003). What will consumers pay for
social product features? Journal of Business Ethics, 42(3), 281–304.
Baker, J., Parasuraman, A., Grewal, D., & Voss, G. B. (2002). The influence of multiple store
environment cues on perceived merchandise value and patronage intentions. Journal
of Marketing, 66(2), 120–141.
Bargh, J. A., Chen, M., & Burrows, L. (1996). Automaticity of social behavior: Direct effects
of trait construct and stereotype activation on action. Journal of Personality and Social
Psychology, 71(2), 230–244.
Baron, R. M., & Kenny, D. A. (1986). Moderator–mediator variables distinction in social
psychological research: Conceptual, strategic, and statistical considerations. Journal
of Personality and Social Psychology, 51(6), 1173–1182.
Barone, M. J., Miyazaki, A.D., & Taylor, K. A. (2000). The influence of cause-related
marketing on consumer choice: Does one good turn deserve another? Journal of the
Academy of Marketing Science, 28(2), 248–262.
Barone, M. J., Norman, A. T., & Miyazaki, A.D. (2007). Consumer response to retailer use of
cause-related marketing: Is more fit better? Journal of Retailing, 83(4), 437–445.
Batson, C. D., & Shaw, L. L. (1991). Evidence for altruism: Toward a pluralism of pro-social
motives. Psychological Inquiry, 2(2), 107–122.
Bergkvist, L., & Rossiter, J. R. (2007). The predictive validity of multiple-item versus singleitem measures of the same constructs. Journal of Marketing Research, 44(2), 175–184.
Berman, S. L., Wicks, A.C., Kotha, S., & Jones, T. M. (1999). Does stakeholder orientation
matter? The relationship between stakeholder management models and firm
financial performance. Academy of Management Journal, 42(5), 488–506.
Bhattacharya, C. B., & Sen, S. (2003). Consumer-company identification: A framework for
understanding consumers' relationships with companies. Journal of Marketing, 67(2),
76–88.
Bhattacharya, C. B., & Sen, S. (2004). Doing better at doing good: When, why, and how
consumers respond to corporate social initiatives. California Management Review,
47(1), 9–24.
Bhattacharya, C. B., Sen, S., & Korschun, D. (2008). Using corporate social responsibility to
win the war for talent. MIT Sloan Management Review, 49(2), 37–44.
Bijmolt, T. H. A., Van Heerde, H. J., & Pietres, R. G. M. (2005). New empirical generalizations
on the determinants of price elasticity. Journal of Marketing Research, XLII, 141–156.
Bitner, Mary J. (1992). Servicescapes: The impact of physical surroundings on customers
and employees. Journal of Marketing, 56, 57–71.
Blattberg, R. C., Kim, B., & Neslin, S. A. (2008). Database Marketing: Analyzing and
Managing Customers. New York: Springer.
Bommer, W. H., Johnson, J. L., Rich, G. A., Podsakoff, P.M., & Mackenzie, S. B. (1995). On the
interchangeability of objective and subjective measures of employee performance —
a metaanalysis. Personnel Psychology, 48(3), 587–605.
Brown, T. J., & Dacin, P. A. (1997). The company and the product: Corporate associations
and consumer product responses. Journal of Marketing, 61(1), 68–84.
Bucklin, R. E., Gupta, S., & Siddarth, S. (1998). Determining segmentation in sales response
across consumer purchase behaviors. Journal of Marketing Research, 35(2), 189–197.
Bustillo, M., & Kesmodel, D. (August 1). ‘Local’ grows on Wal-Mart. Wall Street Journal,
258(26), B1–B5.
167
Cotte, J., & Trudel, R. (2009). Socially conscious consumerism: A systematic review of the
body of knowledge. Network for Business Sustainability.
Du, S. L., Bhattacharya, C. B., & Sen, S. (2007). Reaping relational rewards from corporate
social responsibility: The role of competitive positioning. International Journal of
Research in Marketing, 24(3), 224–241.
Efron, Bradley, & Tibshirani, R. J. (1993). An Introduction to the Bootstrap. Chapman &
Hall/CRC, Monographs on Statistics and Applied Probability. 57.
Fishbein, M., & Ajzen, I. (1975). Belief, attitude, intention, and behavior: An introduction to
theory and research. Reading, MA: Addison-Wesley.
Food Marketing Institute (2010). Supermarket facts: Industry overview 2010. Accessed
on July 30, 2011 at http://www.fmi.org/facts_figs/?fuseaction=superfact
GMA-Deloitte (2009). Finding the green in today's shoppers: Sustainability trends and new
shopper insights. GMA/Deloitte Green Shopper Study.
Godfrey, P. C., & Hatch, N. W. (2007). Researching corporate social responsibility: An
agenda for the 21st century. Journal of Business Ethics, 70(1), 87–98.
Gönül, F., & Srinivasan, K. (1993). Modeling multiple sources of heterogeneity in
multinomial logit-models — methodological and managerial issues. Marketing
Science, 12(3), 213–229.
Greene, W. H. (2003). Econometric analysis (5th ed.)Upper Saddle River, N.J: Prentice
Hall, 440–441.
Hillman, A. J., & Keim, G. D. (2001). Shareholder value, stakeholder management, and
social issues: What's the bottom line? Strategic Management Journal, 22(2), 125–139.
Kamakura, W. A., & Russell, G. J. (1989). A probabilistic choice model for
market-segmentation and elasticity structure. Journal of Marketing Research, 26(4),
379–390.
Klein, J., & Dawar, N. (2004). Corporate social responsibility and consumers' attributions
and brand evaluations in a product–harm crisis. International Journal of Research in
Marketing, 21(3), 203–217.
Krishna, A. (2011). Can supporting a cause decrease donations and happiness? The cause
marketing paradox. Journal of Consumer Psychology, 21(3), 338–345.
Lichtenstein, D. R., Drumwright, M. E., & Braig, B.M. (2004). The effect of corporate social
responsibility on customer donations to corporate-supported nonprofits. Journal of
Marketing, 68(4), 16–32.
Lindquist, J.D. (1974). Meaning of image. Journal of Retailing, 50(4), 29–38.
Luo, X. M., & Bhattacharya, C. B. (2006). Corporate social responsibility, customer
satisfaction, and market value. Journal of Marketing, 70(4), 1–18.
Luo, X. M., & Bhattacharya, C. B. (2009). The debate over doing good: Corporate social
performance, strategic marketing levers, and firm-idiosyncratic risk. Journal of
Marketing, 73(6), 198–213.
Manget, J., Roche, C., & Munnich, F. (2009). Capturing the green advantage for consumer
companies. Boston Consulting Group Report.
Margolis, J.D., & Walsh, J. P. (2003). Misery loves companies: Rethinking social initiatives
by business. Administrative Science Quarterly, 48(2), 268–305.
Mazursky, D., & Jacoby, J. (1986). Exploring the development of store images. Journal of
Retailing, 62(2), 145–165.
McWilliams, A., & Siegel, D. (2001). Corporate social responsibility: A theory of the firm
perspective. Academy of Management Review, 26(1), 117–127.
Orlitzky, M., Schmidt, F. L., & Rynes, S. L. (2003). Corporate social and financial
performance: A meta-analysis. Organization Studies, 24(3), 403–441.
Raghubir, P., Roberts, J., Lemon, K. N., & Winer, R. S. (2010). Why, when, and how should
the effect of marketing be measured? A stakeholder perspective for corporate social
responsibility metrics. Journal of Public Policy & Marketing, 29(1), 66–77.
Rindfleisch, A., Malter, A. J., Ganesan, S., & Moorman, C. (2008). Cross-sectional versus
longitudinal survey research: Concepts, findings, and guidelines. Journal of
Marketing Research, 45(3), 261–279.
Sen, S., & Bhattacharya, C. B. (2001). Does doing good always lead to doing better?
Consumer reactions to corporate social responsibility. Journal of Marketing Research,
38(2), 225–243.
Talukdar, D., Gauri, D. K., & Grewal, D. (2010). An empirical analysis of the extreme cherry
picking behavior of consumers in the frequently purchased goods market. Journal of
Retailing, 86(4), 336–354.
The Kroger Company (2009). Kroger Records Fourth Quarter and Record Full Year 2008
Results. Accessed on January 30, 2011, at http://www.thekrogerco.com/corpnews/
corpnewsinfo_pressreleases_03102009.htm
Urbany, J. E., Dickson, P. R., & Kalapurakal, R. (1996). Price search in the retail grocery
market. Journal of Marketing, 60(2), 91–104.
van Heerde, H., Gijsbrechts, E., & Pauwels, K. (2008). Winners and losers in a major price
war. Journal of Marketing Research, 45(5), 499–518.
Verhoef, P. C., Neslin, S. A., & Vroomen, B. (2007). Multichannel customer management:
Understanding the research shopping phenomenon. International Journal of
Research in Marketing, 24(2), 129–148.
Wedel, M., & Kamakura, W. A. (2000). Market segmentation: Conceptual and
methodological foundations (2nd ed.)Boston: Kluwer Academic.
Zhao, X., Lynch, J. G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and
truths about mediation analysis. Journal of Consumer Research, 37, 197–206.
Intern. J. of Research in Marketing 31 (2014) 168–177
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
The market value for product attribute improvements under
price personalization☆
Garrett P. Sonnier
University of Texas at Austin, 1 University Station, Austin, TX 78712, United States
a r t i c l e
i n f o
Article history:
First received in 18 July 2012 and was under
review for 4 ½ months
Available online 14 October 2013
Area Editor: John H. Roberts
Keywords:
Product management
Pricing
Personalization
Choice models
a b s t r a c t
Personalization of the marketing mix is a topic of much interest to marketing academics and practitioners. Using
discrete choice demand theory, we investigate the aggregate market value for product attribute improvements
when firms are engaged in personalized pricing. Our results provide a theoretically grounded rule for how to
aggregate consumer valuations to assess the overall profitability of attribute improvements under price
personalization. Under common pricing, each consumer contributes the same margin. Profitability of an attribute
improvement is thus driven by inducing more consumers to buy. Consumers with high choice probabilities are
given less weight in the market valuation under common pricing as they are less responsive to attribute
improvements. Under personalized pricing, profitability of an attribute improvement is driven by extraction of
consumer surplus from high valuation consumers. Consumers with higher valuations, and consequently higher
choice probabilities, are given more weight in the market valuation under personalized pricing. Since individual
consumers play a more central role in the market valuation under personalized pricing, estimation of consumerlevel valuations is of increased importance. Under common pricing, the market valuation for an attribute
improvement is robust to extreme estimates of the consumer-level valuations. Through our theoretical and
empirical analyses, we demonstrate that this robustness does not hold under personalized pricing.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
New product development is crucial to sustained firm performance.
Companies that fail to develop new products risk being supplanted by
more nimble competitors responding to shifts in consumer demand.
While new companies often focus on creating disruptive technologies
that alter the competitive landscape, most new product development
activity focuses on incremental innovation devoted to improving
existing products. For example, at Sony, over three quarters of new
product activity is dedicated to improving existing products (Kotler &
Keller, 2006). Bayus (1994) notes the existence of a similar pattern
across a range of industries (Abernathy & Utterback, 1978) as well as
evidence that incremental innovation is more crucial to profitability
than breakthrough technology (Gomory, 1989). While new product
development is undeniably important, it is also risky. Some studies
suggest a failure rate of 95% in the U.S. (Kotler & Keller, 2006). To
improve the odds of success, product managers must carefully assess
how consumers value product attribute improvements and, importantly,
how to aggregate consumer valuations into a market-level valuation
useful for product planning decisions.
From the perspective of an individual consumer, the value for a
product attribute improvement is typically defined as the change in
price that would keep consumer utility constant given the attribute
improvement (Train, 2003). Appealing to discrete-choice theory of
☆ The author wishes to thank Elie Ofek for providing access to data and for helpful
comments.
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.09.002
consumer and firm behavior, Ofek and Srinivasan (2002) derive a
market-level analog to this consumer-level valuation termed the
market value for an attribute improvement (MVAI). MVAI can be
compared to the marginal cost of the attribute improvement, providing
product managers with guidance in assessing the overall profitability of
the improvement. However, the Ofek and Srinivasan (2002) derivation
of MVAI assumes that firms charges a common price to all consumers. In
contrast to a homogenous pricing policy, the notion of personalized
pricing is of great appeal to both marketing academics and managers
(Fay, Mitra, & Wang, 2009). A stream of research in the marketing
literature has considered the personalization of the marketing mix
from both an empirical and theoretical perspective (Chen & Iyer,
2002; Choudhary, Ghose, Mukhopadhyay, & Rajan, 2005; Heilman,
Kaefer, & Ramenofsky, 2003; Khan, Lewis, & Singh, 2009; Knox &
Eliashberg, 2009; Liu & Zhang, 2006; Rossi, McCulloch, & Allenby,
1996; Shaffer & Zhang, 2002). Firms from the apparel, airline, bank
issued credit-card, and enterprise software industries have engaged in
personalized pricing (Choudhary et al., 2005; Montgomery & Smith,
2009; Shaffer & Zhang, 2002). In light of academic and practitioner
attention to the topic of personalized pricing, it is interesting to consider
whether and how price personalization affects the market value for
product attribute improvements.1
1
Rather than focusing on the normative question of whether or not firms should
engage in price personalization, we adopt a positive point of view to understand the
implications of engaging in one-to-one price personalization for estimates of the market
value for a product attribute improvement.
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
The main contribution of this paper is to derive the market value for
product attribute improvements when firms are engaged in price
personalization. Our results generalize the MVAI measure for common
pricing and provide managerial guidance on product planning decisions
under personalized pricing. Similar to Ofek and Srinivasan's (2002)
analysis of MVAI under common pricing, we obtain closed form
expressions for MVAI under personalized pricing in the context of the
widely used multinomial logit demand model. However, two important
differences in MVAI under common versus personalized pricing
emerge from our analysis. First, under common pricing, every consumer
contributes the same margin. Incremental profitability from an attribute
improvement is thus driven by inducing more consumers to purchase.
Consumers with extreme choice probabilities are given less weight in
the aggregate market valuation as these consumers are less responsive
to attribute changes. In contrast, under personalized pricing, the
profitability of an attribute improvement is driven by the extraction of
surplus from consumers with higher valuations and, consequently,
higher choice probabilities. Under personalized pricing, consumers
with high choice probabilities are given greater weight in the market
valuation. The first difference between market-level valuations under
common and personalized pricing (i.e., which consumers matter more
for the aggregate market valuation) relates to the second difference.
As individual consumers matter more under personalized pricing,
extreme consumer-level valuations have a greater impact in this setting.
Unlike the case of common pricing, computing MVAI under personalized
pricing requires more careful attention to the estimation of the
consumer-level valuations, a point underscored by the results of our
empirical application.
Choice models specified with additive linear utility imply that the
consumer-level valuation for an attribute improvement is identified as
the ratio of the estimated attribute and price coefficients (Train,
2003). With a heterogeneous model, the distribution of consumerlevel valuations is specified indirectly as a ratio of random coefficients.
Such an identification strategy may yield distributions of the valuations
that lack finite moments (Daly, Hess, & Train, 2012). Even if finite
moments are assured, the distribution may be prone to yield extreme
estimates (Meijer & Rouwendal, 2006; Ofek & Srinivasan, 2002).
Alternatively, the valuations can be directly identified in the choice
model likelihood which avoids ratio estimation and its associated
problems (Cameron & James, 1987; Jedidi, Jagpal, & Manchanda, 2003;
Sonnier, Ainslie, & Otter, 2007). An interesting and important property
of MVAI under common pricing is its robustness to extreme consumer
valuations (Ofek & Srinivasan, 2002) which renders the estimation of
the consumer-level valuations less important. Our results demonstrate
that robustness to outliers is not a general property of the MVAI
measure and does not hold under personalized pricing. Using Ofek
and Srinivasan's (2002) data set on stated preferences for portable
camera mounts we empirically investigate the MVAI under personalized pricing. Computing MVAI under personalized pricing with ratio
estimates of the consumer-level valuations suggests that nearly every
attribute improvement is profitable for any product. In contrast, using
consumer-level valuations that are directly identified and less prone
to extreme estimates to compute MVAI under personalized pricing
yields estimates that are smaller in magnitude and suggest a smaller
subset of profitable attribute improvements.
The remainder of the paper is organized as follows. We begin with a
discussion of personalized pricing to motivate the study of product
planning decisions under one-to-one pricing. We then review the
derivation of the market valuation for an attribute improvement under
common pricing and extend the derivation to the case of one-to-one
price personalization. In doing so, we also consider the intermediate
case of a discrete segment-based price discrimination strategy. We
then discuss discrete choice demand models and the specification of
consumer-level valuations used to compute the market-level valuation
under personalized pricing. Our empirical application follows. The final
section summarizes and concludes.
169
2. Personalized pricing in marketing
The marketing literature has discussed numerous examples of
personalized marketing in both consumer and business-to-business
markets. Choudhary et al. (2005) discuss examples of firms in the
enterprise software industry, such as IBM, Hewlett–Packard, and Sun
Microsystems, that use personalized pricing discounts for products of
the same quality. In consumer markets, information technology has
enabled firms to develop rich databases of consumer information
giving firms the ability to reach individual consumers and personalize
the marketing mix. Direct marketing firms such as Land's End and
L.L. Bean use promotional discounts to tailor prices to individual
households (Shaffer & Zhang, 2002). Firms in the bank issued credit
card industry, such as Wells Fargo, engage in price personalization
through personalized discounts on card fees (Choudhary et al., 2005).
The consulting firm Accenture offers clients a personalized pricing tool
to assist in implementing a one-to-one price promotion program.2 A
CNN.com report details price variation across consumers for the same
product in a variety of online product categories, including airline
tickets, digital cameras, and personal computers.3 The online data
provision company Lexis–Nexis sells to different consumers at different
prices (Ghose & Huang, 2009). Even when met initially with consumer
resistance, firms such as Amazon continue to find innovative ways to
implement personalized pricing, such as the Gold Box (Choudhary
et al., 2005).
A challenge in implementing a personalized pricing strategy is that
firms must obtain consumer willingness-to-pay for the products in the
competitive set. Fay et al. (2009) consider conditions under which
firms invest in technology to solicit preferences from consumers at the
point of purchase versus technology that allows the firm to infer
preferences based on past observations. Wertenbroch and Skiera
(2002) discuss different methods for determining consumer valuations,
or willingness-to-pay, in market research. These methods include
Vickery auctions, the Becker–DeGroot–Marshak (BDM) elicitation procedure, and discrete choice models applied to either stated preference
data or market transaction data. Cameron and James (1987), Jedidi
et al. (2003), and Ofek and Srinivasan (2002) use discrete choice models
to estimate consumer valuations for product attributes. Most empirical
applications of personalized marketing also utilize discrete choice
models (Ansari & Mela, 2003; Khan et al., 2009; Knox & Eliashberg,
2009; Rossi et al., 1996; Zhang & Krishnamurthi, 2004; Zhang &
Wedel, 2009). An advantage of using discrete choice models is that
with an attribute based utility function (Fader & Hardie, 1996), the
valuation for the product can easily be decomposed into the valuations
for the product attributes. Furthermore, if the valuations can be linked
to consumer characteristics, such as demographics or purchase history,
the model can be used to impute the valuations for new consumers
conditional on the characteristics enhancing the firm's ability to
implement a personalized pricing strategy (Rossi et al., 1996).
In considering the question of whether and how the firm's pricing
strategy affects the market value for product attribute improvements
it is natural to address the problem from the perspective of firms selling
direct to consumers. Shaffer and Zhang (2002) study one-to-one
promotions among competing direct marketing firms. Chen and Iyer
(2002) study competition among firms that offer personalized prices
assuming that firms have an imperfect ability to reach consumers.
Choudhary et al. (2005) consider how price personalization in a
duopoly impacts firm choices over product quality. It is important to
note, though, that selling through a retailer does not preclude the
2
Accenture.com, http://www.accenture.com/NR/rdonlyres/6EFFD307-3CBE-40AEB1929F7FADC5776/0/personalized_pricing_tool.pdf, retrieved on Dec 12, 2009.
3
CNN.com, http://www.cnn.com/2005/LAW/06/24/ramasastry.website.prices/, retrieved
on Dec 12, 2009.
170
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
manufacturer from engaging in personalized pricing. Gerstner, Hess,
and Holthausen (1994) analyze a manufacturer targeting pull discounts
in the form of coupons or rebates to price sensitive consumers
purchasing through a retailer. Liu and Zhang (2006) study personalized
pricing in a channel where both the manufacturer and the retailer can
personalize price and the manufacturer can open a direct to consumer
channel. Silva-Risso and Ionova, (2008) study customized manufacturer
incentives in the automotive industry, where manufacturers spend
approximately $45 billion per year on sales incentives directed at
consumers.
3. Theoretical analysis
3.1. The market value for an attribute improvement under pricing common
to all consumers
We begin by reviewing the derivation of the market value for an
attribute improvement (MVAI) under common pricing (Ofek &
Srinivasan, 2002). Assume a market consisting of i = 1,…,I consumers
choosing amongst a set of m = 0,…,M products (where 0 denotes the
“outside” alternative). Let product m be defined by a vector of
continuously differentiable product attributes, xm, and a common
price, pm.4 The share of consumers predicted to choose product m
I
from the competitive set is Sm ¼ 1I ∑ Pr½yim ¼ 1# where yim = 1 denotes
i¼1
the choice of product m and Pr[yim =1] is the choice probability. Assume
that competing firms sell only one product (such that m also indexes
firms) and that fixed costs are zero. The profits from product m, πm,
I
are given by πm ¼ ∑ Prim ½pm −cm # ¼ I $ Sm ½pm −cm # where cm is the
i¼1
variable cost. Note that the aggregation of the choice probabilities
into market shares prior to multiplication with the margin is
possible in this setting because the prices are common across all
consumers. The firm's first order condition for the pricing decision is
∂πm
∂Sm
¼ Sm þ ∂p
½pm −cm # ¼ 0 . Now consider the total change in
m
∂pm
profitability of product m triggered by a change in the kth product
attribute, xkm. The total derivative of profits with respect to a change in
!
"
dπm
∂Sm
∂cm
xkm is k ¼ I
½pm −cm #−Sm k . After substitution of the pricing
k
dxm
∂xm
∂xm
first order condition, the total derivative of profits with respect to
2 2 3
3
∂Sm
dπm
∂cm 5
∂xkm 5
4
4
the attribute change is given by k ¼ ISm − ∂S − k .
m
dxm
∂xm
∂pm
Under common pricing, incremental profitability hinges on the
changes in market share in response to the attribute improvement
and price. As each consumer contributes the same margin, the
profitability of an attribute improvement will ultimately depend on
inducing more consumers to purchase the product. For the attribute
change to be profitable to the firm, the ratio of market share changes
with respect to the attribute and price must exceed the marginal cost
of the attribute change. Ofek and Srinivasan (2002) term this ratio of
∂Sm
∂xk
market shares, − ∂Sm , the market value for an attribute improvement
m
(MVAI).
5
∂pm
3.2. The market value for an attribute improvement and market
share simulators
Managers often use the parameter estimates from a discrete choice
model to build a market share simulator. The simulator can be used to
assess the sensitivity of market share to changes in price and product
attributes. Under common pricing, simulation techniques can also be
used to compute the price increase given an attribute improvement
that leaves aggregate market share constant. The manager first improves
the level of the product attribute then searches for the price change that
would leave market share unchanged. Ofek and Srinivasan (2002) show
that for common pricing this approach coincides with the MVAI. The
total differential of market share with respect to the kth product attribute
∂Sm
∂Sm
and price is dSm ¼ k dxkm þ
dp . The price change that satisfies
∂xm
∂pm m
dSm = 0, which is the price change given an attribute change that holds
∂Sm
market share constant, is
dpm
∂xk
¼ − ∂Sm which is exactly the MVAI under
m
dxkm
∂pm
common pricing.
Consider now the case of personalized pricing. Rather than finding
the incremental change in the common price that equalizes market
share before and after the attribute improvement, the manager seeks
the incremental change in the personalized price that leaves the
individual's choice probability unchanged. This consumer specific price
change can also be approximated via simulation. The manager changes
the product attribute then searches for the personalized price change
that would leave the individual choice probability unchanged. However,
at the end of this exercise the manager is left with a set of consumer-level
quantities approximate to the consumer-level valuations from the choice
model. The question of how to aggregate these quantities into a marketlevel value to assess the profitability of the attribute improvement
remains. Intuitively, the attribute change will be profitable to the firm if
the sum of the expected incremental prices that can be captured from
consumers exceeds the sum of the expected costs of the improvement.
3.3. The market value for an attribute improvement under
price personalization
Before addressing one-to-one personalized pricing, it is useful to
dwell on whether and how MVAI under common pricing would differ if
the firm engaged in a more discrete price discrimination strategy.
Under such a strategy, the firm might offer product m at different prices
to discrete segments of consumers. Assume there are d=1,…,D segments
of consumers each of size Id and each of which receive a price of pdm. Let Sdm
represent the share of product m in segment d. The profits from product
" d
#
h
i
h
ii
D
I
D h
d
d
m would then be π m ¼ ∑ ∑ Prim pm −cm ¼ ∑ Id $ Sdm pdm −cm .
d¼1 i¼1
4
As with Ofek and Srinivasan's (2002) derivation of MVAI under common pricing, our
conceptual analysis considers continuously differentiable product attributes, such as fuel
economy for automobiles or processor speed for personal computers. Some product
attributes are, of course, discrete. For the case of discrete product attributes, simulation
techniques would be required to assess the profitability of an attribute improvement.
For example, in the multinomial logit model, there is a closed form expression for the
derivative of the choice probability with respect to a change in a continuous product
attribute. For a change in a discrete product attribute, the effect on the choice probability
is computed as the difference in choice probabilities with the new and original level of the
product attribute.
d¼1
The derivative of the profit function is obtained by summing the
segment specific derivatives. After substitution of the pricing first
order condition, the total derivative of profits with respect to the
2
2
2 2 d3
333
∂Sm
attribute change is given by
dπm
61 D 6
6 6 ∂xk 7 ∂cm 777
¼ I4 ∑4I d Sdm 4−4 md 5− k 555 .
k
∂Sm
I d¼1
dxm
∂xm
∂pdm
Thus, the MVAI under a discrete price discrimination strategy is given
5
See Ofek and Srinivasan (2002) for a more complete discussion of MVAI under
common pricing.
171
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
2
2
2
∂Sdm
333
6 D 6
6 ∂xk 777
by 1I 4 ∑4I d Sdm 4− md 555. Each segment MVAI is computed exactly the
d¼1
∂Sm
∂pdm
same as MVAI under common pricing (i.e., as the ratio of market share
derivatives) and is weighted by the segment size and market share
within the segment.
Consider now the profits from product m under personalized
I
I
i¼1
i¼1
pricing, π m ¼ ∑ π im ¼ ∑ Prim ½pim −cm # .6 Unlike common pricing or
segment pricing, under price personalization the choice probabilities
cannot be aggregated into market shares prior to multiplication with
the margin. Profits are obtained in this case by summing over the
product of each individual consumer's purchase probability and the
consumer specific contribution margin. Thus, we may view segment
or common pricing as special cases of the more general case of
personalized pricing, where analysis of the two former situations is
simplified by the ability to aggregate the choice probabilities into shares
prior to multiplication with the common or segment margins. For each
consumer, the firm's first order condition for the pricing decision under
personalization is
∂πim
∂Prim
¼ Prim þ
½p −c # ¼ 0:
∂pim
∂pim im m
ð1Þ
The total derivative of profits with respect to the attribute change is
I
dπ m ∂π m X
∂π m dpim
¼ k þ
:
k
dxm ∂xm i¼1 ∂pim dxkm
ð2Þ
I
Since πm ¼ ∑ πim , the second term in this equation becomes
i¼1
∂πim dpim
dπm
which is zero by the first order condition. Thus,
¼
∂pim dxkm
dxkm
!
"
I
∂Prim
∂cm
½pim −cm #−Prim k . Plugging in the expression for [pim − cm]
∑
k
∂xm
i¼1 ∂xm
from the first order condition and rearranging terms yields the following
condition
I
∑
i¼1
2
3
3
∂Prim
"
#
I
6 ∂xk 7
61 X
dπ m
∂cm 7
6
7
6
7
¼ I6
Prim 6− m 7−Sm
7:
k
4
5
4
I
∂Prim
dxm
∂xkm 5
i¼1
∂pim
2
ð3Þ
Since each consumer has a unique contribution margin under price
personalization, the firm cares much more about which consumers are
induced to purchase. High valuation consumers are more likely to
tolerate higher prices and yield higher margins. Thus, incremental
profits depend not solely on attracting more customers (as is the case
under common pricing) but more on the extraction of consumer surplus
from buyers with larger valuations for the attribute improvement.
Under personalized pricing the ratio of consumer choice probability
derivatives with respect to the attribute and price determines the
MVAI. More specifically, the attribute improvement will be profitable
if the weighted average of these ratios exceeds the marginal cost
weighted by the product's market share. Specifically, MVAI under
personalized pricing is
MVAI
per
¼
I
1X
I
i¼1
2
3
∂Prim
6 ∂xk 7
6
7
Prim 6− m 7:
4 ∂Prim 5
∂pim
3.4. MVAI under personalized pricing and the multinomial logit
(MNL) model
We now discuss expressions for MVAI under personalized
pricing implied by the widely used multinomial logit model. Suppose
we observe the choices of the i = 1,…,I consumers on a set of t =
1,…,T choice occasions. Assume a linear indirect utility function,
V imt ¼ x′ imt ϕ( −α ( pimt þ ε(imt , with error term εimt⁎ ~ EV(0,μ*). It is
well known that the utility function can be multiplied by a constant
without changing the consumer's utility maximizing choice. This
scale identification problem is typically addressed by estimating
(
(
the parameters ϕ ¼ ϕμ ( and α ¼ αμ ( , normalizing utility by the scale
parameter of the error distribution (Swait & Louviere, 1993; Train,
2003). The choice probabilities are
Pr½yimt
2
3
6
7
6
exp½x′ imt ϕ−αpimt # 7
7:
¼ 1# ¼ 6
6
M
X
#
$7
4
5
exp x′ ilt ϕ−αpilt
1þ
ð5Þ
l¼1
Parametric distributions of heterogeneity are easily incorporated
%
&
into the analysis.
Fori example, one could specify θi e MVN θ; Σθ
h
′
where θi ¼ ϕ i ln ðα i Þ . Following Eq. (4), with heterogeneous θ the
MVAI for the kth product attribute under personalized pricing is
given by
per
MVAIθ
¼
" #
k
I
1X
ϕ
Prim i :
I i¼1
αi
ð6Þ
Under personalized pricing the MVAI is the average of the
consumer-level valuations weighted by the choice probabilities.7 In
this specification the distribution of the consumer-level valuations is
identified indirectly as the ratio of the random attribute and price
coefficients, αϕii (Train, 2003). While commonly employed, unfortunately
not much can be said in favor of such an identification strategy in the
context of MVAI under personalized pricing. The heterogeneity
distribution for the attribute and price coefficients implies a distribution
for the ratio which will generally be different from that specified for the
coefficients. For example, a normal distribution on the coefficients does
not imply a normal distribution on the ratio. The implied heterogeneity
distribution may reflect a prior belief that the researcher has no
intention of expressing. Since ratios of random variables are generally
heavy tailed, the researcher using indirect identification is implicitly
(and perhaps unwittingly) imparting a prior belief that the distribution
of consumer-level valuations is heavy tailed. Furthermore, it is not at all
clear that the implied distribution possesses finite moments (Daly et al.,
2012). Even if the heterogeneity distribution does possess finite
moments, it is clear that consumers with estimates of αi tending
towards zero will be problematic in this setting as their valuations will
tend to be very large. Such consumers will inflate the market value for
the product attribute improvement. Indeed, only a handful of such
consumer-level valuations would likely result in the MVAI exceeding
the share weighted marginal cost, suggesting a profitable attribute
improvement.
Given the significance of the consumer-level valuations in the MVAI
under personalized pricing, it seems advantageous to parameterize
(
the model to directly identify the valuations. 8 We can estimate β ¼ αϕ(
(
and μ ¼ αμ ( , normalizing by the price coefficient and directly identifying
ð4Þ
6
We will return to the question of how the firm implements personalized pricing in our
empirical application.
7
Interestingly, the weighted average approach has been suggested as an ad-hoc
aggregation rule for consumer-level valuations (Ofek & Srinivasan, 2002).
8
The interested reader is directed to Sonnier et al. (2007) for a more detailed discussion
of direct and indirect estimation of consumer-level valuations.
172
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
the consumer-level valuation via β (Cameron & James, 1987; Sonnier
et al., 2007). The choice probabilities are
Pr½yimt
2
! ′
" 3
x imt β−pimt
6 exp
7
6
7
μ
7:
¼ 1# ¼ 6
!
"
6
7
M
′
X
4
x ilt β−pilt 5
exp
1þ
μ
l¼1
4. Empirical application
ð7Þ
An advantage of direct identification is that the heterogeneity
distribution is specified directly on the consumer-level valuations.
For example, one could specify normally distributed valuations via
h
i
%
&
λi e MVN λ; Σλ where λi ¼ β′ i ln ðμ i Þ . This would place less
prior mass in the tails of the distribution of the valuations tamping
down on outlier valuations. Alternatively, if the researcher believes the
distribution of valuations is in fact thick-tailed, a heterogeneity
distribution that reflects this belief, such as the t-distribution, may be
utilized. The valuations may also be modeled as a function of
demographics or other consumer-level covariates, allowing for the
prediction of valuations for future consumers conditional on this
information.
For heterogeneous λ, the market value for an improvement in the
kth product attribute under one-to-one price personalization is
computed as
per
MVAIλ ¼
I
h i
1X
k
Prim βi :
I i¼1
profitability from an attribute improvement is driven by the extraction of
surplus from these consumers.
ð8Þ
The expression for MVAI in Eq. (8) makes use of the directly
indentified consumer valuations and avoids potential problems
associated with ratio estimates of the valuations. From Eqs. (7) and (8)
we see that for MVAI under personalized pricing the scale of error
variance, captured by the parameter μ, plays a role similar to that in the
case of common pricing. As the effect of the error variance increases,
the ability of the valuations to explain consumer choices diminishes. In
1
the extreme, as μ → ∞ the value of Prim approaches 1þM
. As the effect of
the error variance decreases, the probability of choosing the alternative
with the highest valuation increases. Under common pricing, MVAI
gives smaller weight to such high value, high probability consumers
since the weight Prim[1 − Prim] reaches its maximum value at Prim =
0.5 (Ofek & Srinivasan, 2002). Under common pricing consumers very
likely to buy product m are given a smaller weight in the market
valuation for an attribute improvement compared with consumers who
are indifferent between product m and the composition of all other
products. Under personalized pricing, the consumer-level valuations
are weighted by the choice probabilities, Prim. Consumers very likely to
buy product m are given a larger weight in the market valuation.
Since the choice probability increases in the valuations, consumers
with higher valuations for the attributes of product m are also more likely
to be consumers with higher choice probabilities for product m. Under
common pricing, the firm can raise the price of a product subsequent to
an attribute improvement to capture surplus from higher value
consumers, but does so at the risk of losing consumers with lower
attribute valuations and choice probabilities. By the nature of the
S-shaped logit probability response curve, consumers with choice
probabilities away from zero or one will exert most of the influence on
the changes in market share with respect to the attribute and price,
which ultimately determines MVAI under common pricing. MVAI under
common pricing reflects the importance of these consumers by giving
them higher weights. However, one-to-one price personalization allows
the firm to capture surplus from higher valuation, higher probability
consumers without driving lower valuation, lower probability consumers
away from the product. Thus, consumers with higher valuations, and
consequently higher choice probabilities, are given more weight in the
market valuation under one-to-one price personalization as incremental
Using a data set on consumer stated preferences for camera mounts
we explore estimates of MVAI under personalized pricing.9 A complete
description of the data and a thorough treatment of MVAI under common
pricing can be found in Ofek and Srinivasan (2002). A total of 302
respondents each rank 18 profiles. Each profile is described by 5
attributes and a price. In addition, respondents completed a holdout
ranking task with 4 profiles. We begin by considering heterogeneity
distributions for the directly identified valuations, βi. We consider normal
and t heterogeneity distributions for valuations. A normal heterogeneity
%
&
distribution, βi e N β; Σβ , places small prior mass on outlier consumer
valuations. In contrast to a normal heterogeneity distribution for the
%
&
valuations, a t distribution of heterogeneity, βi e t ν β; Σβ , permits
more prior mass in the tails. For both cases, we use a log-normal
heterogeneity distribution for μi. For the t-distribution, the degree of
freedom parameter ν must lie on the range (0,∞). We specify a lognormal prior for ν and treat it as an unknown parameter to be estimated.
We estimate the normal and t models with standard Markov–Chain
Monte Carlo methods, running the sampler for a total of 15,000
iterations, keeping the last 5000 for inference. Time series plots of the
model log-likelihoods indicate that this is sufficient for convergence. Insample fit measured by the Deviance Information Criteria (DIC)
(Spiegelhalter, Best, Carlin, & van der Linde, 2004) and holdout fit
measured by the log predictive density (LPD) suggest that the model
based on a t-distribution of heterogeneity outperforms the model based
on normal heterogeneity (DIC of 15,994 vs. 16,689 and LPD of −689 vs.
−691 for the t vs. normal model, respectively).
Table 1 presents the attributes and levels of the five products used in
our MVAI analyses along with the marginal costs of improvement for
each attribute. An issue to consider is how the firms might implement
personalized pricing in practice. An approach widely discussed in the
literature is to offer a personalized discount, via a coupon or rebate,
off of a regular common price (Rossi et al., 1996; Shaffer & Zhang,
2002). However, in a personalized marketing environment, firms
could forgo regular price altogether. Regular prices place an upper
bound on the price a consumer pays which limits the ability of the
firm to extract surplus from high value customers. Of course, this
issue could be resolved by simply charging a high regular common
price, at or above the maximum suggested consumer-level price, and
offering discounts accordingly. Shaffer and Zhang (2002) show that in
a competitive environment lower regular prices perform the important
function of limiting competitive poaching of a firm's high value
customers. From a practical vantage point, a regular common price
coupled with a personalized discount may also be a more feasible
strategy for firms to implement. In light of these issues, we consider a
personalized price discount, denoted by zim, off of the common prices
used in Ofek and Srinivasan (2002) and reported in Table 1.10 We
show in Appendix A that the MVAI under a personalized discount is
equivalent to the expression shown in Eq. (4). To find the optimal
personalized price discounts, we compute for each consumer the vector
of discounts that satisfies the first order condition with respect to zim
by finding the consumer-specific discount vector that minimizes
"'
(2 #
M
Prim
. As in Rossi et al. (1996), we
∑ ½pm −zim −cm #−
∂Prim =∂zim
m¼1
allow for the possibility that the optimal discount may be zero. Once
the optimal discounts are obtained, we compute MVAI accordingly.
9
While our empirical application makes use of stated preference data, computation of
MVAI is not limited to stated preference data. MVAI can also be computed from a choice
model calibrated on revealed preference data.
10
This can be viewed as an approximation to a two stage game where competing firms
choose a regular price in the first stage then price discounts in the second stage.
173
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
Table 1
Marginal costs, attribute levels, and common prices for camera mount product
simulations.
Attribute
Weight (tens of oz.)
Sizeb
Set up time (min.)
Stabilityc
Flexibilityd
Common prices for
competitive sete
Three products
Four products
Five products
Marginal cost
of attribute
improvementa
$4.90
$0.23
$1.41
$0.31
$0.26
UltraPod
Q-Pod
Gorilla
Pod
Camera
Critter
0.17
0.80
0.62
2.50
1.80
0.20
0.98
0.98
1.80
1.96
0.35
0.84
0.84
2.50
2.17
0.46
1.27
0.50
2.30
2.84
$8.84
$7.72
$7.15
$9.89
$9.22
$8.53
$9.53
$8.50
$7.75
Half
Dome
0.57
1.20
0.42
3.00
2.33
$8.22
$7.49
$10.39
Table 2 reports the estimates of MVAI under personalized pricing
implied by the model that directly identifies the valuations with a
t-distribution of heterogeneity. The estimates are reported for the
three product market in Ofek and Srinivasan (2002). Under common
pricing, Ofek and Srinivasan (2002) find that improvements in weight,
size, stability and flexibility are profitable for all three products.
Improvements in set up time are not profitable for any of the three
products. When firms are engaged in price personalization, MVAI
estimates suggest a smaller set of profitable attribute improvements.
As with common pricing, improvements in size and flexibility are
profitable for all three products. However, improvements in weight
fail to generate incremental profits for any of the products. Similarly,
improvements in stability are profitable only for Q-Pod and Gorilla Pod.
Interestingly, under personalized pricing improvement in set up time
is profitable for Gorilla Pod. These results reflect two effects at play
when firms move from common to personalized pricing. On the one
hand, personalization allows firms to capture more consumer surplus
versus common pricing. This implies that finding incremental
profitability from attribute improvements may be more difficult under
price personalization as firms are already wringing much of the surplus
from the market. On the other hand, moving to price personalization
may render some attributes that are unprofitable to improve under
common pricing profitable. This is due to the fact that profitability
under personalization is driven by capturing surplus from high value
Table 2
Market value for attribute improvements under price personalization: direct identification
of consumer-level valuations.
Attribute
UltraPod
Q-Pod
Gorilla Pod
Weight
1.70
(0.10)
0.55a
(0.04)
0.36
(0.03)
−0.31
(0.28)
0.48
(0.03)
1.04
(0.06)
0.63
(0.09)
0.44
(0.04)
1.09
(0.14)
0.57
(0.04)
0.94
(0.05)
0.39
(0.03)
1.26
(0.13)
0.79
(0.08)
1.18
(0.11)
Set up time
Stability
Flexibility
Attribute
UltraPod
Q-Pod
Gorilla Pod
Weight
1.16
(0.08)
0.50a
(0.04)
0.34
(0.02)
0.10
(0.03)
0.40
(0.02)
0.82
(0.05)
0.40
(0.03)
0.31
(0.03)
0.54
(0.10)
0.41
(0.03)
0.87
(0.07)
0.34
(0.03)
0.51
(0.06)
0.83
(0.22)
0.70
(0.08)
Size
Set up time
Stability
Flexibility
a
Marginal cost of weight in dollars to reflect coding of weight in tens of ounces. All
other marginal cost data are in tens of dollars.
b
Where 1 represents a camera mount that fits into a standard pocket and 3in a standard
book bag.
c
Where 1 means stable enough under light-medium wind conditions for a small
camera with a built-in lens and 3 for a full-size camera with a large lens.
d
Where 1 is low flexibility and 3 is high flexibility. Flexibility is the degree to which the
camera mount can be adapted to various terrains and adjusted for height and angle.
e
Prices in tens of dollars, as reported in Ofek and Srinivasan (2002).
Size
Table 3
Market value for attribute improvements under segment pricing: direct identification of
consumer-level valuations.
Table cells report the posterior mean (in tens of dollars) and posterior standard error
(in parenthesis).
a
Bold indicates that 95% of the distribution of the difference between the valuation and
the share weighted marginal costs is positive.
Table cells report the posterior mean (in tens of dollars) and posterior standard error (in
parenthesis).
a
Bold indicates that 95% of the distribution of the difference between the valuation and
the share weighted marginal costs is positive.
consumers. Under common pricing, firms are unable to capture this
value without driving down share. Price personalization frees the firm
from this constraint. As noted, we see both of these effects in our results.
It is interesting to compare the estimates of MVAI under personalized pricing with those under segment pricing. The segments
may be determined according to any of a number of bases (e.g.,
demographics, brand loyalty, or usage). For our illustration, we perform
a two-step cluster analysis on the posterior means of the attribute
valuations. This cluster analysis results in two segments. The first
segment is comprised of 87% of the respondents while the second
segment is comprised of 13% of the respondents. The mean attribute
valuations in the second segment are all higher than the mean
valuations in the first segment. We compute the optimal segmentspecific price discounts and then compute the MVAI under segment
pricing. The results appear in Table 3. The results under segment pricing
suggest largely the same smaller set of profitable attribute improvements as the results under personalized pricing. The exception is that
improvements in set up time are not profitable for Gorilla Pod under
segment pricing. In addition, the MVAI estimates under segment pricing
are, for the most part, smaller in magnitude.
While an indirect identification strategy on the valuations raises a
number of concerns, it is nonetheless instructive to dwell on such a
strategy. Normal heterogeneity distributions are often used
% by&
academics and
alike.11 We may specify θi e MVN θ; Σθ
h practitioners
i′
where θi ¼ ϕ′ i ln ðα i Þ . The log-normal distribution for αi ensures
that the consumer-valuations are well-defined (Daly et al., 2012).
However, the valuations are distributed as the ratio of normal and lognormal random variables and are likely to be heavy tailed. This
specification is thus analogous to using the thick tailed t-distribution
in the case of direct identification. An important difference though is
that under indirect identification, small price coefficients will lead to
valuations tending towards infinity. While such valuations have smaller
prior probability under direct identification, this is not necessarily so
under indirect identification. A draconian fix to this problem is to
restrict the price coefficient to be homogeneous across consumers.
The implied valuations are then identified as the normally distributed
attribute coefficients scaled by the homogeneous price coefficient. This
is analogous to the direct model using a normal heterogeneity
distribution on the valuations. However, such an approach cannot be
recommended as the price of normality in this case is the more
restrictive homogeneous specification on price responsiveness. While
degradation of model fit is a concern, a larger concern is bias in the
estimates of the price coefficient and hence the attribute valuations
(Chintagunta, Jain, & Vilcassim, 1991; Daly et al., 2012).
11
For example, Sawtooth Software's Hierarchical Bayes module for choice based
conjoint analysis uses normal heterogeneity distributions.
174
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
Valuation (tens of dollars)
3.50
Table 4
Market value for attribute improvements under price personalization: indirect
identification of consumer-level valuations.
3.00
2.50
Median
2.00
UltraPod
Q-Pod
Gorilla Pod
Weight
6.55a
(2.42)
2.72
(1.47)
1.96
(0.84)
−0.48
(3.35)
2.65
(1.15)
7.26
(2.20)
3.61
(1.85)
3.01
(1.07)
4.30
(1.61)
4.03
(1.28)
8.06
(2.50)
2.93
(1.87)
4.51
(1.73)
4.04
(2.04)
7.60
(2.93)
Size
1.50
Set up time
1.00
Stability
0.50
Flexibility
0.00
Weight
Size
Set Up Time Stability
Flexibility
Fig. 1. Inter-quartile range of consumer-level attribute valuations based on βi.
We estimate an indirect model with normal heterogeneity on the
attribute coefficients and log-normal heterogeneity on the price
coefficient. In-sample fit measured by the DIC is 16,319 while holdout
fit measured by the LPD is −713. Both in-sample and holdout fit is
inferior to the direct model with the t-distribution of heterogeneity. As
noted, previous research has shown indirect identification of the
consumer valuations to be prone to outliers. Figs. 1 and 2, which present
the median and inter-quartile ranges of the consumer level valuations
implied by the direct and indirect models, respectively, confirm that
this is indeed the case for our data. Fig. 1 corresponds to the direct
model. The 75th percentile values are in the neighborhood of $10–
$30. Fig. 2 corresponds to the indirect model. In Fig. 2, the valuations
are far more dispersed. The 75th percentile values range from about
$80–$250. The median valuations are all above the marginal costs
reported in Table 1. Table 4 presents the MVAI estimates under
personalized pricing for the valuations based on indirect identification.
The MVAI estimates imply that improving any of the attributes for
nearly all of the products is profitable. The MVAI estimates are also
much larger in magnitude compared to those based on the directly
identified valuations. Given the distribution of the estimates of the
consumer-level valuations under indirect identification, it is not
surprising that the MVAI estimates computed with these valuations
are large and suggest that nearly any attribute improvement will be
profitable. The results demonstrate that extreme consumer valuations
have a significant impact on MVAI under personalized pricing.
Table 5 presents the MVAI estimates under segment pricing for the
valuations based on indirect identification. As before, we perform a
two-step cluster analysis on the posterior means of the attribute
valuations which again results in two segments. The first segment is
comprised of 87% of the respondents with lower mean valuations
30.00
Valuation (tens of dollars)
Attribute
Table cells report the posterior mean (in tens of dollars) and posterior standard error (in
parenthesis).
a
Bold indicates that 95% of the distribution of the difference between the valuation and
the share weighted marginal costs is positive.
compared to the smaller second segment comprised of 13% of the
respondents. However, the average valuations for both segments are
much higher compared to the average valuations for the segments
derived from the directly identified valuations. This is to be expected
given the distributions shown in Figs. 1 and 2. The more interesting
issue is how the MVAI estimates under segment pricing compare to
those computed with the directly identified valuations. We compute
the optimal segment-specific price discounts and then compute the
MVAI under segment pricing. Although the indirectly identified
valuations are widely dispersed, as noted in Fig. 2, the MVAI estimates
under segment pricing are not impacted to the degree to which the
MVAI estimates under personalized pricing are affected. However,
under segment pricing, the MVAI estimates computed with the indirectly
identified valuations suggest that more attributes can be profitably
improved compared to those computed with the directly identified
estimates. The estimates imply that improvements in size, stability and
flexibility are profitable for all products and improvements in set up
time are profitable for Q-Pod and Gorilla Pod. In addition, the magnitude
of the estimates computed with the indirectly identified valuations,
although not as explosively large as those under personalized pricing,
are generally larger compared to those computed with the directly
identified valuations. While the MVAI estimates under segment pricing
are more robust compared with those under personalized pricing, it is
still the case that the estimates are adversely impacted by the widely
dispersed valuations resulting from an indirect identification strategy.
4.1. MVAI and competitive entry
As noted in the conceptual analysis, MVAI under both common and
personalized pricing depends upon the consumer choice probabilities,
Table 5
Market value for attribute improvements under segment pricing: indirect identification of
consumer-level valuations.
25.00
20.00
Median
Attribute
UltraPod
Q-Pod
Gorilla Pod
Weight
1.35
(0.12)
0.57a
(0.06)
0.48
(0.04)
0.18
(0.03)
0.66
(0.06)
1.35
(0.14)
0.61
(0.07)
0.56
(0.06)
0.64
(0.07)
0.79
(0.09)
1.68
(0.19)
0.69
(0.09)
0.80
(0.10)
0.73
(0.08)
1.23
(0.15)
15.00
Size
10.00
Set up time
Stability
5.00
0.00
Flexibility
Weight
Size
Set Up Time Stability
Flexibility
-5.00
h
Fig. 2. Inter-quartile range of consumer-level attribute valuations based on ϕ′ i
i
αi .
Table cells report the posterior mean (in tens of dollars) and posterior standard error (in
parenthesis).
a
Bold indicates that 95% of the distribution of the difference between the valuation and
the share weighted marginal costs is positive.
175
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
Table 6
The effect of increased product competition on the market value for attribute improvements under price personalization.
Three products
Four products
Five products
Attribute
UltraPod
Q-Pod
Gorilla Pod
UltraPod
Q-Pod
Gorilla Pod
Camera Critter
UltraPod
Q-Pod
Gorilla Pod
Camera Critter
Half Dome
Weight
1.70
(0.10)
0.55a
(0.04)
0.36
(0.03)
−0.31
(0.28)
0.48
(0.03)
36%
$8.84
54%
$7.45
1.04
(0.06)
0.63
(0.09)
0.44
(0.04)
1.09
(0.14)
0.57
(0.04)
30%
$9.89
54%
$8.30
0.94
(0.05)
0.39
(0.03)
1.26
(0.13)
0.79
(0.08)
1.18
(0.11)
34%
$9.53
41%
$8.48
0.70
(0.04)
0.23
(0.02)
0.16
(0.02)
−0.38
(0.28)
0.30
(0.02)
17%
$7.72
43%
$6.86
0.44
(0.03)
0.36
(0.06)
0.20
(0.02)
0.71
(0.08)
0.39
(0.03)
16%
$9.22
56%
$7.97
0.55
(0.04)
0.23
(0.02)
0.81
(0.11)
0.31
(0.06)
0.94
(0.11)
20%
$8.50
33%
$7.91
1.99
(0.11)
0.76
(0.07)
0.88
(0.07)
0.94
(0.09)
0.61
(0.03)
47%
$8.22
59%
$6.72
0.41
(0.03)
0.19
(0.02)
0.11
(0.01)
−0.41
(0.33)
0.25
(0.02)
13%
$7.15
37%
$6.54
0.29
(0.02)
0.30
(0.06)
0.12
(0.01)
0.19
(0.02)
0.22
(0.03)
11%
$8.53
51%
$7.56
0.40
(0.03)
0.19
(0.02)
0.52
(0.05)
0.10
(0.04)
0.78
(0.09)
15%
$7.75
34%
$7.26
1.62
(0.09)
0.67
(0.07)
0.52
(0.04)
0.34
(0.02)
0.44
(0.02)
39%
$7.49
53%
$6.27
0.95
(0.06)
0.24
(0.02)
0.78
(0.11)
1.37
(0.17)
0.55
(0.04)
21%
$10.39
62%
$8.90
Size
Set up time
Stability
Flexibility
Market share
Regular price
% receiving discount
Average discounted price
Table cells report the posterior mean (in tens of dollars) and posterior standard error (in parenthesis).
a
Bold indicates that 95% of the distribution of the difference between the valuation and the share weighted marginal costs is positive.
although in different ways. Dependence on the choice probabilities
renders MVAI sensitive to competitive entry. The choice model
parameters, of course, do not change. However, competitive entry will
alter equilibrium prices, the choice probabilities and hence the MVAI
estimates. Under common pricing, the effect of expanding the
competitive set on the MVAI of existing products will depend on the
choice probabilities through the expression Prim[1 − Prim]. Ofek and
Srinivasan (2002) show that under common pricing, MVAI may
increase or decrease in response to an expansion of the competitive
set. For example, a price cut in response to entry may attract more
consumers with lower valuations and lower MVAI while products that
can maintain premium pricing in the face of entry may lose some
lower valuation consumers while retaining higher valuation consumers
thereby increasing the MVAI.
Consider now the effect of expanding the competitive set on MVAI
under personalized pricing. Since the consumer-level valuations are
directly weighted by the choice probabilities competitive entry will
reduce the probability of purchase for incumbent products, entry is
likely to reduce the incumbent firms' MVAI estimates. Table 6 presents
MVAI under personalized pricing when the choice set expands from
three to four products and then four to five products.12 Also listed in
Table 6, for each product, is the market share, the regular price, the
percentage of consumers receiving a discount, and the average
discounted price. When Camera Critter is added to the choice set, it
gains a considerable amount of market share at the expense of the
three incumbent products. Camera Critter is the dominant alternative
on weight and size, shares dominance on stability with Q-Pod, and
engages in personalized discounting with broader scope and scale. As
a result, Camera Critter obtains a 47% share upon entry. The MVAI
estimates for all the incumbent firms decrease. Gorilla Pod is the
dominant alternative on set up time and flexibility. Consequently, its
MVAI for these attributes does not decline as sharply. Indeed, its
advantage on flexibility is substantial and the Gorilla Pod MVAI for
flexibility remains the highest even after Camera Critter's entry. Half
Dome enters with dominance on set-up time and stability. Camera
Critter retains dominance on weight and size while Gorilla Pod retains
dominance on flexibility. Half Dome has the highest average discounted
price but still manages to obtain a 21% share. As expected, share declines
bring about declines in MVAI for the incumbent firms. However, Camera
Critter still has the highest MVAI for size and Gorilla Pod the highest
MVAI for flexibility.
12
For the purposes of this analysis we focus only on the model that directly identifies the
consumer-level valuations. See Table 1 for a description of all five products.
4.2. MVAI under asymmetric personalization
Personalization of the marketing mix is costly in terms of
information, computing, and administration (Rossi et al., 1996). In
light of this, firms will likely differ in their willingness and/or ability to
implement personalized pricing strategies. In this section, we examine
the impact of asymmetric personalization on MVAI estimates. We use
the term asymmetric personalization to refer to the situation where
some firms are engaged in personalization while other firms employ
common pricing. This is opposed to the case where all firms personalize,
which we term full personalization. To conduct the analysis, we assume
that after setting regular price, UltraPod and Q-Pod set personalized
discounts while Gorilla Pod sells at the regular price. We then compute
MVAI under personalized pricing for UltraPod and Q-Pod and MVAI
under common pricing for Gorilla Pod. The results are presented in
Table 7.
Under asymmetric personalization, UltraPod and Q-Pod offer
personalized discounts to over 60% of consumers resulting in an average
discounted price of $7.51 for UltraPod and $8.30 for Q-Pod. The average
discounted prices are close to those under full personalization and
considerably lower than Gorilla Pod's regular price of $9.53. As a result,
Gorilla Pod share drops to 28% while UltraPod and Q-Pod shares increase
to 37% and 35%, respectively. The MVAI estimates for UltraPod and Q-Pod
Table 7
Market value for attribute improvements under asymmetric price personalization.
Attribute
UltraPod
Q-Pod
Gorilla Pod
Weight
1.61
(0.10)
0.55a
(0.04)
0.36
(0.03)
−0.29
(0.27)
0.52
(0.03)
37%
$8.84
62%
$7.51
1.20
(0.06)
0.67
(0.09)
0.40
(0.03)
1.22
(0.14)
0.58
(0.04)
35%
$9.89
63%
$8.30
2.77
(0.13)
1.14
(0.07)
1.58
(0.10)
2.32
(0.34)
2.16
(0.12)
28%
$9.53
–
$9.53
Size
Set up time
Stability
Flexibility
Market share
Regular price
% receiving discount
Average discounted price
Table cells report the posterior mean (in tens of dollars) and posterior standard error (in
parenthesis). For m = 1,2 MVAI computed under personalized pricing. For m = 3 MVAI
computed under common pricing.
a
Bold indicates that 95% of the distribution of the difference between the valuation and
the share weighted marginal costs is positive.
176
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
(computed via the rule for MVAI under personalized pricing) increase
slightly, commensurate with the share increases. Consumers choosing
Gorilla Pod at the premium price are likely consumers with high
valuations for the product. Indeed, UltraPod and Q-Pod are unable to
profitably entice these consumers to switch even with a personalized
discount. This intuition is confirmed by the relatively high MVAI's for
Gorilla Pod (computed via the rule for MVAI under common pricing).
Improvements in size, stability, and flexibility are profitable for Gorilla
Pod under asymmetric personalization.
5. Summary and conclusions
Understanding the market value for product attribute improvements is crucial to successful product planning and new product
development. A measure of the consumer's value for an attribute
improvement is the increase in price that would leave utility unchanged
given the attribute improvement. A discrete choice model calibrated on
stated or revealed preference data is a popular method for estimating
consumer valuations. With heterogeneous consumers, the issue of
how to aggregate the consumer-level valuations into a market-level
valuation to assess profitability arises. Ad-hoc methods such as taking
the average may yield misleading results and, empirically, may suffer
from the effect of extreme valuations. Based on micro-economic theory
of consumer and firm behavior, Ofek and Srinivasan (2002) derive the
market valuation for an attribute improvement (MVAI) as the ratio of
changes in market share with respect to the attribute improvement
and price. Their derivation assumes the firm employs a common pricing
strategy, charging the same price to all consumers.
Marketing academics have long been interested in the effects of
personalizing the marketing mix (Rossi et al., 1996). Recently, online
channels have stimulated industry interest in and enabled more
widespread use of price personalization based on purchase history or
other information. We consider the market value for product attribute
improvements for the case of one-to-one price personalization. Our
results demonstrate how to assess the profitability of attribute improvements in this interesting and important setting. Compared with the
market valuation for an attribute improvement under common pricing,
two important differences emerge. First, under common pricing, the
profitability of an attribute improvement is driven by inducing more
consumers, each of whom contributes the same margin, to buy. Thus,
consumers with extreme choice probabilities are given less weight in
the market valuation under common pricing as these consumers are
less responsive to attribute improvements. Under personalized pricing,
the profitability of an attribute improvement is driven by the extraction
of consumer surplus from high value consumers. Thus, higher valuation
consumers with higher choice probabilities are given greater weight in
the market valuation under personalized pricing. Second, because the
individual consumers play a more central role in the market valuation
under personalized pricing, MVAI under one-to-one price personalization is not robust to extreme consumer-level valuations. Therefore,
when engaged in personalized pricing, the identification and estimation
of consumer-level valuations is of increased importance relative to the
case of common pricing.
With additive linear utility, consumer-level valuations are identified
as the ratio of attribute and price coefficients from the discrete choice
model. This identification strategy has been shown to yield distributions
for the valuations that lack finite moments in some cases and is
particularly prone to yield extreme valuations. A simple alternative is
to utilize a choice model directly identifies the valuations. Using a dataset
on consumer stated preferences for camera mounts, we demonstrate the
managerial relevance of our analysis. We estimate choice models that
directly and indirectly identify consumer-level valuations for product
attribute improvements. We then use these models to compute the
MVAI implied by both models under personalized pricing strategies.
Under personalized pricing, models that indirectly identify the
consumer-level valuations result in MVAI estimates that suggest nearly
any attribute improvement for all products considered is profitable. In
contrast, the model that directly identifies the consumer-level valuations
provides a better fit to the data and results in a smaller set of profitable
attribute improvements.
There are a number of avenues for future research. As noted, the
problem of discrete product attributes remains a challenge. Our
expressions for MVAI are based on a logit demand model. Future
research may consider other empirical models of demand. Recent
research investigates the price discrimination across multiple channels
(Wolk & Ebbling, 2010). Investigating product planning decisions in
the context of channel competition where manufacturers and retailers
each have the ability to personalize price would be very challenging
but may yield interesting insights. Lastly, our analysis considers single
product firms. Firms may offer different product attributes via vertically
differentiated product lines (Michalek, Ebbes, Adigüzel, Feinberg, &
Papalambros, 2011). The impact of price personalization on product
attribute decisions in a product line may be an interesting topic
to consider. Sorting out the market value for a product attribute
improvement in these cases should assist firms in making better
product planning decisions.
Appendix A. MVAI with personalized price discounts
Consider our firms engaged in personalized pricing by offering a
personalized discount, zim, off of the regular price, pm, common to all
consumers. Such a discount could be in the form of a personalized
coupon or a rebate. We will abstract away from targeting costs and
I
redemption issues. Profits to firm m are πm ¼ ∑ Prim ½pm −zim −cm #
i¼1
while the MNL choice probabilities in this setting are
2
3
6
7
6 exp½x′ im ϕi −α i ðpm −zim Þ# 7
7
Prim ¼ 6
6
M
X
# ′
$7
4
5
exp x il ϕi −α i ðpl −zil Þ
1þ
l¼1
2
! ′
" 3
x β −ðpm −zim Þ
6 exp im i
7
6
7
μi
¼6
! ′
"7
6
7:
M
X
4
x il βi −ðpl −zil Þ 5
1þ
exp
μi
l¼1
ðA1Þ
We assume the regular prices are observable to all firms when
choosing their personalized discounts. This is consistent with the
notion that regular prices are a high level managerial decision slow to
adjust in practice (Shaffer & Zhang, 2002). For each consumer, the
manufacturer's first order condition for the discounting decision is
∂πim
∂Prim
¼ −Prim þ
½p −z −c # ¼ 0:
∂zim
∂zim m im m
ðA2Þ
The total derivative of manufacturer profits with respect to the
attribute change is
dπm
dxkm
¼
"
I
X
∂Prim
i¼1
∂xkm
½pm −zim −cm #−Prim
∂cm
∂xkm
#
:
ðA3Þ
Plugging in the expression for [pm − zim − cm] from the first order
condition and rearranging terms yields the following condition
2
3
3
∂Prim
"
#
I
6
7
6
dπm
∂cm 7
6 ∂xkm 7
61 X
7
¼
I
Pr
−S
6
7
6
7:
im
m
4 ∂Prim 5
4 I i¼1
dxkm
∂xkm 5
∂zim
2
ðA4Þ
G.P. Sonnier / Intern. J. of Research in Marketing 31 (2014) 168–177
Inspection of Eq. (A4) reveals for heterogeneous choice models
h ki
I
ϕ
parameterized in the space of θ, the MVAI will be given by 1I ∑ Prim αii .
i¼1
For heterogeneous choice models parameterized in the space of λ, the
h i
I
MVAI will be given by 1I ∑ Prim βki .
i¼1
References
Abernathy, W., & Utterback, J. (1978). Patterns of industrial innovation. Technology
Review, 2, 41–47.
Ansari, Asim, & Mela, Carl F. (2003). E-customization. Journal of Marketing Research, 40,
131–145.
Bayus, B. (1994). Optimal pricing and product development policies for new consumer
durables. International Journal of Research in Marketing, 11, 249–259.
Cameron, T. A., & James, M.D. (1987). Estimating willingness-to-pay from survey data: An
alternative pre-test market evaluation procedure. Journal of Marketing Research, 24,
389–395.
Chen, Y., & Iyer, G. (2002). Consumer addressability and customized pricing. Marketing
Science, 21, 197–208.
Chintagunta, P., Jain, D. C., & Vilcassim, N. J. (1991). Investigating heterogeneity in brand
preferences in logit models for panel data. Journal of Marketing Research, 33(4),
417–428.
Choudhary, V., Ghose, A., Mukhopadhyay, T., & Rajan, U. (2005). Personalized pricing and
quality differentiation. Management Science, 51, 1120–1130.
Daly, A. J., Hess, S., & Train, K. E. (2012). Assuring finite moments for willingness to pay in
random coefficients models. Transportation, 39, 19–31.
Fader, P., & Hardie, B. (1996). Modeling consumer choice among SKUs. Journal of
Marketing Research, 33, 442–452.
Fay, S., Mitra, D., & Wang, Q. (2009). Ask or infer? Strategic implications of alternative
learning approaches in customization. International Journal of Research in Marketing,
26, 136–152.
Gerstner, E., Hess, J.D., & Holthausen, D.M. (1994). Price discrimination through a
distribution channel: Theory and evidence. American Economic Review, 84, 1437–1445.
Ghose, A., & Huang, K. (2009). Personalized pricing and quality customization. Journal of
Economics and Management Strategy, 18, 1095–1135.
Gomory, R. (1989). From the ladder of science to the product development cycle. Harvard
Business Review, 89, 99–105.
Heilman, C. M., Kaefer, F., & Ramenofsky, S. D. (2003). Determining the appropriate
amount of data for classifying consumers for direct marketing purposes. Journal of
Interactive Marketing, 17, 5–28.
177
Jedidi, K., Jagpal, S., & Manchanda, P. (2003). Measuring heterogeneous reservation prices
for product bundles. Marketing Science, 22, 107–130.
Khan, R., Lewis, Michael, & Singh, Vishal (2009). Dynamic customer management and the
value of one-to-one marketing. Marketing Science, 28, 1063–1079.
Knox, G., & Eliashberg, J. (2009). The consumer's rent vs. buy decision in the rentailer.
International Journal of Research in Marketing, 26, 125–135.
Kotler, P., & Keller, K. L. (2006). Marketing management (12e). New Jersey: Pearson
Prentice Hall.
Liu, Y., & Zhang, J. (2006). The benefits of personalized pricing in a channel. Marketing
Science, 25, 97–105.
Meijer, E., & Rouwendal, J. (2006). Measuring welfare effects in models with random
coefficients. Journal of Applied Econometrics, 21, 227–244.
Michalek, J., Ebbes, P., Adigüzel, F., Feinberg, F., & Papalambros, P. (2011). Enhancing
marketing with engineering: Optimal product line decisions for heterogeneous
markets. International Journal of Research in Marketing, 28, 1–12.
Montgomery, A., & Smith, M. (2009). Prospects for personalization on the internet. Journal
of Interactive Marketing, 23, 130–137.
Ofek, E., & Srinivasan, V. (2002). How much does the market value an improvement in a
product attribute. Marketing Science, 21, 398–411.
Rossi, P., McCulloch, R., & Allenby, G. (1996). On the value of household purchase history
information in target marketing. Marketing Science, 2, 321–340.
Shaffer, G., & Zhang, Z. J. (2002). Competitive one-to-one promotions. Management
Science, 48, 1143–1160.
Silva-Risso, J., & Ionova, I. (2008). A nested logit model of product and transaction-type
choice for planning automakers' pricing and promotions. Marketing Science, 27,
545–566.
Sonnier, G., Ainslie, A., & Otter, T. (2007). Heterogeneity distributions of willingness to pay
in choice models. Quantitative Marketing and Economics, 5, 313–331.
Spiegelhalter, D., Best, N., Carlin, B., & van der Linde, A. (2004). Bayesian
measures of model complexity and fit. Journal of the Royal Statistical Society B, 64,
583–639.
Swait, J., & Louviere, J. (1993). The role of the scale parameter in the estimation
and comparison of multinomial logit models. Journal of Marketing Research, 30,
305–314.
Train, K. (2003). Discrete choice methods with simulation. Cambridge: Cambridge
University Press.
Wertenbroch, K., & Skiera, B. (2002). Measuring consumers' willingness to pay at the
point of purchase. Journal of Marketing Research, 39, 228–241.
Wolk, A., & Ebbling, C. (2010). Multi-channel price differentiation: An empirical investigation
of existence and causes. International Journal of Research in Marketing, 27, 142–150.
Zhang, J., & Krishnamurthi, L. (2004). Customizing promotions in online stores. Marketing
Science, 23, 561–578.
Zhang, J., & Wedel, M. (2009). The effectiveness of customized promotions in online and
offline stores. Journal of Marketing Research, 46, 190–206.
Intern. J. of Research in Marketing 31 (2014) 178–191
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
Full Length Article
How much to give? — The effect of donation size on tactical and strategic
success in cause-related marketing
Sarah S. Müller ⁎, Anne J. Fries, Karen Gedenk 1
University of Hamburg, Max-Brauer-Allee 60, 22765 Hamburg, Germany
a r t i c l e
i n f o
Article history:
First received in 7 February 2011 and was
under review for 6½ months
Available online 16 October 2013
Area Editor: Zeynep Gurhan-Canli
Keywords:
Cause-related marketing
Donation size
Donation framing
Promotion
Choice experiment
a b s t r a c t
In cause-related marketing (CM), companies promise a donation to a cause every time a consumer makes a
purchase. We analyze the impact of the size of this donation on brand choice (tactical success) and brand
image (strategic success). Our results reveal different effects of donation size on these success measures. For
brand choice, the effect of donation size is moderated by a financial trade-off for consumers, whereas the effect
on brand image is moderated by donation framing. Specifically, we show that donation size has a positive effect
on brand choice if consumers face no financial trade-off; i.e., if they do not have to choose between triggering a
donation or saving money. The effect is negative if a trade-off exists such that higher donations come at higher
costs. Brand image is enhanced by larger donations if the framing is nonmonetary (e.g., the campaign promises
the provision of vaccinations), whereas donation size has a negative effect if donation framing is monetary
(e.g., the campaign states the Euro amount). If campaigns use a combination of both frames, the effect of donation
size on brand image has an inverted U shape. Our results suggest that CM enhances tactical and strategic success
only if firms select the right donation size, taking into account donation framing and financial trade-offs.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
In a cause-related marketing (CM) campaign, Tommy Hilfiger
featured a promotion in which 50% of the price of a specific bag would
be donated to Breast Health International. In another CM promotion,
Starbucks donated $1 to the Global Fund to support people living with
AIDS in Africa for every pound of East Africa Blend coffee sold. Volvic
promoted its “Drink 1, Give 10” campaign in cooperation with UNICEF,
stating that for every liter of water sold, the company would provide
10 l of drinking water in Africa. Procter & Gamble (P&G) promised “1
pack = 1 vaccine” in its CM promotion, in which for every promotional
package sold, the company would donate .054€ to UNICEF, equal to the
cost of one vaccination against tetanus.
In CM campaigns such as these, the firm contributes a specific
amount to a cause if a customer buys the firm's product (Varadarajan
& Menon, 1988). This transactional element is the main characteristic
of CM: The customer must make a purchase to trigger the donation.
Corporate sponsorship of social causes has become very frequent, with
spendings in North America reaching $1.86 billion in 2011 (IEG, 2011).
CM is both a tactical tool that firms employ to increase their sales
and a strategic activity aimed at improving brand image (Ross, Stutts,
& Patterson, 1991). However, whether the investment in CM always
⁎ Corresponding author. Tel.: +49 40 42838 7132.
E-mail addresses: sarah.mueller@wiso.uni-hamburg.de (S.S. Müller),
anne.fries@wiso.uni-hamburg.de (A.J. Fries), karen.gedenk@wiso.uni-hamburg.de
(K. Gedenk).
1
Tel.: +49 40 42838 3748.
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.09.005
pays off is unclear. On the one hand, by triggering a donation through
their purchases, consumers might derive utility from giving, which is
known as “warm glow” (Andreoni, 1989), and thus exhibit favorable
purchase behaviors. On the other hand, CM might raise consumer
skepticism about the company's motivation because the donation is
conditional on sales and ensures the company's own benefit (Barone,
Miyazaki, & Taylor, 2000). These consumer considerations can negatively
impact brand image. Whether positive or negative effects prevail depends
on several success factors (Fries, 2010).
We study one key success factor, donation size, which is particularly
interesting because it is a design element that is directly controlled by
managers; i.e., they can decide how much to give when implementing
CM. Campaigns vary in their donation sizes as indicated by the
introductory examples, in which donations range from 1% of the
product's price in the P&G example to 50% of the price in the
Tommy Hilfiger campaign. The effect of investing in a larger donation
is unclear. On the one hand, consumers may derive more warm glow
when donation size increases, which should make them more likely to
make a purchase. A larger donation could also produce more favorable
evaluations of the brand. On the other hand, consumers who face a CM
offer with a substantial donation may prefer to receive this money for
themselves or may not believe that the company will really donate as
much as promised. Thus, donation size could also have a negative effect
on sales and brand image.
Previous research has studied the influence of donation size on
CM success, but the results are equivocal. Some studies find a positive
effect (e.g., Olsen, Pracejus, & Brown, 2003), others a negative one
(e.g., Strahilevitz, 1999), and others no effect at all (e.g., Human &
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Terblanche, 2012). We therefore analyze the effect of donation size
on CM success in more depth and extend previous research by
focusing on the following three aspects:
First, we acknowledge that firms use CM for both tactical and strategic
purposes and therefore study two success measures: brand choice and
brand image. Previous research has rarely compared these success
measures. We expect the effects of donation size on brand choice and
brand image to differ because of different underlying drivers.
Second, we study two potential moderators of the effect of donation
size on CM success that have not been analyzed before: the presence of
a financial trade-off and donation framing. A financial trade-off occurs
when consumers choose between one brand with a CM campaign
and another brand with a price promotion. We expect that such a
trade-off moderates the effect of donation size on brand choice.
The framing of a CM campaign can be monetary (e.g., 5 cents),
nonmonetary (e.g., one vaccination), or a combination of both (e.g.,
one vaccination, worth 5 cents). We expect framing to moderate the
impact of donation size on brand image.
Third, we vary our independent variable – donation size – over a wide
range and in small intervals, which allows us to test for nonlinear effects.
In a large-scale experimental survey, we systematically vary donation
size and the potential moderators, and ask respondents to make a brand
choice decision and evaluate the image of the focal brand. In an additional
exploratory study, we also measure prospective drivers underlying CM
success to shed light on the differences between tactical and strategic
success.
We find that the effect of donation size is different for brand choice
(tactical success) versus brand image (strategic success). The effect
on brand choice is moderated by the presence of a financial trade-off,
and the effect on brand image is moderated by donation framing.
Furthermore, we find a nonlinear effect of donation size on brand
image for a combined monetary and nonmonetary framing. Finally,
our exploratory analysis suggests that brand choice is driven by warm
glow, whereas brand image mostly depends on what consumers infer
about the company's altruism and about the effectiveness of the
campaign.
Our results have important implications for managers. We show that
spending more money on a larger donation does not always produce
more favorable effects, but rather donation size has to be chosen
carefully, taking into account financial trade-offs and donation
framing.
Our research contributes to the CM literature by clarifying the effects
of donation size: We explain why the effect can be positive, negative, or
null. In particular, we detect differences in tactical versus strategic
success. Furthermore, we investigate the moderating effects of financial
trade-offs and donation framing for the first time and provide new
insights into nonlinear effects.
We proceed as follows. In Section 2, we review existing research on
donation size before presenting our conceptual framework and deriving
hypotheses about the effects of donation size on CM success in
Section 3. We present the research design of our experimental survey
investigating the different effects of donation size in Section 4, and its
results in Section 5. To gain insights into the drivers underlying tactical
and strategic CM success, we report the data and results of an additional
study in Section 6. We conclude by summarizing our work and discussing
its implications for both managers and researchers in Section 7.
2. Literature review
Much previous research has studied the characteristics of successful
CM campaigns (for an overview, see Fries, 2010) and has identified a
broad range of success factors, including the characteristics of the
cause (e.g., Ross et al., 1991), the company (e.g., Strahilevitz, 2003),
the consumer (e.g., Wymer & Samu, 2009), the non-profit organization
(NPO) (e.g., Barnes, 1992), the product (e.g., Strahilevitz & Myers,
179
1998), and the fit among these factors (e.g., Zdravkovic, Magnusson, &
Stanley, 2010).
A success factor that has been analyzed in several past studies is
donation size. As indicated in Table 1, the results of these studies
are equivocal, spanning positive (e.g., Dahl & Lavack, 1995; Pracejus
et al., 2003/04), negative (e.g., Arora & Henderson, 2007, study 3;
Strahilevitz, 1999), and insignificant effects of donation size (e.g., Arora
& Henderson, 2007, study 1; Vaidyanathan & Aggarwal, 2005). Although
several studies incorporate moderating effects (Table 1), these cannot
fully explain the conflicting findings. The (potential) moderators
either do not influence the effect of donation size (e.g., promotion size,
donation recipient), or they merely affect the strength of a positive or
negative effect (e.g., cause involvement, price, product type), but do not
change its direction.
We suggest three possible reasons why the effect of donation size on
CM success can be positive, negative, or null, which have not been
studied systematically thus far: differences in tactical versus strategic
success, moderating and nonlinear effects. First, companies pursue
two main goals with CM: the tactical goal of increasing sales and the
strategic goal of improving brand image (Polonsky & Wood, 2001).
Previous research on donation size mainly uses sales-related dependent
variables such as purchase intention and brand choice to measure
tactical success. Alternatively, a few studies analyze the effects on
attitudes towards the brand to capture strategic success. Only Arora and
Henderson (2007), Holmes and Kilbane (1993), and Olsen et al. (2003)
investigate both types of success measures and find no differences
between them. Yet, a more in-depth analysis of the effect of donation
size might reveal differences regarding its impact on tactical and strategic
success because the underlying drivers of the success measures should
be different. Specifically, purchase decisions should be driven mainly by
the utility that consumers derive from the campaign, whereas changes
in brand image should result mostly from the inferences consumers
make about the brand offering the campaign. These distinct underlying
drivers should also cause the effect of donation size on brand choice
versus brand image to be moderated by different variables, as explained
next.
Second, two potentially relevant moderators have not been examined
so far: the presence of a financial trade-off and donation framing. Some
studies on donation size have provided respondents with decision tasks
that involve choosing between a CM option, in which the money is
donated, and a non-CM option, which offers a price reduction of equal
size (e.g., Arora & Henderson, 2007, study 3; Strahilevitz, 1999). In this
case, respondents face a financial trade-off: they can either do something
good by choosing the CM option or they can gain a financial advantage for
themselves by selecting the competitive offer. Other studies have not
included such a trade-off. Table 1 reveals that studies with a financial
trade-off tend to find that larger donations hurt sales, whereas most
studies without a trade-off report that donation size has a positive or
no significant effect. These findings suggest that larger donations help
only when they come at no increased costs to the consumer. This is in
line with the assertion of Burnett and Wood (1988) that prosocial
behavior depends on the cost of helping; forgoing a price discount
could be an important cost. Thus, differences in utility caused by financial
trade-offs could explain the equivocal effects of donation size on tactical
CM success. To date, the moderating effect of a financial trade-off has
not been studied.
Another new potential moderator is the framing of the donation
in monetary versus nonmonetary terms. So far, two studies have
examined donation size and framing. Olsen et al. (2003) compare CM
campaigns that present the donation as a percentage of the price versus
a percentage of the profit and find no differences in the effect of
donation size between the two frames. Chang (2008) shows that
expressing a donation in absolute monetary value is more favorable
for small donations than a percent of price framing, whereas no
difference exists for large donations. Both of these studies compare
different monetary frames. However, the examples in our introduction
180
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Table 1
Research on donation size.
Study
Donation size†
Financial trade-off
Dependent variable
Result
Moderating effects
Dahl and Lavack (1995)
$0.0025 vs. $0.1
No
+
Promotion size: n.s.
Garretson Folse, Niedrich, and
Landreth Grau (2010)
0.13–32% of price
1.88–67.5% of price
2.5–40% of price
0–6.8% of price
0–40 cents
No
No
No
No
No
Perceived exploitation of NPO
Product appeal
CM participation intentions
CM participation intentions
CM participation intentions
Attitude toward ad
Willingness to pay
Olsen et al. (2003)
5 vs. 40 cents
5 vs. 40 cents
5 vs. 40 cents
1 vs. 10% of price
No
No
No
No
+/n.s.
+/n.s.
+/n.s.
+
Pracejus, Olsen, and Brown (2003/04)
Smith and Alcorn (1991)
0–10% of price
$0.1; $0.25; $0.4
No
No
Willingness to pay
Willingness to pay
Willingness to pay
Attitude toward ad
Attitude toward brand
Purchase intention
Brand choice
Intention to use coupon
Arora and Henderson (2007), study 3
Yes
Brand choice
−
Subrahmanyan (2004)
1 vs. 5% of monthly
credit card charge
5 vs. 25% of price
5 vs. 50% of price
1 vs. 25% of price
1 vs. 25% of price
1–20% of price
No
Yes
Yes
Yes
Yes
Behavioral intention
Brand choice
Brand choice
Brand choice
Purchase likelihood
−
−
−
−
−
Arora and Henderson (2007), study 1
0–45% of price
No
n.s.
Fries, Gedenk, and Völckner (2010)
Holmes and Kilbane (1993)
5 vs. 15% of price
0–6.8% of price
No
No
Human and Terblanche (2012)
$0.18 vs. $1.14
No
Vaidyanathan and Aggarwal (2005)
van den Brink, Odekerken-Schröder,
and Pauwels (2006)
6.3 vs. 12.5% of price
0.1 vs. 25% of price
Yes/no
No
Brand choice
Purchase likelihood
Attitude toward brand
Brand choice
Attitude toward store
Intention to respond
Attitude toward cause alliance
Attitude toward campaign
CM participation intentions
Willingness to buy
Brand loyalty
Holmes and Kilbane (1993)
Koschate-Fischer, Stefan, and Hoyer
(2012)
Chang (2008)
Strahilevitz (1999)
+
+
+
+
+
Price: n.s.
Attitude toward helping: +
Warm glow: +
Cause involvement: +
Cause organization affinity: +
Fit: +/n.s.
Fit: +/n.s.
Fit: +/n.s.
Framing (% of price vs. % of profit): n.s.
+
+
Price: −
Product type: + (hedonic)
n.s.
n.s.
Price: n.s.
n.s.
Donation recipient: n.s.
n.s.
n.s.
Notes: + = positive effect; - = negative effect; n.s. = no significant effect; † = donation sizes were transformed into % of price when possible.
illustrate that companies use not only such monetary frames (e.g., $1 in
the Starbucks example) but also nonmonetary frames in which
the donation is presented as a charitable object or service (e.g., 10 l of
drinking water in the Volvic example), as well as a combined framing
that provides both types of information (e.g., 1 vaccine, worth .054€,
in the P&G example). Research on promotions has shown that a
promotion's value in relation to the product's price is assessed differently
when it is framed in monetary versus nonmonetary terms (e.g., Nunes &
Park, 2003; Palazon & Delgado-Ballester, 2009). For CM campaigns,
the effect of donation size on brand image could also be affected by
monetary versus nonmonetary framing because these frames provide
different information that might influence the inferences consumers
make about the company. The moderating effect of donation framing
in monetary, nonmonetary, or combined terms has not previously been
examined.
Third, donation size could exert nonlinear effects on CM success.
Most previous studies investigate only two different levels of donation
size and the range of donation sizes varies across studies. So far, few
studies have tested for nonlinear effects. Pracejus et al. (2003/04) find
an insignificant quadratic term. However, they only study a range of
donation sizes from 0 to 10% of the price, whereas in actual CM
campaigns firms donate up to 50% of the price (e.g., Tommy Hilfiger).
Koschate-Fischer et al. (2012) also use a quadratic term and report a
positive effect of donation size that is concave (i.e., weaker for larger
donations). However, their dependent variable is willingness to pay,
i.e., they do not vary donation size in relation to the product's price.
Finally, evidence for the nonlinear effects of donation size appears in
the context of charity auctions (Haruvy & Popkowski Leszczyc, 2009),
which reveal a negative effect for very large donations but a positive
effect when a smaller fraction of the auction's final price is donated.
However, whether the same mechanisms apply to both CM campaigns
and charity auctions is unclear. More importantly, all three studies
analyze the nonlinear effects of donation size for a monetary framing.
As we explain in the next section, we expect a nonlinear effect for a
combined framing (monetary and nonmonetary), which has not been
studied, yet.
3. Conceptual framework
3.1. Overview
Fig. 1 depicts our conceptual framework. We analyze the impact of
donation size on both brand choice (tactical success) and brand image
(strategic success). Furthermore, we consider the presence of a financial
trade-off (i.e., non-focal brand on price promotion) as a moderator
of the effect on brand choice, and donation framing (i.e., monetary,
nonmonetary, combination) as a moderator of the effect on brand
image.
We expect that the effects of donation size on brand choice versus
brand image are different and that different moderators are relevant
because we propose that these success measures are affected by
different underlying drivers. More specifically, we assume the effect of
donation size on brand choice to be driven mostly by the utility that
consumers derive for themselves from the CM campaign, whereas the
effect on brand image should be driven primarily by what consumers
infer about the company.
181
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Financial trade-off
Tactical CM success:
Brand choice
Donation size
Strategic CM success:
Brand image
Donation framing
Fig. 1. Conceptual framework.
When consumers make brand choice decisions, they focus on
themselves such that the utility that they derive from the campaign is
crucial. Consumer's utility is determined by the benefits and costs of
the CM campaign and a key benefit of a CM campaign is warm glow.
Warm glow theory postulates that subjects derive utility from the mere
act of giving, which is known as “warm glow” or “moral satisfaction”
(Kahneman & Knetsch, 1992). Consumers are thus more likely to choose
an option if it provides more warm glow (Andreoni, 1990). Triggering a
donation through a product purchase can offer warm glow to consumers
(Strahilevitz & Myers, 1998), and the findings of Fries et al. (2010)
support that warm glow is the main underlying driver of the positive
effect of CM on brand choice. Hence, we assume that the choice of
the CM product depends primarily on the campaign's utility, which is
provided by warm glow, in relation to the costs of engaging in the
campaign. The latter should be affected when there is a financial tradeoff for consumers, i.e., when selecting the CM option comes at the cost
of foregoing savings for oneself.
When consumers assess brand image, they focus on the company, so
the inferences that they derive about the brand from the CM campaign
are crucial. Information integration theory (Anderson, 1981) suggests
that new information is incorporated into prior attitudes, resulting in
updated attitudes that reflect how the stimulus is evaluated. Thus,
how consumers evaluate the company's engagement should affect its
impact on brand image. We consider two aspects of this evaluation as
the main drivers of brand image: perceived altruism and perceived
effectiveness. Perceived altruism captures the degree to which consumers
perceive the company to be motivated by a genuine interest in supporting
the charitable cause. Perceived effectiveness is the degree to which
consumers believe that the company will really donate as much as
promised and that this donation will actually reach the needy recipients.
Perceived altruism and perceived effectiveness have been studied as
drivers of the effect of CM on brand choice and have been found to be
less influential than warm glow (Fries et al., 2010). We assume that
they are more important as drivers of the effect of CM on brand image
because CM should positively affect brand image only if consumers
attribute altruistic motives to the company's efforts and believe the
promises stated in the campaign.
In the remainder of this section, we build on these underlying
drivers when we derive our hypotheses about how donation size affects
CM success.
3.2. Brand choice hypotheses
We expect the effect of donation size on brand choice to be
moderated by the presence of a financial trade-off because brand choice
is driven by warm glow and the cost of giving. We do not predict a
moderating effect of donation framing for brand choice. Instead, the
mere act of triggering a donation through one's purchase should induce
warm glow regardless of the framing of the donation. This is in line with
the notion that when consumers contribute to a cause, they are satisfied
by the fact that something will be done without requiring detailed
information (Kahneman, Ritov, Jacowitz, & Grant, 1993).
Warm glow is an increasing function of what is given (Andreoni,
1989). Accordingly, without a financial trade-off, i.e., when consumers
do not have to choose between doing good and saving money, a
higher donation induces no increased costs to consumers and warm
glow and utility should thus increase if the donation rises. We therefore
propose:
H1a. The effect of donation size on brand choice will be positive, if the
consumer faces no financial trade-off.
In contrast, when consumers choose between a brand with a CM
campaign and another brand with a lower price, they face a trade-off
and buying the CM brand comes at a cost. Donors are price-sensitive,
and their likelihood of helping decreases as the cost of helping increases
(Burnett & Wood, 1988; Eckel & Grossman, 2003). We expect that this
increase in costs outweighs the increase in warm glow for larger
donations. This is supported by the previous research summarized in
Table 1: almost all studies in which consumers face a financial tradeoff find a negative effect of donation size on purchase behavior. Hence,
we hypothesize:
H1b. The effect of donation size on brand choice will be negative, if the
consumer faces a financial trade-off.
3.3. Brand image hypotheses
We expect that the effect of donation size on brand image is
moderated by donation framing because brand image is driven
primarily by what consumers infer about the brand, i.e., by perceived
altruism and perceived effectiveness. We do not predict a financial
trade-off to be a moderator in this context because a financial
trade-off should affect consumers' utility, but not inferences about
the brand.
When donations are framed in monetary terms, larger donations
are likely to decrease perceived effectiveness. With larger monetary
donations, consumers may become skeptical that the company will
really donate this much money, and the complete amount will reach
the needy recipients. Similar effects have been shown for price
promotions, where consumers do not believe that discounts are
really as large as advertised. More specifically, consumers discount
price discounts and do so increasingly as promised savings rise (Gupta
& Cooper, 1992). A similar effect is likely to occur for CM campaigns
with monetary donations such that consumers may assume that
the actual donation will be lower than advertised and will therefore
increasingly discount the advertised donation which lowers perceived
effectiveness as donations become larger. Hence, we posit:
H2a. The effect of donation size on brand image will be negative, if
donation framing is monetary.
182
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
A nonmonetary frame emphasizes the output achieved with the
donation and the good the company does. Because consumers typically
cannot assess the monetary value of public goods (Green, Kahneman,
& Kunreuther, 1994), they most likely cannot judge the value of
donations that are expressed in nonmonetary terms, such that perceived
effectiveness is not affected. However, nonmonetary donations are
perceived to require more effort from companies than monetary
donations (Ellen, Mohr, & Webb, 2000), such that this frame should
indicate more sincere company motives, which is reflected in perceived
altruism. Hence, when companies make larger donations, this should be
perceived as more effort and result in stronger perceived altruism and
thus better brand image. We therefore expect the following:
H2b. The effect of donation size on brand image will be positive, if
donation framing is nonmonetary.
Finally, combined frames include both monetary and nonmonetary
information and emphasize not only how much money the company
donates but also the achieved output. In this case, opposing forces
are at work: On the one hand, larger monetary donations decrease
perceived effectiveness; on the other hand, larger nonmonetary
donations increase perceived altruism. To derive an effect of donation
size on brand image for a combined frame, we consider that this
frame provides maximum transparency, which should affect perceived
effectiveness. Consumers prefer transparency in donation framing
(Landreth Grau, Garretson Folse, & Pirsch, 2007). Hence, in the case of
small to medium donations, larger donations should not negatively
impact perceived effectiveness because the combined frame reveals
not only the monetary amount but also the output. Here, the dominant
effect should be that consumers perceive a larger donation as more
charitable effort by the company, thereby enhancing perceived altruism
such that brand image becomes more favorable with rising donations.
However, this only works up to a certain donation level because in the
case of a very high monetary amount, consumers might again become
skeptical about perceived effectiveness. After this point, we suppose
that the effect of donation size on brand image is dominated by these
negative inferences and becomes negative. In summary, we expect an
inverted U-shaped effect and predict:
H2c. The effect of donation size on brand image will follow an inverted
U shape, if donation framing combines monetary and nonmonetary
information.
Table 2 summarizes our hypotheses.
4. Research design
To test our hypotheses, we conducted a between-subjects experiment
based on a large-scale survey. Different groups of respondents considered
different CM campaigns, made brand choice decisions, and assessed the
image of the CM brand.
4.1. Stimuli
We constructed choice sets with two brands per product category.
Participants read the following scenario: “You would like to buy product
Table 2
Hypotheses.
Brand choice
Effects of donation size
H1a
+, if no financial trade-off
H1b
−, if financial trade-off
Brand image
H2a
H2b
H2c
−, if donation framing monetary
+, if donation framing nonmonetary
∩, if donation framing combined
Notes: + = positive effect; − = negative effect; ∩ = inverted U-shaped effect.
category X. You can choose between Brand A and Brand B. With regard
to your choice, only the brand (Brand A/Brand B) is relevant, not
the depicted flavor.” We presented photos of the two brands and
information about their prices and product sizes. In the control
condition, both brands appeared without a promotion. In the treatment
conditions, one brand offered a CM campaign, and the other did not. In
each product category, the CM campaign was always tied to the same
focal brand.2 The design of the CM campaign varied across treatment
groups. For the treatments with a CM offer and a financial trade-off,
the competitive brand offered a price discount of the same size as the
donation.
We investigated four product categories: chocolate bars, toothpaste,
beer, and detergent. All products were fast moving consumer goods
(FMCG) that varied in their price levels and degrees of utilitarianism
and hedonism to ensure robust results. For each product category,
we chose two well-known national brands and presented products
that were identical in price, size, and flavor. The prices reflected
average prices found in major German supermarkets at the time of
our study. We list the brands, sizes, and prices used in the study in
Appendix A.
We employed the same NPO and cause for all product categories:
all CM campaigns promised a donation to SOS Children's Villages to
support immunization against tetanus. This well-known charity
enjoys a very good reputation (German Fundraising Association,
2009), and immunization against tetanus represents an important and
uncontroversial cause. Tetanus remains a major risk in countries with
low immunization rates (WHO & UNICEF, 2010).
We systematically varied three experimental factors, as specified in
Table 3. The donation size manipulation included eight levels: 1, 2.5, 5,
10, 20, 30, 40, and 50% of product price.3 We used these percentages
to calculate the respective donation amount in Euros and the number
of vaccinations.
The factor donation framing had three levels: In the monetary frame,
the donation amount was presented in Euros, such as .20€. In the
nonmonetary frame, the donation was stated as the number of
vaccinations, such as four vaccinations. We used a price of 5 cents per
vaccination to translate monetary into nonmonetary donations (WHO
& UNICEF, 2010). The combined frame included both the monetary
amount in Euros and the equivalent number of vaccinations.
All three frames were presented without a financial trade-off. In
these treatments, the CM and the competitive brand were priced
equally. We also combined the monetary frame with a financial
trade-off because this frame supports an equal framing of donation
size and savings. In these treatments, the competitive brand offered
a price discount equal in size to the donation promised by the CM
brand (Arora & Henderson, 2007; Strahilevitz, 1999; Vaidyanathan
& Aggarwal, 2005). We provide an example stimulus for a monetary
CM campaign and equivalent competitive price promotion in
Appendix B.
In our experimental set-up, the three frames without financial
trade-offs and the monetary frame with the competitive price
promotion were combined with the eight levels of donation size.
We also included a control group, such that we tested 33 conditions
between-subjects.
4.2. Procedure
Subjects were randomly assigned to one of the 33 conditions.
They assessed up to four product categories, although they answered
questions for a category only if they had made a purchase in that
2
Across categories, there is variance in whether the CM brand has a larger purchase
frequency than the non-focal brand (measured as number among the last three category
purchases before the survey), a smaller or a similar purchase frequency.
3
A systematic research of CM campaigns in Germany revealed 50% as the maximum
donation size.
183
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Table 3
Experimental factors.
Factor
Level
Realization
Donation size
1/2.5/5/10/20/30/40/
50% of product price
Monetary
Nonmonetary
Combined
Converted into Euro amount
and/or number of vaccinations
Amount in Euros
Number of vaccinations
Amount in Euros and equivalent
number of vaccinations
Competitive brand offers price
discount of equal size as donation
by CM brand
No competitive promotion
Donation framing
Financial trade-off
Present
(only in combination
with monetary frame)
Not present
category at least once during the previous year. This filter increases the
response quality because subjects who are familiar with a category
give more valid and reliable answers (Alba & Hutchinson, 1987). The
categories appeared in the same order for all respondents. For each
participant, the experimental treatment was kept constant across all
categories.
To measure brand awareness, the respondents first indicated
whether they were familiar with the two brands in each product
category. They then stated their brand preferences by specifying
the number of times they had bought the two brands in their
last three purchases in the category. In line with previous research
(e.g., Bouten, Snelders, & Hultink, 2011; Simonin & Ruth, 1998), we
assessed brand awareness and preferences before the experimental
manipulation to prevent any influence that the stimulus might
have on these measures.4 After the presentation of the stimulus,
the respondents made their brand choice decision and then revealed
their brand image assessments for the CM brand.5 Finally, we asked
about demographic characteristics.
(p b .05). Therefore, we include these demographics in our models as
control variables. No significant differences emerge for other consumer
demographics (i.e., age, gender, household size, income), previous
donation behavior, or brand-related variables such as brand awareness
and brand preference (p N .10). Brand awareness rates are greater than
92% for all brands used in the study, confirming that we selected wellknown brands.
For the analyses, we pool the data across the four product categories,
resulting in a total of 4686 observations, with a minimum of 93
observations per experimental group.
4.5. Models
For brand choice, we estimate the following binary logit model, which
includes donation size and the moderators as concomitant variables:
Phc ¼
1
;
1 þ eð−Vhc Þ
Vhc ¼
C
X
c¼1
δh ¼ γ h þ
4.4. Sample
We sent the questionnaire to participants in an online access panel
in Germany. As an incentive, respondents could participate in a drawing
to win Amazon gift cards. Of the 1446 respondents who answered
the questionnaire between December 2008 and January 2009, 85 were
excluded from the analysis due to missing values. The final data set
contains 1361 complete observations. Of the respondents, 47% are
women. On average, participants are 34 years of age and live in
households with 2.3 people. The majority has a monthly net household
income ranging from 1000€ to 2000€. More than half (61%) are members
of a church, and 25% have children. Finally, 46% of the respondents are
employed, and 35% are students.
We find significant differences across the 33 experimental groups
with respect to occupation, church membership, and parenthood
Phc
Vhc
CATc
PREPREFhc
CMh
DEMOkh
Xsh
k¼1
c¼1
βc % PREPREFhc % CATc þ δh % CMh ; and
ηk % DEMOkh þ
S
X
s¼1
λsh % Xsh ;
ð2Þ
ð3Þ
Probability that subject h chooses the CM brand in category c,
Systematic utility of the CM brand for subject h in category c,
Category indicator (1 if product category c, and 0 otherwise),
Stated brand preference of subject h in category c
(=number of times the CM brand was bought during the
last three purchases before the survey, minus the number
of times the other brand was bought on these purchases),
CM indicator (1 if subject h sees a CM campaign, and 0
otherwise),
Demographic variable k for subject h, and
CM-related concomitant variable s for subject h.
In the utility function (Eq. (2)), we include category-specific intercepts
and control for stated preference heterogeneity with a PREPREF variable
for each category (Ailawadi, Gedenk, & Neslin, 1999; Horsky, Misra, &
Nelson, 2006). The parameter δh captures the effect of a CM campaign
on utility. It differs across subjects because different respondents receive
different experimental treatments, as described by the concomitant
variables Xsh, and because response is heterogeneous. We control for
demographic variables that vary between the experimental groups
(Section 4.4.). Finally, we use a continuous mixture model to capture
unobserved heterogeneity in all parameters except those for PREPREF
and the demographics, which are household-specific. We assume that
all heterogeneous parameters follow normal distributions and estimate
their means and standard deviations.
For brand image, we estimate the following linear regression model
with the same independent and concomitant variables:
BIMAGEhc ¼
4
These measurements might make initial preferences more salient and thus affect the
measures of our dependent variables. If this were indeed the case, it would make our
hypothesis tests conservative because the experimental treatments would have less of
an effect.
5
We measured brand choice before brand image because we did not want the choice
decision to be biased by consumers' elaboration on the focal CM brand. To test if this order
causes a bias in the measurement of brand image, we counterbalanced the order of the
dependent measures for one experimental treatment in our second study. We found no
evidence that our results for brand choice and brand image were affected by the order
of these two measures.
K
X
C
X
where
4.3. Measures of CM success
We used brand choice to measure tactical success and brand image
to capture strategic success. For each choice set, respondents indicated
which of the two brands they would rather buy. Next, respondents
evaluated the image of the CM brand on six seven-point semantic scales
(Völckner, Sattler, & Kaufmann, 2008; see Appendix C).
α hc % CATc þ
ð1Þ
C
X
c¼1
α hc % CATc þ
C
X
c¼1
βc % PREPREFhc % CATc þ δh % CMh ;
ð4Þ
where
BIMAGEhc Image of subject h of the CM brand in category c.
The models for both brand choice and brand image are developed in
consecutive steps. Starting with demographics, we add the concomitant
variables Xsh stepwise to check for model improvements from adding
184
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
moderating and nonlinear effects. The models are nested, as outlined in
Table 4. Appendix D summarizes the operationalization of the
independent variables.
Model 1, with the main effects of donation size, donation framing,
and a financial trade-off, is our base model (Jaccard, 2001; Jaccard,
Turrisi, & Wan, 1990). In Model 2, we add interactions of donation size
with donation framing and a financial trade-off. With Model 2 we can
test our hypotheses about the moderating effect of a financial tradeoff (H1a and H1b) and donation framing (H2a and H2b). To examine
the nonlinear effects of donation size (H2c), we incorporate quadratic
terms in Models 3 and 4.6 To provide evidence for the significance of
the interaction effects (Jaccard, 2001; Jaccard et al., 1990), we first
include a quadratic term for donation size (Model 3), and then add
quadratic terms for the interactions with donation framing and a
financial trade-off (Model 4). Although our hypotheses do not feature
all possible moderating and nonlinear effects for both brand choice
and brand image, we estimate all four models with both dependent
variables to ensure that we do not miss any effects.
We estimate our models with simulated maximum likelihood (Train,
2009) using the MAXLIK module in GAUSS.7 We test whether pooling
across the product categories is appropriate using likelihood ratio tests
for the logit models of brand choice and Chow tests for the regression
models of brand image. In the Chow tests, the improvement in model
fit when we move from a pooled model to four separate models is not
significant for any of the models (p N .05). In the likelihood ratio tests,
no fit improvement is significant at the 1% level; for Models 2 and 3,
the improvement is significant at the 5% level. However, with more
than 4000 observations, even small differences tend to be significant,
and we find no substantive differences for the effects of our experimental
variables across the four categories. Thus, we consider pooling to be
appropriate.
5. Results
5.1. Brand choice
Table 5 contains the fit measures for our four brand choice models.
Because the models are nested, we use likelihood ratio tests to determine
whether more comprehensive models offer a significant improvement
over simpler ones.
Model 1 includes the main effects of our experimental manipulations
of donation size, donation framing and a financial trade-off. In Model 2,
we add the moderating effects of donation framing and a financial
trade-off to the effect of donation size. The likelihood ratio test shows
that model fit improves significantly, indicating that the effect of
donation size on brand choice is moderated. In Models 3 and 4, we add
quadratic terms to capture the nonlinear effects of donation size but
find no significant improvements. This is in line with our predictions:
we expected nonlinear effects for brand image but not for brand choice.
We rely on Model 2 to test our brand choice hypotheses and present its
parameter estimates in Table 6.
All coefficients for the control variables exhibit plausible signs. The
positive PREPREF coefficients (p b .01) indicate that consumers are
more likely to choose the CM brand when they preferred it over the
competitive brand in their recent purchases. Respondents with children
react more favorably to CM campaigns (p b .01), in line with previous
research (Ross, Patterson, & Stutts, 1992). Respondents who did not
indicate whether they were church members also react more favorably
to CM, but this effect is only weakly significant (p b .10).
6
We also tested for thresholds by allowing the coefficient of donation size to be
different below and above a threshold in our Model 2. We inserted thresholds at donation
sizes of 10 and 30%, which are common for price promotions (e.g., van Heerde, Leeflang, &
Wittink, 2001). However, none of these thresholds improved model fit, neither for brand
choice nor for brand image (p N .10). Details are available from the authors upon request.
7
We rescaled donation size (by dividing it by 100) in the brand image models to
facilitate the estimation (Ailawadi, Gedenk, Lutzky, & Neslin, 2007).
Table 4
Model specification.
Variables
Category variables
CAT_CHOCOLATE
CAT_TOOTHPASTE
CAT_BEER
CAT_DETERGENT
PREPREF_CHOCOLATE
PREPREF_TOOTHPASTE
PREPREF_BEER
PREPREF_DETERGENT
Demographic variables
Child_YES × CM
Church membership_YES × CM
Church membership_ NO_RES
× CM
Occupation_FULL × CM
Occupation_PART × CM
CM variables
CM
DONSIZ
NONMON
COMBI
TRADE-OFF
DONSIZ × NONMON
DONSIZ × COMBI
DONSIZ × TRADE-OFF
Nonlinear effects
DONSIZ2
DONSIZ2 × COMBI
DONSIZ2 × NONMON
DONSIZ2 × TRADE-OFF
Model 1
Model 2
Model 3
Model 4
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Notes: × = interaction effect; ✓ = included in the model.
Regarding our first hypothesis, we find that the presence of a financial
trade-off moderates the effect of donation size, as indicated by λ7. In
support of H1a, donation size has a positive effect on brand choice
when there is no financial trade-off, according to the significant and
positive λ1, which captures the effect of donation size for a monetary
frame without a financial trade-off. As expected, we find no differences
in the effect of donation size across the three frames without financial
trade-offs; neither λ5 nor λ6 is significantly different from zero. To
formally test H1a for the nonmonetary and combined frames, we test
whether the sums of the respective coefficients (λ1 + λ5 and λ1 + λ6)
differ from zero using a Wald test (Greene, 2008). We find a weakly
significant positive effect of donation size for the combined frame
(p b .10), but for the nonmonetary frame, the effect is only close to
significance (p = .11). Thus, the results support H1a for the two frames
with a monetary component and without a financial trade-off. The effect
of donation size on brand choice is negative when consumers face a
financial trade-off: A Wald test shows that the sum of the coefficients
λ1 and λ7 is significantly negative (p b .01), thereby supporting H1b.
To provide a sense of the strength of the effects of donation size on
brand choice, we simulate the changes in brand choice probability for
different donation sizes and frames. For the simulation, we use the
estimated parameter means from Model 2. We assume that a consumer
chooses between two brands that are equally preferred (category
dummies and PREPREF equal zero). For the demographic variables, we
use the most frequent levels (i.e., no children, church member, fulltime occupation). We present the simulation results in Fig. 2.
Fig. 2 reveals that the positive impact of donation size for frames
without financial trade-offs is moderate. For example, with a monetary
frame, a campaign with a donation of 1% of the price increases brand
choice probability by 14.3 percentage points (from 50% without a
campaign to 64.3%). Increasing the donation size to 20% of the price
earns the firm another 5.6 percentage points in brand choice probability,
which is unlikely to offset the loss in margin. For a donation of 50% of the
price, brand choice probability increases to 77.8%.
185
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Table 5
Model fit and improvement.
Model
Brand choice
Log likelihood
Likelihood ratio test:
Reference model
Chi2
(p)
Brand image
Log likelihood
Likelihood ratio test:
Reference model
Chi2
(p)
Model 1
Model 2
Model 3
Model 4
−2360.015
−2318.506
−2317.138
−2314.786
Model 1
83.017
(b.001)
Model 2
2.737
(.255)
Model 3
4.705
(.582)
−6317.384
−6316.166
−6312.979
Model 1
28.210
(b.001)
Model 2
2.437
(.296)
Model 3
6.373
(.041)
−6329.615
Model 4*
Notes: N=4,686. p values are printed in bold if pb.10.
For a monetary frame with a financial trade-off, brand choice
probability is 70.4% for a 1% donation but falls to 6.7% for a 50% donation
and an equal competitive discount. Here, the negative effect of donation
size on brand choice is substantial. The simulation demonstrates that for
donations of up to 13.1% of the price, consumers prefer a CM campaign
over a price reduction of the same size, but when donations increase
Table 6
Parameter estimates for brand choice and brand image models.
Independent variables
CM success measure
Brand choice
Brand image
Parameter estimates (standard errors)
Model 2
Mean
Category variables†
β1 PREPREF_CHOCOLAT
β2 PREPREF_TOOTHPASTE
β3 PREPREF_BEER
β4 PREPREF_DETERGENT
Concomitant variables††
γ Constant
η1 Child_YES
η2 Church membership_YES
η3 Church membership_NO_Res
η4 Occupation_FULL
η5 Occupation_PART
λ1 DONSIZ
λ2 NONMON
λ3 COMBI
λ4 TRADE-OFF
λ5 DONSIZ × NONMON
λ6 DONSIZ × COMBI
λ7 DONSIZ × TRADE-OFF
λ8 DONSIZ2
λ9 DONSIZ2 × COMBI
Model 2
SD
1.752***
(.317)
1.295***
(.213)
1.791***
(.270)
.921***
(.082)
.423
(.311)
.352***
(.135)
.103
(.122)
.580*
(.308)
.051
(.138)
.093
(.163)
.014**
(.006)
.273
(.235)
.148
(.225)
.362
(.255)
−.002
(.009)
−.002
(.009)
−.085***
(.014)
Mean
Model 4*
SD
.204***
(.021)
.195***
(.016)
.395***
(.030)
.230***
(.019)
.832**
(.293)
.005
(.014)
.219
(1.224)
.267
(.703)
1.311***
(.332)
.001
(.019)
.009
(.018)
.020
(.017)
.124
(.123)
.080*
(.048)
.007
(.044)
.064
(.106)
−.024
(.049)
−.043
(.058)
−.540**
(.218)
−.101
(.086)
−.176**
(.080)
−.026
(.078)
1.122***
(.348)
1.314***
(.320)
.027
(.306)
Mean
SD
.203***
(.021)
.197***
(.016)
.396***
(.031)
.231***
(.019)
.101
(.308)
.500*
(.258)
.298***
(.091)
.051
(.105)
.374***
(.080)
.162
(.966)
.721**
(.316)
.010
(.434)
.156
(.134)
.070
(.047)
.014
(.043)
.052
(.105)
−.027
(.049)
−.045
(.058)
−1.203**
(.581)
−.092
(.090)
−.301**
(.105)
−.032
(.101)
1.105**
(.371)
3.525**
(1.097)
.021
(.438)
1.413
(1.096)
−4.703**
(2.185)
.126
(.104)
.253
(.377)
.230**
(.113)
.011
(.114)
.332***
(.096)
.488
(.578)
.613
(.509)
.196
(.323)
.396
(.763)
.957
(1.536)
Notes: N = 4686; * p b .10; ** p b .05; *** p b .01 (two-sided); SD = Standard deviation; Standard errors in parentheses; † Category constants available upon request; †† Donation size
rescaled (divided by 100) for brand image models.
186
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
0.4
Nonmonetary
∆ Choice probabilities
0.3
0.2
Combined
0.1
0.0
-0.1 0
10
20
30
40
50
Monetary
-0.2
-0.3
-0.4
-0.5
Donation size (% of price)
Monetary
with financial
trade-off
Fig. 2. Change in choice probabilities through a CM campaign.
further, CM can no longer compete with an equivalent competitive
price promotion. From this point on, consumers are more attracted by
the competing firm's discount than by the focal firm's CM campaign.
The finding that consumers would rather have the money for themselves
than donate it to a cause is in line with research on willingness to pay for
ethical products, which shows that consumers are willing to pay only a
limited premium for social attributes (Auger, Devinney, Louviere, &
Burke, 2008).
In summary, our results suggest that the moderating effect of a
financial trade-off can explain most of the equivocal findings on the
effect of donation size on tactical CM success in previous research
(Table 1). We find that the effect of donation size is positive if consumers
face no financial trade-off, but becomes negative when larger donations
induce higher costs to consumers.
5.2. Brand image
Table 5 presents model fit for the regression models with brand
image as the dependent variable. We use likelihood ratio tests to
test for improvements in fit in our hierarchy of nested models.
Model 1 includes demographics and the three experimental factors
as concomitant variables. In Model 2, we add interaction effects of
donation size with donation framing and a financial trade-off, and find
that model fit improves significantly. Thus, the moderators affect the
impact of donation size on brand image. Next, we add nonlinear effects
of donation size, but the incorporation of a quadratic term for donation
size in Model 3 does not improve model fit. Hence, the effect of donation
size on brand image is not nonlinear per se. Model 4 includes quadratic
terms for the interactions of donation size with donation framing and a
financial trade-off. In Model 4, though, we encounter problems with
multicollinearity; thus, we exclude the quadratic terms DONSIZ2 ×
NONMON and DONSIZ2 × TRADE-OFF from the model.8 The reduced
Model 4* represents a significant improvement over Model 3. We
therefore use Models 2 and 4* to test our hypotheses and list their
parameter estimates in Table 6. Again, the parameters for all control
variables have plausible signs: Previous preferences relate positively
to brand image (p b .01), and the effect of CM on brand image is more
favorable for respondents with children in Model 2 (p b .10).
In Model 2, we find support for H2a: The impact of donation size is
negative when the frame is purely monetary, regardless of whether
the competitive brand is on promotion or not. Specifically, the
significant negative coefficient λ1 shows that the effect of donation
size is negative for a monetary frame without a financial trade-off, and
the interaction with a financial trade-off (λ7) is not significant. A t-test
further reveals that the sum of λ1 and λ7 is significantly negative
8
This modification does not limit our insights. In several alternative models, we find no
significant parameters for the terms we exclude, and the nonlinear effect of donation size
for a combination frame remains stable.
(p b .05), providing formal support for H2a. The sum of λ1 and λ5 is
significantly positive (p b .05), which supports H2b: The impact of
donation size is positive when the frame is purely nonmonetary.
Finally, we test for a nonlinear effect of donation size for a combined
frame, using Model 4*. With respect to the expected inverted U shape,
we use t-tests pertaining to the sum of the coefficients for donation
size and its interaction with the combination frame (λ1 + λ6), as well
as the sum of the two quadratic terms (λ8 + λ9) (Jaccard et al., 1990).
The sum of λ1 and λ6 is positive and significantly different from zero
(p b .05). The sum of λ8 and λ9 is weakly significant and negative
(p b .10). That is, the effect of donation size follows an inverted U
shape for a combined frame, and H2c is supported.
To illustrate the effects of donation size on brand image for the
different frames and to assess the strength of the effects, we again run
a simulation. We use the estimated parameter means from Model 4*,
and the same assumptions about PREPREF and demographics as in the
brand choice simulation. Fig. 3 presents the results.
Fig. 3 reveals the negative effect of donation size for monetary
frames, the positive effect for the nonmonetary frame, and the inverted
U-shaped effect for a combined frame. When a CM campaign presents
both monetary and nonmonetary information, donation size first has a
positive effect on brand image and then a negative one. The turning
point is reached at a donation of 35.3% of the price — well within the
range of realistic donation sizes.
CM with large monetary donations (N 15.9% of the price) and very
small donations with a combined frame (b6.9% of the price) hurt brand
image. The former finding is in line with our reasoning that high
monetary donations might lower consumers' perceived effectiveness.
The latter indicates that with a combined frame, transparency is counterproductive in the case of very small donations. Revealing the exchange
ratio between money and the charitable object demonstrates how few
resources are necessary to achieve a considerable outcome. In turn,
consumers likely perceive very small combined donations (e.g., 2 cents
equaling 0.4 vaccinations in our study) as paltry, which results in lower
perceived altruism.
Overall, we find changes between +.37 and −.12 on a seven-point
scale; brand image in the control group was 4.95. Given that the wellknown brands in our study possess established images that are unlikely
to change much because of a single experimental treatment, these
effects are substantial. In summary, Fig. 3 shows that donation size can
have substantial effects on brand image and highlights the importance
of considering donation framing when deciding on the size of the
donation.
6. Underlying drivers
To derive our hypotheses about the effects of donation size on tactical
and strategic success, we relied on different underlying drivers. In a
second study, which is exploratory in nature, we collected data on these
187
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
0.4
Nonmonetary
∆ Brand Image
0.3
Combined
0.2
0.1
Monetary
0.0
0
10
20
30
40
50
-0.1
-0.2
Monetary
with financial
trade-off
Donation size (% of price)
Fig. 3. Change in brand image through a CM campaign.
potential underlying drivers, and analyzed how they affect brand choice
and brand image. For this purpose, we regressed the two success
measures on warm glow, perceived altruism, and perceived effectiveness.
6.1. Data
We included a subset of the stimuli from our large-scale survey. In
contrast to our main study, we varied product category betweensubjects because of the longer questionnaire, which now included
measures on potential underlying drivers. To keep the number of
experimental groups tractable, we used fewer donation sizes (2.5, 10,
30 and 50% of product price) and only two product categories (chocolate
bars and toothpaste). We employed the same frames as in our large-scale
study (monetary, nonmonetary, and combined without a financial tradeoff and monetary with a competitive price promotion). The three frames
without financial trade-offs and the monetary frame with the financial
trade-off were combined with the four levels of donation size. We
employed all combinations for the two product categories, resulting in
32 conditions, which we varied between-subjects.
The procedure and measurements for the success variables were the
same as in our first study. In addition, we included measures to capture
the underlying drivers. After making their brand choice decisions and
brand image assessments, participants indicated their warm glow,
perceived effectiveness of the campaign and the perceived altruism of
the company running the campaign (all on seven-point multi-item
scales, see Appendix C).9
We invited members of an online access panel in Germany to
participate in our survey. As an incentive, respondents could participate
in a drawing to win Amazon gift cards. Between August and October
2011, 1402 respondents answered the questionnaire. We excluded 34
participants due to a response time of less than 2.5 min (the mean
was 7.8 min) or because they clicked through all multi-item scales
(straight line response on all scales). The final data set contains 1368
complete observations. Among our respondents, 62% are women. On
average, they are 33 years of age and live in households with 1.2 people.
The most respondents have a monthly net household income between
1000€ and 2000€. More than half (60%) are church members, and 25%
have children. 55% of the respondents are employed, and 31% are
students. Thus, the sample's demographics are very similar to those of
our main study. We do not find significant differences between the
experimental groups on any consumer demographics (i.e., age, gender,
household size, income, church membership, parenthood), previous
donation behavior or brand-related variables such as brand awareness
and brand preference (p N .10).
9
We also measured self-sufficiency and ease of imagination of the donation as
additional potential drivers to test for alternative explanations. Because these did not
prove to be relevant, we excluded them from the analysis. Full results are available from
the authors upon request.
6.2. Results
Our hypotheses are based on the reasoning that brand choice is
driven by warm glow, and brand image is driven by perceived altruism
of the firm and perceived effectiveness of the campaign. Therefore, we
regressed both success measures on these three potential drivers. We
also included category-specific intercepts and controlled for stated
preference heterogeneity with a PREPREF variable for each category.10
A Chow test for the regression model of brand image and a likelihood
ratio test for the logit model of brand choice indicate that pooling across
the product categories is appropriate (p N .10). All variance inflation
factors are below 1.68, indicating no problems with multicollinearity.
The estimation results are displayed in Table 7. For the brand choice
model, fit is good, as indicated by the value of .328 for Nagelkerke's R2. In
line with our reasoning, the only significant driver of brand choice is
warm glow, which is crucial for the utility consumers derive from a
CM campaign. In contrast, perceived altruism and effectiveness do not
disclose significant effects.
For brand image, the linear regression's R2 value of .159 is satisfactory,
given that we study the same well-known brands as in our main study,
for which images have been formed over a long time in the consumer's
mind. In line with our reasoning, perceived altruism and perceived
effectiveness both exert a significant positive effect on brand image.
Warm glow also has a significant effect, which most likely represents a
spillover of the good feeling consumers experience through the campaign
onto the brand's image. However, the coefficient for warm glow is the
smallest. Thus, the primary drivers of brand image are perceived altruism
and perceived effectiveness, which both relate to what consumers infer
from CM about the brand.
In summary, this study suggests that in CM, brand choice and brand
image are indeed affected by different underlying drivers. For brand
choice, the utility the consumer derives from the campaign is crucial,
and this is determined by the warm glow the campaign triggers. In
contrast, brand image is mainly affected by the inferences consumers
make about the brand involved in the campaign. It is critical for
consumers to believe in the company's sincere motives and that the
donation will be used as promised.
7. Summary and implications
We have investigated the impact of donation size on the effect of CM
on brand choice and brand image in a large-scale experimental survey
with different product categories. An additional exploratory study
10
Since our dataset contains only one observation per respondent, we do not model
unobserved heterogeneity.
188
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Table 7
Results underlying drivers.
Independent variables
CM success measure
Brand choice
Brand image
Parameter estimates (standard errors)
Category variables†
PREPREF_CHOCOLATE
PREPREF_TOOTHPASTE
1.190*** (.103)
.704*** (.092)
.162*** (.026)
.190*** (.030)
Underlying drivers
Warm glow
Perceived effectiveness
Perceived altruism
Nagelkerke's R2
.393*** (.053)
.091 (.062)
−.027 (.064)
.328
.046** (.019)
.067*** (.023)
.111*** (.023)
Chi2
(p)
Log likelihood
R2 (Adj. R2)
F
(p)
352.253
(b.001)
−631.189
.159 (.155)
42.741
(b.001)
Notes: N =1,368; * pb .10; ** pb .05; *** pb .01 (two-sided); Standard errors in parentheses; p
values are printed in bold if p b .10. † Category constants available upon request.
provides insights into the underlying drivers of consumer behavior in
the context of CM. Our key findings are the following:
• The effect of donation size on brand choice depends on the presence of
a financial trade-off. If consumers face no trade-off, larger donations
increase brand choice probability. However, if consumers have to
choose between doing good and savings for themselves, larger
donations and larger trade-offs respectively will decrease brand
choice probability.
• The effect of donation size on brand image depends on donation
framing. With a monetary frame, larger donations are less favorable
for brand image and CM campaigns with high monetary donations
can even hurt brand image. For a purely nonmonetary frame, donation
size has a positive effect on brand image, and for a combined frame, the
effect follows an inverted U shape.
• In CM, tactical and strategic success appear to be driven by different
mechanisms. Brand choice depends on consumers' utility and thus
warm glow is the crucial driver. Brand image improves when
consumers make positive inferences about the company and thus is
mainly affected by the perceived altruism of the company and the
perceived effectiveness of the campaign.
Our systematic analysis of donation size helps explain why donation
size can have positive, negative, or no effects on CM success. First, we
show that tactical success (brand choice) and strategic success (brand
image) are affected differently. This is attributed to differences in their
underlying drivers.
Second, we identify two new moderators of the effect of donation
size on CM success. For brand choice, larger donations exert a favorable
effect as long as consumers face no financial trade-offs, whereas the
effect is negative when an alternative brand offers a price promotion
of equal size. This finding goes a long way toward explaining the
contradictory previous results summarized in Table 1. Except for
Chang (2008) all studies that find a negative impact include a financial
trade-off (e.g., Arora & Henderson, 2007; Strahilevitz, 1999) while in
none of the studies reporting a positive effect a larger donation comes
at higher cost to consumers (e.g., Koschate-Fischer et al., 2012; Pracejus
et al., 2003/04). For brand image, the effect of donation size is moderated
by donation framing. Through the introduction of this moderator,
we extend the scope of previous research on donation size, which has
studied only monetary frames.
Third, we find that the effect of donation size is nonlinear for a frame
that combines monetary and nonmonetary information: It follows
an inverted U shape. So far, nonlinear effects have only been studied for
monetary frames (Koschate-Fischer et al., 2012; Pracejus et al., 2003/
04). Our results support the findings of Pracejus et al. (2003/04) in that
we do not find nonlinear effects for monetary frames. We note that
Koschate-Fischer et al. (2012) find a concave effect for a monetary
frame, but with a different success measure, i.e., willingness-to-pay.
Their result may simply reflect consumers' reluctance to pay more for
larger donations, whereas in our study product price remained constant
even with larger donations.
Our results suggest important implications for managers who intend
to use cause-related marketing. First, large donations are not essential for
tactical success. CM can be a cost-effective sales promotion instrument to
increase brand choice because even small donations have a substantial
positive impact. In contrast, most price promotions must pass a 10–
20% discount threshold to significantly affect purchase intentions and
behavior (Gupta & Cooper, 1992; van Heerde et al., 2001). Rising CM
donations increase brand choice probability only moderately, which is
most likely not sufficient to offset their additional costs.
Second, large CM donations may not be able to compete in a
promotion-intensive environment in which consumers face trade-offs
between doing something good and savings for themselves. In many
FMCG categories, price discounts are in the range of approximately 20%
of the product price (e.g., van Heerde, Leeflang, & Wittink, 2000).
Donating that much money to a cause would increase brand choice
probability but not enough to maintain market share when the
competitor offers an equivalent price promotion. However, van Heerde,
Leeflang, and Wittink (2004) observe huge variations in price discounts,
ranging from 5% to 51% of the price. Against smaller discounts, CM is likely
to prevail.
Third, larger donations can help or hurt strategic success depending
on their framing. When CM campaigns use a monetary frame, the effect
of CM on brand image becomes less favorable with increasing donations
and can even turn negative. In contrast, donation size has a positive effect
on brand image for campaigns with nonmonetary framing. Finally,
managers may combine monetary and nonmonetary information in
their donation framing. In this case, a medium donation size is optimal
for brand image. Therefore, managers should carefully align donation
size and donation framing in CM to create a positive effect on strategic
success. If they succeed in this task, CM may be an attractive alternative
to price promotions, which – even if possibly more effective in the short
run – typically hurt brand loyalty in the long run (Neslin & van Heerde,
2009). CM, in contrast, may help managers to increase consumers'
brand loyalty by adding a philanthropic component to their brand.
Finally, managers should pay attention to their communication in CM
campaigns. They should appeal to the warm glow consumers derive from
participating in the campaign to enhance tactical success. At the same
time, to improve brand image, they need to make a credible claim that
the company is committed to help the cause and that the donations
reach their targets.
Our study also has some limitations that provide opportunities for
further research. Our measures may suffer from a social desirability bias
(e.g., Lautenschlager & Flaherty, 1990; Nancarrow, Brace, & Wright,
2001), such that respondents might be more likely to choose the CM
brand and evaluate its brand image more favorably than they would
in a real purchase situation. Even if our data suffered from this social
desirability bias, though, it would be unlikely to affect our results
regarding donation size because the bias would be the same for all
experimental groups. However, our measures of the absolute effect
of CM would be biased, which would make it challenging to derive
specific implications for the optimal donation size. We suspect that
our data are not affected by a strong social desirability bias because
we find negative effects of CM on brand image and a strong effect of a
competitive promotion on brand choice. Nevertheless, it would be
worthwhile to validate our results with field data.
Furthermore, we study only fast moving consumer goods.
Investigating the impact of donation size and its two moderators
for durable goods might lead to further insights. For example, the
189
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
effect of a financial trade-off might be different for higher priced
products.
Previous research has indicated that CM is more successful for hedonic
than for utilitarian products (e.g., Strahilevitz, 1999; Strahilevitz & Myers,
1998). The proposed underlying mechanism is that CM reduces
customers' guilt associated with indulging in hedonic products. We find
no differences for the effects of donation size between hedonic and
utilitarian products, maybe because in our scenarios we explicitly told
consumers that they want to buy a product, such that their purchase
incidence decision was already made. Further research should look
more deeply into the role of category type and test whether the feeling
of guilt indeed plays a role.
Another limitation of our study is that brand image is formed over a
long period of time, such that it could be interesting to validate our
results with a study that repeatedly measures brand image over a longer
time period. This would also present an opportunity to measure how
improvements in brand image translate into brand equity and affect
purchase decisions in the long run (Keller, 1993).
Finally, we have studied a specific type of financial trade-off where the
competitive brand offers a price promotion. Our results are interesting for
firms who consider using CM to compete in a promotion-intensive
environment. Future research may want to study the trade-offs that
occur when the CM brand increases its price. The respective results
might help firms decide whether they can pass on the costs of the
donations to consumers.
Despite these limitations, we think that our study yields interesting
results with important implications for managers and researchers, and
we hope that further research will build on it.
Appendix A
Product stimuli.
Product category
Brand
Size
Chocolate bars
KitKat
Duplo
Colgate
Odol-med3
Bitburger
Warsteiner
Persil
Ariel
200 g
1.99€
75 ml
1.99€
Toothpaste
Beer
Detergent
Notes: Italics = CM brand.
Appendix B
Example stimuli for monetary CM campaign and competitive price promotion.
Price
24 × 0.33 l
12.49€
4.75 kg
12.49€
190
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Appendix
(continued)
D (continued)
Appendix C
Multi-item scales.
Measures
CM success
Brand image
Rating the CM brand on Likert scales:
−3 = bad and +3 = good
−3 = not likeable and +3 = likeable
−3 = low quality and +3 = high quality
−3 = not trustworthy and +3 =
trustworthy
−3 = unpleasant and +3 = pleasant
−3 = unattractive and +3 = attractive
Underlying drivers (Study 2)
Warm glow
Extent to which participants agreed/disagreed
with the following statements:
When I purchase [CM brand name], I feel
good because I do not only spend money for
myself but also for other people.
I feel comfortable if I donate for a good cause
by purchasing [CM brand name].
I am pleased that I do not only get a product
by purchasing [CM brand name], but that I
also do a good deed at the same time.
Perceived effectiveness of CM campaign
Extent to which participants agreed/disagreed
with the following statements:
I believe that the donated money reaches the
needy persons.
I am convinced that little of the donated
money is wasted.
I assume that the donated money will be
distributed in favor of the cause.
I trust in the fact that the donated money will
be used for the cause.
I believe that the company actually donates
as much as stated in the CM campaign.
Perceived altruism of company
Extent to which participants agreed/disagreed
with the following statements:
The manufacturer conducts the campaign in
order to do a good deed.
The campaign is an honest effort.
The manufacturer is not truly committed to
the purpose of the donation.
Source
α
Völckner et al. (2008)†
.95
(S1)
.92
(S2)
Arora and Henderson
(2007), Andreoni
(1989), Fries et al.
(2010), and Monin
(2003)
.93
.93
Fries et al. (2010),
Sargeant and Lee (2004),
and Webb Green, and
Brashear (2000)
Fries et al. (2010),
Nowak (2004),
Strahilevitz (2003), and
Webb et al. (2000)
.93
Multi-item scales.
Appendix D
Variable specifications
Category variables
CAT_CHOCOLATE
CAT_TOOTHPASTE
CAT_BEER
CAT_DETERGENT
PREPREF
Demographic variables
Child_YES
Church membership_YES
Church membership_NO_Res
Occupation_FULL
Occupation_PART
CM variables
DONSIZ
NONMON
Specification
COMBI
1 if the CM campaign includes combined
nonmonetary and monetary donation framing, 0
otherwise
1 if the competitive brand offers a price promotion, 0
otherwise
TRADE-OFF
References
Notes: α = Cronbach’s alpha in our study; † we added the item trustworthy; S = Study.
Variable
Variable
Specification
1 if product category is chocolate bars, 0 otherwise
1 if product category is toothpaste, 0 otherwise
1 if product category is beer, 0 otherwise
1 if product category is detergent, 0 otherwise
Number of times the CM brand was bought in the last
three purchases in the category prior to the survey,
minus the number of times the other brand was
bought/Divided by 3
1 if subject has children, 0 otherwise
1 if subject is member of a church, 0 otherwise
1 if subject did not indicate church membership, 0
otherwise
1 if subject works full time, 0 otherwise
1 if subject works part time, 0 otherwise
Donation size of the CM offer
1 if the CM campaign includes nonmonetary
donation framing, 0 otherwise
Ailawadi, K. L., Gedenk, K., Lutzky, C., & Neslin, S. A. (2007). Decomposition of the sales
impact of promotion-induced stockpiling. Journal of Marketing Research, 44(3),
450–467.
Ailawadi, K. L., Gedenk, K., & Neslin, S. A. (1999). Heterogeneity and purchase event
feedback in choice models: An empirical analysis with implications for model
building. International Journal of Research in Marketing, 16(3), 177–198.
Alba, J. W., & Hutchinson, J. W. (1987). Dimensions of consumer expertise. Journal of
Consumer Research, 13(4), 411–454.
Anderson, N. H. (1981). Foundations of information integration theory. New York: Erlbaum.
Andreoni, J. (1989). Giving with impure altruism: Applications to charity and ricardian
equivalence. Journal of Political Economy, 97(6), 1447–1458.
Andreoni, J. (1990). Impure altruism and donations to public goods: A theory of
warm-glow giving. The Economic Journal, 100(401), 464–477.
Arora, N., & Henderson, T. (2007). Embedded premium promotion: Why it works and
how to make it more effective. Marketing Science, 26(4), 514–531.
Auger, P., Devinney, T. M., Louviere, J. J., & Burke, P. F. (2008). Do social product features have
value to consumers? International Journal of Research in Marketing, 25(3), 183–191.
Barnes, N. G. (1992). Determinants of consumer participation in cause-related marketing
campaigns. American Business Review, 10(2), 21–24.
Barone, M. J., Miyazaki, A.D., & Taylor, K. A. (2000). The influence of cause-related
marketing on consumer choice: Does one good turn deserve another? Journal of the
Academy of Marketing Science, 28(2), 248–262.
Bouten, L. M., Snelders, D., & Hultink, E. J. (2011). The impact of fit measures on the
consumer evaluation of new co-branded products. Journal of Product Innovation
Management, 28(4), 455–469.
Burnett, J. J., & Wood, V. R. (1988). A proposed model of the donation decision process. In
E. Hirschman, & J. Sheth (Eds.), Research in consumer behavior, 3. (pp. 1–47)Greenwich,
CT: Elsevier JAI.
Chang, C. -T. (2008). To donate or not to donate? Product characteristics and framing
effects of cause-related marketing on consumer purchase behavior. Psychology &
Marketing, 25(12), 1089–1110.
Cause-related marketing: Impact of size of corporate donation and size of cause-related
promotion on consumer perceptions and participation. Dahl, D. W., & Lavack, A.M.
(Eds.). (1995). AMA Winter Educators' Conference Proceedings, 6. (pp. 476–481).
Eckel, C. C., & Grossman, P. J. (2003). Rebate versus matching: Does how we subsidize
charitable contributions matter? Journal of Public Economics, 87(3/4), 681–701.
Ellen, P.S., Mohr, L. A., & Webb, D. J. (2000). Charitable programs and the retailer: Do they
mix? Journal of Retailing, 76(3), 393–406.
Fries, A. J. (2010). The effects of cause-related marketing campaign characteristics — A
literature review. Marketing — Journal of Research and Management, 6(2), 145–157.
Fries, A. J., Gedenk, K., & Völckner, F. (2010). Cause-related marketing: Designing successful
campaigns. University of Cologne working paper.
Garretson Folse, J. A., Niedrich, R. W., & Landreth Grau, S. (2010). Cause-related marketing:
The effect of purchase quantity and firm donation amount on consumer inferences
and participation intentions. Journal of Retailing, 86(4), 295–309.
German Fundraising Association (2009). Spendenbilanz ausgewählter Organisationen
2005–2008. Retrieved March 15, 2010, from. http://www.fundraisingverband.de/
fileadmin/pdf_upload/1Spendenbilanz_2005-2008.pdf
Green, D. P., Kahneman, D., & Kunreuther, H. (1994). How the scope and method of public
funding affect willingness to pay for public goods. Public Opinion Quarterly, 58(1), 49–67.
Greene, W. H. (2008). Econometric analysis (6th ed.)Upper Saddle River, NJ: Pearson
Prentice Hall.
Gupta, S., & Cooper, L. G. (1992). The discounting of discounts and promotion thresholds.
Journal of Consumer Research, 19(3), 401–411.
Haruvy, E., & Popkowski Leszczyc, P. T. L. (2009). Bidder motives in cause-related
auctions. International Journal of Research in Marketing, 26(4), 324–331.
Holmes, J. H., & Kilbane, C. J. (1993). Cause-related marketing: Selected effects of price and
charitable donations. Journal of Nonprofit & Public Sector Marketing, 1(4), 67–83.
Horsky, D., Misra, S., & Nelson, P. (2006). Observed and unobserved preference heterogeneity
in brand-choice models. Marketing Science, 25(4), 322–335.
Human, D., & Terblanche, N. S. (2012). Who receives what? The influence of the donation
magnitude and donation recipient in cause-related marketing. Journal of Nonprofit
and Public Sector Marketing, 24(2), 141–160.
IEG (2011). IEG sponsorship report. Retrieved June 2, 2011, from. http://www.sponsorship.
com/About-IEG/Press-Room/Economic-Uncertainty-To-Slow-Sponsorship-Growth-In.
aspx
Jaccard, J. (2001). Interaction effects in logistic regression. Thousand Oaks, CA: Sage.
Jaccard, J., Turrisi, R., & Wan, C. K. (1990). Interaction effects in multiple regression. Thousand
Oaks, CA: Sage.
Kahneman, D., & Knetsch, J. L. (1992). Valuing public goods: The purchase of moral
satisfaction. Journal of Environmental Economics and Management, 22(1), 57–70.
S.S. Müller et al. / Intern. J. of Research in Marketing 31 (2014) 178–191
Kahneman, D., Ritov, I., Jacowitz, K. E., & Grant, P. (1993). Stated willingness to pay for
public goods: A psychological perspective. Psychological Science, 4(5), 310–315.
Keller, K. L. (1993). Conceptualizing, measuring, and managing customer-based brand
equity. Journal of Marketing, 57(1), 1–22.
Koschate-Fischer, N., Stefan, I. V., & Hoyer, W. D. (2012). Willingness to pay for
cause-related marketing: The impact of donation amount and moderating effects.
Journal of Marketing Research, 49(6), 910–927.
Landreth Grau, S., Garretson Folse, J. A., & Pirsch, J. (2007). Cause-related marketing: An
exploratory study of campaign donation structures issues. Journal of Nonprofit &
Public Sector Marketing, 18(2), 69–91.
Lautenschlager, G. J., & Flaherty, V. L. (1990). Computer administration of questions: More
desirable or more social desirability? Journal of Applied Psychology, 75(3), 310–314.
Monin, B. (2003). The warm glow heuristic: When liking leads to familiarity. Journal of
Personality and Social Psychology, 85(6), 1035–1048.
Nancarrow, C., Brace, I., & Wright, L. T. (2001). Tell me lies, tell me sweet little lies: Dealing
with socially desirable responses in market research. Marketing Review, 2(1), 55–69.
Neslin, S. A., & van Heerde, H. J. (2009). Promotion dynamics. Foundation and Trends in
Marketing, 3(4), 177–268.
Nowak, L. I. (2004). Cause marketing alliances: Corporate associations and consumer
responses. Journal of Food Products Marketing, 10(2), 33–48.
Nunes, J. C., & Park, C. W. (2003). Incommensurate resources: Not just more of the same.
Journal of Marketing Research, 40(1), 26–38.
Olsen, G. D., Pracejus, J. W., & Brown, N. R. (2003). When profit equals price: Consumer
confusion about donation amounts in cause-related marketing. Journal of Public
Policy & Marketing, 22(2), 170–180.
Palazon, M., & Delgado-Ballester, E. (2009). Effectiveness of price discounts and premium
promotions. Psychology & Marketing, 26(12), 1108–1129.
Polonsky, M. J., & Wood, G. (2001). Can the overcommercialization of cause-related
marketing harm society? Journal of Macromarketing, 21(1), 8–22.
Pracejus, J. W., Olsen, G. D., & Brown, N. R. (2003/04). On the prevalence and impact of
vague quantifiers in the advertising of cause-related marketing (CRM). Journal of
Advertising, 32(4), 19–28.
Ross, J. K., Patterson, L. T., & Stutts, M.A. (1992). Consumer perceptions of organizations
that use cause-related marketing. Journal of the Academy of Marketing Science,
20(1), 93–97.
Ross, J. K., Stutts, M.A., & Patterson, L. (1991). Tactical considerations for the effective use
of cause-related marketing. The Journal of Applied Business Research, 7(2), 58–65.
Sargeant, A., & Lee, S. (2004). Trust and relationship commitment in the United Kingdom
voluntary sector: Determinants of donor behavior. Psychology & Marketing, 21(8),
613–635.
Simonin, B.L., & Ruth, J. A. (1998). Is a company known by the company it keeps?
Assessing the spillover effects of brand alliances on consumer brand attitudes.
Journal of Marketing Research, 35(1), 30–42.
191
Smith, S. M., & Alcorn, D. S. (1991). Cause marketing: A new direction in the marketing of
corporate responsibility. Journal of Consumer Marketing, 8(3), 19–35.
Strahilevitz, M. (1999). The effects of product type and donation magnitude on willingness to
pay more for a charity-linked brand. Journal of Consumer Psychology, 8(3), 215–241.
Strahilevitz, M. (2003). The effects of prior impressions of a firm's ethics on the success of
a cause-related marketing campaign: Do the good look better while the bad look
worse? Journal of Nonprofit and Public Sector Marketing, 11(1), 77–92.
Strahilevitz, M., & Myers, J. G. (1998). Donations to charity as purchase incentives: How
well they work may depend on what you are trying to sell. Journal of Consumer
Research, 24(4), 434–446.
Subrahmanyan, S. (2004). Effects of price premium and product type on the choice of
cause-related brands: A Singapore perspective. Journal of Product & Brand
Management, 13(2), 116–124.
Train, K. E. (2009). Discrete choice methods with simulation (2nd ed.)Cambridge: Cambridge
University Press.
Vaidyanathan, R., & Aggarwal, P. (2005). Using commitments to drive consistency:
Enhancing the effectiveness of cause-related marketing communications. Journal of
Marketing Communications, 11(4), 231–246.
van den Brink, D., Odekerken-Schröder, G., & Pauwels, P. (2006). The effect of strategic
and tactical cause-related marketing on consumers' brand loyalty. Journal of
Consumer Marketing, 23(1), 15–25.
van Heerde, H. J., Leeflang, P.S. H., & Wittink, D. R. (2000). The estimation of pre- and
postpromotion dips with store-level scanner data. Journal of Marketing Research,
37(3), 383–395.
van Heerde, H. J., Leeflang, P.S. H., & Wittink, D. R. (2001). Semiparametric analysis to
estimate the deal effect curve. Journal of Marketing Research, 38(2), 197–215.
van Heerde, H. J., Leeflang, P.S. H., & Wittink, D. R. (2004). Decomposing the sales
promotion bump with store data. Marketing Science, 23(3), 317–334.
Varadarajan, P. R., & Menon, A. (1988). Cause-related marketing: A coalignment of
marketing strategy and corporate philanthropy. Journal of Marketing, 52(3), 58–74.
Völckner, F., Sattler, H., & Kaufmann, G. (2008). Image feedback effects of brand extensions:
Evidence from a longitudinal field study. Marketing Letters, 19(2), 109–124.
Webb, D. J., Green, C. L., & Brashear, T. G. (2000). Development and validation of scales to
measure attitudes influencing monetary donations to charitable organizations.
Journal of the Academy of Marketing Science, 28(2), 299–309.
WHO, & UNICEF (). Immunization summary: The 2010 edition. Retrieved January 22,
2010, from. http://www.childinfo.org/files/Immunization_Summary_2008_r6.pdf
Wymer, W., & Samu, S. (2009). The influence of cause marketing associations on product
and cause brand value. International Journal of Nonprofit and Voluntary Sector
Marketing, 14(1), 1–20.
Zdravkovic, S., Magnusson, P., & Stanley, S. M. (2010). Dimensions of fit between a brand
and a social cause and their influence on attitudes. International Journal of Research in
Marketing, 27(2), 151–160.
Intern. J. of Research in Marketing 31 (2014) 192–206
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
Full Length Article
Choosing a digital content strategy: How much should be free?☆
Daniel Halbheer a,⁎, Florian Stahl b, Oded Koenigsberg c, Donald R. Lehmann d
a
University of St. Gallen, Department of Economics, Varnbüelstrasse 19, CH-9000 St. Gallen, Switzerland
University of Mannheim, Department of Business Administration, L5, 2, D-68131 Mannheim, Germany
c
London Business School, Regent's Park, London, NW1 4SA, United Kingdom
d
Columbia Business School, Uris Hall, 3022 Broadway, New York, NY 10027, United States
b
a r t i c l e
i n f o
Article history:
First received in 27 August 2012 and was under
review for 8 months
Available online 7 November 2013
Area Editor: Kalyan Raman
Guest Editor: Marnik G. Dekimpe
Keywords:
Information goods
Content pricing
Sampling
Advertising
Dorfman-Steiner condition
a b s t r a c t
Advertising supported content sampling is ubiquitous in online markets for digital information goods. Yet, little is
known about the profit impact of sampling when it serves the dual purpose of disclosing content quality and generating advertising revenue. This paper proposes an analytical framework to study the optimal content strategy
for online publishers and shows how it is determined by characteristics of both the content market and the advertising market. The strategy choice is among a paid content strategy, a sampling strategy, and a free content
strategy, which follow from the publisher's decisions concerning the size of the sample and the price of the
paid content. We show that a key driver of the strategy choice is how sampling affects the prior expectations
of consumers, who learn about content quality from the inspection of the free samples. Surprisingly, we find
that it can be optimal for the publisher to generate advertising revenue by offering free samples even when sampling reduces both prior quality expectations and content demand. In addition, we show that it can be optimal for
the publisher to refrain from revealing quality through free samples when advertising effectiveness is low and
content quality is high. To illustrate, we relate our framework to the newspaper industry, where the sampling
strategy is known as the “metered model.”
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
Digital information goods have been available on the Internet for
almost twenty years. During that time, publishers have developed different strategies to distribute content. Some publishers provide all
their information for free, while some charge consumers for access to
their content. Other publishers employ a hybrid business model, giving
away a portion of their content for free and charging for access to the
rest of their content. Offering free content samples allows publishers
to both disclose their content quality and to generate revenues from
advertisements shown to online visitors. According to Alisa Bowen, general manager of The Wall Street Journal Digital Network, “working with
advertisers to offer open houses has proven to be one of the most
☆ We thank Asim Ansari, Jean-Pierre Dubé, Anthony Dukes, Jacob Goldenberg, Avi
Goldfarb, Raju Hornis, Ulrich Kaiser, Anja Lambrecht, Philipp Renner, Catherine Tucker
and seminar participants at the 11th ZEW ICT Conference (2013), the GEABA 2012
(Graz), the Marketing in Israel Conference 2011 (Tel Aviv), the INFORMS Marketing
Science Conference 2011 (Houston), the University of Hamburg, the HEC Paris, the
University of Passau, the University of Tilburg, and the University of Zurich for helpful
comments and suggestions. Daniel Halbheer gratefully acknowledges support from the
Swiss National Science Foundation through grant PA00P1-129097 and thanks the
Department of Economics at the University of Virginia for its hospitality while some of
this research was being undertaken.
⁎ Corresponding author.
E-mail addresses: daniel.halbheer@unisg.ch (D. Halbheer),
florian.stahl@bwl.uni-mannheim.de (F. Stahl), okoenigsberg@london.edu
(O. Koenigsberg), drl2@columbia.edu (D.R. Lehmann).
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.10.004
valuable and efficient ways to expose our premium content to new
readers and potential subscribers” (GlobeNewswire, 2012). The main
contribution of this paper is to provide a formal analysis of how publishers should choose between different digital content strategies.
Information goods such as digital content are experience goods;
hence consumers must have an experience to value them (Shapiro &
Varian, 1998). Offering free samples is a way for publishers to disclose
their content quality, thereby allowing consumers to have actual experience with the good before purchase. Digital information goods are particularly suitable for sampling because the costs of providing free
samples are negligible and publishers can include advertisements in
the free samples to generate advertising revenues. These two features
distinguish sampling of information goods from the typical sampling
of perishable or durable goods.
Recently, publishers introduced business models that allow consumers to select samples of their choice within a set sample size. A
prominent example of this is the “metered model” in the newspaper industry, where publishers offer a number of articles for free and charge
for access to the rest. Such “customer selected sampling” differs from
the approach where the publisher chooses not only the sample size
but also the sample content, which allows the firm to strategically
manipulate the sample and creates an environment where customers
are likely to discount the sample quality in estimating actual quality. A
recent study by the Newspaper Association of America (2012) shows
that 62% of the publishers employ a metered model, out of which 95%
offer up to twenty free articles monthly. For example, the New York
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
Times currently offers access to ten articles for free on its website each
month. Advertising supported sampling is also employed by distributors of music such as Spotify or Rhapsody. Allowing consumers to choose
which content to sample means that publishers have no control over the
content consumers actually sample. Taking this into account is important for publishers when setting the optimal sample size.
The business model where publishers set a sample size and let consumers choose which content to sample differs from versioning or
“freemium,” where a firm selected low-end version is offered for free
and consumers have to pay for access to the high-end version.1 Such
versioning of information goods is often observed in the software industry (see, for instance, Faugère & Tayi, 2007; Cheng & Tang, 2010). Customer selected sampling, in contrast, does not involve quality
differentiation: Within the set sample size, the publisher allows the consumers to sample any of its content for free.
This paper develops an analytical framework to study the optimal
content strategy for online publishers of information goods. Content
strategy consists of two decisions: the size of the sample and the price
of the paid content. The publisher is assumed to receive revenues
from content sales and from advertisements, which are embedded in
the free content. A key feature of sampling is that it serves the dual purpose of generating revenues from advertising and disclosing content
quality. Consumers have prior expectations about content quality,
which they update in a Bayesian fashion through inspection of the
free samples. The information transmitted through the free samples affects the consumers' posterior expectations about content quality,
which in turn influence expected content sales. Taking the consumers'
quality updating into account, the publisher faces a tradeoff between
an expansion effect (through learning) and a cannibalization effect
(through free offerings) on expected content demand induced by sampling. When the publisher makes its sampling and pricing decisions, it
should take both the effects on content demand and the impact on advertising revenue into account. We assume that the publisher can
choose among a “sampling strategy,” a pure “paid content strategy,” or
a pure “free content strategy.” Importantly, under a sampling strategy,
consumers are allowed to sample some content for free, while they
have to pay to access the remaining content (which corresponds to
the metered model in the newspaper industry).
We derive several important results. First, we show how the
publisher's advertising-sales revenue ratio and hence its optimal content strategy is determined by characteristics of both the content market and the advertising market. The analysis confirms the insight that
the elasticities of content demand and advertising demand jointly determine the publisher's optimal ratio of advertising revenue to sales revenue (Dorfman & Steiner, 1954). Importantly, the elasticity of
consumers' updated expectations with respect to the sample size
plays a key role in determining the ratio of advertising revenue to
sales revenue. This elasticity is positive if sampling increases consumers'
expectations about content quality. In this case, managers can expect
the advertising-sales revenue ratio to be low due to high revenues
from content sales (resulting from the expansion effect). In contrast, if
sampling reduces consumers' expectations about content quality, managers can expect the advertising-sales revenue to be high due to low
revenues from content sales (resulting from the cannibalization effect).
Second, we describe a Bayesian learning mechanism and derive expected content demand from model primitives. This allows us to provide insights into how the effects of pricing and sampling on expected
content demand are intertwined with the consumers' prior beliefs. We
find that managers must consider two demand regimes, depending on
whether consumers' prior expectations are “high” or “low.” In both
cases, expected content demand is decreasing in price and increasing
in expected posterior quality, as expected. Importantly, we show that
sampling can have a demand-enhancing effect through consumers'
1
Bhargava and Choudhary (2008) analyze optimal versioning of information goods.
193
learning when prior expectations are sufficiently low (even though
sampling produces a cannibalization effect). We establish the rule of
thumb that sampling increases content demand if the elasticity of consumers' updated expectations exceeds the ratio of sampled to paid
content.
Third, to study the publisher's optimal pricing and sampling decisions, we bring our model of the content market together with a
standard model of the advertising market. When content quality is
common knowledge, we show that a paid content strategy is optimal
for the publisher only if the effectiveness of advertising is sufficiently
low. For intermediate levels of advertising effectiveness, the publisher should employ a sampling strategy and generate revenues from
both sales and advertising. Once advertising is sufficiently effective,
the publisher should switch to a free content strategy. Thus, it can
be optimal for the publisher to offer content samples even if sampling solely cannibalizes content demand.
In the case where consumers are uncertain about actual content
quality and learn about it through inspection of the free samples, the
optimal strategy is determined by the relationship between advertising
effectiveness and quality expectations. As in the benchmark model, two
cut-off values of advertising effectiveness determine the publisher's
optimal content strategy: a lower bound that depends on prior quality
expectations (separating paid from sampling strategies) and an upper
bound that depends on posterior quality expectations (separating sampling from free content strategies). Interestingly, it can be optimal for
the publisher to generate advertising revenue by adopting a sampling
strategy even when sampling reduces both quality expectations and
content demand. In addition, we find that it can be optimal for the publisher to adopt a paid content strategy and to refrain from revealing high
quality through free samples.
Finally, we explore three extensions of the model to generate additional managerial insights. First, when the publisher also includes advertisements in the paid content, the cut-off values depend not only
on the relationship between advertising effectiveness and quality expectations, but also on the consumers' attitudes towards advertisements. Second, we provide insight on how the publisher's optimal
content strategy is determined when the willingness to pay for advertisements is related to content quality. Third, we introduce competition
in the content market and analyze how asymmetries in prior beliefs affect the equilibrium content strategies.
This paper is related to two literature streams. The first stream is on
media strategy in two-sided markets.2 For instance, Kind, Nilssen, and
Sørgard (2009) analyze how competition, captured by the number of
media platforms and content differentiation between platforms, affects
the composition of revenues from advertising and sales. Godes, Ofek,
and Sarvary (2009) investigate a similar question, focusing on competition between platforms in different media industries. Our paper examines optimal advertising supported content sampling and content
pricing when the firm can generate revenue from both content sales
and advertising. While papers that examine content sampling from different perspectives include Xiang and Soberman (2011) for preview
provision and Chellappa and Shivendu (2005) for piracy-mitigating
strategies, neither consider the impact of sampling on advertising revenues. To the best of our knowledge, optimal content sampling when
sampling impacts revenues from both content sales and online advertising has not been addressed by the literature.
This paper is also related to the broad literature on consumer learning about product attributes. Firms typically enable consumer learning
through disclosing information about their products and services. This
information can be disclosed in various ways, including through informative advertising (see Anderson & Renault, 2006, and Bagwell, 2007
for a comprehensive survey), or product descriptions or third-party reviews (Hotz & Xiao, 2013; Sun, 2011). Another way for firms to disclose
2
See Rysman (2009) for a general review of the two-sided markets literature. Anderson
and Gabszewicz (2006) provide a canonical survey of media and advertising.
194
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
Table 1
Components of the general framework.
General Framework
Explicit form
Variables
Publisher
Content parameters
N … content size
V … maximum content quality
Cost parameters
F … fixed production costs
cs … unit distribution costs
Assumed properties
Expected posterior quality
˜
VðnÞ
Expected content demand
DE ðp; nÞ ≡ Dðp; n; ˜
VðnÞÞ
˜′
V ðnÞ ≷ 0
Section 3.3
Dp b 0, Dn b 0, D Ṽ N 0
DEn N 0 … demand-enhancing sampling
DnE b 0 … demand-reducing sampling
Proposition 2
Lemma 2
Lemma 2
Decision variables
p … content price
n … sample size
Content market
Advertising market
Prior parameters
v0 … minimum estimate of V
α … uncertainty about v0
Posterior parameters
e
v0 ðnÞ
α+n
Preference parameters
θ … valuation of quality
ξ … ad attraction/ad repulsion
x … preferred product characteristic
τ … sensitivity to mismatch
Indirect utility
u(p,n)
Conditional indirect utility
ui(x;pi,ni)
Advertiser parameter
ϕ … advertising effectiveness
Inverse advertising demand
a(n) … ads in free content
ap(N − n) … ads in paid content
e ðnÞ
Endogenous ad effectiveness ϕ
information is through sampling. Heiman, McWilliams, Shen, and
Zilberman (2001) and Bawa and Shoemaker (2004) study how sampling affects demand and the evolution of market shares for consumer
goods, while Boom (2009) and Wang and Zhang (2009) investigate
sampling of information goods. However, when firms sample information goods, they only offer a portion of the good for free to avoid the
“information paradox” (Akerlof, 1970). Consumers' inferences about a
product's attributes are most naturally modeled in a Bayesian framework. Bayesian learning processes based on product experience have
been widely employed in the literature, for instance, by Erdem and
Keane (1996), Ackerberg (2003), and Erdem, Keane, and Sun (2008),
and we follow this approach here.
We organize the remainder of the paper as follows. Section 2
presents the general framework and describes how the publisher
operates in two markets: content and advertising. Section 3 describes
the content market and studies the impact of sampling on expected
content demand. Section 4 describes the advertising market. Section 5
analyzes the publisher's optimal content strategy. Section 6 offers extensions of our analysis. Conclusions and directions for future research
are provided in Section 7. To facilitate exposition, we have relegated
proofs to the Appendix.
2. General framework
We first introduce the three main components of our modeling
framework: the publisher, the content market, and the advertising
market. Next, we define the strategies available to the publisher and
characterize the optimal advertising-sales revenue ratio and thus the optimal content strategy. Table 1 summarizes the components of the general framework (as well as its main assumptions) and indicates where
the reduced-form expressions are replaced by specific functional forms.
2.1. Publisher
We consider a publisher who offers a digital information good with
content of size N N 0 through an online channel. Content size may be
thought of as the number of chapters of a book or movie, the number
Section 3.1
Section 6.3
a′(n) b 0
a′p(N − n) b 0
Section 4
Section 6.1
Section 6.2
of songs on an album, or the number of articles on a news platform.
To mirror the cost properties of information goods, we assume that
the publisher has fixed costs F ≥ 0 and zero unit costs to produce the
content.3 The cost to provide digital access per subscriber is cs ≥ 0 and
the costs of providing free samples are normalized to zero. The qualities
!
"
of the content parts are distributed on the quality spectrum 0; V ,
where V is the publisher's private information. We treat quality V as
an outcome of a previous strategic decision and suppose that the publisher has two decision variables: the sample size n ∈ [0, N] and the
price p at which to sell the good.4 Notice that in the context of the
metered model, the sample size n is the “meter” and p is the price for
the content behind the “paywall,” which separates free content from
paid content.
2.2. Content market
We consider a market with a unit measure of consumers that
observe the publisher's sampling and pricing decisions. Consumers are
uncertain about content quality. We assume that they update their
prior expectations in a Bayesian fashion through inspection of the free
e ðnÞ the consumers' expected posterior quality
samples and denote by V
given the sample size n. Expected content demand depends on price p,
e ðnÞ. Specifically, we assume that the publisher's exsample size n, and V
pected content demand is given by
E
e ðnÞÞ:
D ðp; nÞ ≡ Dðp; n; V
ð1Þ
This representation emphasizes that the sample size has both a direct
effect on content demand and an indirect effect that operates through
e ðnÞ.
the impact of n on expected posterior quality V
3
Throughout the analysis, we assume that the fixed cost do not exceed the product
market profit. Hence they do not change the analysis and can therefore be omitted.
4
The choice of (p, n) is not a multidimensional signal for quality as studied, for instance,
by Wilson (1985) and Milgrom and Roberts (1986). In this strand of the literature, n is an
advertising signal for quality. However, in our setting, the publisher's choice of n allows
the consumers to gain information about the actual content quality through their sample
experience before making the purchase decision.
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
Expected content demand satisfies the following basic assumptions.
b 0, i.e. content demand depends negatively on
First, we assume that ∂D
∂p
b 0, so that a larger sample size has a
price. Second, we impose that ∂D
∂n
direct negative effect on demand for the remaining content accessible
∂D
N 0, i.e. content demand
through the paywall. Third, we require that ∂Ṽ
depends positively on expected posterior quality. The overall effect of
the sample size n on expected content demand is given by
E
∂D
∂D ∂D e ′
V ðnÞ;
¼
þ
e
∂n
∂n ∂V
∂D e ′
where term ∂Ṽ
V ðnÞ captures the indirect effect of the sample size on expected content demand. It is not clear a priori how the sample size afE
fects posterior expectations and hence ∂D
. If Ve ′ðnÞ b 0 , sampling
∂n
reduces posterior expectations and is thus demand-reducing. Even if
e ′ ðnÞ N 0 , that is, if sampling increases posterior expectations, offering
V
an additional sample may be demand-reducing if the direct effect dome ′ ðnÞ is sufficiently large, the ininates the indirect effect. However, once V
∂DE
∂n
N 0 and
direct effect is stronger than the direct effect so that
sampling has a demand-enhancing effect. In line with Bawa and
Shoemaker (2004), we refer to the direct effect of sampling on content
demand as the “cannibalization effect” and to the indirect effect as the
“expansion effect.”
2.3. Advertising market
We consider a market where the advertisements are delivered to the
publisher through a representative advertiser (e.g., an advertising agency). We assume that the publisher includes one advertisement in each
free content part. The inverse advertising demand is denoted by a(n)
and maps the publisher's choice of n into the market price for ads.
Thus, a(n) can be thought of as the advertiser's willingness to pay for
placing n advertisements. We make the natural assumption that the
price for advertisements decreases in sample size, that is, a′(n) b 0.
2.4. Optimal strategy
The publisher receives profits from the content market and from
the advertising market. The expected profit from content sales is ðp−cs Þ%
e ðnÞÞ, while the profit (revenue) from including advertisements in
Dðp; n; V
the free articles is a(n)n. The publisher makes pricing and sampling decisions so as to maximize its (expected) profit from the two markets:
max
p;n
e ðnÞÞ þ aðnÞn
πðp; nÞ ¼ ðp−cs ÞDðp; n; V
s:t: p ≥ 0
0 ≤ n ≤ N:
ð2Þ
As long as the publisher's profit function π(p, n) is concave, standard optimization theory posits that there is a unique constraint global maximizer
(p*, n*) because the constraint set is convex. Depending on the optimal
pricing and sampling decision, the following definition gives the strategies
available to the publisher.
Definition 1. Strategies
Given the optimal pricing and sampling decision (p*, n*), the publisher
adopts either (i) a “sampling strategy” if p* N 0 and n* ∈ (0, N), (ii) a pure
“paid content strategy” if p* N 0 and n* = 0, or (iii) a pure “free content
strategy” if p* = 0 and n* = N.
Notice that both the paid content strategy and the free content
strategy are nested within the sampling strategy: The publisher receives no advertising revenue under a paid content strategy and no
sales revenue under a free content strategy. The following result
195
describes the optimal strategy as the ratio of advertising revenue to
sales revenue.
Proposition 1. Advertising-sales revenue ratio
Under a sampling strategy, the publisher's optimal ratio of advertising
revenue to sales revenue is given by
ηn − ηṼ εṼ
an%
$ ;
¼#
1
Dp%
1−
η
ηa p
ð3Þ
where ηp ≡ −(∂D/∂p)(p/D) denotes the elasticity of content demand with
respect to price, ηn ≡ −(∂D/∂n)(n/D) denotes the elasticity of content dee V=DÞ
e
mand with respect to sample size, ηṼ ≡ ð∂D=∂VÞð
denotes the elastice ′ ðnÞðn=VÞ
e denotes
ity of content demand with respect to quality, εṼ ≡ V
the elasticity of posterior quality expectations with respect to sample
size, and ηa ≡ − n′(a)(a/n) denotes the price elasticity of advertising
demand.
This result has two important managerial insights: First, it shows that
the publisher's advertising-sales revenue ratio and hence its optimal
content strategy are determined by characteristics of both the content
market and the advertising market. Specifically, consumer preferences
determine the characteristics of the content market, captured by the
elasticities of content demand with respect to price, sample size, and
quality. The price elasticity of advertising demand reflects advertiser
preferences. This general result thus provides guidance for managers
seeking to better understand the contributions of (expected) sales and
advertising to total revenue.
Second, Proposition 1 shows how changes in the “market environment,” captured by the various elasticities, affect the publisher's
composition of revenues. Unsurprisingly, if the price elasticity ηp increases, the advertising-sales revenue ratio is lower. Intuitively, for a
given sample size, the optimal price for the content is lower, which
results in a higher sales revenue. In contrast, higher elasticity of content demand with respect to the sample size ηn increases the
advertising-sales revenue ratio. Furthermore, the higher the price
elasticity of advertising demand ηa, the lower is the advertisingsales revenue ratio.
Proposition 1 also highlights the crucial role which the elasticity of
posterior quality expectations with respect to sample size plays. Because the elasticity of content with respect to quality ηṼ is positive,
the impact of sampling on posterior quality determines the sign of
ηṼ εṼ . If εṼ is negative, the ratio of advertising revenue to sales revenue
tends to be high, while it tends to be low if εṼ is positive. Intuitively,
if ε Ṽ b 0, sampling reduces expected content demand as Ṽ ′ ðnÞ b 0, and
hence the advertising-sales revenue ratio is high. In contrast, if εṼ N0,
sampling increases expected content demand as consumers revise their
expectations about quality upwards, resulting in a lower advertisingsales revenue ratio.
Interestingly, the optimal advertising-sales revenue ratio is reminiscent of the well-known Dorfman-Steiner condition, which states that a
monopolist's ratio of advertising spending to sales revenue is equal to
the ratio of the elasticities of demand with respect to advertising and
price (Dorfman & Steiner, 1954). Proposition 1 reduces to this result in
the special case when offering additional samples does not affect posterior quality ðε Ṽ ¼ 0Þ and if the advertising demand is perfectly elastic
(ηa → ∞).
Our general framework is agnostic about how consumers form posterior quality expectations. To shed light on effects of sampling on posterior quality expectations, the next section introduces a Bayesian
learning mechanism in which consumers update their prior expectations about content quality through experience with the sample. Importantly, we show how the publisher should take the consumers' learning
into account to gauge the effects of offering free samples on content
demand.
196
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
(B)
(A)
Fig. 1. Prior expectations about V (where V ≡ 1).
3. Content market
This section analyzes the impact of customer selected sampling on
expected content demand when the qualities of the content parts are
!
"
uniformly distributed on the quality spectrum 0; V . We begin by laying
out assumptions regarding consumer behavior and describing the
Bayesian learning mechanism. Taking consumers' learning into account,
we derive the publisher's expected content demand given its pricing
and sampling decisions. Finally, we provide conditions under which
sampling increases or decreases expected content demand.
3.1. Consumer behavior
Consumers know that the
of the free samples are uniformly
! qualities
"
distributed on the interval 0; V , but they do not know the upper bound
of the publisher's quality spectrum V and are hence uncertain about (average) content quality.5 Consumers have a common prior belief about V
that may stem, for instance, from reviews, ratings or “word of mouth.”
The natural conjugate family for a random sample from a uniform distribution with unknown upper bound is the Pareto distribution (DeGroot,
1970). We capture uncertainty about V by a prior belief that consists of a
minimum estimate v0 of the upper bound V and a level of uncertainty α
about this value. Specifically, we assume that prior beliefs follow a Pareto distribution with density function
f ðvjv0 ; α Þ ¼
8 α
< αv0
αþ1
:v
0;
;
!
" ðα þ nÞe
v0 ðnÞ
:
v0 ðnÞ; α ¼
E V je
α þ n−1
Consumers infer the expected quality of the information good
E½Vjv1 ; …; vn ' from the average quality of the sampled content parts.
Knowing that qualities are uniformly distributed on the quality spectrum offered, the expected quality of the information good is given by
E½Vjv1 ; …; vn ' ¼
for v N v0
otherwise:
ð4Þ
Obviously, prior expectations increase in v0 and decrease in α. Fig. 1
illustrates two prior beliefs along with the corresponding expectations
for different parameter values. Prior expectations are lower than actual
quality in Panel A and higher than actual quality in Panel B. Note that
prior expectations can be higher than actual quality even if v0 b V.
Consumers update their prior belief about the upper bound of the
quality spectrum V by taking the observed qualities of the free samples
into account. Specifically, consumers use the n sample qualities Vi = vi
Note that the upper bound V is monotonically related to the mean, which may be an
alternative way for consumers to think about content quality.
6
Our measure of uncertainty corresponds to the scale parameter α of the Pareto distribution. Hence, when uncertainty is higher, the prior distribution is more spread out.
ð5Þ
%
θNE½Vjv1 ; …; vn ' þ ξn−p;
θnE½Vjv1 ; …; vn 'þξn;
from purchasing at price p
from staying with the f ree samples:
The parameter ξ captures the intensity of ad-attraction (ξ N 0) or adrepulsion (ξ b 0) in the population (Gabszewicz, Laussel, & Sonnac,
2005). Thus, when consumers exhibit ad-loving behavior, the utility
of both options is augmented by ξn, while the utility of both options
is reduced by ξn in the case of ad-avoiding behavior.
The value of the information good is equal to the number of content
parts multiplied by their expected quality.8 This implies that a consumer
will purchase the information good if and only if the indirect utility from
buying exceeds the indirect utility from consuming only the free samples, that is, if
θðN−nÞE½Vjv1 ; …; vn '−p ≥ 0:
ð6Þ
This condition means that the (quality-weighted) expected value of the
content that has not been sampled must exceed the price. Importantly,
purchase does not depend on consumer behavior towards advertising.
7
5
!
"
E V je
v0 ðnÞ; α
:
2
Consumers believe that higher quality is better than lower quality
but differ in the way they value quality. To capture this heterogeneity, we introduce a preference parameter for quality θ, which is uniformly distributed on the interval [0,1]. We assume that each
consumer either purchases the information good at price p or stays
with the n free samples. A consumer's indirect utility from these
two options is given by
uðp; nÞ ¼
We assume α N 1 to ensure existence of prior expectations.6 Based
on the consumers' prior knowledge about v0 and α, their prior expectation about V is
!
"
αv0
E Vjv0 ; α ¼
:
α−1
(i = 1, …, n) to form their posterior belief e
vðnÞ about V. Using standard
Bayesian analysis, e
vðnÞ follows a Pareto distribution with minimum
value parameter e
v0 ðnÞ ¼ maxfv0 ; v1 ; …; vn g and shape parameter
α + n (DeGroot, 1970).7 Hence, the posterior expectation of V is
given by
The proof of this result appears in the Appendix.
This additivity assumption is justified for independently valued content parts. However,
a concave or convex relationship between the value and the number of content parts might
be more appropriate for interrelated content parts, that is, if the content parts are substitutes
or complements.
8
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
v0 b V, sampling may have a demand-enhancing effect. We next address
this possibility.
3.2. Expected content demand
Because consumers do not know the upper bound of the quality
spectrum V with certainty, content demand is influenced by consumers' posterior quality expectations. Thus, when the publisher
makes decisions about the sample size and the price, it has to base
them on expected content demand as consumers have not yet evaluated sample qualities and updated their expectations about content
quality.
In order to compute expected content demand, we assume that the
publisher knows the consumers' prior parameters v0 and α, which can
be learned using standard market research techniques such as surveys.
The calculation involves a three-step procedure: In the first step, the
publisher computes the expected posterior quality by averaging posterior expectations about V as given by (5) across all possible realizations
of sample qualities:
E½E½VjV 1 ; …; V n '' ¼
In the second step, the publisher substitutes the expected posterior
quality into the purchase condition given by (6) to obtain expected content demand:
&
p
2ðαþn−1Þ
:
ðN−nÞ ðα þ nÞE½e
v0 ðnÞ'
ð7Þ
In the third step, the publisher calculates E½ e
v0 ðnÞ' to obtain the expected
content demand as a function of the underlying model parameters. This
calculation leads to the following result.
Proposition 2. Expected content demand
When the publisher sells the information good at price p and offers
n ∈ {1, …, N − 1} samples, then
(a) if v0 b V, expected content demand is given by
(
)
n
p 2ðα þ n−1Þðn þ 1ÞV
:
Dfv0 b V g ðp; nÞ ¼ max 0; 1−
ðN−nÞ ðαþnÞðvnþ1 þ nV nþ1 Þ
E
0
ð8Þ
(b) if v0 ≥ V, expected content demand is given by
E
Dfv0
≥
%
ð
p;
n
Þ
¼
max
0; 1−
Vg
3.3. The role of quality expectations
For a given level of prior expectations about content quality, sampling either increases or decreases expected content demand. Whether
sampling compensates for cannibalization through consumers' learning
depends on the gap between posterior quality and actual quality.
Expected posterior quality is
8
nþ1
>
ðα þ nÞðvnþ1
þ nV Þ
>
0
>
<
n ; if v0 b V
e ðnÞ ¼ 2ðα þ n−1Þðn þ 1ÞV
V
>
> ðαþnÞv0
>
;
if v0 ≥ V
:
2ðα þ n−1Þ
ð10Þ
e ðnÞ− V . Consumers overestimate (unand the quality gap is defined as V
2
derestimate) quality if the expected posterior quality is higher (lower)
than actual quality. This leads to the following result.
ðα þ nÞE½ e
v0 ðnÞ'
:
2ðα þ n−1Þ
%
E
D ðp; nÞ ¼ max 0; 1−
197
&
p 2ðα þ n−1Þ
:
ðN−nÞ ðα þ nÞv0
Lemma 1. Quality gap
When v0 b V , consumers overestimate quality after their sample
experience if
v0
V
N
#
$1
α−1 nþ1
;
αþn
ð11Þ
and underestimate it if the inequality is reversed. If v0 ≥ V , consumers
overestimate quality irrespective of the sample size and their level of uncertainty α N 1.
Prior expectations can be higher than actual quality even if v0 < V (a
high level of uncertainty about v0 is captured by a low α). The condition
in Eq. (11) applies when consumers overestimate quality based on posterior expectations: This is likely to be the case for a low α and when the
publisher offers a small number of free articles n. On the other hand,
consumers underestimate quality if their uncertainty is low and the
sample size is large.
When v0 b V, the shaded area in Fig. 2 illustrates the set of parameters for which consumers overestimate and underestimate quality,
respectively. By construction, where α N 1, condition (11) holds and
consumers overestimate quality. The parameter region for which cone ðnÞ → V as
sumers overestimate quality shrinks as n gets larger since V
2
n → ∞, meaning that consumers learn actual quality once the sample
size gets “large enough.”
ð9Þ
Expected content demand has the intuitive properties that we assumed in our general framework: it decreases in price and increases
in expected posterior quality. In addition, sampling has both a direct
demand-reducing effect and an indirect effect that operates through
its impact on posterior expectations. The direct effect kicks in
1
through the factor N−n
and mirrors the cannibalization effect ∂D
b 0.
∂n
This follows because a larger sample size reduces the utility of the
remaining content consumers have to pay for. Proposition 2 nests the
expected content demand under a paid content strategy (n = 0), in
which case D(p, 0) is a function of prior quality expectations as there
is no learning. In contrast, when the publisher employs a free content
strategy (n = N), consumers do not purchase the information good
as they can download it for free and hence D(p, N) ≡ 0.
Which of the two demand functions reported in Proposition 2
emerges depends on the value of the minimum estimate v0 about V
vis-à-vis the actual upper bound of the quality spectrum V : If v0 ≥ V ,
sampling necessarily reduces expected content demand. In contrast, if
Fig. 2. The quality gap for the case v0 b V (where V ≡ 10). The shaded area indicates where
consumers underestimate quality.
198
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
e ðnÞ allows us to rewrite the expected content deThe definition of V
mand derived in Lemma 2 more compactly as
(
E
D ðp; nÞ ¼ max 0; 1−
p
)
e ðnÞ
ðN−nÞV
;
ð12Þ
which is a specific version of the reduced-form demand function in
Eq. (1). Hence the number of free samples n has both a direct effect on
expected content demand and an indirect effect that operates through
e ðnÞ. The next result uses this demand
posterior quality expectations V
function to identify conditions under which sampling has a demandE
enhancing effect (that is, ∂D
N 0).
∂n
Lemma 2. Effects of sampling
n
, that is,
Offering free samples has a demand-enhancing effect if εṼ N N−n
if the elasticity of consumers' posterior quality expectations exceeds the
ratio of sampled to paid content.
Lemma 2 shows that offering free samples may increase expected
content demand through consumers' learning, even though it produces
a cannibalization effect. Intuitively, the indirect effect dominates the
direct cannibalization effect if sampling results in a sufficiently large upwards revision of consumers' prior expectations.
4. Advertising market
This section derives inverse advertising demand of a representative
advertiser that places advertisements in the free samples offered by
the publisher. The representative advertiser can be thought of as an advertising agency that delivers independent product advertisements to
the publisher. Following Godes, Ofek, and Sarvary (2009), we let the
advertiser's utility from placing n advertisements be given by
n
−an;
2N
where ϕ N 0 is a parameter capturing the marginal benefit of an advertisement and a denotes the unit price set by the publisher to run an advertisement. This specification captures decreasing marginal utility
from placing advertisements caused by decreasing consumer recall
and retention rates for ads (Burke & Srull, 1988).
The advertiser maximizes its utility to determine how many advertisements to run on the publisher's platform. The implied inverse advertising demand has the linear form
aðnÞ ¼ ϕ−
n
;
N
5.1. Strategy with known quality
We first derive content demand under a sampling strategy and subsequently derive the demands for the two boundary strategies when
consumers know content quality. Next, we characterize the optimal
pricing and sampling decisions for each of the three strategies. Finally,
we determine the optimal content strategy.
5.1.1. Content demand
When consumers know the upper bound V and hence the quality
spectrum, the minimum estimate v0 of the upper bound V is equal to V
with certainty and thus α → ∞. Proposition 2 implies that content demand can be expressed as
9
8
>
>
>
>
=
<
p
ð14Þ
Dðp; nÞ ¼ max 0; 1−
>
V>
>
>
:
ðN−nÞ ;
2
for n ∈ {0, …, N − 1}. Recall that D(p, N) ≡ 0 under a free content
strategy.
5.1.2. Optimal pricing and sampling
In the benchmark case with known quality, the publisher makes its
pricing and sampling decisions so as to
#
$ '
p
n(
max πðp; nÞ ¼ p 1−
þ ϕ− n
V
p;n
N
ðN−nÞ 2
s:t: p ≥ 0
0 ≤ n ≤ N;
where content demand is given by Eq. (14) and inverse advertising demand by Eq. (13). From the first-order conditions, the optimal price for
a given sample size is
2
uðnÞ ¼ ϕn−
ϕ N 1.9 For expositional purposes, we normalize the costs of providing
digital access cs to zero and present analysis ignoring the integer constraint on the number of free samples.
ð13Þ
where ϕ is referred to as “advertising effectiveness.” Inverse demand
slopes downward, implying that the publisher's revenue per ad impression is decreasing in sample size. Intuitively, a(n) can be thought of as
the advertiser's willingness to pay for placing n advertisements.
5. Optimal content strategy
This section analyzes the publisher's optimal content strategy with
customer selected sampling. We first analyze the benchmark case
in which the consumers know V and hence the publisher's quality
spectrum. In this case, sampling does not affect the consumers' expectations about quality and simply serves to generating advertising
revenues. Next, we analyze the case where V is not known to consumers. In contrast to the benchmark case, sampling not only
generates advertising revenues but also influences consumers' expectations about quality.
We consider three strategies: the paid content strategy, the sampling strategy, and the free content strategy. In order to study the
tradeoffs between the three strategies, we focus on the case where
pðnÞ ¼
ðN−nÞV
:
4
ð15Þ
This implies that the more free samples the publisher chooses to offer,
the less it will be able to charge the consumer for the remaining content.
The next result summarizes the optimal pricing and sampling decisions
for each of the three strategies.
Lemma 3. Pricing and sampling
Suppose that consumers know the upper bound of content quality
)
*
V . Then, (i)
under* a sampling strategy, p% ¼ NV 8ð2−ϕÞ þ V =64
)
and n% ¼ N 8ϕ−V =16, (ii) under a paid content strategy, p∗ ¼ NV=4
and n∗ = 0, and (iii) under free content strategy, p∗ = 0 and n∗ = N.
The parameters V and ϕ have opposite effects on the optimal price
and on the optimal sample size under a sampling strategy: As expected,
p∗ increases in V while n∗ decreases in quality. In contrast, p∗ decreases in
ϕ, and n∗ increases in advertising effectiveness. Both the optimal price
and the optimal sample size increase in content size N. The following
proposition characterizes the publisher's optimal strategy.
Proposition 3. Optimal strategy
If consumers know the quality spectrum, then: (i) if ϕ ≤ V8, the publisher
'
(
should follow a paid content strategy, (ii) if ϕ ∈ V8 ; V8 þ 2 , the publisher
should employ a sampling strategy, and (iii) if ϕ ≥
optimal strategy is a free content strategy.
V
8
þ 2, the publisher's
9
When ϕ ≤ 1, the free content strategy is never optimal because the advertising revenues and hence the publisher's profit is zero. Thus, the publisher's choice is only between
the paid content strategy and the sampling strategy. The analysis where ϕ N 1 encompasses this special case.
199
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
Proposition 3 shows that the choice of the optimal strategy is driven
by the relationship between content quality V and advertising effectiveness ϕ. Thus, for a given content quality, a paid content strategy is
optimal if the effectiveness of advertising is sufficiently low. For intermediate levels of advertising effectiveness, a sampling strategy that
generates revenues from both sales and advertising on the free samples
is optimal. If advertising is sufficiently effective, the publisher should
switch to a free content strategy.
The effects of ϕ and V on the optimal strategy can also be understood
by inspection of the advertising-sales revenue ratio. The ratio follows
from (3) and is
an%
¼
Dp%
V
ϕ−
8
V
4
!
!
V
þϕ
8
! :
V
þ 2−ϕ
8
The ratio of advertising revenue to sales revenue tends to zero as ϕ
V
8
approaches the lower bound , implying that the publisher should employ a paid content strategy. A sampling strategy is optimal only if advertising is not “too effective,” that is, as long as ϕ ≤ V8 þ 2 . Once ϕ
exceeds this level, the publisher should switch to a free content
strategy.
5.1.3. Summary
When content quality is known to consumers, the publisher's optimal strategy is determined by the relation between advertising effectiveness and content quality. The more effective advertising is, the
more free samples the publisher should offer—even if it solely cannibalizes content demand.
5.2. Strategy with unknown quality
We first determine the optimal content strategy when consumers
do not know content quality and compare our findings to the results
from the benchmark case. Next, we study how the interplay of prior expectations and advertising effectiveness governs the optimal choice of
content strategy.
5.2.1. Optimal pricing and sampling
When content quality is not known to consumers, the publisher
makes its pricing and sampling decisions so as to
max
p;n
p
E
π ðp; nÞ ¼ p 1−
s:t:
!
e ðnÞ
ðN−nÞV
p≥0
0 ≤ n ≤ N;
'
n(
þ ϕ− n
N
In contrast to our benchmark case, it is not possible to characterize the publisher's optimal pricing and sampling decisions (and
hence profits) analytically. Nevertheless, we have the following
result.
Proposition 4. Optimal strategy
Suppose that consumers are uncertain about the upper bound of
content quality V and that the profit function πE(n) is strictly concave.
e ð0Þ−NV
e ′ ð0ÞÞ and ϕ ¼ 2 þ Ṽ ðNÞ
Then, there are cut-off values ϕ ¼ 1 ðV
4
4
such that a paid content strategy is optimal for ϕ ≤ ϕ , a sampling
strategy is optimal for ϕ ∈ ðϕ; ϕÞ, and a free content strategy is optimal
for ϕ ≥ ϕ.
This result is consistent with the insights from the benchmark case
when quality is known: a paid content strategy is optimal only if the advertising effectiveness is sufficiently low, a sampling strategy is optimal
for intermediate levels of the advertising effectiveness, and the publisher should switch to a free content strategy once advertising is sufficiently effective. Fig. 3 illustrates the optimal strategy for varying advertising
effectiveness ϕ and the expected profit for each strategy (π∗SC for the
sampling strategy, π∗PC for the paid content strategy, and π∗FC for the
free content strategy).
Hence prior expectations determine the lower of the two cut-off
values for a sampling strategy to be optimal whereas posterior expectations for sample size n = N determine the upper cut-off value. In effect,
ϕ is determined by the impact of the “first” free content part on posterior expectations, while ϕ is determined by posterior expectations
after inspection of the “last” free content part. The next lemma shows
that the model where quality V is not known to consumers nests the
full information benchmark case (see Proposition 3).
Lemma 4. Cut-off values
Suppose that consumers are uncertain about content quality V and that
the profit function πE(n) is strictly concave. Then, when consumers have
correct quality expectations, that is, if v0 ¼ V and α → ∞, the lower
bound ϕ converges to V and the upper bound ϕ converges to V8 þ 2.
8
5.2.2. The impact of prior expectations
Proposition 4 shows that the optimal strategy depends not only on
advertising effectiveness ϕ and quality V as in the benchmark case,
but also on the specific values of the prior parameters v0 and α (as
well as content size N). Fig. 4 illustrates the optimal strategy for
given advertising effectiveness and prior expectations. Panel A
depicts the cut-off thresholds between the different strategies in
the ðϕ; v0 Þ-space (given α = 2). Similarly, Panel B illustrates the
publisher's optimal strategy in the (ϕ, α)-space (given v0 ¼ 5 ).
where expected content demand is given by Eq. (12) and inverse advertising demand by Eq. (13). The only difference between this expected
profit and the profit when content quality is known to consumers is
the dependence on expected quality rather than actual (average) quale ðnÞ is the posterior estimate of avity. Based on (15) and recalling that V
V
erage quality 2 , we thus obtain that
pðnÞ ¼
e ðnÞ
ðN−nÞV
:
2
Substituting p(n) back into the profit function allows us to rewrite the
profit maximization problem as
max
n
e ðnÞ '
V
n(
þ ϕ− n
4
N
s:t: 0 ≤ n ≤ N:
E
π ðnÞ ¼ ðN−nÞ
ð16Þ
Fig. 3. Optimal strategy with unknown quality (for v0 ¼ 5, α = 2, V ¼ 10, and N = 10).
200
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
(A)
(B)
Fig. 4. Optimal strategy (for V ¼ 10 and N = 10).
Here prior expectations are correct and coincide with actual quality
when v0 ¼ 5 and α = 2. The following observation summarizes our
insights.
Observation 1. Prior expectations
If advertising effectiveness is high and actual content quality is relatively
low, it can be optimal for the publisher to adopt a sampling strategy to generate advertising revenue even when it reduces prior quality expectations
and content demand. In contrast, if advertising effectiveness is low and actual content quality is relatively high, it can be optimal to adopt a paid content strategy and not reveal high quality, even though sampling would
increase content demand.
The shaded areas in Fig. 4 illustrate these important managerial insights. Notice that the larger areas correspond to the case where prior
expectations are high relative to actual content quality. In such market
environments, the publisher should sacrifice content demand to boost
advertising revenues. The figure also shows that when prior expectations are sufficiently low relative to actual content quality (that is, if v0
is small or α is large), the choice is between a sampling strategy and a
free content strategy only. Intuitively, the publisher has an incentive to
reveal its higher than expected quality through free samples, possibly
offering its content for free. In contrast, when prior expectations are sufficiently high, the optimal strategy is either a paid content strategy or a
free content strategy. In such market environment, the publisher has no
incentive to reveal its lower than expected quality when ϕ is low.
5.2.3. Summary
When actual content quality is not known to consumers, the optimal
strategy is determined by the relation between advertising effectiveness,
prior quality expectations, and posterior quality expectations. As in the
benchmark case, employing a paid content strategy is optimal only if
advertising effectiveness is sufficiently low compared to prior quality
expectations. For intermediate levels of advertising effectiveness, the
publisher should use a sampling strategy. The publisher should switch
to a free content strategy once advertising is sufficiently effective compared to posterior quality expectations. Counter to intuition, it can be
optimal for the publisher to generate advertising revenue by adopting
a sampling strategy even when sampling reduces both prior quality expectations and content demand. In addition, it can be optimal for the
publisher to adopt a paid content strategy and to refrain from revealing
high quality through free samples.
6. Extensions
This section relaxes three key assumptions and studies the effects on
our findings. First, we allow the publisher to generate advertising revenues from paid content as well. Second, we allow advertising
effectiveness to depend on content quality. Third, we introduce competition into the model.
6.1. Including advertisements in the paid content
In this section, we extend the model by allowing it to include advertisements in both the free articles and the paid content. To this end, we
assume that the market price for advertisements included in the paid
content is given by
^ Þ ¼ ϕp −
ap ðn
^
n
;
N
ð17Þ
^ ≡ N−n and ϕp N 1 denotes the advertising effectiveness for
where n
ads in the paid content. This inverse demand is a natural counterpart
to the advertising demand a(n) given by Eq. (13) and indicates
^ that have
that the ad price depends on the number of articles n
not been offered as free samples. Differences in the levels of
advertising effectiveness ϕp and ϕ capture differences in reach or
the degree of targeting in the advertising markets for paid and free
content.
When allowing for advertisements in the paid content, a consumer's
indirect utility from the two options is
^ ðp; nÞ ¼
u
%
θNE½Vjv1 ; …; vn ' þ ξN−p;
θnE½Vjv1 ; …; vn ' þ ξn;
from purchasing at price p
from staying with the free samples:
Importantly, the utility of the purchase option now depends on the
overall number of advertisements shown to the consumer rather than
the number of ads contained in the free samples only. This augmented
specification implies that a consumer will purchase the information
good if and only if
θðN−nÞE½Vjv1 ; …; vn ' ≥ p−ξðN−nÞ;
ð18Þ
that is, if the expected value of the content that has not been sampled
exceeds its full price p−ξ(N − n). Intuitively, the full price of the content is lower if consumers are ad-lovers (ξ N 0) and higher if they are
ad-avoiders (in which case exposure to the ads in the paid content results in a nuisance cost ξ(N − n) that has to be added to the price p).
From the purchase condition in Eq. (18), expected content demand
can be derived as
%
^ E ðp; nÞ ¼ max 0; 1−
D
' p
(&
1
−ξ :
E½Vjv1 ; …; vn ' N−n
ð19Þ
Compared to the case where consumers are ad-neutral, content
demand is higher when consumers are ad-lovers and lower when consumers are ad-avoiders.
201
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
The publisher makes its pricing and sampling decisions so as to
'
(
'
(
E
^ E ðp; nÞþ ϕ− n n
^Þ D
max π ðp; nÞ ¼ p þ Rp ðn
p;n
N
s:t: p ≥ 0
0 ≤ n ≤ N;
where expected content demand is given by Eq. (19) and inverse adver^ Þ ≡ ap ðn
^ Þn
^ are the additional
tising demand by Eq. (17). Notice that Rp ðn
revenues from including ads in the paid content. By definition, Rp = 0
under a free content strategy, while Rp = (ϕp − 1)N under a paid content strategy. We summarize our insights as follows.
Observation 2. Ads included in paid content
When the publisher includes advertisements in the paid content and
consumers are neutral about advertisements or ad-lovers, the range of advertising effectiveness ϕ for which the sampling strategy is best expands. In
contrast, when consumers are ad-avoiders, the range of ϕ for which the
sampling strategy is optimal can shrink.
Fig. 5 illustrates the profit effects of including advertisements in the
paid content when consumers are neutral about advertisements. Intuitively, exploiting revenues from advertisements in the paid content increases the unit margin from selling content, which translates into a
higher profit under a sampling strategy. These profit effects are more
pronounced when advertising effectiveness for ads in paid content ϕp
increases. Further, the profit under a sampling strategy is higher when
the consumers are ad-lovers (for given ϕp) and lower when they are
ad-avoiders.
6.2. Endogenizing advertising effectiveness
Up to now, we considered advertising effectiveness to be exogenous
and independent of content quality. In this section, we relax this assumption and allow the advertiser's willingness to pay for advertisements to be positively related to expected posterior quality, in effect
having product quality have a halo effect on ad effectiveness. Specifically, we let inverse advertising demand be given by
e ðnÞ−
aðnÞ ¼ ϕ
n
N
ð20Þ
e ðnÞ on qualityand capture the effect of expected posterior quality V
e ðnÞ in the following way:
adjusted advertising effectiveness ϕ
e
e ðnÞ ≡ ϕ V ðnÞ ;
ϕ
e ð0Þ
V
e ð0Þ
where ϕ is advertising effectiveness (as introduced in Section 4) and V
denotes prior quality expectations. This specification reflects the idea
that free samples become more attractive as an advertising platform
when consumers' posterior expectations exceed their prior expectations,
which in turn increases the advertiser's willingness to pay for the advertisements (and vice versa).
When advertising effectiveness is positively related to posterior content quality, offering free samples affects the willingness to pay for advertisements in two ways: Through changes in the quality-adjusted
e ðnÞ and through changes in the number of
advertising effectiveness ϕ
ads shown to consumers. Therefore, advertising demand can be increasing in the number of free samples when prior expectations are sufficiently low relative to consumers' updated expectations (see Panel A
in Fig. 6 for a graphical illustration). Over the range of where a(n) increases, the marginal benefit of higher advertising effectiveness outweighs the disutility of showing an additional ad to consumers.
The publisher makes its pricing and sampling decisions so as to
max
p;n
p
E
π ðp; nÞ ¼ p 1−
s:t: p ≥ 0
0 ≤ n ≤ N;
!
e ðnÞ
ðN−nÞV
'
(
e ðnÞ− n n
þ ϕ
N
where expected content demand is given by Eq. (12) and inverse advertising demand by Eq. (20). The next observation summarizes our
insights.
Observation 3. Endogenous ad effectiveness
When consumers' prior expectations are low and the advertiser's willingness to pay for advertisements is positively related to the expected posterior content quality, it can be optimal for the publisher to adopt a free
content strategy even when baseline advertising effectiveness ϕ is low.
Panel B of Fig. 6 illustrates this observation. In the figure, the lines
depict the cut-off thresholds between the different strategies in
the (ϕ, v0)-space. The solid lines correspond to the case with qualityadjusted advertising effectiveness and show that a free content strategy
can also be optimal for low ϕ, whereas a sampling strategy yields a
higher profit when advertising effectiveness is exogenous (indicated
by the dashed lines reproduced from Panel A in Fig. 4). This occurs because when prior expectations are low, sampling not only reveals high
quality, but also increases the advertiser's willingness to pay for ads
and thus to boost advertising revenues.
6.3. The impact of competition
Thus far, we have examined a publisher operating in a monopoly
setting. Here, we allow for competition between two publishers that
offer differentiated information goods. We add horizontal product
differentiation to capture the intensity of competition between the
two publishers.10 In the newspaper industry for instance, horizontal
product differentiation may arise due to different political opinions
among consumers (Gabszewicz, Laussel, & Sonnac, 2005).
The analysis of the competitive case indicates that, as in the monopoly case, advertising effectiveness is a key driver of the publishers' strategy choice. Specifically, if consumers' prior expectations about content
qualities are similar and advertising effectiveness is high, it is optimal
for both publishers to adopt a free content strategy in equilibrium. However, two other drivers that affect strategy choice in the competitive
case: the gaps in prior expectations about the quality of the competing
products and the degree of horizontal product differentiation. The
following observation summarizes our insights.
Observation 4. Competition
If the gap in prior expectations regarding the product qualities is large
and advertising effectiveness is high, both publishers should adopt a sampling strategy. As the degree of horizontal product differentiation increases
Fig. 5. Optimal strategy with (dashed lines) and without advertisements in paid content
(solid lines) for ξ = 0, v0 ¼ 5, α = 2, V ¼ 10, N = 10, and ϕp = 1.1.
10
All details of the model and the analysis of the competitive case are in the Appendix.
202
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
(A)
(B)
Fig. 6. Impact of endogenizing advertising effectiveness (for α = 2, V ¼ 10, N = 10; in addition, in Panel A v0 ¼ 2 and ϕ = 2).
and the products become less substitutable, the parameter region in which
the sampling strategy is optimal for both publishers shrinks.
Counter to intuition, our results show that it is not per se optimal to
choose a free content strategy when advertising effectiveness is high.
What matters in addition is the gap in prior expectations between the
two publishers: if it is small, then both firms should adopt a free content
strategy. Instead, if the gap in prior expectations is large, then both firms
should adopt a sampling strategy. Intuitively, the publisher that faces
lower prior expectations from consumers has a stronger incentive to reveal its higher than expected quality—and it is a best-response of the
rival publisher to also adopt a sampling strategy.
The second insight is that the parameter region in which the sampling strategy is optimal for both publishers shrinks when products are
perceived as less substitutable (e.g., when the readers with left-wing
preference consider a right-wing newspaper a less good alternative).
For the publishers, this results in less ability to acquire consumers who
purchase from the competitor. Consequently, the publishers try to generate as much revenue as possible from the advertising market by
employing a free content strategy. This result mirrors the findings from
the monopoly case: as the degree of horizontal product differentiation
increases and publishers tend to become monopolists in their respective
market segment, they should employ a free content strategy once advertising effectiveness exceeds a certain threshold level.
7. Conclusion
This paper analyzed digital content strategies when content sampling serves the dual purpose of disclosing content quality and generating advertising revenue. One of the key features of the model is that
consumers evaluate free samples of their choice within the limit set
by the publisher. Consumers then use the information gathered from
the free samples to update their prior expectations about content quality in a Bayesian fashion. Taking consumers' quality updating into account, the publisher can adopt a sampling strategy, a paid content
strategy, or a free content strategy.
We derived several important results. First, we show in our general
framework how the publisher's advertising-sales revenue ratio and
hence its optimal content strategy is determined by characteristics of
both the content market and the advertising market. We capture the
characteristics of the content market by the elasticity of consumers' updated quality expectations and the elasticities of content demand with
respect to price, sample size, and posterior quality. The corresponding
characteristic in the advertising market is the price elasticity of advertising demand. We show that, all else equal, the advertising-sales revenue
ratio is higher the lower the price elasticity of content demand, the
higher the elasticity of content demand with respect to sample size,
and the lower the price elasticity of advertising demand. In addition,
managers can expect the ratio of advertising revenue to sales revenue
to be low when sampling increases consumers' quality expectations
(resulting from the expansion effect) and high when sampling reduces
quality expectations (resulting from the cannibalization effect).
Second, when consumers make purchase decisions based on the
price of the content behind the paywall, content demand depends on
consumers' posterior quality expectations, which can be influenced by
the publisher through its sampling decision. Expected content demand
has the natural properties that it is decreasing in price and increasing in
expected posterior quality. Further, sampling has a demand-enhancing
effect through consumers' learning when prior expectations are sufficiently low (even though sampling produces a cannibalization effect).
We uncovered the rule of thumb that sampling increases content demand if the elasticity of consumers' updated expectations exceeds the
ratio of sampled to paid content.
Third, we characterize the publisher's optimal content strategy
when consumers are uncertain about actual content quality and learn
about it through inspection of free samples. We identify two cut-off
values that determine the publisher's optimal content strategy: a
lower bound that depends on prior quality expectations (separating
paid from sampling strategies) and an upper bound that depends on
posterior quality expectations (separating sampling from free content
strategies). From a managerial perspective, it can be optimal to reduce
both prior quality expectations and content demand in order to generate advertising revenue. In addition, it can be optimal for managers to
adopt a paid content strategy and to refrain from revealing high quality
through free samples.
We also explore several model extensions that are relevant for managerial decision making. First, when the publisher also includes advertisements in the paid content, the analysis shows that the cut-off values
between the content strategies depend not only on the relation between
advertising effectiveness and updated quality expectations, but also on
consumers' attitudes towards advertisements. Second, when the willingness to pay for advertisements is related to content quality, we show that
a free content strategy can be optimal even when advertising effectiveness is low. Third, we show that under competition the advertising effectiveness is a key driver of the publishers' equilibrium strategy choices. This
analysis also sheds light on recent developments in the newspaper industry and explains why publishers have moved away from pure advertisingfinanced business models to metered models (Abramson, 2010).
Our general framework offers several avenues for future research.
Regarding consumers, we assume they correctly update quality expectations based on their sample experience. One alternative is to assume
203
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
a consistent bias in the consumers' judgments. In addition, in circumstances where the firm selects the samples, consumers are likely to adjust (discount) observed quality, assuming that the publisher has
provided a non-representative set of samples to choose from in order
to persuade them to buy the paid content. Further, one could assume
that consumers do not evaluate the quality of all free samples because
of “sampling costs,” for example due to the opportunity cost of time or
mental costs. One could also enrich the model by allowing for internal
competition, where the publisher offers two websites to serve different
categories of consumers, which relates to the versioning literature.11
Thus, there are several further directions which research in these
areas could take. We view this paper a step in this process and hope
the paper encourages work in these and related directions.
Appendix A
A.1. Sampling from a uniform distribution
A.1.1. The Pareto distribution
A random variable X has a Pareto distribution with parameters w0
and α (w0 N 0) and (α N 0) if X has a density
f ðxjw0 ; α Þ ¼
8
< αwα0
αþ1
:x
0
for x N w0
otherwise:
For α N 1 the expectation of X exists and it is given by EðX Þ ¼
αw0
.
α−1
Re-
garding sampling from a uniform distribution, we use the following
result.
Theorem. (DeGroot, 1970)
12
Suppose that X1, …, Xn is a random sample from a uniform distribution of the interval (0, W), where the value of W is unknown. Suppose
also that the prior distribution of W is a Pareto distribution with parameters
w0 and α such that w0 N 0 and α N 0. Then the posterior distribution of W
when Xi = xi (i = 1, …, n) is a Pareto distribution with parameters w0′ and
α + n, where w0′ = max{w0, x1, …, xn}.
Proof. For w N w0, the prior density function ξ of W has the following
form:
ξðwÞ∝
1
:
wαþ1
Furthermore, ξ(w) = 0 for w ≤ w0. The likelihood function fn(x1, …, xn|w)
of Xi = xi (i = 1, …, n), when W = w (w N 0) is given by13:
f n ðx1 ; …; xn jwÞ ¼ f ðx1 jwÞ⋯ f ðxn jwÞ ¼
(
1
wn
0
for maxfx1 ; …; xn g b w
otherwise:
It follows from these relations that the posterior p.d.f. ξ(w|x1, …, xn) will
be positive only for values w such that w N w0 and w N max{x1, …, xn}.
Therefore, ξ(w|⋅) N 0 only if w N w0′. For w N w0′, it follows from Bayes'
theorem that
ξðwjx1 ; …; xn Þ∝f n ðx1 ; …; xn jwÞξðwÞ ¼
1
wαþnþ1
(the marginal joint probability density function fn(x1, …, xn) of X1, …, Xn is
a normalizing constant). □
11
For instance, The Boston Globe operates the ad-supported site boston.com and the
subscriber-only site BostonGlobe.com.
12
Theorem 1, p. 172.
13
Given W = w, the random variables X1, …, Xn are independent and identically distributed and the common probability density function of each of the random variables is f(xi|w).
A.2. Proofs
Proof of Proposition 1. By strict concavity of the profit function, the
solution to the problem in Eq. (2) must satisfy the necessary and sufficient first-order conditions
e ðnÞ þ ðp−c Þ
Dðp; n; V
s
ðp−cs Þ
e ðnÞÞ
∂Dðp; n; V
þ λ1 ¼ 0
∂p
!
e ðnÞÞ ∂Dðp; n; V
e ðnÞÞ ′
∂Dðp; n; V
e ðnÞ
þ
V
e
∂n
∂V
þ a′ðnÞn þ aðnÞ þ λ2 −λ3 ¼ 0
ðA:1Þ
ðA:2Þ
and the constraints λ1p = 0, λ2n = 0, and λ3(n − N) = 0, where the
λi's are non-negative real numbers (whose existence is assured by the
Kuhn–Tucker theorem). Suppressing the arguments of content demand,
(A.1) can be rewritten as
#
$
p−cs
1
λ
¼
1þ 1 :
ηp
p
D
ðA:3Þ
Dividing Eq. (A.2) through p and substituting from Eq. (A.3) produces
#
$#
$
1
λ
∂D ∂D e ′
a′ðnÞn aðnÞ λ2 −λ3
V ðnÞ þ
þ
þ
¼ 0:
1þ 1
þ
e
ηp
p
p
D
p
∂n ∂V
Recalling that n′ ðaÞ ¼
1
a′ ðnÞ
(from the inverse function theorem) and
using the definitions of the respective elasticities, the preceding equation can be rearranged to obtain
#
$
$
( #
pD 1
λ '
e ¼ 1− 1 þ λ2 −λ3 :
1 þ 1 ηn −η Ṽ V
n
an ηp
ηa
D
a
ðA:4Þ
Under a sampling strategy there is an interior solution and hence the
λk's are zero. Thus, Eq. (A.4) can be rewritten as
an ηn −ηṼ εṼ
( :
¼'
Dp
1− 1 η
ηa
□
p
Proof of Proposition 2. (a) In order to calculate E½e
v0 ðnÞ' when v0 b V ,
we first derive the distribution of e
v0 ðnÞ ¼ maxfv0 ; V 1 ; …; V n g. Before
doing so, we state a preliminary fact: The distribution function of
M = max{V1, …, Vn} is given by
F M ðt Þ ≡ PrfmaxfV 1 ; …; V n g ≤ t g
¼ PrffV 1 ≤ t g∩ … ∩fV n ≤ t gg
# $n
n
t
:
¼ ∏ Pr fV i ≤ t g ¼
V
i¼1
ðA:5Þ
As an immediate implication, the density function of M is given by
f M ðt Þ ¼
nt n−1
V
n
:
ðA:6Þ
v0 ðnÞ
Next, we derive the density function of e
v0 ðnÞ. By definition, e
cannot be smaller than v0 . Therefore, e
v0 ðnÞ ¼ v0 if and only if
maxfV 1 ; …; V n g ≤ v0 . The probability of this event follows from
Eq. (A.5) and it is given by
F M ðv0 Þ ¼
# $n
v0
:
V
For e
v0 ðnÞ N v0 , let e
F ð(Þ denote the truncated distribution function
of e
v0 ðnÞ. After removing the lower part of the distribution, we
204
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
!
"
e
e
have
! F ðt"Þ ¼ F M ðt Þ− F M ðv0 Þ for t ∈ v0 ; V . This implies f ðt Þ ¼ f M ðt Þ for
t∈ v0 ; V , and hence
n−1
ef ðt Þ ¼ nt
n ;
V
if v0 ≤ t ≤ V
by Eq. (A.6). The distribution of e
v0 ðnÞ has a mixed structure with
# $n
v0
Prfe
v0 ðnÞ ¼ v0 g ¼
V
ðA:7Þ
and density
n−1
ef ðt Þ ¼ nt
n ;
V
if v0 ≤ t ≤ V:
ðA:8Þ
The expectation of this mixed distribution is given by
E½ e
v0 ðnÞ' ¼ v0
¼
#
$n Z
v0
V
þ
nþ1
v0
nþ1
strategy is optimal if π∗PC ≥ π∗SC and π∗PC ≥ π∗FC, that is, if ϕ ≤ V8 . A free content strategy is optimal if π∗FC ≥ π∗SC and π∗FC ≥ π∗PC, that is, if ϕ ≥ V8 þ 2. □
Proof of Proposition 4. At an interior solution, the optimal sample size
n∗ satisfies the first-order condition
)
N−n
e ′ ðn% Þ
%* V
4
−
e ðn% Þ
V
2n%
þ ϕ−
¼ 0:
4
N
e ð0Þ
e ð0Þ V
NV
þϕ ≤ 0
−
4
4
⇔ ϕ ≤ ϕ:
At the other extreme, when n∗ = N, the Kuhn–Tucker conditions require
that
Proof of Lemma 1. If v0 b V, the quality gap can be expressed as
nþ1
e ðnÞ− V ¼ v0 ðα þ nÞ−V ðα−1Þ :
V
n
2
2ðα þ n−1Þðn þ 1ÞV
ðA:9Þ
Clearly, the sign of the quality gap depends only on the sign of numerator (A.9). The latter can easily be rearranged to obtain Eq. (11). If v0 ≥ V,
the quality gap can be written as
Proof of Lemma 2. Differentiating Eq. (12) with respect to n yields
'
(
e ′ ðnÞ−V
e ðnÞ p
ðN−nÞV
∂DE ðp; nÞ
:
¼
'
(
∂n
e ðnÞ 2
ðN−nÞV
′
e ðnÞ−V
e ðnÞ N 0, which
Clearly, sampling is demand-enhancing if ðN−nÞV
n
.
N−n
e ðN Þ
V
þ ϕ−2 ≥ 0
4
⇔ ϕ ≥ ϕ: □
e ðnÞ in Eq. (10), the lower
Proof of Lemma 4. Using the definition of V
bound can be expressed in terms of the underlying model parameters as
ϕ¼
ð2α ðα−1Þ þ NÞv0
:
16ðα−1Þ2
Settingv0 ¼ V and letting α → ∞ yields thatϕ → V8. Likewise, we have that
ϕ¼
ðα þ NÞV
þ 2:
8ðα þ N−1Þ
A.3. Analysis of the competitive case
which is strictly positive by our assumptions. □
N
−
Letting α → ∞, we obtain ϕ → V8 þ 2. □
)
*
V ðα þ nÞ v0 −V þ V
e
V ðnÞ− ¼
;
2
2ðα þ n−1Þ
can be rewritten as
2
from Lemma 3 and is given by π∗SC ¼ NðV −16V ðϕ−2Þ þ 64ϕ2 Þ=256.
Employing a sampling strategy is optimal if π∗SC N π∗PC and π∗SC N π∗FC. It
)
*
is immediate that these conditions hold if ϕ ∈ V8 ; V8 þ 2 . A paid content
′
nt n
dt
v0 V
V
Substituting this expression into Eq. (7) produces Eq. (8). (b) If v0 ≥ V,
then e
v0 ðnÞ is equal to v0 , which in turn implies that E½ e
v0 ðnÞ' ¼ v0 .
Substituting this expression into Eq. (7) yields Eq. (9). □
e ′ ðnÞn
V
e ðnÞ
V
Proof of Proposition 3. Using Lemma 3, it is straightforward to derive
the profits under a free content strategy (FC) and a paid content strategy
(PC). The profits are given by, respectively, π ∗FC = (ϕ − 1)N and
π∗PC ¼ NV=8 . Comparing the two profits shows that π∗FC ≥ π∗PC if
and only if ϕ N V8 þ 1. The profit under a sampling strategy (SC) follows
For a corner solution involving n∗ = 0, the Kuhn–Tucker conditions imply
þ nV
n :
ðn þ 1ÞV
nþ1
content strategy, λ1 = λ3 = 0, leading to p∗ ¼ NV=4 and n∗ = 0. Under
a free content strategy, we have that p∗ = 0 and n∗ = N. □
□
Proof of Lemma 3. The optimal decisions on the size of the sample and
on the price follow from solving the Kuhn–Tucker conditions in
Proposition 1.14 Under a sampling strategy, the λk's are zero and it follows
)
*
)
*
that p∗ ¼ NV 8ð2−ϕÞ þ V =64 and n∗ ¼ N 8ϕ−V =16 . Under a paid
14
It is straightforward to show that the objective function is concave for all parameter
values.
We study competition between two publishers indexed by i = 1,2
and suppose that they offer differentiated information goods to a population of consumers through an online platform. We frame the analysis
in terms of the newspaper market and assume that the readers (consumers) can be politically ranked from left to right on the political spectrum captured by the unit interval [0,1] (Gabszewicz, Laussel, & Sonnac,
2005). Horizontal differentiation captures different editorial opinions,
and we assume that the publishers are located at the extremes of the
political spectrum at x1 = 0 and x2 = 1, respectively. Vertical differentiation captures the firms' different content qualities such as the level of
investigative reporting (which are unrelated to the publishers' political
orientation).
We again assume that the perceived quality qi(ni) of information
good i is equal to its number of content parts Ni multiplied by its expected posterior quality, that is, qi ðni Þ ¼ Ni E½V i jv1i ; …; vni '. Consumers make
a discrete choice and decide which of the two information goods to purchase. A consumer's conditional indirect utility from buying information
good i is given by
ui ðx; pi ; ni Þ ¼ qi ðni Þ−τjx−xi j−pi ;
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
where x ∈ [0,1] is the consumer's political orientation and the parameter τ N 0 captures the sensitivity to horizontal mismatch |x − xi|. Intuitively, the mismatch arises because consumers make a discrete choice
and cannot purchase the information good that perfectly matches
their political preferences. Political orientations are drawn independently across consumers from a uniform distribution over the interval
[0,1].
The publishers compete for consumers by making their pricing and
sampling decisions. To derive expected content demands, we determine
the location ^x of the consumer who is indifferent between buying from
publisher 1 and from publisher 2 for given prices p = (p1, p2) and sample sizes n = (n1, n2). Clearly, the location of the indifferent consumer
^xðp; nÞ is a solution to the indifference condition u1 ð^xðp; nÞÞ ¼ u2 ð^xðp; nÞÞ,
which ensures that the indirect utilities from the two information goods
are the same. With linear mismatch, the consumer located at ^xðp; nÞ segments the market: Consumers located to the left of ^xðp; nÞ purchase
from publisher 1, while consumers located to the right of ^
xðp; nÞ purchase from publisher 2.15 Ignoring cannibalization, publisher 1 thus
faces ^xðp; nÞ consumers, while publisher 2 faces 1−^xðp; nÞ consumers.
To capture that sampling not only reveals quality but also cannibalizes
N −n
sales, we let i i denote the conditional purchase probability given
Ni
sample size ni. Hence expected content demands can be expressed as
E
D1 ðp; nÞ ¼
N1 −n1
^xðp; nÞ and
N1
E
D2 ðp; nÞ ¼
N2 −n2
ð1−^xðp; nÞÞ:
N2
Consumers thus choose their preferred publisher based on prices and
posterior quality expectations and subsequently purchase the content
with probability
N i −ni
. Similar to the monopoly case, the sampling deciNi
sion ni therefore has a direct effect on expected content demand
through the conditional purchase probability (a cannibalization effect)
and an indirect effect on ^xðp; nÞ (an expansion effect). Note that publisher i's expected content demand is zero under a free content strategy due
to the cannibalization effect.
Publisher i makes its pricing and sampling decisions so as to
#
$
n
E
E
max π i ðp; nÞ ¼ pi Di ðp; nÞ þ ϕi − i ni
pi ;ni
Ni
s:t: pi ≥ 0
0 ≤ ni ≤ Ni ;
where ϕi is the advertising effectiveness of publisher i's advertising.
Compared to the monopoly case, each publisher now has to take into account the rival's choice of strategy to make its optimal decision. Thus,
there are nine possible outcomes, summarized in Fig. A.1. If both firms
use a paid content strategy, the firms' corresponding (expected) profits
PP
are denoted by πPP
1 and π2 , respectively (and likewise for the other outcomes). For each outcome, the profit levels can be obtained by solving
the publishers' decision problems. The optimal strategy choices are
then obtained as a Nash equilibrium of the matrix game depicted in
Fig. A.1.
To analyze the optimal strategy choice, we focus on a market environment where consumers have different prior beliefs about the content quality of the two publishers. Specifically, we suppose that the
consumers' minimum estimate v01 differs from v02, while the publishers
are actually symmetric and in particular offer the same quality spectrum
(V 1 ¼ V 2). Fig. A.2 illustrates the publishers' optimal strategy choices as
a function of the gap in prior expectations and advertising effectiveness
(ϕi ≡ ϕ). The figure holds v02 constant (at v02 = 1) and plots different
values of v01 on the vertical axes. Notice that where v01 b v02,
15
See Anderson, de Palma, and Thisse (1992) for an in-depth treatment of Hotellingtype models.
205
Fig. A.1. Strategy choices and corresponding profits.
consumers believe that publisher 1 offers a lower content quality than
publisher 2.
The key insights from the analysis of the competitive case can be
summarized as follows: If consumers' prior expectations about content
qualities are similar and advertising effectiveness is high, it is optimal
for both publishers to adopt a free content strategy in equilibrium. Instead, if the gap in prior expectations is large and advertising effectiveness is high, both publishers should adopt a sampling strategy. Further,
the parameter region in which the sampling strategy is optimal for both
publishers shrinks when consumers are more sensitive to horizontal
mismatch and products thus are less substitutable.
To understand these insights, it is important to notice that each point
in the (ϕ,v01)-space in Fig. A.2 corresponds to a (pure-strategy) Nash
equilibrium of the matrix game described above.16 The solid line depicts
the cut-off threshold between the two types of Nash equilibria: sampling (SC, SC) and free content (FC, FC). Counter to intuition, it is not
per se optimal to choose a free content strategy when advertising effectiveness ϕ is high. What matters in addition is the gap in prior expectations among the two publishers: if it is small, then both firms should
adopt a free content strategy. Instead, if the gap in prior expectations
is large, then both firms should adopt a sampling strategy. Intuitively,
publisher 1 has a stronger incentive to reveal its higher than expected
quality—and it is a best-response of publisher 2 to also adopt a sampling
strategy (that is, publisher 2 attains a higher profit under a sampling
strategy than under a free content strategy when taking publisher 1's
strategy as given).
Fig. A.2 also illustrates that the parameter region in which the sampling strategy is optimal for both publishers shrinks when consumers
are more sensitive to horizontal mismatch τ. Intuitively, a higher τ
means that products are less substitutable from the viewpoint of the
consumers (e.g., when readers with left-wing preferences consider a
right-wing newspaper a less good alternative). For the publishers, this
results in a lower ability to engage in business stealing in the content
market. Consequently, the firms try to generate as much revenue as
possible in the advertising market by employing a free content strategy.
This comparative statics result mirrors the findings from the monopoly
case (cf. Fig. 3): When τ increases and publishers become monopolists
on their respective market segment, the firms should employ a free content strategy once ϕ exceeds a certain threshold level—irrespective of
the gap in prior expectations (observe that this threshold level lies at
around ϕ = 2 in Fig. A.2).
Up to now, the focus has been on strategy choices in the upper-right
corner of the matrix game in Fig. A.1. The reason is that these strategy
combinations capture the essence of the debate in the newspaper industry: whether to offer the content for free or to employ a metered
model. Of course, there is also a symmetric industry configuration
16
The code to numerically compute the Nash equilibria is available from the authors upon request.
206
D. Halbheer et al. / Intern. J. of Research in Marketing 31 (2014) 192–206
Fig. A.2. Equilibrium strategies (for ν02 = 1, α = 2, V i ¼ 3, and Ni = 10).
where both publishers employ a paid content strategy. Unsurprisingly,
this equilibrium arises if advertising effectiveness is low. In addition,
there are asymmetric strategy choices such as (PC, SC), which are optimal
when consumers overestimate the content quality of publisher 1
(and hence the incentive to not reveal low quality) and underestimate
the content quality of publisher 2 (and hence the incentive to reveal
high quality through free samples).
References
Abramson, J. (2010). Sustaining quality journalism. Daedalus, 39–44.
Ackerberg, D. A. (2003). Advertising, learning, and consumer choice in experience good
markets: An empirical examination. International Economic Review, 44(3), 1007–1040.
Akerlof, G. A. (1970). The market for lemons. Quarterly Journal of Economics, 84(3), 488–500.
Anderson, S. P., de Palma, A., & Thisse, J. -F. (1992). Discrete choice theory of product differentiation. Cambridge, MA: MIT Press.
Anderson, S. P., & Gabszewicz, J. J. (2006). The media and advertising: A tale of two-sided
markets. In V. Ginsburgh, & D. Throsby (Eds.), Handbook of the economics of art and
culture (pp. 567–614). Elsevier.
Anderson, S. P., & Renault, R. (2006). Advertising content. American Economic Review, 96,
93–113.
Bagwell, K. (2007). The economic analysis of advertising. In M. Armstrong, & R. Porter
(Eds.), Handbook of industrial organization, Vol. III. (pp. 1701–1844). Elsevier.
Bawa, K., & Shoemaker, R. (2004). The effects of free sample promotions on incremental
brand sales. Marketing Science, 23(3), 345–363.
Bhargava, H. K., & Choudhary, V. (2008). Research note: When is versioning optimal for
information goods? Management Science, 54(5), 1029–1035.
Boom, A. (2009). Download for free—When do providers of digital goods offer free
samples? Working Paper. Copenhagen Business School.
Burke, R. R., & Srull, T. K. (1988). Competitive interference and consumer memory for
advertising. Journal of Consumer Research, 15(1), 55–68.
Chellappa, R. K., & Shivendu, S. (2005). Managing piracy: Pricing and sampling strategies
for digital experience goods in vertically segmented markets. Information Systems
Research, 16(4), 400–417.
Cheng, H. K., & Tang, Q. C. (2010). Free trial or no free trial: Optimal software product
design with network effects. European Journal of Operational Research, 205,
437–447.
DeGroot, M. (1970). Optimal statistical decisions. New York: Mc Graw-Hill.
Dorfman, R., & Steiner, P. O. (1954). Optimal advertising and optimal quality. American
Economic Review, 44(5), 826–836.
Erdem, T., & Keane, M. P. (1996). Decision-making under uncertainty: Capturing dynamic
brand choice processes in turbulent consumer goods markets. Marketing Science,
15(1), 1–20.
Erdem, T., Keane, M. P., & Sun, B. (2008). A dynamic model of brand choice when price
and advertising signal product quality. Marketing Science, 27(6), 1111–1125.
Faugère, C., & Tayi, G. K. (2007). Designing free software samples: A game theoretic approach. Information Technology Management, 8, 263–278.
Gabszewicz, J. J., Laussel, D., & Sonnac, N. (2005). Attitudes toward advertising and price
competition in the press industry. In V. A. Ginsburgh (Ed.), Economics of art and
culture (pp. 61–74). Elsevier.
GlobeNewswire (April 5). Wall Street Journal offers digital ‘open house’ today; access includes online, apps for iphone, ipad. http://www.globenewswire.com/-newsroom/
news.html?d=250593
Godes, D., Ofek, E., & Sarvary, M. (2009). Content vs advertising: The impact of competition on media firm strategy. Marketing Science, 28(1), 20–35.
Heiman, A., McWilliams, B., Shen, Z., & Zilberman, D. (2001). Learning and forgetting:
Modeling optimal product sampling over time. Management Science, 47(4),
532–546.
Hotz, V. J., & Xiao, M. (2013). Strategic information disclosure: The case of
multiattribute products with heterogeneous consumers. Economic Inquiry,
51(1), 865–881.
Kind, H. J., Nilssen, T., & Sørgard, L. (2009). Business models for media firms: Does competition matter for how they raise revenue? Marketing Science, 28(6), 1112–1128.
Milgrom, P., & Roberts, J. (1986). Price and advertising signals of product quality. Journal
of Political Economy, 94(4), 796–821.
Newspaper Association of America (2012). Paid digital content benchmarking study.
Rysman, M. (2009). The economics of two-sided markets. Journal of Economic Perspectives,
23(3), 125–143.
Shapiro, C., & Varian, H. R. (1998). Information rules. Boston, MA: Harvard Business School
Press.
Sun, M. (2011). Disclosing multiple product attributes. Journal of Economics and Management
Strategy, 20(1), 134–145.
Wang, C., & Zhang, X. (2009). Sampling of information goods. Decision Support Systems,
4(1), 14–22.
Wilson, R. (1985). Multi-dimensional signalling. Economics Letters, 19(1), 17–21.
Xiang, Y., & Soberman, D. A. (2011). Preview provision under competition. Marketing
Science, 30(1), 149–169.
Intern. J. of Research in Marketing 31 (2014) 207–223
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
Empirical generalizations of demand and supply dynamics for movies
Michel Clement a,⁎, Steven Wu a, Marc Fischer b,c
a
b
c
University of Hamburg, Institute for Marketing and Media, Welckerstr. 8, D-20354 Hamburg, Germany
University of Cologne, Chair for Marketing and Market Research, Albertus-Magnus-Platz, D-50923 Köln, Germany
UTS Business School, Sydney, Australia
a r t i c l e
i n f o
Article history:
First received in 24 May 2012 and was under
review for 5 months
Available online 10 December 2013
Area Editor: Dominique M. Hanssens
Guest Editor: Marnik G. Dekimpe
Keywords:
Generalizations
Movie industry
Screen allocation
Endogeneity
a b s t r a c t
High financial risks in production and marketing, the hedonic nature of products, and the global cultural relevance of movies have encouraged a substantial number of researchers to analyze the success drivers of movies.
This research provides empirical generalizations in managing the supply and demand of motion pictures. Prior
empirical research either ignored the endogeneity of box office and screen allocation or was based on selective
samples, ignoring the large amount of smaller movies released to the market. Using two large and unique samples of all movies released in two major movie markets, the US (2000–2010; n = 2098) and Germany
(2002–2010; n = 1360), we extend prior research and present empirical generalizations and new fields of
research.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
In 2010, the movie Avatar broke all box office records and grossed
more than $2.78 billion worldwide within a few weeks. The movie was
an exceptional success for the motion picture industry. Each movie is an
innovation requiring specific management attention. In addition, the substantial costs to produce the initial first copy of the movie (Avatar was
budgeted at $237 million) and the high prelease advertising costs
(US: $53.14 million and Germany: €1.13 million for the Avatar movie)
are both sunk at the time of release, making it a risky business
(Eliashberg, Jonker, Sawhney, & Wierenga, 2000). Thus, studio managers
face a high financial risk of producing the next gigantic flops, adding to
such legendary examples as The Adventures of Pluto Nash, Stealth, and Gigli.
The hedonic nature of movies, their relevance in global culture, the
high economic importance of the industry, and the public availability
of data have led to a substantial number of academic studies on the success drivers of movies (Eliashberg, Elberse, & Leenders, 2006). Scholars
have analyzed the effect of various variables such as star power (Elberse,
2007), academy awards (Deuchert, Adjamah, & Pauly, 2005), word-ofmouth (Liu, 2006), and age restrictions (Leenders & Eliashberg, 2011),
on success measures such as box office, number of visitors, and screens.
However, the large amount of empirical research has provided conflicting results on the effect of several success drivers. For example, the role
of critics has been addressed by several researchers without consistency
or generalizable results. While some studies show positive effects of
⁎ Corresponding author. Tel.: +49 40 42838 8721.
E-mail addresses: michel@michelclement.com (M. Clement),
steven.wu@uni-hamburg.de (S. Wu), marc.fischer@wiso.uni-koeln.de (M. Fischer).
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.10.007
positive reviews and negative effects of negative reviews on sales (e.g.,
Litman, 1983), we also find studies revealing that even negative reviews
lead to higher sales (e.g., King, 2007; Wallace, Seigerman, & Holbrook,
1993) or to positive distribution effects with respect to the number of
screens (Elberse & Eliashberg, 2003). Furthermore, some authors find
evidence that critics influence sales (e.g., Basuroy, Chatterjee, & Ravid,
2003; Boatwright, Basuroy, & Kamakura, 2007; Kamakura, Basuroy, &
Boatwright, 2006; Moon, Bergey, & Iacobucci, 2010), whereas others
find that critics (actually) only predict sales and that their influence
on sales is rather negligible (Eliashberg & Shugan, 1997). Another field
with conflicting results is the star power research. Hennig-Thurau,
Völckner, Clement, and Hofmann (2013, Appendix A, p.45) present a literature overview of previous research with respect to star power and
identify ten studies that report a positive impact of stars on revenues
or admissions. However, they also identify twelve studies that find no
empirical support for such an effect (six studies find partial support).
The heterogeneous findings may be a result of various data limitations. Many studies are based on outdated data sets or face a substantial
selection bias because the authors sampled only successful movies
(e.g., top 25 in Variety or a pre-defined minimum production budget)
and ignored the large number of “small” movies that entered the market
more or less successfully (e.g., Elberse & Eliashberg, 2003; Ravid, 1999).
Furthermore, most research focuses on the US market and thus ignores
other international markets. Finally, many studies use only a very limited set of variables.
In this research, we focus on the question of whether prior findings
in the motion picture industry can be generalized. The relevance of
generalizations has been regularly highlighted in Marketing Science
(Albers, 2012; Hanssens, 2009). Especially, Bass (1995) and Ehrenberg
208
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
(1995) emphasized the necessity of empirical research focusing on generalizing prior research findings to provide further insights for new research topics. Additionally, several editors (e.g., Goldenberg & Muller,
2012; Winer, 1998) have highlighted the relevance of replications to
investigate the generalizability of earlier research findings.
This research contributes to the literature by generalizing prior empirical findings on the success factors of movies. We rely on the established
theoretical and modeling framework of Elberse and Eliashberg (2003),
which accounts for the interrelationship in the behavior of audiences
and exhibitors. Specifically, their dynamic simultaneous equation models
account for the endogeneity of revenues and screens and incorporate the
need to determine revenues and screens simultaneously. This endogenous relationship has also been identified by Krider, Li, Liu, and
Weinberg (2005) who visualize causal interferences using graphical analysis. They conclude that “the dominant industry pattern is one of movie
exhibitors monitoring box office sales and then responding with screen
allocation decisions” (Krider et al., 2005, p. 625). While Elberse and
Eliashberg (2003) (in the following used as E/E) used a sample of 164
American (co-) productions from 1999 that needed to appear at least
once in the US box office top 25, we base our analysis on a much larger
database covering the full US and German movie markets. Our sample
consists of all 2098 movies released in the US between 2000 and (partially) 2010 and all 1360 movies released in Germany between the summer
of 2002 and the spring of 2010. We collected information on all movies
released during this period from various sources and abandoned any
minimum box office criterion when choosing the movies to avoid selection biases. Thus, our sample represents the general movie market in
two major countries. Further, we extend the findings of E/E by adding
new, important variables such as sequels, MPAA ratings, US productions,
genres, and the highly relevant advertising budget for the German market. We also revise the initially counterintuitive results in E/E's study
about the effect of reviews. They find that unfavorable reviews by critics
correspond with a higher number of opening screens. This finding is surprising in light of the research findings that address the effects of critics
on the box office (e.g., Eliashberg & Shugan, 1997), but can be confirmed
and explained by supply dynamics and revenue sharing models as we
will show later in this paper. Finally, aside from generalizing the results
for the US market, we are able to estimate the model covering the full
German market. Thus, we provide a setting that also allows for generalizations across the two major international movie markets.
Our findings contribute to the literature in two ways. (1) The nature
of our data allows for replication as well as substantial extension and actualization of prior research. Thus, we provide empirical generalizations
of prior US-based research findings. (2) We provide new managerial insights for the German movie market that allow us to compare the empirical findings across two major movie markets to generate further
empirical generalizations. Summarizing, our contribution lies in demand
as well as supply generalizations that are compared across two countries.
In the next section, we provide an overview of our modeling
approach. In Section 3, we discuss our data. The estimation results are
presented in Section 4, followed by a discussion in Section 5. We
conclude with generalizations and avenues for future research.
2. Model
The value chain in the movie business is full of dependencies and conflicting interests (Hennig-Thurau, Henning, Sattler, Eggers, & Houston,
2007). Effectively, two parties are involved in the initial stage of the sequential release strategy (theatrical release) of a movie. Managers of studios and cinemas negotiate with each other, each attempting to enforce
favorable conditions (Eliashberg, Swami, Weinberg, & Wierenga, 2001).
To analyze the behavior of both parties, our model follows E/E and uses
two interdependent equations that cover audience demand (demand
equation) and screen allocation by cinemas (supply equation).
Research on the diffusion of movies has shown that demand for the
majority of movies reaches its maximum during the first week of release
(Ainslie, Drèze, & Zufryden, 2005; Sawhney & Eliashberg, 1996). In particular, first weekend box office results serve as an indicator for the movie's
total success (Joshi & Hanssens, 2009). Thus, the industry is release-driven
and the market players focus on the first week (Karniouchina, 2011). Consequently, we model the dynamics of supply and demand determinants
over time and explicitly differentiate the first week from the following
weeks using separate equations. The dynamic interests of the two parties
are a result of changing profit margins of distributors and cinemas over
time. The expected number of visitors, or at least the distributor's estimate, is best reflected in the number of opening screens, which needs
to be determined before release (Eliashberg, Hegie, Ho, Huisman, Miller,
et al., 2009). Based on the number of expected visitors, the distributor
provides the relevant number of copies to be distributed to the cinemas.
Assuming that a specific market potential of consumers wants to see
a movie, declining profit margins over time imply that strong demand
for a movie right after release is favorable for the distributor. In contrast,
the cinema earns a higher margin if consumers attend the movie in later
weeks. Thus, it can be assumed that the distributor will always attempt
to collect as many screens as possible for the first weeks, and cinemas
will attempt to shift their capacities to later weeks to maximize profits.
Therefore, consistent with E/E, we account for the endogeneity of the
number of screens when estimating revenues, and we assume that in
each week, the errors in the supply and demand equations may be correlated. We also choose a log–log formulation to directly retrieve elasticities that allow us to better compare our results to previous research.
2.1. Model for the US market—week t = 1
Eq. (1) provides the model for the demand (measured in box office)
for movie i in week t = 1.
β
β
β
β
β
β
REVENUESit ¼ e 0 " SCREENSit 1 " STARi 2 " DIRECTORi 3 " AD EXP i 4 " REVIEWSi 5
β
β
β "SEQUELi
β "US
β
" COMP SCR REV it 6 " SEASONit7 " e 8
" e 9 i " MPAAi 10
β 11 "CHILDREN i
β 12 "ACTIONi
β 13 "DOCUMENTARY i
β14 "HORRORi
"e
"e
"e
"e
β "COMEDY i
β "OTHERi
ε
" e 15
" e 16
" e Rit
ð1Þ
Revenues are driven by the number of screens allocated to movie i in
week t = 1 and a set of time-invariant variables (star power, director
power, advertising, critical acclaim, sequel, US production, MPAA rating,
and genre variables) and time-variant variables (competition and an
index variable to measure season). We provide details on the measurement of the variables in Section 3. The error term for the revenue
equation is denoted as εRit.
We model the supply of screens for movie i in week t = 1 as shown
in Eq. (2).
α0
SCREENSit ¼ e
%%α
α
α
α
α
" REVENUESit 1 " BUDGET i 2 " STARi 3 " DIRECTORi 4 " AD EXP i 5
α
α "DISTR MAJORi
α
" REVIEWSi 6 " e 7
" COMP SCR NEW it 8
α
α "SEQUELi
α "US
α
α "CHILDRENi
" COMP SCR ONGit 9 " e 10
" e 11 i " MPAAi 12 " e 13
α 14 "ACTIONi
α 15 "DOCUMENTARY i
α 16 "HORRORi
α 17 "COMEDY i
"e
"e
"e
"e
α "OTHERi
ε
" e 18
" e Sit
ð2Þ
The number of screens allocated to movie i in week t = 1 is a function
of its expected revenues (REVENUESit**), the time-invariant production
budget, the distributor's market power, two time-variant competition
variables that cover the competition for screen space from new releases
and ongoing movies, and, finally, the same variables as listed in Eq. (1),
except that we exclude the season variable in the screen equation because of fixed capacities. We include budget only in the screen supply
model because cinema operators usually know the movie budget and
use it for evaluating the success potential of the movie. In contrast, the average moviegoer is not aware of the production budget of a movie. The
error term for the supply equation is denoted as εSit. We model the
expected revenues by relying on prerelease interest for the movie on
209
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
the professional website IMDb.com. The website measures interest in a
particular movie on its pages and ranks movies accordingly in its
“Moviemeter”. We use the Moviemeter rank of movie i at the time of release as an indicator for revenues in week t = 1 as formulated in Eq. (3).1
The pre-launch rank of the Moviemeter is a powerful predictor
(R2 = .71) of expected opening-week revenues (Karniouchina, 2011):
α0
%%
REVENUESi ¼ e
α
ε
" MOVIEMETERi 1 " e i :
ð3Þ
2.2. Model for the US market—week t N 1
β
β
β
β
REVENUESit ¼ e 0 " SCREENSit1 " COMP SCR REV it 2 " SEASONit3 " WOMit4
β "WEEK it
ε
"e 5
" e Rit
ð4Þ
Here, the variable WOM captures the observed buzz with respect to
the released movie. Analogously, we model screen allocation for weeks
t N 1 in Eq. (5). Compared with E/E, we apply a modification of the supply model. We replace the variable WOM with REVENUES per SCREENS
of the previous week because we have learned from our interviews
with studio and cinema managers that movie exhibitors focus on
these two variables (which can be easily obtained by them) when
making their allocation decisions.
α
α
α
α
3
SCREENSit ¼ e 0 " COMP SCR NEW it 1 " COMP SCR ONGit 2 " REVENUESit−1
α4
a5 "WEEK it
εSit
" SCREENSit−1 " e
"e
ð5Þ
2.3. Model for the German market—week t = 1
We model the demand2 and supply for the German movie market
analogously, and include box office performance of the movie in the
US and an interaction term of the time difference in weeks between
the US and the German release and the US box office performance.
The dummy variable US_LAUNCHi indicates whether the movie was
launched in the US earlier than in Germany. Our demand (supply)
model for Germany is displayed in Eq. (6) (7). In the following coefficients with p b .10 are considered to be significant effects (two sided).
β
β
β
β
β
β
ADMISSIONSit ¼ e 0 " SCREENSit 1 " STARi 2 " DIRECTORi 3 " AD EXP i 4 " REVIEWSi 5
β
β
β US LAUNCHi
" COMP SCR REV it 6 " SEASON it 7 " US PERF i 8
!
"β9 US LAUNCHi β10 "SEQUELi β11 "USi β12 "GERi
" USPERF i " TIMELAGi
"e
"e
"e
β
β "TIP
β "CHILDREN
β "ACTIONi
i
" MPAAi 13 " e 14 i " e 15
" e 16
β18 "HORRORi
β 19 "COMEDY i
β20 "OTHERi
ε
"e
"e
"e
" e Rit
α
α
Analogous to the US market, we model the system of equations for
weeks t N 1:
β
β
β
β
ADMISSIONSit ¼ e 0 " SCREENSit1 " COMP SCR REV it2 " SEASONit3
β
β "WEEK it
ε
"WOM it4 " e 5
" e Rit
ð8Þ
and
Following E/E, we model revenues for weeks t N 1 as denoted in
Eq. (4).
β
2.4. Model for the German market—week t N 1
α
β17 "DOCUMENTARY i
"e
ð6Þ
α
α
SCREENSit ¼ e 0 " ADMISSIONit 1 " BUDGET i 3 " DIRECTORi 4 " AD EXP i 5
α
α DISTR MAJORi
α
" REVIEWSi 6 " e 7
" COMP SCR NEW it 8
α9
α 10 "US LAUNCH i
" COMP SCR ONGit " US PERF i
α "US LAUNCH i
α "SEQUELi
α "US
" ðUS PERF i " TIME LAGi Þ 11
" e 12
" e 13 i
α 14 "GERi
α 15
α 16 "TIP i
α 17 "CHILDREN i
α 18 "ACTIONi
"e
" MPAAi " e
"e
"e
α DOCUMENTARY i
α "HORRORi
α "COMEDY i
α "OTHERi
ε
" e 19
" e 20
" e 21
" e 22
" e Sit
ð7Þ
1
E/E use data from the Hollywood Stock Exchange to construct expected revenues for
the opening week. We could not use their approach because it covers only “wide-opening”
movies. Our data set includes more than just these top range movies.
2
In Germany, demand data are based on admissions and not on box office. Using admissions, however, does not have any effect on our elasticity estimates, because ticket prices
do not differ. Hence, revenues and admission are only different up to a scale factor. Since
we use multiplicative response models, the scale differences are fully absorbed in the estimated regression constant and not the elasticity parameters.
α
α
α
α
3
SCREENSit ¼ e 0 " COMP SCR NEW it 1 " COMP SCR ONGit 2 " ADMISSIONSit−1
α4
α5 "WEEK it
ε Sit
" SCREENSit−1 " e
"e :
ð9Þ
3. Data
The data comprise two data sets. Each data set consists of the complete inventory of all movies released in the US market (n = 3460
movies) and the German market (n = 2598 movies). For the US
market, we cover the time span from 2000 to 2010, and for Germany,
we include releases from 2002 to 2010. Contrary to E/E, the German
sample includes all movies released in Germany (including the movies
in German language). Thus, contrary to many other studies, we do not
face a selection effect by, for example, restricting the sample to a minimum box office requirement or by focusing on movies that have been
released in the US and then transferred to Germany. After adjusting
for missing values that especially occur in information on production
budgets, the samples comprise 2098 movies for the US market and
1360 movies for the German market.3
Table 1 provides an overview of the measurement of the variables,
and Table 2 presents the respective descriptive statistics. Table 2 further
indicates the substantial differences between the sample of E/E and our
sample and supports the generalization of our findings to the total
market.
We follow prior research in measuring our variables. All US-based
measures are obtained from variety.com, IMDb.com, Nielsen, or
metacritic.com. Measures for the German market are primarily collected from the professional website mediabiz.de and enhanced with the
help of executives from Warner Bros., Germany. We received German
advertising data from MediaCom. Compared with E/E, we measure
star and director power differently. For the US market, we use
the IMDb Starmeter ranking to measure director and star power
(Karniouchina, 2011). For the German market, we cannot rely on the
IMDb Starmeter index because many German actors are not listed in
this US-focused service. Therefore, we apply a star power measure of actors and directors that uses the confidence-based weighted mean of two
members of the market research department at Warner Bros., Germany
(van Bruggen, Lilien, & Kacker, 2002). To validate the measures, we correlate the star power measure of IMDb.com with our star measure from
Germany for US movies also released in Germany (n = 912) and find a
high correlation (p b .01) of r = .45 for actors and r = .57 for directors.
For the German sample, we rely on the German age restrictions
(FSK received from Vdf.de; Leenders & Eliashberg, 2011) and measure
critical acclaim using rankings from the movie magazine Cinema. The
magazine also provides a recommendation for the German market
that we include as a separate variable (Tip).
Compared with prior research, our model includes a wide range of
relevant drivers within two large movie markets. We specifically extend
the research of E/E by including additional variables as shown in
Tables 3 and 4 (extended model). Especially for the German market,
3
Because almost all missing values were attributable to the lack of production budget
data, we also estimated all models, including all movies without budget information
(see Table 6), to ensure that the missing data did not cause any biases. The results, which
are discussed in the robustness checks section, are very robust.
210
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
Table 1
Measures.
Variable
Description
Measure US data
Source US-data
Measure German data
Source German data
REVENUES/ADMISSIONS
SCREENS
REVENUES1**/
ADMISSIONS1**
Weekly revenues
Weekly number of screens
Expected revenues first week
Variety.com
Variety.com
Variety.com
IMDb.com
Weekly admissions
Weekly number of screens
Real admissions of first weeka
Mediabiz.de
Mediabiz.de
Mediabiz.de
REVENUES**/
ADMISSONS**
BUDGET
Expected revenues beyond
first week
Production budget
Weekly revenues in US$
Weekly number of screens
Expected revenues of the first week
were forecasted using an OLS
model with Moviemeter as
independent and revenues of the
first week as dependent variablea
Real revenues of the respective
weeka
Production budget in US$
Variety.com
Real admissions of the respective
weeka
Production budget in US$
Mediabiz.de
STAR
Star power
IMDb Starmeter ranking at release
date (highest rank = 1)b
DIRECTOR
Director power
IMDb Starmeter ranking at release
date (highest rank = 1)b
IMDb.com
AD_EXP
REVIEWS
Advertising expenditures
Critical reviews
Advertising expenditure in 000 US$
Weighted mean of US critics
Nielsen
Metacritics.com
DISTR_MAJOR
Major distributor
1 = Paramount, Sony Pictures
(Columbia, TriStar), Disney
(Buena Vista, Touchstone,
Hollywood Pictures), 20th Century
Fox, Universal, Warner (New Line,
Fine Line Features); 0 = other
Variety.com
WOM
Word-of-mouth communication
Variety.com
COMP_SCR_NEW
Competition for “screen space”
from new releases
COMP_SCR_ONG
Competition for “screen space”
from ongoing movies
COMP_SCR_REV
Competition for the attention
of audiences
SEASON
Seasonality
US_PERF
US market performance
Revenues per screen in previous
week
New releases, weighted by
advertising expenditures in 000 US
$ until release, if there were no
advertising expenditures movie
was weighted with 1 US$
Average age of ongoing releases, for
each calendar week excluding the
movie under consideration
Number of similar movies (same
genre, same MPAA), weighted by
runtime of the movie
Index value per week (0 = min,
100 = max)
n.a.
TIME_LAG
MOVIEMETER
Time lag between domestic
and foreign release
Moviemeter ranking
MPAA
MPAA rating
SEQUEL
US
Sequel
US production
GERMANY
German production
DRAMA
CHILDREN
ACTION
DOCUMENTARY
HORROR
COMEDY
OTHER
Genre drama
Genre children
Genre action
Genre documentary
Genre horror
Genre comedy
Other genre
TIP
Movie is marked with
“Tip” (Cinema)
a
IMDb.com
BoxOfficeMojo.com
the-numbers.com
IMDb.com
Variety.com
Nielsen
Varietey.com
Variety.com
Variety.com
n.a.
Moviemeter ranking on the release
date
0 = unrated; 1 = G; 2 = PG;
3 = PG-13; 4 = R; 5 = C-17
IMDb.com
1 = sequel; 0 = no sequel
1 = US production; 0 = other
country
n.a.
IMDb.com
IMDb.com
1 = Drama; 0 = other
1 = Children; 0 = other
1 = Action; 0 = other
1 = Documentary; 0 = other
1 = Horror; 0 = other
1 = Comedy; 0 = other
1 = Other genre; 0 = any of
genre above
Variety.com
Variety.com
Variety.com
Variety.com
Variety.com
Variety.com
Variety.com
IMDb.com
Movies were rated on a 0–5 scale
using WCMEAN
(highest score = 5)c
Movies were rated on a 0–5 scale
using WCMEAN
(highest score = 5)
Advertising expenditure in €
Cinema rating (on 1–5 scale,
5 is best)
1 = Paramount, Buena Vista, 20th
Century Fox, Constantin, Sony
Pictures, United Pictures
International, Warner; 0 = other
(due to market share differences
measure differs slightly from US
measure)
Admissions per screen in previous
week
New releases, weighted by
marketing budget in €, if there were
no advertising expenditures movie
was weighted with 1 €
Average age of ongoing releases, for
each calendar week excluding the
movie under consideration
Number of similar movies (same
genre, same FSK), weighted by
runtime of the movie
Index value per week (0 = min,
100 = max)
Average of revenues per screen in
the US over the first two weeks
Number of days between US
release and release in Germany
n.a.
MPAA reflects FSK rating in
Germany
1 = FSK0; 2 = FSK6; 3 = FSK12;
4 = FSK16; 5 = FSK18
1 = sequel; 0 = no sequel
1 = US production; 0 = other
country
1 = German production;
0 = other country
1 = Drama; 0 = other
1 = Children; 0 = other
1 = Action; 0 = other
1 = Documentary; 0 = other
1 = Horror; 0 = other
1 = Comedy; 0 = other
1 = Other genre; 0 = any of
genre above
1 = “Tip”; 0 = no “Tip”
IMDb.com
BoxOfficeMojo.com
the-numbers.com
Own survey by
interviews
Own survey by
interviews
MediaCom
Cinema.de
Mediabiz.de
Mediabiz.de
Mediabiz.de
MediaCom
MediaBiz.de
MediaBiz.de
Babelsberg Charts
Variety.com
Variety.com
Mediabiz.de
Vdf.de
IMDb.com
Vdf.de
Vdf.de
MediaBiz.de
MediaBiz.de
MediaBiz.de
MediaBiz.de
MediaBiz.de
MediaBiz.de
MediaBiz.de
n
Measure is different from Elberse and Eliashberg (2003).
Variable is reverse coded for estimation, meaning that a larger value represents higher star power.
c
We asked two experts to rate the star power of each movie in our data set. In addition, they had to provide a confidence value for each rating. The star power for a movie was then
generated by applying the confidence-based weighted means method (WCMEAN, van Bruggen et al., 2002).
b
211
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
Table 2
Key descriptive statistics.
Extended model
US
Germany
Elberse/Eliashberg
Attributes
US
Germany
Variable
N
Mean/%
Median
SD
Budget (000 US$)
Star rankinga
Director rankinga
Ad_Exp (000 US$)
Review ratings
Screens (t = 1)
Revenues (t = 1) (000 US$)
Total Revenues (000 US$)
Length of run (weeks)
Comp_Scr_New (t = 1)
Comp_Scr_Ong (t = 1)
Comp_Scr_Rev (t = 1)
Season (t = 1)
Moviemeter ranking (t = 1)
MPAA
Sequel
US productions
Distr_Major
Children
Action
Documentary
Drama
Horror
Comedy
Other
Budget (000 US$)
Star rating
Director rating
Ad_Exp (000 €)
Review ratings
Screens (t = 1)
Admissions (t = 1) (000)
Total admissions (000)
Length of run (weeks)
US_Perfb (000 US$)
Time_Lagb
Comp_Scr_New (t = 1)
Comp_Scr_Ong (t = 1)
Comp_Scr_Rev (t = 1)
Season (t = 1)
MPAA (FSK)
Sequel
US
German productions
Distr_Major
Tip
Children
Action
Documentary
Drama
Horror
Comedy
Other
Budget (000 US$)
Star rating
Director rating
Ad_Exp (US) (000 US$)
Review ratings
Screens (t = 1)
Revenues (t = 1) (000 US$)
Total Revenues (000 US$)
Length of run (weeks)
Screens (t = 1)
Admissions (t = 1) (000)
Total admissions (000)
Length of run (weeks)
US_Perf (000 US$)b
Time_Lag
2098
2098
2098
2098
2098
1917
1917
2098
2098
1917
1917
1917
1917
1917
2098
2098
2098
2098
2098
2098
2098
2098
2098
2098
2098
1360
1360
1360
1360
1360
1335
1335
1360
1360
1360
1360
1335
1335
1335
1335
1360
1360
1360
1360
1360
1360
1360
1360
1360
1360
1360
1360
1360
139
164
164
164
158
164
164
164
164
138
138
138
138
138
138
32,815.00
9172.35
10,678.39
11,438.99
53.87
1525.33
16,252.09
39,106.58
13.30
39,954.59
14.11
11.43
55.25
914.59
2.86
14.06%
79.17%
39.66%
2.53%
26.31%
3.34%
34.70%
3.86%
24.93%
4.34%
36,532.67
1.34
.35
656.30
3.38
243.53
211.37
581.42
10.47
7.06
118.17
2452.83
8.04
8.36
66.19
2.75
11.69%
64.78%
11.54%
50.59%
22.79%
6.54%
21.47%
2.57%
27.50%
5.81%
27.13%
8.97%
36,879.42
46.28
25.28
10,455.01
3.15
1658.73
10,964.91
43,712.51
16.21
276.69
2876.92
9650.27
9.67
8.97
139.83
19,000.00
118.00
1936.00
4731.91
54.00
1694.00
6143.98
13,585.99
12.00
40,919.03
11.84
11.19
48.50
21.00
3.00
39,775.91
81,462.15
41,469.67
13,387.57
17.89
1418.86
26,216.15
65,737.08
10.32
24,851.38
5.45
6.21
17.37
3070.51
1.33
Minimum
1.10
1.00
1.00
.00
3.00
1.00
.01
.10
1.00
.00
1.00
.00
32.47
1.00
.00
Maximum
300,000.00
2,737,503.00
1,338,551.00
55,349.52
100.00
4468.00
196,019.50
839,081.62
220.00
137,029.97
31.95
31.38
100.00
53,569.00
5.00
20,381.77
1.08
.04
193.15
4.00
149.00
64.25
157.56
9.00
1.86
104.00
2248.23
7.78
8.27
68.17
3.00
43,270.39
1.22
.81
945.03
1.17
247.90
422.47
1144.55
6.56
22.82
113.44
1676.37
1.43
4.01
12.36
1.11
.22
.00
.00
.00
.00
1.00
.04
.05
1.00
.00
.00
.00
3.97
.11
40.00
1.00
448,595.00
4.75
5.00
5512.73
5.00
1337.00
3680.04
10,428.18
52.00
719.97
832.00
10,966.11
13.01
25.38
100.00
5.00
30,000.00
48.39
13.82
10,005.90
3.33
1870.00
6947.73
22,059.95
16.00
245.00
1199.07
3400.41
8.00
5.66
124.00
29,762.84
33.67
28.63
6626.67
.84
999.82
12,569.02
58,542.32
6.66
229.23
4445.93
16,187.10
7.34
11.89
97.42
22.00
1.00
1.00
6.20
1.00
1.00
6.81
752.12
2.00
1.00
1.61
3.09
1.00
.78
.00
170,000.00
99.73
97.53
27,827.80
4.67
3309.00
63,674.40
431,088.30
30.00
1001.00
32,236.48
99,859.53
30.00
85.63
529.00
Note: Ad_Exp = advertising expenditure, Comp_Scr_New = competition for “screen space” from new releases, Comp_Scr_Ong = competition for “screen space” for ongoing movies,
Comp_Scr_Rev = competition for the attention of audiences, MPAA(/FSK) = age restriction, Distr_Major = major distributor, US_Perf = US market performance, Time_Lag = time
lag between domestic and foreign release, Tip = movie is marked with “Tip” (in the German magazine “Cinema”).
a
Compared with Elberse and Eliashberg (2003), the variable was reverse coded and represents the ranking positions of the IMDb Starmeter.
b
Includes only those movies that were launched in the US earlier than in Germany.
212
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
Table 3
3SLS estimation results US opening week.
log(Revenues)
Extended model
Elberse/Eliashberg
Coeffi
se
p
Constant
log(Screens)
log(Star)
log(Director)
log(Ad_Exp)
log(Reviews)
log(Comp_Rev)
log(Season)
Extension
Sequel
US
log(MPAA)
Children
Action
Documentary
Horror
Comedy
Other
N
R2
Note: ⁎⁎⁎ p b .01; ⁎⁎ p b .05; ⁎ p b .10 (two sided).
4.99
1.04
−.00
.11
.02
1.05
−.19
.29
.18
−.23
−.01
−.25
−.28
.03
−.39
−.12
−.21
1917
.92
.51
.02
.01
.02
.01
.06
.04
.06
.06
.06
.05
.14
.06
.14
.11
.05
.11
.00
.00
.93
.00
.05
.00
.00
.00
.00
.00
.85
.08
.00
.82
.00
.03
.06
log(Screens)
Extended model
Variables Elberse/Eliashberg
Variables Elberse/Eliashberg
Constant
log(Revenues)**
log(Budget)
log(Star)
log(Director)
log(Ad_Exp)
log(Reviews)
Distr_Major
log(Comp_Scr_New)
log(Comp_Scr_Ong)
Extension
Sequel
US
log(MPAA)
Children
Action
Documentary
Horror
Comedy
Other
N
R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
Coeffi
***
***
***
**
***
***
***
***
***
.27
.81
.10
.00
.20
.77
−.20
.02
se
p
1.22
.04
.04
.03
.07
.03
.06
.27
.82
.00
.01
.91
.01
.00
.00
.95
***
**
**
***
***
*
***
***
**
*
164
.88
Elberse/Eliashberg
Coeffi
se
p
−1.66
.40
.28
.10
−.17
.31
−1.15
.67
−.27
.07
.26
.76
−.49
.60
.55
.64
.93
.43
.17
1917
.76
.89
.03
.03
.02
.03
.02
.10
.07
.02
.08
.11
.11
.09
.23
.10
.24
.19
.09
.19
.06
.00
.00
.00
.00
.00
.00
.00
.00
.39
.01
.00
.00
.01
.00
.01
.00
.00
.36
*
***
***
***
***
***
***
***
***
Coeffi
se
p
−.29
1.41
−.02
.04
−.03
.25
−1.48
.10
−.19
.07
2.14
.08
.10
.05
.05
.11
.28
.19
.21
.16
.89
.00
.87
.47
.57
.02
.00
.61
.36
.65
***
**
***
**
***
***
**
***
***
***
***
164
.81
Note: Ad_Exp = advertising expenditure, Comp_Scr_Rev = Competition for the attention of audiences, Distr_Major = major distributor, Comp_Scr_New = Competition for “screen
space” from new releases, Comp_Scr_Ong = Competition for “screen space” for ongoing movies.
we include major variables (e.g., advertising) to better account for regional market specifics than the prior study by E/E.
4. Estimation
To estimate the dynamic system of equations, we need to compute
estimates for the expected revenues (REVENUES⁎⁎
i ) in the respective
screen equations for the first week (t = 1). For the US market, we generate estimates for first-week revenue expectations by estimating the
parameters of Eq. (3) using OLS.4 For the German market, we use real
admissions in the first week. We assume that the expected revenues
for the first week in Germany do not differ much from the observed
revenues because most movies on the German market have been
4
The double exponential smoothing (DES) procedure used by E/E to estimate the expected revenues for the following weeks could not be applied to our data sets. We find that
DES is only applicable for movies with a very regular (already smooth) diffusion pattern,
which is the case for the top 25 movies. However, our data sets also contain smaller movies
with less regular diffusion patterns, resulting in highly fluctuating DES values for consecutive weeks of a movie if we apply the method proposed by E/E.
previously released in the US. We test this assumption by splitting the
sample in a calibration (all movies released in 2002–2005, N = 614)
and a validation sample (all movies released in 2006–2010, N = 721).
We estimate the first week admissions in the calibration sample using
Eq. (6) applying OLS (R2 = 0.87). Relying on the estimated parameters,
we predict the admissions at t = 1 for the validation sample and compare the estimated and the observed admissions in 2006–2010. The
results are highly correlated with r = .94 (p b .01). Therefore, we
argue that movie exhibitors have very realistic expectations about
how successfully the movie will open.5
We then estimate the demand and supply equations for the US
and German markets using instrumental variable estimation. The
Hausman–Wu specification test (Greene, 2006) provides support for
5
Given the lack of Moviemeter data for many non-US movies, a procedure analogous to
that for the US sample was not possible. It was also not possible to use US box office data
during the first week to obtain a proxy value for the expected revenues for the German
market because not all movies in Germany are previously released in the US. Using either
method would result in many missing cases, causing a severe systematic bias in our data
set.
213
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
Table 4
3SLS estimation results Germany opening week.
log(Admissions)
Variables Elberse/Eliashberg
Constant
log(Screens)
log(Star)
log(Director)
log(Reviews)
log(Comp_Rev)
log(Season)
log(US_Perf)
log(Time_Lag*US_Perf)
Extension
log(Ad_Exp)
Sequel
US
Germany
log(MPAA)
Tip
Children
Action
Documentary
Horror
Comedy
Other
N
R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
log(Screens)
Variables Elberse/Eliashberg
Constant
log(Admissions)
log(Budget)
log(Star)
log(Director)
log(Reviews)
Distr_Major
log(Comp_Scr_New)
log(Comp_Scr_Ong)
log(US_Perf)
log(Time_Lag*US_Perf)
Extension
log(Ad_Exp)
Sequel
US
Germany
log(MPAA)
Tip
Children
Action
Documentary
Horror
Comedy
Other
N
R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
Extended model
Elberse/Eliashberg
Coeffi
se
p
5.30
1.11
.20
.12
.18
−.09
.57
.05
−.02
.01
.21
−.17
.02
−.00
.32
−.19
.00
−.02
.04
.07
.04
1335
.88
.24
.06
.05
.05
.06
.05
.09
.01
.01
.01
.06
.06
.07
.0
.05
.12
.07
.13
.12
.06
.09
.00
.00
.00
.03
.00
.05
.00
.00
.00
.06
.00
.00
.83
.97
.00
.12
.96
.87
.74
.19
.62
***
***
***
**
***
**
***
***
***
*
***
***
Coeffi
se
p
−2.47
1.51
−.03
−.02
.37
−.07
.39
.17
.08
1.01
.07
.04
.03
.23
.02
.18
.08
.90
.02
.00
.51
.56
.11
.00
.03
.04
.90
**
***
***
**
**
***
138
.88
Extended model
Elberse/Eliashberg
Coeffi
se
p
−2.49
.29
.17
.13
−.03
.04
.16
−.00
.17
.02
−.02
.06
.18
.15
.21
−.13
.08
.65
.33
−.05
.43
.17
.27
1335
.85
.40
.09
.03
.06
.05
.06
.04
.01
.09
.01
.0
.01
.07
.05
.07
.04
.06
.11
.07
.10
.11
.06
.08
.00
.00
.00
.03
.56
.51
.00
.75
.07
.08
.01
.00
.02
.00
.00
.00
.21
.00
.00
.60
.00
.00
.00
Coeffi
***
***
***
**
***
*
*
***
***
**
***
***
***
.99
.38
.17
.12
−.20
−.41
−.15
−.13
.34
.95
−.28
se
p
1.58
.07
.10
.05
.25
.37
.16
.06
.36
.16
.13
.53
.00
.07
.01
.44
.26
.36
.02
.34
.00
.03
***
*
**
**
***
**
***
***
***
***
***
138
.50
Note: Comp_Scr_Rev = competition for the attention of audiences, US_Perf = US market performance, Time_Lag = time lag between domestic and foreign release,
Ad_Exp = advertising expenditure, Distr_Major = major distributor, Comp_Scr_New = competition for “screen space” from new releases, Comp_Scr_Ong = competition for “screen
space” for ongoing movies, Tip = movie is marked with “Tip” (in the German magazine “Cinema”).
endogenous relation between REVENUES and SCREENS for all equations
(Eqs. (1), (4), (7), (8)) except for Eq. (6).
Further, we consider advertising expenditures as a potentially endogenous variable in the demand equation. We test for the exogeneity
assumption of advertising expenditures by applying the Hausman–Wu
specification test (Greene, 2006). All exogenous variables that we use
to instrument and identify screens in the demand equation serve also
as instruments for advertising expenditures. Advertising expenditures
in Germany serve as a powerful instrument to identify advertising expenditures on the same movies in the US and vice versa. Expenditures
in both countries are correlated because the same studio makes budget
decisions. Advertising campaigns in one country, however, should not
impact moviegoers in the other country. The Hausman–Wu test does
not reject the exogeneity assumption for the advertising variable. The
coefficients associated with the additional test variables are not significant (p = .433 for the U.S. and p = .374 for Germany).6 Thus, we focus
on the endogenous relation between REVENUES and SCREENS for all
equations.
Analogous to E/E, we do not use additional instrumental variables.
The instrumental variables for the respective equations result from
the combination of the two interlaced supply and demand equations.
For example, in Eq. (6), the first stage is denoted by Eq. (7), which
6
First-stage R2 ranges from .226 (Germany) to .588 (USA) and the associated F-values
exceed the threshold of 10 in both cases (Greene, 2006). Thus, we conclude that our instruments are not weak.
214
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
provides the overidentifying instruments BUDGET, DISTR_MAJOR,
COMP_SCR_NEW, and COMP_SCR_ONG. These variables serve as instrumental variables because we assume that they only affect the exhibitors'
and not the consumer behavior. In Eq. (7), the first stage is denoted by
Eq. (6). In this case, COMP_SCR_REV and SEASON serve as overidentifying
instruments, and we assume that they only affect the consumer, not the
exhibitor.
Based on the tests, we linearize all equations by log-transforming
them and estimating the respective parameters using 3SLS. We account
for time-specific fixed effects in estimating our model for the first
26 weeks by including a set of dummies in the equations (Eqs. (4),
(5), (8), (9)) for the following weeks, which is denoted by the vector
WEEK. All weeks after week 26 are captured by a single dummy because
we do not expect any time-specific effects after this week.
In the appendices, we present the correlation matrices for both markets, which rarely have been presented in prior research. As expected,
we find substantial correlations between various variables that raise
the question of multicollinearity. However, the VIF values remain
below the critical value of 10 and do not indicate severe problems
(see Table 6). We ran various robustness checks with respect to
multicollinearity by dropping some variables from the model. However,
we find that the results are robust, which is depicted in Table 6 where,
for example, advertising spending has been dropped.
our findings are different from the results of E/E that show a significant
effect of stars (.10, p b .05) but no effect of directors (p = .91). With respect to advertising, we find rather small elasticities for movies during
the first week (.02, p b .05). These small elasticities are substantially
lower than the elasticity reported by E/E (.20, p b .05), but reflect the effect of a potential selection bias because their sample covers only “rather successful” movies that were at least one week in the top 25 of
Variety. Very similar to E/E, we find a negative effect of competition
on box office. Similar movies (same genre and/or age restriction) will
decrease the demand for the movie under consideration. The elasticity
of −.19 (p b .01) is close to the − .20 (p b .01) reported by E/E. Contrary to E/E, we find a significant pull effect (.29, p b .01) in high seasons, which substantially increases demand (Radas & Shugan, 1998).
Relying on other findings (Hennig-Thurau, Houston, & Heitjans, 2009),
we find significant higher demand for sequels (.18, p b .01), which
highlights the recent discussion of sequels' relevance for Hollywood.
Ceteris paribus, we find a negative impact of US productions on demand
(−.23, p b .01). With respect to MPAA ratings, we find no significant effects on demand (p = .85). Compared to drama (which serves as the
base category for the genres), we find significantly lower demand for
children (p b .1), action (p b .01), horror (p b .01), comedy (p b .05),
or other movies (p b .01) that potentially target genre-specific audiences representing a subset of the market potential. No significant
effects are observed for documentaries.
5. Results
Tables 3–5 provide the 3SLS results for the supply and demand equations of the first week in the US (Table 3) and Germany (Table 4).
Table 5 presents the estimation results of the subsequent weeks for
both markets.
The tables present our estimates and, for comparison, the findings of
E/E. A comparison of the R2 values indicates a very good fit of all our
models. In particular, the models for the German market show a substantially better fit than the findings of E/E.
First, we compare the estimated elasticities to the prior findings of
E/E. We find substantially more significant influences in the equations
addressing the first week, which can partially be attributed to the
larger sample size (US 164 versus 1917 and Germany 138 versus
1335 movies).7 Generally, our findings support the significant findings of E/E with respect to the direction of the estimated elasticities.
5.1. US demand effects t = 1
Focusing on the first week estimation results for the US, we find that
the most relevant drivers of movie consumers' demand are the number
of screens and movie reviews. We assume that reviews reflect the quality of the movie, which is the strongest driver of consumers' demand.
These findings correspond to the findings of E/E. However, we find a
higher elasticity for reviews (1.05 versus .77, both p b .01). The lower
elasticity reported by E/E points towards a sampling bias. Besides blockbusters our sample also includes low budget movies with lower media
coverage. Thus, reviews may serve as a stronger signal for “smaller”
movies as compared to the successful ones. In addition, we assume
that our more recent data better reflects the effect that reviews are
now easier to obtain via the Internet than before via TV, newspapers,
or magazines (Hennig-Thurau, Wiertz, & Feldhaus, 2013). This is
reflected in the overall higher elasticity finding. We also find a higher
elasticity for screens (1.04 versus .81, both p b .01). This result, however, corresponds to more recent findings of Karniouchina (2011) who
uses data from 2005 and finds a screen elasticity of .94 for t = 1.
Further, we find a significant effect of director power on demand (.11,
p b .01) but no significant effect of stars on demand (p = .93). Thus,
7
We analyzed a comparable subsample of our US data set to test whether we could replicate the findings of E/E. We found very similar results, although we used somewhat different measures (the results can be obtained from the authors upon request).
5.2. US supply effects t = 1
Our results reveal a substantial number of significant drivers that influence the behavior of cinema managers, who have a high incentive to
shift demand from early weeks to later weeks because margins increase
over time. Whereas E/E identified only expected revenues, advertising,
and reviews as significant influencers, we find much more differentiated results. Consistent with E/E, we find that positive reviews reduce the
number of screens during the first week. The elasticity of reviews with
respect to the number of screens is − 1.15 (p b .01), compared with
−1.48 (p b .01) found by E/E. Thus, movies with better reviews receive
fewer screens during the first week. Karniouchina (2011) observed a
similar negative effect with respect to screen allocation for highquality movies and star buzz. Similar to E/E as well as Karniouchina
(2011), we argue that this effect may arise from the larger attractiveness
cinemas have for high-quality movies. As a result, the movie is expected
to have a stronger staying power. Low-quality movies do not have this
staying power, so that cinemas have an incentive to support these
movies by providing relatively more screens in the first week. In addition, sometimes distributors systematically limit screens (limited release) in the opening week in order to increase the hype of a high
quality movie (Karniouchina, 2011). Finally, the result may also be an
expression of the documented negative relationship between customer
satisfaction and market share as noted by Fornell (1995). In this case the
quality of the product valued by movie reviewers (typically highbrow
content) is in contrast to the notion of lowbrow or “mass appeal” content that comes with widely distributed (and, therefore, high marketshare) movies.8
Further, we find a significant effect (.40, p b .01) of the expected revenues of a movie, albeit smaller than E/E (1.41, p b .01). With respect to
the last consistent finding of advertising, we show rather similar results
(.31, p b .01 versus .25, p b .05 of E/E). We also identify a number of additional influences. For example, we find a significant effect of production budget on screens (.28, p b .01), which was not reported by E/E.
Moreover, consistent with the negative elasticity of reviews on screens,
we find that director power reduces significantly the provision of
screens. Contrary to actors, the director works behind the camera and
is responsible for the visualization of the story with respect to the
8
We thank the area editor for pointing this issue to us.
215
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
Table 5
3SLS estimation results US and Germany following weeks.
log(Revenues)
US following weeks
Extended model
Variables E/E
Constant
log(Screens)
log(Comp_Rev)
log(Season)
log(WOM)
N
R2
Adj. R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
log(Screens)
Elberse/Eliashberg
Coeffi
se
.14
1.00
−.01
.15
.85
24,030
.97
.05
.00
.00
.01
.00
p
.01
.00
.05
.00
.00
Coeffi
***
***
*
***
***
Constant
log(Revenues)
log(Comp_Scr_New)
log(Comp_Scr_Ong)
log(WOM)
Ext.
log(Revenues-1)
log(Screens-1)
N
R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
log(Admissions)
Constant
log(Screens)
log(Comp_Rev)
log(Season)
log(WOM)
N
R2
Adj. R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
log(Screens)
Coeffi
se
−2.04
.05
p
.00
***
Coeffi
−.01
.00
.00
.01
.00
1.00
***
.32
.60
24,030
.95
.00
.00
.00
.00
***
***
Constant
log(Admissions)
log(Comp_Scr_New)
log(Comp_Scr_Ong)
log(WOM)
Ext.
log(Admissions-1)
log(Screens-1)
N
R2
Note: ***p b .01; **p b .05; *p b .10 (two sided).
.22
.00
.04
.70
.00
***
**
***
−.59
1.08
−.26
.06
.35
se
p
.38
.05
.02
.05
.09
.12
.00
.00
.27
.00
***
***
***
2489
.74
Germany following weeks
Elberse/Eliashberg
Coeffi
se
−.03
1.03
.03
.19
.86
12,807
.94
.04
.00
.01
.02
.01
p
.51
.00
.00
.00
.00
Coeffi
***
***
***
***
−.55
1.08
.03
.08
.74
1196
.88
.88
se
p
.22
.03
.03
.05
.03
.01
.00
.33
.09
.00
se
p
.23
.01
.08
.10
.16
.00
.00
.00
.11
.05
**
***
*
***
Germany following weeks
Extended model
Variables E/E
p
.24
.02
.02
.06
.04
Elberse/Eliashberg
Extended model
Variables E/E
se
US following weeks
Extended model
Variables E/E
.29
1.01
−.03
.02
1.05
2489
.92
.92
Elberse/Eliashberg
Coeffi
se
p
Coeffi
−1.02
.05
.00
***
−.02
.07
.00
.02
.00
.00
***
***
.20
.75
12,807
.96
.00
.01
.00
.00
***
***
2.94
.09
−.42
.16
.32
***
***
***
*
12,807
.64
Note: Comp_Scr_Rev = competition for the attention of audiences, Comp_Scr_New = competition for “screen space” from new releases, Comp_Scr_Ong = competition for “screen
space” for ongoing movies, Revenues-1 = revenues in previous week, Screens-1 = number of screens in previous week, Admissions-1 = admissions in previous week.
overall content of the script resulting in an immense impact on the overall product (Ainslie et al., 2005). Further, some star directors succeed in
giving their movies their own distinctive style (Hadida, 2010). Thus,
higher director power may serve as an indicator for higher quality
and, eventually, longer staying power of the movie. Therefore, we assume that this effect is in line with the argument that positive reviews
indicate staying power, which leads to greater incentives for cinema
managers to extend the screening period (with higher margins). This
finding is supported by the significant correlation of positive reviews
and director power (Table A1 in Appendix A). Interestingly, the correlation of star power with reviews is negative, indicating that greater star
power does not necessarily lead to high quality movies. Therefore,
ceteris paribus, star power significantly and positively influences the allocation of screens for the first week (.10, p b .01). Further, we find a
strong significant effect of major distributors on screen allocation (.67,
p b .01). Thus, major studios are able to monetize their market position
in negotiating more screens for their movies. With respect to competition, we also find a negative elasticity for screens from newly released
movies (− .27, p b .01), whereas no significant influence is observed
from competition generated by ongoing releases. Sequels also generate
more screens (.26, p b .05). Therefore, they add further relevance to the
total effect of sequels on demand through this substantial indirect effect
216
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
from screens on total demand. We find a negative effect of US productions on demand (− .23, p b .01). This negative influence on demand
may be due to variety seeking motives of moviegoers. The vast majority
of movies shown are produced in Hollywood. A foreign production
might be a welcome increase in variety. The result probably would be
different if the share of non-US productions was higher. In addition,
the finding may reflect the fact that we have a much larger data set including the many “small” US movies released to the market. However,
we find a strong positive effect of this variable on screens (.76,
p b .01). One reason for this contrary effect could be that the production
and distribution of films and television programs is perceived as one of
the “nation's most valuable cultural and economic resources” (MPAA,
2013; http://www.mpaa.org/policy/state-by-state) which is used as an
argument for substantial financial distribution subsidization by US
states, which in turn may lead to a relative high number of screens at
the opening week. Thus, the overall effect of US productions is positive
if the screen elasticity of 1.04 is considered. Adding to recent research by
Leenders and Eliashberg (2011), we find that restrictive MPAA ratings
significantly reduce the number of screens allocated to a movie (−.49,
p b .01). Finally, the genre effects show, compared with drama and except for “other movies,” all significant positive effects with respect to
screens.
5.3. German demand effects t = 1
The results for the first week in the German movie market support
and widely enhance the prior findings of E/E. A strong relevance
of screens on demand (1.11, p b .01) is also observed in our study
(E/E = 1.51, p b .01). The strong influence of reviews observed in the
US market is not present in the German market. However, E/E also
found a lower elasticity (.37) for Germany. In our study, we find that reviews significantly influence demand (.18, p b .01) in the German market. In addition, the variable “Tip” (regression coefficient: .32 resulting
in a multiplier effect of 37.7%, p b .01) includes the magazine's recommendation, which is also highly significant. Thus, reviews have a
major effect on demand in Germany. Interestingly, we find that seasonality has a substantial effect on demand (.57, p b .01), which is higher
than E/E's finding (.39, p b .05) and substantially higher than in the
US. With respect to (German and international) star (.20, p b .01) and
director power (.12, p b .05), we find that both variables significantly
explain demand. Thus, German moviegoers seem to be more attracted
by star and director power than their US counterparts. Consistent
with the US results, we also find small but significant effects from advertising (.01, p b .1) and competition (−.09, p b .05). In addition, our results reveal that successful US movies face greater demand in Germany
(.05, p b .01), although the effect is dampened if the timing of the
release is delayed compared with the US release (− .02, p b .01). Demand is also higher for sequels and the effect size with respect to admissions (.21, p b .01) compares well with the US (.18, p b .01). Further,
US-produced movies face lower demand (− .17, p b .01), which can
be also attributed to the fact that we include all US-produced movies
released in Germany in our sample. We find all other variables to be
insignificant, meaning that we do not find any effect on demand with
respect to age ratings, movies produced in Germany, or genres.
5.4. German supply effects t = 1
We find very consistent results in the supply equation for the
German market compared with our US results. The estimated effects
from expected admissions (.29, p b .01), budget (.17, p b .01), and
star power (.13, p b .05) are very close to the findings of E/E. Similar
to the results of E/E, we also find no significant effect of director
power or reviews (including the German specific “Tip” variable) on
screens. Although this result is consistent with E/E, we found director
power or reviews to be significantly negative in the US equation, indicating regional differences between these two markets. Our findings
add new insights into the positive effect of adverting on screens. We
find a significant but rather small effect (.06, p b .01) of advertising on
screens. Additionally, we find significant positive effects on the number
of screens attributable to the market power of major distributors (.16,
p b .01) that we also identified in the US market, although on a much
larger scale. Unlike E/E, we find no significant effect of competition on
new releases. However, the competition effect of ongoing movies (.17,
p b .1), measured by the average age of the competing movies, significantly influences the number of allocated screens. Thus, greater competition means a lower average age, leading to fewer allocated screens for
the movie under consideration. Moreover, our findings indicate a much
smaller effect of US performance on the success in the German market
compared with E/E (.02, p b .1 vs. .95, p b .01), and the interaction effect on the time lag between US and German release is much smaller
than that of E/E (−.02, p b .01 vs. −.28, p b .05). However, our results
are conclusive because E/E do not account for non-US movies or German
advertising, likely resulting in an overestimation of the US performance
and time lag effect.
The remaining effects point in the same direction as in the US supply
equation. Sequels (.18, p b .05), US and German productions (.15 and
.21, both p b .01), MPAA rating (− .13, p b .01), and genre effects are
consistent in both equations (with the exception that, on the one
hand, we cannot find a significant effect for documentaries in the
German equation and, on the other hand, the genre category “others”
becomes significant).
5.5. US demand effects t N 1
The estimated elasticities for the following weeks after week 1 are
very similar to E/E. In our results, we find that screen elasticity has the
highest effect on demand (1.00, p b .01) and that the result is highly
consistent with E/E (1.01, p b .01). This result corresponds also to the
findings of Karniouchina (2011), who reports elasticities of .94/.89/.93
for t = 2/3/4. We also observe a negative competition effect from
other movies of the same genre and/or age rating (− .01, p b .1) that
is similar to E/E's findings (− .03, p b .05). Our results differ with respect to E/E in that we find a significant effect of season on demand
(.15, p b .01) and a smaller effect of WOM (.85 versus E/E's 1.05, both
p b .01) on demand.
5.6. US supply effects t N 1
With respect to supply effects in the following weeks, we find strong
effects of the previous week's box office result (.32, p b .01) and, especially, the number of screens allocated to the movie in the previous
week (.60, p b .01) on screens. Our results better separate the findings
of E/E that report a significant effect of expected revenues (1.08,
p b .01) on screens. Consistent with E/E but on a smaller scale (− .01
versus − .26, both p b .01), we find a significant negative effect of
competition from newly released movies on screens.
5.7. German demand and supply effects t N 1
The results for screen allocation in Germany are similar to our US results with respect to the relevance of the movie's box office performance
measured by the admissions (.20, p b .01) and screens (.75, p b .01) of
the previous week. However, we find a stronger tendency in the
German market to depend on the previous week's screen allocation. Interestingly, we find stronger effects from competition in the German
market. On the one hand, competition from similar movies has a positive effect (.03, p b .01) on demand. Such a market expansion effect attributable to competition was also identified by Radas and Shugan
(1998). On the other hand, we find a stronger effect (−.02, p b .01) of
competing new releases and ongoing movies (.07, p b .01) on the
supply side than in the US market.
217
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
5.8. Robustness and validation checks
Table 6 shows the results of some of our robustness and validation
checks with respect to the results of the first week in the two markets
(for comparison purposes, we report our benchmark results in the
first column). First, as previously noted, we observe a large number of
missing values for our budget variables. If we exclude the production
budget from our estimation (column 3SLS w/o budget), we can rely
on larger data sets (n = 3460 for the US market and n = 2598 for the
German market). Interestingly, we find no substantial effect in the demand model. In the supply model, the exclusion of the budget only
leads to some minor changes. The omission of this variable results in
an increase in the effect of major distributors on screens (1.01 instead
of .67, both p b .01). Thus, the effect of the budget on screens is picked
up by the major studio dummy variable. We also find less relevance
for US productions on screen allocation (.35 instead of .76, both
p b .01) in the US market. In Germany, we observe consistent effects
and find that the market power of the distributor becomes more
relevant for screen allocation. Interestingly, we find that the dummy
variable indicating German movies becomes insignificant and that documentaries face a significant hurdle in gaining additional screens (−.22,
p b .01). Second, we test whether the omission of advertising results in
substantial changes attributable to multicollinearity. However, we only
find marginal changes in the results (see column 3 in Table 6). Finally,
we also report the results of a simple OLS for comparison with E/E. As
a conclusion, our robustness checks and our verifications of the results
Table 6
Robustness and validation checks.
US first week
3SLS
log(Revenues//Admissions)
Variables Elberse/
Eliashberg
Extension
log(Screens)
Variables Elberse/
Eliashberg
Extension
Constant
log(Screens)
log(Star)
log(Director)
log(Ad_Exp)
log(Reviews)
log(Comp_Rev)
log(Season)
Sequel
US
Germany
log(MPAA)
Tip
Children
Action
Documentary
Other
R2
Adj. R2
Mean VIF
Max VIF
Constant
log(Revenues**//
Admissions)
log(Budget)
log(Star)
log(Director)
log(Ad_Exp)
log(Reviews)
Distr_Major
log(Comp_Scr_New)
log(Comp_Scr_Ong)
log(US_Perf)
Sequel
US
Germany
log(MPAA)
Tip
Children
Action
Documentary
Horror
Comedy
Other
R2
Adj. R2
Mean VIF
Max VIF
N
Germany first week
3SLS w/o
Budget
4.99
1.04
−.00
.11
.02
1.05
−.19
.29
.18
−.23
***
***
***
**
***
***
***
***
***
5.28
1.03
−.01
.13
.03
1.10
−.28
.23
.17
−.17
3SLS w/o
Ad-Exp
***
***
***
***
***
***
***
***
***
OLS
4.98 ***
1.05 ***
.00
.11 ***
1.06
−.17
.29
.18
−.22
***
***
***
***
***
3SLS
8.36
.79
.09
.16
.09
.66
−.50
.37
.30
−.01
***
***
***
***
***
***
***
***
***
−.01
−.03
−.25 *
−.28 ***
.03
−.21 *
.92
−.48 ***
−.29 ***
.04
−.18 **
.94
−.21
−.27 ***
.04
−.20 *
.92
−.19
−.21 ***
.07
−.39 ***
.95
.94
1.52
2.60
−1.66 *
.40 ***
3.05 ***
.44 ***
−8.68 ***
.57 ***
−1.73 *
.37 ***
.28
.10
−.17
.31
−1.15
.67
−.27
.07
***
***
***
***
***
***
***
.10
−.15
.35
−1.25
1.01
−.29
.06
−.02
.01
***
***
***
***
***
***
.47 ***
.12 ***
−.20 ***
−1.07 ***
.90 ***
−.01
.09
.30
.11
−.16
.29
−1.15
.82
−.24
.02
***
***
***
***
***
***
***
.26 **
.76 ***
.36 ***
.35 ***
.20 *
.99 ***
.27 **
.76 ***
−.49 ***
−.52 ***
−.28 ***
−.45 ***
.60
.55
.64
.93
.43
.17
.76
1917
**
***
***
***
***
.67
.59
.18
.69
.21
.09
.81
3460
***
***
*
***
***
.79
.50
.77
1.09
.51
.18
.73
1917
***
***
***
***
***
.60
.54
.65
.98
.43
.18
.76
.76
1.91
4.18
1917
**
***
***
***
***
5.30
1.11
.20
.12
.01
.18
−.09
.57
.21
−.17
.02
−.00
.32
−.19
.00
−.02
.04
.88
3SLS w/o
Budget
***
***
***
**
*
***
**
***
***
***
***
−2.49 ***
.29 ***
.17
.13
−.03
.06
.04
.16
−.00
.17
.02
.18
.15
.21
−.13
.08
.65
.33
−.05
.43
.17
.27
.85
1335
***
**
***
***
*
*
**
***
***
***
***
***
***
***
***
5.72
.98
.22
.22
.03
.19
−.13
.49
.32
−.16
−.05
−.02
.39
−.26
.02
−.28
−.06
.88
3SLS w/o
Ad-Exp
***
***
***
***
***
***
***
***
***
***
***
**
***
−.61
.34 ***
.29
−.02
.06
.03
.31
−.00
.17
.02
.19
.16
.03
−.07
.12
.90
.41
−.22
.47
.22
.27
.85
2598
***
***
***
**
**
***
**
*
***
***
***
***
***
***
OLS
5.31
1.13
.22
.12
***
***
***
**
.18
−.10
.56
.21
−.18
.02
.00
.32
−.19
−.01
−.03
.04
.88
***
**
***
***
***
***
−2.84 ***
.35 ***
.19 ***
.20 ***
−.01
.05
.17
−.00
.01
.03
.16
.14
.27
−.10
.07
.73
.31
−.07
.49
.16
.29
.83
1335
***
**
**
***
***
**
***
***
***
***
***
5.18
1.14
.18
.12
.01
.18
−.09
.56
.20
−.19
.01
.01
.31
−.23
−.02
−.00
.02
.88
.88
1.89
5.04
***
***
***
**
**
***
*
***
***
***
***
**
−3.36 ***
.56 ***
.09
−.02
−.07
.03
−.05
.13
.02
.06
−.01
.02
.16
.10
−.07
−.05
.41
.18
−.04
.18
.06
.11
.89
.89
1.82
5.12
1335
***
*
***
***
**
***
**
**
***
***
***
**
Note: ***p b .01; **p b .05; *p b .10 (two sided).
Note: Ad_Exp = advertising expenditure, Comp_Scr_Rev = competition for the attention of audiences, Distr_Major = major distributor, Comp_Scr_New = competition for “screen
space” from new releases, Comp_Scr_Ong = competition for “screen space” for ongoing movies, US_Perf = US market performance, Tip = movie is marked with “Tip” (in the
German magazine “Cinema”).
218
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
with E/E indicate that our findings are robust and may serve as a
foundation to generate empirical generalizations.
6. Implications and generalizations
High financial risks in production and marketing, the hedonic nature
of products, and the global cultural relevance of movies have attracted a
substantial number of researchers who have focused on a large number
of various success drivers of movies. The findings of previous research
substantially differ with respect to the included variables, the underlying
measurements, data quality, and—as a result—to the findings. Following
the taxonomy of Hubbard and Armstrong (1994), this paper provides an
extension of E/E's research and allows new insights and generalizations.
We test the conceptual relationships involved in the original study with
changes in the initial design by (1) extending the variables included in
the model, (2) using two new samples drawn from a different population, and (3) using partially different measures for the variables.
Table 7 presents an overview of the findings. Our study also helps in understanding whether the passage of time has had an effect on the results
of E/E and whether we can generalize the findings. We find substantial
differences in the effect of marketing variables with respect to the behavior of suppliers and consumers for movies, indicating that a separate
analysis of the two players is fruitful. Given our data and the robustness
of our findings, we present the following set of generalizations.
6.1. Demand generalizations
The analysis of the first week (t = 1) demand drivers in the two markets reveals insights with respect to the relative magnitude of the
elasticities. In the US market, we find that reviews and the number of
screens are by far the most important drivers for demand with elasticities of 1.05 (p b .01; reviews) and 1.04 (p b .01; screens). Next, we observe substantial seasonal primary demand effects (elasticity of .29,
p b .01) in the US market. Genre effects also show a substantial multiplier effect for horror (−48%; p b .01), action (−32%; p b .01), children
(−28%; p b .1), and other movies (−23%; p b .1). Finally, US produced
movies result in a multiplier effect of −26% (p b .01). All other drivers
are equal or below an elasticity of |.2| or a multiplier effect of |20%|.
In the German market, we identify a strong influence of screens
(1.11; p b .01) on demand in the first week. Additionally, we find that
reviews (.18 and an additional multiplier effect of 37% if the movie has
been marked as a “tip” by the leading German movie magazine Cinema;
both p b .01) and seasonality (.57; p b .01) have a strong influence on
demand. We also find sequels with a multiplier effect of 24% (p b .01)
to be of high relevance in the German market. All other drivers are
equal or below an elasticity of |.2| or a lift-up of |20%|.
For later weeks (t N 1), we find consistent results across both markets
with respect to the relative importance of the number of screens (elasticity of 1.00 in the USA and 1.03 in Germany, both p b .01) and WOM (elasticity of .85 in the USA and .86 in Germany, both p b .01). The results lead
to the following generalizations with respect to demand effects.
6.1.1. Demand generalization 1
The demand for a movie in its initial release week (t = 1) is driven
by the quality (measured using aggregated professional reviews) of the
movie. The elasticity for reviews is 1.05 in the US and .18 with an additional 37% if the movie is listed as a “tip” in Germany. The effect size for
Germany is significantly smaller (t-test; p b .01) than for the US.
Table 7
Overview of empirical findings.
Variable
Description
US demand
First week
revenues
REVENUES/ADMISSIONS
REVENUES1**
REVENUESt − 1/ADMISSONSt − 1
SCREENS
SCREENS t − 1
BUDGET
STAR
DIRECTOR
AD_EXP
REVIEWS
DISTR_MAJOR
WOM
COMP_SCR_NEW
COMP_SCR_ONG
COMP_SCR_REV
SEASON
US_PERF
US_PERF* TIME_LAG
SEQUEL
US
GERMANY
MPAA/FSK
CHILDRENa
ACTIONa
DOCUMENTARYa
HORRORa
COMEDYa
OTHERa
TIP
Weekly revenues (US)/admissions (GER)
Expected revenues first week
Revenues (US)/admissions (GER)
previous week
Weekly number of screens
Screens previous week
Production budget
Star power
Director power
Advertising expenditures
Critical reviews
Major distributor
Word-of-mouth communication
Competition for “screen space”
from new releases
Competition for “screen space”
from ongoing movies
Competition for the attention
of audiences
Seasonality
US market performance
US market performance × Time lag
between domestic and foreign release
Sequel
US production
German production
MPAA/FSK rating
Genre children
Genre action
Genre documentary
Genre horror
Genre comedy
Other genre
Movie is marked with “Tip” (Cinema)
US supply
Following
weeks
revenues
First
week
screens
GER demand
Following
weeks
screens
First week
ad-missions
GER supply
Following
weeks
ad-missions
First
week
screens
Following
weeks
screens
+
+
+
+
+
+
+
+
+
+
+
−
+
−
+
n.s.
+
+
+
+
+
+
n.s
+
n.s.
+
+
+
+
+
+
+
−
−
n.s.
−
n.s.
n.s.
+
+
−
−
−
+
+
+
+
+
−
+
+
−
+
+
n.s.
−
−
n.s.
−
−
−
−
+
+
+
+
+
n.s.
+
−
n.s.
n.s
n.s.
n.s.
n.s.
n.s.
n.s.
n.s.
+
Note: + = coefficient is positive at a significance level of p N .01; − = coefficient is negative at a significance level of p N .01; n.s. = not significant.
a
Base level in the model is DRAMA.
+
−
+
+
+
−
+
+
n.s.
+
+
+
n.s.
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
6.1.2. Demand generalization 2
The distribution of the movie, measured by its numbers of screens,
has a major influence on revenues. The screens' elasticity for the
first week is similar and not statistically different (p N .30) for the US
and Germany, with elasticities of 1.04 and 1.11, respectively. The elasticity remains very high in later periods (1.00 in the US and 1.03 in
Germany, the elasticities are statistically different; t-test; p b .01).
Thus, the major managerial challenge for studios and distributors is to
generate distribution power.
6.1.3. Demand generalization 3
The demand in later stages of a movie is strongly influenced by
WOM (elasticity of .85 in the USA and .86 in Germany, the elasticities
are statistically not different; t-test; p b .25).
6.1.4. Demand generalization 4
The elasticities for director power are .11 (USA) and .12 (Germany)
and the elasticities are statistically not different (t-test; p b .94). This
generalization is also supported by Hadida (2010).
6.2. Supply generalizations
The analysis of the supply drivers in the first week (t = 1) in both
markets provides further opportunities for generalizations. We discuss
the insights with respect to the relative magnitude of the elasticities.
In the US market, we observe a negative elasticity of reviews on screens
of − 1.15 (p b .01). Further, we find that investments in production
budgets (elasticity .28, p b .01) and advertising (elasticity .31, p b .01)
positively influence the number of screens allocated to a new movie.
In addition, we find a strong significant effect of major distributors on
screen allocation (regression coefficient of .67 resulting in a multiplier
effect of 95%, p b .01). Thus, the financial power and the large market
share of the major studios result in substantial benefits with respect to
receiving a substantial number of screens; especially for US productions
(regression coefficient of .76 resulting in a multiplier effect of 113%,
p b .01). Finally, we note that the expected revenues have a substantial
impact on screen allocation (elasticity .40, p b .01).
In the German market, we do not find any support for a negative effect of reviews on screen allocation as both variables review and tip are
insignificant. However, in line with the findings regarding the US market we identify a substantial impact of the expected admissions on
screens (elasticity .29, p b .01). Additionally, we find support for the
positive effect of financial power (production budget .17 and advertising expenditures .06, both p b .01) and market share advantages
(major distributor, regression coefficient of .16 resulting in a multiplier
effect of 17%, p b .01) on screens, although both effects are lower
compared to the effects in the US market.
For later weeks (t N 1), we find consistent results across both markets
with respect to the relative importance of screen (elasticity of .60 in the
US and .75 in Germany, both p b .01) and revenue/admission carry-over
effects (elasticity of .32 in the US and .20 in Germany). The results lead to
the following generalizations with respect to supply effects.
6.2.1. Supply generalization 1
High quality movies with excellent reviews receive substantially
fewer screens during the first week only in the USA (elasticity −1.15;
p b .01). This effect is based on the high total revenue expectations of
the cinemas for high quality movies resulting in longer expected staying
power. Due to the fact that the revenue shares for cinemas are higher in
later weeks of the movie's run, the cinemas have, ceteris paribus, an
incentive to shift demand to later weeks of the movie's run.
6.2.2. Supply generalization 2
Advertising has only limited effects on demand (elasticity: USA .02,
p b .05; Germany .01; p b .1; the elasticities are according to a t-test
statistically not different; t-test; p b .80). The limited ad effectiveness
219
might point to the existence of advertising thresholds in the market
for moviegoers. However, we find that advertising has a major influence
on supply (elasticity: USA .31, p b .01; Germany .06; p b .01; the elasticities are statistically different; t-test; p b .01). Thus, advertising serves
as an instrument targeted mainly towards the cinemas. The lower
elasticity in the German market is based on the effect that the German
sample includes all German movies, i.e. also such movies that spent
much less on advertising.
6.2.3. Supply generalization 3
Screen allocation is substantially influenced by the expected revenues (elasticity: USA .40, p b .01; Germany .29, p b .01; the elasticities
are statistically not different; t-test; p b .24). Supply in later periods
is primarily driven by carryover effects in such a way that revenues/
admissions (elasticity of .32 in the USA and .20 in Germany) and, in
particular, screens of the previous week (elasticity of .60 in the USA
and .75 in Germany) substantially influence the number of screens
provided in subsequent periods. Although the effects are statistically
different (t-test, p b .01) between the two countries, they support in
their relative importance the statement by Krider et al. (2005) that
the dominant industry pattern is one of movie exhibitors monitoring
box office sales and then responding with screen allocation decisions.
6.3. Country specific findings and generalizations
Besides the generalizations with respect to demand and supply, we
find effects in the two markets that are of general interest but that
have a lower relative importance with respect to the elasticities.
6.3.1. Generalization of sequel multiplier effect
Sequels result in higher demand (multiplier effect of 20% in the USA
and 24% in Germany) and a higher number of screens (multiplier effect
of 30% in the USA and 19% in Germany) in both markets. The estimated
demand and supply parameters do not statistically differ from each
other between the two countries.
6.3.2. Generalization of major multiplier effect
Major distributors are in a position to roll out their global market
power and receive a higher number of screens during the opening
week than other studios. The multiplier effect in the USA (95%) is
much higher than in Germany (17%). The smaller effect of majors in
markets outside of the USA has also been noted by E/E for France, the
UK, and Spain. The reason is the strong relevance of national contents
that are partially distributed by local firms that compete with the US
majors.
6.3.3. Generalization of competition
The effect of competition on demand and supply is consistently negative for the opening week. The competition of similar movies competing for the attention of audience has a higher impact in the US-market.
We find elasticities of −.19 in the US vs. −.09 in Germany. However, a
t-test shows this difference to be not significant (p = .14). Moreover,
we find different competition effects to be significant in the US and in
Germany. While in the US the competition for screen space is dominated by new releases (− .27, p b .01), in Germany only the ongoing
movies have a significant competition effect (.17, p b .1, higher scores
represent weaker competition). However, in later weeks, the effects
from competition tend to become smaller. The elasticities of the competition of similar movies for the attention of audiences amount to −.01
(p b .1) in the US vs. .03 (p b .01) in Germany, so in Germany the effect
is even slightly positive on admissions pointing towards a market extension effect of competition. Regarding the number of screens, we
find an elasticity of −.01 in the US for the competition of new releases.
In Germany both competition measures for screen space prove to be significant (new releases: − .02, p b .01; ongoing movies: .07, p b .01,
higher scores represent weaker competition).
CHILDREN
1
−.099
−.125
−.358
−.124
−.043
.205
−.076
.024
−.313
.030
ACTION
6.3.4. Generalization of star power
Stars lead to significantly more screens (elasticity .10 in the USA and
.13 in Germany, statistically not different; t-test; p b .68) in both markets. However, German moviegoers seem to be more attracted by star
power (elasticity .2) than their US counterparts where the star effect
is not significant. Contrasted with the findings on reviews, it is likely
to assume that US moviegoers rely more on reviews than on other
quality signals such as star power.
MPAA
1
.386
−.047
.260
−.216
.054
−.061
.013
.064
.355
−.249
−.141
−.190
.128
.648
.367
−.003
.244
−.286
.023
.032
.014
−.142
.449
−.229
−.099
−.231
.068
.441
.646
.568
.272
.112
.345
−.231
−.038
−.071
.009
−.096
.531
−.244
−.053
−.373
.142
1
1
.256
.420
.694
.535
.763
.743
.433
.039
.288
−.232
.099
−.034
.015
−.117
.491
−.199
−.114
−.290
.068
.827
.285
.460
.756
.561
.725
.602
.292
.101
.300
−.175
.076
.011
−.008
−.153
.607
−.239
−.087
−.401
.067
1
.949
.759
.245
.474
.696
.524
.676
.504
.280
.082
.288
−.169
.102
.033
−.016
−.291
.585
−.208
−.080
−.366
−.012
1
.100
.224
.120
.210
.198
.008
.057
.198
−.037
.077
−.051
−.012
−.015
.141
−.059
.012
−.161
.104
1
.243
.301
.460
.294
.244
.028
.092
−.041
.051
.070
−.054
−.227
.342
−.083
−.053
−.139
−.005
1
.416
.325
.281
.063
.144
−.120
.052
−.004
−.006
−.050
.322
.524
.138
.109
−.023
1
STAR
AD_EXP
BUDGET
US
SEQUAL
REVENUES1**
REVENUES
SCREENS
REVENUES
REVENUES1**
SEQUEL
US
BUDGET
AD_EXP
STAR
DIRECTOR
MPAA
CHILDREN
ACTION
DOCUMENTARY
HORROR
COMEDY
OTHER
REVIEW
DISTR_MAJOR
COMP_SCR_NEW
COMP_SCR_ONG
COMP_SCR_REV
SEASON
We thank the reviewers and the editorial team for valuable
comments throughout the review process. We also thank Alexa B.
Burmester for her excellent comments on prior versions of this paper.
Correlations of the US data set (after log transformation, significant correlations on p b .05 level are bold)
Acknowledgments
Appendix A. Correlation matrices for the US and German data
This research has been set up systematically to generate empirical
generalizations in the motion picture industry. The necessity of providing generalized findings has been emphasized by many scholars and has
led to, for example, a special edition of Marketing Science focusing on
empirical generalizations in 1995. Our findings provide such generalizations for both the demand and supply side of a major global industry in
two major markets.
DIRECTOR
Along with these generalizations, we also find interesting avenues
for further research.
(1) We strongly suggest focusing more on profit-driven analyses.
Although we are aware of production costs and advertising
budgets, we are not able to analyze whether stars or directors
are worth their specific costs because the public does typically
not know their salaries.
(2) Further, we encourage additional studies to analyze the side payments of distributors to cinemas in the profit estimation. The
profits of cinemas are also substantially driven by concession
sales (e.g., popcorn, drinks) that go to the exhibitor. The concession profit (which has been assumed to be about 1 EUR per visitor in the Netherlands) depends on the genre of movie and has
been included in prior research in optimizing movie allocation
for the Dutch cinema group Pathé (Eliashberg et al., 2001). However, the relative importance of the concession profits compared
to the overall profit generated per visitor for a cinema is not
mirrored in prior research.
(3) Although the perceived quality of a movie is often included in
empirical studies using proxies such as critical acclaim, additional details with respect to the relevance of the movie's content
would be helpful. Thus, a stronger interdisciplinary approach
using methods grounded in film or, more generally, content analysis is encouraged to extend the prior research of Eliashberg, Hui,
and Zhang (2007). Such new data should be used to engage in
new studies that explicitly study the reviews' influence on screen
allocation. Clearly, more research on this variable is needed.
(4) Although some research has been done on the optimal allocation
of movies to screens (Eliashberg et al., 2001; Swami, Eliashberg,
& Weinberg, 1999), we suggest a deeper analysis of simultaneous
decision-making processes of cinemas and studios (or distributors) with respect to advertising budgets and, more generally,
movie buzz strategies. Our interviews revealed that these issues
are of key relevance during negotiations with respect to screens.
(5) The high motivation of individuals to use hedonic goods as signals to their social system explains the high relevance of WOM
in the market. However, industry dynamics, coupled with short
life cycles of movies in the first window, provide interesting
new avenues for further research that analyzes the differences
between pre- and post-release WOM on supply and demand.
1
−.089
.102
−.222
.084
.018
−.028
−.104
.136
−.093
−.092
.055
−.052
6.4. Research implications
Note: REVENUES1** = expected revenues first week, AD_EXP = advertising expenditure, DISTR_MAJOR = major distributor, COMP_SCR_NEW = competition for “screen space” from new releases, COMP_SCR_ONG = competition for “screen
space” for ongoing movies, COMP_SCR_REV = competition for the attention of audiences.
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
1
−.100
−.026
−.033
−.094
−.032
.077
.067
−.016
.029
−.298
.057
220
Correlations of the US data set (after log transformation, significant correlations on p b .05 level are bold)
DOCUMENTATION
HORROR
COMEDY
1
−.033
−.093
−.032
.068
−.118
.070
−.001
−.011
−.042
1
−.118
−.041
−.151
−.019
−.005
.020
−.074
−.034
1
−.116
−.122
.034
−.015
.000
.072
.020
OTHER
REVIEW
DISTR_MAJOR
1
−.112
.061
.039
.050
.123
1
−.192
−.124
−.275
.090
COMP_SCR_NEW
COMP_SCR_ONG
COMP_SCR_REV
SEASON
1
.179
.007
1
−.149
1
1
.032
−.025
−.005
−.009
−.188
.012
1
.300
.539
−.083
Correlations of the German data set (after log transformation, significant correlations on p b .05 level are bold)
ADMISSIONS
SEQUEL
US
GERMANY
BUDGET
AD_EXP
STAR
DIRECTOR
MPAA (FSK)
CHILDREN
ACTION
DOCUMENTARY
HORROR
COMEDY
OTHER
TIP
REVIEWS
DISTR_MAJOR
COMP_SCR_NEW
COMP_SCR_ONG
COMP_SCR_REV
US_PERF
TL_US
SEASON
SCRE ENS
ADMISSIONS
SEQUAL
US
GERMANY
BUDGET
AD_EXP
STAR
DIRECTOR
MPAA (FSK)
CHILDREN
ACTION
DOCUMENTATION
.927
.262
.397
−.154
.664
.654
.486
.238
−.008
.181
.191
−.212
.058
−.028
.168
.152
.349
.417
−.181
−.031
−.221
.374
.196
.092
1
.272
.360
−.148
.624
.630
.510
.279
.022
.117
.180
−.198
.052
−.026
.178
.203
.408
.375
−.219
−.026
−.211
.391
.186
.145
1
.145
−.095
.193
.128
.090
.060
.069
.057
.098
−.045
.049
−.058
.111
−.032
.119
.143
−.136
−.017
−.145
.086
−.027
.023
1
−.491
.512
.236
.385
.122
.012
−.025
.107
−.115
.089
.106
.059
.042
.190
.420
−.118
.030
−.121
.596
.593
.010
1
−.367
−.083
−.111
−.023
−.071
.058
−.098
.102
−.080
−.050
−.025
.021
−.090
−.151
.072
−.010
.126
−.522
−.456
.035
1
.431
.510
.225
.024
.109
.216
−.283
−.002
−.019
.175
.064
.290
.424
−.177
.011
−.132
.484
.368
.058
1
.386
.211
.050
.074
.102
−.148
.058
−.044
.107
.127
.229
.269
−.100
−.107
−.133
.268
.145
.074
1
.377
.052
−.165
.187
−.150
−.109
.054
.090
.144
.312
.358
−.145
−.127
.020
.307
.240
.076
1
.104
−.074
.082
−.023
−.045
−.077
.103
.150
.249
.179
−.123
−.143
−.012
.126
.026
.064
1
−.325
.334
−.180
.269
−.293
.085
−.002
.093
−.001
−.024
.051
−.223
.046
−.006
−.049
1
−.138
−.043
−.065
−.159
−.083
.036
−.012
.005
.027
−.009
−.159
−.059
−.045
.041
1
−.086
−.131
−.318
−.166
−.016
.114
.071
−.110
−.024
−.202
.077
.029
−.066
1
−.041
−.100
−.052
−.007
−.089
−.137
.053
.005
−.028
−.139
−.113
−.009
221
Note: AD_EXP = advertising expenditure, Tip = movie is marked with “Tip” (in the German magazine “Cinema”), DISTR_MAJOR = major distributor, COMP_SCR_NEW = competition for “screen space” from new releases, COMP_SCR_ONG = competition for
“screen space” for ongoing movies, COMP_SCR_REV = competition for the attention of audiences, US_PERF = US market performance, TL_US = time lag between domestic and foreign release × US market performance.
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
REVENUES
REVENUES1**
SEQUEL
US
BUDGET
AD_EXP
STAR
DIRECTOR
MPAA
CHILDREN
ACTION
DOCUMENTARY
HORROR
COMEDY
OTHER
REVIEW
DISTR_MAJOR
COMP_SCR_NEW
COMP_SCR_ONG
COMP_SCR_REV
SEASON
222
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
SEASON
1
1
.043
1
−.095
−.157
−.116
.288
.234
.019
.164
−.285
.036
−.077
.240
.099
−.003
1
1
.327
.002
−.036
.034
−.036
.177
.100
.040
1
.058
.118
.103
−.044
.081
−.148
.067
−.019
.055
1
−.192
−.116
−.064
.081
−.009
.013
.156
.027
.103
−.011
REVIEW
TIP
OTHER
COMEDY
HORROR
1
−.151
−.079
−.011
−.044
.044
.019
−.040
−.478
.065
.080
−.019
ADMISSIONS
SEQUEL
US
GERMANY
BUDGET
AD_EXP
STAR
DIRECTOR
MPAA (FSK)
CHILDREN
ACTION
DOCUMENTARY
HORROR
COMEDY
OTHER
TIP
REVIEWS
DISTR_MAJOR
COMP_SCR_NEW
COMP_SCR_ONG
COMP_SCR_REV
US_PERF
TL_US
SEASON
Correlations of the German data set (after log transformation, significant correlations on p b .05 level are bold)
Appendix A (continued)
DISTR_MAJOR
1
−.092
.079
−.063
.019
.100
COMP_SCR_NEW
1
−.015
.034
.005
−.110
COMP_SCR_ONG
1
−.069
−.066
.018
COMP_SCR_REV
1
.756
.056
US_PERF
TL_US
Appendix B. Supplementary data
Supplementary data to this article can be found online at http://
www.runmycode.org.
References
Ainslie, A., Drèze, X., & Zufryden, F. S. (2005). Modeling movie life cycles and market
share. Marketing Science, 24(3), 508–517.
Albers, S. (2012). Optimizable and implementable aggregate response modeling for marketing decision support. International Journal of Research in Marketing, 29(2), 111–122.
Bass, F. M. (1995). Empirical generalizations and marketing science: A personal view.
Marketing Science, 14(3), G6–G19.
Basuroy, S., Chatterjee, S., & Ravid, A. S. (2003). How critical are critical reviews? The box office influence of film critics, star-power, and budgets. Journal of Marketing, 67, 103–117.
Boatwright, P., Basuroy, S., & Kamakura, W. A. (2007). Reviewing the reviewers: The impact of individual film critics on box office performance. Quantitative Marketing and
Economics, 5, 401–425.
Deuchert, E., Adjamah, K., & Pauly, F. (2005). For Oscar glory or Oscar money? Academy
awards and movie success. Journal of Cultural Economics, 29, 159–176.
Ehrenberg, A. S.C. (1995). Empirical generalisations, theory, and method. Marketing
Science, 14(3), G20–G28.
Elberse, A. (2007). The power of stars: Do star actors drive the success of movies? Journal
of Marketing, 71, 102–120.
Elberse, A., & Eliashberg, J. (2003). Demand and supply dynamics for sequentially released
products in international markets: The case of motion pictures. Marketing Science,
22(3), 329–354.
Eliashberg, J., Elberse, A., & Leenders, M.A. A.M. (2006). The motion picture industry:
Critical issues in practice, current research, and new research directions. Marketing
Science, 25(6), 638–661.
Eliashberg, J., Hegie, Q., Ho, J., Huisman, D., Miller, S. J., Swami, S., et al. (2009).
Demand-driven scheduling of movies in a multiplex. International Journal of
Research in Marketing, 26(2), 75–88.
Eliashberg, J., Hui, S. K., & Zhang, Z. J. (2007). From story line to box office: A new
approach for green-lighting movie scripts. Management Science, 53(6), 881–893.
Eliashberg, J., Jonker, J. -J., Sawhney, M. S., & Wierenga, B. (2000). MOVIEMOD: an
implementable decision-support system for prerelease market evaluation of motion
pictures. Marketing Science, 19(3), 226–243.
Eliashberg, J., & Shugan, S. M. (1997). Film critics: Influencers or predictors? Journal of
Marketing, 61, 68–78.
Eliashberg, J., Swami, S., Weinberg, C. B., & Wierenga, B. (2001). Implementing and evaluating SilverScreener: A marketing management support system for movie exhibitors. Interfaces, 31(3), S108–S127.
Fornell, C. (1995). The quality of economic output: Empirical generalizations about its distribution and relationship to market share. Marketing Science, 14(3), G203–G211.
Goldenberg, J., & Muller, E. (2012). Editorial. International Journal of Research in Marketing,
29(4), v–vii.
Greene, W. H. (2006). Econometric analysis (6th ed.)Upper Saddle River, NJ: Prentice Hall.
Hadida, A. (2010). Commercial success and artistic recognition of motion picture. Journal
of Cultural Economics, 34(1), 45–80.
Hanssens, D. (2009). Empirical generalizations about marketing impact. Cambridge, MA:
Marketing Science Institute (MSI) Relevant Knowledge Series.
Hennig-Thurau, T., Henning, V., Sattler, H., Eggers, F., & Houston, M. B. (2007). The last picture
show? Timing and order of movie distribution channels. Journal of Marketing, 71, 63–83.
Hennig-Thurau, T., Houston, M. B., & Heitjans, T. (2009). Conceptualizing and measuring
the monetary value of brand extensions: The case of motion pictures. Journal of
Marketing, 73(6), 167–183.
Hennig-Thurau, T., Völckner, F., Clement, M., & Hofmann, J. (2013a). An ingredient branding approach to determine the financial value of stars: The case of motion pictures.
Available at SSRN: http://ssrn.com/abstract=1763547
Hennig-Thurau, T., Wiertz, C., & Feldhaus, F. (2013b). Does twitter matter? An investigation of the impact of microblogging word of mouth on consumers' adoption of new
products. Available at SSRN: http://ssrn.com/abstract=2016548
Hubbard, R., & Armstrong, J. S. (1994). Replications and extensions in marketing-rarely
published but quite contrary. International Journal of Research in Marketing, 11(3),
233–248.
Joshi, A.M., & Hanssens, D.M. (2009). Movie advertising and the stock market valuation of
studio: A case of “great expectations”? Marketing Science, 28(2), 239–250.
Kamakura, W. A., Basuroy, S., & Boatwright, P. (2006). Is silence golden? An inquiry into
the meaning of silence in professional product evaluations. Quantitative Marketing
and Economics, 4(2), 119–141.
Karniouchina, E. V. (2011). Impact of star and movie buzz on motion picture distribution and box office revenue. International Journal of Research in Marketing,
28(1), 62–74.
King, T. (2007). Does film criticism affect box office earnings? Evidence from movies
zreleased in the U.S. in 2003. Journal of Cultural Economics, 31(3), 171–186.
Krider, R. E., Li, T., Liu, Y., & Weinberg, C. B. (2005). The lead-lag puzzle of demand and distribution: A graphical method applied to movies. Marketing Science, 24(4), 635–645.
Leenders, M.A. A.M., & Eliashberg, J. (2011). The antecedents and consequences of restrictive age-based ratings in the global motion picture industry. International Journal of
Research in Marketing, 28(4), 367–377.
Litman, B. R. (1983). Predicting success of theatrical movies: An empirical study. Journal of
Religion and Popular Culture, 16, 159–175.
M. Clement et al. / Intern. J. of Research in Marketing 31 (2014) 207–223
Liu, Y. (2006). Word of mouth for movies: Its dynamics and impact on box office revenue.
Journal of Marketing, 70, 74–89.
Moon, S., Bergey, P. K., & Iacobucci, D. (2010). Dynamic effects among movie
ratings, movie revenues, and viewer satisfaction. Journal of Marketing, 74,
108–121.
MPAA (2013). State-by-State Film & Television Economic Contribution. http://www.
mpaa.org/policy/state-by-state last accessed on December 18th 2013.
Radas, S., & Shugan, S. M. (1998). Seasonal marketing and timing new product introductions. Journal of Marketing Research, 35, 296–315.
Ravid, A. S. (1999). Information, blockbusters, and stars: A study of the film industry.
Journal of Business, 72(4), 463–492.
223
Sawhney, M. S., & Eliashberg, J. (1996). A parsimonious model for forecasting gross
box-office revenues of motion pictures. Marketing Science, 15(2), 113–131.
Swami, S., Eliashberg, J., & Weinberg, C. B. (1999). SilverScreener: A modeling approach to
movie screens management. Marketing Science, 18(3), 352–372.
van Bruggen, G. H., Lilien, G. L., & Kacker, M. (2002). Informants in organizational marketing research: Why use multiple informants and how to aggregate responses. Journal
of Marketing Research, 39, 469–478.
Wallace, T. W., Seigerman, A., & Holbrook, M. B. (1993). The role of actors and actresses in
the success of films: How much is a movie star worth? Journal of Cultural Economics,
17(1), 1–27.
Winer, R. S. (1998). From the Editor. Journal of Marketing Research, 35, iii–v.
Intern. J. of Research in Marketing 31 (2014) 224–238
Contents lists available at ScienceDirect
Intern. J. of Research in Marketing
journal homepage: www.elsevier.com/locate/ijresmar
Full Length Article
Drivers of the cost of capital: The joint role of non-financial metrics
Alexander Himme a,⁎, Marc Fischer a,b,1
a
b
University of Cologne, The Faculty of Management, Economics, and Social Sciences, Chair for Marketing and Market Research, Albertus-Magnus-Platz 1, D-50923 Cologne, Germany
UTS Business School, Sydney, Australia
a r t i c l e
i n f o
Article history:
First received in 1 June 2011 and was under
review for 9 months
Available online 2 December 2013
Area Editor: Koen H. Pauwels
Guest Editor: Marnik G. Dekimpe
Keywords:
Cost of capital
Stock market beta
Credit spreads
Brand value
Customer satisfaction
Corporate reputation
a b s t r a c t
Recent marketing studies suggest that non-financial metrics, such as customer satisfaction and brand value, help
explain the variation in the cost of equity and the cost of debt. These studies typically focus on only one nonfinancial metric and one component of capital cost. In this study, we broaden the understanding of the relevance
of non-financial metrics to the cost of capital. We investigate the joint role of customer satisfaction, brand value,
and corporate reputation for stock market beta and credit ratings, which reflect variation in equity and debt risk
premiums across firms. In addition to the joint direct influence of these metrics on capital cost, we also study their
interaction effects. We develop a conceptual model to explain the effects on capital costs and test the resulting
hypotheses in a broad sample of 344 firms from diverse industries using data from the 1991–2006 period.
Our results suggest that higher satisfaction ratings reduce both the cost of equity and cost of debt, whereas
brand value and corporate reputation only show a negative direct association with the cost of debt. In addition,
both measures moderate the effect of satisfaction on the cost of debt. Brand value attenuates the influence of
satisfaction, whereas corporate reputation amplifies this effect.
© 2013 Elsevier B.V. All rights reserved.
1. Introduction
The weighted average cost of capital (WACC) is an important
financial metric relevant both to members of the financial community,
such as institutional investors, and to the top management of (publicly
listed) firms. Given a stream of future cash flows, a lower WACC indicates a higher present value of that stream. For management, a lower
WACC constitutes lower hurdle rates for investment projects because
investors require less return from the according capital expenditures.
WACC is composed of equity cost and debt cost. Both providers of
capital demand a return for their investment. The larger the risk that
they perceive to be associated with the investment, the higher the
required return. The most important measure for equity holder risk is
systematic risk, whereas credit ratings are the best signal for debt
holders with respect to the default risk of a firm (Brealey, Myers, &
Allen, 2007).
Systematic risk and default risk vary across companies and over
time. The extant accounting/finance literature has thus addressed the
natural question regarding the drivers of such risks (e.g., Beaver,
Kettler, & Scholes, 1970; Blume, Lim, & MacKinlay, 1998). Most studies
focus predominantly on “hard” financial metrics, such as operating
margins, asset growth, leverage, and earnings variability, which are
commonly documented in financial reports or can be derived from
⁎ Corresponding author. Tel.: +49 221 470 8679; fax: +49 221 470 8677.
E-mail addresses: himme@wiso.uni-koeln.de (A. Himme),
marc.fischer@wiso.uni-koeln.de (M. Fischer).
1
Tel.: +49 221 470 8675; fax: +49 221 470 8677.
0167-8116/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijresmar.2013.10.006
corporate or analyst disclosures. Researchers have found that several financial variables serve as drivers of the costs of equity and debt; however, they also acknowledge that their models explain only a fraction of
the observed variance in capital cost (e.g., Elton, Gruber, Agrawal, &
Mann, 2001). Several authors believe that so-called soft or intangible,
non-financial metrics, such as management capabilities and marketing
metrics, contribute to explaining the residual variance (e.g., Blume
et al., 1998; Pinches & Mingo, 1973).
An emerging research stream on the interface between accounting/
finance and marketing provides evidence for the value relevance of
marketing metrics. In particular, recent efforts demonstrate that advertising expenditures, brand value, customer satisfaction, and corporate
social responsibility possess the power to lower the cost of capital (for
an overview, see Srinivasan & Hanssens, 2009). However, all these
studies investigate only a single non-financial driver of capital cost.
We believe that marketing-related non-financial metrics may offer
different informational value for investors and creditors. As a result,
such metrics may impact capital costs above and beyond each other.
Measures such as customer satisfaction, brand value, and corporate
reputation reflect competitive advantages from different domains. Satisfaction focuses on the customer, brand value focuses on the product,
and corporate reputation emphasizes the firm. Therefore, these measures provide different signals to investors regarding the financial
health of a firm that eventually influence the cost of debt and equity.
This study attempts to provide several contributions. First, we
investigate the joint role of the common non-financial measures of
customer satisfaction, brand value, and corporate reputation in the
cost of capital. We call these measures “non-financial” because they
225
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
inform investors about the quality of marketing and management capabilities although they may be measured in monetary units (e.g., brand
value). Specifically, we consider the popular and publicly available
American Customer Satisfaction Index ratings, the financial brand
values by Interbrand, and Fortune's corporate reputation scores. We
develop a novel conceptual model of the informational value and signals
contained in these metrics. From this model, we derive hypotheses
regarding the incremental contribution of each metric in explaining
the risk components of the cost of capital. In addition, we suggest potential moderating effects. Specifically, we suggest that brand value and
corporate reputation moderate the influence of customer satisfaction
on the cost of capital. Customer satisfaction plays this central informational role because it reflects customer experiences with past transactions (Fornell, Johnson, Anderson, Cha, & Bryant, 1996). Financial
accounting is transaction-based and emphasizes historical earnings,
which contain information with the highest certainty level (Kothari,
2001). Brand value and corporate reputation are less transactionbased and rather provide information on a firm's potential for future
growth. Therefore, these information signals influence the interpretation and processing of satisfaction ratings by investors.
Second, we test the hypotheses in a broad sample of 344 firms from
diverse industries in the 1991–2006 period. Our analysis accounts for
the dynamics and the potential endogeneity of our focal non-financial
metrics. Including all three metrics together in the empirical models of
equity cost and debt cost enables us to quantify the relative effect of
each of the measures above and beyond each individual metric. For
managers and investors, it is important to know whether satisfaction
ratings, brand value, and corporate reputation scores provide additional
distinct information. If not, investors and managers could simply substitute one non-financial metric for another to evaluate risk potential.
Third, given that the focal metrics are measured at different scales, it
is difficult to compare their relative importance in driving the cost of
capital. Hence, we transform the estimated coefficients into elasticity
estimates. This study is among the first to calculate elasticities for the
effects of non-financial metrics on the components of capital costs.
These elasticities enable managers and investors to assess precisely
how changes in non-financial metrics influence the cost of capital. In
addition, the results enable us to conduct meta-analyses.
This paper is organized as follows. We briefly discuss the related
literature in the next section. Subsequently, we provide details
about the conceptualization of our key variables, which is important to assess their informational value. In Section 4, we derive
our hypotheses. The next section includes the empirical study and
the estimation results. We discuss these results in the final section
and finish by presenting the conclusions and limitations of our
study.
2. Literature background
In Table 1, we briefly review the related accounting, finance, and
marketing literature. From the marketing literature, we include all
studies that consider either systematic risk (equity cost) or default
risk (debt cost) as a dependent variable and non-financial metrics as
an independent variable.
2.1. Accounting and finance literature
The extant literature examines the effects of various factors on
systematic risk and the cost of equity. Beaver et al. (1970) provide one
of the first contributions within this field of research. Their model
relates systematic risk (measured by beta) to variables that describe
the financial position of a firm. The authors find that greater systematic
risk is related to lower dividend payout, higher growth, smaller asset
size, and greater leverage. Subsequent studies (e.g., Hill & Stone, 1980)
consider similar variables and support the results obtained by Beaver
et al. (1970).
The research of Horrigan (1966) is among the first studies to analyze drivers of credit ratings that reflect the terms of debt financing.
He considers different financial variables (e.g., total assets) to predict
corporate bond ratings. Kaplan and Urwitz (1979) use an ordered
probit model to predict bond ratings. The authors find, as an example, that total assets, the ratio of long-term debt to total assets, and
the stock market beta are relevant. Blume et al. (1998) extend the
approach by analyzing a panel of firms in the 1978–1995 period.
These researchers introduce new variables, such as pretax interest
coverage. We adopt the widely used models by Beaver et al. (1970)
and Blume et al. (1998) as baseline specifications that we extend
using our focal non-financial metrics.
Table 1
Sample of prior research on drivers of the cost of capital.
Author(s)
Accounting/financial variables
Non-financial (marketing) metrics
Advertising
Studies with focus on accounting/financial variables
Beaver, Kettler, and Scholes (1970)
Blume, Lim, and MacKinlay (1998)
Horrigan (1966)
Kaplan and Urwitz (1979)
Pinches and Mingo (1973)
Studies with focus on non-financial (marketing) variables
Agarwal and Berens (2009)
Anderson and Mansi (2009)
Bharadwaj, Tuli, and Bonfrer (2011)
Fornell, Mithas, Morgeson, and Krishnan (2006)
Gruca and Rego (2005)
Johansson, Dimofte, and Mazvancheryl (2012)
Luo, Homburg, and Wieseke (2010)
Madden, Fehle, and Fournier (2006)
McAlister, Srinivasan, and Kim (2007)
Orlitzky and Benjamin (2001)
Osinga, Leeflang, Srinivasan, and Wieringa (2011)
Rego, Billett, and Morgan (2009)
Singh, Faircloth, and Nejadmalayeri (2005)
Tuli and Bharadwaj (2009)
This study
a
b
Brand value
Satisfaction
Cost of capital
Reputation
✓
✓
✓
✓
✓
Debt
✓
✓
✓
✓
✓
(✓)a
✓
✓
✓
✓
✓
✓
(✓)b
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Equity
✓
(✓)a
✓
✓
✓
✓
✓
Authors only investigate one dimension of corporate reputation, which is corporate social responsibility.
Authors investigate one dimension of brand value, which is brand quality.
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
226
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
2.2. Marketing literature
Non-financial metrics provide current and forward-looking information above and beyond the “hard” information that is contained in a
firm's financial statements. Five studies (see Table 1 again) focus on
the effect of customer satisfaction on one component of the cost of capital (stock market beta or credit spreads). The results of these studies
provide strong evidence that customer satisfaction reduces systematic
risk (stock market beta) and leads to better credit ratings.
Compared with customer satisfaction, the effect of brand value on
the components of capital costs is ambiguous. Madden et al. (2006)
and Rego et al. (2009) report that stronger brands improve credit
ratings and lower systematic risk. By contrast, Bharadwaj et al. (2011)
find a positive relationship between strong brand quality and systematic
risk. Johansson et al. (2012) conclude that top brands as measured by
financial brand value (Interbrand) did not show lower systematic risk
than the market as a whole during the stock market downturn in the
fall of 2008. However, brands scoring the highest on a consumerbased brand equity measure (EquiTrend) have lower systematic risk.
To the best of our knowledge, prior research on the effects of
corporate reputation on capital cost is not available. Nevertheless,
some studies focus on corporate social responsibility, although this
dimension is only one of several that contribute to the overall reputation of a company. Orlitzky and Benjamin (2001) as well as Agarwal
and Berens (2009) show that higher corporate social responsibility is
associated with lower financial risk and lower capital costs in general.
Collectively, prior studies provide stronger and clearer evidence
with regard to customer satisfaction compared with brand value and
corporate social responsibility. The findings pertaining to the role of
brands with respect to systematic risk are inconsistent. We are not
aware of any studies that investigate the relationship between corporate reputation and the components of capital cost. As a consequence,
we hope that our joint consideration of all three metrics contributes to
providing insight into their role as drivers of the cost of capital.
3. Conceptualization and measurement of key variables
Customer satisfaction, brand value, and corporate reputation are
multi-dimensional constructs that are not directly observable. Our
hypotheses regarding their influence on capital cost are based on the
distinct informational value that these metrics provide for investors.
Different approaches have been suggested for measuring these constructs. It is beyond the scope of this paper to discuss these approaches
in detail, but it is important to understand the conceptual foundation of
the specific measures that we use in this study.
Following the idea of efficient capital markets, we selected measures
that are publicly available, consistently measured over time, and widely
appreciated by investors. Three measures fulfill these criteria: the
American Customer Satisfaction Index (ACSI), Interbrand's brand
value measure, and Fortune's corporate reputation index. Following
the finance literature, we use credit spreads and the stock market beta
to measure the risk components of capital cost that are responsible for
company-specific differences in the cost of debt and equity.
3.1. Credit spread (default risk)
For debt holders, the default risk of a firm is the most relevant
(Blume et al., 1998). Consistent with the literature (Brealey et al.,
2007; Ederington, Yawitz, & Roberts, 1987), we measure default risk
in terms of credit spreads that are closely related to the credit ratings
that are issued by rating agencies, such as Standard & Poor's (S&P).
S&P defines credit ratings as follows: “[…] ratings express the agency's
opinion about the ability and willingness of an issuer … to meet its
financial obligations in full and on time” (Standard & Poor's, 2011a, 3).
Typically, analysts obtain information from published reports and financial statements as well as from interviews with the issuer's management.
Analysts use such information to assess an entity's financial condition
and risk potential. In fact, credit analysts use varying criteria in the rating
process. Two risk components lead to the final credit rating. First, credit
analysts assess the financial risk of a firm by evaluating “hard” financial
metrics, such as capital structure or profitability. Second, analysts assess
a firm's business risk by considering the company's market position,
cost efficiency, and management and marketing capabilities (Standard
& Poor's, 2011a). Relevant factors may include market share, which reflects a firm's market position and its ability to sustain or increase shares;
the strength of the brand; the degree of operating efficiency; and
management's track record of product innovation and brand building, including the efficiency and effectiveness of marketing spending (Standard
& Poor's, 2011b). All this information is non-financial by nature and relates to marketing capabilities.
3.2. Stock market beta (systematic risk)
Following the Capital Asset Pricing Model (e.g., Brealey et al., 2007),
the covariance between firm i's stock return, ri, and the market return,
rm, relative to the variance of the market return, σ 2rm , measures the
systematic risk or stock market beta, respectively. Using the identity Cov
ðr i ; rm Þ ¼ ρri ;rm σ ri σ rm , we can also write the following equation for beta:
BETAi ¼
ρri ;rm σ ri σ rm
σ 2rm
¼ ρri ;rm
σ ri
σ rm
;
ð1Þ
where ρri ;rm measures the correlation between returns and σ ri and σ rm
are the associated standard deviations. The correlation coefficient
measures how closely firm returns follow the overall market trend.
For example, insurance companies and banks depend quite heavily on
the business cycle, whereas pharmaceutical companies are not greatly
affected by market trends. Hence, ρri ;rm rather reflects differences across
industries but does not vary greatly over time for a single firm. However,
the ratio of the standard deviations of firm and market returns, σ ri /σ rm ,
captures firm-specific differences that also vary over time. Thus, we
focus on this ratio when developing our arguments for the associated
hypotheses. Consistent with fair value theory, stock returns and
earnings/cash flows are highly correlated (Brealey et al., 2007).
A major driver of the variation of cash flows is their growth rate.
Fischer, Leeflang, and Verhoef (2010) provide a formal proof for this
fundamental relationship. Empirical research on capital markets consistently indicates that growth stocks are indeed associated with a higher
beta (e.g., Fama & French, 1992). In the subsequent development of
our hypotheses, we refer to expectations regarding firm growth rates
relative to the market average.
3.3. Customer satisfaction
An individual firm's customer satisfaction represents its current
customers' overall evaluation of their total purchase and consumption
experiences (Fornell et al., 1996). Thus, customer satisfaction is an indicator of the loyalty and the willingness to pay of current customers;
thus, it provides information related to revenues from the current customer base (Anderson, Fornell, & Lehmann, 1994). We use customer
satisfaction ratings (ACSI) from the National Quality Research Center
at the University of Michigan. Fornell et al. (1996) provide details on
how ACSI is measured.
3.4. Brand value
Brand value measures the incremental discounted future cash flows
accruing from a branded product compared with an identical but
unbranded product (e.g., Johansson et al., 2012). We use brand value
data from the Interbrand Group (see www.interbrand.com for a
detailed description). One specific characteristic of the Interbrand
approach is that it forecasts the current and future revenues that
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
are specifically attributable to the branded products. The costs of
conducting business (e.g., operating costs) and intangibles, such as
patents and management strength, are subtracted to assess what
portion of the earnings results from the brand. Interbrand calculates a
brand strength score to measure a brand's ability to secure ongoing
customer demand (e.g., loyalty, retention) and thus sustain future
earnings, translating branded earnings into net present value. In general,
the Interbrand measure is used to reflect the revenue and profit growth
potential of a firm as a result of brand strength.
3.5. Corporate reputation
Following Fombrun (1996, 72), we define corporate reputation as “a
perceptual representation of a company's past actions and future prospects that describe the firm's overall appeal to all its key constituents
when compared to other leading rivals”. Corporate reputation is likely
the most complex metric among our focal metrics. In general, corporate
reputation comprises the credibility and respect that an organization
has among a broad set of constituents (e.g., employees, investors,
regulators, customers). In this study, we follow Fortune's approach to
measure corporate reputation (Fombrun & Shanley, 1990; Fortune,
2009). A firm's overall reputation score is built from ratings of eight
dimensions: financial soundness; innovation; long-term investment;
the ability to attract, develop, and retain talented people; product/
service quality; the quality of management; social responsibility; and
the wise use of corporate assets. The Fortune measure reflects the
potential of a firm to increase its future revenues and operational
efficiency.
4. Hypotheses
4.1. Conceptual model
Fig. 1 shows the conceptual model that underlies our hypotheses.
Information regarding customer satisfaction, brand value, and corporate
reputation is assumed to affect the cost of capital via the risk components beta and credit spreads. In this framework, we assume that
customer satisfaction plays a prominent role in both beta and credit
spreads.
Investors prefer clear, certain, and unambiguous information
regarding the earnings power of companies. As a consequence, accounting earnings measurement rules place great emphasis on transactionbased revenue recognition (Kothari, 2001). Our three non-financial
metrics contain information about both past transactions and potential
future revenues and profits. Therefore, we suggest that these metrics
have a direct effect on beta and credit spreads. For example, reduced
customer satisfaction leads to customer defection in the long term,
which in turn affects firm revenue. Brand value and corporate reputation provide additional information regarding the quality of products
and services and the efficiency of firms. A decline in brand value and/
or corporate reputation is also expected to influence customer purchase
decisions and thus the revenue base of a firm. A decline in customer
satisfaction, brand value, and corporate reputation would influence
the expectations of a firm's stakeholders about current and future
earnings. Because investors immediately react to changes in their
expectations, we postulate a direct current effect of the non-financial
metrics on the cost of capital. Because economic information often
unfolds its full meaning only over time, we also assume carryover effects
in our framework.
In addition, we postulate that brand value and corporate reputation
moderate the influence of customer satisfaction on the cost of capital. By
definition, customers report their satisfaction based on past transactions. Hence, customer satisfaction has the closest link to past transactions among the three non-financial metrics, which suggests that it
plays a central role in our framework. Brand value and corporate reputation expand the information set of investors with additional signals
227
regarding the future earnings potential of firms. These information signals influence the interpretation and processing of satisfaction ratings
by investors and thus moderate the relationship between satisfaction
and cost of capital.
4.2. Information content of satisfaction, brand value, and
corporate reputation
The three non-financial metrics contain information to assist
investors in assessing future risk. The information content of the three
metrics overlaps, but each metric provides unique information.2
Because of this unique information, we believe that each metric has
incremental value for investors in evaluating the potential risk of an
investment in a specific company. Table 2 summarizes the differences
in the information value of the non-financial metrics.
Customer satisfaction signals existing customers' loyalty and willingness to pay (e.g., Anderson et al., 1994). Hence, investors make inferences about revenues and cash flows that stem from existing customers
in the future.
Brand value informs about the strength of a brand. This strength
emanates to a great extent from the innovativeness and the potential
to grow with existing and new products in existing and new markets
(e.g., Barth, Clement, Foster, & Kasznik, 1998; Leone et al., 2006). In
addition, brand value signals how familiar investors are with a firm
(e.g., Rego et al., 2009).
Corporate reputation provides additional non-market-based information that reflects within-firm characteristics (Fombrun & Shanley,
1990). Six of the eight dimensions of Fortune's reputation metric focus
on internal firm processes. Financial soundness and the wise use of
corporate assets provide signals about corporate cost management
and operational efficiency (Fombrun, 1996). In addition, the metric
informs about the quality of management and employees.
4.3. Hypotheses on credit spreads
We begin this subsection by discussing the potential influence of
customer satisfaction, brand value, and corporate reputation on credit
spreads, followed by a discussion of the effects on beta.
Finance research has shown that firms are less able to service their
debt obligations when suffering from higher equity risk as measured
by the stock market beta (e.g., Blume et al., 1998). A higher beta reflects
more vulnerable and volatile cash flows relative to the market average.
The default risk of a firm increases, and the risk premium or credit
spreads, respectively, for corporate bonds consequently increase as
well. Our empirical model to explain variation in credit spreads includes
beta as an important predictor. In the following, we focus on developing
hypotheses regarding the direct influence of satisfaction, brand value,
and reputation on credit spread above and beyond the mediated influence via beta, which we discuss subsequently.
4.3.1. Customer satisfaction
Customer satisfaction positively influences the willingness to pay of
customers while also reducing behaviors with negative economic consequences for firms, such as complaints (e.g., Anderson & Mansi,
2009). Satisfied customers are more likely to buy more of the same
product, to buy additional products, and to make recommendations to
other customers (e.g., Anderson, Fornell, & Mazvancheryl, 2004). Customer satisfaction ratings influence the credit rating process by providing information regarding the behavior of current customers that
determines the size of firm profits (Anderson & Mansi, 2009). Firms
with a higher level of expected cash flows ensure payment and are
viewed as less risky borrowers. Thus, we propose the following
hypothesis:
2
We would like to thank the AE for stimulating the following discussion.
228
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
: Direct influence
: Moderating influence
Fig. 1. Conceptual model.
H1. Customer satisfaction is negatively associated with credit spreads.
4.3.2. Brand value
Strong brands are a signal for excellent marketing, which credit rating
agencies consider to be an important criterion in their rating process
(Standard & Poor's, 2011b). In addition to customer satisfaction, brand
value offers potential to grow the customer base by acquiring new
customers. These growth opportunities provide signals about a firm's capability of generating additional sales in the future to help fulfill its
liabilities. Moreover, it is well known that a significant proportion of a
firm's market value lies in intangible, off-balance sheet assets, such as
brands (e.g., Bahadir, Bharadwaj, & Srivastava, 2008). Brands may serve
as an elementary security for debt holders in case of a firm's financial distress or even bankruptcy. Finally, brands facilitate the access to fresh capital from equity investors, which again reduces the likelihood of financial
distress. For example, investor funds from Abu Dhabi and Qatar provided
fresh equity capital to Daimler and Porsche in 2009 when the cash holdings were tight as a result of the deep financial crisis. Both investor funds
cited the strength of the premium car brands among the reasons for their
investment decisions. Hence, we hypothesize as follows:
H2. Brand value is negatively associated with credit spreads.
4.3.3. Corporate reputation
Corporate reputation leads to greater familiarity of debt holders and
credit rating agencies with a firm (e.g., Fombrun, 1996). Note that
corporate reputation offers unique information signals with regard to
operational efficiency and the quality of management. Naturally, operational efficiency is an important driver of firm profitability and thus of a
firm's ability to fulfill future liabilities (e.g., Singh et al., 2005). The quality of management and employees is also a positive signal for credit rating agencies because it reduces the likelihood of a situation of financial
distress (e.g., Blume et al., 1998). Well-known companies are generally
more successful in attracting and retaining better employees, who are in
turn more productive (Luo & Bhattacharya, 2006). Thus, corporate reputation provides positive signals to credit rating agencies that assess the
credit worthiness of a firm. Standard and Poor's (2011b) mentions the
degree of operating efficiency and management's track record of product innovation among their top factors in providing credit ratings.
Therefore, we hypothesize as follows:
H3. Corporate reputation is negatively associated with credit spreads.
4.3.4. Relative strength of effects
All three non-financial metrics offer unique information value for
the credit rating process. Therefore, we expect that each of these metrics
influences credit spreads. However, considering their different information signals, we assume that these metrics exert effects of different
strength. Corporate reputation is measured across a diverse range of
dimensions involving a broad set of constituents (see Sections 3 and
4.2). In addition, corporate reputation particularly emphasizes the
financial soundness and operational efficiency of a firm. Compared
Table 2
Main information content of focal non-financial metrics.
Main information content (signal)
Loyalty of existing customers
Willingness to pay by existing customers
Potential to grow with product/services
into new markets/customer segments
Innovativeness
Familiarity with product and firm
Operational efficiency
Quality of management and employees
(X) means limited signaling content.
Construct (measure)
Customer satisfaction (ACSI)
Brand value
(Interbrand financial brand value)
X
X
(X)
(X)
X
X
X
Corporate reputation
(Fortune reputation index)
Customer focus
Product focus
X
(X)
X
X
Firm focus
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
with brand value and customer satisfaction ratings, it may be more difficult to improve corporate reputation, as it requires advancements across
several dimensions simultaneously. Credit rating agencies consider all
these dimensions in their rating process and combine them with their
evaluation of a firm's financial soundness. We therefore assume the relative responsiveness (elasticity) of credit spreads to be higher for
corporate reputation than for brand value and customer satisfaction.
Thus, we formulate the following hypothesis:
H4. Compared with brand value and customer satisfaction, corporate
reputation has the strongest negative effect on credit spreads.
Consistent with our conceptual model in Fig. 1, we believe that
both brand value and corporate reputation also moderate the role of
customer satisfaction ratings in the credit rating process.
4.3.5. Moderating effect of brand value
Competing arguments regarding the moderating effect of brand
value can be made: the uncertainty argument and the price premium
argument.
4.3.5.1. The uncertainty argument. Signals from customer satisfaction ratings may be less informative for firms with strong brands. Customer satisfaction ratings are the results of past customer transactions, whereas
brand value informs about the potential to grow a business with revenues from new customers. Hence, brand value informs about a second
source of future revenues that is not fully reflected in satisfaction ratings
from current customers. In addition, future growth from new customers
has a side effect, as it increases uncertainty about the exact level of future cash flows, which is important to evaluate a firm's potential to service its debt obligations. This uncertainty also makes the information
signal from customer satisfaction less powerful. Thus, we hypothesize
as follows:
H5. Brand value attenuates the negative effect of customer satisfaction on
credit spreads.
4.3.5.2. The price-premium argument3. Higher brand value is not only
viewed by consumers as a signal of higher price but also associated
with higher prices (Bharadwaj et al., 2011). If current customers are satisfied, then potential new customers can be expected to show equally
high customer satisfaction. As a result, new customers will be more likely to be willing to pay a price premium and contribute to generating increased revenues. The price premium information provided by high
brand value can make the information signal from customer satisfaction
for credit rating agencies more powerful. Thus, we propose the following hypothesis:
H5(alt). Brand value amplifies the negative effect of customer satisfaction
on credit spreads.
4.3.6. Moderating effect of corporate reputation
The perceived quality of a product or service is a primary driver of
customer satisfaction (Fornell et al., 1996). As a result, companies invest
heavily in implementing systems for customer relationship management
or total quality management to increase customer satisfaction (Anderson
et al., 1994; Mithas, Krishnan, & Fornell, 2005). Naturally, firms differ
with regard to the efficiency of such investments (e.g., Anderson et al.,
1994). In particular, corporate reputation provides information regarding the financial soundness and operational efficiency of a firm. Hence,
the information signal from a firm's customer satisfaction rating becomes more valuable to credit rating agencies if the firm is known for
3
We would like to thank the editor for suggesting this alternative hypothesis.
229
its operational efficiency and financial soundness (Standard & Poor's,
2011b). This increased information value implies that such a firm generates its revenues from satisfied customers at lower costs. Accordingly,
credit rating agencies may evaluate the same customer satisfaction for
firms differently depending on their reputation for operational efficiency. Therefore, we hypothesize as follows:
H6. Corporate reputation amplifies the negative effect of customer
satisfaction on credit spreads.
4.4. Hypotheses on stock market beta
4.4.1. Customer satisfaction
Several studies have shown that customer satisfaction enhances
customer retention and therefore contributes to reducing the volatility
and vulnerability of future cash flows (e.g., Gruca & Rego, 2005). Customers are thus more committed to the firm and less likely to switch
to other firms. In periods of cyclical downturn, cash flows are cushioned
from the downward trend. In upswing periods, the firm probably does
not grow as fast as other companies that lost customers and now
expand with the market. As a result, the systematic risk for firms with
more satisfied customers is lower. Because prior research (e.g., Fornell
et al., 2006; Tuli & Bharadwaj, 2009) provides strong support for this
relationship, we do not repeat the arguments in detail here.
H7. Customer satisfaction is negatively associated with systematic risk
(stock market beta).
4.4.2. Brand value
A strong brand acts as a barrier to competition and increases the
probability of a customer continuing to purchase the brand (McAlister
et al., 2007). The perceived value of a brand prevents customers from
brand switching even if remaining with a certain brand requires
paying a price premium (Rego et al., 2009). Higher brand value results
from higher awareness, which in turn reduces consumer search costs
and facilitates repeat purchases (Johansson et al., 2012). These forces
strengthen the cash flow basis. Compared with average performers,
this basis is less likely to erode for strong brands during an economic
downturn when demand is shrinking. However, it has also been argued
that the opposite is true. According to Bharadwaj et al. (2011), consumers view high brand quality as a signal of high prices. As consumers
become more price-conscious in a downturn, strong brands may lose
market share more rapidly than weaker but less expensive brands. As
a result, cash flows and thus stock returns decline more rapidly than
the market average.
During an economic upswing, strong brand value signals faster
growth. Strong brand value is an indicator of a firm's cross-selling
potential, and consumers of strong brands are more likely to increase
purchases in the future (Rego et al., 2009). These benefits imply that
in an upswing situation, firms with higher brand value may outperform
the market average. Although faster growth in cash flows is positive, it
has a side effect of involving a higher variance of cash flows (Fischer
et al., 2010), i.e., increases in beta (see Eq. (1)). Consequently, arguments favoring both positive and negative relationships between
brand value and beta exist. We leave this question to be solved by the
empirical analysis.
4.4.3. Corporate reputation
Corporate reputation provides insight into the operational efficiency
of a firm and the quality of its management and employees (see
Table 2). During a market downturn, an efficient company is more flexible in managing costs compared with less efficient peers (e.g., Soteriou
& Zenios, 1999). Firms with high-quality management enjoy stable relationships with their stakeholders, including employees and suppliers
(Srivastava, Shervani, & Fahey, 1998). Thus, in economically difficult
times, these firms can expect stakeholders to be more willing to
230
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
cooperate in lowering costs (e.g., by reducing input prices or wages). All
these benefits contribute to stabilizing revenues and costs during market downturns. As a consequence, the variance of firm returns is lower
than the market average, leading to a lower beta.
Although an excellent reputation may insulate a firm's stock from
market downturns, it may also contribute to outperforming competitors
during market upswings. Because of the superior management capabilities and operational efficiency, new markets are entered more rapidly
and easily (Fombrun, 1996). In addition, good reputation increases the
acceptance of new product introductions among consumers and channel partners (Kaufman, Jayachandran, & Rose, 2006). Higher operational
efficiency also implies the potential of firms to increase their revenues at
a lower cost relative to competitors. Hence, such companies offer larger
growth potential. However, faster growth also leads to greater variance
of returns relative to the market average, which in turn increases beta.
Thus, arguments for both positive and negative relationships between
corporate reputation and beta exist, which again leads to an empirical
question.
Table 3 summarizes the hypotheses for the main and interaction
effects of the focal variables on credit spread and beta. We test these
hypotheses subsequently.
5. Empirical study
5.1. Data and measures
Because the database and especially the data alignment influence
the model specification, we begin with this discussion before presenting
the empirical model. We collect data from various databases, including
the Center for Research in Security Prices (CRSP), Standard & Poor's
COMPUSTAT database, Bloomberg, Interbrand, Financial World,
Fortune, and the National Quality Research Center at the University of
Michigan. The data cover the 1989–2006 period. However, we cannot
observe all variables since 1989.
5.1.1. Credit spread (default risk)
Credit ratings are obtained from the COMPUSTAT databases.
COMPUSTAT offers Standard & Poor's long-term domestic issuer credit
ratings, which measure a firm's capacity to meet its long-term financial
commitments. The ratings range from AAA (highest credit standing)
to D (firm is in default) on an ordinal scale. The credit spread for a rating
class is calculated as the difference between the average yield (10-year
maturity) of a bond portfolio including only bonds of that rating class
and the yield of a risk-free bond (10-year US Treasury Bond). These
data are provided by Bloomberg's database.
5.1.2. Stock market beta (systematic risk)
We follow the standard market-model approach (see, for example,
McAlister et al., 2007) to estimate firm-specific betas. We use daily
stock returns for each firm and the market return of the CRSP ValueWeighted-Return Index of all trading days of the specific year to obtain
the estimates.
5.1.3. Customer satisfaction
ACSI produces a customer satisfaction score for each organization
that ranges from 0 to 100. ACSI collects and releases data on an annual
basis. We obtain the customer satisfaction scores from the fourth quarter of 1994 to the fourth quarter of 2006.
5.1.4. Brand value
Interbrand has been publishing the financial value of the Top 100
global brands since 1992. We obtain these data from publications in
Financial World or Business Week or from the website of the Interbrand
Group (www.interbrand.com). For the multi-brand firms in our data
set, we aggregate the available values across individual brands.
5.1.5. Corporate reputation
Corporate reputation scores are obtained from Fortune's annual
report on America's Most Admired Corporations. The reputation score
ranges from 0 to 10. Responses are solicited primarily from company
executives. Reputation data have been published since 1991.
5.1.6. Financial and other control variables
We follow previous studies in defining the financial variables. We
measure growth by the growth rate in total assets (e.g., Beaver et al.,
1970). Dividend payout is the ratio of cash dividends with respect to
earnings available to common stockholders (e.g., McAlister et al.,
2007). We measure leverage by total senior securities (preferred stocks
and bonds) divided by total assets (e.g., Beaver et al., 1970). Liquidity is
the “current ratio” of a firm (e.g., McAlister et al., 2007). Earnings variability is measured as the standard deviation of the earnings–price
ratio (Beaver et al., 1970; McAlister et al., 2007). The log of total assets
determines firm size (e.g., Rego et al., 2009). Pretax interest coverage
is the operating income after depreciation plus interest expense divided
by the interest expense (e.g., Blume et al., 1998). We compute the
operating margin by dividing operating income before depreciation by
firm sales (e.g., Blume et al., 1998). We measure competitive intensity
with the C4 concentration index. This index cumulates the market
shares of the four largest firms at the two-digit North American Industry
Classification System (NAICS) level. Data on these financial control
variables are obtained from COMPUSTAT. Appendix A provides the
exact data definitions that are used in our analysis.
5.1.7. Data merging procedure
Fig. B.1 in Appendix B summarizes the release dates of the financial
and non-financial metrics that are supposed to drive credit spreads
and beta. This figure shows that the release dates differ across years.
This variation implies challenges for model building that are discussed
in detail in Appendix B.
5.1.8. Descriptive and correlation statistics
Table 4 displays the descriptive statistics of our sample. The mean
beta is close to 1. The mean credit spread amounts to 1.24%. Both dependent variables demonstrate large variation. The average brand in our
sample is worth $8337 million. We multiply the brand values with the
firm-specific WACC by period to correct for the discounting applied by
Interbrand (see Appendix C for details). As a result, we obtain average
annual future branded earnings of $620.2 million. The mean satisfaction
and reputation scores are 75.7 (scale 0–100) and 6.6 (scale 0–10),
respectively. The means and standard deviations for the control variables can be obtained from Table 4 and are comparable to those in
other studies (e.g., McAlister et al., 2007).
Table 3
Overview of hypotheses.
Independent variable
Dependent variable
Credit spread
Customer satisfaction
Brand value
Corporate reputation
Beta
Direct impact
Relative strength
Moderating the impact of customer satisfaction
Direct impact
− (H1)
− (H2)
− (H3)
H4: Strongest impact by corporate reputation
−
+/− (H5;H5(alt))
− (H6)
− (H7)
+/−
+/−
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
231
with
Table 4
Univariate statistics.
Variables
N
Mean
Median
Std. dev.
Beta
Credit spread (in percent)
Branded earnings ($m)
Satisfaction (scale: 1–100)
Reputation (scale: 1–10)
Dividend payout
Earnings variability
Growth
Leverage
Ln asset size ($m)
Liquidity
Industry concentration
Pretax interest coverage (in percent)
Operating margin
4940
3196
1164
1893
1732
3992
3598
4785
5171
5130
4516
6153
4712
5053
.95
1.24
620.16
75.71
6.59
1.09
.08
.11
.44
9.08
1.54
.17
24.90
.18
.88
.92
321.16
76.01
6.70
.38
.02
.07
.46
9.24
1.30
.11
5.93
.17
.54
.95
854.82
6.47
1.04
30.00
.21
.29
.18
1.80
.57
.17
249.22
.35
Table 5 displays the correlation matrix. We do not note any excessive correlations that would indicate collinearity issues. Nevertheless,
we check for potential collinearity issues subsequently. Interestingly,
reputation scores are more strongly correlated with satisfaction than
with brand value (.31 vs. .23). There is virtually no correlation between
the WACC-corrected brand value and customer satisfaction (.12;
p N .10). The correlations of brand value and corporate reputation
with beta are negative but are not significant on a practical level. In
contrast, the correlation of customer satisfaction with beta is strongly
significant (− .26; p b .01). All the correlations of the three nonfinancial metrics with credit spread are negative and highly significant
(satisfaction: − .24; p b .01; brand value: − .16; p b .05; reputation:
−.44; p b .01).
Consistent with our theoretical arguments (see Table 2), brand value
and reputation are positively and significantly correlated with growth
(brand value: .12; p b .05; reputation: .20; p b .01), but we find no
significant correlation of growth with customer satisfaction (− .02;
p N .10).
5.2. Model
Building on the extant research on accounting, finance, and
marketing (e.g., Beaver et al., 1970; Blume et al., 1998; McAlister et al.,
2007), we specify the following two equations to explain the
components of capital cost. We adopt the models by Beaver et al.
(1970) and Blume et al. (1998) as baseline specification for Eqs. (2)
and (3), respectively:
SPREADit ¼ α 0i þ α 1 SPREADit−1 þ α 2 BV it þ α 3 SAT it þ α 4 REP it
þ α 5 SAT it % BV it þ α 6 SAT it % REP it þ α 7 BETAit−1
þ α 8 INT it−1 þ α 9 OPERit−1 þ α 10 LEV it−1
L−1
X
þ α 11 ln ðASSET it−1 Þ þ α 12 CONC it−1 þ
α 12þl IDil þ υit ; ð2Þ
i¼1
!
"
!
"
g N 0; σ 2 ; α ¼ α þ φ and φ i:i:d:
g N 0; σ 2 ; Coυðυ ; φ Þ ¼ 0;
υit i:i:d:
υ
0i
φ
it
i
i
i
!
"
!
"
g N 0; σ 2 ; β ¼ β þ κ and κ i:i:d:
g N 0; σ 2 ; Coυðε ; κ Þ ¼ 0;
ε it i:i:d:
ε
0i
i
i
κ
it
i
where
SPREADit Credit spread of firm i in period t
BVit
Brand value of firm i in period t
SATit
Customer satisfaction rating of firm i in period t
REPit
Corporate reputation of firm i in period t
BETAit − 1 Systematic risk of firm i in period t − 1
INTit − 1 Pretax interest coverage of firm i in period t − 1
OPERit − 1 Operating margin of firm i in period t − 1
LEVit − 1 Leverage of firm i in period t − 1
ASSETit − 1 Asset size of firm i in period t − 1
DIVit − 1 Dividend payout of firm i in period t − 1
GROWTHit − 1 Asset growth of firm i in period t − 1
LIQit − 1 Liquidity of firm i in period t − 1
EVARit − 1 Earnings variability of firm i in period t − 1
CONCit − 1 Industry concentration (C4 index) relevant for firm i in
period t − 1
IDil
Industry dummy for firm i and industry l (1 = firm i belongs
to industry l; 0 otherwise)
νit, εit, φi, κi Error terms
σ2υ, σ2φ, σ2ε , σ2κ Variances
α, β
Parameters to be estimated
i = 1, … I (number of firms)
t = 1, … T (number of periods)
l = 1, … L (number of industries).
Beaver et al. (1970) and Blume et al. (1998) provide detailed explanations of the financial control variables and their expected effects,
which we do not present here. We extend their models in several
ways. First, we add our three focal non-financial metrics and their interactions with customer satisfaction to the baseline model. Second, we
include the competitive intensity for each industry (McAlister et al.,
2007) and time-invariant industry dummies to control for heterogeneity at the industry level. More highly concentrated industries signal
opportunities for new competitors to enter the market and threaten
the cash flow stream of incumbents. We therefore expect a positive
effect on beta. For credit spreads, the effect of industry concentration
is not uniform. Firms in concentrated industries have above-average
profits. However, higher concentrated industries signal opportunities
for new competitors to enter the market and threaten the cash
flow stream of incumbents. Hence, we do not make a sign prediction
in this case. Third, we specify a random constant in our models to capture unobserved heterogeneity. This random constant controls for
other firm-specific differences in systematic risk and credit spread,
that we do not observe but that may affect our estimates. Finally,
we capture dynamic effects by including the lagged dependent variable.
Moreover, the inclusion of lagged dependent variables also controls for
inertia, persistence, and different initial conditions (Tuli & Bharadwaj,
2009).
5.3. Estimation
BETAit ¼ β0i þ β1 BETAit−1 þ β2 BV it þ β3 SAT it þ β4 REP it þ β5 SAT it
% BV it þ β6 SAT it % REP it þ β7 DIV it−1 þ β8 GROWTH it−1
þ β9 LEV it−1 þ β10 LIQ it−1 þ β11 EVARit−1
L−1
X
þ β12 ln ðASSET it−1 Þ þ β13 CONC it−1 þ
β13þl IDil þ !it ;
ð3Þ
i¼1
Note that both Eqs. (2) and (3) include a random constant to account
for unobserved firm heterogeneity (Greene, 2008). The terms φi and κi
denote firm-specific deviations of the heterogeneous constant from its
mean (α; β) and are assumed to be drawn from a normal distribution
with zero mean and constant variance. In addition, we acknowledge
the possibility that our non-financial variables are endogenous. Changes
in these metrics are a result of investments, which are in turn influenced
232
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
Table 5
Correlations (number of observations in parentheses).
1
1. Beta
2. Credit spread
3. Brand value
4. Satisfaction
5. Reputation
6. Dividend payout
7. Earnings variability
8. Growth
9. Leverage
10. Asset size (log)
11. Liquidity
12. Industry conc.
13. Pretax int. coverage
14. Operating margin
1.00 (4940)
.09⁎⁎⁎ (3068)
−.09 (1044)
−.26⁎⁎⁎ (1721)
−.03⁎ (1639)
−.03 (3731)
.11⁎⁎⁎ (3475)
.22⁎⁎⁎ (4541)
.17⁎ (4201)
−.08⁎⁎⁎ (4781)
−.21⁎⁎⁎ (3807)
.02⁎⁎ (4896)
.05⁎⁎ (4429)
−.16⁎⁎⁎ (4744)
2
3
4
5
6
1.00 (3196)
−.16⁎⁎ (781)
−.24⁎⁎⁎ (1433)
−.44⁎⁎⁎ (1311)
−.00 (2543)
.24⁎⁎⁎ (2411)
.02 (3049)
.41⁎⁎⁎ (2733)
−.25⁎⁎⁎ (3218)
.05⁎⁎ (2713)
.05⁎ (3167)
−.08⁎⁎⁎ (2951)
−.22⁎⁎⁎ (3047)
1.00 (1164)
.12 (438)
.23⁎⁎ (511)
−.03 (900)
.08⁎⁎ (798)
.12⁎⁎ (1085)
−.18⁎⁎ (943)
.40⁎⁎⁎ (1094)
−.05 (980)
.11⁎⁎ (1162)
.06⁎⁎ (1017)
.26⁎⁎⁎ (1081)
1.00 (1893)
.31⁎⁎⁎ (836)
−.10⁎⁎⁎ (1370)
−.24⁎⁎⁎ (1401)
−.02 (1774)
−.32⁎⁎⁎ (1560)
−.10⁎⁎⁎ (1789)
.10⁎⁎⁎ (1603)
−.35⁎⁎⁎ (1891)
.04 (1669)
−.09⁎⁎⁎ (1762)
1.00 (1732)
−.03 (1343)
−.37⁎⁎⁎ (1301)
.20⁎⁎⁎ (1674)
−.51⁎⁎⁎ (1449)
.08⁎⁎⁎ (1680)
.11⁎⁎⁎ (1447)
−.10⁎⁎⁎ (1730)
.08⁎⁎⁎ (1557)
.20⁎⁎⁎ (1653)
1.00 (3992)
.01 (2743)
−.06⁎⁎⁎ (3736)
−.00 (3412)
−.01 (3992)
−.03⁎⁎⁎ (3135)
−.01 (3980)
−.01 (3721)
−.01 (3939)
⁎ Statistical significance at 10% level (two-tailed).
⁎⁎ Statistical significance at 5% level (two-tailed).
⁎⁎⁎ Statistical significance at 1% level (two-tailed).
by the cost of capital, creating potential simultaneity issues. We follow
Fischer et al. (2010) and adopt their two-step estimation approach. In
the first step, we obtain instrumental variables by regressing the
respective endogenous variable on its instruments. We then estimate
the models with the instrumental variables using the simulated maximum likelihood technique. The estimator is consistent and asymptotically
normally distributed under the usual regularity conditions. Further details
regarding the estimation procedure can be found in Appendix C.
5.4. Estimation results
Tables 6 and 7 summarize the estimation results with respect to
credit spread and beta, respectively. We show the results for models
that include a varying set of predictors. The last column of Tables 6
and 7 displays the results when all predictors of Eqs. (2) and (3) are
incorporated. As mentioned above, this inclusion of all predictors creates the highest demand for joint observations, reducing the sample
sizes significantly. Therefore, including the results for varying predictor
sets helps us to better assess the stability of our results.
Overall, the model fit is very good for this class of data. The pseudo-R2
values, which are based on the squared correlation between predicted
and actual values of the criterion variable, range from .59 to .69 for the
beta regressions and from .57 to .76 for the credit spread regressions.
5.4.1. Results for credit spreads
We begin our discussion with the results of the first column in
Table 6. Here, we estimate a model that includes only the financial
and other control variables. All estimation results show the expected
sign. Our results are largely consistent with the findings of prior studies
(e.g., Anderson & Mansi, 2009; Blume et al., 1998).
The next three columns display the results for models that include
only one non-financial metric. We find a significant negative effect
(p b .05) on credit spreads for all three focal variables. The fifth column
of Table 6 presents the findings for the model that simultaneously
includes all three non-financial metrics. It is noteworthy that although
we have a reduced sample size, the picture does not change substantially.
The effects are again in the expected direction and reach significance.
Hence, the results support all three hypotheses regarding the main
effects: H1, H2, and H3. The findings show that customer satisfaction
(− .005; p b .01), brand value (− 4.9 × 10−4; p b .05), and corporate
reputation (− .072; p b .05) contribute to reducing credit spreads.
Because all variables are included simultaneously, the effects are indeed
incremental with respect to one another. Moreover, the effect of beta
is still positive and significant (.157; p b .05). Hence, non-financial
metrics may also influence credit spreads indirectly via beta.
The moderating effects of brand value (3.3 × 10− 7; p b .10) and
reputation (−5.8 × 10−4; p b .01) with regard to customer satisfaction
are both significant and show the expected sign (sixth column of
Table 6). The likelihood ratio test also supports this model extension
(χ2(2) = 8.68; p b .05). Brand value significantly attenuates the negative effect of customer satisfaction on credit spreads (H5), whereas
corporate reputation amplifies this negative effect (H6).
To summarize, we find support for H1 to H3 as well as for H5 and H6.
Customer satisfaction, brand value, and corporate reputation are found
to significantly decrease credit spreads. We discuss the relative effects of
the focal variables (H4) subsequently when we compute the elasticity
estimates.
5.4.2. Results for stock market beta
We begin our discussion with the results of the first column in
Table 7. Here, we estimate a model that includes only the financial
and other control variables. All effects for the financial control variables
show the expected sign. The results are similar to those in previous
studies (Beaver et al., 1970; McAlister et al., 2007).
The next three columns of Table 7 demonstrate the estimation
results that are observed when we add only one non-financial metric at a time. We find a strong negative effect of customer satisfaction (− .010; p b .01), which strongly supports H7. We have
suggested arguments for both positive and negative effects of
brand value and corporate reputation on beta. In fact, we do not
find a significant influence for these two non-financial metrics
(brand value: − 2.7 × 10 − 5, p N .05; reputation: − .025, p N .05),
which may suggest that both lines of arguments are relevant and that
the opposing effects offset one another. This conclusion does not change
if we jointly estimate the effects for all three non-financial metrics (fifth
column of Table 7). The coefficient associated with customer satisfaction
(−.008, p b .01) is still highly significant and of similar size despite the
substantially smaller sample.
The sixth column of Table 7 presents the results for a model in which
we also consider potential moderating effects of brand value and corporate reputation with regard to customer satisfaction. However, these
additional variables (brand value × satisfaction: −3.6 × 10−7, p N .10;
reputation × satisfaction: .008; p N .10) are not significant and do not
improve the model fit. The likelihood ratio test does not support this
model extension (χ2(2) = 1.36; p N .10).
5.5. Robustness tests
We perform several tests to assess the robustness of our results. First,
we note that the results in Tables 6 and 7 already indicate a relatively
high level of stability of the estimated effects across several models
233
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
7
1.00 (3598)
−.09⁎⁎⁎ (3172)
.07⁎⁎⁎ (3415)
−.43⁎⁎ (3415)
−.04⁎⁎ (2989)
−.02 (3590)
−.02 (3082)
−.12⁎⁎⁎ (3374)
8
9
10
11
12
13
14
1.00 (4785)
−.11⁎⁎⁎ (4073)
−.12⁎⁎⁎ (4785)
.25⁎⁎⁎ (3626)
−.02 (4776)
.04⁎⁎⁎ (4401)
−.28⁎⁎⁎ (4722)
1.00 (4303)
.25⁎⁎⁎ (4302)
−.38⁎⁎⁎ (3923)
−.15⁎⁎⁎ (4291)
−.18⁎⁎⁎ (4203)
.01 (4273)
1.00 (5130)
.37⁎⁎⁎ (3923)
−.19⁎⁎⁎ (5118)
−.03⁎⁎ (4710)
.21⁎⁎⁎ (5052)
1.00 (4059)
.12⁎⁎⁎ (4053)
.30⁎⁎⁎ (3649)
.14⁎⁎⁎ (3901)
1.00 (6153)
−.01 (4700)
−.05⁎⁎⁎ (5041)
1.00 (4712)
.08⁎⁎⁎ (4654)
1.00 (5053)
with varying numbers of predictor variables. To assess the stability of
our focal variables, we calculate the coefficient of variation, which is
the standard deviation of estimates divided by their mean across
different models. For the credit spread regressions, the values are .327
(satisfaction), .251 (brand value), and .197 (corporate reputation). For
the beta regressions, we obtain values of .284 (satisfaction), .141
(brand value), and .137 (corporate reputation). Overall, we observe a
low relative variance in the coefficient estimates across different
models with varying sample sizes and predictor variables. Distributions
with a coefficient of variation below 1/3 are considered low variance
(McKay, 1932).
Second, we determine whether our results are subject to collinearity issues. Following the “artificial orthogonalization” procedure
by Hill and Adkins (2008), we regress customer satisfaction on
brand value as well as corporate reputation and compute the residuals, which are orthogonal to the regressors by definition. We substitute the residuals into Eqs. (2) and (3) separately for each
interaction term and then re-estimate the equations. The results
for our interaction terms are not significantly different from the results in Tables 6 and 7.4
Third, we estimate models that include changes in the stock market
beta and credit spreads as dependent variables and changes in the
non-financial metrics and accounting/finance metrics as independent
variables (e.g., Bharadwaj et al., 2011). A model of such changes is
appropriate to reduce potential problems associated with time-invariant
unobservable factors (Tuli & Bharadwaj, 2009) and multicollinearity.
However, a change regression reduces the power of tests, as sample
size and variation are significantly reduced. We use these change
models for every model in Tables 6 and 7. Overall, the results are similar
to the results of the level models.
Fourth, for the stock market beta, we test for differences in the
effects for the non-financial metrics during economic upswings and
downswings. We follow Lamey, Deleersnyder, Dekimpe, and
Steenkamp (2007) to determine cyclical upturns versus downturns.
We include the non-financial metrics moderated by upturn and downturn dummies in our empirical model. The results are consistent with
our reasoning in that they show the expected sign. However, the
majority of the estimated coefficients are not significant because of
the small sample size.5
4
However, we note that the coefficient of the substituted collinearity-free variable is
consistently estimated, although the coefficients for the other variables are biased (Hill
& Adkins, 2008).
5
The estimation results can be obtained from the authors upon request.
6. Implications and conclusions
6.1. Managerial and research implications
Our results provide interesting insights that should be useful for
both managers and researchers. Conceptually, we provide a detailed
distinction between the information content of the three non-financial
metrics. Empirically, we find a strong effect of customer satisfaction
on both stock market beta and credit spreads. The conclusions differ
for brand value and corporate reputation. Neither metric appears to
affect beta, but both directly influence credit spreads. In addition, our
findings suggest that brand value and corporate reputation significantly
moderate the effect of customer satisfaction on credit spreads. We conclude that customer satisfaction, brand value, and corporate reputation
provide value-relevant information for investors, creditors, and credit
rating agencies above and beyond each individual metric. Hence, stakeholders do not appear to substitute one metric for the other when
assessing the various types of risks associated with the cost of capital.
Strictly speaking, we note that this conclusion relies on information provided by the ACSI, Interbrand, and Fortune's reputation index.
Given that all three non-financial metrics affect components of the
cost of capital, we can use our estimates to assess their relative influence. Because the metrics are measured with different scales, we cannot
compare coefficient estimates directly but must instead transform them
into elasticities. Table 8 summarizes the elasticity estimates. The table
shows the relative (percentage) increase in the cost of equity, cost of
debt, and WACC in reaction to a relative (percentage) increase in
customer satisfaction ratings, brand value, and corporate reputation
ratings. We take sample means for the risk-free rates, the capital
structure, and so forth. We also differentiate between short-term and
long-term effects. This separation has a direct practical implication
because it enables investors to incorporate a time-varying discount
factor into their valuation models. The long-term effect is obtained by
dividing the short-term elasticity by 1 minus the carryover coefficient.
Because we use a common carryover coefficient, differences in the
long-term elasticities are driven by the differences in the short-term
elasticities. The estimation results in Tables 6 and 7 show that dynamic
effects are indeed present, as the carryover coefficient associated
with the lagged dependent variable is always significantly different
from zero (p b .05). To better assess the robustness of elasticity
magnitudes, Table 8 shows both estimates based on the credit
spread models with and without moderators. In the following discussion, we refer to the credit spread model that includes the moderating
effects.
Customer satisfaction shows highly significant (p b .01) short-term
and long-term elasticity with respect to the cost of equity (short-term:
234
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
Table 6
Estimation results (Eq. (2)); dependent variable: credit spread.
Variables
Accounting/financial
Exp. sign Accounting/financial Accounting/financial
control
control variables only control
variables + customer variables + brand value
satisfaction
Accounting/financial
control
variables + corporate
reputation
Accounting/financial
Accounting/financial
control variables + all control variables + all
3 non-financial metrics 3 non-financial
metrics + moderators
3.290 (.618)⁎⁎⁎
.392 (.029)⁎⁎⁎
.448 (.134)aaa
–
–
3.642 (.822)⁎⁎⁎
.335 (.031)⁎⁎⁎
.381 (.121)aaa
−.005 (.002)aaa
−4.9×10−4
(2.7×10−4)aa
−.072 (.040)aa
–
–
Constant
Estimated SD
Lagged spreadb +
Satisfactionc
−
Brand valuec
−
3.412 (.277)⁎⁎⁎
.489 (.016)⁎⁎⁎
.541 (.119)aaa
–
–
3.661 (.341)⁎⁎⁎
.451 (.071)⁎⁎⁎
.437 (.082)aaa
−.007 (.002)aaa
–
Reputationc
Sat.c ∗ br. val.c
Sat.c ∗ reput.c
−
+/−
−
–
–
–
–
–
–
3.472 (.291)⁎⁎⁎
.355 (.081)⁎⁎⁎
.412 (.152)aaa
–
−3.8 × 10−4
(2.0×10−4)aa
–
–
–
Beta
Pretax int. cov.
+
−
.261 (.152)aa
−.009 (.004)aaa
.222 (.117)aa
−.002 (4.2×10−4)aaa
.247 (.068)aaa
.157 (.092)aa
−8.6×10−4 (2.0×10−4)aaa −.005 (9.2×10−4)aaa
Oper. margin
Leverage
Ln asset size
Ind. conc.
Log L
Pseudo R2
N
−
+
−
+/−
.201 (.122)aa
−.006
(3.4 10−4)aaa
−1.471 (.375)aaa
1.321 (.190)aaa
−.114 (.039)aaa
−.055 (.128)
−2067.63
.569
2586
3.716 (1.081)⁎⁎⁎
.342 (.046)⁎⁎⁎
.440 (.065)aaa
−.003 (7.6 10−4)aaa
−2.9 × 10−4
(2.1×10−4)a
−.102 (.052)aa
3.3×10−7 (1.8×10−7)⁎
−5.8 × 10−4
(2.7×10−4)aaa
.155 (.083)aa
−.004 (.001)aaa
−1.329 (.377)aaa
1.219 (.252)aaa
−.133 (.071)aa
.163 (.155)
−1036.93
.621
1262
−1.277 (.317)aaa
1.246 (.415)aaa
−.059 (.022)aaa
.251 (.186)
−353.41
.632
659
−.876 (.286)aaa
.758 (.251)aaa
−.056 (.027)aaa
.115 (.131)
−661.52
.665
1082
−.939 (.585)a
.550 (.337)a
−.014 (.010)a
−.167 (.220)
−36.97
.758
136
−.129 (.031)aaa
–
–
−1.043 (.711)a
.584 (.451)a
−.021 (.007)aaa
−.241 (.502)
−40.59
.735
136
Note: The standard errors are reported in parentheses.
Coefficients of industry dummies are not reported, but can be obtained from the authors.
a
Statistical significance at 10% level for variables with directional hypothesis (one-tailed).
aa
Statistical significance at 5% level for variables with directional hypothesis (one-tailed).
aaa
Statistical significance at 1% level for variables with directional hypothesis (one-tailed).
b
Denotes that the variable is instrumented by its lagged deviation from the firm-specific mean.
c
Denotes that the variables are instrumented by all exogenous variables of model (2) plus dividend payout, growth, liquidity, and earnings variability.
⁎ Statistical significance at 10% level (two-tailed).
⁎⁎ Statistical significance at 5% level (two-tailed).
⁎⁎⁎ Statistical significance at 1% level (two-tailed).
Table 7
Estimation results (Eq. (3)); dependent variable: stock market beta.
Variables
Exp. sign
Accounting/financial
control variables only
Accounting/financial
control
variables + customer
satisfaction
Accounting/financial
control
variables + brand value
Accounting/financial
control
variables + corporate
reputation
Accounting/financial
control variables + all
3 non-financial metrics
Accounting/financial
control variables + all
3 non-financial
metrics + moderators
1.093 (.322)⁎⁎⁎
.341 (.019)⁎⁎⁎
.318 (.059)aaa
−.010 (.004)aaa
–
.844 (.312)⁎⁎⁎
.244 (.017)⁎⁎⁎
.453 (.055)aaa
–
−2.7 × 10−5
(2.3 × 10−5)
–
–
1.236 (.172)⁎⁎⁎
.261 (.011)⁎⁎⁎
1.200 (.602)⁎⁎
.242 (.023)⁎⁎⁎
1.402 (.713)⁎⁎
.277 (.041)⁎⁎⁎
Constant
Estimated SD
Lagged betab
Satisfactionc
Brand valuec
+
−
+/−
.733 (.158)⁎⁎⁎
.233 (.022)⁎⁎⁎
.398 (.121)aaa
–
–
Reputationc
Sat.c ∗ br. val.c
+/−
+/−
–
–
–
–
Sat.c ∗ reput.c
Dividend pay.
+/−
−
–
−.022 (.011)aaa
Growth
Liquidity
Earnings var.
Leverage
Ln asset size
Ind. conc.
Log L
Pseudo R2
N
+
−
+
+
−
+
–
−1.3 × 10−4
(2.9 × 10−4)
.324 (.075)aaa
−.099 (.092)
4.021 (1.342)aaa
.543 (.077)aaa
−.039 (.008)aaa
.144 (.058)aaa
−622.49
.591
3204
−.038 (.058)
−.241 (.154)a
3.215 (1.539)aaa
.367 (.081)aaa
−.020 (.015)a
.288 (.077)aaa
−236.52
.667
1303
–
1.9 × 10−4
(2.3 × 10−4)
.341 (.188)aa
−.062 (.126)
4.251 (1.374)aaa
.122 (.097)a
−.029 (.014)aaa
.133 (.089)a
−122.43
.613
807
.299 (.102)aaa
–
–
−.025 (.016)
–
–
−1.8 × 10−4
(2.4 × 10−4)
.271 (.106)aaa
−.051 (.083)
2.195 (1.302)aa
.406 (.144)aaa
−.025 (.021)
.162 (.079)aaa
−144.79
.657
1184
.217 (.130)aa
−.008 (.004)aaa
−1.9 × 10−5
(1.6 × 10−5)
−.021 (.013)
–
–
−.088 (.035)aaa
.229 (.101)aaa
−.005 (.003)aa
−2.1 × 10−5
(3.5 × 10−5)
−.032 (.029)
3.6 × 10−7
(3.8 × 10−7)
.008 (.007)
−.105 (.052)aaa
.293 (.369)
−.073 (.054)a
4.214 (2.272)aa
.119 (.062)aa
−.052 (.044)
.027 (.151)
−33.60
.682
145
.315 (.278)
−.099 (.050)aa
3.871 (1.948)aaa
.102 (.078)a
−.031 (.038)
.029 (.142)
−32.92
.688
145
Note: The standard errors are reported in parentheses.
Coefficients of industry dummies are not reported, but can be obtained from the authors.
a
Statistical significance at 10% level for variables with directional hypothesis (one-tailed).
aa
Statistical significance at 5% level for variables with directional hypothesis (one-tailed).
aaa
Statistical significance at 1% level for variables with directional hypothesis (one-tailed).
b
Denotes that the variable is instrumented by its lagged deviation from the firm-specific mean.
c
Denotes that the variables are instrumented by all exogenous variables of model (3) plus interest coverage and operating margin.
⁎ Statistical significance at 10% level (two-tailed).
⁎⁎ Statistical significance at 5% level (two-tailed).
⁎⁎⁎ Statistical significance at 1% level (two-tailed).
235
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
Table 8
Comparison of estimated short-term and long-term elasticities.
Due to (percentage change) … N
(Eq. (2))
N
(Eq. (3))
Elasticity (percentage change) of …
Short-term
Cost of equity
Joint impact WITHOUT MODERATORS
(column 5 in Tables 6 and 7)
Brand value
136
145
−0.004 (0.003)
Satisfaction
136
145
−0.205 (0.103)aaa
Corporate reputation
136
145
−0.050 (0.031)
Joint impact WITH SIGNIFICANT MODERATORS
for credit spread equation (column 6 in Table 6 and column 5 in Table 7)
Brand value
136
145
−0.004 (0.003)
Satisfaction
136
145
−0.205 (0.103)aaa
Corporate reputation
136
145
−0.050 (0.031)
Long-term
Cost of debt
WACC
Cost of equity
Cost of debt
WACC
−0.005 (0.004)
−0.067 (0.039)aa −0.032 (0.020)
−0.041 (0.023)aa −0.020 (0.012)
−0.065 (0.026)aaa −0.143 (0.069)aaa −0.263 (0.131)aaa −0.110 (0.046)aaa −0.193 (0.094)aaa
−0.073 (0.041)aa −0.058 (0.035)
−0.064 (0.040)
−0.120 (0.069)aa −0.086 (0.053)
−0.023 (0.016)a
−0.012 (0.008)
−0.005 (0.004)
−0.041 (0.028)a
−0.021 (0.015)
−0.075 (0.019)aaa −0.148 (0.058)aaa −0.263 (0.131)aaa −0.134 (0.034)aaa −0.206 (0.088)aaa
aa
aa
−0.137 (0.071)
−0.083 (0.048)⁎
−0.064 (0.040)
−0.244 (0.128)
−0.137 (0.079)⁎
Note: The standard errors are reported in parentheses. For long-term elasticities, they are approximated with the delta method.
Mean values for cost of equity: 8.85% (total sample) and 9.02% (sample with joint impact of non-financials).
Mean values for cost of debt: 7.24% (total sample) and 7.66% (sample with joint impact of non-financials).
Mean values for WACC: 8.37% (total sample) and 8.61% (sample with joint impact of non-financials).
a
Statistical significance at 10% level for variables with directional hypothesis (one-tailed).
aa
Statistical significance at 5% level for variables with directional hypothesis (one-tailed).
aaa
Statistical significance at 1% level for variables with directional hypothesis (one-tailed).
⁎ Statistical significance at 10% level (two-tailed).
⁎⁎ Statistical significance at 5% level (two-tailed).
⁎⁎⁎ Statistical significance at 1% level (two-tailed).
− .205; long-term: − .263; p b .01). An elasticity of approximately
−.25 suggests decreasing marginal returns for the effect of satisfaction
ratings on equity cost, but the magnitude appears to be quite substantial. Consistent with the estimation results, we do not estimate significant equity-cost elasticities with respect to brand value and corporate
reputation (p N .05).
The picture is different when we consider the effects on credit
spreads. Note that we account for the indirect effects of non-financial
metrics via beta and interaction effects with respect to customer
satisfaction. We apply the delta method to obtain standard errors.
Because of the nature of this Taylor-series approximation of a nonlinear random term, the estimates tend to be inflated. Interestingly,
corporate reputation appears to exert the greatest influence on the
cost of debt; its short-term elasticity is − .137, and its long-term
elasticity − .244 (for both, p b .05). These elasticities are substantial.
The elasticities for customer satisfaction are − .075 (short-term) and
−.134 (long-term) and are highly significant (p b .01). The values are
considerably smaller for brand value (short-term: − .023 and longterm: −.041; both p b .10). Table 8 also shows the ultimate effects of
the non-financial metrics on WACC. Here, we consider the average capital structure. WACC elasticity is the highest with respect to satisfaction
(short-term: −.148 and long-term: −.206; both p b .01), followed by
corporate reputation (short-term: − .083; and long-term: − .137;
both p b .10). The elasticity estimates with respect to brand value are
rather small and are subject to a rather large approximated standard
error.
Table 8 shows that the credit spread elasticity is substantially higher
for corporate reputation compared with that for brand value and satisfaction. However, the difference is statistically significant only with
respect to brand value (p b .01). Therefore, we find partial support for
H4.
The estimated elasticities underline that the non-financial metrics
contain information that significantly drives the cost of capital. These
results may serve as an input for a dialog among senior accounting,
finance, and marketing executives regarding the role of non-financial
metrics in the terms of financing. Our findings can assist in measuring
the full financial impact of improving marketing metrics. Given that
many studies show the contribution of marketing to financial success
via an increase in the level of cash flows, these results complement
the picture via the reduction in capital cost. As a consequence, reducing
investments in marketing, customer relationship management or
reputation-building activities may not only harm a firm's value with
regard to future cash flows but also lead to higher hurdle rates for the
required return on capital.
Our analyses show that all three non-financial metrics are relevant drivers of capital cost components. We find an amazing stability
of coefficients across models with different sample sizes and different predictor sets. We have discussed the joint and distinct informational content with regard to the non-financial metrics. The
moderation analysis with regard to the cost of debt reveals the importance of how to frame the communication and disclosure of customer satisfaction ratings to stakeholders. Whereas strong brand
value dilutes the information content of customer satisfaction, demonstrating operational efficiency makes the information derived
from customer satisfaction much more valuable. Hence, marketing
managers can more effectively communicate the effects of these intangibles and make a stronger case for marketing in the boardroom
(Luo et al., 2010).
6.2. Limitations and future research
We need to mention a few limitations of this study that could
stimulate future research. Although the results are stable across model
estimations with different sample sizes, the full model is estimated
with a rather small sample. Analysis of a larger sample is likely to further
increase the precision of estimates.
Although we have good reasons for the choice of ACSI, Interbrand,
and Fortune to measure customer satisfaction, brand value, and corporate reputation, it may be interesting to investigate the sensitivity of
the results with respect to alternative measures of satisfaction, brand
value, and corporate reputation.
Our elasticity analysis shows that the effect of customer satisfaction
ratings on capital cost, for example, is substantial. However, this finding
does not imply that satisfaction ratings should be improved per se.
Improvements require investments that are likely to be subject to
decreasing returns of scale. Future research may determine the optimal
investment level.
236
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
Appendix A. Variable definitions
Variables
Definition
Measure
COMPUSTAT
Asset size
Dividend payout
Log of total assets
Cash dividends/earnings
Ln(total assets)
Cash dividends/available income
Earnings variability
Standard deviation of earnings/price ratio
Standard deviation of earnings/price ratio
Growth
Industry concentration
Terminal total assets/initial assets
C4-concentration index
Leverage
Total senior securities (preferred stocks
and bonds)/total assets
Current ratio
Operating income before depreciation/
sales
(Operating income after
depreciation + interest expense)/interest
expense
Ln(total assetst/total assetst − 1)
Sum of market shares of the top four firms
in the industry defined at two digits of the
NAICS
Total senior securities/total assets
DATA 6 (total assets)
DATA 21 (cash dividend); DATA 20 (income available for
common stockholders)
DATA 20 (income before extraordinary items — adjusted
for common stock equivalents); DATA 24 (Price — close),
DATA 25 (common shares outstanding)
DATA 6 (total assets)
DATA 12 (sales)
Liquidity
Operating margin
Pretax interest coverage
Current assets/current liabilities
Operating income before depreciation/
sales
(Operating income after
depreciation + interest expense)/interest
expense
Our dependent variables credit spread and beta reflect the level at
the end of each year. The financial control variables are reported by
COMPUSTAT at the annual and quarterly levels, whereas quarterly
information is typically unavailable at the beginning of the time series
in 1989.
The structure of the data has important implications for model
building. First, because the focal metrics are measured only at the annual
level, the year is the periodicity of our empirical analysis. Ideally,
we would align our dependent variables with the release dates of the
financial and non-financial information. For example, beta would be
calculated over the 12 months preceding the release date of new satisfaction ratings for a firm. Unfortunately, this alignment is not possible
Appendix B. Data merging procedure
Fig. B.1 in the appendix summarizes the release dates of the financial
and non-financial metrics that are supposed to drive credit spreads and
beta. The figure shows that release dates differ across years. It also
shows the period over which the variables in our empirical models are
measured. As is evident in this figure, the non-financial metrics are
measured only once per year, whereas the release dates differ across
years. Annual ACSI data are collected in different quarters for different
industries. Interbrand tracks brands across the year and releases new
brand values in the third quarter (usually September). Fortune releases
data on corporate reputation in the first quarter (usually March).
Measurement
Periods
DATA 5 + DATA 9 + DATA 10 (total senior securities);
DATA 6 (total assets)
DATA 4 (current assets); DATA 5 (current liabilities)
DATA 13 (operating income before depreciation); DATA
12 (sales)
DATA 178 (operating income after depreciation); DATA
15 (interest expense)
Financial Controls
Corporate Reputation
Brand Value
Stock market beta
Credit spreads
ACSI 1a
ACSI 2
ACSI 3
ACSI 4
Year 1
Q1
Release Dates
Q2
Year 2
Q3
Q4
Q1
Q3
Q2
Financial controls
Corporate reputation (Fortune)
Brand value (Interbrand)
Customer Satisfaction (ACSI)
Time Index in
Models (2) and (3)
t-1
t
Note:
Denotes time span during which construct is measured
a
ACSI 1 etc. denote the quarters when data are released for different economic sectors
Fig. B.1. Data alignment.
Q4
t
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
with our data because the non-financial metrics are released at different
points in time within a year. Although this lack of alignment may be a
limitation of this database, we believe that any limitation would be offset by the new insights that we generate with respect to the joint role of
the three non-financial metrics. We also estimate models for credit
spreads and beta that account for the different release dates of satisfaction. We find no evidence that the release date moderates the influence
of satisfaction on capital costs.
Second, we note that the financial control variables themselves
could be influenced by the cost of capital, i.e., stock market beta and
credit ratings. To avoid such reverse causality between the dependent
variables and the financial control variables, we include financial
controls of the previous period (t − 1) in our models. Consistent with
our hypotheses, the announcement of the non-financial metrics has
informational value for investors. Hence, we regress credit spread and
beta in period t on the values announced in period t. To account for
potential simultaneity issues, we use an instrumental variable estimation approach that we describe subsequently.
Appendix C. Details about the estimation procedure
C.1. Measurement-induced endogeneity of financial brand value
We acknowledge potential concerns with respect to a financial
brand value measure such as Interbrand's measure, which involves
discounting future brand-induced cash flows. We multiply each brand
value by the WACC of the parent company. The result is a value for
the average brand-induced future cash flow per annum. Note that
dividing annual cash flows by WACC produces the net present value
of this cash flow stream. Through this transformation, we remove the
effect of capital cost on computing the financial brand value measure.
To account for other sources of endogeneity, we still need to instrument
this transformed brand value variable as we do for the other nonfinancial metrics.
C.2. Identification of endogenous non-financial metrics
All predictor variables in Eqs. (2) and (3) that are not endogenous
serve as instruments. Specifically, we consider brand value (i.e., the
transformed variable), satisfaction, corporate reputation, and the interaction of brand value and corporate reputation with satisfaction to be
endogenous in Eqs. (2) and (3). Pretax interest coverage, operating
margin, leverage, asset size (log of total assets), asset growth, dividend
payout, liquidity, earnings variability, and industry concentration,
which are measured in period t − 1, are assumed to be exogenous.
We extend this set of instruments by dividend payout, leverage, the
log of total assets, interest coverage, liquidity, earnings variability, and
operating margin, which are measured in period t − 2. These 7 twoperiod lagged instruments plus dividend payout, asset growth, liquidity,
and earnings variability provide the overidentifying restrictions for
Eq. (2), which includes 5 endogenous variables (brand value, customer
satisfaction, corporate reputation, interaction of brand value and corporate reputation with satisfaction). In Eq. (3), the same 7 two-period
lagged instruments plus interest coverage and operating margin
provide the overidentifying restrictions for 5 endogenous variables.
Conceptually, operating margin, industry concentration, and interest
coverage in particular should provide the identification for the
endogenous non-financial metrics. First, profitability is a determinant
of financial brand value (Interbrand, 2012); thus operating margin
should be a good instrument for brand value. Second, customers in
less concentrated industries are typically more satisfied than those in
heavily concentrated industries (Luo et al., 2010). Moreover, firms in
heavily concentrated industries have higher market shares on average
than firms in less concentrated industries. Rego, Morgan, and Fornell
(2013) show a consistently significant negative relationship between
market share and customer satisfaction. Third, interest coverage is an
237
indicator of the financial resources that are available for investments
in corporate reputation.
C.3. Testing the exogeneity assumption
Although we have good conceptual reasons for our choice of instruments, we test for their exogeneity. We proceed as follows. First, we
examine whether the predetermined predictor variables in Eqs. (2)
and (3) – which serve as instruments – are indeed exogenous. In
Eq. (2), these variables are pretax interest coverage, operating margin,
leverage, the log of total assets, and industry concentration, which are
all measured in t − 1. In Eq. (3), these variables are dividend payout,
asset growth, leverage, liquidity, earnings variability, the log of total
assets, and industry concentration, which are all measured in t − 1.
We apply the Durbin–Wu–Hausman test (Davidson & MacKinnon,
1993) to test the independence assumption for these regressors with
respect to the error term. Specifically, we regress each instrument on
all other exogenous variables and obtain the residuals from this regression. These residuals are then included in Eqs. (2) and (3), and the
significance of the residuals' coefficients is tested. However, we find
that none of these coefficients is significant (p N .10). Detailed results
for the significance of each residual's coefficients can be obtained from
the authors upon request. These test results suggest that the exogeneity
assumption for our predetermined variables in Eqs. (2) and (3) cannot
be rejected.
Second, we apply the specification test (HT test) outlined by
Hausman and Taylor (1981). We use this test to test for the exogeneity
of all other instruments that are not included in the estimation equations. Given a set of exactly identified instruments, the HT test examines
the exogeneity of additional overidentifying instruments. For Eq. (2),
the one-year and two-year lagged exogenous variables pretax interest
coverage, operating margin, leverage, and asset size as well as the
one-year lagged variable industry concentration provide the initial
set of instruments. We test for the following set of overidentifying
instruments: dividend payout, asset growth, liquidity, and earnings
variability, which are measured in period t − 1. The HT test is not
rejected (χ2(4) = 1.52, p = .82). For Eq. (3), the one-year and twoyear lagged exogenous variables dividend payout, growth rate, leverage,
liquidity, earnings variability, and the log of total assets as well as the
one-year lagged variable industry concentration provide the initial set
of instruments. We apply the HT test to the overidentifying instruments
of pretax interest coverage and operating margin. The test is not
rejected (χ2(2) = 1.48, p = .48).
C.4. Strength of instruments
Establishing evidence of the exogeneity of the instruments is necessary but not sufficient, as the instruments may be weak. We therefore
check for the strength of our set of instruments. First-stage regressions
for the endogenous non-financial metrics of brand value, satisfaction,
and corporate reputation show satisfactory levels of R2 and F-values.
The mean R2 is .35, and all F-values exceed the threshold of 10, thus
indicating no issues with weak instruments (Stock, Wright, & Yogo,
2002). Detailed estimation results from the first-stage regressions can
be obtained from the authors upon request. In addition, variables such
as operating margin, industry concentration, and interest coverage
show significant effects in the direction that is consistent with our conceptual arguments of identification (p b .01).
C.5. Identification of carryover effects
The lagged credit spread and beta in Eqs. (2) and (3) measure carryover effects. In the estimation, we use the lagged deviations of credit
spread and beta from the firm-specific mean. This procedure is necessary to isolate the true dynamic effects from the heterogeneity effects
that are associated with the lagged dependent variable in a panel
238
A. Himme, M. Fischer / Intern. J. of Research in Marketing 31 (2014) 224–238
(e.g., Arellano, 2003; Fischer & Albers, 2010). In addition, the lagged
beta is a predictor of the credit spread in Eq. (2). Because the equation
system is recursive, we can use observed values for the lagged beta in
Eq. (2).
References
Agarwal, M. K., & Berens, G. (2009). How corporate social performance influences financial
performance: Cash flow and cost of capital. MSI Working Paper No. 09-001, Cambridge
(Mass.).
Anderson, E. W., Fornell, C., & Lehmann, D. R. (1994). Customer satisfaction, market share,
and profitability: Findings from Sweden. Journal of Marketing, 58(3), 53–66.
Anderson, E. W., Fornell, C., & Mazvancheryl, S. K. (2004). Customer satisfaction and
shareholder value. Journal of Marketing, 68(4), 172–185.
Anderson, E. W., & Mansi, S. (2009). Does customer satisfaction matter to investors?
Findings from the bond market. Journal of Marketing Research, 46(5), 703–714.
Arellano, M. (2003). Panel data econometrics. Oxford: Oxford University Press.
Bahadir, S.C., Bharadwaj, S. G., & Srivastava, R. K. (2008). Financial value of brands in
mergers and acquisitions: Is value in the eye of the beholder? Journal of Marketing,
72(6), 49–64.
Barth, M. E., Clement, M. B., Foster, G., & Kasznik, R. (1998). Brand values and capital
market valuation. Review of Accounting Studies, 3(1–2), 41–68.
Beaver, W., Kettler, P., & Scholes, M. (1970). The association between market-determined
and accounting-determined risk measures. The Accounting Review, 45(4), 654–682.
Bharadwaj, S. G., Tuli, K. R., & Bonfrer, A. (2011). The impact of brand quality on shareholder wealth. Journal of Marketing, 75(5), 88–104.
Blume, M. E., Lim, F., & MacKinlay, C. A. (1998). The declining credit quality of U.S.
corporate debt: Myth or reality? Journal of Finance, 53(4), 1389–1413.
Brealey, R. A., Myers, S.C., & Allen, F. (2007). Principles of corporate finance (9th ed.),
Boston: MA McGraw-Hill/Irwin.
Davidson, R., & MacKinnon, J. G. (1993). Estimation and inference in econometrics. New
York: Oxford University Press.
Ederington, L. H., Yawitz, J. B., & Roberts, B. E. (1987). The informational content of bond
ratings. Journal of Financial Research, 10(3), 211–226.
Elton, E. J., Gruber, M. J., Agrawal, D., & Mann, C. (2001). Explaining the rate spread on
corporate bonds. Journal of Finance, 56, 247–277.
Fama, E. F., & French, K. R. (1992). The cross-section of expected stock returns. Journal of
Finance, 47(2), 427–465.
Fischer, M., & Albers, S. (2010). Patient- or physician-oriented marketing: What drives
primary demand for prescription drugs? Journal of Marketing Research, 47(1), 103–121.
Fischer, M., Leeflang, P.S. H., & Verhoef, P. C. (2010). Drivers of peak-sales for pharmaceutical brands. Quantitative Marketing and Economics, 8(4), 429–460.
Fombrun, C. (1996). Reputation: Realizing value from the corporate image. Boston: Harvard
Business School Press.
Fombrun, C. J., & Shanley, M. (1990). What's in a name: Reputation building and corporate
strategy. Academy of Management Journal, 33(2), 233–258.
Fornell, C., Johnson, M.D., Anderson, E. W., Cha, J., & Bryant, B. E. (1996). The American
Customer Satisfaction Index: Nature, purpose, and findings. Journal of Marketing,
70(4), 7–18.
Fornell, C., Mithas, S., Morgeson, F. V., III, & Krishnan, M. S. (2006). Customer satisfaction
and stock prices: High returns, low risks. Journal of Marketing, 70(1), 3–14.
Fortune (2009). Most admired companies. http://money.cnn.com/magazines/fortune/
mostadmired/2009/index.html (June 2010)
Greene, W. H. (2008). Econometric analysis (6th ed.)Upper Saddle River, NJ: Pearson.
Gruca, T. S., & Rego, L. L. (2005). Customer satisfaction, cash flow, and shareholder value.
Journal of Marketing, 69(3), 115–130.
Hausman, J. A., & Taylor, W. E. (1981). Panel data and unobservable effects. Econometrica,
49(November), 1377–1398.
Hill, R. C., & Adkins, L. C. (2008). Collinearity. In B. H. Baltagi (Ed.), Companion to theoretical
econometrics (pp. 256–278). Chichester: Wiley.
Hill, N. C., & Stone, B. K. (1980). Accounting betas, systematic operating risk, and financial
leverage: A risk-composition approach to the determinants of systematic risk. Journal
of Financial and Quantitative Analysis, 15(3), 595–637.
Horrigan, J. O. (1966). The determination of long-term credit standing with financial
ratios. Journal of Accounting Research, 4, 44–62.
Interbrand (2012). Best global brands: Applications and methodology. http://www.
interbrand.com/de/best-global-brands/2012/best-global-brands-methodology.aspx
(October 2012)
Johansson, J. K., Dimofte, C. V., & Mazvancheryl, S. K. (2012). The performance of global
brands in the 2008 financial crisis: A test of two brand value measures.
International Journal of Research in Marketing, 29(3), 235–245.
Kaplan, R. S., & Urwitz, G. (1979). Statistical models of bond ratings: A methodological
inquiry. Journal of Business, 52(2), 231–261.
Kaufman, P., Jayachandran, S., & Rose, R. L. (2006). The role of relational embeddedness in
retail buyers' selection of new products. Journal of Marketing Research, 43(4),
580–587.
Kothari, S. P. (2001). Capital markets research in accounting. Journal of Accounting and
Economics, 31, 105–231.
Lamey, L., Deleersnyder, B., Dekimpe, M. G., & Steenkamp, J. -B. E. M. (2007). How
business cycles contribute to private-label success: Evidence from the United States
and Europe. Journal of Marketing, 71(1), 1–15.
Leone, R. P., Rao, V. R., Keller, K. L., Luo, A.M., McAlister, L., & Srivastava, R. (2006). Linking
brand equity to customer equity. Journal of Service Research, 9(2), 125–138.
Luo, X., & Bhattacharya, C. B. (2006). Corporate social responsibility, customer satisfaction,
and market value. Journal of Marketing, 70(October), 1–18.
Luo, X., Homburg, C., & Wieseke, J. (2010). Customer satisfaction, analyst stock recommendations, and firm value. Journal of Marketing Research, 47(6), 1041–1058.
Madden, T. J., Fehle, F., & Fournier, S. (2006). Brands matter: An empirical demonstration
of the creation of shareholder value through branding. Journal of the Academy of
Marketing Science, 34(2), 224–235.
McAlister, L., Srinivasan, R., & Kim, M. (2007). Advertising, research and development, and
systematic risk of the firm. Journal of Marketing, 71(1), 35–48.
McKay, A. T. (1932). Distribution of the coefficient of variation and the extended
t-distribution. Journal of the Royal Statistical Society, 95(4), 695–698.
Mithas, S., Krishnan, M. S., & Fornell, C. (2005). Why do customer relationship management
applications affect customer satisfaction? Journal of Marketing, 69(4), 201–209.
Orlitzky, M., & Benjamin, J.D. (2001). Corporate social performance and firm risk: A
meta-analytic review. Business & Society, 40(4), 369–396.
Osinga, E. C., Leeflang, P.S. H., Srinivasan, S., & Wieringa, J. E. (2011). Why do firms invest
in consumer advertising with limited sales response? A shareholder perspective.
Journal of Marketing, 75(1), 109–124.
Pinches, G. E., & Mingo, K. A. (1973). A multivariate analysis of industrial bond ratings.
Journal of Finance, 28(1), 1–18.
Rego, L. L., Billett, M. T., & Morgan, N. A. (2009). Consumer based brand equity and firm
risk. Journal of Marketing, 73(6), 47–60.
Rego, L. L., Morgan, N. A., & Fornell, C. (2013). Reexamining the market share-customer
satisfaction relationship. Journal of Marketing, 77(5), 1–20.
Singh, M., Faircloth, S., & Nejadmalayeri, A. (2005). Capital market impact of
product marketing strategy: Evidence from the relationship between advertising
expenses and cost of capital. Journal of the Academy of Marketing Science, 33(4),
432–444.
Soteriou, A., & Zenios, S. A. (1999). Operations, quality, and profitability in the provision of
banking services. Management Science, 45(9), 1221–1238.
Srinivasan, S., & Hanssens, D. H. (2009). Marketing and firm value: Metrics, methods,
findings, and future directions. Journal of Marketing Research, 46(June), 293–312.
Srivastava, R. K., Shervani, T. A., & Fahey, L. (1998). Market-based assets and shareholder
value: A framework for analysis. Journal of Marketing, 62(1), 2–18.
Standard & Poor's (2011a). Criteria for rating the global branded nondurable consumer
products industry. Available at: http://www.standardandpoors.com/prot/ratings/
articles/en/eu/?articleType=HTML&assetID=1245303899075 (January 2011)
Standard & Poor's (2011b). What are credit ratings and how do they work? Available at:
http://img.en25.com/Web/StandardandPoors/SP_CreditRatingsGuide.pdf (January
2012)
Stock, J. H., Wright, J. H., & Yogo, M. (2002). A survey of weak instruments and weak
identification in generalized method of moments. Journal of Business and Economic
Statistics, 20(4), 518–529.
Tuli, K. R., & Bharadwaj, S. G. (2009). Customer satisfaction and stock returns risk. Journal
of Marketing, 73(6), 184–197.