Marketing Science Institute Working Paper Series 2012 Report No. 12-112 Do Marketing Campaigns Produce Multichannel Buying and More Profitable Customers? A Field Experiment Elisa Montaguti, and Scott A. Neslin, and Sara Valentini “Do Marketing Campaigns Produce Multichannel Buying and More Profitable Customers? A Field Experiment,” Elisa Montaguti, and Scott A. Neslin, and Sara Valentini © 2012; Report Summary © 2012 Marketing Science Institute MSI working papers are distributed for the benefit of MSI corporate and academic members and the general public. Reports are not to be reproduced or published in any form or by any means, electronic or mechanical, without written permission Report Summary One of the most intriguing and managerially relevant findings in the multichannel customer management literature is the positive association between the multichannel customer and profits. The question is whether this is an actionable, causal relationship. Specifically, can marketing campaigns be designed to turn single-channel customers into multichannel customers, and in turn will these multichannel customers become more profitable to the firm? Elisa Montaguti, Scott Neslin, and Sara Valentini conduct a field experiment to investigate this question. In their study, they manipulate two key aspects of campaign design: message and incentives. They investigate two messagesone which directly communicates the benefits of multichannel shopping, the other which emphasizes the core value proposition of the firm. They also investigate the provision of financial incentives in the form of coupons and compare it to the case where no financial incentives are provided. Findings The multichannel message coupled with no financial incentives produced more multichannel customers and increased revenues more than other campaign designs. The authors estimate the profit ROI of this strategy to be 40%. These results suggest that the multichannel message motivated customers to explore the benefits of multichannel buying, and as they did so, they became more satisfied customers and hence purchased more over time. Further, this multichannel message/non-financial incentive campaign worked best on customers who otherwise would be predicted not to become multichannel on their own. Using an a priori estimated model, the authors estimate that targeting just customers with low pre-disposition towards multichannel shopping would generate a 51% ROI. The multichannel message/financial incentive campaign did not significantly increase either multichannel channel behavior or revenues. It had a one-year ROI of -142%. The results suggest that the poor performance of the multichannel/financial campaign resulted from its inability to motivate customers to explore different channels. The campaign forced customers to shop in channels they might not like in order to take advantage of the financial offer, negatively affecting satisfaction with the shopping experience. The value proposition campaigneither with or without incentivesdid not increase either multichannel behavior or revenues. The ROIs for these campaigns were -104% and -106% respectively. Managerial implications A multichannel/nonfinancial campaign raising customer awareness of the availability of a firm’s different points of contact is inexpensive and effective at generating multichannel shoppers and increasing profits. This strategy acts to modify customers’ channel shopping behavior at a lower cost than financial multichannel incentives. The multichannel shoppers who respond to this campaign generate more profit and are therefore more valuable to the firm. Marketing Science Institute Working Paper Series 1 Further, financial incentives might be a loss-generating activity for firms because they are not only costly but do not increase multichannel shopping. Conclusion This study is contingent on the particular firm that participated, as well as the specific messages and incentives. However, the research clearly builds the knowledge base in this critical area by demonstrating the effectiveness of multichannel campaigns, offering guidance on the design of those campaigns, and demonstrating the potential for targeting based on an a priori estimated model of the customer’s multichannel potential. Importantly, it demonstrates that research on the multichannel/profitability relationship can be translated into practice, and verifies the causal link from marketing actions to multichannel customer behavior to customer profitability. Elisa Montaguti is Associate Professor of Marketing, Department of Management, University of Bologna, Italy. Scott A. Neslin is Albert Wesley Frey Professor of Marketing, Tuck School of Business, Dartmouth College. Sara Valentini is Assistant Professor of Marketing, Department of Management, and University of Bologna, Italy. Acknowledgements The authors thank an anonymous retailer for cooperating in the implementation of this field experiment. This research is supported by the Marketing Science Institute and the Wharton Interactive Media Initiative as a winner of the “Modeling Multichannel Customer Behavior” research proposal competition. Marketing Science Institute Working Paper Series 2 Introduction The ever-expanding multiplicity of channels through which customers can purchase from companies has produced the “multichannel customer”. An intriguing finding related to multichannel customers is the growing consensus among academics and practitioners that multichannel customers buy more and are more valuable than single channel customers. This finding is highly important because it suggests a customer management strategy for increasing customer value – undertake marketing campaigns that produce more multichannel customers. Such customers should produce higher revenues and profits, thereby increasing their value to the firm. However, recent literature points out that we still do not know the extent to which this positive association between multichannel behavior and customer value is actionable (Neslin and Shankar 2009). One issue of course is whether the association is causal or due to spurious factors such as self-selection or high levels of marketing directed at multichannel customers (Blattberg, Kim, and Neslin 2008). But even if the result is causal in the statistical sense, the studies that find this positive relationship are conducted in a status quo environment. That environment produces a certain amount of multichannel behavior and a certain amount of profits. The question is whether the firm can proactively intercede in that environment and create more multichannel behavior and more customer value. Hence, the purpose of our work is to answer four main questions: (1) Can newly-acquired customers be turned into multichannel customers? (2) Does the newly created multichannel customer become more valuable to the firm? (3) How in practical terms can this transformation be accomplished, in particular, what types of messages and incentives work best? (4) Can the customers most likely to react favourably be identified and hence targeted in future campaigns? An important theme to the above questions is, can we establish a causal link from marketing actions to customer multichannel behavior to customer profitability. Our approach to establishing this causality is to investigate these questions through a field experiment. The field experiment allows us to establish causality both in the creation of multichannel customers and in the generation of higher profits (Cook and Campbell 1979). And most importantly, it shows whether an active intervention by the firm can produce high value multichannel customers. Marketing Science Institute Working Paper Series 3 The field experiment is conducted with the cooperation of a multichannel book retailer. Two cohorts of newly acquired customers are available. The first contains customers acquired at the end of 2009 and monitored over 2010 and is used to develop a model that predicts the potential for a customer to become multichannel. This model is used as a covariate to analyze the results of the field experiment, and to investigate whether particular customers can be targeted based on their multichannel potential. The second cohort includes customers acquired at the end of 2010. The field experiment is conducted on the second cohort whereby customers are randomly assigned to experimental treatments and multichannel shopping buying behavior is monitored for a full year (January 2011-January 2012). The field test entails two factors – “message” and “incentive”. Both factors contain two levels. The message factor is either an explicit invitation for the customer to become multichannel or a more general message communicating the value proposition of the firm. The incentive factor entails either the provision of price discount coupons or no coupons provided. We find that: 1. The multichannel message coupled with no financial incentive achieves the desired results. It produces more multichannel customers and increases revenues. We estimate the profit ROI of this strategy to be 40%. 2. The multichannel message/no-financial incentive campaign works best on customers who otherwise would be predicted not to become multichannel on their own. These customers can be identified by an a priori estimated model. Profitability of a well-targeted campaign would return more than the 40% ROI achieved by targeting all customers, specifically 51% ROI by targeting just the customers with low pre-disposition towards multichannel shopping. 3. The multichannel message/financial-incentive campaign does not significantly increase either multichannel channel behavior or revenues. As a result it produces a loss, a oneyear ROI of -142%. 4. Both the value proposition campaigns – incentive and no incentive –do not increase either multichannel behavior or revenues. The ROIs for these campaigns are -104% and 106% respectively. The most important result is that indeed, management can proactively create profitable multichannel customers. However, our findings suggest that identifying the right combination of Marketing Science Institute Working Paper Series 4 message and incentive is a challenge. Our particular results highlight the importance of message over financial incentive as a strategy for creating profitable multichannel customers. However, for a different company with a different customer base, the optimal campaign could vary. What we learn crucially is that (1) the multichannel/profitability strategy can work, and (2) companies will want to pre-test the particulars of the message/incentive aspects of their campaign. This pretesting is especially valuable because as we show in conclusion (2) above, the results can be used to develop a targeting strategy for reaching the customers most likely to become multichannel. Theory and Evidence Regarding the Multichannel-Profitability Link Why multichannel customers may become more profitable There are several reasons why the multichannel customer might be more profitable (see Blattberg, Kim and Neslin 2008; and Neslin and Shankar 2009 for a review): (1) self-selection, (2) marketing, and (3) customer loyalty. The self-selection explanation is that high volume customers have more purchase occasions; hence they naturally use more channels if available. If this explanation holds, then the multichannel/profitability relationship is one of reverse causality – more profitable customers become multichannel rather than multichannel causes them to become more profitable. The marketing explanation is that multichannel shoppers naturally receive more marketing simply because they interact with the firm through several channels. For example, the customer who utilizes the Internet and call center channels might be more likely to receive both e-marketing and direct mail solicitations. The customer loyalty explanation views multichannel marketing as additional service that increases customer satisfaction and ultimately loyalty. The firm can increase service by making available the channel/s customers prefer and providing tight integration between channels. In short, the multichannel customer is a happier customer and therefore becomes more valuable. Another, more mechanical way of looking at it is that the multichannel customer incurs a set-up cost (in learning how to use the various channels) and hence would incur a switching cost in defecting to another company. The multichannel-to-profitability link has received considerable empirical verification in the literature. Both academic work and the business press suggest that multichannel buying behavior equals extra profits (e.g. Loftus, Mulliken and Sharp 2008; Myers, Pickersgill, and Van Metre Marketing Science Institute Working Paper Series 5 2004; Thomas and Sullivan 2005; Kumar and Venkatesan 2005; Venakatesan, Kumar and Ravishanker 2007; Ansari, Mela, and Neslin 2008; Boehm 2008; Campbell and Frei 2010; Xue, Hitt and Chen 2011). Thomas and Sullivan (2005) show that multichannel shoppers generate more revenue, purchase more items, purchase in more categories, and purchase more frequently than the single-channel shoppers counterpart. Venkatesan, Kumar, and Ravishanker (2007) model customer profits versus lagged multichannel purchasing and their results show that multichannel shopping is positively and significantly related to profits. Ansari, Mela, and Neslin (2008) find a positive association between multichannel purchasing and customer revenues. They attempt to disentangle the self-selection, loyalty, and marketing effects. They rule out the first two and attribute the relationship to additional marketing and higher responsiveness to marketing. Their findings highlight the need to control for other forms of marketing besides the multichannel campaign we wish to assess. Indeed we will do so in that all customers in our field test will receive the same marketing except for different multichannel campaigns. While these studies are important for establishing the positive link between multichannel buying and customer value, they do not address two issues: (1) Can a proactive marketing campaign geared toward creating multichannel customers succeed in doing so, and (2) which customers are most likely to respond to these efforts. Can marketing create multichannel customers? Previous research has investigated which factors influence customers’ choices of channels. Blattberg, Kim, and Neslin (2008, Chapter 25) and Neslin et al. (2006) proposes six such factors: (1) channel attributes, (2) social influence, (3) channel integration, (4) individual differences, (5) situational factors, and (6) marketing. Channel attributes include convenience, privacy, and risk. Social influence entails the impact that significant “others” have on channel choice. Channel integration refers to coordination among channels that makes it easier for the customer to use multiple channels. Individual differences include demographics, psychographics, and previous channel usage experience. Situational factors comprise the context of the purchase occasion, ranging from the weather to the type of purchase (personal vs. gift). Finally, several studies have shown that marketing can influence channel choice (Thomas and Sullivan 2005; Ansari, Mela, and Neslin Marketing Science Institute Working Paper Series 6 2008), and the adoption of an additional channel (Venkatesan, Kumar, and Ravishanker 2007). Marketing includes e-mails, catalogs, and incentives that encourage customers to use one channel versus another. The marketing efforts investigated in these studies were not designed to create multichannel customers. They do show however that marketing influences channel choice at least of a particular channel. This suggests that marketing that attempts to convince customers to become multichannel has a chance to succeed. Interestingly, the above literature has primarily focused on non-financial communications without considering the different effect that financial versus non-financial incentives might generate on channel choice. Gedenk and Neslin (1999), for example, have shown that financial incentives produce greater sales bumps but are relatively detrimental to brand loyalty, compared to non-financial incentives. In summary, previous research has found a positive association between multichannel buying behavior and customer value. It has also found that marketing influences customer choice of a particular channel, e.g., emails encourage Internet usage. This suggests that a campaign to create multichannel usage has a chance to succeed. No previous research has demonstrated that an actionable campaign could be assembled that: 1) creates multichannel customers, 2) determines the most effective way to do this, and 3) examines the profit implications of such a campaign. This is what we do. Research Design We are interested in determining whether the firm can increase its profit by turning its customers into multichannel buyers using a proactively delivered marketing campaign. We also want to investigate whether customers who are more likely to become multichannel can be effectively targeted to improve profitability. In this section we first describe the setting for our field test. Then we discuss the alternative campaigns we designed and the rationale behind them. Next we describe the research design in detail. Finally we outline the analysis we will use to measure the effectiveness of the campaigns and determine whether customers can be identified who are particularly responsive to them. Marketing Science Institute Working Paper Series 7 Empirical setting We obtained the cooperation of a major multichannel European book retailer for conducting the field experiment. The company sells books through stores, mail-order, phone and the Internet. Each channel shares the same assortment and price. The company operates on a subscription business model, thus each customer must become a member in order to purchase, and all transactions and their timing are tracked. The firm sends its main catalog five times per year, and its other marketing activities are managed around each mailing, e.g., special promotions, price changes, etc. Consequently, customers make purchase decisions in a shopping context created by the current catalog. In our data we monitor five catalogue mailings (henceforth “periods”). Importantly, none of the firm’s marketing activities is targeted according to channel usage. This allows us to consider the firm’s current marketing activities as “baseline” and our field test delivers additional different types of communications and incentives to drive multichannel buying. Additionally, we benefited from a context in which newly acquired customers had never been encouraged to change their channel choice pattern prior to the field test. The cooperating firm never explicitly tried to move customers across channels, so the communications we sent in our test was entirely new to these new customers. Marketing communication treatments We distinguish two key aspects of a marketing campaign designed to induce customers to become multichannel: the message and the incentive. We will create a 2 × 2 design with message at two levels and incentive at two levels. In specifying these levels, we wanted to cover a range from highly explicit, overt urging of the customer to become multichannel, to a more implicit, less overt “nudge.” We accomplished this as follows: The two levels of the message factor are: (1) a “multichannel” message extolling the benefits of multichannel shopping and making sure the customer is aware of the multichannel choices available, and (2) a “value proposition” message emphasizing the key selling points of the company, which entailed assortment, service, and special promotions. The multichannel message overtly urged the customer to become multichannel. The value proposition message encouraged the customer to buy more, which perhaps would result in the customer trying multiple channels. Marketing Science Institute Working Paper Series 8 The two levels of the incentive factor were: financial incentives available versus not available. The financial incentive was the provision of price discount coupons. While the incentive factor is crossed with the message factor, the nature of the financial incentives differed depending on the message. In the spirit of creating a “hard sell” multichannel campaign, the financial incentives for the multichannel message entailed three coupons, one for each channel. The idea was to provide direct incentive to use three channels, and a message that backed up why this was a good idea for the customer. For the value proposition message, there were three coupons but no specifications on which channels they could be used for. This was in the spirit of creating a more soft-sell campaign. In summary then, we have four campaigns identified by message (multichannel versus value proposition), and incentive (financial versus non-financial). We label these “multichannel/financial” (MF), “multichannel/non-financial” (MNF), “value proposition/financial” (VPF), and “value proposition/non-financial” (VPNF). The campaign was delivered via a prominent card that was sent just few days before the catalog mailing. A card reminder was featured clearly at the front of the transparently wrapped catalog. Figure 1 shows these four cards. (Figures and tables follow References throughout.) As suggested above, the multichannel/financial campaign was the most explicit hard-sell approach, while the value proposition/non-financial campaign was the most soft-sell. The intermediate campaigns were also interesting. The multichannel/non-financial campaign emphasized the message and perhaps would be a more effective communication. In terms of the standard hierarchy of effects way of thinking about customer decisions, this campaign might be most successful at creating the attitudes required for the customer to become multichannel. However, it lacked the promotion aspect often necessary to convert attitudes to behavior. The value proposition/financial campaign might convince the customer to buy more, but there is no guidance here on becoming multichannel. That would presumably occur simply because the customer would want to make more purchases, and learn in the process about multichannel shopping. We do not advance a priori hypotheses for the ordering of success for these campaigns. It is tempting to anticipate the financial incentive campaigns will do better. They provide the promotional push needed to change behavior. However, in the multichannel/financial campaign, the coupons required multichannel behavior. The customer could react negatively to being Marketing Science Institute Working Paper Series 9 forced to become multichannel, i.e., this approach might be too hard sell. In the value proposition/financial treatment, the customer could simply use the coupons as needed for purchases that would have been made anyway, hence little incrementality and little multichannel shopping. Among the non-incentive campaigns, the multichannel message would appear to be more effective for what we were trying to accomplish than simply reciting the value proposition of the retailer. Presumably the customer was at least partially convinced of these virtues to begin with, which is why the customer signed up with the retailer. In summary, it is difficult to rationalize how the value proposition/non-financial campaign would be best. However, arguments could be made for the other three campaigns. We leave it to the field test to decide. What is important at this point is that we are demonstrating the need to separate message from incentive, and think strategically about each of these in coming up with a campaign to create multichannel customers. Experimental design and data The field experiment allows us to observe whether, and if so, which of the alternative marketing campaigns induce multichannel buying, then examine whether this translates to higher customer profitability. We obtained two cohorts of customers (Cohort 1 and Cohort 2) who lived within at least one store’s service area and who entered into a subscription agreement with the company after the last catalog mailed in 2009 (Cohort 1) and 2010 (Cohort 2) was sent. We refer to the period in which the customer entered into the subscription as the acquisition period; the latter periods are post acquisition. For Cohort 1 the acquisition period was the fifth and last period of 2009; their behavior was then monitored over the subsequent four periods in 2010. For Cohort 2, the acquisition period was the fifth period of 2010; they were observed over the next five periods in 2011, from January 2011 until January 7, 2012. This means that this last cohort was followed for five post-acquisition periods (see Figure 2). Cohort 1 was used to estimate a multichannel potential model, described subsequently, and that is the only way in which their data are used. Cohort 2 was the experimental cohort, randomly assigned to different marketing campaign treatments. Marketing Science Institute Working Paper Series 10 For Cohort 2 we created four treatment conditions (see Figure 3): multichannel/financial (MF); value proposition/financial (VPF); multichannel/non-financial (MNF), and value proposition/non-financial (VPNF), plus a control group. We randomly selected customers to be included in different treatments and in the control group. On January 7th, 2011, the beginning of period 1 for Cohort 2, customers included in the treatment conditions received one of the above-mentioned cards one to three days before the catalog was mailed to them. A card reminder was included in the catalog cover. By contrast, customers allocated to the control group did not receive any communications except the catalog. A second card was then sent using the same procedure, i.e. card then catalogue, on the 10th of March, the beginning of Period 2, to the same customers. On May, 20; July, 29; and October, 7 respectively, a third, fourth, and fifth catalog was mailed to all customers both in treatment and control conditions without any further communications related to channel usage. The firm recorded all consumer transactions during these five periods. We therefore have information on: which channel was selected by each customer on each purchase occasion, the date of each purchase, and how much was spent. Analysis Analysis approach There are four key parts to our analysis: First, we will estimate and evaluate a multichannel potential model using data from Cohort 1 customers. This can be viewed as a database marketing predictive model (Blattberg, Kim, and Neslin 2008) of whether a customer will naturally become multichannel. We will use this model in two ways: (1) as a covariate for enabling us to discern more clearly the impact of the treatments, and (2) as a variable to interact with treatments to uncover whether response to treatment varies according to multichannel potential. This will allow us to identify customers who should be targeted by specific treatments. Second, we will provide simple descriptive statistics of the results - % who become multichannel, average spending levels, etc. Third, we will use the multichannel potential model to predict whether the customers in Cohort 2 are likely to become multichannel on their own, and use these predictions as a covariate in analyzing the descriptive statistics and interacting these Marketing Science Institute Working Paper Series 11 predictions with treatment. This will provide us with a more precise “read” on the impact of the treatments, and generate guidance on whether there are interactions between treatment and potential that could help in future targeting. Fourth, we will analyze the impact of becoming multichannel on customer profits. Steps 3 and 4 together tell us (1) which if any treatments created multichannel customers, (2) which treatments should be targeted based on multichannel potential, and (3) whether multichannel behavior translates into higher profits. Multichannel potential model As discussed above, we develop a predictive model to identify in advance customers with a high probability of becoming multichannel customers. We estimate and test this model on Cohort 1, and then use it to “score” Cohort 2 customers. This model is not of substantive interest in and of itself. We simply want a model that performs well at predicting whether the customer would or would not become multichannel on her/his own. It is important to emphasize that the multichannel potential model requires only variables obtained during the acquisition period, plus a time trend variable that is defined exogenously. Therefore, the model can be applied to customers in Cohort 2 using only information available by the end of their acquisition period, before they are assigned to treatments and observed. Let i denote each customer (i=1,..., I), t denote time (t=1, …T), and Acquisition and Post subscripts denote the acquisition versus post-acquisition periods. Customer i becomes multichannel in the period in which for the first time she makes a purchase using a channel she never used before. Becoming multichannel is therefore the “event” of interest. We consider the five periods to be five discrete points in time because they correspond to the firm’s marketing cycle of one catalog per period, and rarely did the customer make more than one purchase in a period (only 7% of purchases). Previous work has proposed two alternatives for defining multichannel shopping behavior. One consists of cross-sectional ex-post measures whereby a customer who used more than one channel or made a certain percentage (e.g. 80%) of her purchases using different channels at the end of the observation period is considered a multichannel shopper (e.g. Kumar and Venkatesan 2005). The second alternative consists of time-varying measures where customers are defined as Marketing Science Institute Working Paper Series 12 multichannel when they use more than a channel within a period t (i.e. a year, see Venkatesan, Kumar and Ravishanker 2007). This measure varies over time, hence is able to capture the dynamic nature of multichannel shopping, but it requires a minimum number of purchases per period per customer. Customers need to have made at least two purchases in a given period in two different channels to be classified as multichannel shoppers in that period. Moreover, customers who make purchases across periods using two different channels are not considered as multichannel. Our definition combines the dynamics of the second approach with the more-than-onechannel aspect of the first approach. Specifically, we define the dependent variable to equal 1 in the period t in which the customer i becomes multichannel, if indeed the customer ever becomes multichannel1. Our definition focuses on the point in time when a customer moves from being a single channel to a multiple channel user. Hence, it does not require a minimum number of purchases per period. Customers who adopt a new channel are considered multichannel regardless of their number of purchase occasions. We use our dynamic measure of the multichannel behavior and model the probability the customer becomes multichannel using a discrete-time hazard model formulated as a probit regression function (Allison 1982). We can express, therefore, the potential of becoming multichannel as: (1) Pr (Becoming Multichannel )Post ,it = Φ(X Acquisition ,i , Z Acquisition ,i , C Acquisition ,i , Trend t ) The symbol “Ф” stands for the cumulative distribution function of the standard normal distribution. We used three classes of independent variables that can be observed during the acquisition period (XAcquisition,i, ZAcquisition,i, and CAcquisition,i) and a trend term . XAcquisition,i indicates observable individual variables available at the time the customer is acquired (e.g., gender, age, place of residence, etc.). ZAcquisition,i indicates transactional variables that can be computed during the acquisition period (e.g. RFM, returns). CAcquisition,i indicates the sales channel used by customer i during the acquisition period (i.e. catalog, Internet, store or phone). Finally, we 1 For example if individual i became multichannel at time 3, we would have four different observations for this individual. For the third observation, the multichannel dependent variable would be coded 1, and 0 for the previous two time units. Cohort 1 is observed for a maximum of four time units, after the acquisition period. Hence if a customer never experienced the event she would have four observations and the multichannel variable would be coded as 0 for each of the four observations. Marketing Science Institute Working Paper Series 13 include a Trendt term because the baseline hazard function may not be constant over time (see Table A1 in the Appendix for the complete list of the variables used). We estimated equation (1) with random effects across customers (see Table A2 in the Appendix for results). We evaluated the predictive accuracy of this multichannel potential model through cross-validation. Specifically we randomly select 75% (i.e. 26,583 customers) for in-sample estimation and the remaining 25% (i.e. 8,808 customers) for out-sample prediction. Figure 4 shows a lift chart for predicting which customers become multichannel. Predictions are based on the random effect probit model parameter estimates. As the chart shows customers in the top deciles have a distinctly higher chance of becoming multichannel customers compared to the average customer. There is no degradation in lift in going from the in-sample to out-sample predictions. We conclude that this model produces good predictions of whether a customer will become multichannel. The ultimate test of this, however, will be when we use its predictions as a covariate in analyzing the results for Cohort 2. Specifically, we will use the parameter estimates contained in Table A2 to predict each customer in Cohort 2’s potential to become multichannel. We will only use the acquisition period data for this cohort, because that is all we need to predict their natural propensity to become multichannel. Descriptive analysis We ran the field experiment for five periods encompassing a full year. During this time customers purchased using the channel of their choice. We now present descriptive statistics of the results. Figure 5 illustrates the cumulative percentage of customers who became multichannel, by treatment, over the five periods. Table 1 provides more information (sample size and standard errors) on the cumulative percent of customers who became multichannel by the end of the entire observation period, by treatment. Figure 5 and Table 1 show that for the multichannel/non-financial treatment there is a notable gain in the percentage of customers who became multichannel. The percentages in Figure 5 and Table 1 are not significantly different at conventional levels, but as described earlier, we will use the market potential model predictions as a covariate in order to analyze these data with more precision. Marketing Science Institute Working Paper Series 14 Figure 6 shows the difference in cumulative profits per customer by treatment versus the control group. It appears the customers in the multichannel/non-financial treatment group generate higher profit than the control group; the other treatment groups do not generate higher profits compared to control. Table 2 reports the average cumulative profit over time, by group, and also displays profit ROI. Only the multichannel non-financial communication group is significantly different from the control group, and achieves a positive ROI (40%), whereas all other conditions have a negative ROI. To provide insight on why the multichannel/non-financial treatment performs best, we examine the impact of the treatments on the frequency of purchase and the average profit per purchase occasion. We observe a slight increase in the number of purchase occasions for customers in the multichannel/non-financial treatment group by the end of the observation period (Figure 7). More pronounced, customers who receive the multichannel/non-financial communication cumulatively spend more per purchase occasion than customers included in other groups (Figure 8). Coupled with the finding that these customers were more likely to become multichannel over time, these results suggest that the multichannel message motivated customers to explore the benefits of multichannel buying, and as they did so, they became more satisfied customers and hence purchased more over time. Relatedly, the poor performance of the multichannel/financial campaign results from its inability to motivate customers to explore different channels. The multichannel/financial campaign forces customers to shop in channels they might not like in order to take advantage of the financial offer. This can negatively affect satisfaction with the shopping experience, hence does not have a beneficial effect on the amount spent (see Figure 8). The above descriptive statistics are consistent with our thesis that a marketing campaign can be designed to turn customers multichannel and once they become multichannel they become more profitable. However, we need more precision in order to verify statistical significance. Accordingly, we utilize a statistical model with covariates to assess the magnitude and significance of the campaigns’ effects on multichannel shopping and the impact of multichannel shopping on profits. Marketing Science Institute Working Paper Series 15 6WDWLVWLFDODQDO\VLV Our statistical analysis models the probability of a customer becoming multichannel, and the change in customers’ profits from this time point onward. The profits equation will contain a dummy variable signifying whether the customer becomes multichannel. However, becoming multichannel is endogenous. Therefore, we need to take into account the potential endogeneity of the multichannel dummy (see e.g., Imbens and Angrist 1994; Angrist, Imbens and Rubin 1996; Wooldrige 2002, Chapter 18; Verbeek 2004, Chapter 7). To do so, we formulate an endogenous binary variable model (becoming multichannel) to examine the impact of becoming multichannel on profits. Let i indicate the individual (i=1, …I), t indicates the time (t=1,…5), and j the treatment (j=1,…4). Equations 2 and 3 summarize the specification of this model: (2) (3) 4 4 j =1 j =1 Multichannel *it = ιi + α t + δPotentialit + ∑ λ jTreatmentit + ∑ κ jTreatmentij × Potentialit + ν it Profitit = γ 0i + βMultichannelit + γ 1Seasonalityt + γ 2Trend t + ε it Where: 1 if Mulichannelit = 0 if Multichannel * it Multichannel * it >0 ≤0 And Profitit = Profit (€) generated by customer i during time unit t. Multichannelit = Dummy variable that indicates the event of becoming multichannel. This variable takes value 1 if customer i became multichannel in time unit t, and 0 otherwise. Seasonalityt = Seasonal index for revenues in time unit t (we adapted the Abraham and Lodish’s (1993) method to this context to derive this index). Trendt = Linear trend term. Potentialit = Probability of becoming multichannel. We used the parameter estimates of the multichannel potential model described in equation (1) to predict this probability. Marketing Science Institute Working Paper Series 16 Treatmentij = dummy variables that take value 1 if customer i is in the treatement group j (j=1,…,4), and 0 otherwise. The four groups (MF, VPF, MNF VPNF) identify the different types of campaigns used. The control group (i.e. no campaign) represents the baseline. In the above specification, Multichannelit is the endogenous dummy variable indicating whether the customer becomes multichannel or not. Maddala (1983, p.120-121) derives a twostep estimator procedure in the form of a switching regression model: (4) E ( Profit it | Mutltichannelit = 1 ) = β + γ 1 Sesonalityt + γ 2Trend t 4 φ (α t + δPotentialit + ∑ λ j Incentiveij + +σ j =1 4 Φ (α t + δPotentialit + ∑ λ j Incentiveij + j =1 (5) 4 ∑κ j Incentiveij × Potentialit ) j Incentiveij × Potentialit ) j =1 4 ∑κ j =1 E ( Profit it | Mutltichannelit = 0 ) = γ 1 Sesonalityt + γ 2Trend t 4 φ (α t + δPotentialit + ∑ λ j Incentiveij + −σ j =1 4 ∑κ j Incentiveij × Potentialit ) j =1 4 1 − Φ (α t + δPotentialit + ∑ λ j Incentiveij + j =1 4 ∑κ j Incentiveij × Potentialit ) j =1 Where σ=Cov(εht,νht), and ϕ and Ф stand for the density cumulative distribution functions of the standard normal distribution. The key parameters of interest in these equations are: β which quantifies the impact of the treatment endogenous variable (becoming multichannel) on the outcome variable (profits), δ which indicates the role played by the individual pre-disposition of becoming multichannel, λj which captures the effectiveness of treatments in turning customers multichannel, and κj, which reports whether the effectiveness of the treatments is moderated by the potential to become multichannel. Finally, αt is a set of constants to express the hazard rate as a function of time (Allison 1982). To estimate these parameters we used a random effect probit for the multichannel model, a random effect GLS regression for the profit equation, and the two-stage estimation procedure proposed by Maddala (1983, p.121). Statistical analysis results Table 3 shows the estimation results of the multichannel model (equation 3) and compares Marketing Science Institute Working Paper Series 17 three different nested versions with the full model. Model A represents the basic model. It is just a function of treatments and potential. Model B adds interactions between potential and treatments. Model C relaxes the assumption of a constant hazard rate over time. The full model is a function of potential, treatments, interactions between potential and treatments, and time. Models A-C are nested within the full model. We compared each alternative nested model with the full model using a likelihood ratio test. The LR test statistic indicates that the more general specification fits the data significantly better than the alternative nested specifications (model AC). We summarize the results in Table 3: Time. The set of constants that express the hazard rate as a function of time are significant and indicate that the baseline hazard is not constant over time. No clear time pattern emerges, although it is less likely to become multichannel immediately after the acquisition period (α1= 4.19), then there is a tendency of the hazard rate to increase over time. Treatments. One of the key objectives of this work is to investigate whether a firm can create more multichannel customers. Our results indicate that the multichannel/non-financial campaign (MNF) that communicated the benefits of multichannel shopping but did not offer a financial incentive is the only campaign that achieves this (λ3=0.10). The impact of this campaign on multichannel buying behavior is significant at p < 0.05. Potential: Results show that indeed pre-disposition towards multichannel usage (potential) increases the probability of becoming multichannel (δ=5.89). Note this means that companies can identify in advance customers with different levels of pre-disposition through a targeting model by using just the information observable at the very start of the relationship with its customer base (i.e. the acquisition period). Interaction between treatments and potential: Customers can be more or less receptive to a given multichannel campaigns depending on how strong their pre-disposition is towards multichannel shopping. The superior performance of Model D indicates that these interactions add explanatory power. To investigate the nature of these interactions in more detail we computed for each customer the slope of the predicted effect of each incentive on the probability of becoming multichannel at any value of potential, and then we computed the estimated Marketing Science Institute Working Paper Series 18 standard error for each effect2. The significance test of the impact of the incentives at any given value of potential takes the form of a t-test that divides the coefficient by its estimated standard error (Jaccard and Turrisi 2003). Figure 9 plots on the y-axis the t-test of the impact of the incentive and on the x-axis different values of potential3. This figure summarizes the impact of each treatment at different possible levels of pre-disposition towards multichannel behavior. It shows that at low values of potential the multichannel/non-financial treatment (MNF) has a positive and significant impact on becoming multichannel, but for very high level of potential the magnitude of the impact decreases and the treatment does not significantly affect the probability. Value proposition treatments (both financial and non-financial) do not significantly affect the probability of multichannel shopping at any level of potential. By contrast, the multichannel/financial treatment at high levels of potential is slightly significant, with a negative impact, i.e., it decreases the probability of becoming multichannel among customers with a high potential of becoming multichannel. In summary, these results show that a marketing campaign can trigger multichannel shopping. The multichannel message coupled with no financial incentive (MNF) achieves this, especially for customers with a very low a priori propensity of becoming multichannel. We therefore computed ROI assuming that the firm mails the card with the multichannel/nonfinancial message just to those customers for whom the impact of this treatment is significant (i.e. t-stat >1.96), as the company would do in implementing the program in a new cohort of customers or in a second year. The ROI not surprisingly increases to 51%. By contrast, the financial campaigns not only do not work, but may produce a negative effect in terms of multichannel shopping among customers with a high pre-disposition towards the multichannel usage, who maybe would have become multichannel spontaneously if financial incentives had not been offered. Note that in our cohort of customers the potential prediction assumes on average a value of 0.07 (Standard Deviation= .098, Median=0.04). This means that the multichannel/non-financial campaign has a positive and significant effect for the majority of the customer base. This 2 bj at Potential=λj+κjPotential SE(bj at Potential)=[(var(λj)+Potential2var(κj)+2Potentialcov(λj, κj)]1/2 , where j=1 if MF, j=2 if VPF, j=3 if MNF, j=4 if VPNF. 3 Note that potential is a probability, hence it varies from 0 to 1. We built the chart by taking into consideration all the 0.01 increments starting from 0 till 1. Marketing Science Institute Working Paper Series 19 treatment is not effective just for the customers included in the top-decile of the variable potential. In contrast, the multichannel/financial treatment has a significant and negative impact just for the customers with an extremely high pre-disposition towards multichannel usage, that is, customers included in the top decile in Cohort 2. Multichannel shopping and Profits. We estimated equation (3) to examine the link between multichannel shopping and customer profitability, answering the question of whether newly created multichannel customers become more valuable to the firm. Table 4 shows that multichannel shopping has a positive and significant impact (β=8.45). Hence, when the customer becomes multichannel she generates on average an additional profit of 8.45 € per period. Note that the coefficient that represents the covariance between the error terms of the multichannel and profit models is not significant (σ= Cov(εht,νht) =-0.05). This means that the correlation between the error terms of equation (2) and (3) is close to zero and seems to suggest that individuals do not self-select into the becoming multichannel on the basis of the idiosyncratic component of their response to the treatment (Verbeek 2004). This might suggest that the “self-selection” explanation provided to explain the higher purchase volumes observed among multichannel customers (see Blattberg, Kim and Neslin 2008, p. 639) is not supported. Overall these results confirm the thesis of our research, that a marketing campaign can be designed that produces more multichannel customers, and in turn these customers are more profitable. We show that the particular campaign that does this – a benefits-of-multichannel message not coupled with a financial incentive – works especially well on customers who a priori would not have been expected to become multichannel on their own. This is a sensible result and suggests that while the campaign is profitable when mailed indiscriminately to a group of customers, it can be made even more profitable (on a ROI basis) by targeting customers who otherwise have low potential of becoming multichannel. Conclusions This research sheds light on the relationship between multichannel shopping and profitability. Previous works provided some empirical evidence of their association (e.g. Venkatesan, Kumar, and Ravishanker 2007; Hitt and Frei 2002). In this paper, we show this association can be translated into practice by designing a marketing campaign that produces Marketing Science Institute Working Paper Series 20 more multichannel customers who then are more profitable. We thus establish a causal link from marketing to multichannel to profitability. Our work provides insights into how to design more effective marketing campaigns to induce customers to become multichannel. Customers are responsive to specific claims emphasizing the advantages of multichannel shopping. By inducing customers to explore the benefits of multichannel buying, the non-financial multichannel message makes customers more inclined to spend. These findings corroborate existing literature documenting the positive effect of non-financial communication in multichannel contexts (e.g. Ansari, Mela and Neslin 2008). Moreover, we find that customers are not responsive to a firm’s value proposition message or to financial incentives. Our field experiment allows for the measurement of the impact of customers’ turning multichannel on profit. Again we show that a multichannel/non-financial campaign reminding customers of the availability of a firm’s different points of contact is inexpensive and effective at both generating multichannel shoppers and increasing profits. This strategy acts to modify customers’ channel shopping behavior at a lower cost than financial multichannel incentives. The multichannel shoppers who responded to this campaign generate more profit and are therefore more valuable to the firm. Our field experiment is a longitudinal study. We exposed customers to different treatments for two periods and we monitor them for three periods after the campaign finished. This allowed us to assess how persistent is the effect of the communication treatments on a customer’s likelihood to turn multichannel and in turn on customer profitability in the long-run. This research also shows that customers with a high pre-disposition towards multichannel shopping have little “upside”. For instance, the multichannel/non-financial communication is more effective on low-potential customers, and not effective on high-potential customers. This implies that marketing campaigns should be pre-tested to develop targeting strategy for reaching customers who are more likely to respond to this campaign. Our field test offers some important results for managers. First and foremost, the multichannel customer strategy is viable. Firms can devise marketing campaigns turning customers into multichannel shoppers and making them more profitable. Second, not all communications and incentives are equally effective. Non-price oriented promotional advertising increases the likelihood and accelerates the time to become multichannel. More Marketing Science Institute Working Paper Series 21 specifically multichannel non-financial communications increase revenues by 1.35 € per year at a cost of 0.46 (the cost of two cards) euros per customer. This is a ROI of 40%. Second financial incentives might be a loss-generating activity for firms because they are not only costly and ineffective for the large majority of the customer based, but they can even negatively affect the probability of using multiple channels of customers favorably disposed to multichannel shopping (high potential). Several factors could influence the effectives of marketing campaigns to encourage multichannel shopping behavior beside the nature of the incentives including: frequency, copy, scheduling. Additionally, the industry in which a firm operates along with the products it sells may influence our results. We, therefore, recommend that managers test their campaign before implementing it. Finally, we show that a model of multichannel potential can be used to increase ROI through targeting, because customers with low a priori probabilities of becoming multichannel are the ones who are most responsive to the optimal campaign. Limitations and Future Research There are several limitations to our approach that invite further research and extensions. First we use data from a single company and from a single market. It will be useful to replicate this work in different contexts where more firms are examined; customers are not bounded to the firm by a contract, and/or buy more frequently. Second, while the best campaign is profitable, the percentage of customers who eventually turned multichannel is rather low. This might be related to the low number of purchase occasions observed in this industry, which limit a customer’s opportunity to use multiple channels. Third, we just observed whether customers made a purchase and we do not know whether our treatments encouraged customer to search across channels. Future work could pursue this avenue and try to assess whether marketing campaigns boost channel search and make customers more satisfied with the shopping experience. Fourth, our treatments were sent to customers immediately after the acquisition period (first and second periods). In this way, campaigns reached customers new-to-the-firm, who had never Marketing Science Institute Working Paper Series 22 being encouraged to change their channel usage before the field test, so did not have time to form channel preferences with this firm and were still responsive to marketing (Valentini, Montaguti and Neslin 2011). However, during the acquisition phase, in order to attract new customers the cooperating firm offers several price and quantity discounts. This generally affects the frequency of purchase in the first post-acquisition period and might have reduced the number of multichannel customers in the first period. In summary, this research builds the knowledge base in this critical area by demonstrating the effectiveness of multichannel campaigns, offering guidance on the design of those campaigns, and demonstrating the potential for targeting based on a priori estimated models of the customer’s multichannel potential. This clearly demonstrates the causal link from marketing actions to multichannel customer behavior to customer profitability. Marketing Science Institute Working Paper Series 23 References Abraham, Magid M. and Leonard M. Lodish (1993), “An Implemented System for Improving Promotion Productivity Using Store Scanner Data,” Marketing Science 12 (3), 248–69. Angrist, Joshua D., Imbens, Guido W., and Donald B. Rubin (1996), “Identification and Causal Effects Using Instrumental Variables,” Journal of the American Statistical Association 91, 444–455. Allison, Paul (1982), “Discrete-Time Methods for the Analysis of Event Histories,” Sociological Methodology 13, 61-98. Ansari, Asim, Mela, Carl F., and Scott A. Neslin (2008), “Customer Channel Migration,” Journal of Marketing Research 45 (1), 60-76. Blattberg, Robert C., Kim, Byung-Do, and Scott A. Neslin (2008), Database Marketing: Analyzing and Managing Customers. New York: Springer. Blattberg, Robert C., and Scott A. Neslin (1990), Sales Promotion Concepts, Methods, and Strategies. New Jersey: Prentice Hall. Boehm, Martin (2008), “Determining the Impact of Internet Channel Use on a Customer's Lifetime,” Journal of Interactive Marketing 22 (3), 2-22. Campbell, Dennis, and Frances Frei (2010), “Cost Structure, Customer Profitability, and Retention Implications of Self-Service Distribution Channels: Evidence from Customer Behavior in an Online Banking Channel,” Management Science 56 (1), 4-24. Cook, Thomas D., and Donald T. Campbell (1979), Quasi-Experimentation: Design and Analysis Issues for Field Settings. Boston: Houghton-Mifflin. Gedenk, Karen, and Scott A. Neslin (1999), “The Role of Retail Promotion in Determining Future Brand Loyalty: Its Effect on Purchase Event Feedback,” Journal of Retailing 75 (4) 433-459. Hitt, Lorin M., and Frances Frei (2002), “Do Better Customers Utilize Electronic Distribution Channels? The Case of PC Banking,” Management Science 48 (6), 732-748. Imbens, Guido W., and Joshua D. Angrist (1994), “Identification and Estimation of Local Average Treatment Effects,” Econometrica, 62 467–476. Jaccard James, and Robert Turrisi (2003), Interaction Effects in Multiple Regression. Sage University Paper series on quantitative applications in the social sciences (72). Thousand Oaks, CA: Sage. Marketing Science Institute Working Paper Series 24 Kumar V., and Rajkumar Venkatesan (2005), “Who Are Multichannel Shoppers and How Do They Perform? Correlates of Multichannel Shopping Behavior,” Journal of Interactive Marketing 19(2), 44-61. Loftus, Brad, Mulliken, John, and Jonathan Sharp (2008). The Multichannel Imperative. The Boston Consulting Group, 9/08. Maddala, G. S. (1983), Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press . Myers, Joseph, Pickersgill, Andrew D., and Evan S. Van Metre (2004, October), Steering Customers to the Right Channels. The McKinsey Quarterly. Neslin, Scott. A., Grewal, Dhruv, Leghorn, Robert, Shankar, Venkatesh, Teerling, Marije L., Thomas, Jacquelyn S., and Peter C. Verhoef (2006), “Challenges and Opportunities in Multichannel Customer Management,” Journal of Service Research 9(2), 95-112. Neslin, Scott A., and Venkatesh Shankar (2009), “Key Issues in Multichannel Customer Management: Current Knowledge and Future Directions,” Journal of Interactive Marketing 23 (1), 70-81. Thomas, Jacquelyn S., and, Ursula Y. Sullivan (2005), “Managing Marketing Communications with Multichannel Customers,” Journal of Marketing 69(4), 239 – 251. Valentini, Sara, Montaguti Elisa, and Scott A. Neslin (2011), “Decision Process Evolution in Customer Channel Choice,” Journal of Marketing 75(November), 72-86. Venkatesan, Rajkumar, Kumar, V., and Nalini Ravishanker (2007), “Multichannel Shopping: Causes and Consequences,” Journal of Marketing 71(2), 114-132. Verbeek, Marno (2004), A Guide to Modern Econometrics. Chinchester, UK: Wiley. Verhoef, Peter C., Neslin, Scott A., and Björn Vroomen (2007), “Multichannel Customer Management: Understanding the Research-Shopper Phenomenon,” International Journal of Research in Marketing 24 (2), 129-148. Wooldridge, Jeffrey M. (2002), Econometric Analysis of Cross-Section and Panel Data. Cambridge, MA: MIT press. Xue, Mei, Hitt Lorin M., and Pei-yu Chen (2011), “Determinants and Outcomes of Internet Banking Adoption,” Management Science 52(2), 291-307. Marketing Science Institute Working Paper Series 25 Table 1: Percentage of Customers Who Become Multichannel by the End of the Observation Period, by Treatment Group Treatment Groups MF VPF MNF VPNF Control Total n Percent 7063 7074 7068 7042 3533 31780 SE 7.7% 7.6% 8.4% 7.6% 7.5% 7.8% Marketing Science Institute Working Paper Series .3% .3% .3% .3% .4% .2% 26 Table 2: Average Cumulative Profits per Customer as of the Last Observation Period p-value a Average Cumulative Profits n ROIb 7063 .665 MF € 17.50 -142% 7074 .515 VPF € 17.68 -104% 7068 .078 MNF € 18.34 40% 7042 .524 VPNF € 17.67 -106% 3533 Control € 17.69 a Two sample t-test: we tested the difference between the average cumulative profit of each treatment and the average cumulative profit of the control group. b ROI= [(Unit ProfitTreatment- Unit ProfitControl) – Treatment Cards Cost]/Treatment Cards Cost Marketing Science Institute Working Paper Series 27 Table 3: Estimates for Random Effects Probit Models Predicting the Probability of Becoming a Multichannel Shopper. Variable MF (λ1) VPF (λ2) MNF (λ3) VPNF (λ4) Model B: Model A: Basic Basic Model + Model Interactions .01 .03 Model C: Basic Model +αt .03 Model D: Full Model .05 (.03) (.03) (.05) (.05) .01 .01 .01 .03 (.03) (.03) (.05) (.05) .05* .09* .10** (.05) (.05) .01 .06* (.03) .02 .02 .04 (.03) (.03) (.05) (.03) *** Potential (δ) MF*Potential (κ1) VPF*Potential (κ2) MNF*Potential (κ3) VPNF*Potential (κ4) α1 α2 α3 α4 α5 Constant (α) Observations Log likelihood Likelihood-ratio test of nested vs. full model * p<.10, ** p<.05, ***p<.01 2.81 3.25 (.19) (.39) *** 5.17 (.05) *** (.11) (.72) -.76* - (.43) -1.39* - (.81) -.69 -.39 - (.43) - (.81) -.21 - (.43) -.23 - (.81) -.59 - (.43) - - -.91 -4.18*** -4.19*** (.17) (.17) - - - - - (.81) *** -3.02*** (.12) (.12) -2.98*** -2.99*** -3.03 - 5.89*** (.10) (.11) -3.06*** -3.07*** (.10) (.10) -3.04*** -3.05*** -2.23*** -2.22*** (.10) (.10) (.05) (.03) - - 154455 -12445.9 154455 -12443.6 χ2(9)=689.67*** χ2(5)=685.04*** Marketing Science Institute Working Paper Series 154455 -12105.9 χ2(4)=9.58** 154455 -12101.1 - 28 Table 4: Estimates for Random Effects GLS Regression Predicting Profits Variable Coef Seasonal (γ1) 7.43*** Trend (γ2) -1.31*** Covariance between the -.05 error terms (σ) Multichannel impact (β) 8.45*** Dependent Variable: Profitit Sample (Multichannel=1) Observations:= 6255 Wald χ2(3) =219.28*** * p<.10, ** p<.05, ***p<.01 Marketing Science Institute Working Paper Series SE .80 .10 p-value .00 .00 .27 .86 1.22 .00 29 Figure 1: Marketing Communication Treatments Financial Multichannel Incentive (MF) Financial Value Proposition Incentive (VPF) Non-Financial Multichannel Message (MNF) Non-Financial Value Proposition Message (VPNF) Marketing Science Institute Working Paper Series 30 Figure 2: Experimental Design-Timeline Marketing Science Institute Working Paper Series 31 Figure 3: Test Conditions Multichannel/ Financial (MF) Group MF: Multichannel Financial Group 7,064 customers Value Proposition/ Financial (VPF) Group VPF: Value Proposition Financial Group 7,074 customers Multichannel/ Non-Financial (MNF) Group MNF: Multichannel Non-Financial Group 7,068 customers Value Proposition/ Non-Financial (VPNF) Group VPNF: Value Proposition NonFinancial Group 7,043 customers No-Campaign Control Control Group No-Campaign Control Group 3,534 customers Marketing Science Institute Working Paper Series 32 Figure 4: Lift Chart Performance of Multichannel Potential Model – Cohort 1 Marketing Science Institute Working Paper Series 33 Figure 5: Cumulative Percentage of Customers Who Became Multichannel in Each Period, by Treatment Marketing Science Institute Working Paper Series 34 Figure 6: Difference in Cumulative Profits per Customer vs. Control Group - Full Sample Marketing Science Institute Working Paper Series 35 Figure 7: Differences in Cumulative Average # of Purchase Occasions per Customer vs. Control Group at the End of the Observation Period Marketing Science Institute Working Paper Series 36 Figure 8: Difference in Average Cumulative Revenues per Purchase Occasion per Customer vs. Control Group at the End of the Observation Period Marketing Science Institute Working Paper Series 37 Figure 9: Significance of Treatment Effects at Different Values of Multichannel Potential Marketing Science Institute Working Paper Series 38 APPENDIX Table A1: Independent Variables Used in the Multichannel Potential Model Class Variable Description Age of the customer XAcquisition,h Age Female North Big City Average City Early Email Franchisee Street Agent Nov Acquisition Dec Acquisition CAcquisition,h Initial Catalog Initial Web Initial Store Initial Phone ZAcquisition,h Initial Store Promo Initial Returns Initial Price Cut Initial Revenues Initial Purchase Trendt Trend Dummy variable that takes value 1 if the customer is a female and 0 otherwise Dummy variable that takes value 1 if the customer lives in the north of the country, and 0 otherwise Dummy variable that takes value 1 if the customer lives in a big city (more than 500 thousand inhabitants), and 0 otherwise Dummy variable that takes value 1 if the customer lives in an average city (499 – 100 thousand inhabitants), and 0 otherwise Dummy variable that takes value 1 if the customer provided her email address during the acquisition quarter, and 0 otherwise Dummy variable that takes value 1 if the store closest to the customer place of residence is run by a franchisee, and 0 if the closest store is run directly by the firm. Dummy variable that takes value 1 if the customer was acquired through an on-the-street agent, and 0 otherwise Dummy variable that takes value 1 if the customer was acquired in November, and 0 otherwise Dummy variable that takes value 1 if the customer was acquired in December, and 0 otherwise Dummy variable that takes value 1 if the customer purchased through the catalog during the acquisition time unit and 0 otherwise Dummy variable that takes value 1 if the customer purchased through the Internet during the acquisition time unit and 0 otherwise Dummy variable that takes value 1 if the customer purchased through the store during the acquisition time unit and 0 otherwise Dummy variable that takes value 1 if the customer purchased through the phone during the acquisition time unit and 0 otherwise Dummy variable that takes value 1 if in the closest store was running special store promotions, and 0 otherwise Total value (€) of products returned to the firm by the customer i in the acquisition period. Total value (€) of price discounts used by the customer i in the acquisition period Total amount (€) spent during the acquisition quarter Dummy variable that takes value 1 if the customer purchased during the acquisition quarter, and 0 otherwise Linear trend term Marketing Science Institute Working Paper Series 39 Table A2: Estimates for Random Effects Probit Model Predicting the Probability of Becoming a Multichannel Shopper – Multichannel Potential Model Variable Age Female Street Agent North Early Email Nov Acquisition Dec Acquisition Big City Average City Franchisee Initial Catalog Initial Web Initial Store Initial Phone Initial Store Promo Initial Returns Initial Price Cut Initial Revenues Initial Purchase Trend Constant Coef. .004 .111 -.156 .231 .242 .287 .386 -.081 -.102 -.041 2.198 2.046 .815 1.672 -.053 -.010 -.010 .002 -.475 .092 -3.661 Marketing Science Institute Working Paper Series SE .001 .027 .027 .034 .030 .034 .043 .044 .048 .025 .138 .132 .086 .114 .042 .003 .003 .001 .084 .012 .144 p-value .000 .000 .000 .000 .000 .000 .000 .065 .035 .096 .000 .000 .000 .000 .209 .001 .000 .152 .000 .000 .000 40