1
Technical Appendix
1. Note on implementation of our conceptual framework
The data on product consideration are particular to consumer, product category and country
of origin. We discuss below how consumer differences, such as home-country bias, are taken
into account and focus for the moment on a two-way table/matrix X of countries-of-origin by
product-categories that gives the percentage of all consumers who would consider a product
from each country for each category. The inter-relations of countries and products are explained
by the interaction of countries and product categories, each with estimated locations in M latent
dimensions. Plotting countries and product categories in M dimensions reveals a lot about COO
effects—a country will be considered often for a product if both country and product depart from
the origin in similar directions (product-country “fit” of Roth and Romeo 1992). This plot will be
referred to as the “map” of COO effects.
Given enough dimensions, the map is capable of fitting any matrix X exactly. However,
adding additional terms to the model allows the map to explain only what variability remains.
What terms to add depends upon what the map is meant to explain. For example, consumers will
be choosier about country of origin for automobiles than pencils, but this is a characteristic of
product category that need not be explained by a map of the competitive positions of countries of
origin. Therefore product category main effects are added separately to the model, relieving the
map of the burden of representing these effects.
There is also good reason to expect that some countries will be considered for more product
categories than other countries, a country main-effect. Verlegh and Steenkamp (1999) find that
COO effects are significantly larger for products made in more developed countries than for
products made in less developed countries, and “[i]n Asia, Japan emerges as the quality leader”
2
(Gurhan-Canli and Maheswaran 2000, p. 309). We believe that a map explaining COO effects
must be able to explain why some countries are considered more often than others. Therefore we
“force” the map to explain these differences by not including country main effects separately in
the model.
So far the model for the table X of data on product consideration (which includes column
main effects, no row main effects, and estimated locations of rows and columns on M latent
dimensions) is known as a “columns-regression biadditive model” (cf. Gower and Hand 1996,
Eq. 8.14(ii)). We extend this model of product consideration by adding the effects of nationality
of respondent (considered below) and random effects.
Random effects are added to the model so that the map need explain only systematic
variability in COO effects. Some COO effects are very product specific. For example, Russia
may be associated with poor quality generally, except in the case of caviar where produce
labeled with ‘Russia’ or ‘USSR’ fetches much higher prices than caviar from other countries
(Beverland and Lindgreen 2002). It is asking a lot of a macro-level map of COO effects to
account for this degree of detail. Latent constructs are not expected to explain all variability in
their indicators, only common variation (Harman 1976). Statistical evidence of a latent construct
requires that it add to a model that already allows the indicators to vary randomly and
independently. We take this same approach by adding to our model a random effect for
consideration of each combination of product category, country of origin and nationality of
consumer.
2. Note on Effects of consumer nationality
While the goal is to obtain a single competitive map that explains COO effects in a region,
there are likely to be differences arising from the different nationalities of respondents. As
3
Verlegh and Steenkamp point out, emotional and affective attachment may be formed “in direct
experiences during holidays or encounters with foreigners, but also in indirect experiences with
countries and their citizens through, e.g., art, education and mass media” (1999, p. 526). Such
experiences will vary by individual in ways that are related to their nationality. The model
includes such differences in two ways. The first is random. The random effects already described
for product-country considerations and country-attribute perceptions are applied separately for
each nationality. In addition, the relevance of the common map of COO effects is allowed to
vary by nationality, and separately for product consideration and country perceptions. If these
four provisions for differences by nationality are insufficient, then the result will tend towards of
map of COO effects that fits no nationality well and will lack coherence.
3. Note on error term specification and model identification
In our model, rpcn is the usual error term in a random utility model that allows two
apparently identical respondents to nevertheless give different answers to the same question. pcn
is a random effect, specific to each combination of product, country and nationality, that allows
the expected utility to differ from what is predicted by the model. Inclusion of this random effect
is important. Including the random effects (pcn) means that the map will be statistically
significant if and only if it is needed to account for nonrandom variability. In other words, if
COO effects vary by product, but with no apparent pattern from one product to the next, then the
random effects (pcn) will have a large variance and the map (pc) will be unnecessary.
Model identification was attained by imposing necessary constraints. The logistic error
terms (rpcn and racn) have their variances fixed at 2/3 = 3.29 (a value commonly used for the
error term in logistic regression). The vectors of nationality main effects ((n1) and (n2)),
thresholds ((p) and (d)), and country locations in each dimension of the map ((c1) and (c2))
4
each sum to zero. The geometric means of map relevance over nationalities ((n1) and (n2))
equal one for each question. And finally, the 20  M matrix of product and country attribute
locations in the map (B, consisting of (p) and (a)) satisfies BB = I and the 10  M matrix of
country locations (  (c)) satisfies  = diag(v) with vm > vm+1 for m = 1, …, M – 1.
These last constraints on the map parameters mean that the locations of products and
country attributes in the map (B) are orthogonal, just like the loadings in a principal components
analysis (PCA). If the country attributes are related to product categories, they will load on the
same dimensions; otherwise they will not. The variances in country locations (v1, v2, …, vM)
necessarily decreases with each dimension. The first dimension explains maximum variance in
countries and thus in the data as whole, the second the maximum of remaining variance, etc., and
the importance of a dimension is given by the variance of the countries on that dimension.
4. Note on Bayesian analysis and determining the number of dimensions for COO effects
Bayesian models require the specification of prior distributions for model parameters; in
accordance with common practice we chose uninformative priors so that inferences about
parameters are based solely on the data. Bayesian estimation is accomplished by generating
many simulates that ultimately come from posterior distributions that reflect what has been
learned about the parameters based on the data (Gelman et al. 2004). All parameters are
simulated in unison; each iteration of the simulator generates one simulate for every parameter.
The simulation consists of two phases. Initially, enough simulates must be generated to
determine that they are being generated by the true posterior distributions (“convergence”).
Subsequent simulates are used as the basis for inference.
The model of Equations Error! Reference source not found.-Error! Reference source
not found. was estimated for different numbers of dimensions (M) to the map, beginning with
5
one and adding additional dimensions one at a time as long as it resulted in substantial
improvement in a Bayesian deviance information criterion (DIC) developed by Spiegelhalter
(1998; Spiegelhalter et al. 2002). DIC generalizes the classical Akaike information criterion to
hierarchical modeling. Models with lower DIC values are better supported by the data. Models
with DIC values at least 10 higher than the best DIC can be ruled out from further consideration,
whereas DIC differences less than 3 are inconclusive (Burnham and Anderson 1998; Congdon
2005; Spiegelhalter et al. 2002). For all models, no more than 20,000 iterations were needed to
establish convergence according to the Gelman-Rubin statistic as modified by Brooks and
Gelman (1998). Inference was based on a further 100,000 iterations.
5. Note on assessment of the selection of product categories and country attributes
The first two dimensions of systematic effects appear to explain very well the relations
among countries, product categories and country attributes. But the product categories differ in
terms of the extent of COO effects they exhibit, some country attributes are more strongly related
to country image than others, and the third dimension has identified items that contain systematic
information missed by the first two dimensions. These three issues can be addressed by
considering the last two columns of Table 2 labeled “country variance.” The first of these two
columns shows the variance countries as fitted by the first two dimensions (“the map”), and the
last column shows the variance in countries for these same items in the third dimension. It is
immediately apparent that the third dimension explains little of the variability in the data, and the
total variance in countries over all three dimensions, which is the sum of these two columns, is
nearly identical to the first column.
Small country variances for product categories mean that COO effects for these categories
are small. This does not imply, however, that these categories should be excluded from studies in
6
future. Countries in disadvantaged COO positions are most interested in the short run in such
categories.
One product category — airlines — is appreciably different from zero on the third
dimension. Although this dimension explains little variance among countries overall, it accounts
for some variability for airlines. This is a product category that shows moderately strong COO
effects and is somewhat different from the others. It has the most extreme value on the third
dimension and the value is negative. The only other service (banking) also has a negative
estimated value, as do cars. This suggests perhaps that a fuller understanding of COO effects
would be obtained by including a greater number of services and/or higher-cost (higher risk?)
categories. The extensive COO literature offers theoretical guidance on additional products to
include in future. What is suggested by the results here is that the airline category shows
moderately strong COO effects that are only partly accounted for by the Fashion and Technology
dimensions.
Examination of the last columns of Table 2 also helps evaluate and refine the choice of
country attributes studied. The country variances on the attributes as fitted by the map are
generally larger than for product categories. (Since evaluations and perceptions were measured
on a common scale, these differences in magnitudes are meaningful.) The two attributes showing
the weakest relation in the map to country image are “fun” and “approachable”, with the latter
loading highest on the third dimension. Thus “approachable” is related to country image but not
very strongly to the Technology and Fashion dimensions. It may also be related to the airline
product category but this is uncertain. Here too theoretical guidance from the country image
literature is helpful. (cf. Roth and Diamantopoulos 2009; Zeugner-Roth and Diamantopoulos
2010). The analysis here has revealed two dimensions of country attributes that are related to
7
country image and are relevant for explaining COO effects. The study of country image in
isolation does not assure such relevance (Samiee 2010; Usunier 2006).