Content uploaded by Michael R. Ward
Author content
All content in this area was uploaded by Michael R. Ward on Oct 18, 2019
Content may be subject to copyright.
Netnomics 3, 103–117, 2001
2001 Kluwer Academic Publishers. Manufactured in The Netherlands.
Will Online Shopping Compete More with Traditional
Retailing or Catalog Shopping? ∗
MICHAEL R. WARD ward1@uiuc.edu
Department of Agricultural and Consumer Economics, University of Illinois, Urbana-Champaign,
1301 W. Gregory Dr., MC-710, Urbana, IL 61801, USA
Abstract. This paper examines consumer substitution behavior among three distribution channels: online
shopping, traditional retail and catalog shopping. A model of consumer transactions costs is presented in
which consumer investments in shopping human capital determine distribution channel choice. This model
predicts that consumers who make investments that “spillover” to multiple channels will tend to have lower
transactions costs for those two channels and will tend to consider these two channels to be closer substitutes
than other channels. The model implies that, even if these investments are not measured, they will represent
a component of the regression models’ residual terms, which will predict distribution channel choice. I test
this implication using the GVU Center’s survey data on channel selection for various product categories.
The results strongly suggest that consumers consider online shopping and catalog shopping to be closer
substitutes than any other pair of channels.
Keywords: e-commerce, distribution channel, consumer choice
1. Introduction
The marketing literature has made useful distinctions between distribution channels such
as traditional retailing and direct-to-customer marketing (e.g., Balasubramin [2]). Mar-
keting over the Internet, or e-commerce, represents a potentially important and little
understood new marketing channel.1This paper explores consumers’ preferences across
a set of marketing channels which includes e-commerce. Specifically, I explore the de-
gree to which e-commerce will compete with traditional retailing or direct-to-customer
marketing for a series of product categories. While different patterns are found for dif-
ferent product categories, in general online shopping is shown to be a closer substitute
for direct-to-customer marketing than for traditional retailing.
I present a model of transactions costs in which consumer investments in shopping
human capital, differences in the type of products purchased, as well as their costs of
acquiring human capital, all determine their distribution channel choice. This model
predicts that consumers who make large investments that “spillover” to two channels will
tend to have lower transactions costs for those two channels and will tend to consider
∗I would like to thank Michael J. Lee for excellent research assistance, the participants of the Second Berlin
Conference on Internet Economics and the 27th Annual Telecommunications Policy Research Conference
and an anonymous referee for their comments. The research was supported by C-FAR Project No. 99I-
027-2.
104 WAR D
these two channels to be closer substitutes than other channels. The model implies that if
these specific investments are important they will represent a component of the residual
terms of regression models which can predict distribution channel choice.
This paper employs this variance component methodology using data from Geor-
gia Institute of Technology’s Graphics, Visualization and Usability (GVU) Center on
e-commerce usage GVU [7].2The results of the GVU Center’s World Wide Web User
Surveys, begun in 1994, have been made publicly available for academic research. These
semiannual surveys record users opinions and usage patterns for a large number of Web-
based activities. One part of the survey asks about purchase behavior for three distri-
bution channels (traditional retailing, catalog shopping and online shopping) across a
number of product categories.
Substitution patterns among marketing channels are estimated by employing dis-
crete choice models of consumer choice. The specification includes fixed effects for
individuals and for product categories for each of three distribution channels. Substitu-
tion patterns are uncovered by variance components methods applied to the error terms
of distribution channel choice equations. In particular, investments in human capital that
span two channels, but not the third, will tend to result in a positive correlation in the
residuals for those two channels.
For all but one of the seventeen product categories available, online shopping is es-
timated to be a greater substitute for catalog shopping than for traditional retailing. These
results may only pertain to U.S. consumer behavior since survey responses come pre-
dominately from U.S. respondents. Reasons for differences across product categories are
explored and seem consistent with differences in markets and product attributes across
product categories. For example, this result seems to be stronger for some product cate-
gories in which e-commerce is more established (e.g., computer hardware and software)
but, somewhat surprisingly, not for others (e.g., books and music).
2. Discussion
Some products are better suited for some distribution channels than others. When asked
what products Amazon.com would not sell over the Internet, its Chief Financial Officer
simply replied “Cement”. This anecdote highlights one of the main differences between
online retailing and traditional retailing: when an online shopper commits to purchasing,
the product still must be delivered. For some products, delivery is efficiently integrated
into sellers’ retailing functions (e.g., building materials, white goods) or naturally fol-
lows consumers’ product selection task (e.g., groceries). Generally, products with es-
pecially high delivery costs will tend to have higher transactions costs when purchased
either over the Internet or through catalog shopping. Thus, for products with relatively
high delivery costs, both catalog shopping marketing and online shopping are likely to
be poor substitutes for traditional retailing.
In fact, the important differences between online shopping and traditional retailing
tend to highlight the relative similarities between online shopping and catalog shopping.
First, unlike traditional retailing, for both online shopping and catalog shopping, pos-
DISTRIBUTION CHANNEL COMPETITION 105
session of the product is not transferred when the terms of the transaction are agreed
upon. Second, since online or catalog purchases typically involve payments by credit
card, they both require a certain degree of customer familiarity with credit and financial
matters. Third, for purchases in which minimal search is required, both are likely to
involve lower transactions costs than traditional retailing. And finally, both rely on pho-
tographic images or descriptions of the product being sold rather than direct consumer
contact with the product being purchased. These shared characteristics should suggest
that online retailing is likely to compete more directly with catalog shopping than with
traditional retailing.
To be sure, both online and catalog shopping have their own comparative advan-
tages. Online shopping requires that the shopper interact with a computer, usually at
a fixed location, while catalogs can be browsed through more leisurely. Catalog shop-
pers are limited to those vendors for whom they have catalogs, while online shoppers
can easily search across previously unknown vendors.3Direct marketers incur a con-
stant marginal cost – postage – to inform their potential customers, whereas Internet
marketers’ costs of informing customers are associated with their websites, which are
likely to be more fixed relative to the number of customers they reach. These differences
suggest that both online and catalog shopping have specific niches for specific types of
consumer purchase decisions.
Ideally, one would like to list and measure the relative advantages and charac-
teristics associated with each distribution channel. One could then list the features of
particular products and the endowments of individuals making purchases. Economic
theory would predict that products will be purchased via the channel whose character-
istics tend to minimize the transactions costs incurred due to the product features and
the purchasers’ endowments (e.g., Lancaster [8]). However, this paper does not do this.
Instead, I attempt to draw inferences by modeling the pattern of actual channel choices
that would emerge from differences and similarities in these characteristics across chan-
nels. In this regard, I do not identify which characteristics of the channels lead them to
be substitutes, only that they are substitutes.
3. Economic model
I assume that consumers wish to minimize the expected transactions costs of making
purchases. These costs include both the price of the good being purchased and other non-
pecuniary costs incurred by the shopper. Since the paper’s focus is comparisons between
different marketing channels, differences in these transactions costs across channels will
be most important. These might include differences in delivery costs, search costs, risk-
iness of the seller, enjoyment of shopping and other factors affecting the attractiveness
of using different channels.
I explicitly model how consumer heterogeneity can lead to different consumer
choices and how this can imply distinct substitution patterns across channels. The
source of the heterogeneity in consumers could be with regard to their preferences, as
in Pashigian and Bowen [10]. However, differences in their information gathering costs,
106 WAR D
as in Banabou [2,3] and Carlson and McAfee [6] could also yield similar results. Infor-
mation about an individual’s purchasing patterns in other product categories is used to
model her purchase probability in a particular product category. In this regard, the model
is similar to Ainslie and Rossi [1], Rossi, McCulloch and Allenby [11], and Messinger
and Narasimhan [9].
Consumers attempt to minimize the sum of expected future transactions costs by
choosing human capital investments that tend to reduce these costs. Some investments
are better suited to specific distribution channels. Consumers who invest more heavily in
these types of human capital will, when making a purchase decision, tend to favor those
channels whose transactions costs are reduced the most by these investments. These
channels will tend to be closer substitutes with each other than another channel that is
not affected by these investments.
Denote transactions costs as TC. Consumers expect to make a number of purchases
over the foreseeable future. The transactions costs for these purchases can be reduced
by appropriate investments in human capital, a vector H. For the present, it will be ad-
vantageous for us to think of generic human capital, H, that can simultaneously affect
transactions costs across different product categories and distribution channels to vary-
ing degrees. Individuals with greater levels of human capital that lowers transactions
costs for purchases in product category jvia channel c,TCjc, will tend to have lower
transactions costs of purchasing goods in that product category via that channel. The
investment costs of obtaining a given level of human capital, H, depend on specific ca-
pabilities of the individual that are signaled by her demographic characteristics, ICi(H )
for individual i.
Formally the consumer’s problem is:
Min w.r.t.H, E
i
jc
njcTCjc(H )−ICi(H ),
where njc is the number of purchases made in category jvia channel c. Assuming
certain regularity conditions hold, the first order conditions for a local minimum are
Ei
jc
njcTCjc
H−ICi
H=0,
where the subscripts denote partial derivatives. These yield optimal human capital
choices Hi∗and hence expected optimal levels of transactions costs TCjc(H i∗).Di-
minishing returns to human capital or eventual increasing marginal costs of acquiring
human capital insures that Hi∗will be an interior point. Actual choices, and hence
transactions costs, can differ across individuals due to differences in their costs of accu-
mulating human capital, ICi(H ) (see figure 1).4
The actual transactions costs for a particular purchase may differ from TCjc(H i∗)
due to the heterogeneity of products within categories. Some products will be more
amenable to purchases using one distribution channel while others are amenable to other
channels. Or, knowing the product category is not enough to predict which channel will
DISTRIBUTION CHANNEL COMPETITION 107
Figure 1. Choosing the transactions cost minimizing investment.
Figure 2. Channel choice as a function of investment.
be used. Following the model from above, consumers will evaluate the transactions costs
associated with the characteristics of the purchase and her own level of human capital,
H. The actual shape of the TC curves will depend on the particular purchase. In figure 2
the minimum of transactions costs occurs with different channels for different levels
of H. The regions PR,PDand POrefer to the areas in which minimum transactions
costs are found when using traditional retailing, direct marketing5and online purchasing
respectively.
So far, human capital investments have been treated generally. However, the mag-
nitude of these investments on transactions costs will depend on the particular charac-
teristics of the product category jand channel cin which the purchase is being made.
For example, while investments in human capital in the ability to discern product quality
may lower the transactions costs of traditional retailing in general, they will lower them
108 WAR D
Figure 3. Change in channel choice due to change in investment.
more for product categories that rely more heavily on these skills (e.g., jewelry purchases
rather than airline reservations). This implies that the derivative TCjc
H<TCkc
Hwhere j
and krefer to jewelry and travel arrangements and the relevant His, say, the ability to
discern quality merchandise. Moreover, particular investments are likely to yield returns
in related areas. For example, investments in the ability to use a computer will facilitate
online shopping for multiple product categories. In this case, TCjc
Hand TCkc
Hhave sim-
ilar magnitudes where the particular Hrefers to computer skills. Likewise, an ability
to evaluate products from descriptions rather than in person will facilitate both online
shopping and direct-mail relative to traditional retailing. This would imply that TCjc
H
and TCjd
Hhave similar magnitudes where cand drefer to different distribution channels.
Variation in the effect of human capital on transactions costs, TCjc
H, and in optimal
investments, Hi∗can yield interesting patterns among actual channel choice. In partic-
ular, consumers with high levels of investment Hi∗will tend to have lower transactions
costs for purchases in product categories and channels that are particularly amenable to
this sort of investment. These consumers are more likely to substitute between these two
channels than with a third that does not benefit from these specific investments. Figure
3 depicts the optimal channel choice as a function of an element of H. However, other
investments, perhaps computer skills, have shifted down the TC curves for direct mar-
keting and online shopping more than traditional retailing. Hence, for more values of H,
she will choose direct marketing or online shopping. We expect an increase in purchases
from these two channels at the expense of traditional retailing.
4. Data description
To empirically examine substitution patterns, I make use of data from Georgia Institute
of Technology’s Graphics, Visualization and Usability Center (GVU) on e-commerce
usage. These data are from four semiannual online surveys of e-commerce, traditional
DISTRIBUTION CHANNEL COMPETITION 109
Tab l e 1
Descriptive statistics of shopping behavior from GVU 5-8 (April 1996–Sept. 1997).
Categories Retail Direct Online
Computer hardware (<$50) 59.7% 33.4% 18.1%
Computer hardware (>$50) 58.2% 38.0% 21.2%
Computer software (<$50) 37.9% 37.2% 32.4%
Computer software (>$50) 52.6% 36.9% 23.7%
Home electronics (<$50) 63.5% 17.0% 5.4%
Home electronics (>$50) 62.3% 16.7% 6.0%
Legal services 11.4% 1.0% 1.5%
Food 81.8% 8.2% 3.3%
Investment choices 16.2% 2.8% 9.9%
Videos/Movies 69.1% 15.7% 7.3%
Music (CD, tapes, albums) 75.1% 28.9% 16.7%
Books/Magazines 81.8% 29.7% 23.9%
Concerts/Plays 37.9% 2.9% 6.8%
Travel arrangements 39.2% 3.8% 23.8%
Apparel/Shoes 79.9% 31.2% 6.9%
Web services 7.6% 5.4% 38.0%
Sunglasses/Personal items 67.7% 10.7% 3.0%
This table reports the percentage of respondents who have purchased in each prod-
uct category for the three distribution channels. Sample size is 10,059.
retailing and direct-mail usage from 1996 through 1997. Surveys contain information
regarding sources of product information and choice of channel for product purchases
for up to two-dozen product categories. In total, about 10,000 individuals over four
different time periods answered questions about three channels for 17 product categories.
Of particular interest for this study are answers to whether the respondent has purchased
a product from that category via that distribution channel in the past six months. The
fraction of times respondents claim to have purchased a product from each category via
each channel in the previous six months are summarized in table 1.
These survey data pose two general problems to researchers. First, collected at
high-exposure sites, the data may not fully represent the characteristics of the population
of interest – all individuals who use the World Wide Web. Only those with a disposition
toward highly trafficked sites such as Netscape and Yahoo! are likely to be included in
the samples. Second, only those willing to spend the time filling out the questionnaires
were included. This creates a self-selection problem: those who answer the surveys may
not represent the population of interest.
While inference testing should be robust to these problems of non-probabilistic
sampling and self-selection, forecasts likely are not. Being a Web-based survey causes
the GVU surveys to over-sample consumers who have the option of making Web based
purchases. The self-selection bias may cause an over-representation of users who have
made purchases: those interested in the issues of the survey – i.e., those who have pur-
chased online – are more likely to fill out the often lengthy surveys. Bias in the sample
of Internet users in the GVU surveys should not affect within sample comparisons, such
110 WAR D
as when performing inference tests. However, since the sample is not representative, one
should be cautious about making out-of-sample forecasts from the results.
5. Empirical methodology
Consider an individual’s choice of distribution channel for a purchase decision in a par-
ticular product category. Denote E[TC
ij c ]to be individual i’s expected transactions cost
when purchasing in product category jvia distribution channel c. She will choose the
channel that minimizes transactions costs,
δjcHi∗=1ifTCij c Hi∗TCij d Hi∗,all d,
0 otherwise. (1)
We do not observe actual human capital investments, Hi∗. However, we do observe who
the individual is and which product category she is considering. These two factors is
likely to account for most of the variation in transactions costs. For example, the TC
curve for online purchases of any product is likely to be much lower for an individual
who works in the computer industry. Likewise the TC curve for online purchases of
books appears to be much lower than that for foodstuffs for all individuals. Thus, a
linear approximation of transactions costs can be employed,
TCij c =
I
i=1
αicdi+
J
j=1
βjcdj+εij c ,
where diand djare dummy variables for individual iand product category jrespectively
and αic and αjc are shift parameters. Together with equation (1) above, we have
δij c =1ifαic +βjc +εij c αid +βjd +εij d ,all d,
0 otherwise.
The actual choice is a random variable with mean E(δjc)≡πjc given by
πij c =Pr{αic +βjc +εij c αid +βjd +εij d ,all d}.
Empirically, the observable distribution channel choice, rather than the unobservable
transactions costs, is modeled as depending on individual and category effects. The αic’s
represent the individual’s average propensity to purchase via channel cover channel
dand the βic’s represent the difference in the average propensity of shoppers to use
channel cfor product category j.Theαic’s will depend on individuals’ human capital
investments while the βic’s will depend on the amenability of investments to that product
category. Part of the error term is due to heterogeneity within a product category that
makes some goods better or worse suited for the channel than the average in the category.
Of greater interest is the part of the error term that reflects specific human capital
investments that will “spill over” to some product categories or channels beyond the
average effect. Greater human-capital investments in computer proficiency are likely to
lower the transactions costs of purchasing online more than average in those categories
DISTRIBUTION CHANNEL COMPETITION 111
that lend themselves to online shopping, e.g., books and music. Note that the β’s for
online purchasing of books and music are likely to be larger than for other categories,
but these β’s represent the average over all individuals. Those with greater computer
skills are likely to have a positive component of their error term for all such product
categories (and, conversely, a negative one for categories that do not lend themselves
to online shopping). Thus, a positive correlation between the error terms of categories
suggests that the transactions costs for these categories depend non-trivially on similar
investments in human capital.
A different component of the error term, the component that “spills over” chan-
nels, can be used to draw inferences about substitution across channels. Suppose one
particular form of human capital has a larger effect in two channels than a third. Expe-
rience with credit card transactions or with delivery services may facilitate direct-mail
and online purchasing more than traditional retailing. If these effects are large, they may
be reflected in the error terms across channels being correlated. Again, the different
β’s for the different channels will predict channel usage for individuals with the aver-
age amount of these investments. Greater investment in this form of human capital is
likely to reduce transactions costs more than average for these two distribution chan-
nels. Those with greater investments are more likely to have a positive component of
their error terms for these channels. Thus, a positive correlation between the error terms
of channels, corr (εij c ,ε
ij d )>0, suggests that the transactions costs for these channels
depend on similar investments in human capital.
One complication of this argument for the GVU data is that the analysis assumes
that purchase channels are mutually exclusive, but the data are not. For any given pur-
chase, the selection of one distribution channel necessarily implies that other channels
are not selected. This adding up constraint tends to make the error terms for the differ-
ent channels negatively correlated. Inferences about the relative substitutability between
channels are drawn from the relative magnitudes of the correlations. A larger correlation
between channels A and B than between channels A and C implies that the human capital
that tends to lower the transactions costs for channel A tends to reduce the transactions
costs for channel B more than they reduce the transactions costs for channel C. This in
turn implies that channel B, rather than channel C, is a better substitute for channel A.
However, the survey question asks if respondents had made at least one purchase
in the previous six months in the product category using the channel. Respondents could
have made multiple purchases, using one or more channels. We actually observe whether
the rank-order statistic equals zero or not for each category. Assume that the error terms
for purchases within a category for a distribution channel are independently distributed.
Then the number of purchases follows a binomial distribution with probability πij c and
nij possible events in the previous six months. The probability of making more than zero
purchases is given by:
µij c =1−(1−πij c )nij .
112 WAR D
The actual choice is a random variable with mean E(µjc)≡θjc given by
θij c =1−1−Pr{αic +βjc +εij c αid +βjd +εij d ,all d}nij .(2)
Unfortunately, we do not observe nij . This will cause estimates of the α’s, β’s and,
in particular, the ε’s from equation (2) to be biased. We can expect nij to differ across
individuals. Holding the πij c constant, shoppers who make more purchases in a category
have a higher likelihood of having made at least one purchase in all channels. This causes
their θij c ’s to be above average for all channels and, consequently, the correlations across
channels to tend to be positive.
Despite this “size” effect, comparisons of correlations across channels can still
yield inferences about channels tending to be similarly affected by human capital invest-
ments and thus substitutable across channels. Since the bias induced by the “size” effect
is caused by dispersion in nij , categories in which this dispersion is likely to be small are
more immune from this problem. To gauge the magnitude of the “size” effect, results
from categories in which we expect wide dispersion in the number of purchases (e.g.,
food, videos, music) can be compared to those we would expect no more than a single
purchase (e.g., computer equipment over $50, legal services, investments).
To test the model, I propose a regression to predict the observed µij c ’s, dummy
variables indicating that individual ipurchased a product in category jvia channel c
during the previous six months. These are regressed against individual and category
dummy variables for each channel to yield (possibly biased) estimates of the average
propensity for each individual and for each category to purchase. This entails regressions
of 10,059 individual dummy variables and 17 category dummy variables against 171,003
observations.
Because of the computational burden imposed by such a large model, this was
estimated in two stages.
First, I account for individual effects by taking deviations from the individual
means. For each individual, the average usage of the channel across all 17 categories,
µij c , was calculated,
µic =1
J
J
j=1
µij c .
Deviations about this mean, µij c =µij c −µic, are then used in the subsequent analysis.
Second, these deviations are regressed against the product category dummy variables,
µij c =
J
j=1
βjc +dj+νij c .
The β’s represent the average propensity of consumers to use the channel to purchase
in the category after controlling for individual effects. Finally, the error terms from this
regression are used to calculate a correlation matrix across channels for each product
category.
DISTRIBUTION CHANNEL COMPETITION 113
6. Results
While I do not actually calculate the regression for the individual fixed effects, generat-
ing deviations from means will be equivalent. The R2can be calculated as the explained
sum of squares divided by the total sum of squares. The explained sum of squares equals
one minus ij µ2
ij c while the total sum of squares is ij µ2
ij c . For the GVU data, the
R2thus calculated are 0.78, 0.36, and 0.33 for traditional retail, direct marketing, and
online marketing respectively. This suggests that individual effects are rather important
in predicting purchase behavior and especially important for traditional retailing.
Table 2 reports results from regressions of µij c on dummy variables for each
category. The coefficients are meant to reflect the average consumer’s propensities to
purchase for each category. Categories containing products that lend themselves to the
distribution channel more should have larger propensities. In fact, adding the mean of
the individual effects, µic, to these parameters should equal the fraction of consumers
who have ever purchased in the category (table 1). These means are 0.545, 0.188 and
0.146 for traditional retail, direct marketing, and online marketing respectively.
Table 3 reports correlations between the νij c and νij d for each category jand each
pair of channels cand d. While one might infer from table 1 which channels are likely to
Tab l e 2
Coefficients from linear probability model.
Categories Retail Direct Online
Computer hardware (<$50) −0.079∗0.279∗0.151∗
Computer hardware (>$50) −0.095∗0.273∗0.181∗
Computer software (<$50) −0.056∗0.266∗0.293∗
Computer software (>$50) −0.151∗0.262∗0.206∗
Home electronics (<$50) −0.042∗0.064∗0.023∗
Home electronics (>$50) −0.054∗0.061∗0.030∗
Legal services −0.563∗−0.097∗−0.016∗
Food 0.141∗−0.025∗0.003
Investment choices −0.515∗−0.078∗0.069∗
Videos/Movies 0.014∗0.051∗0.042∗
Music (CD, tapes, albums) 0.074∗0.182∗0.136∗
Books/Magazines 0.141∗0.190∗0.209∗
Concerts/Plays −0.298∗−0.077∗0.037∗
Travel arrangements −0.285∗−0.069∗0.208∗
Apparel/Shoes 0.122∗0.205∗0.039∗
Web services −0.601∗−0.053∗0.349∗
Constant 0.132∗−0.081∗−0.115∗
Sample size 171,003 171,003 171,003
R-squared 0.28 0.15 0.12
Standard error of estimates 0.37 0.32 0.29
F statistic 4125.87 1879.20 1443.13
Significance of F 0.0000 0.0000 0.0000
Coefficients of category effects from linear probability model of deviations from individual means.
Asterisks denote statistical significance at the 1% level.
114 WAR D
Tab l e 3
Correlation coefficients of residuals.
Categories Online/Direct Online/Retail Direct/Retail
Computer hardware (<$50) 0.288∗0.069∗0.100∗
Computer hardware (>$50) 0.202∗−0.054∗−0.061∗
Computer software (<$50) 0.188∗−0.019+0.001
Computer software (>$50) 0.238∗0.034∗0.020+
Home electronics (<$50) 0.190∗−0.014 0.056∗
Home electronics (>$50) 0.221∗−0.026∗0.035∗
Legal services 0.280∗0.027∗0.080∗
Food 0.276∗0.085∗0.085∗
Investment choices 0.155∗0.022+0.048∗
Videos/Movies 0.238∗0.034∗0.053∗
Music (CD, tapes, albums) 0.183∗0.067∗−0.006
Books/Magazines 0.141∗0.048∗−0.020+
Concerts/Plays 0.182∗0.063∗0.011
Travel arrangements 0.069∗−0.005 −0.021+
Apparel/Shoes 0.223∗0.093∗0.178∗
Web services 0.061∗−0.139∗0.225∗
Sunglasses/Personal items 0.267∗0.044∗0.126∗
This table reports correlation coefficients between residuals from equation (3) predicting distrib-
ution channel usage. Asterisks and plus signs denote statistical significance at the 1% and 10%
levels.
be used for the average product in the category, table 3 gives some indication of which
channels are likely to be considered for the marginal purchase in a category. As can
be seen, the correlations are highest between online shopping and direct-mail for all
but one category. This result is consistent with direct marketing and online shopping
both benefitting from similar human capital investments and that the effects of these
investments are large. It suggests that online shopping may compete more directly with
direct marketing than with traditional retailing as a distribution channel.
Interestingly, for the category Web Services, another correlation dominates the one
between online shopping and direct-mail. Understandably, this is the category for which
respondents most often use online shopping (see table 1). This result suggests that there
is more substitution at the margin between direct-mail and traditional retail for these
services. I conjecture that this is caused by online shopping being far superior to any
other channel for those respondents familiar with online shopping. For the rest, direct-
mail and traditional retailing are good substitutes.
Finally, table 4 reports correlation coefficients across equations for each time pe-
riod. While online shopping is currently by no means ubiquitous as a distribution chan-
nel, it was in its true infancy in 1996. It would not be surprising to see that consumers
attitudes and usage of e-commerce would have changed over the four time periods. In
general, table 4 confirms the patterns detected in table 3. However, it also suggests that
the pattern has become stronger over time. This may reflect web usage spreading from a
select group to more wide-spread usage. Interestingly, for product categories that seemed
DISTRIBUTION CHANNEL COMPETITION 115
Tab l e 4
Correlation coefficients of residuals by time period.
Time Period Online/ Online/ Direct/ Online/ Online/ Direct/
Direct Retail Retail Direct Retail Retail
Computer hardware (<$50) Computer hardware (>$50)
Spring 1996 0.176∗0.077∗0.059+0.124∗0.00 −0.091∗
Autumn 1996 0.237∗0.018 0.101∗0.153∗−0.107∗−0.098∗
Spring 1997 0.347∗0.068∗0.065∗0.248∗−0.048∗−0.076∗
Autumn 1997 0.356∗0.123∗0.177∗0.268∗−0.035 0.033
Computer software (<$50) Computer software (>$50)
Spring 1996 0.076∗−0.079∗−0.020 0.134∗0.043+−0.009
Autumn 1996 0.132∗−0.052∗−0.043+0.200∗0.005 −0.024
Spring 1997 0.226∗−0.054∗−0.048∗0.283∗0.020 −0.004
Autumn 1997 0.311∗0.119∗0.144∗0.309∗0.088∗0.131∗
Home electronics (<$50) Home electronics (>$50)
Spring 1996 0.134∗−0.020 0.071∗0.134∗−0.045+0.009
Autumn 1996 0.136∗−0.024 0.070∗0.224∗−0.057∗0.043+
Spring 1997 0.234∗−0.004 0.038+0.264∗−0.029+0.024
Autumn 1997 0.234∗−0.010 0.046+0.217∗0.028 0.063∗
Legal services Food
Spring 1996 0.279∗0.010 0.064∗0.247∗0.104∗0.108∗
Autumn 1996 0.279∗0.009 0.086∗0.284∗0.054∗0.065∗
Spring 1997 0.304∗0.026 0.065∗0.295∗0.075∗0.084∗
Autumn 1997 0.276∗0.090∗0.089∗0.275∗0.125∗0.087∗
Investment choises Videos/Movies
Spring 1996 0.097∗0.022 0.016 0.180∗0.014 0.049+
Autumn 1996 0.177∗0.012 0.044 0.204∗0.003 0.041+
Spring 1997 0.159∗0.040+0.052∗0.285∗0.070∗0.082∗
Autumn 1997 0.169∗0.003 0.070∗0.250∗0.021 0.022
Music (CD, tapes, albums) Books/Magazines
Spring 1996 0.112∗0.066∗−0.010 0.142∗0.012 −0.008
Autumn 1996 0.184∗0.033+−0.035∗0.151∗0.030 −0.030
Spring 1997 0.225∗0.083∗0.015 0.163∗0.078∗0.004
Autumn 1997 0.167∗0.090∗0.008 0.102∗0.051+−0.053+
Concerts/Plays Travel arragements
Spring 1996 0.175∗0.045+−0.009 0.056+0.013 −0.017
Autumn 1996 0.170∗0.064∗0.018 0.040+−0.010 −0.034+
Spring 1997 0.215∗0.074∗0.007 0.073∗0.003 −0.031+
Autumn 1997 0.147∗0.056+0.019 0.100∗−0.022 0.006
Apparel/Shoes Web services
Spring 1996 0.216∗0.105∗0.204∗0.049+−0.147∗0.177∗
Autumn 1996 0.175∗0.080∗0.164∗0.064∗−0.128∗0.217∗
Spring 1997 0.265∗0.093∗0.187∗0.087∗−0.138∗0.246∗
Autumn 1997 0.264∗0.089∗0.144∗0.043+−0.125∗0.222∗
Sunglasses/Personal items
Spring 1996 0.206∗0.016 0.107∗
Autumn 1996 0.247∗0.037+0.139∗
Spring 1997 0.296∗0.042+0.113∗
Autumn 1997 0.306∗0.090∗0.140∗
This table reports correlation coefficients by time period between residuals from equation (3) predicting distri-
bution channel usage. Asterisks and plus signs denote statistical significance at the 1% and 10% levels.
116 WAR D
to achieve the more success online (e.g., books, music), the correlation between online
shopping and direct mail fell in the last period. This may indicate that these product
categories began to appeal to traditional retail customers as well as catalog shoppers.
7. Conclusion
This paper presented a simple model of consumer choice among distribution channels
in order to develop testable implications about channel substitutability. Consumers in-
vest in human capital that reduces their expected future transactions costs up to the point
where marginal investment costs equal expected gains. Since consumers are hetero-
geneous in their investment costs, they can have different channel-specific transactions
costs and make different channel choices. Actual investments can “spillover” to mul-
tiple channels or product categories. Consumers with more human capital will tend to
have lower transactions costs with the channels and will benefit the most from these
investments. Thus, channels with similar characteristics will tend to be substitutes.
An empirical methodology is proposed to investigate GVU Center survey data on
channel choice for a number of product categories. The test of the theory involves finding
a particular pattern among the residuals of channel choice regressions that control for
both individual and category effects (but not interacted). Generally, I find that use of
online shopping and direct-mail tend to be positively correlated to a greater degree than
online shopping and traditional retailing or direct marketing and traditional retailing.
These results are consistent with online shopping being a closer substitute for direct-
mail than for traditional retailing.
While the evidence presented here pertains to consumer substitutability between
online and catalog shopping, there is reason to believe that producers will find these two
channels to be close substitutes also. First, a website can be thought of as portraying the
same information as a catalog but via a different medium. If so, the transition to online
shopping requires a smaller investment on the part of direct marketers than traditional
retailers. Second, online retailing allows target marketing to a finer degree than has
previously been available. Direct marketers have typically led traditional retailers in the
use of target audience information. Third, as mentioned above, both catalog and online
shopping require similar product delivery mechanisms. A delivery system for catalog
shoppers would seem to be equally as useful for online shoppers. Both the analysis
above and these features suggest that direct marketers, or those with direct marketing
skills, are more likely to succeed at online retailing.
Forecasting the future of e-commerce, especially at this early date in its devel-
opment, would be speculative at best. This analysis indicates that, at least in its early
development, e-commerce tended to disproportionately draw customers that would oth-
erwise have purchased from catalogs. Online retailers of products in categories that have
had relatively low penetration to date may continue to draw these customers. However,
retailers of products from categories that are relatively more mature online seem to be
expanding their customer base to those who would have otherwise shopped at traditional
retailers. To be sure, the volume of e-commerce will grow as more consumers have ac-
DISTRIBUTION CHANNEL COMPETITION 117
cess to and develop experience with the Internet. Wider spread adoption of e-commerce
may require advances that make websites more than just catalogs via a different medium.
Notes
1. Currently, the dollar volume of business-to-business e-commerce is thought to exceed business-to-
consumer e-commerce. This paper primarily pertains to business-to-consumer e-commerce, often re-
ferred to as online shopping in the paper.
2. I would like to acknowledge the liberality of Georgia Tech Research Corporation and the GVU Center
for making these data available.
3. Brynjolfsson and Smith [5] find that ease of search can lead to e-commerce markets being more efficient
markets with lower prices and prices that better correspond to costs.
4. In fact, the endogeneity of human capital is not important for what follows. Two key assumptions are that
human capital stocks are heterogeneous and that human capital is not entirely specific to a distribution
channel or product category, but instead spills over to related channels and categories.
5. While the survey actually asks about catalog purchases, for notational convenience, direct marketing and
catalog shopping are taken to be synonymous for the rest of the article.
References
[1] A. Ainslie and P.E. Rossi, Similarities in choice behavior across product categories, Marketing Sci-
ence 17(2) (1998) 91–106.
[2] S. Balasubramian, Mail versus mall: A strategic analysis of competition between direct marketers and
conventional retailers, Marketing Science 17(3) (1998) 181–195.
[3] R. Benabou, Search market equilibrium, bilateral heterogeneity, and repeat purchases, Mimeo, MIT
(1990).
[4] R. Benabou, Search market equilibrium, bilateral heterogeneity, and repeat purchases, Journal of
Economic Theory 60(1) (1993) 140–158.
[5] E. Brynjolfsson and M. Smith, Frictionless commerce? A comparison of Internet and conventional
retailers, working paper http://ecommerce.mit.edu/papers/friction/friction.pdf (January 1999).
[6] J. Carlson and R.P. McAfee, Discrete equilibrium price dispersion, Journal of Political Economy,
91(3) (1983) 480–493.
[7] Graphics, Visualization and Utilization Center GVU’s WWW User Surveys, http://www.gvu.gatech.
edu (1998).
[8] K. Lancaster, A new approach to consumer theory, Journal of Political Economy (1966).
[9] P.R. Messinger and C. Narasimhan, A model of retail formats based on consumers’ economizing on
shopping time, Marketing Science 16(1) (1997) 1–23.
[10] B. Pashigian and B. Bowen, The rising cost of time of females, the growth of national brands, and the
supply of retail services, Economic Inquiry 32 (1994) 33–65.
[11] P.E. Rossi, R. McCulloch and G. Allenby, On the value of household purchase history information in
target marketing, Marketing Science 15(4) (1996) 321–340.
