See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/239604583

Statistics and Marketing

Article in Journal of the American Statistical Association · June 2000

DOI: 10.1080/01621459.2000.10474240

CITATIONS READS
14 3,801

2 authors, including:

Greg Allenby
The Ohio State University
204 PUBLICATIONS 9,759 CITATIONS

SEE PROFILE

All content following this page was uploaded by Greg Allenby on 09 October 2014.

The user has requested enhancement of the downloaded file.

 Statistics and Marketing

Peter E. Rossi and Greg M. Allenby

Statistical research in marketing is heavily influenced by the availability of different

types of data. The last ten years have seen an explosion in the amount and variety of data
available to market researchers. Demand data from scanning equipment has now become
routinely available in the packaged goods industries. Data from e-commerce and direct
marketing is growing at an exponential rate and provide coverage to a wide assortment of
different products. Web-based technology has dramatically lowered the cost of survey
research. Web-browsing data is an important new source of information about consumer
tastes and preferences which is becoming available for a large fraction of the total consumer
population. In this vignette, we explore some of the implications of this data explosion for
the development of statistical methodology in marketing with primary emphasis on the
explosion in demand data.
Scanning equipment has provided the market researcher with a national panel of
stores in addition to panels of households, altering the focus of marketing research. This
data has stimulated a large literature on applied demand and discrete choice modeling.
Demand models at the store level typically take the form of multivariate regression models in
which demand for a vector of products is related to marketing variables such as prices,
displays and various forms of advertising. At the household level, demand is discrete and a
wide variety of multinomial logit and probit models have been applied to the data.
Early experience with scanner data revealed that households have very different
patterns of buying behavior that cannot be explained just by differences in the marketing
environment. Some households, for example, exhibit strong brand loyalties while other
households readily switch brands when prices are lowered. Even at the store level, large
differences have been detected in price and local advertising sensitivity. Initial observations
of store and consumer heterogeneity created considerable interest in models of observed and
unobservable heterogeneity, primarily of the random effects form. The development and
application of random effect models in marketing has been dictated in large degree by the
available inference technology.
The first paper in this area by Kamarkura and Russell (1989) used a finite mixture
model of heterogeneity in a logit framework. Kamarkura and Russell postulate a discrete

1
multivariate distribution for the intercepts and marketing mix variable coefficients in a
multinomial logit model. The ease of computation of finite mixture models produced a
stream of papers applied this idea in a wide variety of other model contexts. Until quite
recently, it was not computationally practical to investigate multivariate continuous models
of heterogeneity in a discrete choice setting. Simulation-based Maximum Likelihood
(Geweke, 1989) and Markov Chain Monte Carlo (MCMC) Bayesian inference methods
(Gelfand and Smith, 1990) have now made this possible. Recent work (Lenk, DeSarbo,
Green and Young (1996) and Allenby and Rossi (1999)) has documented that finite mixtures
provide poor approximations to the high dimensional distributions necessary to capture
heterogeneity not only in the intercepts but also in the slopes of the multinomial choice
models. Rossi, McCulloch and Allenby (1996) use a multivariate normal distribution of
random effects in a Bayesian hierarchical setting. Observable household characteristics such
as demographics are included in the model by allowing the mean of the random effects
distribution to depend on these variables. These observable characteristics account for only
a small amount of the total household heterogeneity. Brand preferences and marketing mix
sensitivities are therefore revealed primarily by household purchase behavior.
A distinguishing characteristic of random-effects applications in marketing is the
interest not only in the hyperparameters of the random-effects distribution but in making
inferences about household parameters. Marketers can use household level parameters to
target households for customized promotions or to cross-sell other products. Even at the
store level, there is considerable interest in the allocation of marketing efforts across
different geographic areas which requires the building of store or market level models.
The sampling-theoretic approaches to random effects models typically average the
model likelihood over the random effects distribution and provide no natural way to make
household level inferences. In contrast, Bayesian hierarchical models provide a natural
setting in which inferences can be made concerning both the parameters of the random
effects distribution and the household or store level parameters. The combination of ability
to make inferences about the random parameter draws and computational tractability for
even very high dimensional problems has made Bayesian hierarchical methods very popular
in the empirical marketing literature (see Allenby and Lenk (1994), Montgomery (1997),
Allenby, Arora and Ginter (1998), Ainslie and Rossi (1998), Boatwright and Rossi (1999),

2
Neelamegham and Chintagunta (1999), Allenby, Leone and Jen (1999), Bradlow and
Schmittlein (1999)).
Just as the current statistical literature in marketing was stimulated by the availability
of panel data, we expect the future course of research to be influenced by the growth in
customer transaction databases. Direct marketers, credit card and financial service
marketers, and retailers are now compiling large databases comprised of customer
transactions and marketing contacts. These databases range from a complete record of all
customer purchases along with information on other alternatives considered to much more
incomplete data in which only the purchases are recorded (as in most frequent shopper or
loyalty programs) or where only indirect evidence of product preferences is available (as in
Internet Web-browsing data).
One of the most basic uses of customer transaction data is to find customers who
are likely to express interest in or purchase a specific product. For example, a marketer
could use credit usage data to find card holders who are likely to be interested in an upscale
Italian restaurant and then target promotion activities at this subset. Another common
example is the use of house file information in catalogue purchases to identify customers
who are likely to respond to a catalogue mailing. The challenge is essentially a classification
task. Can we construct a set of variables summarizing past transaction data and choose a
model form which effectively locates the customers we want to reach? The industry uses a
variety of standard regression and logit models for the most part. Given the extremely large
set of possible variables and functional forms possible, many of the classification problems
have a huge variable selection problem associated with them. Also, there is no particular
theoretical reason to believe that standard linear regression or logistic regression models
have the right functional form. Recent developments of the statistics literature on variable
selection and CART models might find very fruitful application here (George and
McCulloch (1995) and Chipman et al (1998)). It may also be useful for firms to build upon
the current investment in regression models by considering regression models with
coefficients which change depending on the location in the explanatory variable space.
Neural networks models (c.f. Ripley 1996)) have already found application in predictive
model building with marketing data.
In many instances, firms need to predict the response to a new or modified offering,
ranging from new prices for old brands to entirely new offerings. A direct mail merchant

3
who wants to predict the effects of a price change, for example, needs information about
prices of competing brands. While this information is often not available in a firm’s
transaction database, it can be imputed by augmenting the transaction data with data from
other sources. For example, a firm could obtain the data for a subset of customers by
recruiting a panel and having them record their purchases and prices for the other brands
that were considered at the time. Once a joint model of all causal variables is available for a
subset of customers, estimates of price sensitivity and brand preference can be generated
from transaction data by constructing the conditional distribution of these variables given
the available information.
Even when data are available to identify the parameters in a customer demand
model, the resulting estimates are sometimes insufficient for marketing decisions. The brand
intercepts, for example, reflect consumer preference for brands net the effects of price and
the other variables that change through time. These preferences are driven by product
attributes, the perceived performance of the brand on these attributes, and the importance
that consumers attach to them. Furthermore, the importance of product attributes to a
consumer can be traced to the concerns, interests and motives they bring to the problem for
which the brand is relevant. Understanding these more primitive constructs is important in
many marketing decisions, such as market segmentation, advertising and product
development. A valuable area for future research is the merging of survey data with
transaction data. For example, in choice models the brand intercepts reflect preference net
the effect of price and other variable that change over time. Marketers want to relate these
preferences to product attributes. For many products these attributes are defined
subjectively by the consumer. Survey techniques can be used to assess these attribute levels
that can then be entered into choice models to parameterize the intercept.
As more demand data becomes available, demand models can be used to assess the
impact of marketing actions such as changes in price and advertising. Researchers are
increasingly aware that standard regression and choice models are not always adequate for
optimal policy determination as the parameters of these models are not policy invariant.
For example, the price sensitivity parameter in a choice model can give misleading
predictions if the marketing environment is altered significantly. Dynamic models of
consumer choice that take into account price expectations and inventory decisions are
required for policy evaluation (Gonul and Srinivasan (1996)). For example, the effects of a

4
temporary price cut (sale) can be very different in a market environment with frequent,
predictable sales than in an environment with infrequent sales.
Dynamic considerations also affect the interpretation of the intercept parameters in
choice models. In general, the consumer may update dynamically their views about product
quality. For some products, past consumption and/or advertising exposure may lead to
“learning” or updates of the intercept values (Erdem and Keane (1996)). In models with
consumption or advertising based updating of brand preference, marketing interventions
such as temporary price cuts and advertising can have long-run affects which require a
dynamic structural model for policy evaluation.
In addition to dynamic considerations, we must also recognize the fact that the
consumer may not be fully informed about all of the marketing mix variables when making a
purchase decision. Typically, the choice models used in marketing applications assume that
the consumer is aware of the prices of all products in the choice set. Studies have shown
that consumers are not fully aware of prices and engage is a price search process in which
they become selectively informed. This is best modeled in a search theoretic framework in
which the consumer decides to sample or not sample a price of an item based on the
expected return to further search. For example, if the consumer finds a relatively low price,
as judged by the prior, on a preferred item, further search may have a negative expected pay-
off. Estimating and postulating non-trivial choice models with price search presents an
important challenge for the marketing literature (see Mehta, Rajiv and Srinivasan (1999) for a
pioneering effort). A search theoretic framework can also shed light on the role of
information-style advertising such as in-store displays and feature ads in local newspapers. A
major role for such advertising is to inform consumers of prices without incurring the cost
of search. Finally, marketers have long debated the possibility of long-term negative effects
of price promotions (frequent sales). In a search model, greater price variability gives rise to
a greater return on search and heightens price sensitivity. A formal search model can
provide marketers with a sense of the trade-off between the long-term and short-term
effects of price promotions.
Reduction in the cost of acquiring survey data will result in increased interest in
statistical and psychometric methods. For example, web-based consumer satisfaction
surveys present methodological challenges including the problem of respondent-specific
scale usage patterns (see Yorkston (1992), and Rossi et al (1999)). More generally, survey

5
data provides a unique set of statistical challenges which stem from measurement error
problems (see Bagozzi et al (1999) for an excellent discussion of measurement error models)
as well as the multivariate aspects of the data (see DeSarbo (1994) for an overview of
developments in multi-dimensional scaling). Low response rates to many marketing surveys
also present a modeling and measurement challenge (see Bradlow (1999) for a recent
approach to the problem of “no answer” responses).
In summary, the rapid growth in consumer data has caused a major change in the
sorts of statistical models used by marketing researchers. Many standard models of demand
data are often not sufficient to provide insights into consumer behavior, nor to deal with the
often incomplete nature of demand data available to firms. The challenge to researchers
working in marketing is to develop new statistical tools appropriate for this data
environment, while employing models that can shed meaningful insight on consumer
behavior.

6
 References

Ainslie, A. and P.E.Rossi (1998) “Similarities in Choice Behavior Across Product Categories,”
Marketing Science, 17, 91-106.

Allenby, G.M.., R.P. Leone and L. Jen (1999) “A Dynamic Model of Purchase Timing with
Application to Direct Marketing,” Journal of the American Statistical Association, 94, 365-
374.

Allenby, G.M. and P.E. Rossi (1999) “Marketing Models of Consumer Heterogeneity,” Journal of
Econometrics, 89, 57-78.

Allenby, G.M., N. Arora and J.L. Ginter (1998) “On The Heterogeneity of Demand,” Journal of
Marketing Research, 35, 384-389.

Allenby, G.M. and J.L. Ginter (1995) “Using Extremes to Design Products and Segment
Journal of Marketing Research, 32, 392-403.

Allenby, G.M. and P.J. Lenk (1994) “Modeling Household Purchase Behavior with Logistic
Journal of the American Statistical Association, 89, 1218-1231.

Bagozzi, R..P., Y. Yi and K. Nassen (1999), “Representation of Measurement Error in

Marketing Variables: Review of approaches and extension to three-facet designs,”
Journal of Econometrics 89, 393-421.

Boatwright, P. and P.E.Rossi (1999) “Estimating Price Elasticities with Theory-based

Priors," forthcoming, Journal of Marketing Researach.

Bradlow, E.T. and D.C. Schmittlein (1999) “The Little Engines That Could: Modeling the
Performance of World Wide Web Search Engines,” Marketing Science, forthcoming.

Bradlow, E.T. and A. M. Zaslavsky (1999), “A Hierarchical Latent Variable Model for
Ordinal Data From a Customer Satisfaction Survey with “No Answer” Response,”
JASA 94, 43-52.

Chipman, H., E. George, and R. McCulloch (1998), "Bayesian CART Model Search," JASA
93, 935-960.

DeSarbo, W. S., A. K. Munrai, L. A. Munrai (1994), “Latent Class Mulit-dimensional Scaling:

A review of recent developments in the marketing and psychometric literature,” in
Bagozzi, R. P. (ed.), Advanced Methods for Marketing Research, Cambridge: Blackwell.

Erdem, T. and M.P. Keane (1996) “Decision-making Under Uncertainty: Capturing

Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets,”
Marketing Science, 15, 1-20.

7
 Gelfand, A.E. and A.F.M. Smith (1990), "Sampling-Based Approaches to Calculating Marginal
Densities," Journal of the American Statistical Association, 85, 398-409.

Geweke, J. (1989), “Simulation Estimation Methods for Limited Dependent Variable

Handbook of Statistics, Vol 11, G.S. Maddala, C.R. Rao and H.D. Vinod
(eds), Amsterdam: North Holland.

George, E. and R. McCulloch (1995), "Variable Selection Via Gibbs Sampling," JASA 88,
881-889.

Gonul, F. and K. Srinivasan (1996), "Estimating the Impact of Consumer Expectations of

Coupons on Purchase Behavior: A Dynamic Structural Model," Marketing Science 15,
262-279.

Greenleaf, E. (1992), “Improving Ratings Scale Measures by Detecting and Correcting Bias
Components in Some Response Styles,” Journal of Marketing Research 29, 176-188.

Kamakura, W. and G. Russell (1989), “A Probabilistic Choice Model for Market

Segmentation and Elasticity Structure, “ Journal of Marketing Research 26, 379-90.

Mehta, N. S. Rajiv and K. Srinivasan (1999), "Active Versus Passive Loyalty: A Structural
Model of Consideration Set Formation," working paper, Carnegie-Mellon University.

Montgomery, A.L. (1997) “Creating Micro-Marketing Pricing Strategies Using Supermarket

Scanner Data,” Marketing Science, 16, 315-337.

Neelemegham, R. and P.Chintagunta (1999) “A Bayesian Model to Forecast New Product

Performance in Domestic and International Markets,” Marketing Science, 18, 115-136.

Ripley, B.D., Pattern Recognition and Neural Networks, Cambridge: Cambridge University Press.

Rossi, P.E., R.E. McCulloch and G.M. Allenby (1996) “The Value of Purchase History Data in
Target Marketing,” Marketing Science, 15, 321-340.

Rossi, P. E., Z. Gilula and G. M. Allenby (1999), “Overcoming Scale Usage Heterogeneity: A
Bayesian Hierarchical Approach,” working paper, Graduate School of Business,
University of Chicago.

View publication stats