Incorporating Heterogeneity Into Count Data Models Applied to Marketing

In numerous empirical investigations into consumer behavior, the focus of the analysis is centered on explaining a limited dependent variable. Such limitation takes several forms, namely as a variable with limited continuous distribution to the left or to the right. Other limitations include variables with a finite number of values, such as those that take non-negative discrete values. Those data are usually referred to as “count data” and occur frequently in different marketing contexts, because they describe the number of times an event is observed. Examples include the number of purchases in a product category or the number of clients who visit a given store within a specified period of time. In marketing, this type of data has two basic characteristics: 1) excess of zeros (zero-inflation), more than expected in any Poisson distribution and; 2) heterogeneity among observations (buyers or consumers). Such characteristics may represent real problems if we use traditional models (such as the Poisson Regression model) to treat these data. This article deals with these two problems and incorporates some unusual reflection on count data modeling in marketing: 1) models which do not take zero-inflation into account will have poor fit; and, 2) the inclusion of unobserved heterogeneity may significantly improve the model’s goodness of fit, depending on the data (which would support the idea that independent variables are not always crucial to explain the dependent variable); and; 3) in many cases the heterogeneity of consumers may be sufficient (or even necessary) to explain the behavior of the dependent (count) variable.

Citation:

Delane Botelho and Pedro Jesus Fernandez (2006) ,"Incorporating Heterogeneity Into Count Data Models Applied to Marketing", in LA - Latin American Advances in Consumer Research Volume 1, eds. Silvia Gonzalez and David Luna, Duluth, MN : Association for Consumer Research, Pages: 90-91.