Analyzing Qualitative Predictors with Too Few Data: An Alternative Approach to Handling Sparse-Cell Values

Abstract

The marketing manager faces several dilemmas when analyzing multivariate frequency data. If the choice is to analyze a series of two-dimensional condensed tables, the interrelationships between those factors not in the table will be lost and biased inferences can result. If the decision is to analyze the complete multiway table, many of the cells may be sparse. The authors address the issue of how best to handle sparse-cell values in the context of a marketing data set relating store choice behavior to a number of shopper-specific variables. A simple new approach to this problem, which utilizes loglinear modeling techniques, is developed and contrasted with alternative remedies. The results of the comparative analysis show the proposed approach performs well, especially in the correct classification of seemingly unclassifiable shoppers.