880-01Understanding the consumer``s purchase behavior and how it is affected by advertising and promotion has been major concerns for marketing scientists and practitioners. To this end, they developed a number of quantitative models of the consumer``s response to advertising and promotion. In investigating the consumer``s response to promotion, marketing scientists tend to make use of discrete choice models, especially Multinomial Logit (MNL) models. The basic idea behind the MNL model is that at a particular choice occasion the consumer chooses the brand with maximal random utility. Thus, the MNL model is appropriate in investigating the consumer``s choice among a set of brands at a point of time rather than over time. In marketing, however, the MNL model is often applied to scanner data characterized by a series of choices made over time, causing the inconsistency between the scanner data and the MNL model applied to it. In addition, in estimating the parameters of the MNL models marketing researchers tend to employ either Maximum Likelihood Estimation (MLE) methods or Generalized Least Squares (GLS) methods, depending on whether the data set is of binary or proportional choice. In particular, GLS methods have the problem of "empty cell" resulting from the use of log-linear transformation of binary choice data. At this point some questions arise: (Q1) Does the inconsistency between the scanner data and the MNL model applied to it lead to biased estimates for the parameters of the MNL model? If so, how can we alleviate the problem of the inconsistency? (Q2) In estimating the parameters of the MNL model, how can we deal with both binary and proportional choice data without causing the problem of "empty cell"? Using the advertising response models, marketing researchers want to know whether it is best to advertise at a constant rate, referred to as even strategy, or it is best to alternate between periods of greater spending and periods of less spending, referr...