Title
Nonparametric Bayesian Methods For Analyzing Consumer Preference Models
Speaker
Marcus Sobel, Department of Statistics, Temple University
Abstract
Nonparametric Bayesian methods have proved successful in analyzing many, otherwise intractable, statistical problems. We use two of them to examine how different consumer preference determinants help to explain the choice between store and national brands. Dirichlet Process Mixture Models are shown to enhance and extend the usual regression methodology by: (I) distinguishing product category heterogeneity between consumers; (II) supporting the use of this assessed heterogeneity to make coherent, informative predictions about individual consumers and (homogeneous) groups of consumers; and (III) supporting the fitting of models through the efficient selection of hyperparameters. Nonparametric Bayesian methodology, which treats some missing data as 'censored', is shown to further enhance our results. Additional features including the use of Polya Tree Urn Priors for enhanced residual analysis are discussed.
