A Bayesian neural network model of consumer choice

Since Guadagni and Little''s (1983) seminal work three decades ago, the vast majority of discrete choice models within academic marketing for analyzing frequently purchased product categories have generally been built upon two assumptions: (1) a linear relationship between covariates and latent utility, and (2) the limited value of household demographic information. This coupled with the inherent need amongst researchers to propose models that are as parsimonious as possible has lead to limited experimentation with interaction terms which in turn has reinforced the field's stance on the irrelevance of demographic information when dealing with (low priced items in) frequently purchased product categories.; This dissertation explicates the potential usefulness of deviating slightly from these practices by considering a non-linear mapping between (both marketing mix and demographic) covariates and latent utility.; Specifically, this dissertation investigates the advantages and disadvantages of a family of the relatively new Bayesian neural network models (Neal 1996) over the more established multinomial probit discrete choice model on a scanner panel product choice marketing problem. In addition, the usefulness of demographic information under each approach is evaluated.; Despite being more interpretible and much easier to estimate, the results show the multinomial probit lagging behind most of the Bayesian neural networks on all fit metrics.; The results also indicate that demographic information plays a more influential role in the Bayesian neural network model as compared with the multinomial probit.; Finally, extensions to multi-category choice situations are proposed.