Comparing Alternative Models of Heterogeneity in Consumer Choice Behavior

While there is general agreement that consumer taste heterogeneity is crucially important in marketing, there is no consensus on a preferred approach to modeling heterogeneity. In this paper, we assess the performance of five alternative choice models, using ten empirical data sets. We include the popular latent class (LC) model, and the mixed logit (MIXL) model where utility weights are assumed to be multivariate normal. The new generalized multinomial logit (G-MNL) and scale heterogeneity (S-MNL) models are also included. G-MNL generalizes MIXL by allowing for heterogeneity in the scale coefficient. S-MNL is a special case of G-MNL where only scale heterogeneity is present. Finally, we consider the potentially more flexible mixture-ofnormals logit or “mixed-mixed” logit (MM-MNL) model. We find that according to the Bayes information criteria, G-MNL is preferred in 4 datasets while MM-MNL and S-MNL are preferred in 3 datasets each. By further investigating what behavioural patterns each model can capture better than others, we find that: (i) the more flexible heterogeneity distributions of GMNL and MM-MNL allow them to better capture "extreme" (i.e., lexicographic) as well as "random" behaviour; and (ii) which model is preferred depends on the structure of heterogeneity, which differs across datasets.