Understanding unobserved heterogeneity in consumer preferences: A comparative analysis of discrete choice models

This article employs the discrete choice experiments technique to elicit consumer preferences and sources of unobserved heterogeneity by comparing different choice models. In addition to the basic multinomial logit model, the most popular extensions are outlined, namely, the mixed and the latent class logit models to account for unobserved taste heterogeneity; the scaled multinomial logit model to allow for unobserved scale heterogeneity; and the generalized multinomial logit model to explicitly account for both. The application uses data from a survey on preferences for wine in Switzerland. Estimation results confirm the presence of a substantial amount of heterogeneity in preferences for most attributes and differences across willingness to pay estimates. In general, models that do not account for unobserved preference heterogeneity are statistically inferior on goodness-of-fit and lead to considerably higher mean willingness to pay. The Latent class model provides advantages in understanding the nature of heterogeneity of consumers, offering additional interpretation gains. Results confirm the need to test distinct models in a given application, and it stresses that more important than using a better-fitting model, is the understanding of the behavioral implications and the layers of individual heterogeneity revealed by each model. This is particularly relevant for policy recommendations.