What is really uncovered by mixing different model structures: contrasts between latent class and model averaging

Latent class models have long been a tool for capturing heterogeneity across decisionmakers in the sensitivities to individual attributes. More recently, there has been increased interest in using these models to capture heterogeneity in actual behavioural processes, such as information/attribute processing and decision rules. This often leads to substantial improvement in model fit and the apparent finding of large clusters of individuals making choices in ways that are substantially different from those used by others. Such findings have however not been without criticism given the potential risk of confounding with other more modelspecific heterogeneity. In this paper, we consider an alternative approach for exploring the issue by contrasting the findings obtained with model averaging, which combines the results from a number of separately (rather than simultaneously) estimated models. We demonstrate that model averaging can accurately recover the different data generation processes used to create a number of simulated datasets and thus beused to infer likely sources of heterogeneity. We then use this new diagnostic tool on two stated choice case studies. For the first, we find that the use of model averaging leads to significant reductions in the amount of heterogeneity of the type analysts have sought to uncover with latent class structures of late. For the second, results from model averaging show clear evidence of the existence of both taste and decision rule heterogeneity. Overall, however, our results suggest that heterogeneity in the sensitivities to individual attributes rather than the behavioural process per se could be the key factor behind the improvements gained through the adoption of latent class models for heterogeneity in behavioural processes.