Unobserved Heterogeneity in Structural Behavioral Models Using Experimental Data

In this paper we compare a mixed logit model (MLM) and a latent class model (LCM) in the context of behavioral structural estimation using experimental data. By providing an instrument to deal with the intrinsic unobserved heterogeneity that characterizes experimental data, these alternative models have clear advantages compared with a multinomial logit model (MNL) typically used in structural estimation of behavioral models. We carry out our exercise by using experimental data that allows us estimation of distributional parameters related to risk and social preferences. Somehow coherently with the economic theory, the LCM identifies three classes of subjects (risk/ineq. lovers, risk/ineq. neutral, risk/ineq. averse). Moreover, estimates from both MLM and LCM somehow confirm the findings from a MNL model, that under the veil of ignorance (VOI) subjects’ variance aversion mostly reflects risk, rather than distributional concerns. By taking unobserved heterogeneity adequately into account in the estimation of our structural behavioral model, also provides new insights into individual behavior on the interplay between risk and inequality concerns. For example, we find that there is much more variability in individual behavior when subjects face pure inequality than under VOI. Moreover, in the case of pure inequality subjects are also more likely to be inequality lovers than under VOI.