Bayesian and classical estimation of mixed logit: An application to genetic testing

Abstract

Discrete choice experiments (DCEs) in health economics have recently used the mixed logit (MXL) model to incorporate preference heterogeneity. These studies typically use a classical approach to estimation or have specified normal distributions for the attributes. Specifying normal distributions can lead to erroneous interpretation; non-normal distributions may cause problems with convergence to the global maximum of the simulated log-likelihood function. Hierarchical Bayes (HB) of MXL is an alternative estimation approach that may alleviate problems of convergence. We investigated Bayesian and classical approaches to MXL estimation using a DCE that elicited preferences for a genetic technology. The classical approach produced unrealistic results in one of the econometric specifications, which led to an erroneous willingness to pay estimate. The HB procedure produced reasonable results for both specifications and helped ascertain that the classical procedures were converging at a local maximum.