A Monte Carlo Study of Design Procedures for the Semi-parametric Mixed Logit Model

Abstract We determine efficient designs for choice-based conjoint analysis for the semi-parametric mixed logit model which captures latent consumer heterogeneity in a very flexible way. Different methods constructing one or multiple designs are tested. Additionally we apply Halton draws and determine a minimum potential design for prior draws to reduce computation times caused by accounting for latent heterogeneity of consumers. As main efficiency criteria for the construction of designs we consider measures related to D-error and entropy. As additional benchmarks we generate designs both randomly and by an approach which starts from orthogonal designs developed for linear models. We compare these alternative design procedures by simulating choices for different constellations on the basis of the semi-parametric mixed logit model. Using these simulated choices we estimate parameters of the semi-parametric mixed logit model in the next step. ANOVA with root mean squared error between estimated and true coefficient values as dependent variable shows that performance of design procedures depends on dissimilarity and segment size. Following a mean-standard deviation approach we determine which procedure should be used under different constellations or lack of prior information. Overall, either constructing ten designs based on DB-error or constructing one design based on entropy turn out to be preferable.