Maximum Simulated Likelihood Estimation of the Mixed Logit Model

This thesis investigates the estimation of the parameters of a mixed logit model. We use maximum simulated likelihood estimation where we compare two methods to evaluate the integrals in the simulated log likelihood of the mixed logit model. The first method is the pseudo-random Monte Carlo (PMC) method, which uses pseudorandom numbers to evaluate the integrals. The second method is the quasi-random Monte Carlo (QMC) method, which uses Halton draws to evaluate the integrals. We compare the performance of both methods with numerical experiments using data about ketchup brand choices. We find that the QMC method provides better accuracy, although the difference with the PMC method is small. We also considered a latent class mixed logit (LCML) model as an extension of the mixed logit model. However, based on the Bayesian information criterion we found that for our data set the mixed logit model is preferred over the LCML model.