Estimating Mixtures of Discrete Choice Model

In this note, we take up a computational problem observed with fitting such mixtures of discrete choice models as the mixed multinomial logit, the parameter values explode as a numerical optimization algorithm maximizes the logarithm of the simulated likelihood function. We describe two identification issues that can increase the probability of this phenomenon. First, the parameters of the variance-covariance matrix of differences in the latent utility indexes may not be identified. Second, that variance-covariance matrix may be singular.