Estimation of the mixed logit likelihood function by randomized quasi-Monte Carlo

We examine the effectiveness of randomized quasi-Monte Carlo (RQMC) techniques to estimate the integrals that express the discrete choice probabilities in a mixed logit model, for which no closed form formula is available. These models are used extensively in travel behavior research. We consider popular RQMC constructions such as randomized Sobol’, Faure, and Halton points, but our main emphasis is on randomly-shifted lattice rules, for which we study how to select the parameters as a function of the considered class of integrands. We compare the effectiveness of all these methods and of standard Monte Carlo (MC) to reduce both the variance and the bias when estimating the log-likelihood function at a given parameter value. In our numerical experiments, randomized lattice rules (with carefully selected parameters) and digital nets are the best performers and they reduce the bias as much as the variance. With panel data, in our examples, the performance of all RQMC methods degrades rapidly when we simultaneously increase the dimension and the number of observations per individual.

► Support of the use of RQMC for estimating mixed-logit log-likelihood.
► Best performance with rank-1 lattices constructed with selected parameters.
► Such rank-1 lattices outperform Monte Carlo and Halton sequences.
► Reduction of the bias as much as the variance.
► Improvement is smaller, but still notable, in large dimension or with panel data.