The Impact of Varying the Number of Repeated Choice Observations on the Mixed Multinomial Logit Model

Analysts increasingly make use of the Mixed Multinomial Logit (MMNL) model to allow for random variations in sensitivities across respondents. The improvements resulting from this approach have been observed to be especially large in the case of datasets containing multiple observations per respondent. In this paper, the authors aim to address the question as to whether these improvements are primarily due to the increase in the number of respondents (N) or in the number of choices per respondent (S). At the same time, the authors seek to answer the question whether panel data is in fact a de facto requirement in order to guarantee stable retrieval of taste heterogeneity patterns. A large scale simulation exercise making use of nearly 5000 different MMNL models shows that, at small values of S, increases in the number of choices per respondent may be more effective in improving the accuracy of estimation of model parameters than increases in the number of respondents while keeping the number of choices per respondent fixed. At larger values of S, these advantages seem to at least partly disappear.