This research investigated the choice of distributions for the parameters of a mixed logit model, under the risk that the model appears statistically successful when its actual fit to the true behavior is poor. A subsampling technique is used to examine the aptitude of several model specifications to approximate the range of preferences of bus commuters in York, England. Difficulties in replicating the distribution of the willingness to pay are discussed. Various issues are raised, such as the existence of negative willingness to pay, constraining distributions by either estimation or imposition, and the potential of subsampling experiments.