Improving the conditional logit model in discrete choices of individuals

In modeling discrete choices of individuals, many models are possible. The most wellknown model is the conditional logit model, but this model has his limits in the assumptions it makes, namely homogeneity for all the individuals and so assumes the same model for all the individuals and the assumption of independence of irrelevant alternatives. To relief these assumptions, I first use a latent class model. This model construct different segments of individuals with different preferences, so that the not all the individuals have the same parameters anymore. Another model I use to relief the assumption of independence of irrelevant alternatives, is called the nested logit model. This model separate the individuals in different clusters of base-preference and so does not assume independence of irrelevant alternatives between different clusters. I also use past-choice, where I take the choice of the individual in the last visit to estimate the model. The model with latent class performs better than the conditional logit model, and including a past-choice variable improves the out-of-sample performance of the model. The nested logit model performs better than the conditional logit model, but does not improve the out-of-sample performance.