The Mixed Logit model is quickly becoming the dominant discrete choice modelling tool, due to its capacity to incorporate flexible error structures allowing to model complex behavioural patterns. Problems remaining relate precisely to its great flexibility, as interpreting its results requires a non-negligible level of expertise. On the other hand, two methods of estimation have been proposed for the model, classic and Bayesian estimation. Both provide the same type of results for two groups of parameters: (i) the vector of means and the covariance matrix of the distribution of model parameters in the population, and (ii) parameter estimates at the individual level, but it is not obvious which method should be used in practice. To advance on this issue we used simulated and real data and estimated the model with both methods for a wide range of situations. We considered variations in sample size, number of alternatives and number of attributes, as well as the importance of the white noise error in the utility function. We found that in all cases the population parameters gave inferior results than the individual parameters, although in most reported applications only the former are estimated. In the case of the population parameters the classic approach performed better for small samples (the methods converge for large samples), but also depended on the percentage error in the data. More importantly, we also found that in the case of the individual parameters the Bayesian method normally outperformed the classic approach in the case of typical size samples and thus should be preferred in practical applications; this result contradicts previous finding reported in the literature. © 2008 by Nova Science Publishers, Inc. All rights reserved.