Re: Extremely large estimated standard deviation from a Mixed Logit model using a lognormal distribution
Posted: 26 May 2023, 16:56
Hi Shan,
Mixed logit models are much more complex than MNL models. As such, their identification is more complicated (see Walker (2002) for a detailed discussion), and there is no guarantee that your data will support a model with several random coefficients due to empirical identification issues (i.e. your data may not contain enough information to estimate the model you want). So running into the issues you mention is actually quite common when estimating mixed logit models.
My recommendation would be to start estimating a simple MNL model first. When you are confident that you found a good utility specification in your MNL, and that your data is well behaved (i.e. does not seem to contain errors), only then I would start estimating mixed logits. When doing mixed logits, I would also start by introducing randomness to a single coefficient at a time, building up the complexity of the model slowly, so you can more easily identify sources of trouble.
About the starting values, when using log-normals I usually start the mu at -3, and the sd at 0, but there is no rule about it. I would just avoid large positive values because they can lead to numerical breakdowns.
Best wishes
David
Mixed logit models are much more complex than MNL models. As such, their identification is more complicated (see Walker (2002) for a detailed discussion), and there is no guarantee that your data will support a model with several random coefficients due to empirical identification issues (i.e. your data may not contain enough information to estimate the model you want). So running into the issues you mention is actually quite common when estimating mixed logit models.
My recommendation would be to start estimating a simple MNL model first. When you are confident that you found a good utility specification in your MNL, and that your data is well behaved (i.e. does not seem to contain errors), only then I would start estimating mixed logits. When doing mixed logits, I would also start by introducing randomness to a single coefficient at a time, building up the complexity of the model slowly, so you can more easily identify sources of trouble.
About the starting values, when using log-normals I usually start the mu at -3, and the sd at 0, but there is no rule about it. I would just avoid large positive values because they can lead to numerical breakdowns.
Best wishes
David