Relation between error component and random taste parameter in ML model

[mlmoleman]

Dear Mr. Van Cranenburgh,

Hereby you receive a question about the (possible) relationship between the error component and random taste parameter(s) in the Mixed Logit model.

Is it expected that, when a random taste parameter is added to the error component ML model, the random taste parameter will deduct taste captured by the error component? In other words, is it possible for taste heterogeneity to move from the error component to the random taste parameter after adding the random taste parameter? Or will taste heterogeneity of segments of individuals on a certain attribute never be observed by the error component?

Thank you in advance.

Best regards, Milan Moleman

[sandervancranenburgh]

Sorry for my late reply. I must have missed it due to the exam preparations. This is a very good question. In practice, you indeed often see happening what you have just described: i.e. a random taste parameter 'eating' away correlation in the unobserved utilities. One example we have seen in calls. When trying different models to the Swiss metro data, we saw that after accounting for unobserved taste heterogeneity, the error component got much smaller (i.e. the sigma decreased in size). I hope this answers your question. Sander