Handling Nested Logit with Varying Availability in Ranked Choice Data

[Amin Afshari]

Dear Professor Bierlaire,

I hope you are doing well.

I am working on a choice model with five alternatives, two of which (PTMM and PTW) share similar attributes, while the other three are independent. I am considering adding a shared error component to PTMM and PTW, allowing the model to capture their correlation.

Alternatively, due to the high time required to calculate monte_carlo draws, I have also attempted to estimate a Nested Logit (NL) model by grouping PTMM and PTW into a single nest. However, I have encountered a challenge due to the structure of my data. In my survey, respondents ranked all five alternatives. I constructed multiple derived choice situations from each scenario by successively removing the chosen alternative (by setting its availability to zero), resulting in four distinct choice sets per scenario.

The difficulty arises when one or both of PTMM and PTW are unavailable in a given derived choice set: the nest is then left with a single or no available alternative, which leads to model instability and estimation failure.

Is there a recommended way in Biogeme to handle such cases? Specifically, is it appropriate to estimate a Nested Logit model only for the choice sets where both PTMM and PTW are available, and revert to a standard MNL for the remaining choice sets? Or could you suggest an alternative modeling strategy to handle this availability issue more elegantly?

Thank you very much for your time and guidance.

Best regards,
Amin.