Model choice

[gabriellesod]

Hi!

This might be an extremely stupid question, but I cannot seem to find a straight forward answer.

When constructing an efficient design where I state:

;eff = (mnl, d)

is it only possible for me to estimate an MNL then? Can I not estimate a latent class or hybrid model then?

Best regards,
Gabrielle

[Michiel Bliemer]

You need to distinguish between a "design model" and an "estimation model". A design model is the model that you assumed when optimising the experimental design, which includes the choice of utility functions, model type, and priors. The estimation model may deviate from this design model, e.g. it may have somewhat different utility functions (e.g. including interactions or using dummy coding instead of effects coding), it may assume a different model type (e.g. mixed logit instead of MNL), and the parameter estimates will not be identical to the priors assumed. This is not a problem. The general rule is that the more the estimation model deviates from the design model, the more efficiency is lost in the data collection. But you will always lose some efficiency.

So to answer your question directly, yes you can still estimate a latent class or hybrid choice model, even if you optimised your experimental design for the MNL model. This is in fact very common as it is usually not practically possible to optimise an experimental design for the latent class or mixed logit model because priors are usually not easily obtained and the design generation process would be very computationally intensive. One recommendation I would give is to create an experimental design with a sufficient number of rows to ensure enough variation in your data to estimate more parameters.

Michiel

[gabriellesod]

Thank you so much for your answer, that was what I thought!

Best,
Gabrielle