ApolloChoiceModelling forum

Hi Stephane,

following the example in the manual I´ve introduced an error component for heteroskedasticity in one of my models.
Results show that sigma_hsk is a significantly negative coefficient but the estimates of the other parameters only change very slightly in comparison to a model without sigma_hsk. Why is there at all an issue with heteroskedasticity at all as homoskedasticity is not an assumption of logit models?

Can you briefly explain what sigma_hsk does in the estimation (does it show if there is heteroskedasticity in my data at all or does it make the model account for heteroskedasticity?) and how to deal with the mentioned results?

Maybe you would be able to reply to my post until sunday as I have a deadline there and it would help me a lot to include your answer into my submission.
Thank you very much in advance.
Best,
Julia

Julia

difficult to answer without seeing the specific code/results, but a sigma different from 0 would indeed indicate heteroskedasticity. You may want to look at Joan Walker's work on this

Stephane

Okay, thank you very much for clarifying.
If I use a mixed logit the iid assumption of the MNL model is not valid anymore. So why is there a possibility for heteroskedasticity in mixed logit? Can it be due to misspecified variables (e.g. assumed linear but is quadratic)?
Any advice is highly appreciated.

Thanks and best,
Julia

Julia

in mixed logit, the model inside the mixture is still MNL, so still subject to IID and IIA. But the presence of the random terms moves us away from that, but not in a structured way if done solely with random coefficients. Including error components allows you to test specific error structures.

Stephane