ApolloChoiceModelling forum

Hello,

I am trying to use a hybrid choice model with five alternatives. I got a question if Apollo can test the correlation between alternatives in the DCM. Would it be possible to get a test result?

Thank you very much.

Hi,

Yes, it is possible to calculate correlation between alternatives. You can use one of two approaches (there might be other ways, but these are the most common ones):

• Change the kernel of your model (i.e. the choice component) from an MNL to a GEV model (for example a nested logit or cross-nested logit) to estimate the correlation between alternatives. This might be faster, but you need a good understanding of nested or cross-nested logit.

• Keep using an MNL kernel for the choice part, but add error components to the utility of each alternative that are correlated across alternatives. This would be analogous to what is done in example example 15, but instead of making random coefficients correlated across alternatives, it would be an additional error term (different for each alternative, as discuss by Daly & Hess)

After estimating the model, you can check for the significance of the corresponding correlation parameters.

So while it is possible to do in Apollo, it requires some modelling effort. I'm afraid there is no simple option to calculate correlation between alternative's utilities.

Cheers
David

Thank you for your advice, David.

Actually, I'm trying to address the reviewer's comments, and he/she said that "It is likely that the alternatives are perceived as correlated. This should be tested in the DCM."

I think it would be great if I can see that the alternatives are not significantly correlated in the DCM. Is there any way to test with a nested logit using the same variables? In this case, another concern is that the hybrid choice model that I used takes a lot of time to be estimated...

Please let me know if you have any ideas.

Thank you very much.

Hi

as a first step, it would of course be wise to test for the correlations in a simple NL model, so without the hybrid structure. But there is no guarantee that your findings in terms of correlation structure would be the same in a discrete choice only model and a hybrid model

Stephane