choice-metrics.com

Hi everyone!

I understand that when calibrating a model using logistic regression with data from a binary unlabeled choice experiment, I should include no intercept in the model specification, since there should be no alternative specific constant. I usually use GLM in R to fit my models, and then I use the deviance and the null.deviance reported by the GLM object to calculate the pseudo R2 (1 - deviance/null.deviance). I noticed that the model without intercept was resulting in a much higher R2 than the same model with the intercept. This seems to happen because the null.deviance reported by GLM is different (higher) in the model without intercept. So I was wondering:

1- How does GLM calculate the null deviance in the model without intercept? I mean, what could be the "null model" in this case, since it doesn't seem to be the model with only a constant term?

2- Should I fit a model with only a constant term in another GLM object, get its deviance and use it as the null deviance to be compared to the deviance of the model I am calibrating? Or that doesn't make sense since the intercept wouldn't have any meaning?

I'm not very familiar with Stata, but I was using it today and it seems that when I force the model to be without intercept, the software doesn't report any R2 (which it normally does). I unfortunately don't have access to NLOGIT to see if the software reports a R2 for models without constants...

In the end, I think my doubt is if it's possible to calculate a R2 for models that are derived from unlabeled experiments, and, if so, how it's done.

I sincerely apologize if this type of question is not allowed in the forum, but I'm really in doubt and I don't think there's a better place in the internet for me to find an answer.
Since now, thank you very much to anyone who can help me with this one!

I am not familar with GLM in R, but I could recommend using Apollo. Apollo is a free package in R designed by the editor of the Journal of Choice Modelling specifically for estimating choice models.

http://www.apollochoicemodelling.com/

There is also a forum for Apollo where you can ask questions about model estimation.

Regards,
Michiel

Thank you for your answer, Michiel! I'll definitely take a look at Apollo package.

But, as a more theoretical question, regardless of the software used for estimating the model, when calibrating a model for an unlabeled experiment, is still possible to obtain a R2? I mean, does it still make sense to compare the model's deviance with the null model's deviance, since the null model by definition contains an intercept and the model itself will not have one? Or R2 in this case should not be calculated with the null model's deviance, but with something else?

There are three LL values:

LL(final)
LL(constants only), i.e. all coefficients equal to zero except for the constants
LL(0), i.e. all coefficients equal to zero

If there are no constants, then LL(constants only) = LL(0), i.e. the constant is normalised to 0.

Using these three LL values, you can compute rho squared values.

Note that in choice modelling, most people look at:
* adjusted rho squared (which corrects for the number of parameters)
* AIC
* BIC

The latter two for comparing model fit do not use LL(constants only) or LL(0).

Does that clarify?

Michiel

It sure does, Michiel!
Thank you very much for your help, I really appreciate it.

Regards,
Gabriel.