Multinomial logit model, robubstness checks

[Mia Laufer]

Dear members,

I have several questions concerning possible robustness checks for my model. I run a clustered multinomial logit model where the dependent variable has three possible outcomes.

First, does it make sense to run a multinomial probit model as a robustness check? I know that logit and probit lead to similar results, but in the case of the multinomial probit model, the IIA assumption is not as important as in the case of the multinomial logit model. Does the probit model just fulfill expectations or is it a useful robustness check? What do you think?

Second, I divided the time period into two subperiods. The results for the first period are different to the full period, but the second period equals exactly the full period. Even when I divide the period in three parts – the results of the last part still equals the full period. How is that possible or what could be my mistake?

And third, I also would like to run a standard pooled OLS regression – but my dependent variable is a categorical variable. Is it possible to run such an OLS regression and interpret the output in a way like: “An increase in variable X increases the probability of the occurrence of state Y of my dependent variable.”? (where state Y is the highest number of the three possible outcomes of my dependent variable)

Any help is much appreciated! Thank you!

[Mia Laufer]

I have to correct on thing concerning the time periods: When I divide the period in three shorter periods, the results for the two last periods equal those for the whole period. (not only the last)

[Maarten Buis]

First, does it make sense to run a multinomial probit model as a robustness check? I know that logit and probit lead to similar results, but in the case of the multinomial probit model, the IIA assumption is not as important as in the case of the multinomial logit model.

That is incorrect, at least in the way mprobit estimates that model. So mprobit is not useful as a robustness check.

Second, I divided the time period into two subperiods. The results for the first period are different to the full period, but the second period equals exactly the full period. Even when I divide the period in three parts – the results of the last part still equals the full period. How is that possible or what could be my mistake?

Hard to say without seeing what you did...

And third, I also would like to run a standard pooled OLS regression – but my dependent variable is a categorical variable. Is it possible to run such an OLS regression and interpret the output in a way like: “An increase in variable X increases the probability of the occurrence of state Y of my dependent variable.”? (where state Y is the highest number of the three possible outcomes of my dependent variable)

Your variable Y does not make sense: it assumes that your categories are ordinal, but mlogit assumes they are categorical. You could run a linear regression on each outcome category. But comparing that with the mlogit results is extremely tricky. So that will probably create more problems than it solves.

[Mia Laufer]

Thank you very much for your reply! I changed my robustness checks in a way that I think they are now meaningful and correct.

Turning to my regression model (panel data set, multinomial logit regression with "cluster" addition), I included year dummies to see what happens to my results. And nothing happened. Is this because I do not use a panel regression model, but a multinomial logit regression? Is it in this case not necessary to include year or country dummies?

[Maarten Buis]

Unless you did something that did not make sense, "nothing happens" is an empirical finding, not a result that is necessarily true and thus meaningless.With the information you have given us we have no way of determining whether you did something that makes sense, so this is all we can say.