Multinomial Logit Modeling

[brandon zagorski]

Greetings-

I am looking for some advice on a multinomial logit model with marginal effects. The model assumes independent observations, an assumption that is clearly violated in my dataset. How can I adjust for the within subjects correlation? Below is my syntax.

mlogit emply3 $indepvars
margins, dydx($indepvars) predict(outcome(0)) post /*there are 4 unique outcome levels*/

Thank you in advance for responding!

Best Regards, Brandon

[Weiwen Ng]

One way might be to use a multinomial random effects model. There is no memlogit or xtmlogit command, but this can be done through gsem, as the link will document and as you will find if you search the forum for multinomial random effects - the question does tend to come up from time to time!

[Leah Jones]

One way might be to use a multinomial random effects model. There is no memlogit or xtmlogit command, but this can be done through gsem, as the link will document and as you will find if you search the forum for multinomial random effects - the question does tend to come up from time to time!

If you're using Stata 15 there is a command for this.
https://www.stata.com/new-in-stata/m...it-regression/

[Weiwen Ng]

If you're using Stata 15 there is a command for this.
https://www.stata.com/new-in-stata/m...it-regression/

I am not familiar with asmixlogit because I'm not an econometrician, and choice experiments are not my thing. Can you confirm, though, that it would fit a model compatible with repeated measures on the same individual? You clearly do have to organize the data in long format from the perspective of each choice, which is probably not a typical data organization for non-econometrics purposes (although it does allow characteristics specific to each potential choice at each time point).

[brandon zagorski]

Thank you all for your responses. Is it appropriate to use the option vce(cluster varname) to control for within subjects variation?

mlogit employ3 $indepvars, vce(cluster ID)
margins, dydx($indepvars) predict(outcome(0)) post