Xtlogit, cluster se & marginal effects of interactive dummies


I'm just exploring the xtlogit, fe model to model happiness in BHPS.

I was wondering:
1) Is there any way to obtain the marginal effects of interactive dummies in an xtlogit, fe model? There is a method using xtlogit and nlcom, but not with xtlogit,fe. Would I need to use nlcom or is there a better command?

2) I found the only way to cluster standard errors in xtlogit, fe is using -vce(bootstrap)-. I read that it is easily misapplied so I was wondering if anyone can offer some guidance?

Thank you,
(1) Why -margins- doesn't work for you? Stata 11 and Stata 12 recognizes interaction of factor variables.
(2) You could look at -clogit- or the user-written command -gallam-(type -findit gllamm-)
Thanks for replying!
(1) I tried -margins- but it gives me an error: "default predict option not appropriate with margins". I'm not quite sure why, has it got to do with the fixed effects?
(2) I'll look at clogit, it seems similar to xtlogit,fe. Thanks a lot!
Thanks Bukharin that definitely works!

However, I have a question concerning the -margins- command: as margins predicts the probabilities of the effect of the interactive dummy on the dependent variable, for eg, numchild#married on happiness, this means that I can only tell if numchild#married variable has a high or low probability effect on happiness, but I don't know if this effect is positive or negative, right? Or have I understood this wrong? I was wondering if there is a command that interprets the coefficients like mfx for interactive dummies, just in case I'm missing out such effects.

Sorry if I didn't word this question properly and thanks again


Can't you determine that by the interaction's coefficient from the main model? If it's positive then the effect is positive, if it's negative then the effect is negative...

I think perhaps you're understanding it wrong. The -margins- command with the predict(pu0) option is giving you the predicted probability at different combinations of the covariates, under the assumption that the fixed effect is zero. It doesn't give the probability that a covariate (eg numchild#married) has an effect - by including the interaction in the model you've basically forced numchild#married to have an effect. The effect is the coefficient estimated by the model, and -margins- helps you to translate that effect from the logit scale into a probability.