Using margins-command with clogit

[Wilfried Pasmans]

Dear Statalist users,

I received a reviewer comment on an analysis using a fixed-effects logit model via the clogit command in Stata 17.0. The model has a binary outcome care_ind and we are interested in modelling the interaction of two nominal variables care_ind_previous and child_employment_change2.

clogit care_ind  /*
*/ i.care_ind_previous##i.child_employment_change2 b2.child_proximity_change2/*
*/ i.child_gender b3.child_age_cat i.child_education i.child_partner2 i.child_nchild2 /*
*/ , group(mergeidlevel) or vce(cluster country)

The first version of the manuscript reports the regression parameters in terms of odds ratios, but the interpretation is quite complicated and even inappropriate (e.g. the need to test different reference categories) to test the paper hypotheses. The suggestion to present the results by means of predicted probabilities and marginal effects could resolve a lot of issues that are currently present in the analysis. However, I learned from previous forum posts here and this presentation that this makes no sense (given that the predictions assume that the fixed effect is zero), although this is offered in Stata's postestimation options for clogit and xtlogit, fe. Does this hold for categorical covariates as well? In any case, the margins results do seem to make sense and are in line with main conclusions taken from the regression parameters at first glance.

Thanking you in advance for the advice!

[Wilfried Pasmans]

As a follow-up on my own question: could it be a solution to estimate a logit with a dummy approach to estimate the fixed effects (i.e. include k-1 dummy variables for the id "mergeidlevel")? I found out that it gives stable results in the postestimation, whereas results for margins after clogit differ when changing a reference category in the model, which is probably in line with previously-mentioned warnings about postestimation for clogit and xtlogit models.

So, this leads actually to the following question.

Is this fixed-effects model...

clogit care_ind  /*
*/ i.care_ind_previous##i.child_employment_change2 b2.child_proximity_change2/*
*/ i.child_gender b3.child_age_cat i.child_education i.child_partner2 i.child_nchild2 /*
*/ , group(mergeidlevel) or vce(cluster country)

...essentially equivalent to this one...

logit care_ind  /*
*/ i.care_ind_previous##i.child_employment_change2 b2.child_proximity_change2/*
*/ i.child_gender b3.child_age_cat i.child_education i.child_partner2 i.child_nchild2 i.mergeidlevel /*
*/ , or vce(cluster country)

?