Question about post estimation commands in Stata after firthlogit (SSC)

[Marie Kaldahl]

Dear All,

I would like advice on specifying my model and commands to do post estimation in Stata with the user-written program firthlogit (available from SSC).

I have 3850 units in my dataset and my dependent variable (party switching) is binary, but with only 72 “1”’s. It is therefore rare event and on top of that, I have a problem with (quasi-) complete separation, because I have 2 out of 22 legislatures where no switching occur, and Stata issues this error message:

note: 1.legislature != 0 predicts failure perfectly 1.legislature dropped and 156 obs not used
note: 14.legislature != 0 predicts failure perfectly 14.legislature dropped and 170 obs not used

This made me discard logistic regression, and I do not think exlogistic is an opportunity either, because of the large N. Instead, I decided on firthlogit, which uses penalized likelihood.
My problem is that according to this post http://www.statalist.org/forums/foru...406#post809406 I cannot use the margins command in Stata with firthlogit. Do you know other post estimation commands in Stata that runs with firthlogit?

My model looks like this:
firthlogit partyswitcing position intrapartydemocracy intrapartydemocracysquare leftright leftrightsquare cabinetparty shableyShubikindex gender age senority i.legislature

partyswitching, gender, position and cabinetparty are dichotomies. I have put in legislatures as a categorical variable to get a fixed effect for legislature, but I do not want to use it as a predictor.

I would like to be able to plot my main independent variables (intrapartydemocracysquare, leftrightsquare and shableyShubikindex) to ease the interpretation.


Thanks in advance.

[Joseph Coveney]

Do you know other post estimation commands in Stata that runs with firthlogit?

The only one that I'm aware of is predict, and only for the linear prediction (xb).

I would like to be able to plot my main independent variables . . . to ease the interpretation.

Why not fit the model with firthlogit,* and then plot the linear predictions against each of your explanatory variables?

*See also:

search penlogit

[Marie Kaldahl]

Dear Joseph,

Thank you very much. Your answer was very helpful.

I get negative values, when I use the predict, xb but as I can see from your answer here: http://www.statalist.org/forums/foru...ter-firthlogit I can covert it using the invlogit, but I do not fully understand this operation. Why is it predict double pr, xb?

Something like this:

predict double pr, xb
quietly replace pr = invlogit(pr)

[Joseph Coveney]

Why is it predict double pr, xb?

The double type specifies that you want the predictions in double-precision. See

help data_types

for more information.

[Marie Kaldahl]

Thank you for your answer.