post-estimation for firthlogit

[Dimitriy V. Masterov]

Is it possible to trick the community-contributed -firthlogit- to work with -estat classification- or any other post-estimation command that can return the confusion matrix?

[Richard Williams]

You could probably do it. But I suspect the classification table would be a waste. Since you are using firthlogit because events are rare, the predictions will probably all be for the more common event.

[Joseph Coveney]

Is it possible to trick the community-contributed -firthlogit- to work with -estat classification- or any other post-estimation command that can return the confusion matrix?

Yes. The trick is the same as that for obtaining the conventional coefficient standard errors, which is shown in the user-written command's ancillary file SEMatch.do.

do http://fmwww.bc.edu/RePEc/bocode/s/SEMatch.do
estat classification

You could probably do it. But I suspect the classification table would be a waste.

Yep.

[Dimitriy V. Masterov]

That SEMatch approach worked nicely.

On the prediction questions, what approach would you recommend? I have a dataset with 1,127,897 observations where 5,239 units experience the rare event outcome. I have 10 covariates, mostly binary. Based on reading Richard Williams notes, this seems like grey area as far as the choice between the ordinary and the firth logit. I have also considered running the rare event logit.

[Richard Williams]

With that many events, I suspect you do not need Firth logit, but you can compare the logit and firthlogit approaches. I suspect the classification table will get you nowhere but adjusted predictions and marginal effects may show some promise.