I understand that in case of separated data, logistic regression via ordinary MLE has an upward bias in the p values, which implies that any penalized MLE designed to reduce this bias will have more power in such cases. Specifically I'm considering Firth's penalization.

Can someone point me to a reference or an argument that compares the power of Firth logistic regression to ordinary MLE in general (that is, without the assumption of data separation)?


1 Answer 1


With some more research I found this article where several methods are compared for power amongst other features, and they find that Firth regression has a power comparable to ordinary GLM:


  • 1
    $\begingroup$ The too-large-p-value problem applies when unadvisedly using the Wald $\chi^2$ test. If you use the likelihood ratio $\chi^2$ test, complete separation is not a problem, and Firth's approach is not needed. $\endgroup$ Sep 11, 2022 at 18:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.