I think it is better to use r hyphenyted with R commander and Rstudio to tackle some of these problems With R. In general I suggest R for your analysis.

I don't familiar with R but you could do it by SPSS.

plz see attached , hope that helps.

Those who have responded here- how do you INTERPRET the Firth logistic regression? I have not been able to find any resources on the web or in articles on this.  Only papers discussing why Firth regression is ideal for small samples or complete separation.

Lisa Chea If you use R and brglm package, it is very straightforward. Interpreted just like logistic regression.

Interested in what this method adds above and beyond the standard logistic regression. Could you or anyone explain to me?

Katja Rudell if you want a technical explanation in some detail, here it is https://pdfs.semanticscholar.org/4f17/1322108dff719da6aa0d354d5f73c9c474de.pdf

Manuel CF Pontes . Thanks great job. It seems worthwhile to do this to check for convergence anyhow? No?!? How come it is described in SAS, but I had not seen it in SPSS. Are these stats programs inferior to R and SAS? Also it described detail around the dummy coding, but I have rarely seen this method referred to in papers. Is it because it is more recent i.e. in the paper it quotes a reference from 2003 - does it mean this was not described before?

Katja Rudell Yes it is worthwhile and is also recommended by Prof Allinson. R and SAS have I believe have more estimation methods than SPSS but I rarely use SPSS. Not sure of the history, though the first paper by Firth was in 1993. If one is trying to work with rare side effects (< 5%) would definitely use Firth. There are two R packages, logistf and brglm.

Thank you. Very good responses. Now I need to read about estimation. I guess social science does less estimates as we normally have our theories to prove we need real data to to test them. 😁

Manuel CF Pontes Hi! As you said that the interpretation of the brglm function output is the same as logistic regression, could you, or anyone else, please suggest a function which would give me the marginal effects of brglm. I have been using logitmfx but then again I get the warning "glm.fit: fitted probabilities numerically 0 or 1 occurred" so I am back where I was when I was using glm.

Thanks!