On 5/27/08, martholt <martinholt42@yahoo.co.uk> wrote:

> I've always understood that, when doing logistic regression, at least > 10 of the rarer events are required per variable to be entered into > the model. I've been reading "Categorical Data Analysis using the SAS > System" by Maura E. Stokes et al, a SAS Institute book I believe. Here > is written, "Your choice depends partially on the sample size. There > should be at least 5 observations for the rarer outcome per parameter > being considered in the expanded model. Some analysts would prefer at > least 10." > > Does anyone have any information supporting the use of 5 ?

I would not go below 5. When pressed, I have used 5, but those two particular analyses never saw the light of peer-reviewed publication. Hopefully, you know enough about the science to support that model.

You could simulate the data to see what effects 5-10 observations have on the estimates and their variance. Unless I had restrictions, I would push hard to get at least 10 observations, but that cannot alwasy be.

Also, searching the archives you will find that we have discussed this, but as it is a tricky issue, we will continue to discuss it...