I am running a logistic prediction models where one of my categorical (3 categories) independent variables has a 0 cell one of the outcome groups. Is there a way I can deal with this in SAS without having to collapse the categories, since it would not make clinical sense to do so in this situation?
Could you give more detail as to why you can't collapse the groups? And when you say 0 cells, do you mean there were 0 observations for that specific category in the outcome, or it had 0 observations for some of your covariates?
Thursday - last edited Thursday
One of the categories of a predictor I would like to enter into the prediction model has 0 observations for 1 category of the binary outcome. Here are the number for that variable in my data:
X 1 64 42
2 4 13
3 4 0
I do not want to collapse 2 and 3 because it does not make any sense clinically to do so - they are separately meaningful categories of X.
My suggestion would be to drop that category. You can't infer anything from it if you have no response for a category. As well, you're likely to run into some type of problem like quasi-seperation for that category. It maybe something you're really interested in, but it may be best to look only at the other two categories, and then report on the third having too much missing data.
Though, if you really want to keep it and have the time, you could look into multiple imputation. I don't have much experience with that, so I can't give too much input on whether it's even feasible for this situation.
Depending on your research question, sometimes it won't be informative and you would need to drop the category if you can't collapse it with another group. So you could go with what SAS_Rob mentioned and do an exact logistic regression.