Hallo Andre,

Could you please describe your aim and problem more clearly. I understand that you have a binary dependent variable and multiple independent variables, thus logistic regression is the obvious choice. You seem to have a nominal (categorical) independent variable for which one of the categories is empty? But I don't understand how what this has to do with 'mutually exclusive events'.  What is the problem with using logistic regression?

Maybe it would help to clarify the problem if you describe your investigation concretely.

Hi Andre,

the litterature is scarce about the question you ask. If not already read, maybe you will take advantages for the two attached publications. Both are for meta-analysis but may apply for multiple mutually exclusive binary outcomes ? If you merely consider a 0 vs 100% distribution of your outcome ( or even 0 vs a lesser %), I don't think logistic regression models and other classical models for binary outcomes (log-binomial) might converge to provide confidence intervals. One time I have fit GEE-logistic regression model, and it converged. I haven't found my Stata shits. Best pg

Hi all,

Thank you for your answers and help and thank you, Patrick, for the papers. 

Below follows an example of the problem from the Stata output for such situation:

note:

1.exposure

!= 0 predicts success perfectly

1.exposure

dropped and 14 obs not used

note:

2.exposure

omitted because of collinearity

The approach I was taking was to exclude such observations from the analysis but I was looking for a better way to deal with that. I will try the GEE approach.

Thank you for the interest in helping. 

I agree with Jeremy and Subhashisa above. I advise you to consult a statistician and show him or her the data you are working with. Especially if you are considering the use of GEEs for multi-level or repeated measures data.

If the possibility of having a positive (or negative) outcome is based solely on one variable (reading between the lines here, you need to be exposed to an agent in order to be at risk of an event) - then stratify your data based on that variable. So do the regression only on samples that have been exposed to risk and exclude any that have not. This works in inverse if the outcome is related to non-exposure.

This means that any model you create must always be implemented in conjunction with a pre filtering step that determines if exposure has occurred.

I presume this is what you were doing when you say 'The approach I was taking was to exclude such observations from the analysis but I was looking for a better way to deal with that.' I don't understand why this would be a concern. In all statistical analysis you should remove unnecessary decisions from your analysis path to create cleaner models. You just need to specify the pre-analysis decisions clearly so that the analysis can be replicated appropriately.

Don't over complicate the picture just because you feel a single step analysis is somehow more desirable. In screening for disease it is common to have pre-filtering steps to exclude low-risk people from the screening program. These steps are essential for reducing the false positive burden resulting from too widespread an application of the diagnostic model. Your model should be targeted specifically to those who are likely to give a positive outcome (or negative if that is what you target is).