Fisher's Exact test and subjects with multiple observations?

[NenaB]

Hi Everyone,
I was wondering if I could please have some advice on how to deal with data that have multiple observations for some of the subjects (but not all).

We are looking at list of 150 patients who are currently receiving therapy for various illnesses/indications. We are planning to stratify patients based on some characteristics (which would be then our 'patient groups') to see if there are differences in the "proportions of optimal therapies" amongst different patient groups.
E.g. To see if there is a significantly larger proportion of optimal therapy for patients with Indication A compared to Indication B.

However, some (but not all) patients are receiving multiple therapies as they have several illnesses that they are receiving treatment for at once. So looking at the proportion of appropriate therapies means we'd have some observations that are related because they have the same patient source.

Originally I had thought of using Fisher's Exact test to test out the difference amongst proportions of appropriate therapies in different patient groups. But now that I realise some of the observations are related, I dont believe this test is appropriate. I have seen some papers published with the same exact data and using the fishers exact test but I am not sure at this point.

If anyone could provide any advice on how to deal with these data I would very much appreciate it.

many many thanks in advance
Nena

 

[hlsmith]

Fisher would be inappropriate.

 

[NenaB]

Fisher would be inappropriate.

Thanks so much. Would you have any other advice on how to deal with the data and/or test the differences amongst the proportions?

Thanks so much again,
Nena

 

[hlsmith]

How many groups do you have? Also for the dependent observation (multiple obs) how many unique patients are there? Also, if they are contributing multiple data observations, are these observation for the exact same time periods. Or is it for multiple encounters or admissions? If it is not for the same moment, then the ordering of treatments may even create more groups (renal patient then cardio could differ from cardio to renal patient).


My first inclining would be there are not many of these patients. If there are only a handful, I would through them out, I am guessing for your purposes it would be too difficult to attempt to address this, plus if there are only a few, how representative can they be (given sampling variation) to all other patients with multiple conditions. If you throw them out, you then just need to disclose this in your results.

 

[NenaB]

Thank you so much again for your response.

The observations are all taken at once - and just the one time only.
I.e. We are recording the conditions and treatments for all patients on one/same day only.
However, the patients included are complicated cases and unfortunately many have multiple conditions that they are receiving therapy for - I would say approximately half the patients in the study would have 2 or more conditions.

I am trying to understand "Generalized Estimating Equations", as, based on the small readings that I have done so far, it seems that "GEEs" would be a good method to use with this type of data?

Would you have any suggestions?

Thank you so much again, your help is much appreciated.
N