Is it possible to use robust standard errors with exact logistic regression?

[Charlotte Kerr]

Hi,

I have a small sample, with infrequent outcomes, which also has some clustering in the sampling. I would like to use exact logistic regression due to the small sample and sparse outcomes but also clustered robust standard errors to account for the custering. To establish prevalence this was fine ( proportion var1,citype(exact) vce(cluster clustervar) percent). However this does not work in logistic regression (eg exlogistic var1 var2, vce(cluster clustervar or logistic var1 var2, citype(exact) vce(cluster clustervar)).

Does anyone know a way to do this?

Thanks.

Charlotte

[Joseph Coveney]

With -exlogistic-, you can use the -group()- option to stratify on a variable. So, maybe try

exlogistic var1 var2, group(clustervar)

unless you have a lot of incomplete clusters.

Is var2 categorical or continuous?

You also might want to consider a power analysis, if for no other reason than to see how misleading a statistically significant result will be for opportunistic null hypothesis statistical testing with your sample size and prevalence.

[Charlotte Kerr]

Thanks Joseph,

A colleague has said I may not need to perform robust standard errors in exact logistic regression as there is not the assumption of normalcy in exlogistic and so I dont need my SEs to be robust to this. Any thoughts?

Var2 would be categorical although if I use age as an a priori confounding variable it is as a continuous variable.

I think doing a power calculation would be a good idea, so that this could at least be presented alongside results