Group Size for Binary Logistic Regression

[michroze]

I am trying to determine the odds ratio for the effect of C on A. However, my is a sample of n = 45 with 4 = yes and 41 = no. Is this small group size a major issue, and how do test for power in this case?

THANKS!

 

[hlsmith]

Is this multivariate regression, are you controlling for any other variables in your model? Is the independent variable you are examining continous or categorical (if categorical, how many groups)?

 

[michroze]

Right now I am performing a generalized mixed model to determine if a group level binary effect changes an individual level binary outcome. The basic model accounts for both the group (unit) and supra-group (hospital) level as random effects.

 

[hlsmith]

Please better explain the model. Is this all of the covariates and groups ---- a single binary variable you are treating as a level?

 

[michroze]

So right now I am trying to perform a bivariate analysis to determine if group level stress affects an individual's blood pressure. I also want to account the effects of nesting: where individuals are found within groups, which are found within hospitals. This random effect is there to eliminate any potential effect of individuals not being completely independent of each other.

Individual's Blood Pressure (High or Low) = Group Level Stress (High or LOW) + (1 | Unit) + ( 1 | Hospitals)

 

[hlsmith]

I would say you are probably trying to do too much with these data. For example create one of those figures you see in articles, illustrating the study sample from the population source when applying inclusion criteria. So what I am getting at is how many people do you have in all of your unique groups and do those people represent everyone comparable to them across the population that are in your sample. So if you have 1 person from hospital, unit, condition, can you say that they represent all of their peers. Probably not.

 

[GretaGarbo]

Analyse the original variable instead (Individual's Blood Pressure ). Classifying them as High low would be like throwing away a large part of the information. And then you can estimate with small sample statistics and get rid of many of the problems.

(And now we get a quite different impression than in the first post.)