Reporting logistic regression results

#1
Hi,

I have a study that aims to find out the association of a binary outcome (e.g. BP control-Y/N) with categorical independent variables e.g. their dietary and exercise preferences. As I have a very long list of these independent variables (50 qns on dietary and 50 qns on exercise), do I perform one regression for each question on dietary vs BP control, adjusting for their confounding factors like age, gender, medical conditions?

If so, how do I report the summarized results as there would be results from 100 regressions? I cannot possibly put the ORs into one table as I am not including all the variables in one model.

Also, take for example one of the regression results shows significance for age and the dietary preference question I included in the model, can i conclude that the dietary preference question is associated with BP control, and how do I interpret the association with age? Age might not be significant when other dietary preference questions are included.

Thanks for all the help!
 

noetsi

Fortran must die
#3
I don't think you would have a hundred regressions (that is a hundred DV), I think you would have one regression with a hundred predictors. But I am confused what exactly you are doing. Do you have a hundred predictors or do you want to predict a hundred dependent variables. Only in the latter case would you say normally you ran a hundred regressions.

I don't think the idea of having a 100 IV variables (if that is what occuring) in a model is a great one period (it gets worse of course if you have a small sample size). I don't think I have every seen one done with a hundred variables. You could add to any model of course every variable that might contribute and thus model reality, but what have you gained? There is a difference between analysis and a simple modeling of every variable that might impact the results.
 
#4
Hi both,

Thanks for the replies. My sample is 1100. I understand that for logistic regression we need at least 10 observations per predictor added to the model.

Prior to logistic regression, I have done many chi-square tests to check if the association between BP control and all the 100 variables are significant, but these tests are not controlled for the potential confounders such as age, weight, comorbidities, drinking habit etc. Hence I am using logistic regression to control for that, which is why i would end up having 1DV (BP control) against 1 IV (1 of the 100 IVs) + age+weight+comorbidities+drinking habit etc as one regression. Since I have 100 variables, does that mean I have to do the same, replacing the 1 IV tested in the model for 100 times? Which is why I am puzzled how to interpret the results if that were the case.

Hope I have clarified my question. :)

Thanks!

I don't think you would have a hundred regressions (that is a hundred DV), I think you would have one regression with a hundred predictors. But I am confused what exactly you are doing. Do you have a hundred predictors or do you want to predict a hundred dependent variables. Only in the latter case would you say normally you ran a hundred regressions.

I don't think the idea of having a 100 IV variables (if that is what occuring) in a model is a great one period (it gets worse of course if you have a small sample size). I don't think I have every seen one done with a hundred variables. You could add to any model of course every variable that might contribute and thus model reality, but what have you gained? There is a difference between analysis and a simple modeling of every variable that might impact the results.