Multiple logistic regression interpretation help

#1
I'm looking at the association between one binary dependent variable, and several independent variables. I first did univariate analysis, and then put all variables with p<0.05 into a logistic regression model.

My issue is this: a couple of the variables produced odds ratios <1 on regression analysis, despite univariate analysis demonstrating positive associations between the dependent variable and the independent variables in question (i.e. I would have expected OR>1).

Can this be explained by taking into account the confounding effect of other variables? Or does it mean I have done something wrong?

Thanks :)
 

noetsi

Fortran must die
#3
Univariate analysis does not mean anything for multivariate analysis. It is entirely possible the sign can be reversed. It is the impact of x on Y controlling for the other variables (which may be what you mean by confounding).
 

hlsmith

Less is more. Stay pure. Stay poor.
#4
I would also add, how big was it's effect in bivariate analyses versus multiple logistic regression? @noetsi - is correct in that bivariate analyses are not adjusting for other covariates, which may interact, mediate, or unconfound relationships. Using bivariate <0.05 as a screening is frowned upon. Model building should be based on your context knowledge of the underlying data generating process. Also, results from such a model should be report with a high level of reservation on their generalizability if they are not validated using a holdout dataset.
 
#5
Thanks for the responses everyone. For context, I have attached the simplified results tables.

I should add - this is just for an undergraduate dissertation, and I have no prior stats experience. This is why I decided to use bivariate <0.05 for inclusion in multivariate analysis. It seemed like the most straightforward method, and one which I have seen done in several studies similar to mine.

Variables G and M were the ones I had wondered about specifically - I had expected OR>1.
 

Attachments

hlsmith

Less is more. Stay pure. Stay poor.
#6
Of note, most people may say your model could only support 3-4 predictors, given the ~70 outcome events.