Hi all,

I agree with Kandasamy's answer on collinearity. I however would respectfully disagree in letting the software decide for you. This often is only a preliminary step.

The art of logistic regression is to enter the right amount and the right choice of variables in order to test a precise experimental hypothesis. It can be trial and error and collinearity has to be carefully excluded.

Logistic regression is all but a fishing expedition. Unfortunately, as a reviewer, I too often see authors who blindly report what the software provides, without understanding that this is not the best way to support their experimental hypothesis.

As long as an IV is not perfectly correlated with any other IV, or function of other IVs, you will avoid singularity in the IV covariance matrix and, if so, yes is the answer. If, however, there is linear dependency as I have described, then you can not. In this case, the software package will actually drop one of the predictors, but it will do so without consulting you, so you would have no way of controlling which it drops -- a bad idea.

The next question regards multicollinearity. As the correlations of IVs, or functions thereof, increases, multicollinearity increases. This causes larger variance in the sampling distributions of the prediction weights and larger VIFs. Note that VIFs are applicable to logistic regression as well as OLS regression. An important question, although often not considered, regards the purpose of your model. If it is for prediction (herein, accurate classification) rather than explanation (estimation of population weights), then much of my research shows that multicollinearity -- even extreme multicollinearity -- has no effect. That work is distributed in different journals, although some recent pieces are in the General Linear Model Journal.

Simple and strait forward answer is no, you can not include. Entering an independent variable which is related to other independent variable in a logistic regression leads to multicollinearity issues. As pointed out by Dr Morris, software package automatically selects/drops the predictors without consulting you (of course you had instructed the software to follow certain criteria to select the variables) which is not a good idea.

Hi all,

I agree with Kandasamy's answer on collinearity. I however would respectfully disagree in letting the software decide for you. This often is only a preliminary step.

The art of logistic regression is to enter the right amount and the right choice of variables in order to test a precise experimental hypothesis. It can be trial and error and collinearity has to be carefully excluded.

Logistic regression is all but a fishing expedition. Unfortunately, as a reviewer, I too often see authors who blindly report what the software provides, without understanding that this is not the best way to support their experimental hypothesis.

I want to echo Thierry C Gillebert's concern about the increasing dependence on statistical software to make methodological decisions. Analytic plans should be based on authors' perception of exposure-outcome pathways, in the context of prior published work and the research hypothesis.

In the interesting example that Stephanie Yiallourou has outlined, collinearity may rightly be a concern, but in certain cases, you may have a difficult trade off between confounding and collinearity. A useful metric to quantify the severity of collinearity on your regression models (or in other words to determine whether there is a gross violation of regression assumptions) is the variance inflation factor (VIF).

Another point to consider, is the possibility that variables like age or BMI may be effect modifiers. If this is the case, a stratified analysis would address both confounding and interaction effects, with less concern for collinearity.