Variable selection should be blinded to the outcome

When multivariate models are used for confounder adjustment, all potential confounders should be included in the model, or if not, their selection should be blinded to the outcome. Other methods, such as “pre-filtering” variables based on their univariate association with the outcome may give rise to biased regression coefficients, biased standard errors, biased confidence intervals, misspecified test distributions and exaggerated p-values. If variable selection is needed (for instance due to the high number of potential confounders) data reduction methods that are blinded to the outcome, modern approaches such as Bayesian Model Averaging or the application of shrinkage, i.e. penalization (regularization) of the regression are preferable.