An improper use of statistical significance testing in studying covariables

Abstract

How badly a covariable confounds the association between a disease and a primary trait depends on the strength of two other associations: between covariable and disease, and between covariable and primary trait. To assess the confounding potential by testing one of these associations for statistical significance, using a 'traditional' critical level, is inadequate. Such a preliminary test places the burden of proof in the wrong direction, ignores a major aspect of confounding potential, and may be inappropriately influenced by other factors. If it must be used at all, the significance level should be 0.25 or 0.50, possibly even higher.