Adjusting for confounding with statistical software packages.

Abstract

Much computer use in support of clinical research involves the execution of standard statistical packages. Prenatal risks, such as maternal substance use, may be related to outcomes, such as lowered birthweight for gestational age, using multiple regression. In standard stepwise multiple regression, variance shared by two risks is attributed to the first variable entered in the equation, potentially underestimating the effect size of subsequently entered variables. A novel solution to this problem, adapted from the epidemiologic technique of calculating attributable risk, is applied in the multivariate setting and produces more appropriate estimates of the effect sizes of independent risks. An example of this method, using a standard statistical package, is presented.