Interpreting the literature in obstetrics and gynecology: II. Logistic regression and related issues

The goal of epidemiology is accurate measure of the relationship between an exposure and a disease of interest. The control of covariates of the exposure-disease relationship is required to obtain a valid measure. Two types of covariates, confounders and effect modifiers, must be considered. Investigators can design studies to measure and control for the impact of both types of covariates. Design strategies for dealing with covariates include randomization, restriction, and matching. If the impact of a covariate is not eliminated by study design, it must be controlled for during study analysis by use of either stratification or mathematical modeling. Stratified analysis permits an assessment of the exposure-disease relationship for each category of relevant covariates. Although stratification is the best initial approach for controlling covariates, it is often impractical, particularly if more than one or two covariates must be controlled. Multivariate mathematical models are required if multiple covariates are to be controlled. Logistic regression is the mathematical modeling procedure most often used to analyze studies in obstetrics and gynecology. Although there are no uniform rules for building a proper model for regression analysis, useful general strategies are available. It must be emphasized that, though the use of mathematical modeling can control for multiple covariates and thereby improve the chance of obtaining an accurate measure of the exposure-disease relationship, it cannot "fix" data that result from a poorly designed or improperly conducted study.