Distinguishing between confounders and effect modifiers using stratified analysis and logistic regression. A case study in healthcare epidemiology

The bulk of etiological research in clinical epidemiology consists of observational studies aiming to elucidate the effect of an exposure on an outcome of interest. However, several other factors may be associated with the exposure and/or affect the risk of the outcome. There are two main “complications” from such third variables: confounding and effect modification (interaction). The former is a distortion that must be prevented or controlled, whereas the latter is useful information that may enhance understanding of the phenomenon at hand. This case study presents the example of a prevalent cohort study designed to estimate the prevalence of healthcare-associated infections and assess their impact on inpatient mortality in acute-care hospitals. The case explains elements of the research design in relation to study objectives and illustrates how stratified data analysis may reveal otherwise hidden confounding and distinguish it from effect modification. It also contrasts stratified analysis with multivariable logistic regression and explains the relative merits of the two approaches.