Overestimation of Relative Risk and Prevalence Ratio: Misuse of Logistic Modeling

Abstract

The extensive use of logistic regression models in analytical epidemiology as well as in randomized clinical trials, often creates inflated estimates of the relative risk (RR). Particularly, in cases where a binary outcome has a high or moderate incidence in the studied population (>10%), the bias in assessing the relative risk may be very high. Meta-analysis studies have estimated that about 40% of the relative risk estimates in prospective investigations, through binary logistic models, lead to extensive bias of the population parameters. The problem of risk inflation also appears in cross-sectional studies with binary outcomes, where the parameter of interest is the prevalence ratio. As an alternative to the use of logistic regression models in both longitudinal and cross-sectional studies, the modified Poisson regression model is proposed.