Comparison of regression models for binary outcome variables in clinical trials

Abstract

The widely used logistic regression may not be suitable to model binary outcomes in clinical trials. The present study compares and assesses various binary regression models such as logistic, log-binomial, Poisson and Cox proportional models for clinical trials. A dataset obtained from a clinical trial conducted on tuberculosis patients is used to illustrate the models. The estimated odds ratios from logistic regression severely overestimated the relative risks, thereby overestimating the overall relationship. Log-binomial and Poisson with sandwich covariance estimator were found to be suitable for estimating adjusted relative risks in clinical trials