| Physician-randomized trials are clinical trials that randomize physicians to a treatment or control intervention. Patient health outcomes, clustered by physician, are collected to evaluate the effectiveness of the intervention. As an example, we considered a clinical trial of a quality improvement intervention targeted to the physician to improve management of patients at risk of osteoporosis. Several methods exist to analyze correlated outcomes, however the optimal methods for physician randomized trials is unknown.;We performed simulation studies to compare the available methods of analyzing correlated binary and survival outcomes. Using the structure of our study example, we simulated 3 different scenarios: (1) fixed cluster size with normal random effects (2) variable cluster size (3) fixed cluster size with non-normal random effects. We tested conditional, marginal, and nonparametric methods in each of these scenarios. Methods were compared on the basis of power, Type I error, coverage probability, bias, and mean squared error.;Based on our research we recommend the following methods for physician-randomized trials with binary or survival outcomes. For binary outcomes, maximum likelihood estimation and Generalized Estimating Equations perform better than penalized quasi-likelihood and the nonparametric methods tested. For survival outcomes, maximum likelihood estimation and Cox proportional hazards with robust standard errors outperform the nonparametric methods tested. |
