| Statement of the problem. Results from multi-center randomized, observational, and cross-sectional studies involving binary outcomes are often reported in terms that lack meaning and usefulness to audiences of clinicians or health policy analysts. In addition, the statistical methods used for multi-center studies are often applied to data generated by biological or behavioral processes that are unobserved, without regard to the disparity between the underlying form of the data and the model for estimation. Procedures and methods. Computer-based simulations tested the performance of alternative methods for estimating, with appropriate confidence intervals (variance), relative risk for single center studies, and odds ratios and risk difference for multi-center, and the performance of alternative methods for estimating treatment effects with rare outcomes. Results. For single center studies, logistic regression using the delta method or bootstrap resampling outperformed alternative methods, such as Poisson regression, for estimating relative risk. In multi-center studies, logistic regression is also more robust to misspecification than is linear regression when the additive or multiplicative nature of the true disease process is not observed. In the context of repeated cross-sectional and longitudinal cluster-randomization studies with binary outcomes, the bias and coverage of statistical methods depends heavily on the exact study design. Bayesian methods exhibit the best overall performance, especially when patients are followed within centers over time and the number of centers is small. Finally, the performance of statistical methods in the analysis of rare outcomes from multi-center studies also depends on the study design, and "exact methods" do not routinely exhibit the best performance. Conclusions . For the analysis of binary outcomes from both single and multi-center studies, logistic regression is often the method of choice, regardless of whether the goal of analysis is an estimate of the odds ratio or risk difference. But the optimal method of implementing logistic regression depends on the precise study design. When outcomes are rare, no single method of analyzing binary outcomes performs best. |
