Does Exact Test of Poisson Models Yield Reasonable Confidence Intervals When Estimating Relative Risks for Small Samples with Common Binary Outcomes

• Activity Number: 624
• Type: Contributed
• Date/Time: Thursday, August 2, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Wansu Chen*+ and Zoe Li and Michael Schatz and Robert Zeiger
• Companies: Kaiser Permanente and Kaiser Permanente and Kaiser Permanente and Kaiser Permanente
• Keywords: relative risk ; log-binomial regression ; Poisson regression ; robust estimator ; exact test

Abstract

To estimate relative risks or risk ratios for common binary outcomes, the most popular model-based methods are the robust Poisson and the log-binomial regression. However, due to the limitations of Wald-type or likelihood ratio based confidence intervals (CI) for small samples, the performance of the two commonly used methods is not ideal, especially when continuous covariates exist (results presented at JSM 2011). In this study, simulation was conducted to evaluate the empirical coverage of 95% CI based on the exact test of Poisson models. Our findings suggest that the exact test yielded overly conservative 95% CI for all scenarios we examined. Compared to the Wald-type CI generated by the robust Poisson models and both Wald-type and likelihood ratio based CI generated by the log-binomial models, the exact test of the Poisson models did not improve the performance in terms of coverage. Users should be aware of the limitations when using the exact test of the Poisson models, the robust Poisson models and the log-binomial models to estimate relative risks in small samples with common binary outcomes.