Comparisons of Model-Based Approaches to Estimate Relative Risks for Common Binary Outcomes

• Activity Number: 139
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Wansu Chen*+ and Feng Zhang and Michael Schatz and Zoe Li and Robert S. Zeiger
• Companies: Kaiser Permanente Southern California and Kaiser Permanente Southern California and Kaiser Permanente Southern California and Kaiser Permanente Southern California and Kaiser Permanente Southern California
• Keywords: common binary outcome ; relative risk ; risk ratio ; generalized estimating equation ; bias correction ; sandwich estimator

Abstract

To estimate relative risks(RR)for common binary outcomes,the most popular model-based methods are the robust Poisson and the log-binomial regression. A simulation was conducted to evaluate the statistical properties of three existing model-based approaches along with a pseudo-likelihood based method with a bias-corrected sandwich estimator (MBN method) that is available in SAS for various sample sizes. For small samples(n<=50), the MBN method corrected type I errors when the RR and the outcome were large. With continuous covariates, the COPY and non-linear programming (NLP) methods yielded low coverage most of time when sample sizes were small and the RR was one regardless of the type of confidence intervals (CI) used. For models with small sample sizes and a continuous covariate, likelihood ratio-based 95% CI were too narrow. All four methods yielded comparable bias. The findings suggest that the MBN method could outperform other methods in some scenarios in terms of coverage when sample size is small. Users should be cautious when using the log-binomial/COPY method or the NLP method to estimate RR in small samples when the effect is small and continuous covariates are included.