Meta-analysis of rare binary events in treatment groups with unequal variability

Abstract

Meta-analysis has been widely used to synthesize information from related studies to achieve reliable findings. However, in studies of rare events, the event counts are often low or even zero, and so standard meta-analysis methods such as fixed-effect models with continuity correction may cause substantial bias in estimation. Recently, Bhaumik et al. developed a simple average estimator for the overall treatment effect based on a random effects model. They proved that the simple average method with the continuity correction factor 0.5 (SA_0.5) is the least biased for large samples and showed via simulation that it has superior performance when compared with other commonly used estimators. However, the random effects models used in previous work are restrictive because they all assume that the variability in the treatment group is equal to or always greater than that in the control group. Under a general framework that explicitly allows treatment groups with unequal variability but assumes no direction, we prove that SA_0.5 is still the least biased for large samples. Meanwhile, to account for a trade-off between the bias and variance in estimation, we consider the mean squared error to assess estimation efficiency and show that SA_0.5 fails to minimize the mean squared error. Under a new random effects model that accommodates groups with unequal variability, we thoroughly compare the performance of various methods for both large and small samples via simulation and draw conclusions about when to use which method in terms of bias, mean squared error, type I error, and confidence interval coverage. A data example of rosiglitazone meta-analysis is used to provide further comparison.