Identification of Optimal Study Weights in Meta-Analyses with a Binary Outcome

Abstract

Meta-analysis is a method that combines the results of multiple studies, so that the overall treatment effect can be estimated. However, the traditional method of study weight estimation by taking the reciprocals of the estimated variances is biased. For binary outcome data from a clinical trial, the accuracy of estimation of single study weight, summary effect, and variance of summary effect from the developed bias correction factors for log relative risk (RD), log relative risk (lnRR) or log odds ratio (lnOR) were assessed. When sample sizes are small, zero cell frequencies often occur in contingency tables and make parameter estimation more difficult. Methods of dealing with zero-cells were elaborated, which including adding 0.5 to the zero cell, adding 0.5 to all cells in the table if a zero frequency occurs, adding 0.5 to all cells all the time, and adding the reciprocal of the size of the contrasting study arm to each cell when a zero frequency occurs. In addition, for risk difference, adding 0.5 to the zero cells when two zero cells occur, and adding 0.5 to all the cells when two zero cells occur are also considered since the continuity of the weight of risk difference is only affected by double zero frequencies. Impact of bias correction on real meta- analyses from Cochrane Database was demonstrated.