Comparison of Two Methods to Correct for Non-differential Misclassification in Meta-analysis

Abstract

Like all data, administrative health data are susceptible to bias. In this project I focus on bias due to misclassification of disease outcome in the context of a meta-analysis. I propose a novel approach to random effects estimation in the presence of misclassification based on a method proposed in the literature for fixed effects estimation. Both these approaches to meta-analysis in the presence of misclassification adjust the study-specific variance in log odds ratio for the presence of between-study variance in misclassification rates. Monte Carlo simulation is used to compare these variance correction approaches to naïve (non-variance) correction approaches. The simulation demonstrates that, in fact, the naïve correction procedure yields effect estimates that are less biased than those yielded by the variance correction procedure, and its coverage probability is closer to its nominal value. High false negative rates are observed for all homogeneity statistics, while their false positive rates remain low.