Abstract:
|
Current controversies in meta analysis include identifying optimal methods for combining results of clinical trials or observational studies when effects of treatment or exposure are constant on the log odds ratio, log relative risk, or risk difference scales. Investigators must select appropriate statistical analyses when the correct scale of effect of treatment or exposure is unknown and therefore the statistical model might be misspecified We report results of simulations that: 1.) create multi-trial datasets assuming constant log odds ratio, log relative risk, and risk difference and 2.) assess the bias and coverage of common statistical methods used for meta analysis, including: random effects methods of DerSimonian and Laird, and mixed effects regression. For example, computing risk difference, using mixed effects logistic regression and the delta method, appears robust to model misspecification, at least when the control group risk is moderate (p=0.2) and treatment effect is modest (RR=2.0). Variance estimation using both asymptotic (delta) and Bayesian (Gibbs sampling) methods are discussed in terms of performance and ease of implementation.
|