Meta-Analysis of Effect Sizes Reported at Multiple Time Points Using General Linear Mixed Model

• Published in: PloS one Vol. 11; no. 10; p. e0164898
• Main Authors: Musekiwa, Alfred, Manda, Samuel O M, Mwambi, Henry G, Chen, Ding-Geng
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 31-10-2016; Public Library of Science (PLoS)
• Subjects: Algorithms; Analysis; Cancer; Clinical trials; Combined Modality Therapy; Computer simulation; Correlation analysis; Covariance; Economic models; Evidence-based medicine; Glioma; Humans; Likelihood Functions; Linear Models; Longitudinal Studies; Medicine and Health Sciences; Meta-analysis; Meta-Analysis as Topic; Models, Statistical; Neoplasms - mortality; Neoplasms - therapy; Odds Ratio; Parameter estimation; Physical Sciences; Postoperative Care; Research and Analysis Methods; Research Design; Sample Size; Studies; South Africa
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0164898

Meta-analysis of longitudinal studies combines effect sizes measured at pre-determined time points. The most common approach involves performing separate univariate meta-analyses at individual time points. This simplistic approach ignores dependence between longitudinal effect sizes, which might result in less precise parameter estimates. In this paper, we show how to conduct a meta-analysis of longitudinal effect sizes where we contrast different covariance structures for dependence between effect sizes, both within and between studies. We propose new combinations of covariance structures for the dependence between effect size and utilize a practical example involving meta-analysis of 17 trials comparing postoperative treatments for a type of cancer, where survival is measured at 6, 12, 18 and 24 months post randomization. Although the results from this particular data set show the benefit of accounting for within-study serial correlation between effect sizes, simulations are required to confirm these results.