Meta-analysis of longitudinal studies.

BACKGROUND: Longitudinal studies typically report estimates of the effect of a treatment or exposure at various times during the course of follow-up. Meta-analyses of these studies must account for correlations between effect estimates from the same study.
PURPOSE: To describe and contrast alternative approaches to handling correlations inherent to longitudinal effect estimates in meta-analyses.
METHODS: Linear mixed-effects models can account for correlations in a number of ways. We considered three alternatives: including study-specific random-effects, correlated time-specific random-effects or a general multivariate specification that also allows correlated within-study residuals. Data from a review of studies of the effect of deep-brain stimulation (DBS) in patients with Parkinson's disease are used to illustrate the application of these models.
RESULTS: are contrasted with those from a naïve meta-analysis in which the correlations are ignored. Results The data included 46 studies that yielded 82 estimates of the effect of DBS measured at 3, 6, 12 months or later after implantation of the stimulator. Models that accounted for correlations, particularly the full multivariate specification, provided better fit (lower AIC) and yielded slightly more precise effect estimates. This was in part due to a relatively extreme observation from a study that provided similar estimates at other times, which in the naïve approach exerts greater influence since it is treated as an independent observation. LIMITATIONS: Since the true values of the parameters are not known, it is impossible to confirm that estimates from the multivariate approach are necessarily more accurate.
CONCLUSION: Standard meta-analytic models can be readily extended to account for correlations between effects in longitudinal studies. These models may provide better fit and possibly more precise summary effect estimates.