Abstract:
|
Meta-analysis is useful for synthesizing results from multiple studies to evaluate the effect of a treatment on a response variable of interest. A common difficulty occurs when studies report treatment effects in terms of inconsistent, but correlated, outcome measures. We first explore two simple strategies: the 'no pooling' approach, in which each measure is analyzed separately, and the 'complete pooling' approach, in which the measures are analyzed jointly after standardization. The former approach is appropriate if the various measures are uncorrelated, while the latter approach is appropriate if they are perfectly correlated. We next propose a compromise, the 'partial pooling' approach, in which each measure is analyzed separately, but prior information about correlations between measures is used to regularize treatment effect estimates via Bayesian inference. We perform a simulation study to evaluate the relative performance of the different approaches in different contexts.
|