Statistical Power for Moderation in Network Meta-Analysis

Title: Statistical Power for Moderation in Network Meta-Analysis

This presentation focuses on power analysis of individual participant data from multiple studies. Comparative effectiveness involves analytic techniques to determine which of several alternative interventions is expected to have the strongest effect. This question of which intervention is more effective can be posed not only in terms of main effects but also as a function of baseline characteristics, as intervention impact may be moderated within a population. For both main effects and moderator analyses, the strongest comparisons of two interventions would be derived from synthesizing findings from randomized trials the two intervention conditions were separate arms of same trials. Often, there are no randomized trials that directly compare two interventions head-to-head, and even if in cases where such trials exist, there may exist other trials, where each intervention is tested against the same control condition, which allow for an indirect comparison of these interventions. This general approach to making comparisons using all available data is called network meta-analysis.

In this presentation, we extend existing network methods for main effects to examining moderator effects. This type of extended synthesis to examining moderation most often requires individual level data and multilevel analyses, in contrast to main effects comparisons where meta-analytic modeling of effect sizes or other summary statistics is routine. We develop a multilevel model approach to network analysis using individual level data, show the power gains that may be obtained by combining trials while taking into account within- and between-trial heterogeneity, and covariates. These methods are illustrated on both simulation study and real data of a classroom-based randomized trial that involved two interventions that were never tested against one another in the same school.