Which combinations of behaviour change techniques are effective? Assessing interaction effects in meta-analysis

Abstract

Meta-analysis is an important tool to synthesize results from multiple studies in a systematic way. Interaction effects play a central role in assessing conditions under which the relationship between study features and study effect sizes differs in strength and/or direction. For example, the question “which specific combinations of behavior change techniques (BCTs) used by health promotion interventions are effectiveâ€, can be investigated by estimating interaction effects between BCTs in a meta-analysis. However, when several study features are available, regression in meta-analysis lacks sufficient power to detect interactions between them. To overcome this problem, a new approach named "meta-CART" (Dusseldorp et al., 2014) was proposed that introduced classification and regression trees (CART) in the field of meta-analysis. The first version of meta-CART had some shortcomings: the study effect sizes were dichotomized and were not weighted by their accuracies. In the present study, new meta-CART extensions are proposed, without dichotomization of effect sizes and with the use of weights. The performances of all versions of meta-CART were evaluated via an extensive simulation study. The results revealed that: a) the Type I error was low for both meta-classification trees and meta-regression trees; b) meta-regression trees without weights showed the highest detection rates, and c) the required number of studies depended on the number of study characteristics, the magnitude of interaction, and the residual heterogeneity. Based on the results, user guidelines for meta-CART were formulated, such as, the minimum number of studies should be 40 and the number of features may exceed 20.