Exploring the efficacy of psychotherapies for depression: a multiverse meta-analysis

BACKGROUND: Hundreds of randomised controlled trials and dozens of meta-analyses have examined psychotherapies for depression—yet not all points in the same direction. Are these discrepancies a result of specific meta-analytical decisions or do most analytical strategies reaching the same conclusion? OBJECTIVE: We aim to solve these discrepancies by conducting a multiverse meta-analysis containing all possible meta-analyses, using all statistical methods. STUDY SELECTION AND ANALYSIS: We searched four bibliographical databases (PubMed, EMBASE, PsycINFO and Cochrane Register of Controlled Trials), including studies published until 1 January 2022. We included all randomised controlled trials comparing psychotherapies with control conditions without restricting the type of psychotherapy, target group, intervention format, control condition and diagnosis. We defined all possible meta-analyses emerging from combinations of these inclusion criteria and estimated the resulting pooled effect sizes with fixed-effect, random-effects, 3-level, robust variance estimation, p-uniform and PET-PEESE (precision-effect test and precision-effect estimate with SE) meta-analysis models. This study was preregistered (https://doi.org/10.1136/bmjopen-2021-050197). FINDINGS: A total of 21 563 records were screened, and 3584 full texts were retrieved; 415 studies met our inclusion criteria containing 1206 effect sizes and 71 454 participants. Based on all possible combinations between inclusion criteria and meta-analytical methods, we calculated 4281 meta-analyses. The average summary effect size for these meta-analyses was Hedges’ g (mean)=0.56, a medium effect size, and ranged from g=−0.66 to 2.51. In total, 90% of these meta-analyses reached a clinically relevant magnitude. CONCLUSIONS AND CLINICAL IMPLICATIONS: The multiverse meta-analysis revealed the overall robustness of the effectiveness of psychotherapies for depression. Notably, meta-analyses that included studies with a high risk of bias, compared the intervention with wait-list control groups, and not correcting for publication bias produced larger effect sizes.