# Change of d-irection: current limitations and future directions in psychological meta-analysis

## TL;DR

This paper discusses limitations of using standardized effect sizes in psychological meta-analysis and suggests better alternatives like multivariate and imputation-based methods.

## Contribution

The paper introduces and evaluates multivariate and imputation-based approaches as alternatives to traditional standardized effect sizes in psychological meta-analysis.

## Key findings

- Multivariate meta-analysis provides meaningful and precise estimates under missingness at random.
- Imputation techniques offer flexibility for handling non-ignorable missing outcome measures.
- Adopting these methods can reduce bias and improve the interpretability of psychological findings.

## Abstract

Meta-analysis is a statistical tool used to combine the results of multiple studies to answer a research question. In psychology, effects are often measured on different scales (i.e., with different units), and their aggregation is not trivial. The problem is commonly solved using standardized effect sizes such as Cohen’s d. Despite being widely adopted, this approach is flawed. The misunderstanding is that standardized measures are dimensionless by definition—two d do not share the same dimension, as they do not have any. The present work explores alternative approaches to meta-analysis: Multivariate meta-analysis jointly models correlated outcomes while preserving their original unit; Imputation techniques treat the different outcome measures as a missing data problem. We evaluate these approaches through Monte Carlo simulation and an application to real data from psychotherapy studies. Results confirm that under missingness at random, multivariate meta-analysis provides meaningful and precise estimates. Imputation techniques offer an even more flexible alternative for dealing with non-ignorable missing outcome measures. The findings encourage the adoption of multivariate and imputation-based meta-analysis techniques to reduce bias, avoid research waste, and enhance the interpretability of psychological findings.