Aligning One-stage and Two-stage Approaches for Combining Individual Participant and Aggregated Data in Meta-analysis

• In Event: Effect Estimation for Causality and Prediction: Multilevel and Single-case Designs, Machine Learning, and Distributional Assumptions

Abstract

This study evaluates one-stage and two-stage methods for synthesizing individual participant data (IPD) and aggregated data (AD) in meta-analysis, focusing on continuous outcomes in social science. Although both approaches have been long established, limited attention has been given to aligning their distributional assumptions for model parameters. Using data from Zettersten et
al. (2024), I compare approaches under similar distributional assumptions for sample-specific parameters and explore synthesizing treatment group-specific outcomes when AD provides group means. Results show little within-sample moderation and significant between-sample moderation. Between-sample moderation estimates diverge slightly between one-stage and two-stage analyses using AD effect estimates but align closely when using AD group means. Within-sample moderation estimates are consistent across all modeling approaches.

Authors

• Jingru Zhang, University of Wisconsin - Madison

• James E. Pustejovsky, University of Wisconsin - Madison