Estimating interactions in individual participant data meta-analysis: a comparison of methods in practice

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Pharmacogenomics

Background:

• Precision medicine relies on understanding how patient characteristics influence treatment efficacy.
• Individual participant data (IPD) meta-analysis allows detailed examination of treatment-covariate interactions, often preferred over aggregate data meta-analysis.
• Various advanced modeling techniques exist for analyzing these interactions.

Purpose of the Study:

• To apply and compare recent modeling approaches for treatment-covariate interactions.
• To utilize the Perinatal Antiplatelet Review of International Studies (PARIS) Collaboration IPD dataset for this comparison.
• To discuss the practical implications and aid meta-analysts in selecting appropriate complex models.

Main Methods:

• Comparison of a two-stage meta-analysis of interaction terms with one-stage models incorporating multiple random effects.
• Fitting models using nine covariates and five binary outcomes within the PARIS Collaboration IPD dataset.
• Evaluation of model stability, convergence, and consistency of interaction estimates.

Main Results:

• Generally consistent interaction terms were produced across the compared methods.
• Meta-analysis of interactions showed unstable estimates and convergence issues with sparse data and low covariate heterogeneity.
• Varying assumptions in one-stage models (multiple random effects or within-trial information only) had minimal impact on interaction estimates in this dataset.

Conclusions:

• The choice of modeling method for treatment-covariate interactions depends on dataset characteristics and analyst preference.
• While methods are generally consistent, careful consideration is needed for sparse data scenarios to avoid unstable estimates.
• IPD meta-analysis provides valuable insights into personalized medicine by examining treatment effect modifiers.