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

Medical interventions may be more effective in some types of individuals than others and identifying characteristics that modify the effectiveness of an intervention is a cornerstone of precision or stratified medicine. The opportunity for detailed examination of treatment-covariate interactions can be an important driver for undertaking an individual participant data (IPD) meta-analysis, rather than a meta-analysis using aggregate data. A number of recent modelling approaches are available. We apply these methods to the Perinatal Antiplatelet Review of International Studies (PARIS) Collaboration IPD dataset and compare estimates between them. We discuss the practical implications of applying these methods, which may be of interest to aid meta-analysists in the use of these, often complex models.

Models compared included the two-stage meta-analysis of interaction terms and one-stage models which fit multiple random effects and separate within and between trial information. Models were fitted for nine covariates and five binary outcomes and results compared.

Interaction terms produced by the methods were generally consistent. We show that where data are sparse and there is low heterogeneity in the covariate distributions across trials, the meta-analysis of interactions may produce unstable estimates and have issues with convergence. In this IPD dataset, varying assumptions by using multiple random effects in one-stage models or using only within trial information made little difference to the estimates of treatment-covariate interaction. Method choice will depend on datasets characteristics and individual preference.