Assessing heterogeneous causal effects across clusters in partially nested designs.

Area of Science:

• Psychology
• Biostatistics
• Causal Inference

Background:

• Partially nested designs (PNDs) are common in psychological intervention studies.
• Clustering in the treatment arm complicates examining treatment effect heterogeneity.
• Defining causal effects in PNDs is challenging due to the absence of clustering in control arms and potential nonrandomized cluster assignments.

Purpose of the Study:

• To develop methods for defining, identifying, and estimating causal effects of treatment across specific clusters in PNDs.
• To address scenarios where treatment and/or cluster assignments may be nonrandomized.
• To provide a framework for understanding treatment heterogeneity in complex intervention designs.

Main Methods:

• Utilized the principal stratification approach and potential outcomes framework.
• Defined causal estimands for cluster-specific treatment effects under no-interference and within-cluster interference scenarios.
• Employed a multiply-robust estimation method, incorporating machine learning for nuisance model estimation, under the principal ignorability assumption.

Main Results:

• The proposed methods allow for the definition and estimation of cluster-specific causal effects in PNDs.
• The multiply-robust estimator offers protection against model misspecification.
• Simulation studies and an empirical example demonstrated the performance and applicability of the developed methods.

Conclusions:

• The study provides a robust statistical framework for analyzing treatment effect heterogeneity in partially nested designs.
• The developed methods enhance causal inference in psychological intervention studies with complex assignment structures.
• This research offers valuable tools for researchers investigating nuanced treatment effects in clustered settings.