Modern Causal Inference Approaches to Improve Power for Subgroup Analysis in Randomized Controlled Trials

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Health Services Research

Background:

• Subgroup analyses in randomized controlled trials (RCTs) assess treatment effect heterogeneity but are often limited by small sample sizes, reducing statistical power.
• Existing methods to boost power, such as covariate adjustment and borrowing external data, have limitations, especially with small trial samples and practical positivity violations in external data.

Purpose of the Study:

• To develop and present an approach for enhancing statistical power in preplanned subgroup analyses of small RCTs.
• To leverage both baseline predictors and external data to improve the detection of heterogeneous treatment effects across patient subgroups.

Main Methods:

• Proposed de-biased estimators accommodating parametric, machine learning (ML), and nonparametric Bayesian methods.
• Introduced three estimators to address practical positivity violations (PPVs): a covariate-balancing approach, an automated de-biased machine learning (DML) estimator, and a calibrated-DML estimator.
• Evaluated methods through simulations and applied them to a real-world case study involving citalopram for schizophrenia.

Main Results:

• Demonstrated improved statistical power in simulations using the proposed de-biased estimators.
• The introduced estimators effectively handled practical positivity violations, leading to more stable inferences.
• The methods were successfully applied to analyze citalopram effectiveness in first-episode schizophrenia (FES) patients across subgroups defined by duration of untreated psychosis (DUP).

Conclusions:

• The proposed methods offer a robust framework for improving power in subgroup analyses of small RCTs by integrating baseline predictors and external data.
• These techniques provide practical solutions for challenges like model misspecification and positivity violations, enhancing the reliability of subgroup effect estimates.
• The findings have significant implications for designing and analyzing clinical trials, particularly in rare diseases or specialized populations where small sample sizes are common.