RCT_Bench — A Public Benchmark of Real-World RCT Datasets for Covariate Adjustment Research

This repository provides a curated, analysis-ready benchmark of real-world randomized controlled trial (RCT) datasets, together with the R pipelines used to clean and analyze them. It accompanies the study "How should covariates be handled in randomized trials? Empirical evidence from 50 trials and recommendations for practice" (Yulin Shao, Liangbo Lyu, Menggang Yu, Bingkai Wang — Department of Biostatistics, University of Michigan).

The goal of this repo is to give methodologists, trialists, and statisticians a ready-to-use, standardized benchmark for evaluating covariate-adjustment methods, variable-selection strategies, and other estimation procedures on real (not simulated) trial data — without having to track down, clean, and harmonize dozens of individual trial datasets.

What's Included

• Cleaned individually-randomized (non-clustered) RCT datasets — standardized, analysis-ready data from publicly available trials, with harmonized variable naming so datasets can be processed interchangeably
• meta_data.xlsx — trial-level metadata (trial name, publication, journal, sample size, primary outcome and type, study phase, randomization scheme, research area, citation, etc.) for each trial in the benchmark
• meta_data_demo.xlsx — per-trial participant demographic breakdowns (sex and race/ethnicity counts, e.g., Male/Female/Other and Asian/Black/White/Hispanic/etc., by trial)
• meta_data_comparison.xlsx — the full method-comparison results table: for every trial/treatment-outcome comparison and covariate-selection strategy (All / Top-3 / Baseline+), the point estimates, standard errors, p-values, and precision-gain metrics for each estimator (Unadjusted, ANCOVA, ANHECOVA, IPW, g-logistic, DML, TMLE, AIPW, SuperLearner-based, PSWeight-based, etc.)
• RCT_data_cleaning.Rmd — the unified data-cleaning pipeline that transforms raw trial data into the standardized format provided here (included for transparency and so users can apply the same conventions to their own datasets)
• RCT_analysis.Rmd — an end-to-end analysis pipeline demonstrating how to apply and compare covariate-adjustment estimators (e.g., ANCOVA, ANHECOVA, IPW, g-computation, machine-learning-based estimators such as DML and TMLE) across multiple covariate-selection strategies on these datasets

Note: This public release includes only the non-clustered (individually randomized) trial datasets. Clustered-RCT data and other supplementary materials from the original study are not part of this benchmark.

Variable Naming Convention

All cleaned datasets follow a common schema so they can be looped over programmatically:

• YP_* — primary outcome(s)
• YS_* — secondary outcome(s)
• X_* — baseline covariates
• Treatment — treatment assignment

Repository Structure

RCT_Data/
├── cleaned_data/
│   ├── Non_Clustered_RCT/        # Standardized, analysis-ready datasets (one per trial)
│   ├── meta_data.xlsx            # Trial-level metadata (publication, sample size, outcomes, etc.)
│   ├── meta_data_demo.xlsx       # Per-trial participant demographics (sex, race/ethnicity)
│   └── meta_data_comparison.xlsx # Full estimator-comparison results (estimates, SEs, p-values, precision gains)
├── RCT_data_cleaning.Rmd         # Pipeline used to produce the cleaned datasets
└── RCT_analysis.Rmd              # Example pipeline: apply & compare adjustment methods