RCT_Bench-main

RCT Bench

RCT Bench is a curated benchmark of public, participant-level randomized controlled trial datasets for evaluating covariate-adjustment methods.

Dataset

The public cleaned dataset contains 125 individually randomized RCTs:

• cleaned_data/trial1.csv through cleaned_data/trial125.csv
• cleaned_data/trial1.rds through cleaned_data/trial125.rds
• meta_data.xlsx
• data-dictionary.xlsx
• rct_bench_all_data_and_dictionaries.zip for one-click download of all cleaned trial files and per-trial dictionary CSVs

Each cleaned trial is participant-level and follows the same variable contract:

• Treatment: randomized treatment assignment, with the control/reference arm first when identifiable.
• YP_*: primary outcome variables. Each trial has at least one.
• YS_*: secondary outcome variables when useful and available.
• X_*: baseline covariates or pre-treatment measurements.
• Participant_ID: an optional de-identified cluster key retained only for crossover/repeated-measures trials; it is not an adjustment covariate.
• Crossover_Sequence, Crossover_Period, and Assessment_Window: optional trial-design fields that are likewise excluded from the X_* covariate set.

Metadata

meta_data.xlsx contains one public metadata workbook for all 125 trials. The main Sheet1 uses an 18-column metadata schema, including the historical spelling Priamry Outcome for compatibility.

Primary Outcome Type uses the controlled categories Continuous, Binary, Count, Time-to-event, Ordinal, and Categorical; trials with co-primary outcomes of different types use semicolon-separated combinations. Research Area uses 15 broad clinical groupings to keep filtering consistent across trials.

Additional workbook sheets preserve expansion curation context where available:

• Provenance
• Validation
• Audit_Summary
• Audit_Detail
• Duplicate_Screen
• Cleanup_Issues

data-dictionary.xlsx contains one row per variable per trial with the variable role, inferred statistical/storage type, R class, missingness, unique-value count, levels or range, and a short explanation.

For expansion trials 51-125, metadata should describe the primary publication rather than the source dataset record. Trial Number/Name is reserved for official registry identifiers such as ClinicalTrials.gov, ISRCTN, UMIN, CTRI, TCTR, ACTRN, DRKS, or similar registries; it is left blank when no registry ID is discoverable. Paper Link should point to the paper DOI or publisher page, and Citation is stored as a numeric citation count.

Data Curation Rationale And Workflow

RCT Bench is designed for method evaluation, so each trial is curated to expose the participant-level structure needed for covariate-adjusted analyses while keeping the analysis table compact and reproducible. The cleaned files retain randomized treatment assignment, publication-aligned primary outcomes, useful secondary outcomes, and baseline covariates measured before treatment.

Expansion trials are screened before inclusion. A candidate must be an individually randomized trial, provide participant-level data with treatment assignment and at least one outcome, have reuse terms compatible with research reuse, and be linkable to a primary publication or official preprint. Cluster trials, observational studies, simulations, reviews, meta-analyses, and aggregate-only datasets are excluded from the public benchmark.

The curation workflow is:

1. Identify candidate datasets from open repositories and confirm eligibility.
2. Link each dataset to the primary publication, trial registry, DOI, and journal record when available.
3. Clean source data into one participant-level .csv and .rds file per trial using the Treatment, YP_*, YS_*, and X_* contract.
4. Select baseline covariates from the publication baseline table or prespecified analysis variables rather than keeping every raw field.
5. Audit cleaned outcomes against publication values when feasible; otherwise record descriptive summary-statistics audit rows.
6. Apply the reproducible public-contract repair step for known identifier, sentinel, crossover-grain, and variable-role issues.
7. Refresh publication-backed metadata, data dictionary, validation summaries, and provenance outputs from scripts rather than editing generated artifacts by hand.

Raw files, downloaded papers, backup candidates, screening queues, and detailed provenance live under ignored local/ staging folders. The public repository keeps the cleaned benchmark, public metadata workbooks, and reusable preprocessing code.

Preprocessing

Reproducible cleaning, screening, validation, and metadata-generation code lives under preprocessing/.

Important entry points:

• preprocessing/RCT_data_cleaning.Rmd: original trial cleaning workflow.
• preprocessing/RCT_data_cleaning_trials51_125.Rmd: expansion trial cleaning workflow for the current flat 125-trial public layout.
• preprocessing/archive/RCT_outcome_reproducibility_expansion.Rmd: retained expansion outcome reproducibility and summary-statistics audit workflow.
• preprocessing/archive/RCT_analysis.Rmd: downstream method-comparison analysis workflow retained from the original repository.
• preprocessing/archive/build_public_metadata.py: rebuilds meta_data.xlsx.
• preprocessing/archive/clean_metadata_trials51_125.py: refreshes publication-backed metadata for trials 51-125 and writes a cell-level provenance CSV.
• preprocessing/archive/build_data_dictionary.R: rebuilds data-dictionary.xlsx, the 125 data_dictionary/trialN_dictionary.csv files, and data_dictionary/validation_summary.json.
• preprocessing/archive/validate_public_dataset.R: validates the flat public cleaned_data/ layout.
• preprocessing/archive/repair_cleaned_data_quality_issues.R: applies the source-backed public-contract repairs before metadata and dictionary rebuilds.

Local Materials

Raw data, downloaded publications, provenance archives, screening workspaces, backup candidates, generated plots, and legacy staging folders are kept under local/. That folder is intentionally ignored by git so the public repository stays focused on the cleaned benchmark and reproducible preprocessing code.

Use

In R, a cleaned trial can be loaded with:

trial1 <- readRDS("cleaned_data/trial1.rds")
trial51 <- read.csv("cleaned_data/trial51.csv", check.names = FALSE)

Before analysis, validate the public layout:

Rscript preprocessing/archive/validate_public_dataset.R

Regenerate metadata and the data dictionary:

Rscript preprocessing/archive/repair_cleaned_data_quality_issues.R
python3 preprocessing/archive/clean_metadata_trials51_125.py  # requires openpyxl and network access for OpenAlex refreshes
python3 preprocessing/archive/build_public_metadata.py
Rscript preprocessing/archive/build_data_dictionary.R