Replicate-level synthetic benchmark for parametric competing-risks inference under censoring and varying latent dependence
Description
This repository provides a reproducible simulation dataset and benchmark for evaluating parametric competing-risks inference under censoring and varying latent dependence. It includes independent latent failure times and Clayton, Frank and Gumbel copula dependence. The dataset supports the assessment of test size, statistical power, parameter bias, root-mean-square error, confidence-interval coverage, Monte Carlo uncertainty, test availability and boundary behaviour across a structured collection of simulation conditions. The final release is approximately 102.11 MiB and contains 343 files. It includes registries describing 303 block-specific scenarios, 291 unique statistical conditions, and the corresponding null and true cause-specific parameters. The main simulation results are stored in 303 Parquet partitions containing 648,000 replicate-level records. Derived products report test-specific rejection and coverage rates, estimator bias and RMSE, Monte Carlo standard errors, Wilson intervals, failure bounds and bootstrap coverage. A representative subject-level Parquet file is also included. It contains 35,250 rows generated from five replicates for each of 93 selected moderate-profile null scenarios. This file provides a documented example of the subject-level data structure. It is not a complete subject-level copy of all 648,000 simulated studies. The repository also contains validation results, four figures with their source data, a comprehensive data dictionary, a file manifest and SHA-256 checksums. The data dictionary documents the fields, stored data types, units, permitted values and missing-value meanings for 27 file schemas. The manifest records and verifies 341 payload files. The manifest and checksum files are excluded from their own inventory. All records were generated through controlled simulation. They do not represent real individuals, patients or observed populations. The dataset is intended for methodological evaluation, software verification, statistical education and reproducible comparison of competing-risks inference procedures. It should not be used to estimate real-world prevalence, incidence or clinical effects. Archive integrity: The SHA-256 checksum of the deposited release ZIP is `42f07751c4143517503bc64975f424c938a7b9fc153544f63808bba2eceb4f40`. This value verifies the ZIP file exactly as deposited. After extraction, the 341 payload files can be verified using `manifests/file_manifest.csv` and `manifests/sha256_checksums.txt`.
Files
Steps to reproduce
1. Download and extract the release ZIP while preserving its directory structure. 2. Open `code/CRM_competing_risks_REFINED_v2_publication_colab.ipynb` in Google Colab. 3. Verify the downloaded release ZIP using the SHA-256 checksum reported in the Mendeley Data record. After extracting the ZIP without changing its folder structure, verify the 341 payload files using `manifests/file_manifest.csv` and `manifests/sha256_checksums.txt`. The manifest and checksum files are excluded from their own inventory. 4. Set `RUN_MODE` to `QUICK` and run the notebook from beginning to end to confirm that the workflow operates correctly. 5. Restart the Colab runtime, set `RUN_MODE` to `PUBLICATION`, and run all cells sequentially. The notebook uses deterministic seeds and version-pinned Python packages. Generated files are saved to Google Drive by default when `SAVE_TO_GOOGLE_DRIVE` is enabled. 6. Confirm that the completed publication run produces 303 scenario summaries, 465 subject-level replicate groups, 27 independent closed-form checks, 361 target-boundary records and 17,127 scenario-specific nuisance-boundary records. 7. Confirm that all 108 programmed validation records in `data/validation_report.csv` have passed. 8. Before running the packaging cells, save the executed notebook and, if required, enter its saved location in `NOTEBOOK_SOURCE_PATH`. 9. Verify the generated release using `manifests/file_manifest.csv` and `manifests/sha256_checksums.txt`. The manifest should contain 341 payload files; the manifest and checksum files are intentionally excluded from their own inventory. Detailed descriptions of the variables, schemas and generated outputs are provided in `README.md` and `DATA_DICTIONARY.md`. For an independent clean execution, open the notebook in a new Google Colab session and configure a new empty output directory. Run every cell in its original order without changing the master seed or simulation settings. An empty output directory triggers generation of all 303 scenario partitions. If the released partitions are used instead, the notebook verifies them and reports `RESUME_VALIDATION`. The original clean-generation timings are provided in `generation_runtime_report.csv`, while `generation_current_run.csv` documents the final resume-validation execution.