Reproducibility archive: Disparate Impact of Missing-Data Policy in Composite Healthcare Star Ratings

Abstract

Reproducibility archive for the article Disparate Impact of Missing-Data Policy in Composite Healthcare Star Ratings: a pre-registered audit of three missing-data policies (the published CMS Overall Hospital Quality Star Rating design, median impute-then-rank, and an abstaining ranker whose uncertainty interval is built only from exogenous data-support referents) under injected MCAR/MAR/MNAR dropout, read per provider group. Two files. ns-hospital-software-v2.0.0.zip (MIT): the analysis package (data loader, missingness injection, the three rankers with the width-invariance test, the two primary outcomes, the disparate-impact, decision-theoretic, conformal, temporal and second-system modules, the paired Wilcoxon/Holm/bootstrap/permutation statistics), the content-hashed configuration, the three frozen pre-registration documents with the deviations ledger, tests, and REPRODUCE.md. ns-hospital-dataset-v2.0.0.zip (results CC BY 4.0; CMS source files public domain): the SHA-256-pinned CMS Hospital Care Compare release of 2026-04-28 exactly as downloaded (CMS no longer serves it), the 2023 files used for the second-release check, the Nursing Home Care Compare files, the Provider of Services and IPPS impact files with the derived hospital-attribute join, and every result file the manuscript's numbers derive from, including the per-trial grid. Version 2.0.0 adds, relative to the first version: the pinned raw input files; the third pre-registration stage (provider attributes: location, bed size, teaching status, disproportionate-share band; fidelity check of the re-implementation against the published stars; the 2023 replication at 30 repetitions; five-completion multiple imputation with its seed-noise diagnostic) and its post-hoc sensitivity analyses, each labelled as such; and the identity check that the committed summaries regenerate exactly. Every number in the manuscript regenerates from these files with the recorded per-cell seeds.