abigailhaddad/hhs-dab-decisions

HHS Departmental Appeals Board decisions

Every decision the HHS Departmental Appeals Board has published, as typed Parquet with the header and parts of the body parsed into columns.

The Board is an administrative tribunal inside HHS. Its ALJs (Civil Remedies Division) hear a case first and its Appellate Division reviews them, so the two files are two levels of the same tribunal and can be joined on reviews_decision_no / appealed_in.

file	tribunal	decisions	span
dab.parquet	Appellate Division, and the Grant Appeals Board before it	3,309	1974-03-07 – 2026-08-21
alj.parquet	Civil Remedies Division (ALJs)	6,426	1985-05-14 – 2026-07-28
council.parquet	Medicare Appeals Council — a curated selection, see below	224	2001-03-21 – 2015-04-24
slices.parquet	category membership — 8,006 rows over 77 non-empty slices

The Appellate and ALJ corpora are complete against the Board's own published index: of the 9,809 listings on its year-by-year index pages, all 9,486 of the 9,486 that carry an identifiable decision number are here; the other 323 sit outside the check. The Council file is different, and by a wide margin: the Board posts only "certain significant decisions and actions ... selected for posting since they involve the adjudication of issues that may be of interest". For scale, HHS put the Council's annual adjudication capacity at roughly 2,600 appeals in FY2016 and 2,300 at the end of FY2015, against a pending workload above 14,000 (Medicare Appeals Backlog primer). This file holds 224 decisions spanning 2001 to 2015 — on the order of a percent of one decade's output. It is a selection of significant decisions, and a count over it is not a count of what the Council decided.

Derived from Kmisener/HHS-DAB-Decisions, which collected and text-extracted the original corpus. This adds the decisions it was missing, parsed metadata, website-furniture removal, evidence-based category labels and a typed schema. Build code and method: abigailhaddad/HHS_DAB_Collect-Sort (a fork of the repository the source dataset came from).

What the decisions are about

On the ALJ side, most cases run against CMS or the HHS Inspector General: exclusions barring an individual from Medicare and Medicaid, nursing-home enforcement and civil money penalties, revoked billing privileges, CLIA lab sanctions, EMTALA penalties, and FDA tobacco-retailer enforcement. On the Appellate side, and especially before 2000, a large share are grant disputes — Head Start, TANF, child support enforcement, foster care, tribal self-determination contracts.

Columns

column	notes
id, corpus	source filename stem; dab or alj
decision_no, decision_no_source	2740, CR4685, CR10R, RULING2013-2; and which of index / filename / header it came from
docket_nos	a decision can carry several
decision_date, year	from the text; null rather than wrong
tribunal, respondent	which division; who the appeal ran against
judges	a panel on the Appellate side (60% of Appellate, 76% of ALJ)
reviews_decision_no	the ALJ decision an Appellate decision reviews (759)
appealed_in	ALJ only — Appellate decisions that reviewed it
dispositions	affirmed, reversed, remanded, … (43% / 24%)
disposition_text	the conclusion, verbatim
provider_ids	CCN:, NPI:, PTAN:, CLIA: — joins to CMS data (36% of ALJ)
source_url	where it was collected from
had_web_chrome, text_layer_ok, clean_guard_tripped	extraction flags
text	the decision, website furniture stripped

Quick start

import duckdb
duckdb.sql("""
  SELECT unnest(dispositions) AS outcome, count(*) AS n
  FROM 'dab.parquet' WHERE reviews_decision_no IS NOT NULL
  GROUP BY 1 ORDER BY 2 DESC
""").show()

What this does not establish

• Dispositions are partial by design. They are read only from operative first-person phrasing in a decision's own conclusion. Empty means "not stated unambiguously here", never "nothing happened", so a rate over the labelled rows is a rate over a non-random subset. disposition_text carries the conclusion verbatim.
• appealed_in is a join, not a census. An appeal that settled, was withdrawn, or is pending produces no Appellate decision, so empty means "no Appellate decision here names it".
• Category labels are a heuristic. Membership is decided from where a citation appears and how often, after case citations are blanked out. No precision or recall is claimed; there is no hand-labelled sample.
• The slices are not a partition. 6,333 of 9,959 decisions (64%) land in at least one slice and 1,419 in more than one, out of 76 possible category/corpus pairs. There is no residual category.
• Some text is wrong rather than missing. A few decisions extract to about one character per page and are flagged by text_layer_ok. Others extract at normal length and are garbled — DAB No. 88 (1980) reads "Financisl", "yesr" — and nothing flags those. DAB No. 437 (1983) is published by HHS as an empty page and is absent for that reason.
• Decision numbers come from three sources, best first: the Board's index, the source filename, then the header. The header alone agreed with the index only 64% of the time — OCR mangles the label in the older scans, and a failed parse picks up a citation to a different decision. decision_no_source records which was used.

Provenance

Decisions are published at hhs.gov/about/agencies/dab/decisions, which returns 403 to automated clients, so collection routes through the Internet Archive and, for anything published since its last crawl, a browser attached over the DevTools protocol. The decisions are works of the United States government and are not subject to copyright.

Built 2026-09-05 from the corpus in this repository.