Dartmouth Atlas of Health Care

It is the landmark record of how wildly American medicine varies by place - and the proof that the variation is about practice, not patients. Produced by The Dartmouth Institute through its Dartmouth Atlas Project, the Atlas computes Medicare rates across nine measure families: Medicare Reimbursements, End-of-Life Inpatient Care, Care for Chronically Ill Patients (last two years of life), Primary Care Access and Quality Measures, Post-Discharge Events, Mortality, Medical Discharge Rates, Surgical Discharge Rates, and Hospital and Physician Capacity.

Every family breaks out the same four ways - hospital referral region (about 306 HRRs), hospital service area (roughly 3,400 HSAs), state and county - as one row per geography, year and measure, with adjustment already applied so regions compare fairly. The reach is archival: hundreds of published archives spanning up to 28 years, with reimbursement rates alone carrying claims-based price, age, sex and race adjustment back to 1992.

Inside the Datadory catalog - 1,744 datasets averaging a quality score of 7.81 - this slice scores 8/10, carried by verified field documentation and an observed sample row behind every definition. Get a sample of this dataset and we return rows shaped exactly like the one below, cut to your markets.

One verified row, flat and immediately readable:

# one row = one geography x year x measure (end-of-life cohort)
Geo_code      : 1
Geo_label     : HRR
Geo_name      : Birmingham, AL
Population    : 14341
Year          : 1994
Cohort        : EOL_L6
Eventname     : L6_DIS
Event_label   : Hospital Admissions per Decedent During the Last
                Six Months of Life
Adjusted_Rate : 1.442999978
OE_Ratio      : 1.115621198
Percentile    : 86.229

Read the anatomy. One row settles one geography, one year, one measure - here, Birmingham's hospital referral region in 1994 logging 1.443 adjusted hospitalizations per decedent during the last six months of life. The OE_Ratio of 1.116 says Birmingham ran about twelve percent above what its demographics predict, and the 86th-percentile rank places it near the aggressive tail of the national distribution. Three columns, three framings of the same fact - absolute intensity, relative excess, and standing among peers - which is why models rarely need anything more exotic than this triple.

Eleven verified fields carry every rate row, defined below against the Atlas's own documentation. Cohort and Eventname are the machine spine: the pair identifies which measure family a row belongs to and which rate it reports, so a panel filtered on one cohort stays internally consistent across decades. Event_label carries the plain-English version - the column that makes a deliverable readable without a codebook open. Geo_label keeps the geography type explicit on the row itself, so an HRR never masquerades as a county in a careless join.

Supplemental materials travel alongside: ZIP-to-HSA-to-HRR crosswalks (1995 through 2019), HSA-HRR and hospital-to-HRR/HSA files, ESRI shapefiles of HRR, HSA and PCSA boundaries in NAD83, MedPAR coding-trend tables and hospital research data. Display-only labels fold into the request step rather than cluttering every pull.

Three chips summarize the footprint:

• Geography: the United States at four levels - about 306 hospital referral regions, roughly 3,400 hospital service areas, all states and counties - plus hospital-level detail in the Hospital Research Data section and NAD83 boundary files for mapping.
• Temporal: reimbursement rates run 1992 through 2019 (100% sample, with 2003 through 2010 drawn from the 20% sample plus full 2010); end-of-life and chronic-illness cohorts run 1994/2001 through 2019; mortality 1999 through 2019; discharges 1992 through 2015; primary care access 2003 through 2019; post-discharge events 2004 and 2008 through 2019; capacity snapshots sit at 1996, 2006, 2011 and 2012.
• Granularity: one row per geography times year times measure, each carrying the adjusted rate, observed-to-expected ratio and percentile rank, with denominator population attached and missing rates coded -99999 rather than dropped.

Each family closes its own window - the Atlas is a fixed historical archive, deepest where it matters most for retrospective work. Set against the wider catalog's 7.81 average, the verified documentation and long consistent windows are what hold this slice at 8/10.

API, files, or your warehouse. Daily, weekly, or hourly.

You pick the channel and the cadence; the field dictionary above travels unchanged through all three. Flat files suit overnight warehouse loads and BI tools, structured responses suit dashboards that refresh themselves, and warehouse delivery puts three decades of regional utilization one join away from your property or membership tables. Changing cadence is a settings conversation, not a re-integration project - and a sample cut to your markets comes first either way.

• Small-area demand sizing for senior housing and post-acute. Read where an HRR or HSA sits on chronic-illness burden, discharge intensity and end-of-life hospitalization before committing capital to assisted living, skilled nursing or home-health capacity in that market.
• Care-intensity benchmarking across markets. Adjusted rates and observed-to-expected ratios put a Birmingham beside a Bend on equal footing - utilization differences that survive case-mix, which raw counts never do.
• Physician-behavior evidence for end-of-life planning. Admissions per decedent during the last six months of life, tracked from 1994 forward, quantifies how aggressively each region treats its final months - the baseline for hospice and palliative demand models.
• Longitudinal health-services research. Consistent definitions running up to 28 years, backed by persistent DOIs minted for the Atlas's own research deposits, hold up under peer review and reproduce across studies.
• Competitive intelligence on hospital systems. Medical and surgical discharge rates by referral region show which systems grow volume and which shrink it - positioning context for operators, landlords and their lenders.
• Feature engineering for utilization models. A clean geography-year-measure panel with typed rates, ratios and percentiles joins straight onto property, membership or claims tables keyed to county, state or ZIP crosswalk.

Data scientists get a consistent HRR-level Medicare panel that replicates the classic small-area-variation analyses without vintage-stitching - see data scientists use cases. Market researchers and consultants frame service-area opportunity on benchmarks no vendor recomputes. Journalists and academics cite the series behind a generation of health-services literature. Investors and quants read structural care intensity before underwriting utilization-driven revenue. Competitive-intelligence and developer profiles round out the index at lower relevance - regional norms for competitor positioning, and a stable shape for embedded benchmarks.

Four complements cover what pre-computed regional rates cannot:

• CMS Data Hub - Medicare & Medicaid Datasets - the facility-and-service-line lens on the same program - who treated whom, at what payment. This Atlas supplies the population-adjusted regional rates that lens lacks.
• CMS Medicare Provider Charge Data (Inpatient & Outpatient) - hospital-by-DRG and hospital-by-APC dollars for 2013 through 2024 - price where the Atlas gives utilization.
• CDC open data portal - vital statistics, chronic disease and mortality down to county level - the demographic demand side behind these utilization rates.
• vs HCUP - Healthcare Cost and Utilization Project (via AHRQ) - pre-computed regional Medicare rates against all-payer encounter-level discharges, scored field by field.

Browse the ranking on best health care reits datasets, or start from the health care reits data hub.

Three things, stated up front.

First, each measure family owns its window. Reimbursements reach 28 years; capacity exists only at four snapshot years; post-discharge events skip 2005 through 2007. Tell us the years you need and the sample confirms what exists rather than padding the gaps.

Second, missing is coded, not absent. A -99999 cell means no publishable rate for that geography-year-measure combination. We deliver the code as-is so your pipeline decides how to treat it instead of inheriting ours.

Third, adjustment is already in the rate. Age, sex and race adjustment (plus price adjustment on the reimbursement measures) is baked into Adjusted_Rate, which is exactly why cross-region comparisons hold up - and why mixing these rates with unadjusted counts needs one deliberate step, not a silent join. Name your markets and measures, and the sample returns shaped to them.