MEPS data: 493 files of US health spending and monthly coverage, delivered as rows

The query sounds like a checkout decision; the answer is the closest thing American health economics has to a ground-truth instrument. Medical Expenditure Panel Survey (MEPS) is the Agency for Healthcare Research and Quality's continuously fielded survey of households, employers and medical providers, running every year since 1996, and it owns one question nobody else answers at this grain: what gets spent on health care in the United States, and who actually pays.

The published Household Component library spans 493 files across survey years 1996 through 2024: person, family, job, event and condition levels, plus full-year consolidated files, longitudinal panels, NHIS link files and projections aligned to the National Health Expenditure Accounts. It scores a perfect 10 on Datadory's quality rubric against a catalog-wide average of 7.81 across the 1,744 datasets we catalog - one of only two ceiling scores among the sixteen primary records in this slice. Datadory delivers the whole library as typed rows keyed on the person spine; get a sample cut to your years and service categories before anything else.

Person-grain rows after cleaning - one row per person, illustrative values on the real skeleton:

DUID     PID  DUPERSID      PANEL  YEAR  AGE  SEX  RACETHX  MONTHLY COVERAGE      TOTEXP   TOTSLF   PERSON WT
---------------------------------------------------------------------------------------------------------
2320034  001  2320034001       27  2023   34    2        6  PRV x12                 4,870      920     41,203
2320117  002  2320117002       27  2023   67    1        6  MCR x12 + PRV x12      31,450    2,140     12,876
2320149  001  2320149001       28  2023    8    1        1  MCD x12                 1,610        0     58,442

Dictionaries are where MEPS projects quietly die, because the columns are coded, era-specific and load-bearing all at once. The spine is stable: DUID and PID identify the dwelling unit and the person within it, DUPERSID concatenates them into the join key every event file resolves back to, PANEL and DATAYEAR place the row in the rotating design, and AGELAST, SEX and RACETHX carry the edited demographics, with race/ethnicity harmonized to one comparable column across file years.

The signature variables are the coverage flags. Twelve per source per person-year - MCRmm23X for Medicare, MCAIDmmX for Medicaid/SCHIP, TRImm23X for TRICARE/CHAMPVA, PUBmm23X for any public plan, PEGmm23 for employer/union coverage - turning insurance from an annual label into a monthly timeline. Beside them sit the money pairs: TOTEXP and its splits by payer (self/family, Medicare, Medicaid, private, other) at event grain and as annual totals, with TOTSLF isolating out-of-pocket spending.

Then the trap worth naming out loud: the person weight changes name across eras - WTSAQ, WTPER and WTFINL appear in different file years - and any national estimate computed unweighted, or weighted with the wrong vintage, is simply wrong. Datadory ships each weight named, typed and documented per file year, so the lookup happens upstream of you rather than mid-analysis.

Geography - the United States, civilian noninstitutionalized population: national estimates with census-region breakdowns. Nursing-home residents, the incarcerated and active-duty military sit outside the household universe by design, which is precisely when the Human Mortality Database or institution-side records earn their place in the stack.

Granularity - five grains (person, family, job, event, condition) plus consolidated, longitudinal and NHIS-link files, and twelve monthly coverage flags per person per year. Event files split by service category - dental, hospital inpatient, emergency room, office-based, outpatient, prescribed medicines, other medical expenses - so utilization studies should pick the family that matches the question rather than pulling blind.

Two scope notes belong in every plan. Employer-side microdata stay confidential and publish as summary tables at national, regional, state and metro levels, so firm-level premium questions answer at the aggregate the survey releases. And state-level detail on the household side lives in those same summary tables, not in the person files - a structural fact of the survey design that no amount of clever joining works around.

Name the years, service categories and grains when you request the sample and the extract arrives cut to that scope; the standing feed follows the same shape, so anything prototyped survives delivery intact. See how the slice behaves at machine speed in the life & health insurance data APIs, and how citation-grade work stays auditable in citation-grade research.

Four workflows pay for themselves fastest.

Coverage churn and continuity. Twelve monthly flags per person-year turn insurance history into transitions: who moved from employer coverage into Medicaid, how long uninsured spells ran, what continuity looks like by income band. An annual snapshot cannot see this structure at all, and most substitute feeds offer exactly that.

Pricing and utilization benchmarks. Spend distributions by event type and condition give health-plan actuaries an empirical backbone for plan-design and morbidity assumptions, and give investors covering managed care a national baseline to test a portfolio company's reported trends against.

Market sizing that survives diligence. Expenditure totals reconcile to the National Health Expenditure Accounts, so sizing work for payer-, provider- and pharmacy-adjacent markets carries a public denominator; the workflow continues in market-sizing.

MEPS prices care; four sibling records fill the edges it leaves open.

Stacked deliberately, they cover counts, context, mortality, carrier financials and the classroom - the standard configuration for this slice in best life & health insurance datasets.

Ranked by how directly the files answer the day job:

1. Data scientists & modelers train expenditure and utilization models back through 1996 on survey-weighted targets no synthetic panel reproduces; the workflow ranks the slice in life & health insurance data for data scientists.
2. Investors & quants covering managed care benchmark utilization frequency and payment mix against the national baseline before believing a portfolio company's trend line; see investors & quants.
3. Market researchers & consultants size payer, provider and pharmacy-adjacent markets off denominators that reconcile to the National Health Expenditure Accounts; see market researchers.
4. Journalists, academics & students cite national estimates whose methodology, response rates and variance guidance are published end to end; see journalists, academics & students.
5. Actuarial and pricing teams at health plans read event-type spend distributions as the empirical backbone for plan-design pricing and morbidity assumptions.

Because the hard part was never the first extract - it is the tenth. Weight columns that rename themselves across eras. Enum definitions that shift between file years. Two panels hiding inside one calendar year. Event families that join to people only through one concatenated key. Each is survivable once; none is fun to re-solve in every new notebook.

Datadory normalizes before delivery: enums resolved against their file-year definitions, weights shipped named and typed per vintage, panel overlap flagged so deduplication is a column rather than a discovery, and event extracts pre-keyed to the person spine. Suppressed or lagging vintages surface explicitly rather than being found mid-model. When the next survey year lands, it arrives as new rows under the same dictionary - no re-integration project.

Name the years, service categories and grain when you request a sample and it arrives cut to that scope; the production feed follows the same shape, so anything prototyped on the sample survives delivery intact.