US MEPS Survey — Survey Overview 美国 MEPS 调查 — 概览

The Medical Expenditure Panel Survey (MEPS) is run by the Agency for Healthcare Research and Quality (AHRQ) and has collected data every year since 1996. It produces the U.S. government's core estimates of health care use, spending, sources of payment, and insurance coverage for the civilian noninstitutionalized population.

MEPS has two linked parts. The Household Component (HC) interviews a nationally representative sample of households directly — this is the source for everything in this explorer. The Medical Provider Component (MPC) separately contacts the pharmacies, doctors, and hospitals that HC respondents named, and verifies or fills in payment and drug details the household couldn't accurately report themselves — it isn't designed to produce its own national estimates, just to improve the HC data.

Sampled by household, recorded by person. The household is only the entry point: MEPS samples dwelling units to reach people, but the single household respondent then answers questions about every household member individually — this person's age, this person's insurance, this person's doctor visits. Since the facts collected are inherently person-specific, AHRQ builds the Consolidated PUF with one row per person, not one per household — a 4-person household contributes 4 person-records, all from the same interview. Household/family ID variables (DUID, FAMID23, RUSIZE23, etc.) are still on the file, so you can roll person-records back up to the family level if you need to.

Source: AHRQ, 2023 MEPS HC-251 Full-Year Consolidated File documentation, §B 1.0 ("All data for a sampled household are reported by a single household respondent... Data can be analyzed at the person, the family, or the event level," p. B-1). h251doc.pdf →

This explorer covers the 2023 Full-Year Consolidated File (HC-251) — one person per record, 18,640 inscope respondents, 1,374 variables spanning demographics, income, health status, employment, insurance, and health care utilization/expenditure.

Each year AHRQ recruits a new Panel — a fresh cohort of households, drawn as a subsample of households that answered the prior year's National Health Interview Survey (NHIS). For the two panels behind this file, Panel 27 sampled 9,700 households (from the 2021 NHIS) and Panel 28 sampled 9,800 households (from the 2022 NHIS). Every household in a panel is then interviewed 5 times ("Rounds") about 5 months apart, covering roughly 2.5 calendar years total, so each panel's data collection straddles two calendar years.

Because a new panel starts every year while the previous one is still mid-way through its 5 rounds, two panels are always running at once, offset by about a year. Any single calendar year's data — like 2023 — is stitched together from the second half of one panel and the first half of the next. §04 below walks through exactly how those ~19,500 sampled households turn into this file's 18,463 final person-records.

Source: h251doc.pdf §3.1.1, "Panel 27/28 Household Sample Size."

That's the whole trick behind the "R3/1", "R4/2", "R5/3" notation and variable suffixes like REFPRS31: the first digit is Panel 27's round number, the second is Panel 28's round number for that same time window — Round 3 of Panel 27 and Round 1 of Panel 28 happened concurrently, so they're referred to together as "R3/1" (variable suffix 31). Likewise R4/2 → 42, R5/3 → 53. A variable ending in plain 23 (no round pairing) means "status as of December 31, 2023," combining whichever round's data applies for that person.

Source: AHRQ, 2023 MEPS HC-251 Full-Year Consolidated File documentation, §B/C 1.0–2.4 and the "MEPS Panel Design: Data Reference Periods" chart (p. C-2). h251doc.pdf →

Each household has one respondent — usually the person most knowledgeable about the family's health and health care — who answers for everyone in the household across all 5 rounds, via computer-assisted personal interviewing (CAPI). AHRQ distinguishes a few respondent roles that show up repeatedly in the "Survey Administration" variables:

Combining two panels means combining two separate response funnels — households sampled, then whittled down round by round as some don't respond. AHRQ's Table 19 gives the exact 2023 numbers:

4,262 + 5,143 = 9,405 households (technically "Reporting Units," which can be more granular than a physical household) completed every round required of them — an overall unweighted response rate of 26.1% (Panel 27: 24.1%, Panel 28: 27.4%, combined by weighting them 0.40 / 0.60). From RU to person: those 9,405 completing households contained 18,463 people who ended up with a positive final weight (PERWT23F) — averaging just under 2 people per household, which is why the final person count lands close to the household counts above rather than far above them.

Source: h251doc.pdf §3.2, Table 19.

The year-over-year decline below isn't just "a different panel happened to be smaller" — AHRQ documents a specific cause. In its own words: "MEPS has been substantially affected by the pandemic... One effect of the pandemic is the significantly lower response rates... analysts... should continue to exercise caution when interpreting estimates and assessing analyses, especially for data collected from 2020 through 2022. This includes the comparison of such estimates to those of other years and corresponding trend analyses." AHRQ's sample design modifications made to compensate ended in 2022, but response rates themselves haven't recovered to pre-pandemic levels — this is a broader trend across nearly all U.S. government surveys in this period, not unique to MEPS.

Source: h251doc.pdf §3.1.2, "Discussion of Pandemic Effects on Quality of MEPS Data."

This explorer's 18,640 figure is the inscope-respondent count before the final person-weight cut; 18,463 of those have a positive final weight (PERWT23F) and are the set AHRQ says to use for U.S.-population-level estimates. All statistics in this explorer are unweighted — raw respondent counts, not projected national totals. Weighting exists specifically to correct for the panel design's uneven sampling and non-response, so treat any percentage here as "share of respondents," not "share of the U.S. population."

Source: h251doc.pdf §C 2.0.