Self-Rated Health Isn't Sticky. 47% Went From Poor to Good in a Year.

"Would you say your health is excellent, very good, good, fair, or poor?" is one question, asked one way, on nearly every major health survey. It's also quietly load-bearing: risk-adjustment models use it, cost studies control for it, and a single answer to it often stands in for someone's underlying health status for an entire study year. This question puts that stand-in to a direct test:

"What share of people who reported fair or poor general health in panel year 1 improved to good or better in panel year 2?"

MEPS re-interviews the same people across two panel years, so the same respondent's answer to that exact question can be compared to their own answer roughly a year later. VerbaGPT joined the panel table to itself on person ID (DUPERSID) and panel number, kept only respondents with a valid answer in both years, and restricted to the cohort that said "fair" or "poor" the first time — then checked how many of that same cohort said "good," "very good," or "excellent" the second time, weighting every person by their survey weight (PERWT) so the result reflects the U.S. population, not just the sample.

The answer: 47.3%. Not a fringe minority recovering — almost half of everyone who described their health as fair or poor flipped to describing it as good or better just one panel year later.

That's a bigger number than the "fair/poor" label suggests it should be, and it's worth sitting with why. Self-rated general health is often treated, implicitly, as a proxy for chronic health status — the kind of thing that changes slowly, if at all, over twelve months. This says otherwise: for a lot of respondents, "fair" or "poor" was capturing something closer to a bad stretch — a flu, a recent injury, a rough month — than a fixed condition. A single year's self-rating tags nearly half of the people who use it this way with a label they won't apply to themselves again within the year.

Worth noting in passing: this exact question, worded identically, had been run through VerbaGPT once before and rated helpful — the sharer's own note on this chat says as much ("MEPS - test query"). That earlier run's code came back as a 100%-similarity match in 0.36 seconds and was handed to the model as a working reference before it wrote a single new line of SQL. And yet, in the same response, before that query had executed even once, VerbaGPT's own narration guessed the answer out loud: "approximately 23.4%" — almost exactly half the real figure. The number that actually reached the chat, 47.3%, came from the query result, not the guess, because the two are wired to different things: one is commentary, the other is a live variable read off the database. Having the right code sitting right there as a reference didn't stop the model from narrating a different number anyway — which is exactly why VerbaGPT never lets a model's own account of its result stand in for the result itself.

The rest of the run was unremarkable by comparison: schema resolution against MEPS's cached column embeddings finished in 1.43 seconds on gpt-oss-120b, code generation ran on Gemma 4 31b — both open-weight, both served through Cerebras — and the join itself executed against Snowflake in 1.13 seconds.

None of the mechanics here are specific to health surveys. Any table with a repeated subjective rating and a person identifier — a renewal survey, a support ticket's satisfaction score, an employee engagement check-in — can be asked the same before/after question against your own data, and the same retrieve-generate-execute path runs underneath it.

Data source: Medical Expenditure Panel Survey (MEPS), person-level panel data with self-rated general health (RTHLTH31) and survey weights, 2019–2023 (AHRQ). One of VerbaGPT's built-in sample datasources.

Ask your own question about this dataset → The MEPS datasource is live and open — no signup required for a few questions.