CMS asks every discharged patient about the same twenty minutes five different ways. Four of the questions can't see anything. One of them can. Here are all 2,578 hospitals.
Today's newsletter is about retrieval — that the annoyances you noticed in a hallway years ago were the raw material, and the expensive habit you lost was writing them down. This is the version of that argument you can check.
Because somebody did write it down. Since 2006 CMS has been mailing a survey to a random sample of every hospital's discharged patients, and five of its questions are about the same twenty minutes at the end of the stay. They are not equally sharp. Read them in order and watch the resolution climb:
Q19 and Q20 ask whether a thing happened. Q15 asks whether it landed. That is the whole difference, and it is worth about forty percentage points.
Every hospital in the country, five times over, on one 60–100% axis. Each mark is a hospital. Click a rung to put it on the x-axis of the scatter below.
That's the part nobody checks. Below, each hospital is a dot: the question you picked on the x-axis, and its 30-day heart-failure excess readmission ratio on the y — CMS's own risk-adjusted outcome, where 1.00 is exactly as expected and above the line means worse. If the sharper question sees more, its cloud should tilt harder.
Drag Min HF discharges to the right and watch what happens to Q15 — side effects. It doesn't decay cleanly. It wanders: −0.096 at ≥100 discharges, −0.071 at ≥300, −0.034 at ≥700 — and then back up to −0.101 at ≥900, on 149 hospitals. Split into non-overlapping bands it is worse: exactly 0.00 below 100 discharges, −0.14 in the 100–299 band, −0.03 above 700. A coefficient that will not settle as you improve the measurement isn't a weak relationship. It's the absence of one, plus noise: a hospital with 74 heart-failure discharges has an excess readmission ratio built from a handful of events.
Now do the same thing to the Discharge Information composite. It holds — between −0.19 and −0.27 across every threshold, never crossing zero, never collapsing.
The blunt question predicts the outcome better than the sharp one. “Did someone talk to you about help at home?” carries r = −0.22 against readmission; “did they always explain the side effects?” carries −0.09. That is the opposite of what the specificity argument predicts, and it is the honest finding.
Both are weak — 4.9% and 0.8% of the variance. Neither is a lever. What the blunt question is probably measuring is not discharge teaching at all but the kind of hospital that does everything a little better, and that hospital readmits fewer people for a hundred reasons at once.
So the specific thing you noticed in the hallway is worth writing down — it is where the variance is, and the coarse question is at the ceiling and can't see it. But noticing and mattering are two claims. This chart is the second one, and it doesn't go your way. Write it down anyway. Then check it.
The point of the exercise isn't the r. It's that someone wrote the question down twenty years ago, at a resolution fine enough that the answer is still interesting today — and that the four coarser versions of the same question, the ones that made it onto the star rating, tell you almost nothing because everyone is already at 85%.