Where this data disagrees with the literature

Every diagnosis-based outcome loses half or more of its raw signal once you control for who gets checked and who is on Medicare Advantage. The one outcome that is not itself a diagnosis rate — case mortality among the already-diagnosed — loses only a fifth.

The analysis states the limit on this itself: “Caveat I must state plainly: the death-rate bucket is n=7 — too small to trust the ratio on its own, but the direction is the same one the checkup/MA controls were built to test, and it's a >2x gap.”

Concretely, diabetes and hypertension county prevalence: raw +0.378 / +0.358, strict -0.011 / -0.011 (n=2,347, controls: pct_85plus, median_household_income, primary_care_physicians_per_100k, region_south, population_density_per_sqmi, pct_black, norc_ma_share_of_bene_2020, pct_checkup_annual). The registry is largely measuring who gets diagnosed, exactly as hypothesized — this is the strongest, most reproducible finding in the set.

The county and state prevalence-and-incidence signs are all one mechanism (diagnosis-intensity contamination, Finding 1). The state death-rate flip is a different mechanism (the G30 coding-share control, matching the earlier finding that state Alzheimer's death rates are a coding artefact). The country death-rate negative is a third, independent phenomenon (Finding 3) — it is not induced by any control at all. “4 signs” is true as a count but implies one story; there are three.

The ranking

Ranked by the literature weight it opposes: hypertension ≈ diabetes >> air pollution — even though air pollution has by far the largest raw and strict coefficients.

The analysis names this as the one cell it would lead with: “a country-level ecological correlation that robustly (outlier-tested) says higher hypertension and diabetes prevalence goes with lower age/GDP/life-expectancy-adjusted dementia death rates, standing against 113-138 individual-level studies including RCT meta-analyses.”

And it immediately names the competing explanation, which it cannot rule out: “The most likely alternative explanation, per the country level's own documented caveat, is that dementia_deaths_per_100k_age_std is crude, not age-standardized (its own metadata admits this) and death-certification completeness tracks GDP — so this is very plausibly ‘richer countries certify Alzheimer's deaths better and have higher metabolic-disease prevalence,’ a competing-cause/ certification artefact riding on the same GDP axis, not a real biological reversal. I cannot rule that out with what's in this data; the strict control on GDP + life expectancy narrows but does not eliminate it.”

The analysis flags this result as contradicting its own premise: the diagnosis-intensity variables (pct_checkup_annual, norc_ma_share_of_bene_2020) were the a priori suspicion. “They matter, but region (South) and race composition (pct_black) are named more often as the single control whose removal alone restores the original sign.”

Concretely, for diabetes and hypertension county prevalence (CX-26, CX-28), removing median_household_income, region_south or pct_black individually each independently restores the raw positive sign — but no single addition to the primary set flips it alone. The flip is a joint effect of the whole strict bundle, not one control's doing. That names the two candidate confounds as load-bearing, and it also means there is no clean single culprit to control for and then stop worrying about the rest.

KT-2 — GDP per capita against the 2000-2021 change in country dementia death rates

n=142. Pearson r=0.067 (p=0.43, null) but Spearman rho=0.559 (p=5e-13, strong). Trimming the extreme 5% of GDP (Monaco $288k, Luxembourg $147k, Ireland $132k at one end; Burundi/Afghanistan/CAR under $600 at the other) only partially recovers it (r rises to 0.156, still far short of 0.559). This is not an outlier artefact — GDP per capita is intrinsically right-skewed, and the relationship is monotonic but log-shaped, which a linear Pearson genuinely cannot see. A rank statistic is the only reason this relationship is visible at all in this dataset.

LG-1 — allele frequency against effect size, genetics

log10_allele_frequency vs abs_log_or, n=90. Pearson -0.092 (null) vs Spearman -0.426 (p=3e-5, strong). Trimming the tails makes it worse (-0.023), so this is not tail-driven either — it is the well-known GWAS pattern where a few rare variants (TREM2, log10 freq ≈ -2.7) carry large effect sizes while many common variants (CD33, ADAM10) carry small ones. That is a real, biologically expected rank relationship that a straight line across allele frequencies was never going to capture.

A sixth case, below the flag threshold

Not one of the official 5, but real: hypertension at country level (KP-9) is significant by Pearson (p=7e-6) but not by Spearman (p=0.097) — a gap of 0.20, under the 0.25 flag threshold but still a case where the two tests disagree on whether there is anything there at all. Worth a footnote, not a headline.

1. Country-level hypertension and diabetes prevalence are negatively associated with (crude) dementia death rates after age/GDP/life-expectancy controls (r=-0.36 and -0.23, n=189), robust to dropping the 5 most influential countries, standing against 113-138 individual-level studies including RCT meta-analyses. Test: partial correlation + leave-one-country-out. Most likely alternative explanation: the outcome variable is a crude death rate, not age-standardized (confirmed from its own metadata) — richer/higher-life-expectancy countries both certify Alzheimer's deaths more completely and have higher metabolic-disease prevalence, so GDP-linked certification quality is a live confound that 3 country-level controls narrow but can't rule out.
2. No single control drives the county/state sign flips — it is the combination. Test: leave-one-out addition/removal on the exact nested control lists in the report. Region-South and race composition are named more often (7 and 5-6 times respectively) than the diagnosis-intensity variables expected going in. Alternative explanation: this could reflect genuine regional health-system differences rather than confounding — the data cannot distinguish “South is a confound” from “South is itself a real, unmeasured causal pathway.”
3. Diagnosis-based county outcomes lose 50-68% of their raw signal under strict controls; the one non-diagnosis outcome loses 21%. Test: relative decay grouped by outcome type. Alternative explanation and caveat: the death-rate comparison group is n=7, too small to be a stable ratio on its own — the direction is consistent with everything else in this dataset (including the earlier finding that counties with more 65+/85+ residents have lower diagnosed prevalence), but the analysis would not publish the specific 21% number without a larger death-rate n.

Every coefficient on this page is a correlation between places: counties, states or countries. The contradictions ranked above are contradictions between a place-level number and a body of published research on individuals. They are not evidence that high blood pressure or diabetes protects anyone from dementia, and nothing here should change what anyone does about their own health. The analysis's own reading is that the country-level result is very plausibly a death-certification artefact riding on national income, and that it cannot be told apart from a real effect with the data available.

Two further things limit what any of these numbers can mean. The county and state outcome is diagnosed dementia in Medicare claims, which is why diagnosis intensity dominates Finding 1: a place that diagnoses well and a place with more disease produce the same number. And the country outcome column is a crude rate despite its name, so country comparisons carry age structure inside them.

The coefficients themselves and how they were classified are on What held up, and what did not. If you want evidence about a person rather than a place, the individual-level study evidence is on What the research found, and the risk factors with the best individual-level support are ranked on Risk factors, ranked.

Provenance

Analysis

What's analytically interesting in the alzhub correlation data

Written

2026-09-28

Coefficients re-checked

233

Source file

data/deep_dive/deep_dive_findings_20260928.md (15 KB)

The figures on this page are transcribed from that file and from nowhere else; the “could not determine” list is read out of it when the page loads. Nothing here was re-rounded or reinterpreted.