What held up, and what did not

• 38 survive Significant after false-discovery-rate correction, and the partial correlation keeps at least half the raw strength.
• 3 survive with the sign flipped Statistically robust, but pointing the opposite way to what the individual-level literature expects. Read as measurement, not risk.
• 81 collapsed or artefact The association falls below half its raw strength once the confounder is controlled, or it is a product of coding practice or a shared denominator.
• 75 null or negligible No association, or one too small to carry meaning.
• 9 uninterpretable A number was produced, but the two columns cannot be read against each other.
• 12 twins Predictor pairs that are effectively the same column (|r| at or above 0.95), listed so they are not read as two findings.
• 39 checks and documented artefacts Run to test the data itself rather than to answer a question about dementia.

The method

Pairwise-complete Pearson r with Fisher 95% CI and Student-t p, Spearman rho, and partial correlations by OLS residualisation on the named confounder(s) (scripts/pystats.py, identical maths to the audited baseline). Benjamini-Hochberg FDR is applied within each level and used as the raw significance gate. A result 'survives' if BH q<0.05 and the partial keeps at least half the raw |r| with p<0.05; it 'collapses' if the partial drops below half the raw |r|. Blanks are never zero-filled, imputed or interpolated; suppressed NORC cells stay blank. Twin predictors (|r|>=0.95) are listed per frame. Every variable carries unit, year, source and a measured/modelled tag. Every place-level result carries the ecological-fallacy caveat; every claims-based outcome carries the under-diagnosis caveat.

What the earlier baseline already established

The 43-result baseline (scripts/run_correlations.py) tests one exposure against one outcome per card on three cross-sections (51 states, ~3,100 counties, ~190 countries) plus a BRFSS join, with the same pairwise-complete Pearson/Spearman/partial machinery. It already shows: (1) two raw counts correlate through population and collapse when population is controlled (S11: r=0.944 to -0.046); (2) the state education gradient in Alzheimer's deaths halves under age+income control (S4: -0.58 to -0.26) and rurality does the same (S9: 0.54 to 0.27); (3) country health spending vs dementia DALYs collapses under GDP (I3: 0.76 to 0.28); (4) two sign flips against the literature (state hearing loss S2, country diabetes I2, plus country air pollution I5) that it correctly reads as measurement rather than risk; (5) at county level the six NORC-based cards find only weak associations, the largest being pct_black vs diagnosed prevalence (D1: r=0.30, partial 0.27 on income). What it does not do: it uses the unadjusted state death rate, the modelled Alzheimer's Association prevalence, the all-age NORC denominator and a country death rate whose own metadata says it is crude despite its column name; it never tests diagnosis intensity, coding practice, race-stratified rates, the exposure time series, the study library, genetics or drug data; and it applies no multiple-comparison control.

Provenance

Analysis

alzhub.org definitive Alzheimer's correlation analysis: county, state, country and study level

Generated

2026-09-28

Generated by

analysis/run_analysis.py (Claude Fable 5.1 session, 2026-09-28)

Baseline

data/correlation_results.json (unchanged), 43 results

Source file

data/correlation_report.json

State 'Alzheimer's' death rates are mostly a coding-practice league table

The fraction of a state's dementia deaths that certifiers code as G30 (Alzheimer's) rather than F03 (unspecified dementia) explains most of the between-state variation in the age-adjusted Alzheimer's death rate, and predicts the 2010-2019 trend. The baseline's state education and physician-supply gradients (S4, B5) fall by more than half once income and the G30 share are controlled.

Coefficients: SK-1, SK-2, SK-3, SM-1, SM-8, ST-2

There is no age-standardised country outcome, and the GDP-dementia correlation is age structure plus certification

The timeseries column labelled age-standardised is numerically the crude rate. Age structure alone correlates 0.72 with the country death rate. Health spending, physician density, hospital beds, urbanisation, literacy, UHC coverage and immunisation all collapse under age structure (and GDP). GDP itself survives a linear pct_65plus control, which cannot be read as anything more than 'rich countries are older AND certify dementia deaths': no file here can separate those.

Coefficients: KA-1, KA-2, KA-3, KT-3, KP-1, KP-2, KP-3, KP-17

County diagnosed prevalence tracks detection, not just disease

Counties with a higher Medicare Advantage share and higher per-beneficiary spending have higher diagnosed prevalence; counties with high prevalence have a LOWER share of 'highly likely' cases, i.e. the excess sits in the softer tiers. Counties with older populations (more 65+ and 85+) have LOWER diagnosed prevalence among the 65+, the opposite of biology. The all-age rate the baseline used is also diluted by under-65 disability beneficiaries.

Coefficients: CD-3, CD-4, CD-5, CD-1, CX-18, CX-17, CA-1

The county race gradient is a place effect, not (only) composition, and it collapses under place controls

The baseline's D1 (pct_black ~ prevalence, partial 0.27 on income) is real but is not about Black beneficiaries: White beneficiaries in high-pct_black counties also have higher diagnosed prevalence (CC-1), while prevalence among Black beneficiaries does not rise with pct_black at all (CC-2). Adding region (South), density and coding-intensity markers takes the pct_black partial to 0.12. The candidate third thing is Southern place: stroke-belt vascular risk and Southern coding practice, which these files cannot separate.

Coefficients: CX-11, CC-1, CC-2, CC-4, CC-5, CH-1, CP-1

Modelled PLACES risk factors collapse; PM2.5 is the one environmental exposure that survives, with caveats

County diabetes, hypertension, obesity, smoking and fair/poor health (all PLACES small-area MODEL outputs that share ACS inputs with the controls) collapse under age+income+physician+region. Physical inactivity keeps a partial of about 0.23. Modelled 2016-2019 PM2.5 keeps about 0.28 under the strict set, but the MEASURED monitor-based PM2.5 (n=559, urban-selected counties) falls to 0.05 under the same controls, and at state level PM2.5 collapses to 0.10 once race and claims intensity are controlled. The modelled PM2.5 survivor is therefore not confirmed by the measured one.

Coefficients: CX-26, CX-28, CX-30, CX-35, CX-36, CX-37, CH-6, CI-6, CM-1, SP-18, KP-15

Raw counts correlate through population, at every level

Five count-versus-count pairs with raw r between 0.88 and 0.97 fall to 0.14-0.44 once population is controlled. None is a relationship.

Coefficients: CA-2, CA-3, SA-1, SA-2, SA-3

Measured and modelled versions of the same outcome disagree

State diagnosed prevalence (claims) and the Alzheimer's Association modelled prevalence agree at r=0.65; neither agrees with the age-adjusted death rate (r=-0.03 and -0.39) nor with self-reported cognitive decline. The AA estimate correlates 0.77 with pct_race_black because race is an input to that model (baseline B2/B4/S2/S3/S10/S12/A3 all used it as the outcome).

Coefficients: SK-5, SK-6, SK-7, SK-8, SU-1, SU-2, SU-3

Almost every coefficient on this page is a correlation between places: counties, states or countries. It says that places with more of one thing also tend to have more of another. It does not say that a person with the exposure is more likely to get dementia. Treating a place-level correlation as a personal risk is the ecological fallacy, and it can invert the answer entirely — one of the artefacts above shows exactly that, with a county race gradient that holds for White beneficiaries and vanishes for Black beneficiaries in the same counties.

Two further things limit what any of these numbers can mean. The county and state outcome is diagnosed dementia in Medicare claims, so a place that diagnoses well and a place with more disease produce the same number. And the country outcome columns are crude rates despite their names, so country comparisons carry age structure inside them.

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.