Dementia Prevention and Modifiable Risk Factors

Nucleus of the dementia cluster — the domain-opener. Dementia is a top healthspan axis (57 million people worldwide in 2019, projected 153 million by 2050 (Livingston et al., 2024)); this page holds the canonical map of which levers exist and when in life they act. The 2024 Lancet Commission is a consensus triangulation (systematic reviews + new meta-analyses + Mendelian randomization where possible), not a single trial — grade its claims accordingly.

The headline, and how to read it honestly

The Commission estimates that the 14 modifiable risk factors account for a population attributable fraction (PAF) of 45.3% — «nearly half of all cases of dementia being associated with 14 potentially modifiable risk factors». (Livingston et al., 2024)

This is a modelled, population-attributable figure — the same discipline the wiki applies to GBD [inferred from @livingston2024]. Three caveats bind how it may be used:

PAF is not a per-person effect. It is what theoretically would not occur if the whole population’s exposure were eliminated — «this risk modification affects the population but does not guarantee any individual will avoid dementia». (Livingston et al., 2024) Removing a factor does not remove that share of one person’s risk.
The relative risks are borrowed from observational data and assumed causal. The Commission «assumed risk factors cause dementia»; for several factors (hearing, depression, social isolation, obesity in late life) reverse causation over the long prodrome is explicitly a live alternative -> The U-Shaped Association Artifact. (Livingston et al., 2024)
The factors overlap, so the individual PAFs do NOT sum to 45.3% — they are weighted down for communality. A principal-components analysis of the 14 factors «found five principal components, explaining 54 % of the total variance», so each factor’s raw PAF is discounted for its shared variance with the others. (Livingston et al., 2024) The per-factor weighted PAFs live in the report’s Table 1 — a figure NOT recoverable from the held accepted-manuscript text, so this is a named gap pending the published-version PDF (the source slug itself is already held; an token here would false-fire reconcile awaits-arrived, which keys on the slug, not the missing table — re-key only if a distinct published-version row is ever registered).

The 14 factors, their direction, and their life-course window

The two 2024 additions to the 2020 list of 12 are high LDL-C and uncorrected vision impairment («adding vision impairment and high cholesterol as potentially modifiable risk factors»). The 2020 list of 12 carried a PAF of 40%. (Livingston et al., 2024) Relative risks below are the Commission’s cited/new-meta-analysis figures; each is for dementia incidence.

Factor	Effect (Commission’s cited RR/HR)	Life-course window
Less education	educational attainment protective; adulthood cognitive stimulation at work HR 0.77 (de-secondhanded below) -> Cognitive Stimulation at Work and Dementia	early + mid life
Hearing loss	HR 1.35 (1.26-1.45), 50-cohort MA (Yu 2024, first-hand; de-secondhands the Commission’s borrowed 1.37, 1.00-1.87). But the treatment lever is weaker: the ACHIEVE RCT was null overall on 3-yr cognition (0.002 SD, p=0.96), positive only in a pre-specified high-baseline-risk subgroup — the large exposure PAF does not certify that fitting aids prevents dementia -> Hearing Loss and Dementia	midlife
High LDL-C (new 2024)	+8% per 1 mmol/L (1.08, 1.03-1.14); >3 mmol/L HR 1.33	midlife
Depression	RR 2.25 (1.69-2.98), new MA (partly reverse-causal)	midlife
Traumatic brain injury	RR 1.66 (1.42-1.93)	any age
Physical inactivity	activity RR 0.80 (0.77-0.84) protective (Iso-Markku 2022, first-hand; de-secondhanded below) -> Physical Activity Dose and Mortality	midlife+
Diabetes	HR 1.24 per 5-yr earlier onset; midlife-specific	midlife
Smoking	midlife RR 1.30 (1.18-1.45); no excess in former smokers	midlife
Hypertension	untreated HR 1.42 (1.15-1.76) observational; treatment lever RCT OR 0.87 (0.75-0.99) (Peters 2022 IPD, first-hand; interventional de-secondhanding below) -> Blood Pressure Lowering and Cardiovascular Events	midlife
Obesity	midlife RR 1.31 (1.02-1.68)	midlife
Excessive alcohol	>21 units/wk HR 1.22 (1.01-1.48); first-hand dementia dose-response (J-curve, protective arm unadjudicated, harm arm real) now held -> Alcohol and Mortality and Vascular Disease	midlife
Social isolation	RR 1.57 (1.32-1.85) / 1.18 (1.08-1.30)	late life
Air pollution	PM2.5 HR 1.03 (1.02-1.05) per 1 ug/m3 (de-secondhanded below)	late life (lifelong)
Uncorrected vision (new 2024)	RR 1.47 (1.36-1.60)	late life

(Livingston et al., 2024)

The evidence-STRENGTH axis, orthogonal to PAF — Zhang’s credibility tiering (F/C-advance, 2026-09-08)

The Commission’s list answers how much dementia each factor accounts for at the population level (PAF magnitude). It does not answer how believable each association is once study design, heterogeneity, and bias are tested. An umbrella review supplies that missing axis: it grades the strength and validity of the evidence per factor. «We conducted an umbrella review to evaluate the strength and validity of the existing evidence»; the motivating gap is that «hierarchies of evidence have not been determined across various factors» (Zhang et al., 2022).

Scale: 243 unique factor-outcome associations from 70 SR/MAs (64 observational, 5 RCT, 1 both) + 48 MR studies. Each association is graded on pooled effect + p-value + case count + I2 + 95% prediction interval

small-study (Egger) + excess-significance bias into a class I-IV scheme — convincing (p<10^-6, cases>1000, I2<50%, PI excludes null, no bias) · highly suggestive · suggestive (p<10^-3) · weak (p<0.05) · non-significant. Counts across the 243: 12 convincing · 12 highly suggestive · 27 suggestive · 87 weak · 105 non-significant — «Twelve associations were graded as convincing evidence involving education, BMI, depression, gait speed, blood glucose level, atrial fibrillation and benzodiazepine use» (Zhang et al., 2022). Grades are per outcome — a factor can be convincing for one outcome, weak for another.

The two axes are different quantities — a distinction, NOT a tension

The parameter table forces the point: Zhang’s class and Livingston’s PAF measure different things, so the umbrella review neither confirms nor contradicts the PAF headline — it adds a perpendicular reading.

Parameter	Zhang 2022 (VoR)	Livingston 2024	Same quantity?
What the number grades	evidence-credibility class (I-IV), from p + cases + I2 + PI + bias tests	population attributable fraction (% of dementia preventable)	NO — believability of the association vs its population-level magnitude
Physical activity	suggestive (class III)	in the 14; inactivity RR 0.80 (0.77-0.84), part of the 45% PAF	NO — a credibility tier vs a PAF contribution
Diet (fruit & veg)	suggestive (class III)	not itemized in the 14 (diet is not a Commission factor)	NO — different construct entirely
Depression	convincing (class I), RR 1.85 (1.67-2.04) for ACD	in the 14, RR 2.25 (1.69-2.98)	NO — Zhang certifies believability; Livingston sizes impact

(Livingston et al., 2024; Zhang et al., 2022). The per-factor point estimates for the suggestive lifestyle levers (physical activity, fruit & veg, wine) live in the source’s figure panels (Figs 2-4), which the VoR’s OCR text does not recover — so only their credibility GRADE is cited here, not a bare RR the VoR text cannot confirm.

What the strength axis tempers, and what it raises

Tempers several Livingston-14 lifestyle levers. Physical activity and diet (fruit & vegetables) land only at suggestive (class III), not convincing — the observational protection is real but not bias-proof by the umbrella’s bar. Wine reaches only suggestive as well. And the interventional leg is thin: only 9 RCT-derived associations, «the protective effects of physical activity and unsaturated fat supplementation are not significant in meta-analyses of RCTs» (Zhang et al., 2022) — converging with the FINGER/supplement-null story this page already holds (see The RCT leg section below).

Raises new high-grade factors NOT on the Commission’s 14. «There were new modifiable factors firstly raised in our study that reached a high level of evidence like low gait speed and hypoglycemia» (Zhang et al., 2022); also convincing/highly-suggestive but off-list: atrial fibrillation, benzodiazepine use, homocysteine. Gait speed and hypoglycemia are largely prodromal/reverse-causal markers, not obvious lifestyle levers — a credibility grade is not a causal or actionable warrant.

The obesity paradox, life-course-split. Midlife obesity is graded a risk (suggestive); late-life overweight (BMI 23-30) is convincingly protective, RR 0.81 (0.75-0.88) — «late-life high BMI is convincingly protective of dementia and cognitive impairment» (Zhang et al., 2022) — read through leptin/CSF mechanisms and, more plausibly, prodromal reverse causation (late-life weight loss as an early-disease marker). This sharpens the midlife-vs-late-life window the 14-factor table already carries for obesity.

A causal-verification (MR) layer. 48 MR studies / 157 causal analyses give a genetic natural-experiment check on top of the credibility grade. Two factors clear BOTH the top tier AND MR: «Our findings from MR studies verified causal relationships of convincing factors education and plasma glucose with Alzheimer’s disease» (Zhang et al., 2022). MR also flags periodontitis and alcohol as raising AD risk — corroborating (not independently, see below) the periodontal candidate lever this page holds (see Periodontal disease section below).

The whole tiering rests on low-quality underlying reviews — the umbrella’s own stated limitation (AMSTAR-2: only 39% high, 31% low, 6% critically low). The credibility axis grades the evidence, not the world; a convincing class means the association survives bias tests, not that acting on the factor prevents dementia.

Not an independent witness (NOT-E). Zhang is the Qingdao/Fudan group (Yu Jin-Tai corresponding), shares authors with held dementia sources (Xu, Wei is on both this and the omega-3 source), and cites Livingston 2017/2020 while resting partly on the same observational literature. So this is a type-F credibility refinement of Livingston plus a type-C new axis — never [E-independent] against Livingston or the group’s own siblings.

The less education / cognitive stimulation at work cell above carries the Commission’s borrowed cognitive-reserve figure. Its primary source is now held first-hand: Kivimaki 2021, an IPD multicohort study (7 IPD-Work cohorts, 107,896 dementia-free adults, 1143 cases, mean follow-up 16.7 y). Gill Livingston co-authored that paper, so this is the Commission’s own upstream evidence, not an independent second witness — a de-secondhanding (type-F), not an [E-independent] corroboration.

The effect, first-hand: high vs low cognitive stimulation at work, «age and sex adjusted hazard ratio 0.77, 95% confidence interval 0.65 to 0.92, heterogeneity in cohort specific estimates I2=0%» (Kivimäki et al., 2021); fully adjusted 0.82 (0.68-0.98). The secondhand 0.79 the Commission propagates is this study’s reverse-causation-lag estimate (dementia arising >=10 y after exposure assessment, HR 0.79, 0.66-0.95) — a robustness figure, not the headline.
What the first-hand source adds that the secondhand cell cannot show: (i) the association survives a 10-year lag exclusion, so it is not merely prodromal dementia lowering engagement -> The U-Shaped Association Artifact; (ii) it is an adulthood/mid-life lever, distinct from the early-life education factor it is bundled with here; and (iii) residual confounding by childhood IQ is the unremovable alternative the observational design cannot exclude. Full estimate, mechanism arm, and Layer-1 ranking (RR ~1.3, below the cardiometabolic big rocks and below education itself) on Cognitive Stimulation at Work and Dementia. (inferred from Kivimäki et al., 2021; Livingston et al., 2024)

The air pollution cell above carries the Commission’s borrowed PM2.5 figure — HR 1.03 (1.02-1.05) per 1 ug/m3, a confidence interval that excludes 1, reading as an established risk factor. Its primary evidence is now held first-hand: Wilker 2023, a gold BMJ SR+MA («2080 records identified 51 studies for inclusion»; 14 meta-analysable for PM2.5). Elissa Wilker (Harvard Chan) does not overlap the Commission author list, but Wilker cites Livingston 2020 and both rest on the same underlying observational cohort literature — so this is a de-secondhanding (type-F), not an [E-independent] second witness.

The effect, first-hand, is BORDERLINE — the overall CI crosses 1. «The overall hazard ratio per 2 μg/ m3 PM2.5 was 1.04 (95% confidence interval 0.99 to 1.09). The hazard ratio among seven studies that used active case ascertainment was 1.42 (1.00 to 2.02) and among seven studies that used passive case ascertainment was 1.03 (0.98 to 1.07).» (Wilker et al., 2023) Rescaled to the Commission’s per-1-ug increment (multiplicative, justified by the linearity finding below), Wilker’s pooled HR is ~1.02 (0.99-1.04) per 1 ug/m3 (inferred from Wilker et al., 2023) — the same point estimate as the borrowed 1.03, but with a confidence interval that crosses the null. The secondhand cell reads significant; the comprehensive first-hand pool does not. This is the beyond-summary move: the primary evidence is weaker than the propagated figure implies.
The significant signal is an ascertainment-METHOD moderator, NOT a person-level stratum. The only sub-pool clearing significance is active case ascertainment (systematically screening every participant) at 1.42 (1.00-2.02); passive ascertainment (registry/records) is null at 1.03 (0.98-1.07). This moderates how dementia was detected, not who is exposed, so it does not license a route-(b) effect-modification claim -> Baseline Risk and the Relative-Absolute Split. Wilker’s own «best estimate» leans on it — «the best estimate for the effect of a 2 μg/m3 higher concentration of PM2.5 is a hazard ratio of 1.42 (95% confidence interval 1.00 to 2.02) based on the studies that used active case ascertainment. However, given concerns of time trend bias and causally relevant time windows, a more conservative estimate is 1.17 (0.96 to 1.43) after removing four studies for these reasons.» (Wilker et al., 2023) Both the headline 1.04 and the conservative 1.17 include 1, and Wilker concedes «the confidence limits are likely too wide given the number and characteristics of the included studies».
Risk of bias runs toward the null, so borderline may UNDERstate. «Most studies were at high risk of bias, although in many cases bias was towards the null» (Wilker et al., 2023) (ROBINS-E on observational cohort pools). So the null-crossing CI is not a no-effect verdict — a true small positive effect is consistent with the data, and the direction of bias means the estimate is a floor more than a point. Read as insufficient-evidence-tilting-positive, not “no effect.”
Dose-response is essentially LINEAR over the studied range — no knee to exploit. «One other study that explored a possible non-linear dose response association found essentially a linear relation with exposure from 3 μg/m3 to 16 μg/ m3»; two studies suggested a levelling-off «but the concentration at which the levelling started was often where data were more sparse» (~8.5 and 35 ug/m3) (Wilker et al., 2023). This is the dose-response prior working: an apparent plateau sitting at the sparse edge of the data is edge-of-evidence, not a curve feature -> every reduction pays. Sub-standard exposures still carry the association: below the EPA 12 ug/m3 annual standard (n=8) the HR was «also 1.04 (0.97 to 1.11)».
Other pollutants: NO2 and NOx directional, O3 null. «Evidence suggested an association with NO2 (per 10 μg/m3 hazard ratio 1.02 (0.98 to 1.06)) and NOx (1.05 (0.98 to 1.13)), with all studies but one of each showing small but elevated hazard ratio» (Wilker et al., 2023); ozone showed no association (~1.00). So the airborne-pollutant signal is PM2.5-and-nitrogen-oxide-led, each individually borderline.

Parameter table — the borrowed 1.03 vs the first-hand pool [inferred from @wilker2023; @livingston2024]. The Commission’s cell and Wilker’s pool are not the same quantity — different increment, different pooled set, different ascertainment mix — so the comparison bounds and re-grades the cell, it is not a tension.

Parameter	Commission (borrowed)	Wilker 2023 (first-hand)	Same quantity?
Increment	per 1 ug/m3 PM2.5	per 2 ug/m3 PM2.5	NO
PM2.5 estimate	HR 1.03 (1.02-1.05), excludes 1	HR 1.04 (0.99-1.09), crosses 1	NO — per-1 rescale ~1.02 (0.99-1.04)
Statistical significance	significant	borderline / null overall	NO
Pooled set	borrowed prior estimate	14 studies, 2023 systematic pool	NO
Ascertainment	unspecified	active 1.42 (1.00-2.02) vs passive 1.03 (0.98-1.07)	NO

Layer-1: a weakly-modifiable STRUCTURAL/POLICY lever — the de-secondhanding does NOT reorder the personal rocks [inferred from @wilker2023]. The effect is «smaller than those reported for other risk factors for dementia (eg, education and smoking)» (Wilker et al., 2023), and ambient air pollution is only weakly modifiable by an individual — «To some degree, this reduction can be done on a personal level and clinicians should communicate the risks of air pollutant exposures to their patients. More importantly, steps can be taken at a broader public policy level» (Wilker et al., 2023). A person’s own margin is thin (residence, filtration, commute); the lever lives mainly in the population/policy tier -> Layer 1 - Ranking Interventions for a Stratum. So Wilker quantifies and weakens the Commission’s air-pollution PAF slice; it neither overturns the life-course model nor displaces the cardiometabolic big rocks.

The physical inactivity cell above carries the Commission’s borrowed activity figure — RR 0.80 (0.77-0.84). Its primary source is now held first-hand: Iso-Markku 2022, a gold SR+MA + quality assessment of 58 prospective cohorts. The borrowed cell is this study’s exact all-cause estimate — RR 0.80 (0.77-0.84, n=257,983) — and Iso-Markku is the Commission’s own ref 165, so this is a de-secondhanding (type-F), not an [E-independent] second witness.

The effect, first-hand, with the subtype breakdown the single cell cannot show: «PA was associated with a decreased risk of all- cause dementia (pooled relative risk 0.80, 95% CI 0.77 to 0.84, n=257 983), Alzheimer’s disease (0.86, 95% CI 0.80 to 0.93, n=128 261) and vascular dementia (0.79, 95% CI 0.66 to 0.95, n=33 870), even in longer follow- ups (>=20 years) for all-cause dementia and Alzheimer’s disease. Neither baseline age, follow-up length nor study quality significantly moderated the associations.» (Iso-Markku et al., 2022) The protection is strongest for vascular dementia and weakest (but still significant) for AD — consistent with the shared cardiovascular route this page already counts.
It SURVIVES the reverse-causation / long-follow-up check — the payload the assumed-causal cell hides -> The U-Shaped Association Artifact. The whole study exists to answer the objection that «the association between PA and dementia appears absent when PA is measured before the age of 65 … or in follow-ups longer than 10 years» (Iso-Markku et al., 2022) (prodromal dementia lowering activity). Within the 16 studies of >=20 years’ follow-up the estimate holds: RR 0.79 (0.71-0.87), mean baseline age 50.5 y, mean follow-up 27.6 y; AD holds too at RR 0.76 (0.64-0.90, 7 studies). Iso-Markku’s verdict: «we did not find evidence to suggest that reverse causation or regression dilution bias88 affected the observed associations between PA and dementias.» (Iso-Markku et al., 2022) This is the strong-adjudication direction the U-Shaped concept asks for — a protective association that passes a long-follow-up exclusion, not merely covariate adjustment.
Honesty guard — the CLEANEST-design subset loses significance. The point estimate is stable across quality strata, but the three high-quality studies with a young (30-55) baseline AND >20-year follow-up gave RR 0.79 (0.62-1.01) — non-significant (Iso-Markku et al., 2022), and the three high-quality AD studies gave 0.71 (0.42-1.22), also non-significant. So the reverse-causation survival is directional and consistent, but the very cleanest cut is under-powered, not independently significant. Read the 0.80 as a robust-but-observational association, not a demonstrated causal slope.
The residual confound the design CANNOT remove — baseline cognitive reserve. Iso-Markku is explicit that surviving reverse causation is not surviving all confounding: «Physically active individuals may have higher cognitive reserve to start with … higher cogni- tive ability or other unmeasured confounding factors, and not leisure-time PA, may drive the association with a decreased inci- dence of dementia.» (Iso-Markku et al., 2022) Baseline cognitive ability is barely measured or adjusted for in the pool, so a cognitive-reserve confound stays live — the unremovable alternative behind the still-observational grade.
ApoE ε4 does NOT modify the effect (route-b NULL) — reinforces modifiable regardless of genotype. «Most studies that investigated ApoE ε4 interactions found no significant interactions (9 of 11 studies)» (Iso-Markku et al., 2022); carriers RR 0.81 (0.67-0.98) vs non-carriers 0.72 (0.56-0.92), and «Our ApoE ε4 interaction analyses suggest no such modification for all-cause dementia, Alzheimer’s disease or vascular dementia.» (Iso-Markku et al., 2022) So the activity lever is not one to withhold or intensify by genotype — the same route-(b)-fails reading the Commission states generally, now first-hand for PA specifically. (The between-group heterogeneity test was «invalid because of large heterogeneity», so this is absence of demonstrated modification, not a proven identity.)
Work-related PA runs the OPPOSITE way -> The Physical Activity Paradox. «The two studies examining the association of work-related PA and all-cause dementia showed an opposite trend than other PA (RR 1.25, 95% CI 0.98 to 1.59)» (Iso-Markku et al., 2022) — directional only (2 studies, CI crosses 1). It corroborates the domain-flip on a new (dementia) outcome, though Iso-Markku reads it partly as the cognitive-reserve/SES confound above.
Publication bias detected. «Funnel plots showed possible publication bias for all-cause dementia and Alzheimer’s disease» (Iso-Markku et al., 2022), for AD suggesting under-reporting of null results — so the pooled protection may be modestly inflated.

Parameter table — the borrowed cell vs the first-hand source [inferred from @isomarkku2022; @livingston2024]. Unlike the air-pollution de-secondhanding, the headline number is the SAME quantity (the cell IS this source), so this is an identity plus enrichment, never a tension — the value is the facets the single cell cannot carry.

Parameter	Commission cell (borrowed)	Iso-Markku 2022 (first-hand)	Same quantity?
All-cause dementia RR	0.80 (0.77-0.84)	0.80 (0.77-0.84), n=257,983	YES — the cell is this figure
Subtype resolution	none (one cell)	AD 0.86 (0.80-0.93); vascular 0.79 (0.66-0.95)	new — cell cannot show
>=20-year follow-up	not shown (assumed causal)	all-cause 0.79 (0.71-0.87); AD 0.76 (0.64-0.90)	new — the reverse-causation check
Cleanest subset	not shown	3 high-quality young-baseline >20y: 0.79 (0.62-1.01), NS	new — under-powered
ApoE ε4 modification	modifiable regardless of genotype (general)	no interaction (9/11), «no such modification»	new — PA-specific route-b null

Layer-1: this firms an existing big rock, it does not add a factor [inferred from @isomarkku2022]. Physical inactivity is already one of the 14 and already a cardiometabolic big rock; the de-secondhanding adds the reverse-causation robustness and the subtype/ApoE detail, so the dementia benefit is a second patient-important outcome on the same non-substitutable, pleiotropic activity lever (no drug replicates it) — strengthening its Layer-1 rank rather than competing -> Layer 1 - Ranking Interventions for a Stratum, Physical Activity Dose and Mortality. The lever is be active broadly, not a modality (as the section below states).
The counter-source is now held, and the issue is JOINED (type-D) [@kivimaki2019inactivity]. Iso-Markku’s headline directly rebuts Kivimaki’s individual-participant meta-analysis (BMJ 2019, n approx 404,840, 19 cohorts; a different Kivimaki paper from Cognitive Stimulation at Work and Dementia), which found inactivity->all-cause dementia present at <10 y (HR 1.40, 1.24-1.59) but absent beyond 10 years (HR 1.01, 0.89-1.14) once reverse causation is minimised — while its cardiometabolic positive controls (diabetes 1.42, CHD 1.24, stroke 1.16) stayed robust in both periods, so «the lack of association between physical inactivity and dementia in the later follow-up is not an artefact of the methodology».
```
 
```
What the joined clash resolves to — filed as Does Physical Activity Protect Against Dementia Beyond Ten Years. The disagreement turns on ascertainment sensitivity and study-level-vs-IPD pooling, not the reverse-causation window; and on the cleanest cut each offers the two nearly converge (Iso-Markku’s best young-baseline >20 y subset is itself NS at 0.79, 0.62-1.01; Kivimaki’s >=10 y is 1.01). So the direct, dementia-specific effect sits at insufficient evidence, not proven either way, while the vascular-route benefit both studies share is undisputed. This weakens the warrant for reading the borrowed 0.80 cell as a direct neuroprotective slope, WITHOUT dislodging activity as a big rock — it is recommended for cardiometabolic reasons regardless (Kivimaki’s own conclusion supports «this basic tenet of prevention in public health»). (Iso-Markku et al., 2022; inferred from Kivimäki et al., 2019)

The hypertension cell above carries the Commission’s observational risk-factor figure — untreated hypertension HR 1.42. That is the wrong quantity for a decision: it says having high BP raises risk, not that lowering it helps. Peters 2022 supplies the missing INTERVENTIONAL arm first-hand — an IPD meta-analysis of five seminal double-blind placebo-controlled antihypertensive RCTs (HYVET, SYST-EUR, PROGRESS, ADVANCE, SHEP; «28 008 individuals recruited from 20 countries», 861 incident-dementia cases) (Peters et al., 2022). This is the strongest de-secondhanding on the page: the others replace a borrowed observational cell with a better observational source; this replaces the observational risk factor with a randomized treatment effect («Class I evidence» (Peters et al., 2022)).

The effect, first-hand: «Multilevel logistic regression reported an adjusted odds ratio 0.87 (95% confidence interval: 0.75, 0.99) in favour of antihypertensive treatment reducing risk of incident dementia with a mean BP lowering of 10/4 mmHg.» (Peters et al., 2022) A ~13% relative reduction in dementia odds from a sustained ~10/4 mmHg BP fall in older hypertensives (mean age 69.1).
In ABSOLUTE terms — a small per-person gain. «Incident dementia occurred in 403 (2.9%) and 458 (3.3%) of those in active and placebo groups» among those with >=2 years’ follow-up (Peters et al., 2022) — an absolute risk difference of ~0.4 percentage points over median 4.3 years, i.e. NNT ~250 over ~4.3 years. (inferred from Peters et al., 2022) The absolute effect is small and CONSERVATIVE: the trials «stopped early upon achieving the es-timated primary cardiovascular endpoint» (Peters et al., 2022), and dementia accrues slower than the CV events that stopped them, so follow-up under-captures dementia and the true effect is plausibly larger. This is also the LAST such evidence — «it is no longer ethical to recruit to a trial comparing antihypertensive treatment to a placebo group» (Peters et al., 2022).
No effect-modification -> route-(a), treat on baseline risk. «There were no interactions by age, baseline BP, or history of stroke» (Peters et al., 2022), so the relative effect is uniform and stratification runs through baseline risk (absolute benefit), not effect-modification -> Baseline Risk and the Relative-Absolute Split. Consistent with preDIVA (the lever pays where BP is genuinely untreated) and with the treat-on-absolute-risk rule on Blood Pressure Lowering and Cardiovascular Events.
The observational U-shape does NOT survive randomization -> The U-Shaped Association Artifact. The abstract frames the problem: «Observational studies indicate U-shaped associations of blood pressure (BP) and incident dementia in older age, but rando-mized controlled trials of BP-lowering treatment show mixed results on this outcome in hypertensive patients» (Peters et al., 2022) — the apparent harm at low BP in old age. Under randomization the U vanishes: «There was no evidence of a U-shaped re-lation of the effect at any age, nor any increase in risk of dementia with treatment in the oldest age» (Peters et al., 2022), and the dose-response is monotone — «a linear relationship between lower risk of dementia and lower BP, down to at least 100 mmHg systolic and 70 mmHg diastolic» (Peters et al., 2022). Studied-range caveat: 100/70 is the observed floor, not evidence benefit continues below it.
~Half the effect runs through BP itself (mediation). «attributing 53% (95% CI: 27%, 76%) of the difference in dementia seen between the treatment and control groups to the effect of on systolic BP rather than any other aspects of trial participation or pleotropic antihypertensive drug effects» (Peters et al., 2022); controlled indirect (BP-mediated) risk difference −0.218% (−0.311%, −0.109%). So the BP channel is real but not exclusive — some benefit is drug/other.
Outcome-specific — dementia diagnosis yes, cognitive-decline no. «In comparison to the SPRINT-MIND trial, we found no effect of treatment on cognitive decline» (Peters et al., 2022) (attributed partly to MMSE insensitivity). The effect is on the dementia-diagnosis endpoint, not the continuous MMSE decline endpoint — an honest outcome-specificity caveat.

Parameter table — the observational cell vs the interventional source [inferred from @peters2022bp; @livingston2024]. The headline numbers are NOT the same quantity — an observational risk-factor association (having untreated hypertension) versus a randomized treatment effect (lowering BP) — so this is a genuine design-upgrade de-secondhanding, not an identity-plus-enrichment.

Parameter	Commission cell (observational)	Peters 2022 (interventional, first-hand)	Same quantity?
Estimate	untreated HR 1.42 (1.15-1.76)	treatment OR 0.87 (0.75-0.99)	NO — risk-factor RR vs RCT treatment effect
Design	observational cohorts	IPD MA of 5 double-blind placebo RCTs («Class I»)	NO — the design upgrade is the point
What it licenses	high BP is a risk marker	lowering BP reduces dementia (a decision)	NO — marker vs lever
Dose-response	(implied U-shape in old age)	linear to 100/70, no U at any age	new — the interventional refutation
Absolute effect	none given	ARD ~0.4pp / NNT ~250 over 4.3y (conservative)	new — cell cannot show
Effect-modification	not resolved	none by age / baseline BP / stroke history	new — route-(a)

Layer-1: this firms an existing big rock AND sizes the drug-vs-lifestyle rock [inferred from @peters2022bp]. Hypertension is already one of the 14 and a cardiometabolic big rock; the interventional evidence converts the dementia benefit from an assumed PAF into a demonstrated randomized effect, adding a second patient-important outcome to the same BP lever already carried for CV events -> Blood Pressure Lowering and Cardiovascular Events. Because a mature, low-harm antihypertensive class has the RCT dementia evidence while lifestyle BP-lowering does not, the drug is a genuine Layer-1 comparator here (the substitution principle), though the two are additive on the shared BP channel rather than exclusive -> Layer 1 - Ranking Interventions for a Stratum.
NOT [E-independent] of the CV-events evidence. The five trials sit inside BPLTTC’s 48-trial base and the author lists overlap (Chalmers, Woodward, Anderson — George Institute), so the shared RCT substrate defeats independence on the BP-lowering mechanism. What is new is the OUTCOME (dementia), a new-endpoint extension of the same trials, not a second independent witness. (Blood Pressure Lowering Treatment Trialists Collaboration, 2021; inferred from Peters et al., 2022)

What the framework changes about what to do

Most levers are shared with the cardiometabolic big rocks. Hypertension, diabetes, obesity, LDL, smoking, physical inactivity and excessive alcohol are already the vault’s hard-CV levers — so pulling them buys a second patient-important outcome (cognition) on top of CV events, strengthening their Layer-1 ranking rather than competing with it -> Layer 1 - Ranking Interventions for a Stratum, Big Rocks (Elderly). The dementia-specific additions are hearing, vision, education/cognitive activity, depression, TBI protection, social contact and air pollution.
The activity lever is be active, not a specific modality. The RR 0.80 attaches to physical activity broadly; claims that a particular type (mind-body exercise — Tai Chi, yoga) does extra cognitive work over generic activity are not established — the apparent advantage is a cross-review comparator artifact -> Mind-Body Exercise and Cognition. Rank the rock as being active, not as a modality choice.
Timing: earlier and longer. «it is never too early or too late to reduce» dementia risk; the operative rule is «decrease risk factor levels early (the earlier, the better) and keep them low throughout life (the longer, the better)». (Livingston et al., 2024) Duration of exposure matters (mid-life diabetes and hypertension carry the risk; the same conditions arising in late life often do not, partly a shorter-exposure / reverse-causation artifact).
Risk is modifiable regardless of APOE genotype — «Multicomponent interventions … potentially benefit individuals with either high or low genetic dementia risk». (Livingston et al., 2024) Genetic risk is not a reason to withhold the modifiable levers.
Compression of morbidity is a trajectory finding, not only a length one. Healthier lives not only lower dementia risk but push its onset back further than life expectancy extends, compressing the years lived with dementia -> the shape-of-decline outcome the telos centres.

The diet lever — quantified, observational, and probably NOT additive to the cardiometabolic rocks

Diet is not one of the Commission’s headline 14 factors as a standalone — it acts through its components (hypertension, diabetes, obesity, LDL, activity), which are counted separately. A specific Mediterranean-diet -> AD/MCI quantification lets us size the lever and, more importantly, see why adding it on top of those factors would double-count.

The single-COMPONENT arms below (dairy, flavonoids, soy) share one structure the individual sections only state piecewise, lifted on Single Food Components and Cognitive Outcomes: each shows a signal on a soft or surrogate endpoint and none on the hard AD / incident-dementia diagnosis — a Layer-1 ceiling that licenses not chasing individual foods or isolated-component supplements for cognition.

The association (Garcia-Casares 2021, dose-response MA, 11 studies / 12,458 participants). Per one-point rise on the 0-9 MD adherence score: AD RR 0.89 (0.84-0.93) («an 11% reduced risk of developing AD»), MCI RR 0.91 (0.85-0.97), composite 0.89 (0.86-0.92). Restricting to the lower-bias cohort designs attenuates it slightly to RR 0.91 (0.88-0.94) with no heterogeneity. (García-Casares et al., 2021)
Association, NOT established causation — the confound is the story. Every pooled study is observational and most are cross-sectional; the authors state this «limit[s] to infer causality» (García-Casares et al., 2021). Two biases bind hard here: (i) healthy-user — «Adherence to MD is included in a lifestyle pattern inﬂuenced by sociocultural, educational, family and economic factors» (García-Casares et al., 2021), several of which (less education, inactivity, depression) are themselves in the 14; and (ii) reverse causation over the long prodrome — «cognitive impairment patients could obtain MDs that do not accurately quantify their adherence to MD in the past» (García-Casares et al., 2021). Dietary self-report also carries large measurement error -> Measurement Error in Dietary Assessment. So this is a low-to-moderate-certainty association; do not read the RR as «MedDiet prevents Alzheimer».
Mediation / double-counting flag [inferred from @livingston2024; @garciacasares2021]. The source frames the effect as running «directly (by its neuroprotective effects) as well as indirectly (being protective factors of cardiovascular and metabolic diseases, which are themselves risk factors for AD)» (García-Casares et al., 2021). Since the Commission already counts hypertension, diabetes, obesity, LDL and inactivity, a MedDiet dementia benefit is likely substantially mediated by those already-counted levers, not additional to them — so the diet lever is largely a route to pull the cardiometabolic rocks, not a 15th independent PAF slice to add on top.
The RCT leg is separate and weaker. No RCT met this MA’s inclusion; the randomized diet-dementia evidence lives in the multicomponent FINGER family, which shows only a small cognitive-composite benefit and a null on dementia incidence (see Scope guard). A cohort association of this size has not been reproduced as a randomized incidence effect.
Umbrella credibility grade + the design discordance (Dinu 2018, F-refinement). Dinu’s umbrella grades the Med-diet -> neurodegenerative-disease association Convincing (Sofi RR 0.87), Alzheimer’s Convincing (Wu RR 0.60) — «convincing evidence in favour of a positive relationship» — which raises the certainty language on the cohort signal above. But it also bounds it two ways that reinforce the confound story: (i) «the meta-analysis of cross-sectional studies provided no evidence» — the signal is cohort-only; and (ii) dementia is internally discordant — Cao 2015 grades it Convincing (RR 0.69) while Wu 2017 grades the same outcome No evidence (RR 1.07). So “convincing” here is design- and MA-selection-dependent, consistent with a moderate-at-most read. (Dinu et al., 2017)

A second, larger MedDiet -> dementia SR+MA (Nucci 2024; PRISMA-2020, PROSPERO CRD42023444368, Newcastle-Ottawa RoB, 21 datasets / 65,955 participants, elderly >=60 only) firms the direction on ~5x the pooled n but adds no causal-identification route — every constituent is observational (13 cohort + 6 cross-sectional + 1 both + 1 nested case-control), and it deliberately includes the cross-sectional studies prior MAs excluded (its own limitation: «most of the included studies were cross-sectional, in which causality might not be assessed by definition») (Nucci et al., 2024). So it is a magnitude/robustness refinement of the Garcia-Casares leg, not independent replication and not a causal upgrade.

Estimates (contrast = highest-vs-lowest adherence, NOT per-point; carry each I2).
- all-type dementia OR 0.89 (0.84-0.94), n=65,955, I2=69.94 (moderate) — «the highest adherence to the Med-Diet is associated with an approximate 11% reduction in the likelihood of developing dementia» (Nucci et al., 2024). The headline pools ORs (cross-sectional) and HRs (cohort) as one figure — «reported as OR or HR based on the study design».
- cohort-only HR 0.84 (0.76-0.94), n=55,205, I2=89.70 (very high) — the pooled figure averages over highly heterogeneous designs/populations, and on the stricter cohort-only cuts «the association was borderline significant» (Nucci et al., 2024).
- Alzheimer disease OR 0.73 (0.62-0.85), n=38,292, I2=63.85 — «an approxi-mately 27% lower risk of Alzheimer’s disease»; the effect is outcome-specific (stronger for AD than all-dementia), while MCI alone is OR 0.89 (0.79-1.01), non-significant (3 studies) — «the association was no longer significant when looking at MCI alone» (Nucci et al., 2024).
New caveats the incumbent leg did not carry (this is what makes it F, not mere confirmation). (i) Publication bias DETECTED — Egger intercept -1.08, p=0.013 (main); -1.88, p=0.055 (AD); after trim-and-fill the main estimate attenuates to OR 0.92 (0.86-0.99) but stays significant (Nucci et al., 2024). (ii) Very high cohort heterogeneity (above). (iii) Half the pool cross-sectional, so temporality is unestablished and reverse causation stays live. (iv) Elderly-only, so «the effects of recently adopted dietary habits may have less influence on health outcomes such as dementia (which require long-term expo-sure)» (Nucci et al., 2024) — a shorter-exposure / reverse-causation frame that reinforces, not relieves, the confound story -> The U-Shaped Association Artifact.

Parameter table — the 0.89 trap [inferred from @nucci2024; @garciacasares2021]. Garcia-Casares’ figures are per one-point rises on the 0-9 MD score; Nucci’s are high-vs-low category contrasts. They are not the same quantity, so the identical 0.89 is a coincidence, not a replication.

Parameter	Garcia-Casares (incumbent)	Nucci 2024	Same quantity?
Contrast	per +1 point on 0-9 MD score (dose-response)	highest-vs-lowest adherence category	NO
AD estimate	RR 0.89 (0.84-0.93) per-point	OR 0.73 (0.62-0.85) high-vs-low	NO
all-dementia / composite	composite RR 0.89 (0.86-0.92) per-point	OR 0.89 (0.84-0.94) high-vs-low	NO — same number, different quantity
MCI	RR 0.91 (0.85-0.97) per-point, sig	OR 0.89 (0.79-1.01) high-vs-low, NS	NO
Population	mixed adult	elderly >=60 only	NO
Pooled n	12,458	65,955 main / 38,292 AD	NO

The critical cell: a high-vs-low contrast spanning several score-points would give a larger effect than one point if the per-point slope held — so a Nucci all-dementia OR numerically equal to Garcia-Casares’ per-point composite implies Nucci’s per-unit effect is, if anything, shallower, not a confirmation. Every “Same quantity?” cell is NO, so this is a bounding/refinement, never a tension.

Shared-cohort / non-independence note (F, not [E-independent]). Nucci and Garcia-Casares pool the same underlying MedDiet-dementia observational literature — the foundational AD cohorts (Scarmeas 2006a/b and 2009, Féart 2009, Gu 2010) that any MedDiet-AD MA rests on sit in Nucci’s included-studies list (Nucci et al., 2024). Overlapping constituent cohorts defeat independence, so two MAs agreeing is not two witnesses; do not lift confidence: on the agreement.
Mediation corroborated, not overturned. Nucci’s own mechanism section leads with the cardiometabolic pathway — «Cardiovascular risk factors such as hypertension, obesity (mainly abdominal obesity), dyslipidaemia, and type 2 dia-betes are considered to have a significant impact on the risk of dementia» (Nucci et al., 2024) — plus amyloid/tau-biomarker and gut-microbiota routes (mechanism, directional,-grade, not outcomes). This reinforces the double-counting flag above: the MedDiet -> dementia benefit runs substantially through already-counted levers, so Nucci does not convert the diet lever into a 15th additive PAF slice (Livingston et al., 2024; inferred from Nucci et al., 2024).
Confidence: held medium. Nucci firms direction and magnitude on ~5x the n, but adds no causal route, carries very high cohort heterogeneity and detected publication bias, a non-significant MCI leg, and shares constituent cohorts with the incumbent leg — so the diet association stays the weaker low-to-moderate leg and the Commission’s consensus backing still sets the page grade. Not raised on a second SR arrived.

The same Mediterranean pattern has an RCT on cardiovascular hard outcomes (PREDIMED, stroke-driven, at high baseline risk) -> Mediterranean Diet and Cardiovascular Events — which is consistent with the mediation reading: the pattern’s best-evidenced effect is on the vascular channel that feeds AD risk.

Periodontal disease — an observational candidate lever, NOT one of the 14 (2026-08-23)

Periodontal disease (PD) is not one of the Commission’s 14 factors, and the evidence for it is weaker than for any factor on the headline list — it is a candidate lever held at association grade only. A gold SR+MA (Dibello 2024, 46 observational studies) finds PD associated with incident dementia RR 1.22 (95% CI 1.10-1.36) (8 cohorts, n=3,076,684 dementia-free at baseline, mean follow-up 11 y) — «PD was associated to higher risk of incident dementia (RR 1.22, 95% CI 1.10 to 1.36) with sig- nificant heterogeneity across the studies (I2= 95%, p< 0.001)» (Dibello et al., 2024), plus cognitive impairment RR 1.25 (1.11-1.40) and cognitive decline RR 3.01 (1.52-5.95); PD-depression is null (RR 1.07, 0.95-1.21).

Why it stays a candidate, not a 15th factor [inferred from @dibello2024periodontal]. The signal is entirely observational with very high heterogeneity (I2=95%) and the same two biases that discount the diet lever above: (i) reverse causation over the long prodrome — «poor peri- odontal health may not necessarily cause dementia, but rather results from it» (Dibello et al., 2024) -> The U-Shaped Association Artifact; and (ii) shared confounders that are already-counted factors — the discussion names cardiovascular disease, diabetes, low education, smoking, drinking, socioeconomic status and gender as common to both PD and dementia (Dibello et al., 2024), so a raw PD-dementia RR is largely mediated by the cardiometabolic/SES cluster this page already holds. No trial has tested whether treating PD lowers dementia incidence. So the lever is confidence: low, and an interventional periodontal-treatment -> cognitive-outcome study before any dementia-prevention claim. The treatable/causal-grade arm of this exposure is glycaemic control, not cognition -> Periodontitis and Systemic Health holds the full two-arm decomposition.

Omega-3 fatty acids / oily fish — an observational candidate lever, NOT one of the 14 (2026-08-23)

Omega-3 (marine EPA/DHA) is not one of the Commission’s 14 factors, and like the diet and periodontal levers above it is a candidate held at association grade only. A gold SR+MA (Wei 2023; 48 cohorts, 31 pooled, 103,651 participants) finds dietary DHA associated with cognitive decline RR 0.82 (95% CI 0.72, 0.93; Level H) — a 27% lower dementia / 24% lower AD risk — while dietary total omega-3 is weaker (RR 0.91, 0.82-1.00) and dietary EPA/ALA are null (Wei et al., 2023). Fish is the dominant dietary DHA source, so the full decomposition (exposure-form dissociation, the supplement-RCT contrast) lives on Fish and Seafood Consumption.

Why it stays a candidate, not a 15th factor [inferred from @wei2023omega3]. The same three discounts as the diet and periodontal levers apply, plus a fourth that is specific and decisive here:

Entirely observational and FFQ-based — dietary measurement error binds -> Measurement Error in Dietary Assessment, and the marine-n-3 signal travels with overall diet quality and SES (the observed-healthy-population problem).
Reverse causation over the long prodrome — Wei mitigated only partially (sensitivity analyses removing year-1 cases in the ADNI arm).
Largely mediated by / not additive to the cardiometabolic rocks — the plausible mechanism runs through vascular and inflammatory pathways this page already counts, so a raw omega-3 -> dementia RR is not a clean 15th additive PAF slice.
The isolate-supplement RCTs are NULL — the decisive discount. Wei’s own intro concedes «randomized clinical trials have shown limited efﬁcacy of omega-3 fatty acid sup-plementation in reducing cognitive decline and probable AD» (Wei et al., 2023), and the held RCT (MAPT, 800 mg DHA + 225 mg EPA/d) was cognition-null -> Multidomain Lifestyle Intervention and Cognitive Decline. So the lever is confidence: low, and any supplement dementia-prevention claim is refuted at RCT grade; the observational dietary signal is a reason to prefer oily fish within a good diet, not to prescribe capsules.

Ultra-processed food — the first FOOD-GROUP dietary lever, observational candidate, NOT one of the 14 (2026-09-03)

Until now this page held diet only at the pattern altitude (Mediterranean/MIND) plus the omega-3 and periodontal candidate levers — no food-group lever. Ultra-processed food (UPF, NOVA4) is the first, and like the diet-pattern and omega-3 levers it is a candidate held at association grade only, not a 15th factor. A gold SR+MA (Henney 2023; PROSPERO CRD42023388363, Newcastle-Ottawa + NutriGrade, the first SR+MA on NOVA-UPF -> dementia; 10 observational studies, 8 longitudinal, 867,316 participants) finds high (vs low) UPF intake -> all-cause dementia RR 1.44 (95% CI 1.09-1.90), p=0.02, I2=97.0% (Henney et al., 2023). The full outcome arm, the category-vs-composition critique and the modest-magnitude framing live on Ultra-Processed Food and Health Outcomes; only the dementia-lever verdict is here.

Why it stays a candidate, not a 15th factor [inferred from @henney2023upf]. The same discounts as the diet-pattern, periodontal and omega-3 levers apply, and two of them are directly evidenced in the SR:

No robust dose-response. Moderate UPF intake is null (RR 1.12, 0.96-1.31, p=0.13), so «we did not demonstrate a robust dose–response relationship between the quantity of UPFs consumed and dementia prevalence» (Henney et al., 2023); and every dementia subtype is individually non-significant (AD 1.08 (0.79-1.48), vascular 2.05 (0.39-10.90), MCI 2.01 (0.75-5.42), dementia-excl-MCI 1.24 (0.93-1.65)) (Henney et al., 2023). Only the pooled all-cause figure clears significance, at near-total heterogeneity (I2=97%).
Mediation by the already-counted cardiometabolic rocks is DIRECTLY shown, not just plausible. The association was lost when restricted to studies adjusting for type 2 diabetes (1.47, 0.97-2.00, p=0.06) and total energy intake (1.26, 0.95-1.67, p=0.09), while surviving BMI/CVD/SES adjustment (Henney et al., 2023). So a raw UPF -> dementia RR runs substantially through diabetes and total energy — levers this page already counts — rather than being a clean additive PAF slice. It also lost significance at a 10,000-participant sample-size cutoff («results lose significance when performing sensitivity analysis based on a sample size cut-off point of 10,000 par- ticipants») (Henney et al., 2023) — a small-study contribution -> The U-Shaped Association Artifact.
The exposure is reviewer-assigned and FFQ-measured. Only 1 of 10 studies directly referenced NOVA; «nine assessed various foods that were retrospectively defined as ultra-pro- cessed by our research team using NOVA criteria» (Henney et al., 2023), with «no pre-determined cut-off… Moderate intake in one population may exceed high intake in another» (Henney et al., 2023), all FFQ-based -> Measurement Error in Dietary Assessment, Is the Food Category Doing Any Work.
Reverse causation over the long prodrome stays live (the SR does not resolve temporality), and RCTs are ethically ruled out — «RCTs assessing the association between UPFs and incident cognitive impairment/dementia would be ethically unjustifiable» (Henney et al., 2023) — so the observational ceiling is by design.

So the lever is confidence: low, mediated-not-additive, and does not reorder the cardiometabolic big rocks; it reinforces them (reducing UPF is a route to pulling diabetes/obesity/energy-balance, the same mediation reading as the Mediterranean-diet lever above). The broader cognitive-outcome evidence (global cognition, cognitive domains and decline, beyond dementia diagnosis) now lands via a second SR (Smith 2025; narrative synthesis of five observational studies, preprint-provisional — medRxiv, not yet peer-reviewed, published version may differ) and does not change the lever’s grade. It is a vote-count with no pooled magnitude (3/5 studies a significant adverse main effect, all five an adverse subgroup effect), it shares the Li 2022 UK Biobank primary with Henney’s pool (F-refinement, not independent corroboration), and it carries the same diet-quality / composition confounding — «the adverse effects of UPF exposure on cognitive outcomes remained whilst controlling for adherence to a healthy diet in two out of the three studies (Bhave et al., 2024; Li et al., 2022)» (Smith et al., 2026), while diet quality modified it in the third. One decision-relevant frame it surfaces: replacing 10% of UPF weight with equivalent less-processed food was estimated at a 19% lower all-cause dementia risk (HR 0.81, 95% CI 0.74-0.89) in Li 2022 (Smith et al., 2026) — observational, the shared primary, same confounding caveat. The lever stays confidence: low. Full broader-cognition detail: Ultra-Processed Food and Health Outcomes.

Fruit and vegetables — a second FOOD-GROUP dietary lever, observational candidate, NOT one of the 14 (2026-09-04)

Where UPF is the harm-direction food-group lever, fruit-and-vegetable intake is the protective-direction one — and like the diet-pattern, omega-3 and UPF levers it is a candidate held at association grade only, not a 15th factor. A gold SR+MA (Zhou 2022; MOOSE, 16 observational studies — 9 cohort / 6 cross-sectional / 1 case-control, 64,348 older adults >=60, 9,879 cases) finds high (vs low) F&V intake -> cognitive disorders OR 0.82 (95% CI 0.75-0.90), I2 35.3% (Zhou et al., 2022) (the paper’s abstract headlines 0.79, its mixed-gender subgroup, not the Figure-2 pool — carry 0.82). Fruit alone 0.83 (0.77-0.89), vegetables alone 0.75 (0.70-0.80). The full outcome arm and the CVD-arm comparison live on Fruit and Vegetable Intake and Health; only the dementia-lever verdict is here.

Why it stays a candidate, not a 15th factor [inferred from @zhou2022fruit]. The same discounts as the diet-pattern, periodontal, omega-3 and UPF levers apply, and three are directly evidenced in the SR:

Alzheimer’s is NULL, so the signal is not a direct anti-neurodegenerative effect. Cognitive impairment/MCI 0.76 (0.72-0.80) and dementia 0.84 (0.78-0.91) are significant, but AD is null (0.88, 0.76-1.01) (Zhou et al., 2022) — consistent with the association riding the vascular/cardiometabolic route this page already counts, i.e. mediated-not-additive, the same reading as the Mediterranean-diet and UPF levers.
The design gradient is a reverse-causation tell. The prospective cohort subgroup shows the WEAKEST effect (0.83, I2 0%); cross-sectional (0.70) and case-control (0.68) — the designs where prevalent cognitive disease can depress recalled/current intake — the strongest. The association shrinks as the design gets less vulnerable to reverse causation over the long prodrome -> The U-Shaped Association Artifact, so the pooled 0.82 is an upper bound on any causal reading.
No robust dose-response, plus two measurement discounts. The reported linear trend (P=0.03) rests on only 4 studies (the only ones with >=3 exposure categories) with no located knee and no stated studied range — weak evidence of a true curve. Exposure is FFQ-dominated (23 of 31 effect groups), so dietary measurement error binds -> Measurement Error in Dietary Assessment, and the paper flags an OR-as-RR approximation that overstates the RR at these case fractions «using the OR as an approximation of the RR produces progressively larger errors as the outcome rate rises above 1%» (Zhou et al., 2022). No causal claim is made «no casual relationships could be established due to the observational nature of the studies» (Zhou et al., 2022).

So the lever is confidence: low, mediated-not-additive, and does not reorder the cardiometabolic big rocks; it reinforces them (more F&V is a route to the same vascular/metabolic levers, and the AD-null + cohort-attenuation both point that way). No independence (E) is claimed with any other diet lever here. The page grade stays medium on the Commission’s consensus backing; the F&V slice is the weaker low leg.

Flavonoids — the COMPONENT side of the food-group levers, observational candidate, NOT one of the 14 (2026-09-04)

Where fruit-and-vegetable intake (Zhou, above) is a whole-food-group lever, dietary flavonoids are a bioactive component nested inside those foods — the component-side sibling, and like every diet lever above a candidate held at association grade only, not a 15th factor. A gold SR+MA (Peng 2025; 26 observational studies, 269,574 participants, 15 cohorts) finds high (vs low) flavonoid intake -> adverse cognitive events OR 0.90 (95% CI 0.83-0.98), I2 62.2%, and better continuous cognitive test scores (β 0.03, 0.02-0.04) (Peng et al., 2026). The full subclass table, the fitted-linear dose-response and the food-vs-component critique live on Flavonoid Intake and Cognitive Function; only the dementia-lever verdict is here.

Why it stays a candidate, not a 15th factor — and specifically a COMPONENT of the F&V lever, not an additional one [inferred from @peng2025flavonoid]. The same discounts as the F&V lever apply, and three are directly evidenced in the SR:

Both hard diagnostic endpoints are NULL — only the soft endpoint moves. The pooled signal is carried by cognitive decline (OR 0.88, 0.79-0.98); dementia (OR 0.97, 0.79-1.19) and Alzheimer’s (OR 0.90, 0.69-1.17) are both null (Peng et al., 2026), echoing the AD-null in the Zhou F&V pool — consistent with a vascular/cardiometabolic route already counted, i.e. mediated-not-additive. The author reads the null as a power caveat, so dementia/AD is insufficient-evidence, not established no-effect.
Doubly-estimated FFQ exposure + confounding in the healthy-user direction. Flavonoid intake is FFQ food reports multiplied through a food-composition database (two error sources), and meta-regression found «BMI and smoking status were significant confounders in the relationship between dietary flavonoids consumption and cognitive function, potentially overestimating the positive effects if not adjusted for» (Peng et al., 2026) -> Measurement Error in Dietary Assessment, Is the Food Category Doing Any Work.
The flavonoid MA does not isolate the flavonoid. The exposure is computed from the same fruit/veg/tea whose contribution it would need to be separated from, so it is a re-expression of the F&V signal in flavonoid units, not an independent handle — no biomarker, no isolated-flavonoid arm, no MR. So Peng and the Zhou F&V MA are NOT independent corroboration (not type-E): the exposures are nested and share the observational substrate; they are a component/food-group type-F pair leaving the food-vs-component G-gap open -> Flavonoid Intake and Cognitive Function.

So the lever is confidence: low, mediated-not-additive, and does not reorder the cardiometabolic big rocks; it reinforces the F&V lever rather than adding to it (it is a part of it). No independence (E) is claimed. The page grade stays medium on the Commission’s consensus backing; the flavonoid slice is a weaker low sub-leg of the F&V lever.

Dairy — a candidate FOOD-GROUP lever whose only signal is a pooling artifact, NOT one of the 14 (2026-09-05)

Dairy is another food-group diet lever, sibling to UPF and fruit/veg above — and it is the weakest of them: a candidate held at association grade only, and one whose headline association is null. A gold dose-response SR+MA (Villoz 2024; 15 prospective cohorts, 312,580 participants, median 11.4 y follow-up, no RCTs) finds the highest-vs-lowest dairy contrast has «no association with cognitive decline or dementia (RR: 0.94; 95% CI: 0.82, 1.07)» (Villoz et al., 2024). The full effect-estimate arm, the type-B product decomposition and the region-pooling artifact live on the dairy-cluster facet Dairy and Cognitive Decline; only the dementia-lever verdict is here.

Why it stays a candidate, not a 15th factor — and specifically a NON-lever for the default stratum [inferred from @villoz2024dairy]. The same discounts as the other diet levers apply, and three are directly evidenced in the SR:

The only “signal” is a between-population pooling artifact. The apparent nonlinear dose-response («nadir at ~150 g/d», RR 0.88, 0.78-0.99) is drawn by pooling low-intake Asian cohorts (protective, RR 0.83, 0.75-0.92, I2 0%) with high-intake European cohorts (null, RR 1.01, 0.86-1.19) whose intake ranges barely overlap (Asia 29-165 g/d vs Europe 170-711 g/d) (Villoz et al., 2024). The «nadir» sits on the seam between two populations, not at a within-person optimum -> The U-Shaped Association Artifact. For a Western intake range the association is flatly null.
It vanishes on dietary-pattern adjustment — dairy as a diet-quality marker. «studies that took into account other food groups or dietary patterns that could affect the relationship between dairy consumption and cognitive function found no associations» (Villoz et al., 2024), and the plausible mechanism runs through the cardiometabolic levers this page already counts — so any effect is largely mediated-not-additive, not a clean 15th PAF slice. Milk-alone and cheese-alone are both null -> Is the Food Category Doing Any Work.
Observational, single-measurement FFQ, reverse-causation-prone. No RCT exists; differential recall bias (early cognitive change bending reported intake) stays live -> Measurement Error in Dietary Assessment.

So the lever is confidence: low — weaker than the other diet levers, since even its raw headline is null and its one curve feature is a pooling artifact. It does not reorder the cardiometabolic big rocks. A future independent test must avoid the Villoz/Filippini/Vinceti (Modena) author group -> an independent dairy-cognition SR/MA.

Sleep disorders — a non-diet candidate lever, NOT one of the 14 (2026-09-04)

Sleep is the first non-diet, non-activity candidate lever on this page, and — like the diet levers above — it is a candidate held at association grade only, not a 15th factor. A gold SR+MA (Zhang 2025; 76 longitudinal cohort studies, all Newcastle-Ottawa > 7, eight sleep-disorder types x four cognitive outcomes, non-demented adults) finds every major sleep disorder associated with higher dementia / decline risk: SRMD (restless-legs) -> VD 2.53 (1.30-4.93), EDS -> VD 1.85 (1.39-2.47), long sleep > 8 h -> AD 1.66 (1.44-1.91, I2 0%) and dementia 1.43 (1.21-1.69), OSA/SRBD -> AD 1.39 (1.16-1.68), poor sleep quality -> dementia 1.17 (1.03-1.32), insomnia -> dementia 1.13 (1.04-1.23); circadian-rhythm disturbance and RBD are null (J. Zhang et al., 2025). The full disorder x outcome map, the measurement check and the arm-level adjudication live on Sleep and Cognitive Decline; only the dementia-lever verdict is here.

Why it stays a candidate, not a 15th factor [inferred from @zhang2025sleep]. The same discounts as the diet levers apply, plus a duration-arm hazard specific to sleep:

Entirely observational, high heterogeneity, publication bias present. Most pools carry I2 60-90%; publication bias was «found … for multiple sleep disorders» (J. Zhang et al., 2025) (trim-and-fill kept estimates consistent), and meta-regression left the heterogeneity only «partially accounted for».
The long-sleep -> dementia arm is a likely PRECLINICAL MARKER, not a cause -> The U-Shaped Association Artifact. The U is asymmetric and outcome-specific: short sleep (< 7 h) hits cognitive decline but «was not associated with future risk of all-cause dementia and AD», while long sleep (> 8 h) drives the dementia/AD arm and «is closely associated with age» (concentrated in the >= 70 y elderly), a «preclinical marker driven by the APOE ε4 carrier gene» (J. Zhang et al., 2025). Zhang runs only weak checks (baseline-only exposure, fixed follow-up; no MR, no referent-correction), so the long arm is unadjudicated — do NOT read it as sleep less to avoid dementia.
The VD-loaded disorders (SRMD, EDS, OSA) plausibly run through the cerebrovascular route already counted, so their benefit is largely mediated-not-additive to the cardiometabolic rocks, not a clean 15th PAF slice. The insomnia signal actually STRENGTHENS under objective measurement (objective insomnia RR 1.26, I2 26.1% vs pooled 1.09, I2 77.7%), arguing against a pure self-report artifact on that arm.
No interventional -> incidence evidence. No trial here tests whether treating a sleep disorder lowers dementia incidence -> a CBT-I / CPAP -> cognitive-outcome study -> Sleep Aids and Insomnia Treatment, Sleep Apnea Treatment and Cardiovascular Risk.

So the lever is confidence: low, mediated-not-additive on the vascular arm, and screening-relevant (route-a prognostic: a sleep disorder marks higher baseline risk) rather than a demonstrated intervention; it does not reorder the cardiometabolic big rocks. The page grade stays medium on the Commission’s consensus backing; the sleep slice is the weaker low leg.

Second landmark MA now held — Xu 2020 — but it does NOT raise the sleep leg [inferred from @xu2020sleep; @zhang2025sleep]. The prior JNNP landmark Xu 2020 (51 cohorts, 260,915 participants for the insomnia pool) is now held and converges with Zhang on direction and magnitude — insomnia -> cognitive disorders RR 1.27 (1.16-1.39), long-sleep -> AD 1.57-2.47, apnoea 1.29 (1.12-1.48) (Xu et al., 2019). The convergence is not independent (type-E) backing: Zhang updates and cites Xu (ref 17), sharing the observational-cohort primary literature, so it is a prior-vs-updated echo (type-F) that leaves confidence: low unchanged. What Xu adds as an F-refinement is the covariate-sensitivity finding — the insomnia associations «became non- significant when specific factors (hypnotics, APOE4 status and depression at baseline) were included as covariates» (Xu et al., 2019) — which bounds the bare insomnia RR (some signal is shared variance with depression/APOE4/hypnotics), and it confirms the reverse-causal long-sleep arm is left unadjudicated by BOTH landmark MAs (Xu, like Zhang, runs no MR or referent-correction and states outright the associations «were not equal to causal relationships» (Xu et al., 2019)). The full parameter table and F-analysis live on Sleep and Cognitive Decline. NB: Xu Wei here is the same recurring author who co-authored the held Zhang 2022 dementia umbrella cited elsewhere on this page — a second work by that author, not independent corroboration.

The whole map above is observational — modelled PAFs on relative risks “assumed causal.” The one landmark interventional test is FINGER (Ngandu 2015), a 2-year double-blind multidomain RCT (n=1260, at-risk CAIDE >=6 elderly) -> Multidomain Lifestyle Intervention and Cognitive Decline. It confirms the direction (acting on the levers as a bundle improves cognition) at RCT grade, but the two are not the same quantity, so it bounds and de-risks the PAF rather than validating or refuting it:

Parameter	Livingston (observational)	FINGER / Ngandu (RCT)	Same quantity?
Design	consensus of observational SRs/MAs + MR	double-blind multidomain RCT, n=1260	NO
Exposure unit	14 separable per-factor risk factors	one non-decomposable 4-component bundle	NO
Endpoint	dementia incidence (modelled PAF 45.3%)	NTB cognitive composite Z-score (surrogate)	NO
Magnitude	~45% population-attributable	between-group 0.022/yr, Cohen’s d 0.13; incidence NOT measured	NO

(Ngandu et al., 2015)

What the RCT does and does not buy the PAF [inferred from @ngandu2015; @livingston2024]: (i) it raises confidence that the aggregate causal assumption is sound — intervening on the bundle really moves cognition; (ii) but it cannot validate the per-factor decomposition the PAF table depends on (FINGER is bundle-only, author-flagged as needing component-level study); and (iii) it leaves the PAF’s own endpoint — dementia incidence — open (the significant effect is on a cognitive surrogate, and the effect is small, d=0.13). So the 45.3% figure is neither confirmed nor overturned by FINGER; the trial de-risks the direction while the decomposition and the incidence claim stay unproven -> the earlier secondhand “FINGER-family” reading below is the meta-analytic bound.

The diabetes lever splits by DRUG CLASS — the pharmacotherapy channel opened (2026-09-04)

Until now this page counted diabetes only as a shared cardiometabolic incidence rock — pull it (by lifestyle or otherwise) and dementia incidence falls with it. Kuate Defo’s umbrella opens a second, distinct decision inside the already-diabetic stratum: for the large population already on a glucose-lowering drug (T2D affects ~450M; metformin is first-line), which class they are on is itself a cognition-relevant contrast. This is the pharmacotherapy-taper lens the telos holds standard drugs for — a standard drug’s effect on a non-primary outcome (dementia) is a stratum-level decision-change.

The incidence anchor first (confirms the shared-rock framing). Across the 100-review umbrella, diabetes raises AD RR 1.39 (1.16-1.66) to 1.57 (1.41-1.75) and, more strongly, VaD RR 1.91 (1.61-2.25) to 2.49 (2.09-2.97) (Kuate Defo et al., 2023). Two Mendelian-randomization studies did not find a T2D->dementia effect (higher fasting glucose did raise AD odds) — a genetic-instrument caveat that the observational incidence signal may be partly confounded, consistent with this page already counting diabetes through its cardiometabolic company.

The drug-class split (treatment MA of 27 observational studies, N=3,046,661; incident dementia). Random-effects RR vs unexposed/comparator, heterogeneity I2 in brackets:

Lower risk: metformin 0.83 (0.71-0.96) [I2 99.1%]; GLP-1 RA 0.35 (0.16-0.78) [98.5%]; SGLT2i 0.39 (0.20-0.76) [96.1%]; pioglitazone 0.74 (0.55-0.98) [74.7%]; all thiazolidinediones 0.770 (0.593-0.999) [97.8%].
Neutral: DPP-4i 1.04 (0.79-1.38); alpha-glucosidase inhibitors 1.04 (0.89-1.22); insulin 0.97 (0.78-1.20).
Higher risk: meglitinides 1.87 (1.43-2.45) [I2 0.0%, only 2 studies]; sulphonylureas 1.39 (1.04-1.87) [99.8%].

(Kuate Defo et al., 2023)

Confounding by indication is the load-bearing threat — read every number through it. Every contrast is observational, and the comparators are a mix (metformin vs sulphonylurea; drug vs no-drug; drug vs «other antidiabetic monotherapy»). Metformin is first-line (earlier-stage, better renal function, healthier adherers); sulphonylureas / meglitinides / insulin mark later-line or more severe disease. So part of metformin’s «benefit» and the sulphonylurea/meglitinide «harm» is the same disease-severity / line-of-therapy contrast, not a drug acting on the brain. The paper’s own tells align: the harm signals are «in line with their association with severe hypoglycaemia» (Kuate Defo et al., 2023), and metformin’s benefit attenuates with longer follow-up (meta-regression pinteraction=.039) — «the observed benefit of medication may dissipate in longer-term studies and there may exist time-related biases in the efficacy of antidiabetic treatments» (Kuate Defo et al., 2023), the signature of prevalent-user / healthy-adherer bias.

Certainty is LOW-to-VERY-LOW across the board, and the umbrella does not grade it for us. Four deflators stack: (i) no RCT with incident dementia — «most studies … have been cohort studies and not RCTs» (Kuate Defo et al., 2023); (ii) I2 ~ 99% for most classes (within-class studies point opposite ways — metformin ranges from 1.68 harmful to 0.52 protective), so a barely-significant pool hides gross inconsistency; (iii) publication bias detected (Egger p=.041); (iv) mixed review quality («only 53.3% of systematic reviews and 72.9% of meta-analyses were of good quality» (Kuate Defo et al., 2023)). The umbrella used AMSTAR + Newcastle-Ottawa risk-of-bias only — no convincing/suggestive credibility grade per class. And in sensitivity analysis TZD and sulphonylurea lost significance (metformin held) — so two of the ten signals are already fragile on the authors’ own re-analysis. Metformin’s effect was also significant only in US/Western populations, null in Eastern (RR 1.06, 0.79-1.41) — an unexplained region split.

What it changes about what to do. For someone with T2D, the evidence leans toward metformin / GLP-1 RA / SGLT2i / pioglitazone over sulphonylureas / meglitinides on the cognition axis — but the lean is soft (low certainty, heavily confounded) and per-person agent selection is prescriber-zone, out of scope. What is IN scope is the appraisal: the newer agents (GLP-1 RA, SGLT2i) — already ranked high for CV, renal, and weight outcomes -> GLP-1 Non-Cardiometabolic Effects and Safety, SGLT2 Inhibitors — carry a possible additional cognition dividend, not an established one; and the old secretagogues carry a possible cognitive penalty that tracks their hypoglycaemia risk. This is a drug-vs-drug refinement within the diabetes lever, NOT a lifestyle-vs-drug rock-sizing: lifestyle prevention of T2D incidence (the shared rock this page already counts, and the structural lever that removes the driver -> Lifestyle vs Metformin for Diabetes Prevention) is untouched and still ranks above managing which pill.

Scope guard

Multicomponent prevention RCTs (FINGER-family) show only a small cognitive-composite benefit (Cochrane MD 0.03, 0.01-0.06) and wide, null-spanning effects on dementia incidence (RR 0.94, 0.76-1.18) — the population PAF is a modelling claim, not a demonstrated trial effect on incidence. (Livingston et al., 2024) That incidence-null is now held first-hand, not secondhand: preDIVA (Moll van Charante 2016), a 6.7-yr multidomain vascular-care cluster-RCT (n=3526, unselected elderly), found HR 0.92 (0.71-1.19) on clinical dementia incidence — a primary constituent of the pooled null above, consistent with it, and detailed with FINGER’s surrogate-positive contrast on Multidomain Lifestyle Intervention and Cognitive Decline. (van Charante et al., 2016) Treatment, diagnosis, biomarkers and drug therapy (cholinesterase inhibitors, anti-amyloid antibodies) are appraised on the source page, out of this prevention framework’s scope.

And no responder subgroup rescues the incidence-null — the PAF’s implicit target the high-risk hope is tested and fails on the hard endpoint [2026-08-07, Coley]. Pooling the MAPT+preDIVA IPD (n=5205, up to 12 yr), «there was no effect of multidomain intervention on the risk of all-cause dementia (HR 0.98, 95% CI 0.80–1.21)», with no effect in any of 11 pre-specified risk-factor subgroups and none found by a data-driven SIDES search (Coley et al., 2025). This matters for how the PAF is used: the observational per-factor map identifies who is at risk (route-a prognostics — Coley confirms age, APOE ε4, inactivity, low MMSE raise incidence), but it does not license a claim that intervening harder on the high-risk converts to prevented dementia — that route-b step is exactly what the pooled trial search could not find. Coley is the pooled COMPOSITE of the two RCTs already cited here (shared authorship, FINGER excluded), so it is corroboration of the null by re-analysis, not an independent third witness -> Multidomain Lifestyle Intervention and Cognitive Decline holds the full responder verdict. On the surrogate the same orbiter now also ranks the levers (Mendes 2025 NMA, 109 RCTs): exercise + cognitive training is the top combination and the fuller multidomain bundle does not improve on it — more is not better — a within-surrogate hierarchy that does not convert to the incidence benefit the pooled null denies. (Mendes et al., 2025)

Self-critique `[run 2026-08-05, before commit]`

Not laundered / not overclaimed. The headline 45.3% is presented with all three of the source’s own limiting frames (population-not-individual, borrowed-RR-assumed-causal, communality-discounted-so-no-sum) rather than as a per-person prevention promise — the exact misread the maintainer framing flagged.
Now two-source (F-refinement added 2026-08-05). The Garcia-Casares MedDiet -> AD/MCI dose-response quantifies + bounds the diet lever the Commission references; it is folded in with its own confound frame (observational-only, healthy-user, reverse causation) and the mediation/double-counting flag, NOT as a 15th additive PAF. No independence (E) is claimed between the two sources. confidence: medium still reflects the Commission’s consensus-grade backing; the diet association is explicitly the weaker, low-to-moderate leg. Where a held page already estimates an exposure (smoking, LDL, BP, activity, alcohol), the dementia bullet is added there rather than re-litigated here.
RCT leg added (FINGER, F-refinement 2026-08-05). The primary FINGER trial was folded in via a parameter table whose “same quantity?” column is NO on every row — so it is filed as a bounding / refinement of the observational PAF, not a tension (the not-joined guard: an RCT between-group difference on a cognitive composite is not the same quantity as an observational PAF on dementia incidence). No independence (E) is claimed. The reading is deliberately symmetric — FINGER neither confirms nor overturns the 45.3% figure; it de-risks the direction while the per-factor decomposition and the incidence endpoint stay unproven.
Cognitive-stimulation de-secondhanding added (F-refinement 2026-08-14). The Commission’s borrowed 0.79 is replaced by its first-hand IPD source (Kivimaki 2021), with the reverse-causation-lag origin of that number made explicit and the headline HR 0.77 restored. Filed as F, not E: shared authorship (Livingston) and the Commission’s reliance on this evidence defeat independence, flagged at the . The full estimate/mechanism/ranking is hosted on the sibling page, not re-litigated here.
Omega-3 candidate lever added (2026-08-23, Wei 2023). Filed as an observational candidate lever, NOT a 15th factor — parallel to the periodontal and diet levers, and held at confidence: low. Not overclaimed: the dietary-DHA RR 0.82 is stated observational/FFQ/Level-H-per-Wei, and the supplement arm is explicitly refuted at RCT grade (Wei’s own «limited efﬁcacy» concession + the held MAPT null), so the section cannot be misread as licensing omega-3 capsules. Not laundered-E: Wei shares the observational cohort literature with the diet lever and the mechanism with the cardiometabolic rocks, so it is flagged as largely mediated, not additive. No tension filed. The full decomposition lives on Fish and Seafood Consumption, not re-litigated here.
Air-pollution de-secondhanding added (F-refinement 2026-08-27, Wilker 2023). The Commission’s borrowed PM2.5 HR 1.03 (1.02-1.05) per 1 ug/m3 — CI excluding 1 — is de-secondhanded to its first-hand BMJ SR+MA, whose overall pool per 2 ug/m3 is 1.04 (0.99-1.09), CI crossing the null (a per-1-ug rescale ~1.02 (0.99-1.04), marked). Filed as F, not E: Wilker cites Livingston 2020 and shares the underlying observational cohort literature, so no independence is claimed. Honesty guards held: the significant 1.42 is stated as an active-case-ascertainment moderator (a detection-method artifact, explicitly NOT a route-(b) person-level stratum), not headlined; the ROBINS-E bias-toward-null frame keeps the null-crossing CI from reading as a no-effect verdict; the parameter table’s “Same quantity?” column is NO on every row, so it is a bounding/refinement, never a tension. Layer-1: air pollution is stated as a weakly-modifiable structural/policy lever that does not reorder the personal rocks. No confidence: change — the page stays medium on the Commission’s consensus backing; the air-pollution slice is, if anything, weaker first-hand than the borrowed cell implied.
F&V candidate lever added (2026-09-04, Zhou 2022). Filed as a second FOOD-GROUP dietary lever (protective direction, sibling to the UPF harm lever), an observational candidate NOT a 15th factor, held at confidence: low. Not overclaimed: the pooled OR 0.82 is stated observational/FFQ, the abstract’s 0.79 is corrected to the Figure-2 pool, and the AD-null + cohort-attenuation + OR-as-RR discounts are surfaced so it cannot read as a direct anti-dementia effect. Not laundered-E: Zhou shares the observational cohort literature and the vascular mechanism with the other diet levers, so it is flagged mediated-not-additive. No tension filed. Full outcome arm on Fruit and Vegetable Intake and Health, not re-litigated here.
Flavonoid component-side lever added (2026-09-04, Peng 2025). Filed as the COMPONENT side of the F&V lever (nested inside it), an observational candidate NOT a 15th factor, held at confidence: low. Not overclaimed: the any-event OR 0.90 is stated observational/FFQ, and the dementia-null + AD-null + BMI/smoking-confounding discounts are surfaced so it cannot read as a direct anti-dementia effect. Not laundered-E: the flavonoid exposure is COMPUTED from the same fruit/veg/tea as the Zhou F&V lever (nested exposure) and shares the observational substrate, so Peng and Zhou are explicitly a component/ food-group type-F pair, NOT independent corroboration — no [E-independent] minted, and this was the ingest’s crux check. No tension filed. Full subclass/dose-response/food-vs-component detail on Flavonoid Intake and Cognitive Function, not re-litigated here.
Physical-activity de-secondhanding added (F-refinement 2026-09-04, Iso-Markku 2022). The Commission’s borrowed activity cell (RR 0.80, 0.77-0.84) is de-secondhanded to its first-hand SR+MA, whose headline is the identical figure (n=257,983) — so filed as F (an identity-plus-enrichment), not E: Iso-Markku is Livingston’s ref 165 and the Commission rests on this evidence base, flagged at the. Not overclaimed: the reverse-causation survival (>=20y RR 0.79) is stated with the honesty guard that the cleanest young-baseline high-quality subset goes non-significant (0.79, 0.62-1.01) and with the residual cognitive-reserve confound the design cannot remove — so the section cannot read as a demonstrated causal slope. The ApoE ε4 route-b null and the work-related-PA opposite trend (RR 1.25, CI crosses 1) are stated at their own (weak) strength. The parameter table’s headline row is Same quantity? YES, which is the de-secondhanding claim itself (not a tension). The unheld Kivimäki 2019 PA-IPD null it rebuts is named as a G-gap with a registered [AWAITS] placeholder, not asserted as fact. No confidence: change — the page stays medium on the Commission’s consensus backing; PA was already a counted big rock, so this firms it rather than adding a factor. Full extraction on Physical Activity Dose and Mortality’s sibling source page, not re-litigated here.
Sleep candidate lever added (2026-09-04, Zhang 2025). Filed as the first non-diet/non-activity candidate lever, an observational candidate NOT a 15th factor, held at confidence: low. Not overclaimed: every RR is stated observational with its I2, publication bias is disclosed, and the long-sleep -> dementia arm is explicitly flagged a likely preclinical marker (age-dependent, APOE-driven, weak-checks-only) so it cannot read as sleep less. The objective-insomnia strengthening is scoped to insomnia. Not laundered-E: single source, no independence claimed; the VD-loaded disorders are flagged mediated-not-additive via the cerebrovascular route already counted. No tension filed against the unheld Xu 2020 landmark — a G-gap with a registered [AWAITS] handle. Full decomposition on Sleep and Cognitive Decline, not re-litigated here.
Dairy candidate lever added (2026-09-05, Villoz 2024). Filed as a food-group diet lever (sibling to UPF/F&V), an observational candidate NOT a 15th factor, held at confidence: low — and flagged the weakest diet lever, since its headline highest-vs-lowest contrast is null (RR 0.94, 0.82-1.07) and its one dose-response feature (the «nadir at ~150 g/d») is read as a between-population pooling artifact (Asian low-intake cohorts vs European high-intake cohorts, non-overlapping ranges) rather than an optimum. Not overclaimed: no protective claim survives, the Asia RR 0.83 is marked confounded, and the dietary-pattern-adjustment null is used to mark it mediated-not-additive. Not laundered-E: single source, no independence claimed, and the NOT-E caveat for a future 2nd source (Modena author group) is stated. No tension filed. Full extraction on the dairy-cluster facet Dairy and Cognitive Decline, not re-litigated here.
Evidence-strength axis added (F/C-advance 2026-09-08, Zhang 2022 umbrella). Filed as a type-C new axis (per-factor credibility class I-IV, orthogonal to PAF magnitude) plus a type-F bounding of each factor, NOT a 15th factor and NOT a rival PAF. The parameter table’s “Same quantity?” column is NO on every row — Zhang’s credibility class and Livingston’s PAF measure different things — so it is a distinction, never a tension; Zhang neither confirms nor overturns the 45.3% headline, it reads perpendicular to it. Not laundered-E: Zhang is the Qingdao/Fudan group (Yu Jin-Tai), shares Xu, Wei with the held omega-3 source and cites Livingston 2017/2020, so no [E-independent] is minted — F against Livingston, no E against the group siblings; this was the ingest’s crux NOT-E check. Not overclaimed: the whole tiering is held at the umbrella’s own stated LOW study-quality caveat (AMSTAR-2), the credibility class is stated to grade the evidence not the world, and the new convincing factors (gait speed, hypoglycemia) are flagged prodromal/reverse-causal, not actionable levers. The MR causal layer is reported as a genetic check that education + plasma glucose clear, not as a lifestyle warrant. confidence: medium unchanged — the page still rests on the Commission’s consensus backing; Zhang tempers the lifestyle levers toward suggestive rather than adding weight.
PA-dementia counter-source held, tension filed (type-D, 2026-09-10, Kivimaki 2019). The [AWAITS] G-gap on the physical-activity de-secondhanding is cashed: Kivimaki’s IPD (n approx 404,840) is now held first-hand and joins the issue Iso-Markku rebutted, filed as Does Physical Activity Protect Against Dementia Beyond Ten Years. Not overclaimed toward the null: the IPD’s design advantages (within-cohort 10 y split, I2=0%, cardiometabolic positive controls validating the method, both-ascertainment and >60-subgroup nulls) are stated as tilting the weighing, while the honest near-convergence (Iso-Markku’s own cleanest young-baseline >20 y cut is NS at 0.79, 0.62-1.01) is surfaced, so the verdict is insufficient evidence for a direct dementia-specific effect, not proven no effect. Not laundered-E: the two sources DISAGREE (a D, not an E), and Kivimaki’s positive controls corroborate only the cardiometabolic causation. Layer-1 unchanged: activity stays a non-substitutable big rock via the vascular route both sources share; only the direct-neuroprotection warrant weakens. confidence: medium held — the page still rests on the Commission’s consensus backing.
Coherence, not validity (R1): the loop is open — no operation here grades the PAF against a realized dementia outcome. A clean audit of this page is not a validated prevention claim.

References

Blood Pressure Lowering Treatment Trialists Collaboration. (2021). Pharmacological blood pressure lowering for primary and secondary prevention of cardiovascular disease across different levels of blood pressure: an individual participant-level data meta-analysis. The Lancet, 397(10285), 1625–1636. https://doi.org/10.1016/s0140-6736(21)00590-0

Coley, N., Hoevenaar‐Blom, M. P., Shourick, J., van Charante, E. P. M., van Dalen, J., van Gool, W. A., Richard, E., & Andrieu, S. (2025). Searching for responders to multidomain dementia prevention in late life: A pooled analysis of individual participant data from the MAPT and preDIVA trials. Alzheimer’s & Dementia, 21(2). https://doi.org/10.1002/alz.14472

Dibello, V., Custodero, C., Cavalcanti, R., Lafornara, D., Dibello, A., Lozupone, M., Daniele, A., Pilotto, A., Panza, F., & Solfrizzi, V. (2024). Impact of periodontal disease on cognitive disorders, dementia, and depression: a systematic review and meta-analysis. GeroScience, 46(5), 5133–5169. https://doi.org/10.1007/s11357-024-01243-8

Dinu, M., Pagliai, G., Casini, A., & Sofi, F. (2017). Mediterranean diet and multiple health outcomes: an umbrella review of meta-analyses of observational studies and randomised trials. European Journal of Clinical Nutrition, 72(1), 30–43. https://doi.org/10.1038/ejcn.2017.58

García-Casares, N., Gallego Fuentes, P., Barbancho, M. Á., López-Gigosos, R., García-Rodríguez, A., & Gutiérrez-Bedmar, M. (2021). Alzheimer’s Disease, Mild Cognitive Impairment and Mediterranean Diet. A Systematic Review and Dose-Response Meta-Analysis. Journal of Clinical Medicine, 10(20), 4642. https://doi.org/10.3390/jcm10204642

Henney, A. E., Gillespie, C. S., Alam, U., Hydes, T. J., Mackay, C. E., & Cuthbertson, D. J. (2023). High intake of ultra-processed food is associated with dementia in adults: a systematic review and meta-analysis of observational studies. Journal of Neurology, 271(1), 198–210. https://doi.org/10.1007/s00415-023-12033-1

Iso-Markku, P., Kujala, U. M., Knittle, K., Polet, J., Vuoksimaa, E., & Waller, K. (2022). Physical activity as a protective factor for dementia and Alzheimer’s disease: systematic review, meta-analysis and quality assessment of cohort and case–control studies. British Journal of Sports Medicine, 56(12), 701–709. https://doi.org/10.1136/bjsports-2021-104981

Kivimäki, M., Singh-Manoux, A., Pentti, J., Sabia, S., Nyberg, S. T., Alfredsson, L., Goldberg, M., Knutsson, A., Koskenvuo, M., Koskinen, A., Kouvonen, A., Nordin, M., Oksanen, T., Strandberg, T., Suominen, S. B., Theorell, T., Vahtera, J., Väänänen, A., Virtanen, M., … Jokela, M. (2019). Physical inactivity, cardiometabolic disease, and risk of dementia: an individual-participant meta-analysis. BMJ, l1495. https://doi.org/10.1136/bmj.l1495

Kivimäki, M., Walker, K. A., Pentti, J., Nyberg, S. T., Mars, N., Vahtera, J., Suominen, S. B., Lallukka, T., Rahkonen, O., Pietiläinen, O., Koskinen, A., Väänänen, A., Kalsi, J. K., Goldberg, M., Zins, M., Alfredsson, L., Westerholm, P. J. M., Knutsson, A., Theorell, T., … Lindbohm, J. V. (2021). Cognitive stimulation in the workplace, plasma proteins, and risk of dementia: three analyses of population cohort studies. BMJ, n1804. https://doi.org/10.1136/bmj.n1804

Kuate Defo, A., Bakula, V., Pisaturo, A., Labos, C., Wing, S. S., & Daskalopoulou, S. S. (2023). Diabetes, antidiabetic medications and risk of dementia: A systematic umbrella review and meta‐analysis. Diabetes, Obesity and Metabolism, 26(2), 441–462. https://doi.org/10.1111/dom.15331

Livingston, G., Huntley, J., Liu, K. Y., Costafreda, S. G., Selbæk, G., Alladi, S., Ames, D., Banerjee, S., Burns, A., Brayne, C., Fox, N. C., Ferri, C. P., Gitlin, L. N., Howard, R., Kales, H. C., Kivimäki, M., Larson, E. B., Nakasujja, N., Rockwood, K., … Mukadam, N. (2024). Dementia prevention, intervention, and care: 2024 report of the Lancet standing Commission. The Lancet, 404(10452), 572–628. https://doi.org/10.1016/s0140-6736(24)01296-0

Mendes, A. J., Ribaldi, F., Sayin, O., Khachvani, G., Mulargia, R., Volpara, G., Remoli, G., Nencha, U., Gianonni-Luza, S., Cappa, S., & Frisoni, G. B. (2025). Single-domain and multidomain lifestyle interventions for the prevention of cognitive decline in older adults who are cognitively unimpaired: a systematic review and network meta-analysis. The Lancet Healthy Longevity, 6(9), 100762. https://doi.org/10.1016/j.lanhl.2025.100762

Ngandu, T., Lehtisalo, J., Solomon, A., Levälahti, E., Ahtiluoto, S., Antikainen, R., Bäckman, L., Hänninen, T., Jula, A., Laatikainen, T., Lindström, J., Mangialasche, F., Paajanen, T., Pajala, S., Peltonen, M., Rauramaa, R., Stigsdotter-Neely, A., Strandberg, T., Tuomilehto, J., … Kivipelto, M. (2015). A 2 year multidomain intervention of diet, exercise, cognitive training, and vascular risk monitoring versus control to prevent cognitive decline in at-risk elderly people (FINGER): a randomised controlled trial. The Lancet, 385(9984), 2255–2263. https://doi.org/10.1016/s0140-6736(15)60461-5

Nucci, D., Sommariva, A., Degoni, L. M., Gallo, G., Mancarella, M., Natarelli, F., Savoia, A., Catalini, A., Ferranti, R., Pregliasco, F. E., Castaldi, S., & Gianfredi, V. (2024). Association between Mediterranean diet and dementia and Alzheimer disease: a systematic review with meta-analysis. Aging Clinical and Experimental Research, 36(1). https://doi.org/10.1007/s40520-024-02718-6

Peng, Y., Zou, Q., Geng, T., Wang, P., Li, D., Chen, X., Zhang, Z., Wang, F., Xu, X., Sun, L., Gao, X., & Li, Y. (2026). Dietary flavonoids consumption and cognitive function: a systematic review and meta-analysis of observational studies. Food Science and Human Wellness, 15(6), 9250596. https://doi.org/10.26599/fshw.2025.9250596

Peters, R., Xu, Y., Fitzgerald, O., Aung, H. L., Beckett, N., Bulpitt, C., Chalmers, J., Forette, F., Gong, J., Harris, K., Humburg, P., Matthews, F. E., Staessen, J. A., Thijs, L., Tzourio, C., Warwick, J., Woodward, M., & Anderson, C. S. (2022). Blood pressure lowering and prevention of dementia: an individual patient data meta-analysis. European Heart Journal, 43(48), 4980–4990. https://doi.org/10.1093/eurheartj/ehac584

Smith, M., Watson, P., Gallacher, J., & Bauermeister, S. (2026). Ultra-processed food exposure and cognitive outcomes: a systematic review of observational studies. BMJ Nutrition, Prevention & Health, 9(1), 211–219. https://doi.org/10.1136/bmjnph-2025-001325

van Charante, E. P. M., Richard, E., Eurelings, L. S., van Dalen, J.-W., Ligthart, S. A., van Bussel, E. F., Hoevenaar-Blom, M. P., Vermeulen, M., & van Gool, W. A. (2016). Effectiveness of a 6-year multidomain vascular care intervention to prevent dementia (preDIVA): a cluster-randomised controlled trial. The Lancet, 388(10046), 797–805. https://doi.org/10.1016/s0140-6736(16)30950-3

Villoz, F., Filippini, T., Ortega, N., Kopp-Heim, D., Voortman, T., Blum, M. R., Del Giovane, C., Vinceti, M., Rodondi, N., & Chocano-Bedoya, P. O. (2024). Dairy Intake and Risk of Cognitive Decline and Dementia: A Systematic Review and Dose-Response Meta-Analysis of Prospective Studies. Advances in Nutrition, 15(1), 100160. https://doi.org/10.1016/j.advnut.2023.100160

Wei, B.-Z., Li, L., Dong, C.-W., Tan, C.-C., & Xu, W. (2023). The Relationship of Omega-3 Fatty Acids with Dementia and Cognitive Decline: Evidence from Prospective Cohort Studies of Supplementation, Dietary Intake, and Blood Markers. The American Journal of Clinical Nutrition, 117(6), 1096–1109. https://doi.org/10.1016/j.ajcnut.2023.04.001

Wilker, E. H., Osman, M., & Weisskopf, M. G. (2023). Ambient air pollution and clinical dementia: systematic review and meta-analysis. BMJ, 381, e071620. https://doi.org/10.1136/bmj-2022-071620

Xu, W., Tan, C.-C., Zou, J.-J., Cao, X.-P., & Tan, L. (2019). Sleep problems and risk of all-cause cognitive decline or dementia: an updated systematic review and meta-analysis. Journal of Neurology, Neurosurgery & Psychiatry, 91(3), 236–244. https://doi.org/10.1136/jnnp-2019-321896

Zhang, J., Ou, J., Lu, X., Wang, T., Dang, W., Ding, L., Liu, Y., Xu, J., Yan, B., & Yu, H. (2025). Sleep disorders and the risk of cognitive decline or dementia: an updated systematic review and meta-analysis of longitudinal studies. Journal of Neurology, 272(10). https://doi.org/10.1007/s00415-025-13372-x

Zhang, Y.-R., Xu, W., Zhang, W., Wang, H.-F., Ou, Y.-N., Qu, Y., Shen, X.-N., Chen, S.-D., Wu, K.-M., Zhao, Q.-H., Zhang, H.-N., Sun, L., Dong, Q., Tan, L., Feng, L., Zhang, C., Evangelou, E., Smith, A. D., & Yu, J.-T. (2022). Modifiable risk factors for incident dementia and cognitive impairment: An umbrella review of evidence. Journal of Affective Disorders, 314, 160–167. https://doi.org/10.1016/j.jad.2022.07.008

Zhou, Y., Wang, J., Cao, L., Shi, M., Liu, H., Zhao, Y., & Xia, Y. (2022). Fruit and Vegetable Consumption and Cognitive Disorders in Older Adults: A Meta-Analysis of Observational Studies. Frontiers in Nutrition, 9. https://doi.org/10.3389/fnut.2022.871061

Evidenced Health

Explorer

Dementia Prevention and Modifiable Risk Factors

The headline, and how to read it honestly

The 14 factors, their direction, and their life-course window

The evidence-STRENGTH axis, orthogonal to PAF — Zhang’s credibility tiering (F/C-advance, 2026-09-08)

The two axes are different quantities — a distinction, NOT a tension

What the strength axis tempers, and what it raises

The cognitive-stimulation number, de-secondhanded (F-refinement, 2026-08-14)

The air-pollution number, de-secondhanded — and it is weaker first-hand (F-refinement, 2026-08-27)

The physical-activity number, de-secondhanded (F-refinement, 2026-09-04)

The hypertension number, de-secondhanded — the INTERVENTIONAL effect (F-refinement, 2026-09-04)

What the framework changes about what to do

The diet lever — quantified, observational, and probably NOT additive to the cardiometabolic rocks

Nucci 2024 — the same association on a larger elderly pool, with new caveats (F-refinement, 2026-08-20)

Periodontal disease — an observational candidate lever, NOT one of the 14 (2026-08-23)

Omega-3 fatty acids / oily fish — an observational candidate lever, NOT one of the 14 (2026-08-23)

Ultra-processed food — the first FOOD-GROUP dietary lever, observational candidate, NOT one of the 14 (2026-09-03)

Fruit and vegetables — a second FOOD-GROUP dietary lever, observational candidate, NOT one of the 14 (2026-09-04)

Flavonoids — the COMPONENT side of the food-group levers, observational candidate, NOT one of the 14 (2026-09-04)

Dairy — a candidate FOOD-GROUP lever whose only signal is a pooling artifact, NOT one of the 14 (2026-09-05)

Sleep disorders — a non-diet candidate lever, NOT one of the 14 (2026-09-04)

The RCT leg — FINGER bounds the observational PAF (F-refinement, 2026-08-05)

The diabetes lever splits by DRUG CLASS — the pharmacotherapy channel opened (2026-09-04)

Scope guard

Self-critique `[run 2026-08-05, before commit]`

References

Graph View

Backlinks

Evidenced Health

Explorer

Dementia Prevention and Modifiable Risk Factors

The headline, and how to read it honestly

The 14 factors, their direction, and their life-course window

The evidence-STRENGTH axis, orthogonal to PAF — Zhang’s credibility tiering (F/C-advance, 2026-09-08)

The two axes are different quantities — a distinction, NOT a tension

What the strength axis tempers, and what it raises

The cognitive-stimulation number, de-secondhanded (F-refinement, 2026-08-14)

The air-pollution number, de-secondhanded — and it is weaker first-hand (F-refinement, 2026-08-27)

The physical-activity number, de-secondhanded (F-refinement, 2026-09-04)

The hypertension number, de-secondhanded — the INTERVENTIONAL effect (F-refinement, 2026-09-04)

What the framework changes about what to do

The diet lever — quantified, observational, and probably NOT additive to the cardiometabolic rocks

Nucci 2024 — the same association on a larger elderly pool, with new caveats (F-refinement, 2026-08-20)

Periodontal disease — an observational candidate lever, NOT one of the 14 (2026-08-23)

Omega-3 fatty acids / oily fish — an observational candidate lever, NOT one of the 14 (2026-08-23)

Ultra-processed food — the first FOOD-GROUP dietary lever, observational candidate, NOT one of the 14 (2026-09-03)

Fruit and vegetables — a second FOOD-GROUP dietary lever, observational candidate, NOT one of the 14 (2026-09-04)

Flavonoids — the COMPONENT side of the food-group levers, observational candidate, NOT one of the 14 (2026-09-04)

Dairy — a candidate FOOD-GROUP lever whose only signal is a pooling artifact, NOT one of the 14 (2026-09-05)

Sleep disorders — a non-diet candidate lever, NOT one of the 14 (2026-09-04)

The RCT leg — FINGER bounds the observational PAF (F-refinement, 2026-08-05)

The diabetes lever splits by DRUG CLASS — the pharmacotherapy channel opened (2026-09-04)

Scope guard

Self-critique [run 2026-08-05, before commit]

References

Graph View

Backlinks

Self-critique `[run 2026-08-05, before commit]`