Synthesis across the four single-food-component cognition arms the wiki holds — dairy, flavonoids, soy isoflavones, and fruit/vegetables (the whole-food group that carries the first two). Each is appraised in full on its own page; the Dementia Prevention and Modifiable Risk Factors nucleus states, arm by arm, that each is observational, low-certainty, and mediated-not-additive. This page makes the move no single arm page and no single nucleus section makes: it reads the four together and finds one regularity across them, with a decision consequence the per-arm verdicts leave implicit.
The regularity — the signal lives on soft endpoints and vanishes at the hard diagnosis
Line the arms up by endpoint and the same shape appears in every one: a modest protective signal on a soft or composite or surrogate endpoint (a cognitive-impairment/decline score, a neuropsychological test), and null or unestablished evidence on the hard diagnosis — Alzheimer’s disease, or an incident-dementia event.
| Arm (source) | Soft / composite / surrogate endpoint | Hard endpoint (AD / incident dementia) |
|---|---|---|
| Fruit & vegetables (Zhou 2022) | cognitive impairment OR 0.76 (0.72-0.80); any cognitive disorder 0.82 (0.75-0.90) (Zhou et al., 2022) | dementia 0.84 (0.78-0.91) but AD null 0.88 (0.76-1.01) (Zhou et al., 2022) |
| Flavonoids (Peng 2025) | cognitive decline OR 0.88 (0.79-0.98); any adverse event 0.90 (0.83-0.98) (Peng et al., 2026) | dementia null 0.97 (0.79-1.19); AD null 0.90 (0.69-1.17) (Peng et al., 2026) |
| Soy isoflavones (Cui 2020, RCT) | test-score SMD 0.19 (0.07-0.32), memory-carried, over <=2 y (Cui et al., 2019) | no hard-endpoint arm exists — the trials measure test performance, not dementia/decline events |
| Dairy (Villoz 2024) | cognitive decline null RR 1.01 (0.86-1.20) (Villoz et al., 2024) | highest-vs-lowest RR 0.94 (0.82-1.07) null; dementia-alone 0.83 (0.67-1.03) crosses 1 (Villoz et al., 2024) |
Read down the right-hand column: not one arm shows an established benefit on the hard Alzheimer’s or incident-dementia diagnosis. Three arms (F&V, flavonoids, soy) put a signal on a softer endpoint and lose it at the hard one; dairy shows no signal at any endpoint; soy has no hard endpoint at all (its RCT evidence is surrogate-only). The endpoints within each arm share that arm’s design and population, so the soft-vs-hard contrast is a within-arm comparison, not a cross-study one.
Why the gradient is itself the finding
A component-specific neuroprotectant should register on the hard endpoint at least as clearly as on a soft proxy — the diagnosis is the outcome the surrogate is standing in for. Its systematic disappearance at AD/dementia, with the signal surviving only where the endpoint is softer, measured with more error, and more open to reverse causation over the long dementia prodrome, is the signature of a shared bias structure rather than a component effect:
- The same confounds recur in every arm — healthy-user selection, reverse causation over the prodrome, and FFQ-based dietary measurement error (doubly so for flavonoids, computed from food reports through a composition database) -> Measurement Error in Dietary Assessment. Peng’s own meta-regression finds BMI and smoking «potentially overestimating the positive effects if not adjusted for» (Peng et al., 2026).
- The design gradient points the same way. In the F&V pool the reverse-causation-vulnerable designs (cross-sectional 0.70, case-control 0.68) give the strongest effect and the prospective cohort the weakest (0.83) (Zhou et al., 2022) — the association shrinks as the design gets cleaner, the tell the soft-endpoint signal is partly artefactual -> The U-Shaped Association Artifact.
- The arms are nested and mediated, not independent. Flavonoids are a component of the F&V group; both plausibly act through the vascular/cardiometabolic route the Commission already counts as hypertension, diabetes, obesity and LDL. So the arms do not stack — with each other or on top of the 14 factors -> Is the Food Category Doing Any Work, Dementia Prevention and Modifiable Risk Factors.
- The gradient is not component-specific — it recurs at the whole-pattern altitude. The MIND pattern (a Mediterranean-DASH hybrid) shows the same soft-vs-hard shape as the four component arms: pooled observational cognitive function is protectively associated (+0.042 per SD, 0.020-0.065) (Huang et al., 2023), but the harder longitudinal decline signal is non-significant and collapses to null (0.0032, -0.0010-0.0075) once the Morris cohort is removed (Huang et al., 2023), and the one randomized test of the pattern returns a between-group null (Barnes et al., 2023) -> MIND Diet - Observational Benefit vs Randomized Null. So the soft-endpoint-only signal is a property of the diet-cognition observational base as a whole — pattern and component alike — not an artifact of slicing the diet into single components; the shared confound structure operates at both altitudes -> MIND Diet and Cognitive Decline.
A competing reading is simple power: hard-diagnosis events are rarer, so their confidence intervals are wider and cross the null even where a small real effect exists (Peng reads its own dementia/AD nulls exactly this way). What tips the balance toward bias rather than pure power is the design gradient above — power does not explain why the reverse-causation-prone designs give the largest effects, whereas confounding and reverse causation do. So the reading is bias-inflated soft signal more than merely underpowered hard signal, though the two are not exclusive.
The soft-endpoint signals are therefore best read as insufficient-evidence-tilting-null on the outcome that matters, not as small confirmed benefits — the surrogate’s causal transmission to the hard diagnosis is the unmet condition, not a technicality -> Surrogate Outcomes.
The decision consequence — a Layer-1 ceiling, not a menu
The single-food-component route offers no established hard-endpoint dementia-prevention lever. That is a ceiling finding, and reporting it is itself a decision-change: it licenses not optimizing here.
- Do not chase individual foods, and especially not isolated-component supplements (flavonoid extracts, isoflavone pills), for cognition. The one randomized signal (soy) is on a short-term test surrogate in a single domain, with no dementia-event evidence — the weakest warrant for a pill.
- The diet lever’s cognitive value is as a route to the whole-diet pattern and the cardiometabolic big rocks, which do carry hard-outcome evidence, not as a stack of additive component benefits -> Layer 1 - Ranking Interventions for a Stratum. A person eating dairy and berries and soy and more vegetables for compounding brain protection is over-counting overlapping, confounded, soft-endpoint signals.
- This is a candidate-lever gap, not a refutation of diet. A hard-endpoint benefit is
unestablished, not disproven; each arm is low-certainty observational (soy aside), and a future
component with a biomarker handle, an MR arm, or a hard-endpoint RCT could still separate a real
effect from the shared bias. Until one lands, the component approach stays
confidence: low.
Provenance and independence
This is a type-A emergent synthesis (the cross-arm regularity is stated on no single page) resting
on a type-F relationship among the arms: the four are not independent witnesses (flavonoids
nested in F&V; all sharing the observational/FFQ substrate), so their agreement is not
[E-independent] corroboration — it is the same confounded signal seen four times, which is precisely
what makes the shared hard-endpoint null informative rather than reassuring.