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 endpointHard 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.

References

Barnes, L. L., Dhana, K., Liu, X., Carey, V. J., Ventrelle, J., Johnson, K., Hollings, C. S., Bishop, L., Laranjo, N., Stubbs, B. J., Reilly, X., Agarwal, P., Zhang, S., Grodstein, F., Tangney, C. C., Holland, T. M., Aggarwal, N. T., Arfanakis, K., Morris, M. C., & Sacks, F. M. (2023). Trial of the MIND Diet for Prevention of Cognitive Decline in Older Persons. New England Journal of Medicine, 389(7), 602–611. https://doi.org/10.1056/nejmoa2302368
Cui, C., Birru, R. L., Snitz, B. E., Ihara, M., Kakuta, C., Lopresti, B. J., Aizenstein, H. J., Lopez, O. L., Mathis, C. A., Miyamoto, Y., Kuller, L. H., & Sekikawa, A. (2019). Effects of soy isoflavones on cognitive function: a systematic review and meta-analysis of randomized controlled trials. Nutrition Reviews, 78(2), 134–144. https://doi.org/10.1093/nutrit/nuz050
Huang, L., Tao, Y., Chen, H., Chen, X., Shen, J., Zhao, C., Xu, X., He, M., Zhu, D., Zhang, R., Yang, M., Zheng, Y., & Yuan, C. (2023). Mediterranean-Dietary Approaches to Stop Hypertension Intervention for Neurodegenerative Delay (MIND) Diet and Cognitive Function and its Decline: A Prospective Study and Meta-analysis of Cohort Studies. The American Journal of Clinical Nutrition, 118(1), 174–182. https://doi.org/10.1016/j.ajcnut.2023.04.025
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
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
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