Dairy and Cognitive Decline

Facet of the dairy cluster (nucleus Dairy and Cardiometabolic Health) — the cognition/dementia endpoint, distinct from the nucleus’s cardiometabolic outcomes and the fracture endpoint on Dairy and Bone Health. It is also the dairy-side facet of the dementia map Dementia Prevention and Modifiable Risk Factors, where dairy sits as one more observational candidate lever, not one of the Commission’s 14 factors.

The through-line (one gold source, single-witness): Villoz 2024 is the first and only held dose-response SR+MA on dairy -> cognition — 15 prospective cohorts, 312,580 participants, median 11.4 y follow-up, no RCTs, all observational. Its headline highest-vs-lowest contrast is null (RR 0.94, 0.82-1.07), and the one apparent signal — a nonlinear dose-response with a «nadir at ~150 g/d» — is almost certainly a between-population pooling artifact: the descending arm is low-intake Asian cohorts, the flat arm is high-intake European cohorts, and the «nadir» sits exactly at the boundary between the two intake ranges. Every effect estimate is FFQ-confounded and vanishes in the studies that adjust for overall dietary pattern. Decision-change: for the reasonably-healthy Western person, dairy is not a cognition lever to worry about or optimize — the European cohorts show no association, and even the Asian signal does not survive the confounding checks. confidence: low.

The effect estimates — the headline is NULL; the signal lives in subgroups and curve shape

Villoz reports the combined outcome (cognitive decline OR incident dementia) three ways, and the three disagree in a diagnostic pattern:

• Highest-vs-lowest intake: NULL. «the highest intake of dairy products compared to the lowest intake has no association with cognitive decline or dementia (RR: 0.94; 95% CI: 0.82, 1.07)» with high heterogeneity (I2 69.2%) and a wide prediction interval (0.61-1.45) (Villoz et al., 2024). A prediction interval spanning 0.61-1.45 means a future cohort could as easily find harm as benefit — the pooled point estimate carries no decision.
• Cognitive decline alone: NULL (RR 1.01, 0.86-1.20, I2 73.5%). Incident dementia alone: borderline (RR 0.83, 0.67-1.03, I2 63.0%, PI 0.44-1.59) — crosses 1, so an insufficient-evidence lean, not an established protective effect. (Villoz et al., 2024)
• Dose-response, quantity (g/d): a nonlinear inverse with a soft turn at ~150 g/d. «nonlinear association, with an initial decline in risk until 150 g/d (RR: 0.88; 95% CI: 0.78, 0.99), after which a slight change in direction was observed» (Villoz et al., 2024) — 6 studies, 3 Asian + 3 European.
• Dose-response, frequency (times/d): almost linear inverse. «an almost linear negative association when we considered the frequency of consumption (RR for linear trend 0.84; 95% CI: 0.77, 0.92 for 1 time/d increase of dairy products)» (Villoz et al., 2024) — 5 studies.

The shape is measure-specific (nonlinear on grams, linear on frequency) — a tell that the curve is a property of how the exposure was coded across heterogeneous cohorts, not a stable biological dose-response. Two different summaries of the same underlying data give two different shapes.

Why the ~150 g/d nadir is a between-population pooling artifact, not a curve feature

This is the page’s beyond-summary move, and it reads the source against its own headline. Villoz’s quantity dose-response pools 3 Asian and 3 European cohorts whose intake ranges barely overlap:

• The region split IS the signal. Highest-vs-lowest is protective in Asia (RR 0.83, 0.75-0.92, I2 0.0%) and null in Europe (RR 1.01, 0.86-1.19, I2 41.6%); the single Oceania study (high risk of bias) even runs the other way (RR 1.75, 1.17-2.62) (Villoz et al., 2024).
• The regions occupy different intake ranges. «The amount and types of dairy consumption between regions were considerably higher in studies carried out in European countries, with mean value between 170-711 g/d, than studies in Asian countries where total mean dairy intake ranged between 29-165 g/d» (Villoz et al., 2024).
• So the nadir sits on the seam. The descending arm below ~150 g/d is drawn almost entirely by the Asian cohorts (whose whole intake distribution lies below 165 g/d and who show the protective RR 0.83); the flat/rising arm above ~150 g/d is drawn by the European cohorts (170-711 g/d, null). The «nadir at 150 g/d» is therefore the boundary between two populations sampled at non-overlapping doses, not a within-person optimum. Region is confounded with intake range and with everything else that differs between Asian and European cohorts — background diet, dairy-type mix, dietary-pattern context — so the apparent knee carries the whole between-population confound.
• This is a distinct U-shape-artifact mechanism -> The U-Shaped Association Artifact. Not reverse-causation and not sick-quitter: the apparent nonlinearity is manufactured by pooling cohorts whose exposure supports do not overlap, so the spline is stitched across populations rather than estimated within one. It is the same failure as the Seidelmann carbohydrate curve (two cohorts tracing complementary arms of one U), sharpened: here the two arms are also two different populations with two different confounding structures. The «studied range» caveat binds hard — no single cohort spans the full 29-711 g/d, so the curve’s shape at any point is an extrapolation across the seam -> The Underivable Optimum.

Dairy is a type-B category here too — milk and cheese alone are NULL

As on the cardiometabolic nucleus, dairy is not one exposure -> Is the Food Category Doing Any Work. When Villoz stratifies by product, the whole-group inverse dissolves:

• All-dairy-combined is inverse and homogeneous (RR 0.89, 0.83-0.95, I2 0.33%), but milk alone (5 studies) and cheese alone (4 studies) are both null — «the association with specific dairy products was very heterogeneous and inconsistent … with the exception of milk and cheese intake alone, investigated in 5 and 4 studies and reporting both null associations, respectively» (Villoz et al., 2024). Milk’s per-frequency curve is linear-inverse only at high intakes; cheese’s is nonlinear (lower risk at 0.3 times/d, null/positive above). A milk-only or cheese-only eater gets no cognition signal from this data.
• Fat content could not be stratified — Villoz flags full-fat vs low-fat as unresolved, with two cohorts hinting full-fat/whole-cream milk tracks worse cognition, but no pooled contrast. The dairy-desserts sub-signal (higher cognitive decline) is confounded with sugar and is excluded from most dairy guidance. Fermented-vs-unfermented, the axis that carries the only cardiometabolic signal, could not be tested either -> Fermented Foods and Health.

Why the whole signal is discounted — three confounds, all named in-source

• The association vanishes on dietary-pattern adjustment. «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). This is the decisive discount: dairy intake reads as a marker of a broader dietary pattern, and when the pattern is controlled the dairy effect disappears — the observed-healthy-population problem. The plausible mechanistic route (dairy -> cardiometabolic risk -> cognition) also runs through levers the dementia map already counts, so any real effect is largely mediated-not-additive, not a clean independent lever.
• Differential measurement error / reverse causation over the long prodrome. «we cannot discard differential measurement error because of the recall bias, as early symptomatology of cognitive decline could have affected the way people report their diet or their dietary choices» (Villoz et al., 2024) — early cognitive change bends reported intake, manufacturing an association -> The U-Shaped Association Artifact.
• Single-measurement FFQ exposure. Most cohorts measured diet once, at baseline, so the exposure is a snapshot, not a life-course trajectory, and dietary measurement error flattens real gradients while a snapshot cannot capture change -> Measurement Error in Dietary Assessment. Villoz calls for dairy-intake biomarkers to replace recall.

Quality was mostly middling (1 poor, 10 neutral, 4 good on NUQUEST); the funnel plot was symmetric (no small-study bias detected); excluding the one high-RoB cohort strengthened the overall to RR 0.90 (0.82-1.00). The authors declare no conflicts of interest and the work is publicly funded (Swiss SNF + Italian Ministry).

Confidence, decision-relevance, and gaps

• confidence: low — one gold-design source, but every estimate is observational/FFQ, the headline contrast is null, the one dose-response signal is a region-pooling artifact, the product-specific cells (milk, cheese) are null, and the association vanishes on dietary-pattern adjustment. A gold design does not buy a strong finding when the finding is a null wrapped around a confounded curve.
• Decision-change (per stratum): for a reasonably-healthy adult in a Western intake range, dairy is not a cognition/dementia lever — the European cohorts show no association and there is nothing to optimize. The ceiling is the finding: reporting no lever here licenses someone to stop worrying about dairy for brain health. For a low-intake (e.g. Asian-range) adult the data are consistent with a small inverse association, but it is confounded and not adjudicated, so it does not support a “drink more milk to prevent dementia” recommendation either. Attention is an anti-signal applies — dairy-and-brain is discussed far out of proportion to this null.
• NOT independent-E for any future second source (registry-flagged). Villoz is a Modena-group dose-response-MA product (Filippini, Vinceti — the same lineage behind the held greenness/fluoride/ acrylamide dose-response units), and it re-pools the canonical dairy-cognition cohorts. A second source that would raise confidence by independent backing must come from outside this author group.
• Gaps (G):
  • An independent (non-Modena) dairy -> cognition SR/MA to test the region-pooling reading -> a dairy-cognition SR/MA from outside the Villoz/Filippini/Vinceti author group.
  • Fat-content and fermented-vs-unfermented stratification on cognition — Villoz could not run either; the fermented axis is where the cardiometabolic signal lives.
  • Within-population wide-intake-range cohorts (or a dairy-intake biomarker) that would estimate the dose-response inside one population instead of across the Asia/Europe seam.
  • G (needs aggregation): a pooled dairy-cognition dose-response that adjusts for background dietary pattern across cohorts — the magnitude that would separate a dairy effect from a diet-quality marker, which the fabric cannot compute from this single non-stratified source.
  • The high-risk (diabetic/hypertensive) stratum is excluded by design — Villoz restricted to reasonably-healthy adults for primary prevention, so this page does not speak to dairy-cognition among people with the chronic conditions that dominate Western populations.

Self-critique [run 2026-09-05, before commit]

• No dairy halo, no dairy scare. The verdict is stated as null / no lever for the default stratum, not as dairy protects the brain (the abstract’s framing) nor dairy harms cognition. The one protective number (Asia RR 0.83) is presented as confounded and unadjudicated, not headlined.
• Overclaim check. The «nadir at 150 g/d» is never presented as an optimum or a target — it is read as an artifact of the Asia/Europe intake seam, with the studied-range caveat carried. No causal language attaches to any RR; every figure is prefaced observational/FFQ. The dementia-alone RR 0.83 is marked crosses-1 / insufficient-evidence, not protective.
• Not laundered-E. Single source; no independence claimed. The NOT-E caveat for any future second source (Modena author group) is stated explicitly, matching the registry note.
• Region-pooling reading grounded, not imported. The artifact reading is built from Villoz’s own region strata and stated intake ranges (both quoted), then mapped onto the U-shape concept’s non-overlapping-support mechanism — an synthesis flagged as such, not attributed to Villoz.
• Mediated-not-additive flag carried. The dietary-pattern-adjustment null and the cardiometabolic mechanism route are used to mark the lever as largely mediated by already-counted factors, consistent with how the dementia nucleus treats every diet lever.
• Open loop (R1). No operation here grades the dairy-cognition association against a realized dementia outcome; a clean audit is not a validated prevention claim.

References

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