Diet Quality Scores and Cardiovascular Risk

Nucleus of the diet-quality-scores cluster — the canonical owner of what a whole-diet quality score does to CVD and mortality, as distinct from any single food (owned by the food pages) or any branded weight-loss programme (Named Diet Programs Compared, a surrogate/weight question).

The through-line (one large observational source, PURE + 5 replication studies, ~245,000 people, 80 countries): a simple unweighted count of six protective foods — fruit, vegetables, nuts, legumes, fish, and (mainly whole-fat) dairy, range 0-6 — is associated with lower CVD and mortality, consistently across income levels and world regions. The effect is modest per unit and larger at the extremes, the dose-response flattens above a moderate score (most gain is in moving a low consumer up, not a moderate consumer to high), and the authors read the global pattern as under-consumption of protective foods mattering more than over-consumption — a contrarian policy frame that the study’s confounding structure and its partial dairy-industry funding both require holding at arm’s length. Everything here is observational; the headline associations attenuate substantially once lifestyle and socioeconomic status are adjusted, and no hard-outcome RCT exists.

The score and the magnitudes

The score was developed in 147,642 PURE participants (21 countries) and replicated in five independent studies, «six foods each of which has been associated with a significantly lower risk of mortality [i.e. fruit, vegetables, nuts, legumes, fish, and dairy (mainly whole-fat); range of scores, 0-6]» (Mente et al., 2023). Each food scores 1 point above its cohort median intake — an unweighted sum with median cut-offs.

Extreme contrast (PURE, score >=5 vs <=1, median 9.3-y follow-up, multivariable-adjusted):

• Mortality HR 0.70 (0.63-0.77); major CVD 0.82 (0.75-0.91); MI 0.86 (0.75-0.99); stroke 0.81 (0.71-0.93) (Mente et al., 2023).
• Per one quintile (20 percentile) increment: «a 20% (1 quintile) higher PURE healthy diet score was associated with a 6% lower risk of major CVD, and 8% lower risk of mortality» (Mente et al., 2023) — mortality HR 0.91 (0.89-0.93), major CVD 0.94 (0.92-0.97), composite 0.93 (0.92-0.95) (Mente et al., 2023).

Replication holds, same direction, similar size. In three prospective vascular-patient cohorts (ONTARGET/TRANSCEND/ORIGIN): mortality HR 0.73 (0.66-0.81), CVD 0.79 (0.72-0.87); in two case-control studies (INTERHEART/INTERSTROKE): first MI OR 0.72 (0.65-0.80), stroke OR 0.57 (0.50-0.65) (Mente et al., 2023). The consistency across designs and continents is the source’s genuine strength.

But discrimination is weak — a population-level gradient, not an individual predictor. The score’s area under the ROC curve is only 0.61 (mortality), 0.54 (major CVD), 0.58 (composite) (Mente et al., 2023) — barely above chance for CVD. So the score orders populations on risk; it does not tell an individual their fate. A per-quintile 6-8% relative shift on a modest baseline is a small absolute effect for most people (Baseline Risk and the Relative-Absolute Split).

The dose-response knee — the lever is raising LOW consumers to moderate

The decision-relevant shape: the association is not linear, and it is steeper below the median. «the associations of diet scores with events are markedly steeper at lower levels of the diet score (i.e. below the global median)» (Mente et al., 2023), and the restricted-cubic-spline curve reaches diminishing returns around a score of 4 (of 6): score 4 is «the level at which most major gains in health are observed and beyond which there is modest additional health gain» (Mente et al., 2023). Studied range: 0-6; the knee sits mid-range (~3-4), not at the sampling edge.

This is a LOCATED knee in a diet-pattern -> mortality/CVD curve — a data point bearing on the telos’s [PRIOR — CONTESTED] knees-and-plateaus question (lodged here, not adjudicated — that is a separate operation). Two caveats keep it honest, and they are not small:

• The x-axis is an ordinal composite (0-6) built on median cut-offs, not a natural intake unit. A knee on a 7-point integer score assembled from six dichotomized foods is a far weaker object than a knee on grams/day of one food — the median-threshold scoring itself compresses the top of the range.
• FFQ measurement error flattens curves (Measurement Error in Dietary Assessment), so an apparent plateau above moderate intake is partly what the instrument cannot resolve, not certainly a biological ceiling. The authors concede this cuts the other way for magnitude: «variations in reporting might lead to random errors that could dilute real associations between diet scores and clinical outcomes. Therefore, the beneficial effects of a healthier diet may be larger than what we have estimated» (Mente et al., 2023).

Decision consequence (holds regardless of the epistemic caveat): for someone already at a moderate score, pushing the diet score higher buys little; the large remediable gap is in the low-score stratum. This is the ceiling-is-a-finding rule instantiated — and it is why the source’s own emphasis lands on low consumers, not on optimizing an already-adequate diet. -> The Underivable Optimum

The regional “effect modification” is largely position on the curve — and the under-nutrition reframe

The associations are «significantly stronger in South Asia, China, and Africa, regions where the PURE diet score is low» (mean scores 2.1 / 2.6 / 3.1) (Mente et al., 2023). Read naively this is route-(b) effect modification by region. Read against the knee above, it is mostly route-(a)/curve-position: low-consuming populations sit on the steep part of the dose-response curve, so the same score increment moves their risk more — no biological region-interaction is required to explain it. (inferred from Mente et al., 2023)

From this the authors draw their headline policy claim: «an inadequate level of consumption of key healthy foods is a larger problem than over-consumption of some nutrients or foods» (Mente et al., 2023). Two guards: (i) it is an interpretation of an observational gradient, not an experimental result; (ii) it is stratum-dependent — it describes low-consuming (largely lower-income) populations, and does not transport to a calorie-replete population whose binding constraint is over-nutrition (the era/stratum-dependence the telos flags for the appeal-to-nature trade-off). For the reasonably-healthy, adequately-fed default reader, the big protective-food rocks are largely already pulled. The direction-of-guidance question this raises -> Which Objective Moved This Recommendation.

Which score is best? Barely matters — except the Planetary score, which was null

PURE compared its score head-to-head with five others on its own participants (Table 4). «The PURE score appears to be slightly more predictive of composite events than the Mediterranean, HEI-2010, HEI-2015, and DASH diet scores and substantially more predictive than the Planetary diet score» (Mente et al., 2023). On mortality: PURE HR 0.91 / AUC 0.61; Mediterranean 0.96 / 0.55; HEI-2010 0.93 / 0.55; DASH 0.96 / 0.53; Planetary (EAT-Lancet) HR 1.00 (0.98-1.03) / AUC 0.52 — null (Mente et al., 2023).

• The “PURE wins” margin is small and self-serving by construction — every score was tested on PURE’s own FFQ, which captures PURE’s components best and can disadvantage scores built on foods PURE measured coarsely. Read the near-equivalence of Mediterranean/HEI/DASH as the signal and PURE’s slight edge as method-favouring noise. This echoes Named Diet Programs Compared: across well-constructed diet metrics the label barely carries the effect.
• The Planetary-score null is the one real divergence, and the authors attribute it to its being «the most restrictive» — it limits animal foods (dairy, and to a degree fish/legumes) that carry PURE’s signal. Whether that is a real demerit of the EAT-Lancet pattern or an artifact of scoring it on a population where those foods are protective is unresolved here (its non-health / environmental objective is a different axis this wiki does not price — name it, do not net it).

No single component carries it — and the score is reachable many ways

Removing or swapping components barely moved the score’s predictive value: each of the six contributes similarly and modestly, and the two most-contested optional foods are neutral additions — «when we included red meat in the diet score in a sensitivity analysis, the findings were similar (neither stronger nor weaker)», the same for whole grains (Mente et al., 2023). So a moderate score «can be achieved in a number of ways which does not necessarily require either including or excluding animal foods from the diet» (Mente et al., 2023) — a vegetarian reaches it via fruit/veg/legumes/nuts/dairy, an omnivore via any of dairy/fish/meat.

This is the inverse of the usual food-category trap (Is the Food Category Doing Any Work): there the question is whether a broad label hides a load-bearing sub-boundary; here no single sub-component dominates, so the aggregate pattern is the right unit and substituting within it is near-neutral. The whole-fat dairy inclusion is the score’s distinctive, contrarian element (other scores favour low-fat or exclude dairy) -> Dairy and Cardiometabolic Health, where PURE’s dairy signal is weighed against the neutral pooled-cohort evidence and its funding caveat.

Appraisal — healthy-user confounding is the headline caveat

The estimates carry the standard observational-nutrition liability, and PURE shows it in its own numbers:

• The signal roughly halves under adjustment. Mortality for the top vs bottom score group is HR 0.54 (0.50-0.59) age/sex-adjusted, attenuating to 0.70 (0.63-0.77) on full multivariable adjustment (composite: 0.64 -> 0.78) (Mente et al., 2023). A third of the crude association was confounding by lifestyle/SES that the model could measure — which is the signature that residual confounding by the factors it could not measure runs the same direction (toward the null of the causal effect).

• How big can the healthy-user confounder alone be? A causally-clean upper reference — the placebo arm [2026-08-29]. Simpson 2006’s placebo arms (eight blinded trials, 19,633 people) isolate the pure adherer bundle: good-vs-poor adherence to an inert pill carried «lower mortality (0.56, 0.43 to 0.74)» (Simpson et al., 2006) with no causal path — the entire gap is the health-conscious-person bundle (diet, exercise, screening, follow-up). That reference magnitude sits almost exactly on PURE’s crude diet-score signal, and the two are near-commensurable — both are largely-unadjusted behaviour -> all-cause-mortality associations indexing the same bundle:

  Parameter	Simpson placebo arm	PURE diet-score (crude)	Same quantity?
  Outcome	all-cause mortality	all-cause mortality	yes
  Adjustment	unadjusted (bundle intact by design)	age/sex only — no lifestyle/SES	~yes — both largely unadjusted
  Contrast	good vs poor pill-adherence	top vs bottom protective-food score	~no — different behaviour, same health-conscious bundle
  Estimate	OR 0.56 (0.43-0.74)	HR 0.54 (0.50-0.59)	near-identical magnitude

  So PURE’s crude mortality signal is the size of a bundle that Simpson proves has zero causal input. This does not say PURE’s adjusted 0.70 is all confounding — adjustment removes the measurable part, and Simpson’s estimate is the unadjusted ceiling, not the residual — but it sets the reference: a diet -> mortality association at or below OR ~0.56 is inside what the adherer bundle can manufacture unaided, so the ~30% residual after PURE’s adjustment cannot be assumed causal on magnitude alone. -> The Observational-Trial Discordance

• The authors concede exactly this, and tie it to the effect size: «the possibility of residual confounding from unmeasured or imprecise measurement of covariates cannot be completely ruled out—especially given that the differences in risk of clinical events is modest (~10%-20% relative differences)» (Mente et al., 2023). A 10-20% relative difference is well within what residual confounding can manufacture.

• Two biases point opposite ways, so the net direction is genuinely uncertain: residual confounding inflates the protective estimate; FFQ measurement error dilutes it (the dilution quote above). The honest reading is not “the true effect is larger” (the authors’ framing) nor “it is all confounding” — it is that the point estimate sits between two large, opposing, unquantified biases.

• Funding tell (symmetric standards). PURE is partly funded by «an unrestricted grant from Dairy Farmers of Canada and the National Dairy Council (U.S.)» (Mente et al., 2023), and the PURE group holds a known heterodox, dairy-and-SFA-favourable stance. This does not delete the finding, but the whole-fat-dairy-is-protective and SFA-limits-may-not-be- warranted interpretations are exactly where a directional sponsor would push, so they are discounted, not adopted. The score predicts CVD/mortality finding is robust to this; the policy reframe is not.

Decision relevance, confidence, and gaps

• confidence: medium — a large, geographically diverse, internally-replicated observational signal in a consistent direction, but single-source (PURE + its own replication cohorts are not independent — shared authors and PURE infrastructure with Dairy and Cardiometabolic Health’s Dehghan PURE macronutrient paper; a shared confounding or measurement structure moves them together), attenuating under adjustment, discrimination weak, and with no hard-outcome RCT — «large randomized trials are essential to definitively clarify the clinical impact» (Mente et al., 2023).
• Decision-change (per stratum): for a low consumer of protective foods (the study’s own steep-curve stratum), moving toward a moderate score (~4/6) is the high-value action and the one the evidence most supports; for someone already at a moderate/good score, the marginal gain from pushing higher is small by construction (the knee) — a licence to stop optimizing, not a failure to find a lever. Which specific foods get there is close to interchangeable, so adherence and access decide it, not the brand of score.
• Gaps (G):
  • No RCT of adopting a diet-pattern score on hard events — the structural gap the source names; a decades-long whole-diet trial is impractical, so this stays a named absence.
  • G (needs aggregation): a confounding-adjusted, cross-cohort pooled diet-score -> CVD-events magnitude independent of PURE — a number the fabric cannot compute from one source family.
  • The knee’s reality vs the ordinal-score artifact — whether the flattening above score ~4 is a biological plateau or an artifact of median-cut composite scoring + FFQ error is unresolved and would need a natural-unit intake analysis to settle.

Why WHI does not fill the hard-outcome RCT gap — a distinction, not a tension

— it is tempting to read the largest diet RCT (WHI DM Trial, n=48,835, null on CHD/stroke/ CVD) as the missing randomized test of a protective-food pattern. It is not: the two test different exposures, so WHI’s null neither fills the gap above nor speaks against this page’s association. Parameter table (the not-joined check (ii): different exposure/unit, so a distinction, not a filed tension) -> Low-Fat Dietary Pattern and Cardiovascular Disease:

Parameter	WHI DM Trial	PURE diet-quality score	Same quantity?
Exposure the arm defines	total-fat quantity (target 20%E)	count of 6 protective foods (0-6)	NO — a macronutrient % vs a protective-food pattern
Displaced/added	fat calories removed, replaced by carbohydrate (grains)	protective foods added	NO — opposite operation
Fat quality moved?	no specific SFA->unsaturated focus	not a fat-quality lever at all	n/a
Design	RCT, hard events	observational cohort	different by construction

So WHI is a null RCT of fat-quantity-to-carbohydrate, and the diet-quality-score signal is about adding protective foods — the WHI authors say their trial «is not a test of the dietary guidelines cur- rently recommended for prevention of CVD» (the plant-based, protective-food pattern) (Howard et al., 2006). The “no hard-outcome RCT” gap therefore stands — WHI does not close it.

Self-critique [run 2026-08-25, before commit]

• No diet-score halo. The score-predicts finding is stated with its weak discrimination (AUC ~0.52- 0.61), its ~50% attenuation under adjustment, and the authors’ own residual-confounding concession up front — not buried. The contrarian dairy/SFA policy reframe is quarantined from the core finding and tagged with the dairy-industry funding tell.
• Knee lodged, not adjudicated. The located knee is recorded as a data point for the [PRIOR — CONTESTED], explicitly not scored against the telos here (that is a separate operation), and double-caveated (ordinal composite + FFQ flattening) so it is not oversold as a clean natural-unit knee.
• Independence not laundered. PURE’s five replication studies are marked shared-infrastructure (not type-E independent backing), and the shared-author overlap with the Dehghan PURE paper is stated — no [E-independent] claimed anywhere.
• Effect-modification handled with the right route. The regional heterogeneity is read as curve-position (route-a/arithmetic) rather than asserted as biological region-interaction (route-b), matching the false-positive discipline; the reframe is scoped to the low-consuming stratum, not generalized to the replete reader.
• Counter-passage check ran on the two cross-source echoes (Named Diets “label barely matters”; Dairy neutrality) — both are cross-links to pages that own those verdicts, not tensions filed here; PURE’s dairy reading is flagged as needing the Dairy page’s weighing, not asserted as settled.
• Both biases named. Residual confounding (inflates) and measurement error (dilutes) are stated as opposing, so the page does not adopt the authors’ “true effect is larger” spin.
• WHI distinction is joined-checked, not a fake tension. The parameter table shows WHI tested a different exposure (fat-quantity->carb) than the protective-food score, so it is filed as a distinction (not-joined check (ii)) that preserves the RCT gap, not a tension resolving it; the WHI quote is bin/cite.py-verified and not truncated past its qualifier.

References

Howard, B. V., Van Horn, L., Hsia, J., Manson, J. E., Stefanick, M. L., Wassertheil-Smoller, S., Kuller, L. H., LaCroix, A. Z., Langer, R. D., Lasser, N. L., Lewis, C. E., Limacher, M. C., Margolis, K. L., Mysiw, W. J., Ockene, J. K., Parker, L. M., Perri, M. G., Phillips, L., Prentice, R. L., … Kotchen, J. M. (2006). Low-Fat Dietary Pattern and Risk of Cardiovascular Disease: The Women’s Health Initiative Randomized Controlled Dietary Modification Trial. JAMA, 295(6), 655. https://doi.org/10.1001/jama.295.6.655

Mente, A., Dehghan, M., Rangarajan, S., O’Donnell, M., Hu, W., Dagenais, G., Wielgosz, A., A. Lear, S., Wei, L., Diaz, R., Avezum, A., Lopez-Jaramillo, P., Lanas, F., Swaminathan, S., Kaur, M., Vijayakumar, K., Mohan, V., Gupta, R., Szuba, A., … Yusuf, S. (2023). Diet, cardiovascular disease, and mortality in 80 countries. European Heart Journal, 44(28), 2560–2579. https://doi.org/10.1093/eurheartj/ehad269

Simpson, S. H., Eurich, D. T., Majumdar, S. R., Padwal, R. S., Tsuyuki, R. T., Varney, J., & Johnson, J. A. (2006). A meta-analysis of the association between adherence to drug therapy and mortality. BMJ, 333(7557), 15. https://doi.org/10.1136/bmj.38875.675486.55