Decision one-liner (Layer 1): higher dietary flavonoid intake carries a small, borderline inverse association with total and CVD mortality on observational data — and the one meta-analysis the wiki holds cannot tell you the flavonoid itself is what does the work. Its exposure is flavonoid intake estimated from self-reported food consumption (FFQ), so a flavonoid result is the same observational, measurement-error-laden dietary-pattern signal as fruit/veg or tea, re-expressed in flavonoid units. For a reasonably-healthy person this is not a big rock and not a reason to chase flavonoids as such: eat the flavonoid-bearing whole foods (fruit, veg, tea) for the several better- evidenced reasons already on their pages, not for a flavonoid target this evidence cannot isolate.
Effect estimate (Mazidi 2020, MA of 16 cohorts, highest vs lowest intake)
| Outcome | RR (95% CI) | p | n studies | Significant? |
|---|---|---|---|---|
| All-cause mortality | 0.87 (0.77, 0.99) | 0.039 | 9 | yes — but upper CI a whisker from 1.0 |
| CVD mortality | 0.85 (0.75, 0.97) | 0.017 | 15 | yes |
| Cancer mortality | 0.86 (0.65, 1.14) | 0.302 | 4 | no |
- Relative-only, no absolute layer. Highest-vs-lowest category contrast; the paper reports no dose-response curve and no baseline event rates, so no events-avoided-per-1000 figure and no knee / plateau can be stated. A flavonoid target is not derivable from this.
- Discordant denominators: the all-cause estimate pools only 9 of the 16 studies, the CVD estimate 15 — the two headline numbers do not rest on the same study set.
- Cancer null on 4 studies (insufficient, not established no-effect).
- Robustness, the reassuring side: findings «remained robust in sensitivity analyses» (leave-one- out); Egger p=0.170 (NS), trim-and- fill imputed no missing studies, fail-safe N=20; and — unlike the tea MA — «This research received no external funding» (Mazidi et al., 2020), so no beverage-seller COI applies.
Why confidence is LOW despite a gold-tier design
- Heterogeneity asserted, not shown. The paper claims «a very low level of heterogeneity … highlighting the validity of our results» in the same sentence it concedes «some misclassification of flavonoid consumption is inevitable» (Mazidi et al., 2020) — yet reports no I2 anywhere, while the forest-plot study RRs range from ~0.38 to ~1.40 (one leverage outlier, Ivey 2015, near 0.38). A low-heterogeneity claim without the statistic, over a visibly wide spread, is not credited.
- Borderline result. All-cause upper CI = 0.99, on 9 studies.
- FFQ-estimated exposure, measurement error read optimistically. The authors argue misclassification is non-differential, so the pooled RR «could have been underestimated rather than overestimated» (Mazidi et al., 2020). The telos holds that with mismeasured covariates the bias can run in either direction, so the attenuation-to-null reading is not established -> Measurement Error in Dietary Assessment.
- Residual confounding conceded. Macronutrient adjustment (fiber, protein, fat, carbohydrate) «attenuated our results but there was still an inverse link» (Mazidi et al., 2020), and the authors cannot rule out «unmeasured or residual confounding» and «cannot be certain that the inverse association … is casual [causal], representing the effect of flavonoids only» (Mazidi et al., 2020).
The load-bearing move: a flavonoid MA does NOT isolate the flavonoid
The intuitive reading — «Mazidi measured the flavonoid, so it confirms it is the flavonoids, not just the tea/fruit» — is wrong, and is the value this page adds. Mazidi’s exposure is flavonoid intake estimated from FFQ food reports, and flavonoids are «commonly present in vegetables, fruits, herbs and teas» (Mazidi et al., 2020). So the flavonoid intake variable is computed from the very foods whose contribution it would need to be separated from — it is a re-expression of fruit/veg/tea intake in flavonoid units, carrying the same confounding and measurement-error substrate, not an independent handle on the component. No study used a flavonoid biomarker, an isolated-flavonoid trial arm, or Mendelian randomization — the designs that could isolate the component. This is a Test-3 collinearity case exactly like coffee/caffeine and wine/resveratrol: a study crediting flavonoids may be measuring the food (and the healthy pattern it marks) under a flavonoid label -> Is the Food Category Doing Any Work.
Parameter table — Mazidi (flavonoid) vs Chung (tea): same substrate, NOT independent backing
Built to test whether Mazidi corroborates tea’s CVD signal (type-E) or merely re-expresses it. No row on the effect itself is the same quantity, and the exposure OVERLAPS (tea is inside the flavonoid intake), so this is not independent backing:
| Parameter | Mazidi 2020 (flavonoid) | Chung 2020 (tea) | Same quantity? |
|---|---|---|---|
| Exposure | dietary flavonoid intake, FFQ-estimated | habitual tea, FFQ/diet-recall | overlapping — tea is a flavonoid source inside Mazidi’s exposure |
| Exposure ascertainment | self-report (11 FFQ / 4 interview / 1 record); no biomarker | self-report; no biomarker | Yes — shared substrate |
| Contrast / scale | highest-vs-lowest category RR | per-cup linear RR | No — different unit + scale |
| Effect size | all-cause 0.87, CVD 0.85 | all-cause 0.98, CVD 0.96 per cup | No — not comparable |
| Isolates the component? | no (component computed from foods) | no (no decaf/isolate/MR) | Yes — neither does |
| Causal disconfirmation held | none (no MR) | none (no MR, no decaf) | Yes — neither holds one |
(inferred from Mazidi et al., 2020)
Verdict: type A/G, not E and not D. Because the exposures overlap and share the measurement-error + confounding substrate, Mazidi’s inverse flavonoid association and Chung’s inverse tea association are not independent routes to one claim (E fails: volume is not independence, and a shared exposure is a stronger defeater than a shared method). Nor is there a matched-quantity clash to file as a tension (D): the two estimates are on different units and do not contradict. What they jointly establish is a gap: neither design can say whether the mortality signal attached to tea / fruit-veg / flavonoids is the flavonoid, the food matrix, or the healthy-eating pattern. That is the food-vs-component G-gap, now with a second FFQ-derived source confirming it stays open rather than closing it.
The isolating design Mazidi lacked now exists — and its hard-outcome primary is null [2026-08-31, Sesso/COSMOS]
Mazidi’s whole limitation is that its flavonoid exposure is FFQ-derived, so «no study used a flavonoid biomarker, an isolated-flavonoid trial arm, or Mendelian randomization». COSMOS (Sesso 2022) supplies the middle one for a food-borne flavanol: a randomized placebo-controlled cocoa-flavanol EXTRACT on hard CV endpoints. Its primary composite (total CVD events) was null: HR 0.90 (0.78, 1.02; P=0.11) (Sesso et al., 2022), with delivery confirmed by a
3-fold rise in the flavanol biomarker gVLM (ratio 3.23; 2.84, 3.67) (Sesso et al., 2022).
This is NOT the same quantity as Mazidi’s association — build the table before any comparison. The reflex is to weld COSMOS’s 0.90 to Mazidi’s 0.87 as «the RCT confirms/refutes the cohort». They are different exposures, designs, and endpoints; no numeric weld is made:
| Parameter | Mazidi 2020 (flavonoid) | COSMOS / Sesso 2022 (cocoa flavanol) | Same quantity? |
|---|---|---|---|
| Exposure | dietary flavonoid intake, FFQ-estimated | cocoa-extract supplement (500 mg flavanols + 80 mg epicatechin), a whole-bean extract | No — food-derived estimate vs a dosed pill |
| Isolation | none (computed from foods) | randomized vs placebo, but «cannot disentangle» components of the extract | No — partial physical isolation vs none |
| Design | observational MA (16 cohorts) | double-blind RCT (n=21,442) | No |
| Endpoint | all-cause / CVD / cancer mortality | total CVD events (composite); mortality secondary | No — mortality vs CV-event composite |
| Effect | all-cause 0.87 (0.77-0.99); CVD 0.85 (0.75-0.97) | primary 0.90 (0.78, 1.02), null; all-cause 0.89 (0.77, 1.03) | not comparable |
| Dose vs habitual | within FFQ dietary range | ~5x the European dietary flavanol mean («exceeds the mean intake reported in Europe of 105 mg/d») | No |
(Mazidi et al., 2020; inferred from Sesso et al., 2022) Dose figure: (Sesso et al., 2022).
What it does establish (type F/A, decision-form). The one design that physically removes a food-borne flavanol from the FFQ-collinear food signal shrank the flavonoid-mortality association toward the null on a hard CV primary — the direction the food-vs-component/healthy-user critique predicts. It does not prove the flavonoid is inert: COSMOS tested a cocoa extract (not the same molecule set as diet-wide flavonoids), on CV events (not all-cause mortality), at a supra-dietary dose, and «cannot disentangle» its components (Sesso et al., 2022); the authors themselves name «Residual confounding … limits observational studies examining flavanols or chocolate and CVD risk» (Sesso et al., 2022). So the combined reading: Mazidi’s borderline observational signal + COSMOS’s randomized primary null jointly weaken the case for a flavonoid-as-agent lever without a clean isolation of the molecule — reinforcing this page’s LOW confidence and its Layer-3 read (eat the foods, do not chase a flavonoid target or supplement). Effect estimates + the secondary-endpoint caveat: Vitamin and Mineral Supplements for Disease Prevention; the Test-3 collinearity framing: Is the Food Category Doing Any Work. (inferred from Sesso et al., 2022)
The cognition sibling (2026-09-04)
The same food-vs-component logic runs on a second outcome axis: a gold SR+MA of flavonoid intake -> cognitive function (Peng 2025) finds a small inverse association with adverse cognitive events driven by cognitive decline, with dementia and Alzheimer’s both null, on the same FFQ-derived exposure that cannot isolate the molecule -> Flavonoid Intake and Cognitive Function (which holds the figures). It confirms the food-vs-component G-gap stays open on cognition as it does here on mortality.
Where this sits
- Layer 1: low rank. A borderline observational category-contrast, null in the best-powered check it most needs (a component-isolating design, which does not exist here). Not a lever to prioritize.
- Layer 3: if you want the flavonoid-associated benefit, the realistic implementation is eat the whole flavonoid-bearing foods (fruit, veg, tea) — whose own pages carry better-evidenced, overlapping rationales — not a flavonoid supplement or target, which this evidence does not support. Mazidi’s own «recommendations for flavonoid-rich foods intake» (Mazidi et al., 2020) reaches past what the design isolates.
- Confidence: low — single MA, observational, borderline all-cause estimate, heterogeneity claimed without I2, FFQ exposure that cannot separate component from food -> Upgrading Observational Evidence.