— a synthesis induced across the fabric’s own holdings (UPF, coffee, Mediterranean diet, fibre). Every per-instance figure is established and verified on the linked claim page (with its source locus there); this page makes the structural claim across them, which is in no single source.
A recurring shape in nutrition evidence: a large, consistent OBSERVATIONAL signal — often graded convincing on a credibility scale — sits beside a randomized trial that is null, weak, or absent. The reflex is to let the RCT settle it (the design hierarchy). That reflex is wrong as a default, and the reason is the single most useful thing this pattern teaches: the two evidence streams usually are not estimating the same exposure. The discordance is a diagnostic signal to check exposure-commensurability, not a verdict for the trial.
The pattern is real — three worked instances
— the values below are as established on each linked page; the cross-instance structure is this page’s.
| Instance | Observational signal | The randomized / causal test | How it resolves |
|---|---|---|---|
| Ultra-processed food | CVD-mortality association graded class-I convincing on the credibility scale | GRADE very low for that outcome; Hall’s inpatient RCT confirms UPF causes overeating (the mechanism) — but it is the WHOLE direct-diet RCT base (Aramburu’s RCT-only SR: 4 trials, n=455, 30/42 null, all high-RoB) | on the intake surrogate the trial tested the right exposure and AGREES; on hard outcomes the RCT base is too thin to confirm or refute — INSUFFICIENT, not resolved. Hard-outcome RCTs are ethically ruled out -> Ultra-Processed Food and Health Outcomes |
| Coffee | all-cause mortality RR ~0.83, CVD ~0.85 — consistent across a large umbrella | mostly GRADE very-low; Mendelian-randomization finds no genetic causal signal for coffee->T2D (secondhand, power not stated) or mortality (held, not powered to exclude the observational size); no lifetime RCT is feasible | the natural experiment (MR) finds no support for one arm -> that arm is plausibly confounded; hold as insufficient-for-causation, not confirmed and not disconfirmed -> Coffee Consumption and Health, Upgrading Observational Evidence |
| Mediterranean diet | CVD/mortality graded convincing observationally | pooled whole-diet RCTs mostly null except diabetes; the one whole-PATTERN RCT (PREDIMED) moved events, while single-nutrient RCTs (the 54-RCT SFA-events null, Look AHEAD) did not | the RCT that tested the RIGHT exposure (the whole pattern) AGREES; the “null RCTs” tested a DIFFERENT exposure (single nutrients) -> Mediterranean Diet and Cardiovascular Events |
Sugar / SSB / fructose [2026-08-06] | large monotone SSB dose-response cohort signal (T2DM RR 1.19 per 250 mL/d) + higher fructose in NAFLD cases | isocaloric fructose-for-glucose exchange null on liver fat, and isoenergetic sugar exchange null on weight (0.04 kg) | the trials tested the commensurable exposure — the sugar molecule at equal energy — and found null; so the cohort signal is the added-energy package (SSB adds poorly-compensated liquid calories), confounded by energy. Harm tracks the calories, not the molecule -> Free Sugars Intake, What Drives Fat Gain - Energy Balance vs the Carbohydrate-Insulin Model |
Total dietary fat [2026-08-25; Hooper 2012 pooled 2026-09-04] | diet-heart-era observational/ecological signal that total-fat intake raises CHD | WHI DM Trial (n=48,835) null (CHD HR 0.97, stroke 1.02, CVD 0.98) — and the pooled RCT class agrees: Hooper 2012’s fat-reduction subgroup (50,655 pp, containing WHI) is RR 0.97 (0.87-1.08) on CV events, while its fat-modification subgroup moved events (RR 0.82/0.83) | the RCTs that tested the reduction exposure (total-fat cut, fat->carbohydrate) are null as a class, not just in WHI; the RCTs that tested the right exposure (fat modification, SFA->unsaturated) AGREE with the causal lever. Same wrong-exposure resolution, now generalized from one trial to a moderate-GRADE pooled RCT base — WHI under a diluted ~70%-of-design contrast, ~40% power, is the extreme case of it -> Low-Fat Dietary Pattern and Cardiovascular Disease |
Hearing aids -> cognition [2026-08-28; ACHIEVE first-hand 2026-09-04] | 8-cohort pool HR 0.81 (0.76-0.87), I2=0% — a 19% lower cognitive-decline hazard among hearing-aid users vs uncorrected hearing loss (Yeo 2023) | ACHIEVE RCT (N=977, first-hand): null overall on the continuous 3-year cognition slope (diff 0.002 SD [-0.077 to 0.081], p=0.96); a pre-specified 48% subgroup reduction only in the higher-risk ARIC arm (pinteraction=0.010, lenient alpha<0.10) — Lin/ACHIEVE first-hand, loci on Hearing Loss and Dementia | healthy-user self-selection in the intervention arm (mechanism #3) erases the average effect; a signal survives only where absolute baseline risk is high (ARIC). But match quantities first: the overall null is on a continuous slope Yeo never estimated, and on the commensurable incidence-HR quantity ACHIEVE (0.90 [0.61-1.33]) includes Yeo’s 0.81/0.83 — so the RCT is underpowered there, not contradicting. The subgroup interaction is route-(b)-form but hypothesis-generating (contamination: de-novo control drop-in 19.4% vs 7.8%). A confounded-observational + baseline-risk + partly-non-commensurable resolution -> Hearing Loss and Dementia |
Red meat -> T2D [2026-08-29] (the UNTESTED pole; superseded in part [2026-10-06] — one weak MR now held, see Update below) | NutriGrade “high” per-100 g association, RR 1.17 (1.08-1.26) — robust, consistent, dose-responsive across prospective cohorts (Schwingshackl 2017) | none held in either direction — no Mendelian-randomization and no feeding trial isolates red-meat (or heme-iron) -> T2D; the coded exposure is a decontextualized quantity, not a meal or pattern | UNRESOLVED — held open both ways. A robust association with no natural experiment to null it (as coffee’s did) or confirm it (as LDL/BMI’s did); healthy-user + guideline-adherence confounding stay unexcluded. Less resolved than coffee, not exonerated -> Food Groups and Health Outcomes - A Dose-Response Matrix |
PPI adverse outcomes [2026-09-25] (the DRUG case — a commensurable RCT exists) | umbrella of 42 SR/MAs, mostly observational: hip fracture RR 1.20, AKI RR 1.61, C. difficile OR 1.99, gastric cancer OR 2.50 — consistent in direction, GRADE very low (Salvo 2021) | COMPASS RCT (Moayyedi, n=17,598, pantoprazole vs placebo, ~3y) — «PPI use was not associated with any of these adverse events with the exception of enteric infections» (Salvo et al., 2021) | resolves cleanly TOWARD the RCT — because a drug is blindable, the trial tested the same exposure the cohorts did, so no wrong-exposure escape applies; the observational signals are confounding by indication (sicker patients get PPIs), not drug effect. The one signal both streams share — infection — is the believable one -> Proton Pump Inhibitors and Adverse Outcomes |
The instances are not independent confirmations of one claim (that would be a laundered type-E) — they are three instances of one appraisal structure. That is what makes the page a type-A synthesis rather than a corroboration pile.
On the sugar row, a guidance body now states the structure in its own voice [2026-08-06]. EFSA
2022 reports the discordance directly — RCTs on surrogate endpoints support a causal sugar->metabolic-
disease relationship, yet prospective cohorts in isocaloric exchange «do not support a positive
relationship» with any metabolic or pregnancy endpoint — and resolves it the wiki’s way: «excess energy
intake leading to positive energy balance … appears to be the main mechanism». So the resolution above
(harm tracks the calories, not the molecule) is no longer only wiki-induced; a fourth body reached it by
a separate systematic review (loci on Free Sugars Intake). It corroborates the structure, not an
independent primary dataset.
On the UPF row, a fifth body states both halves of this page’s method in an SR’s voice [2026-08-20].
Aramburu 2024 — the first RCT-only systematic review of UPF-reduction interventions — reaches this page’s
two load-bearing moves independently. (1) The surrogate caution: hard-outcome UPF RCTs are infeasible,
so only short-term surrogate trials exist, and a surrogate is not a free pass — Aramburu’s own analogy is
that «reducing the intake of saturated fats has a favorable impact on lipid profile … although it has
not demonstrated a clear association with cardiovascular mortality», the LDL-vs-mortality surrogate gap
the corpus holds -> Surrogate Outcomes. (2) The resolution: «triangulation has been proposed as the
best approach based on integrating evidence from multiple study designs, such as short-term trials,
mechanistic studies, and well-conducted large-scale epidemiological observational studies» — this page’s
triangulate, do not crown the RCT stated by a guidance-grade source. It also makes the
insufficient-not-refute reading concrete: the RCT base is «only four clinical trials, of which only one
directly evaluated the effects of consuming a UPF-based diet … small … short … based on intermediate
outcomes», so the trials «make it difficult to draw definitive conclusions about the true effect of UPFs
on health.» This corroborates the structure (surrogate-validity + triangulation), not an independent
primary dataset — and Aramburu INCLUDES Hall, so it is not independent backing for the intake foothold.
Loci on Ultra-Processed Food and Health Outcomes and Is the Food Category Doing Any Work.
The untested pole — red meat -> T2D, and why it is NOT coffee [2026-08-29]
— the contrast case that keeps the pattern honest in the other direction. Coffee and red meat both carry a large, consistent cohort signal for T2D, so it is tempting to transport coffee’s resolution (MR nulled it, so treat that arm as confounded and insufficient) onto red meat. The parameter table blocks that: the two are in different epistemic states, because only one has a natural experiment.
| Parameter | Coffee -> T2D | Red meat -> T2D | Same quantity? |
|---|---|---|---|
| Observational signal | large umbrella, consistent inverse (T2D high-vs-low RR 0.70, 0.65-0.75, per Coffee Consumption and Health; corrected 2026-10-06: was the all-cause RR ~0.83, an off-outcome value) | NutriGrade “high”, RR 1.17 (1.08-1.26) per 100 g | yes — both are large, consistent prospective-cohort associations |
| Natural experiment (MR) | null, weak — no genetic causal signal for the T2D arm, but that MR is Poole’s secondhand citation (unheld, power not stated), and the held coffee MR (Nordestgaard, CVD/mortality) by its own ~225,000-case statement cannot exclude the observational 0.86 (see Update below) | none held — untested in either direction (superseded in part [2026-10-06]: one weak, unit-less MR, see Update below) | no — coffee is MR-tested with a null not powered at the observational size; red meat is unexamined (corrected 2026-10-06: disconfirmed -> weak null) |
| Resulting state | insufficient-for-causation (no causal support; not a disconfirmation) | unresolved — held open both ways | no — a tested-null arm is not an untested arm, though neither is closed |
So red meat is less resolved than coffee, not exonerated by analogy to it: coffee earned its downgrade from a null MR (weak evidence the arm was confounded — the null is not powered to exclude the observational effect size, see Update below); red meat has no such evidence in either direction, so the honest state is untested for causation, neither established-harmful nor shown-benign -> Food Groups and Health Outcomes - A Dose-Response Matrix. The symmetric error to “the RCT was null, so ignore the observational” is “the food is not proven causal, so treat it as safe” — both skip the missing test.
Update — red meat is now tested once, and the test is too weak to move it [2026-10-06]
The table’s none held cell (and the instance-table row above) is superseded in part: a two-sample MR of UK Biobank intake of processed meat, pork, beef and lamb against T2D, CAD, hypertension and stroke is now held, and reports every IVW estimate as null (Li et al., 2024). The tempting move is to file red meat beside coffee: MR null, so the cohort signal is confounded. A matched comparison of the two MRs blocks it. Neither MR could exclude an effect of the observational size, but only coffee’s says so and says what it can exclude.
| Parameter | Coffee MR (Nordestgaard; IHD/CVD/mortality outcomes) | Meat MR (Li 2024) | Same quantity? |
|---|---|---|---|
| Exposure unit | per allele, ~8% higher coffee intake | not stated (per unit of a genetically predicted UK Biobank intake answer) | No |
| Instrument strength | two-locus allele score (AHR + CYP1A1/2), F = 827 | F>10 cut-off only; per-SNP F in an unheld supplement | No (a value vs a threshold) |
| Power vs observational size | power calculation stated: can exclude OR 0.97 per allele; ~225,000 cases + 225,000 controls would be needed to exclude the observational 0.86 | no power calculation; T2D CIs wide enough to contain the observational RR (processed 0.71-1.66 vs 1.37; beef 0.87-3.75 vs red meat 1.17; unit-mismatched, heuristic only) | No |
| Positive control | ApoE genotype -> IHD recovered | none | No |
| Pleiotropy handling | five variants near three genes gave similar results; funnel plot of 11 caffeine alleles; authors flag AHR/CYP1A1/2 pleiotropy as a limitation | preference SNPs; PhenoScanner screen for outcome diseases and unnamed confounders; MR-Egger intercept | No |
Coffee values (Nordestgaard & Nordestgaard, 2016) — the power statement: «we could exclude odds ratio per allele of 0.97 for 8% higher coffee intake with 80% power and two-sided P < 0.05) compared with observational differences in coffee intake (we observed odds ratio of 0.86 for approximately 350% higher coffee intake), and thus even more individuals and events are required to thoroughly exclude a causal associ- ation from moderate coffee intake to reduced risk of CVD and all-cause mortality using genetic instruments.» (loci on Coffee Consumption and Health); meat values (Li et al., 2024); the observational RRs from (Schwingshackl et al., 2017). The coffee MR tests IHD/CVD/mortality, not T2D, so the comparison is of design features, not of one outcome.
On decisiveness the two MRs are closer than the older table above implies (that table’s coffee MR is the T2D-arm MR Poole cites secondhand, not Nordestgaard): neither excludes the observational effect size. What separates them is how well the null is characterised. Coffee’s states its unit, an F value, a positive control and an explicit power statement, so a reader knows the design was powered (80%) to exclude a per-allele OR of 0.97 and was not powered to exclude the observational one. Li’s gives none of these, so a reader cannot say what it rules out. Red meat moves from untested to tested once, insufficient; the cell stays open both ways. Heuristic (induced from two MRs, not a law): an MR null counts against an observational signal only when it is shown powered to exclude the observational effect size, which needs a stated exposure unit and adequate instrument strength; a positive control strengthens it. Neither MR here meets it. The same instrument-validity weaknesses would discount a positive result from an MR built like Li’s.
Why the streams disagree — five mechanisms, not one
- The blindable trial tests a DIFFERENT exposure (the load-bearing one). You cannot blind a food or
randomize a lifetime, so the trialable form is an isolate, a reformulation, or a short whole-diet
swap under high adherence — a different exposure than the decades-long habitual pattern the cohort
measured. A null on the isolate does not refute the pattern; it refutes the isolate. This is the
telos’s the blindable form is a different exposure rule and the fibre-isolate-out-grades-fibre-food
case -> Is the Food Category Doing Any Work.
- The mechanism has a boundary — and a DRUG lies outside it (the PPI case)
[2026-09-25]. This escape is available only when the exposure cannot be blinded. A drug can be blinded and randomized, so its RCT tests the same exposure the cohorts measured — the wrong-exposure resolution is unavailable, and a null RCT is decisive against the observational signal. PPI harms are the worked instance: the COMPASS RCT nulled nearly every observationally-associated adverse outcome, leaving confounding by indication (mechanism #3 — sicker patients get PPIs) as the whole story -> Proton Pump Inhibitors and Adverse Outcomes. So the drug case is the positive control for confounding standing alone: where food discordances stay open because the trial tested a different exposure, the PPI discordance closes because it did not.[type-F — bounds mechanism #1 by exposure-blindability; the drug case is what the food cases cannot be]
- The mechanism has a boundary — and a DRUG lies outside it (the PPI case)
- Duration / latency mismatch. A months-to-years trial cannot reproduce a lifetime exposure to a long-latency outcome (atherosclerosis, cancer). The trial’s null can be a power/duration null, not a no-effect null -> The Insufficient-Evidence Statement.
- The observational is confounded (healthy-user, reverse causation) — so the big signal may not be
causal. This is where the RCT/MR does win. The adjudication tools are the artifact diagnostics
-> The U-Shaped Association Artifact. A sharpened sub-form is guideline-adherence confounding:
when the exposure is itself the subject of health guidance (red meat, salt, saturated fat),
avoiding it is collinear with the whole adherence bundle (not smoking, exercising, screening,
medication adherence), so the guideline — not the food — can drive the observed benefit. It is a
near-self-fulfilling pattern and unusually hard to adjust away, because the confounder is generated
by the same guidance the study is testing.
- The bundle has a measured size, from a causally-clean estimator — the placebo arm
[2026-08-29]. Simpson 2006 (21 studies, 46,847 people) pooled good-vs-poor drug-adherence -> mortality at «odds ratio 0.56, 95% confidence interval 0.50 to 0.63», and — the load-bearing subgroup — the placebo arms of eight blinded trials (19,633 people) gave the same value: «Good adherence to placebo was associated with lower mortality (0.56, 0.43 to 0.74)» (Simpson et al., 2006). A placebo has no pharmacological path to mortality, so the entire gap is the kind of person who adheres — the healthy-adherer bundle («adherence to drug therapy may be a surrogate marker for overall healthy behaviour»: the same diet / exercise / screening / follow-up cluster) (Simpson et al., 2006). This puts a reference magnitude on manufacturable effect: a behaviour -> mortality association of order OR ~0.56 can be produced by the adherer bundle with zero causal input from the behaviour, so magnitude alone never certifies causation for a guidance-endorsed exposure -> Diet Quality Scores and Cardiovascular Risk.[type-F — quantifies mechanism #3; the placebo isolation is causally clean where a diet cohort is not] - The counterpoint that keeps it honest — adherence to a HARMFUL agent flips to net harm. In the two trials where the active drug was proved harmful, good adherence to it carried «increased mortality (2.90, 1.04 to 8.11)» (Simpson et al., 2006). So the bundle does not paint every adhered-to exposure protective: where the exposure itself does real harm, the harm shows through the bundle (a promoted, adhered-to, genuinely-harmful product — the trans-fat-margarine case — is not laundered benign by its adherent users). The bundle biases toward benefit for a guidance-endorsed behaviour; it does not reverse a true harm — a partial mask, not an omnipotent one.
- The complementary inert-pill isolation — nocebo (Wood, SAMSON)
[2026-08-29]. The placebo arm isolates the adherer bundle on the outcome (mortality); the blinded n-of-1 crossover isolates the nocebo effect on the side-effect (statin muscle symptoms are ~90% reproduced on an inert pill) -> Statins for Primary Prevention and the Power of Zero CAC. Two inert-pill designs isolating two different non-pharmacological effects of taking a pill — one on the endpoint, one on the complaint.[type-A — the inert pill as a general isolation instrument]- The parallel-group RCT confirms the same discordance in the muscle-symptom domain (Reith/CTT)
[2026-09-07]. SAMSON isolates the nocebo within-person (n-of-1); Reith/CTT’s IPD meta-analysis of parallel double-blind statin RCTs is the between-group randomised estimate against which the observational muscle claim breaks. Reith notes that «on the basis of non-randomised observational studies of routine health-care records, it has been suggested that statin therapy is associated with large excess risks of musculoskeletal disorders» — «although such studies are susceptible to statistical bias and confounding» — whereas the pooled randomised excess is only RR 1·03 (year-1 excess 11 per 1000 person-years, flat thereafter). (Reith et al., 2022) So the muscle-symptom domain now holds two randomised-design instances of the discordance — n-of-1 (SAMSON) and parallel-group (Reith) — a large observational excess that the randomised designs shrink to near-zero. -> Statin Muscle Symptoms and the Nocebo Effect. (Not an independent type-E: Reith is the same CTT body whose nocebo literature it cites, so it is a worked instance of the concept, not a separate arrival at it.) - Third independent instance — antidepressant discontinuation
[E-independent][2026-08-30]. The placebo-discontinuation arm isolates the nocebo component of a withdrawal side-effect: ~1 in 6 report discontinuation-like symptoms after stopping an inert placebo, so roughly half of antidepressant discontinuation symptoms trace to expectation/non-specific effects rather than the drug (figures + provenance on -> Antidepressants for Depression). Corroborated by Henssler 2024, a different drug class and research group that does not cite SAMSON — a genuinely independent arrival at the same inert-pill-isolates-nocebo-on-the-side-effect design, strengthening its status as a general isolation instrument.
- The parallel-group RCT confirms the same discordance in the muscle-symptom domain (Reith/CTT)
- The bundle has a measured size, from a causally-clean estimator — the placebo arm
- Measurement error flattens the trial’s contrast (adherence drift narrows the achieved between-arm difference) while the cohort captures habitual intake -> Measurement Error in Dietary Assessment.
- Credibility grade and certainty grade are DIFFERENT AXES that legitimately disagree. An umbrella review grades an association convincing on volume, consistency, and bias screens; GRADE grades the same evidence very low for causation because it starts observational evidence low. Class-I convincing = GRADE very low is not a contradiction — it is two instruments measuring two things (how consistent vs how causally certain) -> Upgrading Observational Evidence.
The resolution — triangulate on the exposure, do not crown the RCT
- Match the exposures before comparing the verdicts. Ask: did the trial test the same exposure the cohort did (whole pattern vs isolate; lifetime vs 12 weeks)? If not, the “disagreement” is an artifact of non-commensurable exposures — a G-gap, not a refutation. (The parameter-table discipline applied to two evidence streams instead of two sources.)
- Where a natural experiment exists (Mendelian randomization), weight it heavily — it removes
healthy-user confounding without needing to blind the food, so an MR null is strong evidence the
observational arm was confounded — but only when the MR is shown powered to exclude the observational
effect size, which neither held diet MR is (coffee->mortality, meat->T2D: weak nulls), nor is the
secondhand coffee->T2D MR shown to be (power not stated) — see the Update above; corrected 2026-10-06: strong evidence … (coffee->T2D) -> qualified) — and an MR
positive is strong evidence it was causal
(LDL/apoB -> ASCVD) -> LDL ApoB and Cumulative Exposure. The LDL/apoB case is the corpus’s cleanest
MR-positive: multivariable MR not only confirms causation but adjudicates which correlated trait
carries it — entered together, only apoB retains a genetic effect while LDL-C reverses to null — the
natural experiment converging with the RCTs rather than clashing with them (the positive-control end of
this page’s pattern, not a discordance) -> LDL Lowering and Cardiovascular Events.
- BMI -> mortality is a second MR-convergence (positive control)
[2026-08-06]. Where coffee’s MR found no support for the observational, BMI’s MR corroborates it: Wade 2018’s genetic instrument reproduces the causal harm the bias-corrected Global BMI IPD-MA found — significant for CVD-cause mortality, directional-but-imprecise for all-cause (MR 1.03, 0.99-1.07) — and shows the observational curve if anything under-estimated the obesity arm while over-estimating the underweight arm (the J’s nadir shifts from ~26 to ~23 kg/m2 under MR; reverse causation is named as the observational J’s driver). So an MR that agrees is as informative as one that nulls — it moves the observational finding from bias-corrected-strong toward genetically-causal (cause-specifically for CVD; all-cause directional). The caveat is independence, not direction: Wade shares two authors (Davey Smith, Sattar) with the Global BMI collaboration, so it is a type-F genetic refinement of the same lineage, not an independent witness -> BMI and All-Cause Mortality, The U-Shaped Association Artifact.
- BMI -> mortality is a second MR-convergence (positive control)
- Believe it where independent method classes CONVERGE. The folic-acid case (observational + trials + biochemistry + genetics all converging) is the standard for a survived nutrition finding; the two classic reversals (beta-carotene, dietary-fat->breast-cancer) had many studies of one method class -> Upgrading Observational Evidence (the triangulation criterion).
The dual-use guard
— the same failure mode as The Estimate-to-Action Gap and The Descriptive-Normative Category Error.
The discordance is a rationalisation engine in both directions, and the honest path names which:
- “The RCT was null, so ignore the observational.” The science-sceptic move — valid only if the trial tested a commensurable exposure over a sufficient duration; a null short-isolate RCT licenses nothing about the lifetime pattern.
- “The RCT can’t capture it, so trust the observational.” The science-booster move — valid only after the confounding checks (MR, artifact diagnostics) are run; un-blindable is not a free pass past healthy-user confounding.
The discriminator: a genuine catch names which mechanism produces the discordance and points to the commensurable test; a motivated one just picks the stream it wanted. Backstop — the uniformity tell: if the discordance is always resolved toward the answer you already held, that pattern is the signature.
Decision relevance
- Do not read a null RCT as refutation until you have matched its exposure to the observational’s. A nutrition RCT null is often a wrong-exposure or under-duration null rather than a no-effect one — check which before letting it overturn a large observational signal.
- Do not read a convincing credibility grade as causal certainty — it is a consistency grade; the causal question needs MR, mechanism, or a commensurable trial.
- Open loop. This grades how to appraise the disagreement, never which answer is true against a realized outcome; the wiki cannot close that loop.