A single-author Viewpoint arguing that the modal nutritional-epidemiology cohort study — small relative risks (typically <2), extensive uncontrollable confounding among correlated dietary exposures, unreliable self-reported FFQ-based intake, and a wide garden of forking analytic paths — cannot support the causal dietary claims routinely drawn from it and folded into public guidelines. Calls for large randomized trials, objective biomarkers, and pre-registration to replace continued observational-cohort output as the field’s primary evidence base.
relevance_note: The named primary position-artifact of the “don’t trust these cohorts” pole of the epistemics debate; sets the terms the defenders (Satija et al. 2015) respond to.
Extracted (structured summary)
Near-universal food–mortality associations
O-69 - Updated cohort meta-analyses find almost all food groups statistically significantly associated with all-cause mortality
A referenced fact about the nutritional-cohort literature (from an external food-group meta-analysis), resting on no data of Ioannidis’s own. Its discriminating value: if nearly every food is “significantly” associated with mortality, that near-universality is more consistent with correlated exposures plus flexible analysis than with each food exerting a real causal effect.
Link to original
Implausible implied life-expectancy effects (incl. eggs)
O-70 - Taken as lifelong causal, meta-analyzed cohort estimates imply implausible life-expectancy effects including minus 6 years for one egg per day
An arithmetic illustration Ioannidis derives from published meta-analyzed cohort estimates (external data, none of his own): the conditional ‘if these hazard ratios are lifelong causal, then…’ is uncontested arithmetic; the egg figure (−6 years/day) is directly on the main question. The implied magnitudes are individually implausible and cannot all hold jointly (one cannot simultaneously gain 12 y from hazelnuts, 12 y from coffee, and 5 y from a mandarin against an 80-year baseline).
Link to original
Reductio: implausibility marks the estimates as bias-driven
A-25 - Implausible, mutually incompatible implied effects mark the cohort estimates as bias-driven not causal
Original
statement: “Because the meta-analyzed cohort estimates imply individually implausible and jointly incompatible life-expectancy effects — a picture made worse, not rescued, by the attenuation defense — they most parsimoniously reflect cumulative bias rather than causal effects.”
Validity (step 6)
status: corrected — reason_if_not_false: checked. Traced the IBE. The internal step “if the estimates were literal lifelong causal effects they would have to be jointly consistent and biologically plausible; they are neither; therefore they are not literal lifelong causal effects” is valid, and the joint-incompatibility half is the sharpest: against a fixed ~80-year baseline the summed positive effects are arithmetically impossible, so they cannot ALL be true as stated — a genuine reductio. What does NOT go through is the further leap to “cumulative bias rather than causal effects.” An undercutting defeater survives without denying any premise: the implausible life-expectancy magnitudes are generated by the “taken as lifelong causal” extrapolation itself (applying a per-increment HR linearly across a lifetime). Small but real causal effects that simply do not extrapolate linearly to lifelong would produce the same implausible numbers. So residual confounding + selective reporting is not the uniquely most-parsimonious explanation; invalid extrapolation of modest-but-real effects is an equally live one. Hence the strong “bias not causal” conclusion overreaches. Corrected to the conclusion the reductio actually licenses: the estimates cannot be read at face value as literal lifelong causal effects and are strongly distorted — leaving the bias-vs-nonlinearity split open. (The near-universal-significance point that would push specifically toward bias is the separate argument A-26.)
Reasoning
Take the meta-analyzed cohort hazard ratios at face value as lifelong causal effects. Each then implies a specific life-expectancy change: roughly +1 year per hazelnut (12/day → +12 y), +12 y for 3 cups of coffee, +5 y for a daily mandarin, −6 y for a daily egg, −10 y for two bacon slices. Two problems compound. (i) Individual implausibility: single small daily food increments do not plausibly move survival by 5–12 years, and a −10-year effect “worse than smoking” for two rashers of bacon is biologically incredible. (ii) Joint incompatibility: against a fixed ~80-year baseline these gains and losses cannot all be simultaneously true — the positive effects alone would sum to implausibly long lifespans. A defender might invoke attenuation from non-differential misclassification (true effects even larger); but that makes the estimates MORE implausible (12 hazelnuts → +20–30 y), and there is no guarantee the self-report error is non-differential. Given that almost every food shows a “significant” association, the parsimonious explanation of a body of implausible-yet-ubiquitous associations is that they are dominated by cumulative bias (residual confounding + selective reporting), not causal signal. This is an inference to the best explanation with a valid internal step: if the estimates were causal they would have to be jointly consistent and biologically plausible, and they are neither.
Link to original
Non-identifiability from correlated exposures and flexible analysis
A-26 - Universal correlation among dietary variables plus unmeasured behavioral confounding and flexible analysis make single-food effects non-identifiable in cohorts
Reasoning
Two mechanisms make cohort diet–disease associations non-identifiable. (a) Confounding by correlation: almost all nutritional variables are correlated with one another, so if even one is causally related to an outcome, many correlated others will also show significant associations in large samples — significance therefore does not localise the causal component. Diet is further entangled with time-varying social/behavioral factors (a healthy-user matrix), and no available cohort measures enough of them to adjust adequately; individuals consume thousands of chemicals in millions of combinations (>250,000 foods, >500 polyphenols in seemingly similar items), so the true causal variable is typically unmeasured. Ioannidis’s genome-linkage analogy: studying single foods is like linking large chromosomal regions to disease with a handful of microsatellite markers — each region (food) contains thousands of variants (chemicals), so the design cannot resolve the causal unit. (b) Analytic multiplicity: the vast space of analyzable associations plus non-prespecified reporting means investigators can and do report a subset of many possible analyses; “meta-analyses become weighted averages of expert opinions,” and single significant slices are published without accounting for the cohort’s other findings. Under (a)+(b), obtaining significant associations for almost any food is the expected null-model behavior of the method, so such associations carry little evidential weight for causation. The inference is valid: a test whose positive rate is high under the no-causal-effect scenario (because of confounding + multiplicity) has low discriminating power, so its positives do not support causal claims.
Validity (step 6)
status: approved — reason_if_not_false: checked. Traced both legs and the join. Leg (a): if dietary variables are pervasively mutually correlated and even one is causal, correlated others acquire significant associations in large samples, so significance does not localise the causal variable; combined with the true causal unit (a specific chemical among thousands) being unmeasured and the healthy-user matrix under-adjusted, a single food’s effect is not identifiable — this follows from the correlation/omitted-variable premises. Leg (b): a large space of non-prespecified analyses with selective reporting inflates the realized positive rate. Join: under (a)+(b) the probability of a significant single-food association is high even when no single-food causal effect exists, i.e. P(positive | no effect) is high, so the likelihood ratio of a positive is near 1 and it barely discriminates — a valid Bayesian/signal-detection step. Probed for an undercutting defeater conditional on premises: “triangulation with MR/RCTs can still identify some foods” — does not break the step, because the claim is explicitly scoped to cohort single-food estimates (“in cohorts”), and appealing to non-cohort designs steps outside that scope rather than defeating the inference within it. The strong words “non-identifiable” / “carry little evidential weight” are exactly what the premises deliver, so no weakening needed. Approved as stated.
Link to original
Objective-assay association still fails in trials (beta-carotene)
O-71 - Serum beta-carotene shows a strong protective mortality association in cohorts (RR 0.69) that randomized trials exclude
A referenced fact resting on no data of the Viewpoint’s own. Discriminating value: even when dietary measurement error is bypassed with a biochemical assay, the cohort association failed to replicate in randomized trials — evidence that the confounding/selection problem, not just measurement error, drives nutritional-cohort associations.
Link to original
Thesis: cohorts cannot support causal dietary claims
H-31 - Nutritional-cohort diet-disease associations reflect cumulative bias and cannot support causal dietary claims
The central position of the Viewpoint and the named artifact of the “don’t trust these cohorts” pole. It is a candidate answer to the meta-level question of whether the egg-cohort evidence can carry causal weight; contested (directly rebutted by Satija et al.), hence a hypothesis, not an observation.
Link to original