A cross-cutting diagnostic, not a claim about one exposure. When observational data shows a U- or J-shaped association — risk lowest at some intermediate exposure, higher at both zero and high — the lower arm (the apparent benefit of a little vs none) is usually the fragile part, and is often not causal — though the artifact can occupy either arm (the nuts -> stroke and grip-strength upper arms below; the muscle-strengthening upper arm is suspect, not adjudicated). Alcohol is the worked case; the same shape sits unadjudicated for sodium, and sleep is adjudicated at arm level below (lede updated 2026-10-28; self-critique).

The recipe that manufactures a false protective arm

The lower arm can be produced with no true benefit at all, by any of:

  • Referent-group contamination / sick-quitter bias — the unexposed group is enriched for people who stopped the exposure because they became ill, so the referent looks unhealthy and everyone else looks protected. In alcohol this is decisive: including former drinkers among abstainers «will bias drinking risk estimates downward, thereby magnifying the appearance of health benefits from low-level drinking», and former drinkers carried «a 38% increased risk» (Stockwell). (Stockwell et al., 2016)
  • Reverse causation — poor health lowers the exposure, not the reverse.
  • Confounding by lifestyle / frailty — the intermediate-exposure group differs systematically (Naimi 2005, reported by Stockwell: «27 (90%) of 30 potential adverse confounders for coronary heart disease were more prevalent among abstainers than among moderate drinkers»).
  • Unequal between-group precision — per the telos, differential reporting/measurement precision can bend a flat relationship into a U with no bias and no confounder, passing both standard checks.
  • Analytic multiplicity / the garden of forking paths — an analyst free to choose the referent, the covariate set, the exposure cut-points, and the subgroup can land on a protective arm among many paths, and selection publishes the significant one. NASEM’s worked case is the post-hoc astrological subgroup that found «Geminis and Libras did not benefit from aspirin, while Capricorns benefited the most» — «This obviously spurious relationship illustrates the dangers of analyzing data with hypotheses and subgroups that were not prespecified.» A curve’s shape is one of those choices, so an un-prespecified U carries the same discount -> P-Hacking and Researcher Degrees of Freedom. (National Academies of Sciences Engineering and Medicine, 2019)

The three adjudication routes — and what they showed for alcohol

RouteWhat it doesAlcohol verdict
Bias-stratified / referent correctionre-analyse with never-drinkers (not ex-drinkers) as referent; keep only bias-free studiesprotection vanishes: RR 0.97 (0.88-1.07) once abstainer biases and study quality are adjusted (0.89, 0.84-0.94, before follow-up length, ill-subject inclusion and race/smoking controls are added); bias-free 0.90 (0.76-1.06), both ns (Stockwell Tables 3-4)
Mendelian randomizationgenetic instruments fix lifetime exposure, immune to reverse causationmonotonic harm for stroke, no protective arm (Millwood)
Exclude early follow-up / sick baselineremove the reverse-causation windowU-shape persisted observationally — so this alone is insufficient (Millwood)

(Stockwell et al., 2016)

The load-bearing lesson: excluding early follow-up is the weak check (the alcohol U-shape survived it); the referent correction and MR are the strong ones, and where a genetic instrument exists it is decisive. A protective arm that survives only the weak check has not been adjudicated.

The decision rule

A U/J-shaped observational association is not, by itself, evidence that an intermediate dose is optimal. Before recommending a little is better than none, require that the lower arm survive a referent-correction or a genetic/quasi-experimental check. If it has only survived covariate adjustment and early-follow-up exclusion, treat the protective arm as unadjudicated, not established — a shape equally consistent with the causal and the artifact explanation has no diagnostic value.

Where it applies in the wiki

Worked-instance catalog — per-exposure applications of the diagnostic; content, not revision history.

  • Alcohol — adjudicated, artifact. The protective lower arm is largely non-causal -> Alcohol and Mortality and Vascular Disease.
    • The BP outcome supplies the interventional check on a DIFFERENT endpoint [2026-08-20, Roerecke] type-E/F. 36 randomised alcohol-reduction trials (Roerecke 2017) show a dose-dependent BP fall with «an apparent threshold eff ect at two drinks per day» and no protective arm — reducing alcohol never raises BP, and below 2 drinks/day it is flat. (Roerecke et al., 2017) This is a third independent (RCT-interventional) leg on the alcohol->BP causal slope, alongside Millwood’s conventional + genetic-MR slopes -> Alcohol and Mortality and Vascular Disease. Distinction, not a re-adjudication: the endpoint is BP, not mortality/IHD, so it confirms no protective arm on a new outcome rather than re-settling the mortality J — a randomised design being the strongest form of the interventional check the decision rule asks for.
      • Qualified 2026-10-22 (resolves the review flag raised by the Filippini instance’s rule). Two parts of the sentence above carry less than the rest. The threshold at 2 drinks is a between-trial contrast: trials are pooled by baseline band and a meta-regression on baseline intake does the work («Alcohol intake at baseline explained 75·4% of the between-study variance in systolic blood pressure, and 93·4% in diastolic blood pressure»), and the <=2 stratum also cut the least alcohol (weighted mean -15 g/day vs -30, -40 and -32 g/day in the higher bands). Below 2 drinks the 5 trials give SBP -0.18 (-1.02 to 0.66) and DBP +0.61 (-0.04 to 1.26); in the band’s high-CVD-risk subgroup (4 trials) DBP is +1.30 (0.14 to 2.47). (Roerecke et al., 2017) So the threshold rests on a between-trial contrast, and flat below 2 means no significant fall, not a shown zero. The no-protective-arm direction stands for SBP: it rests on pooled within-trial randomised contrasts, and no band shows cutting alcohol significantly raising SBP. DBP is weaker: in the <=2 band cutting moved DBP up, non-significantly overall and significantly in the 4-trial high-CVD-risk subgroup (a between-trial split, hypothesis-grade). (inferred from Roerecke et al., 2017)
    • A low-dose cohort curve with no lower arm and an unreported referent [2026-10-22, Di Federico] type-F/G. Seven cohorts (19,548 adults, 4-12 y) give a near-straight fitted BP-change curve from the zero referent with no threshold shown: SBP 1.25 (0.49 to 2.01) mmHg at 12 g/day vs «no alcohol consumption». (Di Federico et al., 2023) The studied range is 0 to ~55 g/day (plotted extent, figure read), with no formal non-linearity test in the main text. The referent’s make-up is not reported in the main text [searched: former / never across chunks 01-02, srcgrep 0 / 0], so the referent check cannot be run. Every point on the curve is a contrast against that referent, and a 3-knot spline pinned at zero cannot isolate a step at zero, so contamination could tilt the whole fitted curve, not only its lowest step; on a BP-change outcome the direction of such a distortion is unsigned. The overall curves are monotone over the studied range (women’s and North Americans’ DBP are not), so the diagnostic has no protective arm to discount here. Set against Roerecke’s <=2-drink trial null it is a distinction, not a joined issue. The two measure different quantities (weeks-scale BP level vs years-scale BP change) and are consistent once matched. Overall DBP intervals fail to overlap only if the cohort difference is read as a level effect reversing within weeks, and they touch in women only (no men-only <=2 trials) -> Does Light Drinking Raise Blood Pressure. (inferred from Di Federico et al., 2023; Roerecke et al., 2017)
    • The CANCER outcome has no lower arm to adjudicate — the J is outcome-specific [2026-08-27, Bagnardi] type-F. Bagnardi 2014’s site-specific dose-response MA (572 studies) finds the aerodigestive and breast curves rise monotonically from zero, significant already at light drinking (breast light 1.04, oral/pharynx 1.13, oesophageal SCC 1.26) with «a clear dose–risk relationship» and no protective arm. (Bagnardi et al., 2014) So alcohol’s protective lower arm is a property of the mortality/vascular endpoint, not of alcohol — on cancer there is nothing to adjudicate. The only inverse cancer arms are single sites (kidney, thyroid, lymphoma) — net-outweighed or plausibly reverse-causation (lymphoma) -> Alcohol and Cancer Risk. Not independent-E (same exposure, new outcome, shared observational base) -> type-F cross-outcome.
    • The CKD outcome shows the artifact signature — unadjudicated, weak checks only [2026-08-29, Kelly] type-F. Kelly 2020 (SR + MA, 104 studies) finds both moderate-alcohol RR 0.86 (0.79-0.93) and high-alcohol RR 0.87 (0.79-0.95) protective against incident CKD — a dose order with no gradient, the classic manufactured-arm shape. The daily-vs-weekly split is diagnostic: daily RR 0.98 (0.82-1.18) is null while weekly RR 0.82 (0.75-0.90) is protective — protection tracks drinking pattern, not dose, and Kelly names the confounder: «social integration as a product of moderate alcohol consumption and overall well- being, which is good for health.» (Kelly et al., 2020) Only the weak check (covariate adjustment) was run — no referent-correction, no Mendelian randomization — so the arm does NOT clear the bar the mortality J had to (where MR removed it). A second exposure-outcome pair now carries the alcohol artifact’s signature, adjudicated in-principle by the concept but not yet by a strong instrument -> Chronic Kidney Disease and Modifiable Exposures. Not independent-E (same exposure, new outcome, shared observational base) -> type-F cross-outcome.
    • A burden-weighted MODEL J is not an adjudication route — it inherits the input arm [2026-09-02, Bryazka] type-F. GBD 2020 (Bryazka) estimates a J-shaped burden-weighted RR curve for ages 40+, non-zero optimum in high-CVD regions «for all regions». (Bryazka et al., 2022) But this J is a weighted composite of the same observational IHD/T2D dose-response RRs whose lower arm the concept already flags — built on a «reference group of non-drinkers» (sick-quitter referent, not referent-corrected), no MR run, and Bryazka concedes residual «measurement bias and selection bias, as well as the potential impacts of reverse causality». (Bryazka et al., 2022) Bryazka’s own cited MR meta-analysis nulls it — «67% of studies on cardiovascular disease and 75% of studies on diabetes reporting a null association». (Bryazka et al., 2022) Lesson for the concept: a J re-surfacing in a downstream model is not a fourth check — it is the artifact arm propagated, and re-passes neither the referent-correction nor the genetic gate. Same-body GBD revision (Bryazka = GBD 2020, Griswold = GBD 2016), so type-F not independent-E.
  • Sleep — adjudicated, and it sharpens the concept. The sleep-duration U-curve’s long-sleep arm (RR 1.30) has no demonstrated mechanism and is «a powerful additional marker of ill-health» (reverse (Cappuccio et al., 2010) causation), while the short-sleep arm (RR 1.12) carries a mechanism and cause-specific evidence -> Sleep Duration and Mortality. New nuance: the artifact can occupy ONE arm of a U-curve while the other arm is causal — so adjudicate the arm, not the curve. The tells that flagged it were the concept’s own: the artifact arm was the larger association, had higher heterogeneity, and strengthened with age (a frailty gradient, not a dose-response).
    • Second sleep instance, different outcome — the arm-level rule replicates. Shan’s sleep-duration U-curve for incident type 2 diabetes (nadir 7-8 h) shows the identical asymmetry: the short arm shows no nonlinearity detected (P=0.22) with a human-corroborated glucose/insulin mechanism, while the long arm is «currently considered more speculative», carries higher heterogeneity, and goes borderline (1.09, 0.99-1.12) when one study (Tuomilehto) is dropped — Shan even allows «long sleep is a consequence of the sleep-inducing effects of the inflammatory state» (reverse causation). (Shan et al., 2015) Shan runs only the weak check (multivariable adjustment, which the long arm survives); no referent-correction or MR, so the long-sleep -> T2D arm stays unadjudicated. Not an independent witness to Cappuccio: the two pools probably share the Gangwisch NHANES I cohort (Shan 2007 row, Cappuccio 2008 row; same group, ages 32-86, 8-10 y follow-up), though the author lists do not overlap -> type-F (same reverse-causation mechanism, new outcome) -> Sleep and Metabolic Health. (Shan et al., 2015) (Cappuccio et al., 2010)

    • Third sleep instance — the per-hour curve QUANTIFIES the asymmetry [2026-08-13, Yin] type-F. Yin 2017’s dose-response SR+MA (3.58 M participants; restricted-cubic-spline, 7 h reference) resolves the mortality U per discrete hour and the asymmetry the concept predicts is now measured, not just argued: the short arm is shallow (6 h 1.01, 5 h 1.04, 4 h 1.08, reaching only 1.12 at the 3 h extreme) while the long arm is steep (9 h 1.15, 10 h 1.32, 11 h 1.53; stroke steeper still — 10 h 1.64). (Yin et al., 2017) The tells are this concept’s own: the long arm is the larger association yet Yin judges «the potential mechanisms underlying the association between long sleep duration and adverse outcomes are considered more speculative» and reports the view that «the elevated risk of long sleep duration most likely represented the confounding effects of subhealthy status» (Yin et al., 2017), and the U «was more pronounced among the studies that reported mortality of total CVD» than incidence (long-arm per-h 1.15 mortality vs 1.00 incidence) (Yin et al., 2017) — the signal strengthens toward death, the frailty/severity gradient. Adjudication is WEAK-checks-only (multivariable adjustment; no referent-correction, no MR), so the long arm stays unadjudicated. Publication bias does not distinguish the arms — Egger flagged both long-sleep/CVD and short-sleep/all-cause at P=0.01 and both survived trim-and-fill (1.22 / 1.04, unchanged). Not independent-E of Cappuccio (same design class, overlapping cohorts, same reverse-causation reading) -> type-F dose-resolution -> Sleep Duration and Mortality.

    • Fourth sleep instance — the MR STRONG check finally lands: short arm survives, long arm not supported (but underpowered, not refuted) [2026-09-02, Wang] type-F. Every sleep instance above ran weak checks only (multivariable adjustment; no referent-correction, no MR), leaving both arms unadjudicated by this concept’s decision rule. Wang 2022 (meta-review + MA of 4 observational MAs + 11 MR studies) supplies the genetic natural experiment on CVD morbidity endpoints. The short arm survives the strong check: genetically-predicted short sleep raises CAD (IVW OR 1.24, 1.15-1.34), MI (1.20, 1.12-1.28) and HF — «evidence is accumulating that short sleep duration is a causal risk factor for CAD and HF» (Wang et al., 2022) — so this arm is now genetically supported, not merely mechanism-argued. The long arm gets no genetic support: «long sleep duration has no causal associations with stroke and CAD in the MR studies» (Wang et al., 2022), matching the concept’s prediction that the larger, mechanism-free arm is the artifact-suspect one (Wang: long sleep «a surrogate risk indicator for poor health status», associations «reflected potential reverse causality»).

      • The discipline the alcohol case did NOT need — insufficient vs no-effect. Unlike Millwood’s decisive alcohol MR (which removed the protective arm on a powered instrument), Wang’s long-sleep MR nulls are underpowered: few genetic instruments for long sleep (one MR ran «no analysis» for want of SNPs), and Wang concedes «MR studies … did not provide sufficient evidence supporting the causal association between long sleep duration» and «No clear experimental evidence shows the harmful effects of long sleep duration». (Wang et al., 2022) So the long arm moves from argued-artifact to genetically-unsupported — a stronger position than Yin/Shan left it, but NOT a demonstrated no-effect. Scope caveat: the MR lands on CAD/HF/stroke incidence, not directly on the mortality U-arm Cappuccio/Yin measured, so it strengthens the arm-asymmetry read across outcomes rather than re-adjudicating the mortality curve itself.
      • Type + independence. The observational arm re-pools the same overlapping Cappuccio-class cohorts (shared base, not independent-E); the MR is a genuinely different method but it triangulates the same question rather than backing a specific held estimate -> type-F (supplies the missing genetic leg, exactly as Wade’s MR did for BMI) -> Sleep Duration and Mortality.
    • Fifth sleep instance, the DEMENTIA/cognition outcome — the asymmetry becomes OUTCOME-SPECIFIC WITHIN the duration curve [2026-09-04, Zhang] type-F. Zhang 2025 (76 cohorts) runs the sleep-duration U on cognition and the arms hit different endpoints: «Sleep duration < 7 h primarily increases the risk of cognitive decline; while sleep duration > 8 h mainly elevates the risk of AD, dementia, and cognitive decline» (Zhang et al., 2025), and the short arm explicitly «was not associated with future risk of all-cause dementia and AD» (Zhang et al., 2025). So on the dementia/AD endpoint the U collapses to a long-arm-only elevation (long > 8 h -> AD 1.66, 1.44-1.91, I2 0%; dementia 1.43, 1.21-1.69) — there is no short arm to defend. The long arm shows this concept’s every tell: it is the larger association, mechanism-poor, and age-graded — «The role of prolonged sleep duration in the development of dementia remains unclear and is closely associated with age» (Zhang et al., 2025), concentrated in the >= 70 y elderly, «a preclinical marker driven by the APOE ε4 carrier gene» (Zhang et al., 2025) (Tang: long sleep tracks higher Aβ40/tau, lower Aβ42/Aβ40, smaller gray matter — pathology as cause of the long sleep, i.e. reverse causation). Adjudication is WEAK-checks-only: baseline-only exposure with a fixed 1-13 y follow-up offered as the lag argument, but NO MR and NO referent-correction — so the long-sleep -> dementia arm stays unadjudicated, matching Wang’s genetic long-arm null on CVD. What it ADDS: the arm-level asymmetry is now within one exposure across outcomes (short->decline, long->dementia), so adjudicate the arm, not the curve sharpens to adjudicate the arm FOR THIS OUTCOME — the same duration curve is a U on decline and a long-arm-only rise on dementia. Independent measurement note: the reassurance the duration arm lacks, insomnia has — objective insomnia is stronger and far less heterogeneous (RR 1.26, I2 26.1%) than self-report (1.09, I2 77.7%), the measurement-precision tell running the causal way on that arm. Not independent-E of Cappuccio/Yin (overlapping cohort base, same reverse-causation reading, new outcome) -> type-F cross-outcome -> Sleep and Cognitive Decline, Dementia Prevention and Modifiable Risk Factors.

  • Muscle-strengthening activity — a J-shape with its own counter-instance built in. In the pooled MA, MSA shows a J-shaped dose-response for all-cause mortality/CVD/cancer (nadir in a wide, imprecise ~30-80 min/week region; the all-cause point estimate reaching 1.0 near ~140 min/week, with its lower 95% band above 1.0 from ~165, figure-read; knot note below) — but an L-shape (no upturn) for diabetes, the one outcome with a clear muscle-glucose mechanism -> Muscle-Strengthening Activity and Mortality. An instance consistent with, not diagnostic of, the adjudicate-the-arm rule: the outcome with a clear mechanism (diabetes) shows no upturn, while the upturn outcomes have only candidate mechanisms (a CV channel such as arterial stiffness; IGF-1 for cancer, per Zhang) — plus very-low GRADE, self-reported exposure, and sparse high-volume data. The upper arm is the artifact-suspect region and is left unbelieved (wording corrected 2026-10-28 from the sharpest single-source instance and upturn lacks a mechanism; self-critique). The lower arm is kept as the action default under the firmness rule (cheap, reversible, low-harm), but on mortality it is itself unadjudicated by this page’s rule: it has passed only covariate adjustment and lags, and in Zhang’s cohort adjustment removes most of the crude association (1-29 min/week: age-adjusted 0.83 (0.81-0.85), 0.90 with covariates, 0.95 (0.92-0.97) with aerobic activity added). (Y. Zhang et al., 2026)
    • A repeated-measures cohort does not show an above-null all-cause category [2026-10-28, Zhang 2026] type-F. With weekly weight-machine/resistance-training time re-asked every 2-4 years and averaged (HPFS, NHS, NHSII; 35,798 deaths), the top category (>=120 min/week, mean 204) has all-cause HR 0.92 (0.86-0.98) on 1093 deaths, and for the aerobic-adjusted CVD curve no non-linearity is detected (P non-linearity 0.98). (Y. Zhang et al., 2026) The all-cause spline point estimate crosses 1.0 only near ~270 min/week (about 1.02 at 300, the sparse top of its axis, interval including 1); no category estimate lies above the null. Beyond each model’s last knot a restricted cubic spline is constrained to be linear, so the shape there (and its band) reflects the model as much as the data; the data beyond the knot still set the slope. This holds for each curve here. Zhang’s last knot is at 150 min/week (Y. Zhang et al., 2026). Momma used a «restricted cubic spline model with three knots at fixed percentiles (10%, 50% and 90%) of time of the exposure» (Momma et al., 2022) without reporting the knots in min/week; its all-cause curve is near-linear from ~85-90 min/week to its end at ~195 (axis to 200; figure-read, PDF p 6). Shailendra, pooling 4 of Momma’s 6 dose-response studies with the same percentile scheme, put its last knot at 150 min/week («3 knots placed at 10, 30, and 150 minutes, which roughly correspond with the 10th, 50th, and 90th percentiles of the distribution of dose values from the included studies») (Shailendra et al., 2022) So Momma’s last knot plausibly sits somewhere in ~75-150 min/week, and the same caveat applies to its above-null segment from ~165 as to Zhang’s tail

      Zhang does reproduce the J’s diminishing benefit at high volume: all-cause curvature is significant (P non-linearity <0.001), neurological mortality is J-shaped (P 0.002), cancer benefit is confined to 1-59 min/week, and the cancer spline rises above 1.0 at the top (to ~1.2 at 300 min/week, interval including 1; P non-linearity 0.05), though that crossing lies mostly beyond the last knot and no cancer category estimate is above 1.0. What it does not reproduce is the all-cause crossing above 1.0 at the category level. That is consistent with an artifactual above-null arm but does not discriminate it: Zhang differs from the pooled cohorts on several axes at once (repeated averaging; an explicit 2-year lag, this page’s weak check; a narrower item; knot placement and linear tails in both models, Momma’s last knot at its 90th exposure percentile and Zhang’s at its 95th; a US health-professional sample, a sample type also present in Momma’s six-study dose-response pool — «Six studies were eligible for the dose–response analysis» (Momma et al., 2022) — whose Figure 3 lists Grøntved 2012 (men; the HPFS, per the nucleus page) and Kamada 2017 (older women; whether it is a health-professional cohort is not stated in the held text) among the six). Aerobic adjustment may differ in depth — in Momma «Adjustment for confounders varied widely across studies» and «All studies considered aerobic or other types of physical activity» (Momma et al., 2022) — but it cannot by itself explain the pooled arm sitting above Zhang’s: in Zhang adding aerobic activity moves every RT category toward the null (>=120 min/week: 0.79 with covariates, 0.92 with aerobic activity added) (Y. Zhang et al., 2026), so weaker aerobic adjustment in the pooled studies would, under the same confounding structure, pull their high-volume estimates down rather than above the null. The mechanism list above names no single-measurement route, and Zhang names no artifact route: it says that a single baseline measure «may introduce measurement error and limit their ability to evaluate the dose-response» and that non-differential mismeasurement would be «generally attenuating the associations» — neither names a route by which one measurement creates an upturn. (Y. Zhang et al., 2026) Its one link between single vs cumulative measurement and an earlier CVD U-shape (its ref 14) is a hedged causal one — training duration («acute vs. long-term») among factors that «may explain these discrepancies», since «acute resistance training may increase arterial stiffness, whereas long-term regular training may reduce it». (Y. Zhang et al., 2026) Held material does name candidate routes for the arm, none tested: a single baseline may over-represent transient high-volume phases (rehabilitation, exercise prescribed when ill), which repeated averaging dilutes, a reverse-causation route via single measurement; and Saeidifard, after causal harm mechanisms of high-intensity training (blood pressure, arterial stiffness, the Valsalva manoeuvre), adds a confounding route specific to high volume — «people performing a high volume of resistance training may be more likely to use anabolic–androgenic steroids or supplements with sub- stances that could adversely affect their health» — before concluding that «the current evidence is not enough to conclude that the relationship between resistance training and survival has a U-shape curve». (Saeidifard et al., 2019) Zhang’s HPFS men are also already in two MA all-cause pools (nucleus page). So the held point estimates conflict above ~140 min/week: Momma’s pooled spline rises above 1.0, as does Shailendra’s 4-study spline (crossing 1.0 at roughly ~140-150, lower 95% band above 1.0 from ~165-170, about 1.3 at 200; figure-read, PDF p 6), which re-pools a subset of Momma’s studies and so is not independent of it; its authors read it as «a possible “U”-shaped relationship between the duration of resistance training and the risk of all-cause mortality; however, because this analysis only included 4 studies, it should be viewed with caution» (Shailendra et al., 2022) while Zhang’s categories, and its spline to ~270, stay below it. At ~140 Zhang’s and Shailendra’s curves are still before their 150-min knots and Momma’s may lie either side of its knot; from ~165 all three sit on their linear segments (Momma’s on the knot inference above), under the caveat above. No held source settles which is right: the all-cause upper arm stays unbelieved under this page’s rule, now with one partly-independent non-reproduction of its above-1.0 part; it is not refuted, and the cancer arm stays open -> Muscle-Strengthening Activity and Mortality. (inferred from Y. Zhang et al., 2026)

  • Iodine excess -> thyroid — an UPPER arm of HARM whose direction the interventional check CONFIRMS [2026-09-17, Katagiri] type-F. The mirror flavour of the protective-lower-arm cases: the fragile part here is the excess arm of iodine’s dose-response U (the deficiency arm -> neurodevelopment is real -> Deficiency Repletion vs Enhancement). Katagiri’s SR+MA finds excess vs adequate iodine raises hypothyroidism — adults OH OR 2.78 (1.47-5.27), SCH OR 2.03 (1.58-2.62) (Katagiri et al., 2017) — off cross-sectional, crude, unadjusted, area-level (ecological) exposure contrasts, the artifact recipe’s confounding leg. But unlike the protective arms above, this harm arm gets the strong interventional check the decision rule asks for: a double-blind RCT (Sang) «proposed that a total intake of 800 μg/ day confers a risk of SCH» in the already-replete (Katagiri et al., 2017), plus a human-corroborated mechanism (Wolff-Chaikoff escape failure via NIS) and a UIC-SCH gradient. So the RCT + mechanism clear the DIRECTION (excess harms) while the observational design caps the MAGNITUDE — adjudicate the arm with the sign believed and the OR discounted. The 300 μg/L cutoff is a WHO definitional cutpoint, not a located knee: the studied excess areas ran 338-1241 μg/L, so «we were unable to clarify the situation in borderline excess areas.» (Katagiri et al., 2017) Not independent-E (a different exposure sharing the one diagnostic) -> type-F worked-case extension -> Excess Iodine Intake and Thyroid Disease.
  • Physical activity -> dementia — a protective signal that SURVIVES the reverse-causation check, but only the WEAK one [2026-09-04, Iso-Markku] type-F. Not a U-curve: PA -> dementia is a monotone protective association (RR 0.80, 0.77-0.84), but it carries the concept’s signature threat — reverse causation over the long dementia prodrome, whereby preclinical dementia lowers activity years before diagnosis, manufacturing a spurious benefit. Iso-Markku 2022 (58-cohort SR+MA) is built as the test: the association «appears absent when PA is measured before the age of 65 … or in follow-­ups longer than 10 years» in some prior work, so it re-estimates within the >=20-year-follow-up subset and the protection holds — all-cause RR 0.79 (0.71-0.87, 16 studies, mean baseline age 50.5), AD 0.76 (0.64-0.90) — with «we did not find evidence to suggest that reverse causation or regression dilution bias88 affected the observed associations». (Iso-Markku et al., 2022)
    • But this is the concept’s WEAK check, not the strong one. Extending the follow-up window is a version of exclude early follow-up / sick baseline — the route the alcohol U-shape survived yet was still artifact. No referent-correction applies (no sick-quitter referent here) and no MR / genetic instrument for PA-dementia is held, so the strong check is absent. Two further tells keep it unadjudicated: the cleanest cut (3 high-quality young-baseline >20y studies) goes non-significant (0.79, 0.62-1.01), and Iso-Markku names the residual confounder the design cannot remove — «Physically active individuals may have higher cognitive reserve to start with … may drive the association». (Iso-Markku et al., 2022) So by the decision rule the causal reading is directionally supported but not established — a protective arm that cleared the weak check and never faced a strong one, like NSS and the sodium low-arm. Not independent-E (a different exposure sharing the one diagnostic) -> type-F worked-case extension -> Dementia Prevention and Modifiable Risk Factors, The Physical Activity Paradox.
  • Weight loss -> mortality — reverse causation manufacturing a false HARM, adjudicated interventionally. The mirror-image direction: here the observational signal is a spurious harm, not a spurious benefit. «Most observational epidemiologic studies have indicated that the rate of death from cardio-vascular and all other causes is increased after weight loss» — because they cannot «distin-guish intentional from unintentional weight loss», so «the observed weight loss might be the consequence of conditions that lead to death rather than the cause». (Sjöström et al., 2007) SOS supplies a fourth adjudication route the alcohol table lacks — a controlled intervention that assigns intentional weight loss (bariatric surgery vs matched conventional care), removing the intentional-vs- unintentional confound by construction — and the sign flips: intentional loss reduces all-cause mortality (adjusted HR 0.71, 0.54-0.92). (Sjöström et al., 2007) The intervention is the strongest reverse-causation check available, even at SOS’s non-randomized grade -> Does Weight Loss Reduce Cardiovascular Events, BMI and All-Cause Mortality.
  • Macronutrients in PURE — confounding-by-income, the machinery without (mostly) the U. Dehghan found higher carbohydrate → higher mortality and higher fat/SFA → lower mortality across an 18-country income gradient. The associations are largely monotone, not U-shaped — so this is not a lower-arm case — but it is a clean instance of the concept’s confounding-by-lifestyle/frailty mechanism in its socioeconomic form: the highest-carb quintiles are the poorest (refined-carb subsistence diets), so the fat-protective / carb-harmful signals are the mirror of an income gradient, which the authors concede («residual confounding… cannot be completely excluded»). (Dehghan et al., 2017) The one candidate U-arm is the authors’ suggestion that «a very low intake (ie, below about 7% of energy) [of saturated fat] might even be harmful» — a low-SFA arm that would be the artifact-suspect region (very low SFA tracks very low animal-food intake, i.e. poverty/frailty), and is unadjudicated: no referent-correction or genetic check isolates it from the income confound. The decision rule applies unchanged — the protective/harmful arm is not believed until it survives a confounder-immune check -> Saturated Fat Intake and Replacement.
  • Carbohydrate -> mortality — a BOTH-arms-harmful U whose arm dissolves into a SUBSTITUTION question (a big-rock-adjacent macronutrient instance) [2026-08-05, Seidelmann]. Seidelmann’s ARIC cohort (n=15 428, 25-yr follow-up) + 8-cohort MA (432 179) finds a U-shaped carbohydrate->all-cause- mortality curve, nadir 50-55%E, both arms elevated: «both low carbohydrate consumption (<40%) and high carbohydrate consumption (>70%) conferred greater mortality risk than did moderate intake … (pooled hazard ratio 1·20, 95% CI 1·09-1·32 for low carbohydrate consumption; 1·23, 1·11-1·36 for high carbohydrate consumption)». (Seidelmann et al., 2018) Unlike alcohol/BMI this is not a protective-lower-arm case — there is nothing to defend as a benefit; the diagnostic interest is in the LOW arm’s cause.
    • The distinctive adjudication route — a substitution DECOMPOSITION, not MR. Seidelmann’s own analysis dissolves the low-carb arm by decomposing what replaces the carbohydrate: «mortality increased when carbohydrates were exchanged for animal-derived fat or protein (1·18, 1·08-1·29) and mortality decreased when the substitutions were plant-based (0·82, 0·78-0·87)». (Seidelmann et al., 2018) The arm’s sign flips with the replacement source — so «low carbohydrate» is not a well-defined exposure, and the harm reads as a replacement-food signal («Low carbohydrate diets have tended to result in lower intake of vegetables, fruits, and grains and increased intakes of protein from animal sources … which has been associated with higher mortality»). (Seidelmann et al., 2018) The low-carb quantile is also confounded-by-lifestyle (more smoking 33% vs 22%, more diabetes, higher BMI) (Seidelmann et al., 2018) — the artifact recipe’s confounding leg.
    • What it ADDS to the concept — the exposure can be under-specified. Alcohol/coffee/BMI adjudicate a confounder on a fixed exposure; here the adjudication reveals the exposure itself is a composite («low carb» = low-carb-plus-whatever-replaces-it), and the composite’s mortality sign is set by the substitution, not the carbohydrate. This is the telos’s frame-as-substitutions rule meeting the U-artifact: decompose the arm into its substitution before believing the curve.
    • Adjudication strength — WEAK checks only, so the causal reading stays not-established. Seidelmann ran a reverse-causation sensitivity analysis («individuals with cardio-vascular disease, diabetes, or cancer at baseline were excluded», findings unchanged) and a time-varying diet update — the weak checks the alcohol artifact also survived. (Seidelmann et al., 2018) No MR / genetic instrument. So strong-but-not-decisive, like coffee/BMI.
    • The PURE reconciliation — two cohorts trace complementary arms of one U. ARIC (mean 49%E) populates the left arm, PURE/Dehghan (mean 61%E, Asian/low-income; the bullet above) the right; overlaid on one reference, «the associations between primarily high carbohydrate intake and mortality in the PURE study still fell within the confidence intervals of those observed in ARIC». (Seidelmann et al., 2018) The two headline-clashing literatures are the two arms of a single curve — and the right arm is the same refined-carb/poverty signal Dehghan’s bullet flags, not carbohydrate per se. Not independent-E of Dehghan (both observational FFQ cohorts, PURE shared between them) -> type-F worked-case extension, reconciling with the Dehghan/PURE bullet above -> Low-Carbohydrate vs Balanced-Carbohydrate Diets, Dietary Protein and Mortality.
    • The pooled version — Qin 2023 (41-cohort SR+MA) quantifies the J but shows it is FRAGILE and outcome-specific; type-F over Seidelmann, not independent-E [2026-08-19]. Qin re-pools Seidelmann’s ARIC AND Dehghan’s PURE and cites Seidelmann as the antecedent it agrees with (finding «consistent with a previous meta-analysis revealing a U-shaped association … [17]») (Qin et al., 2023) — shared studies, so F not [E-independent]. Two refinements it lands on the artifact question: (i) shape is outcome-specific within one source — Qin finds NO non-linearity (linear/monotone) for CVD, CHD and stroke incidence, but a J-shaped curve only for all-cause (Pnon-linearity 0.008) and CV mortality (0.055) (Qin et al., 2023) — the U/J is a mortality- endpoint phenomenon, absent on the incidence endpoints; (ii) the all-cause J is fragile — RR 1.07 (1.00-1.14), «not robust in the sensitivity analysis», going non-significant on removing any single one of Seidelmann/Dehghan/McKenzie/Frisoni, and the per-5%E all-cause slope is null (Qin et al., 2023). Adjudication strength: still WEAK checks only — Qin is cohort-only with no MR/genetic instrument and no substitution decomposition, so it adds pooled magnitude but does NOT advance the low-arm past the unadjudicated status Seidelmann left it in (a pooled J assembled from cohorts that individually do not robustly show it) -> Low-Carbohydrate vs Balanced-Carbohydrate Diets, Is the Food Category Doing Any Work.
  • MCE cholesterol-death — the machinery turned on a CONTRARIAN headline (symmetric standards) [2026-08-04, Ramsden]. Not a U-curve but the frailty/reverse-causation machinery in a within-RCT association: in the recovered Minnesota Coronary Experiment, each 30 mg/dL serum-cholesterol decrease tracked a 22% higher risk of death (HR 1.22, 1.14-1.32) — the figure seed-oil skeptics cite as evidence LA-lowering-of-cholesterol kills. The tells are this concept’s own: the association is entirely >=65-driven (age >=65 HR 1.35, age <65 null 1.01, 0.88-1.16) — a frailty gradient, not a dose-response — and Ramsden concedes it is «observational in nature», a within-trial cholesterol-change association that «did not differ between the intervention and control group» (so it is not the randomized diet contrast). (Ramsden et al., 2016) Ramsden ran only the weak check — a frailty sensitivity analysis adjusting for weight/BP changes, which the association survived — but that is a crude proxy, not an instrument immune to reverse causation (low cholesterol is a marker of the frailty/illness that causes death). So by the decision rule the causal reading (lowering cholesterol raises death) stays unadjudicated: the association is real, its causal interpretation is not established, and it must get the same discount the pro-LA observational benefit signals get -> Linoleic Acid and Cardiovascular Disease. The clean part of MCE — the randomized cholesterol-fell/mortality-null contrast — needs no U-shape machinery and stands on its own -> Surrogate Outcomes.
  • Non-sugar sweeteners — reverse-causation machinery that did NOT dissolve the association (only weak checks exist). Not a U-curve: higher NSS use shows monotone positive associations with type 2 diabetes (HR 1.23-1.34), CVDs (1.32), stroke (1.19) and all-cause mortality (1.12) in long-term cohorts, while short-term RCTs show flat cardiometabolic biomarkers -> Non-Sugar Sweeteners. The textbook reverse-causation story fits (people already heavy/dysglycaemic switch to NSS because of their condition), and WHO ran the reverse-causation battery — BMI adjustment, weight stratification, excluding pre-baseline weight-losers, dropping early follow-up. Its verdict is the honest middle: «although reverse causation and residual confounding may be contributing factors … the associations … cannot be dismissed as being solely a result of reverse causation or residual confounding.» (World Health Organization, 2023)
    • This is the instance where the machinery neither dissolved nor confirmed the signal. Contrast alcohol, where referent-correction + Mendelian randomization removed the protective arm. NSS has had only this page’s WEAK checks (covariate adjustment + early-follow-up exclusion); the STRONG check — a genetic/MR instrument immune to reverse causation — does not exist for NSS. So by the decision rule, the NSS→disease association is unadjudicated, not established: it survived the weak checks (which alcohol’s artifact also survived) and has never faced the strong one.
    • Corroborated as a candidate by Qin 2020 (SSB/ASB dose-response MA) — same unadjudicated status [2026-08-06]. Artificially-sweetened beverages carry a monotone positive dose-response with T2DM (RR 1.15 per 250 mL/d) and obesity (1.21), plus non-linear arms for hypertension and all-cause mortality — and Qin runs no reverse-causation or MR check (cohort-only, «residual confounding cannot be ruled out»). So this adds beverage-form data points to the NSS reverse-causation candidate without adjudicating it: still the arm that has faced no strong check -> Free Sugars Intake.
    • A trap the case exposes: adjustment cannot tell a confounder from a mediator. WHO notes that since the associations «largely persist when body weight is controlled for … increased body weight (resulting from chronic NSS use) may be an intermediary step … rather than a confounding factor» — so the association survived BMI adjustment is ambiguous: if weight is a mediator on an NSS→disease path, adjusting for it wrongly attenuates a real effect; if a confounder, adjusting is correct. The persistence-after-adjustment that reads as robustness cuts both ways. AWAITS a Mendelian- randomization source on NSS — the only instrument that would adjudicate the arm.
  • Sodium — named, argued-but-not-adjudicated. WHO cited a J-shaped hypothesis (harm below ~2 g/day) as the reason for its review, then never cited the two cohort studies behind it again [searched: J-shaped / refs (37, 38) across chunks 01-02], and excluded a priori the strata where a lower-arm harm is most plausible -> Sodium Intake and Blood Pressure. He 2013 rebuts the same J-curve papers invoking «measurement error … confounding … and reverse causality» — this concept’s recipe, named — but only as a critique, from CASH/WASH advocates, with no referent-correction or MR on the low-intake arm. (He et al., 2013) So it is the textbook case of the decision rule: a partisan argument that the arm is artifact is not the strong check, and the arm stays unadjudicated. The concept says exactly what would settle it: a referent-corrected or MR analysis of the low-intake arm, which the corpus does not yet hold. Huang 2020 (same lineage) adds a second such argument, not the check: it frames the low-intake mortality signal as «an artefact attributable to factors such as reverse causation and biased estimation of sodium intake» and directly contradicts PURE’s high-intake BP-association threshold with RCT effects «far below this» — but this is the concept’s recipe named again, still without an MR or referent-correction on the arm. (Huang et al., 2020) Two same-lineage critiques do not compound into an adjudication; the arm stays open. PURE (Mente 2016) supplies the direct low-arm observation the critics lacked — and it is still a pole, not the check. Pooling 133,118 people (>10,000 events, 3 cohorts), the low arm is directly present on hard outcomes: <3 g/day carries HR 1.34 (1.23-1.47) in hypertensives and 1.26 (1.10-1.45) in non-hypertensives vs a 4-5 g/day referent, surviving BP adjustment. (Mente et al., 2016) It directly measures the arm the He/Huang critiques only argued about — but it does not adjudicate it either, because the candidate artifact mechanisms are all live and Mente ran only the weak checks against them: (i) reverse causation / sick-quitter — sick people eat less salt; Mente excluded events in the first 2 years and excluded known CVD/hypertension/diabetes in turn (the pattern held), but neither is a referent-correction; (ii) spot-urine measurement error — PURE estimated intake from one fasting morning urine via the Kawasaki formula, and whether that error is larger at the extremes is untested in the held sources. Mente argues that correcting toward 24-h values would steepen the associations at both the low and the high end (Mente et al., 2016); Graudal, citing He 2018, argues a single estimate gave an outcome almost identical to multiple 24-h measurements (Graudal et al., 2020). Neither tests error by intake level. If the error were unequal across the range, it could bend a flat curve into a U with no confounder and no reverse causation (gate-6), passing exactly those exclusion checks -> Measurement Error in Dietary Assessment; (iii) confounding by frailty — residual, not removed by multivariable adjustment. Mente ran no referent-correction and no Mendelian-randomization on the low arm, which is the strong check this concept requires, and himself concedes observational analysis cannot prove causality. So PURE moves the arm from argued to directly observed while leaving it unadjudicated — the [PRIOR] does not close here. The policy clash this feeds is Should Sodium Reduction Be Population-Wide or Targeted. (inferred from Mente et al., 2016) The opposing camp argues the arm real, also without the strong check [2026-10-22]. Graudal 2020 (Cochrane, RCTs of BP and markers only) defends the cohort low arm with weak checks: a pooled table of adjusted cohorts using the healthiest subgroups (all-cause mortality, RR/OR 0.84 or 0.87, usual vs low intake; the table includes PURE itself, RR 0.62, so it is not independent of the PURE arm above) and NHANES data showing sick and healthy people eat the same sodium, against reverse causation; and against measurement error the claim that misclassification is «not systematic, but random… which contributes to an underestimation of the outcome, but do not cause a change of the direction of the outcome». (Graudal et al., 2020) That covers classical error only; the unequal-precision mechanism above is the one this random-error argument does not reach. So the sodium arm now has weak checks on Graudal’s side, a dismissal on the other, and the strong check (referent correction or a genetic/objective test) on neither. (inferred from Graudal et al., 2020)
  • Coffee CANCER mortality — a worked referent-correction whose confounder pushed the OPPOSITE way to alcohol [2026-08-04, Grosso + Poole] type-F. Grosso 2016 (dose-response MA, 31 cohorts, 1.6M) performs the smoker/non-smoker referent correction the coffee page flagged as pending. Its cleanest, Grosso-attributed instance is cancer mortality, whose sign flips across the correction: pooled, coffee shows no significant cancer-mortality association, but stratified «cancer mortality was significantly decreased only when considering non-smokers, while increased in smokers» (Grosso et al., 2016) (non-smoker linear RR 0.98/cup, 0.96-1.00). Grosso reads the flip as confounding, not interaction: «it is hardly plausible that any biological effect of coffee causally differs by smoking status… residual confounding by smoking is the most likely the explanation». (Grosso et al., 2016)
    • The refinement — the correction removes whatever the confounder manufactured, and that need not be a protective signal. Set the two referent-corrections side by side (same diagnostic move, route 1; note this compares a confounder-manufactured signal, not literally a J-arm in each — coffee-cancer is a confounded sign-flip, not a benefit-then-harm curve):

      ParameterAlcohol -> all-cause (Stockwell 2016)Coffee -> cancer (Grosso 2016)Same quantity?
      Correctionnever-drinker referent (drop sick-quitters)never-smoker stratum (drop smoking confounder)both route-1 referent/confounder corrections — yes
      What the confounder manufactureda spurious benefit (the protective lower arm of the J)a spurious harm (the smoker-stratum increase that masks a real benefit in the pool)both a confounder-made signal the correction targets — yes (as a diagnostic move)
      Direction of the spurious signalbenefitharmno — opposite
      After correctionbenefit vanishes -> monotone harmharm vanishes -> monotone benefit (0.98/cup)no — opposite directions
      Adjudication strengthreferent-correction + MR (Millwood) = decisivereferent-correction only for cancer (observational; no MR for the cancer outcome — Poole cites coffee MRs only for T2D and all-cause/CV mortality, its ref 123 being the held Nordestgaard, a null not powered at the observational size, and calls for MR on further outcomes) (Poole et al., 2017) (Poole et al., 2017) (corrected 2026-10-06: Poole calls for MR studies but reports none -> cites two, neither on cancer)coffee is less fully adjudicated

      The bottom rows are the payoff: the confounder-correction does not “restore the null by killing a protective arm” as a rule — it removes whatever the confounder was creating. Smoking manufactured an apparent coffee-cancer harm (smokers drink more coffee and, because smoking is the dominant cancer risk factor, die more of cancer), so correcting it revealed more benefit — the reverse of alcohol. This extends the adjudicate the arm, not the curve rule: adjudicate the signal and note which direction the confounder pushes it — the artifact is not synonymous with “the protective arm.”

    • All-cause/CVD linearizes too, but Grosso does not attribute THAT to smoking. The overall all-cause curve is a J (nadir RR 0.83 at 3 cups -> 0.90 at 7) (Grosso et al., 2016) while never-smokers show «a linear dose-response… decreased risk by 6 % for each additional cup… for all- cause and CVD mortality (RR = 0.94, 95 % CI = 0.93, 0.96 and RR = 0.94, 95 % CI = 0.91, 0.97, respectively)» (Grosso et al., 2016) — so the correction removes the upper-arm attenuation here as well. But Grosso reports «No differences were found between smokers and non-smokers for all-cause and CVD mortality risk» (Grosso et al., 2016), attributing the smoking artifact explicitly only to cancer. So the all-cause/CVD upper-arm-as-artifact is the wiki’s inference from the linear-vs-J contrast, held against Grosso’s own no-difference statement — suggestive, not established. The clean instance is cancer.

    • Adjudication status: still only partial. Grosso removes the dominant confounder (smoking) but is observational — SES / reverse causation / other residuals remain, and the genetic instrument is null: «genetically coffee intake was not associated with risk of cardiovascular disease or all-cause mortality» (Nordestgaard 2016, now held primary — the coffee->mortality MR Poole cited secondhand). (Nordestgaard & Nordestgaard, 2016) Consistent, not in tension: the per-cup benefit survives the smoking referent-correction yet finds no support from the genetic instrument, so residual non-smoking confounding is the live remaining explanation. By the decision rule the lower-arm benefit stays not established as causal — but smoking is no longer a candidate for the whole association.

      • The genetic check has its own two limits (so it bounds rather than closes the arm): the MR has a strong instrument (F=827) but is not powered to exclude a causal effect as large as the observational one (~225,000 cases + 225,000 controls would be needed; it excludes only a much steeper per-allele OR of 0.97 for 8% higher intake) (corrected 2026-10-06: powered to exclude … as large as the observational one -> not powered, Nordestgaard chunk 01), and «is based on the assumption of linearity … [so] will not be capturing non-linear differences» — i.e. an MR cannot in principle detect a true U, which is exactly the shape under test here. (Nordestgaard & Nordestgaard, 2016) So the genetic null excludes only a steep linear protective effect — it does not disconfirm one of the observed size, and is not proof of zero effect — the adjudicate the arm verdict rests on referent-correction + a bounded genetic check, not a decisive instrument (unlike alcohol’s Millwood MR). -> Coffee Consumption and Health.
    • Type guard: Grosso and Poole share the coffee-cohort evidence base (Poole is an umbrella over MAs of this class), so this is F-refinement of Poole’s mortality arm, NOT independent-E. Against the alcohol instance it is neither E nor a tension — coffee and alcohol do not disagree; they are two exposures exhibiting the one diagnostic. It is the second route-1 referent-correction worked on a dose-mortality curve (after alcohol); sleep, by contrast, was adjudicated by mechanism-presence and heterogeneity tells, not by a referent-correction.

  • BMI -> all-cause mortality — the cleanest DECOMPOSABLE confounder-strip cascade (a big-rock exposure) [2026-08-05, Global BMI]. The 10.6M-participant IPD-MA is built as a stepwise bias-removal cascade, so you can watch the artifact form. The overweight arm (BMI 25-30) — the obesity paradox — walks from apparent protection to clear harm as each bias is stripped: raw 0.96 (0.95-0.97) -> +adjust smoking/exclude baseline disease 0.99 -> +exclude first 5 y follow-up 1.03 -> +restrict to never-smokers 1.11 (1.10-1.11). (Global BMI Mortality Collaboration, 2016) The cleanest single confirmation isolates one confounder: holding the 5-y exclusion + no-baseline-disease constant so only smoking status differs, overweight is 1.07 (1.06-1.07) in never-smokers vs 0.94 (0.94-0.95) in ex/current smokers (between-stratum heterogeneity significant for every BMI group; overweight P=0.0003, obesity I P<0.0001). (Global BMI Mortality Collaboration, 2016) Smoking is the confounder manufacturing the protection -> BMI and All-Cause Mortality.
    • What it adds to the concept. (i) A big-rock adiposity instance, not another peripheral exposure. (ii) Direction: like alcohol, the confounder manufactured a spurious benefit (overweight protection), the opposite of the coffee-cancer harm — reinforcing note which direction the confounder pushes. (iii) A genuinely mixed artifact route — the concept’s self-critique flagged that alcohol+sleep both rest on reverse causation alone; here the arm is bent by smoking confounding AND reverse causation AND prevalent disease, and the cascade decomposes their marginal contributions separately (smoking adjustment 0.96->0.99; +early-death exclusion ->1.03; +never-smoker restriction ->1.11). (iv) The adjudicate-the-arm rule again: the overweight arm is entirely artifact (0.96->1.11), while the underweight arm is only partly so — 1.81 (raw) -> 1.47 (primary) but stays elevated, residual real harm. Not independent-E of alcohol/coffee (a different exposure sharing the one diagnostic, not independent backing of one claim) -> type-F worked-case extension.
    • The strong check now exists — Wade MR closes the BMI arm, and it is the SECOND exposure (after alcohol) with a genetic instrument [2026-08-06, Wade]. Global BMI adjudicated by confounder-removal only; Wade 2018’s Mendelian randomization in UK Biobank (77-SNP GRS, 335,308 people, 9,570 deaths) supplies the genetic natural experiment. Its verdict converges: «The J-shaped BMI-mortality association remained in MR analyses … but with a smaller value of BMI at which mortality risk was lowest (~23 vs. ~26 kg/m2 with observational analyses) and apparently flatter over a larger BMI range» (Wade et al., 2018) — and Wade names the mechanism the concept predicts: «Reverse causality … may be the driver of the characteristic J-shaped association». (Wade et al., 2018)
      • What MR does to the two arms sharpens the arm-level rule. Unlike alcohol (where MR removed the protective arm cleanly), BMI’s low arm does not vanish — the J remains but deflates: the nadir shifts down into the normal range (~23) and the residual J is an extreme-quantile effect (removing the tails -> linear, P=0.999). So MR deflates the underweight arm (reverse causation) while inflating the obesity arm — the observational curve «overestimate[s] the harmful effects of having underweight while underestimating the harmful effects of having overweight or obesity». (Wade et al., 2018) The concept’s adjudicate the arm, not the curve rule is corroborated genetically: the low arm is largely artifact, the high arm is genetically supported (significant for CVD-cause mortality, directional-but-imprecise for all-cause), and severe underweight plausibly keeps a real (non-artifact) residual — matching Global BMI’s 1.47.
      • BUT the MR is NOT independent of the observational IPD-MA — the lineage chase (Richardson lesson). Wade’s senior author George Davey Smith and co-author Naveed Sattar both sit on the Global BMI Mortality Collaboration writing committee (Smith GD, Sattar N), and Wade cites Global BMI as ref 5. Two shared authors including the anchor MR figure -> same-lineage type-F refinement (supplies the missing genetic leg), not independent-E backing. So BMI is MR-adjudicated but by an overlapping group; a genuinely independent MR would upgrade it further. Full parameter table + estimates -> BMI and All-Cause Mortality.
    • Sun MR DECOMPOSES the J — the aggregate curve is a MIXTURE of subgroup shapes; stratify by the effect-modifier [2026-08-19, Sun] type-F. Sun 2019’s non-linear MR (HUNT + UK Biobank, 100 residual-BMI strata) is the third BMI-MR instance and the sharpest decomposition case in the catalog. The J survives the genetic instrument (a causal basis, nadir «around 22-25» (Sun et al., 2019)) — but stratifying by smoking splits it: «an always-increasing relation of BMI with mortality in never smokers and a J shaped relation in ever smokers … the BMI-mortality relation is likely comprised of at least two distinct curves, rather than one J shaped relation. An increased risk of mortality for being underweight was only evident in ever smokers.» (Sun et al., 2019)
      • What it ADDS beyond adjudicate the arm. Alcohol/coffee/BMI adjudicate a confounder on a fixed aggregate curve; Sun shows the aggregate shape is itself a mixture — the J is not a property of BMI-mortality, it is what you get pooling an always-increasing never-smoker curve with a J-shaped ever-smoker curve. So the move is decompose the curve by the effect-modifier before believing its shape — a pooled J (like a pooled “no U”, Jayedi/Aune above) can hide heterogeneous subgroup shapes. This is an MR (strong-check) instance, so the artifact survives into the causal analysis yet still resolves into subgroups.
      • It refines the underweight-arm reading — the residual harm is largely a SMOKING phenomenon. In never-smokers Sun finds «no evidence for a harmful effect of reducing BMI in underweight participants» (clearest in HUNT; in UKB «confidence intervals were wide and compatible with a null effect at all values of BMI») (Sun et al., 2019); the underweight harm concentrates in ever-smokers, where «Increased mortality in underweight smokers might be driven by respiratory diseases» (Sun et al., 2019) — the «other» (non-CVD-non-cancer) cause category carries the only profound J (cancer flat, CVD increasing -> shape is outcome-specific). Candidate mechanism for the residual low-BMI risk: «low lean mass rather than low fat mass» (Sun et al., 2019). This does NOT overturn Global BMI’s never-smoker residual (1.47) — that is an observational association, Sun’s is a genetic causal slope on a small (1-3% of sample) and imprecise underweight stratum — it sharpens the arm-level rule toward condition on smoking before reading the low arm.
      • Adjudication strength + independence. MR = strong check, with two caveats: the non-linear fractional-polynomial method carries a published editor’s-note methodological criticism (note text not in the held chunk -> non-linear shape claims caveated, the smoking-decomposition direction less so), and the smoking split is a collider Sun argues is «likely to be negligible». (Sun et al., 2019) NOT independent-E of the held BMI-MR fabric — co-author Di Angelantonio also leads Global BMI 2016, and Sun cites Wade (ref 33); the HUNT cohort and the smoking decomposition are what is novel -> type-F -> BMI and All-Cause Mortality.
    • Central adiposity has FAR LESS lower arm to defend than BMI — the confound that manufactures the obesity paradox does not load the low-WAIST end [2026-09-08, Jayedi central-fatness] type-F. The BMI instance above is a J with a substantial protective-looking overweight arm that the cascade had to strip. Jayedi 2020’s central-fatness dose-response MA (72 cohorts, 2.53M, 150,164 deaths) runs the same all-cause outcome for waist-based measures and the U largely disappears: WHR and ABSI are monotone with «little evidence of sharp changes at particular cut-off points» (Jayedi et al., 2020); waist circumference and waist-to-height ratio are shallow-J with a negligible protective arm (WC nadir men 90 cm HR 0.96, women flat over 60-80 cm; the lower arm spans only ~0.96-1.01, a threshold-then-rise rather than a benefit-then-harm curve). Only BAI (n=4, large total-fat component) shows a real U.
      • Why the arm shrinks — the mechanism the concept predicts. BMI’s low arm is manufactured by frailty/low-lean-mass (thin people include the sick and sarcopenic); a low waist does not carry that confound: «smaller waist circumference might reflect lower detrimental visceral fat mass and does not necessarily reflect lower lean body mass» (Jayedi et al., 2020). So the artifact-generating confounder (reverse-causation/frailty at the low end) is absent by construction for the waist measure — the curve is closer to monotone not because it was adjudicated away but because the exposure does not encode the frailty signal BMI does. This sharpens the adjudicate the arm rule: a better-specified exposure can lack the arm the crude proxy invented.
      • Adjudication strength: WEAK checks only. Jayedi ran the never-smoker / healthy / >10y-follow-up restriction family (results «approximately similar» (Jayedi et al., 2020)) — the same weak route the alcohol U survived — with no MR and no referent-correction. So the small residual arms stay formally unadjudicated; but unlike BMI there is little protective arm to defend. Not independent-E of the BMI instance (a distribution refinement of the same adiposity->mortality question, shared observational base) -> type-F -> Central Adiposity and All-Cause Mortality, BMI and All-Cause Mortality.
  • Hemoglobin -> CAD — the MR DIRECTLY TESTS the U’s nonlinearity and finds none [2026-09-03, Liu] type-F. A distinctive shape-dissolution instance: most catalog instances infer the artifact from tells (mechanism-absence, heterogeneity, a referent flip), and alcohol/BMI adjudicate via referent- correction or a curve that survives-but-deflates; here the strong check tests the curve shape directly — a genetic nonlinearity test. Liu 2024’s iron-status MR reports «The observational analyses in Figure 1 demonstrated U-shaped associations of hemoglobin levels with CAD, wherein both lower and higher levels of hemoglobin were each associated with higher risks of CAD (reference level 14 mg/dL). However, there was no evidence of nonlinearity in the MR analyses (Cochran Q P=0.853, quadratic test P=0.703)» (Liu et al., 2024). So the observational U — both a low-hemoglobin (anemia) and a high-hemoglobin arm elevated — flattens to a monotone genetic slope: the quadratic term is null. This is the adjudicate the arm rule met by a nonlinearity test rather than a referent-correction, and it is a STRONG (genetic) check, so the low (anemia) arm reads as reverse causation / frailty, not a protective feature of higher hemoglobin.
    • The direction is a SEPARATE matter from the shape — do not read “no U” as “more is better.” The same MR splits by outcome: «modest protective effects of iron biomarkers for CAD (7%-14% lower risk for 1 SD higher levels of iron biomarkers), adverse effects for T2D, but no associations with IS or HF» (Liu et al., 2024), and the hemoglobin instrument runs the other way from the iron-store instruments (linear +8%/SD adverse for CAD) — an internal discordance the source leaves unexplained. The concept’s claim here is only about the shape (the observational U is not causal), not about which direction the monotone slope runs -> that outcome-specific direction question lives on Heme Iron and Cardiometabolic Risk.
    • Same-quantity bound + type. The exposure is systemic iron status / hemoglobin, NOT dietary heme (the cross-link’s nucleus draws exactly that distinction), so this instance adjudicates the iron-burden curve, not the food channel. A new exposure sharing the one diagnostic, not independent backing of a held claim -> type-F worked-case extension -> Heme Iron and Cardiometabolic Risk. (inferred from Liu et al., 2024)
  • Smoking cessation — not a U-curve, but the cleanest worked sick-quitter correction. Jha 2013 (Smoking and Mortality) states the mechanism exactly — «Life-threatening illness can cause smokers (Jha et al., 2013) to quit, which distorts the rates of death among current smokers and among those who have quit smoking recently in opposite ways» — and applies the canonical fix: reclassify anyone who quit within 5 years of death as a current smoker, then check by dropping the first 2 years of follow-up (unchanged). It is the template for the correction this concept keeps invoking (alcohol, sleep): the fix biases against the finding, so a benefit that survives it is conservative, not inflated. A monotone-harm exposure, but the reverse-causation machinery is identical.
  • Dementia — a cross-OUTCOME replication of both adjudicated arms [2026-08-05, Livingston]. The 2024 Lancet Commission’s dementia analysis reproduces the concept on a new outcome for the two exposures it already banks. Alcohol: the light-vs-none dementia J-arm is artifact — it «is probably because many non-drinkers have previously had high alcohol consumption», and AD-Mendelian-randomization says «any relationship between not drinking and AD is due to survivor bias». (Livingston et al., 2024) Sleep: the long-sleep dementia arm is artifact — «prolonged sleep is not a risk factor for dementia, although dementia and its prodrome may cause prolonged sleep», the association «completely attenuated» once the first 5 years of follow-up are dropped, while the short-sleep arm keeps a small mechanism-backed risk. (Livingston et al., 2024) Both are the adjudicate-the-arm rule confirmed on dementia -> Alcohol and Mortality and Vascular Disease, Sleep Duration and Mortality, Dementia Prevention and Modifiable Risk Factors. Not independent-E of the mortality instances (same reverse-causation/sick-quitter mechanism, new outcome) -> type-F cross-outcome replication.
    • The alcohol dose-response, first-hand — the primary MA leaves the arm UNADJUDICATED; the held AD-MR is what clears it [2026-09-04, Xu] type-F. Livingston’s verdict is qualitative; Xu 2017’s dose-response MA (11 prospective cohorts, 73,330 participants, 4586 ACD cases) is the first-hand source that draws the dementia J — protective range 0-12.5 g/day, «lowest effect size … (RR around 0.9)» near 6 g/day, harm above 38 g/day. (Xu et al., 2017) The instance sharpens the concept’s weak-vs-strong-check distinction: Xu ran only the weak checks (prospective design, covariate adjustment, NOS>=8) and concedes every strong-check gap — the referent is the sick-quitter-contaminated «lowest category» («we cannot exclude the potential influences of including former drinkers, who may quit drinking due to underlying diseases and have a high risk of dementia, in the reference group»), and no referent-correction and no MR were run, the missing check named in-source («warrants further validation using more advanced approach, such as Mendelian randomization»). (Xu et al., 2017) So the primary dose-response leaves the arm exactly where the mortality J sat before Stockwell/Millwood — unadjudicated; the «artifact» verdict is carried entirely by the held AD-MR (Livingston’s cited «any relationship between not drinking and AD is due to survivor bias»), not by the MA that produced the curve. The tell reinforcing it: protection is wine-only (0.58-0.67), null for beer/liquor — the healthy-user signature, not a dose-response, and the resveratrol rescue Xu offers is the one the held Semba biomarker null already refutes. Not independent-E (single observational MA, shared cohort literature + sick-quitter machinery) -> type-F -> Alcohol and Mortality and Vascular Disease. (inferred from Xu et al., 2017)
  • Blood pressure -> dementia — the RANDOMIZED interventional check kills the observational U outright [2026-09-04, Peters] type-F. The cleanest form of the diagnostic on this page: the artifact is not merely adjudicated by a covariate correction but refuted by randomization. Observational cohorts show a U — «Observational studies indicate U-shaped associations of blood pressure (BP) and incident dementia in older age, but rando-mized controlled trials of BP-lowering treatment show mixed results on this outcome in hypertensive patients» (Peters et al., 2022) — the apparent-harm- at-low-BP lower arm being the classic frailty/reverse-causation shape (declining BP in the dementia prodrome). Peters 2022, an IPD MA of five double-blind placebo-controlled BP-lowering RCTs, removes it: «There was no evidence of a U-shaped re-lation of the effect at any age, nor any increase in risk of dementia with treatment in the oldest age» (Peters et al., 2022), and the achieved-BP dose-response is monotone — «a linear relationship between lower risk of dementia and lower BP, down to at least 100 mmHg systolic and 70 mmHg diastolic» (Peters et al., 2022) (studied-range floor 100/70, not evidence below it). Treatment lowered dementia odds (OR 0.87, 0.75-0.99). So the protective-looking lower arm of the observational curve is a study-design artifact, exactly as this concept predicts — a randomized design being the strongest interventional check the decision rule asks for. Mirrors the BPLTTC J-curve refutation on CV events (Blood Pressure Lowering and Cardiovascular Events): same exposure, two outcomes, both observational U-arms failing the randomized test. Not independent-E (same exposure, new outcome, and Peters’ trials overlap BPLTTC’s base) -> type-F cross-outcome -> Dementia Prevention and Modifiable Risk Factors.
    • The same randomized check, run inside each age band, on CV events [2026-10-27, Rahimi] type-F. The observational worry is the old-age version of the lower arm: «However, other studies have reported an increased risk of cardiovascular events and death in older patients with lower blood pressure compared with those with higher blood pressure.» (Rahimi et al., 2021) The BPLTTC age re-analysis stratified each of five age bands by baseline SBP (<120 to >=170) and found no evidence that the relative effect varied with baseline BP in any band («all adjusted p-interaction>0·070»; «adjusted p-interaction=1·00» at >=85), and none by DBP down to <70 (Rahimi et al., 2021). On death: «Stratified effects on all-cause death followed a similar pattern, with no evidence to suggest treatment increases mortality in any age group.» (Rahimi et al., 2021) That is about harm; on benefit the Discussion adds that the reduction in all-cause death was marginally larger in the young, «with no obvious effect in people older than 75 years», and the authors caution generally that all-cause effects would be expected to be diluted in a subgroup where non-CV deaths make up more of the deaths (Rahimi et al., 2021). So in the four bands below 85, the randomized data give no evidence of a baseline-BP level, down to <120/<70, below which the relative benefit on CV events is lost: the low-BP harm arm finds no support in a randomized check there. At >=85 the band’s own benefit is not established (HR 0.99, 0.87-1.12, about 1000 events) (Rahimi et al., 2021), and only 121 people in that band started below 120 mmHg (Rahimi et al., 2021), so the arm is unadjudicated there for want of power, not refuted. Three of the four observational studies the authors cite for the low-BP worry are oldest-old cohorts, two aged 85 and over by title (Rahimi et al., 2021), so for the oldest old (>=85) the U is largely untested by this check. Scope: stratifying trials by baseline BP is the closest randomized form of the question the observational U raises (does lowering from an already-low BP harm?); it is not an achieved-BP analysis (Peters, above, is) and not a measured dose-response below 120/70. The authors call interaction tests «notoriously conservative» (Rahimi et al., 2021), said of the age test and applied here by extension: no evidence of variation is weaker than constant. Separately, the relative effect had point estimates that differed from the 55-84 value at both ends of age, stronger at <55 (0.82) and weaker at >=85 (0.99), with 0.91 between; only the <55 interval excludes 0.91; the one relative-scale age interaction test for the primary outcome gives adjusted p=0.050, and the authors do not rule out chance, while absolute reductions were larger in older bands (Rahimi et al., 2021); that is an age gradient, not a low-BP arm. Same collaboration, overlapping dataset (51 vs 48 trials), as the BPLTTC J-curve result -> type-F stratum refinement, not independent backing.
  • Milk -> mortality — a single-cohort confounding artifact (an upper-arm case, no U required) [2026-08-06, Guo]. The high-milk-doubles-mortality scare is the exposure-harm mirror of the protective-arm cases: Guo 2017’s pooled milk -> mortality is null (RR 1.00, 0.93-1.07), and the only thing the Swedish Mammography Cohort (Michaelsson) adds is heterogeneity (I2 = 97.4%; excluding it -> 70.1%, RR 0.99). The confounder is named in-source — the highest milk drinkers had lowest education and «highest percentage of smokers and those living alone» (Guo et al., 2017). The tell that seals it: the same cohort drives the opposite (protective) fermented-dairy/cheese arm — «the inverse associations of fermented dairy and cheese with all-cause mortality or CVD disappeared after removing the study of Michaelsson et al.» (Guo et al., 2017). One confounded outlier manufacturing both poles is the recipe with no causal signal. Adjudication is weak here — leave-one-out sensitivity + Guo’s confounder narrative only; no MR or genetic instrument, and Michaelsson’s own D-galactose mechanism is not imported. So the arm is dismissed as unsupported, not positively refuted -> Dairy and Cardiometabolic Health. Not independent-E of the alcohol/sleep instances (same reverse-causation/confounding machinery, new exposure) -> type-F.
  • Milk/dairy -> fracture — a DESIGN-DISCORDANCE instance: the artifact lives in the study design, not just a confounder [2026-08-06, Malmir] type-F. Malmir 2019’s milk/dairy -> osteoporosis+hip-fracture SR-MA is the cleanest corpus case where the protective signal is a property of the weaker design: milk/dairy looks protective in cross-sectional/case-control studies (milk -> hip fracture RR 0.75, 25% lower) but is null in prospective cohorts (0.93 [0.75-1.15]), and milk even reverses to «a 9% greater risk of hip fracture (RR ¼ 1.09; 95% CI: 1.07–1.11)» per 200 g/day in the cohort meta-regression (Malmir et al., 2019). The author adjudicates by the design hierarchy — «findings from cohort studies are closer to the causal associations than those from cross-sectional and case-control studies» (Malmir et al., 2019) — because cross-sectional/case-control designs are reverse-causation-prone (fracture/osteoporosis can lower dairy intake, or recall differs by disease status; the diagnosis can precede the exposure measurement) -> Dairy and Bone Health, The Observational-Trial Discordance.
    • The tie-in to the milk-mortality artifact (same cohort). The two largest positive, highest-weight points in Malmir’s milk cohort forest plot are Michaelsson’s (female 2014 RR 1.60; 2018 RR 1.55) — the same Swedish cohort family that manufactures the milk -> mortality scare on Dairy and Cardiometabolic Health (where its removal collapses that signal), so one confounded high-milk-drinking population plausibly drives both apparent harms. But Malmir runs no leave-one-out on Michaelsson for fracture and no MR — only the design-hierarchy argument — and the pooled milk-cohort RR is itself null (0.93, NS), the positive signal living only in the meta-regression slope; so Michaelsson is the largest positive contributor is read off the forest plot while Michaelsson drives the signal is untested. By the decision rule the milk -> fracture harm arm stays unadjudicated (not established as causal), exactly as the milk-mortality arm does. A dairy -> osteoporosis dose-response J (>250 g/d harm) sits on the same reverse-causation-prone designs and is likewise unadjudicated. Not independent-E of the Guo milk-mortality bullet (overlapping Michaelsson cohorts, same confounding machinery, new outcome) -> type-F.
  • Eggs -> CVD/CHD — a shallow, near-null protective arm, unadjudicated [2026-08-06, Godos] type-F. Godos 2020’s egg dose-response shows a shallow U/J for CVD and CHD (nadir SRR ~0.95 at 2-5 eggs/wk, attenuating back toward 1.00 at high intake; CHD Pnonlin 0.042) alongside a null stroke curve and a monotone-increasing heart-failure curve -> Eggs Dietary Cholesterol and Cardiovascular Risk. The moderate-intake “protection” is at most ~5% relative with CIs touching 1.00, has survived only covariate adjustment (the weak check — Godos concedes reverse causation was «not investigated»), and reverses in the diabetic subgroup (CVD 1.22). So by the decision rule the lower arm is not established — which here coincides with the page’s own de-escalator read (there is no protection to bank). It sharpens the arm-level rule again: the mechanism-bearing outcome (heart failure) is monotone, the mechanism-thin one (moderate-egg CVD “benefit”) is the U-arm suspect. Not independent-E (a new exposure sharing the one diagnostic; Grosso is a shared author) -> type-F.
    • The protective dip did NOT replicate on mortality endpoints [2026-08-19, Ma] type-F. Ma 2022’s egg -> mortality dose-response MA (24 studies, ~11.9 M) models all-cause / CVD / IHD / stroke mortality as linear over the studied range (~0.07-1.5 eggs/d) with no protective lower arm — the all-cause slope even tilts to harm (per 1-egg/d 1.06, 1.02-1.10). (Ma et al., 2022) Ma’s Table 1 cohorts (NIH-AARP, PURE, CKB, Zhong, WHI…) are substantially overlapping with Godos’s pool, so this is not independent-E (inferred from Godos et al., 2020; Ma et al., 2022) — the overlap is the wiki’s cross-source comparison, not a claim Ma itself makes. A shallow protective U/J that appears on one MA’s incidence-weighted endpoint and disappears on a second overlapping MA’s mortality endpoint is the arm-level rule confirmed: the dip was not established, and shape is again endpoint-specific (protective-dip on inc+mort, flat/linear on mortality) -> Eggs Dietary Cholesterol and Cardiovascular Risk.
  • Fish -> chronic disease — a NULL instance, plus a region-masking nuance [2026-08-06, Jayedi] type-F. The Jayedi 2020 umbrella (34 cohort MAs) ran nonlinear dose-response on 16 associations and found «no evidence of a U- or J-shaped association between fish consumption and the risk of chronic disease» — inverse-linear for all-cause/CVD mortality, CHD, MI, stroke, heart failure, nonlinear (but not U/J) for CHD mortality/HTN/Alzheimer/AMD (Jayedi & Shab-Bidar, 2020). So fish is a worked case where the artifact machinery had nothing to bite on — a protective exposure with no suspect upper arm at population intakes. The nuance that keeps it honest: the pooled no-U/J can mask region-specific U-shapes — two recent MAs found linear-inverse mortality/MI in Asian cohorts but «modest U-shaped associations in Western countries» (Jayedi & Shab-Bidar, 2020), so a global “no U” is a statement about the pooled range, not a guarantee within every stratum -> Fish and Seafood Consumption.
  • Physical activity / steps -> mortality — a NULL-for-U instance where objective measurement REMOVES a self-report distortion [2026-08-06, Ekelund + Paluch] type-F. The device-measured PA dose-response (Ekelund 2019, accelerometry) and the daily-steps curve (Paluch 2022) are monotone-decreasing, then flatter on the point estimates — no U, no harmful upper arm at achievable doses (the high-dose intervals fit a plateau and a shallow gradient alike; superseded 2026-11-13: previously monotone- decreasing to a plateau), so the artifact machinery has no protective lower arm to defend. What the case adds to the concept is the opposite of the sick-quitter trap: a place where objective measurement un-does a self-report artifact rather than creating one. Self-report underestimates the effect ~2-fold (Ekelund: «about twice as large» vs self-report (Ekelund et al., 2019)), so the self-report literature’s flatter curves and occasional high-volume plateau are partly a measurement artifact — the same reading the CRF page reaches from objectively-measured fitness (no plateau) -> Cardiorespiratory Fitness and Mortality, Physical Activity Dose and Mortality. Adjudication is the WEAK check only: all three sources (Ekelund 2019/2016, Paluch) address reverse causation (frail-move-less) by excluding early deaths, and it survives — but attenuates, and Paluch’s effect is stronger at <6 y follow-up (HR 0.32 vs 0.57) (Paluch et al., 2022), a sick-quitter tell. No MR/genetic instrument, so the monotone benefit is not-fully-adjudicated causal but is not purely artifact either. Not independent-E of the other instances (shared reverse-causation machinery, new exposure) -> type-F.
  • Occupational physical activity -> mortality — a spurious-harm-AMPLIFIER via the healthy-worker effect, the SELECTION mirror of sick-quitter [2026-08-14, Coenen] type-F. Coenen 2018’s occupational-PA meta-analysis (men, high vs low, HR 1.18) is NOT a U-curve and the harm is not dismissed as artifact — but one selection mechanism inflates it and belongs in this catalog. The harm looked stronger in relatively healthy study samples (the wrong direction for a dose-response), which Coenen attributes to the healthy-worker effect: «this finding is probably due to so-called healthy worker effect, a form of selection bias were more healthy subjects select into and remain in the most physically strenuous occupations». (Coenen et al., 2018) This is the mirror of sick-quitter: there the referent is enriched for the ill (manufacturing a spurious benefit); here the exposed (strenuous-job) group is enriched for the healthy, so a real harm shows up amplified in healthy subsamples rather than created. Adds a new leg to the recipe — selection into the exposure, not out of the referent — and a direction note: the artifact here strengthens an apparent harm (opposite to alcohol’s manufactured benefit, same direction as milk/weight-loss spurious harms). Adjudication is WEAK/subgroup only — a healthy-vs-unhealthy-sample χ2 contrast, plus Coenen’s argument that SES may be a pathway (over-adjusting it biases conservative), no referent-correction and no MR — so the causal size of the harm stays unadjudicated, but the direction of the selection bias is named -> The Physical Activity Paradox. Not independent-E (a new exposure/mechanism sharing the one diagnostic) -> type-F.
  • Nuts -> stroke — a spurious HARM upper arm that dissolves under an OUTCOME-composition stratification [2026-08-13, Aune] type-F. Aune 2016’s nut dose-response is inverse and flatter at higher intake (point estimates) for CHD/CVD/all-cause, but the stroke curve shows «a slight J-shaped curve with reductions in risk observed up to approximately 10-15 grams per day, but a slight positive association at intakes of 30 grams per day, however, this was not observed when studies were stratified by whether the outcome was stroke incidence or stroke mortality». (Aune et al., 2016) So the upper (harm) arm is a candidate artifact of pooling two different outcomes (incidence + mortality), not a dose-response feature — the split removes it. A new flavour of the recipe’s confounding leg: the arm is manufactured by outcome-composition heterogeneity, adjacent to the adjudicate the arm, not the curve rule and to Jayedi’s region-masking nuance (a pooled “no U” can hide, or here invent, an arm the strata do not share). Adjudication is the WEAK check only — an outcome-stratification, no MR or referent-correction — and the whole benefit side is observational healthy-user (nut eaters slimmer / less-smoking / more-active), so neither arm is causally adjudicated -> Nut Consumption and Mortality. Not independent-E (a new exposure sharing the one diagnostic) -> type-F.
  • Fluoride -> fracture — a U whose protective LOWER arm the source itself flags as sparse-data artifact [2026-08-14, Mazzoli] type-F. Mazzoli’s dose-response MA (37 studies; restricted cubic spline) shows «a clear evidence of a U-shaped curve only in females, with the lowest risk around 0.4 mg/L and a monotonic increase above 0.9 mg/L» (Mazzoli et al., 2025) — an apparent protective dip at low-moderate drinking-water fluoride, then a rise. Applying adjudicate the arm, not the curve: the upper arm is believed (fracture rising above ~1.5 mg/L overall; RR 1.26, 95% CI 1.10-1.46 at 1.0 mg/L in postmenopausal females — the one CI excluding 1.0), while the lower/protective arm is NOT — its CIs all cross 1.0 and the authors themselves attribute it to sparse data: «some types of fractures showed a U-shaped pattern … Such a pattern might be an artifact due to the more limited number of studies on which such [estimates rest]» (Mazzoli et al., 2025). The distinctive mechanism this adds: unequal between-group precision — far fewer studies at low fluoride displace the spline nadir with no bias and no confounder (the gate-6 mechanism), so «lowest risk around 0.4 mg/L» is a sampling artifact, never an optimum -> The Underivable Optimum. Adjudication is WEAK-checks-only (RoB stratification, sensitivity by excluding high-RoB; no referent-correction, no MR) on mostly high-RoB ecological data from a single lab -> Fluoride and Bone Health. Not independent-E (a new exposure sharing the one diagnostic) -> type-F.
  • Grip strength -> mortality — a U whose UPPER arm the source itself names a studied-range-edge/sparsity artifact [2026-08-19, Lopez-Bueno] type-F. Lopez-Bueno 2022’s dose-response MA (48 cohorts, 3.14 M, grip studied over 15-50 kg) reports cancer and CV mortality as «a trend towards a U-shaped association» — inverse at low-moderate grip, then an uptick at the high-strength end. (López-Bueno et al., 2022) Applying adjudicate the arm, not the curve, the authors adjudicate their OWN upper arm as artifact: «the uptick of the dose-response curves at the higher end of the exposure may simply represent lack of data rather than a genuine lack of association … The inversion of the right part of the dose-response curves in this study likely reflect the sparsity of data/events rather than a genuine lack of beneficial association at higher levels of handgrip strength.» (López-Bueno et al., 2022) The distinctive mechanism, as at Mazzoli: unequal between-group precision — thin data/events at high grip displace the spline’s right tail with no bias and no confounder (the gate-6 mechanism), so the apparent high-strength harm is a sampling artifact, never a real upper bound, and the all-cause upper threshold (50 kg) sits exactly at the studied-range edge -> The Underivable Optimum. This is the operative-default direction (a hidden plateau means over-shooting merely fails to help), so the U is read as monotone-inverse with a spurious right tail. Adjudication of the causal (marker-vs-lever) reading is WEAK/none — the paper NEVER addresses reverse causation (srcgrep 0/2: no «reverse caus» / «residual confound»), and its follow-up runs as short as 2.3 y, which amplifies reverse causation relative to a landmark-exclusion cohort -> Grip Strength and Mortality. Not independent-E (a new exposure sharing the one diagnostic; UK Biobank is inside its own pool) -> type-F.
  • Dairy -> cognition — a nadir manufactured by pooling NON-OVERLAPPING population intake ranges [2026-09-05, Villoz] type-F. A new mechanism variant: not reverse causation and not sick-quitter, but a nonlinear dose-response stitched across cohorts whose exposure supports do not overlap. Villoz 2024’s dairy -> cognitive-decline/dementia dose-response MA reports «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» (Villoz et al., 2024) — read at face value, a knee near 150 g/d. But the descending arm is low-intake Asian cohorts (highest-vs-lowest RR 0.83, 0.75-0.92, I2 0%) and the flat arm is high-intake European cohorts (RR 1.01, 0.86-1.19), and the two barely share a dose: «mean value between 170-711 g/d [Europe] … than studies in Asian countries where total mean dairy intake ranged between 29-165 g/d» (Villoz et al., 2024). So the «nadir» is the seam between two populations sampled at different doses, each carrying its own confounding structure (background diet, dairy-type mix), not a within-person optimum. The tell is that the shape is measure-specific — nonlinear on grams but «almost linear» on frequency — a curve that changes with the coding is a property of the pooling, not the biology.
    • This is the Seidelmann two cohorts, two arms of one U instance sharpened. There the two arms were two cohorts on one exposure; here they are also two populations with two confounders, and the protective arm additionally dissolves on dietary-pattern adjustment («studies that took into account other food groups or dietary patterns … found no associations» (Villoz et al., 2024)) — the dairy-as-diet-quality-marker confounder. Adjudication is WEAK/none: no referent-correction (no sick-quitter referent applies), no MR, single-measurement FFQ with differential recall bias conceded. So the lower arm stays unadjudicated, and for the default (Western) intake range the association is flatly null. Not independent-E (a new exposure sharing the one diagnostic) -> type-F -> Dairy and Cognitive Decline. The lesson for the concept: before believing a dose-response knee, check that a single population spans the range — a spline across non-overlapping cohort supports can manufacture a knee at the seam.
  • Very high exercise volume -> mortality / CV outcomes — the extreme-exercise hypothesis upper arm [2026-10-21, Franklin] type-F. The suspect arm here is the upper one: the claim that the most active lose part of the benefit. An AHA scientific statement (consensus, no stated search) walks three cohort cases and offers a hedged explanation for each. (Franklin et al., 2020)
    • Small top groups. «the relatively small sample sizes of the most active sub- groups (typically <5% of the total cohort)» «may» explain the wide intervals; Schnohr’s strenuous joggers were n=36 with 2 deaths, HR 1.97 (0.48–8.14). (Franklin et al., 2020) That is the unequal-between-group-precision mechanism: a thin top cell widens the interval and lets a point estimate drift up without any bias.
    • Confounding. Armstrong’s daily strenuous exercisers (Million Women) had smoking «25.6% versus 13.3%» and lower socioeconomic status «20.1% versus 13.1%» against 2-3x/week exercisers, which «might explain this finding in part». (Franklin et al., 2020)
    • A different exposure, not only a different coding. Lear’s attenuation in the most active «disappeared when recreational and nonrecreational physical activities were combined». (Franklin et al., 2020) Total activity adds occupational activity, which the occupational-activity instance above treats as a separate exposure with its own harm signal, so this result fits the artifact and the causal reading equally well and is not a diagnosis.
    • What the hedges leave unexplained. The statement’s own Table 4 shows the most-active estimate attenuated in every cohort with separate optimal and most-active groups, including two where the top group is not small: Lee 2584 vs 2570 people, HR 0.67 (0.55–0.80) vs 0.77 (0.63–0.92); Maessen 2127 vs 2130, OR 0.31 (0.20−0.48) vs 0.43 (0.28−0.65). (Franklin et al., 2020) Small cells cannot explain a same-direction attenuation in groups of equal size. Most top groups stay below 1.0 against the inactive referent, so the arm is a smaller benefit, not harm.
    • Verdict, as the statement gives it: «In aggregate, there is currently no evidence to reject the curvilinear association (Figure 8A) between exercise volumes and cardiovascular health outcomes, with AF as a possible exception.» (Franklin et al., 2020) AF keeps a J-shape in a nonlinear meta-regression (19 studies, 29,855 AF cases). (Franklin et al., 2020) In cardiac patients the statement calls the evidence U-suggestive but conflicting (Mons, CHD, 7 sessions/wk: HR 1.77 [0.90−3.47]; Stewart, n=15,486, a graded reduction). (Franklin et al., 2020)
    • What it adds to the concept: an upper arm that is attenuation of benefit, not harm vs the referent, and a verdict that splits by outcome (AF) and by stratum (cardiac patients). In the general population the arm is not supported, not refuted: the adjudication is by hedged diagnosis of mechanism, not by the referent-correction or genetic check the decision rule asks of a protective arm, and part of the attenuation stays unexplained. The statement is a narrative review citing these cohorts, not an independent analysis of them -> type-F, no [E-independent]. Applied page: Physical Activity Dose and Mortality.
  • Potassium supplements and BP — a U drawn through randomized trials, whose shape is still between-trial [2026-10-22, Filippini] type-F/G. A 1-stage spline meta-analysis of 32 RCTs (1764 participants) reports a «U‐shaped relationship» with BP rising above an ~80 mmol/d arm difference in urinary potassium. (Filippini et al., 2020) This stretches the concept on two counts: the data are randomized, and the arm under test is a high-dose harm arm, not a protective low arm. Each 2-arm trial contributes one randomized contrast over its own dose interval, so the curve’s shape comes from comparing trials, which randomization does not protect. The recipe’s confounding leg is possible but unsigned: the 6 trials above 80 mmol/d are all crossover, all untreated or normotensive, and 3 come from one group, so the high-dose region is not like-for-like with the rest, and the paper signs those differences both ways. Untreated hypertensives fell more on the forest plot, which would mask a rise: «Among those with hypertension, potassium supplementation was on average more effective in lowering BP in participants not using antihypertensive medication compared with those receiving antihypertensive drug treatment»; though on the achieved-axis spline drug use «did not appear to be associated with major changes in the effect of achieved potassium intake on BP levels»; normotensives (2 of the 6 trials) fell less, which could manufacture one: «a larger BP decrease following potassium supplementation in those with compared with those without hypertension»; and design reverses between analyses: «In an analysis stratified by trial design (crossover versus parallel), the dose‐response analysis showed a larger BP decrease in the latter group, but there was a higher increase in BP in those receiving the largest supplementation, starting at approximately a higher excretion of ≈30 mmol/d, in either the overall population or those with hypertension» (the latter group = parallel), while «The corresponding forest plot analysis showed a larger BP decrease in the crossover studies, in the total sample and analyses restricted to those with hypertension». (Filippini et al., 2020) (inferred from Filippini et al., 2020) These are between-trial averages, not dose-specific, and the one-lab cluster has no known direction, so they neither explain the rise away nor show it to be real. (The paper’s design-stratified spline also reports a rise from ~30 mmol/d among the largest supplementations, read here as the parallel trials, and the >=12-week subset (N=5; the 4 identifiable in Table 1 are all parallel) a BP-raising effect below 60, both figure-only and the latter «statistically imprecise»; so the rise is not confined to the crossover trials, though no parallel trial reaches the >80 region.) The problem that does the work is not on the list: sparse-tail imprecision. Those 6 trials hold 109 participants, and every reported interval at the high end (90 and 120 difference; 150, the plateau’s edge, and 180 achieved) includes 0. That is random error, which leaves the arm unestimated rather than manufactured. The paper also places the rise in different subgroups in its abstract, Results and Discussion, and runs no formal non-linearity test. Verdict: the BP fall from moderate supplementation is robust as a pooled randomized contrast (forest SBP -3.9, -5.2 to -2.6; the spline puts -3.3, -4.9 to -1.6, at a 30 mmol/d difference). The shape of both arms is between-trial. The low end is better populated and its intervals exclude 0 (+3.9, 2.1 to 5.8, at 60 vs 90 mmol/d achieved), though the +9.1 at 30 mmol/d sits below the lowest control arm (34 mmol/d). The upper arm is unadjudicated. (Filippini et al., 2020) The rule this adds: pooled from RCTs does not make a dose-response shape randomized. The strong check is not MR or a referent correction but a within-trial multi-dose comparison. The rule applies back to the alcohol->BP entry above: Roerecke’s closest-to-randomized part is each trial’s own reduction->BP change (some of its reduction groups are reducing control arms: «If a control group in a parallel-group trial reported a reduction in alcohol consumption, we included it as a “reduction of alcohol” group.» (Roerecke et al., 2017)); its threshold at two drinks per day is between-trial too. Applied page: Potassium Intake and Blood Pressure.
  • The open telos prior. This concept is the fabric form of [PRIOR] #2 (U/J-shapes as artifacts): it supplies the mechanism list and the adjudication routes so the prior can be scored against concrete cases, rather than asserted.

Corrections and revision history (dated strata)

Dated self-critique and audit strata, preserved in their original form.

Self-critique [run 2026-07-28, before commit]

  • Not laundered from one source. The mechanism and adjudication routes are induced across Stockwell (bias-correction) and Millwood (MR), and the concept adds what neither states: the ranking of the three checks (early-follow-up exclusion is weak; referent/MR are strong) and the transfer to sodium’s open case. It is not the alcohol page restated under a general title.

  • Not a platitude. Suspect U-shapes alone would be one; the decision rule is specific and falsifiable (which checks a protective arm must survive), and it changes what a reader does with a J-curve.

  • Confirmed on two exposures, and a prediction landed. Alcohol (whole protective arm artifact) and sleep (long arm artifact, short arm causal) are both adjudicated instances; the concept predicted sleep would show the pattern and it did -> Sleep Duration and Mortality. Sodium’s low-intake arm stays open. The concept is now banked on two exposures with an arm-level refinement, not one — though both rest on the same reverse-causation mechanism, so a genuinely different artifact route (e.g. the unequal-precision one) is still untested.

  • Coffee added a directional sharpening [2026-08-04]. Grosso’s smoking referent-correction is the second route-1 instance (after alcohol), and it guards against a latent over-generalization: that the correction always kills a protective signal. It does not — in coffee-cancer the correction dissolved a confounder-manufactured harm and revealed more benefit. So the rule is “the correction removes what the confounder created,” not “the correction removes the protection.” This is a refinement, not a laundered repeat (it changes what the concept claims). It is honestly scoped: the clean instance is the cancer sign-flip (Grosso-attributed), the all-cause/CVD linearization is a flagged inference (Grosso reports no smoker/non-smoker difference there), and the whole thing is partial (referent-correction only, MR null), so the coffee lower arm is left not-established rather than over-sold as adjudicated.

  • Carbohydrate added a distinct refinement [2026-08-06, Seidelmann]. This is the first instance where the adjudication reveals the exposure (not just a confounder) is under-specified: «low carb» mortality flips sign with the replacement source (animal 1.18 vs plant 0.82), so the arm decomposes into a substitution rather than dissolving under a confounder-strip. It is NOT a protective-lower-arm case (both arms harmful), so it stretches the concept from “adjudicate the protective arm” toward “decompose the exposure before believing the curve.” Honestly scoped: weak checks only (no MR), so the causal reading stays not-established; and it reconciles with — not independently backs — the Dehghan/PURE bullet (shared observational-FFQ-cohort base, PURE shared between them), so type-F, not E.

  • Coherence, not validity (R1): the concept says when a protective arm is unwarranted, not that the exposure is harmful; it is a rule about evidence, not about the world.

References

Aune, D., Keum, N., Giovannucci, E., Fadnes, L. T., Boffetta, P., Greenwood, D. C., Tonstad, S., Vatten, L. J., Riboli, E., & Norat, T. (2016). Nut consumption and risk of cardiovascular disease, total cancer, all-cause and cause-specific mortality: a systematic review and dose-response meta-analysis of prospective studies. BMC Medicine, 14(1). https://doi.org/10.1186/s12916-016-0730-3
Bagnardi, V., Rota, M., Botteri, E., Tramacere, I., Islami, F., Fedirko, V., Scotti, L., Jenab, M., Turati, F., Pasquali, E., Pelucchi, C., Galeone, C., Bellocco, R., Negri, E., Corrao, G., Boffetta, P., & La Vecchia, C. (2014). Alcohol consumption and site-specific cancer risk: a comprehensive dose–response meta-analysis. British Journal of Cancer, 112(3), 580–593. https://doi.org/10.1038/bjc.2014.579
Bryazka, D., Reitsma, M. B., Griswold, M. G., Abate, K. H., Abbafati, C., Abbasi-Kangevari, M., Abbasi-Kangevari, Z., Abdoli, A., Abdollahi, M., Abdullah, A. Y. M., Abhilash, E. S., Abu-Gharbieh, E., Acuna, J. M., Addolorato, G., Adebayo, O. M., Adekanmbi, V., Adhikari, K., Adhikari, S., Adnani, Q. E. S., … Gakidou, E. (2022). Population-level risks of alcohol consumption by amount, geography, age, sex, and year: a systematic analysis for the Global Burden of Disease Study 2020. The Lancet, 400(10347), 185–235. https://doi.org/10.1016/s0140-6736(22)00847-9
Cappuccio, F. P., D’Elia, L., Strazzullo, P., & Miller, M. A. (2010). Sleep Duration and All-Cause Mortality: A Systematic Review and Meta-Analysis of Prospective Studies. Sleep, 33(5), 585–592. https://doi.org/10.1093/sleep/33.5.585
Coenen, P., Huysmans, M. A., Holtermann, A., Krause, N., van Mechelen, W., Straker, L. M., & van der Beek, A. J. (2018). Do highly physically active workers die early? A systematic review with meta-analysis of data from 193 696 participants. British Journal of Sports Medicine, 52(20), 1320–1326. https://doi.org/10.1136/bjsports-2017-098540
Dehghan, M., Mente, A., Zhang, X., Swaminathan, S., Li, W., Mohan, V., Iqbal, R., Kumar, R., Wentzel-Viljoen, E., Rosengren, A., Amma, L. I., Avezum, A., Chifamba, J., Diaz, R., Khatib, R., Lear, S., Lopez-Jaramillo, P., Liu, X., Gupta, R., … Mapanga, R. (2017). Associations of fats and carbohydrate intake with cardiovascular disease and mortality in 18 countries from five continents (PURE): a prospective cohort study. The Lancet, 390(10107), 2050–2062. https://doi.org/10.1016/s0140-6736(17)32252-3
Di Federico, S., Filippini, T., Whelton, P. K., Cecchini, M., Iamandii, I., Boriani, G., & Vinceti, M. (2023). Alcohol Intake and Blood Pressure Levels: A Dose-Response Meta-Analysis of Nonexperimental Cohort Studies. Hypertension, 80(10), 1961–1969. https://doi.org/10.1161/hypertensionaha.123.21224
Ekelund, U., Tarp, J., Steene-Johannessen, J., Hansen, B. H., Jefferis, B., Fagerland, M. W., Whincup, P., Diaz, K. M., Hooker, S. P., Chernofsky, A., Larson, M. G., Spartano, N., Vasan, R. S., Dohrn, I.-M., Hagströmer, M., Edwardson, C., Yates, T., Shiroma, E., Anderssen, S. A., & Lee, I.-M. (2019). Dose-response associations between accelerometry measured physical activity and sedentary time and all cause mortality: systematic review and harmonised meta-analysis. BMJ, l4570. https://doi.org/10.1136/bmj.l4570
Filippini, T., Naska, A., Kasdagli, M., Torres, D., Lopes, C., Carvalho, C., Moreira, P., Malavolti, M., Orsini, N., Whelton, P. K., & Vinceti, M. (2020). Potassium Intake and Blood Pressure: A Dose‐Response Meta‐Analysis of Randomized Controlled Trials. Journal of the American Heart Association, 9(12). https://doi.org/10.1161/jaha.119.015719
Franklin, B. A., Thompson, P. D., Al-Zaiti, S. S., Albert, C. M., Hivert, M.-F., Levine, B. D., Lobelo, F., Madan, K., Sharrief, A. Z., & Eijsvogels, T. M. H. (2020). Exercise-Related Acute Cardiovascular Events and Potential Deleterious Adaptations Following Long-Term Exercise Training: Placing the Risks Into Perspective–An Update: A Scientific Statement From the American Heart Association. Circulation, 141(13). https://doi.org/10.1161/cir.0000000000000749
Global BMI Mortality Collaboration. (2016). Body-mass index and all-cause mortality: individual-participant-data meta-analysis of 239 prospective studies in four continents. The Lancet, 388(10046), 776–786. https://doi.org/10.1016/s0140-6736(16)30175-1
Godos, J., Micek, A., Brzostek, T., Toledo, E., Iacoviello, L., Astrup, A., Franco, O. H., Galvano, F., Martinez-Gonzalez, M. A., & Grosso, G. (2020). Egg consumption and cardiovascular risk: a dose–response meta-analysis of prospective cohort studies. European Journal of Nutrition, 60(4), 1833–1862. https://doi.org/10.1007/s00394-020-02345-7
Graudal, N. A., Hubeck-Graudal, T., & Jurgens, G. (2020). Effects of low sodium diet versus high sodium diet on blood pressure, renin, aldosterone, catecholamines, cholesterol, and triglyceride. Cochrane Database of Systematic Reviews, 2021(8). https://doi.org/10.1002/14651858.cd004022.pub5
Grosso, G., Micek, A., Godos, J., Sciacca, S., Pajak, A., Martínez-González, M. A., Giovannucci, E. L., & Galvano, F. (2016). Coffee consumption and risk of all-cause, cardiovascular, and cancer mortality in smokers and non-smokers: a dose-response meta-analysis. European Journal of Epidemiology, 31(12), 1191–1205. https://doi.org/10.1007/s10654-016-0202-2
Guo, J., Astrup, A., Lovegrove, J. A., Gijsbers, L., Givens, D. I., & Soedamah-Muthu, S. S. (2017). Milk and dairy consumption and risk of cardiovascular diseases and all-cause mortality: dose–response meta-analysis of prospective cohort studies. European Journal of Epidemiology, 32(4), 269–287. https://doi.org/10.1007/s10654-017-0243-1
He, F. J., Li, J., & MacGregor, G. A. (2013). Effect of longer term modest salt reduction on blood pressure: Cochrane systematic review and meta-analysis of randomised trials. BMJ, 346(apr03 3), f1325–f1325. https://doi.org/10.1136/bmj.f1325
Huang, L., Trieu, K., Yoshimura, S., Neal, B., Woodward, M., Campbell, N. R. C., Li, Q., Lackland, D. T., Leung, A. A., Anderson, C. A. M., MacGregor, G. A., & He, F. J. (2020). Effect of dose and duration of reduction in dietary sodium on blood pressure levels: systematic review and meta-analysis of randomised trials. BMJ, m315. https://doi.org/10.1136/bmj.m315
Iso-Markku, P., Kujala, U. M., Knittle, K., Polet, J., Vuoksimaa, E., & Waller, K. (2022). Physical activity as a protective factor for dementia and Alzheimer’s disease: systematic review, meta-analysis and quality assessment of cohort and case–control studies. British Journal of Sports Medicine, 56(12), 701–709. https://doi.org/10.1136/bjsports-2021-104981
Jayedi, A., & Shab-Bidar, S. (2020). Fish Consumption and the Risk of Chronic Disease: An Umbrella Review of Meta-Analyses of Prospective Cohort Studies. Advances in Nutrition, 11(5), 1123–1133. https://doi.org/10.1093/advances/nmaa029
Jayedi, A., Soltani, S., Zargar, M. S., Khan, T. A., & Shab-Bidar, S. (2020). Central fatness and risk of all cause mortality: systematic review and dose-response meta-analysis of 72 prospective cohort studies. BMJ, m3324. https://doi.org/10.1136/bmj.m3324
Jha, P., Ramasundarahettige, C., Landsman, V., Rostron, B., Thun, M., Anderson, R. N., McAfee, T., & Peto, R. (2013). 21st-Century Hazards of Smoking and Benefits of Cessation in the United States. New England Journal of Medicine, 368(4), 341–350. https://doi.org/10.1056/nejmsa1211128
Katagiri, R., Yuan, X., Kobayashi, S., & Sasaki, S. (2017). Effect of excess iodine intake on thyroid diseases in different populations: A systematic review and meta-analyses including observational studies. PLOS ONE, 12(3), e0173722. https://doi.org/10.1371/journal.pone.0173722
Kelly, J. T., Su, G., Zhang, L., Qin, X., Marshall, S., González-Ortiz, A., Clase, C. M., Campbell, K. L., Xu, H., & Carrero, J.-J. (2020). Modifiable Lifestyle Factors for Primary Prevention of CKD: A Systematic Review and Meta-Analysis. Journal of the American Society of Nephrology, 32(1), 239–253. https://doi.org/10.1681/asn.2020030384
Liu, Y., Clarke, R., Bennett, D. A., Zong, G., & Gan, W. (2024). Iron Status and Risk of Heart Disease, Stroke, and Diabetes: A Mendelian Randomization Study in European Adults. Journal of the American Heart Association, 13(6). https://doi.org/10.1161/jaha.123.031732
Livingston, G., Huntley, J., Liu, K. Y., Costafreda, S. G., Selbæk, G., Alladi, S., Ames, D., Banerjee, S., Burns, A., Brayne, C., Fox, N. C., Ferri, C. P., Gitlin, L. N., Howard, R., Kales, H. C., Kivimäki, M., Larson, E. B., Nakasujja, N., Rockwood, K., … Mukadam, N. (2024). Dementia prevention, intervention, and care: 2024 report of the Lancet standing Commission. The Lancet, 404(10452), 572–628. https://doi.org/10.1016/s0140-6736(24)01296-0
López-Bueno, R., Andersen, L. L., Koyanagi, A., Núñez-Cortés, R., Calatayud, J., Casaña, J., & del Pozo Cruz, B. (2022). Thresholds of handgrip strength for all-cause, cancer, and cardiovascular mortality: A systematic review with dose-response meta-analysis. Ageing Research Reviews, 82, 101778. https://doi.org/10.1016/j.arr.2022.101778
Ma, W., Zhang, Y., Pan, L., Wang, S., Xie, K., Deng, S., Wang, R., Guo, C., Qin, P., Wu, X., Wu, Y., Zhao, Y., Feng, Y., & Hu, F. (2022). Association of Egg Consumption with Risk of All-Cause and Cardiovascular Disease Mortality: A Systematic Review and Dose–Response Meta-Analysis of Observational Studies. The Journal of Nutrition, 152(10), 2227–2237. https://doi.org/10.1093/jn/nxac105
Malmir, H., Larijani, B., & Esmaillzadeh, A. (2019). Consumption of milk and dairy products and risk of osteoporosis and hip fracture: a systematic review and Meta-analysis. Critical Reviews in Food Science and Nutrition, 60(10), 1722–1737. https://doi.org/10.1080/10408398.2019.1590800
Mazzoli, R., Filippini, T., Iamandii, I., De Pasquale, L., Veneri, F., Birnbaum, L. S., Rothman, K. J., & Vinceti, M. (2025). The association of fluoride exposure with bone density and fracture risk: a dose-response meta-analysis. Environmental Health, 24(1). https://doi.org/10.1186/s12940-025-01226-y
Mente, A., O’Donnell, M., Rangarajan, S., Dagenais, G., Lear, S., McQueen, M., Diaz, R., Avezum, A., Lopez-Jaramillo, P., Lanas, F., Li, W., Lu, Y., Yi, S., Rensheng, L., Iqbal, R., Mony, P., Yusuf, R., Yusoff, K., Szuba, A., … Yusuf, S. (2016). Associations of urinary sodium excretion with cardiovascular events in individuals with and without hypertension: a pooled analysis of data from four studies. The Lancet, 388(10043), 465–475. https://doi.org/10.1016/s0140-6736(16)30467-6
Momma, H., Kawakami, R., Honda, T., & Sawada, S. S. (2022). Muscle-strengthening activities are associated with lower risk and mortality in major non-communicable diseases: a systematic review and meta-analysis of cohort studies. British Journal of Sports Medicine, 56(13), 755–763. https://doi.org/10.1136/bjsports-2021-105061
National Academies of Sciences Engineering and Medicine. (2019). Reproducibility and Replicability in Science. National Academies Press. https://doi.org/10.17226/25303
Nordestgaard, A. T., & Nordestgaard, B. G. (2016). Coffee intake, cardiovascular disease and all-cause mortality: observational and Mendelian randomization analyses in 95 000–223 000 individuals. International Journal of Epidemiology, dyw325. https://doi.org/10.1093/ije/dyw325
Paluch, A. E., Bajpai, S., Bassett, D. R., Carnethon, M. R., Ekelund, U., Evenson, K. R., Galuska, D. A., Jefferis, B. J., Kraus, W. E., Lee, I.-M., Matthews, C. E., Omura, J. D., Patel, A. V., Pieper, C. F., Rees-Punia, E., Dallmeier, D., Klenk, J., Whincup, P. H., Dooley, E. E., … Fulton, J. E. (2022). Daily steps and all-cause mortality: a meta-analysis of 15 international cohorts. The Lancet Public Health, 7(3), e219–e228. https://doi.org/10.1016/s2468-2667(21)00302-9
Peters, R., Xu, Y., Fitzgerald, O., Aung, H. L., Beckett, N., Bulpitt, C., Chalmers, J., Forette, F., Gong, J., Harris, K., Humburg, P., Matthews, F. E., Staessen, J. A., Thijs, L., Tzourio, C., Warwick, J., Woodward, M., & Anderson, C. S. (2022). Blood pressure lowering and prevention of dementia: an individual patient data meta-analysis. European Heart Journal, 43(48), 4980–4990. https://doi.org/10.1093/eurheartj/ehac584
Poole, R., Kennedy, O. J., Roderick, P., Fallowfield, J. A., Hayes, P. C., & Parkes, J. (2017). Coffee consumption and health: umbrella review of meta-analyses of multiple health outcomes. BMJ, j5024. https://doi.org/10.1136/bmj.j5024
Qin, P., Huang, C., Jiang, B., Wang, X., Yang, Y., Ma, J., Chen, S., Hu, D., & Bo, Y. (2023). Dietary carbohydrate quantity and quality and risk of cardiovascular disease, all-cause, cardiovascular and cancer mortality: A systematic review and meta-analysis. Clinical Nutrition, 42(2), 148–165. https://doi.org/10.1016/j.clnu.2022.12.010
Rahimi, K., Bidel, Z., Nazarzadeh, M., Copland, E., Canoy, D., Wamil, M., Majert, J., McManus, R., Adler, A., Agodoa, L., Algra, A., Asselbergs, F. W., Beckett, N. S., Berge, E., Black, H., Boersma, E., Brouwers, F. P. J., Brown, M., Brugts, J. J., … Teo, K. K. (2021). Age-stratified and blood-pressure-stratified effects of blood-pressure-lowering pharmacotherapy for the prevention of cardiovascular disease and death: an individual participant-level data meta-analysis. The Lancet, 398(10305), 1053–1064. https://doi.org/10.1016/s0140-6736(21)01921-8
Ramsden, C. E., Zamora, D., Majchrzak-Hong, S., Faurot, K. R., Broste, S. K., Frantz, R. P., Davis, J. M., Ringel, A., Suchindran, C. M., & Hibbeln, J. R. (2016). Re-evaluation of the traditional diet-heart hypothesis: analysis of recovered data from Minnesota Coronary Experiment (1968-73). BMJ, i1246. https://doi.org/10.1136/bmj.i1246
Roerecke, M., Kaczorowski, J., Tobe, S. W., Gmel, G., Hasan, O. S. M., & Rehm, J. (2017). The effect of a reduction in alcohol consumption on blood pressure: a systematic review and meta-analysis. The Lancet Public Health, 2(2), e108–e120. https://doi.org/10.1016/s2468-2667(17)30003-8
Saeidifard, F., Medina-Inojosa, J. R., West, C. P., Olson, T. P., Somers, V. K., Bonikowske, A. R., Prokop, L. J., Vinciguerra, M., & Lopez-Jimenez, F. (2019). The association of resistance training with mortality: A systematic review and meta-analysis. European Journal of Preventive Cardiology, 26(15), 1647–1665. https://doi.org/10.1177/2047487319850718
Seidelmann, S. B., Claggett, B., Cheng, S., Henglin, M., Shah, A., Steffen, L. M., Folsom, A. R., Rimm, E. B., Willett, W. C., & Solomon, S. D. (2018). Dietary carbohydrate intake and mortality: a prospective cohort study and meta-analysis. The Lancet Public Health, 3(9), e419–e428. https://doi.org/10.1016/s2468-2667(18)30135-x
Shailendra, P., Baldock, K. L., Li, L. S. K., Bennie, J. A., & Boyle, T. (2022). Resistance Training and Mortality Risk: A Systematic Review and Meta-Analysis. American Journal of Preventive Medicine, 63(2), 277–285. https://doi.org/10.1016/j.amepre.2022.03.020
Shan, Z., Ma, H., Xie, M., Yan, P., Guo, Y., Bao, W., Rong, Y., Jackson, C. L., Hu, F. B., & Liu, L. (2015). Sleep Duration and Risk of Type 2 Diabetes: A Meta-analysis of Prospective Studies. Diabetes Care, 38(3), 529–537. https://doi.org/10.2337/dc14-2073
Sjöström, L., Narbro, K., Sjöström, C. D., Karason, K., Larsson, B., Wedel, H., Lystig, T., Sullivan, M., Bouchard, C., Carlsson, B., Bengtsson, C., Dahlgren, S., Gummesson, A., Jacobson, P., Karlsson, J., Lindroos, A.-K., Lönroth, H., Näslund, I., Olbers, T., … Carlsson, L. M. S. (2007). Effects of Bariatric Surgery on Mortality in Swedish Obese Subjects. New England Journal of Medicine, 357(8), 741–752. https://doi.org/10.1056/nejmoa066254
Stockwell, T., Zhao, J., Panwar, S., Roemer, A., Naimi, T., & Chikritzhs, T. (2016). Do “Moderate” Drinkers Have Reduced Mortality Risk? A Systematic Review and Meta-Analysis of Alcohol Consumption and All-Cause Mortality. Journal of Studies on Alcohol and Drugs, 77(2), 185–198. https://doi.org/10.15288/jsad.2016.77.185
Sun, Y.-Q., Burgess, S., Staley, J. R., Wood, A. M., Bell, S., Kaptoge, S. K., Guo, Q., Bolton, T. R., Mason, A. M., Butterworth, A. S., Di Angelantonio, E., Vie, G. Å., Bjørngaard, J. H., Kinge, J. M., Chen, Y., & Mai, X.-M. (2019). Body mass index and all cause mortality in HUNT and UK Biobank studies: linear and non-linear mendelian randomisation analyses. BMJ, l1042. https://doi.org/10.1136/bmj.l1042
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
Wade, K. H., Carslake, D., Sattar, N., Davey Smith, G., & Timpson, N. J. (2018). BMI and Mortality in UK Biobank: Revised Estimates Using Mendelian Randomization. Obesity, 26(11), 1796–1806. https://doi.org/10.1002/oby.22313
Wang, S., Li, Z., Wang, X., Guo, S., Sun, Y., Li, G., Zhao, C., Yuan, W., Li, M., Li, X., & Ai, S. (2022). Associations between sleep duration and cardiovascular diseases: A meta-review and meta-analysis of observational and Mendelian randomization studies. Frontiers in Cardiovascular Medicine, 9. https://doi.org/10.3389/fcvm.2022.930000
World Health Organization. (2023). Use of non-sugar sweeteners: WHO guideline. https://www.who.int/publications/i/item/9789240073616
Xu, W., Wang, H., Wan, Y., Tan, C., Li, J., Tan, L., & Yu, J.-T. (2017). Alcohol consumption and dementia risk: a dose–response meta-analysis of prospective studies. European Journal of Epidemiology, 32(1), 31–42. https://doi.org/10.1007/s10654-017-0225-3
Yin, J., Jin, X., Shan, Z., Li, S., Huang, H., Li, P., Peng, X., Peng, Z., Yu, K., Bao, W., Yang, W., Chen, X., & Liu, L. (2017). Relationship of Sleep Duration With All‐Cause Mortality and Cardiovascular Events: A Systematic Review and Dose‐Response Meta‐Analysis of Prospective Cohort Studies. Journal of the American Heart Association, 6(9). https://doi.org/10.1161/jaha.117.005947
Zhang, J., Ou, J., Lu, X., Wang, T., Dang, W., Ding, L., Liu, Y., Xu, J., Yan, B., & Yu, H. (2025). Sleep disorders and the risk of cognitive decline or dementia: an updated systematic review and meta-analysis of longitudinal studies. Journal of Neurology, 272(10). https://doi.org/10.1007/s00415-025-13372-x
Zhang, Y., Lee, D. H., Rezende, L. F. M., Ma, Y., & Giovannucci, E. (2026). Long-term resistance training with all-cause and cause-specific mortality: assessing dose-response and joint associations with aerobic physical activity. British Journal of Sports Medicine, 60(12), 874–883. https://doi.org/10.1136/bjsports-2025-110503