The bias-corrected shape of the BMI to all-cause-mortality relationship, from the largest individual-participant-data meta-analysis of the question: 239 prospective cohorts, four continents, 10.6 million participants (the primary bias-controlled analysis draws on 189 studies, 3.95M never-smokers, 385,879 deaths). Its design is built to answer the artifact question directly -> The U-Shaped Association Artifact: the pre-specified primary analysis restricts to never-smokers, drops anyone with chronic disease at baseline, and excludes the first 5 years of follow-up — «Primary analyses will also exclude current and former smokers, and the first 5 years of follow-up.» (Global BMI Mortality Collaboration, 2016)
The corrected curve — nadir at 22.5-25, monotone rise above it
In the bias-controlled analysis the curve is J-shaped with a nadir at BMI 22.5-25 kg/m2 (the pre-specified reference category), a shallow, mostly reverse-causation underweight arm to its left, and a steady monotone rise above 25 with no plateau in the studied range. Per 5 kg/m2 above BMI 25, all-cause mortality rises HR 1.31 (1.29-1.33) overall. (Global BMI Mortality Collaboration, 2016)
WHO 6-group hazard ratios, primary analysis (never-smokers, no baseline chronic disease, first 5 years excluded; reference = normal weight 18.5-25):
| BMI group | HR (95% CI) |
|---|---|
| Underweight 15 to <18.5 | 1.47 (1.39-1.55) |
| Normal 18.5 to <25 | 1.00 (reference) |
| Overweight 25 to <30 | 1.11 (1.10-1.11) |
| Obesity I 30 to <35 | 1.44 (1.41-1.47) |
| Obesity II 35 to <40 | 1.92 (1.86-1.98) |
| Obesity III 40 to <60 | 2.71 (2.55-2.86) |
(Global BMI Mortality Collaboration, 2016)
Stratum-dependence (steepness, per 5 kg/m2 above 25) — the relative effect is real but not uniform:
- Age — steepest in the young, attenuating with age: 35-49 y 1.52 (1.47-1.56), 50-69 y 1.37 (1.35-1.39), 70-89 y 1.21 (1.17-1.25). (Global BMI Mortality Collaboration, 2016)
- Sex — steeper in men: men 1.51 (1.46-1.56), women 1.30 (1.26-1.33). (Global BMI Mortality Collaboration, 2016)
- Region — Europe 1.39, North America 1.29, East Asia 1.39, Australia/NZ 1.31; South Asia 1.13 (0.97-1.30) is the one non-significant region — but on only 3 studies / 4,040 deaths, so its flatness is as likely small-n imprecision as a real regional difference (East Asia, with 46 studies, shows the same steep 1.39 as Europe). (Global BMI Mortality Collaboration, 2016)
- Cause — CHD 1.42, stroke 1.42, respiratory 1.38, cancer 1.19 (cancer the shallowest). (Global BMI Mortality Collaboration, 2016)
The obesity paradox is a bias-cascade artifact — watch the overweight arm move
The paper’s pre-specified design is a stepwise confounder-stripping cascade, and the overweight arm (25-30) walks from apparent protection to clear harm as each bias is removed — a worked BMI instance of the artifact concept -> The U-Shaped Association Artifact. Overweight HR vs normal weight, across the eTable 5 cascade (189 studies):
| Analysis (cumulative exclusions) | Overweight HR (25-30) |
|---|---|
| All studies, no exclusions (raw) | 0.96 (0.95-0.97) — apparent protection |
| + adjust for smoking (+ exclude baseline disease) | 0.99 (0.98-1.01) |
| + exclude first 5 years of follow-up | 1.03 (1.01-1.04) |
| + restrict to never-smokers (primary) | 1.11 (1.10-1.11) — clear harm |
(Global BMI Mortality Collaboration, 2016)
(inferred from Global BMI Mortality Collaboration, 2016) The obesity paradox (overweight appearing protective for mortality) is, on this evidence, manufactured by smoking confounding, reverse causation, and prevalent disease — the three biases the cascade removes. Smokers are leaner and die more, so they load the normal-weight and underweight referent with high-risk people and make overweight look protective by contrast; early deaths and baseline illness do the same via reverse causation (illness lowers weight before it kills). Strip all three and the protection inverts to harm. The obesity grades steepen in parallel (Obesity I 1.18 -> 1.44; Obesity III 1.96 -> 2.71 across the same cascade). This is the direction-of-artifact lesson from the concept: a confounder here manufactured a spurious benefit (as in alcohol), the opposite of the coffee-cancer case.
The single cleanest confirmation — the smoking-stratified contrast (eTable 6). Holding the 5-year exclusion and no-baseline-disease constant so only smoking status differs, the overweight arm flips sign by smoking stratum:
| BMI group | Never-smokers | Ex/current smokers |
|---|---|---|
| Underweight 15-18.5 | 1.53 (1.39-1.68) | 1.86 (1.74-1.99) |
| Overweight 25-30 | 1.07 (1.06-1.07) | 0.94 (0.94-0.95) |
| Obesity I 30-35 | 1.39 (1.33-1.44) | 1.13 (1.09-1.17) |
| Obesity III 40-60 | 2.69 (2.43-2.97) | 2.04 (1.85-2.24) |
(Global BMI Mortality Collaboration, 2016)
In ex/current smokers overweight looks protective (0.94) and the whole obesity gradient is flatter; in never-smokers overweight is harmful (1.07) and every grade is steeper. The never-smoker-vs-smoker heterogeneity is significant for every BMI group (underweight P=0.046, overweight P=0.0003, obesity I P<0.0001, obesity II P=0.0004, obesity III P=0.0003). Smoking is the confounder doing the work.
The underweight arm — adjudicate the arm, not the curve
The two arms behave differently, per the concept’s adjudicate the arm rule. The overweight arm is entirely artifact (protection -> harm under correction). The underweight arm is partly reverse causation but not wholly: it falls from 1.81 (raw) to 1.47 (primary) as smokers and early deaths are removed, yet stays elevated at 1.47 — so low BMI carries residual real excess mortality that the corrections do not dissolve, distinct from the fully-artifactual overweight signal. (inferred from Global BMI Mortality Collaboration, 2016) Refined below (Sun MR): this never-smoker residual is itself largely a smoking phenomenon (respiratory-death-driven, evident only in ever-smokers) — condition on smoking before reading the low arm.
The genetic-instrument confirmation — Wade MR cashes the missing strong check [2026-08-06, Wade]
The Global BMI curve above is adjudicated by confounder-removal only (no genetic instrument — the gap flagged in Limits). Wade 2018’s Mendelian randomization in UK Biobank (335,308 White British, 9,570 deaths; a 77-SNP BMI genetic risk score used as an instrumental variable) supplies that check — the genetic natural experiment carries orthogonal biases to the observational cascade (immune to the reverse causation and smoking confounding the cascade removes by exclusion), so it adjudicates the same arms by an independent route.
The causal MR estimates (per 1 kg/m2 higher BMI): «MR analyses supported a causal association between higher BMI and greater risk of all-cause mortality (hazard ratio [HR] per 1 kg/m2: 1.03; 95% CI: 0.99-1.07) and mortality from cardiovascular diseases (HR: 1.10; 95% CI: 1.01-1.19), specifically coronary heart disease (HR: 1.12; 95% CI: 1.00-1.25) and those excluding coronary heart disease/stroke/aortic aneurysm (HR: 1.24; 95% CI: 1.03-1.48), stomach cancer (HR: 1.18; 95% CI: 0.87-1.62), and esophageal cancer (HR: 1.22; 95% CI: 0.98-1.53)». (Wade et al., 2018) The all-cause point estimate (1.03/unit; scaled ~16% per 5 kg/m2, 95% CI -5% to +41%) is directionally supportive but imprecise — its CI crosses the null, and the Durbin-Wu-Hausman test finds no significant observational-vs-MR difference (P=0.96). The CVD arm is where MR reaches significance (1.10, 1.01-1.19), and MR estimates are of «similar or greater magnitude to observational analyses (with wider CIs)». (Wade et al., 2018)
What MR says about the U-curve — the low arm largely deflates, the high arm is genetically corroborated (significant for CVD-cause mortality; directional but imprecise for all-cause). «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) The nadir shifts DOWN from ~26 (observational) into the normal range (~23), and the residual J is driven by the extreme BMI quantiles — removing them yields a linear association (P=0.999 for linear trend). Wade names the mechanism: «Reverse causality is an important source of bias in observational estimates … and may be the driver of the characteristic J-shaped association», so 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 direction of the correction is the payoff: MR deflates the underweight arm (reverse causation) and inflates the obesity arm — the mirror image of the confounder-strip cascade above, reached genetically. Severe underweight plausibly keeps real harm (Wade concedes it «is plausible that individuals considered to have severe and unhealthy underweight have a higher risk of mortality»), converging with the Global BMI residual underweight 1.47. (Wade et al., 2018)
Independence verdict — type-F refinement, NOT independent-E (the lineage chase mattered). Before counting Wade as an independent genetic witness, build the parameter table and chase the authorship:
| Parameter | Global BMI 2016 (corrected observational IPD-MA) | Wade 2018 (Mendelian randomization) | Same quantity? |
|---|---|---|---|
| Adjudication route | never-smoker restriction + drop first 5 y follow-up + drop baseline disease | 77-SNP genetic instrument (IV ratio estimate) | NO — different method class (this IS the F-refinement axis) |
| Nadir (BMI at min mortality) | 22.5-25 kg/m2 (pre-specified referent) | ~23 kg/m2 (MR), vs ~26 observational in the same UKB | YES — same construct, convergent |
| Underweight / low arm | 1.47, residual after correction (partly reverse causation, not wholly) | J flattens, nadir drops; reverse causation named the J’s driver; severe underweight plausibly real | YES — same arm, convergent mechanism |
| Above-nadir (overweight/obese) arm | monotone harm, HR 1.31 per 5 kg/m2 above 25 (corrected association) | causal MR harm; obesity harm underestimated by observational (CVD 1.10/unit) | YES — same arm; Wade adds the causal warrant |
The estimates converge on the nadir, on the low-arm-is-reverse-causation reading, and on causal harm
above the nadir. But they are not independent. Wade cites Global BMI (ref 5) as corroborative
context, and — decisively — 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 in the
2016 author list). Two shared authors, including the anchor MR investigator, means the two estimates come
from an overlapping group, so this is a same-lineage type-F refinement (the later source supplies the
earlier’s missing genetic-instrument leg) — not independent-E backing. The convergence is genuine and
cashes the missing strong check, but it must not be counted as independent corroboration.
(inferred from Global BMI Mortality Collaboration, 2016; Wade et al., 2018)
Sun MR — the single J is at least TWO curves; the underweight arm is largely a smoking phenomenon [2026-08-19, Sun]
A second, non-linear MR (Sun 2019: HUNT Norway 56,150 + UK Biobank 366,385; fractional-polynomial meta-regression across 100 residual-BMI strata) both confirms the nadir and decomposes the curve. The overall MR J agrees with the page above — «The lowest risk was at a BMI of around 22-25» (HUNT nadir ~22-23, UKB ~25) (Sun et al., 2019) — converging with the 22.5-25 corrected-observational nadir and Wade’s ~23. Linear MR (per 1-unit genetically-predicted BMI) shows the same category split: overall 1.04 (1.02-1.06), obese 1.09 (1.04-1.14) harm, but underweight 0.66 (0.52-0.84) — in the underweight stratum raising BMI reduces mortality. (Sun et al., 2019)
The decomposition is the new finding. Stratifying by smoking splits the aggregate J into two different shapes: «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) In never-smokers Sun finds «no evidence for a harmful effect of reducing BMI in underweight participants» — clearest in HUNT (a positive slope throughout under/normal/overweight), while in UKB «confidence intervals were wide and compatible with a null effect at all values of BMI». (Sun et al., 2019)
Refinement of the underweight arm above. The section above (from Global BMI) holds the never-smoker underweight residual (1.47) as partly real harm. Sun sharpens, without overturning, that reading: the underweight-mortality risk concentrates in ever-smokers, where «Increased mortality in underweight smokers might be driven by respiratory diseases» — the non-CVD-non-cancer «other» category that carries the only profound J. (Sun et al., 2019) This is NOT a filed tension with Global BMI: Global BMI’s 1.47 is an observational association within never-smokers (residual reverse causation not removed by the never-smoker restriction), Sun’s is a genetic causal slope on a small (1-3% of sample) and imprecise underweight stratum — the observational-vs-MR discordance the page already carries, now stratified by smoking -> The Observational-Trial Discordance. Candidate mechanism for whatever residual low-BMI risk remains: «the higher risk of all cause mortality in the lower range of BMI might be explained by low lean mass rather than low fat mass». (Sun et al., 2019) The decision reading: for a lean never-smoker, the pooled/observational underweight mortality penalty is not shown to apply causally — condition on smoking before reading the low arm.
Shape is outcome-specific (UKB cause-specific MR). CVD mortality increasing (nadir ~21-22); cancer mortality flat («no strong evidence that BMI affects cancer mortality in any BMI category»); the non-CVD-non-cancer «other» curve «had a profoundly curved J shape, with the lowest risk of mortality at a BMI of 23.0-24.0» (respiratory 27% / digestive incl. alcoholic-liver 18% / nervous 15% / external incl. suicide 11%). (Sun et al., 2019) So the J lives in one cause category — the same outcome-specificity the fabric holds for other exposures.
Independence verdict — type-F, NOT independent-E (the lineage chase again). Sun shares no author with Wade, so the triage-stage independent genetic witness (no excluded authors) reads as a fresh witness — but the chase defeats it: co-author Emanuele Di Angelantonio also leads the Global BMI Mortality Collaboration (Global BMI 2016 ref 6), and Sun cites Wade 2018 as ref 33. Shared author vs the held observational IPD-MA + cites-as-antecedent vs the held MR = type-F refinement, not independent-E.
| Parameter | Global BMI 2016 (corrected obs.) | Wade 2018 (linear MR, UKB) | Sun 2019 (non-linear MR, HUNT+UKB) | Same quantity? |
|---|---|---|---|---|
| Nadir | 22.5-25 | ~23 (MR) vs ~26 (obs.) | ~22-23 (HUNT) / ~25 (UKB) | YES — convergent |
| Low arm | residual 1.47 (partly real) | J deflates, reverse causation | no underweight harm in never-smokers; harm is ever-smoker/respiratory | partly — Sun conditions it on smoking |
| Above-nadir arm | monotone harm 1.31/5u | causal MR harm (CVD 1.10/u) | 1.09/u obese; steeper in women | YES — convergent |
| Independence | — | shares Davey Smith/Sattar w/ Global BMI | shares Di Angelantonio w/ Global BMI; cites Wade | NO — overlapping lineage, all type-F |
Adjudication caveats. MR is the strong check, but (i) Sun’s non-linear fractional-polynomial method carries a published Editor’s Note on methodological criticism + an updated analysis — the note’s text is not in the held chunk, so the non-linear shape estimates carry an added caveat while the smoking-decomposition direction is less exposed; and (ii) stratifying on smoking is a collider (genetically-predicted BMI influences smoking), which Sun argues is «likely to be negligible». (Sun et al., 2019) (inferred from Sun et al., 2019) the decompose the aggregate curve by the effect-modifier reading and the not-a-tension resolution are the wiki’s own synthesis -> The U-Shaped Association Artifact.
Decision relevance
- The nadir is 22.5-25, and above it every increment carries risk — there is no protective or even neutral overweight band once bias is removed. For a lean, non-smoking person the relevant target is holding BMI in the low-mid 20s; overweight is fine / protective does not survive the correction. (BMI is a crude adiposity proxy — for central-adiposity strata see Menopause and the Shifting Levers and waist-based measures; this page is the BMI curve, not a claim that BMI is the right instrument.)
- For the specific decision of who is at cardiovascular (MI) risk, BMI is the wrong instrument — use waist / waist-to-hip ratio. INTERHEART (case-control, 27,098 people) found BMI’s whole MI association disappears after adjusting for fat distribution (top-vs-bottom quintile OR 1.44 -> 1.12 after WHR -> 0.98 after all risk factors), while WHR stays the strongest marker; abdominal adiposity’s population- attributable share of MI is ~3x what the BMI cut-point captures (24.3% vs 7.7%) -> Waist-to-Hip Ratio and Cardiovascular Risk. This cashes the crude-proxy limitation above with a hard endpoint — the all-cause-mortality curve here is a different outcome and design (cohort), and both truths hold at once. (Yusuf et al., 2005) Contested (added 2026-09-04): a pooled analysis of 58 prospective cohorts (ERFC 2011) found BMI, waist and WHR of similar strength for incident CVD and none adding prediction over BP/lipids/diabetes, and refutes the “measure WHR not BMI” recommendation for developed-country populations — the case-control MI ranking above is design-contested -> BMI vs Abdominal-Adiposity Markers - Which Predicts CVD. (Collaboration, 2011)
- Adiposity is a big-rock lever, and the effect is largest where there is most life to lose — the gradient is steepest at 35-49 y (1.52 per 5 units) -> Layer 1 - Ranking Interventions for a Stratum.
- This estimates the association, not the benefit of weight loss. A bias-corrected observational curve says where mortality is lowest across people; it does not establish that reducing an individual’s BMI moves them down it -> Does Weight Loss Reduce Cardiovascular Events holds the intervention evidence (the loop the curve alone cannot close). At the extreme-obesity end the loop is partly closed: SOS (bariatric surgery, ~14-25% sustained loss, matched non-RCT cohort) found reduced all-cause mortality — adjusted HR 0.71 (P=0.01, no CI reported) — in severe obesity (BMI >=34/>=38), the interventional counterpart to this curve’s steep upper arm. (Sjöström et al., 2007) The design is weaker than an RCT (self-selection), so it warrants the arm directionally, not causally-clean.
Limits
- Observational, corrected — not causal-proof. The exclusions remove the known biases (smoking, reverse causation, prevalent disease); residual confounding (fitness, socioeconomic status, unmeasured illness) is not excluded, and there is no Mendelian-randomization arm in this analysis. So the corrected curve is strong but, like coffee and unlike alcohol, adjudicated by confounder-removal only, not by a genetic instrument -> The U-Shaped Association Artifact.
- BMI measurement — a mix of measured and self-reported BMI across cohorts (self-report biases toward the null / mislabels categories) -> Measurement Error in Dietary Assessment; the paper runs a self-reported-vs-measured sensitivity analysis.
- The MR gap is now cashed externally, but not independently
[2026-08-06]. Wade’s genetic instrument (the section above) supplies the strong check Global BMI lacked and converges on the nadir and reverse-causation reading — so the corrected curve is no longer MR-orphaned. The residual limits: the two analyses share investigators (type-F, not independent-E), the all-cause MR is imprecise (CI crosses null), and Wade’s MR is one cohort (UK Biobank, White British) vs Global BMI’s 239 cohorts on four continents -> The Observational-Trial Discordance. - Cannot separate fat mass from lean mass or distribution — BMI is the exposure, and the same BMI spans different body compositions across age, sex, and ethnicity (a plausible contributor to the regional differences, alongside the small-n imprecision of the South Asian estimate).
A guidance family stratifies the target by age — NNR2023 [2026-08-27, NNR revisit]
The Nordic Nutrition Recommendations 2023 recommend maintaining a healthy weight for working-age adults but explicitly decline an optimal-BMI target for older adults — the guidance-family form of this page’s own older-adult attenuation finding (the 70-89 y HR being materially shallower than the 35-49 y one). NNR: «Maintaining a healthy body weight and body weight stability is recommended in non-pregnant adulthood and for healthy growth in childhood, due to the associated health effects and the serious health risks of underweight, overweight and obesity.» (Nordic Council of Ministers, 2023) For older adults it stops short: «For older adults, the associations between overweight and health outcomes are less clear, and the available data are inadequate to make precise recommendations for optimal BMI in this age group.» (Nordic Council of Ministers, 2023)
- Classification: guidance-family attribution (F — a stratification caveat), NOT independent
backing. NNR rests on the same observational base this page appraises (its own citations are
Boushey and Cloetens & Ellegård reviews, not a fresh instrument), so it does not lift the
causal ceiling and adds no confidence via method independence. No
[E-independent]tag. The value is that a guidance body, in its current cycle, mirrors the fabric’s age-stratification rather than issuing a single all-age BMI target — a route-(a)/route-(b) caveat named by guidance. - Counter-passage check. NNR states no all-age BMI optimum and no obesity-paradox endorsement; it declines the older-adult target on data-inadequacy grounds, which does not oppose the page’s MR- adjudicated monotone-rise nadir (22.5-25) for working-age adults — the two are consistent once the age stratum is matched (not-joined check (ii): different unit/stratum). No divergence to file.