BMI and All-Cause Mortality: The J-Shaped Curve and Its Confounders

A 2016 meta-analysis pooled 10.6 million participants from 239 studies across 32 countries. The curve it found is the central fact about BMI and mortality. The lowest all-cause mortality sat at a BMI of 20.0 to 25.0 kg/m².

Below that range, mortality climbed. Above it, mortality climbed again. That shape dominates every discussion of BMI and mortality risk. The question is not whether the curve exists. It replicates across every large observational meta-analysis. The question is what drives it and whether it describes a causal effect of body fat or something else entirely.

Two landmark analyses set the terms. The Prospective Studies Collaboration (2009, The Lancet) analysed 900,000 adults from 57 studies. It found the lowest mortality at a BMI of 22.5 to 25.0 kg/m². Above 25.0, each 5 kg/m² rise associated with roughly a 30% increase in all-cause mortality. The Global BMI Mortality Collaboration (2016, The Lancet) expanded the data to 10.6 million participants and confirmed the same pattern: a nadir between 20.0 and 25.0 kg/m², with a hazard ratio of 1.07 at BMI 25.0-27.5 compared to the reference band.

Both analyses report associations. Neither claims causation. That distinction is the only frame that makes the curve interpretable.

The J-shaped curve has three segments. A low-mortality trough between 20.0 and 25.0 kg/m². A gentle rise on the right into overweight and obesity. A steeper rise on the left into underweight, the left limb.

The left limb generates the most confusion. A BMI below 18.5 kg/m² carries a higher hazard ratio for all-cause mortality than a BMI of 25.0 to 27.5 in many analyses. That looks as though being underweight is riskier than being moderately overweight. That impression is wrong, and the explanation for why it is wrong is where the evidence gets interesting.

What Happens When You Remove Smokers And The Already Ill

The Aune meta-analysis (2016, BMJ) narrowed the nadir to a BMI of 23.0-24.0 kg/m² when it restricted the analysis to never-smokers with no pre-existing disease. That restriction is critical. It controls for two of the largest confounders that inflate the left limb: reverse causation, where pre-existing illness causes weight loss before death, and smoking, which suppresses body weight and independently raises mortality. Once those are removed, the left limb flattens substantially.

Reverse causation means the study design catches people who are already losing weight because they are sick. Their low BMI does not cause their death. Their illness causes both their low BMI and their death. The observational curve cannot distinguish that sequence from a causal effect of thinness.

Smoking confounds the J-shaped curve in two ways. Smokers have a lower average body weight than non-smokers. Smokers also have a substantially higher all-cause mortality. A meta-analysis that pools smokers and non-smokers therefore places a group with high mortality and low BMI at the left end of the curve, pulling the left limb upward.

Prospective Studies Collaboration authors addressed this directly. When they excluded deaths within the first five years of follow-up and restricted the analysis to never-smokers, the elevated mortality at low BMI diminished. The Global BMI Mortality Collaboration applied the same restriction and found the same result. The left limb is not gone, but it is weaker and its remaining elevation may reflect residual confounding by pre-existing illness that was not captured by the exclusion window.

Why One Analysis Made Overweight Look Protective

The Flegal meta-analysis (2013, JAMA) of 97 studies and 2.88 million individuals reported a hazard ratio for overweight (BMI 25.0-29.9) of 0.94 compared to the normal-weight reference. That looks like a protective effect of extra weight. It is not. The normal-weight reference group in that analysis included smokers and people with chronic disease whose BMI was in the normal range precisely because they were ill. The apparently protective effect of overweight is a mirage produced by an unhealthy reference category.

Mendelian randomisation analyses address the association-versus-causation problem by using genetic variants that affect BMI as natural randomisation tools. Because genes are assigned at conception and are not confounded by smoking, illness, or lifestyle choices, a Mendelian randomisation analysis approximates a randomised trial of BMI.

These analyses consistently find a more linear relationship between higher genetically predicted BMI and all-cause mortality. The J-shape flattens or disappears. The left limb does not appear because the genetic variants that lower BMI do not carry the baggage of illness or smoking. The right limb, however, remains: higher genetically predicted BMI predicts higher mortality.

That result is the strongest evidence that the observational curve is confounded. The Mendelian randomisation evidence suggests the causal effect of higher adiposity on mortality is a steady upward slope, not a J. The apparent survival advantage of moderate overweight in observational data is an artefact of study design.

The obesity paradox is the observation that in certain disease cohorts, heart failure, chronic kidney disease, older adults with frailty, a higher BMI is associated with lower mortality. It is a real statistical pattern. It is not a biological license to gain weight.

The paradox is explained by the same confounders that distort the main curve: reverse causation and selection bias. In a cohort of people who already have a serious illness, those who maintain a higher body weight are often those who are less sick. Their weight is a marker of preserved health, not a cause of it. The thinner patients in the same cohort are thinner because their disease has progressed further. Comparing mortality rates across BMI groups in that setting compares sicker people to less sick people and attributes the difference to BMI.

No major clinical guideline recommends that a person with a disease try to gain weight to exploit the paradox. The paradox is a statistical observation in a specific subpopulation. It does not generalise to the population at large, and no Mendelian randomisation analysis has found a protective causal effect of higher BMI in disease cohorts.

The Global BMI Mortality Collaboration and the Prospective Studies Collaboration both place the lowest all-cause mortality in the range of 20.0 to 25.0 kg/m². The Aune meta-analysis, restricted to never-smokers with no pre-existing disease, narrows the nadir to 23.0-24.0 kg/m². Above that range, the hazard ratio rises. The Flegal meta-analysis, which does not restrict the reference group, produces a flatter curve that appears to show a protective effect of overweight, an appearance that is an artefact of confounding.

The hazard ratio per 5 kg/m² above 25.0 in the Aune analysis was 1.31. That means a person at BMI 30.0 has a 31% higher mortality risk than a person at BMI 25.0, all else being equal. The same analysis in the Prospective Studies Collaboration gave a 30% increase per 5 kg/m². The numbers are consistent across the largest analyses when the reference group is properly defined.

BMI Alone Misses Central Fat

Body composition matters beyond the scale. Waist-to-height ratio, recommended by NICE in guideline CG189, captures central adiposity that BMI misses. A person with a normal BMI and a waist-to-height ratio above 0.5 may carry excess visceral adipose tissue that carries independent mortality risk. BMI alone cannot distinguish that person from someone at the same weight with low central fat.

Selection bias operates at the study entry level. People who agree to participate in prospective cohort studies tend to be healthier than the general population. They are less likely to be at the extremes of the BMI distribution, and their mortality rates are lower. That biases the curve toward a flatter shape, particularly at the high-BMI end where participation rates are lower.

The population-attributable fraction for obesity-related mortality is a separate calculation from the shape of the curve. It estimates how many deaths in the population could be avoided if everyone had a BMI in the reference range. It depends on the prevalence of each BMI category and the hazard ratio in that category. Even if the J-shaped curve is partly confounded, the contribution of high BMI to total mortality in the population remains substantial because the prevalence of high BMI is high and the hazard ratio above 30.0 kg/m² is consistently elevated across all study designs.

Prospective Studies Collaboration authors calculated that the increase in mortality above BMI 25.0 accounted for a significant fraction of all deaths in the European and North American cohorts they analysed. The Global BMI Mortality Collaboration extended that finding to a global sample. The confounders do not erase the right limb. They primarily explain the left limb.

The evidence on BMI and all-cause mortality rests on observational data that cannot fully control for reverse causation, smoking, and selection bias. Mendelian randomisation analyses point toward a causal effect that is more linear than the observed J-shaped curve, but even those carry assumptions about genetic pleiotropy and population-specific effect sizes.

No single meta-analysis settles the shape of the curve for every population. The WHO Asian cut-offs (≥23.0 kg/m² for increased risk, ≥27.5 kg/m² for high risk) are based on type 2 diabetes and cardiovascular disease incidence, not all-cause mortality. The optimal BMI for mortality may differ by age, sex, ethnicity, and disease background. The one thing the evidence cannot do is tell an individual what their BMI should be. It can only describe the statistical pattern in a group.

The most important single thing that goes wrong with this evidence is treating a population J-curve as a personal prescription. The curve describes a group. It does not tell you whether you should lose, gain, or maintain.