In 1998, researchers analyzing the Finnish Twin Cohort concluded in JAMA that "leisure-time physical activity is associated with reduced mortality, even after genetic and other familial factors are taken into account." Twins who exercised more lived longer than their genetically identical siblings. Case closed.
In 2025, a reanalysis of the same cohort with 25 more years of follow-up reversed the finding. When followed long enough, even twins who met physical activity guidelines for 15 years showed no statistically significant mortality difference versus their sedentary twins. The exercise-longevity association, the authors wrote, may be "largely due to genetic confounding and reverse causality."
Same dataset. Same twins. Opposite conclusion. The only thing that changed was time.
The Claim
"Exercise adds years to your life" may be the single most repeated health recommendation in the world. It rests on a mountain of observational evidence. Moore et al. (2012), studying 654,827 individuals across six prospective cohorts, calculated that 150 minutes per week of brisk walking adds 3.4 years of life. More activity, more years — up to 4.5 extra years at 450+ minutes per week. This paper is the backbone of global exercise-longevity messaging.
The WHO calls physical inactivity the fourth leading risk factor for noncommunicable disease, attributing 3.2 to 5.3 million premature deaths per year to insufficient movement. The CDC estimates 110,000 American deaths per year could be prevented if adults aged 40+ increased their physical activity. Harvard Health tells readers that "even 75 minutes per week of brisk walking was associated with a gain of almost two years."
There is one problem. Every one of these claims comes from observational studies. And when you apply causal methods — designs that can actually distinguish "exercise causes longer life" from "healthier people happen to exercise more" — the evidence collapses.
Five Lines of Causal Evidence
1. The Finnish Twins
The Sillanpää et al. (2025) study followed 22,750 Finnish twins whose physical activity was assessed in 1975, 1981, and 1990, with mortality tracked through 2020. Twins are nature's closest approximation to a controlled experiment — they share genes and childhood environment. If exercise truly extends life, active twins should outlive their sedentary co-twins even after decades.
They didn't. The greatest observed benefit was a 7% mortality reduction between sedentary and moderately active twins — but no additional benefit from higher activity levels. More importantly, the short-term association (higher activity = lower mortality in early follow-up) disappeared in the long term: no difference between highly active and sedentary twins across the full follow-up period.
Epigenetic clock analysis (Horvath method) showed no biological age differences between active and sedentary twins. After adjusting for smoking (not just status but quantity), education, alcohol, and BMI, the exercise-mortality association was "considerably attenuated."
The authors' conclusion is striking: "An underlying pre-disease state can limit physical activity and ultimately lead to death, not the lack of exercise itself." In other words, the observational association may run backward — sickness causes inactivity, not inactivity causes death.
2. Mendelian Randomization
Mendelian randomization uses genetic variants as natural experiments. If genes that predispose people to higher physical activity also predict lower mortality, that's evidence of a causal effect — because genes are assigned at conception, before any lifestyle choices or health conditions can confound the picture.
A two-sample MR analysis using UK Biobank data found no causal relationship between genetic predispositions to physical activity or sedentary behavior and mortality. The observational arm of the same study confirmed the standard association — people who exercise more do live longer in the raw data. But the causal analysis showed the association is not causal.
This directly corroborates the Finnish twin result using a completely independent method and population.
3. Randomized Controlled Trials
Ballin et al. (2021) critically reviewed all exercise RCTs with mortality outcomes — approximately 50,000 participants across multiple trials. Their conclusion: exercise "does not reduce all-cause mortality" in older adults or people with chronic conditions. "Evidence based on RCTs has failed to confirm that LTPA prevents premature mortality."
RCTs are the gold standard for causal inference. When they're available, they supersede observational data. On exercise and all-cause mortality, they're available. And they don't confirm the claim.
4. The Uppsala Negative Control
In May 2025, Uppsala University researchers found that high-fitness adolescents had a 53% lower risk of dying in random accidents — car crashes, drownings, homicides. Events that physical fitness cannot causally prevent.
This is what epidemiologists call a negative control outcome. If fitness appears to protect against death by random accident, the association isn't measuring fitness — it's measuring unmeasured confounders. Fit adolescents differ from unfit adolescents in ways that affect accident risk: socioeconomic status, risk tolerance, neighborhood safety, access to supervised activities. These same confounders inflate the apparent effect of fitness on all-cause mortality.
The Uppsala finding is powerful precisely because of its absurdity. Fitness cannot prevent a car accident. If the data says it does, the data is telling you about confounding, not causation.
5. Mendelian Randomization — The Nuance
The causal picture isn't uniformly null. MR studies have found that genetically predicted vigorous physical activity causally reduces myocardial infarction risk (OR: 0.24, p=0.007) — a very strong effect — with modest reductions in coronary heart disease (OR: 0.95, p=0.01).
Additionally, the largest GWAS meta-analysis on physical activity (Klimentidis et al., Nature Genetics 2022) found that the apparent benefits of exercise and lower sedentary time on cardiometabolic outcomes are substantially mediated or confounded by BMI. Translation: some of what we attribute to "exercise benefits" may actually be "not being overweight benefits."
This is more interesting than a simple null result. Exercise likely does protect against specific cardiovascular events. But the blanket claim — "exercise adds years to your life" — conflates disease-specific causal effects with all-cause mortality reduction. Fewer heart attacks does not automatically mean longer life overall, especially when competing causes of death are considered.
Why Observational Studies Overestimate
The biases aren't subtle. They're structural:
- Healthy user bias: People who exercise also eat better, drink less, smoke less, and have higher socioeconomic status. Studies adjust for some of these — but never all of them, and never perfectly.
- Reverse causality: People with undiagnosed illness, chronic fatigue, or pre-disease states exercise less. Their subsequent death is attributed to inactivity rather than the underlying condition that caused both the inactivity and the death.
- Smoking confounding: Most studies adjust for smoking status (yes/no). But active smokers who exercise smoke fewer cigarettes per day than sedentary smokers. The exercise group gets a smoking reduction benefit that's misattributed to exercise.
- Genetic pleiotropy: The same genes that make someone inclined to be active may also confer cardiovascular resilience. The association is real — it's just not caused by the exercise.
These aren't theoretical concerns. Each of the five causal methods above was specifically designed to circumvent these biases. All five point the same direction.
The Exception That Clarifies
The CHALLENGE trial (NEJM, 2025) is the first RCT to demonstrate that exercise improves survival — in colon cancer patients post-chemotherapy. An 889-patient, three-year structured exercise program produced a 37% lower risk of death and 28% lower risk of recurrence compared to health education alone.
This result matters because it shows exercise can have causal survival effects in specific disease contexts. The issue isn't that exercise is useless — it's that the general-population claim "exercise adds years to your life" lacks causal support. Specific mechanisms in specific populations can produce real effects. The blanket longevity promise cannot be extrapolated from them.
The Contradiction
Exercise clearly and causally improves cardiorespiratory fitness, muscle strength, insulin sensitivity, blood pressure, mental health, and cognitive function. These are not in dispute. Exercise is likely the single best intervention for quality of life. The evidence for how exercise makes you live is overwhelming.
The evidence for how exercise makes you live longer is not.
Five independent causal methods — twin studies, Mendelian randomization, RCTs, negative control outcomes, and genetic mediation analysis — all undermine the observational consensus that exercise extends lifespan in the general population. The most cited study behind the lifespan claim (Moore et al.) is entirely observational, with no causal design, subject to every bias that the causal methods were built to detect.
Yet the messaging hasn't changed. The WHO still attributes millions of deaths to inactivity. The CDC still counts prevented deaths. Harvard still promises extra years. The language quietly shifts from "associated with" in papers to "adds years" in public communication. The causal claim becomes the default because no one stops to check whether association is causation.
This is the sixth mechanism of knowledge failure I've documented: causal conflation. When the observational evidence is strong enough and the conclusion feels obviously true, the distinction between correlation and causation silently disappears. No one fabricated data. No one buried a study. The evidence simply got promoted beyond what it can support — and because the conclusion feels right, no one objects.
Exercise may be the best medicine ever discovered for how you live. Whether it changes how long you live remains, against all intuition, genuinely uncertain.