Does more exercise mean a longer life?

01The answer first

Among 12,069 US adults aged 40–79 (1,587 deaths, up to 13 years of follow-up), the step that matters most is whether someone does any leisure exercise at all (hazard ratio 0.64); among people who already exercise, we cannot detect any further gradient with dose. But a large share of that 0.64 is not an effect of exercise: once early deaths and people with baseline illness are progressively removed, it shrinks to 0.85 and loses statistical support. The one long-term trial that randomised older adults to a structured activity programme did reduce mobility disability (one case avoided per 18 participants over 2.6 years) — but it did not reduce deaths.

Scope: the strength, shape and credibility of the association between adult leisure physical activity and all-cause mortality, plus the functional outcomes that randomised exercise interventions have actually measured. No personal exercise prescriptions, no intensity or set recommendations, no conversion of hazard ratios into "years of life gained."

02Data: who, measured how, followed for how long

NHANES is a nationally representative survey with a complex sampling design. We pooled the four cycles from 2007–2008 to 2013–2014, kept people aged 40–79 at interview who completed the physical examination and were eligible for mortality linkage, then required complete leisure activity reports and model covariates: 12,069 people, representing about 119 million US adults after weighting. The selection steps were 40,617 pooled participants → 13,999 aged 40–79 → 13,595 examined → 13,571 eligible for linkage → 13,558 with a complete activity report → 12,069 with complete covariates. The 11% lost at the last step were mostly missing the family income-to-poverty ratio; keeping them with missing indicators gives 0.59 (0.52–0.67), slightly stronger than the main estimate, so the complete-case rule did not manufacture the association. NHANES public data

Deaths come from the National Center for Health Statistics public-use linked mortality files, followed through 31 December 2019: 1,587 deaths, median follow-up 102.5 months (8.5 years), maximum 159 months. This public version is de-identified with some values perturbed, and is used according to the official documentation. Public-use linked mortality files

The activity questionnaire follows the WHO GPAQ and asks about each domain separately; the leisure questions explicitly say to exclude the work and transport activities already reported, so the three domains separate cleanly. Intensity conversion uses the metabolic equivalents given in the official documentation appendix: 8.0 for vigorous, 4.0 for moderate. The main exposure is guideline-equivalent minutes per week (moderate minutes + 2 × vigorous minutes), because it maps directly onto the 150–300 minute public health recommendation. The dose cut points come from the 2018 US guidelines and were not chosen from the data. NHANES physical activity questionnaire documentation2018 Physical Activity Guidelines

One fact worth holding onto: on this definition, 50.6% of US adults aged 40–79 report zero minutes of leisure-time moderate-to-vigorous activity per week. That zero is not a fringe group; it is half the population.

Before fitting anything, we recounted every frequency cell printed in the official codebooks: four cycles, 21 variables, 368 cells, all matching. That only proves the import and the special codes (refused, don't know, skipped) were handled correctly. It says nothing about whether respondents' reported minutes are accurate — a point we return to below. Reproduction package

03Result 1: the step is at zero, not along the dose

Crude death rates fall with activity: 19.6 per 1,000 person-years in the zero group, 6.9 in the 600+ group. The two groups have similar mean ages (58.1 and 55.1 years), so this is not simply age.

The weighted Cox models give a sharper picture. With "no leisure exercise" as the reference, under three covariate sets:

Collapsing the six groups into "any versus none" gives the number this article keeps returning to: hazard ratio 0.64 (0.56–0.72). Looking across groups, every active category sits clearly below the zero group, but they do not form an ordered ladder among themselves: 75–149 minutes is higher than 1–74, and 300–599 is higher than 150–299. Compared head to head, ≥600 versus 150–299 gives 0.89 (0.59–1.34) — no evidence of a difference. Among the active only, each doubling of dose gives 0.97 (0.90–1.04).

So these data support "moving versus not moving matters a lot," not "more is better." That matches a pooled analysis of 661,137 adults: meeting the guideline minimum captures close to the maximum associated benefit, a ceiling appears at 3–5 times the minimum, and 10+ times is no better (0.68, 0.59–0.78). Pooled cohort analysis A dose-response meta-analysis covering 94 cohorts and over 30 million participants finds the same shape: the risk differences concentrate between zero and roughly 150 minutes per week, with smaller and more uncertain differences above that. Dose-response meta-analysis

04Result 2: every step that removes reverse causation moves the estimate toward 1

"Illness makes people inactive, then kills them" appears in data as "inactive people die more." That explanation has a testable signature: it should concentrate in early follow-up.

Splitting follow-up in two, the hazard ratio in the first five years is 0.48 (0.39–0.59), and after five years 0.76 (0.64–0.91); formally testing the difference gives a ratio of 1.60 (1.24–2.07), p=0.001. Proportional hazards genuinely does not hold here, so the single 0.64 is a weighted average over follow-up rather than a stable effect.

Following that thread through the pre-specified sensitivity analyses, every step moves the same way:

At the strictest setting — no self-reported cardiovascular disease, cancer, lung disease or diabetes at baseline, self-rated health not "fair or poor", and the first 24 months of deaths dropped — the hazard ratio is 0.85 (0.66–1.09), p=0.26.

Two things have to be said together here; saying either one alone is misleading. First, most of the association was removed: a large share of the "exercise cuts mortality by a third" figure in the observational literature can be explained by being able to move being a consequence of health. Second, the remaining 0.85 is not zero: only 364 deaths remain in that stratum and the lower bound is still 0.66; and putting self-rated health and prevalent disease into the model may itself block the pathway through which exercise would act, which is over-adjustment. The honest statement is: these data cannot separate "exercise makes people less likely to die" from "healthier people are more able to exercise"; what they can establish is that the first is much smaller than the headline number.

The same question can be asked in reverse: how strong would an unmeasured confounder have to be to explain the association away entirely? By the standard sensitivity calculation, moving 0.64 back to 1 requires an unmeasured factor associated with both activity and death by a factor of roughly 2.1 to 2.5, over and above the adjusted age, smoking, BMI and income. That threshold cuts both ways. It is not low: weak confounding cannot explain this association, and a vague appeal to "active people have better habits generally" does not suffice. But it is not out of reach either — the self-rated health we did measure is of exactly this magnitude: 30% of the inactive report fair or poor health versus 10% of the active, and self-rated health strongly predicts death. One more unmeasured health-status variable of similar strength would be enough, and the attenuation series above is evidence that such variables are at work.

A more technical check points the same way. We compared three statistical treatments. Using the survey design correctly gives 0.64 (0.56–0.72); ignoring weights and design entirely gives 0.67 (0.60–0.76); and using the weights while treating the 119 million weighted population as the real sample size gives 0.635 (0.634–0.636) — a confidence interval 0.002 wide. The point estimate barely moves, but the third result's precision is fake. This mistake is not rare in secondary analyses, and it is why the design details are spelled out in the reproduction package.

05Result 3: sport, chores and commuting — opposite social gradients, the same hazard ratio

The other major competing explanation is socio-economic position: people who exercise tend to be more educated, richer and less likely to smoke, and those things predict survival on their own. Adjusting for them only handles the measured part.

There is a stronger test available: find another activity domain whose social gradient runs the other way. In the same questionnaire, the "work" domain asks about paid and unpaid work, household chores and yard work. Among adults aged 40–79, people who report this kind of activity are less educated (26% with a college degree versus 34% of those who do not report it) and smoke more (22% versus 18%), with identical incomes — the mirror image of the leisure exercisers (43% versus 18% college, 13% versus 26% smoking).

If the protective association were mainly confounding by advantage, a domain with the opposite gradient should not show the same association. It does: putting all three domains in one model gives leisure 0.65 (0.57–0.75), work and chores 0.70 (0.62–0.80), transport walking or cycling 0.87 (0.75–1.01); formally comparing leisure with work gives a ratio of 0.93 (0.76–1.14), p=0.48.

This weakens two explanations at once. It does not support "the association is mainly an artefact of social advantage," and it does not support "leisure exercise has some special longevity magic." What the three domains share is being able to do physical activity at all — which contains both the real physiological effect of moving and the state of still being well enough to move. These data cannot separate the two.

One further clue points toward a real effect. Split by cause of death (exploratory, with other causes censored), the association is strongest for heart disease (0.48, 0.35–0.65) and weakest for cancer, where it crosses 1 (0.80, 0.64–1.00). That specificity fits the mechanism through cardiorespiratory fitness. But it fits reverse causation equally well: heart failure and chronic lung disease both prevent movement and kill. Specificity here is evidence in both directions and cannot settle causation on its own.

A question people often ask: does it still help when you are older? Comparing those meeting 150 min/wk with those who do not, the hazard ratio is 0.66 (0.50–0.87) at ages 40–59 and 0.63 (0.53–0.74) at 60–79, interaction p=0.90 — no evidence of a difference by age. All the caveats above apply here too.

One limit must be stated plainly: because this item counts chores and yard work as "work", it cannot be used to test the controversy about occupational physical activity. A systematic review reports higher all-cause mortality among men with high occupational activity, but that compares high versus low occupational load, which is not the same as our any-versus-none contrast. Systematic review on the occupational activity paradox

06What "minutes per week" actually measures

Every exposure number above comes from an interview question. The consequences are more serious than "a bit of noise."

The most direct evidence comes from the LIFE trial itself. Participants had to be sedentary to enrol (self-reporting less than 20 minutes per week of structured activity), and their baseline self-reported walking and strength training had a median of 0 to 30 minutes per week. The same people wearing accelerometers recorded a median of 153 to 161 minutes per week of moderate-intensity activity. LIFE randomised trial

In other words, a self-reported "zero" and a device-measured "over two hours a week" can belong to the same person. This is not simple under-reporting: self-report asks about deliberately taking exercise, while the device records the body moving. They are different quantities. Systematic review evidence likewise shows only moderate agreement between self-report and objective measurement, with biases that vary by age, weight and sex. Self-report versus objective measurement

That explains an apparent contradiction. A harmonised analysis of accelerometer cohorts puts the most active quarter against the least active quarter at 0.27 (0.23–0.32) — far stronger than the roughly 0.6 seen in every self-report study. Accelerometry pooled analysis

The stronger association is easily read as "see, the true effect is bigger and self-report dilutes it." That reading has a problem: devices measure total movement, which is exquisitely sensitive to current frailty, pain and cardiorespiratory capacity, and those accelerometer cohorts have a median follow-up of only 5.8 years — exactly the early window where we find the association at its strongest. Change the measurement and you change the quantity being estimated; a bigger number is not automatically stronger causal evidence.

07The one long-term randomised trial: mobility improved, deaths did not

The only way around all of the above is to randomise. The LIFE trial assigned 1,635 adults aged 70–89 with existing physical limitation (Short Physical Performance Battery ≤9) but still able to walk 400 m to either a structured activity programme (walking-based, target 150 min/wk, plus strength, flexibility and balance training) or a health education comparator, with a mean follow-up of 2.6 years and outcomes assessed blind to assignment.

We recomputed each outcome from the counts printed in the paper:

Our recomputed risk ratio for the primary outcome is 0.85, consistent in direction and magnitude with the published hazard ratio of 0.82 (the first compares the proportion of people with an event over 2.6 years, the second compares instantaneous hazards). The absolute difference of −5.4 percentage points means one case of mobility disability avoided for every 18 people taking part over two and a half years. This is the only causally interpretable evidence in this article, and it is about function, not lifespan.

Deaths went the other way (48 versus 42, risk ratio 1.14, 0.76–1.71). Serious adverse events and hospitalisations also leaned slightly toward the activity arm (risk ratios 1.08 and 1.10), none of them statistically significant. This should not be read as "exercise increases mortality or causes harm." What it does say is that in this frail older population, a supervised activity programme is not a zero-risk undertaking, and it showed no mortality benefit.

Putting the trial and the observational number face to face is worth doing carefully:

The trial had 90 deaths in total. If an effect of the size of 0.64 acted within two and a half years, the activity arm should have had about 35; it had 48, two-sided p=0.007. But what that test actually refutes has to be stated precisely, or it will be over-read.

First, neither arm was inactive. The control arm's baseline accelerometer median was 153 minutes per week of moderate activity; the intervention arm's was 161. Randomisation changed whether someone joined a structured walking and strength programme, not whether they moved at all. So the test refutes "adding a structured programme, on top of the substantial activity both arms already had, cuts deaths by a third within 2.6 years." It does not directly refute "people who do nothing would benefit from starting."

Second, it assumes the effect is immediate. If exercise acts on mortality over a decade-long timescale, seeing nothing in 2.6 years is exactly what one should expect. That assumption favours the test, so the p-value above should be treated as an optimistic bound rather than a verdict.

Third, the population and duration are narrow: frail adults aged 70–89, two and a half years, against an observational contrast of decades of habit in a general population. The two estimands differ and cannot be subtracted.

Conversely, the same data have almost no ability to detect a smaller effect: detecting a 15% mortality reduction with 80% power would need about 1,189 deaths, and 90 deaths give 12% power. So the trial neither supports a large mortality benefit nor demonstrates that there is none.

08Four explanations, judged

09What holds up

Three layers, kept separate:

Function (strongest). Randomly assigned structured activity reduces mobility disability in older adults: about 5 percentage points absolute, one case per 18 participants, with blinded assessment. This is something that can be said to an individual directly.

Association (moderate, and much weaker than the headline). The link between doing any leisure exercise and lower mortality is robust across model specifications, both sexes, both age bands, and the two domains with opposite social gradients (leisure and work/chores; the transport domain is weaker and crosses 1). But its credible magnitude is not 0.64 — it lies somewhere between about 0.85 and 0.64, depending on whether you regard adjusting for baseline health as confounder control or over-adjustment.

Lifespan (does not hold). No randomised evidence shows that exercise extends life; the one long-term trial pointed the other way and was underpowered. Every "exercise adds X years" figure is an extrapolation from observational association, and the analysis here shows such extrapolations are systematically inflated.

The most useful sentence for an individual is not "exercise more and live longer" but: the benefit of exercise that randomised trials actually support is staying able to walk and look after yourself for longer; as an investment in lifespan, the evidence chain breaks at the last step.

10What we know and what we do not

Known: half of middle-aged and older US adults report no leisure exercise at all; that zero is associated with higher mortality across many checks; a substantial part of the association reflects health status itself; among people who already exercise, these data detect no additional mortality benefit from more dose; randomised exercise improves mobility in older adults with a number needed to treat of about 18.

Unknown: how large the causal effect of exercise on mortality is — the observational evidence under the strictest setting gives 0.85 (0.66–1.09) and the only randomised evidence gives 1.14 (0.76–1.71), two intervals that overlap heavily and both include 1, so current evidence cannot locate it; what the dose-response shape is causally (self-report error is enough to flatten a real gradient); how much aerobic, resistance and balance training each contribute, which these data cannot distinguish at all; anything about people over 80 (the public files top-code age at 80); and how far any of this transfers beyond US adults.

Safety boundary: this article gives no exercise prescription. In the LIFE trial, serious adverse events and hospitalisations leaned slightly against the activity arm (risk ratios 1.08 and 1.10, neither significant), which suggests that activity programmes for older adults with existing functional limitation need supervision and gradual progression; people with cardiovascular disease, serious joint problems or a recent fall should start under professional guidance. That is an observation under study conditions, not individual advice.

11Methods note, reproduction and updates

For readers of the previous edition, plainly: this updates the legacy report Exercise and anti-ageing: the best-evidenced prescription for longevity. The "🟢 strong evidence" rating at its head, the framing of exercise as "the best-evidenced prescription for longevity," and the claim that "meta-analyses and RCTs agree that regular exercise slows biological ageing" are retracted in this update and replaced by a layered statement: randomised evidence supports the function layer, not the lifespan layer. This update is also narrower than the original. Its three sections on "telomeres and epigenetic age," "sarcopenia and resistance training," and "brain ageing and cognition," together with the specific training advice under "how to train," fall outside this study's question and have not been checked by this library, so they are not carried over. Removing them is not a finding that those directions lack evidence; it means we have not examined them. This edition likewise offers no exercise prescription.

A single agent framed the question, built the data, ran the analysis, conducted the self-review and checked both language versions. This is not independent peer review, nor clinical professional review. The analysis plan — estimands, covariate sets, dose cut points and all sensitivity analyses — was written before any effect estimate was produced and is published with the reproduction package. Three deviations occurred during execution: the work domain's definition was corrected after reading the questionnaire (it includes chores, so it cannot test the occupational activity controversy), the proportional hazards test had to be replaced (the standard test fails under survey weights), and the quantification of unmeasured confounding was added after seeing the results. All three, with reasons, are recorded at the end of the plan, and exploratory analyses are labelled as such.

The downloadable package contains all code, input file fingerprints, the analysis plan and every output; the raw data are public files and can be re-downloaded and checked following the instructions inside. Third-party article originals are not redistributed; follow the citations.

Evidence cutoff 2026-09-18. Next scheduled review 2027-03-18. Earlier updates will be triggered by: the linked mortality files extending to a later follow-up year; a new large randomised exercise trial reporting mortality or healthspan outcomes; long-term follow-up of the LIFE trial; or important new evidence on the comparability of self-reported and device-measured activity.