Does more exercise mean a longer life?

01What we found

Among 12,069 US adults aged 40–79, with 1,587 deaths, reporting any leisure-time moderate-to-vigorous activity was associated with an all-cause mortality HR of 0.64 (95% CI 0.54–0.75). We did not detect a clear additional dose gradient among active participants. Restricting baseline health and restarting follow-up at a two-year landmark gave HR 0.85 (0.64–1.13). This demonstrates sensitivity to population and follow-up conditions; it neither identifies a causal effect nor bounds it between 0.64 and 0.85. In randomized evidence, LIFE reduced mobility disability in its target older population. Mortality findings from LIFE and Generation 100 do not precisely establish whether their exercise programs reduce deaths. Failure to detect a mortality benefit is not proof that exercise cannot extend life.

Correction, September 19, 2026: The original article incorrectly called LIFE the only long-term randomized exercise trial. We have added the five-year Generation 100 trial. Some Cox intervals used model-based standard errors; all now use survey-design covariance. Sedentary-time missing codes and some model denominators were corrected. We withdrew a purported causal-effect range derived from sensitivity analyses and a cross-population death-count test used to reject an observational effect. Text, figures and downloads have been updated.

02Who was observed, and what was measured

We pooled four cycles, requiring age 40–79 at interview, examination participation, mortality-linkage eligibility, and complete leisure activity and model covariates. Of 40,617 records, 13,999 met the age criterion, 13,595 were examined, 13,571 were linkage-eligible, 13,558 had complete activity reports, and 12,069 were complete cases. The last step excludes approximately 11%, mainly because income-to-poverty ratio is missing. NHANES

The NCHS 2019 public mortality linkage provides a median 102.5 months and a maximum of approximately 159 months of follow-up, with 1,587 deaths. Selected follow-up times and causes are perturbed for confidentiality. Because the models include examination measures such as BMI, we use MEC weights divided by four, with survey strata and PSUs in covariance estimation. Mortality files

The questionnaire separates leisure activity, work/chores/yard activity, and walking or cycling for transport. The main exposure is guideline-equivalent minutes: moderate minutes plus twice vigorous minutes. No reported leisure activity does not mean no bodily movement. A weighted 50.6% reported zero leisure-time moderate-to-vigorous activity. Categories reference public-health guidelines; they are not personal prescriptions generated by this analysis. Questionnaire Guidelines

Raw-data counts matched 368 official codebook cells. This checks input frequencies, not every derived variable, downstream analysis or the truth of self-report. Review still found that sedentary-minute codes 7777/9999 must be treated as refusal/unknown rather than actual minutes; the affected sensitivity model was rerun.

03A larger separation by activity status than by dose

All three models use the same complete-case sample. M1 adjusts for demographic and socioeconomic variables; M2 adds smoking and BMI; M3 adds self-rated health and counted chronic conditions. No reported leisure activity is the reference.

All active-group point estimates lie below the reference, but not every confidence interval excludes one. Combining active participants gives M2 HR 0.635 (0.539–0.749). This describes relative instantaneous mortality hazards during follow-up, not a death-probability ratio or years of life gained.

The direct ≥600 versus 150–299 comparison gives HR 0.89 (0.59–1.34); dose doubling among active people gives 0.97 (0.90–1.04). These intervals accommodate smaller benefits or associations in the opposite direction. Reporting error, few high-dose events and model form can affect the gradient. This cannot justify telling an active person that additional activity has no benefits.

Arem's six-cohort pooled analysis and Garcia's dose-response meta-analysis also show larger association changes from low to moderate activity, with smaller marginal changes thereafter. Populations, activity definitions and reference groups differ, and reviews may share cohorts. They are not independent experiments validating our curve. Arem Garcia

04Why health restrictions change the estimate

Inactivity can follow illness, pain or functional limitation. The weakening association over follow-up is compatible with reverse causation: HR 0.48 (0.37–0.62) in the first five years and 0.76 (0.62–0.93) afterward. The survey-adjusted time-interaction ratio is 1.60 (1.20–2.13), p=0.0013. The overall 0.64 is consequently a model summary, not a constant association at every time.

Landmark analyses retain people still at risk at that time and restart follow-up there. The 12-, 24- and 48-month estimates are approximately 0.66, 0.69 and 0.72. Other restrictions are separate sensitivities, not successive steps in one monotonic removal process.

Excluding counted baseline conditions and fair/poor self-rated health gives HR 0.80 (0.60–1.06). Adding a two-year landmark leaves 6,361 people and 364 deaths, with HR 0.85 (0.64–1.13). Baseline conditions do not cover all diseases. The lung variable specifically captures reported emphysema, and the restriction retains people missing that response; this is not a strictly disease-free sample.

These changes do not quantify reverse causation. Health variables may be confounders or lie on pathways from previous activity to health. Eligibility restrictions, fewer events and hazard-ratio noncollapsibility also change estimates. The causal effect need not fall between 0.64 and 0.85, and the data do not establish that it must be smaller than the observed association.

Retaining incomplete covariates with missing-indicator adjustment gives HR 0.59 (0.51–0.69). This shows persistence under one alternative handling rule. Missing-indicator methods can themselves be biased, so the result does not prove complete-case selection introduced no bias. The E-values in the downloads are approximately 2.1–2.5, depending on the approximate conversion from HR to risk-ratio scale. They describe the strength of unmeasured confounding required under specified assumptions, without proving such confounding exists or addressing every selection and measurement process. Method Conversion

Variance estimation matters too. The survey-design interval is 0.54–0.75; ignoring weights and design gives 0.67 (0.60–0.75). Treating survey-weighted population totals as independent sample counts and using ordinary model variance produces a spuriously narrow interval, approximately 0.635 (0.634–0.636). This correction also addresses a subtler problem: survey-model summaries can contain multiple standard-error columns, so intervals and P values must use the same appropriate covariance.

05Different activity domains, different people

In a joint M2 model, HRs are 0.65 (0.55–0.77) for leisure activity, 0.70 (0.61–0.82) for work/chores/yard activity and 0.87 (0.75–1.01) for transport. The leisure-to-work HR ratio is 0.93 (0.76–1.14), p=0.48. Failure to detect a difference does not establish equivalence.

Leisure-active participants more often have college education and smoke less; the work/chores domain shows different directions for these characteristics. This challenges a simple account in which a single wealth indicator explains everything, but it cannot exclude socioeconomic confounding as an important contributor. Domains differ in measurement, selection and confounding; their comparison is not a natural randomized experiment. Being able to move may also be a shared marker of health.

Exploratory cause-specific HRs are 0.48 (0.35–0.66) for heart disease, 0.80 (0.60–1.05) for cancer, 0.51 (0.28–0.92) for stroke and 0.62 (0.48–0.79) for remaining causes combined. These are cause-specific hazards, not cumulative probabilities. Multiple outcomes and wide intervals limit interpretation. The cardiovascular pattern is compatible with physiological benefit and with illness limiting activity; it cannot establish either mechanism alone.

Comparing at least 150 equivalent minutes with less, HRs are 0.66 (0.50–0.87) at ages 40–59 and 0.63 (0.49–0.80) at 60–79; interaction p=0.90. This does not detect a difference between those age bands, rather than proving equal benefits. Because the work domain includes chores, it cannot refute studies of occupational workload. Occupational-activity review

06Self-report and devices use different measures

In LIFE's baseline table, the full randomized groups reported median walking/strength activity of zero minutes/week in the intervention arm and 30 in the control arm. Accelerometry was available for subsamples of 590 and 581, whose device-measured moderate-activity medians were 161 and 153 minutes. Different group summaries and sample coverage cannot establish that a particular person reporting zero simultaneously recorded 150 minutes on a device. LIFE

Self-report asks about activity contexts, while devices apply measurement thresholds. Their difference is not a fixed underreporting multiplier. A methods review found agreement and error varying with populations and instruments. Measurement review In Ekelund's device-based synthesis, the highest versus lowest quarter of total activity had HR 0.27 (0.23–0.32). Subtracting this from the 0.64 for any versus no leisure activity cannot isolate reporting error or a true causal effect. Ekelund

07What happened after random assignment

LIFE studied older adults with functional limitations. It randomized 1,635 adults aged 70–89 who had limited physical performance but could still walk 400 metres to structured multicomponent activity or health education. Mean follow-up was 2.6 years, with blinded assessment of the main mobility outcome. The table below recomputes published event counts: differences in proportions over the observed period, not fixed-time survival differences reconstructed from individual follow-up. LIFE

The crude risk ratio for major mobility disability is 0.85 (0.74–0.97); the published time-to-event HR is 0.82 (0.69–0.98). RR compares proportions and HR compares instantaneous hazards; neither is a person-time incidence-rate ratio. The inverse risk difference is about 18.4, conventionally rounded up to approximately 19 participants. This is a rough NNT over this trial's observation period, not a guarantee that every group of 19 avoids an event within exactly two and a half years.

Mortality RR is 1.14 (0.76–1.71). The wide interval proves neither harm nor absence of a mortality effect. Serious-adverse-event and hospitalization RRs are 1.08 (0.98–1.20) and 1.10 (0.99–1.22). These count all events; they do not attribute every event to exercise. Screening, supervision and program conditions constrain extrapolation. We do not derive individual training instructions from them.

Generation 100 used mortality as its primary outcome. The Norwegian study enrolled 1,567 volunteers aged 70–77, most reporting good health and already being relatively active. It randomized moderate continuous training, high-intensity interval training and following national physical-activity recommendations over five years. Generation 100

The combined comparison does not precisely establish mortality benefit. The favorable high-intensity point estimate also cannot be upgraded to proven lifespan extension. Controls were encouraged to remain active, some performed high-intensity exercise themselves, and intervention adherence was incomplete. This is not exercise versus complete inactivity. As in LIFE, findings concern a specific population, program and comparator.

We did not obtain individual data from either trial. Counts support crude RRs and risk differences, not reproduction of all survival analyses. The original article inserted the NHANES HR into a binomial model of LIFE deaths and used its P value against an observational causal claim. Populations, exposures, comparators and follow-up differ, so that test no longer supports a conclusion. Event-count planning can still illustrate precision: under equal allocation, constant HR 0.85 and two-sided alpha=0.05, approximately 1,189 deaths are needed for 80% power. This is a hypothetical design calculation, not evidence of no effect.

08What the evidence supports

There is direct randomized evidence for function: LIFE's particular program reduced mobility disability in its target population. Observationally, reported leisure activity accompanies lower mortality, while its within-active dose gradient and causal magnitude remain uncertain. Randomized mortality evidence exists. The two trials reviewed here do not precisely establish a generalizable mortality effect. That is different from having no randomized studies, from proven lifespan extension and from disproven lifespan extension.

Principal limitations are single self-reports, health selection and unmeasured confounding, complete-case handling, time-varying hazard associations, and specific trial populations and comparators. We did not isolate aerobic, resistance or balance training, compare devices with questionnaires in the same NHANES participants, or conduct an exhaustive systematic review of all exercise trials. The observational analysis concerns US adults aged 40–79; the added randomized evidence includes Norwegian older adults. Estimates cannot simply be transferred between these layers and populations.

Codex performed this correction; the original author's reviews remain historical. Revision self-review is not clinical or independent professional review. Reproduction materials contain code, input fingerprints, aggregate outputs and corrected figures. Evidence checking was updated through September 19, 2026. New mortality trials, longer follow-up or objective repeated measurements will trigger further review.