Now anchored on the primary harmonised meta-analyses, not the WHO summary [2026-08-06]. The
device-measured dose maxima this page carries were previously held via the WHO 2020 evidence-profiles
annex — but that annex is a recommendation summary that borrowed its evidence from the underlying
harmonised meta-analysis, Ekelund 2019 (accelerometry). Per the cite-the-underlying-SR-for-an-
effect-claim rule, the effect/certainty claims are re-anchored below on the primary source; the WHO
quotes are retained as the summary that reproduced them. Two further primaries are now held: Ekelund
2016 (sitting x PA interaction) and Paluch 2022 (steps/day). All three are gold harmonised MAs of
prospective cohorts — objective/device or harmonised self-report — the objective-measurement corrective
to the older self-report dose-response literature.
These three are not independent replications of each other. Ulf Ekelund is a named co-author on all
three, and they re-slice overlapping cohort pools (the shared accelerometry/harmonised consortium —
NHANES, the Women’s Health Study, and others recur across them). So their agreement on curve shape is
a set of convergent framings of one evidence base, not three separate confirmations — which is why no
[E-independent] token is claimed on it (the same-quantity check returns NO; the convergence is
qualitative). The measurement reading in the next paragraph does not close that gap either.
The independence gap remains open. A genuinely independent check was hoped for from
Muscle-Strengthening Activity and Mortality (Momma 2022) and is NOT one (same observational
PA-epidemiology lineage, overlapping cohorts). The measurement-artifact reading rests mainly on
within-lineage evidence: Ekelund’s accelerometry finds the activity-mortality effect is ~2x larger than
self-report shows, a comparison against earlier self-report studies made in one of the three base MAs
(Ekelund et al., 2019).
Mandsager’s exercise-treadmill-testing study (estimated METs), from a non-overlapping group, finds objectively-measured
fitness with no upper limit of benefit observed in one referral cohort (elite vs high HR 0.77,
0.63-0.95, significant overall and in the 70-plus and hypertensive subgroups; no significant elite-vs-high
difference in younger patients; non-significant by sex and in the other with-comorbidity subgroups)
(Mandsager et al., 2018) -> Cardiorespiratory Fitness and Mortality. Its bands are age- and
sex-relative percentiles: at age 70+, where the step was significant, elite starts at 8.4-11.5 METs
depending on age band and sex (Mandsager et al., 2018). Mandsager calls its result «notably discrepant
from population-based studies» of physical activity and exercise, and its first candidate explanation is «the objective
measurement of physical fitness in the present study, as opposed to self-reported activity levels»
(Mandsager et al., 2018). That explanation is weakened, not refuted, by Ekelund’s device-measured curve,
which also plateaus (low light intensity excepted).
Ross’s AHA statement places most of the CRF-mortality gradient by ~10-12 METs, with >12 METs having «a
relatively lower impact» (Ross et al., 2016), but that sentence has no citation of its own; the only citation in the three sentences ending at ref 290 (on the closing
sentence) is Arem 2015, titled as a pooled dose-response analysis of leisure time physical activity and mortality (Ross et al., 2016); by its title an activity study rather than a fitness study. In chunk 01 Ross writes that many studies find «CRF levels >8 to 10 METs are associated with relative protection», and that «Relatively less is gained by increasing CRF between moderately fit and highly fit individuals.», stated alongside its Figure, modified from Blair 1989, a study in which fitness was estimated by maximal treadmill testing (Ross et al., 2016); the references cited for the >8-10 MET sentence include Kodama 2009 and Myers 2002 (Ross et al., 2016), by title mostly fitness/exercise-capacity studies.
So Ross’s cited evidence puts relative protection above ~8-10 METs, a sentence stated alongside its Figure (Blair
1989, treadmill-estimated fitness) adds that less is gained beyond moderate fitness, and its 10-12 MET sentence is itself uncited.
The Mandsager leg is consistent-with the measurement reading, not an independent confirmation: fitness !=
activity, its top-band step is significant overall but by age only in the 70+ subgroup, and objective
measurement is only one of Mandsager’s three candidate explanations (genetic/unmeasured-habit contributors
and population differences are the others). Not a banked type-E (corrected 2026-10-22: previously a
genuine [E-independent] corroboration). An RCT-grade mortality outcome is still owed.
(inferred from Ekelund et al., 2019; Mandsager et al., 2018)
The decision this page changes
Most of the mortality benefit arrives at a dose far below what people assume they need — and the curve then flattens. That is a decision-change in the telos’s specific sense: it licenses someone to stop optimizing, which is a result, not a failure to find one.
But the outcome of that activity — cardiorespiratory fitness — is a strong mortality predictor, with no upper limit observed in one referral cohort: «Importantly, there was no upper limit of benefit of increased aerobic fitness.» «The adjusted mortality risk of reduced performance on ETT was comparable to, if not greater than, traditional clinical risk factors (eg, CAD, smoking).» (Mandsager et al., 2018) -> Cardiorespiratory Fitness and Mortality. The two are not one claim: activity dose flattens early here (on the device-measured curve too), while objectively-measured fitness showed a significant step at the top in one referral cohort (elite vs high HR 0.77, 0.63-0.95) (Mandsager et al., 2018) — a discrepancy worth keeping open, not proof that more activity is always better.
This page is about total DOSE (how much); a sibling decision governs its DISTRIBUTION across the week. Once a weekly volume is fixed, whether it is concentrated into 1-2 sessions (the weekend-warrior pattern) or spread across the week appears not to change the mortality benefit -> Weekend Warrior Activity Pattern and Mortality (with a session-duration floor: sub-30-min sessions do not bank it, at low observational certainty).
Scope — this page is LEISURE / TOTAL PA; occupational PA can run the OTHER way [2026-08-14, Coenen]
Every estimate on this page is leisure-time or total physical activity. The «more is better» reading does not transfer to activity performed at work: Coenen 2018 (gold SR+MA, 17 studies, 193,696 participants) finds high occupational PA associated with higher all-cause mortality in men (HR 1.18, 95% CI 1.05-1.34), while women show a non-significant inverse tendency (0.90, 0.80-1.01) (Coenen et al., 2018) -> The Physical Activity Paradox. This is a domain disambiguation, not a contradiction: leisure PA (short intense bouts, recovery) and occupational PA (≥40 h/week, static/repetitive, no recovery) are different exposures under one word, and a same-quantity check returns NO on exposure, pattern, measurement, contrast, and direction. The decision-change for a reader: do not count physically demanding work as the exercise this page evidences — meeting a step/MVPA target via the job may not bank the mortality benefit and, in men, may carry cardiovascular risk. The causal reading is unadjudicated (observational, self-reported exposure, healthy-worker selection), so this is a stratum-level caveat on transferability, not a claim that work activity should be reduced.
Active commuting — a third route to the benefit, and the intensity refinement [2026-08-19, Celis-Morales]
A structural PA lever that corroborates the more-PA-lower-mortality reading from a route neither leisure nor occupational. Celis-Morales 2017 (UK Biobank, 263,540 workers, maximally adjusted incl. occupational PA and some leisure PA) finds cycle-commuting associated with HR 0.59 (0.42-0.83) for all-cause mortality, 0.55 (0.44-0.69) cancer incidence, 0.60 (0.40-0.90) cancer mortality, vs non-active commuting (Celis-Morales et al., 2017). This is a refinement (type-F), not independent corroboration (type-E): UK Biobank overlaps the accelerometry/harmonised cohorts this page already rests on, so it is the same observational PA-epidemiology lineage, not a separate route — it bounds and extends the leisure-PA reading, it does not raise its confidence as an independent instrument would.
The refinement: intensity separates the modes, on exactly the CRF axis. Cycling benefits all five outcomes with a dose-response by weekly distance; walking commuting is null for all-cause and cancer mortality but lowers CVD incidence (HR 0.73, 95% CI 0.54-0.99) and CVD mortality (HR 0.64, 0.45-0.91) overall; within its weekly-distance gradient the lower CVD incidence is evident only among walkers covering «more than six miles a week» (corrected 2026-10-06: lowers CVD risk only above six miles a week -> overall CVD benefit, with the six-mile qualifier scoped to the distance gradient, Celis-Morales 2017 chunk 01) (Celis-Morales et al., 2017). The authors attribute the gap to intensity — «approximately 90% of cycle commuters … achieved current physical activity guidelines, only 54% of walk- ing commuters» did (Celis-Morales et al., 2017). So the mode that reaches a CRF-improving stimulus banks the broad benefit; the sub-threshold mode does not — consistent with the objectively-measured fitness signal above -> Cardiorespiratory Fitness and Mortality and with the domain split on The Physical Activity Paradox.
Decision-change (structural leverage): commuting active — especially cycling — is a way to bank the mortality benefit built into the day, so adherence is high; but treat walking-commuting as a lower-intensity dose that may not reach the plateau this page’s leisure estimates describe. Caveat: observational, self-reported mode, and cyclists were leaner and fitter than non-active commuters at the same baseline visit (BMI 25.22 vs 27.51; fitness, measured in a subsample, 11.79 vs 9.32 METs) but not wealthier: by area deprivation more of them sat in the most-deprived fifth (29.2% vs 17.2%) and fewer in the least-deprived (11.9% vs 21.2%), while household income was similar (>£100,000: 8.6% vs 7.0%; <£18,000: 10.6% vs 9.8%) (Celis-Morales et al., 2017). Leanness and fitness may partly be consequences of cycling rather than markers of who chose it, so healthy-user selection is possible but not shown — a robust association, not a proven causal offset -> Upgrading Observational Evidence. (corrected 2026-10-22: markedly leaner/fitter/wealthier -> leaner/fitter, not wealthier, per Table 1) (inferred from Celis-Morales et al., 2017)
Self-critique [run 2026-08-19]: the commuting arm is classed type-F, not [E-independent] —
UK Biobank shares the observational PA-epidemiology lineage this page rests on, so it does not add an
independent instrument and no [E-independent] token is claimed. The benefit is stated as an
association with the healthy-user caveat explicit (cyclists leaner/fitter at baseline; corrected
2026-10-22: not wealthier: more often in the most-deprived area fifth, similar household income,
Table 1);
walking-commuting is flagged as a sub-threshold dose, not folded into the plateau. No overclaim: the
HRs stay Celis-Morales’s, the third-route framing is.
Where the curve bends — named maxima, device-measured
«Maximal risk reductions for light intensity PA was ~375 min/day, low light intensity PA at ~325 min/day, high light intensity PA at ~80 min/day, and MVPA at ~24 min/day.» (World Health Organization, 2020)
«The maximal risk reduction for total PA was observed at about 300 cpm (adjusted HR = 0.34 [95% CI > 0.27 to 0.43]).» (World Health Organization, 2020)
Certainty: HIGH. The underlying profile is a harmonized meta-analysis of eight prospective cohorts in which all eight used accelerometers rather than self-report, and «A non-linear, dose-response association was found between all exposure variables and mortality (p<0.02 for all exposure variables)». A second profile reports «A curvilinear relationship was found between total PA and all-cause mortality (p non-linearity <0.001).» (World Health Organization, 2020)
| Exposure | Dose at maximal risk reduction |
|---|---|
| MVPA (moderate-to-vigorous) | ~24 min/day |
| High light intensity | ~80 min/day |
| Low light intensity | ~325 min/day |
| Light intensity (total) | ~375 min/day |
| Total PA | ~300 cpm — HR 0.34 (0.27-0.43) |
Read the table as substitution, not as a target menu. These are different routes to the same maximum, so the operative reading is: roughly 24 minutes of moderate-to-vigorous activity, or a much larger volume of light activity, reaches most of the available benefit. WHO states the intensity-independence directly:
«Any physical activity, regardless of intensity, was associated with lower risk of mortality, with a non-linear dose-response.» (World Health Organization, 2020)
Two cautions on the HR 0.34. It is a between-quartile contrast in observational cohorts, so reverse causation (illness reduces activity) is not excluded by design — the profile’s HIGH rating reflects consistency and precision, not randomization. And the exposure is total accelerometer counts, which is not a prescription anyone can follow directly.
And the dose points are regions on a sparse, age-bounded top — not a target menu. Two caveats the table drops belong at the point of display. (i) Studied range: Ekelund’s cohorts «included middle aged and older adults who were at least 40 years old; it is unclear whether the estimates of the absolute physical activity levels associated with maximal risk reduction apply to younger people» (Ekelund et al., 2019) — the maxima are not scoped to under-40s. (ii) The upper arm is thin: each dose maximum carries its own wide interval, the MVPA point most of all («wide confidence intervals were observed at this part of the dose-response curve» because few participants reach high MVPA), and the ~300 cpm total-PA maximum «mirrors the median total physical activity for the third quarter» (Ekelund et al., 2019) — i.e. the apparent optimum sits where the data are densest, the apparent-optimum-equals-sampling-edge hazard, so read the dose column as regions on a flat top, not point targets -> The Underivable Optimum.
The very-high-volume tail in the general population — a smaller benefit, not a reversal [2026-10-21, Franklin]
The maxima above say where the benefit flattens. A separate question is whether it turns back at volumes far above guidance. An AHA scientific statement (consensus; no stated search) reviews the cohorts that reach that far. Arem: «Maximal mortality risk reduction (HR, 0.61 [95% CI, 0.59−0.62]) occurred at an exercise volume equaling 3 to 5 times the recommendations. Although the most active indi- viduals (≥10 times the recommendations) had a lower mortality risk (HR, 0.69 [95% CI, 0.59−0.78]) compared with inactive control subjects, their risk reduction was not as large as the group with optimal exercise volume.» The statement attributes the wide intervals in the top groups to their size: they «may be explained by the relatively small sample sizes of the most active sub- groups (typically <5% of the total cohort).» Its verdict: «In aggregate, there is currently no evidence to reject the curvilinear association (Figure 8A) between exercise volumes and cardiovascular health outcomes, with AF as a possible exception.» (Franklin et al., 2020)
- AF is the outcome where the shape may differ. A nonlinear meta-regression (Ricci; 19 studies) found a J-shape: 5-20 MET-h/wk lower AF risk, 20-55 MET-h/wk similar to inactive, and a trend to higher risk above 55 MET-h/wk (>9.5 h of vigorous exercise per week). In the Physicians’ Health Study, men who jogged 5 to 7 times per week had a 50% higher risk of AF than men who did not exercise vigorously, after adjustment for cardiovascular risk factors. (Franklin et al., 2020) Shape is outcome-specific here as elsewhere, and the AF signal starts at volumes a committed recreational runner can reach.
- Coronary calcium at high volume is a surrogate, not an outcome. In the Cooper Clinic cohort (n=21,758 men), the 432 doing ≥3000 MET-min/wk had a higher adjusted risk of CAC ≥100 but no higher all-cause or CV mortality than men exercising at more moderate levels: «These findings refute the notion that high-volume endurance activity (>1 h/d) increases mortality risk, regardless of CAC level.» (Franklin et al., 2020)
- Cardiac patients are carved out. The same paragraph that rejects a U for the general population adds: «These find- ings might not be applicable to cardiac patients, because high-intensity, high-volume PA may increase all-cause mortality among these patients.» Among CHD patients, Mons found the lowest mortality at 2-4 vigorous sessions/week and HR 1.77 (0.90−3.47) at 7 sessions/week; Stewart (n=15,486) found a graded reduction instead, and the statement calls the findings conflicting. (Franklin et al., 2020)
- Reading the Arem pair: the intervals overlap (0.59−0.62 vs 0.59−0.78), so the smaller benefit at ≥10x is a point-estimate difference, not a demonstrated one. Other top-group estimates the statement tabulates point the same way (Lear 0.98 for the most active vs 0.89 optimal; Schnohr 1.97 with a 0.48−8.14 interval).
- Decision consequence: for people without heart disease the relevant tail question is the
transient risk of a bout, not a long-run harmful arm — that lives on
Exercise-Related Acute Cardiac Events. The adjudication of the extreme-exercise upper arm as a
candidate artifact is catalogued on The U-Shaped Association Artifact. These are secondary
citations through one statement, not primaries held here, so they bound the shape claim without
raising this page’s
confidence:.
The primary numbers (Ekelund 2019) — the full dose-response, and where the benefit is banked
The two maxima above trace to Ekelund 2019, a harmonised MA of 8 accelerometry cohorts (n=36 383, 2149 deaths, model B, least-active referent). The full quartile curve makes the shape legible:
| Exposure | Q1 | Q2 | Q3 | Q4 (most active) |
|---|---|---|---|---|
| Total PA (cpm) | 1.00 | 0.48 (0.43-0.54) | 0.34 (0.26-0.45) | 0.27 (0.23-0.32) |
| Light PA (min/d) | 1.00 | 0.60 (0.54-0.68) | 0.44 (0.38-0.51) | 0.38 (0.28-0.51) |
| MVPA (min/d) | 1.00 | 0.64 (0.55-0.74) | 0.55 (0.40-0.74) | 0.52 (0.43-0.61) |
«Any physical activity, regardless of intensity, was associated with lower risk of mortality, with a non-linear dose-response. Hazards ratios for mortality were 1.00 (referent) in the first quarter (least active), 0.48 (95% confidence interval 0.43 to 0.54) in the second quarter, 0.34 (0.26 to 0.45) in the third quarter, and 0.27 (0.23 to 0.32) in the fourth quarter (most active).» (Ekelund et al., 2019)
The single most decision-relevant feature: the steepest drop is the FIRST step off the floor. Getting out of the least-active quartile roughly halves mortality (total PA Q1->Q2 HR 0.48) — and the increment was small: «broadly equal to … 5 min/day of moderate-to- vigorous intensity physical activity» over the referent.
«the greatest risk reduction for mortality was observed when the second quarter was compared with the referent, for all activity intensities.» (Ekelund et al., 2019) This is the marginal-minute-worth-most-at-the-bottom shape, quantified: the near-sedentary person has the largest lever in the whole domain, and it is a small one to pull.
The plateau, stated primarily — and honestly as a plateau, not a “no knee”. Above the maxima the curve flattens:
«No further risk reductions occurred with higher levels of activity except for low light intensity physical activity where the risk appeared to decrease further.» (Ekelund et al., 2019) So this IS a located plateau (monotone-decreasing then flat — no U/J, no harmful upper arm at achievable doses). It largely survives the measurement caveat the wiki attaches to plateaus: measurement error can hide a knee, and where it grows with activity level it can also manufacture a plateau; here the plateau is on the device-measured curve, where that error is smaller than in self-report — so it is more credible than a self-report plateau, though not immune to it (corrected 2026-10-22: previously error can hide a knee but not manufacture one). The one qualifier that cuts the other way: the plateau’s high-dose arm rests on few participants with wide CIs (the maxima caveat above), so it is well-established that the curve flattens but not precisely where. Cross-check: objectively-measured fitness showed no upper limit of benefit in one referral cohort (Mandsager, in age/sex-relative percentiles) -> Cardiorespiratory Fitness and Mortality. Fitness and activity dose are not the same construct, and Ross reports that many studies find relative protection at >8-10 METs and less gain between moderate and high fitness (Ross et al., 2016), while Mandsager saw no upper limit in one referral cohort; a precise flattening point for measured fitness is not pinned by the held sources (corrected 2026-10-22: previously shows NO plateau).
Why the effect is this large — the measurement corrective (the type-A payoff of holding the primary).
«The observed effect sizes for the associations between physical activity and the risk of death are about twice as large compared with those previously reported in studies assessing physical activity by self report» (Ekelund et al., 2019) Self-report attenuates the true gradient toward the null; objective accelerometry roughly doubles the observed magnitude and resolves the light-intensity shape self-report could not. This is the Measurement Error in Dietary Assessment binding-constraint lesson firing in the activity domain, and the reason a device-measured HR is not comparable to a self-report HR for the same behaviour.
Reverse causation (the frail move less) — handled by the WEAK check only.
«We attempted to minimise bias from reverse causation by excluding all deaths within the first two years in sensitivity analyses. The hazard ratios were materially unchanged for the associations of total physical activity and light intensity or moderate-to-vigorous intensity physical activity with mortality, and slightly attenuated for sedentary time.» (Ekelund et al., 2019) The PA associations survived early-death exclusion (Ekelund adds «bias from reverse causation might persist»); there is no MR/genetic instrument, so this is the weak adjudication in the The U-Shaped Association Artifact sense — enough that the monotone benefit is not purely sick-quitter artifact, not enough to call it causal. Consistent across all three sources here (see the steps and sitting sections).
Steps per day — the practical, wearable-native dose (Paluch 2022)
Steps are the metric people actually have. Paluch 2022 (harmonised MA, 15 cohorts, n=47 471, 3013 deaths; quartile medians 3553 / 5801 / 7842 / 10 901 steps/day) gives the same shape in the unit a fitness tracker reports:
«Compared with the lowest quartile, the adjusted HR for all-cause mortality was 0·60 (95% CI 0·51–0·71) for quartile 2, 0·55 (0·49–0·62) for quartile 3, and 0·47 (0·39–0·57) for quartile 4. Restricted cubic splines showed progressively decreasing risk of mortality among adults aged 60 years and older with increasing number of steps per day until 6000–8000 steps per day and among adults younger than 60 years until 8000–10 000 steps per day.» (Paluch et al., 2022)
Highest vs lowest quartile = 40-53% lower mortality; overall nadir ~7000-9000 steps/day; monotone- decreasing to a plateau (p-nonlinearity <0.0001), same as the MVPA curve.
The age plateau is genuine effect modification (route-b), not just baseline-risk (route-a). The age x steps interaction is significant (p=0.012): older adults (>=60) reach the plateau LOWER (~6000-8000 steps) than younger adults (~8000-10 000) — the sex interaction is not (p=0.11). Because this is positive interaction evidence, stratifying the step target by age clears the higher route-(b) bar (per transportability and effect modification / the five-routes table), not only the route-(a) absolute-benefit scaling. Decision-change: an older adult can bank the mortality plateau at a lower step count than the number a younger adult should aim for.
10 000 steps is a marketing number.
«Although the goal of 10 000 steps per day is widely promoted as being optimal for general health, it is not based on evidence, but instead originates from a marketing campaign in Japan.» (Paluch et al., 2022) The benefit plateaus below 10 000 for most adults and well below it for older adults — so 10 000 is not a threshold to clear, and treating it as one may discourage the person for whom 6000-8000 already banks most of the benefit.
Cadence adds little beyond volume.
«We found inconsistent evidence that step intensity had an association with mortality beyond total volume of steps.» (Paluch et al., 2022) Peak-30/60-min stepping rate stayed significant after adjusting for volume, but time spent at a moderate cadence did not — total steps carry the signal, so count the steps, don’t chase the pace. Reverse causation handled by the weak check (effect stronger at <6 y follow-up, HR 0.32, vs >=6 y, 0.57 — a sick-quitter tell — but the 2-year exclusion left it significant).
What an intervention buys on this curve is far smaller than a quartile. Paluch’s quartile medians sit about 2000-3000 steps/day apart (3553 / 5801 / 7842 / 10 901). Two partly overlapping gold MAs of app and tracker RCTs (GRADE low to low-to-moderate) put the induced gain over control at roughly +600-750 steps/day after a few months, with durability unshown. The quartile HRs compare people who habitually differ in steps. They are not the effect of adding steps to one person. So a device-driven gain cannot be read off the curve as a hazard ratio. Even if the gradient were causal and applied to induced change (neither is shown, and the paragraph above notes a sick-quitter tell), ~600-750 steps is roughly a fifth to a third of one inter-quartile gap, held for a few months. (Brickwood et al., 2019; Laranjo et al., 2020; inferred from Paluch et al., 2022) -> Wearable Activity Trackers and Physical Activity
These three metrics are NOT interchangeable numbers — a same-quantity check
Steps/day, MVPA-min/day and sitting-hours/day are complementary framings of one construct (total movement volume), not convertible doses. The parameter table below is the guard against reading one source’s number as another’s:
| Parameter | Ekelund 2019 (accel) | Paluch 2022 (steps) | Ekelund 2016 (sitting) | Same quantity? |
|---|---|---|---|---|
| Exposure unit | MVPA min/d; total-PA cpm | steps/day | sitting h/day; MET-h/week | NO — different units |
| Measurement | hip accelerometer (device) | step-counting device | self-report questionnaire | NO — device vs self-report |
| Referent | least-active quartile | lowest-steps quartile | least-sitting + most-active | NO — different referents |
| Where benefit banks | ~24 min/d MVPA (plateau) | ~7-9k steps (age-varying) | 60-75 min/d MVPA offsets sitting | measure a DIFFERENT feature |
| Top-vs-bottom effect | HR 0.27 (total PA) | HR 0.47 (steps) | HR 1.04 vs 1.27 (joint) | NO — non-comparable contrasts |
The decision-relevant convergence is qualitative, not numeric: all three show the same shape — a steep early drop, then a plateau, on the objectively/harmonised-measured curve — but the HRs cannot be equated (a device-measured total-PA Q4 HR of 0.27 is not the same as a steps Q4 HR of 0.47; they are different exposures, referents and measurement instruments). Use steps for a wearable-native target, MVPA minutes for a guideline-aligned target, and the sitting interaction for the offset question — do not convert between their hazard ratios. (Ekelund et al., 2016, inferred from 2019; Paluch et al., 2022)
Strength training — a real independent association, never ranked above aerobic
| Finding | Effect | Certainty |
|---|---|---|
| Strength-guideline adherence vs not (Stamatakis 2018, 11 cohorts, N=80,306) | HR 0.80 (0.70-0.91) | MODERATE |
| Any strength-promoting exercise vs none (same) | HR 0.77 (0.69-0.87) | MODERATE |
| Meeting both aerobic AND strengthening guidelines vs neither | HR 0.71 | MODERATE |
(World Health Organization, 2020)
The RT dose is small and the benefit does not require volume. Three RT→mortality MAs (pooling many of these same cohorts — Stamatakis and Siahpush above are constituents) put the all-cause reduction around −15% for any RT vs none, and Shailendra’s minutes/week dose-response (4 studies) finds the «maximum risk reduction of 26% observed at around 60 minutes per week of resistance training (RR=0.74; 95% CI=0.64, 0.86)» with benefit diminishing at higher volumes — so some RT captures the signal and there is no mortality case for high volume. The evidence is observational/self-reported and the upper «U» arm rests on 4 studies (do not act on it). Full appraisal + the U-artifact reading: Muscle-Strengthening Activity and Mortality. (Shailendra et al., 2022)
The operative claim is both, and together — not strength instead of cardio. No profile in the annex ranks resistance training above aerobic activity, and where the two are compared head-to-head (anxiety, 16 RCTs; depression, 33 RCTs; sleep), WHO reports «No significant difference was found between studies examining resistance training vs. aerobic exercise training» — at LOW and VERY LOW certainty. (World Health Organization, 2020)
A distinction that is NOT a tension. Siahpush 2019 (N=68,706) reports «There was no association between all-cause mortality and meeting strength recommendations (and not aerobic PA recommendations)» — which looks opposed to Stamatakis. It is not: Siahpush’s profile is about smokers, Stamatakis pools 11 general-population cohorts. Different population, consistent once matched. Recorded here so the apparent clash is not re-filed later as a tension. (World Health Organization, 2020)
Sedentary time is a separate exposure with its own thresholds
«For all-cause and CVD mortality, a threshold of 6-8 h/day of total sitting and 3-4 h/day of TV viewing was identified, above which the risk is increased.» (World Health Organization, 2020)
The relationship is «non-linear for all-cause mortality» and PA-adjusted — i.e. it survives controlling for activity, so sitting less and moving more are not the same lever.
The dedicated anchor for this sub-question is now Sedentary Behaviour and Chronic Disease Risk
(Patterson 2018, gold dose-response MA), which supplies the per-outcome curve shape underlying WHO’s
6-8h/3-4h summary range, adds incident T2D (TV->T2D the strongest association, 29% PAF) and cancer
mortality, and shows the sitting curve is an accelerating-harm knee — the mirror image of the
activity-benefit plateau on this page. [2026-08-21, Patterson]
TV viewing carries a lower threshold than total sitting (3-4 h/day vs 6-8 h/day) and stronger
associations. Two behaviours inside one category, behaving differently — though what accompanies TV
viewing is not addressed here, so confounding is not excluded.
-> [[Is the Food Category Doing Any Work]] for the same structure in a different domain.
The objective sitting threshold is HIGHER than the self-report one — measurement again. Ekelund 2019 (accelerometry) puts the sitting-mortality inflection at 9.5 h/day, above the 6-8 h/day from self-reported-sitting metas — the same self-report/device gap that doubles the activity effect above, running in the sitting direction. Read the WHO 6-8 h number as a self-report figure, not the device one. (Ekelund et al., 2019)
Does activity OFFSET sitting? Mostly yes — the Ekelund 2016 interaction
The decision this answers: if I must sit 8+ hours (desk job, commute), can being active cancel the risk? Ekelund 2016 (harmonised MA, >1 million adults, joint sitting x PA analysis) is the source, and the answer is high activity eliminates the sitting-mortality association:
«Daily sitting time was not associated with higher all-cause mortality rates among those in the most active quartile. Compared with the referent (<4 h of sitting per day and highest quartile of physical activity [>35.5 MET-hour/week]), there was no increased risk of dying during follow up in those who sat for more than 8h/day but who also reported >35.5 MET-hour/week of activity (HR=1.04; 95% CI, 0.99, 1.10). In contrast, those who sat the least (<4 h/day) and were in the lowest (<2.5 MET- hour/week) activity quartile had a significantly increased risk of dying during follow-up (HR=1.27, 95% CI, 1.22, 1.31).» (Ekelund et al., 2016)
The striking comparison: **the most-active + most-sitting group had LOWER mortality than the least-active
- least-sitting group** — activity dominates sitting when both are pushed. Stratified, the sitting penalty (>8 vs <4 h/day) shrinks stepwise across activity quartiles: 1.27 -> 1.12 -> 1.10 -> 1.04 (ns). The dose that eliminates it is high — 60-75 min/day of moderate activity, above the basic guideline:
«High levels of moderate intensity physical activity (i.e. about 60 to 75 minutes per day) appear to eliminate the increased risk of death associated with high sitting time. However, this high activity level attenuates, but does not eliminate the increased risk associated with high TV viewing time.» (Ekelund et al., 2016)
Two decision-relevant asymmetries. (i) The offsetting dose (60-75 min/d) is higher than the mortality-plateau dose (~24 min/d MVPA) above — so enough activity to bank the mortality benefit is not automatically enough to cancel heavy sitting; the sitting offset asks more. (ii) TV viewing is only attenuated, not eliminated — even the most active kept excess risk at >5 h/day TV (HR 1.16, 1.05-1.28) (Ekelund et al., 2016), whereas total sitting was fully offset. TV is not just sitting: postprandial-evening timing, snacking, and fewer sitting-breaks are the offered mechanisms, so the two sedentary exposures are not one lever (consistent with Willett’s near-zero sitting/activity correlation above). Magnitude anchor: the least-active + >8 h-sitting group’s 58% excess risk «is similar to that of smoking and obesity». (Ekelund et al., 2016)
Caveats (all three sources). Ekelund 2016 uses self-reported sitting and PA at one timepoint (attenuation toward null), mostly >45 y, western; reverse causation addressed only by excluding baseline-ill / early deaths — the weak check, same as Ekelund 2019 and Paluch. So the offset is a robust, biologically-plausible association (1 h moderate activity improves postprandial lipid/glucose after prolonged sitting), not a proven causal cancellation.
The offset dose, device-measured — the asymmetry narrows on a common instrument [2026-09-22, Ekelund 2020] type-F
The self-report offset above (60-75 min/day) is now redone with accelerometry, and the device figure is about half. Ekelund 2020 (harmonised MA, nine accelerometry cohorts, 44 370 middle-aged/older adults, 3451 deaths) cross-tabulates device-measured MVPA x sedentary time — the JOINT-combinations follow-up to the SEPARATE-associations Ekelund 2019, so the offset dose is a new estimate, not a re-run:
«Our joint analysis results suggest that about 30–40 min (median of medians=34 min; IQR: 26–48 min) of MVPA per day (online supplemental table 2) attenuated the risk of death in our most conservative model.» (Ekelund et al., 2020)
In the highest MVPA third (~34 min/day) high sedentary time carried no significant excess mortality (+40%, 95% CI 0.87-2.26, ns); the risk sat entirely in the low-MVPA third (~2 min/day), where every sedentary combination was elevated, worst cell HR 2.63, i.e. +163% (95% CI 1.93-3.57; the paper prints it as 263% against an HR-scale interval, a slip its +65% / +40% cells do not make) (corrected 2026-10-06: +263% -> +163%, Ekelund 2020 chunk 01). (Ekelund et al., 2020)
But the device 34 min and the self-report 60-75 min are NOT the same quantity — the parameter table:
| Parameter | Ekelund 2016 (self-report) | Ekelund 2020 (accelerometer) | Same quantity? |
|---|---|---|---|
| Exposure measurement | self-report questionnaire (leisure/recreational, bouts >=10 min) | hip accelerometer (all movement, multiple domains) | NO — device captures incidental movement self-report misses |
| Offsetting dose | ~60-75 min/day moderate-intensity | ~30-40 min/day MVPA (median-of-medians 34) | NO — read off different instruments, not convertible |
| Verb on the association | «eliminate» the sitting-mortality association | «attenuate» the risk (paper’s own word) | NO — eliminate != attenuate |
| Sedentary contrast | >8 vs <4 h/day (self-report) | tertiles; low <8.5 h/d, high >10.7 h/d (device) | NO — different thresholds and instrument |
| Grouping | activity quartiles (top >35.5 MET-h/week) | MVPA tertiles (high third median 34 min) | NO — quartile vs tertile |
| Design / n | >1,000,000 adults, harmonised joint MA | 44 370, 9 cohorts, harmonised joint MA | both joint harmonised MAs — and Ekelund authors both |
(Ekelund et al., 2016) (Ekelund et al., 2020)
So the offset dose halved is NOT a clean apples-to-apples claim — and the paper itself says the drop «is likely explained in part by differences between self-reported and accelerometer measured physical activity and sedentary behaviour», because «Self-reported physical activity usually only assesses specific domains … whereas accelerometer measures capture more movement across multiple domains». (Ekelund et al., 2020)
The decision-change is the WITHIN-instrument comparison. The held asymmetry above — cancelling heavy sitting (60-75 min) asks substantially more than banking the mortality benefit (~24 min) — was partly an artifact of pitting a self-report offset (60-75) against a device plateau (~24 min, Ekelund 2019). Device-vs-device, the offset dose (~34 min) and the mortality-plateau dose (~24 min) come from the same accelerometry consortium and are close — both near the guideline upper level (150-300 min/week ~= 21-43 min/day), which is exactly what Ekelund 2020 recommends for heavy sitters. So on a common instrument the two doses roughly coincide: enough MVPA to bank most of the mortality benefit is, to a first approximation, also enough to attenuate heavy sitting. The earlier the offset asks more reading survives only against self-report and is largely a measurement gap.
Independence: type-F, not type-E. Ekelund 2020 shares first-authorship and overlapping cohorts with
Ekelund 2016 and 2019 (and co-authors Jefferis / Lee with Paluch 2022), so it is a measurement refinement
of the held self-report offset, not an independent instrument — no [E-independent] token, no
sources: pad as corroboration. Caveats carry over and add: observational, median follow-up 5.5 y in some
cohorts (short), ≥40 y so not generalisable to younger/more active or LMIC populations; E-values 2.1-4.63
mean «uncontrolled confounding would have to be as large in magnitude as smoking to distort our results».
(Ekelund et al., 2016, 2019, inferred from 2020)
Older adults — falls, which is a patient-important outcome
«Long-term exercise is associated with a reduction in falls, injurious falls, and probably fractures in older adults, including people with cardiometabolic and neurological diseases.» (World Health Organization, 2020)
Falls and fractures are outcomes people care about directly, not surrogates — the strongest outcome class the annex carries for this stratum.
Quantified and mechanised by a gold Cochrane SR -> Exercise for Preventing Falls in Older Adults. The WHO annex states the association; Sherrington’s 108-RCT review supplies the effect and its shape: exercise cuts the rate of falls by 23% (RaR 0.77, 0.71-0.83, HIGH certainty), and the active ingredient is balance/functional training, not activity volume — resistance training alone does not reduce falls (Sherrington et al., 2019). This is a refinement, not independent backing — the WHO annex’s falls conclusion rests on largely the same RCT base the Cochrane SR pools, so treat the SR as bounding/mechanising the guideline claim, not corroborating it from a separate route. The fracture arm stays lower-certainty on both.
Second outcome — dementia (same shape: most benefit at the bottom)
Physical inactivity is one of the 14 modifiable dementia risk factors -> Dementia Prevention and Modifiable Risk Factors. A 58-study SR+MA found «physical activity was associated with a decreased risk of all-cause dementia (RR 0.80, 0.77 to 0.84, n=257,983)» across intensities and follow-ups >=20 years, with the reduction «greatest when moving from extreme sedentariness to some physical activity». (Livingston et al., 2024) This replicates the marginal-minute-worth- most-at-the-bottom shape on a second outcome. Caveat unchanged: the link is «likely to be bidirectional» (pre-clinical dementia reduces activity), and the one 5-year RCT of structured exercise found no overall cognition/MCI difference — so the observational RR 0.80 is a decreased-risk association, not a proven prevention effect. Where a multidomain bundle including exercise (FINGER RCT) did move a cognitive composite, the design cannot attribute that to the exercise component -> Multidomain Lifestyle Intervention and Cognitive Decline. Whether single-domain exercise RCTs on cognition are null may depend on the comparator. In that 5-year RCT the control arm was itself active: «At 5 years, 96% of the control group adhered to national guidance for physical activity, and 75% and 76% adhered to the interval training intervention» (Livingston et al., 2024). Livingston reads the overall null as «in line with the small cognitive benefit shown in an umbrella review of RCTs» (Livingston et al., 2024). Against health education, Mendes’s 109-RCT network finds exercise alone improves global cognition, SMD 0.14 (0.05-0.22), a small surrogate effect that, per the Results text, is not significant against an active (stretching / low-intensity exercise) control (the Abstract says the active-control models were «similarly significant»; see Multidomain Lifestyle Intervention and Cognitive Decline) (Mendes et al., 2025). That the comparator reconciles the null with the small positive is the wiki’s reading, not a held source’s (Livingston reconciles the null with a small benefit, but not by comparator). (corrected 2026-10-22: bare single-domain exercise RCTs on cognition stay null -> comparator-scoped.)
The direct dementia effect is now contested at the reverse-causation frontier (Kivimäki et al., 2019): Kivimaki’s individual- participant meta-analysis (n approx 404,840, 19 cohorts) found inactivity->all-cause dementia present at <10 y (HR 1.40, 1.24-1.59) but absent beyond 10 years (HR 1.01, 0.89-1.14), while its cardiometabolic positive controls (diabetes 1.42, CHD 1.24, stroke 1.16) held in both periods — so the borrowed RR 0.80 may be a prodromal marker for dementia specifically, not a causal slope. The clash with the 58-study aggregate MA is filed: Does Physical Activity Protect Against Dementia Beyond Ten Years. This page’s dose-shape claim is unchanged — activity’s biggest marginal payoff sits at the sedentary bottom, driven by the cardiometabolic and mortality outcomes BOTH studies agree on; only the dementia-specific causal reading narrows.
Third outcome — cancer, and the dose-shape may differ from the mortality curve
WCRF’s Third Expert Report grades physical activity protective against several cancers — colorectum (the evidence «is for colon cancer only», matrix FN56), endometrium and postmenopausal breast (probable), with a separate probable judgement for vigorous activity and breast cancer. (World Cancer Research Fund & American Institute for Cancer Research, 2018) Mechanism: activity «reduces body fatness, in particular visceral fat» and lowers circulating insulin, oestrogen and inflammation — the same adiposity/hormone/ inflammation routes that make body fatness carcinogenic, run in reverse. (World Cancer Research Fund & American Institute for Cancer Research, 2018)
The friction worth naming — the curve’s shape is outcome-specific. This page’s headline is that the mortality benefit flattens early (~24 min/day MVPA). WCRF states the opposite shape for cancer: «For cancer prevention, it is likely that the greater the amount of physical activity, the greater the benefit» — no plateau asserted. (World Cancer Research Fund & American Institute for Cancer Research, 2018) This is not a contradiction: it is the same outcome-specific dose-shape pattern the wiki already holds (fruit/veg plateaus on all-cause mortality but keeps a gradient on CV mortality in one source). So «most benefit at the bottom» is the mortality reading; for cancer the report leaves the upper arm open, and a person optimizing specifically for cancer risk cannot bank the mortality plateau as a stopping point. (inferred from World Cancer Research Fund & American Institute for Cancer Research, 2018) Caveat symmetric with the rest of the page: WCRF’s cancer grades are observational cohort judgements, so reverse causation (illness lowers activity) is not excluded by design.
What this page does NOT support
- No compensation analysis exists anywhere in the annex — now held on a sibling page. Whether a given activity dose is offset by reduced activity elsewhere, or by increased intake, is untouched here; Exercise Energy Compensation (Riou 2015) supplies it: compensation is real (~18%, rising to ~84% long-term) but is driven by adiposity/age/duration, not by intensity — so walking is better tolerated than intense exercise remains unsupported (intensity is not a compensation predictor).
- No ranking of modalities. See above.
- No evidence that brief distributed exercise snacks change mortality. For the near-inactive who will not do structured sessions, brief vigorous bouts spread across the day raise CRF and lower LDL-C (Exercise Snacks and Cardiometabolic Health, Wan 2025) — but that evidence is on surrogates over <=12 weeks, with no hard-outcome or mortality data, so snacks are a plausible entry lever onto this curve, not a demonstrated substitute for the doses this page prices against mortality. The authors themselves cast snacks as a complement, «rather than a replacement for these recommendations». (Wan et al., 2025)
- Frailty cannot be targeted from this evidence: WHO reports «A lack of consensus regarding the definition of frailty, and an absence of core measures to assess this means any attempt to create an optimal intervention will be impeded.» An ill-defined construct — distinct from a well-defined measure that decays under targeting. (World Health Organization, 2020)
- A study-level correlation is not an individual-level one. «At the study level, there was a positive correlation between the size of the exercise-induced effect on physical function and on cognitive function (b = 0.41; p = 0.002).» That does not license “improving function improves cognition” for a person — the ecological-inference trap. (World Health Organization, 2020)
Sedentary behaviour is not the inverse of activity — and devices have type-specific bias [2026-07-28, Willett ch.10]
Two findings from Willett’s physical-activity assessment chapter bear on how this page’s estimates should be read.
1. Sitting and moving are close to independent, not two ends of one scale.
«Interestingly, there is little correlation between sedentary behaviors and physical activity (Hu et al., 2003), suggesting that sedentary behaviors are not simply the opposite of physical activity.» (Willett, 2012)
So “sedentary time” and “activity” are two exposures, not one variable read in two directions, and a person can be high on both. This is a unit-of-analysis point with a direct consequence: a recommendation phrased as move more does not automatically address sitting, and evidence about one does not transfer to the other. -> Is the Food Category Doing Any Work (same structure, different domain: the label implies a single underlying quantity and there are two).
2. Device measurement is not error-free, and its error is activity-type-specific.
«(2002) found that compared with portable indirect calorimetry, the Tritrac accelerometer overestimated the energy expenditure of walking and jogging, and underestimated that of stair climbing and stationary cycling in middle-aged women.» (Willett, 2012)
This qualifies rather than undermines the device-measured estimate this page carries. The bias is differential by activity mode — over on ambulatory movement, under on climbing and cycling — so a device-measured total activity variable is a weighted mixture whose weights depend on the population’s activity mix. Two populations with identical true expenditure but different activity profiles will be measured differently, which is a transportability problem, not an attenuation problem. (inferred from Willett, 2012)
Note the vintage: this is 2012 hardware (Tritrac). Modern accelerometry and the processing behind the estimate on this page may behave differently, and this wiki holds no source establishing that either way. Recorded as a caution about the class of instrument, not as a defect in the specific estimate.
Is “24 min/day” false precision? Read it as a flat-region central estimate (deliverable-critique, 2026-08-01)
The ~24 min/day MVPA figure is a real source value - the point where the mortality dose-response is «close to the maximum», i.e. where the curve FLATTENS - not an invented threshold. But the concern is right: it should not be read as a precise minimum effective dose. Two held rules make it a REGION, not a number: a minimum effective dose is a region, not a number (outcome-specific), and required precision scales with local curvature - near a plateau the exact figure carries little decision weight. So 24 min is the central estimate of a broad flattening zone (~20-40 min/day of MVPA); whether it is 20, 24 or 40 barely changes the decision, and the decision-relevant shape is the one this page already gives: the marginal minute is worth the most at the bottom, and most of the benefit is banked well before the number is reached.