How much activity carries most of the risk difference?
All-cause RR 0.77 at half the guideline dose (about 75 min/wk), 0.69 at 150 min, 0.66 at double, vs none. Self-report cohorts: a curve, not proof.
Main caveat: Self-report, often without validation, usually one baseline measure; regression dilution likely, as the authors say.
From 196 articles, reported exposures were harmonised into marginal MET-hours per week of activity outside work, often by assuming session duration and intensity. Splines were fitted to each outcome. Differences in risk were greatest from 0 to 8.75 mMET-h/week, smaller to 17.5, and small and uncertain beyond.
The authors read this as substantial protection from small increases in inactive adults, with reservations about residual confounding and reverse causation. The population figures rest explicitly on a causal assumption.
Protocol
Eligible: prospective cohorts of adults without pre-existing conditions, over 10,000 participants, at least 3 activity levels, follow-up of 3 years or more; search to February 2019.
Analysis: two-stage random-effects dose-response with restricted cubic splines, knots by person-years. Most adjusted estimates were used.
Bias checks: six named sources, compared across studies at 8.75 mMET-h/week.
Funding: MRC, ERC, NIHR and others; competing interests none declared.
Key Results
| Outcome, RR vs 0 (95% CI) | 4.375 mMET-h/wk | 8.75 mMET-h/wk | 17.5 mMET-h/wk |
|---|---|---|---|
| All-cause mortality | 0.77 (0.73 to 0.80) | 0.69 (0.65 to 0.73) | 0.66 (0.62 to 0.70) |
| Cardiovascular mortality | 0.81 (0.77 to 0.85) | 0.71 (0.66 to 0.77) | 0.65 (0.60 to 0.71) |
| Cancer mortality | 0.90 (0.88 to 0.93) | 0.85 (0.81 to 0.89) | 0.82 (0.78 to 0.86) |
| Total cardiovascular disease | 0.83 (0.79 to 0.87) | 0.73 (0.69 to 0.79) | 0.67 (0.63 to 0.72) |
| Total cancer incidence | 0.93 (0.91 to 0.95) | 0.88 (0.85 to 0.92) | 0.85 (0.81 to 0.89) |
| Prostate cancer | 1.00 (0.99 to 1.02) | 1.01 (0.98 to 1.04) | 1.01 (0.96 to 1.05) |
| Kidney cancer | 1.02 (0.92 to 1.13) | 1.03 (0.89 to 1.19) | 1.04 (0.88 to 1.24) |
Data density: 76% of person-years below 17.5, 94% below 35 mMET-h/week.
Impact, if causal: 15.7% of deaths at 8.75, and 10.1% (95% CI 8.3 to 11.9) at half that dose.
Early deaths: 54-59% of studies did not exclude early events, and those that did tended to report weaker associations for some outcomes.
Reporting: the 15.7% CI is 13.1 to 18.2 in the abstract, 13.2 to 18.2 in the text and 13.16 to 18.15 in the table. Results say there was no evidence of bias by follow-up length, while the Discussion lists shorter follow-up among features of stronger associations. The abstract states deaths would have been averted without the causal assumption that the text gives.
Mechanisms
Measured: none.
Authors' puzzle: the composite (all-cause) association is as strong as cardiovascular mortality and larger than cancer, and total cardiovascular disease stronger than coronary disease or stroke alone. They suggest detection bias, misclassification, residual confounding and reverse causation by outcome.
Inference: an implausibly broad composite effect is one sign that part of the gradient reflects health status rather than activity.
Practical Application
How much activity carries most of the risk difference? All-cause RR 0.77 at half the guideline dose (about 75 min/wk), 0.69 at 150 min, 0.66 at double, vs none. Self-report cohorts: a curve, not proof.
Good for: a beginner or a person returning after a break. Even half the guideline sits on the steep part.
Not good for: valuing an athlete's extra hours; above 17.5 mMET-h/week differences were small and uncertain, and the data barely reach training volumes.
Not good for: promising cancer protection; site results range widely, and prostate and kidney associations were not shown.
Limitations
Self-report, often without validation, usually one baseline measure; regression dilution likely, as the authors say.
Harmonisation assumptions were used in many studies (standard values in 39-50%, imputation in 58-64%).
Residual confounding and reverse causation remain; 83-93% of studies adjusted for other morbidity, often alongside excluding more serious conditions.
Search date is February 2019, with a long lag to publication.
How this entry was checked and changed
Why this level
a meta-analysis of cohorts, not of controlled trials, so it stays at the observational cap despite its size; certainty was not rated with GRADE or a validated tool