Two gold sources on the same exposure-outcome pair, both built to defeat reverse causation, reach opposite verdicts. Iso-Markku’s aggregate SR+MA finds the protective association survives long follow-up (>=20 y, RR 0.79); Kivimaki’s individual-participant meta-analysis finds it vanishes beyond 10 years (HR 1.01). Both answer the same decision question — does being active prevent dementia, or does incipient dementia merely make people less active? — so the issue is joined (type-D). The hidden insight is that the clash is not about the reverse-causation window (both agree the short- follow-up signal is prodromal); it turns on outcome ascertainment and pooling level, and on the cleanest cut each offers, the two nearly converge toward the null. (inferred from Iso-Markku et al., 2022; Kivimäki et al., 2019)
View A — the association survives long follow-up (Iso-Markku 2022)
(Iso-Markku et al., 2022) Higher PA is associated with lower all-cause dementia (RR 0.80, 0.77-0.84, n=257,983), AD (0.86) and vascular dementia (0.79), and the effect holds within the 16 studies with >=20 y follow-up (RR 0.79, 0.71-0.87; mean baseline age 50.5, mean follow-up 27.6 y). Iso-Markku’s verdict: «we did not find evidence to suggest that reverse causation or regression dilution bias affected the observed associations between PA and dementias.» The design’s honest edge: its own cleanest cut — 3 high-quality studies with young (30-55) baseline AND >20 y follow-up — gives RR 0.79 (0.62-1.01), non-significant, and funnel plots «suggested some publication bias» for all-cause dementia and AD. So View A is a robust-but-observational association whose very cleanest subset is under-powered.
View B — the association vanishes beyond ten years (Kivimaki 2019)
(Kivimäki et al., 2019) The IPD splits follow-up within cohorts: inactivity->all-cause dementia HR 1.40 (1.24-1.59) at <10 y but 1.01 (0.89-1.14) at >=10 y (AD: 1.36 then 0.96), with no heterogeneity (I2=0%). The HR decays monotonically with the prodromal window (1.87 -> 1.30 -> 1.09 -> 0.87 across follow-up bands). Verdict: «physical inactivity was not associated with all-cause dementia or Alzheimer’s disease, although an indication of excess dementia risk was observed in a subgroup of physically inactive individuals who developed cardiometabolic disease» — and «there was little evidence that targeting physical inactivity alone would prevent dementia.» The positive controls are what make this more than a null: the same method kept inactivity robustly tied to diabetes (1.42), CHD (1.24) and stroke (1.16) in both periods, so «the lack of association between physical inactivity and dementia in the later follow-up is not an artefact of the methodology.»
Parameter table — is this the same quantity?
(inferred from Iso-Markku et al., 2022; Kivimäki et al., 2019)
| Parameter | Iso-Markku 2022 (View A) | Kivimaki 2019 (View B) | Same quantity? |
|---|---|---|---|
| Exposure | leisure-time PA, self-report, mixed categories | physical inactivity, self-report, binary/3-level | YES — self-reported leisure-time PA contrast |
| Outcome | incident all-cause dementia / AD / VaD | incident all-cause dementia / AD | YES — same endpoints |
| Long-follow-up estimate | RR 0.79 (0.71-0.87), >=20 y | HR 1.01 (0.89-1.14), >=10 y | YES — this is the joined clash |
| Design / pooling | aggregate SR+MA, 58 study-level estimates, I2 68.7% | IPD, 19 cohorts, within-cohort split, I2 0% | related, not identical |
| Reverse-causation handle | >=20 y subgroup (across-cohort) | exclude first decade (within-cohort) | related — both long-follow-up exclusion |
| Cleanest low-bias cut | 0.79 (0.62-1.01), NS, 3 HQ studies | 1.01 (0.89-1.14), full >=10 y | near-convergence, both near null |
| Dementia ascertainment | excluded dementia-mortality studies (insensitive) | 14/19 death-register; morbidity subgroup 0.97 (0.76-1.22) | NO — the divergence lever |
The headline estimates ARE the same quantity (a long-follow-up PA->dementia contrast), so this is a real joined clash, not a category mismatch. The rows that differ — pooling level and ascertainment — are the explanation of the divergence, not a reason to call it a false tension.
The issue is genuinely joined — Iso-Markku engages Kivimaki directly
(Iso-Markku et al., 2022) This is not two studies talking past each other. Iso-Markku names Kivimaki and rebuts it: «Our results contrast with those from Kivimaki et al … In that study, no associations were found between PA and all-cause dementia or Alzheimer’s disease when follow-ups were longer than 10 years. Notably, the incidence of all-cause dementia in their meta-analysis was 0.5%. This is an exceptionally low all-cause dementia incidence rate.» Two charges: (1) the cohort was too young — «mean age at the end of follow-up was approximately 60.4 years, but the mean age of all-cause dementia diagnosis in the study was 80.6 years»; (2) under-ascertainment — Iso-Markku «excluded studies with dementia mortality as the outcome because the relatively low sensitivity of death registers to detect dementia cases may underestimate its association.» Kivimaki’s data pre-empt both: the >60 subgroup (diagnosis age ~85) still gave 1.04 (0.90-1.19), and the morbidity-only ascertainment subgroup still gave 0.97 (0.76-1.22). The charges are weakened, but the morbidity arm is under-powered — so the ascertainment question stays genuinely open.
Hidden insight — the clash is ascertainment + pooling, and the clean evidence nearly agrees
(inferred from Iso-Markku et al., 2022; Kivimäki et al., 2019)
- Not a window disagreement. Both use a long-follow-up exclusion and both agree the <10 y signal (Kivimaki’s 1.40; the short-follow-up base of the older MAs) is prodromal. They differ on what remains after the exclusion.
- The IPD has the cleaner handle. A within-cohort 10 y split with harmonised exposure and I2=0% removes reverse causation more directly than an across-cohort >=20 y subgroup drawn from 58 studies at I2 68.7% with detected publication bias. Kivimaki’s positive controls (cardiometabolic disease stayed associated in both periods) validate that its null is not a power failure — a check View A’s aggregate design cannot run.
- On the cleanest cut, they nearly converge. Iso-Markku’s own best-quality young-baseline >20 y subset is 0.79 (0.62-1.01) — non-significant; Kivimaki’s >=10 y is 1.01 (0.89-1.14). The genuinely low-bias long-follow-up evidence in both is compatible with a null-to-modest effect. The headline 0.79-vs-1.01 gap is largely between View A’s larger, lower-quality, publication-biased pool and View B’s harmonised, zero-heterogeneity, ascertainment-robust IPD — a difference in study quality and ascertainment sensitivity, not an irreducible contradiction. Neither rules out a modest true effect the clean cuts are under-powered to find (the insufficient-evidence state, not proven no-effect). -> The U-Shaped Association Artifact
What it changes about what to do
(inferred from Iso-Markku et al., 2022; Kivimäki et al., 2019) The recommendation direction does not move — be active — but its warrant does. Both sources agree inactivity causes the cardiometabolic diseases (diabetes, CHD, stroke) that themselves raise dementia risk; Kivimaki’s one surviving dementia signal runs through that route (post-cardiometabolic HR 1.30, 0.79-2.14). So physical activity keeps its high Layer-1 rank as a non-substitutable, pleiotropic big rock via the vascular channel, and it is recommended for cardiometabolic reasons regardless. What weakens is the claim that activity prevents dementia directly and independently of its cardiometabolic effects — the reading the Commission’s borrowed 0.80 cell invites. RCT evidence sits with View B’s caution: multidomain trials show a small cognitive-surrogate benefit but a null on dementia incidence -> Multidomain Lifestyle Intervention and Cognitive Decline. Confidence in a direct dementia-specific effect is low; confidence in the vascular-route benefit is not in dispute.
Self-critique [run 2026-09-10, before commit]
- Joined, not fabricated. Iso-Markku names and rebuts Kivimaki by DOI-level specificity; the headline estimates are the same quantity (long-follow-up PA->dementia). Not-joined checks all fail to fire: the outcomes are identical (not merely verbal), matching scope sharpens rather than dissolves the clash, and neither source is cherry-picked (both positions read end-to-end, including Iso-Markku’s full Discussion rebuttal and Kivimaki’s pre-emptive subgroup/ascertainment robustness).
- Not overclaimed toward View B. The IPD’s design advantages are stated as tilting the weighing, not settling it; the honest convergence (Iso-Markku’s cleanest cut is also NS) is surfaced, and the residual open question (Kivimaki’s morbidity-only arm is under-powered) is named rather than buried. The verdict is insufficient evidence for a direct effect, not proven no effect.
- No independence laundering. These are NOT type-E corroboration — they disagree; and Kivimaki’s positive controls corroborate only the cardiometabolic causation, not the dementia null. No confidence was lifted on agreement.
- Coherence, not validity (R1): the loop is open — neither source, and not this page, grades PA against a realized dementia outcome under intervention. A clean audit here is not a validated claim that exercise does or does not prevent dementia.