The interventional (RCT) facet of the dementia cluster — orbits the nucleus
Dementia Prevention and Modifiable Risk Factors (which holds the observational per-factor PAF map).
This page holds the three landmark randomised tests (FINGER, MAPT, preDIVA) of whether acting on those
levers as a bundle helps, plus the pooled individual-participant responder analysis of two of them
(Coley 2025) — and the story is that the result is not robust: it depends on the endpoint, the
population, and the intervention’s intensity, and when the two hard-endpoint trials are pooled and searched
for the subgroup that responds, no responder subgroup rescues it (see The pooled responder verdict
below). A network meta-analysis of the whole RCT literature (Mendes 2025, 109 trials) then ranks the
combinations and finds the leanest effective one — exercise + cognitive training — beats the fuller
multidomain bundle: more is not better (see The comparative hierarchy below).
- FINGER (Ngandu 2015) — a 2-yr, double-blind, 6-centre RCT (n=1260, selected at-risk CAIDE >=6 but cognitively near-normal Finns) that moved a cognitive-composite surrogate (below).
- MAPT (Andrieu 2017) — a 3-yr, 4-arm placebo-controlled RCT (n=1680, elderly with a memory complaint, France/Monaco) testing a multidomain lifestyle bundle +/- omega-3 on the same kind of cognitive-composite surrogate FINGER moved — and it was NULL at the whole-population level (see The non-replication — MAPT below). The sharp within-family comparison: same outcome type, closest intervention type, opposite result. (Andrieu et al., 2017)
- preDIVA (Moll van Charante 2016) — a 6.7-yr open-label cluster-RCT (n=3526, unselected community-dwelling Dutch aged 70-78) that showed no reduction in the hard endpoint, dementia incidence (HR 0.92; the incidence anchor FINGER lacked — see The hard-incidence anchor below). (van Charante et al., 2016)
FINGER first: it is a proof-of-concept trial: “a proof-of-concept randomised controlled trial, we aimed to assess a multidomain approach to prevent cognitive decline in at-risk elderly people from the general population.” (Ngandu et al., 2015)
The effect — real, significant, and small
effect_measure: NTB total cognitive-composite Z-score, between-group difference in change per year
= 0.022 (95% CI 0.002-0.042, p=0.030); “Improvement in NTB total score after 24 months was 25%
higher in the intervention group than in the control group”; Cohen’s d 0.13 at 2 years.
population_and_comparator: at-risk (CAIDE >=6), cognitively near-normal elderly (60-77), vs an active
control (general health advice + mailed lab feedback), NOT a do-nothing arm.
outcome: a cognitive-test composite (surrogate), NOT dementia incidence.
uncertainty: CI excludes zero but the lower bound (0.002/yr) is near-null; d=0.13 is a small effect.
certainty: single proof-of-concept RCT; confidence: low (unreplicated; surrogate endpoint).
- Both groups improved (2-yr NTB total change 0.20 intervention vs 0.16 control) — partly practice effects of repeated testing; the between-group gap is the effect. (Ngandu et al., 2015)
- What reached significance: overall cognition (p=0.030), executive functioning (p=0.039, +83%), processing speed (p=0.029, +150%). What did NOT: the prespecified memory domain — the intervention showed no significant change on it (figure p=0.36); only a post-hoc complex-memory score reached p=0.036. (Ngandu et al., 2015)
- Cognitive-decline odds (post-hoc, control vs intervention): NTB total OR 1.31 (1.01-1.71); executive 1.29 (1.02-1.64); processing speed 1.35 (1.06-1.71); memory 1.23 (0.95-1.60) — NS. (Ngandu et al., 2015)
- Safe, adherable: all-four-domain participation 72%; “Adverse events occurred in 46 (7%) participants in the intervention group compared with six (1%) participants in the control group; the most common adverse event was musculoskeletal pain” (exercise-related). No serious intervention-related AEs. (Ngandu et al., 2015)
Three limits that bound what this licenses (all author-stated)
- Surrogate boundary. The endpoint is a cognitive composite; dementia/AD incidence was not measured — a 7-year extended follow-up is planned to assess intervention effects on incidence of dementia and Alzheimer’s disease. So FINGER shows the surrogate moved, not that dementia was prevented. (Ngandu et al., 2015) -> Surrogate Outcomes
- Non-decomposability. “The multimodal intervention model needs to be investigated further, particularly with regard to the contribution of each component.” The design cannot say whether diet, exercise, cognitive training or vascular care did the work — a bundle proven as a bundle. This is itself decision-relevant: the trial licenses the package, not a claim about any single lever, and it cannot validate the observational per-factor PAF decomposition on the nucleus page. (Ngandu et al., 2015)
- Conservative vs active control; public-health not personal. Estimates “could be considered to be conservative” (adherence unadjusted; control also got vascular advice) and benefits “might thus be greater if compared with a do-nothing control group”; but the flip side is that FINGER speaks to public-health, not personal, significance — the trial’s own reading is that public-health significance is not easily translated into clinical or personal significance. A d=0.13 population effect does not promise a given person a measurable cognitive gain. (Ngandu et al., 2015)
What it changes about what to do
- The direction is now RCT-backed, the magnitude bounded. Acting on the modifiable levers as a bundle produces a real but small cognitive benefit in an at-risk stratum — enough to raise confidence that the observational levers are worth pulling, not enough to over-sell a personal cognitive payoff. Most of those levers (BP, glucose, weight, activity, diet) are the cardiometabolic big rocks already pulled for other outcomes -> Layer 1 - Ranking Interventions for a Stratum, Big Rocks (Elderly), so the cognition signal is a second patient-important outcome on levers already ranked high, not a new free-standing intervention to add.
- Transportability is baseline-risk-conditioned (route a). The effect was found in CAIDE-high, cognitively-near-normal elderly. It does not transport to dementia patients (excluded) or to a low-risk group; absolute benefit scales with baseline dementia risk. Dietary self-report in the diet arm carries the usual measurement error -> Measurement Error in Dietary Assessment.
- Single-domain trials had been null; the multidomain bundle was positive — consistent with the vault’s held observation that a lone structured-exercise RCT found no cognition benefit -> Physical Activity Dose and Mortality. But non-decomposability means this is not evidence that exercise (or any one component) works alone.
- Nor does the modality of exercise resolve it. A 2023 umbrella claims mind-body exercise outperforms aerobic/resistance for cognition, but the advantage is a cross-review comparator artifact, not established modality-specific benefit -> Mind-Body Exercise and Cognition. So neither the dose nor the type of the exercise component rescues a single-domain cognition effect here.
The non-replication — MAPT (3-yr, 4-arm RCT, NULL on the same surrogate) [2026-08-07]
Where preDIVA differed from FINGER on both the intervention and the endpoint, MAPT is the sharp comparison: it moved the same kind of cognitive-composite surrogate, with a bundle of the same lifestyle type (physical activity + cognitive training + nutrition) — and found no effect at the whole-population level. «The multidomain intervention and polyunsaturated fatty acids, either alone or in combination, had no significant effects on cognitive decline over 3 years in elderly people with memory complaints.» (Andrieu et al., 2017)
effect_measure: change baseline->36 months on a composite Z score of four cognitive tests (FCSRT free+total
recall, ten MMSE orientation items, DSST, Category Naming Test); between-group difference vs placebo.
population_and_comparator: n=1680 (mITT 1525), community-dwelling >=70, selected on a spontaneous memory
complaint, one-IADL limitation, or slow gait (NOT cardiovascular-risk-selected); 13 memory centres,
France/Monaco. Placebo / no-multidomain comparator (weaker contrast than FINGER’s active control).
outcome: cognitive-composite surrogate, not dementia incidence (same construct-type as FINGER’s NTB).
result: NULL after multiple-comparison correction. «there were no significant differences in 3-year
cognitive decline between any of the three intervention groups and the placebo group. Between-group
differences compared with placebo were 0·093 (95% CI 0·001 to 0·184; adjusted p=0·142) for the combined
intervention group, 0·079 (–0·012 to 0·170; 0·179) for the multidomain intervention plus placebo group,
and 0·011 (–0·081 to 0·103; 0·812) for the omega 3 polyunsaturated fatty acids group.»
(Andrieu et al., 2017)
certainty: single large RCT, null on the primary; confidence: low unchanged (the null lowers, not raises,
confidence in the FINGER surrogate signal).
- The combined arm was a borderline positive that died on correction. Its raw p=0·047 and CI lower bound (0·001) barely excluded zero, but Hochberg correction for the three comparisons pushed adjusted p to 0·142 (Andrieu et al., 2017). And even taken at face value, 0·093 over 3 years is far below the composite’s minimal clinically important difference (an anchor-based estimate put the MCID near –0·3 points per year) (Andrieu et al., 2017) — so the surrogate barely moved, if at all.
- A post-hoc pooled signal, but read it as post-hoc. Pooling all multidomain recipients regardless of arm, «cognitive decline from baseline to 36 months… was significantly less in those who had received the multidomain intervention than in those who did not receive this intervention (data not shown; p=0·015)», whereas pooled omega-3 «did not differ significantly… (data not shown; p=0·715)» (Andrieu et al., 2017). This is a non-randomised factorial contrast (data not shown), not the prespecified primary — suggestive, not confirmatory.
- The omega-3 arm was flat — an omega-3-for-cognition null. Supplementation (800 mg DHA + 225 mg EPA/day for 3 yr) did nothing alone (adjusted p=0·812) and added nothing to the bundle. Separately, low baseline red-blood-cell DHA+EPA (lowest quartile) marked a 0·24-point 3-year decline «similar to the worsening noted in patients with CDRs of 0·5… whereas those with normal DHA and EPA concentrations showed no change» (Andrieu et al., 2017) — a marker-vs-intervention gap: low omega-3 status flags decline, but supplementing omega-3 did not prevent it -> Vitamin and Mineral Supplements for Disease Prevention (marine omega-3 also null for CVD/cancer in VITAL). (inferred from Andrieu et al., 2017) This is the RCT half of a diet-vs-supplement split: the observational cohort SR+MA (Wei 2023) finds dietary DHA protective while its supplement-use signal is confounded and its plasma-DHA biomarker null — the same marker-vs-lever pattern -> Fish and Seafood Consumption. (inferred from Wei et al., 2023)
FINGER vs MAPT — the parameter table (why this is EFFECT-MODIFICATION / a REFINEMENT, not a filed tension).
| Parameter | FINGER (Ngandu 2015) | MAPT (Andrieu 2017) | Same quantity? |
|---|---|---|---|
| Design | 2-yr individually-randomised RCT | «a 3-year, multicentre, randomised, placebo-controlled superiority trial with four parallel groups» (chunk 01) | ~ (same class; 2 vs 3 yr) |
| Population | selected at-risk (CAIDE >=6), near-normal, 60-77 | «aged 70 years or older… spontaneous memory complaint… limitation in one instrumental activity of daily living, or slow gait speed» — NOT CV-risk-selected, older (chunk 01) | NO |
| Intervention | 4-domain: diet + supervised exercise + cognitive training + vascular monitoring | «43 group sessions integrating cognitive training, physical activity, and nutrition, and three preventive consultations» — PA is advice-based, no supervised-exercise domain (chunk 01) | ~ (same TYPE, lower intensity) |
| Primary endpoint | NTB cognitive-composite Z score (surrogate) | composite Z score of four cognitive tests (surrogate) | YES |
| Comparator | active control (general health advice) | placebo / no-multidomain (a weaker contrast that should favour a positive) | NO (MAPT’s is easier to beat) |
| Result | positive: 0.022/yr, Cohen’s d 0.13 (p=0.030) | NULL: multidomain 0·079 (adj p=0·179); combined 0·093 (adj p=0·142) | OPPOSITE |
The endpoint row is YES — the closest same-endpoint comparison of the three trials (preDIVA’s endpoint row was NO). So the surrogate itself is now shaky: FINGER barely moved it (d=0.13) and MAPT did not move it at population level even with a weaker comparator that should have made a difference easier to detect -> Surrogate Outcomes. But the population and intervention-intensity rows are NO, so the not-joined check (ii, different scope/population) fires: the two are consistent once conditioned on population — and MAPT’s own internal subgroups supply the conditioning evidence. «Cognitive decline in participants with a CAIDE score of 6 or greater at baseline was less in the combined intervention group than in the placebo group (p=0·023)» (Andrieu et al., 2017) — i.e. MAPT’s CAIDE>=6 stratum shares FINGER’s high-risk selection criterion (the two differ on age and entry, but both are high-baseline-risk), and there MAPT reproduces a benefit. The amyloid-positive subsample (n=269) showed the same: «Less cognitive decline during follow-up was noted in the combined intervention group (adjusted p<0·0001) and in the multidomain intervention plus placebo group (p=0·003)… in amyloid-positive participants» (Andrieu et al., 2017). The authors read it the same way: the intervention «might help to slow cognitive decline in people most likely to undergo decline—ie, those with a CAIDE dementia risk score of 6 or greater at baseline, and those with a positive amyloid PET scan… all subgroup results should be considered exploratory and need to be confirmed» (Andrieu et al., 2017).
Recorded as a REFINEMENT (type F) + a route-(b) effect-modification hypothesis, NOT a tension
[inferred from @ngandu2015; @andrieu2017]. The deciding checks: the parameter table’s result clashes but its population/intensity rows
differ (not-joined ii), and the clash is internally reconciled — MAPT’s higher-risk subgroups replicate
FINGER’s stratum — which is effect-modification, not contradiction (filing a tension here would be a
fake-tension, since the moderator is evidenced within the null trial itself). The composite finding that
survives: the multidomain cognitive-composite effect does not robustly replicate at the whole-population
level; it appears confined to higher-baseline-risk strata (CAIDE>=6, amyloid-positive), the same stratum
FINGER selected for. MAPT also stated the mechanism of its own null: a higher-education, low-decline
sample («the level of education was higher than expected… the rate of cognitive decline in 3 years would
probably be lower than initially expected») left little decline for the intervention to prevent
(Andrieu et al., 2017) — a floor effect on the surrogate,
reinforcing that baseline risk is the live moderator. Whether that moderation holds on a hard endpoint
is the responder question the pooled MAPT+preDIVA IPD (Coley 2025) was ingested to test — and it does
not (the CAIDE>=6 stratum showed no pooled hard-endpoint intervention benefit: AD-dementia HR 1.03,
0.76–1.40, and all-cause subgroups uniformly null; see The pooled responder verdict — Coley 2025 below).
The hard-incidence anchor — preDIVA (6.7-yr cluster-RCT, NULL) [2026-08-07]
FINGER moved a surrogate; preDIVA (Moll van Charante 2016) is the trial that measured the patient-important outcome directly — clinical dementia incidence over 6.7 years — and found nothing. A pragmatic, nurse-led, multidomain vascular-care intervention (18 visits over 6 years: smoking, diet, activity, weight, BP, plus drug treatment for hypertension / dyslipidaemia / diabetes where indicated) in an unselected population of 3526 Dutch aged 70-78. (Ngandu et al., 2015; inferred from van Charante et al., 2016) Unlike FINGER, this bundle carried no cognitive-training and no supervised-exercise component — inferred from preDIVA’s intervention list (advice + drugs) against the source’s own FINGER contrast («lifestyle interventions, drug treatment, and cognitive training»), the difference that makes the two non-comparable below.
- Primary outcome — no reduction shown. Dementia developed in 121/1853 (7%) intervention vs 112/1601 (7%) control: HR 0.92 (95% CI 0.71-1.19), p=0.54 (van Charante et al., 2016). The second co-primary, disability (ALDS), was also null (adjusted mean difference -0.02, 95% CI -0.38 to 0.42, p=0.93), as were all-cause mortality (HR 0.98, 0.80-1.18) and incident cardiovascular disease (HR 1.06, 0.86-1.31). «A nurse-led, multidomain intervention did not result in a reduced incidence of all-cause dementia in an unselected population of older people.» (van Charante et al., 2016)
- But read the dementia result as not-yet-shown, not shown-absent (the four evidence-states apply). The HR of 0.92 is directionally protective, and the authors say so: it «is consistent with these findings [antihypertensive-MA all-cause-dementia risk reductions of only 2-9%], although our study was underpowered to detect such an effect size» (van Charante et al., 2016). So on the dementia endpoint preDIVA is insufficient-evidence / underpowered-for-the-plausible-small-effect, not a demonstration of no effect — the trial was powered for a 33% reduction, an order of magnitude larger than BP-lowering plausibly delivers. (The CVD, mortality and disability nulls are more genuinely flat given the small achieved contrast — next bullet.)
- The intervention did move the risk factor, modestly. Systolic BP fell more in the intervention arm (adjusted mean difference -2.06 mm Hg, 95% CI -3.21 to -0.90, p=0.0005) (van Charante et al., 2016) — so this is a null despite a measured contrast, not a null from an untouched exposure.
- One subtype signal, small numbers: non-Alzheimer’s dementia was less frequent in the intervention arm (11 [1%] vs 23 [2%]; HR 0.37, 0.18-0.76, p=0.007), consistent with a vascular mechanism, but on tiny counts and not the prespecified primary (van Charante et al., 2016). Alzheimer’s itself was null (HR 1.05, 0.78-1.41).
FINGER vs preDIVA — the parameter table (why this is a DISTINCTION, not a tension).
| Parameter | FINGER (Ngandu 2015) | preDIVA (Moll van Charante 2016) | Same quantity? |
|---|---|---|---|
| Design | 2-yr double-blind, individually-randomised RCT | 6.7-yr open-label cluster-RCT (116 GP practices) | NO |
| Population | selected at-risk (CAIDE >=6), near-normal, 60-77 | unselected community-dwelling, 70-78 (not CV-risk-selected) | NO |
| Intervention | 4-component: diet + exercise + cognitive training + vascular monitoring | nurse-led vascular care only (advice + drugs; no cognitive training, no structured exercise) | NO |
| Primary endpoint | NTB cognitive-composite Z-score (surrogate) | clinical dementia diagnosis incidence (hard) + disability | NO |
| Result | positive: between-group 0.022/yr, Cohen’s d 0.13 (p=0.030) | NULL: HR 0.92 (0.71-1.19); disability, CVD, mortality also null | NO |
Same-quantity column is NO on every row, so the apparent clash (“multidomain works” vs “multidomain fails”) is not a joined issue — the not-joined check (ii, different scope/endpoint/intervention/population) fires decisively. (inferred from Ngandu et al., 2015; van Charante et al., 2016) The two trials answer different questions: FINGER shows a 4-component bundle (incl. cognitive training) moves a cognitive surrogate over 2 yr in a selected high-risk group; preDIVA shows vascular-care-only does NOT lower hard dementia incidence over 6.7 yr in unselected elderly. Recorded as a surrogate-positive / hard-incidence-null distinction, not a tension — neither refutes the other because they never measured the same thing.
Why preDIVA was null — four author-stated reasons, all decision-relevant [inferred from @mollvancharante2016]. «This absence of effect might have been caused by modest baseline cardiovascular risks and high standards of usual care.» (van Charante et al., 2016) Unpacked: (i) small between-group contrast — pragmatic, modest-intensity intervention; (ii) high usual care + Hawthorne effect — 2-yearly screening prompted treatment in the control arm too (BP fell in both), and a 2011 Dutch guideline update pushed more proactive prevention in the over-70s; (iii) unselected, mostly modest-risk population — not chosen for high CV risk, shrinking absolute benefit -> Cardiometabolic Interventions and Hard CV Outcomes in Low-Risk People; (iv) wrong life-course window — «our population was aged 70-78 years, whereas most observational data show an association between midlife (age 40-60 years) vascular risk factors and dementia» (van Charante et al., 2016), so late-life vascular care may simply come too late (the nucleus’s earlier-and-longer timing rule -> Dementia Prevention and Modifiable Risk Factors).
The subgroup — a route-(b) hypothesis, NOT the headline. preDIVA reports a per-protocol/adherent effect concentrated in people with untreated hypertension at baseline (adherent-untreated-HTN: HR 0.54, 0.32-0.92, p=0.02; no-CV-history adherent: HR 0.64, 0.44-0.94); separately, the overall per-protocol adherent group’s 24% lower hazard would translate to an absolute risk reduction of ~1.7% (7.2% -> 5.5%) (van Charante et al., 2016) — note that 1.7% is the overall-adherent figure, not the untreated-HTN cell. «Our results do not rule out clinically meaningful effects in people with untreated hypertension who are adherent to the intervention.» (van Charante et al., 2016) Held as an effect-modification hypothesis (route b) to confirm, not a finding — post-hoc, adherence-selected (adherence is not randomised), small counts. This is the responder question the pooled IPD (Coley 2025, MAPT+preDIVA) was ingested to test — and the untreated-hypertension responder cell did not replicate in the pool (AD-dementia intervention HR 0.72, 0.44-1.18, P-interaction 0.071, NS; all-cause subgroup effects likewise null; see The pooled responder verdict — Coley 2025 below).
The pooled responder verdict — Coley 2025 (MAPT+preDIVA IPD, responder NULL) [2026-08-07]
The route-(b) hypothesis the two cognitive-surrogate trials left open — the multidomain effect is real but confined to higher-baseline-risk strata — now has a direct test on the hard endpoint. Coley pooled the individual-participant data of MAPT + preDIVA (n=5205, aged >=70, up to 12 yr follow-up, 486 dementia cases over 37,782 person-years) and went hunting for the responders two ways: 11 pre-specified risk-factor subgroups AND a data-driven SIDES recursive-partitioning search free to combine any variables at any cut-point (min group >100). It found none.
effect_measure: multidomain vs control, incident all-cause dementia (shared-frailty Cox on pooled IPD).
result: overall NULL — «there was no effect of multidomain intervention on the risk of all-cause
dementia (HR 0.98, 95% CI 0.80–1.21)», and «the intervention effect did not differ between the two trials
(P-interaction: 0.928)» (Coley et al., 2025). A
genuinely flat point estimate on the patient-important endpoint; the pool roughly triples preDIVA’s case
count, so it is a firmer null than preDIVA’s underpowered HR 0.92 alone.
subgroups: no pre-specified subgroup showed an effect — «There was also no difference in the effect of
the intervention in any of the pre-specified subgroups» (Coley et al., 2025). The two the route-b reading singled out showed no intervention benefit on the
tabulated intervention-effect figures (Table 3, AD dementia — multidomain vs control): CAIDE>=6 HR
1.03 (0.76–1.40, P-interaction 0.993) and untreated-hypertension — preDIVA’s own headline responder
cell — HR 0.72 (0.44–1.18, P-interaction 0.071), both non-significant; the all-cause intervention
effects across all 11 subgroups were likewise no-difference (Figure 1)
(Coley et al., 2025). (Table 2’s HR
column is the subgroup-vs-referent incidence ratio — prognostic, e.g. CAIDE>=6-vs-<6 = 1.37 — not the
intervention effect.)
data_driven: «the SIDES algorithm did not detect any subgroup showing a differential intervention effect on
incident all-cause dementia» (Coley et al., 2025).
This is the load-bearing part of the null: not merely these moderators failed but an algorithm free to
combine all of them found no responder group > 100 people — and «we did not adjust for multiple
statistical comparisons», so the search was biased toward a false positive and still returned nothing.
certainty: the strongest single test of the route-b hypothesis the unit holds; confidence: low in a
multidomain benefit is unchanged, but the responder reading moves from open to evidenced-against.
So the route-(b) baseline-risk-responder hypothesis is resolved toward the NULL on the hard endpoint (inferred from Coley et al., 2025). Per the telos, a route-b subgroup claim requires positive effect-modification evidence; Coley supplies a well-powered absence of it — the important, decision-changing null for personalization. The cognitive-surrogate CAIDE>=6 / amyloid signals (FINGER selected on CAIDE>=6 and was positive; MAPT’s subgroups reproduced it) do not carry over to dementia incidence when the two hard-endpoint trials are pooled and searched. The authors read it the same way: «the potential impact of multidomain interventions for dementia prevention at the individual level, even among those at greatest risk of dementia, appears modest at best» (Coley et al., 2025), and pivot to Rose’s prevention paradox — «a more effective approach might be to channel dementia prevention efforts into population-level approaches» (Coley et al., 2025) — i.e. if the responders cannot be found, the lever is public-health, not personalized-clinical. This is a worked route-(b)-null case -> Baseline Risk and the Relative-Absolute Split: the prognostic gradient is real (older age, APOE ε4, inactivity, low MMSE all raised incidence), yet no modifier of the intervention effect exists — route-a predictors present, route-b modifiers absent.
Independence — the pooled COMPOSITE of the two trials above, NOT a third witness (type F/A, never E)
(inferred from Coley et al., 2025). Coley re-analyses the MAPT+preDIVA
participants and shares their authorship completely — Coley/Andrieu are MAPT; Hoevenaar-Blom/Moll van
Charante/van Dalen/van Gool/Richard are preDIVA — so it is not independent evidence and is never counted as
corroboration (no [E-independent]). It also excludes FINGER (the one positive trial): the pooled null
does not incorporate FINGER’s signal, so it bounds what the pool can say about the vascular-heavy
MAPT+preDIVA bundle, not about FINGER’s full 4-component version.
What bounds the null — author-stated, keeping it a strong null not an absolute one. MMSE was the only shared cognitive measure and is «insensitive as an outcome measure, notably in prevention trial settings» (Coley et al., 2025) — so the cognitive-change null is the weak arm; the dementia-incidence null is the firm one. The low-MMSE subgroup (n=410) was underpowered and showed only a non-significant favourable trend. Amyloid status — MAPT’s other positive subgroup — could not be tested (no biomarkers in preDIVA). The authors note higher-intensity, younger / midlife, or biomarker-selected trials «could provide different results» — the responder question is closed for these late-life, low-intensity trials, not for every design.
The comparative hierarchy — Mendes NMA 2025 (109 RCTs; E+C > the full bundle) [2026-09-05]
The three trials above test specific bundles one at a time; none answers which combination is best, or whether a fuller multidomain beats a leaner one. Mendes 2025 — a PROSPERO-preregistered random-effects network meta-analysis of 109 RCTs (23,010 cognitively-UNIMPAIRED older adults, median age 70) on the global-cognition surrogate — is the first to rank single- vs multidomain lifestyle interventions in one framework (Mendes et al., 2025). Its decision-changing result: the leanest effective combination wins, and the most comprehensive bundle does not top the ranking.
- Ranking vs health education: «physical exercise and cognitive training combined (SMD 0⋅26 [95% CI 0⋅10–0⋅42]…); cognitive training alone (SMD 0⋅21 [0⋅08–0⋅33]…); diet, physical exercise, cognitive training, and health education combined (SMD 0⋅14 [0⋅02–0⋅27]…); and physical exercise alone (SMD 0⋅14 [0⋅05–0⋅22]…)» all significantly improved cognition; diet-alone, social-activity-alone and the other multidomain nodes did not (Mendes et al., 2025). Exercise + cognitive training (E+C) is the strongest of all twelve — above the full diet+exercise+cognitive-training+health-education bundle (D+E+C+H, SMD 0·14).
- The two robust single levers are cognitive training and physical exercise: «When compared with active control, only physical exercise and cognitive training combined (p=0⋅038) and cognitive training alone (p=0⋅044) revealed significant improvement» (Mendes et al., 2025) — the full multidomain bundle is NOT significant against an active control, its benefit shrinking exactly as FINGER’s authors warned their active-control estimate would.
- More is not better. «Combining lifestyle interventions might enhance efficacy, but increased number of domains does not automatically translate into greater cognitive benefits» (Mendes et al., 2025); «more is not always better—adding additional interventions does not necessarily enhance efficacy and might introduce adherence challenges, particularly in multidomain protocols with more than three interventions» (Mendes et al., 2025). Author mechanisms: a curvilinear complexity/adherence curve («moderate-complexity interventions yielding the greatest effects»), a within-domain plateau (MAPT «cognitive benefits plateaued after approximately half of the planned training sessions»), dilution (embedded cognitive training is «less intensive, individualised, or tailored» than standalone), and mechanistic overlap reducing additive benefit (Mendes et al., 2025). This is adherence-is-part-of-the-effect and structural leverage made quantitative -> Layer 1 - Ranking Interventions for a Stratum.
Independence — a re-synthesis of the same RCT field, NOT a third witness (type F/C, never E)
[inferred from @mendes2025multidomain]. Mendes’s 109-trial
network is drawn from the lifestyle-cognition RCT literature and explicitly benchmarks its result against
FINGER (ref 13) and MAPT (ref 14) — the trials this page holds — reporting its effect sizes as «similar
to those reported in landmark trials» (Mendes et al., 2025). Whether FINGER and MAPT fall among its 109 included trials is not verifiable from
the held text (the included-study list is in the appendix; the FINGER-aligned diet+exercise+cognitive-
training+health-education node holds n=9, and MAPT’s omega-3 arms would be excluded by the no-supplement
criterion while its multidomain-placebo arm could qualify). But it does not change the classification: an
NMA re-synthesizing this field shares its evidence base and method-family, so it is never independent
type-E corroboration. The author lists are wholly disjoint (Mendes/Ribaldi/Frisoni, Geneva; none overlap
Ngandu/Kivipelto, Andrieu/Coley, or Moll van Charante) — but author disjointness does not buy
independence, and the conservative rule is to never assert it. It is therefore never marked
[E-independent]. Its value is type-F (refines the held bundle claim into a ranked hierarchy) and type-C
(supplies the single-vs-multidomain comparative axis the pairwise trials could not).
FINGER vs MAPT vs Mendes — the parameter table (why this is a REFINEMENT + magnitude-concordance, NOT a tension).
| Parameter | FINGER (Ngandu 2015) | MAPT (Andrieu 2017) | Mendes NMA (D+E+C+H node) | Same quantity? |
|---|---|---|---|---|
| Effect metric | between-group NTB change; Cohen’s d 0·13 at 2 yr | between-group composite diff 0·079–0·093 over 3 yr | pooled SMD 0·14 (0·02–0·27) vs health ed | ~ (all standardized cognitive-composite; FINGER’s is 2-yr d, others total) |
| Population | selected CAIDE>=6, near-normal, 60-77 | memory-complaint elderly >=70 | cognitively unimpaired, 109 trials pooled | NO (Mendes broader/healthier, pooled) |
| Intervention | 4-domain incl vascular monitoring | PA+CT+nutrition, advice-based | D+E+C+H (4th domain = health education) | ~ (close; H replaces vascular) |
| Primary endpoint | NTB cognitive-composite (surrogate) | 4-test composite (surrogate) | global cognition MMSE/MoCA/composite (surrogate) | YES (all surrogate) |
| Comparator | active control | placebo / no-multidomain | health ed / no-int / active control | mixed (bundle NS vs active control) |
| Result | +ve small (d 0·13, p=0·030) | NULL overall (adj p 0·142–0·179) | +ve small vs health-ed/no-int (0·14–0·20); NS vs active control | concordant band; NOT independent |
The endpoint row is YES (all the cognitive surrogate), but population/intervention/comparator rows differ, so the apparent clash between MAPT’s whole-population null and Mendes’s significant bundle is not joined (not-joined check ii, different scope) — Mendes’s network aggregates many trials of overlapping design, so a single trial’s null does not survive as the pooled estimate, and its own author reading is concordance: «although modest by conventional standards, our effect sizes are similar to those reported in landmark trials» (Mendes et al., 2025). The magnitude agreement is drawn from the same RCT field, not an independent route. Recorded as a REFINEMENT (F/C), not a tension.
What bounds the ranking (NMA-specific appraisal) [inferred from @mendes2025multidomain]. Coherence is adequate — «The global inconsistency test (Q=39⋅37, df=29, p=0⋅093) did not show statistically significant inconsistency» — with one local exception, «social activity and health education (p=0⋅029)» (the smallest node) (Mendes et al., 2025). But: 40% of trials are high risk of bias («Risk of bias was high in 44 (40%) studies»); the primary vs-health-education panel carries publication bias («Egger’s test statistically significant (p=0⋅011)… smaller trials appeared more likely to report larger effect sizes») that the vs-no-intervention and vs-active-control panels do not; a sex-distribution imbalance threatens transitivity; and it is all a surrogate in short trials (median 13 weeks) with no follow-up (Mendes et al., 2025). So this refines the shape of the surrogate signal (which lever, comparator-dependent magnitude) and does not reach the hard endpoint — Coley’s pooled incidence null still governs the dementia-prevention question.
A route-b age hint, held not banked. Post-hoc, «more complex multidomain approaches may be particularly beneficial in older populations» (age >=70: E+C+H 0·56 [–0·05 to 1·16], NS; E+C 0·29 [0·03–0·55]) while under-70s did best on standalone cognitive training (0·42) / E+C (0·36) (Mendes et al., 2025). Held as an effect-modification-by-age hypothesis (post-hoc, top cell NS) — not a finding -> Baseline Risk and the Relative-Absolute Split.
What the three trials + the pooled analysis jointly change about what to do [2026-08-07]
(Andrieu et al., 2017; Coley et al., 2025; inferred from Ngandu et al., 2015; van Charante et al., 2016)
- The bundle is worth pulling for its other proven outcomes, not sold as demonstrated dementia prevention. Every lever (BP, glucose, weight, activity, diet, smoking) is a cardiometabolic big rock already ranked high for CV outcomes -> Big Rocks (Elderly), Layer 1 - Ranking Interventions for a Stratum; acting on them is well-warranted regardless of the dementia signal. What is not warranted is promising a person a cognitive or dementia payoff: one surrogate-positive trial (FINGER), one surrogate-null trial on a closer design (MAPT), and one hard-incidence null (preDIVA) — the honest state is that the population-level effect is small-to-absent and unreplicated, though the hard endpoint remains not-yet-shown rather than shown-absent (preDIVA underpowered for the plausible 2-9% effect).
- Selection (baseline risk) is the responder hope on the SURROGATE — and the pooled hard-endpoint search KILLS it. The two cognitive-composite trials converge on one reading: the surrogate effect concentrates in higher-baseline-risk strata (FINGER selected CAIDE>=6 and was positive; MAPT was null overall but its CAIDE>=6 and amyloid-positive subgroups reproduced the benefit). But that route-(b) reading does not survive the hard endpoint: Coley’s pooled MAPT+preDIVA IPD found no dementia-incidence responder — not in CAIDE>=6, not in untreated-HTN, not in any pre-specified subgroup, and not by a data-driven SIDES search (see The pooled responder verdict above). So the honest state is surrogate-level responder signal, hard-endpoint responder null — and per the telos (route-b needs positive effect-modification evidence) the hard-endpoint null governs the decision: targeting by baseline dementia risk is not a demonstrated way to convert this bundle into a dementia-prevention benefit. What remains open is intensity, an earlier (midlife) window, and biomarker selection — untested here -> Dementia Prevention and Modifiable Risk Factors, Baseline Risk and the Relative-Absolute Split.
- If pulling levers for the surrogate, lead with exercise + cognitive training, not a maximal bundle. The 109-trial NMA ranks E+C highest and shows the full diet+exercise+cognitive-training+health-education bundle does not improve on it (and is NS vs an active control) — more is not better, because adherence falls and mechanisms overlap. Cognitive training and physical exercise are the two levers robust even against an active control (Mendes et al., 2025). This ranks within the surrogate; it does not convert to a demonstrated dementia-incidence benefit (Coley null still governs), and the exercise levers are cardiometabolic big rocks already pulled -> Big Rocks (Elderly).
- Omega-3 supplementation is not a cognitive lever on this evidence. MAPT (the longest/largest omega-3 cognition RCT) was flat alone and additive-free in the bundle; low omega-3 status marks decline but supplementing it did not prevent decline -> Vitamin and Mineral Supplements for Disease Prevention.
- BP-lowering in the elderly is cognitively safe. preDIVA is a large, long RCT showing intensive vascular care caused no cognitive or disability harm -> Blood Pressure Lowering and Cardiovascular Events.
Open (G-gaps)
- RCT effect on dementia/AD incidence of the FINGER 4-component bundle — preDIVA answers the vascular-care-only version (null); FINGER’s own 7-yr extended follow-up (the 4-component bundle on incidence) is still not held -> — long-term dementia/AD incidence results for the full FINGER bundle`.
- The pooled responder analysis (the route-b test) — CASHED 2026-08-07 (Coley 2025). Does the baseline-risk effect-modification seen in MAPT’s subgroups and preDIVA’s untreated-HTN cell hold on a hard endpoint when the two trials are pooled at individual-participant level? No — the pooled IPD overall null (HR 0.98) held in every pre-specified subgroup and no SIDES-discovered subgroup showed a differential effect (see The pooled responder verdict — Coley 2025 above). The route-b gap is now resolved toward the null for these late-life low-intensity trials; what stays open is a higher-intensity / midlife / biomarker-selected design, not the responder axis in these data.
- Per-component contribution — which lever(s) drive FINGER’s bundle effect (author-flagged); MAPT’s null and preDIVA’s vascular-only null cannot isolate it either. Partially addressed by Mendes 2025 across the wider RCT literature (physical exercise and cognitive training are the two robust single levers; their combination E+C tops the ranking; diet-alone and social-activity-alone are NS) — but on the SURROGATE, so it decomposes the cognitive-composite signal, not FINGER’s own bundle nor any incidence effect.
(Andrieu et al., 2017; Coley et al., 2025; inferred from Ngandu et al., 2015; van Charante et al., 2016) The loop stays open (R1) even with three RCTs and their pooled IPD in hand: the surrogate moved in one selected-risk population (FINGER) and not in a broader, older one (MAPT); the hard endpoint did not move on a thinner vascular-only intervention (preDIVA); and when MAPT+preDIVA are pooled at participant level the hard endpoint stays null with no responder subgroup (Coley) — so no two of them close each other, and the surrogate-to-incidence link is itself unproven -> Surrogate Outcomes. The one thing the pool does settle is the responder question for these trials (route-b null). A validated multidomain dementia-prevention claim now awaits an incidence result for the full FINGER 4-component bundle () and, beyond it, a higher-intensity or earlier-window trial.
Self-critique [run 2026-09-05, before commit — Mendes NMA weave]
- Not laundered independence. The Mendes NMA is the highest-risk laundered-E candidate on this page —
an NMA numerically concordant with the FINGER/MAPT trials the page already holds. It is filed explicitly
as type-F/C and never
[E-independent]: field-overlap defeats independence, and the pooling claim was demoted to the verifiable form (whether FINGER/MAPT are literally among the 109 is not confirmable from the held text — the included-study roster is in the appendix). Cold-audit DEMOTE resolved in-commit. - Not a fake tension. MAPT’s whole-population null vs Mendes’s significant pooled bundle is filed as a
refinement, not a
tension: the parameter table’s population/comparator rows are NO/mixed (not-joined ii), and a single trial’s null is not expected to survive pooling — no joined issue. - Overclaim watch — surrogate is fenced. The E+C ranking is stated as a within-surrogate hierarchy in short (median 13-wk), 40%-high-RoB trials with health-education-comparator publication bias, and is repeatedly said NOT to convert to a demonstrated dementia-incidence benefit (Coley’s hard-endpoint null still governs). The age-subgroup signal is held as a post-hoc hypothesis, not a finding.
- R1 — loop open. Nothing here is graded against a realized patient-important outcome; the surrogate-to-incidence transmission is the standing G-gap, named not closed.