A Judea Pearl Structural Causal Model across two strata — a robust hearing-disability axis (where treatment plainly helps) and a contested dementia axis (where hearing loss is the largest single modifiable risk factor, yet the randomised evidence that treating it prevents dementia is positive only in higher-risk people). All hazard ratios are real published values; the dementia benefit is confined to the stratum where a trial actually showed it and is never allowed to inflate beyond the evidence.
Causal question. For hearing loss, what is the identifiable causal contribution of hearing treatment to (a) hearing function and (b) incident dementia / cognitive decline, after backdoor confounding control and after restricting each effect to the population in which it was actually demonstrated?
| Stratum | Population | Primary endpoint | Anchor (untreated) | Manipulable node |
|---|---|---|---|---|
| Disability axis (robust) | Adults with hearing loss | Hearing function / communication / quality of life | Untreated hearing loss | Hearing aids; cochlear implant (severe); cerumen removal |
| Dementia axis (contested, risk-stratified) | Older adults with hearing loss, by cognitive-decline risk | Incident dementia / 3-yr cognitive decline | Elevated dementia risk by severity (Lin 2011) | Hearing aids — effect confined to the high-risk stratum (ACHIEVE) |
Secondary, well-established benefits of hearing treatment tracked qualitatively: reduced social isolation, depression, and fall risk. These are not part of the contested dementia claim.
Hearing loss sits upstream of dementia through three proposed pathways — reduced cognitive stimulation, increased listening effort / cognitive load, and accelerated structural brain change. The decisive structural problem is the backdoor: age, vascular risk, and education cause both hearing loss and dementia, and prodromal dementia can itself reduce hearing-aid uptake (reverse causation). Randomisation (ACHIEVE) is the only design that closes these paths — which is why its overall-null / high-risk-positive split is treated as the primary evidence, not the observational associations.
| Operation | Method | Implementation here |
|---|---|---|
| Confounding control | Backdoor criterion (Pearl) | Severity HRs adjusted for age/sex/education/vascular risk in source papers; RCT (ACHIEVE) closes the backdoor by design |
| Effect moderation | do(·) effect varies by stratum | Hearing-aid dementia effect applied only in the high-risk stratum (ACHIEVE ARIC); null in the low-risk stratum |
| Causal vs confounded | Causal-fraction discount | Observational HR 0.81 discounted by a user-set causal fraction f: applied HR = 1 − f·(1−0.81) |
| Population impact | Attributable fraction (Levin) | PAF 8% (self-report, Lancet) to 13–17% (objective, ARIC/NHATS) |
| Robustness | E-value (VanderWeele & Ding) | Computed for the hearing-aid dementia HR |
| Minimum-effective set | Pareto frontier (threshold slider) | Removable excess dementia hazard ranked by baseline severity (Section IV) |
Risk rises log-linearly with severity. Lin 2011 (Baltimore Longitudinal Study of Aging, 639 people, 58 incident dementia cases) is the anchor; the Lancet Commission pooled estimate makes hearing loss the largest single modifiable dementia risk factor in high-income countries.
| Exposure | Dementia HR (95% CI) | Source / grade |
|---|---|---|
| Per 10 dB worse hearing | 1.27 (1.06–1.50) | Lin 2011, Arch Neurol B |
| Mild loss (25–40 dB) vs normal | 1.89 (1.00–3.58) | Lin 2011 B |
| Moderate loss (41–70 dB) vs normal | 3.00 (1.43–6.30) | Lin 2011 B |
| Severe loss (>70 dB) vs normal | 4.94 (1.09–22.40) | Lin 2011 (n=6 severe; very wide CI) C |
| Pooled RR (Lancet Commission) | 1.94 (1.38–2.73) | Livingston 2020, Lancet B |
| Population attributable fraction | ~8% (self-report); 13–17% (objective) | Lancet 2020; ARIC 13.2%, NHATS 16.9% B |
| Evidence | Effect | Source / grade |
|---|---|---|
| ACHIEVE — overall population (RCT, n=977, 3 yr) | NULL | Lin 2023, Lancet A |
| ACHIEVE — high-risk ARIC subgroup (n=238) | −48% cognitive decline | Lin 2023, Lancet (prespecified) A |
| ACHIEVE — top quartile of predicted risk | −58.1% (31.4–90.9%) | ACHIEVE risk-moderation analysis B |
| Hearing-aid use → cognitive-decline incidence | HR ~0.81 (−19%) | Yeo 2023 meta-analysis (observational) C |
| UK Biobank hearing-aid / dementia analysis | do not cite | Lancet Public Health 2023 — RETRACTED R |
The honest reading: a hearing intervention slowed decline only where background decline was fast (high-risk older adults); it did nothing in lower-risk volunteers. Subjective hearing loss in ARIC was not associated with dementia (HR < 1.0), and Marinelli 2022 found no per-10 dB association (HR 0.99) — so the association is real but measurement-dependent and confounding-prone.
Where does treating hearing loss buy the most dementia-axis return? Bars show the removable excess dementia hazard by baseline severity — (HRseverity − 1) × (1 − HRaid) at the current causal fraction. Drag the threshold to set a minimum removable hazard: tiers at or above the line are the priority set; the dimmed tail below yields less per person treated.
ACHIEVE's signal appeared only where background cognitive decline was rapid (high-risk older adults declined ~3× faster). A treatment that slows decline can only be detected against a moving baseline — which means the population most likely to benefit is precisely the one already on a steep trajectory.
Objective audiometry yields PAFs 6–8× larger than self-report, and subjective hearing loss in ARIC was null. Whether hearing loss is a "leading" modifiable dementia factor depends on how it is measured — a caution against headline attributable fractions.
Unlike most dementia risk factors, hearing loss has a cheap, scalable do(·): amplification. If even part of the association is causal, population hearing-care programs act on a node no drug currently addresses.
Even if the dementia link were fully confounded, the disability axis stands on its own: hearing aids reliably restore communication and reduce social isolation, depression, and falls — a healthspan return that does not depend on the contested cognitive claim.
| Challenge | Why it threatens the conclusion | Mitigation here |
|---|---|---|
| ACHIEVE was null overall | The primary, prespecified comparison in the full trial showed no benefit; the 48% is a subgroup result | Dementia effect applied only in the high-risk stratum; the overall-null result is shown beside it |
| Confounded observational data | Hearing-aid users are healthier, wealthier, more engaged; reverse causation (prodromal dementia → less uptake) | Causal-fraction slider discounts HR 0.81; E-value exposes how much unmeasured confounding would explain it away |
| A retracted key study | A major UK Biobank hearing-aid / dementia paper was retracted, weakening the observational base | That study is flagged "do not cite"; the engine never uses it |
| Measurement dependence | Subjective hearing loss was null in ARIC; Marinelli 2022 found no per-10 dB association | Severity HRs labelled by source and grade; PAF given as a self-report-vs-objective range, not a point |
| Wide severe-loss CI | The severe-loss HR 4.94 rests on only 6 people (CI 1.09–22.40) | Graded C; the Pareto's severe tier is the most uncertain bar and labelled as such |
severity HR vs normal: mild 1.89 · moderate 3.00 · severe 4.94 (Lin 2011)
aid HR (dementia) = high-risk stratum: 1 − f·(1 − 0.81) ; low-risk stratum: 1.00 (ACHIEVE community null)
RR_treated = HR_severity × HR_aid
excess removed (%) = (HR_severity − RR_treated) / (HR_severity − 1)
removable excess hazard (Pareto) = (HR_severity − 1) × (1 − HR_aid)
E-value (HR<1): RR* = 1/HR ; E = RR* + √(RR*·(RR*−1))
Worked example: moderate loss (HR 3.00), high-risk stratum, f = 0.5 → aid HR = 1 − 0.5·0.19 = 0.905 → RR_treated = 2.72; excess removed = (3.00 − 2.72)/2.00 = 14%; removable hazard = 2.00 × 0.095 = 0.19. At f = 1 (full causal): aid HR 0.81, RR_treated 2.43, excess removed 28%.
1. Lin FR, Metter EJ, O'Brien RJ, Resnick SM, Zonderman AB, Ferrucci L. Hearing Loss and Incident Dementia. Arch Neurol 2011;68(2):214–220. Per 10 dB HR 1.27 (1.06–1.50); mild 1.89, moderate 3.00, severe 4.94.
2. Livingston G, Huntley J, Sommerlad A, et al. Dementia prevention, intervention, and care: 2020 report of the Lancet Commission. Lancet 2020;396:413–446. Hearing loss pooled RR 1.94 (1.38–2.73); PAF ~8%.
3. Lin FR, Pike JR, Albert MS, et al. (ACHIEVE Collaborative). Hearing intervention versus health education control to reduce cognitive decline (ACHIEVE): a multicentre, randomised controlled trial. Lancet 2023;402:786–797. Overall null; high-risk ARIC subgroup −48% 3-yr cognitive decline.
4. ACHIEVE risk-moderation analysis. Effect of hearing intervention on 3-year cognitive change greatest among older adults at high risk: top risk quartile −58.1% (31.4–90.9%).
5. Yeo BSY, Song HJJMD, Toh EMS, et al. Association of Hearing Aids and Cochlear Implants With Cognitive Decline and Dementia: A Systematic Review and Meta-analysis. JAMA Neurol 2023. Hearing-aid use ~19% reduction in long-term cognitive-decline incidence (HR ~0.81), observational.
6. PAF analyses: Population Attributable Fraction of Incident Dementia Associated With Hearing Loss (ARIC-NCS): objective moderate-or-greater hearing loss PAF 13.2% (ARIC), 16.9% (NHATS); subjective hearing loss not associated (HR < 1.0).
7. Marinelli JP, et al. (Mayo Clinic Study of Aging) reported HR 0.99 (0.89–1.07) per 10 dB — a null association, included for balance. The UK Biobank hearing-aid/dementia analysis (Lancet Public Health 2023) was RETRACTED and is not used.
Longevity Research Initiative — Bayesian Causal Atlas. SCM methodology after Judea Pearl; E-values after VanderWeele & Ding. All effect sizes are published values cited above; none were fabricated. Outputs are upper-bound ceilings for research, not clinical guidance. Report generated .