Bayesian Causal Atlas · Vol. Vision · Pearl Structural Causal Model

Vision-Threatening Eye Disease — Structural Causal Analysis

Vision-threatening eye disease — glaucoma, age-related macular degeneration (AMD), and diabetic retinopathy — affects tens of millions of United States adults. Vision loss is not merely disability: it independently raises the risk of falls / fractures, depression, and mortality. This oracle estimates the causal reduction in vision loss — the shared mediator of those downstream harms — achievable by disease-specific therapies (intraocular-pressure lowering for glaucoma; vascular-endothelial-growth-factor [VEGF] suppression for AMD / diabetic retinopathy). Downstream mortality links are largely observational and flagged.

Method. Structural Causal Model (SCM) with backdoor adjustment (Pearl). Interventions are not assumed independent: their heavy mechanistic overlap (nearly all act by preventing vision loss (via disease-specific mediators)) is removed by an eigenvalue-corrected equicorrelation model at an adjustable mean cross-correlation ρ̄ (default 0.30). Robustness to unmeasured confounding is quantified per intervention with the E-value. Probabilities of Necessity / Sufficiency (PN / PS / PNS) are reported under a monotonicity assumption. Every hazard ratio is cited to its source trial — no effect size is invented.
ρ̄ = 0.30
A–D (all)
A high (RCT/meta) · B cohort · C case-series/modelled · D consensus/provisional. Lower-grade interventions are excluded from the DAG, front-door pooling, Pareto, Monte‑Carlo & sensitivity.

Interventions

Tick the interventions to combine. Each shows its trial hazard ratio, 95% confidence interval (CI), E-value, mechanism, and citation. ★ = in the current Pareto effective set but not yet ticked.

Combined causal estimate

Headline is the front-door estimate: shared vision-loss-mediator overlap removed via dose-response saturation; residual direct-effect overlap removed via the eigenvalue model at ρ̄.

1.00
Combined HR
0%
Relative risk ↓
Pooled E-value
Interventions selected (k)
0
Effective independent dimensions (n_eff)
0
Redundancy discount applied
0%
Baseline 10-yr risk (illustrative anchor)
Absolute risk after intervention
Absolute risk difference (RD)
Backdoor-only HR (no front-door)
Mediator (weight) overlap removed
Number needed to treat (NNT)

Causal attribution

Under monotonicity + exogeneity (E-value bounds the exogeneity assumption).

Probability of Necessity (PN)
Probability of Sufficiency (PS)
Prob. of Necessity & Sufficiency (PNS, lower bound)
Causal DAG
Cross-correlation
Pareto (threshold)
Monte Carlo
Front-door mediation
What-if / If-not-for
Sensitivity
Antithesis

Faithful causal directed acyclic graph (DAG). Disease-specific therapies act through different proximal mediators — intraocular pressure (glaucoma) and VEGF / macular leakage (AMD, diabetic retinopathy) — which converge on the shared mediator vision loss. Vision loss then drives depression, falls / fractures and mortality. Low-vision rehabilitation is vision-independent: it reduces falls without restoring acuity. Named confounders open back-door paths (adjusted). Illustrative of structure, not yet the identification engine.

Confounders U:age · diabetes · disease severity · access to care · cognitive status · fellow-eye status → back-door paths (adjusted)Anti-VEGF injectionPanretinal photocoagulationProstaglandin analogueLaser / SLTCataract surgeryLow-vision rehabilitationIntraocularpressure ↓VEGF / macularleakage ↓Vision loss(acuity / field)DepressionFalls / fracturesMortality (Y)Front-door: through vision preservationVision-independent: falls preventionBack-door confounding (adjusted)

Eigenvalue diagnostics for the selected interventions under an equicorrelation matrix (off-diagonal ρ̄). A large λmax relative to k signals redundancy; n_eff is the effective number of independent interventions actually contributing.

k (selected)
0
λmax
λmin
n_eff = (Σλ)² / Σλ²
Condition number

On mechanistic grounds ρ̄ ≈ 0.30 is defensible for this domain: incretin drugs, surgery, diet, and activity overwhelmingly share the weight-loss / insulin-sensitivity pathway, so ~30% of their nominal effects overlap. A mediation analysis of semaglutide found ~80% of its MACE benefit is not mediated by weight loss, which is why ρ̄ is user-adjustable rather than fixed.

Minimum-effective-set analysis. Set a target combined risk reduction; the model finds the smallest set of interventions — accounting for front-door mediator overlap — that reaches it, and highlights them. If the target exceeds what all interventions together can achieve, the full set is shown (never an empty one). "Apply" ticks exactly that set.

Target combined risk ↓ ≥ 50%

Monte Carlo propagation. Each selected intervention's log-hazard-ratio is sampled from a normal distribution implied by its 95% CI; samples are combined with the same eigenvalue overlap discount. 5,000 draws.

Median combined HR
95% simulation interval
Standard deviation of combined HR
P(combined HR < 0.90)

Antithesis — challenging this oracle's own conclusions

Different diseases, one shared mediator. An intraocular-pressure drug (glaucoma) and an anti-VEGF injection (AMD / diabetic retinopathy) only genuinely combine in a patient who has both conditions. The pooled combined estimate assumes co-occurrence; for a single-disease patient, only the relevant subset applies.
Vision-to-mortality is largely observational. The cataract-surgery mortality benefit (hazard ratio ~0.6) comes from cohorts confounded by healthy-patient selection — sicker patients are less likely to undergo elective surgery. Vision loss may be a marker of frailty rather than a cause of death.
Glaucoma progression is slow. Most treated open-angle glaucoma never progresses to blindness; intraocular-pressure lowering roughly halves an already-low progression rate, so absolute benefit is modest and the number-needed-to-treat to prevent blindness is large.
Injections carry their own risk. Intravitreal anti-VEGF has a small but real endophthalmitis rate (~0.05% per injection, cumulative over years of therapy) and debated systemic thromboembolic signals in diabetics — harms not captured by the vision hazard ratio.
Rehabilitation does not restore sight. Low-vision rehabilitation reduces falls through home modification and mobility training, entirely via the vision-independent pathway. It must NOT be double-counted with sight-restoring therapies — which is exactly why the front-door layer does not discount it against them.

What-if — the do-operator: P(Y | do(S))

Intervening on the selected set S with Pearl's do-operator (setting the interventions, not merely observing them). Contrast against do(∅) = no intervention.

P(mortality | do(∅)) — baseline
P(mortality | do(S)) — intervened
Absolute risk reduction (ARR)
Number needed to treat (NNT)

If-not-for — but-for counterfactual (leave-one-out)

For each intervention: "if not for this one, the combined front-door hazard ratio would be…". Isolates each intervention's marginal causal contribution after mediator-overlap removal, so shared-pathway agents are not double-credited.

If not for…HR without itHR with full setmarginal RRR lost

One-at-a-time sensitivity. Each intervention's hazard ratio is swung across its 95% confidence interval (others held at point estimate); the bar is the resulting swing in the combined front-door HR. A long bar means the combined estimate leans heavily on that single trial's precision.

Optimistic bound (all HRs at CI-low)
Point estimate
Pessimistic bound (all HRs at CI-high)
Pooled E-value (confounding robustness)

Front-door (mediation) decomposition. Most therapies act through one shared mediator — prevention of vision loss (reached via disease-specific proximal mediators: intraocular pressure or VEGF leakage). Each log-effect is split into a vision-mediated (indirect) and a vision-independent (direct) part. Indirect parts are pooled through the mediator with dose-response saturation (vision can only be preserved up to its baseline — you cannot restore the same lost letters twice), removing the mediator cross-correlation; direct parts keep the residual eigenvalue correlation at ρ̄. Low-vision rehabilitation, which reduces falls WITHOUT improving acuity, is therefore NOT discounted for overlap with the sight-preserving therapies.

Mediator saturation cap = 70% vision loss ↓
Sum of standalone vision-loss reduction (naive)
Combined vision-loss reduction after saturation
Mediator overlap removed (1 - saturation)
Direct-effect redundancy removed (1 - n_eff/k)
Front-door combined HR
Backdoor-only combined HR (comparison)
InterventionHR%vision↓vis-medindirect logHRdirect logHR

Which % of cross-correlation is appropriate? Not one number. The mediator overlap is fixed empirically by the vision-preservation saturation (currently removing of the summed vision-mediated effect when sight-preserving therapies are combined). Note an important caveat unique to this domain: an intraocular-pressure drug and an anti-VEGF injection only combine if the same patient has BOTH glaucoma and a retinal disease — otherwise the pooled estimate is hypothetical. Vision-mediated fractions are transparent, adjustable priors.

Front-door caveat (antithesis): the cited endpoints are vision outcomes (acuity / field) — well proven. Their link to mortality is largely observational: the cataract-surgery mortality benefit (hazard ratio ~0.6) is confounded by healthy-patient selection, and vision→mortality associations may reflect frailty rather than causation. Falls→fracture→mortality is better supported. Downstream mortality estimates are upper bounds.

Executive summary

Select interventions to generate a plain-language summary.