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

Anaemia — Structural Causal Analysis (the haemoglobin surrogate trap)

Anaemia affects roughly a third of the world’s population (~2 billion people) and is the single clearest example in this atlas of a surrogate trap: correcting the number (haemoglobin) is not the same as correcting mortality. This oracle estimates the causal reduction in mortality across iron, erythropoiesis-stimulating, HIF-PH-inhibitor and transfusion strategies, framed through the CKD, heart-failure and obstetric pathways where anaemia actually drives death. Hazard ratios are from named trials. For education, not individual medical advice.

Method. Structural Causal Model (SCM) with backdoor adjustment (Pearl). Interventions are not assumed independent: haemoglobin-raising options share the mediator, so their overlap is removed by an eigenvalue-corrected equicorrelation model at an adjustable mean cross-correlation ρ̄ (default 0.30). Iron’s large haemoglobin-independent tissue benefit is separated as a direct path. Robustness to unmeasured confounding is quantified with the E-value. PN / PS / PNS under monotonicity. Every hazard ratio is cited — no effect size is invented. The front door is resolved through an EXPLICIT mediator cascade (underlying cause → iron substrate → erythropoietic → haemoglobin → disease state), not one lumped node: each intervention acts on a specific node, so same-node interventions are substitutes that saturate against each other, while different-node interventions are d-separated given the intermediate node and compose in series. The cross-correlation removal thus follows from the graph structure; the residual ρ̄ cleans up only the mediator-independent (direct) effects.
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 effect estimate, 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 Hb 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 risk (illustrative anchor)
Absolute risk after intervention
Absolute risk difference (RD)
Backdoor-only HR (no front-door)
Haemoglobin overlap removed
Number needed to treat (NNT)
ρ-sensitivity band (ρ 0 → 0.6)
Interpretation

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). Haemoglobin-raising therapies act through the shared mediator haemoglobin / red-cell mass, which drives tissue oxygen delivery and thence mortality (Y). Iron additionally acts independently of haemoglobin (direct myocardial / skeletal-muscle tissue-iron effects). Named confounders — underlying illness (reverse causation), inflammation, CKD stage — open back-door paths (adjusted). The mediator is non-monotonic: over-correcting it with an ESA adds thrombotic harm. Mediator cascade: interventions attach to the node they act on (underlying cause → iron substrate → erythropoietic → haemoglobin), which converge on the disease state and thence the endpoint — drawing the intermediate mediators explicitly is what exposes d-separation and per-channel saturation.

Confounders U:underlying illness (reverse causation: disease → anaemia) · inflammation / hepcidin · CKD stage · bleeding source · comorbidity · nutrition → back-door paths (adjusted)Oral iron(ferrous sulfate)IV iron — CKD/ dialysisIV iron — heartfailureIV iron — obstetric(peripartum)ESA — conservativeHb targetESA — normalisingHb targetHIF-PH inhibitor(oral)RBC transfusion(restrictive)Treat underlyingcauseUnderlying cause(bleeding / disease)Iron substrateavailabilityErythropoieticdriveHaemoglobin(direct)Anaemia(low Hb / O2)Outcome /mortalityFront-door: through haemoglobin correctionHaemoglobin-independent (tissue-iron) direct pathBack-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: the haemoglobin-raising agents (ESA, HIF-PH inhibitor, transfusion) share one final common pathway — raising the number — so stacking them is largely redundant and, past target, harmful. ρ̄ is user-adjustable because iron’s tissue effect and “treat-the-cause” share little mechanism with pharmacological Hb-raising. Most of the overlap is now handled structurally by the mediator nodes (same-node substitutes saturate); ρ̄ governs only the residual correlation among direct effects.

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-effect 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

Haemoglobin is a confounded, non-monotonic mediator. Anaemia is often a marker of underlying illness, not its cause (reverse causation: disease → anaemia → death). So the observational anaemia–mortality association (preoperative anaemia carries ~3-fold higher odds of death) overstates what raising the number can achieve, and the mediator is not protective across its whole range.
Chasing the number can kill. Normalising haemoglobin with an erythropoiesis-stimulating agent did NOT reduce cardiovascular events and INCREASED stroke (TREAT: HR 1.92), with harm signals in CHOIR and CREATE. This is the central anti-target: an intervention that moves the mediator strongly yet worsens the hard endpoint. HIF-PH inhibitors are, at best, CV-neutral oral equivalents — not mortality-lowering.
Iron’s benefit is largely haemoglobin-INDEPENDENT. The heart-failure and dialysis benefits of IV iron act substantially through tissue-iron effects on myocardial and skeletal-muscle energetics, not through raising haemoglobin — which is why they carry a gold, mediator-independent arrow and why IV iron helps HF patients who are barely anaemic. Modelling all benefit through Hb mis-specifies the graph.
Morbidity is not mortality. The strongest iron trials reduced hospitalisations and composite events but were neutral or non-significant on cardiovascular death alone (AFFIRM-AHF primary p=0.059; HEART-FID CV death HR 0.86, CI crossing 1). The composite HRs are upper-bound ceilings for the mortality component.
Transfusion is a stewardship choice, not a survival lever. Restrictive and liberal thresholds are mortality-equivalent in most settings; restrictive is preferred because it uses less blood and avoids transfusion harms, not because it saves more lives. Raising haemoglobin acutely with blood does not improve survival outside active haemorrhage.
Population determines which arm works. These effect sizes are not portable: proactive IV iron’s benefit is a dialysis / iron-deficient-HF finding, the obstetric estimate is a transfusion/anaemia surrogate, and “treat the cause” depends entirely on the cause. Selecting an arm outside its evidenced population is unsupported.

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(outcome | do(∅)) — baseline
P(outcome | 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 effect 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 effect is swung across its 95% confidence interval (others held at point estimate); the bar is the resulting swing in the combined front-door effect. A long bar means the combined estimate leans heavily on that single trial's precision.

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

Front-door (mediation) decomposition. Haemoglobin-raising therapies act through one shared mediator. Each log-effect is split into a haemoglobin-mediated (indirect) and a haemoglobin-independent (direct) part. Indirect parts are pooled through the mediator with dose-response saturation, removing the mediator cross-correlation; direct parts keep the residual eigenvalue correlation at ρ̄. Iron’s haemoglobin-independent tissue benefit is NOT discounted against the ESA / HIF-PHI / transfusion options — and, crucially, a high mediated fraction with a null hazard ratio (the ESA arms) exposes a mediator that moves without protecting. Here mediated effects are pooled WITHIN each cascade node (dose-response saturation of substitutes) and composed in SERIES across nodes (d-separated channels), with the per-node reductions reported so the channel structure is visible.

Mediator saturation cap = 55% Hb-gap corrected
Sum of standalone Hb reduction (naive)
Combined Hb 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%Hb↑med-fracindirect logdirect log

Which % of cross-correlation is appropriate? Not one number. The mediator overlap is fixed empirically by the Hb saturation (currently removing of the summed mediated effect when interventions are stacked). Note a domain caveat: the ESA and HIF-PHI arms have a high haemoglobin-mediated fraction yet a null-to-harmful hazard ratio — the model therefore shows a mediator that is strongly moved but not protective. That is the intended lesson, not a bug.

Front-door caveat (antithesis): textbook front-door identification requires an unconfounded mediator→outcome edge; here it is confounded by reverse causation (illness→anaemia→death), and the mediator is non-monotonic (over-correction with an ESA increased stroke, TREAT HR 1.92). Iron trials were neutral on cardiovascular death alone, and transfusion strategy is mortality-neutral.

Executive summary

Select interventions to generate a plain-language summary.