Bayesian Causal Network — US Mortality Endpoints

Pearl Structural Causal Model using cited, confounder-adjusted effect sizes (backdoor adjustment is the source study’s), do-calculus, E-value sensitivity, PN/PS bounds, and PAF, applied to 18 leading causes of death in the United States (10 infectious, 8 chronic) — covering ~74% of all-cause US mortality. Interactive: toggle interventions, adjust dose intensity, set Pareto risk thresholds, watch combined Pearl-adjusted hazard ratios update in real time.

18 endpoints 28 interventions 20 confounders Backdoor adjustment Do-calculus E-value (VanderWeele 2017) PN/PS bounds (Tian & Pearl 2000) Pareto risk threshold Built 2026-05-26

Total Deaths (Annual, In-Scope)

—

multi-cause; ~74% of US all-cause

Selected Endpoint

Heart Disease

I20–I25, I50

Pearl-Adjusted HR (live)

—

naive (independent): —

Lives Saved (Selected Endpoint)

—

PAF: —

Total Lives Saved (All Endpoints, Live)

—

modelled ceiling (best case); antithesis applies

Dashboard

Causal DAG

Sensitivity Heatmap

Pareto Frontier

Tornado

Evidence Base

Antithesis

Methods & Assumptions

Endpoint18

All

Infectious

Chronic

Quick Actions

Live Pearl-Adjusted Results

Evidence tier → causal graph & combined HR: A–D

A=RCT (vaccines, DAA/ART/PrEP, statin/SGLT2/GLP-1, screening, anti-amyloid) · B=policy/infection-control/environmental observational. Lower grades excluded.

ρ̄ = —

Heart Disease

HR (naive)—

HR (Pearl-adj)—

E-value—

PN (necessity)—

PS (sufficiency)—

PAF—

Lives saved (this endpoint)—

Baseline deaths/yr—

HR Combination Rule: log(HR_combined) = Σᵢ log(HRᵢ) × δ(ρ̄, k) where δ = (1−ρ̄) + ρ̄/k and ρ̄ is the eigenvalue-regularized mean off-diagonal correlation among selected interventions. As more interventions in the same domain are added, the cross-correlation discount grows, preventing additive over-estimation. Per-intervention dose adjusts contribution toward the saturation point reported in the evidence column.

Interventions0 / 28

Pearl Structural Causal Model — Force-Directed DAGdrag to rearrange · scroll to zoom

Intervention Endpoint Confounder / Risk Factor

Pearl-Adjusted HR — Sensitivity Heatmap28 intvs × 18 endpoints

Each cell = single-intervention disease-specific hazard ratio. Darker green = stronger protection. Dot (·) = no effect (HR ≥ 1).

Pareto Frontier — Minimum Effective Intervention Setrisk threshold slider

Greedy frontier construction: at each step, add the intervention that most reduces Pearl-adjusted HR. Selected endpoint: —.

Target HR ≤ 0.30 ⟶ More aggressive ←→ More permissive

Recommended Minimum-Effective Bundle

—

Set size: — Achieved HR: — Lives saved: — vs. all 28 intvs: —

k	HR	Lives saved	Set composition

Tornado — Single-Intervention Effect Rankingselected endpoint

Ranks every intervention by its standalone hazard ratio against the selected endpoint. Stronger green = larger protective effect.

Intervention Evidence Base28 interventions

Intervention	Domain	HR (95% CI)	Saturation	E-value	Targets	Evidence

Endpoint Burden & ICD CodingCDC NCHS 2023

Endpoint	ICD-10	Category	Annual deaths	Underlying	Source

Anti-Sycophancy: Where This Model Could Be Wrong

Per Commandment #7 (Antithesis) and #14 (Anti-Sycophancy), the combined HRs reported here represent an upper bound on plausible benefit. Real-world public-health gain in a typical US population would likely be 30–60% smaller. The principal threats to validity:

I. Effect-Size Optimism (RCT-to-Population Gap)

Healthy-vaccinee bias. Hak (Int J Epi 2002) showed influenza vaccine had no mortality benefit after adjustment for healthy-user proxies. Our HR 0.61 may be confounded upward by 0.05–0.15.
SPRINT generalizability. The 25% MACE reduction from SBP <120 (SPRINT) excluded T2DM and prior stroke patients — populations where ACCORD-BP found null benefit.
Anti-amyloid mAbs (lecanemab/donanemab). Clarity-AD's CDR-SB delta of 0.45 was statistically significant but clinically marginal; ARIA-E/H incidence 13–27% offsets benefit in elderly. HR 0.93 may overstate net mortality benefit.
Statin J-curve in primary prevention. Cholesterol Treatment Trialists' Collaboration HR 0.78 reflects pooled trials weighted to secondary prevention; primary-prevention HR closer to 0.85 (HOPE-3).
Cancer screening 0.80 HR is composite. NLST LDCT for lung HR 0.80 (mortality), mammography meta HR 0.81 (Marmot 2013), colonoscopy HR 0.74 (NordICC at 10 yr null in ITT, 0.70 in per-protocol). Substantial heterogeneity hidden in this single number.

II. Sepsis Double-Counting

Sepsis is a syndrome, not a disease. The 270,000 multi-cause sepsis deaths overlap heavily with pneumonia (~30–40%), CDI (~5%), MRSA (~10%), HCV liver decompensation, and HIV opportunistic infections. Summing prevented sepsis + prevented pneumonia/MRSA double-counts. True net infectious-disease reduction is closer to 60% of the displayed sum.
SEP-1 bundle benefit overstated. CMS national cohort (Townsend Ann Intern Med 2022) found no mortality reduction after risk adjustment; the 26% from Seymour NEJM 2017 reflects time-to-antibiotics, which is endogenous to severity.

III. Confounder & Collider Risk

Collider stratification. Selecting on hospitalization (a collider between age and infection severity) inflates apparent intervention effects in ICU-based studies (SEP-1, CHG bath).
Time-varying confounding. ART for HIV, DAA for HCV, GLP-1 RA in diabetes — all post-diagnosis. The HRs assume parallel-world comparison; in reality, treated cohorts differ in engagement and SES.
Reverse causation in lifestyle measures. Mediterranean diet's mortality benefit may be partly explained by pre-existing health driving food choices, not vice versa (Mendelian randomization shows weaker effects than observational).

IV. Cancer as Composite — A Particularly Fragile Endpoint

The 0.288 combined HR for "all malignant cancer" aggregates ~200 distinct diseases. Lung cancer (mostly smoking) is highly intervention-responsive (HR ≈ 0.50 with cessation), but pancreatic, glioblastoma, and ovarian show near-null effects. Population-level realized HR likely 0.60–0.75.
Cancer screening (HR 0.80) carries overdiagnosis costs: 15–20% of screen-detected breast cancers and 30%+ of low-grade prostate cancers represent over-treatment, with no mortality benefit.

V. Antimicrobial Resistance & Iatrogenesis

Antibiotic stewardship has dual effects. Reducing broad-spectrum use lowers CDI (HR 0.65) but may delay therapy in confirmed sepsis (raising sepsis mortality). The model treats these independently; in reality they are competing risks.
Vaccine serotype replacement. PCV13/PCV20 reduce vaccine-type pneumococcal disease but non-vaccine serotypes (15A, 23B, 35B) rise, attenuating population-level HR by ~15–25% over 5–10 yr.
FMT efficacy for recurrent CDI (HR 0.20) doesn't generalize to first-episode CDI (FDA approval is recurrent-only).

VI. Demographic Drift & Counterfactual Baseline

An aging US population (Age ≥65: 17% → 22% by 2035) will mechanically raise infectious + chronic disease deaths even with these interventions deployed. "Lives saved" is against today's baseline, not against the projected 2035 baseline.
Vaccine hesitancy (43% non-uptake assumed) is rising; realized population effect = ideal HR × coverage × adherence. Current coverage of 50–70% may halve by 2030 in some demographics.
Obesity and physical-inactivity prevalence are still rising; this baseline drift will partially cancel the gains from exercise + Med-diet interventions.

VII. What Would Falsify This Model?

A cluster-randomized national trial of a multi-intervention bundle showing <15% all-cause mortality reduction over 5 years.
Mendelian-randomization null findings for Vit-D, Mediterranean diet, or alcohol reduction at the magnitudes assumed here.
Sepsis mortality failing to fall despite SEP-1 mandate adoption in >90% of US hospitals (this has happened — flat 2015–2022).
Continued ADRD mortality plateau despite anti-amyloid mAb rollout — suggesting HR 0.93 is closer to 1.00.

Pearl SCM — Method & Assumption Inventory

Assumption Tags

A1 · Faithfulness A2 · Causal Markov A3 · SUTVA (Stable Unit Treatment Value) A4 · Positivity / overlap A5 · Exchangeability conditional on L A6 · Monotonicity for PNS bounds A7 · Linear log-HR additivity within ρ̄ discount A8 · Time-homogeneous hazard A9 · Chinn 2000 OR↔HR conversion (where used) A10 · Eigenvalue regularization (λ_min ≥ 0.05) A11 · Domain-prior correlations adjustable A12 · Endpoint mutual exclusivity NOT assumed (overlap acknowledged in Antithesis) A13 · Saturation modeled per-intervention A14 · No reverse causation (treatment → confounder) A15 · E-value computed per VanderWeele & Ding 2017 A16 · PAF per Levin 1953 A17 · Calibration anchored to NCHS 2019–2024 vital statistics A18 · ICD-10 multi-cause vs underlying disambiguated A19 · No interference between participants A20 · Treatment fidelity matches trial protocol (idealized) A21 · Chronic-disease interventions composable with infectious-disease interventions (cross-domain ρ ≤ 0.30) A22 · Cancer endpoint = weighted composite, heterogeneity in Antithesis

Backdoor Adjustment Set

For any intervention X → endpoint Y, the backdoor adjustment set L = {Age, Immunocompromise, Comorbidities (DM, HTN, Dyslipidemia, COPD, CKD, Obesity), Behavioral exposures (Smoking, Alcohol, Sedentary, Poor diet, IVDU), Social determinants (Poverty, LTCF residence), Environmental (Air pollution, Recent hospitalization), Pathogen-side (Abx overuse, AMR carriage), Vaccine non-uptake, Family history}. The do-operator do(X=x) is approximated via stratified hazard ratios from source meta-analyses, each already adjusted for ≥5 of the 20 confounders.

Combination Rule (Pearl's Cross-Correlation Discount)

log(HR_combined) = [Σᵢ log(HRᵢ)] × δ(ρ̄, k)

where  δ(ρ̄, k) = (1 - ρ̄) + ρ̄/k

ρ̄ = mean off-diagonal correlation within the selected set,
     bounded [0, 0.95], computed from the eigenvalue-regularized
     domain-prior correlation matrix M (28×28).
k  = cardinality of selected set
HRᵢ = disease-specific hazard ratio for intervention i, scaled
      by per-intervention dose intensity:
      HR_dosed = HR_raw^(dose × saturation_curve)

Pareto Risk-Threshold Slider — Mechanics

The slider sets a maximum acceptable residual hazard ratio. The Pareto frontier (built greedily by repeatedly choosing the intervention that most reduces the combined HR) is scanned to find the minimum-cardinality set whose HR ≤ threshold. As threshold loosens (slider moves right), fewer interventions are needed; as threshold tightens (slider moves left), more interventions are added — until the floor (full deployment HR) is reached, beyond which no further reduction is possible.