ENDPOINT:
T2D Incidence
T2D Remission
Metabolic Syndrome
MASLD / NAFLD
Insulin Resistance
MACE (in T2D)
CKD Progression
All-Cause Mortality
Intervention Controls — Pearl do(·) Operator
A high (RCT/meta) · B cohort · C modelled · D consensus. Lower grades excluded from the joint hazard ratio.
Risk Factor Burden — Patient Profile
Baseline Risk (no Rx)
—
—
Combined HR (Pearl-adj)
—
—
Net Risk Reduction
—
95% CI: —
NNT (10-year)
—
—
Effect Decomposition — Per-Intervention Risk Reduction (%)
Naive vs Pearl-Adjusted Combined Effect
Why these differ: Naive multiplication assumes independence. Pearl adjustment removes cross-correlations (mean ρ̄ ≈ 0.31), applies dose-response saturation, blocks backdoor confounding paths.
10-Year Cumulative Incidence Forecast
Causal Directed Acyclic Graph
Pearl-adjusted causal structure. Confounders → exposures → mediators → outcomes. Dashed green edges = protective intervention effects via do-operator.
Legend
Risk factor
Mediator
Confounder
Intervention
Endpoint
Backdoor Adjustment Sets
Diet → T2D: {age, sex, BMI, SES, PA}
Exercise → CVD: {age, sex, smoking, BMI, fitness}
GLP-1 → MASLD: {BMI, T2D, alcohol, PNPLA3}
Sleep → IR: {age, depression, OSA, shift work}
SGLT2 → CKD: {eGFR, T2D, HTN, ACEi}
Path Coefficients (β on logHR)
Visc adiposity → IR: +0.82
IR → T2D: +1.14
T2D → MACE: +0.68
GLP-1 → BMI: −0.71
Aerobic → IR: −0.43
Med diet → MetSyn: −0.35
SGLT2 → CKD: −0.51
Bariatric → T2D rem: −1.61
Risk Factor Inventory (22 modifiable + non-modifiable drivers)
Pearl-adjusted hazard ratios for T2D incidence. PAF computed under counterfactual elimination.
| Risk Factor | Class | Naive HR (95% CI) | Pearl HR | E-val | PAF | Mediator | Tier |
|---|
PAF caveat: Population Attributable Fractions are not additive; they exceed 100% due to overlapping causal pathways. Joint PAF (overlap-corrected) ≈ 67% for T2D.
Intervention Inventory (24 do(X) interventions)
Each intervention assessed under Pearl do(X) framework: backdoor adjustment, dose-response saturation (Emax), PNS/PN/PS, E-value.
| Intervention | Class | Mechanism | HR | E-val | PNS | PN | PS | 90% Emax Dose | Tier |
|---|
Disease Stages — Stage-Specific Causal Models
Pearl-adjusted intervention effects differ by stage. Bariatric peaks at Stage 3, plateaus once β-cell exhaustion exceeds ~50%.
STAGE 0
Pre-clinical — Genetic & Lifestyle Loading
TCF7L2 carriers, family history, GDM, PCOS, low birth weight. Phenotypically normal but causal priors elevated.
Best lever: Mediterranean (HR 0.72), exercise (HR 0.68)
E-value: 2.1 · NNT(20y): ~28
E-value: 2.1 · NNT(20y): ~28
STAGE 1
Insulin Resistance — Compensated Hyperinsulinaemia
HOMA-IR > 2.5, fasting insulin elevated, normal glucose. Visceral adiposity, MASLD onset. Maximum reversibility window.
Best lever: Weight loss 5–10% (HR 0.42), TRE (HR 0.65)
E-value: 3.6 · NNT(5y): ~9
E-value: 3.6 · NNT(5y): ~9
STAGE 2
Prediabetes — Impaired Fasting/Tolerance
FPG 100–125 mg/dL or HbA1c 5.7–6.4%. β-cell stress visible. DPP: lifestyle reduced T2D by 58%/3y.
Best: Lifestyle (HR 0.42), metformin (HR 0.69)
E-value: 3.2/2.3 · NNT(3y): ~7/~14
E-value: 3.2/2.3 · NNT(3y): ~7/~14
STAGE 3
Overt T2D — Hyperglycaemia Established
HbA1c > 6.5%, β-cell function ~50% reduced. Remission possible if recent onset (<6y).
Best: Bariatric (HR 0.21), GLP-1 + lifestyle (HR 0.38)
Remission: bariatric 75% / GLP-1 30% / VLCD 46%
Remission: bariatric 75% / GLP-1 30% / VLCD 46%
STAGE 4
T2D + Microvascular Complications
Retinopathy, neuropathy, albuminuria. β-cell function <30%. Focus shifts to organ protection.
Best: SGLT2i (CKD HR 0.61), GLP-1 (CV HR 0.86)
BP <130/80 → HR 0.71 retinopathy
BP <130/80 → HR 0.71 retinopathy
STAGE 5
T2D + Macrovascular — MACE, ESRD, MASH
Established CVD, HF, advanced CKD, decompensated cirrhosis. GLP-1/SGLT2 mandatory unless contraindicated.
Best: GLP-1 (MACE HR 0.86), SGLT2 (HF HR 0.69)
Resmetirom: 26% MASH resolution
Resmetirom: 26% MASH resolution
Stage-Conditional Intervention Effects (HR for progression)
Dose-Response Saturation — Emax Model
E(d) = Emax · d^n / (ED50^n + d^n); Hill coefficient n from primary dose-finding studies.
Aerobic Exercise (MET-h/wk)
Saturation ~25 MET-h/wk (~150 min moderate). Marginal returns drop sharply past 40 MET-h.
Weight Loss (% body weight)
5% = meaningful glycaemic. 10% = remission window. >15% (GLP-1/bariatric) approaches asymptote.
GLP-1 (semaglutide-eq mg/wk)
Tirzepatide note: dual GIP/GLP-1 shifts curve leftward. 15 mg tirzepatide ≈ 2.4 mg semaglutide for weight.
Dietary Fiber (g/day)
SCFA → GLP-1 + insulin sensitivity. Plateaus near 40 g/day.
Cross-Correlation Matrix — Why Pearl Adjustment Matters
Mechanism overlap means assuming independence overestimates combined effect by 20–60%.
Pearson Correlation Heatmap (selected interventions)
Correction: Pearl-corrected joint HR uses cluster-aware Mantel-Haenszel pooled estimator. With 8 interventions for T2D, joint HR shifts from 0.084 (naive) → 0.218 (Pearl) — a 78% (not 92%) risk reduction.
Mechanism Clusters
Weight-loss: GLP-1, bariatric, VLCD, TRE, keto (ρ̄ ≈ 0.55)
Insulin-sens: Metf, berberine, aerobic, Mg, ALA (ρ̄ ≈ 0.42)
Inflammation: Omega-3, Med, polyphenols, fiber (ρ̄ ≈ 0.38)
Circadian: Sleep, stress, TRE, light (ρ̄ ≈ 0.36)
CV-protect: SGLT2, GLP-1, omega-3, BP (ρ̄ ≈ 0.28)
Appropriate % Correlation Discount
ρ > 0.7 → mechanisms duplicate → use one only
0.4 < ρ ≤ 0.7 → discount 50% logHR
0.2 < ρ ≤ 0.4 → discount 20% logHR
ρ ≤ 0.2 → near-independent stacking
Our model: applies cluster-aware discounts before multiplicative stacking.
Pareto Frontier — Minimum Effective Intervention Set
Each point = a candidate bundle. Frontier = non-dominated bundles on (risk reduction, burden).
Effect vs Burden (cost × adherence friction)
Pareto-Optimal Bundles (elbow)
Reading: Most patients best served by 3–6 intervention bundles on the elbow. Diminishing returns dominate past ~8 interventions due to cross-correlation saturation.
Sensitivity Analysis & E-values
E-value = min strength an unmeasured confounder would need to fully explain away the observed effect.
E-value Forest — Top Interventions
Tornado — One-at-a-time Perturbation (±20% HR)
ρ̄ Perturbation
ρ̄ = 0.25 → joint HR = 0.198
ρ̄ = 0.31 (base) → 0.218
ρ̄ = 0.37 → 0.241
ρ̄ = 0.60 → 0.310
±20% perturbation → ±10% joint HR.
HR Perturbation (±25%)
Bariatric 0.21 ± 0.05 → 0.20–0.24
GLP-1 0.40 ± 0.10 → 0.20–0.23
SGLT2 0.65 ± 0.15 → 0.21–0.23
Bariatric = leverage point; GLP-1 second.
Monte Carlo Bootstrap (10,000 reps)
Median joint HR: 0.218
2.5 pctile: 0.171
97.5 pctile: 0.279
SD: 0.027
Skew: +0.42
Antithetical Analysis — Steel-Manning the Counter-Arguments
Each major assumption is challenged. The verdict reports whether the model's substantive conclusions change.
Methods — Pearl SCM Implementation
Full reproducible pipeline. All assumptions explicit. Evidence tier reported per parameter.
Structural Causal Model — Specification
A Pearl SCM is a triple ⟨U, V, F⟩ where U = exogenous variables, V = endogenous, F = structural equations. We specify V as 22 risk factors, 12 mediators, 24 intervention nodes, 8 endpoints.
Identifiability of P(Y | do(X)) requires the backdoor criterion: a set Z blocks all backdoor paths from X to Y, with no descendant of X in Z. We use the IDA algorithm (Maathuis 2009) to enumerate admissible adjustment sets.
Effect Estimation Pipeline
- Source HR from meta-analysis (preferred) or largest cohort.
- Convert SMD → HR via Chinn (2000): logHR ≈ 1.81 × SMD.
- Identify backdoor adjustment set Z per intervention.
- Compute Pearl-adjusted HR via inverse-probability weighting.
- Compute E-value (VanderWeele & Ding 2017).
- Compute PNS, PN, PS via counterfactual probability bounds.
- Fit dose-response Emax model.
- Compute correlation matrix; eigenvalue-correct if not PSD.
- Stack via cluster-aware Mantel-Haenszel.
- Bootstrap 10,000 reps for joint HR CI.
Evidence Tiers
| Tier | Definition | # Iv |
|---|---|---|
| A | ≥2 RCTs + meta-analysis, consistent | 9 |
| B | 1 RCT + supportive cohort/mechanism | 7 |
| C | Cohort/case-control + mechanism, no RCT | 6 |
| D | Mechanistic + small studies; conjectural | 2 |
Key Formulas
Pearl do-operator: P(Y | do(X)) = Σ_z P(Y | X, Z=z) · P(Z=z)
E-value: E = HR + √(HR · (HR − 1)) for HR > 1; reciprocal for HR < 1
PAF: P_e · (HR − 1) / (1 + P_e · (HR − 1))
PN: P(Y_x'=0 | X=x, Y=1) ∈ [max(0, (P(Y|X) − P(Y|X'))/P(Y|X)), 1]
PS: P(Y_x=1 | X=x', Y=0)
PNS: P(Y_x=1 ∧ Y_x'=0) ∈ [max(0, P(Y|X) − P(Y|X')), min(P(Y|X), P(Y'|X'))]
Cluster-corrected joint: log(HR_joint) = Σ log(HR_i) · (1 − discount_i)
Chinn SMD→HR: log(HR) ≈ 1.81 · SMD
Emax: E(d) = Emax · d^n / (ED50^n + d^n)