A Structural Causal Model (SCM) integrating hazard ratios, mechanism-of-action, and dose-response evidence across the three leading domains of preventable self-caused mortality in the United States. Interventions are evaluated under backdoor adjustment, cross-correlation removal (mean off-diagonal ρ̄ ≈ 0.30 across shared confounders), and do-calculus, with E-values, PNS/PN/PS, and population-attributable fractions reported per node.
For each candidate intervention X on outcome Y ∈ {overdose, suicide, unintentional injury}, the analysis verifies a sufficient adjustment set Z under the backdoor criterion using the SCM in §02.
Effects are recovered as do(X=1) − do(X=0). All hazard ratios are extracted from peer-reviewed evidence sources (largely Cochrane reviews, Lancet meta-analyses, BMJ, JAMA, AJPH, Inj Prev), then converted to the standardized scale via Chinn 2000 (SMD → ln HR via π/√3).
Cross-correlations between interventions (e.g., MOUD ↔ naloxone, both downstream of OUD; SSRI ↔ CBT, both via depressive symptomatology) are quantified in a Σ matrix; eigenvalue-corrected to remove redundancy with a mean off-diagonal ρ̄ ≈ 0.30.
Naive multiplication ∏HRi assumes independence, which is biologically false here — many interventions act on shared mediators (intoxication, impulsivity, hopelessness, lethal-means access). The correlation-corrected log-additive estimator uses the eigenvalue-corrected effective sample size:
This derives from multivariate-normal latent-effect variance theory: for n correlated interventions with mean off-diagonal ρ̄, the variance of their summed log-hazard contribution scales as n[1 + (n−1)ρ̄], so neff = n / [1 + (n−1)ρ̄]. With ρ̄ = 0.30 and n = 8 active interventions, neff ≈ 2.58 and κ ≈ 0.32 — the 8-intervention portfolio acts as ~2.6 fully independent interventions after de-duplication.
Saturation thresholds (each intervention's empirical dose-response plateau) further cap marginal contribution at observed ceilings (e.g., MOUD HR cannot fall below ~0.30 even with full coverage).
Selecting "Activate all" toggles every intervention to its dose-response saturation point and applies the cross-domain correlation matrix.
Toggle a counterfactual to observe the marginal effect of removing or adding the single most impactful intervention in each domain. The display recomputes assuming all other interventions remain at their current setting.
| Domain | Intervention | HR | 95% CI | E-value | PN | PS | PNS | PAF | Evidence basis |
|---|
Sweep the mean off-diagonal correlation ρ̄ from 0.0 (full independence — naive product) to 0.6 (heavy mediator overlap). The composite HR at full-portfolio activation moves accordingly.
Each intervention's point HR is jittered ±25% (capturing meta-analytic CI width) over 1,000 Monte Carlo draws. The histogram below shows the resulting composite reduction distribution at full portfolio activation, with the Pearl-adjusted central estimate marked.
Most intervention HRs come from observational cohorts (MOUD enrollees are systematically different from non-enrollees on motivation, social support, healthcare access). Backdoor adjustment can only block measured confounders. The E-values shown bound the strength of unmeasured confounding needed to fully nullify each effect — for MOUD with HR 0.50, an unmeasured confounder would need RR ≥ 3.4 on both exposure and outcome, which is implausible but not impossible.
Population-average HRs hide enormous individual variation. Lithium's anti-suicidal effect is concentrated in bipolar I (HR ~0.3); applying it to unipolar depression with mild ideation gives a much weaker effect (HR ~0.8). The dashboard's domain-level rollup understates this stratification.
Means restriction can be partially offset by substitution to other methods, though the bridge/firearm-storage literature shows substitution is incomplete (Yip 2012, Lancet). PDMP enforcement reduced prescription-opioid deaths but contributed to street-fentanyl substitution post-2014. Neither effect is currently mediated in the DAG.
The κ correction is a first-order linear discount. True effect aggregation under shared mediators is non-linear; with full saturation of multiple interventions acting on the same mediator (e.g., intoxication), the marginal contribution can collapse to zero rather than gradually diminish. The composite ceiling shown here is therefore an upper bound, not a point estimate.