Skip to main content

California Freight Cleanup → Investigation 3-6

Which input drives the most uncertainty in how far ahead the top portfolio wins?

Slug renamed 2026-05-04: cascade-sobol-gsa → cascade-sobol-margin

A 40,960-call Saltelli/Sobol design over 4 cascade inputs decomposes the variance of the dominance margin of portfolio D_all_in_4B: how much better it is than the next-best alternative, per draw. Emissions inventory scale is the top total-order driver (ST = 0.453), followed by the PM2.5 concentration response function (ST = 0.375). D’s competitive lead is robust: P(margin > 0) = 99.96%.

We know the all-in $4B portfolio (D_all_in_4B) wins — it dominates all alternatives with zero expected regret. That answers “which portfolio wins?” A CEC evaluator’s follow-on question is sharper: which uncertain input, if resolved, would most strengthen our confidence in that answer?

To answer that, we need to know which inputs drive the variance in how far ahead D wins. Investigation 3-6 runs a global sensitivity analysis over the full cascade, focused on the dominance margin — how much D outperforms the next-best alternative per simulation draw. Variance in that margin tells us which inputs most deserve additional research investment.

Six candidate inputs were scoped; two correctly excluded. βO3 enters via Investigation 4-2’s ozone MFMC and is baked into Investigation M-1 mortality means before reaching Investigation 6-6’s NPV math — running Sobol on it here yields S1 = ST = 0 by construction. The surrogate residual σ from Investigation 3-4’s MFGP would double-count the CRF estimation uncertainty already captured by βPM2.5. The ozone-channel gap is addressed separately by Investigation 4-1’s 5-D ozone Sobol.

The four inputs that genuinely vary inside Investigation 6-6’s NPV pipeline:

InputDistributionSource
βPM2.5Normal(μ=0.02439, σ=0.00447)Investigation 6-3 L3 hierarchical Bayesian posterior
VSLTriangular($7.4M / $10.8M / $13.4M)US EPA 2024 mortality risk valuation (narrow band)
emissions_scaleLognormal(μlog=0, σlog=0.20)EPA NEI 2023 TSD §3.6 (±40% at 95% CI)
cost_overrunLognormal(μlog=0, σlog=0.30)DOE/RAND construction-cost overrun literature

The Saltelli design (SALib v1.5.2, scrambled, seed 20260430) generates N × (2D+2) = 40,960 wrapper calls. Each call evaluates the byte-identical NPV function from Investigation 6-6, with cost_overrun broadcast as a multiplier on each portfolio’s deterministic cost. The quantity of interest per draw:

margin_d = NPV(D_all_in_4B) − max{NPV(A), NPV(B), NPV(C), NPV(E)}

First-order (S1), total-order (ST), and pairwise second-order (S2) indices are estimated with 100-resample bootstrap CIs. The VSL band here is the EPA-narrow envelope; Investigation 3-10 uses the canonical-wide ($5M–$20M) envelope. Differences in VSL’s ST share between the two are expected and informative — the wider envelope gives VSL a proportionally larger variance contribution.

Emissions inventory accuracy matters most (44% of total variance)

44.3% of the variance in D’s dominance margin is attributable to emissions inventory uncertainty (σlog = 0.20, EPA NEI 2023 TSD §3.6). Better emissions data is the single highest-leverage acquisition for reducing uncertainty in the competitive strength of the recommendation.

The health-risk coefficient is close behind (37% of total variance)

The Investigation 6-3 hierarchical posterior (μ = 0.02439, 95% CI 0.01560.0331) contributes 36.7% of total-order variance — the second-priority data acquisition target, and the only lever also touched by Investigation 8-5’s VOI surface.

Life-valuation assumptions contribute 15%; construction cost uncertainty is negligible

VSL contributes 15.2% under the EPA-narrow band. Under Investigation 3-10’s canonical-wide VSL ($5M–$20M), that share rises to 27% — the largest single difference between the two companion studies. Cost overrun at σlog = 0.30 contributes under 3.9%; construction-cost uncertainty does not threaten D’s position.

The inputs act mostly independently — interactions are minor

Interaction fraction is 2.3%. The largest pairwise term is βPM2.5 × emissions_scale (S2 = 0.014) — physically expected, since both multiply deaths_avoided in the NPV formula. All other pairs are at or below noise. The NPV structure is effectively additive in variance-space at this dimension.

D’s dominance is robust: P(margin > 0) = 99.96%

Mean dominance margin: $7.63 billion across 40,960 draws (P5 = $3.40B, P95 = $12.82B). D’s NPV falls below all alternatives in only 99.96% of draws. Even at the joint 5th percentile, D outperforms the next-best portfolio by $3.4 billion.

ItemSHA-256 (12-char)
results.json1d1b387f29bd
analysis.md
scenario.md
Upstream: Investigation 6-3 (CRF posterior) investigations/21_crf-hierarchical-bayes/latest/results.json 3104ba850408
Upstream: Investigation M-1 (portfolio frontier) investigations/15_portfolio-frontier/latest/results.json 145dbfd826d0
Upstream: Investigation 6-6 (CRF-conditional decision) investigations/44_crf-conditional-decision/latest/results.json 5ce9bcd8b87b
Run timestamp 2026-05-04T07:48:05   Sobol N = 4096   SALib v1.5.2   seed = 20260430

Note: analysis.md records stale-upstream warnings for Investigation M-1 and Investigation 6-6 (sha256 drift since last Investigation 3-6 run). The dominance-margin finding (P(margin > 0) = 99.96%, top driver = emissions_scale) is not expected to change materially on re-run; the stale-upstream flag is informational only.