California Freight Cleanup → Investigation 6-1

What does PM2.5 cost California — and does more data change the call?

$179.6B mean — 10,000-draw MC, EPA 2024 VSL, full CRF+surrogate uncertainty

Every cost and equity number in the California Freight Cleanup cascade rests on a defensible dollar figure for what air pollution does to California’s 32 million residents. This investigation builds that number end-to-end under full uncertainty, then asks whether spending more to resolve that uncertainty would change the portfolio recommendation. The answer: no — at $15M against a $2B decision, additional analysis would not pay for itself.

Decision context

Before you can say whether an air-quality program is worth its cost, you need a defensible number for what PM2.5 and ozone actually cost in lives and dollars. We build that number bottom-up: spatially varying exposures across 21,164 California grid cells, attributable deaths and hospitalizations, and monetization at the EPA 2024 value of a statistical life. A 10,000-draw Monte Carlo propagates uncertainty in the dose-response function, the VSL, and the air-quality model simultaneously. The second half asks whether spending more to resolve any of that uncertainty would change which policy to choose.

Methodology

Spatial setup. A grid–tract crosswalk maps 21,164 model cells to census tracts. PM2.5 baselines are AQS regional means (LA Basin 14 μg/m³, SJV 16, Bay Area 10, Sacramento 9, rest-CA 8); O3 uniform at 42 ppb; NO2 at 25/15/12 ppb by region. Baseline mortality rates from CARB/EPA evaldata v1.6.1 (2013 vintage). Population 32,031,832; DAC share 22.9%.

Monte Carlo (10,000 draws, seed 42). Three independent uncertainty groups propagated simultaneously: (a) CRF (βPM2.5, βO3, βNO2, Di/Krewski 50/50 switch); (b) VSL drawn from Triangular($5M, $11.6M, $20M) with income-elasticity Uniform(0.4, 1.0); (c) surrogate GP noise z-score at 10% CV placeholder (pending fitted surrogate from Investigation 3-4). All four tracked outputs pass convergence diagnostics.

Joint sampler (Phase D2, 200 draws). A correlated draw pass via rfaq.uncertainty.joint_sampler encodes rank correlations from literature: corr(βPM2.5, βO3) ≈ +0.3 (Krewski/Turner cohort overlap); corr(emissions_scale, surrogate_σ) ≈ +0.4 (ECHO-AIR / ISRM model documentation (Tessum & collaborators; see echo-air-model.github.io for current model docs)). βPM2.5 reads live from the Investigation 6-3 posterior. The 3.8× higher joint-sampler mean (55,863 vs. 14,665 deaths) is not an error: positive correlations cause the high-uncertainty draws to cluster together rather than cancel. When CRF, emissions, and surrogate noise all land in their upper tails simultaneously, the joint mean is driven up dramatically. The independent MC averages over uncorrelated extremes; the joint MC averages over correlated worst cases.

Value of Information. EVPI = E[max NB(d,θ)] − max E[NB(d,θ)] over five policy alternatives. EVPPI for five parameter groups uses the Strong M., Oakley J.E., Brennan A. (2014), “Estimating multi-parameter partial Expected Value of Perfect Information from a probabilistic sensitivity analysis sample: a non-parametric regression approach.” Medical Decision Making 34(3):311–326 GAM regression estimator.

Headline results

QuantityValue
Deterministic baseline deaths/yr11,665
Deterministic monetized value$135.5B
MC mean total deaths14,665 (P5: 7,679 / P95: 23,217)
MC mean monetized value$179.6B (P5: $78.3B / P95: $322.5B)
PM2.5 deaths (MC mean)5,596
O3 deaths (MC mean)9,069
DAC burden share (MC mean)20.8% (P5–P95: 19.8%–22.1%)
Joint-sampler deaths (200 draws)55,863 (P5: 44,461 / P95: 66,909)
Joint-sampler monetized$660.2B (P5: $355.8B / P95: $961.8B)
Optimal decision (independent MC)accelerate_ev
EVPI$0.015B
EVPPI(surrogate)$0.005B (dominant group)
EVPPI(CRF, emissions, monetization)≈$0 (all numerically zero)

EVPI = $0.015B: the portfolio decision is robust to all remaining uncertainty

$15M against a $2B policy choice. The expected cost of the wrong decision is negligible. The dominant EVPPI component is surrogate GP noise ($5M)—the 10% CV placeholder for Investigation 3-4 is the one group where additional information would most shift expected benefit. CRF, emissions, and monetization EVPPI are numerically zero under the current linear policy-perturbation model.

DAC burden share (20.8%) is lower than DAC population share (22.9%): equity gap is real but small under current PM2.5 model

Disadvantaged Communities bear 20.8% of monetized mortality burden against a 22.9% population share—a 2.1-point under-attribution. The gap direction is most sensitive to whether South Coast Basin DAC tracts receive above- or below-regional-average PM2.5 from the fitted surrogate. This caveat propagates to Element 5.

Caveats