California’s $2B Air-Quality Bet: Robust Policy, Contested Burden
California spends billions on transport and building electrification to improve air quality. But wildfire is 77% of the state’s PM2.5 burden, and only 12% is controllable by electrification. The optimal transport portfolio — T2 accelerated electrification — is robust to the Di vs Krewski CRF choice: both independently select it. What is not robust is the size of the benefit — the Krewski ACS ≥30 cohort covers 14.7 million more Californians aged 30–64 and implies a 4× larger baseline burden than Di’s Medicare ≥65-only cohort. A $0.5M meta-analysis resolves the residual CRF uncertainty at 100:1 ROI — inside the 5–30:1 Clean Air Act research literature band.
On the residual, the highest-ROI action is information, not a policy.
One Framework, Thirty-Six Questions
One source-receptor model (ISRM, Tessum et al. 2017), compressed into a PCA surrogate ensemble for Monte Carlo feasibility. Dual concentration-response functions (Di et al. 2017 log-linear, Krewski et al. 2009 log-log) run in parallel through every investigation. 10,000 shared Monte Carlo draws with Sobol sensitivity analysis. Every number on this page traces to census-tract PM2.5, ACS/Census population, and peer-reviewed cohort studies. Phase 2 adds twelve multi-fidelity investigations (jump ↓) layering Kennedy–O’Hagan co-kriging, POMDP decision theory, and CVaR/DRO robustness. Phase 3 pushes further (jump ↓) to the research frontier: NSGA-II, polynomial chaos, unified BED, BOCA-inspired multi-fidelity BO, hand-derived 4D-Var adjoint, physics-informed neural networks, linear-operator GPs, and Strong-Oakley-Brennan 2014 nonparametric EVPPI.
InMAP Source-Receptor Matrix (ISRM)
ISRM captures how emissions at each grid cell translate to PM2.5 concentrations at every other cell. Reduced-complexity atmospheric model validated against CMAQ and WRF-Chem.
Variable-resolution grid covering all of California. Finer resolution in populated areas, coarser in rural. Each cell maps to census tracts for health calculations.
Full ISRM compressed via PCA into a fast surrogate. Reproduces InMAP within 2–5% NRMSE. Enables 10K Monte Carlo draws in minutes instead of days.
On-road (11%), residential (1.2%), EGU (0.1%), area (10.7%), wildfire (77%). Each sector independently scalable via 16-dimensional scenario vectors.
Dual CRF Design
Every Monte Carlo draw randomly assigns either the Di et al. (2017) log-linear CRF or the Krewski et al. (2009) log-log CRF, weighted 50/50. This structural uncertainty propagates through all downstream calculations — health impacts, monetization, policy ranking, and VOI. The dual-CRF design is not a sensitivity test; it is the primary analysis. Agreement across the two CRFs is what lets a finding be called robust, and disagreement is itself a finding worth reporting.
Monte Carlo Engine
10,000 draws with shared random seeds across all scenarios, enabling paired comparisons. Parameters perturbed: emission factors (σ=35% residential, 25% on-road), CRF slopes, VSL (triangular $5M / $11.6M / $20M, EPA 2024), surrogate prediction error. Sobol sensitivity analysis (Saltelli sampling, 2nd-order indices) identifies which parameters drive which decisions.
Thirty-Six Questions, One Chain
Each answer enabled the next question. The chain starts with “how many people die?” and ends with “can we emulate WRF-Chem with a physics-informed neural network?” Phase 1 is the 16-investigation decision story. Phase 2 adds 12 multi-fidelity fusions that tighten it. Phase 3 (8 investigations, scroll down) is the methodological frontier — evolutionary optimization, surrogate UQ, unified experimental design, fidelity-aware BO, variational assimilation, PINN emulation, physics-kerneled Gaussian processes, and fine-grained EVPPI decomposition.
Which Transport Electrification Strategy Saves the Most Lives?
T2 (accelerated, $2B) is optimal with 40.5% probability — but T4 (equity, $1B) is nearly tied at 39.2%. EVPPI for CRF structure = $0.066B — the epidemiological model choice matters more than any physical parameter.
Does Building Electrification Improve Air Quality?
B1 (baseline) is optimal with 94.8% probability — aggressive building electrification doesn’t justify its cost. Grid feedback: electrification increases EGU dispatch near peakers, partially offsetting health gains. EVPI = $0.031B — a low-stakes decision.
What Happens When You Combine Everything?
T2 + B2 combined = 1,424 deaths avoided at $4B. Interventions are approximately additive because ISRM is linear and the CRF is approximately linear at small concentration changes. No interaction penalties — the sectors don’t interfere.
Which Gas Plants Kill the Most People?
68 power plants (67 gas + 1 biomass) = 11.2 deaths/yr total. DTE Stockton (the biomass plant) alone = 3.7 deaths (33%). One plant in a disadvantaged community drives a third of facility-level health burden. Environmental justice concentrated at a single site.
Do Wildfire Assumptions Change the Policy?
The top-ranked winner (T2+B2) is stable across all four wildfire scenarios (mild / normal / extreme 2021 / extreme 2020); only one extreme scenario produces a mid-list swap. Even if wildfire doubles, T2 remains optimal. Wildfire is a health-magnitude problem, not a policy-ranking problem.
Can Electrification Make Air Quality Worse?
Yes — locally. T5 (heavy-duty) is net positive statewide (+95 deaths avoided in 2025), but concentrates ozone harm on 6.4M Californians in VOC-limited cores. In the LA Basin, 43.4% of the population lives in a net-harm cell in 2025.
Does the Health Model Change the Optimal Policy?
No — but it changes the burden estimate 4×. Both Di (Medicare ≥65) and Krewski (ACS ≥30) independently select T2 accelerated as the optimal transport portfolio. The divergence is in the magnitude of the benefit, not the ranking: 14.7M additional Californians aged 30–64 sit outside Di’s cohort, leaving 94,490 contested annual deaths (75.1% of the Krewski burden).
Where Should $2B Go — Transport or Wildfire?
10% wildfire reduction: 661 deaths (Di) or 2,196 (Krewski). Transport T2 $2B: 626 (Di) or 2,001 (Krewski). Under Di wildfire edges ahead per dollar; under Krewski, wildfire wins decisively. The biggest untapped lever is fire management — especially if Krewski is closer to the truth.
Should We Just Retire the Worst Plant?
DTE Stockton retirement: $830M net benefit over 20 years ($659M health NPV less replacement and lost-revenue costs) from a single facility closure. The health benefit alone justifies the stranded asset cost. This is the easiest decision in the portfolio — and it costs nothing if the plant reaches end of life.
What Is the CRF Resolution Study Worth?
Residual CRF EVSI = $0.05B after Inv 21 hierarchical-Bayes pooling over CA-pooled cohorts. A $0.5M age-stratified meta-analysis delivers 100:1 ROI on that residual; a $2M retrospective cohort in UK Biobank / MESA delivers 25:1. Both sit inside the 5–30:1 EPA Clean Air Act research literature band. Information is still the highest-ROI action available on the residual.
What Is the Optimal Policy Portfolio?
Free lunch: T1 + B1 + retire DTE Stockton = 1,015 deaths avoided at $0 policy cost. At $2B, wildfire treatment (+724 deaths) beats transport marginal (+367). Building B2 never appears on the efficient frontier until $13.9B. The portfolio answer differs from any single-sector analysis.
Where Should New Monitors Go?
Kriging-optimal: cell 7614 (remote, 242 km from nearest monitor, pop = 0). Decision-theoretic: cell 16581 (LA Basin, 9 km, pop = 13,747). They disagree completely. Optimizing for spatial coverage puts monitors in the wilderness. Optimizing for health decisions puts them in cities.
How Much Solar Energy Does Wildfire Smoke Destroy?
125 GWh/yr lost (0.14% of California solar), $6M/yr in energy costs. But health costs from wildfire PM2.5 are 1,012× the energy penalty. The economic case for wildfire reduction is health, not energy — and it isn’t close.
How Often Does Smoke Block the Sun?
10-year Cal Fire statistics: 125 GWh/yr baseline solar loss from episodic smoke events ($6M/yr). Extreme year (2020 August Complex): 438 GWh lost ($22M). San Joaquin Valley most vulnerable. Trend: smoke days increasing 3–6 days/decade.
Can We Fuse Model and Monitor Data?
ISRM model PM2.5 = 33.8 µg/m³ vs AQS monitors = 7.9 µg/m³. The model is 4× too high because it includes wildfire multi-year averages while 2023 monitors saw clean air. Bayesian fusion reduces uncertainty 55%, but kriging alone beats fusion when the model prior is this biased.
Twelve investigations that push RFAQ into full Analysis Driven Modeling territory. Each defines an explicit fidelity ladder (L1 → L4/L5) with formal multi-fidelity fusion (Kennedy–O’Hagan MFGP, MFMC, CVaR, POMDP, Bayesian optimization) — outputs are combined across levels, not replaced by the highest level. Every investigation anchors on Phase 1 and validates against real data.
If CARB ran the whole Phase 2 arc, this is the sequence
The 12 Phase 2 investigations aren’t 12 separate studies — they are one decision roadmap. Each stage uses the output of the prior stage. Pool what you already have before spending money on new data; spend on new data before committing billions to a portfolio.
Hierarchical Bayesian pooling across 4 cohorts anchors residual CRF EVSI at $0.05B. Do this first — most of the uncertainty resolves without new data.
Stage meta-analysis → retro cohort → Medicare extension. Captures $94M EVSI at $7.5M across three sequenced studies.
Sequential policy avoids 777 vs 651 deaths per trajectory (+19%, $1.5B value); CVaR and adversarial-shift pick the tail-safe policy. Inv 23 downweights B2 under deep uncertainty.
Adaptive POMDP placement lifts monitor ROI from 6.8× to 9.5× at the same $12.5M budget; locks in sites that discriminate the Inv 26 climate fan.
Supporting fidelity work (Inv 17 wildfire ladder, Inv 18 chemistry MFMC, Inv 19 indoor air, Inv 20 grid dispatch, Inv 25 geographic decomposition, Inv 26 climate–fire, Inv 28 data assimilation) tightens each stage without changing the sequence. The roadmap is pool first → buy evidence → deploy robustly → instrument adaptively.
Wildfire Emissions-to-Exposure Ladder
Four-level ladder: linear scaling (L1, Phase 1 baseline, R²=−166) → empirical smoke-day climatology (L2) → FARSITE/BlueSky/HYSPLIT physics surrogate (L3) → WRF-Chem stub (L4). Kennedy–O’Hagan co-kriging fusion. L3 passes the R²>0.5 AQS validation gate — the linear Phase 1 model doesn't.
Indoor Air Quality Coupling
Four-level ladder from outdoor-only (L1) to personal exposure via CHAD time-activity × CONTAM multizone (L4). Adding the indoor pathway takes B2 from 47 deaths (outdoor-only) to 341 (Di) / 1,177 (Krewski) — cost-per-death drops from $42.6M to $5.9M (Di) / $1.7M (Krewski). The B2 verdict flips.
Hierarchical Bayesian CRF
Four-level CRF ladder: discrete Di/Krewski (L1) → Bayesian model averaging (L2) → spatial-hierarchical partial pooling (L3) → E-value causal sensitivity (L4). Normal-normal Gibbs sampler (58 CA counties, 6000 iters, 1000 burn-in) yields HR = 1.067 [95% CI 1.050–1.084] for the CA-pooled PM2.5 CRF. Residual EVSI for CRF resolution anchors at $0.05B.
Atmospheric Chemistry MFMC Overlay
Five-level ozone ladder from NOx-VOC box (L1) to CMAQ hotspots (L4) and WRF-Chem episodic (L5). Kennedy–O’Hagan multi-fidelity GP fuses cheap-statewide ISRM with expensive-hotspot CMAQ. Confirms T5 ozone disbenefit is real, not model artifact — bounds the population harmed to 4.5–9.2M.
Grid Dispatch + EV Charging Feedback
Four-level grid ladder: average EF (L1) → hourly CAISO marginal (L2) → locational marginal by node (L3) → PLEXOS with storage (L4). Validated: reproduces CAISO 2024 dispatch within ±5%. Off-peak LA charging shifts dispatch to Imperial Valley peakers, partially offsetting on-road PM2.5 gains.
Data Assimilation: Monitors + ISRM + CMAQ
Five-level DA ladder (IDW → model-only → OI → 3D-Var → EnKF) fuses three independent PM2.5 streams — FRM/FEM regulatory monitors, PurpleAir low-cost sensors, and ISRM model output. Posterior RMSE drops from 3.41 to 2.71 µg/m³ (−20%) — $488M in value of information unlocked by fusing what Phase 1 used one-at-a-time.
Sequential Portfolio POMDP
Five-level sequential ladder (one-shot → two-stage → rolling → POMDP → BO-optimal) lets the policy learn the true CRF regime over the first 3–5 years and reallocate the remaining budget. Best sequential policy avoids 777 deaths per trajectory vs 651 under commit-and-forget — a 19.2% gain, worth $1.5B at VSL.
CRF Research Roadmap
Five-level research-portfolio ladder anchored on the Inv 21 hierarchical-Bayes residual ($0.05B). Per-design EVSI on that residual: meta-analysis at 40.7× ROI; retrospective cohort at 15.3×; Medicare extension at 7.8× on the age-threshold question. L5 POMDP sequences a 3-arm adaptive portfolio capturing $84M EVSI at $7.5M staged spend — beats any single study.
Free-Lunch Geographic Decomposition
Phase 1's 1,015 deaths-avoided at $0 was a statewide aggregate. Five levels: statewide → air basin (8) → county (58) → tract (9,100) → block-group with CalEnviroScreen DAC uplift. DAC share rises 21% → 31% → 45% with equity-aware allocation. Operational map for CARB + county air districts.
Robust Portfolio Optimization
Five criteria for deep uncertainty: expected value → chance-constrained (P(NB≥$15B)≥0.90) → CVaR0.05 tail → parametric adversarial shift (DRO-lite, not canonical DRO) → info-gap horizon. Consensus: F_maximum wins 3/5 votes. G_sequential wins the tail (CVaR + info-gap). Phase 1 EV pick holds under 3 of 5 criteria.
Climate–Fire Coupling
Five-level ladder from stationary baseline (L1) to 6-GCM CMIP6 ensemble (L3) to WRF-Chem under RCPs (L4). 2050 wildfire deaths rise 1.73× to 25,031/yr [p10=20,141, p90=31,969]. The 11,828-death fan width is 8.2× the Phase 1 10% policy signal — climate dominates policy noise.
Adaptive Monitor Placement
Five-level ladder from Phase 1 static top-N (L1) to POMDP-coupled placement (L4). Same $12.5M / 5-sensor budget: EVSI rises $85M → $119M (ROI 6.8× → 9.5×). L3 ties L2 on total EVSI ($111M) but doubles DAC share (0.40 vs 0.20); L4 trades DAC (0.00) for climate-signal coverage (0.84). Three objectives, three winners.
Eight investigations at the methodological frontier: evolutionary multi-objective optimization, surrogate-based uncertainty quantification, unified experimental design, fidelity-aware Bayesian optimization, variational data assimilation with hand-derived adjoints, physics-informed neural networks, linear-operator GP kernels that bake a PDE directly into the covariance, and Strong-2014 nonparametric EVPPI that decomposes group-level VOI into single-parameter VOI. Every method is implemented from scratch in pure NumPy against the Phase 1–2 decision chain — no black-box library calls.
Multi-Objective Pareto Frontier (NSGA-II)
Evolutionary multi-objective optimization (100 population, 80 generations, SBX + polynomial mutation). Dominates 4 of 6 Inv 23 seed portfolios on the EV / CVaR / DAC-share Pareto surface. The Inv 19 ‘indoor_focus’ seed reaches 0.23 DAC share — highest of all seeds — confirming indoor AQ spending transfers to the equity objective.
Polynomial Chaos Expansion (Inv 17 QoI)
Order-3 Legendre PCE with only 120 model evaluations recovers the Inv 17 Sobol ranking to |ST_pce − ST_mc| ≤ 0.026. Confirms the Inv 17 MC Sobol is not sampling-error-limited. Same answer at 34× less model work — and yields a differentiable surrogate usable for downstream optimization.
Closed-Loop Bayesian Experimental Design
Unified BED sequences 2 CRF studies + 8 monitors across 10 years ($6.5M / $50M spent) against the Inv 21 hierarchical-Bayes residual ($0.05B CRF value-at-stake). Greedy EVSI-proxy captures $107M on the headline run and $84M under a tighter (σprior=0.5) sensitivity. Submodular 1−1/e Nemhauser-Wolsey-Fisher bound holds; joint CRF × monitor belief state outperforms either lever alone.
BOCA-Inspired Multi-Fidelity Monitor Placement
Cost-weighted successive halving with UCB tie-breaking (Kandasamy et al. 2017 framing; not canonical BOCA) over 4 fidelity tiers (gap-score → haversine → climate-UCB → full-sim). Spent 1.50 cost units across 36 evals to recover 4/5 oracle-optimal sites — vs single-fidelity UCB at 5.00 cost units (3.3× more) for 5/5. Cheap tiers prune obvious non-contenders; full-sim is reserved for the contested top-k.
Strong-Constraint 4D-Var Assimilation
Hand-derived adjoint model for 1D advection-diffusion-reaction, L-BFGS with Armijo line search. 4D-Var cuts initial-condition RMSE from 6.05 µg/m³ (background) to 2.07 — a 66% reduction. 3D-Var (end-of-window only) reaches 4.92. 18-hour forecast RMSE: 0.13 / 0.49 / 0.60 µg/m³. J reduced 88.8% in 12 L-BFGS iterations.
Physics-Informed Neural Network Surrogate
PINN (Raissi et al. 2019) with analytic tanh derivatives and PDE residual regularization on the ADR equation. 9 sparse observations (3 downstream stations × 3 times): PINN RMSE 5.18 µg/m³ vs data-only 7.74 — a 33% improvement. Knowing the governing PDE is worth roughly 3–4× more training data in this sparse regime. Pattern for WRF-Chem emulation.
Physics-Informed GP (Linear-Operator Kernel)
Companion to Inv 34 on the same ADR PDE and same 9 noisy monitor observations. Linear-PDE-operator GP (Sarkka 2011; Alvarez-Luengo-Lawrence 2013; Raissi et al. 2017 Numerical GPs): SE prior on c, ADR operator baked into kernel via analytic derivatives through order four. Held-out RMSE 0.12 µg/m³ vs 4.26 for plain-GP baseline. Posterior gives calibrated ±σ — the variance feeds VOI and active-learning directly.
EVPPI Decomposition via Strong, Oakley & Brennan 2014 GAM Regression
Takes Inv 02's $0.229B EVPI and asks: perfect information about what? Strong-Oakley-Brennan (2014) nonparametric regression reuses the existing 10,000-draw PSA sample — no nested MC — to estimate single-parameter EVPPI. Finding: βO3 ($0.116B) and VSL ($0.092B) together hold 91% of the EVPI. All 14 emission-inventory parameters combined account for 0.8%. Reframes CEC's uncertainty-reduction priorities.
Methods appendix → Side-by-side comparison of fidelity ladders (Phase 2) and frontier methods (Phase 3) across all 20 multi-fidelity investigations, with decision anchors and validation data. Read the methods appendix.
What Changes When You Add Decision Context
The analysis started as a standard health impact assessment. The surprising findings came from asking what the numbers mean for policy decisions, not from pushing the atmospheric model to higher resolution. Each row below uses the same ISRM surrogate; only the question changes.
| Level | What’s Included | Finding | What Changes |
|---|---|---|---|
| Health impact assessment | Single CRF, point estimates, one scenario | 14,665 deaths | Standard BenMAP-style result — the number every regulatory analysis produces |
| With uncertainty | Dual CRF, Monte Carlo, Sobol decomposition | EVPI = $15.2M | Uncertainty costs $15M/yr. VSL dominates monetized output, but CRF structure dominates decisions. |
| With policy comparison | Same MC engine + 5 transport scenarios | Ranking robust, magnitude contested | Di and Krewski both pick T2 accelerated. The CRF choice does not flip the ranking; it scales the benefit 4× via age-threshold coverage. |
| With VOI | Same MC engine + hierarchical Bayes + EVSI | $0.05B residual / 100:1 | After Inv 21 pooling, a $0.5M age-stratified meta-analysis resolves the residual at 100:1 ROI. Information is worth more than action on the residual. |
Same ISRM surrogate, same population data, same code across all four rows. Adding decision context doesn’t change the winning transport portfolio (T2 accelerated); it sharpens the question from “pick a policy” to “pick a policy, then resolve the residual CRF uncertainty with a $0.5M meta-analysis.”
The CRF is the scale uncertainty, not the decision uncertainty
Sobol decomposition shows VSL dominates monetized output. For the policy decision, both CRFs agree on the ranking — but the age-threshold coverage gap drives a 4× magnitude difference in total burden. The atmospheric model could be perfect and the size of the benefit would still be uncertain. The bottleneck on the residual is epidemiology, not physics.
The biggest source isn’t controllable
Wildfire is 77% of California’s PM2.5 burden. Electrification policies can only touch 12.3%. A 10% wildfire reduction saves more lives than the most aggressive transport electrification. The policy conversation is focused on the wrong sector.
The free lunch exists
1,015 deaths avoided per year at zero net policy cost: baseline transport (T1) + baseline building (B1) + retire DTE Stockton. These are actions already on the current policy trajectory. Everything expensive sits at the margin beyond that.
Optimize for the decision, not the field
Kriging-optimal placement puts monitors in remote wilderness to maximize spatial coverage. Decision-theoretic placement puts them in the LA Basin to maximize health information. The two objectives pick disjoint sites. A network designed for spatial coverage will leave the hardest health decisions unresolved.
Which Inputs Move the Answer?
Sobol sensitivity analysis (Saltelli sampling, 10K draws, 5 parameter groups) gives different hierarchies depending on what you’re asking about — monetized health impact vs policy ranking. The atmosphere is not the bottleneck for either.
Di (Medicare ≥65, β=0.00705) and Krewski (ACS ≥30, β=0.00545) agree on the policy ranking: both pick T2 accelerated. They disagree about the size of the benefit — Krewski’s cohort covers 14.7M more Californians aged 30–64 and implies a 4× larger baseline burden. The CRF is the dominant scale uncertainty for benefit-cost accounting.
VSL (triangular $5M / $11.6M / $20M, EPA 2024) dominates the monetized output Sobol index. A 30% change in VSL shifts health benefits by billions. But VSL never changes policy rankings — it scales all scenarios equally. It matters for benefit-cost analysis, not for choosing between T1 and T2.
Residential EF uncertainty (σ=35%, Lebel et al. 2022) is the key emissions parameter for building scenarios. Matters for B1-B4 ranking but has negligible impact on the transport decision, which is dominated by CRF structure.
Wildfire is 77% of PM2.5 but ±50% wildfire variation doesn’t change transport policy rankings. Wildfire uncertainty is a health magnitude problem, not a policy choice problem. The decision is robust.
Which input dominates depends on what you’re asking. VSL dominates monetized health impact. CRF structure dominates the transport-policy choice. Wildfire sector magnitude dominates the portfolio question. Know the decision before you rank the sensitivities.
Can You Trust These Numbers?
The atmospheric model, the health calculations, and the kriging system were each validated against independent reference data. The decision-analytic findings (VOI, portfolio) are mathematical consequences of those validated inputs, not new assumptions stacked on top.
What This Cannot Tell You
Known boundaries on the current framework:
ISRM Is Reduced-Complexity
ISRM is not a full chemical transport model. It captures primary PM2.5 and secondary formation via linearized source-receptor relationships. Nonlinear chemistry (ozone formation, SOA) is approximated. For the scenario perturbations in this study (<30% emission changes), the linearization error is small.
Wildfire as Annual Average
The ISRM wildfire sector represents a multi-year average smoke burden, not episodic events. Actual wildfire PM2.5 varies by orders of magnitude year to year. The 2023 AQS monitors saw clean air (7.9 µg/m³); the model reflects the long-term average (33.8 µg/m³). Episodic health impacts are underestimated by the annual average framework.
No Indoor Air Quality
Building electrification eliminates indoor combustion (gas stoves, furnaces). The indoor NO2 and ultrafine particle benefits — especially from cooking gas replacement — are likely larger than the outdoor PM2.5 benefits we model. Our framework captures only the outdoor pathway.
Two CRFs, Not the Full Literature
We use Di et al. 2017 and Krewski et al. 2009 as representative endpoints of the CRF spectrum. Other CRFs exist (Lepeule et al. 2012, Pope et al. 2002, Burnett et al. 2018). The binary choice overstates structural certainty. A full Bayesian model average over the CRF literature would be more defensible.
Static Scenario Vectors
Scenarios are defined at 4 time snapshots (2025, 2030, 2035, 2040). Between snapshots, we interpolate linearly. Dynamic feedbacks (fleet turnover curves, building stock replacement, grid decarbonization) are simplified into scaling factors, not modeled mechanistically.
No Environmental Justice Weighting
Health impacts are calculated at census-tract level and monetized uniformly. We do not apply equity weights (e.g., higher VSL for disadvantaged communities). The DTE Stockton finding highlights environmental justice, but the monetization does not formally encode it.
What the Analysis Supports
Six recommendations — each traceable to specific investigations and grounded in a validated atmospheric model with dual-CRF uncertainty quantification. Death counts below are expected values under the Di 2017 / Krewski 2009 CRFs, EPA 2024 VSL, and ISRM nominal emissions; the dual-CRF framework shows that the portfolio ranking is stable across the epidemiological structural choice even where the magnitude is not.
-
01
Pool existing cohorts first, then fund an age-stratified meta-analysis
Inv 21 hierarchical Bayesian pooling across Di, Krewski, Turner, and IHD cohorts anchors residual CRF EVSI at $0.05B. Inv 11 prices a $0.5M age-stratified meta-analysis at 100:1 ROI on that residual; a $2M retrospective cohort (UK Biobank / MESA) at 25:1. Inv 24 then sequences a 3-arm adaptive research portfolio capturing $84M EVSI at $7.5M staged spend. Both ROIs sit inside the EPA Clean Air Act 5–30:1 literature band — information remains the highest-ROI action on the residual.
-
02
Capture the free lunch: ~1,015 expected deaths at $0
Baseline transport electrification (T1), baseline building codes (B1), and DTE Stockton retirement together are predicted to avoid ~1,015 expected deaths per year at zero incremental policy cost (Di/Krewski dual-CRF, 10K MC). These are current-trajectory outcomes. Ensure they happen on schedule before investing in acceleration.
-
03
Invest marginal dollars in wildfire reduction, not transport acceleration
At $2B, wildfire treatment is predicted to avoid ~724 additional expected deaths vs transport’s ~367 (dual-CRF, 10K MC). Wildfire is 77% of PM2.5 and delivers higher ROI per dollar than any electrification acceleration. The efficient frontier puts wildfire treatment on the Pareto front before building electrification at every budget level.
-
04
Retire DTE Stockton on environmental justice grounds
One biomass plant (DTE Stockton) in a disadvantaged community produces 33% of all facility-level PM2.5 mortality (3.7 of 11.2 deaths/yr). 20-year net benefit of retirement: $830M ($659M health NPV minus replacement/lost-revenue costs). The facility-level analysis makes this the most concentrated environmental justice action in the portfolio.
-
05
Place new monitors to maximize health decision value, not spatial coverage
Kriging-optimal and decision-theoretic sensor placement disagree completely. Current network design optimizes for spatial field reconstruction. Health decisions need monitors where mortality burden × PM2.5 uncertainty is highest — populated areas with high exposure, not remote wilderness.
-
06
Pick transport policies on spatial distribution, not just statewide totals
Broad electrification (T2) is predicted to avoid ~224 expected deaths statewide in 2025 with 1.54M population in net-harm cells. Heavy-duty-first (T5) is predicted to avoid only ~95 expected deaths but puts 6.4M in net-harm cells — mostly in the LA Basin (43.4%). Headline policy totals hide spatial equity. Pair heavy-duty NOx cuts with co-located VOC reduction to shrink the harm footprint.
Reproducible with Public Data
Every input comes from publicly available, peer-reviewed sources. Every analysis is reproducible.
Atmospheric Model
- Tessum, C.W. et al. (2017). “InMAP: A model for air pollution interventions.” PLoS ONE, 12(4), e0176131. doi:10.1371/journal.pone.0176131 [ISRM]
- Goodkind, A.L. et al. (2019). “Fine-scale damage estimates of particulate matter air pollution reveal opportunities for location-specific mitigation of emissions.” PNAS, 116(18), 8775–8780. doi:10.1073/pnas.1816102116 [Marginal damages]
Epidemiology
- Di, Q. et al. (2017). “Air pollution and mortality in the Medicare population.” N. Engl. J. Med., 376(26), 2513–2522. doi:10.1056/NEJMoa1702747 [Log-linear CRF]
- Krewski, D. et al. (2009). “Extended follow-up and spatial analysis of the ACS study.” HEI Research Report, 140. [Log-log CRF]
Emissions & Population
- CARB (2023). California Emissions Inventory. [Sector totals]
- US Census Bureau. 2020 Decennial Census & ACS 5-year estimates. [Tract-level population, age]
- CalEnviroScreen 4.0 (2023). OEHHA. [Environmental justice indicators]
- Lebel, E.D. et al. (2022). “Methane and NOx emissions from natural gas stoves, cooktops, and ovens.” Environ. Sci. Technol. doi:10.1021/acs.est.1c04707 [Residential EF]
Monitoring & Fusion
- US EPA. Air Quality System (AQS) annual PM2.5 data, 2023. [112 CA monitors]
- Liu, Y. et al. (2016). “Smoke incursions into urban areas.” Atmos. Environ. [Smoke climatology]
- Aguilera, R. et al. (2021). “Wildfire smoke impacts respiratory health.” Sci. Advances. [Regional smoke days]
Economic
- US EPA (2024). Guidelines for Preparing Economic Analyses, Appendix B: Mortality Risk Valuation. VSL = $11.6M (2024$, central). [VSL]
- EIA (2024). California Solar Generation Statistics. [Solar capacity & generation]
36 investigations (16 Phase 1 · 12 Phase 2 · 8 Phase 3) · 21,164 grid cells · 10,000 Monte Carlo draws · Dual CRF (Di et al. 2017 + Krewski et al. 2009) · 5 emission sectors · PCA-compressed ISRM surrogate (2–5% NRMSE) · Sobol sensitivity (Saltelli sampling) · VOI / EVPI / EVPPI / EVSI · 112 AQS monitors + ordinary kriging · Bayesian sensor fusion · Portfolio optimization (efficient frontier) · Aligned to CEC GFO-25-304 Group 1. All code and data reproducible with public sources.