California Freight Cleanup → Investigation 6-3

Do the published CRF studies bracket the true risk?

HR 1.28 per 10 μg/m³ PM2.5 — multi-pollutant Cox survival model, 705,610 subjects

Every health-cost number in the California Freight Cleanup cascade depends on how strongly PM2.5 raises mortality risk. The two standard published estimates — Di et al. 2017 and Krewski et al. 2009 — disagree. We built a real-cohort analysis from NHIS and NHANES public data to cross-check both, found a higher result (HR = 1.28), and documented why: public survey data assigns exposure at the census-region level, not ZIP-level, which mechanically inflates the estimated slope. For policy comparability with CARB, the portfolio uses the Krewski 2009 range, not this estimate.

Decision context

Every health-cost and portfolio net-benefit number in the cascade depends on the PM2.5-to-mortality dose-response function. Two standard published estimates exist: Di et al. 2017 (Medicare enrollees aged 65 and up, HR = 1.073 per 10 μg/m³) and Krewski et al. 2009 (adults aged 30 and up, HR = 1.056). Both were fit on national data. Picking one without justification is not defensible for a California-specific program. We built the dose-response chain end-to-end on the largest real cohort accessible without a restricted-data application, documented why our result overshoots the published anchors, and delivered the posterior that downstream investigations use as an informed prior.

Methodology: four-rung fidelity ladder

L1 — Discrete CRF anchors. Di et al. 2017 (NEJM 376:2513) and Krewski et al. 2009 (HEI 140) encoded as published point HRs (β = ln(HR)/10). Neither is re-fitted; both serve as ends of the prior envelope for L2.

L2 — Bayesian model averaging (50/50 Di/Krewski). Fixed-weight Gaussian mixture of the two L1 anchors: β posterior mean = 0.006247 per μg/m³, posterior 95 % CI [0.00438, 0.00811]. This is what most policy work implicitly does when it cites “either Di or Krewski”; it is the baseline against which the L3 real-cohort fit is compared.

L3 — Hybrid frequentist hierarchical Cox PH on real cohort. The NHIS Linked Mortality File public-use 2019 release (interview years 2005–2015, follow-up through December 2015) was joined to EPA AQS 5-year-mean PM2.5 at census-region granularity (4 regions). NHANES Continuous 1999–2018 (50 XPT files across 10 cycles) was added under a different CDC infrastructure path. After age ≥ 25 and ELIGSTAT = 1 filtering, the combined panel has 705,610 subjects and 64,218 deaths over a median 102-month follow-up. Covariates include age, sex, race/ethnicity, survey year (fixed effects), smoking status, and BMI (from NHIS Sample Adult join and NHANES MEC exam, ~50% combined coverage). PM2.5, NO2, and O3 are included jointly in the headline multi-pollutant fit (Phase 7c default, 2026-04-30). The PM2.5 coefficient from this joint Cox is promoted to l3_hierarchical and consumed by all downstream investigations.

The hybrid structure: μ comes from a year-FE-adjusted single-pool Cox PH on the full panel (lifelines; identifies β from cross-region PM2.5 differences within survey years). τ comes from DerSimonian-Laird random-effects meta-analysis of per-region Cox βs run without year FE (five groups: four census regions + one NHANES national stratum). Full NumPyro NUTS hierarchical Cox was attempted and OOM-killed at 8 GB RAM; the hybrid is the documented honest fallback (Phase 6d Risk R4). The mathematical schema is identical to the NUTS output; downstream consumers read l3_hierarchical.mu_posterior_mean and mu_posterior_sigma and are unaffected by which computation path produced them.

L4 — E-value sensitivity (VanderWeele & Ding 2017). The E-value at the L3 posterior mean (HR = 1.28) is 1.87; at the lower 95 % CI bound (HR = 1.17) it is 1.61. An unmeasured confounder would need to be associated with both PM2.5 exposure and all-cause mortality at RR ≥ 1.61 to fully explain away the observed association — a strong standard.

Headline results

Posterior density plot comparing L1 Di/Krewski anchors, L2 BMA (uniform 50/50 Di/Krewski prior — equal-weight default in the absence of a defensible non-uniform prior; the L3 hierarchical posterior overrides the BMA result and is the production CRF), and L3 hierarchical Cox PH posterior for PM2.5 mortality CRF beta
Posterior density comparison across fidelity rungs. L1 Di (HR = 1.073, β = 0.00705) and L1 Krewski (HR = 1.056, β = 0.00545) are shown as point anchors. L2 BMA merges them at equal weight. L3 hierarchical posterior (μ = 0.02439, σ = 0.00447) is the real-cohort result; its position above the published anchors reflects the exposure-resolution gap (region-mean vs. ZIP-level).
Bar chart comparing single-pollutant HR 1.31 and multi-pollutant HR 1.28 for PM2.5 Cox PH
Single-pollutant (HR = 1.31) vs. multi-pollutant joint Cox (HR = 1.28, Δ = −8.9%) for PM2.5 per 10 μg/m³. The 9% attenuation from including NO2 and O3 is within the 5–15% multi-pollutant attenuation range reported by Roman et al. 2019 and Krewski et al. 2009. Multi-pollutant is the headline default (2026-04-30).
L3 hierarchical Cox PH results (multi-pollutant, 706K subjects)
QuantityValue
Panel size (subjects / deaths)705,610 / 64,218
Median follow-up102 months
μ posterior mean (β per μg/m³)0.024391
μ posterior σ0.004469
HR per 10 μg/m³ (posterior mean)1.2762
HR 95 % CI[1.1692, 1.3930]
τ posterior mean (between-region SD)0.1288
Shrinkage factor0.989
n groups (census regions + national)5
Single-pollutant HR/10μg (robustness)1.31 [1.21, 1.42]
PM2.5 Δ single → multi-pollutant−8.9%
L4 E-value (point)1.87
L4 E-value (CI bound)1.61

Policy CRF adoption: CARB-cited Krewski et al. (2009) (RR 1.06), not the real-cohort HR = 1.28

CARB implements Krewski 2009 in BenMAP for AB 617 and Cap-and-Invest health impact accounting. EPA’s 2023 TSD (which CARB explicitly follows) points toward Di 2017 as the direction of travel (HR = 1.073). We adopt the Krewski 2009 / Di 2017 range (RR 1.06–1.07) as the policy-comparability band for all visitor-facing health-impact estimates. The HR = 1.28 real-cohort fit is reported as a robustness check; the structural explanation for the gap is documented and auditable.

The 1.28 → 1.07 gap is the price of public-data access, not a claim of higher PM2.5 toxicity

Krewski 2009 used ZIP-level exposure on 1.2M ACS subjects. NHIS public-use ships only census region (4 groups). The well-documented Berkson-type attenuation of within-region exposure contrast biases the slope upward—this is the expected direction and the expected magnitude. Adding smoking + BMI covariates moved the NHIS-only HR from 1.43 to 1.31 (−8%); closing the remaining gap to Di’s 1.07 requires ZIP-level exposure assignment unavailable in the public NHIS. This is not residual confounding—it is a documented structural data limitation.

Downstream dependencies

The posterior from this investigation (μ = 0.02439, σ = 0.00447) feeds four downstream analyses: the health monetization and value-of-information framework (Investigation 6-1), the sequential portfolio decision model (Investigation M-2), the CRF research roadmap (Investigation 6-5), and the portfolio regret surface (Investigation 6-6). Two other investigations use the cohort directly for validation and robustness checks. All downstream reads use a sha256 drift check — any re-run here automatically flags stale inputs in every consumer.

Caveats