California Freight Cleanup → Investigation 6-5
Which CRF studies, in what order, are worth funding next?
666× ROI — meta-analysis ($0.5M); $4.9B net — L3 staged ($7.5M total)The earlier CRF investigation settled which concentration-response function to use. This one answers what comes next: which research studies, in what order, are worth funding to sharpen future decisions on sub-state allocations and post-2030 portfolio refinements.
Decision context
The CRF analysis (Investigation 21) quantified $0.25B in residual information value still available from further concentration-response research — but also showed the Phase 1 $4B portfolio choice is already clear: the all-in option wins every scenario. So why fund more CRF research? Because the program doesn’t stop at Phase 1. Sub-state allocations, AB 617 community-level work, and post-2030 refinements all face decisions where a tighter CRF estimate will matter. This investigation computes the ROI on each candidate study against those future decisions and recommends the cheapest sequence that captures most of the value.
Methodology: five research designs evaluated
All five rungs read the Investigation 6-3 posterior parameters live via
upstream_value("21", ...) (drift-detected). The $0.25B
residual EVSI from Investigation 6-3 is the upper bound on what a single
one-shot study could deliver against the current portfolio decision.
The L3 staged pathway ($4.87B) exceeds this ceiling because each sequential
stage tightens the posterior, generating compounding EVSI against the broader
future decision landscape (Q3 sub-programs, AB 617 sub-state work); the
$0.25B ceiling applies only to the current Phase 1 decision.
L1 — Point-estimate EVSI ceiling. The $0.25B Investigation 6-3 residual EVSI is the theoretical perfect-information maximum (cost = $0). Reported for reference; excluded from best-level ranking (comparing a costless ceiling against cost-bearing L2–L5 designs is a tautology).
L2 — Per-design pre-posterior EVSI. Five candidate designs evaluated independently (meta-analysis, retrospective cohort, Medicare extension, prospective CA, pooled consortium). For each: EVSI = p_success × shrinkage × residual_EVSI. Net value subtracts cost. ROI = net/cost.
L3 — Staged sequential pathway. Studies run in cost-ascending order; each stage consumes the posterior σ left by the prior stage. Tests whether cheap-first ordering captures most value at a fraction of the full-portfolio cost.
L4 — Multi-arm bandit (Gittins-index greedy, $25M / 5yr). Arms selected greedily by expected information yield per dollar. Four arms selected; total cost $22.5M.
L5 — POMDP adaptive portfolio (500 MC rollouts, greedy policy). Sequential learning with early-stop option. Greedy heuristic lower-bounds true adaptive value. Mean trajectory: 1.94 arms, $6.3M cost.
Rungs fused via precision-weighted Bayesian pooling to produce a summary posterior (mean $2.92B, σ $1.53B).
Headline results
| Design / Pathway | Cost | EVSI ($B) | Net ($B) | ROI |
|---|---|---|---|---|
| Meta-analysis (L2) | $0.5M | 0.333 | 0.332 | 666× |
| Retrospective cohort (L2) | $2.0M | 0.678 | 0.676 | 339× |
| Medicare extension (L2) | $5.0M | 0.152 | 0.147 | 30.5× |
| Prospective CA cohort (L2) | $15.0M | 0.000 | −0.015 | 0× |
| Pooled consortium (L2) | $25.0M | 0.625 | 0.600 | 25× |
| L3 staged (meta→retrospective→Medicare ext.) | $7.5M | 4.870 | 4.862 | — |
| L4 bandit (4 arms, $22.5M) | $22.5M | 3.680 | 3.658 | — |
| L5 POMDP (mean 1.94 arms, $6.3M) | $6.3M | 2.231 | 2.225 | — |
Run the meta-analysis first, gate the retrospective cohort on the results — $4.86B net
L3 staged beats L4 bandit by 33% and L5 POMDP by 118% in expected net value. The information surface is smooth enough that a 2-rung staged pathway captures it. Run the meta-analysis first ($0.5M), gate the retrospective cohort ($2.0M) on its results—that is the dominant strategy. L4 and L5 add procurement complexity without proportional value.
This research is for future decisions — the Phase 1 choice is already settled
Investigation 6-6 confirms D_all_in_4B as uniquely optimal at P(optimal) = 1.000: this staged research does not change that recommendation. We recommend it because the Q3 $300M scale—Investigation 6-8, AB 617 air-basin sub-state work, and post-2030 portfolio refinements—will face CRF precision requirements where the Di/Krewski gap is decision-controlling. A $7.5M staged investment delivers $4.9B in expected future-decision value.
Caveats
- Shrinkage fractions are estimates, not measurements. The fraction of posterior variance eliminated by each design is drawn from cohort-design literature. Actual shrinkage depends on achieved sample size and exposure-assessment quality — unknown ex ante. If a study underperforms, downstream EVSI is proportionally overstated.
- p_success is a subjective probability. It encodes the analyst’s judgment that a study will deliver a usable CRF estimate. Meta-analysis (p = 0.85) has strong precedent; prospective CA cohort (p = 0.70) faces enrollment and timeline risk.
- The $4.86B net is the optimistic upper bound. Real meta-analyses suffer from publication bias; retrospective cohorts from confounding by indication. A conservative estimate is $2.5–$3.5B — still a strong ROI on the full $7.5M staged-pathway spend.
- Stale upstream from Investigation 6-3 re-run. The 2026-05-02 multi-pollutant update changed βPM2.5 from 0.02677 to 0.02439 (−8.9%); this shifted all L2–L5 EVSIs by −59–−61% proportionally. The current numbers reflect the updated posterior.