Methodology & Definitions

Definitions used in this methodology

Defined terms (§1–§7)

The SNRS Formula

SNRS is a single 0–10 score per carrier produced by a five-factor B-method weighted composition. Higher = riskier safety pattern.

Bands: LOW (< 2.5) / MODERATE (< 5.0) / ELEVATED (< 7.5) / HIGH RISK (≥ 7.5).
Smoothing: exponential moving average with α = 0.40 against the previous month's smoothed score. Half-life of crash-event influence ≈ 730 days (independent of EMA smoothing — applied at the rate-input level inside the Crash factor).
Exposure: power-unit-days (active days × power_units, capped at 1,825 days). We do not use self-reported VMT from MCS-150 because it has obvious manipulation incentive after a fatal incident.

Bayesian Shrinkage

Small carriers with a single catastrophic event would otherwise dominate their cohort's percentile rank. A one-truck owner-operator with one fatal crash has an observed crash rate of 555 events per million power-unit-days — three orders of magnitude above any real-world fleet average. That's noise, not signal.

We apply Bayesian shrinkage: pull each carrier's rate toward the cohort mean using a conjugate prior with a strength parameter and an exposure-floor guard for very small fleets.

Worked example: a one-truck carrier with one fatal crash and 1,825 days of activity.

Calibration source: empirical sanity check on the May 2026 FMCSA snapshot found 26.9% of 1-truck carriers with one fatal crash were scoring SNRS ≥ 7.0 at prior_strength = 5. Tightening to prior_strength = 10 plus the exposure floor dropped that false-positive rate to 0.74% — below our 2% target. See Reproducibility for the exact SQL.

The SSR Formula

SSR is a statistical settlement-range estimate, not a verdict prediction for any specific case. We model the distribution of comparable-incident outcomes as log-normal per (state, severity) with a per-carrier negligence shift.

National severity baselines (log-space μ for the median comparable verdict before state/carrier adjustment): FATAL μ = 14.7323 (≈ $2.5M), SEVERE_INJURY μ = 12.8992 (≈ $400k), INJURY μ = 11.2898 (≈ $80k), PROPERTY_DAMAGE μ = 9.3927 (≈ $12k). σ ranges 0.65–0.95 by severity.

State tier table

Tier classification uses each state's shift relative to the national baseline for the same severity. Midpoint thresholds for the four tiers are 0.225 / 0.075 / −0.10.

States not listed (DC, territories) fall back to the national baseline (Tier C equivalent) inside compute_ssr_interactive.

Cohort Definitions

Every percentile rank and every Bayesian prior in the model partitions the carrier population by 5 × 4 = 20 cohort buckets: fleet-size × safety-event-count.

Both axes are used in two places: (a) the cohort partition for SMS BASIC percentile rank (since FMCSA's public files publish raw MEASURE values only — see Limitations), and (b) the prior mean for Bayesian crash-rate shrinkage. Carriers in the 0-events bucket receive cohort-baseline percentiles regardless of their MEASURE because there's no observable behavior to rank.

Data Sources

All inputs are public FMCSA records. Snapshot timestamps below auto-populate from the ingestion audit log; each refresh updates the corresponding row.

Bulk FMCSA Source SMS Data files refresh monthly. The L&I bulk file (ActPendInsur) refreshes daily on catalog.data.gov. We re-ingest on the documented cadence; new monthly snapshots are loaded within ~7 days of FMCSA publication.

Night-Driving Pattern Detection

Each carrier's 5-year crash record is classified by Light_Condition_Desc and tested against the national dark-condition crash share. The page surfaces the share and a one-sample binomial p-value when the carrier has ≥5 documented crashes in the window; below 5, the section reports insufficient sample.

Data source

FMCSA SMS Crash File 2026-04 Snapshot, field Light_Condition_Desc. Approximately 259k national records; 99.6% non-null on the field. Source values fall into eight categories: Daylight, Dawn, Dusk, Dark - Lighted, Dark - Not Lighted, Dark - Unknown Roadway Lighting, Unknown, Other.

Night definition

Night aggregates the three Dark-* categories: Dark - Not Lighted, Dark - Lighted, Dark - Unknown Roadway Lighting. National combined share: 25.7% (the binomial baseline used in the per-carrier test).

Excluded from night: Dawn (2.4% national) and Dusk (1.4%). These transition periods carry separately elevated risk in the NHTSA Large Truck Crash Causation Study literature, but the underlying mechanisms (glare, reduced contrast during a brightness transition) differ from sustained-darkness fatigue and are not collapsed into the same metric.

Baseline

The 25.7% baseline is computed directly from the source distribution in the same dataset we use for per-carrier measurement. The dataset is both the measurement source and the comparison reference — internal consistency by construction. No external estimate is imported.

Significance test

One-sample binomial against the 25.7% baseline, two-sided. The p-value is computed via normal approximation with Postgres's erf() (Abramowitz & Stegun erf-form of the standard normal CDF). Sufficient fidelity for decision-support; not a publication-grade significance test, particularly for the smallest qualifying samples.

Sample-size gate

Carriers with fewer than 5 documented crashes in the 5-year window get an insufficient sample message instead of a share + p-value pair. The 5-crash floor is conservative — at smaller sample sizes the normal approximation degrades and an apparent elevation is more likely noise than signal.

Regulatory connection

49 CFR § 392.14 requires extreme caution when hazardous conditions (including reduced visibility) adversely affect visibility or traction. The federal Hours-of- Service regime at 49 CFR § 395 limits consecutive driving hours and mandates rest periods; sustained-darkness over-representation in a carrier's crash record can correlate with fatigue-window operations.

Limitations

Long-haul carriers may have legitimately higher dark-driving exposure because their operations type concentrates over overnight hauls. The exposure-type confound is not corrected for here. The page displays deviation and significance together precisely so the reader can judge whether elevation reflects an operational pattern or a safety pattern — this is an investigative signal, not a fault determination.

Self-Insurance Detection (BMC-91X)

FMCSA permits large motor carriers to self-insure their bodily injury & property damage (BIPD) exposure rather than carry a third-party policy. The regulatory hook is 49 CFR § 387.309 (Self-insurance authorization for motor carriers and freight forwarders). Self-insurers file via the BMC-91X form and must demonstrate financial capacity to FMCSA on an ongoing basis.

Why the public dataset looks like "$0 coverage"

In the FMCSA L&I bulk feed, self-insured BIPD filings appear with a third-party liability_limit of $0. The carrier is not without coverage — they have regulatory authorization to satisfy claims from their own balance sheet — but the bond amount underwriting that authorization is not exposed in the public dataset.

Detection rule

A carrier is presumed self-insured when all three conditions hold against the most recent ingest:

1. Fleet size power_units ≥ 100 — floor that distinguishes BMC-91X self-insurers from the residual multi-subsidiary filing pattern.
2. An active BIPD filing on record (coverage_type starts with BIPD, is_active = true).
3. The filing's liability_limit is $0 or NULL.

The rule is implemented as a Postgres function public.apply_self_insurance_detection() invoked after each monthly recompute so the classification is refreshed alongside SNRS / SSR.

What changes on the carrier page

• confidence_level = 4 (the regulatory authorization is itself a high-confidence signal).
• underinsured_flag = FALSE — self-insurance is not under-coverage.
• Coverage Gap is not computed; the public dataset does not expose the bond amount underwriting the authorization.
• Insurance section label reads Self-insured (BMC-91X, presumed).

Sample size + limitations

Approximately 92 carriers match the detection rule on the current snapshot, including FedEx Corp, Walmart Transportation, Old Dominion, Schneider National, Estes Express, XPO Logistics, Penske Logistics, and the FedEx Freight Inc operating subsidiary. The count refreshes on each ingest cycle as filings change.

The (presumed) qualifier is intentional: we infer the regulatory status from the data pattern (zero-limit BIPD on a large fleet), not from a direct FMCSA self-insurance roster call. A small residual set of multi-subsidiary carriers (parent corporation files BIPD under a different USDOT) may still produce the zero-limit pattern at fleet sizes below the 100-PU floor; those carriers stay outside the self-insured set and continue to surface the multi-subsidiary narrative from v6.4.1.

Attorney Value Score

SafeNY ranks carriers by Attorney Value, a composite score reflecting the expected utility of a personal-injury attorney engaging with a given carrier's record. The score drives the top-1,500 carrier pages pre-rendered at build time and the top-70,000 admitted to the sitemap.

Formula

The foundation term $F(c, t)$ is the recency-aware fatal/severe-injury volume, weighted so that recent incidents count for more than historical ones:

$\tau$ is days since the crash; $f_i$ is the fatality count of crash $i$. The 180-day half-life keeps an event "visible" in the ranking for roughly a year, and the 1.5× boost on the first 60 days surfaces freshly-relevant signals.

Multipliers

Justification of weights

The multiplier ranges are heuristic, not empirically calibrated. The reasoning is consistent across the six multipliers:

• $M_{\text{verdict}}$ uses a $5M denominator so a national-median P90 fatal verdict produces a 2× multiplier, and a top-decile $25M nuclear verdict produces a 6× ceiling. SSR LogNormal P90 already underestimates the tail; the multiplier compounds intentionally.
• $M_{\text{state}}$ mirrors the four-tier classification in the SSR formula above, scaled so a Tier A carrier is roughly 2.5× a Tier D carrier of identical record — consistent with the ATRI Nuclear Verdicts report's state-shift estimates.
• $M_{\text{recovery}}$ rewards the $2–30M sweet spot most heavily because below $2M the case rarely justifies an attorney's contingency arrangement and above $60M the chance of full recovery from a single defendant collapses (multi-defendant strategy required).
• $M_{\text{differentiation}}$ favours small fleets because Phase 0 empirical analysis showed small carriers (7–50 power units) have the highest HIGH_RISK rate (3.2%) and the lowest aggregator competition — SafeNY's marginal information value is largest there.

Sensitivity

Top-50 ranking retains 96–100% overlap when individual multipliers are perturbed by ±10–25% (recovery-band edges, small-fleet boost, narrative thresholds). The model is not load-bearing on any single weight choice.

Limitations

• No outcome data yet — the multipliers are heuristic. Phase 2 will re-calibrate against settled-case outcomes once ≥100 are collected from the attorney-network pilot.
• The state-tier multiplier assumes within-state behavioural homogeneity. Sub-state venue variance (e.g. Los Angeles County vs. rural Northern California within Tier A) is not modelled.
• Self-insured carriers are down-weighted via $M_{\text{recovery}} = 0.5$ but not excluded — a self-insurer with a sustained recent fatal pattern can still rank highly (e.g. FedEx Corp at top-10).
• The freshness term $R(\tau)$ assumes uniform decay across crash severities. A more nuanced model would let multi-fatal incidents decay more slowly — left for Phase 2.

Reproducibility

An independent reviewer with public FMCSA data can recompute every published SNRS and SSR within rounding error. The pipeline is open in structure; the only proprietary pieces are operational (database choices, refresh cadence). All formulas above are implemented directly in the steps below.

1. Download FMCSA Source SMS for the target month from ai.fmcsa.dot.gov/SMS/Tools/Downloads.aspx. Six files: Motor Carrier Census, Crash, Inspection, Violation, SMS_AB_PassProperty, SMS_C_PassProperty.
2. Download L&I active filings actpendins_allwithhistory.txt from catalog.data.gov.
3. Compute SMS percentile ranks per cohort. Public SMS files publish raw MEASURE values; we apply PERCENT_RANK partitioned by (fleet_cohort, safety_event_count_bucket).
4. Compute Crash Factor: for each carrier, weight crashes by severity (0.50 fatal + 0.30 injury + 0.20 towaway) and recency (exp(-ln(2) × days_since_crash / 730)). Apply Bayesian shrinkage with prior_strength = 10 and exposure_floor = 0.005. Rank within fleet cohort.
5. Compose SNRS per the B-method formula. Smooth against the prior month's value with α = 0.40.
6. Compute SSR: for each (state, severity), shift the national baseline μ by the tier shift (table above), then add 0.11 × negligence_factor. Quantiles via inverse-CDF on LogNormal(μ, σ).

Reference SQL for the SNRS computation is in our public migration files (supabase/migrations/*_snrs_*.sql and *_bayesian_shrinkage_calibration.sql). The staged computation script pipeline/scripts/recompute_staged.py implements the full pipeline in 13 deterministic stages.

Disclaimers

• Not legal advice. SNRS is a decision-support score for case selection. SSR is a statistical estimate of comparable-case outcome ranges. Neither is, or substitutes for, legal advice. Consult a licensed attorney for case-specific evaluation.
• Statistical estimate only. P10 / P50 / P90 quantiles describe a distribution; individual case outcomes vary widely with facts, jurisdiction, evidence admissibility, and counsel quality. We do not predict any specific case's verdict.
• Not affiliated with the State of New York or any government agency. SafeNY.com is an independent, privately-operated research resource. Our data is sourced from public US Department of Transportation (FMCSA) records.