Rate & confidence intervals

Every compute response reports a rate with a 95% confidence interval, plus the numerator and denominator it was built from. This page explains how those numbers are derived.

The rate

rate = N / (D_miles / 1_000_000)

• N: the weighted count of crashes matching your selections (outcome, vehicle class, road type, weather, light, crash year, …). If under-reporting adjustment is on, Blincoe multipliers are applied here.
• D_miles: calibrated vehicle-miles travelled for the selected road groups and vehicle classes. D_billions is the same value in billions.
• rate: incidents per million miles (IPMM).

Numerator details

• Outcome is a nested filter: each outcome (observed_any_injury, airbag, ka, fatal, …) is a subset of police_reported.
• Under-reporting: the adjusted option applies Blincoe et al. 2023 (NHTSA) multipliers: PDO ×2.48, non-fatal injury ×1.47, fatal ×1.0.
• Crash year: geofence cities carry a crash_year column. Selecting multiple years pools their crashes and scales the denominator by the number of years, so the rate stays an annual mean.

Denominator details

• Calibrated VMT (default) blends state-DOT roadway-inventory totals with FHWA VM-4 per-class fractions.
• denominator_vmt (California only) toggles between the Caltrans CPRD total and an HPMS functional-class total.
• multiplier_vmt selects the exposure source inside the operator-weighted multiplier (only relevant when robo_taxi_weighting=operator_weighted).

Dynamic operator weighting (geofence)

When robo_taxi_weighting=operator_weighted, the rate is paired with a Chen 2024 spatial multiplier that re-weights county exposure toward where an ADS operator actually drives (using Waymo rider-only mileage). It only applies when every selected road type is a surface street. rate_non_dyn is always returned for reference; rate_dyn and multiplier are null when the operator mileage layer is not loaded.

Confidence intervals

Routes and depots return a ci_method choice:

• Fay-Feuer (default): a weighted-Poisson interval with a phantom-event correction for zero-crash legs.
• Empirical Bayes: a Gamma-Poisson interval using between-cell rate variance.

Geofence intervals use the same weighted-Poisson machinery. rate_low / rate_high are the 95% bounds; route responses also expose the propagated variance.

Data & licensing

Outputs are derived statistics only (rates, counts, bounds); the API never redistributes raw third-party crash records. Crash data comes from public state sources (TxDOT CRIS, CHP/CCRS SWITRS, ADOT ALISS, Nevada DOT); exposure from FHWA HPMS and state DOT inventories; weather from NOAA ISD; ADS exposure from the Waymo Dynamic Benchmark.

Litigation caveat

Safety data compiled for federal highway-safety programs is shielded under 23 U.S.C. § 409 in many states. Do not market these outputs as litigation evidence, and do not use crash data to solicit involved parties.

Attribute every published figure. Full per-source citations live in the project's data license register.