Crash reports per million driverless miles
Per mile driven, how often does this operator file a crash report?
Reporting obligations behind it
2
The filings a company is required to make that this figure is built from
Known biases ledgered
4
Each carrying the direction it pushes the figure, rather than a warning that data may be imperfect
Of those, unsigned
1
Entries where the record does not say which way the error runs
Formula
Unit: reports / million miles
reports(operator, months, state ∈ divisor.jurisdiction) ÷ (miles(operator, months, divisor) ÷ 1,000,000)
What it is built on
- Federal crash reporting for automated driving systems, under Standing General Order 2021-01
- California quarterly driverless mileage reporting
Bias ledger
Each known problem with this number, and which way it pushes it. A signed direction is more useful than a warning that data may be imperfect.
| Effect | ||
|---|---|---|
| Self-reported numerator | direction unknown | Operators file their own crash reports and NHTSA does not verify them before publication. Reporting practice differs sharply between operators — see the severity mix metric. |
| Reports, not crashes | overstates | One collision can generate more than one report, and a report is filed on an allegation. The numerator counts filings. |
| Jurisdiction clip | understates | The numerator is restricted to California because the divisor is. An operator's crashes elsewhere are excluded, so this is a California rate, never a national one. |
| Filing lag | understates | Recent months are still accruing. The two most recent months are hatched everywhere they appear and will rise. The mileage side lags further than the crash side, so the newest quarter with a rate is never the newest quarter with reports. |
| No rows match this filter. | ||
How it is computed
Every step that changes the answer, in the order it happens. Nothing here is a summary of a longer method: this is the method.
- Take every crash report the operator filed for the months in the period.
- Keep only the latest version of each report, because a report is amended rather than replaced, and drop the filings that add no new information.
- Drop any report from outside the states the divisor covers, so the count and the mileage describe the same territory. This is the step almost every published rate skips.
- Divide by the mileage the same operator filed for the same months under the same divisor, in millions.
Not for
Ranking operators against each other unless the comparability test passes. Different operators file under different thresholds, cover different road types, and drive in different places.