Normalizing Crash Risk of Partially Automated Vehicles under Sparse Data

Goodall, Noah J. · 2023 · Crossref

DOI: 10.31224/osf.io/m8j6g

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of evaluating the safety of partially automated vehicles (PAVs) given the lack of standardized crash reporting and exposure metrics. Manufacturers often report crash rates using proprietary definitions and limited data, making it difficult to compare PAV safety against baseline human-driven statistics. The research aims to establish a methodology for normalizing these crash rates by controlling for confounding variables, specifically road type and driver demographics, using Tesla’s quarterly safety reports as a case study. The methodology utilizes Tesla’s self-published crash data from July 2018 through March 2021, which reports average miles between crashes for vehicles using Autopilot (SAE Level 2) versus those using only active safety features (manual control). To normalize this data, the author controls for road usage using mileage ratios from a naturalistic driving study of Tesla vehicles, which indicated that 93% of Autopilot usage occurred on freeways compared to only 30% for manual driving. Driver age demographics were controlled using ownership survey data, which showed Tesla owners are predominantly aged 50–70, a lower-risk group compared to the general population. Crash severity thresholds were aligned with the Strategic Highway Research Program 2 Naturalistic Driving Study (SHRP 2 NDS), focusing on severe and police-reportable crashes. Statistical adjustments were applied to estimate what the crash rates would be if Autopilot and manual driving shared identical exposure profiles. The results demonstrate that unadjusted data significantly overstates the safety benefits of automation. While Tesla’s reports claimed Autopilot had a 43% lower crash rate than manual driving with active safety features, this advantage reduced to just 10% after controlling for the higher proportion of freeway driving associated with Autopilot. Freeway driving carries a lower crash risk per mile than non-freeway driving; thus, the apparent safety gain was largely attributable to where the system was used rather than the system itself. Additionally, controlling for the older, lower-risk demographic of Tesla owners increased the estimated crash rates by approximately 11%. Direct comparison with general public crash rates remained impossible due to discrepancies in crash severity definitions between the manufacturer and national datasets. The study concludes that raw manufacturer crash statistics are misleading without normalization for environmental and demographic factors. The findings highlight the critical need for regulators and manufacturers to provide detailed, standardized data on crash definitions, exposure conditions, and driver demographics. Accurate safety assessment requires isolating the performance of the automated system from the biases introduced by its operational design domain and user profile. This methodology offers a framework for future analyses to more rigorously evaluate the true safety impact of automated driving technologies.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• incidence prevalence
• naturalistic crash near crash
• induced exposure
• telematics crash prediction
• sex gender
• adas effectiveness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes, observational prevalence
• Methodological Resource: dataset resource