Four data-hygiene checks every engineer needs before citing SWITRS or CCRS crash clusters
Public California collision files contain 2,631,409 collisions and 5,478,771 party records between 2020-01-01 and 2026-10-27, but unmapped rows, street-inferred coordinates, directional ramp aliases, and 93,955 fragmented vehicle labels will distort naive SQL queries unless audited explicitly.
- Observation window
- Multi-year warehouse (2020-01-01 to 2026-10-27), 90-day Pulse & 30-day Map Summary
- Primary records
- CHP CCRS / SWITRS public records & Evidence for Zero standardization ledger
- Editorial date
- Published 2026-11-12 (Updated 2026-11-12)
Across 2,631,409 California collisions (2020-01-01 to 2026-10-27); 42.1% standardized by make-model rules.
Consolidating 93,955 raw make-model strings into 81,833 clean values without altering raw fields.
Collisions mapped via OpenStreetMap road-pair inference out of 74,533 90-day statewide records.
14,962 of 15,701 crashes (13,734 source GPS + 1,228 street-inferred; 739 unmappable).
Check 1 & 2: Separate agency GPS coordinates, street-inferred points, and unmapped denominators
Transportation engineers transitioning between UC Berkeley TIMS / legacy SWITRS extracts and live California Highway Patrol CCRS tables encounter an immediate spatial trap: not every valid collision report arrives with accepted California latitude and longitude coordinates, and coordinate completeness varies sharply by reporting agency and year.
In the latest 30-day statewide map summary (2026-09-28 to 2026-10-27), 15,701 collisions were recorded: 13,734 carried accepted source coordinates (87.5%), 1,228 were assigned estimated display coordinates from officer-recorded primary and secondary road names via OpenStreetMap (7.8%), and 739 could not be plotted (4.7%). Across the 90-day Pulse window (74,533 collisions), 15,061 records relied on OpenStreetMap street-intersection inference. In local 365-day municipal profiles, unmapped shares can be much larger: San Diego had 6,818 mappable crashes (81.0%, comprising 4,930 source GPS and 1,888 street-inferred) alongside 1,601 unmapped crashes (19.0%), while Anaheim was 78.7% mappable and Santa Ana was 88.9% mappable.
Two audit rules follow immediately. First, jurisdiction-level rates and year-over-year trends must always query the full reported collision denominator rather than counting only plotted map pins. Second, micro-scale engineering studies (such as sight-triangle or crosswalk setback redesigns) must distinguish officer-supplied GPS coordinates from intersection-centroid inferences so inferred pins are not mistaken for exact point-of-impact measurements.
| Clean vehicle model | Clean party records | Raw string variants | Largest single raw string | Records hidden by fragmentation |
|---|---|---|---|---|
| Honda Civic | 169,425 | 68 | 104,053 (HOND CIVIC) | 65,372 (38.6%) |
| Toyota Camry | 167,609 | 104 | 100,774 (TOYT CAMRY) | 66,835 (39.9%) |
| Honda Accord | 161,069 | 77 | 96,517 (HOND ACCORD) | 64,552 (40.1%) |
| Toyota Corolla | 131,149 | 112 | 80,650 (TOYT COROLLA) | 50,499 (38.5%) |
| Chevrolet Silverado | 85,647 | 72 | 52,680 (CHEV SILVERADO) | 32,967 (38.5%) |
| Toyota Tacoma | 80,232 | 58 | 50,682 (TOYT TACOMA) | 29,550 (36.8%) |
| Toyota Prius | 74,698 | 48 | 47,482 (TOYT PRIUS) | 27,216 (36.4%) |
| Ford F-150 | 72,345 | 44 | 46,717 (FORD F150) | 25,628 (35.4%) |
| Nissan Altima | 69,532 | 47 | 40,027 (NISS ALTIMA) | 29,505 (42.4%) |
| Tesla Model 3 | 35,131 | 80 | 7,463 (TESL 3) | 27,668 (78.8%) |
Check 3 & 4: Normalize directional road aliases and free-text party strings deterministically
Free-text field entry fragments both roadway corridor queries and vehicle-fleet involvement studies. On state highways, officers record the same physical corridor as I-405, RTE 405, SAN DIEGO FWY, or 405 S/B, and unincorporated county filings (758,730 collisions and 1,476,228 party records between 2020-01-01 and 2026-10-27) frequently mix county route numbers with local highway names.
The fragmentation is equally severe in party-level vehicle records. Across 5,478,771 party records in the 2020-01-01 to 2026-10-27 warehouse snapshot, raw officer entries produced 93,955 distinct make-model combinations. Applying 3,275 deterministic, auditable mapping rules (222 on vehicle_make, 68 on vehicle_make_v2, and 2,985 on vehicle_model, with zero rules mapping to null) standardized 2,306,545 party records (42.1%).
Without standardization, vehicle rankings invert whenever NCIC make codes change or officers split abbreviations. For example, Tesla Model 3 records arrive split across 80 raw strings including TESL 3 (7,463), TESL MODEL 3 (7,199), TSMR 3 (5,851), TSMR MODEL 3 (5,610), and TESLA MODEL 3 (4,258), hiding 78.8% (27,668 of 35,131 records) from a single-string query. Commercial Freightliner Cascadia tractors similarly fragment across 32 raw model spellings. Auditing collision clusters responsibly requires versioned, deterministic lookup tables that preserve raw strings side by side with canonical labels.
- Always inspect raw-versus-clean mapping tables before publishing vehicle or commercial-truck involvement comparisons.
- Lock reporting windows and record the warehouse snapshot version on every exported chart or CSV, because CHP CCRS continuously updates preliminary reports and deletes duplicates.
- State missing exposure denominators plainly: neither SWITRS nor CCRS contains registered-vehicle miles by model or intersection turning-movement counts.
Methodology boundaries and exposure limitations
- Street-inferred coordinates place unmapped road-pair records near the matched OpenStreetMap intersection or segment and must never be used to measure lane-offset or curb-line impact geometry.
- Standardization rules only consolidate unambiguous spelling and code variants; genuinely blank or unknown make-model fields remain unstandardized rather than imputed.
- Fault and primary collision factor (PCF) fields represent reporting-officer assessments at the time of filing, not adjudicated engineering or legal causation.
Review full source, coordinate-derivation, and report-revision rules on our Sources, Methodology, and Privacy page.
How to cite this brief
Evidence for Zero (Meru AI), "How Engineers Audit SWITRS & CCRS Crash Clusters," built from CHP CCRS public records (warehouse window 2020-01-01 to 2026-10-27; 90-day Pulse & 30-day Map Summary), https://crashes.meruai.co/insights/how-engineers-audit-switrs-ccrs-collision-clusters.
Inspect the underlying corridors and records
Inspect all 3,275 vehicle rules, raw variants, and jurisdiction cleanliness trends.
Read our constructive proposal for upstream field validation.
Review coordinate derivation, report deletion handling, and privacy controls.
Build source-attributed map embeds with reporting windows attached.