← crashstats.com.au

Vehicle Fatality Trends

In a two-vehicle injury crash, how much does the class of the other vehicle change what happens to the car driver? Every figure on this page has a written estimand, a stated sampling frame, and a query behind it.

Victoria · 2012 – Oct 2025DTP Victoria Road Crash Data (CC-BY-4.0)Severity given injury — never risk“Involved in”, never “caused”AI-built · human-reviewed
Injury crashes analysed
198,839
police-reported, all severities
Deaths
3,564
person records, injury level 1
Vehicle records
362,687
linked to crashes
Car drivers analysed
81,735
two-vehicle, both driven

What the other vehicle does to a car driver

Victorian two-vehicle injury crashes, 2012 – Oct 2025, where both vehicles were being driven. The bar is the share of car drivers killed or hospitalised. Whiskers are 95% Wilson intervals.

Another car12.7%SUV / station wagon16.1%Ute / light commercial21.5%Truck35.4%
Estimand. P(car driver injured to level 1 or 2 | two-vehicle injury crash, both vehicles driven, opponent class t). Unit: the driver, not the vehicle involvement. Class: severity given injury. Victoria records injury crashes only, so this is not a risk of crashing and cannot be made into one.

Read the car–car row carefully. In a car–car crash both drivers are car drivers, so each such crash contributes two units to that row: 48,398 index drivers from 24,199 distinct crashes. Every other row is one driver per crash. The percentages are per driver and are comparable across rows; the counts are not, and n here means drivers, never crashes.

associational A car driver is killed or hospitalised in 12.7% of collisions with another car, 16.1% against an SUV or station wagon, 21.5% against a ute or light commercial, and 35.4% against a truck. Under this pipeline the four Wilson intervals do not overlap. Disjoint sampling intervals do not imply the remaining gaps would survive adjustment for speed zone, road class or reporting threshold.

descriptive In 2,873 coded car–truck injury crashes with both vehicles driven, 100 car drivers and 3 truck drivers were recorded killed. That is a raw driver-fatality count under this filter. It is not a causal responsibility split and not a model-level vulnerability rating.

Conditional on an injury crash, so this is not population risk. Not adjusted for speed zone or road class, and speed-zone exposure rises with opponent class — see the measured breakdown below. Vehicle class is a body-type proxy, not measured collision mass. This is not a per-model safety rating. Rating how well a model protects its own occupants, or how much harm it does to the other party, needs a risk denominator this dataset does not contain. MUARC Used Car Safety Ratings remain the reference.

The evidence that carries the claim: it holds in all four configurations

A single crash configuration can be explained by geometry. Four cannot. These are crash-code bundles, not instrumented collision types.

How the bundles are built. Head-on is DCA 120 (head-on, not overtaking) only — DCA 150, head-on while overtaking, is reported separately below because pooling them would hide a reversal. The adjacent bundle pools DCAs 110, 111, 112, 113, 114, 116, 119, 121 and 123; its three largest components are shown separately below so you can see whether the bundle is hiding heterogeneity. Rear-end rows are DCAs 130, 131 and 132 only — same-direction rear-end codes, not all rear-ish contact — split by the car's coded initial impact point, front or rear.

Every row requires both vehicles to have a driver person record. Parked and unattended opponents are excluded (1,863 car, 769 SUV/wagon, 614 ute, 149 truck pairs). This filter applies to every percentage on this page and is the reason our figures are reproducible.

Not every crash is bundled. Driveway emergence, U-turns, lane changes and sideswipes, leaving parking, striking a parked vehicle, running off the carriageway, reversing and pulling out while overtaking carry no bundle. They are 19.7% of the pairwise pool (16,061 of 81,735) and are not in this table.
Crash-code bundleAnother carSUV / station wagonUte / light commercialTruck
Head-on (DCA 120)31.8%
29.9–33.8 · n=2,230
38.7%
35.9–41.7 · n=1,092
44.8%
41.4–48.3 · n=781
75.0% ⚠
68.9–80.2 · n=224
Side/intersection14.4%
13.9–14.8 · n=23,262
17.7%
16.9–18.5 · n=9,271
25.6%
24.3–26.9 · n=4,452
45.1%
41.1–49.1 · n=597
Car ran into their rear6.8%
6.2–7.4 · n=7,107
10.2%
9.2–11.3 · n=3,196
16.6%
14.5–18.8 · n=1,159
53.4% ⚠
45.7–60.9 · n=163
Car struck from behind7.7%
7.1–8.4 · n=6,844
9.5%
8.4–10.7 · n=2,537
10.3%
9.1–11.7 · n=2,147
17.8%
14.4–21.7 · n=422

Cells show the percentage, the 95%% Wilson interval and n. ⚠ marks cells with fewer than 300 car drivers: the truck bundles are small (head-on n=224, car-into-rear n=163) and those percentages are unstable relative to the car cells.

descriptive The ordering is monotone in every bundle we defined, with no sign flip. The gradient is largest in the head-on group and smallest when the car is struck from behind. That pattern is consistent with a mass or geometry contribution; it does not identify one. Vehicle class is a body-type proxy. Mass, structural stiffness and impact speed are not measured in this table, and neither is speed zone or road class.

The confounder we can measure, measured. We used to assert that trucks concentrate on high-speed routes. Here it is instead: of the car–car crashes in this frame with a recorded speed zone, 25.9% were in an 80 km/h or higher zone. For car–SUV/wagon it is 28.9%, car–ute/LCV 37.0%, and car–truck 60.6%. Speed-zone exposure rises across exactly the same ordering as the outcome, so part of the gradient above is road environment rather than the other vehicle. This table does not separate the two, and nothing on this page should be read as if it did.
The one exception, stated rather than buried. Head-on while overtaking (DCA 150) is not monotone: Another car 26.9% (n=108) · SUV / station wagon 26.2% (n=42) · Ute / light commercial 33.3% (n=33) · Truck 57.1% (n=7). Cars and SUV/wagons invert, and the truck cell rests on seven drivers. We report it separately instead of pooling it into head-on, because pooling would have concealed the only reversal in the analysis. On these sample sizes the inversion is well inside noise, but a reversal disclosed is worth more than a monotone table built by choosing the bundle that produced one.

Is the adjacent bundle hiding heterogeneity?

The adjacent set pools nine DCA codes with different geometry, so it is the most likely place for a pooled result to fray. Its three largest components, shown separately:

Largest adjacent componentsAnother carSUV / station wagonUte / light commercialTruck
Cross traffic (110)13.6%17.4%26.4%48.3%
Right near (113)16.7%19.3%25.9%50.3%
Right through (121)14.5%18.4%26.1%38.2%

All three are monotone on their own, so the bundle is not concealing a reversal here.

Second split: geography

AreaAnother carSUV / station wagonUte / light commercialTruck
Melbourne11.8%14.3%18.6%30.2%
Rest of state15.2%20.2%25.9%44.6%

Monotone in both, and the geography split carries the same caveat as the bundles: it is not adjusted for speed zone or road class, and the non-metropolitan rows carry more high-speed-zone exposure.

On the rear-end split. We split on the car's coded initial impact point rather than the DCA vehicle-slot code. Impact point is still an officer-coded field, not a measured crash pulse; we use it because it tracks which vehicle struck which more directly than the slot does when the two disagree. In rear-end crashes the two agree closely (slot 1: 31,883 front against 158 rear; slot 2: 30,801 rear against 354 front) but not perfectly, and the disagreements are exactly where a slot-only analysis would mislabel the role. This is a construct tightening for role. It is not a measurement of mass, and it does not make anything else on this page identified.

Vehicle age, within driver age band

Share of light-vehicle drivers killed or hospitalised, by the age of the vehicle, split by driver age band — because pooling the two produces a bigger number than either band supports.

07142027340-410-1420+drv 16-25drv 26-59drv 60+
Vehicle agedriver 16-25driver 26-59driver 60+
0-4 years11.96%10.45%19.66%
5-9 years12.64%11.55%21.75%
10-14 years14.97%13.92%24.13%
15-19 years16.88%16.83%26.48%
20+ years19.12%20.29%28.99%
Estimand. P(driver injured to level 1 or 2 | injury crash, light vehicle, vehicle age band, driver age band). Vehicle age is crash year minus year of manufacture; the three are never entered together.

associational Crashes of older vehicles are more often severe, and the gradient holds within every driver age band — so it is not simply an artefact of who drives old cars.

descriptive Correction to our July 2026 report. That page said a 20-year-old vehicle's crash was “more than twice as likely to be fatal” as a near-new one's. Within driver age bands the ratio is 1.6×, 1.94× and 1.47×. The pooled figure was inflated by driver mix and is withdrawn.

Drivers with no recorded age band are excluded: their rates (2.5–8.3%) index record incompleteness, not survival. The level shift for drivers aged 60 and over reflects occupant frailty and is not a statement about the vehicle.

What is in Victoria's fatal crashes, by body type

Share of light-vehicle involvements in fatal crashes, by the crash dataset's own vehicle-type field. A composition share, not a rate.

014284256702012201420162018202020222024CarWagonUtilityVan/LCV

descriptive Between 2012 and 2024 the car share fell from 57.1% to 44.2% while station wagons rose from 21.6% to 29.8%, utilities from 15.2% to 17.9% and vans/LCVs from 6.1% to 8.2%.

Why there is no SUV line here. This dataset has no SUV category. Our July 2026 page identified SUVs with a fixed 34-nameplate list. Tested against the dataset's own vehicle-type field, only 34,709 of the 74,820 vehicles the dataset codes as station wagons were on that list — fewer than half, while 2,546 listed models were coded as cars, 664 as utilities, 219 as light commercials and 81 as panel vans. The two constructs disagree in both directions and the list has no external validation, so it is retired and no SUV-specific trend is published. “Station wagon” here contains both wagons and SUVs.

Who is dying on Victorian roads

Deaths per year by road user type. A count, not a rate — no exposure denominator is available and none is implied.

035701051401752012201420162018202020222024DriversPassengersMotorcyclistsPedestriansBicyclists

descriptive Motorcyclist deaths rose from 38 in 2012 to 61 in 2024. Driver deaths fell from 146 to 130. Pedestrian deaths moved from 35 to 47.

What we withdrew from the July 2026 version, and why

This page previously carried brand and model comparisons. They are gone. Each had a specific defect that a caveat could not repair.

If you used the withdrawn figures to compare brands or models, please stop. They are not reliable for that use.

Brand fatal-crash involvement counts

Cumulative counts are fleet share, exposure and era mixed together. The chart's own caption bridged readers into the severity table below it, which compounded the problem rather than qualifying it.

Brand severity rate — “when it crashes, how often does someone die?”

The share of a brand's involvements that were fatal is a composite quantity. It aggregates single-vehicle, multi-vehicle, multi-occupant and pedestrian events, which are not exchangeable crash roles across brands. The 95% intervals we published quantified sampling error around a number whose dominant error was structural, which made it look more precise than it was.

Fatal involvements per 10,000 registered vehicles

Not reproducible. No full Victorian registration denominator is joined into our pipeline, so the two-decimal figures could not be regenerated from the cited source. It was published as an “indicative band”, but a band you cannot reproduce is not a band.

The vulnerability asymmetry scatter

Held pending a written death-attribution specification and sensitivity checks under alternate attribution rules, because attribution choices can move a model across the diagonal. No claim from that figure is in force on this page.

The mass–severity ladder (per model)

Superseded by the section at the top of this page. A pooled per-model fatality rate mixes single-vehicle, multi-vehicle and pedestrian events, so it could not identify the mass claim it was making

Ask the data yourself

The assistant answers only from this report's tables: it drafts read-only SQL and a deterministic verifier checks every number against the query results before you see it. If it can't verify, it refuses. Use the button at the bottom-right, or start from a question:

Scope note, 2026-08-11. The assistant's tables were rebuilt against this version of the report, so it queries the figures shown above and nothing else. The brand, model and fleet-adjusted tables were removed along with their sections and it will decline questions about them rather than answer from withdrawn numbers. Its previously curated answer set was cleared for the same reason — every stored answer quoted a figure that is no longer published.

Method & honesty

Report vic-vehicle-fatality-2026-08. Supersedes vic-vehicle-fatality-2026-07.

Source

Victoria Road Crash Data (Department of Transport and Planning, CC-BY-4.0), ACCIDENT / VEHICLE / PERSON / NODE tables, crashes from 1 January 2012 to 31 October 2025, portal snapshot of 8 June 2026. 198,839 crashes, 362,687 vehicle records, 3,564 deaths. The dataset records police-reported casualty crashes; property-damage-only crashes are absent.

The constraint that shapes everything here

Victoria records injury crashes only. There is no uninjured denominator, so nothing on this page is a risk of crashing — every figure is severity conditional on an injury crash having occurred. MUARC's Used Car Safety Ratings exclude Victoria from risk calculation for exactly this reason. We do not publish per-model protection or harm-to-others ratings, because both require the component this dataset does not have.

Involvement, not fault

Crash data does not attribute legal fault, and neither do we. There is no fault field. The vehicle-slot code in the DCA classification identifies which vehicle's movement defines the crash type; it is not fault, and no published figure on this page uses it. Where we needed to know which vehicle struck which, we used the recorded physical point of impact instead.

Data quality, including what we got wrong before

Make and model are free-text fields truncated to six characters. Our July page disclosed roughly 9% unknown make; for light vehicles the correct figure is 1.38% (3,989 of 289,896), and that figure needed re-deriving. Those records are not missing at random: they carry 54.8% missing year of manufacture against 1.5% for the rest, a mean tare weight of 835 kg against 1,515 kg, and a driver killed-or-hospitalised rate of 8.22% against 15.16%. Unknown make indexes incomplete recording in less severe crashes. That biases any brand statistic, and is one reason the brand sections are withdrawn rather than repaired.

Make strings are also fragmented, and our harmonisation was incomplete: Hyundai appears as HYNDAI (14,977) and HYUNDA (161); Volkswagen as VOLKS (9,477), VOLKSW (96) and VW (2); Mercedes-Benz as MERC B (7,102) and MERCED (49); BMW as “B M W” (5,552) and “B.M.W.” (32); Land Rover as L ROV (1,354), LAND R (14) and LANDRO (6). Between 0.6% and 1.6% of each marque sat outside its largest string. Small, real, and not previously disclosed.

How this page is gated

Every section carries a written estimand — the quantity in symbols, the sampling frame, and the unit of analysis — before any chart. Every primary contrast is tested under at least one split that could reverse it; a sign flip holds publication. Every sentence is tagged descriptive, associational or external. Every number traces to a logged query.

External context

external The VIAS Institute (Belgium) has published on mass difference and injury outcomes for the lighter party, and the European Transport Safety Council campaigns on large-vehicle risk in urban areas. Those are separate studies on other data, with controls this analysis does not have. They are context for the direction found here, not evidence for it. No external body has reviewed or endorsed this page.

Standing disclaimer

Involvement rates do not indicate fault, vehicle defect, or manufacturer liability. Nothing here is engineering, legal or consumer advice. For controlled per-model safety ratings use MUARC UCSR and ANCAP.

Vehicle Fatality Trends AI Assistant
Answers only from this report's data — every number machine-verified before display.
G'day — ask me about vehicle age and crash severity, the body-type composition of fatal crashes, road deaths by user type, or the report's methods and caveats. The brand, model and fleet-adjusted sections were withdrawn on 11 August 2026 and I no longer answer from them.
Educational report assistant by Sabour Khosravi (SafeFuture Lab). Independent of DTP Victoria, BITRE, all manufacturers and all governments. Not engineering advice. Questions may be reviewed to improve accuracy.