Week 3: Spatial and temporal patterns of vehicle strikes

Why a map of recorded strikes looks like a finding before anyone has checked what the dots represent — and what survives once you check.

8 March 2027

Open the slides

A Red Dot Is Not a Research Finding.

Last week you found that a time axis can mislead you before you’ve checked what produced each row. This week does the same exercise on a space axis, and the failure mode is, if anything, more persuasive: reporting practices that create apparent peaks in time can just as easily create apparent hotspots in space, and a map sells the illusion harder than a bar chart ever did.

Why a map is more convincing than it has earned#

Put collision records on a map and something happens that a spreadsheet column doesn’t do on its own: it looks like evidence. A cluster of red dots on one road reads as a finding, immediately, before anyone has asked what a dot represents, how it got there, or what else might produce a cluster that has nothing to do with danger.

A red dot is not a research finding. Twelve red dots are not automatically twelve times more convincing — they’re twelve rows in a table that happen to share a road_name field. The question this lecture asks is the same one week 2 asked of a time-of-day peak: what does the pattern actually measure, and what would have to be true for the tempting explanation to be the right one?

What the dataset can and can’t show you#

Before going further, a boundary worth stating plainly: the collision log records road_name and region for each incident. It does not record coordinates, a road network, or anything that would let you plot an exact collision location on a real map. Nothing in this unit asks you to invent that precision. Every “map” in this week’s material is a schematic built from road-level counts — useful for comparing roads against each other, not for pinpointing where on a road an animal was struck.

Three ways to count the same collisions#

Take the raw count of recorded collisions per road. This is the number everyone reaches for first, because it needs nothing else — just a tally.

Coorabin Road tops this list by a clear margin. Congratulations: you have found the road with the most records. Whether it is the most dangerous road remains under review — because a raw count answers exactly one question, “where were the most collisions recorded,” and that question doesn’t yet account for how long each road is or how much traffic uses it.

Two better-adjusted measures need a denominator: something to divide the count by, so that a long, busy road isn’t unfairly compared against a short, quiet one on count alone.

• Collisions per kilometre divides the raw count by road length, so a 38 km highway and a 9 km back road are compared per kilometre of road rather than as whole roads.
• Collisions per million vehicle-kilometres divides further, by how much traffic actually used that road over the period being studied — so a road nobody drives isn’t compared on equal terms with one that carries thousands of vehicles a day.

The map has excellent visual authority and no opinion about your denominator. It will happily show you eighteen dots on Coorabin Road and let you draw whatever conclusion you like about what “eighteen” means, without ever mentioning that Coorabin Road is also the longest, busiest road in the dataset.

A companion dataset, built for this comparison#

The collision log alone can’t support either adjusted measure — it has no road length or traffic figures. This unit supplies a second, small, synthetic dataset built specifically to make that comparison possible: the road network table. Like the collision log, every value in it is invented teaching data, built to match the seven roads that already appear in the collision log — not a survey of any real road. Its data dictionary and reporting-coverage categories are detailed on this week’s session page, which is where you’ll join it to your own counts.

The worked example: busy road, quiet road#

Here is the comparison this week turns on. Coorabin Road, in the raw count above, looks like the obvious hotspot. Old Bindari Road, with 11 recorded collisions, doesn’t stand out at all on that list.

The road network table gives Coorabin Road a length of 38 km and an average daily traffic figure of 2,800 vehicles — a long, busy regional link. It gives Old Bindari Road a length of 9 km and an average daily traffic figure of 220 vehicles — a short, quiet back road.

Using the observation window covered by the collision log — 4 January 2023 to 12 November 2026, counting both dates: 1,409 days — the same period for both roads, which matters, since the formula below only works if every road is measured over the same span, the exposure-adjusted rate is:

Recorded collisions per million vehicle-kilometres = recorded collisions / (average daily traffic × days observed × road length in kilometres) × 1,000,000

For Coorabin Road:

18 / (2,800 × 1,409 × 38) × 1,000,000 ≈ 0.12 collisions per million vehicle-kilometres

For Old Bindari Road:

11 / (220 × 1,409 × 9) × 1,000,000 ≈ 3.94 collisions per million vehicle-kilometres

Old Bindari Road has fewer recorded collisions than Coorabin Road, and an exposure-adjusted rate more than thirty times higher. The busy road looks worse on a map with dots sized by count. The quiet road is where a comparable stretch of driving is, on this measure, far more likely to end in a recorded collision.

What the adjustment assumes — and doesn’t fix#

Treat that rate as informative, not final. It rests on assumptions worth naming out loud:

• average daily traffic is treated as constant across the whole 1,409-day window, which real traffic never is — seasonal and long-term changes are smoothed away by using a single average figure;
• it still only counts recorded collisions — exposure adjustment changes the denominator, not the numerator, so it cannot correct for collisions that were never reported in the first place.

That second point matters enough to say plainly: exposure adjustment does not eliminate reporting bias. It corrects for how much a road is used, not for how completely its collisions are logged. A road with genuinely excellent reporting coverage and a road with genuinely poor reporting coverage can produce the same exposure-adjusted rate for very different underlying reasons, and the formula above has no way to tell you which situation you’re in.

Which is exactly why reporting coverage needs to be checked as its own factor, not assumed away.

Reporting coverage: a road nobody is watching#

Two roads can have identical traffic and identical collision risk and still produce very different numbers of recorded collisions, because reporting depends on someone being in a position to make a report. The road network table’s reporting_coverage column names three categories, each with a short in-universe reason:

• High — roads with routine patrols, fixed traffic cameras, or highway patrol coverage (Coorabin Road, Merricourt Highway).
• Medium — roads with irregular patrols or an active local wildlife-carer network, so most incidents eventually get logged, just not quickly or consistently (Bunjalung Ridge Road, Grellan Creek Road, Tindarra Loop Road).
• Low — roads with no routine patrol at all, where a report depends on a rare passing motorist or a wildlife carer happening to visit (Old Bindari Road, Wombaroo Pass).

Notice which road just posted the highest exposure-adjusted rate in the dataset: Old Bindari Road, in the low reporting-coverage category. If anything, its true rate could be even higher than 3.94 — a low-coverage road is the one place a genuinely dangerous stretch could be systematically under-counted. The exposure adjustment and the reporting-coverage category point in the same direction here, which is a stronger result than either alone; they won’t always agree so neatly.

Choosing a hotspot definition, out loud#

“Hotspot” is not a fixed, self-evident category — it’s a threshold someone chooses, and different thresholds answer different questions. This unit uses one explicit, simple definition for teaching purposes:

A road is a hotspot if its exposure-adjusted rate is at least double the median rate across the roads being compared.

Across the seven roads in the dataset, the median rate is 1.06 collisions per million vehicle-kilometres (Bunjalung Ridge Road’s figure, sitting squarely in the middle of the ranked list). Double that is 2.13. Exactly one road clears it: Old Bindari Road, at 3.94. Under a naive “most raw collisions” definition, the answer was Coorabin Road. Under this definition, it isn’t — a genuinely different road, reached by making the definition explicit rather than leaving it implied.

A different, equally defensible threshold — top three roads by rate, or “anything above the mean” rather than the median — would draw the line in a different place. That’s not a flaw in the method; it’s why stating your threshold, and defending it, is exactly what the week 6 assignment’s hotspot analysis asks you to do.

From this week’s session to the assignment#

The session hands you the same seven roads and asks you to build both rankings yourself, then have another pair try to break your hotspot definition using the exposure figures and reporting-coverage categories above. That exercise — producing a defensible threshold and surviving a peer’s challenge to it — is a direct rehearsal for the hotspot analysis half of the week 6 assignment, which asks you to state your exposure adjustment and defend your threshold in exactly this way.

Where this goes next#

A hotspot ranking, however carefully built, only tells you which roads collect more recorded collisions — it says nothing about what happens when one occurs. A road with fewer collisions can still produce far worse outcomes per collision than a busier one. Next week turns from how often to how hard: the physics of what speed does to a collision once it’s already happening.

Reading#

Shilling, F. M., & Waetjen, D. P. (2015). “Wildlife-vehicle collision hotspots at US highway extents: scale and data source effects”, Nature Conservation, 11, 41–60. Two independently collected data sources for the same US roads produced hotspot clusters with very little overlap — real-world evidence that “where the hotspots are” depends heavily on which records you started with, not only on where collisions actually happened.