Week 2 Session: Reading the collision data

Getting hands-on with the shared dataset before this week's lecture on macropod behaviour asks you to explain what's in it.

1 March 2027

The dataset#

Download the collision log (CSV, 90 records, 2023–2026). It’s a synthetic dataset built by the teaching team for this unit — modelled on how Australian jurisdictions structure animal-strike reports (police, insurance, wildlife-carer, and road-authority logs) — not a record of real incidents. Treat every figure in it as teaching material, not evidence about any real road.

Data dictionary

The log contains at least one deliberate recording artefact — a pattern that reflects how a record got made, not when the incident happened. Finding it is part of this session.

Before the session#

Skim the data dictionary above so you know what the columns mean — but don’t open the CSV itself yet. The first time you look at an actual row happens in the session, right after you’ve written down a prediction with nothing to bias it.

In the session#

This runs as one 50-minute sequence. Work in the same pairs throughout.

1. 5 minutes — predict. Before opening the dataset, each pair predicts which day period (dawn, day, dusk, night) will contain the most records, and writes one sentence saying why. No looking yet.
2. 10 minutes — build the first distribution. Produce a count of records by day_period, and a second breakdown by month or season. Compare with another pair: the shapes should roughly agree.
3. 5 minutes — write the tempting explanation. In one sentence, write the most tempting explanation for your first graph. Don’t hedge it — the point of the next step is testing this exact sentence, not a cautious version of it.
4. 15 minutes — audit. Look past the day-period label to the exact values behind it. Work through these in order, stopping as soon as one of them gives you something:
   • Are any exact times unusually common — not “around dusk,” but the same clock time, repeated?
   • Which report_source produced those specific rows?
   • What does data_quality_flag say on those rows, if anything?
   • Does the timestamp on those rows describe when the collision happened, or when someone’s reporting process started?
5. 10 minutes — revise. Decide, as a pair, how to handle whatever you found: remove those records, separate them into their own category, or annotate them and keep them visible. Rebuild your distribution with that decision applied, and write a revised one-sentence conclusion.
6. 5 minutes — exit ticket. Submit your original sentence, your revised sentence, and the one thing that changed your mind.

Afterwards#

This session’s output is the Suspiciously Punctual Kangaroo Memo — a one-page pair submission. It’s formative and ungraded: it’s practice at revising a conclusion when the data turns out to need it, not a mark on whether your first graph was right.

The Suspiciously Punctual Kangaroo Memo

• The original graph (your day-period distribution from step 2).
• The revised graph (same distribution, with the audited records handled).
• The original explanation (your one sentence from step 3).
• The recording artefact you identified (what it was, and which column revealed it).
• The revised conclusion (your one sentence from step 5).
• One additional type of data — not in this dataset — that would help separate animal activity from traffic exposure as the cause of any remaining pattern.

Keep a copy of both distributions — the revised one is the starting point for the hotspot mapping in week 3, where the same question about records versus reality comes back in a different shape.