Towards Vision Zero

An analysis of the collision reports published by Quebec's Société de l'assurance automobile (SAAQ), asking whether the circumstances a police officer records at a collision say anything useful about how badly it turned out.

The short answer is that they say something real and small, that almost all of it concerns who was involved rather than the road or the conditions, and that the limit belongs to the data rather than to the model. The long answer is the book.

Read the book · Command reference

The book opens with an interactive explorer over all 1,717,407 published reports. Tick the conditions recorded on a report and it tallies what became of every collision that matches. It fits nothing and predicts nothing.

What is here

src/vision_zero/	The package: data preparation, modelling, and the bounds
book/	The Quarto book, one document per chapter
conf/experiment/	Experiment specifications, one YAML file each
data/raw/	The twelve published extracts, 2011 to 2022
app/	The interactive explorer, a single static page
tests/	Unit, integration and end-to-end tests

The raw extracts are committed, so a clone carries the data as well as the code. Nothing needs downloading and no account is needed.

Installing

Requires uv and Python 3.12. uv will fetch the interpreter if it is missing.

git clone https://github.com/tmfreiberg/towards-vision-zero
cd towards-vision-zero
uv sync

That installs from a lockfile, so the dependency versions are the ones the published figures were produced with.

Then compile the twelve extracts into a single table:

uv run tvz data prepare

This is the only step that reads the raw CSVs. Every later command reads its output. Check that it worked with:

uv run tvz info

Trying it

uv run tvz examples montreal_ternary_complete   # collisions recorded identically, outcomes differing
uv run tvz ceiling montreal_ternary_complete    # how well any method could do
uv run tvz sweep montreal_ternary_complete      # five estimator families under identical conditions
uv run tvz ablate montreal_ternary_complete     # what each group of variables contributes

The command reference covers all of them.

Building the book

make render-fresh

Requires Quarto. Rendered results are cached under _freeze/ and that cache is committed, so continuous integration publishes the book without executing anything.

Use make render when only prose has changed and make render-fresh whenever the package has, because Quarto tracks documents and not the modules they import.

Development

make check     # lint, type-check, and run the tests
make test      # tests only
make format    # apply formatting

make help lists the rest. The same targets work on Windows through make.bat.

Acknowledgement

This began as a team project at the Erdős Institute Data Science Boot Camp in Spring 2024, with Elizabeth Kelly, Jack Wagner and Supriyo Ghosh. The present repository is a rewrite, and the errors in it are my own.

Licence

Code under the MIT licence; see LICENSE. The collision data is published by the SAAQ under the terms stated on Données Québec.