Road accidents in France and in Haut-Rhin, 2019–2024

Guided tour of the analysis. Every table comes from a SQL query on the DuckDB database built by src/build_db.py; every chart is produced by a function of src/analysis.py.

Data: ONISR national road-accident file (BAAC), INSEE population estimates. Scope: metropolitan France.

import sys; sys.path.insert(0, '../src')
from IPython.display import Image
import pandas as pd; pd.set_option('display.precision', 3)
from analysis import *
from model import fit, fig_odds

0. Can we trust the data?

Share of missing values on the fields we use, then our yearly death counts compared with the official ONISR figures.

data_quality()

	field	missing_%
0	Accident: GPS coordinates	0.00
1	Accident: hour	0.00
2	Accident: weather	0.01
3	Place: speed limit (valid 10-130)	2.23
4	Person: severity	0.05
5	Person: age	1.50
6	Person: sex	1.43
7	Person: travel mode	0.53

validation_vs_onisr()

	year	killed_ours	killed_onisr
0	2019	3244	3244.0
1	2020	2541	2541.0
2	2021	2944	2944.0
3	2022	3267	3267.0
4	2023	3167	3167.0
5	2024	3193	NaN

Our counts match the official figures exactly for 2019–2023 (2024 was not yet published as a final figure when this was written), so the cleaning keeps every accident.

1. Trend

Accidents fell during the 2020 lockdowns and have not returned to 2019 levels; deaths have.

tr = yearly_trend(); tr

	year	accidents_fr	killed_fr	accidents_68	killed_68
0	2019	56016	3244	437	25
1	2020	45121	2541	317	19
2	2021	53540	2944	373	29
3	2022	52380	3267	319	30
4	2023	51641	3167	367	36
5	2024	51058	3193	351	27

fig_trend(tr); Image('../figures/01_trend.png')

Haut-Rhin counts are small (20–36 deaths a year), so year-to-year swings are mostly noise.

2. When

Busiest at the evening rush hour; deadliest between 3 and 6 a.m. (about 11% of accidents fatal, vs 4.5–5.5% during the day).

hw = hour_weekday(); fig_when(hw); Image('../figures/02_when.png')

3. Who

km = killed_by_mode(); fig_modes(km)

	2019	2024	change_%
mode
Car	1658	1571	-5.0
Motorcycle	615	597	-3.0
Pedestrian	482	454	-6.0
Bicycle	172	192	12.0
E-scooter	11	47	327.0
E-bike	15	32	113.0

Image('../figures/03_modes.png')

ag = killed_rate_by_age(); fig_age(ag); Image('../figures/04_age.png')

Young men (20–24) and people over 80 have the highest death rates.

4. Where

dep = by_department(); dep.head(10)

	dep	pop	killed_per_year	acc_per_year	killed_per_million	rank_safest
0	92	1.642e+06	22.667	2394.500	13.804	1
1	94	1.427e+06	21.833	2297.167	15.303	2
2	93	1.683e+06	27.167	2765.833	16.144	3
3	75	2.103e+06	37.667	4905.000	17.914	4
4	95	1.274e+06	27.333	865.167	21.448	5
5	78	1.462e+06	37.000	676.667	25.316	6
6	91	1.316e+06	34.000	1349.833	25.835	7
7	69	1.921e+06	54.500	2119.167	28.370	8
8	59	2.607e+06	86.333	972.333	33.120	9
9	67	1.168e+06	39.333	636.500	33.664	10

dep[dep.dep == '68']

	dep	pop	killed_per_year	acc_per_year	killed_per_million	rank_safest
10	68	769231.0	27.667	360.667	35.967	11

fig_map(dep); Image('../figures/05_map_departments.png')

profile_68_vs_france()

area	Haut-Rhin	France (other depts)
in_town	0.533	0.639
at_night	0.241	0.265
on_motorway	0.091	0.098
with_motorcycle	0.189	0.226
with_bike_or_escooter	0.167	0.138
with_pedestrian	0.148	0.164
fatal	0.072	0.056
killed_per_accident	0.077	0.059

fig_haut_rhin(haut_rhin_points()); Image('../figures/06_haut_rhin_map.png')

A warning about the denominator

Deaths are almost always recorded; minor accidents are not. Haut-Rhin has a higher share of fatal accidents than France, yet fewer deaths per inhabitant: it records fewer non-fatal accidents per inhabitant. The city counts make it visible.

rc = recording_completeness(); fig_completeness(rc); Image('../figures/07_recording_completeness.png')

cities().pivot(index='year', columns='city', values='accidents')

city	Colmar	Mulhouse	Strasbourg
year
2019	48	22	369
2020	20	18	215
2021	31	19	263
2022	24	16	279
2023	19	11	243
2024	24	15	225

Mulhouse records fewer accidents than Colmar, a smaller city, and about 15 times fewer than Strasbourg, which is only about three times its size. That points to under-recording of non-fatal accidents in Mulhouse rather than safer streets, a hypothesis to check with local police or city data.

5. What makes an accident fatal?

Logistic regression on ~300,000 accidents. Odds ratios above 1 increase the odds that someone dies; they are associations, not causes.

ors, stats = fit(); stats

{'n_accidents': 301290,
 'fatal_rate': 0.055524577649440736,
 'auc_holdout': 0.7616546574231433,
 'pseudo_r2': 0.108045078390962}

ors.round(2)

	odds_ratio	ci_low	ci_high	p_value
E-scooter involved	0.75	0.64	0.87	0.0
Rain, snow, fog or wind	0.84	0.80	0.88	0.0
Night, street lights on	1.09	1.03	1.16	0.0
Weekend	1.17	1.13	1.21	0.0
Dusk or dawn (vs daylight)	1.19	1.12	1.27	0.0
Single vehicle, no collision	1.20	1.15	1.27	0.0
Speed limit 110-130	1.23	1.13	1.35	0.0
Motorcycle involved	1.28	1.23	1.34	0.0
Speed limit 70 (vs <=50)	1.52	1.41	1.65	0.0
Bicycle or e-bike involved	1.52	1.43	1.63	0.0
Night, no street lights	1.75	1.67	1.82	0.0
Outside built-up area	1.91	1.78	2.05	0.0
Speed limit 80-90	2.12	1.97	2.28	0.0
Head-on collision	2.56	2.45	2.69	0.0
Fixed obstacle hit (tree, wall...)	2.62	2.53	2.72	0.0
Heavy vehicle involved	2.99	2.84	3.15	0.0
Pedestrian involved	3.09	2.92	3.26	0.0

fig_odds(ors); Image('../figures/08_odds_ratios.png')

Conclusions

• About 3,000 people die every year on metropolitan French roads; the number is flat since 2019.
• Two thirds of deaths happen outside built-up areas, which hold only a third of accidents.
• Deaths of e-scooter riders went from 11 to 47 a year; injured e-scooter users quadrupled.
• Haut-Rhin is among the safer departments (36 deaths per million, rank 11 of 96).
• Pedestrians, heavy vehicles, fixed obstacles and head-on collisions multiply the odds of a fatal outcome by 2.5 to 3.