⚙️ How-To Road crashes & traffic safety Data cleaning Notebook By · Updated

How to Analyse Speeding-Related Crashes with FARS in Python

On this page

Problem statement

Speed is named in a large share of fatal crashes, but "speeding" in crash data is not a measured speed. FARS records, for each vehicle, whether the police report says the driver was speeding: exceeding the limit, going too fast for conditions, racing, or speeding with no detail. The actual travel speed is recorded for a minority of vehicles. Analysing speeding therefore means working at the vehicle level, rolling up to crashes, and being clear that the measure is an officer's judgement written into a report.

This page uses the FARS 2023 vehicle file to flag speeding-related vehicles and crashes, counts the deaths in those crashes, splits them by road type, links drivers' ages from the person file, and checks how often a travel speed was recorded at all.

Quick answer

Flag vehicles with SPEEDREL 2 to 5, and call a crash speeding-related if any of its vehicles is:

V["speeding"] = V.SPEEDREL.isin([2, 3, 4, 5])                          # too fast for conditions, over the limit, racing, unspecified
crash_speeding = V.groupby("ST_CASE").speeding.any()
A["speeding"] = A.ST_CASE.map(crash_speeding).fillna(False).astype(bool)
print(f"speeding-related crashes: {A.speeding.sum():,} of {len(A):,}; deaths in them: {A.FATALS[A.speeding].sum():,} ({A.FATALS[A.speeding].sum() / A.FATALS.sum():.1%})")
print(V.SPEEDRELNAME[V.speeding].value_counts().to_string())
speeding-related crashes: 10,657 of 37,769; deaths in them: 11,913 (29.0%)
SPEEDRELNAME
Yes, Too Fast for Conditions    4555
Yes, Exceeded Speed Limit       4115
Yes, Specifics Unknown          2043
Yes, Racing                      145

10,657 of 2023's 37,769 fatal crashes, 28%, had at least one driver coded as speeding, and 11,913 people died in them, 29.0% of all deaths. Too fast for conditions is the most common code, ahead of exceeding the limit.

Bar chart of the share of drivers in US fatal crashes coded as speeding by age in 2023: 34 percent at 15 to 19, 30 percent at 20 to 24, 24 at 25 to 34, 19 at 35 to 44, 14 at 45 to 54, 12 at 55 to 64, 9 at 65 to 74 and 7 percent at 75 and over, with driver counts from 3,463 to 11,980 per band. Bar chart of the share of drivers in US fatal crashes coded as speeding by age in 2023: 34 percent at 15 to 19, 30 percent at 20 to 24, 24 at 25 to 34, 19 at 35 to 44, 14 at 45 to 54, 12 at 55 to 64, 9 at 65 to 74 and 7 percent at 75 and over, with driver counts from 3,463 to 11,980 per band.
Speeding by driver age.

Kinds of speeding, road types, driver ages

1. See what kind of speeding

The four "yes" codes mean different things: too fast for conditions can be below the limit on a wet road. Count vehicles by code, and how many have no answer:

print(V.SPEEDRELNAME.value_counts().to_string())
unknown = V.SPEEDRELNAME.isin(["Reported as Unknown", "No Driver Present/Unknown if Driver Present"])
print(f"vehicles with no usable answer: {unknown.mean():.1%}")
SPEEDRELNAME
No                                             43971
Yes, Too Fast for Conditions                    4555
Yes, Exceeded Speed Limit                       4115
Reported as Unknown                             3313
Yes, Specifics Unknown                          2043
No Driver Present/Unknown if Driver Present      366
Yes, Racing                                      145
vehicles with no usable answer: 6.3%

6.3% of vehicles have no usable answer, which the crash-level flag treats as not speeding; the share of speeding-related deaths is a floor in that sense.

2. Speeding by road type

Join the crash's functional class and compare the share of deaths in speeding-related crashes:

by_road = A.groupby("FUNC_SYSNAME").apply(lambda d: pd.Series({"deaths": d.FATALS.sum(), "speeding_share": d.FATALS[d.speeding].sum() / d.FATALS.sum()}),
                                          include_groups=False)
print(by_road[by_road.deaths >= 1000].sort_values("speeding_share", ascending=False).round(3).to_string())
                                 deaths  speeding_share
FUNC_SYSNAME                                           
Local                            4133.0           0.362
Minor Collector                  1257.0           0.356
Major Collector                  6711.0           0.317
Other Freeways and Expressways   1651.0           0.311
Interstate                       5236.0           0.279
Minor Arterial                   9374.0           0.276
Other Principal Arterial        12378.0           0.257

Speeding is coded most often on local roads and minor collectors, 36% of deaths, and least on principal arterials, 26%, though those carry three times as many deaths.

Share of deaths in speeding-related crashes by road functional class, US 2023. deaths in speeding-related crashes, by road typelocal roads36.2% of 4,133minor collectors35.6% of 1,257major collectors31.7% of 6,711freeways, expressways31.1% of 1,651interstates27.9% of 5,236principal arterials25.7% of 12,378020%40%All roads: 29.0% of 41,025 deaths spatialworkflow.io
Speeding by road type.

3. Speeding by driver age

Driver ages are in the person file, one row per person, with PER_TYP 1 for drivers; join on crash and vehicle number:

d = V.merge(P[P.PER_TYP == 1][["ST_CASE", "VEH_NO", "AGE"]], on=["ST_CASE", "VEH_NO"], how="inner")
d = d[d.AGE < 998]
bands = pd.cut(d.AGE, [15, 20, 25, 35, 45, 55, 65, 75, 100], right=False)
print(d.groupby(bands, observed=True).speeding.agg(["size", "mean"]).rename(columns={"size": "drivers", "mean": "speeding"}).round(3).to_string())
           drivers  speeding
AGE                         
[15, 20)      3841     0.339
[20, 25)      6446     0.297
[25, 35)     11980     0.236
[35, 45)      9898     0.190
[45, 55)      8142     0.136
[55, 65)      7585     0.115
[65, 75)      4962     0.094
[75, 100)     3463     0.070

The share falls steadily with age, from 33.9% of drivers aged 15 to 19 to 7.0% of those 75 and over.

How often a speed was recorded

The data used on this page

NHTSA's FARS 2023 national file: vehicle.csv (speeding, travel speed, posted limit, driver drinking), accident.csv and person.csv.

import zipfile
from pathlib import Path

import pandas as pd
import requests

Path("data").mkdir(exist_ok=True)
F = Path("data/FARS2023NationalCSV.zip")
if not F.exists():
    F.write_bytes(requests.get("https://static.nhtsa.gov/nhtsa/downloads/FARS/2023/National/FARS2023NationalCSV.zip", timeout=600).content)
Z = zipfile.ZipFile(F)
V = pd.read_csv(Z.open("FARS2023NationalCSV/vehicle.csv"), encoding="latin1", low_memory=False,
                usecols=["ST_CASE", "VEH_NO", "SPEEDREL", "SPEEDRELNAME", "TRAV_SP", "VSPD_LIM", "DR_DRINK"])
A = pd.read_csv(Z.open("FARS2023NationalCSV/accident.csv"), encoding="latin1", low_memory=False,
                usecols=["ST_CASE", "FATALS", "FUNC_SYSNAME", "HOUR"])
P = pd.read_csv(Z.open("FARS2023NationalCSV/person.csv"), encoding="latin1", low_memory=False, usecols=["ST_CASE", "VEH_NO", "PER_TYP", "AGE"])
print(f"{len(V):,} vehicles in {len(A):,} crashes")
58,508 vehicles in 37,769 crashes

Example 1: how often a speed was recorded

TRAV_SP is the travel speed in mph when known; 997 means over 151 mph, 998 not reported and 999 unknown:

known = V.TRAV_SP < 997
print(f"vehicles with a recorded travel speed: {known.mean():.1%}; among speeding-related vehicles: {known[V.speeding].mean():.1%}")
s = V[known & V.speeding & (V.VSPD_LIM < 98)]
print(f"speeding-related with speed and limit: {len(s):,}; median over the limit {(s.TRAV_SP - s.VSPD_LIM).median():.0f} mph")
print(f"speeding-related drivers also coded as drinking: {(V.DR_DRINK[V.speeding] == 1).mean():.1%}; other drivers: {(V.DR_DRINK[~V.speeding] == 1).mean():.1%}")
vehicles with a recorded travel speed: 38.5%; among speeding-related vehicles: 39.2%
speeding-related with speed and limit: 4,205; median over the limit 22 mph
speeding-related drivers also coded as drinking: 36.3%; other drivers: 12.7%

Only 38.5% of vehicles have a travel speed, and speeding-related ones are no better recorded, at 39.2%. Where speed and limit are both known, speeding drivers were a median 22 mph over. 36.3% of speeding drivers were also coded as drinking, against 12.7% of the rest.

Histogram of recorded travel speed minus posted limit for 4,205 speeding-related vehicles in 2023, with a median of 22 mph over the limit; the distribution spikes at multiples of 5 mph, and a small tail lies below the limit, where drivers were coded as too fast for conditions. Histogram of recorded travel speed minus posted limit for 4,205 speeding-related vehicles in 2023, with a median of 22 mph over the limit; the distribution spikes at multiples of 5 mph, and a small tail lies below the limit, where drivers were coded as too fast for conditions.
How far over the limit, where it is known.

Where the speeding variable comes from

What the FARS speeding variable is and is not. what SPEEDREL recordsIt isthe police report, coded by FARSper vehicle: 10,858 flaggedover the limit or too fast for conditionsconsistent nationallyIt is nota measured speedTRAV_SP known for 38.5%only over the limitfree of judgementRoll up with a stated rule: a crash is speeding-related if any vehicle is spatialworkflow.io
What SPEEDREL records.

FARS's speeding variable comes from the police crash report as interpreted by the FARS analyst: the officer's statement that a driver was exceeding the limit, driving too fast for conditions or racing, or a charge that implies it. It is a judgement, made after the crash, often without speed measurement. That makes it consistent enough for national comparisons and weak for fine distinctions, especially where states' report forms differ in how they ask.

Rolling up from vehicles to crashes needs a rule. Here a crash is speeding-related if any vehicle in it was, the usual convention for counting deaths in speeding-related crashes. The resulting share is about three in ten deaths, and the patterns in it are consistent: young drivers are coded as speeding far more often than older ones, and speeding overlaps with drinking. Where a travel speed is recorded, it says how far above the limit drivers were; but it is recorded for a minority of vehicles, about two in five, speeding or not.

Codes, posted limits and crash or vehicle level

Analysing speeding-related crashes in FARS: problems and fixes. speeding data trapsTrapInsteadvehicle and crash counts mixedsay which: 10,858 vehicles, 10,657crashesSPEEDREL read as over the limittoo fast for conditions is the topcodeTRAV_SP averaged with codesdrop 997, 998 and 999unknowns counted as noreport them: 6.3% of vehiclesFARS 2023 spatialworkflow.io
Speeding data traps.
  • Codes: SPEEDREL 0 is no, 2–5 are the kinds of yes, 8 is no driver present and 9 is unknown; check the codes in each year's manual before stacking years.
  • Posted limit: VSPD_LIM 98 and 99 mean not reported and unknown.
  • Crash or vehicle level: a two-vehicle crash with one speeding driver counts once at crash level and once of two at vehicle level; say which.
  • Too fast for conditions: this code includes speeds under the limit; "over the limit" alone is a narrower measure.
  • Other years: stack several years before splitting by state or age band; single-year counts in small groups are noisy.

FAQ

Read vehicle.csv, flag vehicles with SPEEDREL 2 to 5, and mark a crash as speeding-related if any of its vehicles is flagged.

In FARS 2023, 29.0% of deaths were in crashes where at least one driver was coded as speeding.

Does FARS record how fast vehicles were going?

Sometimes: TRAV_SP has a travel speed for a minority of vehicles; for most it is not reported.

Which drivers are most often coded as speeding?

The youngest: in 2023, a third of drivers aged 15 to 19 in fatal crashes were coded as speeding, against under one in ten of those 65 and over.