"""Rank every night of the year by alcohol-impaired traffic deaths, FARS 2015-2024.
Alcohol-impaired death (NHTSA definition): a death in a crash where at least one
driver had BAC >= .08 g/dL. FARS fills in unknown driver BACs by multiple
imputation (MIDRVACC file, A1..A10 = highest driver BAC in the crash, in hundredths
of g/dL). NHTSA's count = FATALS x share of the 10 imputations with BAC >= .08.
Night N(d) = 6:00 p.m. on date d through 5:59 a.m. on d+1 (crash hour 18-23 on d
plus hour 0-5 on d+1). Crashes with unknown hour (99) are left out of nights.
Usage: python -I analyze.py
"""
import csv, io, sys, zipfile, datetime as dt, statistics as st, json
from collections import defaultdict
src, out = sys.argv[1], sys.argv[2]
YEARS = range(2015, 2025)
ai_by_hour = defaultdict(float) # (date, hour) -> alcohol-impaired deaths
all_by_hour = defaultdict(float) # (date, hour) -> all deaths
unknown_hour_ai = 0.0
total_ai_by_year = defaultdict(float)
total_all_by_year = defaultdict(float)
def read(z, base):
for n in z.namelist():
if n.split('/')[-1].lower() == base:
return list(csv.DictReader(io.TextIOWrapper(z.open(n), encoding='latin-1')))
raise KeyError(base)
for y in YEARS:
z = zipfile.ZipFile(f'{src}/FARS{y}.zip')
mi = {}
for r in read(z, 'midrvacc.csv'):
vals = [int(r[f'A{i}']) for i in range(1, 11)]
mi[r['ST_CASE']] = sum(v >= 8 for v in vals) / 10.0
for r in read(z, 'accident.csv'):
f = int(r['FATALS'])
p = mi.get(r['ST_CASE'], 0.0)
ai = f * p
total_ai_by_year[y] += ai
total_all_by_year[y] += f
h = int(r['HOUR'])
if h > 23:
unknown_hour_ai += ai
continue
d = dt.date(int(r['YEAR']), int(r['MONTH']), int(r['DAY']))
ai_by_hour[(d, h)] += ai
all_by_hour[(d, h)] += f
def night(d, table):
n = d + dt.timedelta(days=1)
return sum(table.get((d, h), 0) for h in range(18, 24)) + sum(table.get((n, h), 0) for h in range(0, 6))
# all complete nights: 2015-01-01 .. 2024-12-30
nights = []
d = dt.date(2015, 1, 1)
while d <= dt.date(2024, 12, 30):
nights.append((d, night(d, ai_by_hour), night(d, all_by_hour)))
d += dt.timedelta(days=1)
def nth_weekday(y, m, wd, n): # wd Mon=0
d = dt.date(y, m, 1)
d += dt.timedelta(days=(wd - d.weekday()) % 7)
return d + dt.timedelta(weeks=n - 1)
def last_weekday(y, m, wd):
d = dt.date(y, m + 1, 1) - dt.timedelta(days=1) if m < 12 else dt.date(y, 12, 31)
return d - dt.timedelta(days=(d.weekday() - wd) % 7)
def thanksgiving(y): return nth_weekday(y, 11, 3, 4)
def superbowl(y):
# Super Bowl Sunday dates (NFL): first Sunday of Feb through 2021, second Sunday from 2022
return {2015: dt.date(2015,2,1), 2016: dt.date(2016,2,7), 2017: dt.date(2017,2,5), 2018: dt.date(2018,2,4),
2019: dt.date(2019,2,3), 2020: dt.date(2020,2,2), 2021: dt.date(2021,2,7), 2022: dt.date(2022,2,13),
2023: dt.date(2023,2,12), 2024: dt.date(2024,2,11)}[y]
HOLIDAYS = {
'Thanksgiving Eve (Wed.)': lambda y: thanksgiving(y) - dt.timedelta(days=1),
'Thanksgiving night (Thu.)': thanksgiving,
'Saturday of Thanksgiving weekend': lambda y: thanksgiving(y) + dt.timedelta(days=2),
"New Year's Eve": lambda y: dt.date(y, 12, 31),
"New Year's Day night": lambda y: dt.date(y, 1, 1),
'July 3': lambda y: dt.date(y, 7, 3),
'Fourth of July': lambda y: dt.date(y, 7, 4),
'Halloween': lambda y: dt.date(y, 10, 31),
"St. Patrick's Day": lambda y: dt.date(y, 3, 17),
'Cinco de Mayo': lambda y: dt.date(y, 5, 5),
'Super Bowl Sunday': superbowl,
'Saturday of Memorial Day weekend': lambda y: last_weekday(y, 5, 0) - dt.timedelta(days=2),
'Sunday of Memorial Day weekend': lambda y: last_weekday(y, 5, 0) - dt.timedelta(days=1),
'Saturday of Labor Day weekend': lambda y: nth_weekday(y, 9, 0, 1) - dt.timedelta(days=2),
'Sunday of Labor Day weekend': lambda y: nth_weekday(y, 9, 0, 1) - dt.timedelta(days=1),
'Christmas Eve': lambda y: dt.date(y, 12, 24),
'Christmas night': lambda y: dt.date(y, 12, 25),
}
by_date = {d: (a, t) for d, a, t in nights}
# rank within each year (1 = most alcohol-impaired deaths)
rank = {}
for y in YEARS:
# competition ranking: 1 + number of nights that year with strictly more deaths (ties share a rank)
ys = [(round(a, 6), d) for d, a, t in nights if d.year == y]
for a, d in ys:
rank[d] = (1 + sum(b > a for b, _ in ys), len(ys))
rows = []
for name, f in HOLIDAYS.items():
per = []
for y in YEARS:
d = f(y)
if d in by_date:
a, t = by_date[d]
per.append({'year': y, 'date': d.isoformat(), 'weekday': d.strftime('%a'), 'ai': round(a, 1), 'all': int(t),
'rank': rank[d][0], 'nights_in_year': rank[d][1]})
rows.append({'night': name, 'years': len(per), 'avg_ai': round(st.mean(p['ai'] for p in per), 1),
'avg_all': round(st.mean(p['all'] for p in per), 1),
'median_rank': st.median(p['rank'] for p in per), 'per_year': per})
def avg_where(pred):
v = [(a, t) for d, a, t in nights if pred(d)]
return round(st.mean(a for a, t in v), 1), round(st.mean(t for a, t in v), 1), len(v)
baselines = {
'Average night (all nights)': avg_where(lambda d: True),
'Average Wednesday night': avg_where(lambda d: d.weekday() == 2),
'Average Friday night': avg_where(lambda d: d.weekday() == 4),
'Average Saturday night': avg_where(lambda d: d.weekday() == 5),
'Average Saturday night, June-Aug.': avg_where(lambda d: d.weekday() == 5 and d.month in (6, 7, 8)),
'Average Saturday night, Oct.-Nov.': avg_where(lambda d: d.weekday() == 5 and d.month in (10, 11)),
'Average Wednesday night, Oct.-Dec. (not Thanksgiving Eve)':
avg_where(lambda d: d.weekday() == 2 and d.month in (10, 11, 12) and d != thanksgiving(d.year) - dt.timedelta(days=1)),
}
# top nights each year
top = {}
for y in YEARS:
ys = sorted([(a, d) for d, a, t in nights if d.year == y], reverse=True)[:10]
top[y] = [(d.isoformat(), d.strftime('%a'), round(a, 1)) for a, d in ys]
# Thanksgiving period, NHTSA window: Wed 6pm - Mon 5:59am (4.5 days = 108 hours)
def period(start_dt, hours, table):
s = 0.0
for k in range(hours):
t = start_dt + dt.timedelta(hours=k)
s += table.get((t.date(), t.hour), 0)
return s
per_period = []
for y in YEARS:
tg = thanksgiving(y)
start = dt.datetime.combine(tg - dt.timedelta(days=1), dt.time(18))
per_period.append({'year': y, 'ai': round(period(start, 108, ai_by_hour), 1), 'all': int(period(start, 108, all_by_hour))})
# share of total alcohol-impaired deaths by night share
result = {
'total_ai_by_year': {y: round(v) for y, v in total_ai_by_year.items()},
'total_all_by_year': {y: int(v) for y, v in total_all_by_year.items()},
'unknown_hour_ai_total': round(unknown_hour_ai, 1),
'holidays': rows, 'baselines': baselines, 'top10_by_year': top, 'thanksgiving_period_nhtsa_window': per_period,
}
json.dump(result, open(f'{out}/results.json', 'w'), indent=1, default=str)
print('Totals AI by year:', result['total_ai_by_year'])
print('Unknown-hour AI deaths (excluded from nights):', result['unknown_hour_ai_total'])
print('\nNamed nights, avg alcohol-impaired deaths per night (2015-2024), median rank of 365:')
for r in sorted(rows, key=lambda r: -r['avg_ai']):
print(f" {r['night']:<38} {r['avg_ai']:6.1f} all={r['avg_all']:6.1f} medRank={r['median_rank']} n={r['years']}")
print('\nBaselines:')
for k, v in baselines.items():
print(f' {k:<55} ai={v[0]} all={v[1]} n={v[2]}')
print('\nThanksgiving Eve per year:')
for p in rows[0]['per_year']:
print(' ', p)
print('\nTop 10 nights each year:')
for y, t in top.items():
print(y, t)
print('\nThanksgiving period (Wed 6pm-Mon 5:59am):', per_period)