The 2021 penalty shrank every weather coefficient to zero. That says weather did not help predict a single trip once zones and hours were known. It does not say weather has no effect. This page asks the question directly, for rain and for permitted street events, with intervals, effect sizes and the confounders each comparison does and does not handle.
The outcome
Compare the same trips, not the same day
People ride differently in the rain (more short hops, fewer walks), so raw daily averages mix two things. Each trip is compared instead with the 2019 median of its own route and hour: the index is the mean of log(minutes ÷ that median) over a day, and exp(index) − 1 reads as “the same trips took this much longer than usual”. Route-hours with at least 20 trips (129,930 of them) cover 98.1% of the 74,910,889 trips.
dry light wet (0.1 in or more, right of the dashed line)One dot per day: how much longer than its route-hour median the day's trips took.
Rain
Rain days: about 3% slower
333 days with at least 1,000 trips and no snow (6 sparse days in early January and 12 snow days left out). A day is wet with at least 0.1 inch at Central Park. Bootstrap intervals resample days within each group (B = 4,000, seed 20190101), and the regression uses HC3 standard errors, checked against Newey–West errors because neighbouring days are alike.
Wet (77 days) vs dry (204 days), raw
+0.48 min
95% CI +0.15 to +0.82 (bootstrap) · Welch +0.14 to +0.82, p 0.006 · Hedges' g 0.39
Same trips, wet vs dry
+4.5%
95% CI +2.0% to +7.1% · Hedges' g 0.47
Same trips, adjusted for month, weekday and holidays
+3.4%
95% CI +1.9% to +4.8% · OLS with HC3, n = 333 days, 21 parameters, R² 0.73 · Newey–West (7 lags) +2.0% to +4.8%
Wet vs dry, two-sample77 vs 204 days
+4.5% (+2.0% to +7.1%)
Wet vs dry, regressionmonth, weekday, holiday fixed effects
+3.4% (+1.9% to +4.8%)
Trace to 0.1 in52 days vs dry
+1.4% (−1.2% to +4.1%)
0.1 to 0.5 in36 days vs dry
+5.2% (+1.5% to +8.7%)
0.5 to 1 in29 days vs dry
+3.5% (−0.4% to +7.4%)
1 in or more12 days vs dry
+4.9% (−0.5% to +10.5%)
0%5%10%
Change in duration index, % (95% CI)Vertical line: no difference
Permitted events
Event-heavy days, matched
Permits follow the calendar (summer weekends are full of them), so comparing event days with all other days would mostly compare seasons. Each borough-day is compared only with days in its own window: same borough, same weekday, within 4 weeks, same wet or dry weather. A day is event-heavy with at least 1.5× the median permit count of its window, and its control is the nearest window day at or below that median. Holidays are excluded and controls may be reused. The outcome is the duration index for pickups in that borough (borough-days with at least 200 indexed trips).
31 matched pairs on 16 dates
−0.6%
95% CI −2.8% to +1.8% (bootstrap over dates, B = 4,000, seed 20190101)
Paired t test, one mean per date
p 0.947
mean difference in the log index -0.0008 (-0.0265 to 0.0248) · d_z -0.02
Sign test, one mean per date
7 slower, 9 faster
exact two-sided p 0.804
All boroughs31 pairs, 16 dates
−0.6% (−2.8% to +1.8%)
Manhattan8 pairs
−3.2% (−6.3% to +0.4%)
Brooklyn10 pairs
+1.3% (−0.9% to +3.5%)
Queens13 pairs
−0.6% (−3.7% to +2.5%)
-5%-2.5%0%2.5%
Event-heavy day minus matched control, duration index % (95% CI)Vertical line: no difference
Balance of the matched pairs (means)
Borough
Permits: heavy / control
Days apart
TAVG diff (°F)
Rain diff (in)
All boroughs
137 / 38
15.8
+0.8
+0.02
Manhattan
194 / 65
17.5
−0.3
−0.05
Brooklyn
125 / 29
15.4
+1.6
+0.02
Queens
112 / 28
15.1
+1.0
+0.06
Caveats
What these comparisons cannot rule out
Observational data. Rain and permits were not assigned at random. The estimates are associations after the stated adjustments, not causal effects.
Selection within a route. The index fixes the route and hour, not who rides or why. If rain makes people take the same route at more congested moments within the hour, part of the “slowdown” is that.
One rain gauge, one number a day. Precipitation is Central Park's daily total. It says nothing about when it rained or how wet Queens was.
Collisions are a mediator. Rain causes crashes, and crashes cause delays. Adjusting for collisions would remove part of the effect being measured, so neither comparison does.
Permits are not crowds. NYC Open Data has revised the 2019 events since 2021 (see conditions), and a block party and a marathon each count as one permit.
Neighbouring days are alike. Residuals of the day-level rain regression are serially correlated (lag-1 autocorrelation 0.55, Durbin–Watson 0.88), which HC3 errors ignore. Newey–West errors with 7 lags give +2.0% to +4.8%, so the headline holds.
Exploratory cuts. The dose-response bins and the per-borough results are uncorrected for multiple comparisons. The two headline estimates are the rain regression and the overall matched event effect. I chose the wet-day threshold and the event definition while building the analysis, not in a plan written beforehand.