New York taxi screens compute suggested tips on the all-in total, which includes every government surcharge on the meter. When policy adds a fee, every suggested tip rises with it.
At exactly 4:00 on weekdays, the $2.50 rush-hour surcharge switches on, assigned by the pickup minute. Nothing about the ride or the driver changes. The suggested tips update on their own.
A tip is a voluntary payment for service. The surcharge changes nothing about the service, so economic theory says the answer should be zero. Pure button arithmetic says seventeen cents, the average tip rate. The estimates lie between these benchmarks.
What we find: about eleven cents of every statutory surcharge dollar becomes tip; among trips the meter actually charges, fourteen cents per surcharge dollar, about four-fifths of the rate at which riders tip the fare itself. Riders do not switch to cash or re-time their hails, and when the surcharge was repriced 150% overnight they undid about a fifth of it.
Sources, in order: card trips in the estimation windows (descriptive table of the paper); the June-2025 base test; annualized 2025H1 component totals on card trips; the pooled estimate per statutory dollar, 0.108 (SE 0.001).
Take the 2.18 million credit-card tips of June 2025 and ask what number each tip is a percentage of. Toggle between the two candidates.
Share of tips at each half-point percentage
June 2025 · standard-rate credit-card trips with positive tips
Against the all-in total, 57.3% of tips sit at exactly 20%, with further spikes at 25 and 30, the three buttons. Against the fare alone, the same tips scatter; no bin holds more than 3.6%. Tipping is organized by the base the buttons use.
Sample: June 2025 yellow-cab trips, standard rate (RatecodeID 1), credit-card payment, metered fare $3–200, positive tip, zero recorded tolls (2.18M trips). For each trip the tip is divided by (a) the all-in pre-tip total (total_amount − tip_amount) and (b) the metered fare; ratios are binned in half-percentage-point bins over 5–40%. Bars show each bin’s share of trips. “Exactly 20%” means equality within half a cent. Source: build_spikes.py; Figure 3 of the paper.
The rush surcharge is assigned by the pickup minute, so trips minutes apart face different tip bases. Because fares also drift at rush hour, the estimates difference the 4:00 PM jump against weekends and exempt holidays, where the same clock carries no surcharge.
Mean credit-card tip by pickup minute
Weekdays, Sep 2021 – Sep 2024 · hover for values
The legislature repriced the surcharge from $1.00 to $2.50 on December 19, 2022, and the step at 4:00 grew with the dose. Per statutory dollar the pooled estimate is 0.108 (SE 0.001); dropping every control moves it about a cent.
Points are trip-weighted mean credit-card tips by minute of pickup: weekdays excluding legal holidays, September 2021 to September 2024, standard-rate card trips aggregated to date-by-minute cells. The series split at December 19, 2022, when the statutory dose changed from $1.00 to $2.50. No smoothing is applied. Cells: build_cells_v3.py.
The estimand is a partial derivative: ρ = ∂E[tip] / ∂(statutory surcharge): how many cents the average credit-card tip moves when the legislature puts one more dollar on the meter. Step through what it takes to earn the word causal.
The stepper reproduces the paper’s estimating equation: T = α + πD + ωW + δ(D×W) + γB + f(r)×W + ε, where D indicates minutes past the cutoff, W indicates weekdays, B is the non-surcharge base, and f(r) is a local-linear trend in pickup minute on each side, separately by day type. Cells are weighted by card trips and standard errors are clustered by date; the bandwidth is ±60 minutes. The headline divides δ by the statutory dose. Source: 22_coefplot.do, logged in logs/coefplot.log.
The doses range from $0.50 to $5.00; surcharges switch both on and off; three clock cutoffs are used; the holiday calendar switches the same clock off on exempt days and leaves it on for Juneteenth; the JFK flat fare removes the fare confound by construction; and the toll boundary provides identification with no clock involved. The one failure is reported, in grey, with its diagnosis. Two further sources of variation follow in their own sections: the December 2022 repricing, which tests whether riders re-optimize when the dose jumps 150%, and the adoption events, the congestion surcharge in February 2019 and the congestion toll in January 2025, which show the effect switching on in event time.
Six sources of quasi-experimental variation
Four are plotted here; the repricing and the adoption events follow below. Each source is immune to a different objection. Hover a source to isolate its estimates.
Fifteen estimates on one per-dollar axis: every design that should detect the effect lands between 0.09 and 0.15, and the placebos do not.
The dashed line is the pooled clock mean, 0.108. The 2025 congestion toll, identified purely from geography, lands at 0.107. JFK’s $5.00 flat-fare surcharge gives 0.149 with no controls at all. The holiday calendar checks both directions: exempt holidays, where the same clock carries no surcharge, serve as a second control group (0.096 per dollar at both doses), and Juneteenth, the one holiday still charged, restores the treated step (0.096). The one failure (6 AM weekdays) has a diagnosis: dawn airport composition, and its weekend twin behaves.
Clock rows: the reduced-form jump at each cutoff from a local-linear regression in pickup minute (bandwidth ±60), conditioning on the non-surcharge base, weighted by card trips, clustered by date, then divided by the statutory dose. Calendar triangles: the same difference-in-discontinuities re-estimated with exempt legal holidays replacing weekends as the control group, 0.239 (0.028) at the $2.50 dose (wild cluster bootstrap p < 10−5) and 0.095 (0.038) at $1.00 (p = 0.035), each divided by its dose; the Juneteenth row is the fare-conditioned 4 PM step on the one weekday holiday the statute still charges, 0.241 (0.067) at the $2.50 dose; and the exempt-holiday placebo is the fare-conditioned step on exempt holidays themselves, 0.055 (0.030), divided by the dose (computed in 24_holidays.do). JFK squares: the identical specification on RatecodeID-2 flat fares at doses of $4.50 and $5.00; because the fare is flat, the estimate is unchanged without any controls. Toll diamond: an origin–destination difference-in-differences (a trip is eligible if pickup or dropoff lies in the zone) with route and month fixed effects, clustered by pickup zone; the coefficient $0.080 (0.005) is divided by the $0.75 statutory toll. Hollow circles: weekend and exempt-holiday placebo steps divided by the matched regime’s dose. Grey: the 6 AM weekday cutoff, where the dawn shift toward airport traffic breaks the fare conditioning; its weekend twin is normal. The dashed line is the precision-weighted mean of the six primary changes. Clustering follows the assignment mechanism (dates for the clocks, pickup zones for the toll); designs with few clusters use the wild cluster bootstrap; and minimum detectable effects, 2.8 times the standard error, are reported for every design (0.005 to 0.039 per dollar) so that null results are interpretable. Source: logs/coefplot.log and logs/stata-master.log.
Hypothesis testing, drawn: the permutation distribution
The null hypothesis, an empirical null distribution, and the rejection region, using this paper’s own test
The parametric route reaches the same verdict: the estimate divided by its clustered standard error gives t ≈ 30, far beyond any conventional critical value, and a wild cluster bootstrap rejects with p < 10−6. The permutation version needs no distributional assumption at all; the placebo cutoffs build the null distribution from the data.
On December 19, 2022 the rush surcharge rose 150% overnight, adding $1.50 to the same bill. Two habits predict two lines. Riders who keep tipping their usual percentage show a flat tip rate; riders who hold their tip dollars fixed must cut the rate by 1.1 points.
Weekly tip rate: rush window vs. surcharge-free pre-rush window
Weekdays, 26 weeks around December 19, 2022
The rate fell 0.22 points, one-fifth of the way to the dollar-defending benchmark. Riders kept their customary percentage on the bigger bill; four-fifths of the increase passed through, immediately and permanently. The one-fifth that didn’t? Almost entirely the round-dollar tippers ("$5 every time"), whose tips don’t scale with the bill by construction.
Weekly tip rate = the sum of card tips divided by the sum of pre-tip totals, weekdays, computed separately for the rush window (minutes 960–1050) and the surcharge-free pre-rush window (870–955), for weeks −15 to +12 around December 19, 2022. The benchmark line: holding tip dollars fixed while the base rises $1.50 implies a 1.1-point decline in the rate. The 0.22-point estimate nets the pre-rush series from the rush series; the paper’s regression version clusters by date. Source: 17_lp_dec22.do and the cells above.
Each of these papers varies one element of the same mechanism. Our paper varies the element a legislature controls.
The rush surcharge is a fare component. The January 2025 congestion toll is a tax, collected from riders and remitted by drivers to the MTA. Routes differ in how often the toll is actually charged, and the tip effect climbs step-for-step with that dose.
Tip effect by share of route’s trips charged the 2025 toll
Fare-conditioned, relative to routes charged <10% · 95% CIs
Zone drift would have to reproduce this gradient by coincidence. In the paper, the district also emerges on a map at the true toll date and not at a placebo date. The toll is also the second adoption of a congestion charge in this market; pooling it with the February 2019 congestion surcharge gives an instrumental-variables estimate of 0.163 (SE 0.021) per recorded surcharge dollar, in line with the clocks’ 2SLS of 0.140.
Routes are pickup–dropoff zone pairs. The dose is the share of a route’s post-period trips actually charged the fee, measured from recorded fee incidence (February to June 2025). Effects come from a regression of route-month mean tips on post-period dose-bin indicators (relative to routes charged under 10%), with route and month fixed effects and the non-surcharge base as a control, weighted by card trips and clustered by pickup zone. A note on pre-trends: the toll event study looks flat month by month, yet a joint Wald test on the twelve pre-period coefficients rejects (p = 0.0007); eyeballing pre-trends is not a test, so the paper reports the rejection and rests the spatial claim on this dose gradient, which slow drift cannot mimic. Source: 14_final_analyses.do.
Pass-through is stable from $0.50 to $5.00, so the arithmetic ρ × fee × volume prices any meter fee’s tip leak in advance. Drag the toll.
Arithmetic: induced tips = ρ × fee × charged card trips, with volume held at the 2025 toll’s observed 22.1M charged card trips per year; rider cost per trip = fee × (1 + ρ). The bounds use ρ = 0.108 (the pooled estimate per statutory dollar) and 0.150 (per recorded dollar, the aggregation-consistent rate for collected fees). Dose linearity is estimated in the paper up to $5.00; holding volume fixed is consistent with the estimated demand response of −1.0% (SE 0.9). The actual 2025 toll is $0.75. The 2019 congestion surcharge, still in force at $2.50, moves about $8–11M a year by itself.
In 2012, one payment vendor computed suggestions on the fare and surcharge; the other included tax and tolls. By 2025 every vendor had converged on the all-in total, and with that convergence came a fiscal channel nobody voted on: about $21 million a year of New York’s congestion charges continue past the MTA to drivers as gratuities. The transfer runs from riders to drivers, plausibly progressively. The point is not that the transfer is bad, but that no one chose it. The dial that sets it is one line of payment-terminal arithmetic.