I was reticent to post about this topic; I polled it on Patreon in December and it got just under 50% while the two topics I did blog, difficult urban geography and cross-platform transfers, got 64% and 50% respectively. However, between how close the vote was and the conversation about the current state of the subway in New York, I felt obligated to explain what’s been going on. The short version is that practically the entire change in subway ridership in New York over the last generation or two has come from the off-peak, and the way American cities set their frequency guidelines off-peak amplify small changes in demand, so that a minor setback can lead to collapse and a minor boost can lead to boom.
The good news is that by setting frequency to be high even if it does not look like ridership justifies it, cities can generate a virtuous cycle on the upswing and avoid a vicious one on the downswing. However, it requires the discipline to run good service even in bad times, when bean counters and budget cutters insist on retrenchment. The Chainsaw Al school of management looks appealing in recessions or when ridership is falling, and this is precisely when people who run transit agencies must resist the urge to cut frequency to levels that lead to a positive feedback loop wrecking the system.
The key to the frequency-ridership spiral is that cutting frequency on transit makes it less useful to passengers, since door-to-door trip times are longer and less reliable. The size of this effect can be measured as the elasticity of ridership with respect to service: if increasing service provision by 1% is demonstrated to raise ridership by e%, we say that the elasticity is e.
Fortunately, this question is fundamental enough to transit that there is extensive published literature on the subject:
- In a classical TRB paper, Armando Lago, Patrick Mayworm, and Matthew McEnroe look at data from several American cities as well as one British one, disaggregating elasticity by frequency, mode (bus or commuter rail), and period (peak or off-peak). The aggregate average value is e = 0.44 for buses and e = 0.5 for commuter rail, but when frequency is better than every 10 minutes, e = 0.22 on average.
- Todd Litman of the advocacy organization VTPI has a summary mostly about fare elasticity but also service elasticity, suggesting e is in the 0.5-0.7 range in the short term and in the 0.7-1.1 range in the long term.
- A paper by Joe Totten and David Levinson includes its own lit review of several studies, including the two above, finding a range of 0.3 to 1.1 across a number of papers, with the lower figures associated with urban service and the higher ones with low-frequency suburban service. The paper’s own research, focusing on transit in Minneapolis, finds that on weekdays, e = 0.39.
One factor that I have unfortunately not seen in the papers I have read is trip length. Frequency is more important for short trips than long ones. This is significant, since when the headway is shorter relative to in-vehicle trip time we should expect lower elasticity with respect to the headway. Waiting 10 minutes rather than 5 minutes for an hour-long trip is not much of an imposition; waiting 30 minutes rather than 15 for the same trip is a greater imposition, as is waiting 10 minutes rather than 5 for a 20-minute trip.
In New York, the average unlinked subway trip is 13.5 minutes long, so the difference between 10 and 5 minutes is very large. Lago-Mayworm-McEnroe cite research saying passengers’ disutility for out-of-vehicle time is 2-3 times as large as for in-vehicle time; the MTA’s own ridership screen states that this penalty is 1.75, the MBTA’s states that it is 2.25, and a study by Coen Teulings, Ioulina Ossokina, and Henri de Groot says that it is 2 in the Netherlands. Figuring that this penalty is 2, the worst-case scenario for off-peak weekday wait time in New York, 10 minutes, has passengers spending more perceived time waiting for the train than riding it, and even in the average case, 10/2 = 5 minutes, it is close. In that case, higher values of e are defensible. Lago-Mayworm-McEnroe have less data about in-vehicle time elasticity and do not attempt to aggregate in- and out-of-vehicle time. But adding everything together is consistent with e = 0.8 relative to speed averaged over the total wait and in-vehicle time, and then e is maybe 0.4 relative to frequency.
The impact of service cuts
If the elasticity of ridership relative to frequency is 0.4, then cutting service by 1% means cutting ridership by 0.4%. If half the operating costs are covered by fares, then revenue drops by 0.2% of total operating expenses, so the 1% cut only saves 0.8% of the total subsidy. Achieving a 1% cut in operating costs net of fare revenue thus requires a 1.25% cut in service, which reduces ridership by 0.5%.
This may not sound too bad, but that’s because the above analysis does not incorporate fixed costs. Rail comes equipped with fixed costs for maintenance, station staffing, rolling stock, and administration, regardless of how much service the agency runs. Lisa Schweitzer uses this fact to defend Los Angeles’s MTA from my charge of high operating costs: she notes that Los Angeles runs much less service than my comparison cases in the US and Europe and thus average cost per train-km is higher even without undue inefficiency. In contrast, bus costs are dominated by driver wages, which are not fixed.
New York does not keep a headcount of transit employees in a searchable format – the Manhattan Institute’s See Through New York applet helps somewhat but is designed around shaming workers who make a lot of money through overtime rather than around figuring out how many people work (say) maintenance. But Chicago does, and we can use its numbers to estimate the fixed and variable costs of running the L.
The CTA has somewhat more than 10,000 workers, split fairly evenly between bus and rail. The rail workers include about 800 working for the director of maintenance, working on the rolling stock, which needs regular servicing and inspections regardless of how often it’s run; 550 working for facilities maintenance; (say) 400 out of 800 workers in administrative capacity like communications, general counsel, purchasing, and the chief engineer’s office; 600 workers in power and way maintenance; nearly 1,000 customer service agents; and 450 workers in flagging, switching, and the control towers. Only 500 workers drive trains, called rapid transit operators or extra board, and there may charitably be another 200 clerks, managers, and work train operators whose jobs can be cut if there is a service cut. A service cut would only affect 15% of the workers, maybe 20% if some rolling stock maintenance work can be cut.
In New York the corresponding percentage is somewhat higher than 15% since trains have conductors. Train operators and conductors together are about 13% of the NYCT headcount, so maybe 20% of subway employees, or 25% with some extra avoidable maintenance work.
What this means is that achieving a 2% cut in subsidy through reducing service requires a service cut of much more than 2%. If only 25% of workers are affected then, even without any frequency-ridership elasticity, the agency needs to cut service by 8% to cut operating costs by 2%.
The Uber effect
The combination of elasticity and fixed costs means that rail ridership responds wildly to small shocks to ridership. For a start, if the agency cuts service by 1%, then operating costs fall by 0.25%. Ridership falls by 0.4%, and thus revenue also falls by 0.4%, which is 0.2% of total operating costs. Thus operating costs net of revenue only fall by 0.05%. The only saving grace is that this is 0.05% of total operating costs; since by assumption fare revenue covers half of operating costs, this saves a full 0.1% of the public subsidy.
Read the above paragraph again: taking fixed costs and elasticity into account, cutting service by 1% only reduces the public subsidy to rail service by 0.1%. A 2% cut in subsidy in a recession requires a brutal 20% cut in service, cutting ridership by 8%. And this only works because New York overstaffs its trains by a factor of 2, so that it’s plausible that 25% of employees can be furloughed in a service cut; using Chicago numbers this proportion is at most 20%, in which case revenue falls one-to-one with operating costs and there is no way to reduce the public subsidy to rail operations through service cuts.
Of course, this has a positive side: a large increase in service only requires a modest increase in the public subsidy. Moreover, if trains have the operating costs of Chicago, which are near the low end in the developed world, then the combined impact of fixed costs and elasticity is such that the public subsidy to rapid transit does not depend on frequency, and thus the agency could costlessly increase service.
This is relevant to the Uber effect – namely, the research arguing that the introduction of ride-hailing apps, i.e. Uber and Lyft, reduces transit ridership. I was skeptical of Bruce Schaller’s study to that effect since it came out two years ago, since the observed reduction in transit ridership in New York in 2016 was a large multiple of the increase in total taxi and ride-hailing traffic once one concentrated on the off-peak and weekends, when the latter rose the most.
But if small shocks to ridership are magnified by the frequency-ridership spiral, then the discrepancy is accounted for. If a shock cuts ridership by 1%, which could be slower trains, service disruptions due to maintenance, or the Uber effect, then revenue falls 1% and the subsidy has to rise 1% to compensate. To cover the subsidy through service cuts requires a 10% cut in service, further cutting ridership by 4%.
Off-peak service guidelines
The above analysis is sobering enough. However, it assumes that service cuts and increases are uniformly distributed throughout the day. This is not the actual case for American transit agency practice, which is to concentrate both cuts and increases in the off-peak.
Unfortunately, cuts in off-peak service rather than at rush hour do not touch semi-fixed labor costs. The number of employees required to run service is governed by the peak, so running a lot of peak service without off-peak service leads to awkward shift scheduling and poor crew utilization. Higher ratios of peak to base frequency correlate with lower total service-hours per train driver: in addition to the examples I cite in a post from 2016, I have data for Berlin, where the U-Bahn’s peak-to-base ratio is close to 1, and there are 829 annual service-hours per driver.
I discussed the fact that the marginal cost of adding peak service is several times that of adding off-peak service in a post from last year. However, even if we take rolling stock acquisition as a given, perhaps funded by a separate capital plan, marginal crew costs are noticeably higher at the peak than off-peak.
In New York, the rule is that off-peak subway frequency is set so that at the most crowded point of each route, the average train will be filled to 125% seated capacity; before the round of service cuts in 2010 this was set at 100%, so the service cut amounted to reducing frequency by 20%. The only backstop to a vicious cycle is that the minimum frequency on weekdays is set at 10 minutes; on weekends I have heard both 10 and 12 minutes as the minimum, and late at night there is a uniform 20-minute frequency regardless of crowding.
Peak frequency is governed by peak crowding levels as well, but much higher crowding than 125% is permitted. However, the busiest lines are more crowded than the guidelines and run as frequently as there is capacity for more trains, so there is no feedback loop there between ridership and service.
The saving grace is that revenue is less sensitive to off-peak ridership, since passengers who get monthly passes for their rush hour trips ride for free off-peak. However, this factor requires there to be substantial enough season pass discounts so that even rush hour-only riders would use them. Berlin, where U-Bahn tickets cost €2.25 apiece in bundles of 4 and monthly passes cost €81, is such a city: 18 roundtrips per month are enough to justify a monthly. New York is not: with a pay-per-ride bonus a single ride costs $2.62 whereas a 30-day pass costs $121, so 23.1 roundtrips per month are required, so the breakeven point requires a roundtrip every weekday and every other weekend.
New York subway ridership evolution
The subway’s crisis in the 1970s reduced ridership to less than 1 billion, a level not seen since 1918. This was on the heels of a steady reduction in ridership over the 1950s and 60s, caused by suburbanization. In 1991, ridership was down to 930 million, but the subsequent increase in reliability and fall in crime led to a 24-year rally to a peak of 1,760 million in 2015.
Throughout this period, there was no increase in peak crowding. On the contrary. Look at the 1989 Hub Bound Report: total subway ridership entering Manhattan south of 60th Street between 7 and 10 am averaged about 1 million, down from 1.1 million in 1971 – and per the 2016 report, the 2015 peak was only 922,000. Between 1989 and 2015, NYCT actually opened a new route into Manhattan, connecting the 63rd Street Tunnel to the Queens Boulevard Line; moreover, a preexisting route, the Manhattan Bridge, had been reduced from four tracks to two in 1986 and went back to four tracks in 2004.
Nor was there much of an increase in mode share. The metropolitan statistical area’s transit mode share for work trips rose from 27% in 2000 to 30% in 2010. In the city proper it rose from 52% in 1990 to 57% in 2016. No: more than 100% of the increase in New York subway ridership between 1991 and 2015 was outside the peak commute hours, and nearly 100% of it involved non-work trips. These trips are especially affected by the frequency-ridership spiral, since frequency is lower then, and thus a mild positive shock coming from better maintenance, a lower crime rate, and perhaps other factors translated to a doubling in total ridership, and a tripling of off-peak ridership. Conversely, today, a very small negative shock is magnified to a minor crisis, even if ridership remains well above the levels of the 1990s.
The way out
Managers like peak trains. Peak trains are full, so there’s no perception of wasting service on people who don’t use it. Managers also like peak trains because they themselves are likelier to ride them: they work normal business hours, and are rich enough to afford cars. That current NYCT head Andy Byford does not own a car and uses the city’s transit network to get around scandalizes some of the longstanding senior managers, who don’t use their own system. Thus, the instinct of the typical manager is to save money by pinching pennies on off-peak service.
In contrast, the best practice is to run more service where possible. In Berlin, nearly all U-Bahn trains run every 5 minutes flat; a few lines get 4-minute peak service, and a few outer ends and branches only get half-service, a train every 10 minutes. At such high frequency, the frequency-ridership spiral is less relevant: an increase to a train every 4 minutes would require increasing service by 25%, raising costs by around 5% (Berlin’s one-person crews are comparable to Chicago’s, not New York’s), but not result in a significant increase in ridership as the shorter headway is such a minute proportion of total travel time. However, New York’s 10-minute off-peak frequency is so low that there is room to significantly increase ridership purely by running more service.
In 2015 I criticized the frequency guidelines in New York on the grounds of branching: a complexly branched system must run interlined services at the same frequency, even if one branch of a trunk line is somewhat busier than the other. However, the frequency-ridership spiral adds another reason to discard the current frequency guidelines. All branches in New York should run at worst every 6 minutes during the daytime, yielding 3-minute frequency on most trunks, and the schedules should be designed to avoid conflicts at junctions; non-branching trunk lines, that is the 1, 6, 7, and L trains, should run more frequently, ideally no more than every 4 minutes, the lower figure than in Berlin following from the fact that the 1 and 6 trains are both local and mostly serve short trips.
Moreover, the frequency should be fixed by a repeating schedule, which should be clockface at least on the A train, where the outer branches would only get 12-minute frequency. If ridership increases by a little, trains should be a little more crowded, and if it decreases by a little, they should be a little less crowded. Some revision of schedules based on demand may be warranted but only in the long run, never in the short run. Ideally the system should aim at 5-minute frequency on every route, but as the N, R, and W share tracks, this would require some deinterlining in order to move more service to Second Avenue.
This increase in frequency is not possible if politicians and senior managers respond to every problem by cutting service while dragging their feet about increasing service when ridership increases. It requires proactive leadership, interested in increasing public transit usage rather than in avoiding scandal. But the actual monetary expense required for such frequency is not large, since large increases in frequency, especially in the off-peak, mostly pay for themselves through extra ridership. The initial outlay required to turn the vicious cycle into a virtuous one is not large; all that is required is interest from the people in charge of American transit systems.