Deinterlining: Some Quantitative Evidence

If you have ever crawled on a 2, 3, 4 or 5 train around Franklin Avenue, sat interminably on an R train at 34 St, or waited for the ever-irregular F train, you have likely suffered from interlining. Unlike most metro systems where routes are isolated on their own infrastructure, in New York most trains share tracks (“interline”) with at least two other routes over the course of a given run. This network design offers many one-seat rides, but comes at a significant cost to the subway’s operational health. Since the mid-2010s, many transit advocates in New York have been calling for a change in the system’s approach to routing, producing a number of proposals on how best to simplify these routes (“deinterline”). As visible as these delays may be to the everyday rider, advocates’ reform efforts have as of yet failed to produce a body of public quantitative research on the potential benefits of simplification, making more specific articulations of its benefits rather difficult. Luckily, a set of service changes over the past two years provides us with natural experiments to fill this gap. These case studies’ results further evidence the potential benefits of restructuring our subway routes.

Conceptual Benefits

Generally, deinterlining is proposed as a solution to three linked operations issues in the system. In their simplest forms, the arguments go something like this:

  • Reliability. By reducing the number of merge points (and thus potential merge delays) in the system, a deinterlined service plan may provide significantly more consistent service than today.
  • Speed. Reducing merge delays might also have a positive impact on train speeds — if riders aren’t often waiting for a train in front of them to merge, average trip times might improve.
  • Capacity. With faster and less fragile service, deinterlining may allow NYCT to run more trains on congested subway lines without having to make significant capital investments.

Onto these, one can additionally layer some predictions about where the savings from deinterlining will be greatest. Most obvious are places where deinterlining eliminates diverging movements, which often require trains to slow well below track speed. Additionally, one should expect deinterlining to be a powerful tool at busier junctions. Logic would have it that the probability of a merge delay rises with the number of trains, thus merges which see more trains across the day (and also have more even ratios between merging services) should have the largest returns to simplification.

One final note: estimating the impact of service interventions on capacity is difficult, as unless train volumes are set near the limit of infrastructure capability, the sorts of ‘stress testing’ required to get a good read of capacity are hard to come by. I will instead focus on speed and reliability, with the understanding that capacity is strongly related to these two variables.

Assessing Deinterlining

To ascertain the existence and magnitude of these hypothesized effects, one can analyze train performance data from New York City Transit’s GTFS-RT data feeds. Philippe Vibien has been maintaining an archive of these since March of 2019, allowing a retrospective view into service performance on all lines in that time period. These data, intended more for use by app developers wishing to inform users about the status of trains, are not perfect, but can provide a good-enough approximation of how trains moved through the system. For each of these studies, I isolated a period of about three months bracketing each service change to provide a large sample of train movements while minimizing the probability that other service or infrastructure changes (like speed increases) confound the analysis. I also only used weekday daytime (7A-7P) data, so as to limit the impact that disruptions related to nighttime and weekend service changes have on estimated impacts. Along with the all-day impact estimation, these tables will also break out rush hours as a means of confirming the theorized relationship between train volume and merge impact, and to further control for the impacts of service changes.

In structuring the measurements in the following case studies, a few finer points about the data’s format are relevant. On the A division (the numbered lines and the 42 St Shuttle) and the L train, train position data are derived from the signal system, allowing a relatively high level of accuracy. On the B division (the letter lines), data come from Bluetooth beacons mounted on the train which ‘ping’ receivers in stations when in range. Data from these pipelines is presented to users with station arrival and departure times presented separately, meaning that measuring running times from arrival at the first station to arrival at the next may yield different results than departure to departure analyses. For merges where delays may manifest as excessive time spent at the station(s) before the merge, I use arrival-arrival running times and/or extend the limits of analysis further upstream to capture all associated congestion. Likewise, at locations on the B division where a train held just outside a station by a merge may ‘ping’ the station’s receiver and register as having arrived, I use departure-departure running times so as to capture the full effect of a merge delay.

Finally, a word about metrics. The obvious ones to use in a study of speed and reliability are average travel time and the standard deviation of travel time. However, research has shown that using median travel times for speed and the difference between the median and the 90th percentile for reliability is more predictive of peoples’ transportation choices than using means and standard deviations (a less obfuscated way to think about the latter metric is that it is the amount of padding you need to add to normal travel times so that you are not late on more than one of every ten trips). This study will use those metrics in its analysis.

Case Study #1: 4th Avenue Structural Repairs

From July 30, 2018 to July 29, 2019, NYCT took the express tracks of the Fourth Avenue Subway in Brooklyn out of service south of 36 St station for structural repairs. This forced all N trains, which usually use those tracks, to merge and share tracks with the local R train from 36 St to 59 St in both directions, temporarily altering merging patterns on the corridor. Specifically, the change added a merge between southbound N and R trains north of 36 St and a merge between northbound N and R trains south of 59 St, and moved the merge between northbound D and N trains from south of 36 St to north of 36 St.

No major service level adjustments were made as a part of this project; the small changes in N train frequencies during the PM shoulder period shown in the second chart resulted from diversions unrelated to the 4 Avenue project. Ridership was also relatively stable: at 36 and 59 Sts (the stations most implicated in the analyses described below) weekday ridership through the summer of 2019 remained essentially constant.

To assess the impacts of these additional merges and test some of the hypotheses laid out previously, I took measurements over the following line segments.

Route/DirectionStation PairMerge AssessedDiverging Move?Analysis Type
R/SouthProspect Av-36 StSouthbound N/R (removed)NoDeparture-Departure
N/SouthAtlantic Av-36 StSouthbound N/R (rem.) YesDeparture-Departure
R/NorthBay Ridge Av-59 StNorthbound N/R (rem.) NoDeparture-Departure
N/North8 Av-59 StNorthbound N/R (rem.) YesDeparture-Departure
N/North36 St-Atlantic AvNorthbound D/N (rem.) YesArrival-Arrival
D/North36 St-Atlantic AvNorthbound D/N (rem.) NoArrival-Arrival

Using data from June 15th to September 15th, 2019, I ran analyses on the AM peak (7A-10A), PM peak (4P-7P), and overall day (7A-7P). These yielded the following results (for a graphic version, see here):

Route/DirectionStation Pair Change in Median Running TimeChange in 90-50 Spread
R/SouthProspect Av-36 St-28s-32s
N/South Atlantic Av-36 St -25s-70s
R/North Bay Ridge Av-59 St 10s21s
N/North 8 Av-59 St 6s14s
N/North 36 St-Atlantic Av -75s-33s
D/North 36 St-Atlantic Av -22s-66s
All-day results

Route/DirectionStation Pair Change in Median Running TimeChange in 90-50 Spread
R/SouthProspect Av-36 St-29s-26s
N/South Atlantic Av-36 St -35s-63s
R/North Bay Ridge Av-59 St 0s0s
N/North 8 Av-59 St -15s-21s
N/North 36 St-Atlantic Av -50s-50s
D/North 36 St-Atlantic Av -27s-82s
AM peak results

Route/DirectionStation Pair Change in Median Running TimeChange in 90-50 Spread
R/SouthProspect Av-36 St-40s-42s
N/South Atlantic Av-36 St -35s-50s
R/North Bay Ridge Av-59 St 1s3s
N/North 8 Av-59 St -1s-2s
N/North 36 St-Atlantic Av -70s-17s
D/North 36 St-Atlantic Av -25s-70s
PM peak results

Per these analyses, deinterlining yielded significant rider time savings and reliability gains, especially where diverging moves were eliminated. However, these estimated benefits were variable; 59 St showed little change, and estimations at 36 St were rather inconsistent. The sources of heterogeneity in this dataset are difficult to pin down conclusively, but a few potentially confounding factors are worth mentioning.

The first of these is dispatcher interventions. The role they play in confounding analysis is most relevant to the cases of the southbound R at 36 St and the northbound N and R at 59. The stations at the ends of both these analysis segments are express stations where dispatchers often hold trains to facilitate cross-platform connections. Critically, when the N and R were sharing tracks, these cross-platform connections would not have been possible (or at least advisable, for fear of holding up following trains); the resumption of normal service also reenabled holding at 36 and 59. For both segments, I had originally elected to use a departure-departure methodology to calculate merge impacts so as to minimize the probability that trains pinging beacons in the station (and thus registering as arrived) while waiting outside the station for a clear merge slot would cause an understatement of merge impacts in my calculations. It remains plausible that mis-pings are a source of error, but an arrival-arrival methodology in these cases yields significantly greater estimated impacts of deinterlining. Across the 7A-7P time period, running arrival-arrival analysis increased the 36 St R 90-50 estimation from -32 to -65 seconds, and the running time savings from -28 to -35 seconds. At 59 St, the 90-50 goes from 14 to -25 and running time from 6 to -15 when swapping techniques on the N, and an increase in 90-50 benefits from 21 to -27 and in running time savings from 10 to -15 on the R. Together, these results suggest that there was a significant increase in the time those trains spent at 36 and 59 St after the conclusion of structural work. It seems likely that the resumption of normal service patterns brought a corollary increase in the number of station holds to make connections, leading to an illusion of increased running times.

Of course, even with those adjustments for holds, there remains variability in these estimations. To take the example of the revised N figure at 59 St, the estimated running time savings from merge elimination remain much smaller there than at 36 St. Schedule adherence and merge typology seem to hold at least part of the explanation for this gap. When active, 59 St was the first merge on northbound N and R trains, so trains arrived at the merge relatively close to schedule. During the project, northbound peak-hour N trains arriving at 59 St exhibited 30% less variability (measured here with standard deviations) relative to their scheduled arrival times than they did at 36 St. This is not to say variability at 59 St was by any means insignificant, but instead that some increment of better merge performance may have been realized through this relatively higher adherence to schedule. And alongside this consistency advantage, the merge at 59 St may have also benefitted from its simplicity. In both directions, the 36 St merged involved the N diverging from one service and immediately merging with another, whereas 59 St was a rather simple convergence of two lines. This difference in the number of variables and constraints faced by dispatchers may have aided the better showing from 59 St.

All told, the 36 St case study yields a few important (tentative) conclusions on the hypotheses presented earlier. First, its benefits do, in fact, exist. In all cases, eliminating merges yielded significant running time and reliability gains: median travel times fell generally by about 30-60 seconds, and reliability improved by a similar magnitude. Second, its benefits do seemingly scale with whether or not trains make a diverging move — witness the savings on northbound N trains at 36 St. Finally, benefits during peak hours tended to be slightly greater than the all-day average, evidencing a moderate relationship between train frequency and merge delays.

Case Study #2: COVID on the Broadway Line

During the first wave of the 2020 COVID-19 pandemic, New York City Transit reduced transit service due to a shortage of available crews. This essential service plan reduced frequencies on all subway services, changed service patterns on some, and eliminated three routes entirely: the B, W and Z trains. These service alterations impacted merging patterns on a number of key corridors, including the Broadway Line. On the corridor, the suspension of the W train eliminated the merge between northbound R and W trains north of Whitehall St. Further up the corridor, N trains were routed onto the local tracks beginning at Prince St to compensate for the loss of W service, shifting the N’s express local crossover point to that station from 34 St. When (more) normal service resumed between May 27, 2020 and June 8, 2020, these merges were restored or shifted to their original locations.

As in the first case study, I have composed a series of segment-level analyses to assess merge impacts.

Route/DirectionStation PairMerge AssessedDiverging Move?Analysis Type
R/NorthCourt St-Rector StNorthbound R/W (added)NoArrival-Arrival
R/NorthCanal St-8 StNorthbound N/R (removed)NoArrival-Arrival
R/North28 St-49 StNorthbound N/R/W (added)NoArrival-Arrival
N/North34 St-49 St Northbound N/R/W (added) YesArrival-Arrival
N/South49 St-34 StSouthbound N/Q (added)YesDeparture-Departure
Q/South57 St-34 StSouthbound N/Q (added) NoDeparture-Departure
Q/South14 St-Canal StSouthbound N/Q (removed)NoDeparture-Departure

Running the same analyses as before yielded the following results (for a graphic version, see here):

Route/DirectionStation Pair Change in Median Running TimeChange in 90-50 Spread
R/North Court St-Rector St 20s90s
R/North Canal St-8 St -8s-77s
R/North 28 St-49 St 33s105s
N/North 34 St-49 St 50s55s
N/South 49 St-34 St 60s55s
Q/South 57 St-34 St 20s80s
Q/South 14 St-Canal St 0s-40s
All-day results

Route/DirectionStation Pair Change in Median Running TimeChange in 90-50 Spread
R/North Court St-Rector St 20s99s
R/North Canal St-8 St -2s-48s
R/North 28 St-49 St 36s117s
N/North 34 St-49 St 57s78s
N/South 49 St-34 St 60s65s
Q/South 57 St-34 St 20s65s
Q/South 14 St-Canal St 0s-25s
AM peak results

Route/DirectionStation Pair Change in Median Running TimeChange in 90-50 Spread
R/North Court St-Rector St 21s98s
R/North Canal St-8 St -13s-98s
R/North 28 St-49 St 35s110s
N/North 34 St-49 St 50s50s
N/South 49 St-34 St 65s55s
Q/South 57 St-34 St 20s75s
Q/South 14 St-Canal St 0s-50s
PM peak results

The results from these analyses indicate significant returns to deinterlining. All show large reliability impacts, and with the exception of the Prince St merge, all show significant running time savings as well. Excepting Prince St, these data moreover seem to be slightly less heterogeneous than the 4th Avenue case: where diverging moves were not eliminated, running time improved by 20-35 seconds; where they were, it improved by 50-60 seconds. However, the deeply abnormal nature of the COVID pandemic means that certitude about these results’ validity requires engagement with the potentially confounding conditions in the system’s operating environment during the analysis period.

The first potential hidden variable problem is that of ridership. Long dwell times at stations are normally a major source delay on the subway, especially during peak hours. With total system entries at approximately 10-20% of normal levels around the time of the change, these delays had all but disappeared. We can see this impact by comparing summer of 2019 running times with those from 2020 after the resumption of normal patterns: in 2020, R trains’ running times in Midtown were 10-15% faster, and 90-50 spreads were as much as 60% smaller. At merges, these different conditions had somewhat more ambiguous effects. All of the above analysis segments saw trains move more quickly in 2020 than they did in 2019, but most also saw their variability increase. The result on variability is likely unrelated to merge performance: dispatchers’ efforts to correct trains’ chronic earliness with holds — which increase some trains’ running time — often took place at 34 St, Canal St and Whitehall St. Nevertheless, it is important to adjust our conclusions about merges’ impact on running times slightly upwards, and their impact on variability slightly downwards.

The summer resurgence in subway use is also relevant to our analysis. These deinterlinings took place alongside a partial reopening of the city after the worst of the first wave of the pandemic had passed. Consequently, ridership increased dramatically (in relative terms) from May to June 2020, introducing another potential inconsistency. Evidence suggests that this resurgence had little impact on performance, however. Ridership levels have significant impacts on speeds only above some critical level where the amount of time it takes people to get on and off crowded trains rises above the time it takes to make announcements and operate doors at each stations. That train running times (save for merge impacts) remained largely unchanged after the resumption of normal service despite a doubling of daily ridership suggests increases in passenger loads failed to reach that critical point.

Turning to service level considerations, these deinterlining test cases took place alongside major service frequency changes. The N, Q and R trains experienced throughput reductions of various magnitudes during the first wave of the pandemic to accommodate COVID-driven crew shortages and reduced ridership. Levels were restored to normal alongside the service pattern changes analyzed here, introducing yet another confounding factor into analysis. Looking at a chart comparing weekday train performance before and after the June resumption of W service will show that changes in running time were, for the most part, localized in areas with merges. However, northbound N and R trains did suffer tangible increases in travel time in the section of track shared between the N, R and W trains between 42 St and Queens. These delays seem to have been moderated by infrastructure conditions; running time increases were localized in the 49 St-5 Av/59 St area, which has rather restrictive signaling. Because few of the analysis segments contain signaling as constraining as is present in the 49 St-5 Av stretch, it seems reasonable to rule out much potential error from line congestion.

Finally, there is the matter of the Prince St merge. This segment was the only one not to show significant running time savings from deinterlining — though there was a large reduction in running time variability. This seems to be related to low service levels during the pandemic, when the merge was active. N and R trains’ reduced schedules ran 5-6 trains per hour all day, rather than those lines normal 10 tph peaks and 6 tph middays. Running fewer trains in turn reduces the probability of merge delays, which might explain the relatively smaller degree of impact. Verifying the significance of this effect is trivial: look at a merge that existed unchanged during reduced service. The evidence from those merges bears out this conclusion. At Dekalb Avenue, for example, running times on the northbound N and Q increased by 30-45 seconds (and variability by 25-30) following the resumption of normal service frequencies in June.

Putting all these pieces together, the Broadway case furthers the conclusions reached on 4th Avenue. As before, deinterlining realized tangible speed and reliability benefits for riders. Clearer in this case was the relationship between time savings and the elimination of diverging moves — note that N trains benefitted more than Qs or Rs at 34 St. The frequency-merge impact relationship further gained evidence, albeit mostly through the weakness of change at the Prince St merge rather than the peak/off peak difference as on 4th Avenue. It is possible that some of the additional variables considered to explain the heterogeneity in the earlier case played a role here as well, but whatever role they did play in creating inter-merge differences was minimal; Broadway’s results were significantly more consistent than 4th Avenue’s.


Unsurprisingly, deinterlining works. Reducing merges at even the moderate frequencies seen in these two case studies can directly save riders 20-70 seconds per trip, and indirectly save them as much as another two minutes by reducing variability that needs to be accounted for in travel planning. Setting aside some of the micro-level nuances to which I have devoted much text, these figures are likely underestimates of the general benefits of deinterlining. When trains suffer a delay, they often make other trains late by arriving at merges and terminals off schedule. This cascading effect means that merges have network-scale impacts. For lack of sufficient and high-quality data, I have not attempted to quantify the scale and nature of these cascades here, but it remains a critically important element of deinterlining’s case to keep in mind: deinterlining is a local intervention with regional implications.

All of these benefits must be weighed against the costs of additional transfers for some riders, and the additional operating spending that may be required to implement deinterlined service plans. Yet merge reduction is, in the grand scheme of transit investments, incredibly easy and cheap. If I may make an reductive comparison, CBTC on the 7 line reduced running time on local trains by about 8 percent, or 3 minutes. Speed improvements are only one facet of the case for CBTC, but it is notable that deinterlining’s time savings can be nearly half the size of those realized through the installation of a $600 million dollar signal system. Of course, in reality, CBTC’s benefits scale with line length, and merge elimination is hardly a substitute for long overdue signal modernization (I would in fact classify CBTC and deinterlining as complements, as CBTC works best in simple operating environments). However, if we wish to maximize transit improvement, deinterlining is a potent and readily available tool by which we may advance our efforts.

Expanding American Rail Capacity

Updated 2021-04-09 @ 15:32 with new information on San Diego’s timetabling efforts.

Amtrak recently released a map outlining a $25 billion proposal for expanded passenger rail service in the US. With a focus on improving corridor services, this map would increase Amtrak’s focus on the sorts of short and medium-distance markets where well-operated rail services can be most competitive with driving and flying. Though certainly too limited an intervention relative to the overwhelming need for low-carbon, car-free transport in this country, the map does identify a number of important corridors for rail service expansion in the US, providing a useful starting framework for further improvement.

The largest barrier to realizing this vision is the fact that Amtrak does not own most of the routes on which they seek to improve service. Outside the Northeast, the national passenger carrier relies on rights to run its trains over freight trackage. This saves Amtrak the hassle of maintenance, but the resulting freight-passenger mixing (which takes place on the freight railroads’ terms) is a major source of delays. It also presents a limit on network growth: if adding (more) passenger service to a line will have a significant negative impact on freight operations, freight railroads may deny Amtrak operating rights for that additional service. Amtrak seemingly is testing a more aggressive approach towards these operating rights, but it seems a fair assumption that Amtrak will perpetually face some level of frequency and reliability constraint from freight intermingling in the current rail operations paradigm.

This begs the question: how do we go about implementing Amtrak’s map? Though the median rail mile in the US is lightly utilized, commencing service on a number of key corridors (for example the Los Angeles-Phoenix-Tucson line) will require Amtrak to gain access to extraordinarily busy freight arteries. If the passenger carrier is to run trains at any reasonable frequency over these routes — much less expand beyond them — it will either have to pay its own way with new infrastructure, or find some way to ‘expand the pie’ of rail capacity. We currently tend to accept the former approach, but in the latter there lies massive, unrealized potential. Undertaking national rail capacity planning and beginning to shift America’s railroads towards fully scheduled operations could facilitate a massive expansion of medium-speed passenger service without the construction of much new infrastructure.

The Issue, in Greater Detail

The root of the freight-passenger mixing issue lies not in some nefarious desire on the part of the freight railroads to delay passenger trains, or one on the part of Amtrak to inconvenience their hosts. Instead, the problem stems from the fact that Amtrak and freight carriers exist in fundamentally different operating paradigms.

Most obvious is the fact that American freight trains are slow. The fastest of freights on the country’s best maintained lines top out at 70mph, with most train-miles spent cruising below 50. Passenger trains — theoretically — can run much faster. On relatively tangent track, conventional trains can reach speeds well over 100mph. This inevitably causes conflicts: passenger trains will catch up with freights, and without another available track with which to overtake the freight, the passenger train may have to slow down.

Yet this difference in speed is not the core of the issue. Rather, it is scheduling which drives the fundamental inability of railroads to make these interactions work. Passenger trains are generally run on an exact timetable. Trains are scheduled down to the minute, and are expected to meet those schedules with little room for error. Precision is important not just because us humans dislike waiting, but also because reliable passenger operations generally demand an exact operating plan: schedules lay out where trains moving in opposite directions on single track will pass, where others will overtake, and where still others will make cross-platform connections.

American freight railroads operate in a different environment. Most modern railroads run at least some trains with schedules, but these schedules are hardly the same as those used for passengers. Hours-wide error bars around scheduled arrival times are typical; schedules here serve more as a rough guideline for customers and a means of organizing connections between trains, rather than a precise operating plan for each individual train movement.

Mixing these two operating models often leads to delays and capacity loss for both freight and passenger users. I made a graphic to help visualize some of these issues; the upshot is that it is very expensive to build infrastructure that will allow passenger operations to be resilient to highly variable freight movements. This is especially true on busy or capacity-constrained lines, where dispatchers have fewer opportunities to fiddle with freight train movements to clear a path for passengers.


There are essentially three ways of fixing this problem:

  1. Build lots of new infrastructure along freight lines
  2. Separate passenger and freight traffic onto separate lines
  3. Run freight trains on passenger-style schedules

Let’s take a close look at each.

Building New Trackage

Historically, this model has been the one pursued when adding passenger service. The Chicago-Saint Louis corridor provides a useful example. Upgrades to that corridor (whose freight traffic load is not terribly heavy) in the 2010s involved building 24 miles of additional trackage just to accommodate the greater interference that would come with higher speeds — service levels have remained constant since the aughts. Even with this new infrastructure, trouble continues: before the pandemic, the Lincoln Service’s on-time percent hovered in the mid-70s. This mixed approach is the best in freight-owned corridors with light traffic and significant excess capacity (eg. the portion of the Chicago-Des Moines corridor which runs on Iowa Interstate trackage), but where both passenger and freight volumes will be nontrivial, a different strategy is needed. High frequency/high(er) speed passenger service and high variability freight service simply do not mix well; on these busier corridors, we will either end up building massive amounts of new infrastructure to facilitate mixing, and/or suffer delays in perpetuity.

Separating Passenger and Freight

The North American rail network was built up by a multitude of railroads, with little centralized coordination or infrastructure sharing. This mode of development means that we have built significant amounts of duplicative rail infrastructure. Though some such lines have been abandoned over the years, significant amounts of parallel trackage remain — and often continue to be operated by separate railroads with little coordination. These corridors present a unique opportunity: by separating passenger and freight traffic onto parallel lines, we can realize the benefits that come with exclusive passenger ownership without severely impeding climate-friendly rail goods movement.

Take the example of the Buffalo-Cleveland corridor. The two large eastern roads (CSX and Norfolk Southern) own lines between these two cities, serving same intermediate markets, and often running but feet from each other. If we could encourage the two railroads to work together, we could run the handful of daily Norfolk Southern trains which use their line over CSX’s tracks, and gain a (nearly) freight-free passenger corridor between two large cities — one of the pairs identified for improvements on Amtrak’s map.

This is a relatively extreme example of duplication, but hardly the only such case. Across North America, opportunities for this sort of redistribution abound; whether it be Chicago-Milwaukee, Chicago-Porter, Portland-Spokane, Detroit-Toronto or otherwise, myriad corridors present opportunities for planners to realize passenger capacity through increased inter-railroad cooperation. This is not to say that investment needs magically disappear when you put passenger and freight on separate routes — network restructuring will almost always involve investments in capacity and connectivity — but when the dust settles, you are left with a passenger corridor whose future service levels and service consistencies are unconstrained by freight movement, and a freight corridor which can operate without the delays and capacity losses associated with passenger service.

Notably, this is an approach which has been implemented with great success before. Until the late 1970s, Amtrak’s Northeast Corridor was extremely busy with freight traffic, handling dozens of daily freight trains for the Pennsylvania Railroad (and later, Penn Central and Conrail) on top of the Corridor’s heavy passenger traffic. For all the reasons discussed above, this was a rather inconvenient arrangement.

When the federal government took over six bankrupt northeastern railroads in 1976 and created Conrail, they made it a priority to reroute freight off of the corridor. After upgrading roughly parallel assemblage of routes through the region, and restructuring yards and interchanges to feed it, Conrail shifted freight traffic off of the Corridor, and today, most of Amtrak’s route is free of through freight movements.

Another example comes from Ontario. In the 1960s, Canadian National used a combination of existing trackage and new construction to create a circumferential line around Toronto, allowing its trains to access a new yard north of the city and avoid congestion in Downtown Toronto. This freed up significant capacity on the legacy routes radiating out of downtown, which in turn allowed the commencement of commuter rail service and the eventual purchase of many of those corridors by the regional transit agency (Metrolinx) in support of rail service expansion.

That both these examples come from railroads rerouting trains within their network should not be lost on us: in corridors where the duplication comes from parallel routes of different railroads, implementing traffic separation would require gaining buy-in from multiple freight railroads and would necessitate a willingness to share infrastructure. This is certainly not without precedent or existing frameworks, but it does add another level of complexity in negotiations.

For this reason, legislating a greater governmental role in the management of the rail network is likely necessary to realize the full potential of duplication, or any other major intervention in rail capacity planning. Inter-railroad disputes are simply much easier to handle and forecast within a stronger planning framework, a model which separates infrastructure management from rail operations, or simply a nationalized network. As we evaluate those frameworks, simply exploring these opportunities where they exist seems a worthwhile investment of our time and energy; it is likely that paying railroads to play nice with each other in key corridors will still provide a less expensive (and more reliable) path to network growth than our current mixed traffic model.

Scheduling Freight

Another approach to this problem comes through scheduling freight movements. The United States is not alone in running freight trains flexibly — some Eastern European countries, for example, do as well — but in nations where high volumes of freight and passenger traffic must mix, scheduling is the solution. Switzerland provides an instructive example. Swiss rails handle about twenty percent of intercity ground passenger-kilometers, as well as about thirty five percent of freight ton-kilometers. The former figure is (by nearly a factor of two) the highest in Europe; the latter is about equivalent to the US’s freight rail mode share. Swiss railroads manage these interactions by scheduling everything. Instead of running trains on a flexible schedule like in the US, freight movements are rigidly timetabled like passenger trains. This allows precise, high-capacity and high-reliability freight and passenger movements — saving the country the expense of building costly new infrastructure, and increasing the competitiveness of both rail services’ offer.

An example of such an implementation in North America comes to us from San Diego. After I posted the original version of this piece, Brendan Dawe flagged some fascinating materials about service coordination and design on the rail line linking Orange County to San Diego. In response to planning goals of increasing passenger rail frequencies and allowing more freight movement to and from the Port of San Diego, BNSF and the local commuter operator developed short, medium and long term timetable plans for the route, and are building their future investment priorities around those efforts. By tightly scheduling freight movements, the corridor’s users will realize increased capacity with comparatively low levels of new investment. Needless to say, this plan (which was implemented in consultation with Deutsche Bahn), provides a useful domestic proof of concept in strict freight timetabling — one which can serve as a guide for similar efforts.

In analyzing cases like this one, we should not lose sight of potential barriers to timetabling. Fully structured freight operations may cause freight railroads’ operating costs to increase in the short run as their ability to make operating plan adjustments for demand is reduced, and as their trains are forced to follow stricter performance guidelines (exemplified in the second slide above). A shift to timetabled operations would also require the development of new network planning competencies among American railroads; presumably, in the long run, freight carriers would not want to rely on Deutsche Bahn’s planning capacity.

Yet San Diego’s experience also speaks to the broader potential upsides of scheduled operation. With timetabling, passenger movement over freight trackage becomes significantly easier to coordinate and cheaper to implement. Moreover, freight railroads likely would realize a medium and long run commercial benefit from a move to fully structured operations. Other American research has (unsurprisingly) found that timetables increase track capacity, reduce long-term investment needs, and perhaps most importantly, reduce the service variability that so hobbles’ railroads ability to compete for freight in the age of ‘just-in-time’ logistics. Implementing fully scheduled rail operations is unquestionably a significant change in the short run, but is a change which may end up being key to affecting a shift away from road transport in both passenger and freight movement.


America needs a robust intercity mass transit system, and fast. To decarbonize and shift the nation away from its dependence on automobiles, we will need to spend extensively (and economically) on expanded rail and bus service to better knit American communities together. Unlocking passenger rail capacity on our nation’s freight trackage is only one part of a broader set of necessary changes for our surface transportation system, but it is nevertheless essential. These corridors will not provide us with the high speed network we so sorely need, but they will remain a cost-effective means of expanding the reach of medium-speed rail, and the most expedient means of entering urbanized areas. As we develop infrastructure roadmaps for the coming decades, transforming rail capacity planning thus should be an item near the fronts of our minds.

Industrial Sprawl

Back in September, I went on the podcast Well There’s Your Problem to talk about the relationship between freight transportation and industrial sprawl. Ever since, I have been meaning to follow up with some more extended written thoughts on the subject; this post is my much-belated attempt at that.

Of late, there has been much discussion about e-commerce, warehousing, essential work, and the future of cities. Whether it be retail space becoming pop-up warehouses, or the general uptick in demand for urban industrial space, these shifts would suggest that warehousing and distribution are likely to become ever more visible presences in the cores of America’s cities in the coming years.

What makes this trend remarkable, of course, is the degree to which it constitutes a dramatic movement against the past century’s currents in the geography of American goods. As is true for much else in America through those hundred years, the story of freight, warehousing, and industry has been one of dramatic (and destructive) sprawl — both within and between metropolitan areas. Take Chicago, for example. In 1947, about 75% of manufacturing employment in the region was located within the city limits of Chicago. Fifty years later, in 1997, that proportion had almost flipped — all while the Chicago region had lost more than half of its total manufacturing employment.

Though certainly not without nuance (I think, for example, most would agree that moving more noxious industries away from people is a good thing), this transition has been a damaging one; industrial sprawl has done great damage to equity and the environment. This dramatic locative shift is one which thus should be central to our understanding of the postwar American city, and is one that I see as being an essential framing for discussions about freight movement. For though industrial sprawl was driven by a number of factors, its extent and form was fundamentally shaped by a regime change in transportation. Much as automobility enabled modern suburbia, trucking defined industrial sprawl.

A Short History of American Industrial Geography

As the proliferation of the steam engine (and cheap coal) freed manufacturing from the constraints of water power during the mid-18th century, industrial enterprises began to centralize. Whereas businesses requiring any significant mechanical power were once forced to locate along streams or rivers with adequate size and elevation change to rotate a water wheel with some given energetic output, the portability of fossil fuels allowed enterprises to cluster around larger markets — permitting them to deliver goods to those markets cheaply, while placing themselves within walking (or streetcar, or subway) distance of a massively expanded pool of labor.

Reinforcing this industrial centralization were changes in transportation geography associated with the coming of railroads. On the most basic level, the US’s rail network was, roughly, a radial construction, designed to link various sorts of hinterlands with urban markets. Urban industry thus enjoyed radically improved market access, reinforcing urbanization, and, by all accounts, railroads’ radiality. Rail service quality was also extremely context-dependent. Because trains in that era almost always carried traffic for multiple customers, proximity to aggregation points (rail classification yards) translated into reduced shipment travel times and increased service reliability and reduced costs; your car of stuffs would have to brave fewer intermediate handlings before being placed on a long distance train if your factory was clustered with others near the origin of a long distance train.

Finally, even if you did choose to place your factory in ‘the sticks,’ the nature of railroads simply never allowed real industrial sprawl. The American rail network peaked at an impressive, and likely excessive, 250,000 route miles around 1915 — but even in this time of abundance, constraints remained. A boxcar cannot sprout tires and drive; to access rail service, one needed to be served by a spur off a rail line, and to achieve that access economically, all but the largest manufacturing concerns (which could justify their own branches) essentially needed to be within sight of — if not actually next to — a rail line. The resulting patterns of linear industrial location are visible even without leaving central cities: the blocks of industry abutting the Northeast Corridor in Philadelphia and the Bay Ridge Branch in Brooklyn are illustrative of this dynamic.

What evolved, then, were urbanized industrial metropoles. Though certainly not without a degree of polycentricity — Paterson, Newark, Elizabeth and Jersey City all once were dense industrial hubs of their own right, complementing New York — production was relatively concentrated. We see this in the data presented in the first chart, showing relative center-city dominance of metropolitan in 1947, or in the incredible statistic that, in 1910, about sixty-seven percent of manufacturing employment in New York City was located in Manhattan below Fourteenth Street. We see its transportational dimension in a survey of New York-area manufacturers which found that fully seventy-one percent of structures in Northern New Jersey built before 1920 had rail spurs. And we see its manifestation in the multistory and densely packed physicality of older industrial buildings — a reflection of contemporary technical constraints, to be sure, but also of the simple economics of urban land.

Industrial Sprawl

This order was not to last. Much as the postwar decades saw a fundamental reorganization of urban America’s human geography (to the suburbs and the Sunbelt), they, too, saw a transformation of industrial geography. The transformation which took place did so on two axes: one between metropolitan regions, and one inside them. On the former, the postwar decades saw a marked shift of industry to the South and West, as well as to non-metropolitan areas of the nation. On the latter, the same decades saw a shift of manufacturing out of cities into their suburbs. Though both ‘halves’ of this story are critically important to understanding the fates of industrial cities in the postwar years, it is the latter on which we will focus. For one, it seems to have been the dominant effect: even in metropolitan areas whose manufacturing employment grew, center city employment declined in both percentage and absolute terms (see below). For another, in contemporary discussions about those left behind by capital flight, city dwellers are remarkably absent. Understanding the discontents of the white working class is indubitably an important subject, but to tell deindustrialization’s story without regard to urban cores is to, bluntly, tell it incompletely: after all, from peak to present, New York City has lost more manufacturing jobs than the entire state of Ohio.

Sprawl Before Trucks: Space, Labor, Policy

Industrial sprawl began well before industrial employment in cities peaked. among these was the simple scarcity of urban space. As cities built up and out, expanding industries rapidly found themselves without open or lightly used land on which to build their factories. Though redevelopment of built-up land could be — and was — pursued, the difficulties of assembling lots large enough for modern industrial facilities at reasonable costs and in a reasonable time frame were sufficiently great to discourage industrial redevelopment and expansion in the city. This search for industrial space was made all the more urgent by contemporary technological shifts. As machinery came to be powered electrically, rather than by intricately linked belts and shafts, industrial engineers were afforded massively increased flexibility in the organization of a production line. No longer did processes have to be spatially clustered close to a source of mechanical power; long, continuous-flow production lines (some thousands of feet long) became possible. The ramifications of this should be obvious: in a built-up city, it is rare for there to be hundreds — let alone thousands — of feet of unbroken space where one could erect such a line. Even manufacturers which had an established building in a city, then, had an incentive to look outwards.

The search for space was made all the more difficult by the politics of land use in an inhabited city. As the 20th Century wore on and the dangers of living next to especially more polluting industries became more known, those around urban industry became increasingly loath to allow its expansion. Through zoning codes and other means of regulation, it became the project of many to rein in urban industrial growth, to keep factories away from residences. This was a worthy quest in many cases, but it nevertheless contributed the dynamic of decentralization; hostile cities made for constrained plants, and an outwards orientation for growth.

Labor politics are also central to understanding sprawl. As union strength grew in the first four decades of the 20th Century, it increasingly became the strategy of industrial America to pursue decentralization in order to move production into more ‘business-friendly’ climates (read: those with weaker unions and lower prevailing wages), to atomize the workforce, and to provide duplication of facilities. Perhaps the most famous examples of this behavior come from the auto industry. Starting in the 1920s, the ‘big three’ and their suppliers began to spin small, usually skilled parts manufacturing operations off from their large urban plants into Detroit’s hinterland. In the 1930s, larger operations followed; after a 1935 strike at a parts plant in Toledo which shut every single Chevrolet assembly in North America, General Motors rapidly began duplicating ‘bottleneck’ facilities, taking care to locate new plants in suburban or rural centers with weak union movements.

These ‘runaway shops’ were aided in their flight by contemporary policies. Postwar defense largesse was overwhelmingly one directed outwards from urban centers: for example, per Sugrue’s The Origins of The Urban Crisis (which is fantastic, and which, along with Prof. Sugrue’s other writings, heavily informed this post), over ninety percent of federal investment in the Detroit region was allocated to areas outside the core. Though some of this disparity was likely a manifestation of the other, preexisting anti-urban biases in industrial location, there equally were more active federal interventions. As concerns about a Soviet attack grew through the late 1940s and 1950s, the DoD began to actively encourage the duplication and decentralization of production to reduce the potential productive risk associated with any single nuclear strike. This policy, known as the “parallel plant” policy, led to the scattering of production across metropolitan regions, and indeed across the country.

Less directly, but perhaps more importantly, the government subsidized industrial sprawl through its housing policies. The subsidization of white suburban homeownership and automobility, and the subsequent decentralization of a large portion of the American workforce, drew employers outwards. This was done not just to increase their proximity to suburbanites, but also to provide space for those suburbanites’ cars: as commutes shifted from transit to cars, parking needs only increased the pressure to find more space.

There is finally the question of taxation. Legacy urban centers tended (at that time) to have higher local tax rates than the suburban municipalities of their hinterlands; in some metro areas which bridged state lines, this difference was additionally accentuated by state-level tax disparities. The dominant industrial-geographic impact of tax differences was likely some minor hastening in the interstate movement of industry (especially to the South and West), but it did have some impact on intra-metropolitan location decisions as well, aiding this process of suburbanization.

It is incontrovertible that the above set of economic and political forces would have reshaped American industrial geography even within a rail-centric model of freight transportation. Indeed, GM’s policy of decentralization and duplication began before trucks became such a powerful force in transportation, creating a peculiarly rail-centric variety of sprawl around Detroit. However, it seems unlikely that industrial sprawl would have taken place to the degree that it did, and in the form that it did, had it not been for the rise of American roads. Though these forces added to the logic of sprawl, their ability to shape geography was fundamentally limited by structurally centripetal rail networks; highways (and especially the trucks they carried) removed that final barrier.

Much like a car, a truck is a largely ‘atomic’ mode of transportation. In contrast with trains, whose operating paradigm is either primarily (through the existence of sufficient traffic on a corridor to fill an entire train) or secondarily (through the assembly of individual freight cars with similar origins and destinations into a train) reliant on traffic density, trucks need not rely on aggregation; they generally move a single load on a point to point route. This allows them to serve light-density markets and decentralized industrial geographies without much of an efficiency loss.

The other important parallel between cars and trucks is their incompatibility with dense urban environments. Long, unwieldy, and without efficient and generally dedicated infrastructure like trains, trucks have long had a difficult time navigating narrow and crowded urban streets — something becoming only more true as truck lengths have increased. Though the desire of trucking and industrial interests to improve access to urban manufacturing hubs became a key force behind several early urban highway projects (for example, New York’s Holland Tunnel and Gowanus Expressway), this was a necessarily Sisyphean task. Beyond the operational issues that come with large trucks on small streets, trucks simply take up more space than their steel-wheeled counterparts. Truck-centric industrial design (with its loading docks, parking, and maneuvering space) made the dense, street-fronting industrial buildings necessary for urban location all but impossible; the factories of today overwhelmingly lie in fields of parking in areas where such expansive plots may be found — in the hinterland.

The final piece of the puzzle here is the systematic subsidization of truck transportation in the United States. We have already touched on some such mechanisms, whose subsidies to trucks existed insofar as they subsidized truck-friendly decentralization, but there, too, existed a vast amount of direct subsidy to the mode. This is most obviously evident in the creation of a federally funded network of highways. Though user fees initially recouped a large fraction of highway operating and maintenance costs, recent evidence on the allocation of costs between road users would suggest that truckers underpay significantly relative to their impact on road maintenance. Before partial transportation deregulation in the 1970s and early 1980s, this subsidy asymmetry was also compounded by government rate policies. And these elements of subsidy are to say nothing of the divergence between highway costs and revenues seen since the 1960s, or the manifold other unpriced (and perhaps unpriceable) impacts trucks and highway-centricity have had on us.

From this convergence of industrial and transportation policy evolved a new industrial landscape, centered around trucks, and in the suburbs. Boston’s Route 128 beltway is illustrative. The years following its completion saw an explosion in investment. Through the mid-1950s, about forty percent of all spending on new industrial facilities in the Boston region was directed to facilities along this corridor. Defined in more than their location by this highway, these new plants had about eight-tenths of a parking space per employee, overwhelmingly generated car commutes, negligibly prioritized access to freight rail, and were overwhelmingly larger and flatter than their predecessors. Though some of these businesses were new to the region, or were ones which grew in moving, a large proportion of corridor activity was nevertheless attributable to relocations, and fully seventy-five percent of those relocated businesses had previously resided in Boston’s urban core. Route 128 assuredly built on a preexisting set of industrial instincts towards flight, but its construction is nevertheless what unlocked and shaped them. With the access provided by its belt, Boston’s manufacturers need not renovate their (rail-centric) urban facilities to accommodate growth; they could discard them for a new, more accessible, seemingly more modern truck-oriented facility on the burgeoning urban periphery.

The relationship between freight transportation and these outwards movements is emphasized even further by evidence from the New York region. Before 1920, about seventy percent of industrial properties in northern New Jersey were built with rail spurs. These plants were those of an earlier generation of decentralization; they were the mills of Paterson, the refineries of Elizabeth, the sprawling factories of Kearny. By the 1950s, New Jersey’s industrial growth had turned outwards, to the industrial parks along the New Jersey Turnpike and other arterials. Towns like South Brunswick and Edison became home to an endless sprawl of single story industrial parks, providing expansive suburban locations for industry, and thus contributing to the precipitous decline in New York’s industrial employment over that time. Unsurprisingly, this highway-led industrial development had deep ramifications for manufacturers’ choice of shipping mode: by the mid 1950s, the share of new industrial construction in Northern New Jersey which was built with rail spurs had dropped below forty percent. Forty percent is by no means an insignificant figure, but it is nevertheless a far cry from the seventy percent of mere decades before — and is indicative of how here, too, highways and trucks increasingly shaped sprawl.

Patterns like these are visible everywhere in America. It would seem no accident that, on the above charts of industrial employment patterns, the steepest declines in the urban share of regional industrial employment followed the buildout of highway systems through the 1940s, 50s and 60s. As Chicago (to take another example) built out its highways and beltways, industrial growth in the periphery exploded; today, the highest concentrations of industrial employment in the region lie in industrial parks in the urban periphery, near highway interchanges and along arterials. Similar stories can be told in Atlanta, San Francisco, and beyond: the modern industrial landscape is indelibly tied to the geography of our roads — roads which now carry the plurality of American freight.


This new, truck-oriented industrial geographic regime has, of course, had ramifications which go far beyond simple matters of placement in space. In the world of transportation, industrial decentralization hastened rail traffic declines; reliant as they are on traffic density and aggregation, the dispersed goods movement patterns and lower traffic densities that came with sprawling industrial geographies reduced traffic beyond the basic shift wrought by the appearance of the truck and its subsidies. Railroads have attempted with some success to adopt to this new regime with the introduction of intermodal services, which combine the last-mile flexibility of trucks with the long-haul efficiencies of rail. However (and I will delve into this more in the future) this marriage has been somewhat of a Faustian bargain; sprawl has made rail structurally less competitive.

More important than these matters of network bias have been sprawl’s socioeconomic ramifications. American industrial history is riven with racism and racial disparity; the story of sprawl is no different. As Sugrue has discussed at length, the outward migration of high-paying industrial jobs after World War II worked in conjunction with rampant, institutionalized racism in the housing market, discrimination within the labor market, and anemic investment in high-quality public transit to interdict Black access to new factories at the metropolitan edge. Without mobility, seniority, or access to more skilled job classifications, capital flight (and the concomitant rise of factory automation) translated into economic disenfranchisement for Black communities in many industrial cities. In 1960, the unemployment rate for white autoworkers in Detroit was 5.8%. For their Black counterparts? 19.7%. Deindustrialization was obviously only one of many interlinked factors working against urban prosperity in the postwar years. However, insofar as (unionized) production jobs were the route into the American middle class for many, this set of institutional discriminations implicated in industrial migration away from cities is difficult to read as anything but a great denial of opportunity — one whose ramifications for urban economies in post-industrial metropoles are visible to this day.

The impacts of sprawl-induced urban deindustrialization has left long shadows. So, too, has the infrastructure that supported it. Americas urban highways were violently constructed, and ever since, have been perpetrating another, slower violence of pollution against the communities alongside them. Though the coalitions supporting urban highway plans generally problematized automotive access more so than they did trucking, the latter use did, too, play a role in supporting these constructions. Once built, the trucks using these roads have disproportionately contributed to the pollutants which emanate from them. The average heavy-duty diesel truck emits about nineteen times more PM2.5s per vehicle-mile than the average gas powered car — PM2.5 being a pollutant which has been linked to, among other things, increased COVID-19 morbidity. Railroads are by no means corporate exemplars of good environmental stewardship, but even with their current fleet of diesel locomotives, they offer radically reduced emissions relative to trucks. This is to say that, on top of all else, sprawl has radically increased the environmental footprint of freight in this country, especially in vulnerable communities. I think it’s critically important to note, here, that keeping certain types of industry in densely populated areas is also bad environmental policy, given the very real pollution risks associated therewith, but to banish all to the suburban fringe does not seem altogether wise, either. A nuanced, justice-oriented approach to these questions is what was and is required, and is anything but what has been provided.


America clearly needs improved surface transportation policy. This is perhaps most apparent to the average person in our country’s passenger transportation systems, but freight needs attention too. As we enter the 2020s and fitfully seek to to confront issues of climate, economy, and justice, assembling the sorts of policies that may begin to right these failures of planning and incentives seems critical. Next time, we shall discuss what a ‘new freight policy’ could and should be.

Inefficient Success

Before COVID among the — if not the — hottest topic in New York area transit infrastructure was the Gateway project. With dire (if questionable) predictions of massive capacity cuts from an impending Hudson River Tunnel closure and a constant drumbeat about the centrality of the tunnels to New York-New Jersey commutation, you would be forgiven for thinking that the tunnels leading into Penn Station comprise the busiest transit crossing of the Hudson River. I say “forgiven” because, of course, they are not.

Carrying over 35,000 bus riders under the Hudson River and (mostly) into the Port Authority Bus Terminal during the AM peak hour, the bores of the Lincoln Tunnel hold the crown of busiest Hudson River transit crossing, trouncing the NJT Rail tunnels to their south by a healthy margin of 11,000 peak-hour riders. Indeed, the Lincoln Tunnel is among the busiest of any Manhattan transit entry points; only the Manhattan Bridge’s four subway tracks carry more people. But their story isn’t one of success. The buses that travel through the Lincoln are a testament to our transit planning failures, and threaten to cost our region untold billions in the years to come.


In New Jersey, buses using the Lincoln Tunnel arrive from across the northeastern quadrant of the state, feeding from a dense transit grid in the inner suburbs and from tendrils reaching far into the the state’s interior. These routes are enormously significant to NJ Transit’s network. Trips serving the Port Authority Bus Terminal (PABT herein) consume thirty four percent of all weekday bus service-hours operated by NJT, and in many North Jersey counties, are the majority of bus trips available to residents.

If you slice and dice the data a bit differently, we can begin to gain an even clearer picture of these buses’ catchment. About half of all PABT bus trips stop in Hudson and/or Bergen Counties, with Passaic and Middlesex holding distant third and fourth places. This pattern is repeated if you look at route level data — a map of the top ten routes (by number of weekday trips) serving the PABT shows a dense cluster of routes in Hudson and Bergen, and the 139 bus reaching down into Monmouth.

Unsurprisingly, these service allocations track with the geography of bus use in North Jersey. While transit use in the region is strong overall, it is locally fragmented by mode. Rail is strongest in wealthy suburbs along electrified (read: faster) commuter rail corridors and along PATH. Bus use is most common in northern Hudson County and in Bergen (PABT-land) where rail options are either limited or slow, giving buses a time advantage.

I should note: not all of these bus use clusters are driven by the relative speeds of radial transit. Buses’ large share in the Oranges, for example, is likely driven by high rail ticket prices and low frequencies making it less accessible and useful to lower income transit riders. However, the split between places with electrified rail and those without holds in general.

Finally, it’s worth remarking on the density of the areas served by these buses. Manhattan’s economic gravity means that it’s possible to generate relatively strong transit use in areas with low intensity land use patterns, but, with the exception of some of park-and-ride routes, most Lincoln Tunnel bus ridership seems to be generated by riders from dense inner ring suburbs and secondary cities (I say “seems to be” because I am not aware of any public route-level NJT bus ridership statistics, so must infer rider origins from relative service levels). Indeed, the first, second and third densest municipalities in the United States all lie within the PABT catchment zone — those three being Gutenberg, Union City and West New York.


The PABT and associated NJT bus routes are an unquestionably successful pieces of transit infrastructure. By providing fast service from North Jersey to Manhattan, they likely divert thousands of car trips, and help support Manhattan’s (transit-friendly) regional economic preeminence. But I am here to tell you that while successful and beneficial, the network is, in fact, a massively underrated transit problem.

I see three motivating dynamics here: operating efficiency, service efficiency and capital efficiency. Let’s dissect them.

Operating Efficiency

Transit modes can be roughly stratified into capacity buckets. If you need a little bit of capacity, you use a van. If you need more, an infrequent (and perhaps small) bus. More, and you enter the province of frequent and/or articulated buses. Even more, and you’d probably want to look at light rail or light metro. Beyond that, you want subways, or regional rail. The way these buckets work is simple: in absolute terms, it costs more to operate a subway train than a bus, which costs more to operate than a van. However, it costs less to operate two buses than five vans (or five trains instead of thirty buses, and so on), so as your ridership increases, larger vehicles with more involved infrastructure become more economical by providing more capacity at a lower unit cost. These basic principles are generally borne out in US per passenger-mile transit cost data, despite the US’s issues with underutilized and poorly planned rail.

Lincoln Tunnel buses fall squarely in the capacity range best served by rail. Average loads per bus lie in the 40-50 passenger range that normally defines the upper limit of bus capacity, and their aggregate daily demand profile resembles that of a four track subway line. Yet the way we deal with these riders is by squeezing nearly one thousand buses per hour through the Lincoln Tunnel during peaks. Even at American (read: high) operating costs, commuter rail can deliver these passengers to Manhattan for less money, as NJT bus service costs $0.78 per passenger-mile to run, but rail costs only $0.47. Those savings could, in turn, be reinvested in service expansions across the NJT network.

Service Efficiency

That final point, about reinvestment, is key here. Setting aside the above-identified potential monetary redistribution, some future rail-based Lincoln Tunnel bus network replacement would free up a large number of bus service hours for use elsewhere — recall that thirty four percent of NJT bus service is spent on PABT routes. Instead of running routes oriented towards radial travel, we could give Northern New Jersey a strong BRT network, or a dense, frequent, Toronto-esque grid of local routes for little additional money. These (re)investments could (coupled with land use changes) begin chipping away at auto-dependence in intra-suburban travel, advancing environmental and equity goals. Getting Northern New Jersey even to Canadian levels of transit use in suburb-suburb travel would be a big win.

Spending service hours on a strong local grid is also a necessary complementary investment to rail. Part of what makes it difficult to work our way out of this bus problem is that Lincoln Tunnel buses cover Hudson and Bergen Counties quite comprehensively; you can get a PABT bus almost anywhere, at least during peaks (see earlier maps). Bergen County is not in any way lacking in rail corridors, but these lines are not nearly dense enough to put most places in the county within walking distance of a train; you need (frequent) local buses meeting (frequent) trains at timed and fare-neutral transfers wherever possible to extend the reach of the rails.

Capital Efficiency

By far the most important facet of the Lincoln Tunnel bus issue is its impact on regional transit capital planning, for the PABT is coming due for replacement. The Port Authority is contemplating a number of project alternatives, the best of which do little more than incrementally improve rider experience and transit utility, and the worst of which move parts or all of the (already peripheral) bus terminal further away from Midtown job density and transit connectivity to the Javits Center area. These replacements also come at a massive cost; the Port Authority is projecting PABT replacement to cost somewhere between $7.5 and $10 billion.

The project’s price tag is more a testament to New York’s cost disease than any inherent issues with buses, but, as denizens of ‘Transit Twitter’ have been doing for years, we should ask why we are planning to spend this much on a project that is fundamentally about preserving an inefficient transportation system. Precious few global cities with high transit ridership and job densities are currently building large bus terminals to serve them; for the efficiency reasons illustrated above, buses are simply not a best practice here. Indeed, some cities (for example, Ottawa) are even converting former bus infrastructure to rail on high-ridership corridors in order to realize more capacity for less money. While it’s likely unrealistic to expect a full replacement of the PABT, lower bus volumes would be more conducive to a much less costly transitway based replacement. So to the extent we can, we should be following the lead of other cities, spending on better, higher functioning transit infrastructure rather than preservation.

Ways Forward

What planners must task themselves with, then, is finding ways to better serve the PABT/Lincoln Tunnel catchment area with rail transit. Any such conversation should start with low-hanging fare and operating reforms. It should not, for example, be more expensive to take a train to New York from Paterson than a bus, nor should our regional fare structure penalize people for transfers transfer between, say, NJT and PATH. Especially in light of the pandemic’s impact on rail ridership, now may also be the time to begin working towards (long overdue) commuter rail reform; to whatever extent possible given the limitations of existing infrastructure and equipment, we should seek to rectify NJT’s low service levels, complicated service patterns and high operating costs.

Fundamentally, however, reforming travel patterns will require us to redirect the billions planned for PABT replacement towards investments in rail speed, capacity and coverage in Northern New Jersey. For existing rail corridors (which mostly serve Bergen County) the infrastructure prescriptions to these ends are simple, and have been discussed at length and in more detail by others: New Jersey Transit should electrify its existing lines through northwestern New Jersey to support higher speed, lower cost operations with less environmental impact, and convert stations along them to high platforms to reduce dwell times and staffing requirements. In tandem with those improvements, NJT should begin the construction of the dense local/feeder bus grid so critical to rail’s success, and should likely contemplate expanding rail service onto currently freight-only corridors, for example CSX’s formerly quad-track River Line. Further, we need to invest in more, better designed core network capacity, whether that be some sort of Hoboken-Atlantic Terminal tunnel, a Penn-Grand Central connection in conjunction with Gateway, or an investment in higher performance signaling and equipment so we can run 30+ tph on the tracks under the Hudson as is done in Paris.

For Hudson County, things are more complicated. Unlike Bergen, its bus-dense towns do not have rail rights of way on which one may incrementally improve service. A fix for the area will likely require some combination of subway extensions from Manhattan, improvements to and realignments of the HBLR network (which is currently somewhat underutilized relative to population density north of Jersey City, and can provide a two-seat ride to Manhattan with PATH), and preservation of Manhattan-bound bus service. Yet perhaps more than anywhere else in the New York metro area, transit investments in Hudson County can self-justify with growth: Hudson County has the among the most pro-housing policies of any part of the region.

New York’s truly exceptional construction costs force us to make decisions that we simply should not have to make, and to suffer the climatic and economic disbenefits of glacial transit expansion. While agencies should always be cognizant of opportunities to spend money in ways that will increase operating efficiency, New York’s cost bloat (and operating cost efficiency issues) means this focus should only be stronger; the ethos of doing more with less should dominate transit discussions in this region. Rail replacement of the PABT is perhaps the greatest such opportunity in New York today, one which could positively transform transportation for millions in New Jersey. Let us not pass it up for another half-century of inefficient transit.

Subway Operating Efficiency and the MTA Budget Crisis

As most of you are likely aware, the MTA is currently facing a large deficit. Thanks to the ridership and cost impacts of the COVID pandemic, the agency needs $12 billion to cover operating losses through the end of 2021. It is hoped that the federal government will cover the shortfall, lest New York face a massive round of service cuts.

Given the magnitude of this budget shortfall, it is likely unrealistic to expect the MTA to recover fully without federal aid. However, whether to mitigate the impacts of a no-funding situation or to cover a shortfall in provided funds (for example, if Congress grants, say, $10.5 rather than $12 billion), it is important for the MTA to have identified ways of extracting efficiencies from its operations that do not involve the broad-based service cuts, layoffs and wage freezes threatened.

The path to ‘better’ savings isn’t especially murky. Patrick O’Hara laid out ways to save on the operation of LIRR service in this post of a few weeks ago, and I have tweeted a good bit on NYCT’s operating cost issues in the past. That said, given current events, it’s worth discussing the issue in depth. To be clear: I do not mean for this post to be an all-encompassing list of potential ways to save; my aim here is to present a few frameworks/cost centers of interest when discussing a financial path forwards.

The 80/20: Facility Maintenance

As should immediately stand out on the chart above, NYCT’s operating cost issues are driven by maintenance, specifically facility maintenance (think: track, tunnels, signals, yards, stations). If you look at cost-efficiency numbers on a per track-mile basis, the picture becomes even more dismal: NYCT is spending about three times as much per unit of maintenance as domestic peers. (NOTE: to get a comparison to true best practices, NYCT should benchmark to the likes of Paris or London, but I have not been able to obtain granular cost data for those systems)

There are some caveats here, of which it is important to be aware:

  • Facility maintenance expenditures are somewhat correlated with use intensity (as measured by car-miles per track-mile), even after you exclude the outliers of NYCT and PATH (these analyses include light rail system data to increase sample size beyond 14). However, even when including PATH and NYCT in the regression, NYCT’s maintenance costs are about $800,000 more than predicted.
  • Most US agencies classify some maintenance-like items as capital expenditures; for example, NYCT put spending on switch replacements in the capital budget. While it is possible to access regularized capital spending data through the National Transit Database, the NTD data do not offer sufficient detail to distinguish, say, an ADA upgrade from a laundry list of deferred maintenance items being treated as a capital project. The upshot: operating budget expenditures may not reflect the full extent of maintenance spending, but including capital spending may end up overstating costs.

With that said, let’s break down the issue. Maintenance cost structures across all US systems are dominated by labor. NYCT is no different in this regard; its cost disease seems to be driven almost entirely by a serious maintainer productivity deficit. NYCT’s labor cost per facility maintainer-hour (including salary and fringe benefits) is actually about average for US systems at $65; its issue is that it uses about four times more maintainer hours per track mile than peers.

The obvious question here is “why.” I am, honestly, not entirely sure. I have long thought that some combination of the agency’s complex track access and roadway worker protection (“flagging”) protocols and work rules may be what is driving up costs and labor expenditures, but I cannot present evidence beyond anecdote and my own observations. Nevertheless, this area holds massive potential for further investigation: bringing NYCT’s facility maintenance costs down to the national average could save $1.3 billion dollars per year, or about twenty percent of the projected 2021 deficit.

Other Ways to Save: OPTO and Hidden Time

One of the more commonly proposed ideas for cost reduction at NYCT is movement to one-person train operation, or OPTO. OPTO is, unequivocally, an international best practice in rapid transit, even on systems (ex: Thameslink) with long, crowded trains and curvy platforms. While NYCT’s overall vehicle operations spending isn’t extreme compared with those of other US systems — most of which use OPTO — OPTO would indeed be a significant saving, and would more importantly reduce the incremental cost of subway service, thus making cuts less and expansions more attractive. However, outside the G and L lines, any OPTO expansion would require significant investments in CCTV infrastructure and crew training. This would be money well spent, but the need for investment introduces a relatively inflexible minimum time-to-savings. And, of course, this is all assuming that the TWU would even concede OPTO, which seems unlikely (and honestly, who can blame them?)

Making a unit of service cheaper is only one of the two broad levers transit managers have at their disposal right now. They can also cut service. Even with COVID ridership levels, service frequency/span/coverage cuts should, as a general rule, be avoided (for reasons ranging from equity to the difficulty of re-hiring and re-training operations staff, to the downwards pressure this would have on ridership’s recovery), but there are ways to trim service with small impacts relative to their returns. These savings are all about ‘hidden time,’ which in my mind comes in two varieties: excess scheduled runtime and non-revenue train movement.

Most costs associated with train operation scale with the time it takes to run the length of a line. The longer the round trip runtime, the more trains and crews you will need to run a service at a given frequency. The corollary here is that when you shorten scheduled running times, you can make a service (hours) cut without actually impacting frequencies or coverage: crews and trains are doing the same work in less time. One cannot wake up one morning and magically speed up trains, but NYCT’s incremental speed efforts have borne fruit and COVID ridership losses are making trains run faster; if the agency expects these gains to hold, it may be worth adjusting schedules to reflect them. There are also likely savings to be had in schedules for diversions, though those savings would be more likely to accrue to the capital budget, which pays their cost.

The issue of non-revenue time is more straightforwards. In the course of operating a transit system, you inevitably end up with vehicles running without passengers — for example, on trips to and from train yards. For its size and complexity, NYCT’s non-revenue proportion benchmarks well compared to other US heavy rail systems, but savings likely exist. We could ask, for example, whether every AM rush D train need originate from Stillwell Avenue; on the N, some enter service at 86 St which minimizes time from yard departure to entry into northbound service (this may require slightly rearranging yard patterns in the Coney Island complex).

Hidden time cuts likely will not add to much in the grand scheme of the transit budget. It is not even possible to say for certain whether, after accounting for the resources that would have to be invested in rewriting schedules, they would net much financial benefit in the short run. However, cuts like these have return beyond finances: a faster, more tightly scheduled system is better for riders and operations, and I moreover think we would be remiss not to look at low impact savings before reducing mobility in a crisis.

Difficult Decisions

What New York and its transit system face today is unprecedented and unpredictable there is no certain way out of this mess. I want to be clear in saying that the immediacy of this potential crisis may make the systemic work I suggest too time-intensive to be helpful. But we need to be cognizant of our actions’ long term ramifications for the agency and for New York. Indiscriminate cuts and wage freezes will likely erode the MTA’s knowledge bases, internal networks and organizational capacity, whereas targeting specific cost centers reduces institutional risk while enhancing process reform capabilities. And every transit service cut made threatens climate, equity and the mere survival of the transit-dependent urban landscape that is New York. So, if funding does not come through, it is imperative that we try pursuing more targeted approaches to the budget.

Home Signal

Ever since I (Uday Schultz, or @a320lga) stumbled into the reeds of ‘Transit Twitter,’ I have been planning a blog. Amazed by the content on Alon Levy’s Pedestrian Observations, Sandy Johnston’s Itinerant Urbanist, Michael Noda’s Sic Transit Philadelphia and Clem Tillier’s Caltrain HSR Compatibility Blog (to name just a few) and so frequently frustrated by Twitter’s character limits, it seemed but a natural progression. So, finally, after 18 months of procrastination, I am…getting a blog!

Much like my Twitter account, the focus here will be on transportation, urbanism and industrial history. More so than on my Twitter, I plan to make an active effort to write on a mix of detail-heavy operations/operations planning subjects, and more high level planning, history and background subjects. Given the geography of my knowledge, posts will likely tie back to issues in the Northeast or Industrial Midwest, but I hope to bring in material from the world beyond.

I should note that this blog will be as much a space for experimentation as for exposition: this is a new format for me, and I generally count many of my ideas as works in progress. I will hopefully have a ‘real’ post ready sometime later this week — see you all then!