Journal of Public Transportation 26 (2024) 100083

Contents lists available at ScienceDirect

Journal of Public Transportation

journal homepage: www.journals.elsevier.com/journal-of-public-transportation

Bus stop spacing statistics: Theory and evidence

Saipraneeth Devunuri a, Lewis J. Lehe a, *, Shirin Qiam a, Ayush Pandey a, Dana Monzer b
a
Department of Civil and Environmental Engineering, University of Illinois Urbana-Champaign, United States
b
Northwestern University Transportation Center, United States

A R T I C L E I N F O A B S T R A C T

Keywords: Discussions of bus stop consolidation sometimes refer to average stop spacings, but there are no reliable statistics
Public Transit about spacings, nor methodologies for calculating them. This paper aims to clarify discussions of bus stop
Stop Spacings spacings by introducing clear definitions, a methodology for creating statistics from General Transit Feed
GTFS
Specification (GTFS) files, and a python package, gtfs-segments, which splits bus networks into isolated ‘seg­
Bus stop
Transit Planning
ments.’ With the package, we calculate national-level statistics from 539 US transit providers and 83 Canadian
providers, as well as agency-level statistics for 30 providers in the US, 10 in Canada, and a sample of 38 providers
from other countries. Our estimates of US and Canadian mean spacings are both around 350 m (slightly wider
than five stops per mile). US spacings are wider than sometimes claimed but narrower than those in other
countries. Finally, the paper gives examples of metrics created by combining GTFS with data from other sources
and proposes research ideas and applications to transit planning involving fine-grained stop spacing data.

1. Introduction 2016). The stakes in this enterprise have risen over the last decade:
buses have lost a non-trivial number of riders to ridehailing (Graehler
“Bus stop spacing” refers to the distance between consecutive stops et al., 2019; Grahn et al., 2021; Erhardt et al., 2022), which is more
on a bus route. The choice of bus stop spacing involves a trade-off: wider expensive but faster than the bus. In addition, stop consolidation has
spacings (fewer stops) save passengers in-vehicle time (time spent sometimes been pursued as a means to increase reliability: deVries
traveling on the bus) but raise passengers access time (time spent trav­ Kehoe (2004) found stop consolidation improved on-time performance
eling to/from stops). This trade-off is at the heart of a robust academic for King County bus routes. At the same time, there is concern that
literature1 about the choice of spacings that has been active for about spacings that are too wide may be especially inconvenient for elderly
fifty years (e.g., Vuchic and Newell, 1968; Mohring, 1972; Wirasinghe passengers or others with difficulty walking (Wu et al., 2022).
and Ghoneim, 1981; Li and Bertini, 2009; Ouyang et al., 2014; Wu et al., In spite of the vigorous theoretical and practical interest in stop
2022). spacing, reliable data about stop spacings is scarce. What is the average
In the United States, there is a general perception that current American stop spacing? Some studies2 reference Reilly (1997) (p. 4): “It
spacings are too narrow. Thus, over the past two decades, many US bus is common European practice to have stops spaced at 3 or 4 per mile [ ~
providers have widened their spacings via stop consolidation: the sys­ 400–540 m/stop] in contrast with 7–10 stops per mile [ ~ 160–230
tematic practice of removing large numbers of stops. Stop consolidation m/stop], which is common in the United States. European bus-stop
campaigns have been carried out in Portland (El-Geneidy et al., 2006), distances are comparable to rail rapid transit stop distances in the
San Francisco (Gordon, 2010), Cincinnati (LaFleur, 2019), Seattle United States.” Likewise, a Transit Cooperative Research Program
(Kehoe et al., 2021), Pittsburgh (Blazina, 2020), Dallas (Garnham, (TCRP) report states: “US transit bus operators…place stops about every
2020), Denver (RTD Denver, 2023), the Bronx (Moloney, 2021) and 200 m, creating five stops per kilometer [8 stops per mile]” (Trans­
elsewhere. Researchers have supported the consolidation trend with portation Research Board, 2001, p. 4). But read in context, both quotes
various practical techniques for deciding which stops to remove (Li and are intended more as impressions than exact statistics, and neither study
Bertini, 2009; Wagner and Bertini, 2014; Stewart and El-Geneidy, references any data collection.

* Corresponding author.
E-mail address: [email protected] (L.J. Lehe).
1
See Tirachini (2014, Sec. 2) for a thorough review.
2
For example: Furth and Rahbee (2000, p. 1730), El-Geneidy et al. (2006, p. 33), Morency et al. (2011, p. 800), Wu et al. (2022, p. 2).

https://doi.org/10.1016/j.jpubtr.2024.100083
Received 17 May 2023; Received in revised form 24 January 2024; Accepted 21 February 2024
Available online 7 March 2024
1077-291X/© 2024 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-
nc/4.0/).
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Statistics about real spacings can inform decision-making and introduces an approach called “traversal-weighting” (explained below)
communication. van Nes and Bovy (2000) used an analytical model to and uses GTFS data from 43 US public bus systems to report summary
optimize spacings toward various objectives, then compared the results statistics calculated using the gtfs2gps R package (Pereira et al., 2023).
to their own calculations of average spacings for seven European cities. This paper extends the work begun in Pandey et al. (2021). This paper’s
The real spacings are much shorter than what the study’s model rec­ goal is to facilitate further research and discussion of real stop spacings.
ommends, which gives the theory a practical lesson: that cities ought to To this end, this paper has four contributions
consider wider spacings. The Queens Bus Network Redesign Draft Plan
(MTA, 2022, p. 9) uses averages to communicate a point to the public: (i) Clarify conceptual issues around stop spacings via new defini­
tions and weighting schemes for the calculation of means.
New York City has too many bus stops, resulting in shorter distances (ii) Devise a methodology for calculating stop spacings using GTFS
between stops than most other major cities. With an average of 805 files, which is implemented in a new Python package: gtfs-
feet [245 m] between stops, buses are often stopping as frequently as segments.
every one or two blocks. In Queens, the average is slightly higher at (iii) Provide new evidence about bus stop spacings—mainly in the
909 feet [277 m]. Both are shorter than the distance between stops in United States and Canada but also in select cities elsewhere. This
international peer transit systems around the world, which typically evidence is culled from an open-access database of spacings from
range from 1000 to 1680 feet [305–512 m]. 539 providers in the United States and 83 in Canada.
(iv) Demonstrate two alternative spacing statistics: the average
The point made here is that the City’s status quo is unusual among
spacing experienced by a passenger and the number of traffic sig­
world cities, so consolidation would not be a harsh experiment. (It is not
nals traversed between two stops.
said where the statistics about the peer transit systems come from or
how the averages for Queens and New York were computed.) The in­
Note that the scope of the paper is limited, essentially, to description
ternational comparison dovetails with the claims from Reilly (1997) and
and clarification. There is no optimization, nor statistical identification
Transportation Research Board (2001) that spacings in other countries
(e.g., whether wider spacings cause higher ridership). One motivation
are much wider than US spacings. Communications like this are
for the paper is to provide methods and data by which scholars or
important because changes to stops can be politically fraught (Flint
practitioners can conduct further statistical and operations research and
et al., 2014; Berez, 2015). The nonprofit TransitCenter has even pub­
in the conclusion (Section 6) we propose several ideas. Another is the
lished a report (Miatkowski and Hovenkotter, 2019) devoted entirely to
more abstract notion that, apart from optimization and statistical
public communication about consolidation.
identification, data about a topic that affects so many people’s transport
At first blush, the dearth of hard data about stop spacings is sur­
experience is inherently valuable and worthwhile to collect accu­
prising in light of the trend toward consolidation. The spread of the
rately—even when its applications are not immediately obvious.
General Transit Feed Specification (GTFS) (McHugh, 2013) has made
public transit data easier to analyze, while various APIs and websites (e.
2. Definitions
g., MobilityData (2024)) have made it possible to download GTFS data
from transit providers around the world. But even with GTFS data, it is
This section discusses conceptual questions that arise in the calcu­
challenging to compute spacings. In the first place, one cannot simply
lation of statistics and distributions of stop spacings.
compute the Haversine distance between each pair of stops, because a
particular spacing is (to the degree it speaks to the access/in-vehicle
time trade-off) supposed to be the driving distance (i.e., along the 2.1. Terminology
route) between two stops. Thus, calculating spacings involves some
complex calculations to obtain driving distances between stops. Also, To begin, we propose the following terminology for the analysis of
many agencies do not publish an optional GTFS file (shapes.txt) giving stop spacings:
the paths of routes.
Beyond technical challenges, the calculation of stop spacing statistics • Stop: A physical location where buses stop to pick up and drop off
invites conceptual (or definitional) questions. A single route’s average passengers4.
spacing can be calculated by dividing its length by its number of stops, • Segment: The edge that a bus travels between visiting two consecu­
but it is not obvious how to take an average for a network. Many stops tive stops. A segment is characterized by: (i) a “beginning” stop (the
are served by overlapping routes, or only at certain times, or on certain first stop the bus visits on the segment); (ii) an “ending” stop (the
days. Express routes skip stops. Some routes have portions served only at second stop the bus visits on the segment); and (iii) a path in space
certain times. Reasonably, frequency could be considered, too. Suppose that the bus travels between the two stops. If two routes visit the
a bus system has a “flyer” that runs thirty kilometers directly from a same two stops consecutively, traveling the same path between
downtown stop out to an airport twice per day, and also runs a circulator them, then we say two routes “traverse” the same segment. Fig. 1
within the downtown that stops every block with ten-minute headways. shows an example segment in black (with yellow stops).
Should an “average spacing” for this system weight the large distance to • Traversal: The event of a bus stopping at each of the segment’s two
the airport the same as it does one of the circulator’s spacings? Perhaps stops and traveling along the designated path between them. A
not: few resources are tied up in the airport route, and its distance is so segment on a route with a headway of ten minutes (frequency of six
large that reducing it would not, at the margin, decrease any passengers’ buses per hour) will experience twice as many traversals as one on a
access times. route with a headway of twenty minutes.
Pandey et al. (2021)3, a preliminary attempt to address these issues, • Distance metric: The function that calculates the segment’s spacing.
Below we focus on distance along the segment’s path, which takes into
account how the path meanders along the road. Section 5.3
3
The work by Pandey et al. (2021) presented only traversal weighted stop
spacings data for 43 US cities. This work presents stop spacings data for cities
4
from across the globe. The gtfs-segments package has a more accurate way of Note that in Fig. 1, we do not distinguish between stops in opposite di­
calculating the spacings, accounting for directionality and correcting rections for simplicity. However, within GTFS specification, stops in opposite
out-of-order stop sequences. This study also introduces segment, route, and directions of a route are considered unique stops, as they represent different
passenger weighted spacings along with alternate distance metrics such as locations for boarding or alighting and have a unique stop_id. Additionally,
traffic signals stops with the same name but different stop_id’s are considered distinct stops

2
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Fig. 1. A segment on the K-Ingleside bus route from the San Francisco Municipal Transportation Agency (SFMTA).

demonstrates another distance metric: the number of traffic lights • Segment-weighted: Each segment receives the same weight: i.e., wi = 1
traversed along the segment’s path. ∀ i.
• Measurement interval: The part of the schedule during which stop • Route-weighted: Each segment is weighted by the number of routes
spacing statistics are calculated. Bus schedules vary across days (e.g., that include it. The mean route-weighted stop spacing could be
Sundays typically have less service than weekdays) and within days calculated by summing up the lengths of all routes (ignoring dead­
(e.g., some routes and frequencies may only be offered at peaks), so head portions) and then dividing by the total number of segments.
any measurement is specific to a chosen period. This is the mean distance between stops that someone would expe­
• Deadhead: Ground covered by a bus when the bus does not carry rience if they rode every route in the schedule once.
passengers—when the bus is “not in service.” A typical case would be • Traversal-weighted: Each segment is weighted by the number of times
travel to or from a bus depot where there are no boardings or a bus traverses it—i.e., by its total number of traversals in the
alightings. We exclude deadhead mileage from our spacing statistics schedule. Stops on high-frequency routes are thus weighted more
because they do not reflect the passenger experience. heavily. The mean traversal-weighted stop spacing can also be
• Threshold: In calculating mean spacings, it is sometimes worthwhile calculated by summing the cumulative distances all buses travel
to ignore spacings above a certain level which we call the (ignoring deadhead portions) and dividing by the total number of
“threshold.” The discussion in section 2.3 below offers a justification. segments. The mean traversal-weighted stop spacing is also the
average distance that a bus travels between consecutive stops.
2.2. Weighting schemes
Figure 2 shows a simple network to illustrate the weighting schemes.
We now use our terminology to describe weighting schemes of use in There are two routes: Green and Blue. There are five stops, named by
making distributions and summary statistics (e.g., mean, standard de­ numbers, and one depot named D. The “measurement interval” chosen is
viation, etc.) for a bus system’s stop spacings. Suppose that a system has some hypothetical day (e.g., a weekday), during which Green runs 60
N segments. Let si be segment i’s “spacing” (most typically, the distance buses and Blue 120 buses per day. There are eight segments—each
along the bus’ path between the two stops) and wi be segment i’s depicted as an arrow on the map, with colors representing what route
“weight” (according to a chosen weighting scheme). In this case, the serves the segment. Both routes traverse the segments from 3 to 4 and
system’s mean spacing is from 4 to 1, so these segments are dual-colored. Each segment’s spacing
is written alongside its arrow. The figure is not drawn to scale. Dashed
∑
N
1
s = ∑N ⋅ wi si , (1) arrows to/from the depot indicate deadhead segments. The table beside
i=1 wi i=1 the figure shows the sequence of stops on each route.
Table 1 shows segments and their weights in different weighting
Now consider the following three weighting schemes. These are not
schemes. The segments are named by the scheme a_b, where a is the
the only possible weighting schemes, but they are intuitive and can be
beginning stop and b is the ending stop. Because segments are direc­
calculated directly from most GTFS files.
tionally specific, stops such as 1 and 4 are linked by two different seg­
ments—one of which (4_1) is traversed more often. The last three

3
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

walking to/from stops and in-vehicle time for marginal changes in

Buses Stop
Route per day Sequence spacing along this segment. Setting aside the fact that no one walks
Blue along Interstate 30, the segment is so long that no one would walk to a
120 1-2-3-4-1
20 stop from the middle of the segment, so somewhat increasing or
0m 0m Green
decreasing the spacing would not substantially change anyone’s walk
30 60 5-3-4-1-4-3-5
times. Of course, the entire route 383 need not be excluded: upon
leaving the interstate, the bus traverses a handful of closely-spaced
100m 100m segments; their spacings are both typical for DART and relevant to the
trade-off between access and in-vehicle time.
While there are other defensible ways to remove long segments (such
as outlier detection), we favor a constant threshold as it is transparent
0m and intuitive. In particular, we can apply the same threshold to every
90
city whereas outlier-based techniques would apply different thresholds
to each city. Also, for cities with narrow spacings, techniques such as
Inter Quartile Range (IQR) filtering might eliminate fairly ordinary
segments5. Alternatively, express or regional routes could be filtered out
if the GTFS feed uses extended GTFS route types6. However, in practice,
Fig. 2. Example network with two bus lines (not drawn to scale).
very few agencies use this extension.

Table 1 3. Methodology
Mean spacings for the example network.
Weight of Segment
Following the definitions proposed above, this section describes a
methodology for data collection, processing, and analysis to use GTFS
Segment Route Spacing [m] Segment- Route- Traversal-
files for stop spacing statistics. We have created a python package called
1_2 Blue 100 1 1 120 gtfs-segments (Devunuri and Lehe, 2023) which replicates the method­
2_3 Blue 100 1 1 120
ology in this paper to download the latest sources, process data, and
3_4 Blue & Green 200 1 2 120+60
4_1 Blue & Green 300 1 2 120+60
calculate stop spacings. The package is publicly hosted on GitHub at htt
5_3 Green 900 1 1 60 ps://github.com/UTEL-UIUC/gtfs_segments. The GitHub repository
4_3 Green 200 1 1 60 hosts detailed documentation for readers interested in using the
1_4 Green 300 1 1 60 package.
3_5 Green 900 1 1 60
The GTFS data for a transit provider is organized as a folder (usually
Average Stop Spacing (m) 375 350 300
stored as a zip file) which contains various tabular text files created in
accordance with the GTFS Schedule7 specification. Of these, the files
columns show each segment’s weight according to each of the weighting required for gtfs-segments to run are stops.txt, routes.txt, trips.txt,
schemes described above. The bottom row shows the mean spacings as stop_times.txt, calendar.txt (or calendar_dates.txt), and shapes.txt. The
calculated by each weighting scheme. Note the means are substantially last of these, shapes.txt, is listed as “optional” in the GTFS Schedule
different from one another. The traversal-weighted mean is the smallest specification, and some agencies do not include it in their public feeds. It
because the shorter segments have more traversals. The route-weighted contains trip shapes: latitude/longitude coordinates for the paths of all
mean is the next largest, and the segment-weighted is the largest. routes. These shapes are used to calculate the distances between stops
Which of the three means is the best? Ultimately, there is no “true,” and to split routes into segments. Every shape has a unique ID that
“natural” or “genuine” mean. But among the three, traversal-weighting corresponds to a particular trip.
has the advantage of being tied to a physical event: the traversal- The GTFS feed has schedules for all days of operations between the
weighted mean is the mean distance that a bus travels between stops. feed start and end dates. To compare spacings between different cities,
If planners decided to cosmetically divide a route into two routes with we only use trips served during the busiest day of the schedule, defined to
different names, then the route-weighted mean spacing would change be the day with the most service_ids. In every case, the busiest day is a
even if nothing about the passenger experience changed. Also, unlike weekday. Using calendar.txt and calendar_dates.txt, we identify the
segment-weighting, the traversal-weighted mean favors the parts of a busiest day in the schedule and obtain the corresponding service_ids.
network that have the most service. So if a provider occasionally runs a These service_ids are used to filter the trips in trips.txt, and segments for
bus out to a distant suburb, the traversal-weighted mean is less affected each trip are created based on consecutive stops in stop_times.txt. The
than the segment-weighted is. stop_times.txt file includes optional columns called ‘pickup_type’ and
‘dropoff_type’, which indicate trips and scheduled stops for which ser­
2.3. Threshold vice is not provided. We use these to try and eliminate deadhead seg­
ments. Trips with only one stop time listed are also eliminated as non-
The last concept we propose is the threshold. The threshold is the revenue trips.
upper bound of spacing used in calculating a “truncated” mean (or other With the data downloaded and relevant trips identified, we proceed
statistic). For example, later, we give a mean spacing of all segments to split trip shapes into ‘segments’ (as defined in Section 2) which can be
shorter than 2 km for real bus service providers. The benefit of using a measured. This splitting is challenging because GTFS does not pair the
truncated mean to describe a network is best captured by an edge case.
Fig. 3 shows a segment of a DART express bus route running from a
suburban bus station into downtown Dallas. The segment is about 21 km 5
Based on IQR-based filtering, the Q3 + 1.5 × IQR value for Chicago Transit
long and runs almost entirely on Interstate 30. In considering stop pol­
Authority is 358 m, which is around the typical mean spacing for the US
icies for DART, or in comparing DART to other networks, it may be 6
This extension has added categories that distinguish between regular (Code-
helpful to use a mean that excludes segments such as this one. In the first 700), regional (Code − 701), express (Code-702), and shuttle buses (Code-711)
place, the segment is an extreme statistical outlier that will skew results. to name a few. Find more info at https://developers.google.com/transit/gtfs/re
Second, the utility of using stop spacing statistics is largely related to the ference/extended-route-types
question of consolidation, and there is no trade-off between time spent 7
See https://gtfs.org/schedule/for more details.

4
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Fig. 3. 20.9 km segment on DART Express Bus Route 383 - Lake Ray Hubbard Express.

stop locations (latitude/longitude pairs) in the stops.txt file to points on of Fig. 4, the increase in resolution and using k-nearest neighbors
each trip’s shape. Within the official GTFS documentation, there is no with k = 3 snaps the stops in the correct order.
stipulated maximum distance between stop locations and the associated • Finally, once every stop has been snapped to a coordinate on the trip
route or trip shapes. However, GTFS best practices recommend that this shape, the intermediate points between stops are collected into
distance be no more than 100 meters. For the top 10 US cities (by segments. These are stored as LineStrings that represent the
ridership), we noticed that the stops were on average, less than 20 m geographic path of the segment. All LineStrings are projected onto
away from the nearest point on the trip shapes. Nonetheless, there is a the Mercator projection (EPSG:4326/WGS 84) to ensure consistency.
huge variation in this distance and we notice stops that are farther than The distance metric of each LineString is then calculated and
100 m from their respective shapes. Stops thus need to be “snapped” recorded as the segment’s spacing.
onto the trip’s shape before segmentation. Intuitively, one could project
each stop’s location onto the nearest point on the trip shape or match The process above takes about three to five minutes on a standard
each stop to the shape’s nearest coordinate. However, both approaches desktop PC for a large transit system such as the Chicago Transit Au­
may lead to errors whereby a stop is mapped to an out-of-order point on thority. When processing is finished, we produce a table in the form of a
the trip shape. This happens especially when trips involve loops. GeoPandas GeoDataFrame (Jordahl et al., 2020). The GeoDataFrame is a
Fig. 4 shows an example route with five stops in the sequence 1 → 5. Python object with methods for filtering, sorting, grouping, and visu­
The coordinates of the trip shape are represented as yellow dots, and the alizing geospatial data. We use this to compute statistics.
coordinates of the stops are represented as numbered dots. In the first
panel, snapping each stop to the nearest given coordinate leads to two 4. Results
problems: First, the stop ordering is incorrect, because the bus visits
stops 3 and 4 before stop 2. Second, stop 5 is snapped to a point that is far Following the methodology detailed above, we have created open-
off from its projection onto the trip shape. To overcome these challenges, source datasets containing all segments in the United States and Can­
we use the following procedure. ada and hosted them on the Harvard Dataverse service. The US re­
pository is hosted at Devunuri (2023), and the Canadian one at Devunuri
• First, we increase the resolution of the trip shape by adding points et al. (2023). All feeds used were updated in late 2022 or early 2023 and
(via linear interpolation between the given coordinates) until no two not listed as “deprecated” on the MobilityData Database (MobilityData,
consecutive points are separated by more than five meters. Added 2024). Each provider’s data is gleaned from GTFS feeds updated in late
points are indexed appropriately to indicate the ordering of points 2022 or 2023. In the dataset, each provider has one folder. Each folder
along the route. In the second panel of Fig. 4, the added points are contains both tabular and GeoJSON files for examining and mapping the
represented by grey dots. Increasing the resolution successfully snaps provider’s data, histograms of spacings, and summary statistics calcu­
5 to a closer point along the route, but it does not solve the problem lated using each of the weighting schemes described above. In total, our
of out-of-order stops. data consisted of 660 feeds from 539 providers in the US, 83 in Canada,
• Next, to order stops correctly, we transform the shape’s coordinates and 38 in cities in other countries.
(including the added coordinates) into a k-dimensional tree data National means and medians, calculated using our three weighting
structure 8. As a first pass, we use k = 3. Using this tree, the k-nearest systems, appear in Table 2 for the 539 US and 83 Canadian bus providers
neighbors (i.e., the three closest points on the shape) are identified in our database. The “threshold” chosen for calculating these statistics is
for each stop location in the trip. Among the k neighbors, we choose 2 km: all segments longer than 2 km are excluded. We chose this
the closest one to the stop (using Haversine distance) which is also in threshold because it excludes fewer than 2% of segments in nearly every
the correct order. If, for any stop, there are no neighbors on the trip city (thereby giving a complete view) but still excludes segments too
shape in the correct order, then we double the k value and start over. long for most people to walk from halfway between the stops (i.e.,
This process continues either until we have a suitable ordering or we segments for which the marginal change in spacing would not result in
discard the trip if no solution is found even with k set equal to the shorter realized walk times).
number of coordinates given for the entire route. In practice, for a Note the national statistics reported in Table 2 are weighted by their
typical agency only a handful of trips are discarded9, and upon corresponding measures: i.e., the traversal-weighted mean is not the
manual inspection we have noticed these are almost always mis­ mean traversal-weighted mean of all cities, but rather the mean spacing
labeled somehow (i.e., cutting through lots). Notice in the last panel traversed in the US overall. The three means and medians for both
countries are ordered as they were in the example network above: the
traversal-weighted means are narrower than the route-weighted means,
which are narrower than the segment-weighted means. This ordering is
8
We used the ‘spatial.KDTree’ function from the scipy package intuitive: relative to suburban routes, segments located in dense urban
9
In the case of the top 30 largest agencies in the United States, on average areas are shorter, traversed by more routes, and traversed more
only about 0.1% of all trips are discarded for failing to snap.

5
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Snapped
k-Nearest
Original Coords
Added Coords
5 5 5

2 2 2
4 4 4

3 3 3

1 1 1
Original Increased Resolution Increased Resolution
k=1 k=1 k=3
Fig. 4. Improvement in snapping due to increase in resolution and k-nearest neighbors.

5. Discussion
Table 2
Summary statistics for the United States (539 providers) and Canada (83 pro­
This section discusses uses of the methodology and results derived
viders) with threshold of 2 km.
above.
Attribute Value

US Canada 5.1. Advantages of isolating segments

Traversal-weighted Mean [m] 352 347
Route-weighted Mean [m] 389 353 In Section 3, we described a methodology and introduced a python
Segment-weighted Mean [m] 401 367
package for splitting bus networks into segments. However, it is possible
Traversal-weighted Median [m] 295 287
Route-weighted Median [m] 317 290
to obtain some means without isolating segments. The average spacing of
Segment-weighted Median [m] 326 298 each trip in the schedule can be calculated by measuring the distance of
% Segments Excluded (2 km Threshold) 1.89 1.50 each trip’s shape and then dividing that distance by the number of stops
on the trip. By averaging the average spacing of all trips, one obtains a

Table 3
Unit conversions.
stops/mile 10 9 8 7 6 5 4 3 2

meters/stop 161 179 201 230 268 322 402 536 805
feet/stop 528 587 660 754 880 1056 1320 1760 2640
stops/km 6.2 5.6 5.0 4.3 3.7 3.1 2.5 1.9 1.2

frequently. Table 3 converts between units which may be more intuitive. traversal-weighted spacing. By counting the average spacing of a trip on
Table 4 lists summary statistics spacings for thirty of the largest bus each route once, one obtains a route-weighted spacing. What advan­
service providers by ridership in the United States, and Table 5 does the tages, then, are there to isolating segments individually? Isolating seg­
same for the ten largest providers in Canada. The cities are listed in ments permits:
increasing order of their traversal-weighted mean spacing. Note that, to
save space, we have left route-weighted statistics out of Tables 4 and 5, (i) statistics which by their nature depend on counting individual
because—as noted above—this system is more arbitrary than the other spacings, such as medians, quartiles, standard deviations and the
two, being dependent on how routes are named10 in the GTFS files. segment-weighted mean;
Table 6 reports mean spacings for 38 cities outside the US and (ii) particular spacings to be filtered (e.g., to discard segments which
Canada. The list of cities in Table 6 is very much a convenience sample. are longer than a threshold or which fall outside some political
We were limited by what providers’ GTFS files were possible to obtain, boundary);
suitable for comparison and published with the optional shapes.txt file (iii) the creation of histograms of the complete distribution of spacing.
needed to calculate spacings. Obstacles to obtaining international
comparisons are discussed further in Section 5. Figure 5 shows spacing distributions for traversal-weighted histo­
grams and kernel densities of stop spacings for Chicago, Kansas City,
Montreal, and Toronto’s bus systems. In every city we have examined,
the distribution has a similar shape: unimodal with a right skew.

10
One way to ameliorate this arbitrariness is to group trips by their ‘route_
short_name’ property, but many agencies do not provide this property.

6
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Table 4
Stop Spacings for top 30 transit providers in the United States (Threshold - 2 km).
Provider Urbanized Area Traversal Segment Weighted Traversal Weighted Segment Weighted % Segments Excluded
Weighted Mean Mean [m] Median [m] Median [m] (2 km Threshold)
[m]

Southeastern Pennsylvania Philadelphia, 217 274 172 193 0.313

Transportation Authority (SEPTA) PA
Chicago Transit Authority (CTA) Chicago, IL 230 242 206 206 0.142
San Francisco Municipal San Francisco, 261 264 219 209 0.368
Transportation Agency (SFMTA) CA
Pittsburgh Regional Transit (PRT) Pittsburgh, PA 273 304 209 221 1.112
Massachusetts Bay Transportation Boston, MA 284 301 239 239 0.461
Authority (MBTA)
MTA - New York City Transit (NYCT) New York, NY 289 317 236 244 0.62
& MTA Bus Company
Metro Transit Minneapolis, 298 294 227 226 0.635
MN
Washington Metropolitan Area Washington, DC 300 321 245 257 0.416
Transit Authority (WMATA)
Metropolitan Atlanta Rapid Transit Atlanta, GA 308 315 263 265 0.572
Authority (MARTA)
METRO Houston Houston, TX 309 326 261 268 0.758
Dallas Area Rapid Transit (DART) Dallas, TX 310 319 257 263 0.504
Bi-State Development Agency St. Louis, MO 316 334 264 277 0.368
(METRO)
Milwaukee County Transit System Milwaukee, WI 326 342 296 303 0.187
(MCTS)
TriMet Portland, OR 335 349 285 294 0.868
Maryland Transit Administration Baltimore, MD 339 393 276 305 0.765
Alameda-Contra Costa Transit District Oakland, CA 348 377 294 307 0.54
(AC Transit)
New Jersey Transit (NJ Transit) Newark, NJ 350 448 265 331 2.194
VIA Metropolitan Transit (VIA) San Antonio, TX 353 349 275 276 1.227
TheBus Honolulu, HI 359 376 283 286 1.379
Broward County Transit Fort Lauderdale, 362 371 307 309 0.829
FL
Miami-Dade Transit Miami, FL 372 399 299 319 1.254
Metro Los Angeles, CA 381 396 352 363 0.926
Regional Transportation District Denver, CO 390 412 346 353 1.147
(RTD)
King County Metro Transit Seattle, WA 391 406 345 348 0.876
San Diego Metropolitan Transit San Diego, CA 406 429 341 355 1.763
System (MTS)
Valley Metro (VM) Phoenix, AZ 433 451 401 405 0.449
Orange County Transportation Orange County, 441 451 386 393 0.731
Authority (OCTA) CA
Regional Transportation Commission Las Vegas, NV 444 441 362 367 0.24
of Southern Nevada (RTC)
Santa Clara Valley Transportation San Jose, CA 463 483 395 404 1.331
Authority (VTA)
Capital Metro Austin, TX 484 506 396 402 1.944
Summary 346 366 290 299 0.831

Table 5
Stop Spacings for top 10 transit providers in Canada (Threshold - 2 km).
Provider Urbanized Traversal Weighted Segment Weighted Traversal Weighted Segment Weighted % Segments Excluded
Area Mean [m] Mean [m] Median [m] Median [m] (2 km Threshold)

Winnipeg Transit Winnipeg, MB 263 274 219 226 0.192

Societe de transport de Montreal Montreal, QC 281 308 240 251 0.532
Toronto Transit Commission Toronto, ON 322 337 271 273 0.437
Reseau de transport de la Capitale Quebec, QC 325 304 285 263 0.559
Edmonton Transit Service Edmonton, AB 330 341 272 282 1.116
Ottawa-Carleton Regional Transit Ottawa, ON 333 351 261 273 2.164
Commission (OC Transpo)
Calgary Transit Calgary, AB 373 390 309 318 1.541
MiWay Mississauga, 374 377 303 305 1.004
ON
TransLink Vancouver Vancouver, BC 378 396 308 322 1.468
Brampton Transit Brampton, ON 426 406 327 312 0.632
Summary 341 348 280 283 0.965

5.2. Comparisons consolidation has often compared average spacings among cities (even if
there have not been clear methodologies or definitions for how to do so).
As pointed out in the introduction, communication around stop Thus, the most significant finding of our research is probably that US bus

7
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Table 6
Stop Spacings for 38 international cities (Threshold - 2 km).
Provider Urbanized Country Traversal Segment Traversal Segment % Segments
Area Weighted Mean Weighted Mean Weighted Median Weighted Median Excluded (2 km
[m] [m] [m] [m] Threshold)

Azienda Napoletana Mobilita Napoli Italy 293 315 256 262 0.551
Colectivos Buenos Aires Buenos Aires Argentina 309 306 256 244 1.007
Empresa de Transportes e Transito Belo Brazil 325 319 266 260 0.43
de Belo Horizonte (BHTRANS) Horizonte
Empresa Municipal de Transportes Madrid Spain 332 359 293 307 0.651
de Madrid (EMT Madrid)
Santiago DTPM Santiago Chile 339 362 302 307 0.391
Transports Metropolitans de Barcelona Spain 346 348 324 315 0.176
Barcelona (TMB)
Etufor Fortaleza Brazil 349 352 284 271 0.161
EMT Valencia Valéncia Spain 351 341 307 296 0.95
Azienda Trasporti Milanesi (ATM) Milan Italy 357 415 312 343 0.286
Dublin Bus Dublin Ireland 360 414 315 337 0.408
Toei Bus Tokyo Japan 364 372 335 338 0.057
Azienda Tramvie e Autobus del Rome Italy 369 407 306 321 0.52
Comune di Roma
SBS Transit (SBST) Singapore Singapore 369 373 345 340 0.943
Tours Métropole Val de Loire Tours France 377 415 336 348 0.274
Adelaide Metro Adelaide Australia 378 413 335 350 1.045
Wiener Lokalbahnen (WLB), Wiener Vienna Austria 378 414 345 370 0.007
Linien
Metro Christchurch Christchurch New 379 401 329 338 0.565
Zealand
City Transport Secretary of Rio de Rio de Brazil 379 427 312 333 1.022
Janeiro Janeiro
dBus San Sebastian Spain 392 399 345 344 0.061
Turku region public transport Föli Turku Finland 401 521 350 430 0.427
STIB / MIVB Bruxelles Belgium 402 415 379 389 0.188
Tisséo Toulouse France 407 464 354 381 0.646
Carris Metropolitana Lisboa Portugal 410 453 334 363 1.155
Berliner Verkehrsbetriebe Berlin Germany 412 433 376 389 0.395
Kochi buses (KSRTC) Kochi India 455 502 405 424 0.123
Transport Canberra Canberra Australia 465 438 370 346 3.21
Aachener Straßenbahn und Aachen Germany 466 512 415 433 1.291
Energieversorgungs
Budapesti Közlekedési Központ Budapest Hungary 476 498 422 433 1.079
(BKK)
Gemeente Vervoerbedrijf (GVB) Amsterdam Netherlands 489 525 416 445 1.425
KVB Kölner Verkehrs - Betriebe AG Cologne Germany 508 540 453 470 0.977
Warszawski Transport Publiczny Warsaw Poland 516 580 458 502 0.609
(ZTM Warszawa)
MPK Wroclaw Wroclaw Poland 532 625 477 528 1.008
Klaipeda Transport Klaipeda Lithuania 556 701 505 589 1.177
Kauno viešasis transportas (KVT) Kaunas Lithuania 594 644 557 588 0.828
Rigas Satiksme Riga Latvia 596 620 520 539 1.024
Helsingin seudun liikenne (HSL) Helsinki Finland 611 484 555 404 1.712
Zarzad Transportu Publicznego w Krakow Poland 616 674 550 604 1.528
Krakowie (ZTP Krakow)
Vilnius Transport Vilnius Lithuania 629 672 545 596 1.713
Summary 429 459 377 392 0.79

real spacings are generally wider than the 160–230 m range (7–10 stops While US cities generally have wider spacings than thought, in
per mile) offered by Reilly (1997) or the 200 m offered by Trans­ another way the conventional wisdom (as reflected by the three quotes
portation Research Board (2001). SEPTA (Philadelphia) and CTA (Chi­ given in the introduction) seems to be right. In our sample of 38 inter­
cago) are the only large US systems with traversal-weighted mean national cities, spacings abroad are generally wider than in the US (and
spacings shorter than 230 m. This is so even though we have excluded all Canada). The mean spacing in the international sample is wider than all
spacings longer than 2 km. The traversal-weighted mean spacing for ten Canadian systems, and all but five of the thirty largest US systems.
United States and Canada (truncated from below at 2 km) are even Fig. 6 displays a “strip plot” for the 78 cities displayed in Tables 4, 5 and
wider: about 350 m (about 5 stops per mile). The national medians are 6. Each city’s spacing appears as a dot. Among cities in the international
lower: between 5 and 6 stops per mile. sample, Eastern Europe (Poland and the Baltic countries) have the
A cursory look at city-level data suggests that cities largely built widest spacings—above 500 meters (just fewer than three stops per
before the automobile area have narrower spacings than those that have mile).
developed since. In the US, Philadelphia, Chicago, Boston, Pittsburgh, Still, while the international data are suggestive, we would caution
and San Francisco have much shorter-than-average spacings while Las against drawing confident conclusions from the data depicted in Table 6.
Vegas, Austin, and Santa Clara County have much wider spacings. Our databases contain hundreds of cities in the US and Canada, but data
Similarly, in Canada, Mississauga and Brampton have wider spacings
than their immediate neighbor Toronto. The relationship between stop
spacing, urban form, and development history is a question worth
investigating in further research.

8
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

(a) CTA, Chicago (b) RideKC, Kansas City

(c) Société de transport de Montréal, Montreal (d) Toronto Transit Commission, Toronto
Fig. 5. Traversal-weighted distributions of stop spacings.

bus system comparable to US and Canadian systems. In some places, this

is because one agency serves a large region: e.g., in Norway, Ruter AS
345.7 serves Oslo and Akershus counties, including vast rural areas. In the
United Kingdom, by contrast, cities are served by numerous private12
providers which operate across cities or even nationally. In either case,
340.5 one could ostensibly filter all segments from within some political
boundary, but we have not done so.

428.6
5.3. Alternative metrics

Before concluding, we now consider two alternative ways to calcu­

late stop spacings. These are intended as proofs-of-concepts showing
what can be accomplished by combining GTFS data with other data
sources.
Fig. 6. Comparison of stop spacings between US, Canada, and other countries. The first is a new weighting scheme: passenger-weighting. By this
scheme, segments are weighted by the average number of passengers
from other countries proved much harder to obtain for two major rea­ who traverse the segment. Just as the traversal-weighted mean is the
sons. First (and most importantly), providers outside the US and Can­ average spacing “experienced” by a bus, the passenger-weighted mean is
ada11 rarely include the optional shapes.txt file needed to determine the average spacing experienced by a passenger. The MBTA (in Boston)
driving distances between stops. There is little relationship between an publishes the bus loads for a large sample of trips on their Open Data
agency’s size and whether it publishes a shapes.txt: e.g., Paris’ Ile-de- Portal (MBTA, 2019). Upon request, we obtained the same data from
France Mobilité does not but Toulouse’s Tisséo does. The second RTD in Denver. Fig. 7 shows each agency’s passenger-weighted distri­
obstacle was conceptual: many countries have no equivalents of a city bution. The MBTA’s passenger-weighted mean is 304 m; RTD’s is 401 m.

11 12
Inquiring with Google employees, we learned that navigation apps such as Having many private providers is not fatal to spacing comparisons. For
Google Maps generally use a combination of real-time GPS data and manual example, Buenos Aires organizes its privately-provided routes into a unified
operations to draw bus route shapes when shapes.txt is unavailable. GTFS file at https://data.buenosaires.gob.ar/dataset/colectivos-gtfs

9
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

(a) MBTA, Boston (b) RTD, Denver

Fig. 7. Distributions of passenger-weighted stop spacings.

However, neither agency samples all scheduled trips, so these means are the United States and Canada nationally, as well as thirty large US
not comparable to the other means reported. It would be possible, how­ providers, ten Canadian ones, and thirty-eight in other countries. Evi­
ever, to obtain comparable means by interpolating loads for missing dence shows that US (and Canadian) spacings are not quite as narrow as
trips. some quoted statistics suggest, but do seem to be somewhat narrower
The next proof-of-concept is an alternative distance metric: traffic than those in other countries—although the sample of international
signals traversed. The number of traffic signals, while usually not under providers is not large enough to make confident comparisons. Finally,
the control of a bus agency, also communicates information about how the paper presented, as proofs-of-concept, two spacing statistics that
much time it takes the bus to travel between stops. From various gov­ combine GTFS with outside data sources.
ernment sources, we have obtained data that includes the coordinates of In any discussion of stop spacings, a note of caution is in order: the
traffic signals in twelve cities. To calculate signals-traversed, we find the very idea of comparing spacing statistics should be taken with a grain of
number of signals falling within a 5.5 m buffer around each segment. salt. For example, Chicago’s mean spacing would be wider if the Chicago
Table 7 shows segment-, route- and traversal-weighted mean values of Transit Authority (which operates only in the dense City of Chicago) and
signals-per-segment for each of the twelve cities. Fig. 8 shows the full Pace (which operates in Chicago’s suburbs) were a single agency. Still,
traversal-weighted distributions of signals-per-segment. Most segments both agencies do have average spacings shorter than the US mean, so
have no signal, but the number can be as large as six in Austin. Using aggregates do reflect “composition effects” but also provide (at least
Maximum Likelihood Estimation, we fit a geometric distribution to each suggestive) evidence about genuine policy differences. Even though the
data, with the predicted probability masses shown in red in Fig. 8. The paper is premised on the idea that statistics are useful, one very general
geometric distribution fits surprisingly well. point that the authors hope the paper has made clear is how synthetic any
particular statistic is—dependent on choices, weightings, exclusions,
6. Conclusion etc. Hence, it is of foremost importance to be precise about how exactly
statistics were aggregated.
6.1. Summary

6.2. Further applications of the methodology

US transit agencies are currently reconsidering their stop policies,
and discussions about these policies sometimes reference “typical” or
This paper has focused on the use of our methodology to produce
“average” values for comparison. One task of the paper has been to
averages that can be used to compare agencies, but our methodology has
provide conceptual clarity to discussions of stop spacing statistics by
several other potentially-useful applications to transit planning and
introducing terminologies and weighting schemes. Another task has
operations. While we hope that the methodology will be applied in other
been to describe a methodology (implemented via the python package
ways we cannot anticipate, here we will describe two promising ones.
gtfs-segments) for isolating individual segments from GTFS data. The
First, for a single agency, averages can be used to track how spacings
concepts and tools together were used to provide spacing statistics for
change over time using historical GTFS files. This is useful because many

Table 7
Average signals-per-segment for 12 US cities.
Provider Urbanized Area State Segment-Weighted Mean Route-Weighted Mean Traversal-Weighted Mean

Capital Metro Austin TX 0.83 0.91 0.91

Bloomington Transit Bloomington IN 0.32 0.33 0.33
Chicago Transit Authority (CTA) Chicago IL 0.31 0.32 0.34
Jacksonville Transportation Authority (JTA) Jacksonville FL 0.39 0.40 0.45
Transit Authority of River City (TARC) Louisville KY 0.47 0.53 0.68
Miami-Dade Transit Miami FL 0.37 0.37 0.41
Central Florida Regional Transit Authority (LYNX) Orlando FL 0.23 0.23 0.22
Southeastern Pennsylvania Transportation Authority (SEPTA) Philadelphia PA 0.63 0.61 0.61
Pittsburgh Regional Transit Pittsburgh PA 0.33 0.41 0.50
San Francisco Municipal Transportation Agency (SFMTA) San Francisco CA 1.06 1.14 1.26
Hillsborough Area Regional Transit (HART) Tampa FL 0.28 0.28 0.33
King County Metro Seattle WA 0.38 0.56 0.66

10
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

(a) Austin, TX (b) Bloomington, IN

(c) Chicago, IL (d) Pittsburgh, PA

Fig. 8. Traversal-weighted distributions of signals-per-segment.

agencies have explicit targets for bus stop spacing that they aim to meet package produces a GeoDataFrame with rows representing isolated
by altering stop locations. Hence changes in averages over time can segments. Ordering these segments by spacing identifies the narrowest
show progress toward these goals. For instance, the Regional Trans­ or widest spacings in a network. In the context of stop consolidation, this
portation District (RTD) in Denver13 undertook a long-term stop ranking can focus planners’ attention on candidate segments for stop
consolidation initiative, aiming for a 20% reduction in stops and a target removal. Ranking is also useful to identifying potential errors in GTFS
spacing of approximately 402 meters (1/4th of a mile). Fig. 9 illustrates files. As an example, Fig. 10 shows the stop at 35th Street and Western
the changes in both segment-weighted and traversal-weighted stop on route 35 (CTA - Chicago). One of the westbound stops at 35th Street
spacings for each runboard of RTD from January 2017 to May 2022, and Western Avenue is placed incorrectly in the GTFS file. As a result,
based on historical GTFS data. Prior to the consolidation efforts, the the segment between the consecutive stops named ‘35th Street and
traversal-weighted spacing was around 378 meters, which gradually Western’ and ‘35th Street and Western Avenue’ (in the westbound di­
increased toward 400 meters by January 2020. rection) measures only 38 m instead of the actual 87 m. We spotted this
The data after January 2020 also has a story to tell. Observe in Fig. 9 error when, after ranking segments by spacing, we noticed a 38 m
there is a substantial reduction in spacings with the onset of the COVID- spacing which seemed suspiciously narrow. It is also easy to appreciate
19 pandemic, as RTD reduced its service by 30%. Since they primarily how this error was made: Western Avenue has a large median (the green
eliminated peripheral service with large spacings, both the segment- area in the photograph), and the stop was placed at the corner of the
weighted and traversal-weighted means fell immediately. The median rather than the upstream block. Fortunately, we have not
segment-weighted mean exhibited a much higher drop than the noticed many such errors while working with GTFS14 files, but the
traversal-weighted mean, as they tended to cut low-frequency routes. example goes to show a potential use for validation.
Afterwards, a period of recovery shows a rise in both the means, with the
traversal-weighted mean eventually reaching pre-pandemic levels and
the desired target spacing. The segment-weighted mean has never 6.3. Further research
reached its prior peak, though, because the agency did not restore some
low-frequency routes with wide spacings. To conclude, we offer ideas for further research.
A second promising use of the methodology is in the identification of A broad question that these statistics pose is: Why are US spacings
particularly narrow or wide spacings. As noted above, the gtfs-segments narrower than those abroad? Per optimization models, spacings should
generally be narrower when the rate of demand (measured per hour per

13 14
See also the Maryland Transit Administration’s Bus Stop Design Guide which Note that, in spite of this error, the overall mean spacing for this route
sets explicit average spacing targets for different areas (Maryland Transit remains accurate since the segment upstream (between ‘35th Street and Oakley’
Administration, 2019). and ‘35th Street and Western’) is commensurately longer.

11
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

Fig. 9. Changes in stop spacings means over time for the Regional Transportation District (RTD) in Denver.

Fig. 10. Less than 50 m segment on CTA Bus Route 35 Westbound.

kilometer) is higher (Wirasinghe and Ghoneim, 1981). Yet many US The ideas and tools presented in this paper can also be used for more
cities have much less land-use density and ridership than European cities general research. One way to proceed is to use stop-by-stop passenger
and thus should display (at least intuitively) a lower rate of demand. data, such as the passenger count data provided by MBTA and RTD in
From conversations with transit providers, our impression is that the Section 5.3, to identify relationships between changes in ridership and
modern American emphasis on soliciting community input at all stages changes in bus stop spacing. When a stop is removed, do all passengers
of decision-making makes it difficult to remove stops. Another possible use nearby stops, or do any trips vanish? Do the quicker bus speeds (and,
explanation lies in the fact that the US developed its horsebus and trolley potentially, shorter headways) afforded by consolidation boost demand
systems somewhat earlier than Europe. American bus systems are the on consolidated routes? Because GTFS and ridership data can be ac­
heirs of these systems, so either the stop locations themselves or a quired over time, it should be possible to use time-series techniques such
general sense of proper spacing may have been cemented in an era with as regression discontinuity and difference-in-differences to answer such
low expectations for speed. These are only speculations, however, and a questions. Furthermore, while Section 5.3 has illustrated how to create a
thorough research into this question would undoubtedly be welcome passenger-weighted metric from data on bus loads, it would also be
and timely. possible to weight segments by the number of passengers who board and
A more technical avenue for future research is to incorporate the alight at each segment.
GTFS Realtime specification, which supplies live bus location informa­ Finally, the fine-grained data provided by our methodology is well-
tion online. One use of this data would be to compute metrics of realized suited to the geospatial analysis of stop changes. One might identify
bus stop spacings. While GTFS Schedule data records the locations of local correlates of stop removal, by linking schedule information to
stops, typically buses only actually visit a stop if a passenger wants to demographic and geographic data. Which types of locations and
alight or if there is someone waiting at the stop. By collecting real-time neighborhoods tend to have stops removed? Or, apart from removal, one
bus locations, it would be possible to record when buses actually stop might simply identify correlates of current spacings. For example, many
and thereby measure the actual distances traveled from one stopping US agencies seem to put a stop on every or every other block in urban
operation to another. This is important to understanding stop removal, settings, and in this case stop spacing should be highly correlated with
because on low-demand routes where buses only actually visit a small block length. Time will tell whether these or other questions are of in­
fraction of stops (as in smaller communities), spacings may greatly un­ terest to the research community, but we offer them as examples of how
derestimate how far the bus can travel freely. Similarly, GTFS Realtime data on isolated stop spacings may be used to address broader topics.
data could be used to measure (weighted in various ways) travel time
between stops as a distance metric.

12
S. Devunuri et al. Journal of Public Transportation 26 (2024) 100083

CRediT authorship contribution statement deVries Kehoe, O., 2004. Effects of Bus Stop Consolidation on Transit Speed and
Reliability: A Test Case. Ph.D. thesis. University of Washington. 〈https://kehoe.
org/owen/docs/Effects_of_stop_consolidation.pdf〉.
Lewis J. Lehe: Conceptualization, Project administration, Supervi­ Kehoe, O., Podolsky, M., Kappes, G., 2021. Transit Speed & Reliability: Guidelines &
sion, Writing – original draft, Writing – review & editing. Saipraneeth Strategies. 〈https://kingcounty.gov/~/media/depts/metro/about/planning/sp
Devunuri: Conceptualization, Data curation, Methodology, Software, eed-reliability-toolbox.pdf〉.
LaFleur, P., 2019. Metro just removed 400 bus stops.Was yours one of them? 〈https
Visualization, Writing – original draft, Writing – review & editing. Dana ://www.wcpo.com/news/transportation-development/public-transit/cincinnati-m
Monzer: Conceptualization, Methodology. Ayush Pandey: Conceptu­ etro-program-removed-hundreds-of-bus-stops-over-the-weekend〉.
alization, Methodology, Software. Shirin Qiam: Conceptualization, Li, H., Bertini, R.L., 2009. Assessing a Model for Optimal Bus Stop Spacing with High-
Resolution Archived Stop-Level Data. Transp. Res. Rec.: J. Transp. Res. Board 2111,
Data curation, Methodology, Software, Validation. 24–32. https://doi.org/10.3141/2111-04.
Maryland Transit Administration, 2019. Bus Stop Design Guide ∣ Maryland Transit
Declaration of Competing Interest Administration. 〈https://www.mta.maryland.gov/bus-stop-design-guide〉.
MBTA, 2019. MBTA Bus Ridership by Trip, Season, Route/Line, and Stop. 〈https://mbta-
massdot.opendata.arcgis.com/datasets/MassDOT::mbta-bus-ridership-by-trip-sea
There are no conflicts of interest. This is also stated in the declaration son-route-line-and-stop/explore?filters=eyJzZWFzb24iOlsiRmFsbCAyMDE5Il19〉.
of interest form. McHugh, B., 2013. Pioneering Open Data Standards: The GTFS Story. Beyond
Transparency: Open Data and the Future of Civic Innovation. Code for America Press
San Francisco, pp. 125–135. 〈https://beyondtransparency.org/part-2/pioneerin
Acknowledgments g-open-data-standards-the-gtfs-story/〉.
Miatkowski, P., Hovenkotter, K., 2019. Bus Stop Balancing: A Campaign Guide for
Agency Staff. 〈https://transitcenter.org/wp-content/uploads/2019/07/BusS
The authors would like to thank Rafael H. M. Pereira at the Institute
topBalancing_Final_061719_Pages-1.pdf〉.
of Applied Economic Research in Brasília, who gave advice throughout MobilityData, 2024. Mobility Database. 〈https://database.mobilitydata.org/〉.
the project. The authors would also like to thank Jonathan Wade and the Mohring, H., 1972. Optimization and Scale Economies in Urban Bus Transportation. Am.
team from RTD Denver who provided us the passenger count data. Econ. Rev. 62, 591–604. 〈https://www.jstor.org/stable/1806101〉.
Moloney, S., 2021. Final Bronx Bus Redesign Plan May Involve the Removal of 18
Percent of Bronx Bus Stops - Norwood News. 〈https://www.norwoodnews.org/final-
References bronx-bus-redesign-plan-may-involve-the-removal-of-18-percent-of-bronx-bus-sto
ps/〉.
Berez, D., 2015. The Bus Stops Here: Best Practices in Bus Stop Consolidation and Morency, C., Trépanier, M., Demers, M., 2011. Walking to transit: An unexpected source
Optimization 〈https://escholarship.org/uc/item/4dd8h1vs〉. of physical activity. Transp. Policy 18, 800–806. https://doi.org/10.1016/j.
Blazina, E., 2020. Port Authority’s initial bus stop eliminations showing on-time tranpol.2011.03.010. 〈https://www.sciencedirect.com/science/article/pii/S09670
improvements. Pittsburgh Post-Gazette 〈https://www.post-gazette.com/news/t 70×11000631〉.
ransportation/2020/02/23/Port-Authority-bus-stops-on-time-performance-impro MTA, 2022. Queens Bus Network Redesign: New Draft Plan. 〈https://new.mta.info/
vements-efficiency-transit/stories/202002230032〉. queens-bus-redesign-draft-plan-hi-res〉.
Devunuri, S., 2023. Bus Stop Spacings for Transit Providers in Canada. 10.7910/DVN/ Ouyang, Y., Nourbakhsh, S.M., Cassidy, M.J., 2014. Continuum approximation approach
QFTAPM. to bus network design under spatially heterogeneous demand. Transp. Res. Part B:
Devunuri, S., Lehe, L., 2023. GTFS Segments: A fast and efficient library to generate bus Methodol. 68, 333–344. https://doi.org/10.1016/j.trb.2014.05.018. 〈https://www.
stop spacings. 〈https://github.com/UTEL-UIUC/gtfs_segments〉, 10.5281/ sciencedirect.com/science/article/pii/S0191261514001076〉.
zenodo.10019419. Pandey, A., Lehe, L., Monzer, D., 2021. Distributions of Bus Stop Spacings in the United
Devunuri, S., Qiam, S., Lehe, L., 2023. Bus Stop Spacings for Transit Providers in the US. States. Findings. https://doi.org/10.32866/001c.27373.
10.7910/DVN/SFBIVU. Pereira, R.H.M., Andrade, P.R., Vieira, J.P.B., 2023. Exploring the time geography of
El-Geneidy, A.M., Strathman, J.G., Kimpel, T.J., Crout, D.T., 2006. Effects of bus stop public transport networks with the gtfs2gps package. J. Geogr. Syst. 25, 453–466.
consolidation on passenger activity and transit operations. Transp. Res. Rec. 1971, https://doi.org/10.1007/s10109-022-00400-x. 〈https://github.com/ipeaGIT/gtfs2g
32–41. https://doi.org/10.1177/0361198106197100104. ps〉.
Erhardt, G.D., Hoque, J.M., Goyal, V., Berrebi, S., Brakewood, C., Watkins, K.E., 2022. Reilly, J.M., 1997. Transit service design and operation practices in western european
Why has public transit ridership declined in the United States? Transp. Res. Part A: countries. Transp. Res. Rec. 1604, 3–8. https://doi.org/10.3141/1604-01.
Policy Pract. 161, 68–87. https://doi.org/10.1016/j.tra.2022.04.006. 〈https://www RTD Denver, 2023. Bus Stop Consolidation. 〈https://rtd-denver.vercel.app/about-rtd/p
.sciencedirect.com/science/article/pii/S0965856422000945〉. rojects/bus-stop-consolidation〉.
Flint, T., Ben-Amos, A., Ellis, P., Krykewycz, G., 2014. Piloting low-cost transit service Stewart, C., El-Geneidy, A., 2016. Don’t stop just yet! A simple, effective, and socially
enhancements through agency collaboration. World Transit Res. 〈https://www.wor responsible approach to bus-stop consolidation. Public Transp. 8, 1–23. https://doi.
ldtransitresearch.info/research/5393〉. org/10.1007/s12469-015-0112-9.
Furth, P.G., Rahbee, A.B., 2000. Optimal bus stop spacing through dynamic Tirachini, A., 2014. The economics and engineering of bus stops: Spacing, design and
programming and geographic modeling. Transp. Res. Rec.: J. Transp. Res. Board congestion. Transp. Res. Part A: Policy Pract. 59, 37–57. https://doi.org/10.1016/j.
1731, 15–22. https://doi.org/10.3141/1731-03. tra.2013.10.010. 〈https://www.sciencedirect.com/science/article/pii/S096585641
Garnham, J.P., 2020. To fight huge drop in bus riders, North Texas transit agency faces 3001900〉.
hard choices about who gets service. 〈https://www.texastribune.org/2020/02 Transportation Research Board, 2001. Making Transit Work: Insight from Western
/25/dallas-bus-ridership-plummeting-so-dart-wants-redraw-bus-routes-2020/〉. Europe, Canada, and the United States. Technical Report. National Academy Press.
Gordon, R., 2010. Muni may reduce stops to increase speed, save cash. 〈https://www. Washington, DC. 〈https://nap.nationalacademies.org/catalog/10110/making-tran
sfgate.com/bayarea/article/Muni-may-reduce-stops-to-increase-speed-save-cash- sit-work-insight-from-western-europe-canada-and-the〉, 10.17226/10110.
3168017.php〉. van Nes, R., Bovy, P., 2000. Importance of objectives in urban transit-network design.
Graehler Jr, M., Mucci, R.A., Erhardt, G.D., 2019. Understanding the Recent Transit Transp. Res. Rec. 1735, 25–34. https://doi.org/10.3141/1735-04.
Ridership Decline in Major US Cities: Service Cuts or Emerging Modes?, In: Vuchic, V.R., Newell, G.F., 1968. Rapid transit interstation spacings for minimum travel
Transportation Research Board 98th Annual Meeting Transportation Research time. Transp. Sci. 2, 303–339. https://doi.org/10.1287/trsc.2.4.303.
Board. 〈https://trid.trb.org/View/1572517〉. Wagner, Z., Bertini, R.L., 2014. Benefit–cost evaluation method for transit stop removal.
Grahn, R., Qian, S., Matthews, H.S., Hendrickson, C., 2021. Are travelers substituting Transp. Res. Rec. 2415, 59–64. https://doi.org/10.3141/2415-06.
between transportation network companies (TNC) and public buses? A case study in Wirasinghe, S.C., Ghoneim, N.S., 1981. Spacing of bus-stops for many to many travel
Pittsburgh. Transportation 48, 977–1005. https://doi.org/10.1007/s11116-020- demand. Transp. Sci. 15, 210–221. https://doi.org/10.1287/trsc.15.3.210. 〈htt
10081-4. ps://www.jstor.org/stable/25768017〉.
Jordahl, K., Bossche, J.V.D., Fleischmann, M., Wasserman, J., McBride, J., Gerard, J., Wu, T., Jin, H., Yang, X., 2022. To what extent may transit stop spacing be increased
Tratner, J., Perry, M., Badaracco, A.G., Farmer, C., Hjelle, G.A., Snow, A.D., Cochran, before driving away riders? Referring to evidence of the 2017 NHTS in the United
M., Gillies, S., Culbertson, L., Bartos, M., Eubank, N., Maxalbert, Bilogur, A., Rey, S., States. Sustainability 14, 6148. https://doi.org/10.3390/su14106148. 〈https://
Ren, C., Arribas-Bel, D., Wasser, L., Wolf, L.J., Journois, M., Wilson, J., Greenhall, A., www.mdpi.com/2071-1050/14/10/6148〉.
Holdgraf, C., Filipe, Leblanc, F., 2020.Geopandas/geopandas: V0.8.1. Zenodo.〈
https://zenodo.org/record/3946761〉, 10.5281/ZENODO.3946761.

13