What’s eating public transit in the United States? Reasons for declining transit ridership in the 2010s

Abstract

Despite ever-increasing public and financial support for public transportation in recent decades, transit ridership has dropped substantially in the 2010s across the United States. While many theories that explain the reasons behind the decline in ridership have been proposed, no consensus on the primary causes has been reached. By employing gradient boosting decision tree and counterfactual simulations, we examine the complex relationships between transit ridership and key internal and external factors from 2002 to 2017 in 85 of the largest urbanized areas in the US. Among several contributing factors, the declining cost of driving, measured by the decreasing share of carless households combined with lower gasoline prices, was the salient, most influential factor behind the recent decline. Neighborhood change in high-density neighborhoods also led to a moderate net loss in ridership. The only factor that has been mitigating further decline was an increase in transit services. Had this increase not occurred, loss of ridership would have been more than double the actual loss during the study period.

Full text

1
1
What’s Eating Public Transit in the United States? 2
Reasons for Declining Transit Ridership in the 2010s 3
4
5
6
7
8
9
10
11
12
Yongsung Lee 13
Assistant professor 14
Department of Geography 15
University of Hong Kong 16
Rm 10.35 10/F Jockey Club Tower 17
Centennial campus 18
Pokfulam Road, Hong Kong 19
Office: +852-3917-7107 20
Email: yongsung@hku.hk 21
22
23
24
Bumsoo Lee* 25
Associate Professor 26
Department of Urban and Regional Planning 27
University of Illinois, Urbana-Champaign 28
Temple Buell Hall Room #M206 29
611 Lorado Taft Drive 30
Champaign IL 61820 31
Office: +1-217-333-3601 32
Email: bumsoo@illinois.edu 33
34
35
36
37
* Corresponding author. 38
39
40
41
Lee, Y., & Lee, B. (2022). What’s eating public transit in the United States? Reasons for declining transit 42
ridership in the 2010s. Transportation Research Part A: Policy and Practice, 157, 126-143. doi: 43
https://doi.org/10.1016/j.tra.2022.01.002 44
45

2
Abstract 1
2
Despite ever-increasing public and financial support for public transportation in recent decades, transit 3
ridership has dropped substantially in the 2010s across the United States. While many theories that 4
explain the reasons behind the decline in ridership have been proposed, no consensus on the primary 5
causes has been reached. By employing gradient boosting decision tree and counter-factual simulations, 6
we examine the complex relationships between transit ridership and key internal and external factors from 7
2002 to 2017 in 85 of the largest urbanized areas in the US. Among several contributing factors, the 8
declining cost of driving, measured by the decreasing share of carless households combined with lower 9
gasoline prices, was the salient, most influential factor behind the recent decline. Neighborhood changes 10
in high-density neighborhoods also led to a moderate net loss in ridership. The only factor that has been 11
mitigating further decline was an increase in transit services. Had this increase not occurred, loss of 12
ridership would have been more than double the actual loss during the study period. 13
14
Keywords 15
16
public transit ridership, Uber, neighborhood change, peak Millennials, car ownership, machine learning, 17
gradient boosting decision tree 18
19

3
1. Introduction 1
Public transit is experiencing a major disruption in both operation and demand due to the COVID-19 2
pandemic since March, 2020 (Liu et al., 2020). However, transit demand was already declining 3
substantially many years before the pandemic. Ridership by bus in the United States (U.S.) dropped by 4
16.8% from 2008 to 2017 (for 85 Urban Areas excluding New York) and rail ridership by 5.28% from 5
2014 to 2017 (for 19 Urban Areas excluding New York) (FTA, 2020), which appears to be inconsistent 6
with the ever-increasing public and financial support for public transportation in recent decades. The 7
decline in transit ridership in the 2010s is particularly concerning to many transit agencies and planners 8
because it came right after a decade of strong ridership growth in the 2000s and the decline is ubiquitous 9
across small and large cities, and across transit modes. In response, a number of transportation researchers 10
have attempted to identify the primary causes of declining ridership, ranging from transit service related 11
elements such as bus service cut (Boisjoly et al., 2018) and fare increase (Mallett, 2020) to 12
macroeconomic factors, including economic recovery since the Great Recession and lower gasoline 13
prices. Some others note the importance of emerging mobility services such as ride-hailing and bike-14
sharing, yet it is inconclusive whether these services substitute or complement transit (Malalgoda & Lim, 15
2019; Shaheen & Cohen, 2020). Shrinking young adult population after the peak Millennial (Myers, 16
2016) and neighborhood changes/gentrification in transit rich inner city neighborhoods have also been 17
blamed for eroding the customer base of public transit (Bereitschaft, 2020; Berrebi and Watkins, 2020; 18
Wasserman et al, 2020). Finally yet importantly, it has been reported that the share of no vehicle 19
households especially among minority populations is decreasing in recent years (Manville et al., 2018; 20
Blumenberg et al., 2020). 21
22
Little or no consensus, however, has been reached on the statistical and substantive significance of these 23
various factors largely because most of the previous studies tended to narrowly focus on particular 24
determinants of transit ridership or on the context of certain metropolitan areas. Understanding the 25
relative contribution of each of these internal and external factors to the recent decline in ridership is very 26
important in developing effective strategies to reinvigorate public transit systems in the post-COVID 19 27
era when the pandemic will not pose a serious health risk anymore. Therefore, this study sheds light on 28
the following unsettled questions. How have primary internal and external determinants of transit 29
ridership changed from 2002 to 2017? To what extent did each factor contribute to the decline in ridership 30
in the 2010s? We employ a predictive algorithm, the gradient boosting decision tree (GBDT), to capture 31
complex relationships between these factors and transit ridership, complicated by non-linearity, 32
interactions, and heterogeneity. By performing counterfactual simulations based on GBDT modeling, we 33
assess the relative contributions of various trends in society and transportation systems and identify the 34
main culprits responsible for declining transit ridership in the 2010s. We investigate a wide set of factors 35
from 2002 and 2017 in a comprehensive manner and study transit systems in large and small urban areas 36
in the U.S. that account for 84% and 89% of total bus and rail ridership, respectively (excluding New 37
York). 38
39
This paper is organized in the following way. In the next section, we review recent findings in the 40
literature and identify important research gaps related to key factors behind transit ridership. In the 41
following two sections, we introduce our main data and chosen method, and then examine complex 42
relationships via several plots that we generate via GBDT. After we discuss the implications of this work, 43
we conclude with limitations and directions for future research. 44
45
2. Literature Review 46
We begin by discussing what the literature suggests as key determinants of aggregate transit ridership in 47
general before reviewing the studies specifically looking at driving forces behind the ridership decline in 48
the early 2010s. As Taylor and his colleagues (2009) suggested, the demand for public transit is largely 49

4
shaped by both external and internal factors from the perspective of a transit agency. Internal factors that 1
the agency can control include fares, service level typically measured by vehicle revenue miles (VRM) 2
and vehicle revenue hours (VRH), network coverage, service frequency, and transit modes. Given 3
financial and other regional conditions, the policies and strategies of an agency largely determine the 4
quantity and quality of transit services. A series of publications from the Transit Cooperative Research 5
Program—TCRP Report 95, Traveler Response to Transportation System Changes—report a 6
comprehensive survey of empirical research on how transit demand responds to changes in transit pricing 7
and fares, scheduling, frequency, and coverage (NASEM, 2004a, 2004b, 2004c). In general, transit 8
ridership is inelastic with regard to transit fare and is relatively more responsive to the supply of transit 9
services (Litman, 2019). 10
11
A number of external factors that are beyond the control of a transit agency also determine transit 12
ridership. External determinants include demographics, economic conditions, and the physical forms of 13
urban areas and transportation infrastructures. For demographic characteristics, population size, income 14
(poverty), and racial/ethnic and immigrant composition affect aggregate transit ridership. In particular, 15
college students’ share of the population is a strong determinant (Taylor et al., 2009). In addition to these 16
conventional demographic factors, recent studies indicate the possibility of a generational shift in 17
mobility-related preferences. The downturn in driving and the growth of transit ridership in the early 18
2000s were largely attributed to the large number of Millennials and their lifestyles (McDonald, 2015). 19
The new generation of young adults shows a preference for central locations with urban amenities (Y. Lee 20
et al., 2019) and adopted multimodality more often than previous generations (Vij et al., 2017). 21
22
Economic factors are multifaceted. While more robust economic activity involves more trips in general, 23
higher income can lead to a smaller mode share by transit, especially bus transit. In particular, car 24
ownership, which is largely associated with income, is a powerful predictor of mode choice on both 25
individual and aggregate levels. The price of driving, a main substitute for transit, also influences transit 26
demand. For example, growing transit ridership in the mid-2000s and early 2010s was attributed to rising 27
gasoline prices (B. Lee & Lee, 2013). Finally, a large body of literature shows that urban areas with more 28
compact urban form enjoy high transit ridership (Ewing & Cervero, 2010; Guerra & Cervero, 2011). 29
30
Which factors are more important, internal or external? Empirical studies have shown mixed results. 31
While one study that accounted for the endogeneity of transit supply found that external factors beyond 32
the reach of an agency matter more (Taylor et al., 2009), another study that did not control for the 33
simultaneous relationship between transit supply and consumption found the opposite (Alam et al., 2018). 34
Other studies compare the elasticities of transit ridership with regard to various factors (Holmgren, 2007; 35
Litman, 2019; NASEM, 2004a, 2004b, 2004c). Although these studies do not offer definitive reasons 36
behind the recent ridership decline, they provide a guide to the variables that we should investigate. We 37
should also note that the factors explaining the variation in ridership in a cross-section of urban areas do 38
not necessarily coincide with those that cause changes in ridership over time. 39
40
Turning to the studies that seek to determine the main culprits of the recent ridership decline, we review 41
four themes found in previous studies. The first group of studies accuse internal factors. A study of 42
ridership from 25 transit agencies in the U.S. and Canada from 2002 to 2015 suggested that reduced 43
(especially bus) transit service was likely the main cause of declining ridership (Boisjoly et al., 2018). It 44
pinpointed the possible shift of resources away from bus towards rail operations despite the argument that 45
spending on bus and rail services is not a zero-sum game in the long term (Levine, 2013). Ederer and his 46
colleagues (2019) showed that the links between transit services and ridership may differ across modes 47
and metropolitan areas. While changes in ridership for dedicated right-of-way transit services are strongly 48
related to changes in VRMs, the correlation among mixed right-of-way services (including bus) is 49
significant only in small- and medium-sized metropolitan areas. The studies attributing declines in 50
ridership to cuts in transit service, however, do not address the possibility that transit agencies might have 51

5
reduced bus operations in response to shrinking (bus) transit demand. Moreover, as we will discuss, 1
transit service increased for both bus and rail in the mid-2010s in our sample urbanized areas (UAs). 2
Another significant internal factor that potentially deterred transit ridership was the increase in transit 3
fares, which had been faster than inflation (Mallett, 2020). 4
5
Second, several researchers have noted recent trends in the access to cars. After the prematurely lauded 6
“peak car” in the 2000s (Goodwin & Van Dender, 2013; Zhong & Lee, 2017), vehicle ownership in the 7
U.S. continued to increase. Even more concerning is that in some regions, the reduction in the share of 8
carless households from 2000 to 2015 was more pronounced among those who traditionally use public 9
transit more often; for example, Southern California saw a 30% reduction among all households, 42% 10
among foreign-born, and 66% among foreign-born from Mexico (Manville et al., 2018). Following the 11
same vein, the San Francisco Bay Area (SFBA) underwent a 54% drop in no-vehicle households among 12
Hispanic immigrants from 2000 to 2017 (Blumenberg et al., 2020). Note that the costs of owning and 13
driving private vehicles have decreased in recent years (Bureau of Transportation Statistics, 2019), which 14
may have allowed former carless households to gain access to private vehicles. Nevertheless, to our best 15
knowledge, we are not aware of any studies that examine the role of car ownership on the recent nation-16
wide decline in transit use. 17
18
Third, researchers point to the continued expansion of emerging mobility services as a key factor eroding 19
the customer base of public transit. Whether these services, however, substitute or complement buses and 20
rail is as of yet inconclusive (Malalgoda & Lim, 2019; Shaheen & Cohen, 2020), suggesting substantial 21
heterogeneity across modes, locations, and/or user groups. For instance, ride-hailing appears to provide 22
first- and last-mile access to long-distance trips by rail, but it tends to replace some bus trips (Clewlow & 23
Mishra, 2017). Similarly, another study presents net complementarity for small transit agencies in large 24
cities but net substitution for other cases (Hall et al., 2018). The substantial gaps in per-ride fares between 25
transit and ride-hailing especially in US cities suggests a possibility that many of ride-hailing patrons are 26
coming from outside of the regular transit rider pool. Nevertheless, a recent TCRP Report 231 finds that 27
ride-hailing is the largest contributor to the decline in bus ridership across 209 metropolitan areas, causing 28
10 to 14% decreases on the net (Watkins et al., 2021). Further research on the extent of its effect 29
compared to other major internal and external factors is guaranteed. 30
31
Finally, several studies have suggested possible connections between neighborhood change and ridership 32
decline. A number of studies have noted gentrification or an upward shift in the socioeconomic status of 33
residents in inner city neighborhoods of large US metropolitan areas since 2000 (Baum-Snow & Hartley, 34
2020; Couture & Handbury, 2020). Low-income groups with limited/no access to cars tend to live in 35
central city neighborhoods that have decent transit services (Glaeser et al., 2008). Thus, to the extent that 36
gentrification leads to the displacement of existing residents that frequently use public transit and these 37
transit dependent populations relocate to areas with limited/poor access to transit, agencies will lose their 38
customer base. In Southern California, Paul and Taylor (2021) have reported that the spatial match 39
between “high-propensity” residents and neighborhoods has worsened since 2000. Bereitschaft (2020) 40
found that Urban Cores (neighborhoods within 3 miles from the center of the central city in large 41
metropolitan areas) with increased shares of whites and young adults aged 18-39 from 2000 to 2015 42
experienced a further drop in the transit commuter share while enjoying a rise in active/tele-commuter 43
shares. Berrebi and Watkins (2020) also reported that bus stops in census tracts with larger shares of 44
white, college-educated, or carless residents in Miami and Atlanta lost more boarding and alighting than 45
those with smaller shares. In short, studies have reported noticeable socioeconomic shifts in transit-rich 46
neighborhoods in the central city (Blumenberg et al., 2020; Manville et al., 2018; Wasserman et al., 47
2020); nonetheless, we still lack studies investigating their contributions to the recent decline in ridership. 48
49
A broader demographic shift in the central city has implications for transit use as well. While researchers 50
have documented an increasing number of young adults in the central city since 2000 (H. Lee, 2020; Y. 51

6
Lee et al., 2019; Myers, 2016), they do not agree on whether this trend will continue in the near future. 1
Myers (2016) predicted that the concentration of young adults in cities in the early 2010s would soon 2
diminish as the cohort size each birth year continued to shrink after its peak in 1990 (i.e., age 25 as of 3
2015), jobs in the suburbs would recover, and entry-level houses would become more available in the 4
suburbs than in cities after the Great Recession. Given the strong association between college/graduate 5
students and transit use, their declining number in the central city implies a loss in ridership. By 6
comparison, other studies have suggested that not as many young adults will move away from the central 7
city as did preceding generations at the same age/life stage (H. Lee, 2020; Y. Lee et al., 2019). In 8
addition, Lee (2021) has found that the youngest generation, Gen Z, are filling up the space that 9
Millennials have left. The literature, nevertheless, does not include studies exploring the role of 10
demographic shifts in recent ridership trends. 11
12
3. Research Design 13
3.1. Data and variables 14
To examine the complex relationships between transit ridership and various factors, we have compiled 15
relevant variables from diverse data sources. In so doing, we include all relevant variables generally 16
known to be the determinants of transit ridership and additional variables that are suggested to have 17
caused the recent ridership decline as reviewed in the previous section. Table 1 presents the summary 18
statistics of all variables tested in the modeling stage along with their sources (not all of them are included 19
in the final specification). Note that our unit of analysis in the final sample is the annual summary, either 20
the sum or the average of several variables, at the urban area level. 21
22
We retrieve transit ridership, supply, and related attributes from the National Transit Database (NTD), 23
maintained by Federal Transit Administration (FTA). Transit agencies in the U.S. report data on financial, 24
operating, and asset conditions to the FTA, which reviews the data and allocates funding accordingly. 25
Each record in the NTD represents a combination of an agency, a mode, and a type of service either 26
directly operated (DO) or purchased transportation (PT). Note that in part, because of changes in the 27
reporting requirements and unique challenges that individual agencies face, not all reported data are 28
complete or reliable. We start with those agencies in the largest 150 UAs, according to the 2010 U.S. 29
Decennial Census. We select eight fixed route transit modes—commuter rail, heavy rail, light rail, street 30
car, commuter bus, bus, bus rapid transit, and trolley bus—and group them as bus or rail, including 31
records with both DO and PT. Next, because of the possible reallocation of resources and services across 32
agencies within the same UA (e.g., closing, opening, rebranding, and merger & acquisition of agencies), 33
we aggregate records at the UA level instead of the agency level. In so doing, we also aggregate records 34
at the annual level because most explanatory variables are available only at this level. After testing 35
aggregated and separate trends, we exclude New York because of its unique, distinctive nature, which 36
differs from that of other UAs in terms of the rising and falling timing and pace (refer to Figure 1). To 37
ensure data quality, we manually inspect individual UA data (i.e., their annual sums), identify possible 38
reporting errors, and remove UAs with one or more years of reporting errors. The final dataset used in this 39
study includes 16 annual observations (2002 – 2017) of bus ridership in 85 UAs and rail in 19 UAs (both 40
excluding New York). As of 2017, these UAs accounted for 84% of total bus ridership and 89% of rail 41
ridership in the US (excluding New York), according to the National Transit Database. 42
43
To assess changes in transit ridership, we use annual unlinked passenger trips (UPT). We check and 44
remove records with possible reporting errors for which, for example, the fare revenue is non-zero, but 45
UPT is zero, or UPT differs substantially from its preceding or following years (except New Orleans in 46
2005). After filtering out UAs with suspicious reporting errors, we log-transform UPT to minimize the 47
effects of extreme right skewness of its distribution. To examine the effects of transit supply, after testing 48
a few candidate measures such as subsidies, vehicle revenue hours, and maximum service, we choose 49

7
vehicle revenue miles (VRM). As another key internal factor, we compute fare per ride by dividing the 1
total fare revenue by UPT in the same year. 2
3
Figure 1. Public transit ridership trends 2002-2017 (Left: Bus, Right: Rail). 4
Notes: We include the annual unlinked passenger trips reported by transit agencies without apparent reporting errors in the 5
National Transit Database (NTD). 6
7
We collect several UA level attributes critical to travel mode choice from various sources. We gather 8
annual gasoline prices from the Bureau of Labor Statistics and adjust for inflation, using the Consumer 9
Price Index for each UA. From the same source, we also obtain the annual unemployment rate of the 10
primary city of each UA. The Federal Highway Administration’s (FHWA) Highway Statistics reports the 11
total road miles per 1,000 residents in UAs each year, which we use as a proxy for the amount of 12
infrastructure and political support for driving. As for the number of years since the launching of ride-13
hailing services in individual UAs, we borrow the Uber entry data compiled by Hall et al. (2018). In the 14
absence of service quantity data such as ride-hailing ridership and the number of drivers or vehicles by 15
city, we use the number of years since service launching as a proxy for the market maturity. Thus, the 16
result for this variable needs to be interpreted with caution. Urban form has been well documented as an 17
important factor for transit use, so we prepare two UA-level variables by processing data from the two 18
Decennial U.S. Censuses (2000 and 2010) and the American Community Survey (ACS) five-year 19
estimates (their ending years from 2009 to 2019). One is total population within the 2010 UA boundary, 20
and the other is the population-weighted density. Population-weighted density is known to account for 21
travel behavior better than the oft-used conventional density measure (S. Lee & Lee, 2020). Since both 22
measures are highly skewed to the right, we log-transform them. 23
24
Socioeconomic and demographic variables come from the same U.S. Census data products mentioned 25
above. In so doing, we prepare and test nine variables, for both the entire UA and the top 10% densest 26
census tracts (boundary matched to the 2010 US Census). We use the latter as a proxy for central city 27
transit-rich neighborhoods, many of which have allegedly undergone neighborhood changes in terms of 28
residents’ socioeconomic status in recent years (Baum-Snow & Hartley, 2020; Hwang & Lin, 2016; 29
Landis, 2016). The nine variables include the percentage of non-Hispanic whites, the percentage of 30
Hispanics, the percentage of foreign-born, the percentage of young adults aged 18 to 34, the percentage of 31
college-educated adults 25 years and older, the percentage of college/graduate students 18 years and 32

8
older, the percentage of low-income households (i.e., under 25k/year at each Census/ACS year), the 1
percentage of those under poverty, and the percentage of no-vehicle households. Since the U.S. Census 2
does not provide estimates from 2002 to 2006, we linearly interpolate them with the 2000 decennial 3
census and the 2005-2009 ACS 5-year estimates, the earliest available. Last but not least, using the U.S. 4
Census ACS Public Use Microdata Sample (PUMS) data, we also compute two additional UA-wide 5
vehicle-access measures: the percentage of carless households among poor households and among young 6
households. 7
8
Table 1. Summary statistics. 9
Bus (N=1,360 from 85 UAs)
Rail (N=304 from 19 UAs)
Variable
Mean
Std. Dev
Min
Median
Max
Mean
Std. Dev
Min
Median
Max
National Transit Database
Unlinked passenger trips
39,315,353
77,868,060
239,533
12,318,157
604,412,689
79,335,013
98,717,600
340,626
20,888,099
317,924,502
Vehicle revenue miles
15,010,724
22,238,786
208,085
5,955,839
163,102,580
23,810,059
30,937,555
326,243
7,682,136
121,485,330
Fare/ride ($)1)
0.9
0.3
0.1
1.0
2.0
1.6
0.8
0.5
1.3
4.5
FHWY
Road miles/1000
4.7
1.3
1.2
4.6
10.1
3.8
1.0
2.0
3.9
7.2
BLS
Gas price2)
227.1
60.4
105.2
220.5
418.6
231.5
61.9
105.2
226.3
418.6
% unemplo yment
6.3
2.3
2.5
5.7
16.9
6.1
2.0
2.5
5.7
12.4
Hall et al. (2019)
# of years w ith Uber
0.6
1.3
0.0
0.0
5.0
0.7
1.4
0.0
0.0
5.0
US Census - Urban form
UA population
1,468,976
1,879,714
142,885
746,181
12,347,430
3,395,639
2,761,729
832,635
2,243,090
12,347,430
UA weighted density
1,819
896
491
1,644
6,225
2,697
1,244
978
2,132
6,225
US Census - Socioeconomics/demographics at the UA level
% non-Hispanic white
61.8
16.4
11.0
65.3
94.8
58.6
16.0
26.8
62.9
85.2
% Hispanic origin
16.0
16.1
1.0
9.4
85.2
16.1
11.9
1.0
13.3
47.5
% foreign born
12.2
8.1
2.2
9.5
41.1
18.0
10.7
3.5
15.5
41.1
% 18-34
25.3
3.0
15.4
25.0
37.7
24.3
1.8
20.1
24.6
29.2
% student s
8.6
2.7
3.4
8.0
25.0
8.0
1.0
5.8
7.9
10.1
% college graduate
60.6
7.1
42.4
60.7
80.4
63.4
5.9
51.2
63.5
74.5
% under $25k/year
23.8
5.2
10.0
23.6
40.7
20.5
4.7
10.0
20.6
33.3
% under poverty
14.7
3.4
7.7
14.5
27.7
12.4
2.3
7.7
12.3
17.6
% carless house holds
8.9
2.6
4.2
8.6
17.2
10.4
3.2
5.2
9.6
16.2
% carless among poor2)
21.5
6.5
7.9
20.2
42.5
25.7
7.0
12.9
24.4
39.3
% carless among
young
2)
6.9
2.8
1.1
6.4
16.2
8.9
3.3
2.6
8.1
16.2
US Census - Socioeconomics/demographics among the top 10% densest tracts
% non-Hispanic white
42.1
20.6
3.8
41.0
88.7
39.3
18.1
10.6
37.0
74.9
% Hispanic origin
25.5
23.1
0.8
18.3
92.6
27.0
20.5
1.9
20.0
78.0
% foreign born
17.0
12.5
0.2
13.8
65.1
24.3
15.5
4.2
22.7
65.1
% 18-34
35.4
11.1
19.7
32.2
82.8
34.4
5.8
22.7
34.5
46.1
% student s
14.3
12.6
2.6
9.4
75.9
11.7
4.2
5.2
10.8
24.5
% college graduate
50.5
13.0
21.7
49.1
88.6
56.4
12.7
29.0
57.4
85.6
% under $25k/year
38.9
8.9
12.3
39.2
63.2
33.4
8.4
16.7
32.2
53.2
% under poverty
29.0
8.0
7.2
28.4
54.2
23.9
6.1
13.8
23.3
40.9
% carless house holds
19.8
8.7
5.8
17.7
46.1
25.6
10.4
8.4
24.3
45.5
Notes: 1) Fare/ride is adjusted to real dollar in 2019. 2) For the gas price, we use its consumer price index, which accounts for 10
inflation and whose value for 1982-84 is set to 100. 11
12
3.2. Methods 13
14
We employ gradient boosting decision tree (GBDT), a predictive algorithm whose high performance has 15
been well examined in the machine learning literature (J. Friedman et al., 2000; J. H. Friedman, 2001, 16
2002). While machine learning methods are increasingly used in recent transportation research, most 17
studies have been applying the methods to predict transit ridership at station (Ding et al., 2016; Ding et 18
al., 2019; Shao et al., 2020) or route level (Egu and Bonnel, 2021). As summarized in Table 2, most 19
existing studies on aggregate transit ridership, at either MSA/UA or transit agency level, use traditional 20
inferential statistics, including least square regression and panel data models. However, since one of our 21
main goals is to perform counterfactual simulations as described below, GBDT with a high predictive 22
power is better suited for this study than traditional methods. 23

9
1
The machine learning method is more flexible at capturing complex relationships among variables, non-2
linearity, interactions, and heterogeneity (Ding et al., 2019; Shao et al., 2020). A number of recent 3
changes that have occurred in both society and transportation systems interact; that is, they have not 4
occurred in isolation, and the interactions have evolved over time. For example, the growth of ride-hailing 5
services, shifts in the lifestyle and mobility choices of young generations, and technological changes are 6
all interconnected, implying complex relationships between various determining factors and the outcome 7
of our interest (transit ridership). While researchers need to specify which functional forms such complex 8
relationships take a priori in conventional modeling methods, the machine learning algorithm, as a non-9
parametric modeling approach, allows data to reveal the underlying patterns of the complex relationships. 10
In addition, machine learning approaches with good predictive accuracy better serve our research goal. 11
While many previous studies have attempted to establish causality between various determinants and 12
transit ridership through econometric techniques (Diao et al., 2021; Hall et al., 2018; B. Lee & Lee, 2013; 13
Taylor et al., 2009), we aim to achieve a balance between reasonable conceptualization of complex 14
relationships and the predictive accuracy of chosen models. 15
16
Table 2. Methods to predict aggregate transit ridership used in previous studies. 17
18
Study
Study areas
Methods
Key findings
Taylor et al.
(2009)
265 US urbanized
areas
2-stage least
square (SLS)
regression
Significant impacts of internal factors on ridership,
but larger impacts by external factors
Lee & Lee
(2014)
67 US urbanized
areas
2-SLS random-
effects panel
model
Impacts of gasoline prices on ridership are larger in
more compact urban areas
Alam et al.
(2018)
358 MSAs in US
Log-log ordinary
least square
(OLS)
Internal factors are more significant predictors of bus
ridership than external factors
Boisjoly et al.
(2018)
25 transit agencies
in U.S. and
Canada
Longitudinal
multilevel mixed-
effect regression
Reduction in bus vehicle revenue kilometers (VRK) is
the main cause of ridership decline
Malalgoda &
Lim (2019)
Top 50 US transit
agencies
Data envelopment
analysis and OLS
Ride-hailing increased rail ridership; neither a
supplement nor a substitute of bus
Ederer et al.
(2019)
27 MSAs in US
Correlation
analysis by MSA
cluster
Changes in population and service levels partly
explain ridership change only for transit in dedicated
right of way
Graehler et
al. (2019)
22 US agencies,
monthly ridership
Random-effects
panel data model
Service levels, gas prices, and auto ownership are
important but not sufficient reasons for ridership
decline
Watkins et al.
(2021)
209 MSAs in US
Fixed-effects
panel model
Ride-hailing service is the largest contributor to the
ridership decline between 2012 and 2018
19
Gradient boosting significantly improves prediction accuracy while retaining unique merits as an off-the-20
shelf method (Hastie et al., 2009). In general, boosting refers to a procedure that combines outputs from 21
many simple, but low-accuracy models and makes highly accurate predictions. When applied to decision 22
tree, boosting sequentially fits a simple tree (e.g., a stump) to given data in a way that the following trees 23
identify and reduce errors from the prediction of a combination of all preceding trees. In this context, 24
gradient boosting builds a new tree in the direction that reduces prediction errors the most until the last 25
iteration, so the iterative process becomes more efficient. 26
27
As a predictive algorithm, GBDT has weaknesses, most notably over-fitting and the lack of a formal 28
procedure for the test of significance, so we employ the following approaches to overcome them. First, 29

10
we vary meta-parameters to avoid over-fitting, which helps us find the global optimum. Specifically, we 1
perform a grid search for four meta-parameters, including the size of trees, a learning rate, a subsampling 2
rate, and the minimum number of cases at each terminal node. Hastie, et al. (2009) recommended 3
choosing a very small learning rate and stopping early, which helps avoid overfitting. Accordingly, if we 4
see no further improvement after ten consecutive iterations, we stop the iterative process. Second, we 5
refer to regression model results when selecting predictors in the final model. Note that achieving the 6
highest prediction accuracy is not the main goal of this study. Instead, for interpretation and a scenario 7
analysis, we need to select appropriate predictors among similar alternatives (e.g., % carless households 8
among all households in an UA vs. those only among young/poor households), and avoid potential 9
multicollinearity. Thus, starting with a list of predictors suggested in the literature, we let inferential 10
statistics to guide the process of predictor selection. 11
12
Using various charts from the gradient boosting results, we uncover complex relationships between transit 13
ridership and various factors for the entire study period, 2002-2017. Relative importance plots present the 14
extent to which each predictor contributes to a reduction in prediction errors. Partial dependence plots 15
(PDP) display modeled relationships between the outcome and a predictor while accounting for the other 16
predictors. By creating the individual conditional expectation (ICE), a variant of PDP, we not only 17
examine non-linear relationships but also present heterogeneity across cases. In addition, multi-predictor 18
PDP visualizes interactions between predictors, which indicate the presence and strength of synergy or 19
cancellation effects among the predictors. 20
21
Most importantly, we perform a scenario analysis to determine the extent to which individual factors have 22
contributed to the recent decline in transit ridership. In so doing, we take a two-step approach. First, we 23
develop hypothetical scenarios in recent years, in which transit ridership started to fall (e.g., “What if 24
Uber/Lyft had not been available?” or “What if vehicle ownership had remained the same?”). Second, via 25
gradient boosting, we predict transit ridership for these years under the scenarios and examine alternative 26
ridership trends under each of the counterfactual scenarios. 27
28
4. Results 29
30
4.1. Trends of key variables 31
Before we fit gradient boosting to the data, we examine the trends of key predictors for bus and rail 32
ridership from 2002 to 2017, and we compute the annual averages separately for the UAs for bus 33
ridership (n=85) and for rail ridership (n=19). Each of Figures 2 and 3 presents the eight predictors, 34
whose trends we discuss briefly in this section. 35
36

11
Figure 2. Trends in key predictors of bus ridership in the UAs, 2002-2017 (N=85). 1
Notes: Smoothing lines are drawn via LOESS (Locally Weighted Scatterplot Smoothing) with a 95% confidence interval. Units 2
for the vertical axis are million miles for vehicle revenue miles, real dollars in 2019 for the fares per ride, the consumer price 3
index for gasoline prices (adjusted for individual UAs and years), number of years of Uber operation, and percentage of the four 4
variables: ages 18-34, those under poverty in the top 10% densest tracts, carless households, and carless young (age 18-34) 5
households. 6
7
4.1.1. Bus 8
The annual bus ridership of the 85 Urban Areas (except New York) marked the highest in 2008, mainly 9
because of the great recession, and since then it has sharply decreased with short-lived rebound in 2011 10
and 2012 (see Figure 1). Related to this, Vehicle Revenue Miles (VRM, a proxy for transit supply) also 11
peaked in 2008 and substantially decreased until 2012, but it has gradually increased since then. Thus, 12
reduction in supply does not seem to be a factor for the recent decline in bus ridership. The fare per ride
13
has constantly risen since the early 2000s, which may have negatively affected bus ridership. Inflation-14
adjusted gasoline prices peaked in 2012 and kept declining since then, making driving more affordable
15
for low-income households. In the U.S., Uber has continuously entered new cities and has been providing
16
ride-hailing services since they started services in NY in August 2012. By 2017, all UAs in the sample 17
have Uber in operation with the average duration about four years. Only three UAs in our sample had 18
remained unserved until 2016. 19
20
Among those who take transit frequently are young adults, whose share in our sample UAs peaked around 21
2013 and 2014 and started to keep decreasing thereafter. In the aftermath of the Great Recession, the UA-22
level poverty rate peaked in 2012 (not shown above for brevity), and has been decreasing ever since.
23
More economic/financial resources likely allow households to have better access to private vehicles and 24
resort less to public transit. What is more interesting is the reduction in the poverty rate is greater in the 25
top 10% densest neighborhoods than urban area wide: 32-27=5 percentage point vs. 17-14=3 percentage 26
point at the UA level. While the overall poverty reduction is largely due to economic recovery, the faster 27
reduction in dense neighborhoods with better transit access seems to be a result of the inflow of young 28
professionals and neighborhood changes. 29
30
The share of households without access to private vehicles started to increase from 2007; however, that 31
trend got reversed in 2012 continued to fall, in part because the cost of car ownership and operation 32
became cheaper after the recession (Bureau of Transportation Statistics, 2019). While the share of those 33
without access to private vehicles among young households (i.e., household heads below 35) shows the
34
same trend of a temporary increase from 2006 to 2012 followed by constant decline, no-vehicle 35
households among the poor have continued to decrease in its share for the entire study period (not shown 36
above for brevity), 37
38
4.1.2. Rail 39
Rail transit ridership of the 19 Urban Areas (except New York) peaked in 2014, and since then it 40
continued to lose its customer base (see Figure 1). While key predictors in those UAs with rail show 41

12
similar trends to those in the bus UA sample, there are a few notable differences. Unlike bus service, rail 1
VRM continued to increase for the entire period from 2002 with just a flatter slope during the economic 2
recession. Again, service cut is not likely to be a reason for declining ridership. Overall, poverty rate in 3
top 10% densest neighborhoods is lower and percent carless households is higher in the rail sample than 4
in bus sample UAs. Better transit services in rail UAs, which tend to be larger than UAs with only bus 5
service, seem to enable residents to choose car-light or carless lifestyles, either by choice or by necessity. 6
7
Figure 3. Trends in key predictors of rail ridership in the UAs, 2002-2017 (N=19). 8
Notes: Smoothing lines are drawn via LOESS (Locally Weighted Scatterplot Smoothing) with 95% confidence interval. Units for 9
the vertical axis are million miles for vehicle revenue miles, real dollar in 2019 for the fares per ride, the consumer price index 10
for gasoline prices (adjusted for individual UAs and years), number of years for Uber duration, and percentage for the four 11
variables: age 18-34, those under poverty in top 10% densest tracts, carless households, and carless young households. 12
13
14
4.2. Relative importance of key determinants 15
16
Now, we present gradient boosting results. We fit gradient boosting to a 70% random subsample of the 17
data and check its performance for the unused 30% of the data, separately for buses and rail (N=1,360 and 18
304). In so doing, we search for the best specification by comparing gradient boosting with fixed-effect 19
panel regression results on the same samples. 20
21
4.2.1. Bus 22
The top two panels in Figure 4 present two importance plots for bus ridership. The values in these plots 23
are computed based on the amount of prediction errors reduced by each factor. The left plot shows the ten 24
most important factors and the right the top ten factors excluding the variables that vary greatly across 25
UAs, such as VRM and density or total population. The dominance of these three variables in the left 26
column plots is likely to be because of cross-sectional variations, but it is possible that they are not the 27
main causes of ridership decline over time. It should also be noted that we include the three variables 28
back in the simulation analysis presented in the next section. Not surprisingly, VRM is the most important 29
factor, reducing more than 80% of prediction errors alone, followed by weighted population density. Both 30
vary greatly cross-sectionally, but not longitudinally. In the top right panel, excluding the two 31

13
variables, % carless young households reduces the largest amount of prediction errors, followed by % 1
foreign-born, % students in dense tracts, % UA-wide young adults, and another vehicle-access variable. 2
Interestingly, the presence/maturity of ride-hailing services account for only a very small portion of the 3
variation in bus ridership. 4
Figure 4. Importance plots for bus (top) and rail (bottom) ridership. 5
Notes: Plots in the left column present the top 10 most prediction-error-reducing variables, and plots on the right column present 6
such variables excluding VRM, population, and density that vary mostly across urban areas, but not much longitudinally. Note 7
that we still include these three variables in the following analysis. 8
9
4.2.2. Rail 10
In the case of rail ridership on the bottom row of Figure 4, a slightly different set of factors stand out. 11
When excluding variables with mostly cross-sectional variations (again, omitted from the bottom right 12
plot, but not from its underlying model), we find that transit fare is the most important, followed by 13
variables measuring vehicle access, neighborhood change, and UA-level demographic changes. Again, 14
Uber does not appear to contribute much to the reduction in prediction errors. 15

14
1
Although importance plots show the factors that generally matter in determining transit ridership, they do 2
not offer direct insights into the reasons for the sharp decline in bus ridership since 2012 and rail ridership 3
since 2014 because importance values and ranks are determined based on their aggregated performance 4
throughout the entire study period. Therefore, we employ another approach to identify the factors 5
responsible for the decline in ridership in later years. 6
7
4.3. Relationships between transit ridership and individual variables 8
9
4.3.1. Bus 10
Figure 5 presents ICE plots for selected predictors of bus ridership. In the figure, the black lines trace 11
predicted log-transformed UPT, lnUPT, for given observations as the predictor of interest changes from 12
the minimum to the maximum in the data, and the red line shows the average trajectory of all cases in 13
each panel. Note that all observations are adjusted to start at the origin, and values on the vertical axis 14
indicate the percentages of changes from the prediction with the minimum value (e.g., 0.1 indicates a 10% 15
change). 16
The presence of ride-hailing services appears to reduce bus ridership by an average of about one percent 17
(i.e., -0.01) after five years of operation. One major difference from conventional regression is that 18
gradient boosting does not assume a constant global coefficient. Instead, it allows the effects of ride-19
hailing to vary by the length of operation. It suggests that ride-hailing does not make much difference in 20
early years, but it does have noticeable (but still small) effects as the market matures. The relationships 21
between bus ridership and other determinants are non-linear as well. The unemployment rate presents a
22
nuanced relationship with bus ridership, with a slightly positive association until the 7-8% levels, but a 23
negative association at higher levels, likely because of a reduction in transit commuters. The share of 24
those born outside the U.S. supports bus ridership but only to about the 15% level. After this threshold, it
25
negatively affects bus ridership until it reaches 30%. Not surprisingly, a concentration of young adults
26
between 18 and 34 significantly increases bus ridership. 27
28
29

15
Figure 5. Individual conditional expectation (ICE) of selected predictors of bus ridership, 2002-2017. 1
Notes: Black lines trace predicted lnUPT for given observations as the predictor of interest changes from the minimum to the 2
maximum value. The red line shows the average trajectory of all cases in each panel. The vertical axis indicates percent changes 3
from the prediction with the minimum value (e.g., 0.1 indicates a 10% change). 4
5
As expected, the percentage under poverty in the densest tracts of the UA is positively associated with 6
bus ridership. The consistent uphill slopes suggest that low-income people in central city neighborhoods 7
with better transit access rely on transit on a regular basis. Thus, decreasing poverty rates in high-density 8
areas since the early 2010s (see Figure 2) is likely to account for some of the recent ridership decline. 9
Similarly, the shares of carless households and young carless households show strong positive
10
correlations with bus ridership. Interestingly, bus ridership exhibits steep slopes within certain ranges of 11
these vehicle-access variables—from 7% to 10% among carless household UA-wide and from 4% to 8% 12
among carless young households. We suspect that increasing car ownership is another important factor 13
responsible for the recent ridership decline. 14
15
Last, but most importantly, the time trend presents a concerning picture. It captures the remaining effects 16
not accounted for by the predictors controlled for in the model. The average effect stayed in the positive 17
territory until 2008 (i.e., time trend = 6); then it started to drop and suddenly plummeted below zero from 18
2015 to 2017, ending 7% lower than it was in 2002. We can only speculate significant changes in some 19
other factors such as attitudes and preferences and/or in the way these factors interact with observed 20
predictors in recent years. Such conjecture would require further investigation that determines the specific 21
mechanisms triggering such changes. 22
23
Next, we check to see if a particular pair of predictors present “strong” interaction effects in boosting bus 24
ridership by analyzing two-way partial dependence plots (PDPs). In Figure 6, the two-way PDPs show the 25
highest predicted lnUPT, coded in blue, and the lowest predicted lnUPT, coded in red; the other values 26
between the two are color graded. Figure 6 presents four pairs of factors that show relatively strong 27
interactions. The top left panel indicates that without a sufficient share of carless households (who are 28
likely transit-dependent) in a UA, high gasoline prices do not automatically lead to more bus trips. 29
Maximum effects are found when both the price of gasoline and the share of carless households are high. 30
Likewise, the top right panel reveals similar interactions between gasoline prices and vehicle access 31
among young households. Not surprisingly, as shown in the bottom left panel, the two vehicle access 32
variables present strong interaction effects. Finally, the bottom right panel suggests the potential impact of 33
neighborhood change on bus ridership. When dense inner-city neighborhoods lose poor households or 34
students, an UA experiences a fall in bus trips. If the shares of both groups decline in dense 35
neighborhoods, their effects would become more consequential. 36
37
38

16
1
Figure 6. Two-Way Partial Dependence Plots for Bus Ridership. 2
Notes: Not to mislead readers via extrapolation, we present only areas with actual data points; therefore, each panel does not 3
cover the entire XY plane. 4
5
4.3.2. Rail 6
Compared to the bus ridership plots, the ICE plots of the rail model exhibit more fluctuating lines with 7
considerable noise that may be due to the small sample size (training set n=212, 70% of N=304). Unlike 8
bus ridership, which shows small net substitution effects in the medium run, rail ridership appears to be 9
barely affected by ride-hailing services. The average positive effect for five years in operation is only
10
about 0.4%. The results are consistent with a previous study showing that ride-hailing complements 11
heavy rail service but competes with bus service (Clewlow & Mishra, 2017). Note that bus and rail likely 12
have differing customer bases, mainly serving different locations and populations; hence, the ways in 13

17
which ride-hailing affects the two modes may also differ. The pattern of the poverty rate for the rail plots 1
also differs strongly from that of the bus plot. The poor-household share in high-density neighborhoods is
2
negatively associated with rail ridership when it exceeds about 20%. It is likely that socio-economic 3
changes in transit-rich neighborhoods in large U.S. cities in recent years have led to a gain in rail ridership 4
at the cost of bus ridership. 5
6
Some other variables show patterns similar to those of the bus ridership model. The unemployment rate, 7
especially when it exceeds 6%, begins to lower rail ridership. Percentage foreign born shows an inverted
8
U-shape pattern, whereby its association with rail ridership changes from positive to negative at around 9
25%. We speculate that UAs with high shares of foreign-born populations tend to have high shares of 10
recent immigrants who occupy low-skill jobs and are more likely to take bus than rail. Unlike the case for 11
buses, the share of young adults aged 18-34, evidenced by somewhat flat slopes, does not present a clear
12
relationship with rail ridership. 13
14
15 Figure 7. Individual Conditional Expectation (ICE) of selected predictors of rail ridership, 2002-2017. 16
Notes: Black lines trace predicted lnUPT for given observations as the predictor of interest changes from the minimum to the 17
maximum value. The red line shows the average trajectory of all cases in each panel. The vertical axis indicates percent changes 18
from the prediction with the minimum value (e.g., 0.1 indicates a 10% change). 19
20
The share of carless-households is relatively less significant for rail than it is for bus ridership. The UA-21
wide share of no-vehicle households shows a positive slope to about 12.5% and then it shows a 22
downward slope. We speculate that the portion of choice riders is larger among rail patrons while bus 23
riders are more likely to be transit dependent. Interestingly, rail ridership is more responsive to carless
24
young households. Note that their share has continued to decline since 2012, when the economic
25
recession ended. Thus, one would expect that increased access to private vehicles, especially among 26
young adults, either because of improved economic conditions or stronger demand for vehicles as they 27
pursue once-postponed life events (H. Lee, 2021), has resulted in declining ridership in recent years. Like 28
the trend shown in the bus model, the time trend of the rail model captures aggregate effects from 29

18
unobserved sources, presenting continuous decreasing patterns with a four-percent loss between 2002 and 1
2017. 2
3
Figure 8 presents the interaction effects of four pairs of predictors that show similar patterns, to those 4
found in the bus model, with some noise possibly due to the small sample size. Interestingly, carless 5
young households and gasoline prices present more consistent interaction effects (in the top right panel) 6
than carless households of any age and gasoline prices do (in the top left panel). In addition, we find 7
strong additive effects of concentrated students and poor households in the densest census tracts (in the 8
bottom right panel); however, the direction of their interactions is opposite to that in the bus model. 9
10
Figure 8. Two-Way Partial Dependence Plots for Rail Ridership. 11
Notes: Not to mislead readers via extrapolation, we present only those areas with actual data points; therefore, each panel does 12
not cover the entire XY plane. 13
14
15

19
4.4. Simulated ridership trends 1
We take a two-step counter-factual simulation approach to estimating the unique contributions of key 2
factors to the recent decline in transit ridership: (1) We develop scenarios in which we assume that each 3
of the factors remains at the reference year level in recent years, and (2) under these counter-factual 4
scenarios, we predict ridership via gradient boosting and compare these predictions with a baseline 5
prediction. For instance, “What would bus ridership have been like without Uber?” To answer this 6
hypothetical question, we estimate and compare transit ridership numbers, one using actual values of 7
predictor variables (i.e., the baseline prediction) and the other in which the number of years since the 8
inception of Uber is held at zero. Similarly, we examine the extent to which each key variable has 9
contributed to the recent decline in ridership by estimating the difference between the baseline prediction 10
and a simulated prediction obtained by holding the variable value of interest at the reference-year level. 11
The reference years are those in which key predictors peaked or bottomed out or when transit ridership 12
began to decline (e.g., 2012 for bus and 2014 for rail). Table 2 includes scenarios, their reference years, 13
and predicted changes since 2012 for buses and 2014 for rail, and Figure 5 plots predicted ridership under 14
the scenarios. 15
16
Table 3. Predicted ridership under various counterfactual scenarios. 17
Bus
Rail
(A)
(B)
(C)
(D)
(E)
(F)
Scenario
% change
from 2012
to 2017
% point
difference
from the
baseline
Reference
year
% change
from 2014
to 2017
% point
difference
from the
baseline
Reference
year
(1)
No-change-in-gas-
&-car
1)
0.05% 7.39% 2012 -1.20% 3.14% 2014
(2)
No-change-in-gas-
price
-2.60% 4.74% 2012 -3.12% 1.21% 2014
(3)
No-change-in-SED-
top10
3)
-6.24% 1.10%
2010 &
2012
2)
-3.49% 0.85% 2014
(4)
No-Uber-
introduction
-6.92% 0.43% 2012 -4.68% -0.35% 2012
(5)
No-change-in-fare
-6.77%
0.57%
2012
-3.73%
0.60%
2014
(6)
No-peak-Millennial4)
-7.44%
-0.10%
2014
-4.31%
0.02%
2014
(7)
Baseline prediction
-7.34%
0.00%
-
-4.33%
0.00%
-
(8)
No-change-in-
supply
-15.33% -7.99% 2012 -9.33% -5.00% 2014
Notes: 1) In this scenario, gas prices and two vehicle-access variables, % carless and % carless among young, are held constant 18
at their values in 2012. 2) % students in the densest tracts taken from 2010 and % under poverty from 2012. 3) “SED top 10” 19
refers to socioeconomic and demographic attributes only for top 10% densest tracts in each urban area. 4) No change in % 20
young adults (from 18 to 34) in the UA since the reference years. 21
22

20
1
Figure 9. Predicted Ridership under the Scenarios (bus on the left and rail on the right). 2
Notes: Gray dotted horizontal lines are the baseline prediction, or the predicted percentage change in ridership between base year 3
(2012/2014) and 2017 without any modifications to the data. 4
5
4.4.1. Bus 6
The results of the simulation suggest that the recent declines in ridership are the result of a number of 7
factors but that private vehicle-related factors are, by far, the most influential. This finding is consistent 8
with a recent transit ridership study on the Los Angeles metropolitan area (Manville et al., 2018). 9
Scenario (1) examines what bus ridership would have been if gasoline prices and % zero-vehicle 10
households (both at the UA level and among young households) had remained constant at the 2012 level 11
in each UA. The prediction under this scenario reveals that, without changes to the major costs of owning 12
and operating cars, bus ridership would have remained almost the same as that in 2012 (i.e., 0.05% in 13
Column (A) of Table 2). The substantial difference between this finding and the baseline prediction, 7.39 14
percentage points in Column (B), highlights the contribution of increasing access to car and decreasing 15
costs of driving to the recent decline in ridership. Gas prices alone appear to account for a substantial 16
portion of the gap. When holding only gas prices constant since 2012 (scenario (2)), the gap from the 17
baseline prediction is still as large as 4.74 percentage points. 18
19
Socioeconomic and demographic changes have also contributed to declines in bus ridership, but their 20
impacts appear to be much smaller. Scenario (3), no neighborhood change (i.e., no reduction in the shares 21
of college/graduate students and those under poverty) in high-density neighborhoods would have 22
sustained about a 1.10 percentage point higher bus ridership than the baseline. Likewise, both the no-23
change-in-fare scenario (5) and the no-peak-millennials scenario (6) would have made only marginal 24
changes to ridership, 0.57 and -0.10 percentage points, respectively. 25
26
Interestingly, without Uber, bus systems would still have lost riders, but fewer only by a 0.43 percentage 27
point (than the baseline) from 2012 to 2017 on average. That is, the negative effect of Uber on bus 28
ridership does not appear as substantial as many researchers suspected, at least for our study period (i.e., 29
up until 2017). As several studies have documented (Schouten et al., 2021; Sikder, 2019; Young & 30
Farber, 2019), the socioeconomic profiles of typical bus riders and ride-hailers differ markedly. Although 31

21
some former bus riders might substitute Uber for some bus trips, their share is not large. After all, the 1
price ranges of the two services differ to a great extent, and the intensity of competition between them is 2
likely to vary geographically and temporally (Young et al., 2020). 3
4
Contrary to the claim of a recent study (Boisjoly et al., 2018), transit supply turned out to be the only 5
factor that substantially mitigated a further decline in bus ridership in the 2010s. As shown in a previous 6
section (see Figure 2), VRM in our sample UAs continuously increased since 2012. Without increased 7
services, bus systems would have lost even more riders, about twice as many as the actual loss. 8
9
4.4.2. Rail 10
The results of the simulation of rail ridership are similar to those of bus ridership, with a few exceptions. 11
Similar to bus ridership, decreasing costs of driving and increasing access to private vehicles are the most
12
significant factors in the decline in ridership. Without reductions in both gas prices and carless 13
households, rail ridership would have declined by only 1.20% from 2014 to 2017, that is, 3.14 percentage 14
points above the baseline prediction. Only by holding gas prices constant at the 2014 level would rail 15
systems have experienced a 3.12% drop since 2014, or about two fifths of the combined effects of gas and 16
vehicle access. In addition, the increase in transit fares and neighborhood changes in high-density
17
neighborhoods have also contributed to the loss of rail ridership. Their contributions, however, are not 18
substantial, each less than one percentage point from the baseline prediction. The exit of peak Millennial
19
cohorts out of young adulthood does not seem to have any impact on rail ridership. 20
21
Unlike bus ridership, for which the substitution effect prevails, rail ridership appears to be augmented by 22
Uber although the impact of this ride-hailing service is marginal at best. Last but most importantly,
23
without the increased supply of transit, rail transit would have suffered substantial ridership loss. Again, 24
VRM for rail transit increased by 6.6 percent from 2014 to 2017 (see Figure 3). Without this service
25
improvement, rail systems would have lost ridership by 5 percentage points from the baseline prediction. 26
27
5. Discussion 28
Macroeconomic factors as well as several socio-demographic trends appear to have eroded the customer 29
base of public transit, challenging the efforts by transit agencies to promote transit ridership. We found 30
that increased access to private vehicles and lower gasoline prices in recent years were the primary factors 31
for the decline in both bus and rail ridership in the early- and mid-2010s. According to the American 32
Automobile Association (Bureau of Transportation Statistics, 2019), the costs of owning and operating a 33
car fell by 1.8% per year from 2013 to 2017, mainly the result of falling gasoline prices. During the same 34
period, the share of carless households in our study areas dropped in light of the declining costs of 35
driving. Unfortunately, this trend is the opposite of what peak-car proponents in the U.S. had hoped for 36
(Kuhnimhof et al., 2013; van Wee, 2015). That is, the period of sliding ownership of cars, a decreasing 37
number of cars per capita and a growing share of carless households, was short-lived, ending around 38
2012. Note that the key determinants of personal costs of driving, such as gasoline prices, depreciation,
39
and interest rates, are largely dictated by macroeconomic conditions and policies, which transportation 40
planners have no control over. Fortunately, since our study ended, the personal costs of driving have 41
returned to normal, trending upward, and the trend of long-term car ownership appears to have stabilized 42
at around 10% of the carless household share. 43
44
One important finding highlighted in this study is that people are sensitive to the price of driving. It 45
appears as if transit ridership is as affected by, if not more, the costs of driving as transit fares. While 46
planners cannot control macroeconomic factors, they can increase the costs of driving by internalizing 47
associated social costs via policy interventions such as congestion charges and carbon taxes. Prioritized
48

22
traffic signals and dedicated lanes for public transit would also alter the relative (time) costs of transit 1
rides and driving. 2
3
In addition, it is encouraging that the increased supply of transit services during the study period 4
mitigated further ridership decline. Transit agencies are currently facing fiscal challenges because of 5
reduced ridership and fare revenue, while operating at higher costs for safety measures (e.g., frequent 6
sanitization and capacity limits for social distancing) owing to the COVID-19 pandemic. We expect fiscal 7
stress to continue even during the recovery period because the public, concerned about health, may 8
continue to avoid mass transit (Conway et al., 2020; Matson et al., 2021; Shamshiripour et al., 2020). 9
Despite this setback, transit agencies should work to attain pre-pandemic levels of transit services, with 10
the provision of proper resources. Transit has long served low-income households: the many essential 11
workers with limited access to cars and those who have suffered greater economic distress caused by 12
pandemic shutdowns. 13
14
Changes in the demographic and socioeconomic composition also accounted for ridership loss, albeit not 15
as significantly as macroeconomic factors. The shrinking share of young adults after the “peak 16
Millennial” has not yet affected transit ridership to any extent. Although a smaller cohort size than 17
Millennials, Generation Z (i.e., born from 1997 to 2012 according to Dimock (2019)) has started to 18
occupy a part of the urban housing market formerly occupied by early Millennials (i.e., those born in the 19
early 1980s), who have been moving away from central city neighborhoods (H. Lee, 2021). Thus, the 20
share of young adults in central cities is not shrinking as fast as the peak Millennial hypothesis implied. 21
Still, the 18-34 age group is expected to diminish over time; therefore, transit agencies and cities need to 22
find ways to retain the young adult population in central cities. 23
24
Also causing a moderate net loss of transit ridership are neighborhood changes in high-density transit rich 25
areas, decreasing shares of college/graduate students and those under poverty. Consistent with the 26
findings of recent studies (Paul & Taylor, 2021; Watkins et al., 2020), trends imply a growing spatial 27
mismatch between transit-rich neighborhoods and “high-propensity” transit riders in cities. To prevent 28
further mismatch, planners and policymakers need to advise transit agencies to monitor and take proactive 29
measures such as updating bus routes and schedules in accordance with the needs of transit-dependent 30
populations while protecting and promoting affordable housing in transit-rich neighborhoods. 31
32
Finally, the impact of ride-hailing services in their current form on public transportation systems appear to 33
be minor. Consistent with existing studies (Hall et al., 2018; Diao et al., 2021; Babar & Burtch, 2020), our 34
study found a small substitution effect on bus transit and a very small, yet complementary effect on rail 35
transit. However, it is still too early to draw a conclusion as we use too crude a proxy measure of ride-36
hailing supply up until 2017. Emerging autonomous vehicle technology has a strong potential to reshape 37
the landscape of both ride-hailing and all public transportation systems, calling for more research in this 38
area. Given the current price gap between ride-hailing and traditional mass transit modes, only certain 39
groups of people in certain locations are likely to be inclined to replace transit trips by ride-hailing. By 40
examining the personal characteristics and land-use attributes that account for diverging patterns of 41
response to ride-hailing, we should be able to identify a window of opportunity within which to promote 42
more sustainable modes and multimodality (Alemi et al., 2018; Chen et al., 2021; Etezady et al., 2021). 43
44
45
6. Conclusion 46
Using a predictive algorithm GBDT, we have examined the complex relationships of transit ridership 47
with key internal and external factors in large UAs in the U.S. (except New York), 85 UAs for bus and 19 48
UAs for rail, from 2002 to 2017. For both modes, a reduction in the carless-household share combined 49
with cheaper gasoline prices were the most influential factors responsible for the recent decline in 50

23
ridership; in the last several years, however, an increased transit supply has offset some of the losses in 1
ridership. Neighborhood changes in dense inner-city neighborhoods also have led to moderate attrition in 2
transit ridership, particularly bus ridership. However, the impact of ride-hailing services in their current 3
form on public transportation systems appear to be minor. We also found non-linearity, heterogeneity, 4
and interactions in the complex relationships between transit ridership and its determinants. We discussed 5
implications of our findings that shed light on effective planning and policy responses. In response to the 6
needs of transit agencies, we recommend continued support for transit services, the promotion of shared 7
and active modes via investment in infrastructure and pricing schemes, and improvement in the spatial 8
match between transit services and transit patrons, especially those residing in inner-city communities. 9
10
This study contributed to both the literature and practice in several ways. For one, we covered a long 11
study period, 2002-2017, with a comprehensive set of explanatory variables that included measures for 12
neighborhood change and “peak Millennial,” a recently coined term used in public discourse and 13
academic literature, but one that has not yet been tested empirically to any extent. In addition, we 14
employed a unique research design and a new analytical method, gradient boosting, which enabled us to 15
examine complex relationships between ridership and internal/external factors and to predict 16
counterfactual ridership under hypothetical scenarios. In so doing, we referred to the results from 17
conventional statistical models, which helped us establish a balance and synergy across a realistic 18
framework, predictive accuracy, and interpretability. Lastly, we separately analyzed bus and rail ridership 19
and found evidence that several factors such as ride-hailing affect ridership in contrasting directions 20
between the two modes. 21
22
We acknowledge a few limitations of this study and suggest directions for future research. One limitation 23
is lacking sufficient data: we did not include the last few years (2018 and later), so we could not evaluate 24
the out-of-sample performance of our trained machine learning models. Thus, readers are advised to use 25
caution when applying the findings of this study to ridership trends after 2017. In addition, this study does 26
not shed light on ridership trends that occurred during the COVID-19 pandemic, which imposed an 27
unprecedented shock on public transit systems and their use. The perception of the risk faced by transit 28
riders and changes in their daily travel patterns are critical factors to an understanding of transit use under 29
pandemic and recovery conditions. Nonetheless, we believe that awareness of the primary causes of 30
transit decline before a pandemic will definitely help us design effective strategies that promote transit 31
when a pandemic does not present a critical health threat. After all, underlying, longer-term trends will 32
not disappear overnight; instead, they will continue to interact with other factors, including changes 33
during a pandemic. 34
35
We see two promising streams of future research. For one, deriving policy implications at more granular 36
scales requires studies at disaggregate levels such as neighborhoods and transit catchment areas. In 37
addition, as we presented only the presence and patterns of heterogeneity across UAs, future research 38
could examine their causes, which are key to customized strategies serving individual UAs with varying 39
contexts and challenges. 40
41
42
Acknowledgments 43
44
An early version of this study was presented at the Association of Collegiate Schools of Planning Annual 45
Meeting in Greenville, South Carolina, in November 2019. We appreciate comments and feedback from 46
the session attendees, including Brian Taylor and Evelyn Blumenberg. This study was partially funded by 47
the Department of Urban and Regional Planning at the University of Illinois at Urbana-Champaign 48
through the Tschangho John Kim Endowment Fund. The authors claim no conflict of interest. 49
50

24
References 1
2
Alam, B. M., Nixon, H., & Zhang, Q. (2018). Factors Affecting Travel Demand by Bus: An Empirical 3
Analysis at U.S. Metropolitan Statistical Area Level. Transportation Research Record, 2672(8), 4
817–826. https://doi.org/10.1177/0361198118798714 5
Alemi, F., Circella, G., & Sperling, D. (2018). Adoption of Uber and Lyft, Factors Limiting and/or 6
Encouraging Their Use and Impacts on Other Travel Modes among Millennials and Gen Xers in 7
California (No. 18–06713). Article 18–06713. Transportation Research Board 97th Annual 8
Meeting. https://trid.trb.org/view/1497403 9
Babar, Y., & Burtch, G. (2020). Examining the heterogeneous impact of ride-hailing services on public 10
transit use. Information Systems Research, 31(3), 820–834. 11
https://doi.org/10.1287/isre.2019.0917 12
Baum-Snow, N., & Hartley, D. (2020). Accounting for central neighborhood change, 1980–2010. Journal 13
of Urban Economics, 117, 103228. 14
Bereitschaft, B. (2020). Gentrification and the evolution of commuting behavior within America’s urban 15
cores, 2000–2015. Journal of Transport Geography, 82, 102559. 16
https://doi.org/10.1016/j.jtrangeo.2019.102559 17
Berrebi, S. J., & Watkins, K. E. (2020). Who’s ditching the bus? Transportation Research Part A: Policy 18
and Practice, 136, 21–34. https://doi.org/10.1016/j.tra.2020.02.016 19
Blumenberg, E., Garrett, M., King, H., Paul, J., Ruvolo, M., Schouten, A., Taylor, B. D., & Wasserman, 20
J. (2020). What’s Behind Recent Transit Ridership Trends in the Bay Area? Volume I: Overview 21
and Analysis of Underlying Factors. https://doi.org/10.17610/T6PC7Q 22
Boisjoly, G., Grisé, E., Maguire, M., Veillette, M.-P., Deboosere, R., Berrebi, E., & El-Geneidy, A. 23
(2018). Invest in the ride: A 14 year longitudinal analysis of the determinants of public transport 24
ridership in 25 North American cities. Transportation Research Part A: Policy and Practice, 116, 25
434–445. https://doi.org/10.1016/j.tra.2018.07.005 26
Bureau of Transportation Statistics. (2019, September 16). Average Cost of Owning and Operating an 27
Automobile. Bureau of Transportation Statistics. https://www.bts.gov/content/average-cost-28
owning-and-operating-automobilea-assuming-15000-vehicle-miles-year 29
Chen, Y.-H., Lee, Y., Circella, G., & Mokhtarian, P. L. (2021, January 28). Substitution or 30
Complementarity? A Latent-Class Cluster Analysis of Ridehailing Impacts on the Use of Travel 31
Modes in Southern U.S. Cities (TRBAM-21-03318). Article TRBAM-21-03318. Transportation 32
Research Board 100th Annual Meeting. https://trid.trb.org/view/1759542 33
Clewlow, R. R., & Mishra, G. S. (2017). Disruptive transportation: The adoption, utilization, and impacts 34
of ride-hailing in the United States. 35
Conway, M. W., Salon, D., da Silva, D. C., & Mirtich, L. (2020). How Will the COVID-19 Pandemic 36
Affect the Future of Urban Life? Early Evidence from Highly-Educated Respondents in the 37
United States. Urban Science, 4(4), 50. https://doi.org/10.3390/urbansci4040050 38
Couture, V., & Handbury, J. (2020). Urban revival in America. Journal of Urban Economics, 119, 39
103267. 40
Diao, M., Kong, H., & Zhao, J. (2021). Impacts of transportation network companies on urban mobility. 41
Nature Sustainability. https://doi.org/10.1038/s41893-020-00678-z 42

25
Dimock, M. (2019, January 17). Defining generations: Where Millennials end and Generation Z begins. 1
Pew Research Center. https://www.pewresearch.org/fact-tank/2019/01/17/where-millennials-end-2
and-generation-z-begins/ 3
Ding, Chuan, Xinyu Cao, and Chao Liu. 2019. "How does the station-area built environment influence 4
Metrorail ridership? Using gradient boosting decision trees to identify non-linear 5
thresholds." Journal of Transport Geography 77:70-78. 6
Ederer, D., Berrebi, S., Diffee, C., Gibbs, T., & Watkins, K. E. (2019). Comparing transit agency peer 7
groups using cluster analysis. Transportation Research Record, 2673(11), 505–516. 8
Ederer, David, Simon Berrebi, Chandler Diffee, Taylor Gibbs, and Kari E. Watkins. 2019. "Comparing 9
transit agency peer groups using cluster analysis." Transportation Research Record 2673, no. 11: 10
505-516. 11
Egu, Oscar, and Patrick Bonnel. 2021. "Medium-term public transit route ridership forecasting: What, 12
how and why? A case study in Lyon." Transport Policy 105:124-133. 13
Etezady, A., Mokhtarian, P., & Circella, G. (2021, January 28). Investigating the Modal Impacts of 14
Ridehailing and Their Association with Shared Ridehailing. Transportation Research Board 100th 15
Annual Meeting. 16
Ewing, R., & Cervero, R. (2010). Travel and the built environment: A meta-analysis. Journal of the 17
American Planning Association, 76(3), 265–294. 18
Federal Transit Administration. (2020). June 2019 Raw Database. Retrieved February 7, 2020 from 19
https://www.transit.dot.gov/ntd/ntd-data 20
Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. Annals of 21
Statistics, 1189–1232. 22
Friedman, J. H. (2002). Stochastic gradient boosting. Computational Statistics & Data Analysis, 38(4), 23
367–378. 24
Friedman, J., Hastie, T., & Tibshirani, R. (2000). Additive logistic regression: A statistical view of 25
boosting (with discussion and a rejoinder by the authors). Annals of Statistics, 28(2), 337–407. 26
Glaeser, E. L., Kahn, M. E., & Rappaport, J. (2008). Why do the poor live in cities? The role of public 27
transportation. Journal of Urban Economics, 63(1), 1–24. 28
https://doi.org/10.1016/j.jue.2006.12.004 29
Goodwin, P., & Van Dender, K. (2013). ‘Peak car’—Themes and issues. Transport Reviews, 33(3), 243–30
254. 31
Guerra, E., & Cervero, R. (2011). Cost of a Ride. Journal of the American Planning Association, 77(3), 32
267–290. https://doi.org/10.1080/01944363.2011.589767 33
Hall, J. D., Palsson, C., & Price, J. (2018). Is Uber a substitute or complement for public transit? Journal 34
of Urban Economics, 108, 36–50. 35
Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning: Data mining, 36
inference, and prediction. Springer Science & Business Media. 37
Holmgren, J. (2007). Meta-analysis of public transport demand. Transportation Research Part A: Policy 38
and Practice, 41(10), 1021–1035. 39
Hwang, J., & Lin, J. (2016). What have we learned about the causes of recent gentrification? Cityscape, 40
18(3), 9–26. 41

26
Kuhnimhof, T., Zumkeller, D., & Chlond, B. (2013). Full article: Who Made Peak Car, and How? A 1
Breakdown of Trends over Four Decades in Four Countries. Transport Reviews, 33(3), 325–342. 2
https://doi.org/doi.org/10.1080/01441647.2013.801928 3
Landis, J. D. (2016). Tracking and explaining neighborhood socioeconomic change in US metropolitan 4
areas between 1990 and 2010. Housing Policy Debate, 26(1), 2–52. 5
Lee, B., & Lee, Y. (2013). Complementary Pricing and Land Use Policies: Does It Lead to Higher Transit 6
Use? Journal of the American Planning Association, 79(4), 314–328. 7
https://doi.org/10.1080/01944363.2014.915629 8
Lee, H. (2020). Are Millennials Coming to Town? Residential Location Choice of Young Adults. Urban 9
Affairs Review, 56(2), 565–604. https://doi.org/10.1177/1078087418787668 10
Lee, H. (2021). Are Millennials leaving town? Reconciling peak Millennials and youthification 11
hypotheses. International Journal of Urban Sciences, 0(0), 1–19. 12
https://doi.org/10.1080/12265934.2020.1871061 13
Lee, S., & Lee, B. (2020). Comparing the impacts of local land use and urban spatial structure on 14
household VMT and GHG emissions. Journal of Transport Geography, 84, 102694. 15
https://doi.org/10.1016/j.jtrangeo.2020.102694 16
Lee, Y., Lee, B., & Shubho, M. T. H. (2019). Urban revival by Millennials? Intraurban net migration 17
patterns of young adults, 1980–2010. Journal of Regional Science, 59(3), 538–566. 18
https://doi.org/10.1111/jors.12445 19
Levine, J. (2013). Is Bus Versus Rail Investment a Zero-Sum Game? The Misuse of the Opportunity-Cost 20
Concept. Journal of the American Planning Association, 79(1), 5–15. 21
Litman, T. (2019). Transit price elasticities and cross-elasticities. Victoria Transport Policy, 6. 22
Liu L, Miller HJ, Scheff J (2020) The impacts of COVID-19 pandemic on public transit demand in the 23
United States. PLoS ONE 15(11): e0242476. https://doi.org/10.1371/journal.pone.0242476l 24
Malalgoda, N., & Lim, S. H. (2019). Do transportation network companies reduce public transit use in the 25
US? Transportation Research Part A: Policy and Practice, 130, 351–372. 26
Malalgoda, N., & Lim, S. H. (2019). Do transportation network companies reduce public transit use in the 27
US?. Transportation Research Part A: Policy and Practice, 130, 351-372. 28
Mallett, W. J. (2020). Reauthorization of the Federal Public Transportation Program. 29
Manville, M., Taylor, B. D., & Blumenberg, E. (2018). Falling Transit Ridership: California and Southern 30
California (No. 01661707; pp. 1–83). University of California, Los Angeles, Southern California 31
Association of Governments. https://trid.trb.org/view/1500403 32
Matson, G., McElroy, S., Lee, Y., & Circella, G. (2021). Longitudinal Analysis of COVID-19 Impacts on 33
Mobility: An Early Snapshot of the Emerging Changes in Travel Behavior (TRBAM-21-04310). 34
1–15. https://trid.trb.org/view/1759552 35
McDonald, N. C. (2015). Are millennials really the “go-nowhere” generation? Journal of the American 36
Planning Association, 81(2), 90–103. 37
Myers, D. (2016). Peak Millennials: Three Reinforcing Cycles That Amplify the Rise and Fall of Urban 38
Concentration by Millennials. Housing Policy Debate, 26(6), 928–947. 39
https://doi.org/10.1080/10511482.2016.1165722 40
NASEM. (2004a). Traveler Response to Transportation System Changes Handbook Third Edition: 41
Chapter 9, Transit Scheduling and Frequency. The National Academies Press. 42
https://doi.org/10.17226/23433 43

27
NASEM. (2004b). Traveler Response to Transportation System Changes Handbook Third Edition: 1
Chapter 10, Bus Routing and Coverage. The National Academies Press. 2
https://doi.org/10.17226/23330 3
NASEM. (2004c). Traveler Response to Transportation System Changes Handbook Third Edition: 4
Chapter 12, Transit Pricing and Fares (Third Edition). The National Academies Press. 5
https://doi.org/10.17226/13800 6
Paul, J., & Taylor, B. D. (2021). Who lives in transit-friendly neighborhoods? An analysis of California 7
neighborhoods over time. Transportation Research Interdisciplinary Perspectives, 10, 100341. 8
Schouten, A., Taylor, B. D., & Blumenberg, E. (2021). Who’s on Board? Examining the Changing 9
Characteristics of Transit Riders using Latent Profile Analysis. Transportation Research Record, 10
0361198120987225. https://doi.org/10.1177/0361198120987225 11
Shaheen, S., & Cohen, A. (2020). Chapter 3 - Mobility on demand (MOD) and mobility as a service 12
(MaaS): Early understanding of shared mobility impacts and public transit partnerships. In C. 13
Antoniou, D. Efthymiou, & E. Chaniotakis (Eds.), Demand for Emerging Transportation Systems 14
(pp. 37–59). Elsevier. https://doi.org/10.1016/B978-0-12-815018-4.00003-6 15
Shamshiripour, A., Rahimi, E., Shabanpour, R., & Mohammadian, A. (Kouros). (2020). How is COVID-16
19 reshaping activity-travel behavior? Evidence from a comprehensive survey in Chicago. 17
Transportation Research Interdisciplinary Perspectives, 7, 100216. 18
https://doi.org/10.1016/j.trip.2020.100216 19
Shao, Qifan, Wenjia Zhang, Xinyu Cao, Jiawen Yang, and Jie Yin. 2020. "Threshold and moderating 20
effects of land use on metro ridership in Shenzhen: Implications for TOD planning." Journal of 21
Transport Geography 89:102878 22
Sikder, S. (2019). Who Uses Ride-Hailing Services in the United States? Transportation Research 23
Record, 0361198119859302. https://doi.org/10.1177/0361198119859302 24
Taylor, B. D., Miller, D., Iseki, H., & Fink, C. (2009). Nature and/or nurture? Analyzing the determinants 25
of transit ridership across US urbanized areas. Transportation Research Part A: Policy and 26
Practice, 43(1), 60–77. 27
van Wee, B. (2015). Peak car: The first signs of a shift towards ICT-based activities replacing travel? A 28
discussion paper. Transport Policy, 42, 1–3. https://doi.org/10.1016/j.tranpol.2015.04.002 29
Vij, A., Gorripaty, S., & Walker, J. L. (2017). From trend spotting to trend’splaining: Understanding 30
modal preference shifts in the San Francisco Bay Area. Transportation Research Part A: Policy 31
and Practice, 95, 238–258. 32
Wasserman, J., Taylor, B. D., Blumenberg, E., Garrett, M., King, H., Paul, J., Ruvolo, M., & Schouten, 33
A. (2020). What’s Behind Recent Transit Ridership Trends in the Bay Area?Volume II: Trends 34
among Major Transit Operators. https://escholarship.org/uc/item/96w4g18f 35
Watkins, K., Berrebi, S., Diffee, C., Kiriazes, B., Ederer, D., Transit Cooperative Research Program, 36
Transportation Research Board, & National Academies of Sciences, Engineering, and Medicine. 37
(2020). Analysis of Recent Public Transit Ridership Trends (p. 25635). Transportation Research 38
Board. https://doi.org/10.17226/25635 39
Watkins, K., Berrebi, S., Erhardt, G., Hoque, J., Goyal, V., Brakewood, C., ... & Kressner, J. (2021). 40
Recent Decline in Public Transportation Ridership: Analysis, Causes, and Responses. TCRP 41
Research Report, (231). 42

28
Ding, Chuan, Donggen Wang, Xiaolei Ma, and Haiying Li. 2016. "Predicting Short-Term Subway 1
Ridership and Prioritizing Its Influential Factors Using Gradient Boosting Decision 2
Trees." Sustainability 8 (11). doi: 10.3390/su8111100. 3
Young, M., & Farber, S. (2019). The who, why, and when of Uber and other ride-hailing trips: An 4
examination of a large sample household travel survey. Transportation Research Part A: Policy 5
and Practice, 119, 383–392. https://doi.org/10.1016/j.tra.2018.11.018 6
Young, M., Allen, J., & Farber, S. (2020). Measuring when Uber behaves as a substitute or supplement to 7
transit: An examination of travel-time differences in Toronto. Journal of Transport Geography, 8
82, 102629. 9
Zhong, L., & Lee, B. (2017). Carless or Car Later?: Declining Car Ownership of Millennial Households 10
in the Puget Sound Region, Washington State. Transportation Research Record, 2664(1), 69–78. 11
https://doi.org/10.3141/2664-08 12