A latent class analysis to understand riders’ adoption of on-demand mobility services as a complement to transit

Abstract

On-demand app-based shared mobility services have created new opportunities for complementing traditional fixed-route transit through transit agencies’ efforts to incorporate them into their service provision. This paper presents one of the first studies that rigorously examine riders’ responses to a pilot aimed at providing such a transit-supplementing service. The study conducts latent class analysis on riders of the Via to Transit program, a mobility pilot in the Seattle region where on-demand service was offered to connect transit riders to light rail stations. The analysis identifies three distinct rider groups with heterogenous responses to the on-demand service: (1) riders who previously used private cars or ride-hailing; (2) riders who were pedestrians and bikers but switched likely because of safety concern; (3) mostly socio-economically disadvantaged riders who previously relied on the bus, but switched to the new service for the convenience and speed. These results point to rich transportation policy implications, which can inform decision-making by public transit agencies as they are exploring alternative ways to deliver the mobility services.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11116-022-10351-3.

Introduction

The emerging app-based on-demand shared mobility modes (e.g., ride-hailing, bike-sharing, micro-transit) have greatly impacted how transportation services are provided in the cities. Historically, fixed-route transit has long been the primary form of public transportation service. But for a long time, public transit has been struggling to attract and retain riders despite increasing investments (Lee and Lee 2022; Manville et al. 2018; Watkins et al. 2019). The recent growth of shared mobility services, with their convenience and flexibility, may have the potential to be better integrated with public transit and fill in the gaps where traditional fixed-route transit is too costly to operate (Feigon and Murphy 2018).

Transit agencies in the United States have started to explore incorporating shared mobility modes to serve as first-mile/last-mile solutions to transit, guaranteed ride home, or replacement of some high-cost transit routes (Gifford et al. 2021; Grellier 2020; Miller et al. 2021; Shen et al. 2021). These efforts often require building partnerships with the mobility service companies and incentivizing transit riders to adopt the new services. They may also integrate fare payment and/or trip planning of the two modes. In this paper, we use the term Transit Incorporating Mobility on Demand (TIMOD) to be consistent with the naming of Federal Department of Transportation and accurately capture the supplementary roles played by shared mobility services in enhancing public transit. Similar terminologies, including multimodal integrations, innovative mobility programs, shared mobility public–private partnerships (King County Metro 2022; Wang et al. 2022), have appeared in the literature, each with a somewhat different emphasis.

Early evaluations suggest many existing TIMOD pilots achieved promising results. For example, several evaluations reported that the on-demand services were popularly adopted by riders (Gifford et al. 2021; King County Metro 2020; Miller et al. 2021). However, the exact impacts of incentivizing on-demand shared mobility on rider’s travel behavior remain largely unknown, and more rigorous examinations are needed for the following reasons.

First, introducing and incentivizing on-demand mobility services offers riders in the service area a competitive option for travel, and draws riders from different travel modes. However, most of the evaluations of existing TIMOD pilots were limited to simple descriptive statistics, and little is known about riders’ heterogeneous responses to the pilots and subsidies. For example, the policy implications of adopting on-demand modes are different between riders who switch from walking and those who switch from driving, between commuters and non-commuters, or between riders who value convenience and riders who value cost. Without a granular understanding of whom TIMOD programs primarily serve, it is difficult for transit agencies to tailor future mobility policy innovations to specific rider groups.

Second, TIMOD programs often invest a substantial amount of funds to incentivize on-demand shared modes, with the hope that such subsidies can benefit those who are socio-economically disadvantaged and mobility challenged (Gifford et al. 2021; King County Metro 2020; Miller et al. 2021). However, the social equity implications in the distributions of the investments remain a question, as the nature of the app-based travel may create technological barriers to access and use of the service (Moudon 2020). For example, Shen et al. (2021) found that incentivizing app-based carpooling for commuting in the Seattle region had the unintended consequence of disproportionally benefiting high-income employees. Therefore, it is important to analyze and compare among different groups of riders who adopt the on-demand services, and make sure that socio-economically disadvantaged riders at least benefit equally from the pilots.

This research aims to fill in these gaps as one of the first studies examining the impacts of TIMOD on riders’ travel behavior. The study uses a TIMOD pilot, Via to Transit in the Seattle region, as a case study. Specifically, the study aims to answer the following questions:

1. Are there distinct patterns in the adoption and usage of on-demand mobility services among riders?

2. Can these distinct differences be explained by riders’ socio-demographic statuses and built-environment characteristics?

3. Learning from the answers to the above two questions, what should transit agencies consider when designing and implementing TIMOD programs in the future?

To answer these questions, this study conducts latent class analysis (LCA) using survey data on Via to Transit riders. LCA is a statistical modeling technique that can systematically identify latent (unobserved) subgroups that share certain (observed) commonalities within a population. The paper proceeds with a literature review of how riders perceive and use the on-demand services as complements to traditional transit. Then the paper introduces Via to Transit program, data, and the methodological details of LCA. Next, the paper presents and interprets the latent classes identified by the LCA. The paper concludes with discussions and policy implications derived from the LCA results.

Literature review

On-demand shared mobility modes as complements to public transit

Even without transit agencies’ explicit efforts to integrate and incentivize shared mobility modes, riders have ‘naturally’ adopted them as complementary modes to public transit, most commonly to mass transit such as light rail and bus rapid transit. At the individual level, associations between a person’s usage of shared mobility modes (especially app-based ride-hailing) and public transit use were often reported (Grahn et al. 2020; Tirachini 2020; Young and Farber 2019). At the aggregate level, the introduction of app-based ride-hailing was shown by researchers to be positively associated with transit ridership in some cities, which implied the use of app-based ride-hailing as a complement mode for transit (Hall et al. 2018). However, these studies did not provide evidence that these two modes were taken in junction with each other.

A few recent studies directly modeled the mode choices of access and egress modes for transit using discrete choice models. For example, Zgheib et al. (2020) modeled the choices of feeder modes to transit with stated-preference data collected from Beirut, and their results show that app-based ride-hailing was perceived as a popular choice, especially among younger commuters. Azimi et al. (2021) employed data from an onboard transit rider survey in Orlando, FL, and found positive associations between using app-based ride-hailing as a feeder to transit and household income, trip distance, and specific trip purposes such as going to airports and universities.

Why are TIMOD programs desirable?

The results from the above literature demonstrate the potential of integrating on-demand shared mobility modes with public transit. However, the literature suggests that only ‘natural’ or purely market-driven adoptions may fail to fully achieve the transit-supplementing potential of on-demand shared mobility modes.

Both Zgheib et al. (2020) and Azimi (2021) indicated that the ‘natural’ adoption of on-demand modes as feeder modes was constrained to a niche market. The riders of TIMOD largely overlapped with the early adopters and frequent users of new transportation technologies in general, who were young, central-city dwellers, and tech-savvy (Circella et al. 2018; Clewlow and Mishra 2017; Vij et al. 2020; Young and Farber 2019). Even among them, the use of on-demand modes was still occasional (Tirachini 2020), and only for specific trip purposes such as getting to airports or avoiding drunk-driving (Azimi et al. 2021; Young and Farber 2019). A major barrier to choosing on-demand modes to access/egress transit is the high market-priced fare of on-demand modes, as Zgheib et al. (2020) demonstrated. Other barriers may include the lack of an integrated payment system and difficulties in multi-modal trip planning. It is thus interesting to know whether, with the subsidies provided and better-integrated services, TIMOD programs can overcome some of the barriers and provide more inclusive mobility access.

In addition, a ‘natural’ adoption of using on-demand shared mobility modes may not be sufficient to promote the use of transit for societal benefits. From an urban economics perspective, public transit is often advocated for its promise of positive externalities through reducing congestion, environmental pollution, and road collisions when compared with private car use (O’Sullivan 2012, Chapter 11). TIMOD programs are essentially a new form of transit subsidies that reduce a rider’s total generalized cost for accessing/egressing public transit. They may therefore help realize and internalize the positive externalities of the transit.

On the other hand, the literature pointed out that the level of mobility sharing (i.e., the passenger occupancy rate) of app-based ride-hailing, the most widely adopted shared mobility mode, tended to be low (Henao and Marshall 2018; Shen et al. 2021). As a result, it exacerbated road congestion instead of reducing it (Diao et al. 2021). TIMOD programs can thus be opportunities for transit agencies to strategically select alternative mobility service providers (e.g., ride-pooling, micro-transit) to partner with, and thus encourage “deeper sharing” (Shen et al. 2021). They can also purposefully deploy vehicles with a high level of capacity (Tirachini et al. 2020). This is particularly promising in the case of first-mile/last-mile solutions, as riders would share the same origins or destinations (i.e., the transit stops), and thus the trips can be more ‘sharable’.

Riders’ preferences and adoptions of TIMOD programs

As most of the pilots of TIMOD in the United States are still under development, few studies were able to use real-world observed data collected from these pilots to examine riders’ adoption. Most studies instead have used stated preference surveys to collect riders’ preferences and choices under the scenarios where a hypothetical on-demand, transit-supplementing service was offered. Yan et al. (2019) forecasted the adoption of a proposed TIMOD pilot on a college campus using survey data collected from students, staff, and faculty. Yan et al. (2021) and Wang et al. (2022) were a series of efforts that analyzed the attitude of residents in low-income neighborhoods toward TIMOD. They found associations between preferences towards the proposed TIMOD pilot and the respondent’s characteristics, and identified major barriers to the adoption of pilot if implemented. These studies, although offering timely insights for transit agencies, were largely exploratory because they did not observe riders’ behavior in a real-world TIMOD pilot.

Background, data, and methodology

This study examines riders’ adoption of a TIMOD service using survey data collected from Via to Transit program, supplemented by riders’ complete service usage records.

Via to transit program

Via to Transit is a TIMOD program implemented by King County Metro (KCM), the primary transit operator in the Seattle region, in partnership with Via, a private mobility service provider. Before the COVID-19 pandemic in 2020, the pilot provided first-mile/last-mile solutions to five Link Light Rail stations (Mount Baker, Columbia City, Othello, Rainier Beach, and Tukwila International Blvd Station) in south Seattle. Figure 1 shows the service areas and the locations of the five Link light rail stations served. The service was on-demand, operated using Via minivans, and could be booked through the Via app or phone calls. Riders would be requested to walk for a short distance for easier pick-up or drop-off,1 and the trips were often shared by multiple groups of riders. The service only cost riders standard transit fares ($2.75 for most riders). Based on the information provided by KCM, census tracts within two-mile buffers of the five light rail stations had a population of 279,487. The percentage of racial minorities of the population ranged from 60% in Mount Baker station area to 75% in Rainier Beach station area. The percentage of low-income households that are below the poverty level ranged from 15% in Mount Baker station area to 23% in Rainier Beach station area. Appendix A and Appendix B contain additional information, including the neighborhood income and the land use, of the study area.

Fig. 1.

Via to Transit service areas and Link Light Rail stations

We believe that studying the adoption and usage of riders of Via to Transit offers timely, transferrable insights for both KCM and other transit agencies. KCM is, in many ways, a representative transit agency that serves a medium-to-large sized U.S. metropolitan area, and the Link Light Rail is currently the primary rail transit and the ‘backbone’ of the public transit system in the region. Via to Transit thus provides meaningful mobility access in the five-station service areas. Prior to the COVID-19 pandemic, Via to Transit lasted for a year from April 16, 2019, to March 23, 2020, carrying about 230,000 trips. Its outcomes can shed light on the long-term impacts of such mobility pilots on riders. In addition, the five-station service areas consist of neighborhoods with relatively high percentages of racial minorities and low-income populations. Studying Via to Transit thus can generate invaluable knowledge regarding the social equity implications of TIMOD programs.

Data

This research primarily uses Via to Transit rider survey data while integrating some information from Via to Transit trip records. Via rider survey was administrated by King County Metro and distributed to all 8154 Via to Transit riders by sending an email with the survey link. The survey collection started on December 3, 2019, about 8 months after the launch of the Via to Transit project, and ended on January 20, 2020. Survey respondents were offered a chance to draw a $100 Visa gift card. The survey questionnaire consisted of two sections, one asking questions related to Via to Transit service and Link Light Rail usage, and the other covering respondents’ basic personal and household information. Survey respondents can be linked to Via trip data through a unique, anonymous ID. The raw data contains 1272 samples with a response rate of 15.6%.2 We followed a data cleaning procedure as in Table 1 and obtained an effective sample size of 925.

Table 1.

Via rider survey data cleaning process

	Step	N (remaining)
1	Original raw data	1272
2	Riders who had at least one completed Via trip recorda	1208
3	Riders who did not skip the socio-demographic section of the survey	1110
4	Riders who took Via for first-mile/last-mile solution for Link Light Railb	988
5	Riders who responded to all survey questions except for household incomec	925

^aThe rest could be riders who used the app and requested rides, but did not complete the trip. Or it could be riders who used credit card instead of the ORCA card (the transit smart card in the region) to pay Via to Transit, which results in survey data unable to be linked to the trip records

^bThe rest did not use Via to Transit as the access/egress mode to public transit. It is possible that their travel origins/destinations are near the light rail stations. These riders did not answer questions related to Link Light Rail usage, and thus needed to be dropped

^cDue to the fact that a large number of survey respondents (n = 227) chose not to disclose their household income, we tested two models, one with N = 698 and the household income variable containing three levels: 0—$49,999, $50,000—$100,000, and > $100,000), and the other one with N = 925 and the household income variable having four income levels: 0—$49,999, $50,000—$100,000, > $100,000, and prefer not to answer income. The latent classes identified in LCA between the two models are largely consistent. The one with larger sample comes with smaller standard errors, which helps interpret and explain the identified classes. In this paper, we present the model with N = 925

Methodology

This study employs LCA, a statistical technique that identifies latent, unobserved subgroups (or classes) within a population using manifested, observed characteristics (Vermunt and Magidson 2004). LCA has two components, as illustrated in Fig. 2, a measurement model that determines latent classes from observed indicators, and a membership model that explains the identified latent classes using a series of covariates. In this study, the indicators in the first component are variables related to the service usage and travel behavior changes associated with Via to Transit, and therefore the latent classes represent heterogeneity in the riders’ responses to the TIMOD service provided. The measurement model detects latent groups by maximizing the differences in indicators across latent classes. In the second component, the covariates are riders’ socio-demographics and built-environment characteristics, and thus the membership model explains the associations between these covariates and the probability of a rider belonging to a specific class. The two components of LCA are estimated jointly.

Fig. 2.

LCA model framework

LCA and its extensions are increasingly popular in travel behavior research, and in particular, on the attitude, adoption, and usage of new transportation technology (Alemi et al. 2018; Lee et al. 2022; Vij et al. 2020; Wang et al. 2022). However, its pros and cons compared to commonly used regression models, need to be more thoroughly discussed. The most distinct feature of LCA is that it is ‘person-oriented’ (Weller et al. 2020). Instead of being ‘variable-oriented’ and finding associations between variables, LCA finds associations across individuals and groups them. This approach thus can better inform the outcomes of transportation policy at the individual level. Second, as a direct result of the above feature, instead of narrowly focusing on a limited number of dimensions of the travel behavior (i.e., mode choice or service use frequency), the results of the LCA give us an elegant, interpretable representation of individuals’ variations in much greater dimensions. With that being said, unlike traditional regression models, the indicators are jointly explained by the class membership, and it is not easy to obtain from LCA models some quantities that might be of interest to policymakers, such as marginal effects and elasticities between variables.

A few additional features of the LCA make it an appropriate method for our study: (1) LCA uses individuals’ responses to categorical indicator variables, which works well with data collected from survey questionnaires; (2) LCA determines class from the data, and does not require prior assumptions regarding the model specifications (Alemi et al. 2018; Weller et al. 2020).

Tables 2 and 3 show the list of indicators and covariates used in the LCA of this study, respectively. Most indicators came directly from the Via to Transit rider survey. We regrouped the categories of several indicators from raw survey data to ensure a balanced distribution of observations among categories. Two additional indicators, the number of Via to Transit trips and average Via to Transit trip distance (which is the in-vehicle distance travelled with Via to Transit), were obtained from the Via trip data and transformed from a continuous scale to a categorical one, as LCA requires the indicators in the measurement model to be categorical. Regarding covariates, in addition to using the information from the Via to Transit rider survey, we obtained four additional built-environment measures for each Via to Transit rider’s most frequent travel location3 from the Smart Location Database (Chapman et al. 2021). These built-environment variables are at the Census Block Group (CBG) level and represent the location’s density, land use diversity, street design, and frequencies of transit services. We log-transformed three of them because the original distributions (shown in Appendix C) were severely right-skewed. Variance inflation factors (VIF) of all covariates in the membership model were screened, and all variables had VIF of less than 3, which suggested that the extent of multicollinearity was moderate.

Table 2.

List of indicators—variables used to determine class membership

	Categories	Source
Number of via to transit trips (one-year period)	K = 5 (1–4; 5–14; 15–49; 50–99; > 100)a	Via trip data
Avg. via to transit trip distance of the rider	K = 2 (< 1.5 miles; >  = 1.5 miles)	Via trip data
Mode replaced	K = 5 (Personal car; bus; walk and bike; ride-hailing; others)	Via survey data
Days of using the link light rail per week	K = 4 (Less than once; 1 or 2 days; 3 or 4 days; 5 or more days)	Via survey data
Via usage if the price doubles	K = 3 (Won’t use anymore; use less; use the same)	Via survey data
Safety is one of the reasons for switching to via	K = 2 (Yes; No)	Via survey data
Convenience/faster travel is one of the reasons for switching to via	K = 2 (Yes; No)	Via survey data
Cost is one of the reasons for switching to via	K = 2 (Yes; No)	Via survey data
Has utilitarian trip purpose (work, school, errands)	K = 2 (Yes; No)	Via survey data
Has recreational trip purpose	K = 2 (Yes; No)	Via survey data

^aWe chose K = 5 to make sure that there were sufficient numbers of samples in each category (i.e., K was not too large), and categories together represented rich variation in number of Via to Transit trips (i.e., K was not too small)

Table 3.

List of covariates – variables used to explain class membership

	Variable type	Source
Rider’s socio-demographics
Gender	Categorical, K = 2 (Female; male and other)	Via survey data
Age	Categorical, K = 4 (< 25; 25–44; 45–64; > 65)	Via survey data
Race	Categorical, K = 2 (White; other)	Via survey data
Household size	Categorical, K = 3 (1; 2; 3 or more)	Via survey data
Household income	Categorical, K = 4 (< $50,000; $50,000–$100,000; > $100,000; *Prefer not to answer)	Via survey data
Car available for the trip	Categorical, K = 2 (Yes; No)	Via survey data
Having a checking account	Categorical, K = 2 (Yes; No)	Via survey data
Disability	Categorical, K = 2 (Yes; No)	Via survey data
Built-environment characteristics of rider’s most frequent travel location at Census Block Group level
Population density (log)	Continuous Gross population density (people/acre) on unprotected landa	EPA smart location data, 2021
Employment and household entropy	Continuous This variable uses five-tier employment categories and the number of occupied housing units to calculate the entropy. It represents land use mix	EPA smart location data, 2021
Street intersection density (log)	Continuous Number of street intersections (auto-oriented intersections eliminated) per square mile	EPA smart location data, 2021
Frequency of transit service per square mile (log)	Continuous Aggregate frequency of transit service per hour per square mileb	EPA smart location data, 2021

^aUnprotected land means land that is not protected from development, for example, a park or conservation land. It is used by EPA to calculate density-related measures

^bThe frequency of transit service was calculated by aggregating the frequency per hour of all transit routes within 0.25 miles crow-fly distance of the boundary of the CBG during weekday peak hours. The frequency was then divided by land area to get frequency per square mile

For LCA, the number of latent classes needs to be pre-determined. Model fit and interpretability are two important factors when making such a decision. One common approach is to run LCA with different numbers of classes and choose the one with the best model fit, measured by the Akaike information criterion (AIC) or Bayesian information criterion (BIC). Table 4 shows the model fit with different class numbers, where a smaller AIC or BIC suggests a better fit. Based on the results, AIC favored 4 classes and BIC favored 3 classes. After examining the model outputs in these two models, we chose the one with 3 classes because its classes are more distinctive from each other and thus the results are more interpretable, which lead to relatively clear policy implications. The model was estimated using the ‘poLCA’ package in R.

Table 4.

Model fit with different number of classes

	2 classes	3 classes	4 classes
AIC	15,024	14,710	14,589
BIC	15,275	15,125	15,169

Results

Results of the measurement model

Table 5 shows the three classes identified by the measurement model.4 The first row shows that the sizes of Class 1, 2, and 3 are 34%, 39%, 27% of the total sample size (N = 925), respectively. Each value starting from the second row is the posterior probability of riders within the class being in the corresponding category of the indicator. For example, the first number, 0.59, indicates that 59% of riders in Class 1 made 1–4 Via to Transit trips during the one-year pilot. The bold numbers are the highest values for each row. For example, among three classes, Class 1 has the highest percentage of riders that made 1–4 Via to Transit trips. Table 5 thus presents an easy-to-interpret way to understand the three classes by reading the class-specific distribution of indicators.

Table 5.

Results of the estimated LCA measurement model (N = 925)

Indicators	Categories	Class 1	Class 2	Class 3
Class size		34%	39%	27%
Number of via to transit trips (one-year)	1–4	0.59	0.19	0.10
	5–14	0.31	0.25	0.17
	15–49	0.08	0.30	0.25
	50–99	0.01	0.14	0.19
	> 100	0.00	0.11	0.30
Avg. via to transit trip distance	< 1.5 miles	0.64	0.99	0.25
	≥ 1.5 miles	0.36	0.01	0.75
Mode replaced	Personal car	0.41	0.13	0.31
	Bus	0.12	0.21	0.46
	Walk/bike	0.27	0.60	0.08
	Ride-hailing	0.14	0.01	0.09
	Other	0.05	0.05	0.06
Days of using the link station per week	Less than once	0.82	0.14	0.15
	1 or 2 days	0.17	0.21	0.20
	3 or 4 days	0.02	0.30	0.32
	5 or more days	0.00	0.35	0.34
Via usage if the price doubles	Won’t use it anymore	0.19	0.39	0.39
	Use less	0.48	0.45	0.43
	Use the same	0.32	0.16	0.18
Reason for switching—safety	Yes	0.19	0.44	0.33
	No	0.81	0.56	0.67
Reason for switching—convenience & faster travel	Yes	0.70	0.78	0.84
	No	0.30	0.22	0.16
Reason for switching—cost	Yes	0.22	0.07	0.22
	No	0.78	0.93	0.78
Trip purpose: utilitarian	Yes	0.52	0.99	0.98
	No	0.48	0.01	0.02
Trip purpose: recreation	Yes	0.70	0.46	0.37
	No	0.30	0.54	0.63

Bold numbers indicate the highest value for each row

Class 1 has the highest share of riders who used Via to Transit occasionally, with 59% of them only taking 1–4 Via to Transit trips, and 31% of them taking 5–14 trips. The class has the highest share of riders who switched from private motorized modes, i.e., personal car (41%) and ride-hailing (14%). It is the class that was least dependent on the Light Rail, with 82% of riders only using the Light Rail less than once a week and 17% of them using it 1 or 2 days a week, which also explained the infrequent use of Via to Transit. The class is least sensitive to price with the highest percentage of riders indicating they would still use Via to Transit even if the price increased. This class is least likely to adopt Via to Transit because of safety (19%), least likely to travel for utilitarian purposes (52%), and most likely to travel for recreational purposes (70%).

Class 2 consists of a higher percentage of riders whose Via to Transit trip distance was on average less than 1.5 miles (99%) than the other two groups, which makes sense as this class has the highest share of riders who previously walked or biked to the station (60%), although about 35% of the class still took motorized modes. More riders in this class needed to access the Light Rail frequently as 35% of them rode the Link 5 or more days a week. The class is more price-sensitive than Class 1, with 39% of them would stop using Via to Transit if the price increased. The class has the highest share of riders who switched because of safety (44%). Almost all of them had utilitarian trip purposes (i.e., work, school, or personal errands) (99%).

Class 3 contains a higher percentage of frequent Via to Transit riders than the two groups, as 19% of them took 50–99 Via to Transit trips and 30% of them took more than 100 trips in a year. They were more likely to be riders whose trip distance on average exceeded 1.5 miles (75%). They were likely to be frequent transit riders. The class has the highest share of riders who previously took the bus to access Link stations (46%), although 31% of them drove to the station. The class has a relatively high percentage of riders who took Link Light Rail 3 or 4 days (32%) or 5 days or more (34%) in a week. Similar to Class 2, Class 3 is also price sensitive as 39% of them would stop using Via to Transit if the price increased. A large proportion of riders (84%) in Class 3 switched because of Via to Transit’s convenience and savings in travel time. Regarding trip purpose, almost all riders in Class 3 had utilitarian purposes (98%), and they were the least likely ones to have recreational purposes (37%).

To summarize, three latent classes identified in the measurement model were distinctive from each other and represented substantial variations in their Via to Transit usage, mode choices, reasons for switching, and trip purposes. We will further discuss the role Via to Transit played in their travel and the corresponding policy implications in the discussion section.

Results of the membership models

Table 6 presents the second component of the model, where socio-demographics and built-environment covariates were used to explain the membership of three latent classes identified from the measurement model. In such a way, the membership model resembles a multinomial logistic regression, where the dependent variable is a categorical variable with three categories, and the estimated coefficients represent the associations between a one-unit change of the independent variables and the changes in the log-odds of belonging to a class relative to the reference class, when holding other variables constant. The reference class is Class 1.

Table 6.

Results of the estimated LCA membership model (N = 925)

Coefficient	Class 2 (Ref. class 1)	Class 3 (Ref. class 1)
Gender: female	− 0.011	− 0.36
Age: < 25 (ref. 25–44)	0.54	0.23
Age: 45–64 (ref. 25–44)	− 0.48*	− 0.78**
Age: > 65 (ref. 25–44)	− 2.35***	− 1.63***
Race: racial minorities (ref. white)	0.31	1.20***
HH size: 1 (ref: 2)	− 0.26	0.265
HH size: 3 or more (ref: 2)	0.61**	− 0.00
Household income: < $ 50,000 (ref: $50,000–$100,000)	0.51	0.13
Household income: > $ 100,000 (ref: $50,000–$100,000)	− 0.20	0.32
Household income: prefer not to answer (ref: $50,000–$100,000)	− 0.1	− 0.118
Car available: yes	− 0.59**	− 0.82**
Having a checking account: yes	− 0.27	− 1.04*
Disability: yes	− 0.12	− 0.36
Population density (log)	0.79**	− 0.37
Employment and household entropy	2.29***	− 4.71***
Intersection density (log)	− 0.003	− 2.10***
The total frequency of transit service per sq. mile (log)	0.83***	− 0.98***
Intercept	− 6.34***	18.35***

*p < 0.1; **p < 0.05; ***p < 0.01

Among socio-demographic variables, the model finds that riders who were older (over 45 years old) and who had greater access to cars were significantly less likely to be in Class 2, while riders with larger household sizes were significantly more likely to be in Class 2. Riders who traveled to places that were denser with greater land-use mix and transit service were more likely to belong to Class 2 than to Class 1, and these associations were statistically significant. These findings are in general consistent with the travel behavior of Class 2, as Class 2 consists of more pedestrians and bikers than Class 1, and their origins or destinations were closer to the Light Rail stations as well.

Regarding the coefficients for Class 3 versus Class 1, the model finds that older riders were significantly less likely to be in Class 3. Being a racial minority was significantly associated with a greater likelihood of belonging to Class 3. Having access to a car and access to a checking account were both negatively associated with the likelihood of belonging to Class 3. All these results suggested that riders in Class 3 were more likely to be more disadvantaged compared to Class 1, which is consistent with their higher dependency on transit and higher price sensitivity, as reflected in the measurement model. The results for the built-environment covariates for Class 3 are particularly interesting, as riders who traveled to places with lower land use mix, lower street intersection density, and lower frequencies of transit services were significantly more likely to belong to Class 3. One possible explanation is that Via to Transit greatly expanded the travel destination options for those who were previously more transit-dependent (i.e., Class 3). Before Via to Transit was available, their travel was constrained to narrow corridors along the bus lines. And Via to Transit allowed them to freely travel to places that existing buses could not reach. Another possible explanation is that such associations are a reflection of Class 1 and Class 3’s distinct trip purposes. Riders in Class 1 took more Via to Transit trips for non-utilitarian purposes and therefore they traveled to places where social and commercial activities took place. While riders in Class 3 took Via to Transit trips for utilitarian purposes, and more of their trips started or ended at their homes, where the built environments were more residential. Either way, the results strongly suggest that Via to Transit better connected Class 3 to Link stations from places with limited convenient access to jobs and services, unfriendly street design to active travel, and relatively poor transit services, as shown by the results for the built-environment covariates in Table 6.

Mapping three classes

Figure 3 shows the spatial distributions of the most frequent travel locations of riders belonging to each class, as well as the median household income of census block groups in the area.

Fig. 3.

Heatmaps of three identified classes (darker color = higher rider counts) and median household income. (Color figure online)

The travel locations of riders in Class 1 were in general dispersed across the area, reflecting their travel was less spatially constrained because of their better access to personal vehicles. There were a few concentrations of riders belonging to Class 1 in relatively high-income neighborhoods along the waterfront on the right of the map. Compared to Class 1, more of Class 2’s travel locations were in areas immediately surrounding the Light Rail stations, reflecting the fact that Class 2 consists of higher percentage of riders who previously walked or biked to stations. Class 3, which consists of higher percentage of riders who were previously bus riders, had travel needs from and to areas in the bottom-right of maps, where many neighborhoods were with lower median household income. Their locations were also further away from the Light Rail stations. Such clustering of riders of Class 3 further demonstrates the equity implications of Via to Transit, which enables riders in socio-economically disadvantaged neighborhoods to get better connected to the Light Rail stations.

Discussions and policy recommendations

Rich policy implications can be drawn from the above results. Class 1 is the class with the least frequent use of Via to Transit. It consists of more riders who were previously personal car or ride-hailing users, infrequent Link Light Rail riders, less price-sensitive, less safety-concerned, and primarily recreational travelers. Socio-demographically, they tended to be older and had better access to cars compared to the other two groups. Via to Transit converted a higher percentage of riders in Class 1 from driving alone and ride-hailing to Via to Transit. For travelers who did not take Link Light Rail prior to the pilot, such conversion came with many social and environmental benefits because it could reduce the use of vehicles and the number of vehicle trips. And for travelers who previously used personal cars as first-mile/last-mile solutions to Link, Via to Transit reduced the demand for parking facilities near the Light Rail stations. This class’s profile to some degree aligns with the early adopters of on-demand shared mobility modes (Tirachini 2020; Vij et al. 2020; Young and Farber 2019), as members were likely to be wealthier and to use Via to Transit for recreational purposes, likely, for avoiding drunk-driving. However, the fact that riders in Class 1 were older shows that with subsidies and greater integration between transit and on-demand modes, TIMOD programs have the potential to expand rider groups beyond what can be achieved in a natural, market-driven adoption, where in general riders are young and tech-savvy.

Riders in Class 2 and Class 3 shared some similarities. For example, they were more sensitive to prices, accessed the Light Rail more frequently, and were more likely to use Via for utilitarian purposes. However, the two classes adopted Via to Transit for different reasons. For riders in Class 2, they were more likely to be pedestrians and bikers who were concerned about safety5 when walking and biking to or from the stations. Therefore, Via to Transit offered a safer travel means for this class of riders. By doing this, Via to Transit took riders away from non-motorized, active modes, which may have adverse environmental consequences. To address riders’ safety concern, encouraging mode switch might be an effective temporary solution, but long-range planning and transportation policies are also needed. For example, providing and maintaining adequate streetlight in the neighborhoods and creating mixed-use streets that are welcoming to pedestrians and bikers. If such safety issues can be properly addressed, alternative on-demand modes (e.g., bike-sharing or scooter-sharing) might be more suitable to meet the needs of these riders, given the fact that Class 2’s trip distance were relatively short. We recommend that future TIMOD pilots should explore these alternatives.

Riders in Class 3 were more likely to be those who faced mobility challenges in travel. On one hand, they lacked access to personal vehicles and previously relied on buses to access to/egress from the Light Rail. On the other hand, they had greater needs to travel to/from places that were hard to access without a personal vehicle. In addition, this class consists of a higher percentage of racial minorities. These findings suggested that Via to Transit provided a more convenient and faster first-mile/last-mile solution for riders in Class 3. Furthermore, Via to Transit expanded the places that they could access beyond corridors along the bus lines, although admittedly Via to Transit trips were still bounded by the service area boundary that KCM designated. However, Via to Transit service directly took riders away from existing bus lines in the service areas as it provided a much more convenient service. When implementing future TIMOD programs, transit agencies should investigate ways to better design and deploy TIMOD so that the newly provided services can better fill in the gaps of existing bus transit, and the comparative advantages of both bus and on-demand modes can be realized.

These empirical results provide timely insights for transit agencies that are exploring new shared mobility options to supplement traditional fixed-route transit. However, because the dataset collected from the transit agency was primarily focused on the first-mile/last-mile travel, it is not sufficient for gaining a more general understanding of TIMOD’s impacts on riders’ travel behavior. Future research should consider designing a more comprehensive rider survey, and perhaps employing travel logs, to fill in the remaining gaps. A more comprehensive rider survey can also help better understand riders’ price sensitivity to TIMOD services. Furthermore, future data collection and analysis should include TIMOD programs that serve other trip purposes/destinations in addition to access/egress to rail transit stations. For example, it is conceivable that TIMOD can replace an entire low-efficiency bus route or high-cost paratransit services. Similarly, transit agencies can greatly benefit from research that comparatively examines multiple TIMOD pilots implemented in different cities and evaluates various design and implementation strategies characterized by on-demand mode, incentive amount and structure, and service integration scheme. A broader understanding of how different TIMOD programs may generate different efficiency and equity outcomes will be essential for transit agencies to develop their programs to best serve the targeted population groups.

Conclusion

This work is one of the first studies that use methods beyond simple descriptive statistics to evaluate riders’ responses to a completed TIMOD pilot. Employing LCA on survey data collected from riders of the Via to Transit program serving several light rail station areas in the Seattle region, the study revealed transportation policy trade-offs for each of the three latent rider classes identified. We found that the TIMOD service in Seattle converted some riders who previously used private cars or ride-hailing to a public transit service, provided a safer mobility option for those who used to be pedestrians and bikers, and made accessing Light Rail stations more convenient for those who previously relied on the bus.

Given the findings that TIMOD service impacted the travel behavior of different rider groups and benefited them in different ways, transit agencies must clearly define their policy goals before designing and implementing their TIMOD programs. This requires an in-depth understanding of the socio-demographic and built-environment characteristics of the areas the programs intend to serve. Transit agencies should also actively engage with riders to better understand their travel needs and barriers.

The empirical research presented in this paper focused on a TIMOD pilot designed to serve the first-mile/last-mile travel of light rail riders in the Seattle region, and therefore the insights provided here, while timely and informative, are far from being sufficient to support decision-makings for different kinds of TIMOD programs. Future research must aim at achieving a more general understanding of TIMOD’s likely impacts on population groups in diverse contexts. It will necessitate many studies of TIMOD programs designed to serve different trip purposes, implemented in different cities, and operated by various mobility service providers.

Supplementary Information

Below is the link to the electronic supplementary material.

Biographies

Dr. Yiyuan Wang

is a recent graduate of the Interdisciplinary PhD Program in Urban Design and Planning at the University of Washington. He is interested in applying economic theories and models to investigate the impact of new transportation technologies on individual travel behavior and public policymaking. His most recent works include studying public transit agencies’ initiative to supplement tranditional transit with on-demond shared mobility modes.

Dr. Qing Shen

is Professor and Chair of the Interdisciplinary PhD Program in Urban Design and Planning at the University of Washington. His primary areas of interest are urban economics and metropolitan transportation planning and policy. Professor Shen has developed new methodological frameworks for analyzing urban spatial structure, examined the social and environmental consequences of automobile-oriented metropolitan development, and investigated the differential impacts of information and communication technologies on various population groups. His current work focuses on new challenges and opportunities for transportation demand management in the age of shared mobility. He earned his PhD in City and Regional Planning from University of California, Berkeley.

Author contributions

Conceptualization: QS, YW; Data curation: QS, YW; Software: YW; Methodology: QS, YW; Formal analysis and investigation: QS, YW; Writing - Original Draft: QS, YW; Writing - Review & Editing: QS, YW; Visualization: YW; Project administration: QS, YW; Supervision: QS

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.