How Does Ride-Hailing Influence Individual Mode Choice? An Examination Using Longitudinal Trip Data from the Seattle Region

Abstract

This study investigates the impacts of ride-hailing, which we define as mobility services consisting of both conventional taxis and app-based services offered by transportation network companies, on individual mode choice. We examine whether ride-hailing substitutes for or complements travel by driving, public transit, or walking and biking. The study overcomes some of the limitations of convenience samples or cross-sectional surveys used in past research by employing a longitudinal dataset of individual travel behavior and socio-demographic information. The data include three waves of travel log data collected between 2012 to 2018 in transit-rich areas of the Seattle region. We conducted individual-level panel data modeling, estimating independently pooled models and fixed-effect models of average daily trip count and duration for each mode, while controlling for various factors that affect travel behavior. The results provide evidence of substitution effects of ride-hailing on driving. We found that cross-sectionally, participants who used more ride-hailing tended to drive less, and that longitudinally, an increase in ride-hailing usage was associated with fewer driving trips. No significant associations were found between ride-hailing and public transit usage or walking and biking. Based on detailed travel data of a large population in a major US metropolitan area, the study highlights the value of collecting and analyzing longitudinal data to understand the impacts of new mobility services.

INTRODUCTION

In recent years, the exponential growth of mobile information and communication technology-enabled mobility services offered by transportation network companies (TNCs), such as Uber and Lyft, has drawn much attention of transportation scholars and practitioners [1–3]. TNCs provide a new form of ride-hailing services, which have long existed in major cities, primarily in the traditional form of taxis. Despite the fact that the new form is app-based, it shares many similarities with traditional taxis where drivers are service providers who offer mobility services to riders [3–5].

Attempts have been made to understand the impacts of the new form of ride-hailing on traditional transportation modes, especially on whether it is substituting for or complementing other modes. Previous studies have employed various data and methodologies to investigate this relationship. Some studies have analyzed aggregated metropolitan-statistical-area (MSA) or transit-agency-level data [6–9] to examine the relationship between TNCs’ entry and regional travel demand. However, most studies lacked data on actual service ridership. In contrast, individual-level studies have used data from intercept surveys targeting ride-hailing users [4,10,11]; from household travel surveys targeting ride-hailing users [2,12]; and from regional or national household travel surveys that used travel log [13–15]. However, few of these studies were able to apply rigorous statistical methods that controlled for the various confounders affecting mode choice. And for those that did, most only used one-year cross-sectional data. As a result, in order to reach a consensus on the impacts of ride-hailing on other modes, more research is needed to better prepare cities and the transportation sector for the era of shared mobility [16–18].

This study attempts to advance knowledge on this question by utilizing a unique longitudinal travel log and survey dataset from the ACTION (Assessing Choices in Transportation in Our Neighborhoods) project. ACTION is an NIH-funded study that collects primary data on the location and activity of more than 400 individuals in the Seattle region every two years for three times since 2012. These rich data offer a great opportunity to study trends in ride-hailing usage and other individual travel patterns.

The paper begins with a review of recent studies on the relationship between ride-hailing and other modes, focusing on their research methods and findings. The subsequent section introduces the data and methodology for this study. Next comes a presentation of the models and estimation results, followed by a discussion and conclusions.

LITERATURE REVIEW

The lack of consensus on the impacts of app-based ride-hailing on traditional modes is partly explained by the different research methods used in current literature. Therefore, the first section of the literature review summarizes different methods adopted in previous studies and the potential effects of such choices of methods on their results. The second section of this review synthesizes previous research findings.

Methods to study the impacts of app-based ride-hailing on other modes

A common approach is to use aggregated data to examine the impact of TNC entry on regional mobility trends such as public transit ridership, vehicle ownership, and average commuting time [6–9]. These studies often take MSA or transit agencies as the unit of analysis. The greater availability and consistency of the data at the aggregated level enable these studies to develop statistical models that can single out the effects of app-based ride-hailing from various confounders. For example, Boisjoly et al. [8] applied a mixed effect model to investigate factors affecting transit ridership in the service areas of 25 transit agencies in North America. Hall et al. [6] used a difference in difference model on ridership data for transit agencies in the US to estimate the impact of Uber entry on transit ridership. However, most of the studies measured the effects of app-based ride-hailing using a dummy variable representing whether or not the service is present in the city [6,8], or a variable capturing the years since the entry [7,9]. These studies do not present data on actual TNC ridership. Only Hall et al. [6] used Google search engine ‘intensity’ as a proxy for the ridership. Furthermore, aggregate level analyses cannot reveal mode choice at the individual level, and therefore are not able to identify different impacts of this new form of ride-hailing on heterogeneous population groups.

An alternative approach is to collect and analyze data at the individual level. A convenient way to do so is to survey ride-hailing riders and gather a convenience sample [4,10,11,19]. Such surveys may take place in ride-hailing vehicles during the services, or near ride-hailing hubs, or the ride-hailing service providers may distribute the surveys on their app platforms. This approach can gather data from a large pool of ride-hailing users and contribute to a better understanding of ride-hailing users' characteristics. However, such survey techniques are not likely to generate a representative sample. Of note, most of these studies asked what modes to choose if ride-hailing was not an option to test the substitution effects, but they did not probe potential complementarity effects.

Clewlow and Mishra [2] conducted a study on shared mobility users, and they acquired a representative sample of households in seven metropolitan areas. They asked respondents to not only report what alternative modes they would use when app-based ride-hailing was not available, but also to self-estimate changes in the use of transit, walking and biking after the adoption of ride-hailing, thus potentially getting information on mode complementarity. Coy et al. [12], which is a similar effort, gathered representative cross-sectional travel data with over-sampled app-based ride-hailing users.

There have been efforts to research the impacts of ride-hailing using national or regional household travel surveys [13–15]. Although not created to study the impact of shared mobility solely, many of these travel surveys involved travel logs for more complete and accurate records of participants’ travel activities. Some of them, such as the Puget Sound Regional Council (PSRC) Household Travel Survey in the Seattle region, even utilized GPS devices or smartphone apps to ensure accuracy [20]. Therefore, using data from these sources have a greater potential to overcome bias in stated-preference surveys.

Many studies at the individual level only applied summary statistics, except for a few efforts that developed regression models to better control for various confounders [11,13,15]. However, partly due to the fact that the emergence of Uber and Lyft is a recent phenomenon, most of the regression models were based on cross-sectional data. It is not clear whether the variation in the mode choice among participants is a disruptive effect from TNC services, or a self-selection effect that participants with certain characteristics are more likely to use such services [13]. Therefore, there is a need to develop models using longitudinal data, which helps to examine how the relationship between ride-hailing and other modes changes over time.

Impacts of ride-hailing services on driving

It is more likely that ride-hailing serves as a substitute for driving. For example, In Feigon and Murphy’s study [19], after the adoption of app-based ride-hailing, 35% of the respondents reported driving less for commuting, versus 4% who reported driving more; and 32% reported driving less for errands or recreation, versus 10% who reported driving more. They also found respondents reducing vehicle ownership after the adoption. Similarly, Henao and Marshall [10] found that, when asked what mode to use if app-based ride-hailing was not available, 19% of respondents in Denver chose driving, ranking second after public transit. Sun et al. [13] also reported significant substitution effects of ride-hailing on driving in their models. However, other research suggests weaker substitution effects. For example, Rayle [4] also found few ride-hailing trips in San Francisco replaced driving.

Impacts of ride-hailing services on public transit

It is likely that app-based ride-hailing services, with its flexibility and convenience, take some transit customers away in areas where transit services are not efficient. However, such services can also serve as a first-and-last mile solution and thus complement the transit system.

In studies that applied aggregated data, Boisjoly et al. [8] reported no significant effect of TNC entry on transit ridership, while Graehler et al. [9] used more recent data and reported a substitution effect. Sadowsky and Nelson [7] developed a more complicated model and found an initial complementary effect of TNCs after the entry of the first TNC company, but an increasing substitution effect since the entry of the second TNC company. Hall et al. [6] investigated the heterogeneous effects of app-based ride-hailing on different transit agencies’ ridership, and found that such service was more likely to complement small transit systems in large MSAs, and substitute for big transit systems in small MSAs.

In studies based on individual-level survey data, Feigon and Murphy [19] found that 43% of the respondents reported using public transit more after the adoption of shared mobility options versus 28% who reported less, which suggested a complementary effect. Clewlow and Mishra [2], using a representative sample, found a 6% reduction in the usage for bus, a 3% reduction for light rail, and a 3% increase for commuter rail after the adoption of app-based ride-hailing. These results indicate that the effects are heterogeneous on different transit modes. Also, respondents frequently rank public transit at the top of the list when asked which modes to take if app-based ride-hailing was not available [4,10–12], suggesting potential substitution effects.

Among studies that used regional or national travel surveys, Young and Farber [14] found that the new form of ride-hailing has no substantial effects on public transit in the Toronto region, but Sun et al. [13], using data from Seattle, found that transit and ride-hailing are likely to substitute for each other.

Impacts of ride-hailing services on walking and biking

Ride-hailing can also impact walking and biking. On one hand, some forms of shared app-based ride-hailing services (e.g., Uber Pool and Lyft Line) ask riders to walk to/from nearby intersections for pick-up and drop-off. It is also possible that ride-hailing allows people to walk and bike for their outgoing trip and return by ride-hailing should they prefer to do so. On the other hand, ride-hailing trips may directly substitute for some walking and biking trips. Feigon and Murphy [19] found that 54% of the respondents reported being more physically active since the adoption of shared mobility options. Only 5% reported being less physically active, suggesting a complementary role of ride-hailing on active travel. Young and Farber [14] reported an association between app-based ride-hailing and an increase in active modes for specific travel demand in the Toronto region. However, Gehrke et al. [11] found that certain ride-hailing trips, for example, ones that are less expensive and presumably shorter, are likely to substitute for active transportation trips.

RESEARCH QUESTIONS

This research aims to utilize a longitudinal dataset to investigate the relationship between ride-hailing, including both traditional taxis and app-based services, and other travel modes, and to overcome some shortcomings of cross-sectional studies. Specifically, this research aims to answer the following questions:

Question 1: What are the relationships between ride-hailing and other travel modes, such as driving, public transit, and walking and biking? Does ride-hailing substitute for or complement these modes?

Question 2: How do these relationships change over time?

DATA AND METHODOLOGY

Data

This study uses data from the ACTION project, an NIH funded study designed to study the impacts of the new E and F Bus Rapid Transit (BRT, called RapidRide service in Seattle/King County) lines on individual physical activity. The study included three waves of data collection, taking place before (2012–2014), soon after (2015–2016), and two years after (2017–2018) the new BRT lines started operation.

Recruitment for ACTION began in July 2012 and continued through January 2014. About half of the participants were selected from among people living within ½ mile of an E or F line BRT stop. The other half were selected from among people living in other neighborhoods within the Seattle region that matched BRT stop-adjacent neighborhoods on household income, racial/ethnic composition, residential property values, residential density, land use mix, bus ridership, and housing type. Households in eligible areas were contacted using the address and phone information obtained from MSG (Marketing Systems Group), a consumer marketing company. Enrollment processes ensured participants were adults 18 years old or older who gave informed consent, were able to walk unassisted for at least 10 min, and completed the travel log and survey in English. The Seattle Children’s Research Institute Institutional Review Board approved the study. Figure 1 shows the home locations of participants in our sample at the beginning of the study. These individuals are ideal for studying the relationship of new mobility options with other modes, especially with public transit, because: 1) for most of them, transit serves as a viable travel option; and 2) these individuals were selected through random sampling, thus are more likely to be representative.

Figure 1. Participants’ home locations at wave 1 (n=460).

Map Reference: Open Street Map

In the extensive survey focusing on socio-demographic and attitudinal factors, participants reported their age, gender, race/ethnicity, height, and educational attainment at baseline (wave 1). At baseline and each follow-up wave, they reported annual household income, weight, number of other household members, and number of motor vehicles available to the household. For measures of travel behavior, participants were asked at each wave to wear an Actigraph GT3X accelerometer and a QStarz BG-1000XT global positioning system (GPS) data logger for seven days while also recording their travel in a paper diary over the same period (as shown in Figure 2). The study team timed contact with participants to maximize monitoring at the same time of the year as prior waves.

Figure 2. ACTION travel log sample page and codes for travel mode.

This extensive study generated a series of longitudinal data at three periods, although the number of observations varied somewhat because of participant attrition over the course of the study. Table 1 presents the ACTION’s timeline and sample size. We obtained a total of 58,901 trips for the three waves.

Table 1.

ACTION study dates and longitudinal sample size

Wave	Year	Number of participants	Number of trips in the travel log data
1	2012–2014	460	21890
2	2015–2016	371	18591
3	2017–2018	361	18420
Total			58901

Based on online records, Uber launched the service in the Seattle region in 2011 [21], and Lyft first entered the market in 2013 [22]. Since then, the two companies rapidly expanded their market in Seattle and other regions; their annual ridership is estimated to pass 20 million in 2017 [3]. The timeline of ACTION’s three-wave data collection, which started in 2012 and ended in 2018, aligned well with the emergence and rapid expansion of app-based ride-hailing in the region.

Methods

This study uses the travel log data to calculate the total trip count and duration for each mode and examines how the share of ride-hailing has changed over time. Trip count is commonly used in transportation studies to analyze mode choice, while this study also calculates the trip duration to reflect the actual time that participants spent using each mode, which is an important aspect when considering relationships between modes. We followed a data cleaning procedure to process the travel log data, as shown in Table 2. First, we removed about 20% of total trips with mode = NA or = blank or = invalid code. The majority of the dropped trips are movements within one’s house (e.g., going to the garage, working in the yard). Then, for trip duration statistics, we removed additional trips with duration = NA or = negative values, which tend to be mistakes when participants filled in the travel log. We also excluded a small number of trips with extremely long duration (greater than 5 hours).

Table 2.

ACTION travel log data cleaning process

Step	Description	Trips
		Count	% removed
1	Original number of trips in three waves	58,091
For both trip count and trip duration statistics
2	Removing trips with mode = NA or = blank or invalid code	46,598	19.8%
For trip duration statistics only
3	Removing trips with duration = NA	43,835	5.9%
4	Removing trips with duration = negative	41,155	6.1%
5	Removing trips with duration > 5 hours	41,078	0.2%

To answer the research questions, we develop a series of panel models at the person-wave level. The key variables in the model, the travel behavior variables, are derived from the travel log data. The modes are determined according to the codes in Figure 2. Note that ride-hailing in the ACTION survey is described as “taxi, shuttle bus, limousine” and does not explicitly distinguish between the new app-based ride-hailing services (e.g., Uber and Lyft) and traditional taxis. Therefore, the results of the models in the paper should be interpreted with caution, as they represent impacts from both traditional taxi services and new ride-hailing services. We will further discuss this limitation in the ‘Discussion and Limitation’ section, and elaborate on why we believe this mode effectively captures the emerging app-based ride-hailing trips in addition to traditional taxis.

For each participant at each wave, we calculate their average trip count per day and the average trip duration per day for each of the following modes: driving, public transit, walking and biking, and ride-hailing. We run the models using the average trip count/duration of driving, public transit, and walking and biking as dependent variables and the average trip count/duration of ride-hailing as independent variables. We control for other factors that affect mode choice: demographic factors (age, gender, race, household size, and the number of children within the household), socioeconomic factors (educational attainment, income level, employment status, and vehicle ownership), attitudinal factors (whether a participant likes taking public transit and whether s/he likes driving), built environment factor (residential unit density), and time-specific effects of the second and the third waves. Control variables come from the ACTION survey, except for residential unit density, which is calculated as the number of residential units within an 800-meter radius of each individual's home location based on county assessor’s data [23]. We calculate density in 2013, 2015, and 2017 for the three data collection waves, respectively. These control variables are commonly adopted in travel behavior and new mode choice adoption studies [14,24]. Before running the model, we checked multicollinearity by calculating the variance inflation factor (VIF). All the variables included in the model have VIF less than 3, which suggested that multicollinearity is moderate.

We use negative binomial models that are well-suited for modeling data that are right skewed rather than normally distributed. Three types of panel models are commonly adopted while using longitudinal data: independently pooled model, fixed-effect model, and random effect model. The independently pooled model treats longitudinal data as cross-sectional, and ignores the fact that repeated measures from the same participant tend to be correlated (i.e., serial correlation). Thus, the estimation might be biased due to its inability to account for such correlation. Instead, the fixed-effect model adds an individual effect to the model for each participant and assumes that such an effect is non-random. By doing this, it better controls for the serial correlation. However, the fixed-effect model comes with cost. It purges all the cross-sectional variation from the model and only models the time-variant part of the dependent variables. In our case, the fixed-effect model can only tell us why a person’s usage of transit (or other modes) changes over time, but cannot tell us why person A uses public transit more than person B. Therefore, we first run independently pooled models, and then we run fixed-effect models to verify if the associations between modes still hold after controlling for individual effect. The random effect model is not suitable for our case, because it assumes that the individual effect is not correlated with the X_s (independent variables) in the models, and such an assumption is not met in our models.

RESULTS

Longitudinal Descriptive Statistics

Table 3 presents the total trip count and trip duration for each mode. The ACTION travel log has more than ten modes. We grouped the modes into six broader categories, as listed in the first column in Table 3.

Table 3.

Trip count and trip duration by mode

Category	Mode in ACTION travel log	ACTION	Trip count	% in all trips	Trip duration (min)	% in all trips
		Wave
Drive alone	Auto/ truck	1	11,099	62.3%	165,947	64.5%
		2	9,355	64.8%	143,351	68.8%
		3	8,868	61.2%	130,046	65.1%
	Motorcycle/ moped	1	94	0.5%	1,334	0.5%
		2	92	0.6%	1,491	0.7%
		3	67	0.5%	994	0.5%
	Total	1	11,193	62.8%	167,281	65.0%
		2	9,447	65.4%	144,842	69.5%
		3	8,935	61.7%	131,040	65.6%
Walking/ biking	Walk	1	4,815	36.3%	55,570	21.6%
		2	3,561	34.9%	35,133	16.9%
		3	3,789	37.2%	35,901	18.0%
	Bike	1	327	1.8%	6,434	2.5%
		2	289	2.0%	5,138	2.5%
		3	291	2.0%	5,400	2.7%
	Total	1	5,142	38.1%	62,004	24.1%
		2	3,850	36.9%	40,271	19.4%
		3	4,080	39.2%	41,301	20.7%
Carpool/Vanpool	Carpool/Vanpool	1	204	1.1%	3,189	1.2%
		2	143	1.0%	3,116	1.5%
		3	263	1.8%	5,285	2.4%
Public Transit	Bus	1	865	4.9%	17,578	6.8%
		2	607	4.2%	12,796	6.1%
		3	769	5.3%	16,051	8.0%
	Bus Rapid Transit	1	0	0.0%	0	0.0%
		2	43	0.3%	695	0.3%
		3	61	0.4%	1,014	0.5%
	Light rail	1	77	0.4%	1,612	0.6%
		2	92	0.6%	1,618	0.8%
		3	125	0.9%	1,913	1.0%
	Monorail/ trolley	1	39	0.2%	616	0.2%
		2	21	0.1%	412	0.2%
		3	19	0.1%	246	0.1%
	Heavy rail	1	32	0.2%	699	0.3%
		2	4	0.0%	37	0.0%
		3	1	0.0%	30	0.0%
	Total	1	1,013	5.7%	20,505	7.9%
		2	767	5.2%	15,558	7.4%
		3	975	6.7%	19,254	9.6%
Ride-hailing	Taxi/ shuttle bus/ limo	1	61	0.3%	787	0.3%
		2	68	0.5%	1,333	0.6%
		3	67	0.5%	1,301	0.7%
Others	Others	1	196	1.1%	3,432	1.3%
		2	161	1.1%	3,379	1.6%
		3	161	1.1%	2,073	1.0%
Total		1	17,809	100.0%	257,198	100.0%
		2	14,436	100.0%	208,499	100.0%
		3	14,481	100.0%	199,674	100.0%

Several trends can be observed. First, there had been growth in the percentages of ride-hailing trip count (from 0.3% to 0.5%) and trip duration (from 0.3% to 0.7%). These percentages aligned well with the National Household Travel Survey (0.5% in 2017) and were higher than those in PSRC Household Travel Survey conducted in the Seattle region (0.3% in 2017). However, the percentages were relatively small, suggesting that the new form of ride-hailing was still at the early stage of development. This finding is consistent with results from previous research that uses data from regional household travel surveys [14]. Second, Table 3 also shows that among ACTION participants, the shares of walking and biking, as well as transit, were high, 2 to 3 times above the national average [25]. Third, the share of driving peaked at wave 2, and then decreased at wave 3. Correspondingly, the share of public transit as well as walking and biking decreased at wave 2 and increased at wave 3. These trends suggest joint effects of the opening of BRT lines and other major transit improvements, and possibly the emergence of new mobility options.

Although ride-hailing’s share was relatively small, it almost doubled within six years in terms of trip count and more than doubled in terms of trip duration. If the new form of ride-hailing continues to grow at the current rate, it may play a more substantial role in the future. The next section will examine whether the changes in mode choices observed in the data can be partly attributed to the emergence of app-based ride-hailing.

Model Outcomes

Table 4 presents the summary statistics for the variables in the panel models that examine whether ride-hailing is substituting for or complementing other modes. The travel behavior variables are average trip count and duration per day per person. Some of the control variables show time-variance over the three waves. For example, the average number of children per household decreased while the average number of vehicles owned by households increased. Household income also increased. Over time, participants show a more positive attitude toward public transit, and less positive attitude toward driving. Using these data, we run both independently pooled models and fixed-effect models to examine the association between ride-hailing and other modes. Table 5 presents the results of the independently pooled models. Models (1) and (2) use the average count and duration per day by driving as dependent variables, Models (3) and (4) are for public transit, and Models (5) and (6) are for walking and biking. The estimations show changes in the trip count and trip duration that are associated with a one-unit change in the independent variables.

Table 4.

Summary statistics by ACTION waves

		Wave 1			Wave 2			Wave 3
		n	mean	s.d.	n	mean	s.d.	n	mean	s.d.
Travel behavior variables: average trip count or duration (min) per day	Driving count	460	3.09	2.00	371	3.25	2.12	361	3.11	2.21
	Driving duration	460	48.99	32.06	371	52.60	37.25	361	48.89	33.04
	Public transit count	460	0.26	0.61	371	0.26	0.58	361	0.34	0.70
	Public transit duration	460	6.09	16.74	371	5.16	12.57	361	7.22	16.68
	Walking and biking count	460	1.43	1.65	371	1.36	1.55	361	1.45	1.49
	Walking and biking duration	460	18.63	22.94	371	15.09	20.26	361	15.80	18.53
	Ride-hailing count	460	0.01	0.10	371	0.02	0.14	361	0.02	0.12
	Ride-hailing duration	460	0.20	2.02	371	0.50	3.09	361	0.55	3.81
	Total trip count	460	4.90	2.22	371	4.99	2.28	361	5.05	2.39
	Total trip duration	460	76.01	35.06	371	75.75	40.02	361	74.86	33.64
Other variables	Age	453	54.27	12.92	371	56.33	12.85	333	58.70	12.54
	Female (=1)	460	0.62	0.49	358	0.64	0.48	346	0.64	0.48
	White (=1)	460	0.88	0.33	358	0.89	0.32	346	0.88	0.32
	Education: less than college degree (=1)	424	0.30	0.46	368	0.26	0.44	317	0.26	0.44
	Education: college degree (=1)	424	0.37	0.48	368	0.40	0.49	317	0.40	0.49
	Education: graduate degree (=1)	424	0.33	0.47	368	0.35	0.48	317	0.34	0.47
	Household size	449	2.29	1.30	368	2.29	1.41	357	2.29	1.59
	Number of children under 18	450	0.46	0.89	363	0.44	0.89	359	0.38	0.81
	Number of vehicles owned by household	447	1.23	1.02	369	1.68	1.00	359	1.70	1.10
	Income: < $50k (=1)	455	0.38	0.49	371	0.35	0.48	360	0.34	0.47
	Income: $50k – $100k (=1)	455	0.38	0.49	371	0.35	0.48	360	0.35	0.48
	Income: >$100k (=1)	455	0.24	0.43	371	0.29	0.45	360	0.31	0.46
	Work full time (=1)	460	0.54	0.50	371	0.50	0.50	337	0.52	0.50
	Like taking public transit (1 = strongly agree or somewhat agree, 0 = neutral, somewhat disagree or strongly disagree)	456	0.43	0.50	359	0.45	0.50	359	0.46	0.50
	Like driving (1 = strongly agree or somewhat agree, 0 = neutral, somewhat disagree or strongly disagree)	454	0.56	0.50	365	0.54	0.50	358	0.49	0.50
	Residential unit density	456	15.94	11.95	369	17.09	14.24	357	17.26	14.16
Total N		460			371			361

Table 5.

Independently pooled model results

	Driving		Public transit		Walking and biking
	(1)	(2)	(3)	(4)	(5)	(6
	Count	Duration	Count	Duration	Count	Duration
Ride-hailing count	−0.36**	na	0.329	na	−0.079	na
Total trip count	0.155***	na	0.229***	na	0.193***	na
Ride-hailing duration	na	−0.04***	na	0.024	na	−0.01
Total trip duration	na	0.013***	na	0.023***	na	0.016***
Age	0.002	0.003	−0.003	−0.003	−0.002	0.001
Female	0.091**	0.149***	−0.145	−0.38***	−0.144**	−0.17*
White	−0.01	0.027	−0.48**	−0.491*	0.327***	0.305**
Education: college degree	0.008	0.089*	−0.209	−0.052	0.194***	0.155
Education: graduate degree	0.111**	0.222***	−0.32	0.097	0.079	0.037
Household size	−0.06***	−0.04*	0.037	0.243**	0.049***	0.045
Number of children under 18	0.114***	0.068*	−0.291***	−0.499***	−0.111***	0.024
Number of vehicles owned	0.141***	0.173***	−0.608***	−0.755***	−0.226***	−0.21***
Income: $50k – $100k	0.063	0.068	−0.166	−0.143	−0.103	−0.07
Income: > $100k	−0.002	−0.02	0.065	−0.308	0.015	0.048
Work full time	−0.01	0.038	0.231	0.389*	0.005	0.021
Like taking transit	−0.28***	−0.27***	1.314***	1.474***	0.464***	0.474***
Like driving	0.145***	0.169***	−0.337**	−0.283	−0.174***	−0.31***
Residential unit density	−0.01***	−0.01***	−0.001	0.01	0.008***	0.017***
ACTION wave 2	−0.06	−0.06	0.346**	0.484**	0.043	−0.11
ACTION wave 3	−0.09**	−0.05	0.355**	0.267	0.05	−0.02
Constant	0.198	2.534***	−2.036***	−0.339	−0.987***	0.97***
N	983	983	983	983	983	983
Log-Likelihood	−1,743.99	−4,567.02	−476.94	−1,655.11	−1,306.32	−3,437.44
AIC	3525.98	9172.04	991.89	3,348.21	2,650.64	6,912.88

Note: AIC = akaike information criterion, na = not applicable

^*p<0.1

^**p<0.05

^***p<0.01

Models (1) and (2) show that, respectively, when controlling for the count and duration of total trips and other control variables, more ride-hailing usage is significantly associated with less driving at 0.05 level. This is consistent with previous research findings that suggest ride-hailing can serve as a viable alternative to driving [10,19]. Total trip count and duration show an expected positive association with driving, as individuals with stronger travel demand tend to drive more. Other control variables also help explain the count and duration of driving. Females tend to drive more than males. It is likely that females share more household chores such as chauffeuring children or grocery shopping that often require driving. Higher educational attainment is associated with a greater likelihood to drive. Household composition is a significant factor, as shown by more driving by participants from smaller households, and households with more children. To elaborate, when controlling for the number of children, the household size variable captures the number of adults in the household. Therefore, the negative sign of the household size variable suggests that when there are more adults in the households, each individual adult needs to shoulder less household responsibility and drive less. Attitudinal variables show significant and strong effects. Participants who have a more positive attitude toward driving and a less positive attitude toward transit are more likely to drive. Participants who live in denser built environments are also less likely to drive. This may either indicate that a denser built environment allows individuals to access destinations with modes other than driving, or that self-selection is taking place where individuals who do not like to drive choose to live in denser neighborhoods. Lastly, the wave dummies capture the second and third waves' time-specific impacts on the entire sample. The significant negative estimation at wave 3 in the count model suggests a reduction in driving. It is possibly a reflection of the opening of the BRT lines. There were other infrastructure, service, and policy changes in the region during the time period, which all could have contributed to fewer driving trips. However, such an effect is not observed in the duration model, suggesting that participants tend to make fewer but longer driving trips at wave 3.

Models (3) and (4) both show a positive association between ride-hailing and public transit, which suggests complementarity between ride-hailing and public transit. However, this relationship is not statistically significant. This is different from the results reported in many previous studies based on convenience sample surveys. Many other variables that are significant in these two models are also significant in Models (1) and (2), but with opposite directions of effect, reflecting that driving and transit may be competing modes. One exception is the race variable, which is not a significant variable in the first two models. The negative association between being white and public transit usage suggests that the non-whites are more likely to rely on transit in the Seattle region, and thus transit serves as an important mobility option for them.

Models (5) and (6) suggest a weak and not significant negative association between ride-hailing and walking and biking. Again, many control variables show the expected directions of effect that are opposite to those in Models (1) and (2), because driving and active transportation modes are options typically taken by different population groups with distinctive demographic and socioeconomic characteristics. Racial differences also exist in the mode choices for walking and biking, indicating that being white is associated with more walking. Also, household size and the number of children are significantly related to walking and biking in the count model, but not in the duration model. It shows that individuals from smaller households and with more children tend to make fewer walking trips. Lastly, the positive estimation of residential density suggests that participants who live in denser neighborhoods are more likely to choose walking and biking.

Independently pooled models assume that observations are drawn independently and ignore serial correlation in the data. Therefore, we estimate additional fixed-effect models to test if the identified relationships between ride-hailing and other modes still hold after controlling for serial correlation. Table 6 presents the results, which show that how, at the participant level, a one-unit change in an independent variable would affect the dependent variable over time. The models include all variables shown in Table 5, except for those that are time-invariant. Table 6 displays the results for the travel behavior variables. Most control variables show the same directions of effect as in the independently pooled models and hence are not included in this table to save space.

Table 6.

Fixed-effect model results

	Driving		Public transit		Walking and biking
	(1)	(2)	(3)	(4)	(5)	(6
	Count	Duration	Count	Duration	Count	Duration
Ride-hailing count	−0.35*	na	−0.02	na	0.0	na
Total trip count	0.16***	na	0.01***	na	0.19***	na
Ride-hailing duration	na	−0.03**	na	−0.01	na	0.00
Total trip duration	na	0.01***	na	0.01***	na	0.01***
N	983	983	983	983	983	983
Log-Likelihood	−203.97	−2,379.63	−137.58	−400.88	−484.43	−1,649.65
AIC	435.94	4,787.26	301.15	829.76	996.85	3,327.30

Note: Variables in the control group were included in all models as in Table 5, except for those that are not time-variant, but are not shown here; AIC = akaike information criterion, na = not applicable

^*p<0.1

^**p<0.05

^***p<0.01

Table 6 confirms the key findings from independently pooled models. As shown in Models (7) and (8), ride-hailing is negatively associated with driving, and the relationship is statistically significant. These results tell us that for a specific participant, over time, an increase in the average daily ride-hailing trip count and duration is associated with a decrease in the average daily driving trip count and duration. No statistically significant association is found between ride-hailing and transit use or walking and biking in Models (9) to (12).

DISCUSSION AND LIMITATION

The estimated models show that cross-sectionally, participants who take more ride-hailing trips tend to make fewer driving trips. The association remains longitudinally in that for a specific participant, an increased number of ride-hailing trips over time is also associated with a decreased number of driving trips. This finding provides evidence that ride-hailing may be a viable substitute for driving. However, the models do not show any statistically significant impact of ride-hailing on public transit or walking and biking.

These empirical results are based on data that bundle traditional taxi and new (e.g., Uber and Lyft) forms of ride-hailing into a single variable. The inability to distinguish among Uber, Lyft, or taxi trips is a limitation because it is not possible to confirm changes in the use of the different ride-hailing options. However, it appears that the data effectively capture the new app-based ride-hailing trips for the following reasons:

1) Average daily count and duration of ride-hailing trips in the ACTION data doubled within the six years, which implies the growing presence of the new form of ride-hailing, as traditional taxi ridership first declined after the entry of the new ride-hailing services [26].

2) While it is possible that some participants enter their app-based ride-hailing trips into the survey’s last category ‘other’, the observed data suggest this is unlikely to be the case, because the mode share of ‘other’ did not increase during the study time period.

3) The shares of trip count and trip duration for ride-hailing in the ACTION data aligns with those of the National Household Travel Survey (NHTS) data, and are higher than those of the PSRC travel survey data. Ride-hailing options in NHTS and PSRC surveys include separate entries for trips taken by traditional taxi or by Uber/Lyft. This alignment is another indication that the ACTION data adequately capture both traditional and new ride-hailing options.

4) It should be noted that taxi companies in the Seattle region quickly became app-based after Uber and Lyft entry into the marketplace [26]. Since 2014, they function in a way that greatly resembles Uber and Lyft.

Finally, this study data come from populations living in transit-rich areas of a major metropolitan region. As a result, research findings cannot be generalized and further analyses will be required using data from different population groups and geographical areas.

CONCLUSIONS

This study uniquely attempts to apply longitudinal data to probe the impacts of ride-hailing, consisting of both taxis and app-based services, on other modes at the individual level. The findings deepen our understanding of the substitution and complementarity effects of new mobility services. The results suggest that ride-hailing substitutes for driving, but does not have a significant relationship with transit or walking and biking. In other words, these services so far generate rather limited direct impacts on the more environmentally sustainable travel options. The identified substitution effect of ride-hailing on driving is two-fold: first, among participants, those who use more ride-hailing tend to drive less; and second, over time, the increase of ride-hailing usage is also associated with fewer driving trips.

We believe that this study is informative for transportation professionals who want to better understand how new mobility options fit within the range of available mode choices. This knowledge is essential for policymaking in the era of app-based mobility sharing. Since ride-hailing shows a significant substitution effect on driving, it may create opportunity to reduce dependency on private automobile as the new form of ride-hailing becomes more prevailing. Planners and policymakers should actively monitor the changing usage of app-based mobility services, and be prepared to facilitate the transition. On the other hand, our study does not find a substitution effect of ride-hailing on public transit or walking and biking. This result suggests that transportation planners should explore the potential for integrating on-demand, app-based mobility options into urban transportation systems.

To gain a deeper understanding of the impacts generated by the growing use of app-based ride-hailing, existing survey collection methods need to be updated. Survey instruments should be modified to explicitly ask about the usage of app-based ride-hailing as a mode choice. In addition, longitudinal data should be collected for monitoring urban mobility trends, which can better inform future transportation planning and policymaking.