Available via license: CC BY
Content may be subject to copyright.
EJTIR Issue 17(3), 2017
pp. 360 −383
ISSN:1567 −7141
tlo.tbm.tudelft.nl/ejtir
Understanding future mode choice intentions of transit riders as
a function of past experiences with travel quality
Andre Carrel1
Dept. of Civil, Environmental and Geodetic Engineering, The Ohio State University, Columbus, USA
Joan L. Walker 2
Dept. of Civil and Environmental Engineering, University of California at Berkeley, Berkeley, USA
This paper empirically investigates the causes for transit use cessation, focusing on the influ-
ence of users’ personal experiences, resulting levels of satisfaction, and subsequent behavioral in-
tentions. It builds on a novel data set in which observed, objective measures of travel times are
mapped to smartphone-based surveys where participants assess their travel experience. An inte-
grated choice and latent variable model is developed to explain the influence of satisfaction with
operations (travel times) and satisfaction with the travel environment (e.g., crowding) on behav-
ioral intentions. Satisfaction is modeled as a latent variable, and the choice consists of participants’
stated desire and intention to continue using public transportation. The results show how delays,
in particular in-vehicle delays but also transfer times and being left behind at stops, contribute to
passengers’ intentions to cease transit use. Furthermore, a number of critical incidents, i.e., par-
ticularly memorable negative experiences, are found to have negative and significant effects on
overall satisfaction and on willingness to continue using public transportation. The usefulness of
the framework is demonstrated in a set of simulations in which the effect of three types of delays
on passengers’ willingness to remain transit riders is modeled. This work highlights the value and
potential of using new data collection methods to gain insights on complex behavioral processes,
and it is intended to form the basis for new modeling tools to understand the causes of transit use
cessation and the impact of various strategies and service quality improvements to reduce rider-
ship turnover.
Keywords: Latent Variable Choice Model, Mode Choice, Public Transportation, Rider Loyalty, Satisfaction,
Service Quality.
1 Introduction and motivation
Public transit is a key element to efficient and sustainable urban transportation, and in the past
decades, numerous public policies have been designed to increase its mode share in urban areas
through subsidies, service expansions, and land-use zoning. Yet, as is noted by Perk, Flynn, and
Volinski (2008), US transit agencies continue to see high levels of ridership turnover; in many cases,
a steady influx of new users into the system is offset by similarly high rates of transit use cessation.
On the individual level, these shifts are not trivial: As is explained by Vij, Carrel, and Walker (2013),
travelers tend to build their lifestyle around the use of certain travel modes, and decisions between,
for instance, an auto-oriented lifestyle and a transit-oriented lifestyle are relatively stable. In other
words, users who quit using a transit system are often unlikely to return unless a major upgrade to
the transit system is made.
1A: 470 Hitchcock Hall, 2070 Neil Avenue, Columbus, OH 43210, USA. T: +1.614.292.2771 E: carrel.20@osu.edu
2A: 111 McLaughlin Hall, Berkeley, CA 94720, USA. T: +1.510.642.6897 E: joanwalker@berkeley.edu
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 361
The fact that little is known about the reasons for which people shift away from transit-oriented
lifestyles is primarily due to a lack of suitable data. Authors have generally identified changes in
lifestyles associated with events such as marriage, or having children, as causes of transit use ces-
sation. However, Carrel, Halvorsen, and Walker (2013) identified a variety of negative experiences
with service quality (e.g., delays, high crowding levels) as further potential drivers. This paper
aims to quantify the effects of negative experiences on transit users’ future intentions of transit
use. It is based on the San Francisco Travel Quality Study data set, presented in section 3.1, and
uses a latent variable choice model to understand the link between their individual experiences,
satisfaction and future intentions.
2 Literature review
There are very few publications that have investigated transit passenger loyalty. The only studies
we are aware of that have done so in a generalizable fashion are by Trépanier, Habib, and Morency
(2012) and Ma et al. (2013). Both made use of automatically collected passenger data from smart
cards, but they did not link the observed usage patterns to riders’ experiences with the service.
This paper is concerned with making that link, and with describing the influence of individual
experiences on transit rider satisfaction and on future behavior. The framework laid out in this
paper has several components:
1. The link between individual experiences with transit service quality and reported levels of
satisfaction on a daily level.
2. The link between satisfaction on a daily scale and overall satisfaction reported at the end of
an extended period of time.
3. The link between overall satisfaction and future behavior.
The first item was the subject of a previous paper (Carrel et al., 2016) and will be expanded on in
this paper. The pertinent literature is presented in more depth there and is only summarized in the
paragraph below.
Satisfaction surveys are most valuable if a link can be made between satisfaction and objective
service quality measures (Davis and Heineke, 1998). So far, this has not been done in transit sat-
isfaction literature, with the exception of work by Friman and Fellesson (2009) and Carrel et al.
(2016). Only in the latter was the link made on an individual rather than an aggregate level. In
fact, customer satisfaction is a function of personal use experience (Woodruff, Cadotte, and Jenkins,
1983; Anderson and Sullivan, 1993), and in transportation, it is generally formed through multiple
repeated experiences over time. To control for memory distortions, the analyst needs to be knowl-
edgeable of the subject’s usage history and needs to limit the time frame covered by the satisfaction
survey (Fredrickson and Kahneman, 1993; Kahneman et al., 2004). In Carrel et al. (2016), several
separate ordinal logit models were estimated, linking satisfaction with individual travel time com-
ponents to observed travel times. It was found that the disutility of scheduled in-vehicle travel time
was much lower than in-vehicle delay time, and that in-vehicle delays appear to be strong drivers
of passenger dissatisfaction. Under certain circumstances, the latter might be perceived as more
onerous than out-of-vehicle wait time. Furthermore, it was found that the baseline satisfaction
with transit services and subjective well-being on the day of the survey were important covariates
in the measurement of daily satisfaction.
The second item is the link between satisfaction on a daily level and satisfaction reported at the end
of an extended period. It is recognized in the marketing literature that satisfaction is a dynamic
phenomenon and can change over time (Mittal, Kumar, and Tsiros, 1999), and that this change is
a function of personal experiences a decision-maker has made with the service or product in ques-
tion (Anderson and Sullivan, 1993; Davis and Heineke, 1998). This is consistent with the findings of
Kahneman et al. (1993), who found that subjects’ ratings of a repeated experience were dependent
on their history of previous experiences. Bates et al. (2001) extend this finding to the transportation
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 362
realm and argue that personal experience is very important in the context of travel time variability,
and that travelers will not choose a route based on average travel times on that route, but rather
on travel times they have experienced in the past - i.e., their personal travel time distribution. That
distribution is updated every time a person makes a trip, so a person’s satisfaction reported at the
end of an extended period should be a combination of their satisfaction at the beginning of that
period and their satisfaction with experiences during the period. Abou-Zeid et al. (2012) based an
experimental design aimed at capturing travelers’ subjective well-being before and after a week of
transit use on this notion, though satisfaction with daily experiences was not separately measured.
Furthermore, Abou-Zeid et al. (2012) argue that travelers may be less cognizant of their well-being
and satisfaction when they carry out routine behavior, such as choosing a car for commuting. This
is consistent with findings by Lanken et al. (1994). We propose a further distinction. The effect
described by Abou-Zeid et al. (2012) may be stronger when travel times and service quality are
consistent, and less pronounced when the experienced travel times or other service quality aspects
are variable, since delays or service failures can trigger negative emotions that influence future
decision-making. In public transportation, negative emotions caused by service failures can be ex-
acerbated by the passengers’ feeling of not being in control of their experience (Anable and Gater-
sleben, 2005). Therefore, we postulate that even with habitual transit riders, satisfaction levels can
change as a consequence of positive or negative experiences when service quality is variable. This
is supported by work by Friman, Edvardsson, and Gärling (2001), who found a measurable impact
of “critical incidents”, i.e., memorable positive or negative experiences, on customer satisfaction
with public transportation reported in a post-study survey.
The third item is the link between satisfaction with travel modes and future travel behavior. With
specific regard to public transportation, Pedersen, Friman, and Kristensson (2011) have inves-
tigated the influence of satisfaction on mode choice with a statistical path analysis and found
a positive association between remembered satisfaction and current choices. Satisfaction with
travel modes has also been successfully included in discrete choice models of mode choice (Abou-
Zeid and Ben-Akiva, 2010; Friman et al., 2013). In the application reported on in this paper, fu-
ture choices were not directly observed, so to link satisfaction with behavior, an alternative set of
variables was required. These variables were formulated in accordance with the Model of Goal-
Directed Behavior (MGB), a theoretical model of behavior change that is grounded in psychology
and the behavioral sciences. The MGB is a refinement of the widely used Theory of Planned Behav-
ior (TPB) (Ajzen, 1991). In the TPB, satisfaction is considered part of the set of attitudes toward the
behavior in question. Along with norms and beliefs, attitudes are linked to a person’s intention to
carry out a future behavior, which in turn leads to observed behavior. An example in which the TPB
has been applied directly to mode choice is Bamberg, Ajzen, and Schmidt (2003). The MGB adds
two elements to this framework: Anticipated emotions and behavioral desire. It postulates that
there are three steps to behavior change, as shown in Figure 1: First, a person develops a desire to
change behavior, followed by the formulation of an intention to change behavior. Lastly, the person
will actually change behavior. The development of a behavioral desire and the transition to the two
next steps are governed by a variety of factors. For more information on the model and influencing
factors, see Perugini and Bagozzi (2001). A discussion of the difference between behavioral desire
and intention can be found in Perugini and Bagozzi (2004). The Model of Goal-Directed Behavior
was chosen over the Theory of Planned Behavior due to the explicit recognition of emotions, as the
original study design included several questions on travelers’ subjective well-being. Following the
MGB, the outcome variables covered behavioral intentions as well as desire, in order to capture
transit users in both stages of the decision-making process and to approximate the (unobserved)
future choice as closely as possible.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 363
Figure 1. The Model of Goal-Directed Behavior (Perugini and Bagozzi, 2001).
3 Methodology
3.1 Data source
The data were collected during the San Francisco Travel Quality Study, a large-scale study of tran-
sit service quality, which ran from October 21 to December 22, 2013. The design and organization
of the study is described in detail in Carrel, Sengupta, and Walker (2016); what follows is a sum-
mary. The study involved an initial total of 856 participants recruited from the general public in
San Francisco. It focused on usage of the San Francisco Municipal Transportation Agency network
(commonly called “Muni” in San Francisco; this term is used in the remainder of the paper). As
an incentive, participants received a free one-month transit pass, valid for unlimited travel on the
Muni network. Participants were asked to complete an online entry survey, in which sociodemo-
graphic and mode access information was collected, as well as an exit survey at the end of the
study. All participants received the entry survey on October 21. Depending on when participants
received the transit pass, they were divided into two cohorts. Cohort 1 received the exit survey
on December 8; cohort 2 received it on December 22. Response times to the surveys varied. Af-
ter completing the entry survey, participants were asked to download a survey app for Android
phones. They were instructed to keep location services enabled on their phones, and if they did so,
the app collected location information from the phone every 30 seconds. Once per day at a time
set by the participants, generally in the evenings, they received a survey prompt on their phones
asking them whether they had used transit on that day. If they responded yes, they were presented
with a survey on their phone (hereafter called “daily mobile survey”) in which they were asked
to rate their satisfaction with the transit service they had experienced that day. Participants were
asked to use Muni on at least five days during the study period and fill out the corresponding daily
mobile surveys. The time window for responding to the daily mobile surveys was between Octo-
ber 27 and December 1 for cohort 1 and between October 27 and December 15 for cohort 2. Since
the daily mobile surveys were filled out only once per day, they referred to all transit trips made by
the participant on that day, regardless of the number.
This paper focuses on the link between experienced travel times, satisfaction, and future transit
use. The analysis builds on the following data collected during the study:
•Satisfaction with transit services: The online entry and exit surveys measured satisfaction
with nine variables, each on a five-point Likert scale from ’very dissatisfied’ to ’very satis-
fied’: Overall reliability, in-vehicle travel time, wait time at the origin stop, transfer time (if
applicable), crowding, cleanliness, safety, pleasantness of other passengers, and the accuracy
of real-time information. The prompt in these surveys asked respondents to rate their overall
satisfaction with Muni services. In the daily mobile surveys, respondents were asked to rate
their satisfaction with the same nine variables, but only with respect to the transit service they
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 364
had experienced that day. An exploratory factor analysis on the results confirmed that there
were strong correlations within two groups of variables: The first four can be summarized as
satisfaction with operations, whereas the following four can be thought of as satisfaction with
the travel environment.
•Experienced travel times: During the study, the vehicle locations of all transit vehicles through-
out the city were continuously recorded. The phone location data from the participants were
then matched to the transit vehicle location data to automatically identify whether the par-
ticipant had used transit on that day and to extract wait times, in-vehicle travel times and
transfer times. The measured travel times were then compared to timetable information to
identify deviations from the timetable. This methodology provided objective measurements
of the transit travel times and delays experienced by the participants. An in-depth descrip-
tion of the methodology can be found in Carrel et al. (2015). For the purposes of the model
presented in this paper, since the satisfaction surveys concerned an entire day, participants’
travel times were aggregated on a daily level.
•Future transit use: In the entry and the exit surveys, the participants were asked a set of
questions regarding their future transit use. In accordance with the MGB, they were asked
about their behavioral intentions and their behavioral desires, with the following question
prompts:
◦Question 1: “In 2014, do you intend to use [Muni] more or less than you do now, or the
same way as you do now?”
◦Question 2: “Ideally (regardless of whether you intend to do so), do you want to be able
to travel in San Francisco by [Muni] more or less than you do now, or about the same
way as you do now?”
After the entry survey had been distributed, it became apparent that the formulation of these
questions was not optimal. Some respondents were confused by the difference between be-
havioral intention and desire, and others stated that they did not know in advance what
their mode choices during an entire year would be. Therefore, five additional questions were
added in the exit survey:
◦Question 3: “Compared to how often you used Muni in the month before the study, you
anticipate using Muni in January...”. Responses were on an 8-point Likert scale from
“not at all anymore” to “much more”. (This measured short-term intention compared to
before the study)
◦Question 4: “Compared to how often you used Muni during the study, you anticipate
using Muni in January... ”. Responses were on an 8-point Likert scale from “not at all
anymore” to “much more”. (Short-term intention compared to during the study)
◦Question 5: “Compared to how often you used Muni during the study, you would prefer
to use Muni in January...”. Responses were on an 8-point Likert scale from “not at all
anymore” to “much more”. (Short-term desire)
◦Question 6: “As soon as my circumstances permit, I would like to use public trans-
portation more”. Responses were on a 5-point Likert scale from “strongly disagree” to
“strongly agree”. (Long-term desire)
◦Question 7: “As soon as my circumstances permit, I would like to use public transporta-
tion less”. Responses were on a 5-point Likert scale from “strongly disagree” to “strongly
agree”. (Long-term desire)
The five added questions incorporate three changes: First, the time frame for the intention
questions was shortened from one year to one month, since it was assumed that participants
would have a clearer sense of their mode use in the month following the study. Second, the
word “intend” was removed, and instead, participants were simply asked how they were
going to travel in January. Third, the cumbersome formulation “...do you want to be able
to...”, which was the original measure for desire, was replaced with “...would you prefer
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 365
to...”. After the exit survey was distributed, no emails were received asking for clarification
on these questions, suggesting they were better understood than the original set. Participants
were also presented with a list of possible reasons for their response to question 5 and were
asked to rate the level of influence of every item on their behavioral desire. A descriptive
analysis of the results is provided in section 4. Lastly, it should be noted that the question on
short-term desire was only asked in comparison with the respondents’ transit use during the
study, but not in comparison with their transit use before the study. This was an unfortunate
oversight on the part of the survey designer.
The questions on future transit use were intended as the set of outcome variables. In the original
survey design, it was planned to use questions 1 and 2 to measure changes in behavioral desire
and intentions during the study. The responses to the entry survey would have provided a base-
line. When questions 3 through 5 were added, it was no longer possible to measure a baseline, so
instead, they were formulated as comparisons such that the baseline was self-reported. The ex-
plicit distinction between behavioral desire and intention was introduced in the surveys to ensure
consistency with the MGB. In total, the data set used for developing the model presented here in-
cluded 449 respondents who had filled out the entry and exit surveys, and from whom at least one
daily mobile survey response and relevant travel time data was recorded. More information on the
characteristics of the study population can be found in Carrel, Sengupta, and Walker (2016).
3.2 Satisfaction response patterns
Before introducing the model, we will first consider an interesting observation with respect to the
satisfaction reported in the exit survey. All participants were asked to complete the online exit sur-
vey, but in addition to that, an optional mobile exit survey was also distributed. Its formatting was
identical to the daily mobile surveys, with the exception that the survey prompts asked respondents
to indicate their satisfaction with their overall Muni experience during the study. Participants were
asked to fill out the mobile exit survey in addition to the online exit survey, but it was made clear
that that was not mandatory. The 5-point response scales in the daily mobile surveys were labeled
only at the maximum and minimum with a frowny and a smiley face due to space constraints,
where as the scale in the online survey was labeled with words since the online survey engine did
not allow graphical labels. A distinct difference was noticed between users’ responses with respect
to their daily satisfaction and their overall satisfaction in the online and mobile exit surveys. In
Figure 2c, the distributions of responses to the nine satisfaction items in the daily mobile survey
are shown. It should be noted that this sample includes multiple responses per participant, and
no correction has been made to account for correlation between responses from the same person.
For each item, the darkest bar shows the proportion of “very dissatisfied” responses and the light-
est bar shows the proportion of “very satisfied” responses, with the remaining bars showing the
responses in between. Figures 2a and 2b show the distributions for the online exit survey and the
mobile exit survey, respectively. These samples include only one observation per participant.
Whereas users were willing to state that they were “very satisfied” with their daily experiences on
public transportation, it can be seen that:
•Participants were less willing to state that they were “very satisfied” in the exit surveys
•This effect was more pronounced in the online exit survey than in the mobile exit survey.
The sample size for the mobile exit survey was smaller than the online exit survey since the former
was optional, but all participants who took the mobile exit survey also took the online exit survey.
We propose four possible reasons for the observed discrepancies:
1. The service quality experienced by study participants between the end of the daily survey
prompts and the time they filled out the exit surveys was markedly worse than the service
quality on the days for which daily surveys were filled out. While possible, this explanation
is not plausible.
2. The different time frames to which the questions are referring to: When asked about their
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 366
(a) Online exit survey (N = 482) (b) Mobile exit survey (N = 353)
(c) Mobile survey (N = 6846)
Figure 2. Distribution of survey responses. RTI stands for Real-Time Information.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 367
overall satisfaction, participants may recall negative events that occurred before the study
period.
3. The different presentation of the questions, i.e., the fact that the labels differed between the
mobile and the online survey versions. While this may account for some of the differences
between the online and the daily mobile survey responses, the fact that the mobile exit survey
response patterns also differ from the daily mobile survey response patterns indicates that
this cannot be the only factor at play.
4. The different environments in which the surveys may have been filled out. It is more likely
that the online survey was taken by people at home or at the office, whereas the daily mobile
survey may have been taken anywhere. However, one could argue that the mobile exit survey
may also have been taken anywhere.
The definitive reason for this discrepancy cannot be elicited without further investigation. Until
that is possible, researchers designing future studies should be aware that the medium through
which the survey is delivered (in this case, smartphone vs. online survey engine) can have an effect
on the response patterns. The fact that the “very satisfied” category in the online exit survey had
very few responses would have introduced nonlinearity in the latent variable measurement model
and caused estimation problems since the measurement equations assume a linear relationship
between the latent variable and the indicator variables. To avoid these problems, the satisfaction
ratings for the online exit survey were re-scaled to a four-point scale where “satisfied” and “very
satisfied” were included in one category.
3.3 Model development
Carrel et al. (2016) described the link between experiences and satisfaction in a static context and
considering the individual travel time components separately. This paper extends that work in sev-
eral ways and embeds it in a broader modeling framework. The primary purpose is to link personal
experiences to future behavior via the intermediate construct of satisfaction. In this case, the over-
all satisfaction with travel times and operations was of interest, and not the satisfaction with the
individual travel time components. Following the exploratory factor analysis described in section
3.1, we assume that there are two underlying and unobserved latent satisfaction variables - satis-
faction with operations and with the travel environment. The observed variables, i.e., the recorded
responses, are indicators of that underlying latent variable. To make the links between satisfaction
in the entry survey, experiences during the study, reported daily satisfaction, satisfaction in the
exit survey and future mode choice behavior, a latent variable choice model (Walker, 2001) was
developed, as shown in Figure 3. In the figure, ellipses denote latent variables, rectangles denote
observed variables, and the arrows show the directionality of effects being modeled.
The latent entry and exit satisfaction are shown as “entry satisfaction with operations” and “exit
satisfaction with operations” with the respective indicator variables I1through I6. The indicator
variables were the reported satisfactions with the in-vehicle travel time, wait time and overall re-
liability. The daily satisfaction was modeled using the same latent variable constructs; those are
shown as d1through dr. The indicator variables for the daily satisfaction are omitted in the figure
due to space limitation, but every daily satisfaction item had four indicator variables which were
the aforementioned three satisfaction measurements plus satisfaction with transfer time. There
were five variables for daily satisfaction: the four most recent responses for which travel time data
were available, labeled d1through d4, plus a fifth variable including the average of all remaining
daily observations, labeled dr. This structure was chosen since there were variable numbers of
responses per participant. Every participant with at least one daily mobile survey response was
included in this data set. The structural model for satisfaction with operations reflects the temporal
dependencies between the individual surveys: The daily satisfaction ratings are influenced by sat-
isfaction reported in the entry survey as well as by travel times experienced on that day and by the
user’s general feeling on that day. The daily responses in turn feed into the satisfaction reported in
the exit survey. The assumption is made that the exit satisfaction depends on the entry satisfaction
only by way of the daily satisfaction. All coefficients relating the daily latent satisfaction variables,
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 368
d1through d4and dr, to the exit satisfaction are constrained to be the same. The coefficients of the
measurement equations of d1through d4are also constrained to be the same, but the coefficients
for the measurement equation of drare allowed to differ to reflect the fact that dris averaged over
a number of days. In addition to satisfaction with operations, the exit satisfaction with the travel
environment is included as a separate latent variable, labeled “exit satisfaction with environment”.
The indicator variables for the latter are the reported satisfaction with crowding, cleanliness, safety
and other passengers in the exit survey. No objective measurements were available for the travel
environment. Both latent variables feed into the utility for the choice model.
In addition, several variables on negative critical incidents during the study which were self-
reported in the exit survey were included as explanatory variables affecting the exit satisfaction
directly. These were:
•The number of times a participant arrived late at work or school (“Late arrival at work”).
•The number of times a participant arrived late at a leisure activity (“Late arrival at leisure”).
•The number of times a participant reported that he or she wanted to use Muni but was not
able to because of a delay on the system (“Could not travel due to delay”).
•The number of times a participant was left behind at a stop because the vehicle was full (“Left-
behind”).
In Figure 3, the Greek letters denote groups of coefficients corresponding to the notation in Table 1.
Different letters are assigned to different groups of coefficients to improve readability of the model.
As explained at the beginning of this section, the original outcome variables on behavioral intention
and desire proved to be problematic. Nonetheless, the exploratory data analysis and first model
specifications used the responses to those questions. Since they were asked both in the entry and
the exit survey, they would have provided a more objective measure of changes in participants’
intended/desired future behavior. Unfortunately, we found that the data for these two questions
contained a lot of noise and showed little correlation, both between the entry and exit responses
and between the responses and other variables. Therefore, three of the alternative variables that
had been added in the exit survey were used: Questions 3, 5, and 6. These three variables served as
(imperfect) indicators of the latent choice that was of interest but that could not be observed (Bollen,
2014): Whether or not a person was going to use public transportation less in the future. Though
the responses were recorded on five-point Likert scales, the indicator variables were reduced to a
binary choice between desiring/intending to use Muni less and desiring/intending to use it the
same or more. This was done to reduce model complexity.
Question 3 contains a self-reported baseline and was formulated to ask the respondent about dif-
ferences between pre-study transit use and post-study transit use. It can therefore be considered a
reasonable replacement for question 1. In the figure, it is labeled as “change in intentions”. Ques-
tion 5 also contains a baseline, though it is with respect to transit use during the study rather than
before. However, as 93% of all survey respondents used Muni more than 2 days per week before
the study (see Carrel, Sengupta, and Walker (2016) for details), negative responses such as “[I will]
not use Muni at all anymore” or “[I will] use Muni much less” can nonetheless be seen as indicative
of decreases compared to the pre-study baseline. Although the different baselines may reduce cor-
relations between the two variables, it was decided that this was outweighed by the ability to have
indicators for both short-term intention and desire. Question 6, on the other hand, was formulated
as a statement with which respondents could disagree or agree. This did not require them to make
a comparison with a baseline, which had the advantage that even if respondents did not know
exactly how much they would use transit in the far future, they could express a general sentiment.
Questions 4 and 7 were not included due to the strong similarity to questions 3 and 6.
The model specification assumes a set of causal relationships as shown by the arrows in Figure
3. Most importantly, it assumes that the participants’ desire and intention to stop using transit is
primarily a function of their satisfaction with transit services reported in the exit survey, which in
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 369
Figure 3. Model structure.
turn depends on the quality of transit service experienced during the study and their reported sat-
isfaction during the study. Their sociodemographic characteristics and entry satisfaction influence
the satisfaction reported during the study, but not the exit satisfaction directly. In other words, it
is assumed that experiences during the study were significantly more important to participants’
behavioral intentions and desires than experiences before the study, such that any direct influence
of the latter on the latent choice can be disregarded. In addition, because participants were not
required to fill out a daily mobile survey after every day on which they used Muni (as long as they
submitted the minimum number), the estimation results can only be interpreted properly if it is
assumed that the experiences reported through the daily mobile surveys are representative of the
participants’ average experiences during the study. We recognize that these are limitations of the
model and the data, as is further discussed in section 3.4.
In what follows, the model specification is presented. The interpretation of all coefficients used in
the model specification is shown in Table 1. The structural equation for the entry satisfaction with
operations was:
(1)Satentry,ops =γentrymean_ops +ηentry ·ωentry
The structural equation for the daily satisfaction with operations (denoted diin Figure 3) was:
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 370
Satdaily,o ps =αa ge ·age+αincom e ·income+αunknownincome ·unknown_income+αl onguser ·longuser+αentrysat
·Satentry,ops +αmood ·mood +αivtt ·ivtt +αearly ·e arly +αdelay ·delay +αwaito ver5·waitover5
+αwaitunder5·waitunder5 + αnowait ·no_wait +αtran s f ertime ·tr ans f er +αunobser vedtra ns f er
·unobserved_tra ns f er +αno trans f er ·no_tr ans f er +αle f tb ehind ·l e f t_behind +ηdaily ·ωdail y
(2)
The structural equation for the exit satisfaction with operations was:
(3)
Satexit,o ps =∑βdailysat ·Satdaily,o ps +βl e f tbehi nd_1_9 ·le f tbehind_1_9 + βle f tb ehind_10
·le f t behind_10 + βlatework ·latework +βl ateleisure ·l ateleisure
+βnotravel ·notravel +ηexitops ·ωexito ps +ηerrorcorr ·ωerrorcorr
The summation term above is the summation over all latent daily satisfaction variables. The struc-
tural equation for the exit satisfaction with operations was:
(4)Satexit,env =ηexitenv ·ωexitenv +ηerrorcorr ·ωerrorcorr
And finally, the choice model was:
(5)
V= (µshortint +µshortdes +µlongdes)
·(ASCshortint +ASCshortdes +ASCl ongdes +γops ·Satexit,o ps +γenv ·Satexit,env)
The µand alternative-specific constant (ASC) terms above are specified such that they only enter
into the equation if the choice being modeled relates to the respective outcome variable, and they
are zero otherwise. The measurement equations all had the same functional form. For example, the
conditional probability for the indicator “satisfaction with in-vehicle travel time” (IVTTSat_entry)
of the entry satisfaction with operations is:
(6)
P(IV TTSatentry |IIVT T ,δIVTT,entry ,λIVT T,entry) = 1
σIVTT,entry
·φ(IV TTSatentry −δIVT T,entry −λIV T T,entry ·IIV TT
σIVTT,entry
Table 1. List of coefficients of the structural equations.
Coefficient Meaning
ηentry Error term (entry survey satisfaction with operations)
αage Age (in year brackets)
αincome Income (10,000 USD brackets)
αunknownincome Unknown income (Binary)
αlon guser Long-time user (Binary: System user > 2 years)
αentrysat Entry satisfaction with operations (4 pt. Likert)
αmood General mood (5 pt. Likert)
αivtt In-vehicle travel time (Minutes)
αearl y Early arrival at destination stop (Minutes)
αdelay Late arrival at destination stop (Minutes)
αwaitover5Wait time greater than 5 minutes
αwaitunder5Wait time less than or equal to 5 minutes
αnowait No wait time inferred from location data (Binary)
Continued on next page
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 371
Coefficient Meaning
αtran s f ertime Transfer time (Minutes)
αnotr ans f er No transfer inferred from location data (Binary)
αle f tb ehind Denied boardings (Inferred from location data)
αunob servedt rans f er Transfer reported but
not observed in location data (Binary)
ηdail y Error term (daily mobile satisfaction with operations)
βdail ysat Daily satisfaction with operations (5 pt. Likert)
βle f tb ehind_1_9 Between 1 and 9 denied boardings (Self-reported)
βle f tb ehind_10 10 or more denied boardings (Self-reported)
βlatework Arrived late at work or school (Self-reported)
βlatelei sure Arrived late at a leisure activity (Self-reported)
βnotravel Wanted to use Muni but
could not due to delay (Self-reported)
ηexit op s Error term (exit survey satisfaction with operations)
ηerrorcorr Error term correlation
ηexit env Error term (exit survey satisfaction with the travel environment)
ASCshortint ASC intended Muni use short-term
ASCshortdes ASC desired Muni use short-term
ASClongd es ASC desired Muni use long-term
γenv Coefficient for exit satisfaction with the travel environment
γops Coefficient for exit satisfaction with operations
µshortint Scale parameter intended Muni use short-term
µshortdes Scale parameter desired Muni use short-term
µlon gdes Scale parameter desired Muni use long-term
3.4 Limitations
The model is subject to a few limitations which are discussed here. First, the choice indicator
variables were only measured in the exit survey. Two of the three indicator variables referred to
self-reported baselines, one of which was relative to the traveler’s behavior prior to the study and
one of which was relative to behavior during the study. The three indicators covered different time
scales and different stages of the decision-making process. This design was inspired by the MGB,
a behavioral theory that is a theoretical model, in the hopes of developing indicators that were as
correlated as possible with the future outcome. Nonetheless, the model cannot determine whether
there is a causal relationship between the events during the study, the indicator variables, and the
choice. The assumptions stated in section 3.3 are necessary to interpret the correlation observed
in the data as a causal relationship, but any interpretation of the results should be done with this
caveat in mind.
The uncertainty regarding whether the observed correlation truly represents causality could have
been reduced by a control group. Even though a control group was not available, the assumption
that other, external influences on the choice (aside from experiences with service quality) could
largely be ignored still appears reasonable for several reasons: The study covered less than two
months, and it was carried out in San Francisco, where winter weather is relatively stable and
temperate. We are not aware of any notable events during the study that could have impacted
transit use. Participants were asked about major lifestyle changes or moving plans, and those
factors were controlled for. Finally, as will be shown in the results, the estimated coefficients were
statistically significant.
The model estimation was done via maximum likelihood procedure. If the model were misspeci-
fied, i.e., if it omitted variables relevant to the choice or if the causal relationships were incorrect,
this would lead to biases in the coefficient estimates, though it is not possible to quantify any po-
tential biases, as this would require knowledge of the true model (Bollen et al., 2007). In a latent
variable model, structural misspecification in one part of the model can cause biased coefficient
estimates in correctly specified parts of the model as well. Furthermore, a misspecification would
also impact the efficiency of the estimation technique and the accuracy of the hypothesis tests.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 372
Table 2. Cross-tabulation of two questions regarding cessation of transit use.
In January, I would like to use transit...
less the same more Sum
“As soon as my circ- Agree 96 81 13 190
umstances permit, I Neutral 42 80 32 154
want to use transit less” Disagree 49 210 84 343
Sum 187 371 129 687
A second limitation is that the coefficients of the structural model relating the daily satisfaction
to the exit satisfaction needed to be constrained to be equal, due to estimation difficulties if they
were unconstrained. As a consequence, all daily satisfaction ratings were weighted equally. One
possible cause is that participants’ satisfaction was not measured at the same times with respect
to the time of the exit survey, as they were only required to give five responses over the course
of the study. By specifying separate coefficients for each of the past satisfaction ratings, it was
hoped that a time dependency would be observable, for example, that the most recent experience
would have a stronger influence than more distant experiences. The failure to observe such a
time dependency in this model does not mean it is not present, but is more likely due to the data
limitations. In future research, satisfaction should be measured at the same time intervals between
the daily mobile surveys and the exit surveys for all participants, and the model should be re-
estimated with those data to discover potential time dependencies. In addition, if it is possible to
sample satisfaction for all participants on consecutive days, the data would allow the observation
of serial dependency between measurements of different days.
4 Results
4.1 Descriptive analysis
Table 2 shows a cross-tabulation of the responses to two questions: “Would you prefer to use Muni
less/the same/more in January 2014” and “As soon as my circumstances permit, I would like to
use public transportation less” (responses to the latter were levels of agreement). Out of the 687
participants who filled out the exit survey, 187 stated that they would prefer to use public trans-
portation less in January, and they were subsequently asked for the reasons for that statement. Out
of those 187, 96 also agreed with the statement that they would like to use public transportation less
as soon as their circumstances permitted. Figure 4 shows the stated reasons on a scale from “not
at all influential” (1) to “very influential” (5). It can be seen that only 17 out of the 96 participants
stated that negative experiences during the study did not influence their desire to reduce their use
of transit, and 16 said such experiences were slightly influential. The remaining 63 were split evenly
between “somewhat influential”, “moderately influential” and “very influential”. Among the spe-
cific reasons mentioned, the most important ones were overall unreliability, crowding levels, wait
time unreliability and unreliability of in-vehicle travel times. Unreliability of transfer times was
mentioned less frequently, but that was in part due to the fact that not all participants transferred.
Unreliability and crowding levels are of course linked due to bus bunching (Daganzo and Pila-
chowski, 2011). It can also be seen that out of the other environmental variables, cleanliness, safety,
comfort, the friendliness and competence of staff and the pleasantness of other passengers were re-
ported to be much less influential than crowding. Travel times and service frequencies when there
are no delays were asked about separately, and as can be seen, were reported to be less influen-
tial by participants than travel time reliability variables. Lastly, the least influential variables were
related to the cost of travel and the fare payment system.
For participants who responded to the question “Compared to how much you used Muni during
the study, you anticipate using Muni in January...” either by saying that they were going to increase
or decrease their use of Muni, a follow-up question was asked regarding their anticipated mode
shifts. Participants who said they anticipated decreasing their Muni use were asked what modes
they were going to shift their trips to, and participants who said they anticipated increasing their
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 373
(a) In order from left to right: Negative experiences during the study, positive experiences during the study, overall reliability,
wait time reliability, travel time reliability, transfer time reliability, crowding, cleanliness, safety, pleasantness of other
passengers.
(b) In order from left to right: Ability of Muni to meet daily travel needs, cost, on-board travel times when there are no delays,
frequencies of service, comfort, friendliness of staff, competence of staff, accuracy of real-time information, reliability of fare
payment system, ease of use of fare payment system.
Figure 4. Stated reasons for wanting to use transit less or not at all anymore
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 374
Figure 5. Decrease in Muni use: Travel modes substituted for Muni trips. (1 - applies to none of the trips, 5 -
applies to all trips.)
Figure 6. Increase in Muni use: Travel modes substituted with Muni trips. (1 - applies to none of the trips, 5 -
applies to all trips.)
Muni use where asked what modes they were shifting their trips from. The results are shown in
Figures 5 and 6. “None” refers to trips that the participant did not make before or would cease
making when the shift to or from Muni occurred. It can be seen that increased Muni use drew
mostly from walk trips, followed by trips that were not made before and auto trips. On the other
hand, people who decreased their Muni use primarily either ceased making those trips, began
using the automobile or walking.
4.2 Modeling results
The parameters of the structural model are explained in Table 1. The estimation results for the
structural model are found in Table 3, and the estimation results for the measurement models are
in Table 5 in the appendix. Based on a mix of theoretical considerations regarding model identifi-
ability (Bollen, 2014; Ben-Akiva et al., 2002) and empirical work with the data, it was determined
that several of the coefficients needed to be constrained in order to ensure identification. Generally
speaking, the constrained value can be freely chosen by the modeler, but a value of 1 is a common
approach to support the interpretability of the remaining coefficients. This concerned two of the
error terms (ηdail y,ηentry). Furthermore, one of the scale parameters of the choice model and one
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 375
of the coefficients of each measurement model for latent satisfaction in the exit survey had to be
constrained to set the scale for the remaining coefficients. µl ongdes and two of the λcoefficients were
chosen for this and set to 1. These coefficients are marked by asterisks in Tables 3 and 5. While
the estimation was conducted, it was observed that the likelihood function along the dimension
of the ASC for short-term intention appeared to be very flat, which caused erratic coefficient esti-
mates for that ASC. Due to this empirical identification issue, the ASC for short-term intention was
constrained to 1 as well. As in all cases where coefficients were constrained, we verified that the
constraint did not impact the fit of the model, i.e., that the final value of the likelihood function
remained approximately the same when the constraint was introduced. This ensured that the nor-
malizations stabilized the parameter estimates but did not impact the fit and interpretation of the
model.
Table 3. Estimation results for the structural equations of the latent variable choice model.
Coefficient Value Robust Std. Error Robust t-stat p-value
ηentry 1.00 * * *
αage 0.01 0.00 1.64 0.10
αincome -0.01 0.01 -0.80 0.42
αunknownincome -0.02 0.65 -0.04 0.97
αlon guser -0.25 0.17 -1.47 0.14
αentrysat 1.02 0.16 6.54 0.00
αmood 0.38 0.08 5.02 0.00
αivtt 0.00 0.01 0.18 0.85
αearl y 0.01 0.06 0.24 0.81
αdelay -0.20 0.04 -4.88 0.00
αwaitover5-0.02 0.03 -0.74 0.46
αwaitunder50.01 0.03 0.55 0.58
αnowait -0.28 0.09 -3.04 0.00
αtran s f ertime -0.05 0.01 -3.28 0.00
αnotr ans f er -0.21 0.12 -1.83 0.07
αle f tb ehind -0.12 0.10 -1.20 0.23
αunob servedt rans f er -0.51 0.19 -2.70 0.01
ηdail y 1.00 * * *
βdail ysat 0.10 0.01 8.54 0.00
βle f tb ehind_1_9 -0.10 0.06 -1.85 0.06
βle f tb ehind_10 -0.29 0.12 -2.45 0.01
βlatework -0.06 0.01 -4.48 0.00
βlatelei sure -0.04 0.02 -2.75 0.01
βnotravel -0.07 0.03 -2.59 0.01
ηexit op s 0.42 0.07 6.51 0.00
ηerrorcorr 0.16 0.18 0.88 0.38
ηexit env 0.57 0.05 12.56 0.00
ASCshortdes 1.07 0.12 9.06 0.00
ASClongd es 0.51 0.14 3.68 0.00
ASCshortint 1.00 * * *
γenv 0.23 0.10 2.23 0.03
γops 0.39 0.10 3.89 0.00
µshortint 1.78 0.18 9.95 0.00
µshortdes 2.33 0.69 3.38 0.00
µlon gdes 1.00 * * *
In what follows, the model estimation results are presented and discussed. For all variables that
influence satisfaction, positive coefficient estimates mean that an increase in the respective vari-
able leads to increased satisfaction, whereas negative coefficient estimates mean that an increase
in the variable leads to decreased satisfaction. When comparing these results to those in Carrel
et al. (2016), one should be cognizant of the fact that the assumptions underlying this model are
different from those underlying the models in Carrel et al. (2016). In the latter, the link between
the individual travel time components and satisfaction with those components were modeled. No
assumptions were made about the relationship between the models, and it was assumed that the
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 376
reported satisfaction ratings were true measures of the participant’s satisfaction. On the other hand,
the model results presented here assume that there is an underlying, latent satisfaction with oper-
ations (and thus, travel times). The individual satisfactions with travel time components serve as
indicators of that latent underlying satisfaction.
4.2.1 Effects of sociodemographics, baseline satisfaction and mood
First, we investigate the effect of sociodemographic attributes on people’s reported daily satisfac-
tion. It can be seen in Table 3 that age has a positive effect (0.01 per year of age) and is significant,
whereas having been a long-time user (αlon guser) has a negative effect (-0.25). The latter, however,
is not significant. The effect of income on reported daily satisfaction is also negative (-0.01 per
$10,000), as is the non-response variable to income. The results with respect to income are intu-
itive, as higher income is associated with a higher value of time. Both effects, however, are not
significant at p= 0.42 and p= 0.97, respectively. The effect of the entry satisfaction (αentrys at = 1.02),
which again is a latent variable, and of the participant’s general mood on the day of the survey
(αmood = 0.38) are positive and significant at p<0.01. In other words, these two variables have a
markedly stronger effect on satisfaction than age, income and the length of Muni use. Of course we
would expect the entry satisfaction to depend on sociodemographic variables as well, but it was
not possible to include the mood and entry satisfaction in both the daily and the entry survey as
this caused estimation problems. These results are in line with findings from Carrel et al. (2016),
though the difference between the age and income variables and the mood and entry satisfaction
variables is more pronounced in the latent variable model.
4.2.2 Travel time variables
The scheduled in-vehicle travel time (IVTT) is found to have virtually no effect on overall satisfac-
tion with operations (αivtt). Delays with respect to the scheduled IVTT have a significant negative
effect (αdelay =−0.20 per minute, p<0.01) and earlier arrivals have an insignificant effect on sat-
isfaction (αdelay = 0.01 per minute, p= 0.81). Unfortunately, this model was not able to capture the
effect of wait times on a joint, latent satisfaction variable, as both coefficients for wait times below
5 minutes (αwaitunder5) and wait times above 5 minutes (αwaitover5) are insignificant. The reason for
this merits further investigation, but as discussed in Carrel et al. (2016), it might be linked to the
fact that the majority of wait times was short, with an average around 2 minutes. It must be noted
that the observed wait times used in the model estimation only capture time actually spent stand-
ing at the origin stop. They exclude additional schedule delay times (i.e., the time between when a
person wanted to leave and the time of a transit vehicle departure), which some participants may
have chosen to spend elsewhere, such as in their homes. While these may have been perceived as
wait times by some participants, it was not possible to automatically identify them with location
tracking data from the phones, and therefore, the link between the latent satisfaction with travel
times and the wait times could not be established in this case. As passengers rely more on real-time
information, the strategy of spending wait time at locations other than the stop and of going to the
stop only when an arrival is predicted will most likely become more prevalent.
Wait times could not be identified from the tracking data alone for approximately 50% of observa-
tions in the data set. The observations with missing wait times are denoted by a binary variable,
the coefficient of which (αnowait) is negative and significant. To the best of our knowledge, a missing
wait time observation could be due to one or more of the following causes:
1. If there was insufficient location tracking data available for that portion of the trip. This
includes both smartphone and vehicle location data.
2. If the tracking data showed the participant getting on a vehicle immediately, with no time
spent at the stop.
3. if the participant was carrying out an activity near the stop (e.g. work, shopping) which made
it impossible to distinguish activity time from wait time. This distinction was particularly in
parts of San Francisco where the transit network is very dense.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 377
4. If the wait time was incurred when the participant transferred from BART (regional rapid
transit) to a local metro train inside an underground metro station.
It is not known whether the missing observations skewed the distribution of observed wait times
in any particular way. The insignificance of the wait time coefficients suggests that the observed
wait times at the stop were generally in a range that did not significantly affect riders’ overall
satisfaction with operations. Together with the low sensitivity toward wait times observed by
Carrel et al. (2016), the results of this model suggest that in future research, a different approach
should be taken to identifying wait times. There are three possible avenues:
•Adding other sensor data such as accelerometer in order to better identify when a person
actually walked to a transit stop.
•Directly asking a participant about the perceived wait time and where it was spent.
•Tracking the use of real-time arrival information on the phone in order to determine when
a participant first looked at upcoming departures. This could serve as an indicator of the
beginning of a wait.
Unlike the coefficient of the wait time at the origin stop, the transfer time coefficient (αtran s f ertime )
is negative and significant at p<0.01. A comparison of the transfer time coefficient and the IVTT
coefficient shows that according to the model, one minute of in-vehicle delay causes as much dis-
satisfaction as four minutes of transfer time. The model further includes two binary variables
related to transfer time: αnotr ans f er captures cases where no transfer was identified from the loca-
tion tracking data and the participant did not report a transfer, and αunob served trans f er captures cases
where the participant reported having transferred but the transfer could not be identified from the
location tracking data. Both are negative and significant at p<0.10. While this result is intuitive
for the latter coefficient, it is not intuitive for the former, as it suggests that in general, passengers
who transfer tend to report a higher satisfaction than passengers who do not transfer. This merits
further investigation with a different data set.
4.2.3 Effect on exit satisfaction and critical incidents
The coefficient βdailysat in Table 3 links daily satisfaction with operations to the exit satisfaction with
operations. As expected, it is positive (0.1), and it is also significant at p<0.01, showing a positive
correlation between daily satisfaction and exit satisfaction.
Interestingly, all five coefficients related to self-reported critical incidents have negative and signif-
icant estimates at p<0.1. The first two are the number of times a person arrived late to work or
school (βlatework) due to a transit delay and the number of times a person arrived late to a leisure
activity due to a transit delay (βlateleisure ). Both were self-reported in the exit survey. βnotravel cap-
tures cases where participants reported on their daily mobile surveys that they had wanted to use
public transportation that day but could not due to a delay and were forced to choose a different
mode. However, there was no obligation to report these incidents, so it must be assumed that the
reported numbers are a lower bound. Therefore, the estimated coefficient is an upper bound on the
impact of such incidents. A special case of critical incidents were denied boardings: These were
captured both through self-reports in the exit survey and through automated detection. The latter
affect the daily satisfaction with operations via αle f tb ehind . The automated detection was only based
on location data: If a participant was observed to be at a stop and not board a passing vehicle but
board the following one, it was recorded as a denied boarding. There is a risk of misclassification,
as the participant may have been carrying out an activity and may not have intended to board the
first vehicle. Therefore, the number of such automatically detected incidents in the data set is an
upper bound, and the coefficient estimate is a lower bound for the impact. In Table 3, αle f tb ehind is
negative but not significant. On the other hand, the effect of self-reported denied boardings was
found to exhibit some nonlinearity. There were 11 possible answers to the self-reported question:
From 0 to 9, and then “10 or more”. The model was found to produce the best fit if these two cate-
gories were separated, as in Table 3. Both coefficients, βl e f tbehin d_1_9 and βle f tbe hind_10, are negative
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 378
and significant at p<0.10, but the latter is approximately three times larger than the former. It
is possible (and plausible) that the larger coefficient for 10 or more denied boardings is capturing
protest responses.
4.2.4 Effects on future behavior
The final component is the choice model. It is specified with only two inputs: The participant’s
exit satisfaction with operations and the participant’s exit satisfaction with the travel environment.
All other variables in the model, including the daily satisfaction with operations, the critical inci-
dents, the travel time experiences and the entry satisfaction with operations, affect future behavior
through the exit satisfaction with operations. The indicator variables for the satisfaction in the exit
survey are on a four-point Likert scale, whereas the choice indicators are binary. Given that the
indicator variables had different reference points, one has to make the assumption that the partici-
pants did not significantly change their pre-study frequency of transit use during the study in order
to interpret the results. Given that assumption, the choice is between (a) continuing to use public
transportation at the same frequency as before and during the study or using it more frequently,
and (b) using public transportation less frequently or discontinuing it altogether. The choice indi-
cator for the former is 1, and for the latter it is 0. Thus, positive coefficients mean there is a positive
correlation between the input variable and the participant’s willingness to use public transporta-
tion the same or more in the future. As can be seen in Table 3, the effects of both satisfaction with
operations and satisfaction with the travel environment in the exit survey are positive (0.39 and
0.23, respectively) and significant at p<0.05. This is intuitive, as higher satisfaction leads to a
higher willingness to continue using transit in the future.
Of particular interest here is the relative difference between the two coefficients. The coefficient
of the latent satisfaction with operations is approximately 1.7 times the coefficient of the latent
satisfaction with the travel environment. Since these are latent variables, their exact values cannot
be calculated, but the comparison can be made with the help of two of the indicator variables:
A change in the latent satisfaction with operations variable that causes a one-point increase in
satisfaction with overall reliability has 1.46 times the effect of a change in the satisfaction with travel
environment variable that causes a one-point increase in satisfaction with crowding. This confirms
that for the present group, overall satisfaction with operations has a stronger influence on future
mode choice decisions than overall satisfaction with the travel environment. This is consistent with
the results of the descriptive analysis in section 4.1. In future research, it would be interesting to
add objective measurements to the travel environment variables. Crowding would be of particular
interest, given the importance reported by participants in section 4.1; this could be calculated with
data from automatic passenger counting or fare payment systems.
With the help of the final choice model, it is now possible to calculate the relative influence of var-
ious experiences on passengers’ willingness to remain transit riders in the future. As a calculation
example, the effect on the choice utility of one incident of not being able to travel due to a delay on
the transit network is βnotravel ·γo ps =−0.07 ∗0.39 = −0.027.
4.2.5 Calculation of trade-offs
If the coefficient for wait times were negative and significant, it would be possible to calculate the
impact of negative critical incidents in terms of the equivalent amount of dissatisfaction caused by
wait times. For instance, if αwaitover5and αw aitunder5were -0.02 and significant and αle f t behind were
-0.12 and significant, one could state that the dissatisfaction caused by one instance of a denied
boarding would be equivalent to the dissatisfaction caused by approximately 6 minutes of wait
time at the origin stop. Such trade-offs provide good rules of thumb for transit planning profes-
sionals, as is illustrated by the popularity of the rule of thumb that a minute of out-of-vehicle travel
time is twice as onerous as a minute of in-vehicle travel time (Wardman, 2004). Therefore, a goal
of future research should be to derive significant wait time coefficients in order to calculate such
trade-offs.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 379
Table 4. Results of the simulation.
10 min delay 10 min transfer, Left behind,
on board late to work 10 min wait
Change in probability per person -0.006 -0.003 -0.002
Potential systemwide change -1628 -745 -504
% of SFMTA yearly turnover 5.4 2.5 1.7
4.3 Simulation
To illustrate a potential use of this type of model, a simulation was conducted. Three hypothetical
scenarios were simulated, and the group of subjects from whom the data had been collected served
as a convenience sample. Three hypothetical simulation scenarios were designed:
1. Impact of every participant experiencing one additional ten-minute on-board delay.
2. Impact of every participant having one additional experience of being left behind at a stop
with corresponding ten minutes of additional wait time.
3. Impact of every participant experiencing one additional 10-minute transfer wait time and
arriving late at work.
The scenarios measured the impact of one additional event since the baseline was the choice prob-
ability calculated from the set of experiences that the participants had had during the study. The
output of each simulation run was an average probability of remaining a transit rider in the future,
which was contrasted with the baseline probability. First, we discuss the assumptions of the simu-
lation scenarios in the context of the San Francisco transit system. The first simulation scenario is on
the high end of typical in-vehicle delay times on Muni. The maximum delay observed in the data
set was 9:16 minutes. Therefore, this simulation scenario captures the effect of a major, system-wide
disruption. The second simulation scenario is also on the pessimistic side; out of 449 participants,
90 (20%) were observed to have been left behind at a stop at least once during the study. This
scenario assumes that it happened to each participant one additional time. The third simulation
scenario is more typical of day-to-day operations on Muni. The average transfer time experienced
by participants who transferred was 7:21 minutes, and on average, participants reported being late
to work or school due to difficulties on transit 2.43 times during the study.
Table 4 shows the simulation results. In the base case, on average, participants had a 0.77 proba-
bility of remaining transit riders in the future. The first data row shows the change in probability
due to the simulated incident. In the second row, the change in probability was extrapolated to the
SFMTA’s entire ridership of 280,000, and the potential loss of riders due to the simulated event was
calculated. To put the numbers into context, the third row shows what percentage of the SFMTA’s
yearly turnover (approximately 30,000 passengers) the simulated losses would represent. Since the
scenarios are limited to single events that differ in the likelihood of occurrence and severity, this is
not intended to model actual events observed on the SFMTA’s network. Rather, this is intended to
demonstrate how this type of model can help analysts understand the potential impact of various
operating strategies and capital investment programs on ridership turnover. To analyze a proposed
operational or infrastructure change, two pieces of input are required: First, the analyst needs to
know the anticipated frequencies of delays and critical incidents before and after the changes. Sec-
ond, an approximate knowledge of the rider population that will be impacted by the change can
help in constructing the sample used for the simulation.
5 Discussion of results
The results show the link between service quality problems and loss of ridership from two different
angles. First, in section 4.1, we selected participants who reported a behavioral desire or intention to
reduce their use of public transportation, and investigated the self-reported reasons for which they
wanted to do so. It was seen that travel time reliability was mentioned as the overall most important
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 380
factor. In terms of the number of “very influential” responses, crowding came in second, but in
terms of the average of all responses, the second-most important factors were wait time reliability
and travel time reliability. Both had an average response of 3.78, compared to 4.13 for overall
reliability and 3.64 for crowding. In a broader sense, even crowding can be considered a reliability
variable since the crowding of vehicles is related to vehicle bunching, and passengers do not know
ahead of time whether they will be able to find a seat (Polydoropoulou and Ben-Akiva, 2001).
Overall, travel time variables were more influential than travel environment variables. However,
it should be noted that this study concerned users who were already regular transit users, and
therefore, it may be a self-selected group that might be less concerned with the travel environment
than, for example, a comparison group of auto users.
Second, in section 4.2, we presented model estimation results and applied them to a simulation
in section 4.3. The model results are generally in line with previous findings presented in Carrel
et al. (2016), showing that neither the scheduled in-vehicle travel time nor early arrivals at the
destination have a significant effect on satisfaction, but that in-vehicle delays are an important
driver of dissatisfaction. By extension, it is shown that in-vehicle delays also have a strong impact
on overall satisfaction after an extended time period and on passengers’ desire to stop using public
transportation. The transfer time coefficient was also negative and significant, but the origin wait
time coefficients were not significant. It is assumed that this may be partly related to the difficulties
associated with properly identifying origin wait times, as explained in section 4.2, and partly to the
fact that participants appeared to be choosing to spend their wait times at locations other than the
stop and rely on real-time information to go to the stop when an arrival was predicted. Therefore,
the wait times detected from location data may not necessarily have corresponded to the wait times
as defined by the user. It is also interesting that the coefficient of the binary variable for missing
wait time observations was estimated to have a negative and significant impact on satisfaction.
Unfortunately, since it is unknown what reasons led to a missing wait time measurement in any
specific case, this result is difficult to interpret. Based on the reasons for missing wait time data
discussed in section 4.2, the following may be occurring:
•It is possible that the wait times that were not observed due to missing or insufficient data
were on average significantly longer than the observed wait times. However, we are currently
not aware of any plausible reasons why this may be the case.
•It is possible that if there was a large time gap between participants’ desired departure times
and the next departures, and participants chose to spend that time carrying out other activities
nearby, they still perceived it as wait time, leading to lower satisfaction.
There may be other explanations which we were not aware of at the time of writing. In future
research, these data shortcomings should be addressed in order to solidify our understanding of
the impact of various delay times on passengers. The model presented in this paper goes beyond
previous satisfaction models by explicitly linking critical incidents and personal experiences with
travel times to future behavioral intentions by way of customer satisfaction, and the significance
of the relevant coefficients in Table 3 demonstrates that this link is present. It is shown that for
the present group of participants, which consisted mostly of regular transit users, satisfaction with
travel times and operational aspects is more important in determining their willingness to remain
transit riders than satisfaction with the travel environment. Furthermore, even though the influence
of wait times relative to in-vehicle delay times and transfer times requires further research, the
results clearly demonstrate the value of developing models using participants’ personal experiences
with service quality as a means of understanding future mode choice intentions and the influence
of various factors related to service quality. Besides travel times, we find that several types of
critical incidents have measurable negative effects on participants’ overall satisfaction in the exit
survey and thus on their willingness to remain transit riders.
The value of the latent variable modeling framework used in this paper was that it permitted us
to summarize an individual’s overall satisfaction with operations in one variable and to determine
the influence of a variety of experiences with travel times on that variable. It is flexible, and its
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 381
specification can accommodate variables collected on different time scales, such as the daily sat-
isfaction with operations and the entry and exit satisfaction. Most importantly, it allowed us to
account for correlation between a participant’s satisfaction ratings with respect to different travel
time components.
6 Conclusions
In this paper, we presented an analysis and model results to understand the link between service
quality, satisfaction, and transit ridership loss. This work emphasizes the importance of riders’
personal experiences; an innovative procedure was used to map location data from users’ mo-
bile phones to vehicle location data in order to automatically identify personal experiences and
use them in the estimation of the model. This demonstrates the value and potential of such new
data collection methods in answering complex questions and observing phenomena that require
panel data. The insights gained from these data help establish the link between travel time vari-
ability/critical incidents, satisfaction, and transit ridership loss. For the first time, this framework
makes it possible to directly model the effect of negative personal experiences on future mode
choice decisions and thus on ridership loss due to delays and system management strategies. In
the future, it can be further refined with additional data in order to form the basis for new opera-
tional tools that would enable a move from system-based to person-based performance metrics for
transit agencies.
References
Abou-Zeid, Maya and Moshe Ben-Akiva (2010). A model of travel happiness and mode switching. In:
Choice modelling: The state-of-the-art and the state-of-practice (proceedings from the inaugural international
choice modelling conference), pp. 289–306.
Abou-Zeid, Maya et al. (2012). Happiness and travel mode switching: Findings from a Swiss public
transportation experiment. In: Transport Policy 19.1, pp. 93–104.
Ajzen, Icek (1991). The theory of planned behavior. In: Organizational Behavior and Human Decision Pro-
cesses 50.2, pp. 179–211.
Anable, Jillian and Birgitta Gatersleben (2005). All work and no play? The role of instrumental and
affective factors in work and leisure journeys by different travel modes. In: Transportation Research
Part A: Policy and Practice 39.2, pp. 163–181.
Anderson, Eugene and Mary Sullivan (1993). The antecedents and consequences of customer satisfaction
for firms. In: Marketing Science 12.2, pp. 125–143.
Bamberg, Sebastian, Icek Ajzen, and Peter Schmidt (2003). Choice of travel mode in the theory of planned
behavior: The roles of past behavior, habit, and reasoned action. In: Basic and Applied Social Psychology
25.3, pp. 175–187.
Bates, John et al. (2001). The valuation of reliability for personal travel. In: Transportation Research Part E:
Logistics and Transportation Review 37.2, pp. 191–229.
Ben-Akiva, Moshe et al. (2002). Integration of choice and latent variable models. In: Perpetual motion:
Travel behaviour research opportunities and application challenges. Elsevier Science, pp. 431–470.
Bollen, Kenneth (2014). Structural equations with latent variables. John Wiley & Sons.
Bollen, Kenneth et al. (2007). Latent variable models under misspecification: Two-stage least squares
(2SLS) and maximum likelihood (ML) estimators. In: Sociological Methods & Research 36.1, pp. 48–86.
Carrel, Andre, Anne Halvorsen, and Joan Walker (2013). Passengers’ perception of and behavioral adap-
tation to unreliability in public transportation. In: Transportation Research Record 2351, pp. 153–162.
Carrel, Andre, Raja Sengupta, and Joan Walker (2016). The San Francisco Travel Quality Study: Tracking
the trials and tribulations of a transit taker. In: Transportation. Advance Online Publication.
Carrel, Andre et al. (2015). Quantifying transit travel experiences from the users’ perspective with high-
resolution smartphone and vehicle location data: Methodologies, validation, and example analyses.
In: Transportation Research Part C: Emerging Technologies 58, pp. 224–239.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 382
Carrel, Andre et al. (2016). In pursuit of the happy transit rider: Dissecting satisfaction using daily sur-
veys and tracking data. In: Journal of Intelligent Transportation Systems 20 (4), pp. 345–362.
Daganzo, Carlos and Josh Pilachowski (2011). Reducing bunching with bus-to-bus cooperation. In:
Transportation Research Part B: Methodological 45.1, pp. 267–277.
Davis, Mark and Janelle Heineke (1998). How disconfirmation, perception and actual waiting times
impact customer satisfaction. In: International Journal of Service Industry Management 9.1, pp. 64–73.
Fredrickson, Barbara and Daniel Kahneman (1993). Duration neglect in retrospective evaluations of af-
fective episodes. In: Journal of Personality and Social Psychology 65.1, p. 45.
Friman, Margareta, Bo Edvardsson, and Tommy Gärling (2001). Frequency of negative critical incidents
and satisfaction with public transport services. In: Journal of Retailing and Consumer Services 8.2, pp. 95–
104.
Friman, Margareta and Markus Fellesson (2009). Service supply and customer satisfaction in public
transportation: The quality paradox. In: Journal of Public Transportation 12.4, pp. 57–69.
Friman, Margareta et al. (2013). Psychometric analysis of the satisfaction with travel scale. In: Transporta-
tion Research Part A: Policy and Practice 48, pp. 132–145.
Kahneman, Daniel et al. (1993). When more pain is preferred to less: Adding a better end. In: Psycholog-
ical Science 4.6, pp. 401–405.
Kahneman, Daniel et al. (2004). A survey method for characterizing daily life experience: The day re-
construction method. In: Science 306.5702, pp. 1776–1780.
Lanken, Bas et al. (1994). Attitude versus general habit: Antecedents of travel mode choice. In: Journal of
Applied Social Psychology 24.4, pp. 285–300.
Ma, Xiaolei et al. (2013). Mining smart card data for transit riders’ travel patterns. In: Transportation
Research Part C: Emerging Technologies 36, pp. 1–12.
Mittal, Vikas, Pankaj Kumar, and Michael Tsiros (1999). Attribute-level performance, satisfaction, and
behavioral intentions over time: A consumption-system approach. In: The Journal of Marketing 63.2,
pp. 88–101.
Pedersen, Tore, Margareta Friman, and Per Kristensson (2011). The role of predicted, on-line experienced
and remembered satisfaction in current choice to use public transport services. In: Journal of Retailing
and Consumer Services 18.5, pp. 471–475.
Perk, Victoria, Jennifer Flynn, and Joel Volinski (2008). Transit ridership, reliability and retention. Tech. rep.
FDOT BD 549-32. Tampa, FL: National Center For Transit Research.
Perugini, Marco and Richard Bagozzi (2001). The role of desires and anticipated emotions in goal-
directed behaviours: Broadening and deepening the theory of planned behaviour. In: British Journal of
Social Psychology 40.1, pp. 79–98.
– (2004). The distinction between desires and intentions. In: European Journal of Social Psychology 34.1,
pp. 69–84.
Polydoropoulou, Amalia and Moshe Ben-Akiva (2001). Combined revealed and stated preference nested
logit access and mode choice model for multiple mass transit technologies. In: Transportation Research
Record 1771, pp. 38–45.
Trépanier, Martin, Khandker Habib, and Catherine Morency (2012). Are transit users loyal? Revelations
from a hazard model based on smart card data. In: Canadian Journal of Civil Engineering 39.6, pp. 610–
618.
Vij, Akshay, Andre Carrel, and Joan Walker (2013). Incorporating the influence of latent modal prefer-
ences on travel mode choice behavior. In: Transportation Research Part A: Policy and Practice 54, pp. 164–
178.
Walker, Joan (2001). Extended discrete choice models: Integrated framework, flexible error structures,
and latent variables. Massachusetts Institute of Technology.
Wardman, Mark (2004). Public transport values of time. In: Transport Policy 11.4, pp. 363–377.
Woodruff, Robert, Ernest Cadotte, and Roger Jenkins (1983). Modeling consumer satisfaction processes
using experience-based norms. In: Journal of Marketing Research 20.3, pp. 296–304.
Appendix A.
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 383
Table 5. Estimation results for the measurement equations of the latent variable choice model. The
abbreviations are explained at the end of the table.
Latent variable Indicator variable Value Robust Std err Robust t-test p-value
δEntry sat. ops. Overall reliability 2.90 0.06 47.39 0
δEntry sat. ops. IVTT 2.87 0.05 52.22 0
δEntry sat. ops. Wait time 3.08 0.06 55.68 0
δDaily sat. ops. Overall reliability 2.34 0.12 19.93 0
δDaily sat. ops. IVTT 2.31 0.10 22.21 0
δDaily sat. ops. Wait time 2.24 0.11 19.87 0
δDaily sat. ops. Transfer time 2.22 0.12 19.33 0
δRem. daily sat. ops. Overall reliability 2.44 0.09 27.87 0
δRem. daily sat. ops. IVTT 2.23 0.09 24.01 0
δRem. daily sat. ops. Wait time 2.30 0.08 28.44 0
δRem. daily sat. ops. Transfer time 2.06 0.11 18.26 0
δExit sat. ops. Overall reliability 1.87 0.11 17.56 0
δExit sat. ops. IVTT 2.17 0.08 27.64 0
δExit sat. ops. Wait time 1.99 0.09 22.83 0
δExit sat. env. Crowding 1.59 0.04 41.79 0
δExit sat. env. Safety 2.29 0.03 69.36 0
δExit sat. env. Other Pax 2.15 0.03 69.20 0
δExit sat. env. Cleanliness 1.91 0.04 48.37 0
λEntry sat. ops. Overall reliability 0.74 0.06 11.99 0
λEntry sat. ops. IVTT 0.63 0.06 10.70 0
λEntry sat. ops. Wait time 0.64 0.05 13.86 0
λDaily sat. ops. Overall reliability 0.57 0.04 15.85 0
λDaily sat. ops. IVTT 0.51 0.03 18.87 0
λDaily sat. ops. Wait time 0.52 0.03 17.21 0
λDaily sat. ops. Transfer time 0.53 0.04 15.05 0
λRem. daily sat. ops. Overall reliability 0.37 0.05 7.42 0
λRem. daily sat. ops. IVTT 0.41 0.04 9.65 0
λRem. daily sat. ops. Wait time 0.35 0.05 7.07 0
λRem. daily sat. ops. Transfer time 0.25 0.08 3.34 0
λExit sat. ops. Overall reliability 1.00 * * *
λExit sat. ops. IVTT 0.73 0.05 14.02 0
λExit sat. ops. Wait time 0.82 0.03 24.34 0
λExit sat. env. Crowding 1.00 * *
λExit sat. env. Safety 0.94 0.11 8.94 0
λExit sat. env. Other Pax 0.81 0.12 6.96 0
λExit sat. env. Cleanliness 1.12 0.24 4.61 0
σEntry sat. ops. Overall reliability 0.74 0.03 24.55 0
σEntry sat. ops. IVTT 0.75 0.02 30.87 0
σEntry sat. ops. Wait time 0.78 0.02 41.99 0
σDaily sat. ops. Overall reliability 0.76 0.03 29.10 0
σDaily sat. ops. IVTT 0.88 0.02 47.85 0
σDaily sat. ops. Wait time 0.99 0.02 66.03 0
σDaily sat. ops. Transfer time 0.94 0.04 25.35 0
σRem. daily sat. ops. Overall reliability 0.58 0.03 18.01 0
σRem. daily sat. ops. IVTT 0.63 0.03 23.57 0
σRem. daily sat. ops. Wait time 0.73 0.03 21.25 0
σRem. daily sat. ops. Transfer time 0.88 0.09 9.81 0
σExit sat. ops. Overall reliability 0.75 0.05 16.05 0
σExit sat. ops. IVTT 0.76 0.02 35.25 0
σExit sat. ops. Wait time 0.72 0.03 21.66 0
σExit sat. env. Crowding 0.81 0.03 27.19 0
σExit sat. env. Safety 0.56 0.02 30.34 0
σExit sat. env. Other Pax 0.58 0.03 21.43 0
σExit sat. env. Cleanliness 0.68 0.08 8.54 0
Continued on next page
EJTIR 17(3), 2017, pp. 360 −383
Carrel and Walker
Understanding Future Mode Choice Intentions of Transit Riders 384
Latent variable Indicator variable Value Robust Std err Robust t-test p-value
Legend:
Entry sat. ops.: Satisfaction with operations in entry survey.
Exit sat. ops.: Satisfaction with operations in exit survey.
Exit sat. env.: Satisfaction with travel environment in exit survey.
Daily sat. ops.: Satisfaction with operations in daily surveys d1through d4.
Rem. daily sat. ops.: Satisfaction with operations in daily survey dr.
