You are on page 1of 9

Research in Transportation Economics 68 (2018) 2–10

Contents lists available at ScienceDirect

Research in Transportation Economics


journal homepage: www.elsevier.com/locate/retrec

The conditional effects of social influence in transportation mode choice T


a,∗ b
Susan Pike , Mark Lubell
a
University of California, Center for Environmental Policy and Behavior, Institute for Transportation Studies, Davis, United States
b
University of California, Center for Environmental Policy and Behavior, Davis, United States

A R T I C LE I N FO A B S T R A C T

JEL classification: Recent research suggests that an individual's transportation mode choice is partially influenced by the mode
R41 choice of other people in their social network. This paper advances this basic idea by testing the hypothesis that
Keywords: the extent of social influence is conditional on external context factors such as commute characteristics. Through
Travel behavior an online survey of students at the University of California, Davis, we collected information about students'
Social networks transportation decisions for campus travel and their social networks. For each participant as an ego, we gathered
Social influence information about their ego-network including up to five social contacts or alters. A series of complementary
Transportation mode choice statistical models find the strength of social influence is lower for those with longer commute distances where
Ego-networks
biking is more costly than driving or taking the bus, and is also lower at distances where walking has higher
utility than biking. Social influence is most important when the external commute characteristics entail rela-
tively equal travel costs for different modes. As social influence and other social processes are evaluated as
potential policy instruments, these and other heterogeneous effects should be taken into account.

1. Introduction which individuals may respond to social influence is conditional on


three categories of variables: characteristics of the individual, attributes
Social networks serve as a foundation for multiple social processes, of the network/relationship, and the individual's choice context. First,
such as cooperation, resource sharing and social influence. There is individual characteristics may make some people less likely to conform
evidence that social networks are important in a broad range of out- to social influence; for instance, people who experienced a bicycle ac-
comes including academic achievement (Carrell, Fullerton, & West, cident or have non-conforming personalities may be less susceptible to
2009; Sacerdote, 2001), happiness (Fowler & Christakis, 2008), and social influence. Social relationships are also important mediators of
health issues such as obesity (Koehly, Loscalzo 2009). Social influence social influence; individuals are expected to be more likely to conform
affects environmental actions (Axelrod & Lehman, 1993), and social to the behaviors of those with whom they are closer or trust. In this
networks are important in co-authorship and collaboration (for ex- study, we are particularly interested in the effect of contextual factors
ample Freeman, 1984). Social networks also provide access to resources known to be important in travel behavior. We investigate how commute
such as employment opportunities, through network connections distance constrains or enhances the effect of social influence on trans-
(Granovetter, 1973). portation mode choice. We look at distance because of the importance
The focus of this study is social influence, whereby the knowledge, of distance and related variables (travel time and cost) in travel beha-
actions and/or opinions of one individual affect those of another. For vior. Our results suggest that those with longer commute distances are
example, social networks act as avenues for the diffusion of informa- less responsive to social influence.
tion. Social groups may discuss their use of alternative transportation The next section reviews the quickly growing literature on social
modes, changes to transit routes or schedules, or new bicycle infra- influence in travel behavior research, followed by the development of
structure; influencing shifts in the behavior or attitudes of individuals in our hypotheses regarding the conditional effects of social influence. We
the group. Social networks also provide pathways for the establishment then describe our research design in the context of student travel be-
of social norms where individuals conform to the behavior of those with havior in Davis, California, where we used an online survey to measure
whom they are socially connected. The role of social influence is a individual mode choice, travel distance, and the travel behaviors of
quickly growing area of research in the context of transportation mode social networks. We estimate the extent to which social influence is
choice (Maness, Cinzia, & Dugundji, 2015). conditional on commute distance using statistical models that control
We contribute to this line of research by arguing that the extent to for the endogeneity problem of “common environment” that causes a


Corresponding author.
E-mail addresses: scpike@ucdavis.edu (S. Pike), mnlubell@ucdavis.edu (M. Lubell).

https://doi.org/10.1016/j.retrec.2018.05.010
Received 15 February 2016; Received in revised form 15 May 2017; Accepted 24 May 2018
Available online 18 June 2018
0739-8859/ © 2018 Elsevier Ltd. All rights reserved.
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

Short to medium commute distance

Effect of social influence on decision to bike


Probability of choosing to bike
Higher effect of social influence

Very short commute distance

Long commute distance

Lower effect of social influence

Social influence on choice to Commute distance

Panel a: theore cal linear model of social Panel b: theore cal change in social
influence; slope/effect varies by distance influence with respect to commute distance

Fig. 1. Conceptual model of the impact of distance on social influence effects.

correlation between the mode choices of an individual and her network. moderates responses to social influence. Life-stage or age and sociali-
zation also affect responses to social influence; for instance, one study
1.1. Social influence and other factors in travel behavior finds college students are more susceptible than homemakers to the
influence of social reference group on brand selection for a number of
There is a growing literature exploring social influence in trans- products (Park & Lessig, 1977). The broad literature on social learning
portation mode choice and methods used to measure such effects suggests that some individuals are more likely to use social learning
(Dugundji & Walker, 2005; Maness et al., 2015; Pike, 2014; Páez & strategies rather than individual learning (for example Efferson, Lalive,
Scott, 2007; Walker, Ehlers, Banerjee, & Dugundji, 2011). Social in- Richerson, McElreath, & Lubell, 2009).
fluence is important in transportation decisions such as whether or not Properties of one's social relationships are expected to mediate re-
to telecommute (Scott, Dam, Páez, & Wilton, 2012; Wilton, Páez, & sponses to social influence. Sherwin, Chatterjee, and Jain (2014) pro-
Scott, 2011). Neighborhood transit use is related to an individual's pose three levels of social influence, based on the type of reference
likelihood of using transit (Goetzke, 2008) and social and cultural group; family, friends, and broad social-cultural context. It is expected
context affect bicycle ridership (Goetzke & Rave, 2010). Perceptions of that those with closer relationships are more influential. These differ-
and attitudes towards new technologies such as electric vehicles are ences have been formalized in models with distance-decay functions
affected by social network exposure (Axsen & Kurani, 2011). This line (where distance is defined by social, rather than geographical, proxi-
of research represents a major extension of the standard model of travel mity) with shorter distance, or closer relationships exerting more in-
behavior that mainly examines individual costs/benefits of different fluence than longer/less close (for example see Páez, Scott, & Volz,
modes. 2008).
As travel behavior researchers explore new factors such as social Contextual factors are also expected to affect individual responses to
influence it is important to consider how they relate to and are im- social influence. For example schedules, household responsibilities, and
pacted by more traditional factors, such as trip and mode character- other constraints may limit one's potential to adopt the behaviors of
istics, socio-demographics, attitudes (Mokhtarian & Salomon, 1997), others in their social network. Indeed, Sherwin et al. (2014) find that
land use and the built environment (Ewing & Cervero, 2010). There is social influence is an important predictor of bicycling behavior, but is
growing evidence of social and spatial interdependencies in transpor- often intertwined with other factors, and call for future research that
tation decision making; however, these have primarily been explored in explores the “… relative strength of the different levels of social in-
relation to activity travel (for example Carrasco & Miller, 2006 and fluence and other interacting factors.” Here we explore just that; the
2009, Larsen, Urry, & Axhausen, 2008) or to the impacts of land use interaction between distance and social influence. We hypothesize that
characteristics on social interactions (Farber & Páez, 2009). In this individuals with longer commute distances will respond to social in-
study, we investigate the interplay between social influence and com- fluence to a lesser extent than those with shorter commute distances. As
mute distance. We expect reductions in the effects of social influence for commute distance increases, the response to social influence decreases.
those with longer commute distance. In the next section we describe this concept in more detail.

1.2. Factors affecting responses to social influence 1.3. Conceptual model

We expect individual, relational, and contextual factors to affect Fig. 1a presents a conceptual model of the conditional nature of
responses to social influence. The idea that response to social influence social influence we expect for a sample segmented based on commute
may differ among individuals has been explored within transportation, distance. The probability an individual chooses to bike increases as
and other areas. For example, Park and Lessig (1977) note, “… a per- social influence (the extent that others in ones network bike) to bike
son's dependence on social referents is inversely related to his con- increases. At the same time, we propose those with shorter commute
fidence in his competence to judge the issue.” Similarly, in a more re- distances are more affected by social influence (steeper curve) and
cent example related to travel behavior, Bamberg, Hunecke, and those with longer commute distances are less affected (flatter curve). At
Blobaum (2007) find that knowledge of or involvement with a subject very short commute distances the effect of social influence to bike may

3
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

also be small since the convenience of walking may reduce the effect of (Popovich, 2014). The CTS does not ask for the residential location of
social influence for these individuals. Fig. 1b illustrates how this re- students who live on campus; however, in the social networks survey
lationship might change continuously with respect to commute dis- we asked everyone, including students that live on campus, to list the
tance. The exact shape of this curve is not known, though we expect the cross streets nearest to their home. For those on campus we computed
effect of social influence to bike to be low at very short commute dis- the straight line distance from their home to a central campus location.
tances and to have the greatest effect for those with short to medium This is a realistic estimation of their distance to central campus, since
commute distances. As commute distance increases the effect of social there is a dense network of campus bike paths and many possible
influence decreases and eventually tapers off. A similar, or inverse re- walking or biking routes through campus, that likely differ little from a
lationship may occur for other factors. straight line distance.
The primary dependent variable in our analysis is transportation
2. Data and methods mode choice. We asked respondents “What mode of transportation do
you usually use to travel to campus or your normal off-campus location
It is important to note the relationship between our conceptual for school or work?” The possible choices are: bike, walk, skate/board,
model and traditional models of travel behavior that examine the ex- motorcycle/scooter, drive alone, get a ride, carpool, take the bus or take
pected utility of different mode choices, where the expected utility is the train. The social networks survey also asked participants to indicate
evaluated in terms of individual costs and benefits such as travel dis- the importance of various factors in their mode choice decisions and
tance, safety and other variables. The intuition behind our conceptual which modes of transportation are available to them. To measure social
model is that social influence is stronger when there are only small networks, we asked each respondent to provide the information ne-
differences between the expected utility of different mode choices. In cessary to construct their ego-network; a subset of their social network
these cases, the individual may be persuaded by the social influence of consisting of the respondent, referred to as the ego and a selection of
peers. For example, if it takes about the same amount of time to bike or their social contacts, referred to as alters. We used the following name
drive to work, one might be more persuaded by a friend's arguments generator question where each ego was asked to name up to five alters:
about the benefits of biking. However, if there is a very large difference In this question, think about all the people who have been in your
in the expected utility of different mode choices, social influence will be social circle over the past six months; this includes people with whom
minimal. you live, work or attend class, socialize or participate in activities
etc. or people you speak with over the phone or internet.
2.1. Campus travel survey and social networks survey List the first names of the five contacts you have had the most
frequent regular interaction with over the past six months.
Our study area is the University of California, Davis where extensive We elected to gather information on five alters, as this limits re-
bicycle infrastructure throughout the city and the university campus, spondent burden, and past work has found this to be about the size of
the mild climate, and flat topography all contribute to high shares of one's “core reference group” (Axhausen, 2008). Egos were also asked a
bicycling within the city and the campus community. Survey data was series of questions about their alters, including each alter's usual com-
collected at UC Davis in coordination with the 2013-14 academic year's mute mode. To confirm egos correctly reported about their alters we
annual campus travel survey (CTS), conducted in October of 2013. checked whether ego reports about alters matched self-reported in-
Travel data as well as socio-demographic information and attitudes formation from alters. This was possible for a small subset of the alters,
towards transportation and the environment were collected in the CTS identified as follows. Each year of our survey we utilized a snowball
(the complete survey is contained as an appendix in Popovich, 2014). sampling method; we asked the egos to provide contact information for
A stratified (according to university role; student, faculty, staff, etc.) their alters, or to forward a short survey to their alters (for those who
sample of the UC Davis campus community, consisting of 27,798 in- would not share their alters' contact information). Due to very limited
dividuals, was invited to participate in the CTS, with a target response success of this effort, the data collected has been utilized primarily to
rate of 10.11% or a final sample of 2811 (Popovich, 2014). Ultimately validate ego reports.
3,663 (13.2%), including 2,671 students, gave usable responses In order to validate the mode choices of the alters, we use data from
(Popovich, 2014). At the end of the survey, students were asked if they the previous year (2012–13), since there were more respondents in our
would like to receive information about a related survey involving so- snowball survey in that year. There were 149 alters (listed by egos) that
cial networks and transportation; 1,396 indicated they were interested participated in the snowball survey or happened to be respondents to
and of those, 966 participated in our social networks and travel survey; that year's CTS. For these alters, 90 provided their usual mode of
about 70% of those who indicated an interest, and one-third of the transportation, allowing us to cross-check the alters' self-reported
students who participated in the CTS. The social network survey was modes against the ego-reported modes. Egos identified the correct
implemented in November of 2013. We are able to match each response mode for 70 of these alters, or 78%. Because egos are reasonably ac-
in our survey to that individual's response in the CTS, and our analysis curate when indicating alter modes of transportation, we rely on the
draws on information provided in both surveys. ego-reported typical mode of transportation for each alter. For each
In earlier work (Pike and Lubell, 2016) we compared the sample of ego-network we compute the proportion of alters that uses each of nine
respondents to our social networks survey, with the sample of re- modes of transportation.
spondents to the CTS, and found few significant differences. Further,
the magnitudes of these differences were small; typically a few per- 2.2. Samples used in analysis
centage points, and not expected to substantively alter our results.
Notable differences were class standing, with more sophomores parti- It is necessary to exclude some survey participants from the analysis
cipating in the CTS and more seniors participating in the social net- presented here. In this section we describe these exclusions and the re-
works survey. Related to travel behavior; 53% of CTS respondents bike sulting samples. First, respondents are excluded from the sample if they do
as their usual mode, while 59% of the social networks survey re- not live in Davis. Davis is surrounded by agricultural lands and very few
spondents bike. For bus these numbers are 26%, and 21% for CTS and people live between five and ten miles from campus. This distorts the
social networks survey respondents, respectively. distribution of distances to campus since any individual living more than
The CTS collects cross-streets for respondents' home locations and five miles from campus in effect lives at least ten miles from campus. In
this data was geocoded to determine the approximate road-network addition, since the mode choices of the ego and their alters are key vari-
distance to campus for each respondent. The road-network distance was ables in our analysis, we exclude respondents if they did not provide a
estimated for the route to campus with the shortest commute time usual mode of transportation for themselves and/or their alters.

4
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

Respondents are also removed from the sample if they are missing variable of interest – the choice of the ego to bike. We utilize the
geographic information. Lastly, we restricted our sample to those stu- neighborhood biking environment of the alters as our instrumental
dents who reported that they typically bike, take the bus or drive to variable. The alters' neighborhood and commute characteristics affect the
campus. The numbers of students using other modes are too small for mode choices of the alters but we do not expect characteristics of the alters'
inclusion in model estimations. In our first set of models (two-stage neighborhoods to directly affect the mode choice of the ego. One may be
residual inclusion models) it is necessary to include information about concerned that, when an ego lives in the same neighborhood as one of
the neighborhood characteristics of the alters. Therefore, we exclude his alters, they experience the same neighborhood characteristics.
alters outside of the Davis area because we have limited neighborhood However, we calculate the effect of the alters' neighborhood char-
information outside of the Davis area. The use of the alters' geographic acteristics on alters' mode choices for the group of alters collectively.
locations also excludes from the sample any ego that did not provide Alters with neighborhood characteristics conducive to biking are more
alter geographic information. There are 396 students that live within likely to bike, and alters with neighborhood characteristics more con-
Davis, and for which we have geographic, social network, and mode use ducive to other modes are less likely to bike. Therefore, the average of a
information. This sample is used in the analysis presented in Figs. 2 and neighborhood characteristic across an ego's alters is different from that
3, as well as in our two stage residual inclusion models. The sample is neighborhood characteristic for the ego, and can be used to predict the
further reduced for our multinomial logit model estimations if they are proportion of alters that bike, but not the ego's mode choice.
missing data on variables included in the models. This sample is used Our instrument is the alters' neighborhood characteristic: neigh-
for our multinomial logit models and, since we present many of the borhood biking density. All of the approximately 3,300 individuals who
variables used in these models, this sample is used for our descriptive participated in the CTS, our social networks survey, or are alters whose
statistics. usual commute mode and cross-streets were noted by an ego are po-
tential neighbors (no individual is his own neighbor). For each alter, we
3. Modeling approaches compute neighborhood biking density as the percentage of neighbors
within 0.5 miles of the alter, that bike. If there are 20 neighbors within
We use two modeling approaches to explore the effects of the in- 0.5 miles of an alter and 15 of them bike that alter has a neighborhood
teraction between commute distance and social influence on usual biking density of 0.75. For each ego-network we calculate the average
commute mode. First we use a two-stage residual inclusion model neighborhood biking density among the alters, who live in and around
(described in section 3.1), to model the binary choice to bike or not as Davis. We exclude alters outside of the Davis area because we have
usual commute mode. We estimate two-stage residual inclusion models limited neighborhood information for these alters; there may be only a
for each of three sample segments, based on commute distance. Next, few neighbors who participated in the surveys noted above, and these
we use multinomial logit models to estimate the interaction effects of few neighbors are not likely to accurately represent neighborhood
commute distance and social influence. In these models, we include biking density.
three transportation modes; bike, bus and drive.
3.1.2. Sample segments
3.1. Two stage residual inclusion model For the 2SRI model estimations, we divide the sample into three
segments based on commute distance; those within 1½ miles of
Egos who bike tend to have higher proportions of alters that bike, campus, those living 1½ to 3 miles of campus, and those living 3 to 5
than egos who drive or take the bus (Table 1, below). A similar pattern miles from campus. The last segment is slightly broader in terms of
occurs for egos that take the bus and drive (they have higher propor- distance, covering the remaining individuals who live within the City of
tions of alters that take the bus and drive, respectively) and is likely Davis; however, this segment has the fewest observations. Because
evidence of social influence. However, it may also be the result of other segmenting the sample results in fairly small samples we utilize only the
mechanisms. For example, an ego and his alters may make the same focal variables in these estimations; the instrumental variable, the
mode choice because they commute about the same distance, or have proportion of alters that bike and the binary choice of the ego to bike or
similar access to transit. The “shared environment” is a well-known not.
source of endogeneity in social influence models, which may entail a The average neighborhood biking density of the alters is used to
positive correlation between the behavior of the individual and her predict the proportion of alters that bike in the first stage of the 2SRI
network (Manski, 1993). model. The residuals from the first stage are saved, and entered into the
To control for the shared environment source of endogeneity, we second stage as an explanatory variable along with the endogenous
use a two-stage residual inclusion (2SRI) model, described as an in- variable, and any other predictors. The stage-1 residual acts as a control
strumental variable or control function approach for discrete choice variable for the proportion of alters biking, and captures the portion of
models (for a detailed discussion of the model and its properties see the error term that would be correlated with the social network variable
Rivers & Vuong, 1988, Wooldridge, 2010 chapter 15, and Train, 2009 due to shared environment.
Chapter 13). Using this approach, we incorporate the portion of the
error term that may be correlated with our explanatory variable of 3.2. Multinomial logit model
interest (mode choice of alters), as an additional explanatory variable,
such that the remaining part of the error term is not correlated (Train, Following the 2SRI model, we present a multinomial logit (MNL)
2009). When compared to related or similar modeling approaches, the model (see Ben-Akiva & Lerman, 1985) of the ego's mode choice.
two-stage residual inclusion model is preferred in many cases (see Though the endogeneity challenges noted above are still relevant in this
Rivers & Vuong, 1988; Terza, Basu, and Rathouz, 2008 and Lewbel, framework, in this model we are primarily interested in estimating the
Dong, and Yang, 2012). Guevara and Ben-Akiva (2010) compare this interaction between commute distance and social influence. Future
model to methods using latent variables; while the latent variables work in this area should incorporate interaction terms into discrete
approach is preferred, the use of both methods is supported by their choice models that simultaneously address endogeneity. In the MNL
results. model, we include the effect of distance on mode choice, the effect of
social influence on mode choice and the effect of the interaction of
3.1.1. Instrumental variable these two terms. Interaction terms are used when the effect of one in-
Any instrumental variable approach requires the identification of an dependent variable, x1 is expected to vary with respect to another in-
instrument that is correlated with the endogenous variable, but not dependent variable, x2. The coefficient on the interaction term is ne-
correlated with the unobserved (or observed) variation in the outcome gative if larger values of x2 result in a smaller effect of x1. The

5
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

coefficient is positive if larger values of x2 result in a larger effect of x1. Transportation and Parking Services program. GoClub members
We expect the coefficient on the interaction of commute distance commit to using a mode of transportation other than driving for their
and social influence to be negative for the mode choice of bike. Since commutes to campus and receive benefits depending on which mode of
our base alternative in the MNL model is bike, the coefficient value is transportation they commit to. For example, those outside of Davis who
expected to be positive for the other two modes in the model. That is, at commit to taking the train receive discounted tickets. Social network
longer commute distances, the effect of social influence on the choice to variables are the proportion of alters using each mode of transportation.
bike is smaller, and respondents are more likely to choose to take the Attitudes and preferences are based on five-point Likert-type scales. The
bus or drive, than bike, even when they have high proportions of alters ‘importance of’ survey questions asked respondents to indicate, “How
that bike. Although we also hypothesize at very short commute dis- important to you are the following factors when choosing your usual
tances the effect of social influence is small, the interaction term does means of transportation to travel to work or school?” (1 = not im-
not capture the potentially reduced effect of social influence for those portant, 5 = extremely important). The ‘agreement with’ CTS questions
who have very short commute distances. We also include socio-demo- asked respondents to indicate “To what extent do you agree or disagree
graphic variables, attitudes and trip characteristics as explanatory with the following statements?” (1 = strongly disagree, 5 = strongly
variables. agree).
For the variables in Table 1, we conducted a chi-squared test for
each binary variable; to determine whether the distributions across
4. Results categories is significantly different with respect to usual transportation
mode. For example, a significantly higher proportion of women drive or
4.1. Descriptive statistics take the bus than bike. We conducted analysis of variance for each
continuous variable; on average, those who bike have a significantly
Summary statistics of variables in our analysis, including socio-de- shorter distance to campus than those who drive or take the bus.
mographics, commute behaviors, preferences and attitudes are pre- In our sample, 72% are female and 71% are undergraduates. About
sented in Table 1. Binary variables are equal to 1 if the respondent has 20% of the respondents are members of the goClub. The mean distance
the noted characteristic (0 otherwise). Annual parking permits allow to campus is slightly less than two miles across the full sample. The
parking in designated areas on campus. The goClub is a UC Davis mean age of 22 is expected for a sample of university and graduate
students, as is the average number of years at UC Davis; two. The
Table 1 proportion of alters using each mode of transportation is computed with
Selected sample characteristics by respondent usual commute mode. only the alters for which the ego noted a usual mode of transportation,
Bike; 66% Drive; 7% Bus; 26% Total and live in the Davis area. On average, respondents have higher pro-
(N = 250) (N = 23) (N = 91) (N = 364) portions of alters that bike than proportions of alters that drive or take
the bus, and those that take the bus have higher proportions of alters
Demographics and Trip Characteristics – Binary
that take the bus. Some of the attitudes and preferences are different
Female (p = 0.003) 162 (67%) 20 (87%) 73 (84%) 255 (72%) across mode choices; drivers indicated a greater importance of “the
Has annual parking permit 2 (1%) 11 (48%) 1 (1%) 14 (4%) time it takes to make the trip”, and a lower importance of “the cost of
(p < 0.001) owning a car or other vehicle”. Agreement with liking biking and liking
Member of GoClub 52 (21%) 3 (14%) 11 (13%) 66 (19%)
(p = 0.202)
transit are highest for bikers and bus riders, respectively.
Undergraduate Student 165 (66%) 9 (39%) 85 (93%) 259 (71%)
(p < 0.001) 4.2. Commute distance, social influence and bike mode share
Demographics and Trip Characteristics – Continuous
Next we examine the relationships between the variables central to
Distance to Campus 1.77 2.31 2.19 1.83
(p < 0.001) our analysis; bicycle mode share among egos, proportions of alters
Age (p = 0.001) 22.77 24.57 20.84 22.27 biking, and commute distance. The ego bike mode share was computed
Years at UC Davis (p = 0.459) 2.14 2.87 2.22 2.02 by constructing commute distance bins by rounding ego commute dis-
Social Network tance to campus to the nearest 10th of a mile. Bins with fewer than 10
Proportion of alters biking 61% 42% 31% 53%
egos were combined and we computed the mean distance to campus for
(p < 0.001) egos in the resulting bins. For all bins we calculated the ego bike mode
Proportion of alters taking bus 16% 9% 52% 25% share; the percentage of egos in the bin that commute by bike as their
(p < 0.001) usual mode. In general, the higher the commute distance the lower the
Attitudes and Preferences
ego bike mode share (Fig. 2).
Importance of: In order to detect social influence, we must confirm that there is
“Time it takes to make the 4.09 4.65 4.35 4.19 variance in the modes of alters, regardless of the distance an ego lives from
trip” (p = 0.003) campus. If all students who live far from campus only have friends who
“Cost of owning a car or other 3.39 2.61 3.66 3.40 drive, it would be impossible for their friends to exert social influence for a
vehicle” (p = 0.297)
“Safety” (p < 0.001) 3.26 3.91 3.86 3.45
different mode choice. For each distance bin, we found the mean pro-
“Environmental impacts” 3.21 2.74 3.00 3.13 portion of alters biking among the egos in that bin. Fig. 2 presents the
(p = 0.085) relationships between ego bike mode share and the mean proportion of
“Cost of transit” (p < 0.001) 2.97 1.87 3.90 3.13 alters biking with respect to the each bin's mean distance to campus (in
“Getting physical exercise 3.03 2.26 2.18 2.77
miles). Linear models were estimated for both relationships. Model results
during my commute”
(p < 0.001) are plotted and the equations are presented below. All coefficients in both
Agreement with: models are significant at 0.1%, however in the model of the proportion of
“I like biking” (p < 0.001) 4.44 3.18 3.56 4.14 alters biking, the coefficient for distance is very small. There is a fairly
“I like transit” (p < 0.001) 3.10 2.30 3.81 3.23 high overall ego bike share as well as a moderate relationship between
“I need a car to do many of the 3.06 4.35 3.55 3.26
things” (p < 0.001)
distance and bike share. Distance explains 35% of the variation in ego bike
share. The proportion of alters that bike is relatively stable with respect to
p-values: ANOVA for continuous variables or chi-squared tests of distributions ego distance to campus; the coefficient is very small, and the model ac-
for binary variables, across modes. counts for only 16% of the variation.

6
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

average proportion of alters taking bus


≈ 0.14 + 0.05 × distance bin (adj. R2 = .13); N = 396

Figs. 2 and 3 show ego mode shares for bike and bus, and mean
proportions of alters that bike or take the bus relative to ego commute
distance. Although the mean proportion of alters using each mode
changes with ego distance to campus, there is a weaker relationship
between ego commute distance and alter behaviors, than there is be-
tween ego commute distance and ego behaviors. Although egos who
live further from campus tend to have more alters that take the bus, and
fewer that bike, there is more variation in ego mode choices, than alter
behaviors, with respect to ego commute distance. Therefore, there must
be other relevant variables to explain similarities in ego and alter be-
havior.

4.3. Two-stage residual inclusion model

The 2SRI model results are presented in Table 2 and demonstrate


how the effect of social influence varies with respect to distance from
campus. Model estimations were carried out using the statistical soft-
ware R (R Core Team, 2013). The model results in coefficients scaled by
2 1/2 2
Fig. 2. Share of egos/mean proportion of alters biking by ego commute dis- a factor of (1 + θˆ τ 2) ; where θ̂ is the coefficient estimate for the
ρ1 2 ρ1
tance.
stage 1 residuals in stage 2, and τ22 is the variance of the stage 1 re-
siduals (Wooldridge, 2010). Both the scaled and unscaled parameter
ego bike share ≈ 0.80 − 0.10 × distance bin (adj. R2 = .35); N = 396 estimates of the intercept and the proportion of alters biking are shown
in Table 2. The coefficient for the stage 1 residuals is not scaled. We
average proportion of alters biking utilize bootstrapped standard errors for the scaled parameter estimates
≈ 0.64 − 0.06 × distance bin (adj. R2 = .16); N = 396 since estimates are not, otherwise, strictly valid (for details see
Wooldridge, 2010). The proportion of alters biking is significant for the
A similar pattern occurs, though in the opposite direction, for bus. first two models, and likely not significant for the third segment due to
Fig. 3 shows ego bus share and the mean proportion of alters taking the sample size. These results confirm earlier findings that social influence
bus for each commute distance bin. Ego bus share and the mean pro- is relevant to transportation mode choice, even when accounting for
portion of alters taking the bus were computed for bus as they were for endogeneity.
bike. As distance increases, ego bus share increases, as does (though to a The significance of the stage 1 residual coefficient estimate may be
lesser extent) the mean proportion of alters that take the bus. The linear used as a test of exogeneity for the explanatory variable of interest
models explain 28% of the variation in ego bus share and only 13% of (Wooldridge, 2010); in our case the proportion of alters biking. Even
the variation in the mean proportion of alters that take the bus. The though the stage 1 residual is not a significant predictor in the three
coefficient in the model of the mean proportion of alters taking the bus sample segment models, we don't take this as evidence of exogeneity
is very small, though all coefficients are significant at 0.1%. since this test for exogeneity fails for the full sample model. The F
ego bus share ≈ 0.09 + 0.09 × distance bin (adj. R2 = .28); N = 396 statistic, a criteria for determining the validity of an instrument (Staiger
& Stock, 1997), is less than 10 for the second model, though it is greater
than 10 in the other three model estimations. In general this indicates
that neighborhood biking density is a valid instrument for the propor-
tion of alters that bike. The stage 1 residuals are not strictly normally
distributed, likely because the proportion of alters that bike is not
strictly continuous.
We cannot directly compare the coefficient values across models,
since they each have a different sample (Train, 2009). In order to
evaluate differences in the importance of social influence across the
segments, we compute the marginal effects of social influence on the
choice to bike, for each segment (Table 3). Marginal effects were
computed for each individual, and then the average was taken. We used
bootstrapped standard errors to compute statistical tests for the mar-
ginal effects values. We cannot be certain of the marginal effects for all
segments, because of the lack of significant results. However we can
conclude that for those who live closest to campus, the decision to bike
is less affected by social influence than it is for the full sample. This
result suggests that there are different responses to social influence for
the three sample segments.

4.4. Multinomial logit model results

Multinomial logit model (MNL) results are summarized in Table 4.


Fig. 3. Share of egos and proportion of alters taking bus by ego commute dis- The base alternative in this model is bike. The mode shares are some-
tance. what imbalanced; the model for drive has 8 variables for only 23 cases.

7
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

Table 2
Two-stage residual inclusion models of Ego's choice to bike.
Model - Segment Bike share Stage 2 Binary Probit Model Coefficients: F statistic (stage 1 instrument)
N (%)
Intercept Z (p-value) alters biking Z (p-value) stage 1 residual
unscaled (scaled) unscaled (scaled)

0 to 1½ miles 109 (76%) −0.468 −0.874 1.835 2.498 −1.330 22.79


(N = 142) (−0.508) (p = 0.382) (1.992) (p = 0.013) (p = 0.166)
1½ to 3 miles (N = 207) 126 (61%) −0.936 −2.536 2.629 3.537 −2.00 5.89
(−1.164) (p = 0.011) (3.270) (p < 0.001) (p = 0.175)
3 to 5 miles 22 (47%) −0.784 −1.171 1.585 1.050 −1.769 10.93
(N = 47) (−0.897) (p = 0.241) (1.812) (p = 0.294) (p = 0.215)
Full sample 257 (65%) −0.788 −4.525 2.233 7.78 −1.552 48.327
(N = 396) (−0.897) (p < 0.001) (2.541) (p < 0.001) (p = 0.007)

Table 3 factors, and household size. We also ran estimations with travel time for
Two-stage residual inclusion model marginal effects.a each mode, however, for our sample travel time was highly correlated
Model - Segment Bike share Stage 2 Binary Probit Model Marginal
with distance, and to avoid multi-colinearity issues, we do not include
N (%) Effects travel time in our final estimations. In our final estimations we retained
those variables which are expected to be important controls for factors
alters biking other than distance and social influence.
0 to 1½ miles (N = 142) 109 (76%) 0.575 (p = 0.074)
Of primary interest are the effects of social influence and commute
1½ to 3 miles (N = 207) 126 (61%) 1.111 (p = 0.243) distance, and the interaction of these two variables. Those with higher
3 to 5 miles (N = 47) 22 (47%) 0.696 (p = 0.246) proportions of alters biking are more likely to bike than either of the
Full sample (N = 396) 257 (65%) 0.850 (p < 0.001) alternative modes (drive or bus). Living further from campus may in-
crease the probability of choosing bus or drive, rather than bike, though
a
Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1.
in this model, estimates are not significant. The coefficient on the in-
teraction of distance and the proportion of alters biking is positive for
The ρ2 value of 0.35 indicates that the model explains a moderate
the choice of bus. Related to biking, this term indicates that as distance
amount of information compared to a model with constants only. The
increases, the effect of the proportion of alters that bike, on the choice
likelihood ratio test compares the model presented to a model without
to bike rather than bus, decreases. The p-value on this coefficient is only
the interaction term (but with distance and the proportion of alters
0.111, and the standard error is 0.512; small, compared to the coeffi-
biking), and indicates that the interaction term adds some explanatory
cient value of 0.815. The p-value may be affected by the non-linear
power to this model.
relationship we expect for the interaction of distance and social influ-
In general, the coefficients have the expected signs. Those who in-
ence to bike; at very small commute distances the effect of social in-
dicate transit costs are important are more likely to take the bus than
fluence should theoretically, be small. Nonetheless, the estimate likely
bike, while those who indicated lower importance are more likely to
indicates the true overall direction of the effect, and with a reasonable
drive than bike. The higher the importance of safety the more likely a
level of confidence. Not so for the coefficient on the interaction term for
respondent drives, and possibly takes the bus. Undergraduates are more
the choice of drive; the p-value and standard error are very large and
likely to take the bus than graduate students. Those who need a car are
we expect this coefficient to be positive.
more likely to drive than bike. Other variables included in model esti-
We estimated two alternative specifications of this model; one with
mations, but not found to have significant relationships with mode
only the choices of bike and bus, because the drive share is very small,
choice for our sample, included gender, income, additional attitudinal
and one where we relaxed the restriction that respondents must live in

Table 4
Model 1: Multinomial logistic regression of mode choice (bike, drive, bus).
Full Model Sample is 364 Drive Bus

Base Alternative Bike; N = 23 (6%) N = 91 (25%)


N = 250 (69%)

Variables in Model (Coefficients reflect effect Estimate (Std. error) Estimate (Std. error)
on the utility of choosing drive or bus)
Proportion of alters biking 0.983 (2.188) −3.561** (1.169)
Ego's distance to campus 0.943 (0.578) 0.458 (0.290)
Interaction: alters biking × distance −0.821 (0.917) 0.815 (0.512)
Ego is an undergraduate student −1.184. (0.635) 2.307*** (0.491)
Importance of “Safety” 0.572* (0.257) 0.22. (0.132)
Importance of “Transit costs” −0.677** (0.243) 0.465*** (0.112)
Ego has annual parking permit 4.449*** (1.017) 1.342 (1.305)
Intercept −4.293* (1.773) −5.313*** (1.007)

Model Diagnostics

Log likelihood: Rho-squared Likelihood ratio test statistic: 3.77;


−183.902 (market share Pr(> Chisq) < 0.151
base): 0.35

Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1.

8
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

Table 5
Select terms from models 2 and 3: Alternative MNL models of mode choice.
Model 2: Model Including Respondents Outside of Davis

Model Sample is 385 Drive Bus


Base Alternative Bike; N = 252 (65%) N = 37 (10%) N = 96 (25%)
Variables in Model (Coefficients reflect effect Estimate (Std. error) Estimate (Std. error)
on the utility of choosing drive or bus)
Proportion of alters biking −2.195. (1.188) −3.426*** (0.746)
Ego's distance to campus 0.403*** (0.111) 0.176. (0.106)
Interaction: alters biking × distance 0.641* (0.31) 0.728* (0.288)

Model 3: Model with Bike and Bus Only

Model Sample is 341 Drive Bus


Base Alternative Bike; N = 250 (73%) – N = 91 (27%)
Variables in Model (Coefficients reflect effect - Estimate (Std. error)
on the utility of choosing drive or bus)
Proportion of alters biking – – −3.746** (1.18)
Ego's distance to campus – – 0.412 (0.287)
Interaction: alters biking × distance – – 0.906. (0.514)

Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

Davis. This allows individuals with longer commute distances to be in but for most respondents, these are not very accessible. Lastly, though
the sample; however, commute distance is not continuously distributed we did correct for potential endogeneity between the social network
since anyone who lives outside of Davis lives at least 10 miles from behaviors and the ego's behavior in our 2SRI model, we were not able to
campus. The variables of primary interest are presented in Table 5. The correct for this issue in our MNL models. There are also other potential
results of these alternative specifications provide further evidence that sources of endogeneity, such as homophily, that are difficult to control
the effect of social influence on transportation mode choice is impacted for with cross-sectional data. However, shared environment is probably
by commute distance. the most important source of endogeneity in travel behavior.

5. Discussion and conclusions


5.2. Directions for future work
Our results provide support for the hypothesis that the effect of
social influence is impacted by contextual factors, such as commute Future work should explore this question in areas with a more
distance. In our 2SRI models, for the segment with the shortest com- continuous distribution of commute distances and more balanced mode
mute distances the marginal effect of social influence was moderate. shares. Further, we found that social influence has most effect on in-
The model for those with a medium commute distance had the greatest dividuals with medium commute distances; it may be of interest to
marginal effect of social influence, and the segment furthest from more precisely identify bounds within which social influence to bike is
campus had the smallest. In our multinomial logit models, the inter- most important, or how these bounds might change depending on
action term of distance with social influence is not significant for all transportation mode. Future work should also investigate how other
models, but is in the expected direction. In model 1, the low p-value relevant factors in transportation mode choice interact with and impact
supports our hypothesis that as commute distance increases, response the effects of social influence. Responses to social influences may vary
to social influence is reduced. Taken together with the results of models by attitudes, land use characteristics, travel time, car ownership, gender
2 and 3, our findings corroborate the hypothesis that as distance in- or other factors. Interactions with the other categories of variables ex-
creases, the effect of social influence on the choice of bike decreases. pected to affect social influence should also be explored. These include
individual traits, such as familiarity with biking that may reduce the
5.1. Limitations of the present study effects of social influence, and relationship characteristics. Egos may be
more influence by those alters with whom they are closer.
Though our findings suggest that social influence is affected by As more complex modeling frameworks are developed for discrete
commute distance, there are important limitations of the study. Our choice models, these advances should be applied to this question, and
findings are limited by the study area; no one in our sample lives be- explore the interaction of social influence with other factors, while
tween 5 and 10 miles from campus, distances that might be particularly addressing potential endogeneity problems. One avenue for future re-
relevant. The mode shares in our sample are not typical for most areas search would be to conduct panel studies, and observe changes in
of the US, or even California; UC Davis is known for high levels of bi- networks as well as travel behavior over time. This would allow for
cycling, and more than 50% of our sample bikes as their usual commute improved tracking of social influence, and provide a means for over-
mode. In other words, Davis provides a strategic research site for ex- coming the endogeneity challenges mentioned above.
amining our hypotheses because there are a range of conditions where A better understanding of how social influence is related to trans-
the costs/benefits of different mode choices are relatively similar, such portation mode choice is impacted by external factors will allow sus-
that social influence may be observed. In travel contexts dominated by tainable transportation programs, such as May is Bike Month, or pro-
only one mode, where the costs of alternative modes are exceptionally grams where social networks are used to recruit participants (e.g. UC
high, social influence is less likely to be observed. This unique sample Davis goClub) to more effectively utilize social influence as a tool in
made it possible to explore the research question of interest, but the promoting the use of alternative modes of transportation. For example,
generalizability of these results may be limited. if social influence to bike is only relevant for individuals expected to
Another limitation is that in our multinomial logit model that in- have commute distances within a specific range, these are the in-
cludes all respondents, both inside and outside of Davis, we did not dividuals that should be targeted by social network-based programs
account for choice set differences experienced by these respondents. aimed at increasing bicycle commuting.
Those living outside of Davis have some public transportation options,

9
S. Pike, M. Lubell Research in Transportation Economics 68 (2018) 2–10

References network societies. Information, Communication & Society, 11(5), 640–658.


Lewbel, A., Dong, Y., & Yang, T. T. (2012). Comparing features of convenient estimators
for binary choice models with endogenous regressors. Can. J. Econ. 45(3), 809–829.
Axelrod, L. J., & Lehman, D. R. (1993). Responding to environmental concerns: What Maness, M., Cinzia, C., & Dugundji, E. R. (2015). Generalized behavioral framework for
factors guide individual action? Journal of Environmental Psychology, 13(2), 149–159. choice models of social influence: Behavioral and data concerns in travel behavior.
Axhausen, K. W. (2008). Social networks, mobility biographies, and travel: Survey Journal of Transport Geography, 46, 137–150.
challenges. Environment and Planning B: Planning and Design, 35(6), 981–996. Manski, C. (1993). Identification of endogenous social effects: The reflection problem. The
Axsen, J., & Kurani, K. S. (2011). Interpersonal influence within car buyers' social net- Review of Economic Studies, 60, 531–542.
works: Applying five perspectives to plug-in hybrid vehicle drivers. Environment and Mokhtarian, P., & Salomon, I. (1997). Modeling the desire to telecommute: The im-
Planning A, 44(5), 1047–1065. portance of attitude factors in behavioral models. Transportation Research Part A,
Bamberg, S., Hunecke, M., & Blobaum, A. (2007). Social context, personal norms, and the 31(1), 35–50.
use of public transportation; two field studies. Journal of Environmental Psychology, Páez, A., & Scott, D. M. (2007). Social influence on travel behavior: A simulation example
27(2007), 190–203. of the decision to telecommute. Environment and Planning A, 39(3), 647–665.
Ben-Akiva, M., & Lerman, S. R. (1985). Discrete choice analysis. Cambridge MA: MIT Press. Páez, A., Scott, D. M., & Volz, E. (2008). A discrete-choice approach to modeling social
Carrasco, J. A., & Miller, E. J. (2006). Exploring the propensity to perform social activ- influence on individual decision making. Environment and Planning B: Planning and
ities: A social network approach. Transportation, 33(5), 463–480. Design, 35, 1055–1069.
Carrasco, J. A., & Miller, E. J. (2009). The social dimension in action: A multilevel, Park, C., & Lessig, V. (September 1977). Students and housewives: Differences in sus-
personal networks model of social activity frequency between individuals. ceptibility to reference group influence. Journal of Consumer Research, 4.
Transportation Research Part A: Policy and Practice, 43(1), 90–104. Pike, S. (2014). Travel mode choice and social and spatial reference groups; comparison
Carrell, S. E., Fullerton, R. L., & West, J. E. (2009). Does your cohort matter? Measuring of two formulations. Transportation research record: Journal of the transportation re-
peer effects in college achievement. Journal of Labor Economics, 27(3), 439–464. search board, No. 2412 (pp. 75–81). .
Dugundji, E. R., & Walker, J. L. (2005). Discrete choice with social and spatial network Pike, S., & Lubell, M. (December 2016). Geography and social networks in transportation
Interdependencies: An empirical example using mixed generalized extreme value mode choice. Journal of Transport Geography, 57, 184–193.
models with field and panel effects. Transportation research Record: Journal of the Popovich, N. (2014). Results of the 2013–14 campus travel surveyResearch Report UCD-ITS-
transportation research board, No. 1921 (pp. 70–78). Washington, D.C: Transportation RR-14-14. Davis: Institute of Transportation Studies, University of California.
Research Board of the National Academies 1921(1). R Core Team (2013). R: A language and environment for statistical computing. Vienna,
Efferson, C., Lalive, R., Richerson, P. J., McElreath, R., & Lubell, M. (2009). Conformists Austria: R Foundation for Statistical Computing. http://www.R-project.org/.
and mavericks: The empirics of frequency-dependent cultural transmission. Evolution Rivers, D., & Vuong, Q. H. (1988). Limited information estimators and exogeneity tests for
and Human Behavior, 29(2008), 56–64. simultaneous probit models. Journal of Econometrics, 39, 347–366.
Ewing, R., & Cervero, R. (2010). Travel and the built environment. Journal of the American Sacerdote, B. (2001). Peer effects with random assignment: Results for dartmouth
Planning Association, 76(3), 265–294. roommates. Quarterly Journal of Economics, 116(2), 681–704.
Farber, S., & Páez, A. (2009). My car, my friends, and me: A preliminary analysis of Scott, D. M., Dam, I., Páez, A., & Wilton, R. (2012). Investigating the effects of social
automobility and social activity participation. Journal of Transport Geography, 17(3), influence on the choice to telework. Environment and Planning A, 44(5), 1016–1031.
216–222. Sherwin, H., Chatterjee, K., & Jain, J. (2014). An exploration of the importance of social
Fowler, J. H., & Christakis, N. A. (2008). Dynamic spread of happiness in a large social influence in the decision to start bicycling in England. Transportation Research Part A:
Network: Longitudinal analysis over 20 years in the framingham heart study. British Policy and Practice, 68, 32–45.
Medical Journal, 337(a2338), 1–9. Staiger, D., & Stock, J. H. (1997). Instrumental variables regression with weak instru-
Freeman, L. C. (1984). The impact of computer based communication on the social ments. Econometrica, 65(3), 557–586.
structure of an emerging scientific specialty. Social Networks, 6(3), 201–221. Terza, J. V., Basu, A., & Rathouz, P. J. (2008). Two-stage residual inclusion estimation:
Goetzke, F. (2008). Network effects in public transit use: Evidence from a spatially au- addressing endogeneity in health econometric modeling. Journal of Health Economics,
toregressive mode choice model for New York. Urban Studies, 45(2), 407–417. 27(3), 531–543.
Goetzke, F., & Rave, T. (2010). Bicycle use in Germany: Explaining differences between Train, K. E. (2009). Discrete choice methods with simulation. Cambridge University Press.
municipalities with social network effects. Urban Studies, 48(2), 427–437. Walker, J. L., Ehlers, E., Banerjee, I., & Dugundji, E. R. (2011). Correcting for endogeneity
Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), in behavioral choice models with social influence variables. Transportation Research
1360–1380. Part A: Policy and Practice, 45(4), 362–374.
Guevara, C. A., & Ben-Akiva, M. (2010). Addressing endogeneity in discrete choice Wilton, R. D., Páez, A., & Scott, D. M. (2011). Why do you care what other people think?
models: Assessing control-function and latent-variable methods. In Stephane Hess, & A qualitative investigation of social influence and telecommuting. Transportation
Andrew Daly (Eds.). Choice Modelling: The State-of-the-Art and The State-of-Practice Research Part A: Policy and Practice, 45(4), 269–282.
(pp. 353–370). . Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd ed.).
Koehly, L. M., & Loscalzo, A. (2009). Adolescent obesity and social networks. Preventing Cambridge, Massachusetts and London, England: The MIT Press.
Chronic Disease, 6(3), 1–8.
Larsen, J., Urry, J., & Axhausen, K. (2008). Coordinating face to face meetings in mobile

10

You might also like