Content uploaded by Francesco Piras
Author content
All content in this area was uploaded by Francesco Piras on Dec 21, 2020
Content may be subject to copyright.
1
Do psycho-attitudinal factors vary with individuals’ cycling
1
frequency? A hybrid ordered modeling approach
2
3
Francesco Piras
4
Corresponding author
5
University of Cagliari
6
Via San Giorgio 12, Cagliari, 09124, Italy
7
Tel: (+39) 070-6756405; Email: francesco.piras@unica.it
8
9
Eleonora Sottile
10
University of Cagliari
11
Via San Giorgio 12, Cagliari, 09124, Italy
12
Tel: (+39) 070-6756405; Email: esottile@unica.it
13
14
Italo Meloni
15
University of Cagliari
16
Via San Giorgio 12, Cagliari, 09124, Italy
17
Tel: (+39) 070-6756403; E-mail: imeloni@unica.it
18
19
20
2
Do psycho-attitudinal factors vary with individuals’ cycling
1
frequency? A hybrid ordered modeling approach
2
3
HIGHLIGHTS
4
5
• We examine the relationship between psycho-attitudinal factors and cycling frequency
6
• Psycho-attitudinal variables positively influence propensity to cycle
7
• Link between different socio-demographic variables, built environment characteristics and bike usage
8
• Improvement of cycling infrastructure might not be sufficient to encourage bike use
9
10
ABSTRACT
11
The purpose of the present study was to investigate specifically whether psycho-attitudinal factors could
12
differ for people with different cycling frequency levels and to quantify the determinants influencing the
13
propensity to cycle. To perform our analysis, we developed a hybrid choice modeling approach with a
14
generalized ordered probit choice kernel, using the information collected in 2016 for 2,128 individuals in
15
two mid-size urban areas in Sardinia (Italy). Our results indicate that the latent variables perception of
16
cycling benefits, perception of cycling comfort and perceived importance of bike infrastructure positively
17
influence the propensity to cycle, supporting the idea of a relationship between attitudes and cycling
18
frequency. In addition, the model shows a link between different socio-demographic variables (gender, age,
19
Body Mass Index, education level, number of cars per household, number of household members), built
20
environment characteristics and bike usage. Computation of the pseudo-elasticity effects indicates that
21
strategies focusing only on the physical part of the problem, such as the expansion and improvement of
22
proper infrastructure, might not be sufficient to encourage bike use. At the same time our findings stress the
23
importance of considering people’s psychological characteristics when implementing policies aimed at
24
promoting cycling. This can be helpful for identifying, depending on the population segment that is targeted,
25
the most appropriate advertising/information strategy for convincing people to cycle, as well as the most
26
effective marketing messages.
27
Keywords: cycling behavior, cycling frequency, psycho-attitudinal factors, hybrid choice models,
28
generalized ordered probit
29
30
31
32
3
1. INTRODUCTION
1
In recent years, with a view to reducing car dependence and encouraging the use of more sustainable modes
2
of transport, cycling mobility has been receiving considerable attention in many countries. A growing body
3
of literature has investigated which factors influence the propensity to use the bike (Heinen et al., 2010;
4
Handy et al., 2014; Muñoz et al., 2016; Ton et al., 2019). Indeed, the transport sector is a primary cause of
5
the observed deterioration in urban air quality (European Environmental Agency, 2019), road transport
6
having increased significantly. For example, in 2018 in Italy, where the number of cars saw an increase of
7
4.1% over 2014, air quality worsened, with levels of PM10 and PM2.5 far higher than the standards set by both
8
the European Union and the World Health Organization (ISFORT, 2018). Not less importantly,
9
transportation also generates several other issues that impact on the environment and urban life, including
10
noise pollution, public health and safety.
11
One particularly interesting aspect of recent research on cycling mobility is the recognition that, in
12
addition to objective variables (such as socio-demographics, built environment and trip characteristics),
13
psycho-attitudinal factors can contribute to influencing the choice to travel by bike (Ewing and Cervero,
14
2010; Willis et al., 2015; Muñoz et al., 2016; Arroyo et al., 2020; Gutierrez et al., 2020). This recognition
15
may be attributed to the general finding that people may have different perceptions of cycling infrastructure
16
attributes (Ma et al., 2014). Also, Heinen et al. (2011) observed that individuals with similar socio-economic
17
and cycling environment characteristics are likely to show quite different cycling behavior.
18
Most studies focusing on cycling attitudes and perceptions have examined the difference between
19
cyclists and non-cyclists (Gatersleben and Appleton, 2007). However, it may be useful to distinguish among
20
different types of cyclists (Dill and McNeil, 2013). For instance, one would expect that individuals who are
21
more aware of the benefits associated with cycling, do so more frequently. Some empirical studies have
22
investigated, from a modeling perspective, the association between the effect of psycho-attitudinal variables
23
and cycling frequency. However, several of these have considered the frequency variable as a continuous
24
variable, despite it being measured in ordinal discrete categories, which is inappropriate from an econometric
25
point of view (Bhat et al., 2017). In other works, psycho-attitudinal factors were obtained through some
26
scoring scheme or directly included as exogenous variables in the empirical models, an approach that can
27
potentially lead to measurement errors and result in inconsistent estimates (Walker, 2001).
28
A much more systematic approach would be advantageous for both policy makers and planners. In fact,
29
the implementation of policies aimed at promoting bike use, such as marketing/information campaigns, can
30
benefit from a deeper understanding of this phenomenon. This in turn would help to avoid misspending
31
financial resources on inadequate measures and to diminish the risk of failures that result in reduced public
32
support (Handy et al., 2014).
33
In light of these considerations, the aim of our work is to study whether psycho-attitudinal factors vary
34
among people with different frequencies of cycling for any purpose and to explore the effects of socio-
35
demographic and built environment characteristics on the propensity to cycle. In an effort to better
36
understand to what extent psycho-attitudinal variables affect cycling frequency, we estimated a Hybrid
37
4
Choice Model (Vij and Walker, 2016) with a generalized ordered probit choice (Greene and Hensher, 2010)
1
kernel. Generalization of the ordered response model renders the thresholds themselves dependent upon both
2
the objective and psycho-attitudinal variables, allowing us to account for systematic heterogeneity across
3
individuals. Further, in our approach, psycho-attitudinal variables are specified as latent variables dependent
4
upon some sociodemographic variables. This approach to modeling the latent variables allowed us to
5
understand the direct and indirect effects of sociodemographic variables on the propensity to cycle. Running
6
different test scenarios, revealed that the probability to cycle with a certain frequency changes following a
7
change in the sociodemographic variables defining the latent variables.
8
The econometric model is estimated using a dataset collected in the urban areas of Cagliari and Sassari,
9
main cities in Sardinia (Italy), where, despite the implementation of policies encouraging bike use, cycling
10
levels are still low. The model clearly provides evidence that individuals with a greater perception of cycling
11
benefits and comfort and who attach greater importance to the presence of proper infrastructure, cycle more.
12
In addition, the model reveals a link between different socio-demographic variables, built environment
13
characteristics and cycling frequency.
14
The remainder of the paper is structured as follows. In section 2 we provide a literature review to set the
15
context of the current investigation. Section 3 presents the exploratory data analysis. Next, in sections 4 and
16
5, we illustrate the methodological approach used to perform our analysis and discuss the model estimation
17
results. Finally, section 6 provides conclusions and identifies the study’s limitations.
18
2. CONTEXT OF CURRENT INVESTIGATION
19
The relationship between psycho-attitudinal factors and cycling frequency has been reported in different
20
studies. Some works investigated how individuals’ propensity to cycle is influenced by their perceptions of
21
traffic risks and cycling facilities. Gatersleben and Appleton (2007) investigated attitudes and perceptions of
22
people in different behavioral stages of change of the Transtheoretical Model in relation to cycling to work.
23
They found that as the level of cycling experience increases so their perceptions of various personal and
24
external barriers change, and they are more likely to recognize the benefits of cycling. Sener et al. (2009)
25
showed that the perception of the quality of cycling facilities significantly affects commute cycling
26
frequency, but this variable was not relevant for non-commuting. Handy et al. (2010) indicated that people
27
with higher levels of cycling comfort are more likely to cycle regularly for transportation than non-cyclists
28
and non-regular cyclists. Ma et al. (2014) tried to disentangle the effect of objective built environment and
29
perception thereof. They showed that the perception of the cycling environment has a positive and direct
30
impact on frequency, while the direct effect of the objective environment became insignificant when
31
including the perception effect. Regarding the perception of safety, Sallis et al. (2013) found that this
32
construct correlated with cycling frequency, and individuals would have cycled more had it been safer than
33
the car. Recently, Kelarestaghi et al. (2019) indicated that the latent risk factor, related to indicators such as
34
theft and road safety, has a negative effect on cycle-to-campus frequency for university students in the
35
Maryland Metropolitan Area.
36
5
Other works have focused on the effects of cycling attitudes, which refer to how people generally view
1
bike use. Heinen et al. (2011) studied differences in attitudes and norms between full-time and part-time
2
commuter cyclists, for different distance categories. They found that habits, subjective norms and the
3
recognition of cycling benefits positively influence the likelihood of cycling full-time to work. Further, they
4
showed that individuals who commute over longer distances have, on average, a more positive attitude
5
towards cycling. A study by Swiers et al. (2017) of a sample of University students in the UK found that
6
regular cyclists (daily/weekly) were significantly more likely to perceive health benefits as a motivator than
7
monthly/annual cyclists. Interestingly, Kroesen et al. (2017), using a panel dataset gathered in the
8
Netherlands, indicated that cycling behavior and attitudes mutually influence each other over time, and the
9
effect of cycling on attitudes is stronger than vice versa. Kaplan et al. (2019) found that cycling has the
10
potential for making people feel better about themselves from a physical and psychological perspective.
11
Recently some papers have introduced into their analysis the notion of experience, that represents a
12
concept closely related with frequency of use. A person who rides frequently can be considered an
13
experienced cyclist who is likely to feel more confident (Namgung and Jun, 2019). The same authors
14
examined attitudes towards cycling of cyclists with different experience levels among a sample of students
15
and staff members at Ohio University. They showed that the more experienced group of cyclists exhibited
16
more positive attitudes, while users with less experience are more likely to perceive barriers to cycling.
17
Thigpen (2019) explored whether attending a bike-friendly university, like UC Davies, led to high levels of
18
cycling and a change in cycling attitudes, and to what extent changes are influenced by personal cycling
19
experience. He found that riding a bike at any point during college increases both pro-bike attitudes and
20
cycling skills, while exposure to high levels of cycling appears not to influence attitudes or skills.
21
From a methodological viewpoint, different approaches have been used to explore the link between
22
exogenous variables and cycling frequency (see Table 1). Some studies have employed linear regression
23
models for their analysis, the dependent variable being the number of trips made by bike in a certain period
24
of time (Sallis et al., 2013, Stinson et al., 2014) or the weekly miles of transportation and recreation cycling
25
(Xing et al., 2010). Other studies have used econometric models including multinomial logit, ordered logit
26
and ordered probit. Stinson and Bhat (2004) were the first to estimate an ordered logit model to investigate
27
which factors influence frequency of bike use for commute to and from work. Similarly, Noland et al.
28
(2011), estimating an ordered probit model, investigated which factors are associated with cycling behavior
29
in New Jersey. Manaugh et al. (2017) categorized cycling frequency into four different classes of never,
30
rarely, usually, and always and used a multinomial logit to run their analysis. Interestingly, Bhat et al. (2017)
31
proposed a new spatial generalized ordered response model with skew-normal kernel error terms and applied
32
it to the analysis of cycling frequency. Recently, Oliva et al. (2019) estimated a latent class ordered logit to
33
segment neighborhoods, according to the cycling behavior observed in them, as a function of built
34
environment attributes.
35
The first to include psycho-attitudinal characteristics in their modeling framework were Sener et al.
36
(2009). They estimated two panel ordered probit models for analyzing the frequency of bike commuting and
37
6
non-commuting. In a similar vein, Fu and Farber (2017) used an ordered probit model to study the influence
1
of cycling safety, benefits and comfort on cycling frequency for commuting. Handy et al. (2010) estimated a
2
nested logit to examine the objective and subjective factors influencing bike ownership, use and frequency.
3
Heinen et al. (2011) and Namgung and Jun (2019) employed multiple binary logit models to explore the
4
difference in attitudes between cyclists and non-cyclists, and between experienced and inexperienced ones.
5
One limitation of the studies described above is the use of a two-stage sequential approach for their model
6
estimation. Typically, they first obtain the value of psycho-attitudinal factors, identified by an exploratory
7
factor analysis, through some scoring scheme, for example Principal Component Regression (PCR), and then
8
include them in the estimation phase of the discrete choice model. This methodology may result in biased
9
estimators for the parameters involved (Walker, 2001) or in estimators with a statistical significance greater
10
than their real contribution to the model (Raveau et al., 2010).
11
The other main approach used in the field is structural equation modeling (SEM). In structural equation
12
modeling psycho-attitudinal variables, specified as a linear combination of observed variables, are not
13
directly observed from individuals but are considered as functions of original statement variables. Examples
14
of research employing SEM include Ma et al. (2014), Kelarestaghi et al. (2019) and Zhang et al. (2019). The
15
main issue with these studies is that they treat the frequency of cycling as a continuous variable, despite
16
being measured as an ordered variable, which is improper from an econometric point of view as they assume
17
that the numerical distance between each set of categories is the same.
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
7
Table 1. Methodology used for modeling cycling frequency
1
Authors
Year
Include
psycho-attitudinal
factors
Methodology
Location
Stinson and Bhat
2004
No
Standard ordered logit
USA
Sener et al.
2009
Yes
Standard ordered probit
USA
Handy et al.
2010
Yes
Nested logit
USA
Xing et al.
2010
Yes
Linear regression
USA
Heinen et al.
2011
Yes
Binary logit
Netherlands
Noland et al.
2011
No
Standard ordered logit
USA
Sallis et al.
2013
Yes
Linear regression
USA
Ma et al.
2014
Yes
Structural equation
modeling
USA
Stinson et al.
2014
No
Linear regression
USA
Bhat et al.
2017
No
Spatial generalized
ordered probit
USA
Fu and Farber
2017
Yes
Standard ordered probit
USA
Manaugh et al.
2017
No
Multinomial logit
USA
Kelarestaghi et al.
2019
Yes
Structural equation
modeling
USA
Namgung and Jun
2019
Yes
Binary logit
USA
Oliva et al.
2019
No
Latent class ordered logit
Chile
Zhang et al.
2019
Yes
Structural equation
modeling
China
2
The modeling approach adopted in the current paper is distinct from earlier works for three important
3
aspects. First, the use of a Hybrid Choice Model allowed us to include psycho-attitudinal variables in a
4
discrete choice analysis, in which the choice model and the latent variable model are integrated into a single
5
structure that is estimated simultaneously. The simultaneous estimation yields consistent and efficient
6
estimates of the parameters (Walker, 2001). Further, with this approach we are able to correlate psycho-
7
attitudinal variables with observed socio-economic variables. This feature is particularly relevant because it
8
can help to identify which marketing/information strategy is more effective in convincing people to use the
9
bike along with which messages to deliver to different segments of the population. For example, if the study
10
finds that women are less likely to cycle because they have a greater perception of the limitations of the bike
11
as a means of transport than men, then a marketing campaign promoting bike use could be adapted to target
12
women. For example, a video campaign showing a number of women who commute or go shopping by bike
13
or using as testimonials celebrities who cycle. Second, the estimation of an ordered choice model overcomes
14
the limitations of structural equations modeling (SEM) in handling ordered variables. In fact, it has been
15
demonstrated that using an ordinal approach, instead of a continuous numerical approach, yields more
16
accurate results (Allen et al., 2018). Third, the use of a generalized ordered probit as kernel permits to
17
express the model thresholds as a function of explanatory variables. The standard ordered logit/probit
18
assumes that the threshold values are fixed across observations, which might be not appropriate (Eluru et al.,
19
8
2008; Balusu et al., 2018). Imposing this restriction might produce inconsistent latent propensity and
1
threshold values, and consequently inconsistent effects of the independent variables on the likelihood of
2
different categories of frequency (Eluru et al., 2008).
3
3. DATA ANALYSIS
4
The data used in this study come from a survey conducted by the Regional Government of Sardinia and the
5
Research Centre for Mobility Models (CRiMM) at the University of Cagliari (Italy) in the urban areas of
6
Cagliari and Sassari, main cities in Sardinia (Italy). The survey, called “BIKE I LIKE YOU”, was carried out
7
between 2014 and 2016 and targeted local authority employees. We intercepted potential respondents both
8
via mailing lists and through a promotional campaign. The mailing lists were provided by the universities of
9
Cagliari and Sassari, the Regional Government of Sardinia and the municipalities of Cagliari and Sassari
10
(around 9,600 invitation mails were sent). The promotional campaign was conducted via traditional
11
communication channels and social media, inviting people to complete an on-line questionnaire using the
12
WUFOO survey platform (for more details see Sottile et al., 2019). In particular, the questionnaire was
13
organized into 4 sections:
14
1. Bike use section aimed to identify for what purpose and how frequently people choose to cycle.
15
2. Cycling perceptions section (Likert scale from 1 to 5, 1=Totally disagree to 5 = Totally agree,
16
intended to:
17
• Measure positive and negative aspects of cycling in general.
18
• Measure the perception of safety of bike lanes and paths.
19
• Measure the perceived importance of context characteristics, intended as the importance
20
assigned to policies for increasing bike use.
21
3. Description of home-work commute trip.
22
4. Socio-demographic information section.
23
A total of 4,691 individuals completed the survey. However, after careful screening - excluding records
24
with incomplete socio-economic information and residence location - the final sample size included 2,128
25
individuals (corresponding to 45.4% of respondents).
26
Regarding individual characteristics (Table 2) , the analysis revealed that the sample is almost equally
27
divided between males and females (48.4% vs 51.6%). 73.3% of those surveyed are aged between 41 and 60.
28
The majority of individuals are highly-educated (57.7% have a bachelor’s degree or higher). Average
29
number of household members is 2.88. As for personal monthly income, 6.6 % stated they earned less than €
30
1,000 a month, 64.9% € 1,001-2,000, 9.6% € 2,001-3,000, 14.1% >€ 3,000. Note that, in terms of socio-
31
economic characteristics, the sample is not representative of the entire working population in Sardinia vis-a-
32
vis gender (for the share of male individuals we have 48.4% in our sample vs 58.0%), age (for the age
33
category 41-60 we have 73.3% in our sample vs 58.3% in Sardinia) and the level of education (only 18.9%
34
of workers in Sardinia have a bachelor’s degree). However, the sample can be considered as representative
35
9
of employees in the service sector, which in Sardinia accounts for 58.7% of the entire working population
1
(Sardegna Statistiche, 2018).
2
Participants were asked to report their cycling frequency for the three different purposes -commuting,
3
errands and leisure/sport - in five ordinal categories from “I never cycle” to “I cycle every day” (the same
4
categorization has been used in other works, e.g. Noland et al., 2011; Bhat et al., 2017). Analysis of the
5
questionnaire revealed the following share by frequency of people using the bike: (1) I never use the bike
6
(50.1% of the sample); (2) I use the bike 1-10 times per year (14.6% of the sample); (3) I use the bike 1-5
7
times per month (14.2% of the sample); (4) I use the bike more than once a week (14.7% of the sample); (5)
8
I use the bike every day (6.5% of the sample). Note that the trips considered here are not restricted to
9
utilitarian cycling alone but include recreational cycling as well.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
10
Table 2. Socio-demographic characteristics
1
Variables
N.
[%]
AVG.
Total sample
2,128
Gender
Male
1,029
48.4%
-
Female
1,099
51.6%
-
Age
-
-
48.02
Age 18-30
82
3.9%
-
Age 31-40
341
16.0%
-
Age 41-60
1,559
73.3%
-
Age > 60
146
6.9%
-
Level of education
Low (High school and lower)
901
42.3%
-
Medium (Graduate)
738
34.7%
-
High (Higher than Master’s degree)
489
23.0%
-
Marital status
Married
1,550
72.8%
-
Not married
578
27.2%
-
Presence of children in the household
Yes
1,159
54.5%
-
No
969
45.5%
-
# of household members
-
-
2.88
Driving license
Yes
2,098
98.6%
-
No
30
1.4%
-
Personal car available
Yes
1,930
90.7%
-
No
198
9.3%
-
# of cars per household
-
-
1.72
# of bikes per household
-
-
1.54
Personal income per month
Income 0-1,000 €
140
6.6%
-
Income 1,001-2,000 €
1,382
64.9%
-
Income 2,001-3,000 €
205
9.6%
-
Income >3,000 €
301
14.1%
-
Cycling frequency
Never
1,065
50.1%
-
1-10 times per year
310
14.6%
-
1-5 times per month
303
14.2%
-
More than once a week
311
14.6%
-
Every day
139
6.5%
-
2
3
4
5
11
3.1. Comparisons: socio-demographic characteristics for different levels of cycling
1
Table 3 gives the socio-demographic variables of respondents for different levels of cycling. Some
2
differences were detected among different categories of cycling frequency. Males and low-educated
3
individuals tend to use the bike more frequently. The most interesting difference concerns car ownership and
4
bike ownership per household. Indeed, bike users tend to own more bikes and less cars in their households
5
than individuals who do not choose to cycle.
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
12
Table 3. Socio-economic characteristics for different levels of cycling
1
Frequency of bike use
Never
1-10 times per
year
1-5 times per
month
More than once a
week
Every day
N.
%
N.
%
N.
%
N.
%
N.
%
Total
1,065
50.1%
310
14.6%
303
14.2%
311
14.6%
139
6.5%
Gender
Male
423
39.7%
128
41.3%
159
52.5%
212
68.2%
107
77.0%
Female
642
60.3%
182
58.7%
144
47.5%
99
31.8%
32
23.0%
Age (average)
49.07
-
46.13
-
46.49
-
47.46
-
48.82
-
18-30
24
2.3%
20
6.5%
14
4.6%
19
6.1%
5
3.6%
31-40
164
15.4%
59
19.0%
53
17.5%
44
14.1%
21
15.1%
41-60
786
73.8%
215
69.4%
223
73.6%
232
74.6%
103
74.1%
>60
91
8.5%
16
5.2%
13
4.3%
16
5.1%
10
7.2%
Body Mass Index
(average)
23.82
-
23.06
-
23.19
-
23.75
-
23.93
-
Level of education
Low (High school and
lower)
451
42.3%
107
34.5%
94
31.0%
172
55.3%
77
55.4%
Medium (Graduate)
381
35.8%
110
35.5%
125
41.3%
85
27.3%
37
26.6%
High (Higher than
Master’s degree)
233
21.9%
93
30.0%
84
27.7%
54
17.4%
25
18.0%
Presence of children in the
household
Yes
581
54.6%
170
54.8%
163
53.8%
171
55.0%
74
53.2%
No
484
45.4%
140
45.2%
140
46.2%
140
45.0%
65
46.8%
Marital status
Married
757
71.1%
230
74.2%
224
73.9%
236
75.9%
103
74.1%
Not married
308
28.9%
80
25.8%
79
26.1%
75
24.1%
36
25.9%
# of household members
(average)
2.83
-
2.99
-
2.88
-
2.96
-
2.81
-
Driving License
Yes
1,049
98.5%
308
99.4%
300
99.0%
305
98.1%
136
97.8%
No
16
1.5%
2
0.6%
3
1.0%
6
1.9%
3
2.2%
Personal car available
Yes
978
91.8%
276
89.0%
274
90.4%
275
88.4%
127
91.4%
No
87
8.2%
34
11.0%
29
9.6%
36
11.6%
12
8.6%
# of cars (average)
1.71
-
1.82
-
1.72
-
1.73
-
1.52
-
# of bikes (average)
0.97
-
1.98
-
2.07
-
2.19
-
2.32
-
Monthly personal income
Income 0-1,000 €
52
4.9%
37
11.9%
16
5.3%
22
7.1%
13
9.4%
Income 1,001-2,000 €
706
66.3%
183
59.0%
205
67.7%
206
66.2%
82
59.0%
Income 2,001-3,000 €
148
13.9%
46
14.8%
41
13.5%
50
16.1%
20
14.4%
Income >3,000 €
159
14.9%
44
14.2%
41
13.5%
33
10.6%
24
17.3%
2
13
One important issue concerns the association between built environment and cycling behavior (Cervero
1
and Duncan, 2003; Wang et al., 2016; Yang et al., 2019). In particular one key factor in the choice to use the
2
bike is that in many cases built-environment may represent a barrier to cycling. To analyze this correlation,
3
the micro-environments of the place of residence (presence of bike lanes and percentage of green areas) were
4
assessed within the GIS environment using a buffer of 400 m radius. Using the digital land use maps
5
downloaded from the Sardinian Government website
6
(<http://www.sardegnageoportale.it/areetematiche/databasegeotopografico>), it was possible to calculate the
7
characteristics of the residence location (urban or suburban/rural). Table 4 provides a summary of
8
respondents’ built environment characteristics.
9
However, as can be seen from Table 4, almost no differences were found among cycling frequency
10
categories, since the majority of individuals live in urban areas and fewer than half have access to a bike lane
11
within 400m from home.
12
Table 4. Built environment characteristics
13
Variables
Cycling frequency
Never
1-10 times
per year
1-5 times per
month
More than
once a week
Every day
Total
Total
1,065
310
303
311
139
2,128
Residential location
Urban
842 (79.1%)
224 (72.2%)
225 (74.3%)
247 (79.6%)
116 (83.0%)
1,654
Suburban and rural
223 (20.9%)
86 (27.8%)
78 (25.7%)
64 (20.4%)
23 (17.0%)
474
Presence of bike paths within 400m from home
Yes
519 (48.7%)
151 (48.7%)
137 (45.1%)
146 (47.1%)
82 (59.2%)
1,035
No
546 (51.3%)
159 (51.7%)
166 (54.9%)
165 (52.9%)
57 (40.8%)
1,093
Average % of green areas within 400m from
home per individual
5.2%
5.3%
5.1%
4.9%
5.1%
n/a
Average distance from home to the nearest bus
stop [m]
260 m
258 m
268 m
300 m
237 m
n/a
n/a not applicable
14
3.2. Analysis of psycho-attitudinal characteristics
15
Psycho-attitudinal characteristics were measured by means of the questions with the 5-point Likert scale (1 =
16
Totally disagree to 5 = Totally agree). As explained in paragraph 3, the survey included questions aimed at
17
measuring three latent variables:
18
1. perception of the bike as a means of transport (items A1-A12);
19
2. perception of safety of bike lanes and paths (items B1-B4);
20
3. perceived importance of context characteristics (items C1-C8).
21
To detect any differences in psycho-attitudinal variables among cycling frequency groups, we computed
22
the means of each item for all the subsamples and then tested whether subsample means were statistically
23
different conducting a z-test.
24
Table 5 shows the results of the statistical test analysis for individuals with different levels of cycling
25
experience with regard to the attitudinal factors. To avoid clutter, Table 5 only shows the results of the test
26
conducted between two consecutive groups (e.g. “never” group vs “1-10 times per year” group).
27
14
There are items that are not statistically significantly different among categories. For example, no
1
differences were detected for the item regarding the benefits of cycling in terms of cost, accessibility and
2
reduced level of pollution, the majority of the sample recognizing the positive aspects of cycling.
3
Interestingly, there were no significant differences between the groups as far as the implementation of certain
4
measures was concerned, such as a greater extension of limited traffic zones and presence of end-of-trip
5
facilities. The latter result suggests that all the respondents consider the existence of such facilities important,
6
regardless of how frequently they cycle.
7
However, in general, the z-test analyses revealed that more experienced cyclists have more positive
8
perceptions of cycling than the less experienced, as found in other works (Heinen et al., 2011; Namgung and
9
Jun, 2019). More specifically, frequent cyclists have a greater perception of cycling comfort than occasional
10
cyclists and non-cyclists (items A8, A10 and A12). We also found differences in the perception of safety
11
(items A2 and B1). In particular, the results suggest that more experienced cyclists tend to be less bothered
12
by mixed traffic situations. Nevertheless, it should be noted that, in general, cyclists and non-cyclists agree
13
about the inadequacy, in terms of safety, of the current cycling network. The z-test analyses also showed that
14
the willingness to cycle in the case of integration with the public transit service (items C4 and C7) is greater
15
among non-cyclists and occasional cyclists.
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
15
Table 5. Psycho-attitudinal characteristics for different levels of cycling (0 = never, 1=1-10 times per year, 2=1-5 times per month,
1
3= more than once a week, 4 = every day).
2
Avg
tot
Avg
1
Avg
2
Avg
3
Avg
4
Avg
5
Z-stat
1-2
Z-stat
2-3
Z-stat
3-4
Z-stat
4-5
PERCEPTION OF THE BIKE AS A MEANS OF TRANSPORT
A1. It is a rapid means of
transport
4.12
3.99
3.90
4.15
4.46
4.83
1.12
-2.79**
-3.72**
-5.31**
A2. Cycling in traffic is not
dangerous
1.75
1.72
1.56
1.75
1.86
2.22
2.47**
-2.38**
-1.32
-2.98**
A3. The bike is not likely to be
stolen and there are adequate
parking areas
2.22
2.32
2.25
2.12
2.03
2.11
0.80
1.35
0.86
-0.60
A4. It is not expensive
4.69
4.66
4.65
4.72
4.73
4.84
0.16
-1.01
-0.26
-1.78*
A5. It does not imply
exposure to bad weather and
air pollution
2.44
2.24
2.49
2.65
2.69
2.78
-3.31**
-1.79*
-0.41
-0.64
A6. It avoids wasting time
looking for parking
4.38
4.30
4.24
4.46
4.55
4.78
0.79
-2.61**
-1.24
-2.70**
A7. It is healthy
4.69
4.57
4.67
4.85
4.85
4.86
-1.93*
-3.53**
-0.04
-0.16
A8. It is easy to carry heavy
items
1.88
1.70
1.83
2.01
2.18
2.41
-1.91*
-1.99**
-1.81*
-1.91*
A9. It allows one to
appreciate historic centers
and increases accessibility to
city services
4.36
4.31
4.25
4.38
4.50
4.63
0.96
-1.54
-1.59
-1.50
A10. No need for cycling gear
2.96
2.83
2.83
2.95
3.18
3.70
0.10
-1.35
-2.36**
-3.86**
A11. It contributes to
reducing polluting emissions
4.84
4.83
4.79
4.86
4.86
4.93
0.78
-1.31
0.04
-1.78*
A12. It does not hamper daily
activity patterns
3.09
2.89
3.08
3.08
3.39
3.91
-2.35**
-0.05
-3.18**
-4.24**
PERCEPTION OF SAFETY
B1. Existing bike lanes are
not useful for traveling
3.41
3.57
3.42
3.22
3.14
3.16
1.73*
1.77*
0.73
-0.10
B2. Existing bike lanes and
crossings are safe,
comfortable and well-marked
2.11
2.02
2.16
2.21
2.22
2.16
-1.98**
-0.55
-0.07
0.47
B3. It is better to ride in
traffic than use the existing
bike paths
1.95
1.97
1.86
1.81
1.98
2.17
1.29
0.56
-1.76*
-1.34
B4. Car drivers do not
respect dedicated bike lanes
and often invade them
4.00
3.88
3.93
4.13
4.19
4.30
-0.63
-2.08**
-0.78
-0.95
PERCEIVED IMPORTANCE OF THE CONTEXT. I would cycle more with:
C1. An extensive network of
dedicated bike lanes in urban
area
4.39
4.23
4.44
4.61
4.61
4.61
-3.19**
-2.51**
0.06
0.01
C2. The presence of racks
and secure parking for bikes
4.26
4.15
4.23
4.42
4.43
4.46
-1.11
-2.49**
-0.16
-0.26
C3. A greater extension of the
LTZ or pedestrian zones
3.77
3.67
3.78
3.81
3.94
4.12
-1.28
-0.28
-1.36
-1.50
C4. A bike-sharing station
close to home or at public
transport stops
3.73
3.87
3.72
3.53
3.57
3.42
1.71*
1.71*
-0.30
1.01
C5. If other people use it
2.95
2.99
2.83
2.75
3.08
3.13
1.79*
0.69
-2.81**
-0.29
C6. Dedicated services at
work / study (parking,
showers, lockers for
equipment, etc.)
3.95
3.86
3.93
4.04
4.17
4.01
-0.91
-1.25
-1.32
1.19
C7. An integrated ticket for
bike-sharing and public
transport services
3.84
3.92
3.57
3.81
3.90
3.78
4.21**
-2.22**
-0.85
0.90
C8. A Combination with
public transport services
3.97
3.98
3.72
3.97
4.15
4.04
3.19**
-2.50**
-1.76*
0.88
C9. An Increase of car
parking fees
2.27
2.24
2.20
2.16
2.40
2.53
0.50
0.36
-2.17**
-0.87
* Significant at 90% confidence. ** Significant at 95% confidence.
3
16
3.3. Factor analysis
1
Prior to the modeling phase, an explorative factor analysis (Bollen, 1989) was performed to identify the
2
latent constructs underpinning the set of our attitudinal statements. The factor loadings are estimated using
3
Principal Axis Factoring (PAF) with varimax rotation. To establish if the dataset is suitable for exploratory
4
factor analysis, sample adequacy and strength of the intercorrelation of items must be examined. The Kaiser-
5
Meyer-Olkin (KMO) measure is used for sample adequacy: KMO values between 0.7 and 1 indicate the
6
sampling is adequate. The Bartlett test of sphericity is used to test the hypothesis that the correlation matrix
7
is an identity matrix, which would indicate that variables are unrelated and therefore unsuitable for structure
8
detection. Furthermore, to examine reliability Cronbach's alpha value is used. A Cronbach's alpha value
9
higher than 0.6 indicates that the dataset is reliable and acceptable.
10
Only two of the three psycho-attitudinal constructs turned out to be suitable for factor analysis,
11
perception of the bike as a means of transport (items A1-A12) and perceived importance of the context
12
(items C1-C9), whose KMO was 0.765 and 0.806 respectively. Instead, for the construct perception of safety
13
of bike lanes and paths (items B1-B4) we obtained a KMO of 0.572, which is below the reliability threshold
14
of 0.7. Thus, we proceeded with factor analysis only for the two constructs with KMO higher than 0.7.
15
Factor analysis generated two factors for the latent construct perception of the bike as a means of
16
transport (factor LV1 and factor LV2) and one factor for the latent construct perceived importance of the
17
context (factor LV3). Table 6 gives the results of factor analysis showing the loadings of the survey items on
18
each of the three identified factors.
19
The latent variable LV1, Perception of cycling benefits, expresses the agreement related to generally
20
recognized positive features of bikes, while the LV2, Perception of cycling comfort, expresses the agreement
21
related to generally recognized negative features of bikes (exposure to bad weather, carrying heavy items,
22
limitations in daily activity patterns, fatigue). The latent variable LV3, Perceived importance of bike
23
infrastructure, uses indicators capturing the appeal of some factors that would facilitate bike use. Most of the
24
Cronbach's alpha values are above 0.6, except for LV2 that is just acceptable since it is around the “criterion-
25
in-use” of 0.6.
26
Table 6. Factor scores of the psycho-attitudinal factors towards the bike mode (values below 0.4 are not reported)
27
Factor
Variables
Loading
Cronbach’s
alpha
LV1
A1. It is a rapid means of transport (avoids queues and traffic)
0.582
0.690
A4. It is not expensive
0.540
A6. It avoids wasting time looking for parking
0.594
A7. It is healthy
0.744
A9. It allows one to appreciate historic centers and increases accessibility to
city services
0.699
A11. It contributes to reducing polluting emissions
0.621
LV2
A5. It does not imply exposure to bad weather and air pollution
0.554
0.597
A8. It is easy to carry heavy items
0.629
A10. No need for cycling gear
0.629
A12. It does not hamper daily activity patterns
0.693
LV3
C1. An extensive network of dedicated bike lanes in urban area
0.907
0.778
C2. The presence of racks and secure parking for bikes
0.861
C3. A greater extension of the LTZ or pedestrian zones
0.721
17
4. METHODOLOGICAL FRAMEWORK
1
To perform our analysis we employ a hybrid choice modeling (HCM) approach (Vij and Walker, 2016) with
2
a generalized ordered probit choice (Greene and Hensher, 2010) kernel (Figure 1). The HCM provides a
3
framework for incorporating psycho-attitudinal variables into our model of the decision-making process.
4
5
6
Figure 1. Conceptual model
7
Following the framework of hybrid choice models, each latent variable is assumed to be determined by
8
a structural equation, in turn assumed to be a linear function of respondents’ individual and household
9
characteristics. Hence, for the person q we get:
10
11
where is the (vector of latent variables, is the (vector of constants, is the (
12
vector of individual background characteristics; is the (matrix of coefficients associated with
13
these characteristics, and is a normal distributed error term with a diagonal covariance matrix .
14
The R items reported in Table 6 are used as indicators of latent variables and are linked to them through
15
measurement equations. As in the typical HCM theory, indicators are modeled using the following set of
16
equations:
17
18
19
Eqs. (2), (3) represent a system of ordered logit models for measuring , where
is the (vector of
20
continuous measurement indicators of the latent variables, with elements
(we
21
assume that measurement elements); is the ( vector of constants, is the (
22
18
matrix of parameters denoting the estimated effect of on the indicators and is a logistic distributed
1
disturbance with a diagonal covariance matrix(and are normalized to 0 and 1 respectively for one of
2
the indicators of each latent variable for identification purposes, as suggested by Ben-Akiva et al., 2002).
3
is a categorical indicator with categories and is a vector of threshold parameters
4
. For identification purposes we impose the condition
5
. The probability for a certain response s to the indicator r is thus given by:
6
7
8
9
The latent propensity underlying the ordered response observation, that is the cycling frequency
10
reported for each individual q, has been specified as a function of observed and latent variables:
11
12
where xq is the (vector of explanatory variables, LVq is the (vector of individual specific
13
latent variables, β and β* are the (and (vectors of unknown parameters to be estimated and εq
14
is the error term capturing the effects of unobserved factors on the latent propensity. We assume that is
15
normally distributed across observations with mean = 0 and variance =1. Because the distribution of is
16
univariate, the covariance matrix contains only one term, equal to 1. In the usual ordered-response
17
fashion, the latent propensity
is linked to the observed level through a set of threshold parameters
18
as follows:
19
20
In our specific case K = 5. To allow for heterogeneity (across observations) in the thresholds, they are
21
parametrized as a function of both socio-demographic and latent variables as in Eluru et al. (2008):
22
23
where is a scalar, and are (and (vectors of coefficients associated with level k=1, 2,
24
…, K-1. For identifications purposes, we impose the normalization for all q. The formulation of
25
the thresholds in Eq. (6), where a higher threshold is specified as the sum of its preceding threshold
26
and a non-negative term , guarantees the increasing order of the thresholds. Note
27
that if none of the dependent variables influence the value of the thresholds, the generalized ordered probit
28
collapses to the standard ordered probit formulation.
29
19
Because we assumed that is normal distributed, the unconditional probability that decision-maker
1
q belongs to category k is given by:
2
Prob (
3
where is the cumulative distribution function of the error term .
4
Assuming the error components are independent, the joint likelihood function for
5
individual q may be written as follows:
6
7
where denotes the probability function of observing that the individual belongs to the frequency category
8
k, the density function for the indicators of the latent variables corresponds to the measurement equation of
9
the latent variable model, and the density function of the latent variables corresponds to the structural
10
equation of the latent variable model.
11
Simulation techniques are applied to approximate the multidimensional integral in the likelihood
12
function, and the resulting simulated log-likelihood function is maximized. The models were estimated using
13
Python Biogeme software (Bierlaire, 2016).
14
5. MODEL RESULTS
15
Table 7 presents the results of the structural equation component of the model. It emerged that the
16
Perception of cycling benefits construct is positively affected by the number of bikes in the household.
17
Further, individuals with a lower level of education have a better perception of the positive aspects of
18
cycling. Note that we found a high value of the constant, which suggests that this latent variable is truly
19
latent and only a few explanatory variables are able to explain it.
20
The second construct is the Perception of cycling comfort. As found in other studies (Akar et al., 2013;
21
Habib et al., 2014), males are more comfortable with cycling than females. The positive sign associated with
22
the variable Age indicates that younger people place less importance on the limitations of the bike mode. In
23
line with other studies (Handy et al., 2010; Noland et al., 2011), we also found that the presence of children
24
at home makes people less inclined to cycle. The number of cars per household negatively influences the
25
latent construct. This effect could reflect the fact that car-addicted individuals, since they are used to the
26
car’s comfort, tend to stigmatize the disadvantages of cycling. On the other hand, the number of bikes per
27
household has a positive effect on the latent variable. Furthermore, people with a high level of education are
28
more likely to recognize the negative aspects of cycling.
29
The structural model related to the Perceived importance of bike infrastructure shows that women
30
consider proper infrastructure more important than men. One possible interpretation of this effect could be
31
the fact that women are more concerned about traffic and safety conditions, a finding consistent with the
32
literature (Akar et al., 2013; Bhat et al., 2015; Manton et al., 2016). Furthermore, people with a lower
33
education (high school or lower) are more likely to consider as important the presence of facilitators
34
20
encouraging cycling than those with a higher level of education. Not surprisingly, the presence of children in
1
the household negatively affects the latent variable, suggesting the existence of other kinds of barriers for
2
this segment of individuals in their choice to cycle.
3
Table 7. Determinants of latent constructs.
4
Explanatory Variables
LV1 – Perception of
cycling benefits
LV2 - Perception of
cycling comfort
LV3 - Perceived
importance of bike
infrastructure
Coef.
R T-stat
Coef.
R T-stat
Coef.
R T-stat
Age ∙ 10-1
--
--
-0.037
-1.85
--
--
Gender (male=1, female=0)
--
--
0.045
1.29
-0.158
-2.38
Bachelor's degree or higher (yes=1, no=0)
-0.332
-3.02
-0.093
-2.56
-0.173
-2.50
# of bikes per household
0.271
5.20
0.128
6.71
0.132
4.24
# of cars per household
--
--
-0.081
-3.20
--
--
Presence of children (yes=1, no=0)
--
--
-0.134
-3.43
-0.078
-1.09
Constant
6.010
33.33
1.670
14.97
2.660
28.77
Variance
1.790
10.71
0.540
13.29
1.220
20.24
“--”in a cell indicates that the variable in the corresponding row does not have a significant impact on the utility of the alternative in
the corresponding column.
5
Table 8 presents the results of the measurement model. Several indicators were considered in the latent
6
variable measurement model, which linked the latent variables to the responses to the qualitative attitudinal
7
survey questions. The ζ parameters that indicate the associations between the responses to the items and the
8
associated latent variable all have the expected signs. For example, a more positive perception of cycling
9
benefits will lead respondents to agree more with the statements about the bike as a healthy and
10
environmentally friendly mode of transport.
11
12
13
14
15
16
17
18
19
20
21
22
23
24
21
1
Table 8. Impact of latent variables on non-nominal dependent variables.
2
Latent Variable
Indicators
Const. δ
R T-stat
Coef. ζ
R T-stat
Perception of cycling
benefits
A1. It is a rapid means of transport
-0.629
-1.43
0.801
8.75
A4. It is not expensive
0.152
0.37
0.750
8.66
A6. It avoids wasting time looking for
parking
-0.702
-1.66
0.788
8.80
A7. It is healthy
-1.780
-2.75
1.380
8.83
A9. It allows one to appreciate historic
centers and increases accessibility to city
services
-0.651
-1.71
0.853
10.61
A11. It contributes to reducing polluting
emissions
0.000
n/a
1.000
n/a
Perception of cycling
comfort
A5. It does not involve exposure to bad
weather and air pollution
-2.080
-5.85
2.260
8.77
A8. It is easy to carry heavy items
-3.970
-8.58
2.680
8.50
A10. No need for cycling gear
0.000
n/a
1.000
n/a
A12. It does not limit daily activity
patterns
-0.576
-2.55
2.130
12.59
Perceived importance
of bike infrastructure
C1. Presence of an extensive network of
dedicated bike lanes
-0.765
-1.90
3.430
8.99
C2. Presence of racks and secure parking
for bikes
-0.864
-2.11
3.550
9.83
C3. Greater extension of the LTZ or
pedestrian zones
0.000
n/a
1.000
n/a
“n/a” not applicable
3
The estimation results of the discrete part of the model are reported in Table 9. For reasons of
4
identification, there is no constant in the latent propensity to cycle and the first threshold is constant, with no
5
socio-demographic attributes or latent variables in its specification.
6
Table 9 contains three columns. The first column corresponds to the estimate of the parameters
7
characterizing the latent propensity to cycle. The second column corresponds to the estimates of constant and
8
the parameters for the threshold (threshold between the frequency levels “1-10 times per year” and “1-5
9
times per month”). The third and fourth columns correspond to the constant and the parameters linked to the
10
third threshold, delimiting the “1-5 times per month” and “more than once a week” categories, and to the
11
fourth threshold, demarcating the “more than once a week” and “every day” cycling categories.
12
Some socio-economic variables were found to have a significant effect on the propensity to cycle. In
13
agreement with several previous studies (Stinson and Bhat, 2004; Noland et al., 2011; Fu and Farber, 2017),
14
males are more likely to cycle. In addition to the effect on the propensity to cycle, the gender variable also
15
impacts the value of the thresholds. If the coefficient associated with a variable in the vector μk is negative,
16
the corresponding threshold shifts to the left. In this case, the pattern of threshold effects indicates that men
17
are more likely than women to belong to the category “1-5 times per month”, which differs from the result
18
obtainable with a standard ordered probit model. Regarding age, older individuals are less inclined to cycle,
19
a finding consistent with other studies (Ma et al., 2014; Bhat et al., 2017). Interestingly, the Body Mass
20
Index (a biometric datum used as an indicator of ideal weight) negatively affects the propensity to bike,
21
indicating that people with BMI higher are less inclined to cycle. The causal relationship underlying this
22
correlation could go in either or both directions: healthier people are more likely to use the bike or people
23
who cycle are likely to be healthier due to the benefits of physical activity. The level of education also
24
22
impacts the propensity to cycle. In contrast with other works (Handy et al., 2010; Bhat et al., 2017)
1
respondents with a bachelor’s degree are less likely to cycle. One possible explanation could be that
2
university graduates usually have a job where a dress code is often required, incompatible with bike usage in
3
a workplace not equipped with locker rooms, as in our context.
4
Different household demographics play a crucial role in the propensity to cycle frequently: number
5
of cars, number of bikes, number of household members. As expected, the number of bikes per family
6
positively affects the latent propensity. By contrast, the number of household members negatively impacts
7
the propensity to cycle. This might be because people with children often have to chain trips and do drop
8
offs/pick ups, which might be difficult if traveling by bike. Also, the number of cars per household has a
9
significant negative impact on bike use.
10
Among the built environment variables, living in an urban neighborhood and the presence of bike
11
paths within 400m from home positively affect cycling frequency. The effect on the threshold here indicates
12
that, compared to the standard probit, the generalized probit predicts a lower probability of individuals living
13
in urban areas being occasional cyclists. Instead, no association was found with the percentage of green areas
14
within 400m from home. It is possible that the variable urban neighborhood captures this factor, as in our
15
context the majority of parks and recreational areas are located in urban areas. We also tested the effect of
16
the variable distance from home to the nearest bus stop, given the importance of the integration between
17
public transport and cycling (Martens, 2007, La Paix et al., 2020), but the variable was not found to be
18
statistically relevant. This effect can be explained by the fact that in our context it is not possible to take
19
bikes on buses or park bikes safely close to bus stops or train stations
20
All the attitudinal factors had a positive impact on the propensity to cycle. The positive influence of
21
the latent variable LV3 Perceived importance of bike infrastructure emphasizes the importance of providing
22
bike facilities such as bike paths or bike lanes, rental bikes, and bike parking. People who are aware of
23
cycling benefits, such as protecting the environment as well as staying healthy, (latent variable LV1
24
Perception of cycling benefits) were more likely to be cyclists. This finding suggests that if the benefits of
25
cycling are better understood, then more people would be likely to cycle more frequently. The latent variable
26
LV2 Perception of cycling comfort is also positively associated with the propensity to cycle. This may be
27
because the more people feel that cycling is easier and more accessible, the more they cycle. Interestingly,
28
the effects on the thresholds show that the generalized probit predicts a lower probability, with higher values
29
of LV2 and LV3, of being occasional cyclists, validating the idea of using this modeling structure for our
30
analysis.
31
23
Table 9. Estimation results of the hybrid generalized ordered probit.
1
Variables
Latent propensity to cycle
Threshold between “1-10 times
per year” and “1-5 times per
month”
Threshold between “1-5 times
per month” and “More than once
a week”
Threshold between “More than
once a week” and “every day”
Estimate
R t-stat
Estimate
R t-stat
Estimate
R t-stat
Estimate
R t-stat
Threshold between not being a cyclist
and cycling “1-10 times per year”
1.140
3.31
n/a
n/a
n/a
n/a
n/a
n/a
Threshold constants
n/a
n/a
0.413
1.53
-0.246
-2.14
-0.005
-0.10
Age 10-1
-0.099
-3.07
--
--
--
--
--
--
Gender (male=1, female=0)
0.543
8.17
-0.359
-3.42
-0.125
-1.22
--
--
Bachelor’s degree or higher (yes =1,
no=0)
-0.130
-2.19
--
--
--
--
--
--
Body Mass Index 10-1
-0.030
-3.23
--
--
--
--
--
--
# of bikes per household
0.811
25.26
--
--
--
--
--
--
# of cars per household
-0.063
-1.33
--
--
--
--
--
--
# of household members
-0.281
-9.82
--
--
--
--
--
--
Residential location (urban = 1,
suburban and rural = 0)
0.082
1.09
-0.212
-1.83
-0.210
-1.82
--
--
Presence of bike paths within 400m
from home (yes = 1, no = 0)
0.125
2.28
--
--
--
--
--
--
LV1 – Perception of cycling benefits
0.114
4.80
--
--
--
--
--
--
LV2 - Perception of cycling comfort
0.473
5.30
-0.269
-2.03
--
--
--
--
LV3 - Perceived importance of bike
infrastructures
0.086
2.70
-0.099
-1.85
--
--
--
--
“--”in a cell indicates that the variable in the corresponding row does not have a significant impact on the utility of the alternative in the corresponding column. “n/a” not applicable
24
5.1. Pseudo-elasticity effects
1
The coefficients in Table 9 do not provide a sense of the magnitude and direction of the effects of each
2
variable on each cycling frequency category. But we can compute aggregate-level “pseudo-elasticity effects”
3
of socio-demographic variables (Bhat et al., 2017; Hirk et al., 2017), that can be calculated as:
4
5
where all elements of
are equal to except for the socio-demographic variable v, which is equal to
6
because of the discrete change in the variable . Note that Q is the dimension of our sample.
7
For the dummy variable, we first set the value of the dummy variables to zero (note that all other exogenous
8
variables keep their original values) and we then compute the probability for each frequency
9
category of each independent variable. Next, we change the value of the dummy variable from zero to one
10
for each individual and we calculate the new probability
. To obtain the disaggregate pseudo-
11
elasticity effect we compute the difference
. For the count variable (e.g. the
12
number of cars per household) we change the value of the count variable by the value of one and then we
13
calculate the percentage of change for each category. For the continuous variable, we decrease the value of
14
the variable by 20% for each individual. It is important to note here that, since the latent variables
15
influencing each level of cycling frequency are a function of socio-demographic variables, as socio-
16
demographic variables change so too does the value of the latent variables.
17
Table 10 shows the pseudo-elasticity effects. The numbers in the table can be interpreted as the
18
percentage change in the probability of each cycling frequency category after a change in the socio-
19
demographic variable. For example, the first entry in the table indicates that, ceteris paribus, the probability
20
of a male not being a cyclist is 14.7% lower than for a female. For the variable bike lanes, the probability of
21
an individual not being a cyclist decreases by 3.27% if bike lanes are provided in the individual's residential
22
neighborhood. The directions of the elasticity effects of the model are consistent with the discussion of the
23
results presented in the previous section.
24
25
26
27
28
29
30
31
32
33
34
35
25
Table 10. Pseudo-elasticity effects
1
Variables
I never cycle
1-10 times per
year
1-5 times per
month
More than once
a week
Every day
Gender (male)
-14.70%
-2.96%
+2.26%
+7.96%
+7.44%
Age ∙ 10-1
+2.91%
-0.28%
-0.56%
-1.04%
-1.03%
Body Mass Index∙ 10-1
+3.68%
-0.49%
-0.66%
-1.26%
-1.26%
Education (bachelor's degree or higher
level of education)
+5.95%
-0.15%
-1.14%
-2.27%
-2.38%
# of cars per household
+2.66%
-0.08%
-0.55%
-1.02%
-1.01%
# of bikes in the household
-22.73%
-0.07%
+2.29%
+7.61%
+12.91%
# of household members
7.34%
-1.07%
-1.39%
-2.51%
-2.36%
Presence of children in the household
+1.84%
+0.33%
-0.46%
-0.85%
-0.86%
Residence location choice (urban)
-0.81%
-1.72%
-1.13%
+1.95%
+1.71%
Presence of bike lanes within 400m of
home
-3.27%
+0.39%
+0.56%
+1.12%
+1.19%
2
5.2. Impact of the latent variables on choice probability
3
It is not possible to analyze the direct impact of the latent variables on the latent propensity by computing the
4
pseudo-elasticity effects. In fact, it would be meaningless to artificially increase by some percentage points
5
the value of the latent variables, since they were estimated as a function of socio-economic variables.
6
Instead, as suggested by Hess et al. (2018), we tested what would happen if all the individuals were assigned
7
a value of the latent variables equal to that of a given segment of population. No changes were made to the
8
discrete part of the model, i.e. the direct impact of socio-demographics on latent propensity and thresholds.
9
The results are reported in Table 11. We ran different test scenarios for each of the three latent
10
variables. The scenarios concern gender (two possible values), education (two possible values), household
11
composition (two possible values), car ownership (one possible value) and bike ownership (one possible
12
value). Note that for car and bike ownership it is possible to develop more than one scenario, e.g. one for
13
each different ownership level. But, to keep the presentation simple, we only focused on those scenarios that
14
assign to everyone the perception of an individual not owning a car or bike.
15
We found that changes in the latent variable LV3, Perceived importance of bike infrastructure, only
16
have a minor impact on choice probability, because of the low value of the parameters on the latent
17
propensity to cycle. The greatest changes in choice probability can be observed in the latent variable LV2,
18
Perception of cycling comfort, either adopting the psycho-attitudinal variable of the individuals owning no
19
cars or bikes. Finally, for the latent variable LV1, Perception of cycling benefits, a shift in bike ownership
20
level produces the greatest impact in choice probability.
21
22
23
26
Table 11. Test Scenarios
1
Latent variable LV1: Perception of cycling benefits
I never cycle
1-10 times per
year
1-5 times per
month
More than
once a week
Every day
Graduate degree
+0.41%
-0.04%
-0.06%
-0.14%
-0.17%
No graduate degree
-0.58%
+0.08%
+0.11%
+0.20%
+0.19%
No bikes
+1.30%
+0.00%
-0.13%
-0.49%
-0.67%
Latent variable LV2: Perception of cycling comfort
I never cycle
1-10 times per
year
1-5 times per
month
More than
once a week
Every day
Graduate degree
+0.48%
+0.08%
-0.10%
-0.21%
-0.24%
No graduate degree
-0.67%
-0.10%
+0.17%
+0.31%
+0.28%
No bikes
+2.57%
+0.80%
-0.47%
-1.31%
-1.59%
Male
-0.29%
-0.02%
+0.10%
+0.13%
+0.08%
Female
+0.27%
+0.06%
-0.04%
-0.12%
-0.17%
No car
-1.71%
-0.29%
+0.40%
+0.79%
+0.81%
Children
+0.78%
+0.08%
-0.21%
-0.34%
-0.31%
No children
-0.41%
-0.09%
+0.09%
+0.20%
+0.20%
Latent variable LV3: Perceived importance of bike infrastructure
I never cycle
1-10 times per
year
1-5 times per
month
More than
once a week
Every day
Graduate degree
+0.16%
+0.07%
-0.05%
-0.09%
-0.10%
No graduate degree
-0.23%
-0.10%
+0.08%
+0.13%
+0.11%
No bikes
+0.47%
+0.31%
-0.10%
-0.30%
-0.38%
Male
+0.19%
+0.07%
-0.08%
-0.11%
-0.06%
Female
-0.17%
-0.09%
+0.03%
+0.10%
+0.13%
Children
+0.08%
+0.03%
-0.03%
-0.05%
-0.04%
No children
+0.38%
+0.03%
-0.08%
-0.16%
-0.19%
2
It is also interesting to explore how the value of the latent variables varies among the different frequency
3
categories. In doing this, we segment the sample based on the value of the N latent variables. In particular,
4
for each latent variable n we split the sample into tertiles comprised of the top, middle and bottom third:
5
• the bottom third includes individuals with the lowest values of the n-th latent variable;
6
• the middle third consists of individuals with medium values of the n-th latent variable;
7
• the top third includes individuals with the highest values of the n-th latent variable.
8
Figures 2, 3 and 4 depict the predicted share of individuals in each cycling frequency category by the
9
value of latent variables. Note that the sum of the values reported at the top of each bar with the same color is
10
100%. The graphs in the figures are consistent with the results reported in the previous sections, indicating
11
that as the value of the perception of cycling benefits and cycling comfort increases, so too does the
12
probability of using the bike more frequently. For example, the probability of not being a cyclist for
13
individuals with a low perception of cycling comfort is 78.6%. Conversely, for individuals with a high level
14
of perception of cycling comfort the model indicated that the probability of being a frequent cyclist is 42.3%
15
27
(24.5% + 17.8%). Further, we found that for the segment with a medium perception of cycling comfort the
1
probability of choosing never to cycle is 49.9%, while the distribution of the choice probability for the
2
remaining cycling categories is almost flat.
3
We observed a similar trend when looking at the choice probability for each cycling frequency for the
4
latent variable 3 class. However, note that, for individuals who perceive bike infrastructure as less important,
5
the probability of choosing not to cycle is significantly lower than for individuals with a lower value of the
6
latent variables 1 and 2. One possible interpretation of this outcome may be that some experienced cyclists
7
are more aware of the presence of barriers to cycling.
8
9
Figure 2. Choice probability for each cycling frequency for LV1 class
10
11
Figure 3. Choice probability for each cycling frequency for LV2 class
12
13
28
1
Figure 4. Choice probability for each cycling frequency by LV3 class
2
3
5.3. Data fit
4
Evaluating the data fit of hybrid choice models is not an easy task. Following the approach suggested by
5
Walker (2001), we determine the log-likelihood value of the null model, which considers each choice option
6
equally likely, and the log-likelihood value of the discrete model obtained by removing the measurement
7
model of the latent variables. We examine the data fit of the estimated hybrid generalized ordered probit with
8
its restrictive version, a hybrid standard ordered probit. Then, as proposed by Bhat et al. (2017), we calculate
9
the percentage of share of each category for each model and compare the predicted shares with the observed
10
shares.
11
The results of the data fit comparisons are shown in Table 12. As can be observed, the hybrid
12
generalized ordered probit exhibits a slightly higher value of the likelihood ratio index than the standard
13
model, suggesting that the generalized model offers a superior goodness-of-fit. Instead, in terms of
14
probabilities, the generalized model tends to overestimate choice probabilities compared to the standard one.
15
The two models can also be compared with a likelihood ratio test. The χ2 test statistic of the
16
likelihood ratio test between the generalized model and the standard model is statistically significant at any
17
degree of confidence.
18
19
20
21
22
23
24
25
29
Table 12. Model Goodness of fit
1
Hybrid Generalized
ordered probit
Hybrid Standard
Ordered probit
-
Log-likelihood value at
convergence
-31,451.15
-31,465.17
-
Number of parameters
90
84
-
Log-likelihood value of discrete
model
-2,401.02
-2,415.92
-
Log-likelihood of null model
-3,424.88
-3,424.88
-
Log-likelihood of constant
model
-2,902.34
-2,902.34
-
Likelihood ratio index
0.299
0.294
-
Likelihood ratio test
χ2 = −2[-31,465.17 – (– 31,451.15)] =28.04, 6 df, p = 0.0001
Cycling frequency
Predicted percentage
Predicted percentage
Observed percentage
I never do
51.0%
50.8%
50.0%
1-10 times per year
14.1%
14.0%
14.6%
1-5 times in the past 30 days
13.1%
13.3%
14.2%
1-5 days per week
14.0%
14.1%
14.6%
Every day
7.9%
7.7%
6.5%
2
6. CONCLUSIONS
3
The purpose of the current research was to examine whether psycho-attitudinal factors toward bike use vary
4
in a sample of local authority employees with different cycling frequency and to quantify the determinants
5
influencing the propensity to cycle. Some papers have studied the impact of subjective variables on cycling
6
frequency in different contexts. However, the majority of these treat the problem employing an inappropriate
7
methodology from an econometric perspective, either adopting a two stage sequential approach or assuming
8
the frequency variable as a continuous variable, despite it being measured in ordinal discrete categories. By
9
contrast, we estimated a hybrid choice model with a generalized ordered probit choice kernel, where we
10
defined the thresholds as a function of both objectives and psycho-attitudinal variables. This approach
11
allowed us to overcome the limitations of previous studies and gain further insight into the decision-making
12
process underlying bike usage and frequency.
13
According to the estimated parameters in latent propensity and threshold functions, several factors
14
affect the frequency of using the bike, especially gender, education level, number of cars per household and
15
number of household members. Residential location and presence of bike facilities variables were also found
16
to influence the latent propensity, suggesting that a good urban design and the provision of bike
17
infrastructure are essential for increasing cycling levels. However, computation of the pseudo-elasticity
18
effects showed that the impact of some socio-demographic factors is much greater than the effect of the built
19
environment characteristics. Therefore, strategies focusing simply on the physical part of the problem, such
20
as expanding and improving bike infrastructure, might not be sufficient to encourage cycling.
21
30
The results also indicate that, besides objective characteristics, the latent propensity to cycle is
1
positively influenced by psycho-attitudinal variables, supporting the idea of a relationship between attitudes
2
and perceptions and cycling experience. These conclusions are consistent with the findings of previous
3
studies indicating that a greater inclination to cycle leads individuals to use the bike more frequently.
4
Further, our model specification, with the threshold function of psycho-attitudinal variables, suggests the
5
presence of individual heterogeneity among people with different cycling frequency. In particular, it has
6
been found that a significant variation of the probability of being an occasional cyclist depends on the value
7
of the latent variables perception of cycling comfort and perceived importance of bike infrastructure.
8
In view of these results, considering people’s psychological characteristics is essential when
9
implementing policies and strategies aimed at increasing bike use. In fact, they reinforce the idea that
10
promoting cycling, through the implementation of awareness campaigns and educational programs intended
11
to improve individuals’ perceptions of the bike mode, can persuade them to consider the bike as an
12
alternative means of transport to private motorized vehicles. Further, our findings about the link between
13
socio-economic characteristics and latent constructs revealed that some promotional strategies would have a
14
stronger impact, depending on the target of the intervention. For example, a marketing campaign involving
15
the distribution of material informing people of the location of safe bike paths and parking spots or that
16
cycling involves greater health benefits than risks (Rojas-Rueda et al., 2011) would have a greater effect
17
among women, who are more concerned with safety issues.
18
The study, however, is not without limitations. An issue that was not addressed was that we analyzed
19
the frequency of cycling for all purposes, mixing utilitarian and recreational trips. In fact, the determinants
20
(both objective and subjective) affecting the choice to travel by bike may be different, depending on trip
21
purpose (Félix et al., 2017; Sottile et al., 2020). Further, a relationship may exist between cycling for leisure
22
and the choice to cycle for utilitarian purposes. Further work needs to be conducted, through the estimation
23
of multivariate models, to establish the interplay between the use of the bike for different purposes.
24
In addition, the sample is composed predominantly of public sector employees, who are mainly aged
25
between 41-60, are highly educated and live in households with a larger number of components.
26
Understanding the decision to cycle of different categories of people is crucial for policy makers who intend
27
to implement effective strategies for promoting bike use. In fact, it is possible that individuals with different
28
socio-economic characteristics have different attitudes and propensities compared to our sample, so that our
29
results cannot be generalized to the entire population.
30
Lastly, it should be pointed out that a major concern with hybrid choice models is the derivation of
31
transport policies because of the endogenous nature of psycho-attitudinal variables (Chorus and Kroesen,
32
2014). Nevertheless, it is worth highlighting that, in some contexts, the explicability of a phenomenon is far
33
more important than predictability. In our specific case, a black-box model may have been superior in terms
34
of predicting probabilities in the short-term under different scenarios but it would have deprived us of the
35
possibility of understanding which psycho-attitudinal factors come into play when an individual decides to
36
cycle and their significance in the process of choosing to cycle.
37
31
1
AKNOWLEDGEMENTS
2
This research was funded by the Sardinian Regional Government and Azienda Regionale Sarda Trasporti
3
SpA. The second author gratefully acknowledges financial support provided by the Italian Ministry of
4
University and Research (MIUR), within the Smart Cities framework (Project Netergit, ID:
5
PON04a200490). The authors are grateful to two anonymous reviewers who provided valuable comments to
6
an earlier version of the manuscript.
7
8
Credit author statement
9
Francesco Piras: Conceptualization, Methodology, Formal analysis, Writing - Original Draft, Writing -
10
Review & Editing. Eleonora Sottile: Investigation, Data curation, Writing - Original Draft, Writing -
11
Review & Editing. Italo Meloni: Investigation, Funding acquisition, Project administration, Writing -
12
Review & Editing.
13
32
REFERENCES
1
Akar, G., Fischer, N., & Namgung, M. (2013). Bicycling choice and gender case study: The Ohio State
2
University. International Journal of Sustainable Transportation, 7(5), 347-365.
3
Allen, J., Eboli, L., Mazzulla, G., & Ortúzar, J.d.D. (2018). Effect of critical incidents on public transport
4
satisfaction and loyalty: an Ordinal Probit SEM-MIMIC approach. Transportation, 1-37.
5
Arroyo, R., Ruiz, T., Mars, L., Rasouli, S., & Timmermans, H. (2020). Influence of values, attitudes towards
6
transport modes and companions on travel behavior. Transportation research part F: traffic psychology
7
and behaviour, 71, 8-22.
8
Balusu, S. K., Pinjari, A. R., Mannering, F. L., & Eluru, N. (2018). Non-decreasing threshold variances in
9
mixed generalized ordered response models: A negative correlations approach to variance reduction.
10
Analytic methods in accident research, 20, 46-67.
11
Ben-Akiva, M., Walker, J., Bernardino, A. T., Gopinath, D. A., Morikawa, T., & Polydoropoulou, A. (2002).
12
Integration of choice and latent variable models. Perpetual motion: Travel behaviour research
13
opportunities and application challenges, 431-470.
14
Bhat, C. R., Dubey, S. K., & Nagel, K. (2015). Introducing non-normality of latent psychological constructs
15
in choice modeling with an application to bicyclist route choice. Transportation Research Part B:
16
Methodological, 78, 341-363.
17
Bhat, C. R., Astroza, S., & Hamdi, A. S. (2017). A spatial generalized ordered-response model with skew
18
normal kernel error terms with an application to bicycling frequency. Transportation Research Part B:
19
Methodological, 95, 126-148.
20
Bierlaire, M. (2016). PythonBiogeme: a short introduction. Report TRANSP-OR 160706, Series on
21
Biogeme. Transport and Mobility Laboratory, School of Architecture, Civil and Environmental
22
Engineering, Ecole Polytechnique Fédérale de Lausanne, Switzerland.
23
Bollen, K. A. (1989). A new incremental fit index for general structural equation models. Sociological
24
Methods & Research, 17(3), 303-316.
25
Cervero, R., & Duncan, M. (2003). Walking, bicycling, and urban landscapes: evidence from the San
26
Francisco Bay Area. American journal of public health, 93(9), 1478-1483.
27
Chorus, C. G., & Kroesen, M. (2014). On the (im-) possibility of deriving transport policy implications from
28
hybrid choice models. Transport Policy, 36, 217-222.
29
Dill, J., & McNeil, N. (2013). Four types of cyclists? Examination of typology for better understanding of
30
bicycling behavior and potential. Transportation Research Record, 2387(1), 129-138.
31
Eluru, N., Bhat, C. R., & Hensher, D. A. (2008). A mixed generalized ordered response model for examining
32
pedestrian and bicyclist injury severity level in traffic crashes. Accident Analysis & Prevention, 40(3),
33
1033-1054.
34
European Environmental Agency (2019). Air quality in Europe — 2019 report
35
www.eea.europa.eu/publications/air-quality-in-europe-2019.
36
Ewing, R., & Cervero, R. (2010). Travel and the built environment: A meta-analysis. Journal of the
37
American planning association, 76(3), 265-294.
38
Fu, L., & Farber, S. (2017). Bicycling frequency: A study of preferences and travel behavior in Salt Lake
39
City, Utah. Transportation research part A: policy and practice, 101, 30-50.
40
Félix, R., Moura, F., & Clifton, K. J. (2017). Typologies of urban cyclists: review of market segmentation
41
methods for planning practice. Transportation research record, 2662(1), 125-133.
42
Gatersleben, B., & Appleton, K. M. (2007). Contemplating cycling to work: Attitudes and perceptions in
43
different stages of change. Transportation Research Part A: Policy and Practice, 41(4), 302-312.
44
Greene, W. H., & Hensher, D. A. (2010). Modeling ordered choices: A primer. Cambridge University Press.
45
33
Gutiérrez, M., Hurtubia, R., & Ortúzar, J.d.D. (2020). The role of habit and the built environment in the
1
willingness to commute by bicycle. Travel behaviour and society, 20, 62-73.
2
Habib, K. N., Mann, J., Mahmoud, M., & Weiss, A. (2014). Synopsis of bicycle demand in the City of
3
Toronto: Investigating the effects of perception, consciousness and comfortability on the purpose of biking
4
and bike ownership. Transportation research part A: policy and practice, 70, 67-80.
5
Handy, S. L., Xing, Y., & Buehler, T. J. (2010). Factors associated with bicycle ownership and use: a study
6
of six small US cities. Transportation, 37(6), 967-985.
7
Handy, S., Van Wee, B., & Kroesen, M. (2014). Promoting cycling for transport: research needs and
8
challenges. Transport reviews, 34(1), 4-24.
9
Heinen, E., Maat, K., & Van Wee, B. (2011). The role of attitudes toward characteristics of bicycle
10
commuting on the choice to cycle to work over various distances. Transportation research part D:
11
transport and environment, 16(2), 102-109.
12
Heinen, E., Van Wee, B., & Maat, K. (2010). Commuting by bicycle: an overview of the literature.
13
Transport reviews, 30(1), 59-96.
14
Hess, S., Spitz, G., Bradley, M., & Coogan, M. (2018). Analysis of mode choice for intercity travel:
15
Application of a hybrid choice model to two distinct US corridors. Transportation Research Part A: Policy
16
and Practice, 116, 547-567.
17
Hirk, R., Hornik, K., & Vana, L. (2017). mvord: an R package for fitting multivariate ordinal regression
18
models. R package vignette, 2018a. URL https://cran. r-project.
19
org/web/packages/mvord/vignettes/vignette_mvord. pdf.
20
ISFORT (2018). 15° Rapporto sulla mobilita` in Italia. www.isfort.it.
21
Kaplan, S., Wrzesinska, D. K., & Prato, C. G. (2019). Psychosocial benefits and positive mood related to
22
habitual bicycle use. Transportation research part F: traffic psychology and behaviour, 64, 342-352.
23
Kelarestaghi, K. B., Ermagun, A., & Heaslip, K. P. (2019). Cycling usage and frequency determinants in
24
college campuses. Cities, 90, 216-228.
25
Kroesen, M., Handy, S., & Chorus, C. (2017). Do attitudes cause behavior or vice versa? An alternative
26
conceptualization of the attitude-behavior relationship in travel behavior modeling. Transportation
27
Research Part A: Policy and Practice, 101, 190-202.
28
Ma, L., Dill, J., & Mohr, C. (2014). The objective versus the perceived environment: what matters for
29
bicycling?. Transportation, 41(6), 1135-1152.
30
La Paix, L., Cherchi, E., & Geurs, K. (2020). Role of perception of bicycle infrastructure on the choice of the
31
bicycle as a train feeder mode. International Journal of Sustainable Transportation, 1-14.
32
Manaugh, K., Boisjoly, G., & El-Geneidy, A. (2017). Overcoming barriers to cycling: understanding
33
frequency of cycling in a University setting and the factors preventing commuters from cycling on a
34
regular basis. Transportation, 44(4), 871-884.
35
Manton, R., Rau, H., Fahy, F., Sheahan, J., & Clifford, E. (2016). Using mental mapping to unpack
36
perceived cycling risk. Accident Analysis & Prevention, 88, 138-149.
37
Martens, K. (2007). Promoting bike-and-ride: The Dutch experience. Transportation Research Part A:
38
Policy and Practice, 41(4), 326-338.
39
Muñoz, B., Monzon, A., & Daziano, R. A. (2016). The Increasing Role of Latent Variables in Modelling
40
Bicycle Mode Choice. Transport Reviews, 36(6), 737-771.
41
Namgung, M., & Jun, H. J. (2018). The influence of attitudes on university bicycle commuting: Considering
42
bicycling experience levels. International Journal of Sustainable Transportation, 1-14.
43
Noland, R. B., Deka, D., & Walia, R. (2011). A statewide analysis of bicycling in New Jersey. International
44
Journal of Sustainable Transportation, 5(5), 251-269.
45
34
Oliva, I., Galilea, P., & Hurtubia, R. (2018). Identifying cycling-inducing neighborhoods: A latent class
1
approach. International journal of sustainable transportation, 12(10), 701-713.
2
Raveau, S., Álvarez-Daziano, R., Yáñez, M. F., Bolduc, D., & Ortúzar, J.d.D. (2010). Sequential and
3
simultaneous estimation of hybrid discrete choice models: some new findings. Transportation Research
4
Record, 2156(1), 131-139.
5
Rojas-Rueda, D., de Nazelle, A., Tainio, M., & Nieuwenhuijsen, M. J. (2011). The health risks and benefits
6
of cycling in urban environments compared with car use: health impact assessment study. Bmj, 343, d4521.
7
Sallis, J. F., Conway, T. L., Dillon, L. I., Frank, L. D., Adams, M. A., Cain, K. L., & Saelens, B. E. (2013).
8
Environmental and demographic correlates of bicycling. Preventive medicine, 57(5), 456-460.
9
Sardegna Statistiche (2018). Sardegna in cifre 2018
10
http://www.sardegnastatistiche.it/documenti/12_103_20181212133014.pdf
11
Sener, I. N., Eluru, N., & Bhat, C. R. (2009). Who are bicyclists? Why and how much are they bicycling?.
12
Transportation Research Record, 2134(1), 63-72.
13
Sottile, E., Sanjust di Teulada, B., Meloni, I., & Cherchi, E. (2019). Estimation and validation of hybrid
14
choice models to identify the role of perception in the choice to cycle. International journal of sustainable
15
transportation, 13(8), 543-552.
16
Sottile, E., Diana, M., Piras, F., Meloni, I., & Pirra, M. (2020). To play but not for travel: Utilitarian, hedonic
17
and non-cyclists in Cagliari, Italy. In Mapping the Travel Behavior Genome (pp. 209-228). Elsevier.
18
Stinson, M. A., & Bhat, C. R. (2004). Frequency of bicycle commuting: internet-based survey analysis.
19
Transportation Research Record, 1878(1), 122-130.
20
Stinson, M. A., Porter, C. D., Proussaloglou, K. E., Calix, R., & Chu, C. (2014). Modeling the impacts of
21
bicycle facilities on work and recreational bike trips in Los Angeles County, California. Transportation
22
Research Record, 2468(1), 84-91.
23
Swiers, R., Pritchard, C., & Gee, I. (2017). A cross sectional survey of attitudes, behaviours, barriers and
24
motivators to cycling in University students. Journal of Transport & Health, 6, 379-385.
25
Thigpen, C. (2019). Do bicycling experiences and exposure influence bicycling skills and attitudes?
26
Evidence from a bicycle-friendly university. Transportation research part A: policy and practice, 123, 68-
27
79.
28
Ton, D., Duives, D. C., Cats, O., Hoogendoorn-Lanser, S., & Hoogendoorn, S. P. (2019). Cycling or
29
walking? Determinants of mode choice in the Netherlands. Transportation research part A: policy and
30
practice, 123, 7-23.
31
Vij, A., & Walker, J. L. (2016). How, when and why integrated choice and latent variable models are latently
32
useful. Transportation Research Part B: Methodological, 90, 192-217.
33
Walker, J. L. (2001). Extended discrete choice models: integrated framework, flexible error structures, and
34
latent variables (Doctoral dissertation, Massachusetts Institute of Technology).
35
Wang, Y., Chau, C. K., Ng, W. Y., & Leung, T. M. (2016). A review on the effects of physical built
36
environment attributes on enhancing walking and cycling activity levels within residential neighborhoods.
37
Cities, 50, 1-15.
38
Willis, D. P., Manaugh, K., & El-Geneidy, A. (2015). Cycling under influence: summarizing the influence of
39
perceptions, attitudes, habits, and social environments on cycling for transportation. International Journal
40
of Sustainable Transportation, 9(8), 565-579.
41
Xing, Y., Handy, S. L., & Mokhtarian, P. L. (2010). Factors associated with proportions and miles of
42
bicycling for transportation and recreation in six small US cities. Transportation research part D:
43
Transport and Environment, 15(2), 73-81.
44
Yang, Y., Wu, X., Zhou, P., Gou, Z., & Lu, Y. (2019). Towards a cycling-friendly city: An updated review
45
of the associations between built environment and cycling behaviors (2007–2017). Journal of Transport &
46
Health, 14, 100613.
47
35
Zhang, C. Q., Zhang, R., Gan, Y., Li, D., & Rhodes, R. E. (2019). Predicting transport-related cycling in
1
Chinese employees using an integration of perceived physical environment and social cognitive factors.
2
Transportation research part F: traffic psychology and behaviour, 64, 424-439.
3