Gender Differences in the Perception of Safety in Public Transport

Abstract

Concerns over women's safety on public transport systems are commonly reported in the media. We develop statistical models to test for gender differences in the perception of safety and satisfaction on urban metros and buses by using large‐scale unique customer satisfaction data for 28 world cities over the period 2009–2018. Results indicate a significant gender gap in the perception of safety, with women being 10% more likely than men to feel unsafe in metros (6% for buses). This gender gap is larger for safety than for overall satisfaction (3% in metros and 2.5% in buses), which is consistent with safety being one dimension of overall satisfaction. Results are stable across specifications and robust to inclusion of city level and time controls. We find heterogeneous responses by sociodemographic characteristics. Data indicate that 45% of women feel secure in trains and metro stations (and 55% in buses). Thus the gender gap encompasses more differences in transport perception between men and women rather than an intrinsic network fear. Additional models test for the influence of metro characteristics on perceived safety levels and find that more acts of violence, larger carriages and emptier vehicles decrease women's feeling of safety.

Full text

Gender Differences in the Perception of Safety in1
Public Transport2
Laila Ait Bihi Ouali1, Daniel J. Graham2, Alexander Barron2, and Mark3
Trompet2
4
1Corresponding author: Imperial College London, London, SW7 2AZ, UK. Email:5
laitbihi@ imperial. ac. uk.6
1,2Transport Strategy Centre, Imperial College London7
Accepted Version [04-February-2020]8
Abstract9
Concerns over women’s safety on public transport systems are commonly reported in the10
media. In this paper we develop statistical models to test for gender differences in the11
perception of safety and satisfaction on urban metros and buses using large-scale unique12
customer satisfaction data for 28 world cities over the period 2009 to 2018. Results in-13
dicate a significant gender gap in the perception of safety, with women being 10% more14
likely than men to feel unsafe in metros (6% for buses). This gender gap is larger for safety15
than for overall satisfaction (3% in metros and 2.5% in buses), which is consistent with16
safety being one dimension of overall satisfaction. Results are stable across specifications17
and robust to inclusion of city-level and time controls. We find heterogeneous responses18
by sociodemographic characteristics. Data indicates 45% of women feel secure in trains19
and metro stations (respectively 55% in buses). Thus the gender gap encompasses more20
differences in transport perception between men and women rather than an intrinsic net-21
work fear. Additional models test for the influence of metro characteristics on perceived22
safety levels and find that that more acts of violence, larger carriages, and emptier vehicles23
decrease women’s feeling of safety.24
Keywords: Gender; Safety; Public Transport; Metros; Buses; Customer Satisfaction; Be-25
havioural Responses26
1 Introduction27
Customer satisfaction and passenger safety and security are priorities for most public trans-28
port providers. Aside from the obvious moral, social and equity concerns; lower safety levels29
are associated with reductions in passenger ridership and thus revenues (e.g. Lynch and Atkins30
1988, Carter 2005, Delbosc and Currie 2012). This provides ample incentive for operators to31
intervene by manipulating the public transport characteristics to raise customer satisfaction32
(Zelinka and Brennan 2001). But safety measures can have different impacts on women rela-33
tive to men (Yavuz and Welch 2010), and existing research finds that there can be significant34
mismatches between female customers’ safety needs and strategies implemented by public35

transport operators (Loukaitou-Sideris and Fink 2009). The introduction of women-only car-1
riages is a case in point which has been widely trialled, but has received mixed reactions from2
female passengers.3
A well established strand of literature finds that perceived safety in urban areas varies by4
gender. Women typically have lower perceived safety in public spaces and a greater fear5
of being alone in concealed spaces and with strangers (O’Brien 2005, Jorgensen et al. 2013,6
Steinmetz and Austin 2014, Madge 1997). This points to the conclusion that women may be7
on average more likely to perceive public spaces as less safe (Pain 1997). However, men are8
also found to have similar fears, but perhaps under-report in surveys partly to abide by social9
norms (Yavuz and Welch 2010). For these reasons, it is useful to explore gender differences in10
the perception of safety rather than women’s absolute perception of safety.11
While womens’ fear of crime in public spaces has been widely studied (e.g. Hall 1985, Riger12
and Gordon 1989, Valentine 1990, Gilchrist et al. 1998, Koskela and Pain 2000, Pain 2001),13
safety in public transport has received little statistical attention. Does the perception of safety14
in public transport differ by gender, and to what extent? If so, what are the potential drivers15
of these differences? These are the questions we seek to answer in this paper.16
This paper focuses on the individuals’ perception of their own security. We define it as the17
estimated risk or threat of an intentional personal attack or aggression1. Dependent variables18
used in this study are ordinal thus our empirical strategy relies on ordered probit specifica-19
tions. To the best of our knowledge, this study is the first cross-country statistical analysis of20
gender differences in the perception of both safety and satisfaction in public transport. The21
first analysis uses data from two original Customer Satisfaction Surveys (CSS) respectively22
conducted for (i) urban metros (over 2014-2018) and (ii) buses (over 2009-2018). The richness23
and abundance of the data is ideal for our study as it allows us to provide robust results, gen-24
erate conclusions with wide geographical relevance, and compare gender gaps across different25
transport modes.26
Following quantification of gender differences in the perception of safety in public transport,27
the second aim of this paper is to disentangle potential drivers of observed differences. This28
is relevant as it assesses whether the design and other attributes of public transport help29
improve perceived levels of safety. The motivation behind this analysis is developed in the30
urban social geography and planning literatures which explain that initiatives to “design out”31
fear of crime in urban areas have little effects on perceived safety (Koskela and Pain 2000).32
We draw on Key Performance Indicators data (KPI) that covers extensively 25 metro systems33
for 2014-2018.34
Results suggest a significant gap between men and women both in perceived satisfaction and35
safety. We find that women are 10% more likely than men to feel unsafe in metros and 6%36
more likely to feel unsafe in buses. This gender gap is smaller for overall satisfaction, as37
women are 3% more likely to be dissatisfied with the general service in metros and 2.5% in38
buses. This smaller magnitude is consistent with safety being only one important dimension39
of overall satisfaction (Oliver 1997). Despite this gap, women are on average satisfied with40
both safety and the service. Results are stable across specifications and robust to city and41
1We acknowledge that from a practitioners’ viewpoint, definitions of safety and security differ where safety
refers to infrastructure failure and security defines unlawful acts interfering with individuals. However, to
align our interpretations with the academic literature in urban and transport economics, we will refer to the
perception of safety and security identically.
2

time controls. Our results show heterogeneity with age and by geographical area. On the1
influence of metro characteristics on safety levels, we find that more acts of violence, larger2
carriages, and emptier vehicles decrease the feeling of safety among women.3
2 Data4
The data used in this study have been collected annually over the period 2009 to 2018 via5
responses to Customer Satisfaction Surveys (CSSs) from users of urban metros and buses.6
Data collection is facilitated through collaboration between the Transport Strategy Centre7
(TSC) at Imperial College London and several major public transport operators, organised in8
the form of two consortia of urban metros, CoMET (Community of Metros) and Nova, and9
with bus operators through the IBBG (International Bus Benchmarking Group). The CoMET10
and Nova groups cover 25 cities across Europe, Americas and Asia, while the IBBG group11
is a consortium of 14 bus networks across Europe, Asia and North America. The detailed12
composition of each consortium is presented in Appendix A.13
Both the metro and bus CSSs have a similar structure (Trompet et al. 2013, 2018). The first14
part of the questionnaire contains statements relating to eight customer service areas defined15
under European Norm 13816: availability, time, information, comfort, security, customer16
care, accessibility and environment (see EN13816 2002). In addition, there is one general17
question on overall satisfaction. The questionnaires are produced and disseminated via an18
online survey building and hosting tool. Where necessary, translations of the survey are19
provided by operators into their home languages. Participating operators posted the link(s)20
to their own survey(s) on their home page for the same 4-week period each year, or via their21
social media pages or email bulletins. The important argument in favour of comparability is22
contained in the consistency of this method: face-to-face or phone interviews are never led in23
order to avoid interaction with staff that could bias the results of the respondents.24
Respondents are asked to provide their opinions on normal service operations. In the first sec-25
tion, answers are given on an increasing scale going from “1−Agree Strongly” to “5–Disagree26
Strongly”. Questions on safety in transport follow this structure. The scale of possible27
responses on the perception of overall satisfaction is similar and ranges increasingly from28
“1−Strongly Dissatisfied” to “5–Strongly Satisfied”. The second section of the survey asks29
respondents to select, in order of preference, the three most important customer service areas30
to them. Finally there are four demographic questions to understand the sample frame.31
Data available for estimation encompass information yearly for a group of respondents who32
are customers of urban metros and buses. Metro data records a total of 137,513 observations33
for 25 cities over the period going from 2014 to 2018. Bus data contains a total of 189,89034
observations for 14 cities between 2009 and 2018. Descriptive statistics are in Tables 9 and35
10 in Appendix respectively for metros and buses. The distribution of genders is rather equal36
for both datasets for all years; in addition, respondents are often daily commuters who travel37
for work. Note that due to the need to respect confidentiality it is not possible to reveal the38
name of particular transport systems in our analyses.39
The CSS dataset’s main advantage is in the abundance of observations: data covers a critical40
number of cities worldwide over a consequent time period, which conveys our results robustness41
and external validity of our results. Another unique advantage is in the representativeness42
3

of the data as transport providers sample their respondents every year that way. This cross-1
sectional dataset does not follow respondents over time – yet, the representativeness of the2
data over each year allow us to assume that characteristics do not change much over time.3
Smartcard datasets often contain more information on individuals, such as the metro line4
they take, and OD pairs can enable us to pinpoint daily commuters and infer where they live.5
However, despite the lack of side variables of this sort, survey data still remains until today6
the only resource that can tell us about perceptions and customer satisfaction. Therefore,7
survey data still remains essential to complement demand analysis and observe behaviours8
through the prism of satisfaction surveys.9
3 The Model10
All outcome variables in our study are discrete, representing individual opinions on safety11
and satisfaction in public transport. As such, they have non-normal error distributions and12
so we use Generalised Linear Models (GLMs) for analysis, and specifically ordered probit13
specifications for ordinal variables.14
While the dependent variables are ordinal, they are not continuous in the sense that the metric15
used to code the dependent variables encompass different satisfaction levels. Satisfaction16
levels are represented on a 5-point scale and assigns the numbers {1, ..., 5}to the categories17
{“Strongly Disagree”,..., “Strongly Agree”}. The metric relating numbers (from 1 to 5) is18
linear whereas the metric underlying the satisfaction scale is not. For instance, the difference19
between 1 and 3 (“Strongly Disagree” to “Neither Agree nor Disagree”) is likely to be quite20
different from the difference between 2 and 4 (“Disagree” to “Agree”). The scale is the same21
for the outcome variable measuring overall satisfaction with the service, but ranges from22
“Strongly Dissatisfied” to “Strongly Satisfied”.23
The main GLM we consider is a standard response model in which the cumulative probabilities24
of the discrete outcome are related to an index of explanatory variables. Let yibe the observed25
ordinal variable, then we model Pr[yi≤j|x] = φ(αj−x0
iβ) with j={1, ..., 5}, where αjand β26
are model parameters to be estimated and φis the standard normal cumulative distribution27
function. We assume there is a latent continuous metric behind the observed ordinal responses.28
The observed dependent variable yican take values from 1 to 5 such that29
yi=j⇐⇒ αj−1< y∗
i< αj
30
with j={1, ..., 5}, and where αdesignates the cutpoints estimated by the data. Cutpoints31
help in matching the probabilities associated with each discrete outcome. However, the metric32
of the observed variable yiis linear but the satisfaction scale is not. Therefore, we assume33
there is a latent continuous metric behind the observed ordinal responses.34
The latent continuous variable y∗represents the satisfaction scale. It is a linear combination35
of predictors and an error term and can be written36
y∗
i=xiβ+
N
X
k=1
δkck+
M
X
m=2009
γmtm+ui(1)
4

where y∗
iis the dependent variable (i.e. a statement on safety or satisfaction in public trans-1
port), uiis the error term, xiis a vector of observable individual characteristics, PN
k=1 ckis2
a set of city specific dummy variables, and PM
m=2009 tmis a set of time dummy variables for3
each year.4
Following this model, the probability of observing outcome jcorresponds to the probability5
that the estimated linear function, plus random error, is within the range of the cutpoints6
estimated for the outcome7
P r(yi=j) = P r(αj−1< y∗
i≤αj)
=P r(αj−1< xiβ+ui≤αj)
=P r(αj−1−xiβ < ui≤αj−xiβ)
=φij −φij−1
(2)
where φrepresents the cumulative distribution function in the standardised normal distri-8
bution. The estimation of the regression coefficients in vector βis achieved by maximum9
likelihood.10
Covariates representing sociodemographic characteristics in Equation 1 contribute to a better11
understanding of the drivers of perceived safety. Our covariates also helps distinguishing12
groups of people who are more or less likely to perceive a high (or low) level of safety. First,13
we include measures of travel frequency and travel purpose (e.g. work, education). These14
two dimensions characterise several profiles and whether travel is constrained or for travel at15
different dates, times and frequencies. We also add time dummies to account for potential16
exogenous shocks and evolutions over time. Finally, city fixed effects are deemed helpful to17
account for network (buses or metro) characteristics, city structure and safety levels and also18
country civic values2.19
The choice of an ordered probit over an ordered logit model relies on the assumption that the20
underlying distribution of the observed outcome is normal. In an ordered probit specification,21
the observed outcome (e.g., “Agree”) reflects a threshold is met for the underlying latent22
variable which is normally distributed. The plotted distribution of satisfaction in our data23
provided a relatively normal distribution with a mean satisfaction at about scale 4 (and of 3.2524
for both security in trains and security in stations). Thus, normal distributions appear more25
suited to define satisfaction levels. Finally, when it comes to the computation of marginal26
effects, both logit and probit models make essentially similar predictions3.27
Ordered probit models rely on a couple of assumptions that appear as quite strong: (i) the28
constant threshold assumption; (ii) the distributional assumption. The distributional assump-29
tion assumes there is no additional individual heterogeneity between individual realisations.30
In our case, women may have heterogeneous behaviours as unmeasured variables can affect31
the chances of feeling more or less safe in public transport (e.g., risk aversion). Therefore, to32
counteract this limitation, we compute multilevel mixed-effects ordered probit models that ac-33
count for this individual heterogeneity (see Tables 16 and 11 for random intercept and random34
coefficient specifications).35
Mixed-effects ordered probit models contain both fixed and random effects. As such, they36
2Small differences between bus and metro surveys exist as each group sets their own rules within our
framework. This is at the root of the differences in the “Main Travel Purpose” variable: the bus consortium
grouped ‘Education’ with ‘Work’ as they considered them as similar types of public transport uses.
3We computed estimates using logit models. Results are widely comparable in sign and significance.
5

allow for many levels of nested clusters of random effects. Using the original framework1
presented in equation (2), we now assume a series of K independent clusters. This method2
accounts for unobserved heterogeneity at the city level. Therefore, conditional on a set of3
fixed effects xik, and a set of random effects uk, we can derive the probability of observing4
outcome jas5
P r(yik =j|α,uk) = P r(αj−1< y∗
ik≤αj)
=P r(αj−1< xik β+zik uk+eik ≤αj)(3)
where y∗
ik is the latent continuous response (i.e., a statement on safety or satisfaction in6
public transport). eik is the error term which follows a standard normal distribution and is7
independent of uk.8
This multilevel mixed effect model accounts for city-level heterogeneity which is due to the9
nature of our dataset. The Customer Satisfaction Survey (CSS) is a cross-sectional dataset10
at the individual level. As such, one observation line represents the level of satisfaction at11
time t for a given individual i in a given city. Therefore, the city level is the smallest level at12
which we have unobserved heterogeneity we can control for. We acknowledge that only city13
level heterogeneity is accounted for, not individual level heterogeneity. However, since random14
effects cannot adjust for confounding in any case, this should not materially affect the overall15
conclusions drawn from this study.16
Coefficients remain of the same order of magnitude and indicates that in spite of its strength,17
this assumption does not impair the quality of our estimations. Also, although responses can18
differ by subgroup, the aim of this paper is essentially to observe the difference in perception19
on a worldwide scale while clearing effects of socio-demographics (age, habits) and allows us20
to treat satisfaction levels as a non-linear scale, which is a crucial condition.21
A second limitation of ordered probit models is contained in marginal probability effects22
whose signs’ change once when moving from the smallest to the largest outcome. We accept23
this assumption as it is compatible with the fact that the satisfaction scale, despite being24
non-linear, is an increasing one.25
4 Results26
4.1 Base Results27
4.1.1 Quantification of the gender effect28
Table 1 below presents ordered probit estimates for safety-related questions, estimated with29
covariates for sociodemographic characteristics included along with time and city fixed effects.30
Dependent variables are the discrete representation of statements on safety and overall satis-31
faction. Sensitivity tests show that all estimates, both for buses and metros, are stable across32
specifications and robust to covariate inclusion.33
Table 1 shows a negative and significant gap by gender for both metros and buses. This gap34
is observed for both safety and overall satisfaction statements. This corroborates the urban35
economics and crime literature (Hall 2005, Riger and Gordon 1989) as women’s bigger fear36
6

Metros Buses
The train is a secure Stations are a secure How satisfied are you The bus is a secure How satisfied are you
place for me place for me overall with the metro service? place for me overall with the bus service?
Female (Yes=1) -0.235*** -0.235*** -0.0452*** -0.105*** -0.0372***
(0.00539) (0.00539) (0.00548) (0.00516) (0.00515)
P-value of coeft. female <0.001 <0.001 <0.001 <0.001 <0.001
McFadden Adj.R20.043 0.045 0.081 0.019 0.029
Observations 169,582 169,658 169,831 179,773 180,208
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova and IBBG, own calculations.
Note: All estimates from regressions computed above contain city and year dummies.
Table 1: Ordered Probit Estimates: Gender Coefficients on Perception of Safety
of crime in public spaces is also now observed for public transport. This result also extends1
this literature (Twinam 2017, Phillips and Sandler 2015) by indicating that women and men2
differ not only in propensity to be targeted by crime but also in their perceptions. In addition,3
since the data covers a wide range of cities worldwide and results are robust, this observation4
is applicable to cities with a strong external validity.5
These first results highlight a first contribution of our analysis. The issue of women’s safety6
in public spaces has been included in the public discourse since the 1980s (Shaw 2002). How-7
ever, the literature has been limited thus far by a lack of no homogeneous, standardised way8
to record violences against women, as data is often incomplete or inaccurate and is subject to9
under-reporting (UN Women 2016). Yet we now find, using a homogenous dataset recording10
perceptions at a worldwide scale, that perceptions of safety are commonly share across the11
globe and that the gap in perception is significant. However, the magnitude of coefficients12
presented in Table 1 cannot be interpreted directly - marginal effects are presented and ex-13
tensively discussed in the rest of this section. Yet, these coefficients are still relevant. They14
are negative and significant at the 1% level: as such, they show there is a significant gap in15
the perception of safety between men and women.16
Assessing the magnitude of the difference between men and women in the perception of safety17
requires the computation of marginal effects. The interpretation of the marginal effects is18
straightforward for continuous variables, with a one unit change in the explanatory variable19
resulting in an increase or decrease in the probability equal to the size of the marginal effect20
(all other things equal). Since y∗cannot be observed and is purely artificial, its interpretation21
is not of interest.22
The most natural way to interpret ordered response models (and discrete probability models23
in general) is to determine how a marginal change in one regressor changes the distribution24
of the outcome variable, i.e. all the outcome probabilities. The main focus in the analysis of25
ordered data should be put on the conditional cell probabilities given by26
P r[y=j|x] = F(µj−x0β)−F(µj−1−x0β). (4)
with Fbeing the variance of the distribution function and βthe vector of coefficient attached27
to the vector of observable variables x. In order to identify the parameters of the model we28
have to fix location and scale of the argument in F, the former by assuming that xdoes not29
contain a constant term, the latter by normalizing the variance of the distribution function F30
M P Ejl(x) = δP r[y=j|x]
δxl
. (5)
7

The marginal effects represent the variation in the probability of picking one given response1
(e.g. “Agree”) if the individual is a woman. Marginal effects are computed for every possible2
response ranging from “Strongly Disagree” to “Strongly Agree”. Figure 1 indicates that3
women are 4% more likely than men to “Disagree Strongly” with the statement that they feel4
secure in trains. Marginal effects for gender are presented in Figures 1 and 2 below.5
In the case of a binary variable, the marginal effect is the change in predicted probability6
based on whether a respondent falls into that category or not. When calculating marginal7
effects all remaining variables assume their respective average values. As such, the marginal8
effects show the change in the predicted probability for each gender for an average respondent,9
according to the variable being considered.10
Figure 1: Marginal Effects for Safety-Related Questions - Metro Data
8

Figure 2: Marginal Effects for Safety-Related Questions - Bus Data
Figures 1 and 2 shows that overall, women are 10% more likely to make a negative statement1
on security than men. This figure shows that the two statements on security present marginal2
effects of a similar magnitude, which is in line with the fact that these statements are very close3
both in label and in nature. However, we note that we find a much larger effect magnitude for4
statements regarding safety compared to overall satisfaction. For comparison, Tables 12 and5
13 in the Appendix shows gender differences for all other dimensions of customer satisfaction6
available in the Customer Satisfaction Survey (e.g. Availability, Information, Customer Care),7
respectively for metros and buses. Results indicate that there is a negative and significant8
gender gap in the perception of all dimensions of customer satisfaction. However, the magni-9
tude of the gender gap is much larger for safety than it is for any other dimensions of customer10
satisfaction. These results are indicate that safety is the main vector of gender differences in11
perception.12
While we observe this gender gap, the data also reveal that women are generally happy with13
the overall quality of service. Figures 8 and 9 in the Appendix show the general probability14
of women in picking a given answer for each statement. These results suggest that although15
the gap between men and women is significant, the levels of satisfaction and safety of female16
customers still remains reasonably high.17
4.1.2 Analysis of other socio-demographic variables18
Table 2 below shows coefficients for the socio-demographic and economic variables for both19
transport modes. Stated feelings of safety in metros appear to decrease with age, although20
this feeling of safety is mostly increasing with age for buses. However, satisfaction with the21
general service evolves similarly for buses and metros and appears to increase from 40 years22
old onwards. Responses based on the main travel purpose appear to evolve similarly for both23
9

buses and metros: people travelling mostly for shopping and leisure purposes declare signifi-1
cantly higher safety levels than people who travel mostly to go to work or school. Individuals2
who travel for leisure or shopping are more likely to travel on their own terms, pick their3
destinations and routes as well as travel dates and times; they are also more likely to travel4
with people they know. In contrast, travelling for work or to go to school is more likely to5
abide by constraints.6
Travel frequency is also highly correlated with stated levels of safety, although perceptions7
by subgroups are different for buses than for metros. In both cases, individuals who travel8
“often” feel safer than those who travel “very often”. On the other hand, individuals using9
public transport “Very Rarely” are more likely to feel less safe.10
Metro Buses
The train is a secure Stations are a secure How satisfied are you The bus is a secure How satisfied are you
place for me place for me overall with the service? place for me overall with the service?
Female (Yes=1) -0.235*** -0.235*** -0.0452*** -0.105*** -0.0372***
(0.00539) (0.00539) (0.00548) (0.00515) (0.00515)
Age (ref: 18-29)
Less than 18 0.238*** 0.197*** 0.359*** 0.241*** 0.426***
(0.0144) (0.0144) (0.0148) (0.0126) (0.0127)
30-39 -0.0279*** -0.00773 -0.0442*** 0.0329*** -0.0666***
(0.00703) (0.00703) (0.00716) (0.00692) (0.00690)
40-49 -0.00685 0.0141* 0.0556*** 0.0701*** 0.0217***
(0.00826) (0.00826) (0.00841) (0.00784) (0.00783)
50-65 0.0251*** 0.0505*** 0.152*** 0.0857*** 0.0984***
(0.00857) (0.00857) (0.00875) (0.00771) (0.00771)
Over 65 0.152*** 0.178*** 0.325*** 0.212*** 0.282***
(0.0158) (0.0158) (0.0162) (0.0133) (0.0133)
Main travel purpose (ref: Work)
Education 0.0497*** 0.0451*** 0.0761***
(0.0116) (0.0116) (0.0118)
Shopping 0.0963*** 0.0866*** 0.155*** 0.0421*** 0.197***
(0.00853) (0.00853) (0.00869) (0.0134) (0.0134)
Leisure 0.159*** 0.151*** 0.252*** 0.104*** 0.231***
(0.00994) (0.00994) (0.0102) (0.00833) (0.00835)
Doctor 0.0614*** 0.0378*** 0.189*** -0.135*** 0.0448***
(0.0128) (0.0128) (0.0131) (0.0170) (0.0170)
Other 0.0445** 0.0296 0.128*** 0.0143 0.0237
(0.0211) (0.0211) (0.0215) (0.0158) (0.0157)
Frequency use trains (ref: very often)
Often 0.0836*** 0.0696*** 0.128*** 0.0717*** 0.115***
(0.00733) (0.00734) (0.00748) (0.00653) (0.00653)
Sometimes 0.0640*** 0.0394*** 0.155*** 0.0811*** 0.116***
(0.0104) (0.0104) (0.0106) (0.00880) (0.00881)
Rarely 0.0541*** 0.0231* 0.198*** 0.00361 0.0131
(0.0134) (0.0134) (0.0138) (0.0129) (0.0129)
Very Rarely 0.00238 -0.0320* 0.170*** -0.145*** -0.214***
(0.0171) (0.0171) (0.0176) (0.0169) (0.0169)
P-value of coeft. female <0.001 <0.001 <0.001 <0.001 <0.001
Observations 169,582 169,658 169,831 179,773 180,208
McFadden Adj.R20.043 0.045 0.081 0.019 0.029
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova and IBBG, own calculations.
Note that for buses, the reference category for the main travel purpose is Work/Education
Note: All estimates from regressions computed above contain city and year dummies.
Table 2: Estimates for sociodemographic variables - Bus and Metro Data
4.2 Heterogeneity Checks11
4.2.1 Heterogeneous responses by area12
Figures 3, 4, and 5 show the marginal coefficients of the gender gap by geographic region.13
This enables us to assess whether there is heterogeneity at the continent level.14
10

Figure 3: Regional Marginal Effects for Safety-Related Statement: “Stations are a secure
place for me”
Figure 4: Regional Marginal Effects for Safety-Related Statement: “Trains are a secure place
for me”
11

Figure 5: Regional Marginal Effects for Safety-Related Statement: “The bus is a secure place
for me”
To preserve anonymity and respect the confidential nature of our data we aggregated the1
information to the continent level. We find that for safety statements the evolution is the2
same for all continents but with a different magnitude by continent. Women are found have3
negative experiences in public transport such as harassment (Condon et al. 2007, European4
Commission 2014): our results suggest that they also share the same perceptions of safety.5
For metros, results suggest that both Europe and North America have a bigger gender gap in6
satisfaction compared to Asia or South America. For buses, we observe that Europe and North7
America have a comparable gender gap, with Asia presenting more differences and a higher8
propensity of women to be ”Neutral” regarding their perception of safety. This homogeneity9
is not observable for the “overall satisfaction” statement (see Figures 6 and 7 in Appendix).10
4.2.2 Heterogeneous responses by age11
Table 3 presents results on the heterogeneity of the gender effect by age. We find that overall12
satisfaction with public transport increases with age and that satisfaction increases at an13
increasing rate with age for safety statements. We find that the perception of safety in buses14
also increases with age. However, this effect is not observed for metros, where the perception15
of safety decreases at a decreasing rate with age. Since younger women are over represented16
in our data relative to the population, our estimated effect is likely conservative.17
12

Metro Buses
The train is a secure Stations are a secure How satisfied are you The bus is a secure How satisfied are you
place for me place for me overall with the service? place for me overall with the service?
Female (Yes=1) -0.291*** -0.312*** -0.0405*** -0.145*** -0.0123
(0.00829) (0.00830) (0.00844) (0.00878) (0.00877)
Age (ref: 18-29)
Less than 18 0.272*** 0.224*** 0.411*** 0.300*** 0.506***
(0.0167) (0.0167) (0.0171) (0.0164) (0.0165)
30-39 -0.0603*** -0.0625*** -0.0624*** -0.00595 -0.0832***
(0.00936) (0.00937) (0.00952) (0.00991) (0.00987)
40-49 -0.0626*** -0.0586*** 0.0481*** 0.0403*** 0.0388***
(0.0112) (0.0112) (0.0114) (0.0113) (0.0113)
50-65 -0.0498*** -0.0335*** 0.179*** 0.0334*** 0.146***
(0.0118) (0.0118) (0.0120) (0.0113) (0.0113)
Over 65 0.0895*** 0.0994*** 0.381*** 0.165*** 0.343***
(0.0202) (0.0202) (0.0208) (0.0179) (0.0180)
Age # Female
Female & (age: <18) -0.192*** -0.178*** -0.211*** -0.169*** -0.197***
(0.0324) (0.0324) (0.0332) (0.0257) (0.0259)
Female & (age: 30-39) 0.0685*** 0.117*** 0.0410*** 0.0749*** 0.0382***
(0.0133) (0.0133) (0.0135) (0.0137) (0.0137)
Female & (age: 40-49) 0.116*** 0.151*** 0.0166 0.0562*** -0.0295*
(0.0157) (0.0157) (0.0160) (0.0154) (0.0154)
Female & (age: 50-65) 0.151*** 0.171*** -0.0518*** 0.0957*** -0.0803***
(0.0160) (0.0160) (0.0164) (0.0147) (0.0147)
Female & (age: over 65) 0.147*** 0.184*** -0.135*** 0.0949*** -0.115***
(0.0302) (0.0302) (0.0311) (0.0246) (0.0247)
P-value of coeft. female <0.001 <0.001 <0.001 <0.001 0.156
Observations 169,582 169,658 169,831 179,616 180,047
McFadden Adj.R20.043 0.045 0.082 0.019 0.029
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1 Source: Customer Satisfaction Surveys, Comet/Nova and IBBG, own calculations.
Note: All estimates from regressions computed above contain city and year dummies.
Table 3: Estimates for heterogeneity based on age - Metro and Bus Data
4.2.3 Heterogeneous responses by travel purpose1
Table 4 presents estimates testing for heterogeneity by travel purpose. Results are similar for2
both statements on safety. There is significant heterogeneity between men and women only3
for shopping and leisure, with the latter being significantly less satisfied.4
13

Metro Buses
The train is a secure Stations are a secure How satisfied are you The bus is a secure How satisfied are you
place for me place for me overall with the service? place for me with the service?
Female (Yes=1) -0.212*** -0.206*** -0.0223*** -0.120*** -0.0415***
(0.00686) (0.00686) (0.00697) (0.00595) (0.00594)
Main travel purpose (ref: Work)
Education 0.0624*** 0.0571*** 0.0973***
(0.0145) (0.0145) (0.0148)
Shopping 0.146*** 0.148*** 0.173*** 0.0140 0.181***
(0.0112) (0.0112) (0.0114) (0.0202) (0.0203)
Leisure 0.183*** 0.181*** 0.286*** 0.0695*** 0.218***
(0.0121) (0.0121) (0.0124) (0.0108) (0.0108)
Doctor 0.0587*** 0.0426** 0.220*** -0.150*** 0.0721***
(0.0175) (0.0175) (0.0179) (0.0270) (0.0270)
Other 0.0812*** 0.0597** 0.185*** 0.00126 0.0253
(0.0288) (0.0288) (0.0295) (0.0233) (0.0233)
Main travel purpose # Female
Female & Education -0.0273 -0.0245 -0.0499**
(0.0218) (0.0219) (0.0223)
Female & Shopping -0.103*** -0.128*** -0.0353** 0.0486* 0.0287
(0.0153) (0.0153) (0.0156) (0.0256) (0.0258)
Female & Leisure -0.0515*** -0.0625*** -0.0756*** 0.0705*** 0.0272*
(0.0158) (0.0158) (0.0162) (0.0140) (0.0140)
Female & Doctor 0.00372 -0.0108 -0.0618*** 0.0265 -0.0414
(0.0227) (0.0227) (0.0232) (0.0335) (0.0336)
Female & Other -0.0756* -0.0612 -0.120*** 0.0244 -0.0106
(0.0413) (0.0413) (0.0422) (0.0306) (0.0306)
P-value of coeft. female <0.001 <0.001 0.001 <0.001 <0.001
Observations 169,582 169,658 169,831 179,615 180,046
McFadden Adj.R20.042 0.045 0.081 0.019 0.029
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1 Source: Customer Satisfaction Surveys, Comet/Nova and IBBG, own calculations.
Note: For the bus data, the reference category for the main travel purpose is ”Work/Education”.
For metro data, the reference category is ”Work” since the main travel purpose categories are more detailed.
Note: All estimates from regressions computed above contain city and year dummies.
Table 4: Estimates for heterogeneity based on travel purpose - Metro and Bus Data
5 Supplementary analyses1
The customer satisfaction data we use in this paper originate from an annually repeated cross-2
sectional survey. So far, our results have been estimated conditional on time and city fixed3
effects. However, as the underlying data comes from a repeated cross-section, it cannot be4
assumed that the participating respondents remain identical and constant over time. Our5
ordered probit estimates help assess the impact of explanatory variables on an ordinal vari-6
able. However, ordered probit estimates cannot account for unobserved heterogeneity among7
respondents over time. In our previous regressions, we try to disentangle the extent to which8
women’s perception of safety differs from that of men via inclusion of socio-demographic co-9
variates for gender, age, and travel conditions (purpose and frequency). But while these are10
helpful, they do not adjust comprehensively for other potential important time-varying or11
time–invariant effects, for instance marital status or the number of children. Since such char-12
acteristics could be relevant for women’ perception of safety, their omission from our model13
negates a causal interpretation of results.14
In order to address this issue, we adopt a pseudo-panel methodology which was first devel-15
oped by Deaton (1985) for individual level data. Similar to a standard panel methodology, the16
pseudo-panel approach carefully accounts for unobserved time-invariant heterogeneity facili-17
14

tating an improved understanding of the causal mechanisms by which outcomes are produced.1
We adopt the pseudo-panel approach as a means of estimating individual preferences in the2
absence of true individual panel data. It enables us to get adjust for unobserved heterogene-3
ity among cohorts and to corroborate our previous ordered probit estimates with a different4
methodology.5
The pseudo-panel method restructures the data and forms cohorts that are consistent over6
time. The cross-sectional data used to construct pseudo-panels must include information on7
at least one observable and time invariant variable by which observations can be grouped into8
cohorts. Then, cohort means are computed for all included variables and tracked over time9
to form a matrix of all cohorts as a pseudo-panel. This pseudo-panel of means does not suffer10
from attrition.11
We use the city identifier and the response date available for the metro customer satisfaction12
survey to identify the cohorts and generate the mean variables. For the bus data, response13
dates are not available for the whole period and we therefore construct our psuedo-panel only14
for metros.15
Under this approach the dependent variable represents the satisfaction scale as a continuous16
variable. It is defined as17
yct =xctβ+αc+δt+uct (6)
where crepresents a city / response date cohort, trepresents the date at which individuals fill18
the survey, xct is a vector of observable characteristics, αcare cohort fixed effects, δtare time19
fixed effects, and uct the error term. Assuming the size of the cohorts is sufficiently large and20
the composition relatively stable over the years, the daily cohort average of the firm-specific21
time-invariant effect can be transformed into a city time invariant effect αcwhich allows us22
to control for unobserved heterogeneity between cohorts.23
Each observation in the following analysis is thus the mean individual response of a city cohort24
at time t, hence allowing us to estimate the effect of the average share of women on the level25
of safety perceived by individuals.26
Table 5 shows the results from our pseudo-panel models. We find a significant and negative27
gap between men and women in our sample, indicating that women are less likely to feel safe28
in metros compared to men. This is applicable to both the perception of safety in trains and29
stations. While it is difficult to compare the magnitude of the coefficients in the Fixed Effects30
(FE) estimation model with those of the ordered probit model, we note that all coefficients31
linked to gender have the same negative sign and are highly significant in this specification.32
The pseudo-panel method we use relies on the assumption that the covariates are endogeneous33
and does not adjust for potential time-varying unobserved heterogeneity. We test whether re-34
sults are robust to these assumption via a Generalized method of moments (GMM) method-35
ology using instrumental variables derived for our pseudo-panel. We compute estimators from36
two-stage least-squares generalizations of simple panel-data estimators for exogenous vari-37
ables. In this study, we use the lagged differences of variables as instrumental variables for38
levels:39
yct =xctβ+αc+δt+uct
with xct = ∆xct−1γ+ ∆xct−2φ+vct
(7)
15

where ∆xct−1and ∆xct−2represent the vectors of instrumental variables, respectively the1
first and second lags of the growth of variables; vct is the error term. Table 6 presents the2
results from the GMM specification. We find that a larger share of women in a given city3
decreases the level of the results are robust to this specification. Regarding the magnitude4
of coefficients, we observe that GMM estimates record that a larger share of women decrease5
even more the perception of safety compared to the regular pseudo-panel specification. This6
appears to indicate that not controlling for endogeneity biased estimates upwards.7
The train is a secure Stations are a secure How satisfied are you
place for me place for me overall with the metro service?
Share of female respondents -0.216*** -0.200*** -0.147*** -0.138*** -0.00795 -0.000838
(0.0420) (0.0428) (0.0419) (0.0428) (0.0480) (0.0483)
Socio-demographic controls No Yes No Yes No Yes
Observations 2,955 2,899 2,955 2,899 2,954 2,898
R-squared 0.009 0.024 0.004 0.016 0.000 0.043
Number of cities 25 25 25 25 25 25
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Table 5: Pseudo Panel Estimates: Gender Coefficients on Perception of Safety - Metros
The train is a secure Stations are a secure How satisfied are you
place for me place for me overall with the metro service?
Share of female respondents -3.243* -2.948** -3.615* -3.236* -4.887* -4.489*
(1.697) (1.492) (2.150) (1.875) (2.728) (2.421)
Year dummies No Yes No Yes No Yes
Observations 2,189 2,189 2,189 2,189 2,189 2,189
R-squared 0.11 0.13 0.05 0.06 0.14 0.16
Number of cities 25 25 25 25 25 25
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Table 6: GMM Pseudo Panel Estimates: Gender Coefficients on Perception of Safety - Metros
Finally, Tables 11 and 16 in Appendix present results of a multilevel mixed effects model with8
both a random intercept and a random coefficient. This specification accounts for unobserved9
heterogeneity at the city level and includes a random coefficient for the gender variable at the10
city level. Estimates indicate that our results are robust to this more general specification11
for the safety-related outcomes, which confirms that safety in transport is subject to a gap12
between men and women.13
6 Analysing potential drivers of perceived safety in metros14
Having found convincing evidence for a gender gap in public transport safety we now move15
on to the second aim of this paper, which is to disentangle potential drivers of these observed16
differences. To do this we use extensive data collected by the TSC over the period 2014 to17
2018 for metro systems in 25 cities worldwide. An equivalent version of data is available for18
16

buses, but it does not cover the same time span, and is not as complete; therefore, our analysis1
is restricted to urban metros.2
The choice of covariates is based on the physical characteristics in the previous literature that3
affect individuals’ perception of risk and fear (Atkins 1990, Valentine 1990). The metro data4
allow us to test for the influence of the following covariates on the perceptions of safety: (i)5
number of staff members in metro stations (both own staff and contractors), (ii) number of6
cars per train, (iii) the average number of stations served by line (on average), (iv) the number7
of violent acts committed in the metro, (v) the total car capacity (seating and standing), and8
(vi) the metro ridership.9
The number of staff members encompasses here the metro providers’ own staff present in10
metro stations. Customers can refer to station employees in case of emergency, and they are11
easy to notice as they potentially have a marker (uniform, badge) that can indicate that they12
work for the metro company. Therefore, we expect a potential positive or neutral effect of staff13
members. Indeed, more staffing does not necessarily means constant staffing or staff members14
dedicated to the personal security of individuals (Atkins 1990).15
The total car capacity is the sum of the total seating and the total standing capacity. The total16
standing capacity is measured as the number of standing individuals per m2, and provides a17
good indication of the crowding levels while taking into account crowding expectations per18
city. Individuals tend to be more fearful in environments where they do not have a clear view19
of their surroundings (Loukaitou-Sideris and Fink 2009, Zelinka and Brennan 2001): hence we20
expect that more car capacity increase the level of perceived safety.21
Metro ridership defines the number of journeys done by passengers (including fare evaders).22
The average number of lines by stations and the average metro line length are computed using23
the 2018 World Metros Figure produced by the International Association of Public Transport24
(UITP)4. An increased metro ridership is expected to have a positive effect on perceived25
security. It follows Jacobs (1961) which explains that reaching a critical mass of individuals26
around reduces crime, as there are more ”eyes on the street”, the area then becomes more27
self-policing. Finally, more acts of violence increase the individual perceived risk of accident,28
and is thus expected to decrease the level of perceived safety. Individuals also appear to be29
fearful of empty train wagons of Transport (1997). To summarise, the metro ridership and30
the car capacity respectively indicate actual and potential crowding.31
The number of violent acts is the sum of two events that are (a) the number of robberies32
and (b) the number of acts of violence that occur per year in the metro. The impact of the33
each subcategory on the perception of safety is included in additional regressions in Appendix34
(Tables 14 and 15), as incivilities are predictive of fear in urban spaces (Rohe and Burby35
1988).36
The penultimate control is the average number of stations by line. More stations by line implies37
more time potentially spent in the metro, therefore a longer time when individuals wait in38
transports. A strand of the literature (Loukaitou-Sideris 1999) finds that longer waits lead to39
more anxiety and concerns for personal safety. In addition, this anxiety effect is magnified by40
individuals’ tendency to overestimate waiting times (Fan et al. 2016). The number of cars per41
train is a feature that indicates the potential to exit between two stations in case of emergency;42
4See more about the data: https://www.uitp.org/world-metro-figures-2018. This dataset covers a time span
going from 2013 to 2017 included.
17

therefore, we expect a positive effect of cars per train on perceived safety.1
The specification we use has the same general form as equation 1, but incorporates more2
independent variables as follows:3
y∗
i=x1
iβ1+x2
iβ2+
N
X
k=1
δkck+
M
X
m=2009
γmtm+ui(8)
where x1
iis a vector of observable individual characteristics as the one used in Table 2 (age,4
gender, frequency of use, motive of use). The general influence of the metro characteristics is5
included in the vector x2
i.6
Table 7 presents the ordered probit estimates from the model in equation 8. The table also7
uses as an outcome the likelihood of recommending the service to someone known to the8
respondent5. The coefficient related to women remains stable to the inclusion of metro char-9
acteristics (columns 1 and 2). However, overall satisfaction and the likelihood to recommend10
the metro to another person encompass a smaller gender difference than security items, which11
is in line with security being one dimension of overall satisfaction.12
We find that more acts of violence, less cars per train, more stations served by line are13
associated with lower overall customer satisfaction and perceived safety levels. However, the14
presence of more staff members (both regular staff and police force) has no significant impact15
on perceived safety. Metro ridership coefficients indicate that individuals feel more secure in16
metros with more human activity, which is in line with the literature. We note that questions17
on overall customer satisfaction differ slightly from security questions.18
Stations are secure Trains are secure Overall Likelihood to recommend
for me for me satisfaction to another person
(1) (2) (3) (4)
Female (Yes=1) -0.241*** -0.239*** -0.0422*** -0.0156**
(0.00673) (0.00673) (0.00684) (0.00658)
Total staff in stations -0.0146 0.0274 -0.118*** -0.0967***
(0.0211) (0.0210) (0.0215) (0.0210)
Metro ridership 0.00157*** 0.00218*** 0.00419*** 0.00349***
(0.000214) (0.000214) (0.000218) (0.000209)
Nb of cars per train 0.00726*** 0.0141*** 0.00593** -0.000416
(0.00258) (0.00257) (0.00264) (0.00252)
Avg number stations by line -0.0453*** -0.0663*** -0.0910*** -0.0751***
(0.00547) (0.00546) (0.00554) (0.00532)
Total car capacity 0.00450*** 0.00484*** 0.0128*** 0.0125***
(0.000653) (0.000652) (0.000665) (0.000637)
Nb of violent acts -0.000205*** -0.000161*** -0.000130*** -0.000136***
(2.70e-05) (2.70e-05) (2.76e-05) (2.67e-05)
P-value of coeft. female <0.001 <0.001 <0.001 0.018
Observations 109,630 109,548 109,743 108,057
McFadden Adj.R20.037 0.033 0.076 0.036
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Note: All estimates from regressions computed above contain city and year dummies, as well as sociodemographic controls.
Table 7: Effect of metro characteristics on the perception of safety by gender
Results suggest that the presence of more staff in stations is associated with lower overall19
5The likelihood of recommending the service to someone is on a increasing 10-point scale, where 10 indicates
that the individual will recommend the service for sure
18

satisfaction. This is a counter-intuitive result, perhaps indicative of lower safety satisfaction1
in less modern / user-friendly systems. We also include a measure of the total number of acts2
of violence which have a significant negative effects on perception of safety.3
6.1 Testing for Gender Heterogeneity4
Table 8 presents estimates for heterogeneity in the perception of metro characteristics by5
gender. Testing for gender heterogeneity is motivated as responses to metro providers’ policies6
differ by subgroup (Yavuz and Welch 2010). To derive results, we run regressions on the7
subgroup of women.8
Stations are secure Trains are secure Overall Likelihood to recommend
for me for me satisfaction to another person
(1) (2) (3) (4)
Female (Yes=1) - - - -
Total staff in stations 0.0160 0.0614* -0.0855*** -0.0323
(0.0322) (0.0322) (0.0331) (0.0319)
Metro ridership 0.00167*** 0.00251*** 0.00680*** 0.00469***
(0.000416) (0.000416) (0.000426) (0.000408)
Nb of cars per train 0.0142*** 0.0229*** 0.0163*** 0.0121***
(0.00384) (0.00383) (0.00395) (0.00374)
Avg number stations by line -0.0716*** -0.103*** -0.144*** -0.126***
(0.00962) (0.00962) (0.00977) (0.00937)
Total car capacity 0.00658*** 0.00811*** 0.0196*** 0.0180***
(0.00101) (0.00101) (0.00103) (0.000982)
Nb of violent acts -0.000292*** -0.000261*** -0.000295*** -0.000226***
(4.43e-05) (4.43e-05) (4.55e-05) (4.38e-05)
Observations 50,866 50,816 50,930 50,155
McFadden Adj.R20.036 0.031 0.084 0.036
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Note: All estimates from regressions computed above contain city and year dummies, as well as sociodemographic controls.
Table 8: Heterogeneity by gender in the perception of metro characteristics
Questions regarding safety have similar outcomes: the positive effects on the perception of9
total car capacity and the negative effect of acts of violence are larger. The coefficient related10
to the number of acts of violence is in line with findings presented in the economics of crime11
literature. Crouch (2009) argues that harassment in public transportation constrains women’s12
freedom of movement: our results suggest that this mechanism goes through the channel of13
perceptions, as more acts of violence decreases women’s feeling of safety. However, the number14
of staff in stations has a significant and positive effect on safety for women, which indicates15
different preferences and needs for this subgroup. Finally, the effect of metro ridership effect16
is still observed: more riders make women feel safer. As found previously, results from general17
customer satisfaction questions differ slightly from the safety ones. The significant elements18
are again the same as those observed for the whole population. In a second step, Tables 1419
and 15 in Appendix provide a detailed specification for acts of violence, respectively for all20
individuals and for women. It appears that the total number of robberies particularly affects21
women’s perception of safety negatively.22
Estimates in Tables 7 and 8 indicate that metro characteristics have a significant impact23
on individuals as well as a differentiated impact on women. Our results contribute to the24
19

growing strand of literature which shows that environmental factors highly influence women’s1
perception of safety (Whitzman 2012). Our findings also corroborate previous results from2
the literature indicating heterogeneous responses to urban characteristics by gender (Yavuz3
and Welch 2010, Loukaitou-Sideris and Fink 2009). The magnitude of the coefficients is quite4
small, which is in line with Koskela and Pain (2000), and indicates that the gap between5
women and men in perceived safety cannot exactly be bridged only by changing the metro6
characteristics as it would be likely to come at a high cost. Therefore, results indicate that7
“designing out” the fear of urban crime will depend essentially on what type of customer is8
targeted by a given policy. Yavuz and Welch (2010) and Loukaitou-Sideris and Fink (2009)9
make a case of the existence of gender discrepancies in responses to the evolution of public10
transport characteristics. Yet, our estimates provide a more precise understanding of these11
discrepancies as they indicate that the women’s responses are always of a larger magnitude12
compared to men’s. Therefore, this indicates that initiatives to make transport and urban13
spaces safer for women would also in turn make them safer for everyone (Viswanath and Basu14
2015).15
By capturing a gap in the perception of safety between men and women, our estimates pro-16
vide evidence that in an inner-city context, it is not just crime (Twinam 2017, Phillips and17
Sandler 2015) but the perception of crime that decreases with the improvement of public18
transportation. This suggests that there are ways in which operators can intervene to im-19
prove perceptions of safety with gender specific concerns in mind, but also help ruling out20
some tools highlighted by qualitative analyses (Vanier and d’Arbois de Jubainville 2017). We21
also note that inclusion of the additional variables does not bridge the gap in stated safety22
levels between men and women.23
Overall, women are not deeply dissatisfied with the levels of safety in public transport. The24
general distribution of safety responses in Fig. 8 and 9 indicate that women are on average25
satisfied with metro and bus services. We find that 45% of the female population feel safe26
in metro stations and trains, while 55% of women feel secure in buses. Therefore, the gap27
between men and women likely encompasses general differences in the overall perception of28
transport and the urban environment, aside from the intrinsic network characteristics.29
7 Conclusion30
This paper quantifies gender differences in the perception of safety and satisfaction in public31
transpor, using large-scale unique customer satisfaction surveys worldwide. Results indicate32
a significant gender gap in the perception of safety: women are 10% more likely than men to33
feel unsafe in metros and 6% more likely to feel unsafe in buses. This gender gap is larger for34
safety than for overall satisfaction (3% in metros and 2.5% in buses), which is consistent with35
safety being one important dimension of overall satisfaction. Results are stable and robust36
across various model specifications and we find that effects are heterogeneous with respect to37
age and country. Metro characteristics appear to have an influence on perceived safety: more38
acts of violence, less cars per trains and emptier vehicles decrease the feeling of safety among39
women; while the presence of more staff in metros does not significantly increase female safety.40
These results suggest grounds for intervention for service providers (e.g., more staff in stations41
or information campaigns to reduce the number of violent acts, see Vanier and d’Arbois de42
Jubainville (2017)).43
20

This study analysed unique datasets on customer satisfaction with buses and metros as well1
as on metro performance, which contributes to the existing literature by showing statistical2
evidence of links between characteristics of transport provision and the evaluation of safety by3
gender. Metro characteristics have a significant impact on individuals’ perceived safety and4
satisfaction, and the magnitude of those effects is always magnified for women. However, while5
metro characteristics are important, we also find that they cannot fully explain the gender6
gap that we observe in perceptions of safety which is likely due to other general differences in7
gender perceptions of the urban environment.8
21

References1
Atkins, S. (1990). Personal security as a transport issue: a state-of-the-art review. Transport2
Reviews 10 (2), 111–125.3
Carter, M. (2005). Gender differences in experience with and fear of crime in relation to public4
transport. Research on women’s issues in transportation 2, 100.5
Condon, S., M. Lieber, and F. Maillochon (2007). Feeling Unsafe in Public Places: Under-6
standing Women’s Fears.7
Crouch, M. (2009, 7). Sexual Harassment in Public Places. Social Philosophy Today 25,8
137–148.9
de O˜na, J. and R. de O˜na (2014). Quality of service in public transport based on customer10
satisfaction surveys: A review and assessment of methodological approaches. Transportation11
Science 49 (3), 605–622.12
Deaton, A. (1985). Panel data from time series of cross-sections. Journal of econometrics 30 (1-13
2), 109–126.14
Delbosc, A. and G. Currie (2012). Modelling the causes and impacts of personal safety15
perceptions on public transport ridership. Transport Policy 24, 302–309.16
EN13816 (2002). Transportation −Logistics and Services −Public Passenger Transport −17
Service Quality Definition, Targeting and Measurement (04), European Parliament and18
the Council, 2007. Regulation (EC) N. 1371/2007 of the European Parliament and of the19
Council of 23 October 2007 on rail passengers’ rights and obligations. Off. J. Eur. Union20
50 (L 315), 14–41.21
European Commission (2014). She moves - Women’s Issues in Transportation. Technical22
report.23
Fan, Y., A. Guthrie, and D. Levinson (2016). Waiting time perceptions at transit stops and24
stations: Effects of basic amenities, gender, and security. Transportation Research Part A:25
Policy and Practice 88, 251–264.26
Fellesson, M. and M. Friman (2012). Perceived satisfaction with public transport service in27
nine european cities. In Journal of the Transportation Research Forum, Volume 47.28
Fiorio, C. V., M. Florio, and G. Perucca (2011). Consumers’ satisfaction and regulation of29
local public transport: evidence from european cities.30
for Transport, U. D. (2004). People’s perceptions of personal security and their concerns about31
crime on public transport: research findings. Department for Transport.32
Gilchrist, E., J. Bannister, J. Ditton, and S. Farrall (1998). Women and the ‘fear of crime’33
challenging the accepted stereotype. The British Journal of Criminology 38 (2), 283–298.34
Hall, A. R. (2005). Generalized Method of Moments. Advanced Texts in Econometrics. Oxford:35
Oxford University Press.36
Hall, R. E. (1985). Ask Any Woman: A London Inquiry Into Rape and Sexual Assault. Falling37
Wall Press; Bristol.38
22

Hernandez, S. and A. Monzon (2016). Key factors for defining an efficient urban transport1
interchange: Users’ perceptions. Cities 50, 158–167.2
Jacobs, J. (1961). The death and life of great American cities (1 ed.). New York: Random3
Books.4
Johnson, M. D., A. Herrmann, and A. Gustafsson (2002). Comparing customer satisfaction5
across industries and countries. Journal of Economic Psychology 23 (6), 749–769.6
Jorgensen, L. J., G. D. Ellis, and E. Ruddell (2013). Fear perceptions in public parks: In-7
teractions of environmental concealment, the presence of people recreating, and gender.8
Environment and Behavior 45 (7), 803–820.9
Koskela, H. and R. Pain (2000). Revisiting fear and place: women’s fear of attack and the10
built environment. Geoforum 31 (2), 269–280.11
Loukaitou-Sideris, A. (1999). Hot spots of bus stop crime: The importance of environmental12
attributes. Journal of the American Planning association 65 (4), 395–411.13
Loukaitou-Sideris, A. and C. Fink (2009). Addressing women’s fear of victimization in trans-14
portation settings: A survey of us transit agencies. Urban Affairs Review 44 (4), 554–587.15
Lynch, G. and S. Atkins (1988). The influence of personal security fears on women’s travel16
patterns. Transportation 15 (3), 257–277.17
Madge, C. (1997). Public parks and the geography of fear. Tijdschrift voor economische en18
sociale geografie 88 (3), 237–250.19
Mannering, F. L., V. Shankar, and C. R. Bhat (2016). Unobserved heterogeneity and the20
statistical analysis of highway accident data. Analytic methods in accident research 11,21
1–16.22
Namgung, M. and G. Akar (2014). Role of gender and attitudes on public transportation use.23
Transportation research record 2415 (1), 136–144.24
O’Brien, E. (2005). Publics* and woodlands in england: well-being, local identity, social25
learning, conflict and management. Forestry 78 (4), 321–336.26
of Transport, D. (1997). Crime concern and transport and travel research: Perceptions of27
safety from crime on public transpor. London:Author.28
Oliver, R. L. (1997). Satisfaction: A behavioral perspective on the consumer. Irwin/McGraw-29
Hill, New York.30
Pain, R. (2001). Gender, race, age and fear in the city. Urban studies 38 (5-6), 899–913.31
Pain, R. H. (1997). Social geographies of women’s fear of crime. Transactions of the Institute32
of British geographers, 231–244.33
Phillips, D. C. and D. Sandler (2015, 5). Does public transit spread crime? Evidence from34
temporary rail station closures. Regional Science and Urban Economics 52, 13–26.35
Riger, S. and M. Gordon (1989). The female fear.36
Rohe, W. M. and R. J. Burby (1988). Fear of crime in public housing. Environment and37
Behavior 20 (6), 700–720.38
23

Ross, A. (2010, 1). Violent crime, entrepreneurship, and cities. Journal of Urban Eco-1
nomics 67 (1), 135–149.2
Schulz, D. and S. Gilbert (1996). Women and transit security: A new look at an old issue. In3
Women’s Travel Issues Second National Conference, pp. 550–562. Citeseer.4
Shaw, M. (2002). Gender and crime prevention. Technical report, International Observer.5
St-Louis, E., K. Manaugh, D. van Lierop, and A. El-Geneidy (2014). The happy commuter: a6
comparison of commuter satisfaction across modes. Transportation research part F: traffic7
psychology and behaviour 26, 160–170.8
Stanko, E. (1990). Everyday violence: Women’s and men’s experience of personal danger.9
London: Pandora.10
Steinmetz, N. M. and D. M. Austin (2014). Fear of criminal victimization on a college campus:11
A visual and survey analysis of location and demographic factors. American Journal of12
Criminal Justice 39 (3), 511–537.13
Trompet, M., R. J. Anderson, and D. J. Graham (2018). Improved understanding of the14
relative quality of bus public transit using a balanced approach to performance data nor-15
malization. Transportation Research Part A: Policy and Practice 114, 13–23.16
Trompet, M., R. Parasram, and R. J. Anderson (2013). Benchmarking disaggregate customer17
satisfaction scores of bus operators in different cities and countries. Transportation Research18
Record 2351 (1), 14–22.19
Twinam, T. (2017, 7). Danger zone: Land use and the geography of neighborhood crime.20
Journal of Urban Economics 100, 104–119.21
Tyrinopoulos, Y. and C. Antoniou (2008). Public transit user satisfaction: Variability and22
policy implications. Transport Policy 15 (4), 260–272.23
Ulfarsson, G. F. and F. L. Mannering (2004). Differences in male and female injury severities24
in sport-utility vehicle, minivan, pickup and passenger car accidents. Accident Analysis &25
Prevention 36 (2), 135–147.26
UN Women (2016). Gender Statistics - Violence against Women.27
Valentine, G. (1990). Women’s fear and the design of public space. Built Environment (1978-),28
288–303.29
Vanier, C. and H. d’Arbois de Jubainville (2017, 9). Feeling unsafe in public transportation:30
A profile analysis of female users in the Parisian region. Crime Prevention and Community31
Safety 19 (3-4), 251–263.32
Viswanath, K. and A. Basu (2015, 1). SafetiPin: an innovative mobile app to collect data on33
women’s safety in Indian cities. Gender & Development 23 (1), 45–60.34
Washington, S. P., M. G. Karlaftis, and F. Mannering (2010). Statistical and econometric35
methods for transportation data analysis. Chapman and Hall/CRC.36
Wekerle, G. R. and C. Whitzman (1995). Safe cities: Guidelines for planning, design, and37
management. Van Nostrand Reinhold New York.38
24

Whitzman, C. (2012, 5). The Handbook of Community Safety Gender and Violence Prevention.1
Routledge.2
Yavuz, N. and E. W. Welch (2010). Addressing fear of crime in public space: Gender differences3
in reaction to safety measures in train transit. Urban studies 47 (12), 2491–2515.4
Zelinka, A. and D. Brennan (2001). SafeScape. Creating Safer, More Livable Communities5
through Planning and Design.6
25

A Appendix A: CoMET, Nova and IBBG groups1
The CoMET group is a consortium of some of the world’s largest urban metros including:2
Berliner Verkehrsbetriebe (BVG, Berlin), Delhi Metro Rail Corporation (DMRC), Mass Tran-3
sit Railway (MTR, Hong Kong), the Underground (London), Sistema de Transporte Colectivo4
(STC, Mexico City), Metro de Madrid, Moscow Metro, New York City Transit (NYCT, New5
York), the R´egie Autonome des Transports Parisiens (RATP, Paris) that include both the6
Metro and the R´eseau Express R´egional (RER), Metro de Santiago and the Singapore Mass7
Rapid Transit (SMRT, Singapore). The Nova group is a consortium of small to medium sized8
metros including: Buenos Aires Metrovias, Transports Metropolitans de Barcelona (TMB,9
Barcelona), Soci´et´e des Transports Intercommunaux de Bruxelles (STIB, Brussels), Bangkok10
Expressway and Metro Public Company (BEM, Bangkok), Docklands Light Railway (DLR,11
London), Istanbul Ulasim, RapidKL / Prasarana (Kuala Lumpur), Metropolitano de Lisboa,12
Soci´et´e de Transport de Montr´eal (STM, Montr´eal), Newcastle Nexus, Metro Rio (Rio de13
Janeiro), Toronto Transit Commission (TTC, Toronto) and Vancouver SkyTrain (Vancouver,14
Canada).15
The IBBG group is a large consortium of major bus networks including: Transport Metropoli-16
tans de Barcelona (TMB, Barcelona), Soci´et´e des Transports Intercommunaux de Bruxelles17
(STIB, Brussels), Dublin Bus (Dublin), IETT Isletmeleri Genel M¨ud¨url¨ug¨u (IETT, Istan-18
bul), Rapid Bus Sdn Bhd (Rapid KL, Kuala Lumpur), Companhia Carris de Ferro de Lisboa19
(Carris, Lisbon), London Buses (LBSL, London), Soci´et´e de Transport de Montr´eal (STM,20
Montr´eal), MTA New York City Transit and MTA Bus (New York), the R´egie Autonome21
des Transports Parisiens (RATP, Paris), King County Metro (KCM, Seattle), SMRT Buses22
(Singapore), State Transit (Sydney), and Coast Mountain Bus Company (CMBC, Vancouver).23
26

B Appendix B: Descriptive Statistics1
Transport Mode: Metro
Years
2014 2015 2016 2017 2018 2014/18
Women 0.46 0.43 0.47 0.49 0.46 0.46
(0.50) (0.50) (0.50) (0.50) (0.50) (0.50)
Age
Less than 18 0.04 0.04 0.04 0.04 0.04 0.04
(0.19) (0.19) (0.20) (0.19) (0.20) (0.19)
18-29 0.42 0.39 0.40 0.41 0.37 0.40
(0.49) (0.49) (0.49) (0.49) (0.48) (0.49)
30-39 0.24 0.26 0.24 0.23 0.23 0.24
(0.43) (0.44) (0.43) (0.42) (0.42) (0.43)
40-49 0.14 0.15 0.15 0.15 0.16 0.15
(0.35) (0.36) (0.35) (0.36) (0.36) (0.36)
50-65 0.13 0.14 0.14 0.15 0.15 0.14
(0.34) (0.34) (0.35) (0.35) (0.36) (0.35)
More than 65 0.03 0.03 0.03 0.03 0.04 0.03
(0.17) (0.16) (0.17) (0.17) (0.20) (0.18)
Frequency of use
Very Often 0.62 0.64 0.63 0.64 0.61 0.63
(0.49) (0.48) (0.48) (0.48) (0.49) (0.48)
Often (min:3times/week) 0.20 0.19 0.19 0.19 0.20 0.19
(0.40) (0.39) (0.39) (0.39) (0.40) (0.39)
Sometimes (min:once/week) 0.10 0.09 0.10 0.09 0.11 0.10
(0.30) (0.29) (0.30) (0.29) (0.31) (0.30)
Rarely (min:once/month) 0.05 0.05 0.05 0.05 0.05 0.05
(0.22) (0.22) (0.22) (0.22) (0.22) (0.22)
Very rarely (less than once/month) 0.03 0.03 0.03 0.03 0.03 0.03
(0.16) (0.16) (0.17) (0.17) (0.17) (0.17)
Most frequent travel motive
Work 0.59 0.61 0.60 0.57 0.56 0.58
(0.49) (0.49) (0.49) (0.50) (0.50) (0.49)
Education 0.07 0.06 0.07 0.08 0.06 0.07
(0.25) (0.24) (0.25) (0.27) (0.24) (0.25)
Shopping 0.14 0.13 0.13 0.17 0.14 0.14
(0.34) (0.34) (0.33) (0.37) (0.35) (0.35)
Leisure 0.14 0.13 0.14 0.12 0.14 0.13
(0.35) (0.33) (0.34) (0.32) (0.35) (0.34)
Doctor 0.05 0.06 0.05 0.05 0.07 0.06
(0.22) (0.24) (0.22) (0.22) (0.25) (0.23)
Other 0.01 0.01 0.02 0.02 0.02 0.02
(0.12) (0.10) (0.13) (0.12) (0.14) (0.12)
Standard errors in parentheses.
Source: Comet/Nova data.
Table 9: Descriptive Statistics - Metro Data
27

Transport Mode: Bus
Years
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2009/18
Women 0.50 0.47 0.54 0.54 0.52 0.53 0.53 0.53 0.51 0.51 0.52
(0.50) (0.50) (0.50) (0.50) (0.50) (0.50) (0.50) (0.50) (0.50) (0.50) (0.50)
Age
Less than 18 0.04 0.06 0.05 0.04 0.03 0.04 0.03 0.05 0.05 0.06 0.05
(0.19) (0.24) (0.22) (0.20) (0.18) (0.19) (0.18) (0.22) (0.22) (0.24) (0.21)
18-29 0.37 0.38 0.36 0.36 0.32 0.32 0.30 0.35 0.34 0.31 0.34
(0.48) (0.49) (0.48) (0.48) (0.47) (0.47) (0.46) (0.48) (0.47) (0.46) (0.47)
30-39 0.23 0.25 0.21 0.23 0.22 0.22 0.23 0.23 0.24 0.21 0.22
(0.42) (0.43) (0.41) (0.42) (0.41) (0.42) (0.42) (0.42) (0.42) (0.41) (0.42)
40-49 0.17 0.17 0.16 0.15 0.16 0.15 0.17 0.15 0.16 0.17 0.16
(0.37) (0.37) (0.36) (0.36) (0.37) (0.36) (0.37) (0.36) (0.36) (0.37) (0.37)
50-65 0.19 0.14 0.18 0.18 0.21 0.20 0.21 0.17 0.17 0.19 0.19
(0.39) (0.35) (0.38) (0.38) (0.41) (0.40) (0.41) (0.37) (0.37) (0.39) (0.39)
More than 65 0.00 0.00 0.04 0.04 0.06 0.06 0.05 0.05 0.05 0.06 0.05
(0.02) (0.00) (0.20) (0.19) (0.23) (0.24) (0.22) (0.22) (0.21) (0.24) (0.22)
Frequency of use
Very Often 0.55 0.52 0.59 0.57 0.52 0.55 0.59 0.60 0.58 0.56 0.57
(0.50) (0.50) (0.49) (0.50) (0.50) (0.50) (0.49) (0.49) (0.49) (0.50) (0.50)
Often (min: 3 times/week) 0.20 0.19 0.20 0.22 0.23 0.24 0.23 0.24 0.24 0.25 0.23
(0.40) (0.40) (0.40) (0.41) (0.42) (0.43) (0.42) (0.43) (0.43) (0.43) (0.42)
Sometimes (min: once/week) 0.13 0.12 0.13 0.13 0.15 0.13 0.11 0.11 0.13 0.13 0.13
(0.34) (0.33) (0.33) (0.33) (0.35) (0.34) (0.32) (0.31) (0.33) (0.34) (0.33)
Rarely (min: once/month) 0.07 0.07 0.05 0.06 0.07 0.05 0.04 0.04 0.04 0.04 0.05
(0.25) (0.26) (0.21) (0.23) (0.25) (0.22) (0.20) (0.19) (0.19) (0.20) (0.21)
Very rarely (less than once/month) 0.06 0.09 0.03 0.03 0.03 0.03 0.02 0.02 0.02 0.02 0.03
(0.23) (0.28) (0.17) (0.17) (0.18) (0.17) (0.14) (0.14) (0.13) (0.14) (0.16)
Most frequent travel motive
Work/Education 0.74 0.69 0.75 0.73 0.70 0.72 0.79 0.77 0.75 0.73 0.74
(0.44) (0.46) (0.43) (0.44) (0.46) (0.45) (0.41) (0.42) (0.43) (0.44) (0.44)
Shopping 0.04 0.03 0.05 0.05 0.05 0.05 0.03 0.04 0.04 0.04 0.04
(0.20) (0.17) (0.21) (0.21) (0.21) (0.21) (0.18) (0.20) (0.21) (0.20) (0.20)
Leisure 0.16 0.16 0.15 0.18 0.20 0.17 0.13 0.14 0.15 0.17 0.16
(0.37) (0.37) (0.36) (0.38) (0.40) (0.38) (0.34) (0.34) (0.36) (0.38) (0.37)
Doctor 0.02 0.03 0.02 0.02 0.02 0.03 0.02 0.03 0.03 0.03 0.03
(0.15) (0.16) (0.14) (0.13) (0.15) (0.17) (0.15) (0.16) (0.17) (0.17) (0.16)
Other 0.03 0.10 0.03 0.03 0.02 0.03 0.03 0.02 0.03 0.02 0.03
(0.17) (0.30) (0.17) (0.16) (0.15) (0.18) (0.16) (0.15) (0.17) (0.16) (0.17)
Standard errors in parentheses.
Source: IBBG data.
Table 10: Descriptive Statistics - Bus Data
28

C Appendix C: Distribution of Probabilities1
Stations are secure Trains are secure Overall Likelihood to recommend
for me for me Satisfaction to another person
(1) (2) (4) (3
Share of female respondents -0.409* -0.661*** -0.134 -0.179
(0.218) (0.239) (0.213) (0.185)
Cutpoint 1 -2.417*** -2.553*** -2.001*** -2.104***
(0.254) (0.275) (0.358) (0.352)
Cutpoint 2 -1.624*** -1.810*** -0.731** -1.941***
(0.254) (0.241) (0.361) (0.334)
Cutpoint 3 0.703*** 0.540** 1.100*** -1.797***
(0.264) (0.240) (0.293) (0.298)
Cutpoint 4 2.904*** 2.752*** 3.437*** -1.401***
(0.329) (0.275) (0.287) (0.260)
Cutpoint 5 -0.919***
(0.257)
Cutpoint 6 -0.222
(0.272)
Cutpoint 7 0.574**
(0.284)
Cutpoint 8 1.622***
(0.308)
Cutpoint 9 2.661***
(0.283)
Cutpoint 10 3.464***
(0.280)
Var (city level) 0.423*** 0.388*** 0.467*** 0.391**
(0.105) (0.0848) (0.0933) (0.157)
Var (female) at city level 0.423*** 0.388*** 0.467*** 0.453**
(0.105) (0.0848) (0.0933) (0.194)
Observations 2,899 2,899 2,898 2,898
Number of groups 25 25 25 25
Prob chi2(2) 0 0 0 0
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Table 11: Gender Coefficients on the Perception of Safety - Multilevel mixed-effects ordered
probit estimates (Random Intercept and Random Coefficient)
29

Dimension Availability/Convenience Ease Information Reliability Comfort Customer Care
Variable Freq. Hours Network Interchange Disruption During Journey Plan Journey Station Train Pay Ticket Answer Questions Staff Train
(A1) (A2) (A3) (B1) (C1) (C2) (C3) (D1) (E1) (E2) (F1) (F2) (F3)
Female (Yes=1) -0.0428*** -0.00270 0.0326*** -0.0560*** -0.0243*** -0.0148*** -0.0444*** -0.0986*** -0.0898*** -0.0943*** -0.0421*** -0.0562*** -0.0242***
(0.00545) (0.00547) (0.00543) (0.00541) (0.00548) (0.00539) (0.00556) (0.00546) (0.00541) (0.00539) (0.00549) (0.00570) (0.00550)
Age (ref: 18-29)
Less than 18 0.200*** 0.495*** 0.238*** 0.232*** 0.330*** 0.315*** 0.254*** 0.249*** 0.213*** 0.230*** 0.211*** 0.292*** 0.234***
(0.0146) (0.0150) (0.0146) (0.0146) (0.0146) (0.0146) (0.0153) (0.0145) (0.0145) (0.0144) (0.0149) (0.0153) (0.0149)
30-39 -0.0457*** 0.0453*** -0.0638*** -0.113*** -0.0890*** -0.0719*** -0.149*** 0.00409 0.00444 0.00399 -0.0697*** -0.0953*** -0.0649***
(0.00711) (0.00713) (0.00708) (0.00706) (0.00718) (0.00704) (0.00726) (0.00713) (0.00708) (0.00705) (0.00716) (0.00744) (0.00718)
40-49 0.0273*** 0.153*** 0.0287*** -0.0798*** -0.0326*** -0.0248*** -0.158*** 0.110*** 0.0730*** 0.0742*** -0.0639*** -0.0775*** -0.0333***
(0.00836) (0.00841) (0.00834) (0.00830) (0.00840) (0.00826) (0.00852) (0.00837) (0.00831) (0.00828) (0.00842) (0.00869) (0.00841)
50-65 0.113*** 0.230*** 0.119*** -0.0246*** 0.0132 0.0167* -0.140*** 0.205*** 0.160*** 0.118*** 0.0304*** -0.0186** 0.0328***
(0.00868) (0.00876) (0.00868) (0.00863) (0.00868) (0.00858) (0.00886) (0.00868) (0.00862) (0.00859) (0.00878) (0.00903) (0.00873)
Over 65 0.284*** 0.452*** 0.267*** 0.0735*** 0.0928*** 0.107*** -0.0690*** 0.356*** 0.355*** 0.289*** 0.133*** 0.121*** 0.198***
(0.0161) (0.0164) (0.0161) (0.0159) (0.0159) (0.0158) (0.0164) (0.0160) (0.0158) (0.0157) (0.0164) (0.0166) (0.0161)
Main travel purpose (ref: Work)
Education 0.0690*** -0.0677*** 0.00820 -0.00925 0.101*** 0.0762*** -0.000490 0.114*** 0.0120 0.0599*** -0.0556*** 0.0641*** 0.0398***
(0.0117) (0.0118) (0.0117) (0.0116) (0.0117) (0.0116) (0.0119) (0.0117) (0.0116) (0.0116) (0.0118) (0.0121) (0.0118)
Shopping 0.0959*** 0.0406*** 0.0676*** 0.114*** 0.127*** 0.0921*** 0.142*** 0.128*** 0.0819*** 0.0825*** 0.0663*** 0.131*** 0.0915***
(0.00864) (0.00866) (0.00860) (0.00860) (0.00871) (0.00856) (0.00889) (0.00866) (0.00857) (0.00854) (0.00873) (0.00907) (0.00878)
Leisure 0.178*** 0.00825 0.150*** 0.143*** 0.220*** 0.175*** 0.187*** 0.236*** 0.107*** 0.136*** 0.101*** 0.164*** 0.135***
(0.0101) (0.0101) (0.0100) (0.0100) (0.0100) (0.00997) (0.0103) (0.0101) (0.00997) (0.00993) (0.0102) (0.0106) (0.0102)
Doctor 0.127*** 0.00680 0.0927*** 0.0552*** 0.195*** 0.147*** 0.114*** 0.168*** 0.0628*** 0.0989*** 0.0315** 0.113*** 0.0846***
(0.0130) (0.0131) (0.0130) (0.0129) (0.0130) (0.0129) (0.0133) (0.0130) (0.0129) (0.0129) (0.0131) (0.0134) (0.0131)
Other 0.0327 -0.0996*** 0.0551*** -0.0196 0.131*** 0.103*** 0.0473** 0.102*** 0.0234 0.0522** -0.0104 0.0502** 0.0475**
(0.0213) (0.0215) (0.0213) (0.0211) (0.0214) (0.0212) (0.0219) (0.0213) (0.0212) (0.0211) (0.0216) (0.0223) (0.0215)
Frequency use trains (ref: very often)
Often 0.148*** 0.0658*** -0.0512*** 0.0104 0.110*** 0.0860*** 0.0251*** 0.167*** 0.0822*** 0.124*** 0.00285 0.0767*** 0.0682***
(0.00741) (0.00746) (0.00740) (0.00738) (0.00743) (0.00734) (0.00759) (0.00742) (0.00736) (0.00733) (0.00750) (0.00775) (0.00749)
Sometimes 0.205*** 0.103*** -0.102*** 0.00983 0.140*** 0.0980*** 0.00598 0.213*** 0.112*** 0.168*** -0.0269** 0.0728*** 0.0530***
(0.0105) (0.0106) (0.0105) (0.0105) (0.0105) (0.0104) (0.0108) (0.0105) (0.0104) (0.0104) (0.0106) (0.0110) (0.0106)
Rarely 0.273*** 0.184*** -0.120*** 0.0220 0.206*** 0.153*** 0.0249* 0.269*** 0.162*** 0.216*** -0.0189 0.138*** 0.115***
(0.0136) (0.0137) (0.0135) (0.0135) (0.0136) (0.0135) (0.0139) (0.0136) (0.0135) (0.0134) (0.0137) (0.0144) (0.0138)
Very Rarely 0.319*** 0.176*** -0.143*** 0.0305* 0.296*** 0.218*** -0.0989*** 0.255*** 0.180*** 0.234*** -0.0780*** 0.184*** 0.125***
(0.0174) (0.0175) (0.0172) (0.0172) (0.0176) (0.0174) (0.0176) (0.0174) (0.0172) (0.0171) (0.0174) (0.0184) (0.0177)
P-value of coeft. female <0.001 0.622 <0.001 <0.001 <0.001 0.006 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001
Observations 169,813 169,697 169,667 169,385 166,097 167,753 169,431 169,728 169,931 169,951 168,984 150,099 161,733
McFadden Adj.R20.087 0.049 0.029 0.023 0.055 0.040 0.030 0.090 0.094 0.074 0.027 0.044 0.033
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Note: All estimates from regressions computed above contain city and year dummies.
Table 12: Ordered Probit Estimations: Other Dimensions for Metros
Dimension Availability Ease Information Reliability Comfort Customer care
Variable Network Interchange Move in bus Time Services Goo d ride Spacious Goo d Light Well Dressed Helpful Pay Ticket Sort Pbs
(A1) (B1) (B2) (C1) (C2) (D1) (E1) (E2) (E3) (E4) (F1) (F2) (F3)
Female (Yes=1) -0.0429*** -0.159*** -0.124*** -0.0194*** -0.0499*** -0.0554*** -0.103*** 0.0128** -0.132*** 0.0879*** -0.0254*** -0.0380*** -0.0335***
(0.00515) (0.00526) (0.00509) (0.00509) (0.00513) (0.00515) (0.00508) (0.00511) (0.00506) (0.00533) (0.00513) (0.00523) (0.00538)
Age (ref: 18-29)
Less than 18 0.287*** 0.0936*** 0.119*** 0.315*** 0.227*** 0.300*** 0.360*** 0.173*** 0.198*** 0.331*** 0.217*** 0.190*** 0.363***
(0.0127) (0.0129) (0.0124) (0.0125) (0.0127) (0.0126) (0.0125) (0.0125) (0.0124) (0.0135) (0.0127) (0.0129) (0.0131)
30-39 -0.0646*** -0.0923*** -0.0732*** -0.00869 -0.0355*** 0.0109 -0.0812*** -0.0413*** -0.0231*** -0.198*** -0.0611*** 0.00265 -0.101***
(0.00691) (0.00705) (0.00685) (0.00685) (0.00688) (0.00690) (0.00682) (0.00688) (0.00681) (0.00719) (0.00690) (0.00700) (0.00723)
40-49 -0.0119 -0.0613*** -0.0434*** 0.0865*** 0.0153** 0.145*** -0.0528*** -0.0629*** -0.0126 -0.289*** -0.0416*** 0.106*** -0.109***
(0.00784) (0.00801) (0.00773) (0.00775) (0.00781) (0.00784) (0.00772) (0.00776) (0.00770) (0.00812) (0.00781) (0.00794) (0.00818)
50-65 0.0485*** -0.0919*** -0.0873*** 0.126*** 0.0114 0.251*** -0.0568*** -0.106*** -0.0316*** -0.388*** -0.0345*** 0.206*** -0.121***
(0.00771) (0.00788) (0.00759) (0.00761) (0.00766) (0.00773) (0.00758) (0.00761) (0.00755) (0.00798) (0.00767) (0.00784) (0.00804)
Over 65 0.168*** -0.0959*** -0.0621*** 0.175*** 0.0479*** 0.377*** 0.0585*** 0.0663*** 0.152*** -0.378*** 0.150*** 0.379*** -0.00356
(0.0133) (0.0135) (0.0130) (0.0131) (0.0132) (0.0133) (0.0130) (0.0131) (0.0129) (0.0136) (0.0132) (0.0136) (0.0139)
Main travel purpose (ref: work/education)
Shopping 0.149*** -0.0369*** 0.0413*** 0.166*** 0.0727*** 0.158*** 0.0612*** 0.125*** 0.131*** 0.0584*** 0.0684*** 0.0329** 0.137***
(0.0134) (0.0136) (0.0131) (0.0132) (0.0133) (0.0134) (0.0132) (0.0132) (0.0131) (0.0138) (0.0133) (0.0136) (0.0140)
Leisure 0.135*** 0.103*** 0.118*** 0.137*** 0.111*** 0.169*** 0.0985*** 0.142*** 0.127*** 0.0712*** 0.104*** 0.0818*** 0.118***
(0.00833) (0.00856) (0.00820) (0.00822) (0.00829) (0.00834) (0.00820) (0.00822) (0.00815) (0.00861) (0.00830) (0.00844) (0.00881)
Doctor 0.0331* -0.248*** -0.130*** 0.0566*** -0.0345** 0.0464*** -0.0365** -0.0231 0.0537*** 0.0108 -0.0439*** -0.0750*** -0.0152
(0.0170) (0.0170) (0.0167) (0.0168) (0.0169) (0.0170) (0.0167) (0.0168) (0.0167) (0.0176) (0.0169) (0.0172) (0.0176)
Other -0.0113 -0.112*** -0.0840*** 0.00534 -0.0506*** 0.0300* -0.0384** -0.00137 0.0389** 0.0195 0.00727 -0.0553*** 0.00684
(0.0157) (0.0160) (0.0155) (0.0156) (0.0156) (0.0158) (0.0155) (0.0155) (0.0154) (0.0163) (0.0157) (0.0159) (0.0165)
Frequency of use (ref: very often)
Often -0.0412*** 0.119*** 0.102*** 0.0586*** 0.0313*** 0.103*** 0.0980*** 0.119*** 0.127*** 0.0276*** 0.0744*** -0.0615*** 0.113***
(0.00653) (0.00667) (0.00644) (0.00645) (0.00650) (0.00653) (0.00643) (0.00647) (0.00641) (0.00676) (0.00651) (0.00663) (0.00682)
Sometimes -0.132*** 0.184*** 0.190*** 0.0635*** 0.0310*** 0.132*** 0.179*** 0.179*** 0.200*** 0.0412*** 0.112*** -0.106*** 0.153***
(0.00878) (0.00903) (0.00867) (0.00869) (0.00875) (0.00881) (0.00867) (0.00870) (0.00862) (0.00911) (0.00878) (0.00891) (0.00933)
Rarely -0.308*** 0.186*** 0.223*** 0.0330** -0.00907 0.102*** 0.232*** 0.184*** 0.215*** 0.0177 0.0883*** -0.179*** 0.150***
(0.0129) (0.0134) (0.0128) (0.0128) (0.0129) (0.0130) (0.0128) (0.0128) (0.0127) (0.0134) (0.0130) (0.0131) (0.0141)
Very rarely -0.581*** 0.0363** 0.208*** 0.0150 -0.143*** 0.00812 0.207*** 0.154*** 0.214*** -0.0506*** -0.00438 -0.306*** 0.0735***
(0.0169) (0.0173) (0.0167) (0.0169) (0.0168) (0.0171) (0.0167) (0.0167) (0.0166) (0.0174) (0.0169) (0.0170) (0.0183)
P-value of coeft. female <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.012 <0.001 <0.001 <0.001 <0.001 <0.001
Observations 180,003 180,339 180,302 177,514 179,387 180,073 180,236 180,263 180,208 176,802 178,089 177,149 159,596
McFadden Adj.R20.014 0.043 0.051 0.019 0.021 0.029 0.020 0.043 0.013 0.054 0.022 0.026 0.025
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Note: All estimates from regressions computed above contain city and year dummies.
Table 13: Ordered Probit Estimations: Other Dimensions for Buses
30

Figure 6: Regional Marginal Effects for the Overall Satisfaction with the Metro System
Figure 7: Regional Marginal Effects for the Overall Satisfaction with the Bus System
31

Figure 8: General distribution of safety responses for women - Metro Data
Figure 9: General distribution of safety responses for women - Bus Data
32

D Appendix D: Influence of Metro Characteristics1
Stations are secure Trains are secure Overall Likelihood to recommend
for me for me satisfaction to another person
(1) (2) (3) (4)
Female (Yes=1) -0.241*** -0.239*** -0.0431*** -0.0162**
(0.00673) (0.00673) (0.00684) (0.00658)
Total staff in stations -0.0256 0.0188 -0.185*** -0.143***
(0.0221) (0.0221) (0.0226) (0.0221)
Metro ridership 0.00157*** 0.00218*** 0.00417*** 0.00348***
(0.000214) (0.000214) (0.000218) (0.000209)
Nb of cars per train 0.00742*** 0.0142*** 0.00692*** 0.000283
(0.00258) (0.00258) (0.00264) (0.00252)
Avg number stations by line -0.0443*** -0.0654*** -0.0846*** -0.0708***
(0.00551) (0.00550) (0.00558) (0.00536)
Total car capacity 0.00449*** 0.00484*** 0.0128*** 0.0125***
(0.000653) (0.000652) (0.000665) (0.000637)
Nb of acts of violence -0.000230*** -0.000181*** -0.000285*** -0.000243***
(3.12e-05) (3.12e-05) (3.19e-05) (3.10e-05)
Nb robberies -0.000111* -8.77e-05 0.000441*** 0.000244***
(6.42e-05) (6.42e-05) (6.53e-05) (6.20e-05)
P-value of coeft. female <0.001 <0.001 <0.001 0.014
Observations 109,630 109,548 109,743 108,057
McFadden Adj.R20.037 0.033 0.077 0.036
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Note: All estimates from regressions computed above contain city and year dummies, as well as sociodemographic controls.
Table 14: Heterogeneity by gender in the perception of metro characteristics - Detailed Spec-
ification for Acts of violence
Stations are secure Trains are secure Overall Likelihood to recommend
for me for me satisfaction to another person
(1) (2) (3) (4)
Female (Yes=1) - - - -
Total staff in stations 0.0158 0.0641* -0.169*** -0.0873***
(0.0337) (0.0337) (0.0346) (0.0334)
Metro ridership 0.00167*** 0.00250*** 0.00705*** 0.00487***
(0.000417) (0.000417) (0.000427) (0.000409)
Nb of cars per train 0.0142*** 0.0228*** 0.0180*** 0.0132***
(0.00384) (0.00383) (0.00396) (0.00375)
Avg number stations by line -0.0716*** -0.103*** -0.133*** -0.119***
(0.00970) (0.00970) (0.00986) (0.00945)
Total car capacity 0.00658*** 0.00812*** 0.0193*** 0.0178***
(0.00101) (0.00101) (0.00103) (0.000983)
Nb of acts of violence -0.000293*** -0.000253*** -0.000535*** -0.000384***
(5.25e-05) (5.26e-05) (5.40e-05) (5.22e-05)
Nb robberies -0.000291*** -0.000284*** 0.000442*** 0.000241**
(9.82e-05) (9.82e-05) (0.000100) (9.50e-05)
Observations 50,866 50,816 50,930 50,155
McFadden Adj.R20.036 0.031 0.085 0.036
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Note: All estimates from regressions computed above contain city and year dummies, as well as sociodemographic controls.
Table 15: Heterogeneity by gender in the perception of metro characteristics - Detailed Spec-
ification for Acts of violence
33

Stations are secure Trains are secure Overall Likelihood to recommend
for me for me Satisfaction to another person
(1) (2) (3) (4)
Share of female respondents -0.226*** -0.430*** -0.00905 -0.128***
(0.0566) (0.0685) (0.0456) (0.0405)
Avg. nb. staff in stations -0.118*** -0.231*** -0.414*** -0.190***
(0.0237) (0.0260) (0.0320) (0.0292)
Avg. metro ridership 0.0106*** 0.0264*** 0.0600*** 0.0480***
(0.00220) (0.00261) (0.00350) (0.00304)
Avg. stations by line -0.0841*** -0.0889*** -0.216*** -0.212***
(0.00764) (0.00740) (0.00799) (0.00722)
Avg. total capacity 0.00533*** 0.00464*** 0.0114*** 0.0122***
(0.000573) (0.000635) (0.00120) (0.00113)
Avg. number violent acts -0.000124*** -9.20e-05* 0.000711*** 0.000522***
(4.42e-05) (4.71e-05) (5.81e-05) (5.34e-05)
Avg. metro ridership -0.00221*** -0.00398*** -0.0196*** -0.0224***
(0.000766) (0.000792) (0.00141) (0.00132)
Cutpoint 1 -3.289*** -3.553*** -4.880*** -4.944***
(0.178) (0.183) (0.250) (0.228)
Cutpoint 2 -2.692*** -2.927*** -3.640*** -4.808***
(0.177) (0.183) (0.249) (0.227)
Cutpoint 3 -0.364** -0.520*** -1.655*** -4.660***
(0.176) (0.182) (0.248) (0.227)
Cutpoint 4 1.824*** 1.808*** 0.630** -4.372***
(0.176) (0.182) (0.250) (0.226)
Cutpoint 5 -3.984***
(0.226)
Cutpoint 6 -3.214***
(0.226)
Cutpoint 7 -2.366***
(0.226)
Cutpoint 8 -1.252***
(0.226)
Cutpoint 9 -0.240
(0.226)
Cutpoint 10 0.443*
(0.226)
Var (city level) 0.544*** 0.989*** 0.203*** 0.164***
(0.0699) (0.110) (0.0405) (0.0377)
Var (female) at city level 0.298*** 0.379*** 1.686*** 1.687***
(0.0375) (0.0464) (0.202) (0.201)
Observations 1,748 1,748 1,747 1,747
Number of groups 20 20 20 20
Prob chi2(2) 0 0 0 0
Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Source: Customer Satisfaction Surveys, Comet/Nova, own calculations.
Table 16: Effect of Metro Characteristics on the Perception of Safety by Gender - Multilevel
Mixed-effects Ordered Probit Estimates (Random Intercept and Random Coefficient)
34