Transport Policy 69 (2018) 10–19

Contents lists available at ScienceDirect

Transport Policy
journal homepage: www.elsevier.com/locate/tranpol

Analyzing willingness to improve the resilience of New York City's T

transportation system
Chen Wang, Jiayi Sun, Roddy Russell Jr., Ricardo A. Daziano∗
School of Civil and Environmental Engineering, Cornell University, USA

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

Keywords: Hurricane Sandy revealed the higher-risk vulnerability to natural hazards of civil infrastructure systems in
Willingness to pay coastal megacities such as New York. Traditional sources of funding for both recovering from disasters and
Contingent valuation preventing future damages are not only limited, but also do not account for benefit transfers of the externalities
Discrete choice experiment induced by the provision of resilient infrastructure. In principle, property owners should be willing to pay (WTP)
Hurricane Sandy
an amount equal to the perceived benefit, if this positive externality is internalized by them following some
pricing mechanism. In this paper, we analyze the willingness of residents to financially support improvements in
the resilience to extreme events of the transportation system in New York City. Choice microdata was collected
for over 1500 residents of the metropolitan NYC area. Several logit-type models were estimated and the pre-
ferred model was a discrete-continuous heterogeneity mixture that allows for the derivation of flexible dis-
tributions of willingness to pay. Using hypothetical scenarios of recovery, the annual willingness to pay for
individuals who missed work ands self identify as politically liberal ranges from $120 to $775, whereas for those
individuals that were not directly affected the range is $15-$50.

1. Introduction: extreme weather and disruptions to
transportation
New York City has one of the largest and busiest transportation
systems, both in the US and in the world. Any disruption to this system
has far-reaching and serious ramifications (NYS 2100 Commission,
2013). One of the largest disruptions in the system's history came in
October of 2012 in the form of Hurricane Sandy. According to the
National Centers for Environmental Information, Sandy was the second
costliest and second deadliest tropical cyclone between 1980 and 2014
(FEMA, 2013), with Hurricane Katrina being first in both categories
(NCEI, 2016). Although it is difficult to predict how climate change will
affect future storms, the change in sea levels can be predicted. One
estimate puts the mean annual sea level rise between 12 and 23 inches
by the decade 2080 (Climate Risk Information, 2009). According to the
NYS 2100 Commission, the maximum rise could be as high as six feet in
NYC and Long Island (NYS 2100 Commission, 2013). Not only are sea
levels themselves increasing, but so are their rates of change. As of now,
sea level rise increases at a rate of 0.86–1.5 inches per decade; 150
years ago, these rates were as low as 0.34–0.43 inches per decade
(Climate Risk Information, 2009).
Several measures were taken before Sandy made landfall to try and
prevent damage to transportation infrastructure, including moving
mobile units such as trains and buses out of low-lying areas, placing
sandbags and tarps at the entrances of subways and over vent to pre-
vent flooding, and preemptively clearing debris from drains and pumps
(Kaufman et al., 2012). Bus services, subway services, and commuter
and regional rail train services were shut down the night before Sandy
made landfall in order to allow the city time to prepare. This left many
people no other option but to either walk or use a car or taxi. At the
same time, many of these precautions proved successful: almost all
subway and bus services were restored just a few days after the storm,
and all three major airports in the area (LaGuardia, Newark, and JFK)
opened within two days after the storm (Kaufman et al., 2012). Despite
these efforts, much damage was still done; for example, some PATH
train stations were still closed weeks after the storm. Infrastructure
especially in Lower Manhattan and parts of Brooklyn and Queens near
the shore line was greatly affected, with tunnels and subways inundated
with storm and sea water, and surface and overground transport se-
verely damaged by high winds and torrential rain. Some sections of
subway lines were closed for up to 14 months. At one time, the MTA
alone estimated about five billion dollars in damage was done after
Sandy passed over (MTA, 2012).
Plans, such as MTAs ‘Fix & Fortify’, have been made to improve the
resilience of the transport infrastructure in the metro area. However,
projects like these are very costly. Some of the funding comes from the

∗
Corresponding author.
E-mail address: daziano@cornell.edu (R.A. Daziano).

https://doi.org/10.1016/j.tranpol.2018.05.010
Received 26 November 2017; Received in revised form 5 April 2018; Accepted 20 May 2018
Available online 28 May 2018
0967-070X/ © 2018 Elsevier Ltd. All rights reserved.
C. Wang et al.
state and federal government, some from the Department of
Transportation. Some also comes from revenue. Shutting down the
entire system not only loses revenue, but also causes a loss of money the
system already has. One way the MTA, for example, makes up for a
deficit is to introduce a fare hike. On March 22, 2015, commuters saw
the fourth fare hike for New York City Transit up to $2.75 per ride since
2009. Citizens complained saying that the paycheck they receive stays
the same even when the fares increase. However, the fare hike is in-
tended to help solve the financial obstacles MTA faces, and to support
projects such as ‘Fix & Fortify’ to make the system more resilient to
inclement weather.
The damages from Sandy come at great cost for all of the areas
affected. In New Jersey alone, over $3 billion in transportation-related
recovery expenditures were predicted between 2012 and 2015 (Mantell
et al., 2013). There were also additional, non-monetary costs. In a
survey of eight different residence locations including the five bor-
oughs, New Jersey, the Northern suburbs, and Long Island, six of the
eight locations reported an increase in travel time immediately after
Hurricane Sandy, sometimes by as much as two and a half hours
(Kaufman et al., 2012). These increased travel times correlated to an
increase in the frustration levels of commuters. All of the locations
surveyed reported some level of frustration due to transportation issues,
with the most extreme frustration levels being experienced by those
from Staten Island.
All of these factors – monetary costs due to damages from the storm,
time lost due to an impaired transportation system, and increased levels
of frustration – demonstrate why it is important and necessary to make
the transportation system of New York City and the surrounding areas
more resilient.
Traditional sources of funding for both recovering from disasters
and preventing future damages are not only limited, but also do not
account for benefit transfers of the externalities induced by the provi-
sion of resilient infrastructure. For instance, the construction of massive
structures such as surge barriers to protect coastal urban areas provokes
a positive externality on the residential value of the properties in the
area. This positive externality results from lower expected damage
coming from lower flood risks. In principle, property owners should be
willing to pay an amount equal to the perceived benefit if this positive
externality is internalized by them following some pricing mechanism.
Monetizing these benefit transfers can be used as a tool not only to
leverage scarce public resources, but also to achieve a socially optimal
resource allocation. An essential element is then the estimation of
willingness to pay, because this measure can be exploited to determine
the cost share residents are willing to cover to secure infrastructure
systems as well as to receive the benefits from minimizing potential
damage.
The goal of this paper is to provide statistical inference for the
private willingness to pay for improvements in the resilience to extreme
events of the transportation system in New York City. To make in-
ference on the willingness to pay for flood risk reductions, this paper
adopts an approach based on discrete choice experiments (Hensher
et al., 2005). This objective seeks to provide tools for better informing
planning investments to improve both resilience and security of trans-
portation infrastructure and services.
A fundamental, specific goal is to collect microdata using a unique
discrete choice experiment specifically designed for this project, where
resilience is understood as recovery time of the transportation system.
The population of interest for this study is those coastal communities in
the NYC area facing increased risks of flood damage.
There are multiple contributions in this work. There is the design of
an innovative discrete choice experiment that presents recovery time
after a disaster in an incremental way that is then used to derive in-
cremental willingness to pay estimates to support that recovery. The
estimates are derived no only using standard discrete choice models,
but also the state-of-the-art mixed-mixed logit model that allows for a
semiparametric representation of unobserved preference heterogeneity.
11
Transport Policy 69 (2018) 10–19
Empirical results indicate that the distribution of willingness-to-pay is
bimodal.
2. Literature review: financially supporting resilience investments
Although not directly related to the specific case study of this paper,
which we find to be unique, there is a well-established literature that
looks at the pre and post-disaster challenges in the catastrophe risk
insurance market (Jaffee and Russell, 1997; Grace et al., 1998, 2003;
Froot, 2001; Kunreuther et al., 2002; Kleindorfer and Klein, 2003;
Kunreuther, 2006; H. C. Kunreuther and Michel-Kerjan, 2009; Kousky,
2011; Paudel, 2012; Wang et al., 2017). Regarding demand-side dy-
namics of catastrophe risk insurance, there is strong evidence that
property owners often do not fully insure their property nor do they
invest in pre-event mitigation activities that can reduce losses
(Kunreuther, 1996; Kriesel and Landry, 2004; Kunreuther and Pauly,
2004; Dixon et al., 2006; Wang et al., 2017). Property owners fre-
quently do not have adequate financial resources to recover losses they
do experience, and may demand relief from the government
(Kunreuther and Pauly, 2004). In fact, evidence shows that major dis-
asters are often followed by large, unplanned government expenditures
that create major difficulties for local and state government budgets
(Kunreuther and Pauly, 2006). Over-reliance on post-disaster relief
from the government may create serious stress on the private insurance
market (Kunreuther and Pauly, 2004).
Using discrete choice theory for modeling insurance decisions by
property owners has appeared as a novel avenue of research. In parti-
cular, recent studies have looked at the determination of willingness to
pay for flood insurance, usually in the Netherlands (Brouwer and Akter,
2010; Botzen and Van Den Bergh, 2012; Brouwer and Schaafsma, 2013;
Dekker et al., 2016) but also in the US has emerged as well (H.
Kunreuther and Michel-Kerjan, 2009; Wang et al., 2017). These studies
analyze behavioral response in terms of willingness to pay (premium)
for a given insurance cover, which may have an associated deductible.
In addition to review past studies of flood and homeowners insurance,
Wang et al. (2017) specify mixed logit models for the choice of pur-
chase of flood and wind insurance, using stated-preference data from
North Carolina. The study finds that purchasing insurance is more likely
for higher income and younger homeowners with recent previous
hurricane experience and living in floodplains or closer to the coast.
Past damage also increases the probability of buying insurance. In terms
of willingness to pay estimates, the authors find a median of $0.55–0.65
as marginal premium to reduce the deductible by $1.
In addition to the analysis of the dynamics of the catastrophe risk
insurance market, there is a developing literature that looks at the role
of other funding mechanisms for improving resilience to floods. In
particular, the following two examples analyze willingness to pay
higher taxes. In the Netherlands, Koetse and Brouwer (2015) used
choice experiments to compare different economic measures of the
same change in welfare and to examine the effect of reference point on
preferences. To this end, four experiments were designed: two mea-
suring willingness-to-pay (WTP) and two measuring willingness-to-ac-
cept (WTA). They found a difference between WTP and WTA values and
the difference was affected by flood probability. Dekker et al. (2016)
estimated a model for WTP for flood risk reductions that included the
effects of preference certainty. They found that preference uncertainty
affects the randomness of decision making and/or the tendency to
choose the status quo choice(the adoption of a simplifying heuristic). In
addition, they also concluded that the preference certainty has no effect
on the marginal WTP.
3. The data
3.1. Survey instrument
Given the costs of recovery and infrastructure improvements, a
C. Wang et al. Transport Policy 69 (2018) 10–19

Table 2
Sample demographic statistics (N = 1552).
Respondent characteristics Percentage Income Percentage

Male 44.01 Personal Income ≤ $10,000 10.63%

Age 18-24 8.25 Pers Income >$10,000 and ≤ $19,999 6.89%
Age 25-34 19.78 Pers Income >$20,000 and ≤ $29,999 7.93%
Age 35-44 17.91 Pers Income >$30,000 and ≤ $39,999 8.96%
Age 45-54 17.78 Pers Income >$40,000 and ≤ $49,999 8.31%
Age 55-64 19.78 Pers Income >$50,000 and ≤ $59,999 9.73%
Age 65-74 13.14 Pers Income >$60,000 and ≤ $69,999 7.41%
Age 75 or older 3.35 Pers Income >$70,000 and ≤ $79,999 8.96%
Living in evacuation zones 22.81 Pers Income >$80,000 and ≤ $89,999 6.38%
Own pets 48.00 Pers Income >$90,000 and ≤ $99,999 6.05%
Single 26.61 Pers Income >$100,000 and ≤ $149,999 12.50%
Married 49.42 Personal Income >$150,000 6.25%

Full-time (≥ 30 h per week) job 52.45 Household Income ≤ $10,000 4.77%

Part-time/casual job 14.56 HH Income >$10,000 and ≤ $19,999 4.57%
Homemaker 3.67 HH Income >$20,000 and ≤ $29,999 5.67%
Full-time student 3.80 HH Income >$30,000 and ≤ $39,999 5.80%
Other job status 25.52 HH Income >$40,000 and ≤ $49,999 7.09%
High School Graduate 11.79 HH Income >$50,000 and ≤ $59,999 7.47%
Some College 20.30 HH Income >$60,000 and ≤ $69,999 7.93%
College Graduate 34.60 HH Income >$70,000 and ≤ $79,999 9.09%
Post-graduate degree 22.49 HH Income >$80,000 and ≤ $89,999 6.38%
White/Caucasian 77.00 HH Income >$90,000 and ≤ $99,999 7.41%
Black or African American 10.31 HH Income >$100,000 and ≤ $149,999 19.14%
Asian 7.28 Household Income >$150,000 14.69%
Hispanic 11.60

survey was designed to explore if New Yorkers are willing to financially
support investment to make the transportation system more resilient to
extreme weather. After a first draft of the survey was created, a round of
two focus groups was set to pretest the instrument. After the two focus
groups, the final design of the survey was completed. The survey
comprised the following parts:
1. Screening: to make sure the respondent was an adult living in the
New York Metropolitan area.
2. Daily transportation patterns: this section asked the respondent to
describe his or her daily trips in the city, including frequency of
transportation modes, average commuting time, and parking pre-
ferences.
3. Past flood/evacuation experience: this section collected data about
past flood and extreme weather events, including hurricanes, hur-
ricane evacuation, and property damage.
4. Initial Sandy disruptions: this section included a filter to know
whether the respondent was in New York City or other affected area
when superstorm Sandy made landfall. If that was the case, then it
was asked whether the respondent missed work and for how long,
and whether he or she was paid for the missed work-days.
5. Evacuation: if the respondent evacuated, it was registered when he
or she did so. Additional questions included whether the re-
spondent suffered any personal loss or knew someone who did.
6. Post Sandy Commute: data about patterns of disruption for the post
Sandy commute was registered in this section, including days be-
fore the commute returned to normal.
7. Commute disruptions: this section collected more details about the
disruptions experienced in the daily commute.
8. Other Sandy disruptions: in this section, the respondent had the
opportunity to provide information about other transportation
disruptions (such as plans to leave the city; subway line closures;
abnormal traffic congestion) as well as other overall impacts and
inconveniences (such as difficulty getting food, loss of cellphone
signal, blackouts, lack of heating).
9. Risks associated with extreme weather events: this section asked for
an evaluation of how likely a list of extreme weather events (heat
wave, heavy precipitation, storm surge, nor'easter, among others)
would harm vital transportation infrastructure.
10. Responsibility: assessment of the degree of responsibility for being
prepared for hurricanes.
11. Contingent valuation: this section first introduced text about the
mechanisms (infrastructure improvements) that can be im-
plemented to prevent or reduce damage to the transportation
system when floods and hurricanes occur and improve system re-
silience. Then there was a series of contingent valuation question to
determine how much the respondents were willing to pay for
projects that would reduce the transportation recovery time.
12. Funding mechanisms: this section collected data about the like-
lihood of supporting different funding schemes (such as increased
tax on vehicle sales, increase tax on gas, increased subway fares and
increased parking fees).
13. Discrete choice experiment: a series of discrete choice questions
about the respondents' willingness to pay to improve the recovery
time of the transportation system.
14. Sociodemographic data: the last section of the survey contained a
set of question to gather sociodemographic characteristics of the
respondents.
3.2. The sample
The data was collected in January 2015 using an online panel of
1552 adult respondents living in the NYC metropolitan area. A
minimum sample of 250 respondents was determined as part of the
design of the discrete choice experiment (Rose and Bliemer, 2009), but
we decided to work with a sample well above the minimum to make
sure to have enough representation of different segments so that het-
erogeneity in preferences could be analyzed. Table 2 summarizes the
sociodemographic characteristics of the sample. Comparing the sample
to the population resident in NYC, our sample is more female, wealthier
(median household income of $78,290 in the sample vs. $55,7521 in the
population in dollars of 2015) and biased toward less diversity (10.31%
in the sample vs. 25.1% in the population are black or African
1
American Community Survey estimate.

12
C. Wang et al. Transport Policy 69 (2018) 10–19

Americans; 11.6% vs. 27.5% Hispanic; 7.28% vs. 11.8% Asian).2 The Table 3
sample being of an opt-in online panel (Qualtrics) explains some of the Tobit model for willingness to pay for improving resilience (open-ended con-
differences. Since the purpose of the analysis is the derivation of will- tingent valuation question).
ingness to pay estimates controlling by sociodemographics, non re- Variables Estimate s.e. t-stat
presentativity of the sample is not a main concern as all groups have
enough representation and variability; use of weights did not change Male 253.28 69.60 3.64
Lives in Evac Zone 478.27 91.03 5.25
willingnes-to-pay estimates.
Lives in brownstone 476.84 107.13 4.45
In terms of disruptions provoked by Sandy (Appendix A), re- Children: 1-2 178.30 79.33 2.25
spondents declared difficulties getting food, water, fuel, loss of cell Children: over 3 639.61 132.70 4.82
phone signal, and electric power. For example, 72.99% experienced Education college −175.04 72.64 −2.41
Hispanic 236.11 105.99 2.23
difficulty in getting fuel and 70.88% were affected by malfunctioning
Liberal 240.66 71.02 3.39
traffic signals. Fuel shortages lasted more than 3 days for 74.13% of the Times experienced hurricane −28.49 10.57 −2.70
affected respondents, whereas problems with traffic signals laste more Times evacuated 151.94 47.13 3.22
than 3 days for 59.01% of those experiencing that specific disruption. Was paid for lost workdays 176.99 70.58 2.51
65% of the respondents stated that their normal commute was dis- Know people who experienced loss −164.81 76.53 −2.15
Experienced surge at home 219.82 99.13 2.22
rupted during or immediately following Sandy. 24% of the respondents
Experienced surge in neighborhood −272.06 84.51 −3.22
stated that the disruption lasted longer than 7 days, with an average Used social media 334.70 80.25 4.17
commute disruption duration of 5.38 days. Freq user: subway 207.14 82.39 2.51
Freq user: rail 478.82 125.35 3.82
Constant −220.66 106.54 −2.07
4. Model and results
Loglikelihood −1411
4.1. Contingent valuation BIC 2918.73

As stated above, the survey considered an open contingent valuation

question, where respondents were directly asked to elicit their will-
ingness to pay. Considering the question of a one-time payment to
support investments in making subway infrastructure more resilient,
81.3% of the respondents indicated that they would pay less than
$1000 (as a one-time payment). Only 3.61% would pay more than
$4000. When asked for a monthly payment, 63.23% of the respondents
declared that they would pay less than $10 per month, whereas only
0.19% would pay more than $200.
From the answers to this contingent valuation question, a tobit
model was estimated. Estimates (Table 3) can be interpreted as additive
components of the total willingness to pay. Taking into consideration
that the model has a negative constant of -$220, males, young people
(under 25 years), those living in evacuation zones, those with 3 or more
children, non-whites, and those with a higher experience in previous
evacuation are all willing to pay a significant, positive amount of
money to support resilience investments. Frequent users of subway, and
then even more frequent users of rail, also have significant positive
willingness-to-pay components. Interestingly, both suffering personal
loss and surge at home are associated with a positive willingness to pay,
but knowing someone who experienced loss and having experienced
surge in the neighborhood both have a negative parameter. The largest
willingness-to-pay component is for those having children over 3 years
of age, then come those who live in an evacuation zone, those living in a
brownstone, and frequent users of rail. These three covariates have very
similar point estimates.
4.2. Discrete choice experiments – design
A fundamental research problem in environmental economics and
engineering is to determine the consumer valuation of environmental
goods and services for which there is no market price. Building on the
theory of random utility maximization (McFadden, 2001), choice
modeling is more flexible than other stated-preference techniques, such
as the contingent valuation method used above, because implicit prices
are determined within a controlled experiment that offers better
framing and scope. In transportation, choice modeling has become the
standard technique for estimating the willingness to pay for travel-time
savings (used for establishing fares and tolls) and for deriving demand
functions, especially in the case of new alternatives.
2
Population statistics from the 2010 Census.
In this paper, the stated-preference discrete choice experiment was
designed using a Bayesian efficient design, using the software Ngene
(Rose and Bliemer, 2009). Efficient designs go beyond traditional ways
of determining fractional factorial designs. An efficient design max-
imizes the information extracted from each choice situation by mini-
mizing asymptotic standard errors. In the case of a Bayesian design,
there is the possibility of assuming prior parameters, which were ob-
tained from pretests of the experiment.
In the choice experiment, two hypothetical scenarios of infra-
structure improvement as well as a status-quo scenario were described
in terms of the percentage of the transportation system being operative
1–2 days, 3–5 days, 1 week, and 2 weeks after a highly disruptive ex-
treme weather event. Fig. 1 presents a sample of one of the discrete
choice scenarios as seen by respondents. The extreme weather event
was described as being at least as strong as hurricane Sandy. We will
refer to these percentages as operative levels in the remainder of the
paper. The two scenarios of infrastructure improvement required fi-
nancial support in the form of a hypothetical annual payment. The
attribute levels are shown in Table 4.
4.3. Discrete choice experiments: inference on willingness to pay
For modeling purposes, we considered the operative levels in-
crementally. For example, if the operative level 1–2 days after the storm
was 25%, and then increased to 35% for days 3–5, then we considered
25% for the first 1–2 days, and 10% for days 3–5. The consideration of
incremental recovery makes it possible to derive incremental will-
ingness to pay estimates for supporting a faster recovery. All estimation
was performed using the gmnl package in R (Sarrias and Daziano,
2017).
As base models we estimated a multinomial logit (MNL) model, a
mixed multinomial logit model (MIXL-N) with normally distributed
parameters, and a hierarchical mixed multinomial (H-MIXL-N) logit
model where some interactions with sociodemographic variables were
introduced. Fit of these competing models is presented in Table 5.
Because the models with unobserved preference heterogeneity
provided better fit and significant evidence of heterogenous pre-
ferences, after preliminary tests, a mixed-mixed logit (MM-MNL) model
with two discrete components of the Gaussian mixture was selected as
the best specification. Even though solutions with 2–5 mixture com-
ponents were tried, overfitting and overidentification can be expected
with a large number of components. For this dataset, more than 2

13
C. Wang et al. Transport Policy 69 (2018) 10–19

Fig. 1. Sample of discrete choice scenario.

Table 4 Table 6
Attributes and attribute levels for the WTP for improving resilience experiment. MM-MNL (mixture of normals logit) estimates.
Attribute Levels Variable Estimate SD t-stat

Willingness to pay for resilience improvement 100, 120, 250, 300, Fixed coefficients
400, 500 (in $) Inc. operative level: 3–5 days 0.1054 0.0377 2.796
Percentage of operational transportation system in 1–2 0%, 25% Inc. operative level: 1 week 0.0823 0.0301 2.734
days Random coefficients
Improvement in percentage of operational 5%, 15%, 35%, 50% Class 1
transportation system in 3–5 days Annual payment (mean) −0.0855 0.0241 −3.548
Improvement in percentage of operational 25%, 40%, 60%, 85% Operative level: 1–2 days (mean) 0.0556 0.0350 1.589
transportation system in 1 week Inc. operative level: 2 weeks (mean) 0.0824 0.0216 3.815
Improvement in percentage of operational 70%, 85%, 90%, 100% Annual payment (SD) 0.0434 0.0124 3.500
transportation system in 2 weeks Operative level: 1–2 days (SD) 0.0975 0.0704 1.385
Inc. perative level: 2 weeks (SD) 0.0349 0.0181 1.928
Class 2
Annual payment (mean) −0.0111 0.0031 −3.581
Table 5
Operative level: 1–2 days (mean) 0.1240 0.0394 3.147
Fit of competing models. Inc. operative level: 2 weeks (mean) 0.0590 0.0240 2.458
Model Loglikelihood BIC Annual payment (SD) 0.0078 0.0069 1.130
Operative level: 1–2 days (SD) 0.0284 0.0085 3.341
MNL −11921.0 23889.94 Inc. perative level: 2 weeks (SD) 0.0197 0.0064 3.078
MXL-N −9462.5 19009.83 Assigntment to class 2
H-MXL-N −9384.0 18899.88 Male 0.0951 0.0658 1.445
MM-MNL −7449.4 15172.15 Income below $70,000 −0.4821 0.3145 −1.533
Missed work after Sandy 0.4640 0.2088 2.222
Liberal 0.4280 0.1918 2.231
Conservative −0.1262 0.0734 −1.719
classes always led to a group of either non-significant or wrong-signed African American 0.0956 0.4534 0.211
parameters. The shares of the two components are the following, class Asian 0.4199 0.3282 1.279
1: 68%, and class 2: 32%. The estimates of relevant parameters are Hispanic 0.5579 0.3030 1.841
presented in Table 6.
Loglikelihood −7449.4
All the parameters have the expected signs. The annual payment
BIC 15172.15
parameter represents the marginal utility of income and is negative,
meaning that individuals value saving money. All the operative levels
are positively valued. The order of the valuation is decreasing, as faster significantly more likely to belong to class 2. As we discuss below, in-
recovery is preferred. dividuals in class 2 are willing to pay more for improving recovery of
Note how whereas class 1 (68%) exhibits more heterogeneity in the the transportation system. Hispanic and Asian males with a household
payment attribute, class 2 (32%) shows more heterogeneity in the in- income greater then $70,000 are also more likely to belong to class 2,
cremental improvements in the operative levels. Regarding assignment although the effect is not significant at the 95% level.
to classes (or weights of the mixture components), those respondents The estimated mixing distributions for the random parameters are
that are liberal and missed work after Sandy, are statistically shown in Fig. 2. These graphs are based on a normal kernel function of

14
C. Wang et al. Transport Policy 69 (2018) 10–19

Fig. 2. Estimated mixing densities of the random coefficients.

the point estimates. The estimated mixing distributions of the marginal calculation of these willingness-to-pay estimates that reside the policy
utility for both the recovery 1–2 days after the event (βopera12 ) and 2 implications of this study. To derive the WTP distribution, conditional
weeks after (βincopera2w ) are Gaussian-shaped. This result indicates that, estimates of the marginal rate of substitution between recovery time
instead of a mixture, using a normal distribution to approximate the and the annual payment – the ratio of the parameter of recovery time of
taste variations of these two attributes would be enough. The estimated interest and that of the annual payment – were produced at the in-
mixing distribution of the marginal disutility of the payment (βprice ), dividual level from the population distribution.
instead, shows two clear separated peaks in correspondence with the Individuals are willing to pay more for a faster recovery of the
values of the means of each component. Hence, the heterogeneity dis- transportation system after an extreme weather event. Note that the
tribution of the marginal disutility for the annual payment clearly willingness to pay is considerably higher for class 2 (32% share in the
seems to be multimodal (at least bimodal), a result that would have not sample), which is more likely to include individuals who missed work
been revealed by using a simple normal density as mixing distribution. during superstorm Sandy. Those individuals who missed work suffered
To monetize the estimates, we also derived post-processed will- directly the impacts of Sandy.
ingness-to-pay (WTP) measures for each attribute (Table 7). It is in the To have a better idea of the derived estimates, Table 8 shows the
total annual willingness to pay for investments in the transportation
system under differing hypothetical situations. Each scenario describes
Table 7 different recovery times for the transportation system with respect to a
Summary of individual post-processed WTP. hypothetical base level. The base scenario considers that the transpor-
WTP [$/percent-point] Mixture Class 1 Class 2 tation system is shut down 1–2 days after a highly disruptive event;
only 5% in operative 3–5 days after; one week after the system is 25%
Est. SD Est. SD Est. SD operative; and 2 weeks after the event, 70% of the system is in opera-
tion. The scenarios are ordered in increasing recovery speed. The
Operative level: 1–2 8.6377 604.44 1.1813 2.4203 14.2089 3.0806
days after event willingness to pay for each scenario is incremental with respect to the
Incremental operative 7.7826 512.22 1.5671 1.9079 12.0162 2.8634 base condition. Policy implications of these estimates are discussed
level: 3–5 days below.
Incremental operative 6.0721 457.40 1.1904 1.2859 9.6324 3.4012 Finally, density plots for the WTP measures are shown in Fig. 3. All
level: 1 week
Incremental operative 4.5678 404.62 1.0938 0.6868 6.7232 2.5661
4 post-processed WTP measures are bimodal, which explains the overall
level: large standard deviations for the mixture WTP of Table 7.

15
C. Wang et al. Transport Policy 69 (2018) 10–19

Table 8 two main groups are identified, differentiated in their degree of fi-

Simulated annual WTP for differing scenarios of recovery. nancial support. The willingness to pay for class 1 ranges from about
Situation Annual WTP ($) Scenarios $15 to $50, whereas that of class 2 ranges from $120 to $775. As a
mixture of the two groups, the range of variation is $75-$450. The class
Operative levels certain period after(%) with a higher willingness to pay (class 2, representing 32% of our
sample) is more likely to be composed by individuals who missed work
Mixture Class 1 Class 2 1–2 3–5 1 week 2 weeks
days days
during Sandy (odds ratio of 59%), self-identified as liberals (odds ratio
of 53%), with annual household income above $70,000. In the con-
Base Scenario NA NA NA 0 5 25 70 tingent-valuation question, where respondents were asked how much
Situation 1 76.80 14.27 122.79 0 10 40 80 they would pay to “support investments that would reduce the recovery
Situation 2 169.66 30.91 272.04 0 20 60 90
Situation 3 149.64 19.88 250.58 10 30 60 80
time from 3 weeks to only 3 days”, the average annual willingness to
Situation 4 305.29 51.22 490.90 10 50 80 100 pay was $192, with a standard deviation of $305. Future work should
Situation 5 261.68 31.60 438.43 30 50 70 90 explore the preferred mechanism to collect these additional funds.
Situation 6 322.40 43.50 534.75 30 50 80 100
Situation 7 295.88 27.65 506.12 50 60 70 90
Situation 8 388.75 44.29 655.38 50 70 90 100
Situation 9 377.28 29.40 655.84 70 80 90 90
5. Conclusions
Situation 10 455.11 45.07 776.00 70 90 100 100
New York City is known to many as the “city that never sleeps”.
With over eight million citizens, it is the most populated city in the
4.4. Policy implications United States. The area surrounding the city itself is known as the New
York Metropolitan Area, which includes some New York State counties
Sandy revealed some critical deficiencies in the area's transportation north of the city, parts of New Jersey, Pennsylvania, and Connecticut as
system, and, unfortunately, experts predict that future sea level rise and a combined statistical area (CSA). This CSA is also the most populated
storms will exacerbate the problems caused by these deficiencies. in the US at over 23 million people. An area with this large number of
Luckily, there are many engineering solutions to the problem of residents needs a well-built and well-maintained public transportation
transportation resilience that are currently being explored. However, system. Most people living outside New York City commute to the city
the construction, implementation and management of these solutions for both business and leisure. Not many can afford a car, or afford the
will require locating new sources of funding, possibly including taxes extremely high prices of parking in the city. Therefore, agencies such as
and usage fees, as well as guaranteeing that taxpayers and users are New Jersey Transit (NJT), Port Authority of New York and New Jersey
willing to partially support the associated costs. (PANYNJ), and Metropolitan Transportation Authority (MTA) have
This study provides evidence of individuals willingness to provide worked together to provide vital bus and rail services in the area. If
private economic resources to fund mechanisms to increase the resi- these services were to cease operations for any amount of time, the
lience of New York Citys transportation system. As described above, entire area would come to a halt as well.

Fig. 3. Estimated mixing densities of the willingness-to-pay measures.

16
C. Wang et al. Transport Policy 69 (2018) 10–19

In this project, a survey was designed to collect data on the dis-
ruptions that individuals experienced during and after superstorm
Sandy. 1552 adults living in the metropolitan area of New York City
participated in the online survey. For those who were in an area af-
fected by Sandy, more than 20% experienced subway line and station
closures, and limited bus service for at least one week. 30% of those
affected stated that they considered superstorm Sandy a major disrup-
tion.
The empirical dataset was complemented with a unique discrete
choice experiment. In the experiment, respondents faced differing sce-
narios of the expected transportation system recovery after an extreme
weather event (as strong as superstorm Sandy) under a hypothetical
annual payment. Recovery was expressed in terms of the percentage of
the system being operative in discrete times after the disruptive shock.
With a nonlinear specification, the preferences with respect to both
recovery and payment turned out to be heterogeneous in the sample.
Whereas some components of recovery seem to be normally dis-
tributed, payment heterogeneity is better represented by a bimodal
distribution that can be reasonably approximated by a mixture of two
normal distributions. The estimate of the weights of the mixture –
modeled as a logit-type function of subject covariates – revealed that
individuals who missed work after Sandy hit the area – which acts as a
proxy for those individuals who experienced major disruptions in their
routines, reducing their earnings in some cases – are more likely to pay
more for improving recovery of the transportation system. A similar
effect is observed for individuals who identified themselves as politi-
cally liberal. These variables defined a specific class of residents, which
we labeled as ‘class 2’. Using hypothetical scenarios of recovery, the
willingness to pay as an annual for ‘class 1’ ranges from about $15 to
$50, whereas that of class 2 ranges from $120 to $775. For the mixture,
the range of variation is $75-$450. These estimates provide evidence of
the existence of a positive willingness to pay for funding resilience in-
vestments in New York city.
Acknowledgements
This research was partly supported by UTRC/RF Grant No: 49198-
21-26 (data collection and main analysis) as well as by the Center for
Transportation, Environment, and Community Health (CTECH) (tech-
nical work to produce the postprocessed willingess to pay estimates).
Assistance in reviewing previous studies and in suggesting some of the
survey questions by Melissa Fickel as part of her undergraduate summer
research is appreciated.

Appendix A. Supplementary data

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.tranpol.2018.05.010.

Appendix A. Reported Sandy-provoked disruptions

Table 1
Stated disruptions and their duration.

Not at all A few hours 1–2 days 3–5 days 1 week 1–2 weeks 2 + weeks

Subway line closures 42.00% 6.97% 17.90% 12.82% 7.97% 6.26% 6.08%
Subway station closures 42.65% 6.67% 16.77% 11.70% 8.74% 7.03% 6.44%
Closed bridges 49.20% 7.38% 18.25% 11.16% 6.91% 4.19% 2.89%
Closed tunnels 50.27% 5.55% 16.72% 11.05% 7.56% 4.49% 4.37%
Abnormal traffic congestion 31.42% 7.32% 18.37% 15.06% 13.17% 9.04% 5.61%
Limited bus service 41.29% 6.62% 17.66% 13.88% 10.40% 7.03% 3.13%
Commuting 2X + longer 37.27% 7.21% 16.60% 15.53% 11.16% 7.56% 4.67%
Forced to change routes 31.19% 6.85% 16.48% 16.60% 12.40% 9.74% 6.73%
Working at alternate sites 59.60% 4.13% 10.22% 9.75% 7.62% 4.96% 3.72%
Severe transit crowding 45.36% 5.43% 13.29% 11.40% 10.63% 7.97% 5.91%
Difficulty getting fuel 26.99% 5.79% 13.11% 17.48% 18.02% 11.64% 6.97%
Difficulty getting food 46.01% 9.21% 20.26% 10.63% 6.91% 4.31% 2.66%
Difficulty getting water 56.53% 8.45% 14.83% 8.51% 5.73% 4.02% 1.95%
Loss of cell signal 38.81% 17.90% 15.95% 12.64% 8.09% 4.67% 1.95%
Loss of electric power 28.59% 11.52% 12.17% 17.07% 12.99% 12.52% 5.14%
Malfunction of traffic signals 29.12% 8.68% 20.38% 17.84% 13.05% 7.68% 3.25%

17
C. Wang et al. Transport Policy 69 (2018) 10–19

Appendix B. Discrete choice models and unobserved preference heterogeneity

Because preferences are likely to vary across individuals, econometric models of choice have moved toward flexible ways of incorporating
unobserved preference heterogeneity. In general, consider the following logit-type random indirect utility of individual i for alternative j and time (or
‘choice situation’) t:
⊤
Uijt = x ijt βi + εijt , (1)
where x ijt is a K × 1 vector of alternative attributes; εijt is an i.i.d. type 1 extreme value tasgte shock; and the preference parameter vector βi is
unknown but is assumed to vary randomly in the population.
The distribution of βi is known as heterogeneity or mixing distribution and can be either discrete, continuous or a combination of the two.
Assuming a parametric continuous mixing density f (βi θ ) , where θ are the parameters of this distribution. For example, the most common mixing
specification is to consider multivariate normally distributed preferences, i.e. βi ∼ N (β, Σ) which leads to the common mixed logit with normally
distributed parameters. Note that in this case, βi = β + Lηi , where ηi ∼ N (0, I) , and L is the lower-triangular Cholesky factor of Σ .
To accomodate both observed and unobserved preference heterogeneity, the following hierarchical specification of βi can be used:
βi = β + Πwi + Lηi , (2)
where wi is a set of M individual characteristics, and Π is a K × M matrix of additional parameters.
The mixing distribution can also be discrete, leading to a latent class logit model (Kamakura and Russell, 1989; DeSarbo et al., 1995; Bhat, 1997;
Greene and Hensher, 2003; Shen, 2009). In a latent class logit model, an individual i has a set of fixed preference parameters βq with probability ωiq –
i.e., Pr(βi = βq) = ωiq , where ∑q wiq = 1 and ωiq > 0 . The discrete mixing distribution (or class assignment probability) is unknown to the analyst, but
the number of classes Q (segments with differing preferences) is set before estimation. The model is completed with a logit specification for the class
assignment probabilities:
exp(hi⊤γq)
ωiq = Q
,
∑q = 1 exp(hi⊤γq) (3)
where hi denotes a set of socioeconomic. The parameters of a refernce class – usually the first class – are normalized to zero (γ1 = 0 ) for identifi-
cation.
Let yijt = 1 if individual i chooses j on occasion t, and 0 otherwise. Then, the unconditional probabilities of the sequence of choices by individual i
for MIXL and LC are respectively given by:
yijt
T J ⊤
⎧ ⎡ exp(x ijt βi )⎤ ⎫
Pi (θ ) = ∫ ⎨∏ ∏ ⎢ ∑
J
exp(x ⊤
β )
⎥ f (β ) dβ
⎬ i i
t j ⎢
⎣ j=1 ijt i ⎥
⎦ ⎭
⎩
yijt
Q ⎧T J ⎡ ⊤
exp(x ijt βq)⎤ ⎫
Pi (θ ) = ∑ ⎨∏ ∏ ⎢ J ⎥ ωiq .
q t j ⎢ ∑ exp(x ijt βq) ⎥ ⎬
⊤
⎩ ⎣ j=1 ⎦ ⎭
Note that both expressions are basically the same, with the mixed logit being a continuous mixture; and the latent class logit, a discrete mixture.
On the one hand, mixed logit models impose a very specific shape for the distribution of preferences (such as the normal, which unimodal,
unbounded, and symmetric); on the other hand, latent class logit models are flexible and easy to interpret but within a class there is no unobserved
preference heterogeneity. To exploit the benefits of both specifications, researchers have proposed the use of a discrete-continuous mixture model.
The idea behind the double mixture model is simple: there is a discrete number of clusters of individuals (just as in the latent class logit model),
but within each cluster preferences are modeled according to a mixed logit model with a multivariate normal heterogeneity distribution. From a
statistical point of view, this discrete-continuous representation of preference heterogeneity is interpreted as a Gaussian mixture. Gaussian mixtures –
i.e., a linear combination of normal distributions – can approximate any continuous distribution, including multimodal distributions, to a desired
level of accuracy. The resulting logit-type model with a Gaussian mixture is known in the recent choice modeling literature as the mixed-mixed logit
(MM-MNL) model (Bujosa et al., 2010; Keane and Wasi, 2013; Greene and Hensher, 2013).
The main assumption of the mixed-mixed logit model is:
βi ∼  (βq, Σq) with probability ωiq for q = 1, …, Q, (4)
which leads to the following expression of the unconditional probability of the sequence of choices made b individual i:
ijty
Q ⎛ T J ⊤ ⎞
⎧ ⎡ exp(x ijt βi ) ⎤ ⎫
Pi (θ ) = ∑ ωiq ⎜
⎜
∫ ⎨∏ ∏ ⎢ J ⊤ ⎥
∑ exp(x ijt βi ) ⎥ ⎬
fq (βi ) dβi ⎟,
⎟
q=1 t j ⎢
⎣ j=1 ⎦ ⎭ (5)
⎝ ⎩ ⎠
where fq (βi ) = N (βq, Σq) . Due to the complex expression of the choice probability, estimation in the gmnl package (Sarrias and Daziano, 2017) uses
a Monte-Carlo approximation of the sequence of individual choices as well as an analytical expression of the gradient.

References intercity travel. Transport. Sci. 3, 3448.

Botzen, W.J.W., Van Den Bergh, J.C., 2012. Monetary valuation of insurance against flood
risk under climate change. Int. Econ. Rev. 53 (3), 1005–1026.
Bhat, C., 1997. An endogenous segmentation mode choice model with an application to Brouwer, R., Akter, S., 2010. Informing micro insurance contract design to mitigate

18
C. Wang et al.
climate change catastrophe risks using choice experiments. Environ. Hazards 9 (1),
74–88.
Brouwer, R., Schaafsma, M., 2013. Modelling risk adaptation and mitigation behaviour
under different climate change scenarios. Climatic Change 117 (1–2), 11–29.
Bujosa, A., Riera, A., Hicks, R.L., 2010. Combining discrete and continuous representa-
tions of preference heterogeneity: a latent class approach. Environ. Resour. Econ. 47
(4), 477493.
Climate Risk Information, 17 Feb. 2009. New York City Panel on Climate Change. http://
www.nyc.gov/html/om/pdf/2009/NPCC%5c%5fCRI.pdf.
Dekker, T., Hess, S., Brouwer, R., Hofkes, H., 2016. Decision uncertainty in multi-attri-
bute stated preference studies. Resour. Energy Econ. 43, 57–73.
DeSarbo, W., Ramaswamy, V., Cohen, S., 1995. Market segmentation with choice-based
conjoint analysis. Market. Lett. 6, 137147.
Dixon, L., Clancy, N., Seabury, S.A., Overton, A., 2006. The National Flood Insurance
Program's Market Penetration Rate: Estimates and Policy Implications. RAND, Santa
Monica, CA 2006.
Federal Emergency Management Agency (FEMA), 2013. Hurricane Sandy FEMA After-
action Report. . http://www.fema.gov/media-library-data/20130726-1923-25045-
7442/sandy%5c%5ffema%5c%5faar.pdf.
Froot, K.A., 2001. The market for catastrophe risk: a clinical examination. J. Financ. Econ.
60 (2), 529–571.
Grace, M.F., Klein, R.W., Kleindorfer, P.R., 1998. Overview of Catastrophe Insurance
Markets in the U.S., the Wharton Catastrophe Risk Management Project.
Grace, M.F., Klein, R.W., Kleindorfer, P.R., Murray, M.R., 2003. Catastrophe insurance:
consumer demand, markets and regulation. 1 (1), 76–91.
Greene, W.H., Hensher, D.A., 2003. A latent class model for discrete choice analysis:
contrasts with mixed logit. Transp. Res. Part B Methodol. 37 (8), 681698.
Greene, W.H., Hensher, D.A., 2013. Revealing additional dimensions of preference het-
erogeneity in a latent class mixed multinomial logit model. Appl. Econ. 45 (14),
1897–1902.
Hensher, D.A., Rose, J.M., Greene, W.H., 2005. Applied Choice Analysis: a Primer.
Cambridge University Press.
Jaffee, D.M., Russell, T., 1997. Catastrophe insurance, capital markets, and uninsurable
risks. J. Risk Insur. 205–230.
Kamakura, W., Russell, G., 1989. A probabilistic choice model for market segmentation
and elasticity structure. J. Market. Res. 26, 379390.
Kaufman, S., Qing, C., Levenson, N., Hanson, M., 2012. Transportation during and after
Hurricane Sandy.
Keane, M., Wasi, N., 2013. Comparing alternative models of heterogeneity in consumer
choice behavior. J. Appl. Econom. 28 (6), 1018–1045.
Kleindorfer, P.R., Klein, R.W., 2003. Regulation and markets for catastrophe insurance.
In: Advances in Economic Design. Springer Berlin Heidelberg, pp. 263–279.
Koetse, M.J., Brouwer, R., 2015. Reference dependence effects on WTA and WTP value
functions and their disparity. Environ. Resour. Econ. 1–23 Koetse2015.
19
Transport Policy 69 (2018) 10–19
Kousky, C., 2011. Managing natural catastrophe risk: state insurance programs in the
United States. Rev. Environ. Econ. Pol. 5 (1), 153–171.
Kriesel, W., Landry, C., 2004. Participation in the national flood insurance program: an
empirical analysis for coastal properties. J. Risk Insur. 71 (3), 405–420.
Kunreuther, H., 1996. Mitigating disaster losses through insurance. J. Risk Uncertain. 12
(2–3), 171–187.
Kunreuther, H., 2006. Disaster mitigation and insurance: learning from Katrina. Ann. Am.
Acad. Polit. Soc. Sci. 604 (1), 208–227.
Kunreuther, H., Meyer, R., Zeckhauser, R., Slovic, P., Schwartz, B., Schade, C., Luce, M.F.,
Lippman, S., Krantz, D., Kahn, B., Hogarth, R., 2002. High stakes decision making:
normative, descriptive and prescriptive considerations. Market. Lett. 13 (3),
259–268.
Kunreuther, H.C., Michel-Kerjan, E.O., 2009. At War with the Weather: Managing Large-
scale Risks in a New Era of Catastrophes. MIT Press.
Kunreuther, H., Pauly, M., 2004. Neglecting disaster: why don't people insure against
large losses? J. Risk Uncertain. 28 (1), 5–21.
Kunreuther, H., Pauly, M., 2006. Rules rather than discretion: lessons from hurricane
Katrina. J. Risk Uncertain. 33 (1–2), 101–116.
Kunreuther, H., Michel-Kerjan, E., 2009. At War with the Weather: Managing Large-scale
Risks in a New Era of Catastrophes. MIT Press, Cambridge.
Mantell, N.H., Seneca, J.J., Michael, L.L., Irving, W., 2013. The Economic and Fiscal
Impacts of Hurricane Sandy in New Jersey. Rep. Rutgers. http://www.esf.edu/glrc/
library/documents/EconomicandFiscalImpactsofSandy%5c_2013.pdf.
McFadden, D., 2001. Economic choices. Am. Econ. Rev. 91 (3), 351–378.
Metropolitan Transportation Authority (MTA), 2012. MTA 2013 Final Proposed Budget,
November Financial Plan 2013-2016. http://web.mta.info/mta/budget/nov2012/
November%5c%5f2012%5c%5fFinancial%5c%5fPlan%5c%5fVol%5c%5f1.pdf.
National Centers for Environmental Information (NCEI), 2016. Billion-dollar Weather and
Climate Disasters: Table of Events. http://www.ncdc.noaa.gov/billions/events.
NYS 2100 Commission, 2013. Recommendations to Improve the Strength and Resilience
of the Empire States Infrastructure. Albany, NY January, 11.
Paudel, Y., 2012. A comparative study of public–private catastrophe insurance systems:
lessons from current practices. Geneva Pap. Risk Insur. - Issues Pract. 37 (2),
257–285.
Rose, J.M., Bliemer, M.C., 2009. Constructing efficient stated choice experimental de-
signs. Transport Rev. 29 (5), 587–617.
Sarrias, M., Daziano, R.A., 2017. Multinomial logit models with continuous and discrete
individual heterogeneity in R: the gmnl package. J. Stat. Software 79 (2), 1–46.
Shen, J., 2009. Latent class model or mixed logit model? A comparison by transport mode
choice data. Appl. Econ. 41 (22), 29152924.
Wang, D., Davidson, R.A., Trainor, J.E., Nozick, L.K., Kruse, J., 2017. Howeowner pur-
chase of insurance for hurricane-induced wind and flood damage. Nat. Hazards 88,
221245.