Global Environmental Change 53 (2018) 137–145

Contents lists available at ScienceDirect

Global Environmental Change

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

A human-environmental network model for assessing coastal mitigation T

decisions informed by imperfect climate studies
Mitchell J. Smalla,b, , Siyuan Xianc
⁎

a
Anderson Family Visiting Professor in Energy and the Environment, Princeton University, United States
b
Civil & Environmental Engineering and Engineering & Public Policy, Carnegie Mellon University, United States
c
Civil & Environmental Engineering, Princeton University, 59 Olden St., Princeton, NJ, 08540, United States

ARTICLE INFO ABSTRACT

Keywords: A Bayesian network model is developed to explore the interaction between physical and social processes that
Bayesian network influence mitigation decisions and outcomes for extreme events. The network includes statistical relationships
Risk perception for event occurrence and magnitude; uncertainty in the parameters of these models; a high degree of variability
Mitigation regret in the sequence of events that occurs in any given time interval, and the possibility of long-term trends in the
Coastal storm flooding
frequency, magnitude and impact of events. The model is applied to coastal storm surge events in the New York
Climate studies
New York City
City (NYC) area. A 50 cm increase in sea level is predicted to approximately double the expected cumulative
damage over a 40-year period. A 20% increment in storm frequency yields a further predicted increase of about
18% in the cumulative damage. The uncertainties in long-term trends associated with climate change may be
reduced by scientific studies. However the value of this information is affected both by study accuracy and the
extent of its trust, acceptance and utilization by decision makers. Implications of this are assessed in the model,
showing that the probability of regret is notably reduced when climate study results are used to support miti-
gation decisions. This is demonstrated even when the studies have relatively low accuracy, moreso when they
exhibit good or perfect accuracy. Based on model insights and limitations, further research needs are identified
to better understand extreme event risk perception and management in coupled human-environmental systems.

1. Introduction different inferences and policy recommendations can result (Stiber

et al., 1999). Among conflicting experts, decision makers, or other in-
In many cases the capacity to understand and manage long-term terested parties, additional studies and knowledge can help to build
natural and anthropogenic risks is found to improve with time (Brody, consensus for management plans, but not in all cases (Xian et al.,
2003; Brody et al., 2009; Albright and Crow, 2015; Reyers et al., 2015; 2018a). Depending on how these studies are perceived, they may have
Jongman et al., 2015; Bouwer and Jonkman, 2018). However, the path little or no effect on stakeholder preferences for management options,
to greater scientific knowledge, social learning, and improved capacity and may even increase conflict (Small et al., 2014).
for risk management is often indirect, and rarely rapid. As noted in the Modern statistics provides a quantitative framework for assessing
literature, risk knowledge may be subject to “negative learning” the extent to which inferences from limited samples might misrepresent
(Oppenheimer et al., 2008) in which natural variability or biased study both an underlying population and the distribution of a future sample
and appraisal cause scientific beliefs regarding risks to diverge from of observations from that population (Cox and Hinkley, 1979; Geisser,
their true values. This may also be accompanied by misguided per- 1993; Gelman et al., 2014). This problem is especially relevant when
ceptions of reduced uncertainty in risk model predictions, referred to as characterizing infrequent events, such as severe hurricanes or earth-
“false precision.” (Small and Fischbeck, 1999; Bistline, 2015) A third quakes, since their rarity dictates that many historical records, even
case occurs when new information results in “disconcerting learning those of long duration, include only a small number of events for model
“(Hannart et al., 2013), where uncertainty about the science increases selection and fitting (Bier et al., 2004; Embrechts et al., 2013). Esti-
as a result of new or improved knowledge. As new scientific findings mation is made more problematic when methods for recording and
emerge experts may apply different weights to the conflicting studies. characterizing events change over the period of observation, and when
Even absent motivational bias (Montibeller and Winterfeldt, 2015), the underlying random processes exhibit uncertain trends or other

⁎
Corresponding author at: Engineering & Public Policy, Baker Hall 129, Frew Street, Carnegie Mellon University, Pittsburgh, PA, 15213, United States.
E-mail addresses: [email protected] (M.J. Small), [email protected] (S. Xian).

https://doi.org/10.1016/j.gloenvcha.2018.09.006
Received 9 May 2017; Received in revised form 4 September 2018; Accepted 16 September 2018
Available online 05 October 2018
0959-3780/ © 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/BY/4.0/).
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

nonstationary behavior (Katz et al., 2002; Tokdar et al., 2011). In ad- quantitative methods, including expert elicitation (Stiber et al., 1999;
dition, various biases affect how people view and perceive rare event Catenacci and Giupponi, 2013; O’Hagan et al., 2006) and statistical
occurrences (Tversky and Kahneman, 1975; Wachinger et al., 2013). To analysis of observed data or mechanistic models for event relationships
what degree might estimates based on either formal statistical methods (Barton et al., 2008; Xu et al., 2010; Yang et al., 2012). Various soft-
or empirical heuristics result in misleading inferences, with the poten- ware packages are available for building and implementing Bayesian
tial to support or encourage decisions that are later subject to regret? networks (Korb and Nicholson, 2003) and we utilize one such program,
Hurricane induced storm surge events (e.g. Hurricanes Katrina in Netica, in this study (Norsys, 2016). The executable version of the NYC
2005, Sandy in 2012 and Irma in 2017) have caused significant recent storm damage network model is available from the authors (execution
damage to coastal regions (Pistrika and Jonkman, 2010; Xian et al., will require downloading Netica from Norsys Software, most recent
2015; Hatzikyriakou et al., 2015; Xian et al., 2018b). Concurrently, price: $285 academic / $785 commercial).
scientific advances have improved our understanding of the physical The Bayesian network model developed in this study includes many
and human processes affecting event occurrence and damage, as well as of the physical and engineering science elements found in previous
methods for deciding among alternative mitigation options. Examples studies of coastal storm surge, but adds new dimensions to account for
include: an improved capability to detect changes in storm surge dis- information potentially gained from (imperfect) climatic risk studies
tributions (Resio et al., 2017; Rogers et al., 2018; Lin and Emanuel, and projections; the extent of dissemination of study results; and the
2016; Lin et al., 2016); advances in the ability to predict changes in distribution of perceived risks across decision makers. Physical risk and
mean sea level (Slangen et al., 2017; Hulbe, 2017; Oppenheimer and probability models are available for some of the model elements. For
Alley, 2016); and new methods for multiobjective robust decision other parts, especially involving information flow and human behavior,
making with uncertain future outcomes and learning (Schneider et al., predictive models are not yet available. For these elements we perform
2000; Mochizuki et al., 2017; Kim et al., 2017; Kwakkel et al., 2016). scenario and sensitivity analysis with the network to determine the
Many probabilistic risk assessment models have been developed and effect of different assumptions on predicted model outcomes.
applied to coastal storm surge flooding (e.g. Purvis et al., 2008; Aerts
et al., 2013, 2014; Lickley et al., 2014; Zwaneveld and Verweij, 2014; 2.1. Bayesian network model for New York City hurricane surge damage
Oddo et al., 2017). For the most part these models focus on normative and protection options
decision support, advancing methods for optimization of coastal miti-
gation projects. Examples include solutions that address uncertain The Bayesian network model for New York City (NYC) coastal
storm damage and dike performance (Slijkhuis et al., 1997) and the use protection decisions is summarized in Fig. 1. The model incorporates
of dynamic programming with uncertain future sea level rise information on historic hurricane frequency, simulated surge elevation,
(Brekelmans et al., 2012; van der Pol et al., 2014). In contrast, this resulting economic damage, and an uncertain range of future sea level
study incorporates behavioral elements of decision making that are and storm outcomes following a mitigation decision. Fig. 1 includes
“positive,” seeking to understand how people do behave, rather than inputs and summary results for variables with historic data (on the left
how they should behave under limited normative criteria. These ele- portion of the network) while results for nodes reflecting uncertain
ments have been considered in an increasing number of recent studies, decisions and outcomes are shown in the middle and right side of the
for example, searching for compromise solutions among multiple sta- network, including alternative outcomes that could occur. A graph of
keholders with different preferences and conflicting objectives (Oddo the full Netica network, including all computational nodes and variable
et al., 2017; Wong-Parodi et al., 2018) and consideration of behavioral states, is shown in Figure A1.1 in Appendix A1 of the Supplementary
responses by coastal decision makers confronted with alternative future Material.
climate change outcomes and events (Kwadijk et al., 2010; Haasnoot The major information sources and analyses underlying the network
et al., 2013). Like a number of recent studies addressing coastal or in- model include:
land flood protection, we adopt the avoidance of post-hoc regret from
either insufficient or unnecessary funds devoted to mitigation as a A A coupled hurricane-hydrodynamic-inundation model that simu-
central measure of outcome utility (Aissi et al., 2009; Rosner et al., lates Atlantic tropical storms and specifies their storm surge eleva-
2014; Butler et al., 2016; van der Pol et al., 2016; Casal-Campos et al., tions and inundation areas and depths across the NYC study area
2018). (Aerts et al., 2014).
B An economic damage model that calculates the damage to coastal
2. Bayesian networks housing and infrastructure for each storm simulated in part A.
C A damage function fitted to the results of Part A and B that predicts
The coupled human-environmental systems model developed in this the fraction of potentially impacted housing and infrastructure that
study is implemented in a Bayesian network. A Bayesian network is a is damaged by an event, as a function of its peak storm surge ele-
directed acyclic graph that connects a set of random variables (nodes), vation.
indicates the direction and magnitude of causal influence between these D A probabilistic representation of the number of events that occur
nodes, and allows derivation of joint and conditional probability dis- during a baseline assessment period (conditioned on the estimated
tributions for them. Direct relationships between “parent” nodes and baseline Poisson rate) and the associated cumulative damage from
their “child” nodes are captured by the conditional probability table these events (dependent on the storm surge heights and the damage
(CPT) for the latter. The CPT gives the probability of each possible state function).
in the child node for each possible combination of states in its parent E Sensitivity analyses for the effects of future climate change on event
nodes. Indirect influences between nodes are subsequently inferred frequency and reference elevation (sea level rise), the implications
through numerical propagation of Bayes Rule through the network. for storm damage, and the accuracy of scientific studies conducted
Bayes Rule is the fundamental equation of conditional probability by to predict these changes.
which prior probabilities of events are updated to posterior prob- F A probabilistic model of stakeholder and decision makers’ perceived
abilities given new observations or evidence (Lee, 2012; Stone, 2013) risks and selected levels of coastal protection (dependent on the
Bayesian networks have been used to characterize system un- observation period outcome in D and the results of the scientific
certainty and compare risk management options in a wide range of studies in E).
technology, risk, safety, health, and climate applications (e.g., Borsuk G A probabilistic characterization of the potential cumulative damage
et al., 2004; Landuyt et al., 2013; Richards et al., 2013). Parameter during the future outcome period, determining the likelihood of
(CPT) estimation may be approached using a number of qualitative or regret based on the probability that protection levels chosen in part

138
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

Fig. 1. Hurricane Damage and Protection Model (DFT refers to cumulative damage fraction over a time interval of duration T). Node inputs and statistics are shown
for current storm risk factors (on the left side of the network diagram).

F are under-, properly-, or overly-protective when compared to the mitigation option at the end of the 40-year observation period, decision
damage outcome that ensues during the outcome period. makers may have access to and decide to utilize scientific studies that
predict changes in sea level and event frequency.
To implement the framework, damage outcomes and mitigation
efforts sufficient to protect against them are discretized into three le-
2.2.1. Event frequency uncertainty
vels: low, medium and high. This discretization is illustrated in Fig. 1,
The uncertainty in the point estimate of the event frequency is
Node 14: Current DF(T = 40) Summary; and Node 35: Future DF
characterized using a Bayesian approach, adopting a gamma conjugate
(T = 40) Summary.
distribution for the Poisson rate λ (Lee, 2012). Bayesian parameter es-
timation begins with a subjective prior distribution that reflects the
2.2. Network model variables and influences analyst’s prior beliefs. This distribution is combined with the likelihood
function for the observed data to yield the posterior distribution of the
Damaging hurricanes are modeled to occur as a Poisson process uncertain parameter. These calculations are generally carried out using
with rate λ (yr−1). The number of events, y, that occur in a time period numerical methods or simulation procedures (such as Markov Chain
T (e.g., T = 40 yr) follows a Poisson distribution with parameter λT (= Monte Carlo), but in special cases where the chosen prior distribution
mean number of events). The Poisson model is fit and applied over has the form of a conjugate, the posterior distribution of the parameter
three time periods. The first is the full period of observation available maintains the same mathematical form as the prior, with updated
for the area of interest (T = 163 years). This is used to establish a parameters that are functions of both the prior parameters and appro-
baseline distribution of the current underlying event rate. The second priate statistics of the observed dataset. In the case of the uncertain
time interval is a recent assessment period (here assumed to be 40 parameter λ of a Poisson process, the conjugate distribution for λ is
years) during which decision makers and stakeholders observe storm given by a gamma distribution. In some cases (including here) the
outcomes and determine their beliefs regarding the level of current risk. parameters for the conjugate prior distribution are chosen to be diffuse
This assessment period is shorter than the historical period of record, or “informationless”, so that the prior distribution for the parameter is
reflecting the need for more detailed surge elevation and damage in- flat over the range of possible values. We use an informationless gamma
formation available only for more recent events, as well as reflecting prior distribution for λ and combine this with the observed sample
the tendency of stakeholders and decision makers to weigh their recent y = 55 and T = Thistorical = 163 yr, with the result that the posterior
experience more heavily (Atreya et al., 2013; Shao et al., 2017). At the distribution of λ is gamma distributed with parameters a = 55 and
end of this observation period decision makers choose a single miti- b = 1/T = 6.135 * 10−3 yr-1. Further details on the calculation of the
gation option which is assumed to be implemented without delay, re- conjugate gamma parameters are found in Appendix A. The resulting
ducing the risk of storm surge damage throughout the ensuing outcome uncertainty distribution for λ is computed and shown in discretized
period. While the future outcome duration need not be the same as that intervals in Fig. 1, Node 3: Current Event Rate (#/yr). Based on the
of the second assessment period, we again assume a period of 40 years 163-year record, the uncertainty in the value of λ extends from about
for ease of comparison. Storm surge elevation and event frequency may 0.25 to 0.45 yr-1.
change during the subsequent outcome period; these changes are as- Next an effective assessment period is assumed to occur in which
sumed to be driven by uncertain climate change. These effects are in- stakeholders and decision makers evaluate the hurricane risk. In Node
cluded in the model through consideration of alternative future sea 4: Design Period, this period is chosen to be 40 years. Given the epis-
level rise and event occurrence rates. Note that prior to choosing the temic uncertainty in the event rate for the NYC area, the overall

139
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

uncertainty (epistemic + aleatory, formally refered to as the predictive

distribution1) for the number of events that will occur in an assessment
period of T = 40 yr is calculated and displayed in Figure A1.1, Node 5:
No. Events in Design Period. As shown in Node 5 the predicted number
of events that could occur in the design period of 40 years ranges from
approximately 6 to 22 (encompassing about 94% of the outcome
probability), with an expected value of 13 events. As shown below,
depending on the magnitude of these events, this can motivate the
choice of very different levels of protection for the subsequent period.

2.2.2. Event surge height elevation

The next set of calculations, implemented in the bottom left portion
of the network, involves the application of a fitted distribution function
for the maximum storm surge height of each event, determination of
the relationship between surge height and the regional damage frac-
tion, and the use of this relationship to derive the probability dis-
tribution function of damage fraction for both individual events and
cumulatively, for all events that occur during the future outcome period
Fig. 2. Comparison of empirical and fitted shifted exponential distribution for
(see Fig. 1, Node 13: Cumulative DF(T = 40 yr)).
maximum event storm surge elevation simulated for New York City.
The model of Lin et al. (2012) was used to simulate a set of re-
presentative storm events passing within 200 km of NYC and impacting
the study area. The 549 simulated storms, their tracks, and associated A dataset of the maximum surge height and the calculated damage for
storm surge elevations cover a range from extremely low-probability/ each of the 549 simulated events is provided in Appendix B.
high-severity events to more frequent storms. The distribution of the Given the surge height and the damage fraction for each event, a
maximum surge height, S, associated with each event was fit using a polynomial damage fraction relationship was fit as follows:
shifted exponential distribution with probability density function DF = g XSh (2)
fS (s ) = exp[ (s-Smin )]; s Smin where XS = S – Smin and the parameters g and h are estimated using
=0; otherwise (1) nonlinear least-squares regression. Fig. 3 compares simulated values of
surge height and damage fraction to that predicted by the polynomial
where Smin is the location (shift) parameter, corresponding to the
relationship with estimated parameters: g = 8.61 × 10−4 and
minimum storm surge height, below which no significant damage oc-
h = 2.491. As indicated, the fitted relationship provides a good overall
curs, and β is the exponential parameter (m−1). The parameters were
representation for the simulated storms, especially for the high surge –
estimated using nonlinear least-squares cdf regression with the 549
high damage storms that matter most. A fair amount of random scatter
simulated storm surge elevations, yielding Smin = 0.901 m, and
is nonetheless evident for the predicted damage fraction, especially for
β = 1.553 m−1. The fitted and observed cumulative distribution
intermediate to large storms with storm surge elevations ranging from 2
functions of S are compared in Fig. 2, indicating that the shifted ex-
to 3.5 m above the mean sea level (MSL). This variability is not sur-
ponential provides an excellent fit to the simulated storm surge eleva-
prising as different storms with similar maximum surge heights could
tions. This particular parametric form for the storm surge height has the
occur with very different tracks, onshore resources at risk, and in-
added benefit that it allows the model for individual storms to be di-
undation profiles across the NYC area.
rectly aggregated for calculation of the statistics of cumulative damage
over a fixed time interval (e.g., 40 years) during which multiple events
are expected to occur (see below and Appendix A). 2.2.4. Derived distribution of event and cumulative damage fraction
Given the random storm surge height for each event (Eq. (1) and
Fig. 2) and the fitted damage relationship (Eq. (2) and Fig. 3), the
2.2.3. Storm surge damage function
distribution of the event damage fraction for a single storm i, DFi, can
For each of the 549 events simulated by Lin et al. (2012), economic
be derived. However, as now demonstrated, only the moments of DFi
damage in the NYC area was computed based on the extent of in-
are needed to derive an approximate solution for the distribution of the
undation and the predicted building losses. Flood inundation-depth
key output variable of the model, the cumulative damage fraction, DFT,
maps for the synthetic storms take into account the flood hazard and
over a time period T. DFT is simply the sum of the individual event
topographic elevation (Aerts et al., 2014). The damage associated with
damage fractions for the NT events that occur during T:
each inundation scenario was then calculated using the building in-
formation in the MapPLUTO database from the NYC Department of City DFT =
NT
DFi
Planning and the damage-water depth function of each building type i=1 (3)
from FEMA’s HAZUS model. Infrastructure damages were added using a As shown in Appendix A, DFT is a compound Poisson process and its
scale-up factor following Aerts et al., 2014. The damage estimates were mean and variance can be directly computed from the fitted parameters
next converted into a fraction of the NYC housing and infrastructure of the storm surge distribution and the surge height – damage re-
damaged by each event. The damage fraction, DF, is defined as the loss lationship:
divided by the total exposure of the NYC area. The total exposure of
NYC was computed based on previous studies (Aerts et al., 2013, 2014). E [DFT ] = T g (h + 1)/ h
(4)

Var [DFT ] = T g (2h + 1)/ 2h

(5)
1
Note that the predictive distribution for the Poisson-gamma conjugate can
also be derived analytically as a negative binomial distribution. This model is where Γ() is the gamma function.
equivalent to that obtained in our network where the rate of the Poisson dis- The distribution function of DFT is more difficult to derive and ex-
tribution is first sampled from the conjugate gamma distribution, then the press in simple form. When λT is small there is a significant probability
number of events in the predictive time interval is sampled from a Poisson with of zero events, yielding zero cumulative damage, so that the compound
the sampled rate. Poisson process DFT is a mixed random variable with probability mass

140
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

results in a cdf that is quite similar to the empirical bootstrap, though

the former slightly overestimates the probability of very small cumu-
lative damage fractions (over a 40-year period) relative to the boot-
strap.

2.2.5. Climate change factors

Recent studies have identified a warming climate as predictive of
increases in cyclonic storm surge frequency and severity (Grinsted
et al., 2013; Lee et al., 2017). Other studies (see Appendix A4) have
yielded inconsistent predictions for the effects of climatological change
on storm frequency, with results sensitive to the particular climate
models used. Given the high degree of uncertainty regarding changes in
future storm frequency and the relatively short timeframe of our
modeled projection period (40 years), we explore the potential impact
of changes in the storm occurrence rate by considering two scenarios in
Node 15, Event Rate Multiplier: i) no change; and ii) a 20% increase in
the occurrence rate, representative of the range of outcomes predicted
by Emanuel (2013).
Fig. 3. Simulated and fitted relationship between event surge height and da- Multiple recent studies have addressed potential sea-level rise
mage fraction. globally and for the NYC area, and a number of probability estimates
have been derived as a function of climate change into the future (ty-
pically until 2100). However, given the rapidly shifting estimates in
at DFT = 0, and continuous probability density for DFT > 0. When λT is
recent years, due in part to the high uncertainty associated with pos-
large, many events are expected to occur so that the Central Limit
sible loss of the Antarctic ice sheet (Kopp et al., 2014; Deconto and
Theorem can be applied to Eq. (3) and the distribution of DFT converges
Pollard, 2016; Oppenheimer and Alley, 2016) we elected to refrain
to a normal. For intermediate values of λT, a distribution that provides
from choosing from among available uncertainty estimates for future
a good empirical approximation to simulated distributions may be
SLR. Instead we consider the three values shown in Node 21 Sea Level
chosen, and the gamma distribution provides such an approximation
Rise: i) no change; ii) 20 cm (slow) increase; and iii) a 50 cm (rapid)
(Bagchi et al., 1984; Dominey and Hill, 2004; Ramaswami et al., 2005;
increase as alternatives for sensitivity analysis. Consideration of sea-
Husak et al., 2007; Mikosch, 2009; Truong and Trück, 2016).
level rise in the damage fraction calculation requires further mod-
To check the adequacy of the gamma approximation for DFT, the
ifications to the equations used to predict its moments and gamma
distribution was derived using the fitted parameters presented above.
parameters, as shown in Appendix A4.
Assuming a 40-year time period with λ = 0.337 events/yr, T = 40 yr,
The results of scientific studies to predict whether there will be an
g = 8.61 × 10−4, h = 2.491, and β = 1.553 m-1, the mean and stan-
increase in storm frequency (in this case, by 20 percent) and/or an
dard deviation of DFT are calculated from Eqs. (4) and (5) as: E
increase in baseline sea level (of 0, 20 or 50 cm) are depicted in Fig. 1 in
[DFT] = 0.0128; and StdDev[DFT] = Var1/2[DFT] = 0.0114, respec-
Nodes 20 Rate Multiplier Study and 23 SLR Study, respectively. The
tively. The corresponding parameters of the gamma distribution are
accuracies of these studies are specified in Node 19 Quality of Event
given by a = 1.2538 and b = 98.2. The resulting cumulative distribu-
Rate Study and Node 22 Quality of SLR Study. When the storm fre-
tion function for DFT = 40 yr is plotted in Fig. 4 as the derived gamma
quency studies are weak/good/perfect, we assume that they yield the
cdf (solid black line). The derived gamma distribution is compared to
correct prediction for the rate multiplier (either 1.0 or 1.2) with
an empirical bootstrap for 100 simulated 40-yr periods in which the
probability 0.6/0.8/1.0. That is, a perfect study always indicates the
number of storm events, N40, for each period is sampled from a Poisson
correct result, a good study does so with probability 0.8, and a weak
distribution with parameter λT= 0.337*40 = 13.48. The damage
study has a 60% chance of yielding the correct answer. For the sea-level
fractions for the individual events in each 40-year storm set are then
rise prediction, the weak/good/perfect study is again assumed to yield
generated by a bootstrap simulation from the 549 storms modeled by
the correct prediction (sea-level rise = 0, 20, or 50 cm) with respective
Lin et al., then summed to compute the cumulative damage fraction for
probabilities 0.6/0.8/1.0, with the remaining probability distributed
each 40-year sample. As indicated in Fig. 4, the gamma approximation
among the other two outcomes. These study accuracy assumptions may
be modified as part of a broader sensitivity and value-of-information
analysis to determine the relative importance of better climate fore-
casts.

2.2.6. Human behavioral components

The progression from stakeholder knowledge to risk perception,
decision making, and the evaluation of regret is addressed in a number
of steps in the network model, including: the option to conduct addi-
tional scientific studies to reduce uncertainty in future storm event
frequency and sea level rise ; deciding whether to access and consider
the results of these scientific studies in mitigation decisions ; the miti-
gation decision that follows; and, depending on the storm events that
ensue, possible regret in the mitigation decision due to under- or
overprotection. While the physical components of the overall model are
Fig. 4. Estimated distribution of cumulative storm damage fraction in a 40-yr highly idealized, they are nonetheless based on fundamental physical
period under best-estimate current conditions, comparing derived gamma dis- models and/or empirical data. In contrast, general theories and ob-
tribution with a bootstrap simulation using a Poisson number of events and servational studies for the behavioral elements and decisions have only
each event’s damage fraction sampled from the set of 549 storm surge events begun to emerge in recent years. In this study only simple behavioral
modeled by Lin et al. (2012). options are considered, exploring their importance through integration

141
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

Table 1 Table 2
Assumed Levels of Hurricane Protection Chosen (last column) as a Function of Determination of the Assumed Degree of Under- or Overprotection for the
Its Parent Nodes (first four columns). For the Rate Multiplier and the SLR Study Future Outcome based on the Level of Hurricane Protection Chosen at the be-
the values listed indicate the predicted future values from the scientific studies. ginning of the 40-yr Outcome Period and the Subsequent Cumulative Damage
(Selections are logically consistent, though illustrative for the model results that Outcome (estimated absent mitigation) during this Period.
follow.).
Level of Hurricane Protection Chosen
Node 14 Node 24 Node 20 Node 23 Node 36 (at beginning of 40-yr outcome period, Node 36)
Current DFT Climate Studies Rate SLR Hurricane
Summary Considered? Multiplier Study Protection Low Medium High
(cm) Level Chosen
Cumulative Low: Proper Overprotective Very
DFT < 0.5% No – – Low Damage DFT < Overprotective
0.5% < DFT No – – Medium Fraction in 0.5%
< 2.0% Future Medium: Underprotective Proper Overprotective
DFT > 2.0% No – – High Outcome 0.5% <
Low Recent Damage Period DFT <
DFT < 0.5% Yes 1 0 Low (T = 40- 2%
DFT < 0.5% Yes 1 20 Low yr, High: Very Underprotective Proper
DFT < 0.5% Yes 1 50 Medium estimated DFT > Underprotective
DFT < 0.5% Yes 1.2 0 Low absent 2%
DFT < 0.5% Yes 1.2 20 Medium mitigation,
DFT < 0.5% Yes 1.2 50 High Node 35)
Medium Recent Damage
0.5% < DFT Yes 1 0 Low
< 2.0% frequency and sea level rise. As new hurricane mitigation decisions are
0.5% < DFT Yes 1 20 Medium
made, studied and analyzed for coastal areas, such data may allow
< 2.0%
0.5% < DFT Yes 1 50 Medium better-supported parameterizations and estimates for which Hurricane
< 2.0% Protection Level decisions are preferred and implemented by in-
0.5% < DFT Yes 1.2 0 Medium dividuals and communities. The assumed influences and options shown
< 2.0%
in Table 1 could provide an initial framework for structuring such
0.5% < DFT Yes 1.2 20 High
< 2.0%
studies.
0.5% < DFT Yes 1.2 50 High The unprotected cumulative storm damage estimate for the future
< 2.0% outcome period is aggregated into three categories in Node 35, corre-
High Recent Damage sponding to low (DFT < 0.5%), medium (0.5% < DFT < 2%) and high
DFT > 2.0% Yes 1 0 Medium
(DFT > 2%) cumulative damage. The DFT outcome is then compared to
DFT > 2.0% Yes 1 20 High
DFT > 2.0% Yes 1 50 High the level of hurricane protection implemented at the beginning of the
DFT > 2.0% Yes 1.2 0 Medium 40-yr outcome period (Table 1, Node 36) to determine the extent of
DFT > 2.0% Yes 1.2 20 High under-, proper-, or over protection in Nodes 37: Appropriate Protec-
DFT > 2.0% Yes 1.2 50 High tion? and 38: Protection Summary. Both of these nodes assign the state
Proper when the unmitigated damage fraction matches the protection
level chosen at the end of the assessment period. The full set of as-
in the overall physical-social assessment model. To better bridge the
signments for Node 37 is provided in Table 2, indicating the level of
gap between scientific studies and decision makers’ knowledge we hope
regret assumed to result from different levels of mitigation effort im-
in future studies to incorporate the growing set of findings on stake-
plemented at the beginning of the 40-yr outcome period, compared to
holder engagement, risk perception and improved risk communication
the amount of storm damage that enfolds during this period (as esti-
(e.g., Jungermann et al., 1996; NRC (National Research Council), 1996;
mated absent mitigation).
Feldman and Ingram, 2009; Gaillard and Mercer, 2013; Wong-Parodi
As indicated, a Very Underprotective outcome results from Low
et al., 2014, 2018; DeLorme et al., 2018; Maskrey et al., 2018).
Hurricane Protection and a subsequent high cumulative damage frac-
While many of the nodes in the network use probabilistic CPTs to
tion during the future outcome period, while Very Overprotective is
predict the probability of child node states conditioned on those of their
assessed when a high level of protection is implemented, but the sub-
parent nodes, some employ deterministic decision rules for initial, il-
sequent estimated damage (absent mitigation) is low. To facilitate
lustrative estimates. We illustrate this for two of the key behavioral
summary reports for the model, Node 38 combines the Very
variables in the model: Node 36 for the Hurricane Protection mitigation
Underprotective and Underprotective outcomes into a single
decision, and Node 37 for the Appropriate Protection outcome. The
Underprotective category, and similarly for the Very Overprotective
child node representing the choice of the level of Hurricane Protection
and Overprotective outcomes. The degree of under- or overprotection
(Node 36) has four parents: Node 14 DFT (cumulative recent damage);
selected in Table 2 and Node 37 determines the level of decision regret
Node 24 Consider Climate Studies?; Node 20 Rate Multiplier Study
that decision makers are assumed to incur.
Results; and Node 23 SLR Study Results. An assumed set of selection
rules for Node 36 is summarized in Table 1. When scientific climate
3. Results
studies are either unavailable or not utilized by decision makers (Node
24 = No), the level of mitigation is assumed to depend solely on the
To highlight the type of insights that can be provided by the model,
recent cumulative damage, with low, medium or high values of DFT
a number of alternative scenarios are explored. These include:
during the assessment period motivating low, medium or high levels of
mitigation, respectively. This corresponds to the first three rows of
i A demonstration of how the event rate multiplier and different le-
Table 1. The remainder of Table 1 addresses cases where the scientific
vels of SLR affect the expected cumulative damage in the future; and
climate studies are available and considered by decision makers. As
ii An illustration of the effect that scientific studies of different quality
indicated the selection of higher levels of hurricane protection becomes
and subsequent utilization have on the likelihood of outcomes with
more likely in all cases as the recent damage experience increases, and
under-, proper-, or over-protection in the future period.
moreso following scientific studies that infer higher future event

142
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

Fig. 6. Influence of scientific studies on the probability of outcomes with dif-

ferent degrees of under-, proper, and over-protection.

more likely to occur, with a 12 and 34 percent chance of a very un-

Fig. 5. Effect of increased storm frequency and sea level rise on the expected derprotective or an underprotective outcome, respectively. These
value and standard deviation of the 40-year cumulative damage fraction (with numbers drop considerably when scientific projections are considered,
no mitigation).
more so for perfect and good studies, but even for weak studies with an
accuracy of 60 percent. Concomitantly, the probability of proper pro-
3.1. Effect of future change in storm frequency and sea level rise on storm tection increases from 35 percent (no studies considered) to 47 percent
surge damage (perfect studies). The results in Fig. 6 thus indicate that the probability
of regret can be reduced significantly but not eliminated, even with
The effects of storm frequency and sea level rise on cumulative perfect studies. This suggest that a coordinated set of scientific, en-
storm damage are summarized in Fig. 5. An increase in mean sea-level gineering and risk communication programs is needed to develop the
of 50 cm causes the mean of the cumulative damage fraction to ap- information needed for coastal mitigation decision support and to en-
proximately double from 0.013 to 0.028 for the case where there is no sure that it is accessible, clear and compelling.
increase in storm frequency, and from 0.017 to 0.033 for the case where
λ increases by 20 percent. The standard deviation of the cumulative
4. Implications and discussion
damage fraction also increases with sea-level rise and increased storm
frequency, but proportionally less so than the mean. While these
The analyses presented herein are highly idealized and illustrative,
changes are clearly important, what if they were anticipated by scien-
demonstrating how the physical, social and scientific components of
tific studies of different accuracy? If these studies were available and
hazard assessment and management behave in a coupled manner. As
considered, how might they affect the likelihood of different protection
illustrated in the network for New York City storm surge events, data
decisions being made and the likelihood that they are properly-, under-,
and models are available for key aspects of the system, especially those
or over-protective?
involving the physical and engineering sciences. In contrast, human
behavioral models for the performance and utilization of scientific
3.2. Potential effect of scientific studies on the likelihood of low-regret studies and the selection of mitigation levels are more speculative. For
outcomes these elements, we hope that this and related models might help to
provide a platform for integrating recent economic and behavioral re-
As noted above, the storm frequency and sea-level rise studies are search involving flood risk communication, perception, insurance, and
assumed to yield increasingly accurate predictions as their quality im- mitigation (e.g., Ge et al., 2011; Lindell and Hwang, 2008; Lindell and
proves from weak (60% chance of yielding the correct prediction), to Perry, 2012; Lo, 2013; Shao et al., 2016, 2017; Haer et al., 2016; Xian
good (80% chance of correct prediction), to perfect (100% chance of et al., 2017; Aerts and Botzen, 2011; Czajkowski et al., 2013; Wong-
correct prediction). Decision makers may have a sense of the study Parodi et al., 2018). Our results suggest that further studies of risk
quality and accuracy, especially if these are communicated clearly by perception, communication and behavior are needed to improve the
participating scientists, and this may help to inform the extent to which knowledge base for interdisciplinary systems modeling and behavio-
they consider the climate predicitions. rally-informed mitigation efforts. Nonetheless, the high variability of
Decision makers are assumed to base their protection decisions on a coastal storm outcomes, the imperfect accuracy of scientific studies,
combination of evidence from their previous experience and the sci- and the tenuous nature of stakeholder access and trust in these studies
entific studies. Access and trust in the studies is needed for them to be dictate that some level of regret will continue to be experienced. Con-
considered on an equal footing with recent experience, otherwise only currently, the importance, interest, and interdisciplinary nature of risk
the recent cumulative damage informs decision makers' risk beliefs and management for extreme events should continue to attract a high level
their subsequent choice of mitigation levels. Once their protection de- of effort through the development of integrative studies.
cision is made and subsequently compared to the storm severity and
potential damage that occurs in the ensuing period, the degree to which 5. Conclusions
the protection decision was proper is assessed.
The resulting probabilities of the different under-, proper, and over- In this study, a general Bayesian network model is developed to
protection outcomes for different cases of scientific accuracy and in- explore coastal storm mitigation outcomes for a system with highly
fluence tested in this study are shown in Fig. 6. When the studies are uncertain storm characteristics, imperfect scientific studies that predict
ignored (white bars in Fig. 6), outcomes with the highest regret are future storm risks, and differing degrees of trust in, and access to, these

143
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

scientific studies. The method includes statistical models for event oc- Bouwer, L.M., Jonkman, S.N., 2018. Global mortality from storm surges is decreasing.
currence and magnitude, epistemic uncertainty in the parameters of Environ. Res. Lett. 13 (1) p.014008.
Brekelmans, R., Hertog, D.D., Roos, K., Eijgenraam, C., 2012. Safe dike heights at minimal
these models, and aleatory uncertainty regarding the sequence of costs: the nonhomogeneous case. Oper. Res. 60 (6), 1342–1355.
events that occurs in any given period of time. Depending on the Brody, S.D., 2003. Are we learning to make better plans? A longitudinal analysis of plan
physical storm, damage, and behavioral parameters that characterize quality associated with natural hazards. J. Plan. Educ. Res. 23 (2), 191–201.
Brody, S.D., Zahran, S., Highfield, W.E., Bernhardt, S.P., Vedlitz, A., 2009. Policy learning
an area, significant changes are predicted in the likelihood of decisions for flood mitigation: a longitudinal assessment of the community rating system in
that yield regret as a result of under- or overprotection for the storm Florida. Risk Anal. 29 (6), 912–929.
sequence that ensues. Butler, W.H., Deyle, R.E., Mutnansky, C., 2016. Low-regrets incrementalism: land use
planning adaptation to accelerating sea level rise in Florida’s Coastal Communities. J.
The model also considers possibilities for nonstationary system be- Plan. Educ. Res. 36 (3), 319–332.
havior that further complicate the risk management decision process. In Casal-Campos, A., Sadr, S., Fu, G., Butler, D., 2018. Reliable, resilient and sustainable
the application to NYC storm surge protection decisions, these include urban drainage systems: an analysis of robustness under deep uncertainty. Environ.
Sci. Technol. 52 (16), 9008–9021.
possible increases in the frequency of events and sea-level rise, both of
Catenacci, M., Giupponi, C., 2013. Integrated assessment of sea-level rise adaptation
which result in increased cumulative damage. A number of recent and strategies using a Bayesian decision network approach. Environ. Model. Softw. 44,
ongoing research projects are attempting to obtain improved forecasts 87–100.
for these coastal conditions. Scientific studies that can further reduce Cox, D.R., Hinkley, D.V., 1979. Theoretical Statistics. CRC Press.
Czajkowski, J., Kunreuther, H., Michel-Kerjan, E., 2013. Quantifying riverine and storm-
the uncertainty in future climate-related trends are needed. Such stu- surge flood risk by single-family residence: application to Texas. Risk Anal. 33 (12),
dies can help to reduce the probability of regret resulting from under- or 2092–2110.
overly-protective mitigation, moreso if they are accurate, but only if DeConto, R.M., Pollard, D., 2016. Contribution of Antarctica to past and future sea-level
rise. Nature 531 (7596), 591.
decision makers have access to, understand, trust and assimilate the DeLorme, D.E., Stephens, S.H., Hagen, S.C., 2018. Transdisciplinary sea level rise risk
study results. communication and outreach strategies from stakeholder focus groups. J. Environ.
Stud. Sci. 8 (1), 13–21.
Dominey, M.J.G., Hill, R.M., 2004. Performance of approximations for compound Poisson
Acknowledgements distributed demand in the newsboy problem. Int. J. Prod. Econ. 92 (2), 145–155.
Emanuel, K.A., 2013. Downscaling CMIP5 climate models shows increased tropical cy-
We would like to thank Ning Lin for providing the simulated storm clone activity over the 21st century. Proc. Natl. Acad. Sci. 110 (30), 12219–12224.
Embrechts, P., Klüppelberg, C., Mikosch, T., 2013. Modelling Extremal Events: For
surge data, Jeroen Aerts for the NYC inundation maps, and the three Insurance and Finance Vol. 33 Springer Science & Business Media.
anonymous reviewers for their numerous constructive comments. The Feldman, D.L., Ingram, H.M., 2009. Making science useful to decision makers: climate
paper benefited from helpful discussions with Ning Lin, Kelly Klima, forecasts, water management, and knowledge networks. Weather. Clim. Soc. 1 (1),
9–21.
Howard Kunreuther, Michael Oppenheimer and Gabriel Wong-Parodi.
Gaillard, J.C., Mercer, J., 2013. From knowledge to action: bridging gaps in disaster risk
M. Small was supported in 2015-2016 by the Anderson Family Visiting reduction. Prog. Hum. Geogr. 37 (1), 93–114.
Professorship in Energy and the Environment in the Department of Civil Ge, Y., Peacock, W.G., Lindell, M.K., 2011. Florida households’ expected responses to
and Environmental Engineering and the Andlinger Center for Energy hurricane hazard mitigation incentives. Risk Anal. 31 (10), 1676–1691.
Geisser, S., 1993. Predictive Inference Vol. 55 CRC press.
and the Environment, Princeton University. Continued support was Gelman, A., Carlin, J.B., Stern, H.S., Rubin, D.B., 2014. Bayesian Data Analysis Vol. 2
provided by the NSF Center for Climate and Energy Decision Making Chapman & Hall/CRC, Boca Raton, FL, USA.
(CEDM) at Carnegie Mellon University (SES-1463492). S. Xian was Grinsted, A., Moore, J.C., Jevrejeva, S., 2013. Projected Atlantic hurricane surge threat
from rising temperatures. Proc. Natl. Acad. Sci. U. S. A. 110 (14), 5369–5373.
supported by National Science Foundation (NSF) grant: EAR-1520683. Haasnoot, M., Kwakkel, J.H., Walker, W.E., ter Maat, J., 2013. Dynamic adaptive policy
pathways: a method for crafting robust decisions for a deeply uncertain world. Glob.
Appendix A. Supplementary data Environ. Change 23 (2), 485–498.
Haer, T., Botzen, W.W., Aerts, J.C., 2016. The effectiveness of flood risk communication
strategies and the influence of social networks—insights from an agent-based model.
Supplementary data associated with this article can be found, in the Environ. Sci. Policy 60, 44–52.
online version, at https://doi.org/10.1016/j.gloenvcha.2018.09.006. Hannart, A., Ghil, M., Dufresne, J.L., Naveau, P., 2013. Disconcerting learning on climate
sensitivity and the uncertain future of uncertainty. Clim. Change 119 (3-4), 585–601.
Hatzikyriakou, A., Lin, N., Gong, J., Xian, S., Hu, X., Kennedy, A., 2015. Component-
References based vulnerability analysis for residential structures subjected to storm surge impact
from Hurricane Sandy. Nat. Hazards Rev. 17 (1), 05015005.
Hulbe, C., 2017. Is ice sheet collapse in West Antarctica unstoppable? Science 356 (6341),
Aerts, J.C., Botzen, W.W., 2011. Climate change impacts on pricing long-term flood in-
910–911.
surance: a comprehensive study for the Netherlands. Glob. Environ. Change Part A 21
Husak, G.J., Michaelsen, J., Funk, C., 2007. Use of the gamma distribution to represent
(3), 1045–1060.
monthly rainfall in Africa for drought monitoring applications. Int. J. Climatol. 27
Aerts, J.C., Lin, N., Botzen, W., Emanuel, K., de Moel, H., 2013. Low-probability flood risk
(7), 935–944.
modeling for New York City. Risk Anal. 33 (5), 772–788.
Jongman, B., Winsemius, H.C., Aerts, J.C., de Perez, E.C., van Aalst, M.K., Kron, W.,
Aerts, J.C., Botzen, W.W., Emanuel, K., Lin, N., de Moel, H., Michel-Kerjan, E.O., 2014.
Ward, P.J., 2015. Declining vulnerability to river floods and the global benefits of
Evaluating flood resilience strategies for coastal megacities. Science 344 (6183),
adaptation. Proc. Natl. Acad. Sci 201414439.
473–475.
Jungermann, H., Pfister, H.-R., Fischer, K., 1996. Credibility, information preferences,
Aissi, H., Bazgan, C., Vanderpooten, D., 2009. Min–max and min–max regret versions of
and information interests. Risk Anal. 16 (2), 251–261.
combinatorial optimization problems: a survey. Eur. J. Oper. Res. 197 (2), 427–438.
Katz, R.W., Parlange, M.B., Naveau, P., 2002. Statistics of extremes in hydrology. Adv.
Albright, E.A., Crow, D.A., 2015. Learning processes, public and stakeholder engagement:
Water Resour. 25 (8), 1287–1304.
analyzing responses to Colorado’s extreme flood events of 2013. Urban Climate 14,
Kim, Y., Eisenberg, D.A., Bondank, E.N., Chester, M.V., Mascaro, G., Underwood, B.S.,
79–93.
2017. Fail-safe and safe-to-fail adaptation: decision-making for urban flooding under
Atreya, A., Ferreira, S., Kriesel, W., 2013. Forgetting the flood? An analysis of the flood
climate change. Clim. Change 145 (3-4), 397–412.
risk discount over time. Land Econ. 89 (4), 577–596.
Kopp, R.E., Horton, R.M., Little, C.M., Mitrovica, J.X., Oppenheimer, M., Rasmussen, D.J.,
Bagchi, U., Hayya, J.C., Ord, J.K., 1984. Modeling demand during lead time. Decis. Sci.
Tebaldi, C., 2014. Probabilistic 21st and 22nd century sea-level projections at a
15 (2), 157–176.
global network of tide-gauge sites. Earth's Future 2 (8), 383–406.
Barton, D.N., Saloranta, T., Moe, S.J., Eggestad, H.O., Kuikka, S., 2008. Bayesian belief
Korb, K.B., Nicholson, A.E., 2003. Bayesian Artificial Intelligence. Chapman & Hall/CRC
networks as a meta-modelling tool in integrated river basin management—pros and
computer science and data analysis. http://www.csse.monash.edu.au/bai/book1e/.
cons in evaluating nutrient abatement decisions under uncertainty in a Norwegian
Kwadijk, J.C., Haasnoot, M., Mulder, J.P., Hoogvliet, M., Jeuken, A., van der Krogt, R.A.,
river basin. Ecol. Econ. 66 (1), 91–104.
van Oostrom, N.G., Schelfhout, H.A., van Velzen, E.H., van Waveren, H., de Wit, M.J.,
Bier, V., Ferson, S., Haimes, Y.Y., Lambert, J.H., Small, M.J., 2004. Risk Of Extreme And
2010. Using adaptation tipping points to prepare for climate change and sea level
Rare Events: Lessons From A Selection Of Approaches. Risk Analysis And Society: An
rise: a case study in the Netherlands. Wiley Interdiscip. Rev. Clim. Change 1 (5),
Interdisciplinary Characterization Of The Field. pp. 74–118.
729–740.
Bistline, J.E., 2015. Fat-tailed uncertainty, learning, and climate policy. Clim. Chang.
Kwakkel, J.H., Haasnoot, M., Walker, W.E., 2016. Comparing robust decision-making and
Econ. 6 (02) p.1550009.
dynamic adaptive policy pathways for model-based decision support under deep
Borsuk, M.E., Stow, C.A., Reckhow, K.H., 2004. A Bayesian network of eutrophication
uncertainty. Environ. Model. Softw. 86, 168–183.
models for synthesis, prediction, and uncertainty analysis. Ecol. Modell. 173 (2),
Landuyt, D., Broekx, S., D’hondt, R., Engelen, G., Aertsens, J., Goethals, P.L., 2013. A
219–239.
review of Bayesian belief networks in ecosystem service modelling. Environ. Model.

144
M.J. Small, S. Xian Global Environmental Change 53 (2018) 137–145

Softw. 46, 1–11. Shao, W., Xian, S., Keim, B.D., Goidel, K., Lin, N., 2016. Understanding perceptions of
Lee, P.M., 2012. Bayesian Statistics: An Introduction. John Wiley & Sons. changing hurricane strength along the US Gulf coast. Int. J. Climatol. https://doi.org/
Lee, B.S., Haran, M., Keller, K., 2017. Multidecadal scale detection time for potentially 10.1002/joc.4805.
increasing atlantic storm surges in a warming climate. Geophys. Res. Lett. 44 (20), Shao, W., Xian, S., Lin, N., Kunreuther, H., Jackson, N., Goidel, K., 2017. Understanding
10617–10623. the effects of past flood events and perceived and estimated flood risks on individuals’
Lickley, M.J., Lin, N., Jacoby, H.D., 2014. Analysis of coastal protection under rising flood voluntary flood insurance purchase behavior. Water Res. 108, 391–400.
risk. Clim. Risk Manag. 6, 18–26. Slangen, A.B., Meyssignac, B., Agosta, C., Champollion, N., Church, J.A., Fettweis, X.,
Lin, N., Emanuel, K., 2016. Grey swan tropical cyclones. Nat. Clim. Chang. 6 (1), Ligtenberg, S.R., Marzeion, B., Melet, A., Palmer, M.D., Richter, K., 2017. Evaluating
106–111. model simulations of twentieth-century sea level rise. Part I: global mean sea level
Lin, N., Emanuel, K., Oppenheimer, M., Vanmarcke, E., 2012. Physically based assessment change. J. Clim. 30 (21), 8539–8563.
of hurricane surge threat under climate change. Nat. Clim. Chang. 2 (6), 462–467. Slijkhuis, K.A.H., Van Gelder, P.H.A.J.M., Vrijling, J.K., 1997. Optimal dike height under
Lin, N., Kopp, R.E., Horton, B.P., Donnelly, J.P., 2016. Hurricane Sandy’s flood frequency statistical-construction-and damage uncertainty. Struct. Saf. Reliabil. 7, 1137–1140.
increasing from year 1800 to 2100. Proc. Natl. Acad. Sci. 113 (43), 12071–12075. Small, M.J., Fischbeck, P.S., 1999. False precision in Bayesian updating with incomplete
Lindell, M.K., Hwang, S.N., 2008. Households’ perceived personal risk and responses in a models. Hum. Ecol. Risk Assess. 5 (2), 291–304.
multihazard environment. Risk Anal. 28 (2), 539–556. Small, M.J., Güvenç, Ü., DeKay, M.L., 2014. When can scientific studies promote con-
Lindell, M.K., Perry, R.W., 2012. The protective action decision model: theoretical sensus among conflicting stakeholders? Risk Anal. 34 (11), 1978–1994.
modifications and additional evidence. Risk Anal. 32 (4), 616–632. Stiber, N.A., Pantazidou, M., Small, M.J., 1999. Expert system methodology for evalu-
Lo, A.Y., 2013. The role of social norms in climate adaptation: mediating risk perception ating reductive dechlorination at TCE sites. Environ. Sci. Technol. 33 (17),
and flood insurance purchase. Glob. Environ. Chang. Part A 23 (5), 1249–1257. 3012–3020.
Mikosch, T., 2009. Non-Life Insurance Mathematics: An Introduction with the Poisson Stone, J.V., 2013. Bayes’ Rule: A Tutorial Introduction to Bayesian Analysis. Sebtel Press.
Process. Springer Science & Business Media. Tokdar, S., Grossmann, I., Kadane, J., Charest, A., Small, M., 2011. Impact of beliefs about
Mochizuki, J., Keating, A., Liu, W., Hochrainer-Stigler, S., Mechler, R., 2017. An overdue Atlantic tropical cyclone detection on conclusions about trends in tropical cyclone
alignment of risk and resilience? A conceptual contribution to community resilience. numbers. Bayesian Anal. 6 (4), 547–572.
Disasters. Truong, C., Trück, S., 2016. It’s not now or never: implications of investment timing and
Montibeller, G., Winterfeldt, D., 2015. Cognitive and motivational biases in decision and risk aversion on climate adaptation to extreme events. Eur. J. Oper. Res. 253 (3),
risk analysis. Risk Anal. 35 (7), 1230–1251. 856–868.
Norsys, 2016. Netica, Norsys Software Corporation. https://www.norsys.com/. Tversky, A., Kahneman, D., 1975. Judgment under uncertainty: heuristics and biases.
NRC (National Research Council), 1996. Understanding Risk: Informing Decisions in a Utility, Probability, and Human Decision Making. Springer, Netherlands, pp.
Democratic Society. National Academies Press. 141–162.
Oddo, P.C., Lee, B.S., Garner, G.G., Srikrishnan, V., Reed, P.M., Forest, C.E., Keller, K., van der Pol, T.D., van Ierland, E.C., Weikard, H.P., 2014. Optimal dike investments under
2017. Deep uncertainties in sea-level rise and storm surge projections: implications uncertainty and learning about increasing water levels. J. Flood Risk Manag. 7 (4),
for coastal flood risk management. Risk Anal. 308–318.
O’Hagan, A., Buck, C.E., Daneshkhah, A., Eiser, J.R., Garthwaite, P.H., Jenkinson, D.J., van der Pol, T.D., Gabbert, S., Weikard, H.P., van Ierland, E.C., Hendrix, E.M.T., 2016. A
et al., 2006. Uncertain Judgements: Eliciting Experts’ Probabilities. John Wiley & minimax regret analysis of flood risk management strategies under climate change
Sons. uncertainty and emerging information. Environ. Resour. Econ. 1–23.
Oppenheimer, M., Alley, R.B., 2016. How high will the seas rise? Science 354 (6318), Wachinger, G., Renn, O., Begg, C., Kuhlicke, C., 2013. The risk perception para-
1375–1377. dox—implications for governance and communication of natural hazards. Risk Anal.
Oppenheimer, M., O’Neill, B.C., Webster, M., 2008. Negative learning. Clim. Change 89 33 (6), 1049–1065.
(1-2), 155–172. Wong-Parodi, G., Fischhoff, B., Strauss, B., 2014. A method to evaluate the usability of
Pistrika, A.K., Jonkman, S.N., 2010. Damage to residential buildings due to flooding of interactive climate change impact decision aids. Clim. Change 126 (3-4), 485–493.
New Orleans after hurricane Katrina. Nat. Hazards 54 (2), 413–434. Wong-Parodi, G., Fischhoff, B., Strauss, B., 2018. Effect of risk and protective decision
Purvis, M.J., Bates, P.D., Hayes, C.M., 2008. A probabilistic methodology to estimate aids on flood preparation in vulnerable communities. Weather. Clim. Soc. 10 (3),
future coastal flood risk due to sea level rise. Coast. Eng. 55 (12), 1062–1073. 401–417.
Ramaswami, A., Milford, J.B., Small, M.J., 2005. Integrated Environmental Xian, S., Lin, N., Hatzikyriakou, A., 2015. Storm surge damage to residential areas: a
Modeling–pollutant Transport. Fate and Risk in the Environment. John Wiley & Sons quantitative analysis for Hurricane Sandy in comparison with FEMA flood map. Nat.
Inc. Hazards 79 (3), 1867–1888.
Resio, D.T., Asher, T.G., Irish, J.L., 2017. The effects of natural structure on estimated Xian, S., Lin, N., Kunreuther, H., 2017. Optimal house elevation for reducing flood-related
tropical cyclone surge extremes. Nat. Hazards 88 (3), 1609–1637. losses. J. Hydrol. 548, 63–74.
Reyers, B., Nel, J.L., O’Farrell, P.J., Sitas, N., Nel, D.C., 2015. Navigating complexity Xian, S., Yin, J., Lin, N., Oppenheimer, M., 2018a. Influence of risk factors and past events
through knowledge coproduction: mainstreaming ecosystem services into disaster on flood resilience in coastal megacities: comparative analysis of NYC and Shanghai.
risk reduction. Proc. Natl. Acad. Sci. 112 (24), 7362–7368. Sci. Total Environ. 610, 1251–1261.
Richards, R., Sanó, M., Roiko, A., Carter, R.W., Bussey, M., Matthews, J., Smith, T.F., Xian, S., Feng, K., Lin, N., Marsooli, R., Chavas, D., Chen, J., Hatzikyriakou, A., 2018b.
2013. Bayesian belief modeling of climate change impacts for informing regional Brief communication: rapid assessment of damaged residential buildings in the
adaptation options. Environ. Model. Softw. 44, 113–121. Florida keys after Hurricane Irma. Nat. Hazards Earth Syst. Sci. 18 (7), 2041–2045.
Rogers, R., Cheung, K., Elsberry, R., Kohno, N., Leroux, M.D., Otto, P., 2018. The world Xu, J., Small, M., Fischbeck, P., VanBriesen, J., 2010. Integrating location models with
meteorological organization’s fourth international workshop on tropical cyclone Bayesian analysis to inform decision making. J. Water Resour. Plan. Manag. 136 (2),
landfall processes (IWTCLP-IV): a summary. Trop. Cyclone Res. Rev. 7 (2) pp.http- 209–216.
tcrr. Yang, Y.-M., Small, M.J., Ogretim, E.O., Gray, Donald, D.O., Wells, A.W., Strazisar, B.R.,
Rosner, A., Vogel, R.M., Kirshen, P.H., 2014. A risk-based approach to flood management Bromhal, G.S., 2012. A Bayesian Belief Network for combining evidence from mul-
decisions in a nonstationary world. Water Resour. Res. 50 (3), 1928–1942. tiple CO2 leak detection technologies. Greenh. Gases Sci. Technol. 2 (3), 185–199.
Schneider, S.H., Easterling, W.E., Mearns, L.O., 2000. Adaptation: sensitivity to natural Zwaneveld, P., Verweij, G., 2014. Safe Dike Heights at Minimal Costs: An Integer
variability, agent assumptions and dynamic climate changes. Societal Adaptation to Programming Approach (No. 277). CPB Netherlands Bureau for Economic Policy
Climate Variability and Change. Springer, Dordrecht, pp. 203–221. Analysis.

145