Chapter 78

Risk, Reliability, and Return

Periods and Hydrologic Design
BY
RICHARD M. VOGEL, AND ATTILIO CASTELLARIN

ABSTRACT risks. We describe the various reasons why metrics, such as risk and reliabil-
ity over a planning horizon, may be more effective for communicating flood
The concepts of risk, reliability and return period are fundamental elements
risk than the traditional notion of an average return period (also see Read and
to the analysis of extreme hydrological events for the fields of water resource
Vogel, 2015; Serinaldi, 2015).
systems planning, and management as well as for flood and drought risk
Traditional probabilistic approaches for defining risk, reliability, and return
assessment and mitigation. This chapter reviews the main criteria for select-
periods under stationary hydrologic conditions assume that extreme events
ing the level of infrastructure protection and for defining hydrologic design
arise from a serially independent time series with a probability distribution
variables within a risk-based framework, considering both univariate and
whose moments and parameters are fixed over the design life of a project.
multivariate design conditions. Approaches to hydrologic design under both
Most existing hydrology texts and handbooks provide a review of hydrologic
stationary and nonstationary conditions are summarized.
design procedures under stationary conditions (see Stedinger et al., 1993;
78.1 INTRODUCTION CEH, 1999). Normally, hydrologic design is based on a single random vari-
able X, such as the annual maximum flood (AMF), or the annual minimum
The concepts of risk, reliability, and return periods are widely used in the 7-day discharge. Normally X is assumed to arise from a process which is
analysis of extreme events in the field of hydrology as well as for numerous independent from 1 year to the next with a probability distribution function
other natural hazards including wind loads, sea levels, earthquakes, tempera- (PDF) denoted by fx(x) with cumulative distribution function (CDF) denoted
tures, and other phenomena. For example, the U.S. National Flood Insurance by Fx(x). A stationary PDF is one whose model parameters are assumed fixed
Program defines the floodplain in terms of the T=100-year flood, which is over time. See Chapters 21 and 22 for definitions of commonly used PDFs
defined as the annual maximum river flood discharge (and associated flood and associated CDFs along with further information on how to fit a PDF and
elevation) that is exceeded with an annual exceedance probability (AEP) of CDF to observations.
1% (q=0.01). If flood infrastructure were designed to protect against such an Consider a hydrologic design problem in which a structure is built to pro-
event, the structure would be 99% reliable, in any given year. If river condi- tect against an extreme event with an annual non-exceedance probability,
tions are expected to remain unchanged (stationary) in the future, we show p=Fx(x). The design event for such a structure may be a river discharge for
later in this chapter that the reliability of that structure over a 50-year plan- flood or low flow control. Such a design discharge is also called the pth
ning horizon would only be 60.5%, which is much lower than 99%. It is quantile of X, denoted as xp, which is the value of X with non-exceedance
important to realize that the discharge with average return period T=100 probability p. In flood frequency analysis, one often estimates the design
years may arrive earlier (or later) than 100 years, because that is only the aver- event xp from a series of annual maximum streamflows (see Chapter 21) in
age return period, or the average arrival time of the next 100-year flood. which case q=1–p is referred to as the AEP. Choice of an appropriate design
Under stationary conditions, the return period is a random variable with an discharge forms the basis of hydrologic design. In sections 78.2.1 and 78.2.2,
exponential distribution, so that it is much more likely that the 100-year flood we describe two general approaches to selection of a design discharge: (1) the
will occur in the next 100 years, than in the subsequent 100 years, in fact there traditional approach which assumes an AEP equal to q=1–p, and then com-
is a 63.4% chance that the 100-year flood will arrive before the first 100 years putes the corresponding design discharge xp and (2) a risk-based decision
are over. Similar statements could be made for other natural hazards; how- approach which computes the value of xp as the discharge which maximizes
ever, the focus of this chapter is on statements of risk, reliability, and return the net benefits of the proposed water infrastructure.
periods relating to river discharge. The above statements concerning the
likelihood of future flood events are the type of concepts which hydrologists 78.1.2  Working Hypothesis: Stationarity versus
need to be familiar with and which form the basis of this chapter. Nonstationarity
Once a level of infrastructure is built to protect against a design event xp, in
78.1.1  Selecting the Level of Protection and Effective Risk any future year, a flood will either exceed that design event with probability
Communication: Reliability versus Average Return Period
q=1–p, or not, with probability p. If the flood series is independent in time
It is extremely important that hydrologists are able to communicate the prob- and stationary, then the series of exceedances and non exceedances is said to
ability of flood (and other natural) hazards in a manner which is clearly follow a Bernoulli series. Some hydrologic processes are known to exhibit
understood by those populations who will actually experience the impacts of temporal persistence such as low flow series (Douglas et al., 2002) and water
such hazards. For example, the U.S. Geological Survey (Holmes and Dinicola, supply failures (Vogel, 1987) both of which can be more accurately modeled
2010) and other agencies, issue fact sheets, general information announce- using a Markov model, instead of a Bernoulli model. Thus the values of AEP
ments, and videos which attempt to communicate and clarify the meaning of could change from 1 year to the next for a stationary Markov process.
the T-year flood. One of the goals of this chapter is to describe our current However, for a stationary Bernoulli process, the exceedance probability
understanding of the most effective approaches for communicating flood associated with the design discharge should remain constant over time.

78-1

78_Singh_ch78_p78.1-78.10.indd 1 01/07/16 4:28 PM

 78-2     Risk, Reliability, and Return Periods and Hydrologic Design

Historically, the most common assumption in flood and drought frequency 78.2.1  Existing Methods for Hydrologic Design Based
analysis has been that flood and drought series are temporally independent on Stationarity
and stationary, in which case the non-excedance probability p, and its con- Return Periods of Design Events under
verse the exceedance probability q=1–p are both constant, over time. Stationary Conditions
There is now widespread acceptance in the field of hydrology that most
Without loss of generality, we focus on the design of flood-risk mitigation
hydrologic processes exhibit nonstationary behavior due to changes in land-
measures; analogous concepts hold for other water-related problems and
use, climate, and water infrastructure. In spite of this nearly widespread
hydrological design variables. Although we begin our discussions with a
acceptance combined with the popular quotation that “Stationarity is Dead”
background on the use of return periods in hydrologic design, we emphasize
(Milly et al., 2008), there is still very good reason to employ traditional meth-
here and in subsequent sections that risk-based decision making (RBDM) is
ods based on stationary hydrologic processes as emphasized by Matalas
now a well-established methodology which can be used in place of the tradi-
(2012), Montanari and Koutsoyiannis (2014), and Vogel et al. (2015). This
tional design-event approach which selects a particular average T-year event
chapter documents approaches to hydrologic design problem considering
usually specified by regulation, and then designs the necessary infrastructure
both stationary and nonstationary conditions as well as univariate and multi-
to protect against the hydrological event with that specified average return
variate processes. We emphasize and present an integrated risk-based
period. We further emphasize here that, in spite of the widespread usage of
approach to hydrologic design, considering uncertainties arising from all
return period nomenclature in hydrology, there are good reasons to temper
relevant sources, which in turn enable hydrologists to consider both station-
their use as a metric of communicating flood risk (Read and Vogel, 2015; and
ary and nonstationary conditions together.
Serinaldi, 2015).
Assume a series of AMF discharges arise from a stationary process that is
78.1.3  Selecting the Hydrological Design Variable: independent in time; then in each year, an AMF exceeding some design
Univariate versus Multivariate threshold event xp will occur with probability q=1–p. Since in this case the
Many hydrologic design problems involve several random variables, thus a AMF process is stationary, we expect the exceedance probability of the design
focus on river discharge alone may not be sufficient. For example, in the event xp to be constant over time. The waiting time (also known as the return
design of stormwater, best management practices, such as detention basins, period) to the occurrence of the next AMF which exceeds xp is defined as τ ,
hydrologists are concerned with both hydrograph peaks and volumes. and follows a geometric probability mass function (PMF) so that:
Similarly modern levee design considers the probability of levee failure and t −1
P [τ = t ] = q (1 − q ) t=1, 2, … ,n (78.2.1)
overtopping and often uses several design variables in addition to peak dis-
charge (e.g., water levels and duration of the event, flow velocities, levee fra- where, τ and t are the theoretical and observed values of the waiting time
gility, etc.). A multivariate context arises when the hydrological load for the respectively, n is the planning horizon and q is the exceedance probability. In
structure of interest is a function of two or more hydrological variables that Eq. (78.2.1), the return period is assumed to be a discrete random variable,
are correlated with each other (e.g., flood peak and volume). resulting in a geometric PMF; if it were assumed to be a continuous variable
Such multivariate concerns and nonstationarity introduce additional the resulting PDF of the return period would be exponential. The expected
uncertainty into the process of decision-making. In reference to such multi- waiting time (or average return period) to the next AMF which exceeds xp is
variate hydrologic processes Yue and Rasmussen (2002) describe conditions given by:
in which a single-variable frequency analysis considering only flood peak or n
1
volume alone can lead to estimates of flood events, and associated costs, E [τ ] = ∑t ⋅ P [τ = t ] = = T (78.2.1)
which are much greater than necessary. The chapter ends by examining how t =1
q
existing approaches may be adapted for use under nonstationary conditions The average waiting or recurrence time T is often referred to, incorrectly, as
and in a multivariate context. the return period. In reality, T =1 q is the average return and τ is the return
period. Furthermore, we note that most hydrology texts do not give a deriva-
tion of why T =1 q nor do they distinguish between the properties of τ and T.
78.2  PROBABILISTIC AND RISK BASED APPROACHES TO Given the widespread usage of the relation T =1 q in hydrology, it is surpris-
HYDROLOGIC DESIGN ingly difficult to find examples of the derivation in Eqs. (78.2.1) and (78.2.2)
(Douglas Fernandez and Salas, 1999; Douglas et al., 2001; Wigley, 2009), in
This chapter focuses on the application of probabilistic and risk based spite of the fact that Fuller (1914) and Gumbel (1941) first introduced the idea
approaches to hydrologic design, yet it is important to realize that other non- of the average return period.
probabilistic approaches exist. For example, in the design of large dams, some There are two interpretations of the average return period T under station-
hydrologists still resort to the use of a deterministic interpretation of envelope ary conditions. As shown above, T is the average number of years one must
curves (see section 78.2.1). Castellarin et al. (2005) and Castellarin (2007) pro- wait until the occurrence of the next flood with exceedance probability q.
vide a probabilistic interpretation of envelope curves which Vogel et al. (2007) Cooley (2013) also shows that under stationary conditions, the expected
extended to enable probabilistic statements to be made regarding estimates of number of flood events (exceeding the AMF flood with exceedance probabil-
the probable maximum flood (PMF). Although deterministic and stochastic ity q) is equal to unity over the next T years.
approaches are available for estimation of extraordinary flood magnitudes The variance of the return period is also easily derived from:
[National Research Council (NRC) 1988], much greater attention has been
n
given to the development of deterministic methods, than stochastic methods, 1 p
Var [τ ] = E τ 2  − E [τ ] = ∑t 2 ⋅ P [τ = t ] − 2 = 2 (78.2.3)
2
for estimating extraordinary flood magnitudes including the probable maxi-
t =1
q q
mum flood and the probable maximum precipitation (PMP) (Cudworth, 1989).
Although previous hydrology research relating to hydrologic design For very rare floods, the non-exceedance probability p is roughly unity, which
assumes both a stationary past and future, there is now a nearly pervasive implies that Var [τ ] ≈ 1 q 2 = T 2, so that the mean and standard deviation of
awareness of potential nonstationarity in hydrologic processes due to anthro- the return period are roughly equal. Since the PMF of τ contains only a single
pogenic influences on climate, land use, and water infrastructure. The notion parameter, q, the mean of its distribution T =1 q is said to be a “sufficient”
that “stationary is dead” is now pervasive as indicated by over 1,000 citations, statistic for summarizing the complete probabilistic behavior of the return
to date, of Milly et al. (2008). We should not be so quick to dispense with the period. This fact is predicated upon the assumption that the values of p and
notion of stationarity given that to date, most of our water infrastructure was its complement q, do not change from 1 year to the next, which is only true
designed under the assumption of stationary conditions yet there have been under stationary conditions. Under nonstationary conditions, the average
“very few failures of the nation’s water management infrastructure—that is, return period is an extremely complex function of properties of the PDF of
where the infrastructure failed before its design capacity was exceeded” the flood magnitudes and the degree of nonstationarity so that T is no longer
(Stakhiv, 2011). Matalas (2012) provides ample reasons for questioning “the “sufficient” for describing the behavior of τ (see Read and Vogel, 2015;
degree to which real or perceived nonstationarities in hydrologic processes Serinaldi, 2015).
(should) affect the underlying processes and methods of making water plan-
ning and management decisions.” He argues that “the assumption of station- Return Periods for Dependent or Persistent
arity has not yet been pushed to the limit of its operational usefulness in the Hydrologic Events
face of a changing climate.” In the following section we outline methods of The average return period can be defined in different ways depending on
hydrologic planning under both stationary and nonstationary conditions. whether the hydrologic design depends on initial project conditions or not.
Our approach to nonstationarity involves adapting traditional risk-based Two definitions of return period are possible, the time between two hydro-
approaches which have served us well under stationary conditions. logic events of interest, and the unconditional time to the next event of

78_Singh_ch78_p78.1-78.10.indd 2 01/07/16 4:29 PM

 Probabilistic and Risk Based Approaches to Hydrologic Design      78-3 

interest. We term these two return periods as the conditional and uncondi- instead use the mode, median, or quantiles to describe its distribution in lieu
tional return periods. For example, if one’s interest is in water supply failures, of its moments for stationary processes. Gumbel (1961) gives the geometric
droughts, and low streamflow sequences, all of which exhibit temporal cor- mean TG, median Tmedian, and mode Tmode of the waiting time to the next
relation, ones interest would focus on the conditional return period (see record flood as:
Lloyd, 1970). The definition of return period introduced in (78.2.1–78.2.3) is  n 1
the unconditional return period which is much more widely adopted than the TG = exp∑  ≅ γ + ln (n) = 1.78n (78.2.5a)
conditional return period introduced by Lloyd (1970) and others. In practice,  j=1 j 
the unconditional return period is more useful than the conditional return
period, because it does not require an assumption regarding initial condi- 21 n n 1
Tmedian = ≅ + = 1.44n + 0.5 (78.2.5b)
tions; hence, it more closely corresponds to design conditions. In other words,
one does not usually assume that a flood or drought has just occurred, when
( 1n
)
2 − 1 ln ( 2) 2

planning for the next flood or drought. n +1 (78.2.5c)

Tmode =
An advantage of the conditional return period is that it has been shown to 2
be insensitive to streamflow persistence or correlation (Lloyd, 1970). The same Note that in general, Tmode < n < Tmedian < TG . Clearly these measures of central
cannot be said for the unconditional return period, T =1 q , defined above tendency of the waiting time to an observation which exceeds that of the larg-
which is commonly used in hydrology, so that hydrologic persistence can have est value in a sample of size n vary over a significant range from roughly 0.5n
an important impact on the average return period associated with many to 1.8n.
hydrologic processes which exhibit persistence (see Douglas et al., 2001).
To account for persistence in hydrologic processes on the properties of the Risk and Reliability under Stationary
return period, numerous approaches have been adopted including the use of Conditions
a two-state Markov model (Vogel, 1987; Fernandez and Salas, 1999; Sen,
1999). Vogel (1987) defined the average return period of a reservoir system The concept of reliability is one of the most widely used design criteria in
failure as the expected number of years before the first occurrence of a system water resources planning and management. For example, the concept of reli-
failure (i.e., flood or drought). Vogel (1987) showed that the average return ability is widely used in irrigation and water supply planning (Hirsch, 1979;
period defined in this way is indeed affected by streamflow persistence. Vogel, 1987; Harberg, 1997; Tung, 1999; Loucks, Loucks and van Beek, 2005)
Fernandez and Salas (1999) and Sen (1999) developed more general formula- and many other fields (Kottegoda and Rosso, 1997; Modarres et al., 2009).
tions for estimating the average return period of design events in the presence Perhaps some of the earliest work which related the concept of annual reli-
of persistence. ability to the reliability over an N-year planning period associated with the
It is not only return periods that are impacted by hydrologic persistence. design of flood control structures were introduced by Gumbel (1941),
Estimates of hydrologic design events, the subject of Chapters 21 and 22, can Thomas (1948), and Yen (1970). Hashimoto et al. (1982) suggest that reli-
also be impacted by hydrologic persistence. For example, Potter (1992) ability alone is not sufficient for understanding the performance of a water
describes a procedure for accounting for persistence in the determination of resource system because it does not reflect the consequences of a failure if it
the probability distribution of annual maximum water levels. Tasker (1983) should occur. Instead, their now classic paper suggests that the three metrics
recognized the effect of persistence (temporal correlation) in streamflow of reliability, resilience, and vulnerability are needed to fully understand water
records on the reliability of design flood estimates and developed a method resource system performance. Hashimoto et al. (1982) define reliability as the
for quantile estimation (see Chapter 21 for further information on quantile probability or likelihood that a system remains in a satisfactory state.
estimation) using the effective record length, ne, which, in most cases, is less Reliability is usually defined as the converse of the probability of failure,
than the actual historical record length, n, when persistence is present in the which in the case of AMF discharges, would be the exceedance probability q
streamflow record. Vogel and Kroll (1991) describe a method for accounting associated with the design event of interest. Thus if the annual probability of
for both the temporal and spatial correlation of flood and low flow series, failure is defined as q, then the annual reliability Ra=1-q=p. Hashimoto et al.
when ones interest is in extending short streamflow series. (1982) emphasize that neither the probability of failure nor reliability reflect
the consequences of an extreme event. The notions of resilience and vulner-
Return Periods of the Flood of Record and ability (see Hashimoto et al., 1982 for definitions) are needed to reflect the
Envelope Curves consequences associated with an extreme event.
A flood discharge which exceeds all previous flood discharges is said to be Historically, some hydrologists have referred to the probability of failure q,
the flood of record (FOR). Envelope curves are regional plots describing the as the risk of an event. However, more recently, in the context of ecological,
upper bound on the FOR for many watersheds in a region versus drainage area. health, and other risk assessment studies, the term risk has been defined as
Envelope curves representing the current bound on flood experience have the product of the probability of an event and its consequences. Since that
limited use because of our inability to assign to them an exceedance probabil- definition of risk is now much more commonly used in most fields, we rec-
ity. Castellarin et al. (2005) and Vogel et al. (2007) developed approaches to ommend no longer referring to the probability of failure as the risk of an
estimate the exceedance probability of the expected regional envelope curve event. Instead, we focus on the related concept of reliability of a project over
as well as an individual regional envelope curve, respectively. its planning horizon n.
Since the FOR is of interest in hydrologic investigations which employ Most hydrology textbooks contain expressions which relate the reliability
envelope curves, it is important to understand the theoretical properties of the of a water project over an n-year planning horizon, Rn, to its annual reliability
return period associated with an FOR. Wilks (1959) and Gumbel (1960) Ra. One can relate the annual reliability Ra, to the n-year reliability, Rn, as fol-
derived various properties of Tr,n, defined as the return period associated with lows. The n-year reliability Rn is simply the probability of no flood event for
the rth order statistic, where r = 1 represents the largest observation in a sample the first n-years, which can be derived from the PDF of the return period τ in
of length n. Wilks (1959) and Gumbel (1961) show that the expectation of Tr,n Eq. (78.2.1) as follows:
associated with the rth order statistic in a sample of size n is given by: n
Rn = P [τ ≥ n] = 1 − P [τ ≤ n] = 1 − ∑ P [τ = t ] = (1 − q ) = Ran (78.2.6)
n
n
E Tr ,n  = for r ≥ 2 (78.2.4)
r −1 t =1

Interestingly, the average return period associated with the largest observa- Note that Eq. (78.2.6) assumes that flood events in each year are independent;
tion (the FOR) is infinite under stationary conditions. From Eq. (78.2.4), if the hydrologic process of interest exhibits serial correlation Eq. (78.2.1) no
regardless of how large the FOR is, or how long ago it occurred, on average, longer applies (see previous section on dependent or persistent processes).
one will need to wait until eternity, to experience a flood greater than that We recommend the use of the concept of n-year reliability Rn for future plan-
record. This result should raise some questions concerning the suitability of ning, because it reflects the likelihood of failure over the entire planning
using the average return period in flood planning, even under stationary horizon of interest and is used in many other fields (see Table 1 in Read and
conditions (see Read and Vogel, 2015; and Serinaldi 2015; for further discus- Vogel, 2015).
sion). Similarly, from Eq. (78.2.4) all higher order moments of the return When a hydrologist performs a design based on the T-year flood, its reli-
period associated with the FOR (r = 1) are infinite, though moments do exist ability over the next n-year planning period is obtained from Eq. (78.2.6) as:
for floods smaller than the FOR (i.e., r ≥ 2).  1
n
In contrast to the expectations of the recurrence time of the next FOR Rn = 1 −  (78.2.7)
which is infinite; other measures of central tendency do exist, such as the  T
mode, median, and geometric mean of the recurrence time of the FOR... The above relationships between the annual reliability and the reliability
Since the moments of the recurrence time of the FOR do not exist, one could over an n-year planning period were first introduced by Thomas (1948) and

78_Singh_ch78_p78.1-78.10.indd 3 01/07/16 4:29 PM

 78-4     Risk, Reliability, and Return Periods and Hydrologic Design

further analyzed by Gumbel (1941) and Yen (1970). Those relationships

Annual cost
depend upon the fundamental assumption that flood flows are independent
and identically distributed variables. These relations are in widespread use as
evidenced by their inclusion in both hydrology handbooks (Chow, 1964;
IACWD, 1982; Stedinger et al., 1993; Tung, 1999) as well as in many textbooks Annual total
(Bras, 1990; Viessman and Lewis, 2003; Mays, 2005) and journal papers expected cost
(Gumbel, 1941; Thomas, 1948; Yen, 1970; Wigley, 2009; Salas and Obeysekera,
2014; Read and Vogel, 2015).
One concern with using the average return period T to denote risk of
flooding is that it does not capture the impact of the planning horizon on the
Annual installation
project of interest, as does the concept of reliability over a planning horizon cost
Rn. Note that according to Eq. (78.2.7) the reliability is only 60.5% when the
design event is based on the 1% exceedance (q = 0.01) event (T = 100-year
flood) for n = 50-year project life. Considering that the design life of much of
our water infrastructure is greater than 50 years, and that reliability decreases Annual expected
damage cost
as project life increases for a given design average return period T, careful
attention should be given to how the reliability is impacted by the planning
horizon for structures which have been designed on the basis of an average Optimal Project size, X
return period. project size
Other fields concerned with hazard planning ensure a much higher reli-
ability over typical planning horizons than corresponding reliabilities associ- Figure 78.1  The Risk-Based design method for flood management. [Source: Tung, 2005.]
ated with the 100-year flood so commonly used in hydrologic planning (see
Table 1 in Read and Vogel, 2015). For example, earthquake design regulations are described in most introductory textbooks in statistical decision theory
suggest protection against a “less than 2% chance of failure (collapse) and decision sciences as well as in most textbooks on water resource systems
occur(ring) in a 50-year project life” (National Earthquake Hazard Reduction analysis (Loucks and Van Beek, 2005).
Program, 2010). This level of protection corresponds to a reliability over 50
years of 98% and a corresponding design earthquake magnitude with an aver-
age return period of T = 2,475 years. By comparison, traditional flood fre- 78.2.2  Probabilistic and Risk-Based Hydrologic Design
under Nonstationary Conditions
quency analysis which often bases designs on an average return period of T =
100 years corresponds to reliabilities of 78%, 61%, and 37% over a range of Olsen et al. (1998) first described the theoretical properties of the various
n = 25, 50, and 100 year planning horizons, respectively. hydrologic design indices described in the previous section, under nonsta-
When designing for protection against a flood with an average return tionary conditions. Several investigators have sought to extend the results of
period T, of interest is the reliability over a design life equal to its average Olsen et al. (1998) including the concepts of return period and reliability to
return period n = T, in which case one obtains: nonstationary conditions. Interestingly, probably due to the tremendous
T attention and need to understand how to adapt to climate change, most of the
 1 key developments extending hydrologic design indices to nonstationary
Rn=T = 1 −  (78.2.8)
 T conditions have appeared primarily in the statistical and climate change lit-
1 erature (Wigley, 1988; 2009; Katz, 1993; Olsen et al., 1998; Parey et al., 2007;
Yen (1970) points out that in the limit, as T →∞, Rn=T → = 0.368 which is an Cooley, 2009; Katz, 2010; Parey et al., 2010; Cooley, 2013). Salas and
e
Obeysekara (2014) and Read and Vogel (2015) provide further extensions to
extremely low reliability when compared with other fields of engineering.
the work of Cooley (2013) and Olsen et al. (1998), focusing on water resource
Analogous to the conditional and unconditional return periods discussed
applications of the various measures of return period and reliability which we
earlier, the reliability over an n-year planning horizon, Rn, can be defined
summarize as follows.
using either a conditional or unconditional approach. Douglas et al. (2002),
and Fernandez and Salas (1999) use a two state Markov model to derive
expressions for the unconditional n-year reliability Rn, for events which Return Periods of Design Events under
exhibit serial persistence. Similarly, Sen (1999) uses a two state Markov model Nonstationary Conditions
to derive expressions for the conditional n-year reliability Rn, for events which Recall from the previous section on stationary methods that we defined the
exhibit serial persistence. annual non-exceedance probability associated with a particular flood event as
p=Fx(x) along with its associated AEP q=1–p. If flood magnitudes are nonsta-
Risk-Based Decision Making under Stationary Conditions
tionary and either increase or decrease systematically through time the
Risk-Based Decision Making is a well-established methodology that exceedance probabilities q, will correspondingly increase or decrease over
determines appropriate levels of infrastructure based on the expected dam- time to form a series q1 , q2 , q3 ,..., qt . Under nonstationary conditions, the time
ages avoided versus the cost of the infrastructure required (National Research to the occurrence of the next flood t, follows the distribution:
Council, 2000; Tung, 2005) and is now standard practice by U.S. Federal agen-
cies (see Stakhiv 2011 for references). RBDM can be used in place of the tra- P [τ = t ] = fτ (t ) = (1 − q1 )(1 − q2 )(1 − q3 )...(1 − qt −1 )qt
ditional design-event approach which selects a particular average T-year (78.2.9)
event usually specified by regulation or experience, and then designs the Equation (78.2.9) is the PMF of the unconditional waiting time, recurrence
necessary infrastructure to protect against the hydrological event with that time, or return interval until the occurrence of the first flood to exceed the
specified average return period. Instead, the goal of RBDM is to choose a level design flood. Equation (78.2.9) describes a nonhomogeneous geometric ran-
of infrastructure protection that minimizes the total expected annual cost dom variable (see Mandelbaum et al., 2007), which is completely analogous to
(including flood damage costs) of the infrastructure, which is labeled “Annual the homogeneous geometric variable described in Eq. (78.2.1). When the
Installation Cost” in Fig. 78.1. Alternatively, RBDM may maximize the project non-exceedance probabilities in Eq. (78.2.9) are all equal, as is the case for
net benefits. stationary series, then Eq. (78.2.9) reduces to the homogeneous geometric
Thus an RBDM process may lead to flood-risk mitigation measure against result for stationary series in Eq. (78.2.1). Salas and Obeysekara (2014)
a flood event either larger or smaller than, say, the 100-year flood, which is a describe two separate cases that must be treated independently, one in which
common design event considered in the traditional analyses. magnitudes of floods are increasing and one in which they are decreasing.
One of the most common approaches to performing RBDM for sequential Since increasing floods are the primary concern in hydrologic engineering
decision problems is to use a decision tree. A decision tree is the graphical design and for brevity, we only summarize the case of increasing floods here
equivalent of a stochastic dynamic program which is a well developed math- (see Salas and Obeysekara, 2014; for the case of decreasing floods).
ematical programming method used in the areas of operations research, For the case of increasing floods, there will eventually be a year tmax, in
industrial engineering, and management science. A decision tree describes which the AEP of the flood qt, is unity so that Eq. (78.2.9) becomes:
the sequence of possible decisions for numerous alternatives along with their
t −1
probabilistic and economic outcomes. It is a very powerful approach because
it combines a graphical representation of the overall set of alternatives and P[τ = t ] = fτ (t ) = qt ∏(1 − q y ) t=1,2,…,tmax (78.2.10)
y =1
decisions, with a framework for making risk-based decisions. Decision trees

78_Singh_ch78_p78.1-78.10.indd 4 01/07/16 4:29 PM

 Probabilistic and Risk Based Approaches to Hydrologic Design      78-5 

Similarly, the CDF of the return period τ for the nonstationary case conditions. Salas and Obeysekera (2014) suggest that under nonstationary
becomes: conditions, the n-year reliability Rn is given by:
t t y −1 t n

P[τ ≤ t ] = Fτ (t ) = ∑ fτ ( y ) = ∑q y ∏(1 − qt ) = 1 − ∏ 1 − q y t=1, 2,…, tmax ( ) Rn = ∏ (1 − qt ) (78.2.17)

y =1 y =1 t =1 y =1 t =1
(78.2.11)
As was shown in Section 78.2.1, under stationary conditions, when the AEP
where, Fτ(1) = p1 and Fτ(tmax) = 1. Salas and Obeysekera (2014) describe sev- is constant, the n-year and annual reliabilities are related via Rn = Ran.
eral possible forms of nonstationarity which have different implications con-
cerning the use of the above expressions and the behavior of tmax. If Use of the Traditional Decision-Making
nonstationarity continues and the exceedance probabilities qt continue to Process under Nonstationary Conditions
increase, they will eventually reach the upper limit of unity at tmax. However, Given the tremendous interest in the subjects of climate change, urbanization
if the exceedance probabilities asymptotically approach unity, or if for some and their subsequent impact on hydrologic processes, a risk-based decision
reason the nonstationarity ends, then there will be no upper limit to the dis- framework under nonstationary conditions is of vital importance. Perhaps
tribution of τ in which case t max →∞. the most common example of nonstationarity in hydrologic design involves
Analogous to the above expressions, Mandelbaum et al. (2007) introduced impacts of urbanization. See Chapter 132 “Storm Water Management, Best
the nonhomogeneous geometric probability distribution within the context of Management Practices, and Low-impact Development” and Chapter 149
birth and death processes. They provide a convenient recursive expression for “Human Impacts on Hydrology” for further background on deterministic
computing ft(t) from Ft(t–1): approaches for handling urbanization and other forms of nonstationarity.
fτ (t ) Using a traditional decision-making process under nonstationary condi-
qt = (78.2.12) tions, a trend is first evaluated for statistical significance separately from the
1 − Fτ (t − 1) economic project evaluation. First, a hypothesis test is performed and the
where Fτ(0) = 0. Mandelbaum et al. (2007) also show how Eq. (78.2.12) can be
statistical significance α of the trend is estimated. If α is below some
used to define the structure of future sequences of exceedance probabilities
prespecified critical value, usually αcritical=0.05, the economic analysis is per-
which arise from various forms of the PDFs ft(t) and CDF Ft(t).
formed to evaluate the economic viability of a proposed flood management
Analogous to the expressions for the moments of the return period for
plan. If α exceeds the critical value the trend may be dismissed and the eco-
stationary conditions in Eqs. (78.2.2) and (78.2.3), the PDF of the return
nomic analysis is often not performed.
period τ in Eq. (78.2.10) can be used to derive the moments of the return
Normally, under either stationary, or nonstationary conditions, two possi-
period under nonstationary conditions so that:
ble economic analyses are considered, one which is based solely on an arbi-
t max tmax t −1 trary design event, such as the 100-year flood, and an RBDM approach which
T = E [τ ] = ∑ t ⋅ fτ (t ) = ∑ t ⋅ qt ∏(1 − q y ) (78.2.13) would choose a level of infrastructure to protect against that design event
t =1 t =1 y =1 which maximizes the net benefits of adaptation (the estimated damages
Cooley (2013) provides a useful computational simplification of Eq. (78.2.13) as avoided less the cost of the proposed adaptation). In either case, whether an
RBDM approach or a fixed design event is employed, an economic analysis of
tmax t
nonstationary conditions would normally be performed only if a trend was
T = E [τ ] = 1 + ∑ ∏(1 − q y ) (78.2.14) found to be statistically significant. If the analysis concluded that no trend is
t =1 y =1 evident, the consequences of under-preparation would normally, neither be
Similarly, Salas and Obeysekera (2014) suggest computing the variance of the computed nor ever even considered in a traditional analyses. The reminder of
this section describes the critical need to adapt existing RBDM approaches for
return period τ using the fact that Var [τ ] = E τ 2  − E [τ ] so that
2
use under nonstationary conditions, because the added uncertainty concern-
ing the existence (or not) of a trend needs to be considered in the RBDM
t max t −1  tmax t 
2
process.
Var [τ ] = σ τ2 = ∑ t ⋅ qt ∏(1 − q y ) − 1 + ∑ ∏(1 − q y ) (78.2.15)
2 There is a wide range of possible situations which exist relating to uncer-
t =1 y =1  t =1 y=1  tainty in future flood projections as well as the myriad of consequences asso-
ciated with either protecting or not protecting against projected flood
Recall, that under stationary conditions the coefficient of variation of the damages. A full RBDM analysis is needed in order to integrate all relevant
return period τ is roughly equal to unity for rare floods, so that the mean information (both costs and benefits) concerning future regrets which result
return period provides an excellent summary measure of the distribution of from various sources of economic, social, natural, and other forms of uncer-
future return periods. tainty. For example, within a flood control context, “regret” would be defined
Parey et al. (2007; 2010) and Cooley (2013) discuss another interpretation as the difference between the flood damages corresponding to a particular
of the return period as the number of years one must wait until the expected project design and the minimum damages if one had perfect information
number of events is unity. Cooley (2013) shows that both definitions are about the future. See Rosner et al. (2013) for a nonstationary RBDM analysis
equivalent under stationary conditions. Let N be the number of exceedances which considers the concept of regret. Also see Stakhiv (2011) for a discussion
that occur in the T years beginning with year y=1 and ending with year y=T. and references to a “no regrets” approach to RBDM in water resources.
Under stationary conditions, N follows a binomial distribution, but that is no
longer the case under nonstationary conditions. Cooley (2013) shows that A Risk-Based Approach to Flood Management
under nonstationary conditions, when the expected number of flood events under Nonstationary Conditions
to exceed the T year event is equal to 1, T can be computed as the upper limit Flood (or drought, or other water resources) management decisions in a non-
in the following sum: stationary world are inherently sequential decisions, which depend critically
T
on the uncertainty inherent in future projections of flood scenarios and their
E [ N ] = 1 = ∑(1 − F ( x , y )) (78.2.16) corresponding consequences. Nevertheless, there are remarkably few exam-
y =1
ples of the application of RBDM to nonstationary water resource problems in
the scientific literature. There are also very few examples of the use of decision
where, F(x,y) is the nonstationary CDF of the AMF series, x, where y repre- trees under nonstationary conditions. This is particularly surprising because
sents the year. Read and Vogel (2015) found that for a nonstationary lognor- nonstationary decision processes are sequential decision processes which are
mal model, the definition of T in Eq. (78.2.16) led to nearly identical results subject to uncertain economic outcomes, and a decision tree or a stochastic
as the traditional definition of T given in Eqs. (78.2.13) and (78.2.14), for a dynamic program are the natural approaches to apply to such decision prob-
broad range of hydrologic conditions, thus we do not discuss the definition in lems. Fiering and Matalas (1990) provide one of the earliest examples of a
Eq. (78.2.16) further. sequential statistical decision process for evaluating various alternatives in the
context of nonstationarity (climate change). Chao and Hobbs (1997) give a
Risk and Reliability under Nonstationary Conditions brief history of decision analysis applications to climate change; and apply a
Under stationary conditions, the relationships between the n-year reliability mathematical version of a decision tree known as a stochastic dynamic pro-
Rn, annual reliability Ra, and planning horizon n are given in Eq. (78.2.6) and gram for evaluating breakwater adaptation under possible climate change
between n-year reliability Rn, average return period T, and planning horizon impacts on Lake Erie. Hobbs et al. (1997) were the first to apply a decision tree
n are given in Eq. (78.2.7). Such relations are in widespread usage, thus one approach to water resources management under climate change, and recently
expects that analogous relationships could be useful under nonstationary the method has been resurrected (see Gersonius et al., 2013).

78_Singh_ch78_p78.1-78.10.indd 5 01/07/16 4:29 PM

 78-6     Risk, Reliability, and Return Periods and Hydrologic Design

A critical challenge in the application of decision trees to the problem of considers societal regret associated with under-preparation against future
RBDM under nonstationary conditions involves estimation of the necessary flood risks.
probabilities associated with various outcomes (branches of decision tree). Concerns about NHST are crucial to the fields of geophysics, climate sci-
Hobbs et al. (1997) demonstrate use of a Bayesian approach to analyzing the ence, and water resources engineering, where the trend analysis could have an
necessary probabilities in the decision tree for evaluating alternative adapta- impact on major infrastructure decisions. Remarkably, it is only very recently
tion strategies for climate change for the Great Lakes. Similarly, Manning et al. and rarely that researchers have raised concern over the importance and
(2009) describe a Bayesian analysis consisting of aggregating predictions from impacts of Type II errors in the climate and hydrologic sciences (Ziegler et al.,
suites of model predictions, such as Global Circulation Models (GCMs) or 2003, 2005; Cohn and Lins, 2005; Trenberth, 2008; Morin, 2011; Vogel et al.,
Regional Climate Models (RCMs). 2013). Some journals have even banned the use of NHST (Trafimow and
Another approach for applying an RBDM using a decision tree under Marks, 2015). Though many studies have discussed the importance of consid-
nonstationary conditions would be to integrate the uncertainty inherent in ering Type II errors in the analysis of trends and other hypotheses, they did
projections of future trends. Such an analysis must incorporate the uncer- not consider the resulting impacts on infrastructure decisions and societal
tainty associated with our ability to detect changes in flood series. Rosner preparedness, as was considered by Rosner et al. (2014) and is discussed
et al. (2013) introduced an RBDM approach for use under nonstationary below. A Type II error in the context of an infrastructure decision implies
conditions, using a decision tree, with the outcome probabilities based on under-preparedness, which is often an error which is much more costly to
Type I and Type II error probabilities associated with statistical trend society than the Type I error (over-preparedness) which the NHST focuses on.
hypothesis test outcomes. Their approach integrates numerous consider- The physical implication of a Type I or over-preparedness error in adapta-
ations including: a nonstationary GEV model of flood frequency, the uncer- tion decisions for flood management is wasted money on unneeded infra-
tainty inherent in the trend detection process, natural hydroclimatic structure. The physical repercussions of a Type II or under-preparedness
uncertainty and a detailed economic analysis associated with the various error, on the other hand, are major flood damages due to inadequate protec-
infrastructure alternatives under consideration. The resulting process tion. Decision makers are poorly served by statistical and/or decision meth-
enables the decision maker to ask the question when enough information is ods that do not carefully consider both sources of error.
available to warrant making a particular flood management adaptation deci- There is a continuing need to acknowledge the tremendous uncertainty
sion under nonstationary conditions. associated with our ability to discern trends from other natural phenomenon
such as persistence (see Cohn and Lins, 2005) as well as complications due to
Integration of Trend Detection and Decision-Making
seasonality, censoring, and other issues (see Chapter 11 in Helsel and Hirsch,
Adaptation planning in the context of flood management under nonstation- 2002). One of the main arguments against Null-Hypothesis Significance
ary conditions depends critically on trend detection and a subsequent trend Testing (NHST) is its adherence to a single critical value αcritical (often chosen
model; hence, it is important to understand the limitations and concerns equal to 0.05). Use of NHST implicitly places disproportionate emphasis on
surrounding tests of statistical significance of trends. Studies which seek to the Type I error probability α, while the power 1-β is rarely reported, despite
identify trends in flood series are now widespread. All of the many previous the importance of Type II underdesign error probabilities, for example, in
flood studies we have reviewed which have sought to determine whether a flood management applications.
trend exists in flood series, have employed a null hypothesis, Ho, of no trend There is very little attention given to the power of trend tests in the water
and most have chosen an associated significance level of α=0.05. A signifi- and climate literature. Lettenmaier (1976) first introduced to the water
cance level of 0.05 implies that if there really is no trend (that is assumption resources literature analytical expressions for the power of a hypothesis test
of Ho), we will only (mistakenly) report trends 5% of the time. The societal based on ordinary least squares (OLS) linear regression in the context of
consequences of making such a mistake is that we will prepare for a trend, trend detection in water quality management. Bowling et al (2000) per-
when it does not exist, which we term over-preparedness. Shouldn’t society formed a similar analysis to determine the minimum detectable difference or
also be interested in the likelihood of under-preparedness? Surely there are the smallest trend one could discern to be statistically significant.
situations in which society will regret having been under-prepared for conse- Interestingly, even though exact analytical expressions exist for computing
quences of events which could have been avoided. See Vogel et al. (2013) for the power of a trend test based on OLS regression we found it quite difficult
a complete discussion of this issue. to locate textbooks or primer papers which document such analyses. This is
Null hypothesis analysis, termed Null-Hypothesis Significance Testing especially surprising given the widespread use of linear regression for per-
(NHST), focuses only on our understanding of conditions of no trend, forming trend analyses. Examples of primer papers on the power studies for
because all such hypothesis tests were derived under conditions of no trend. trend detection based on linear regression are common in the medical sci-
Thus the alternative hypothesis, HA, when trends do exist, is usually ignored ences (Dupont and Plummer, 1990; 1998) though we could only find a few
along with its probability of occurrence known as the probability of a Type II examples of analytical power studies in the water literature (Lettenmaier,
error which is termed β. The decision matrix for the general trend detection 1976; Bowling et al., 2000; Ziegler et al., 2003; 2005; Vogel et al., 2013;
decision problem is depicted in Fig. 78.2. Statisticians would define the term Prosdocimi et al., 2014; Rosner et al., 2014).
“power” of this hypothesis test as the likelihood of detecting a trend, when it Remarkably, of the hundreds and possibly thousands of studies which have
exists. Of particular concern to us are the likelihood of Type II errors, which examined trends in hydroclimatic variables, we could only find a few studies
is both out of our control, and involves significant associated societal conse- which computed either the probability of Type II errors or the power. For
quences because they imply no societal response is necessary when one is example, Ziegler et al. (2005) used GCMs to predict trends in annual precipi-
warranted (see Vogel et al., 2013). tation on the Mississippi basin, and then performed Monte Carlo simulations
A true RBDM approach would avoid the need to define a critical value for to determine the minimum length of record which would be needed to detect
either the Type I or Type II error probabilities. Numerous fields, including trends of those magnitudes. They found that between 82 and 143 years would
psychology, economics, social sciences, meteorology, and medical research, be required to detect a trend corresponding to Types I and II error probabili-
have called into question the value of NHST tests due to its focus on its ties of α=0.05 and β=0.10, respectively. Ziegler et al. (2003; 2005) employed
dependence upon a single, often arbitrary, significance level (Cohen, 1994; the simple analytical approximation to the power of a t-test introduced earlier
Nicholls, 2000; Ziliak and McCloskey, 2008). Such concerns over the use of by Lettenmaier (1976). Morin (2011) performed a similar analysis using
NHST are now widespread, though remarkably, none of those studies we Monte Carlo simulations to estimate the minimum magnitude of change in
have reviewed well on the most important criticism of all, that of ignoring annual precipitation at over 9,000 stations globally, that could be detected
the probability of Type II errors, a concern of paramount interest when one over a 50-year period. He reports minimum detectable trends given Types I
No trend (Ho) Trend (HA) and II error probabilities of α=0.05 and β=0.50.
Although Vogel et al. (2011) and Rosner et al. (2014) employ a simple
β nonlinear model (fit using linear regression) to characterize trends in flood
No societal response 1–α Type II error levels as a function of time, they and others recommend use of physical
(under-preparedness) covariates for explaining trends in hydrologic processes. More complex trend
α analyses are possible by incorporating other covariate predictors of the trend
Type I error 1–β such as precipitation (Prosdocimi et al., 2014); climatic indices (Kwon et al.,
Societal response (over-preparedness) Power 2008; Steinschneider and Lall, 2015), large scale atmospheric or oceanic
spatial fields (Renard and Lall, 2014) and/or trends in other moments
Figure 78.2  Decision matrix for the General Trend Detection Decision Problem, (Villarini et al., 2009). Vogel et al. (2011, see appendix) found that an expo-
with Null Hypothesis, Ho, and Alternate Hypothesis, HA, shown. nential model obtained by relating the logarithm of instantaneous annual

78_Singh_ch78_p78.1-78.10.indd 6 01/07/16 4:29 PM

 Multivariate Probabilistic and Risk-Based Approaches to Hydrologic Design     78-7 

maximum streamflow to its year of occurrence provided an excellent 78.3.1  Return Periods of Multivariate Design Events
approximation for thousands of river gages across the continental U.S. (See The theory of copulas offers a convenient and useful approach for mathemat-
Prosdocimi et al. (2014) for a similar analysis in the United Kingdom.) Even ically representing the dependence structure of correlated hydrological design
for highly nonlinear trends, OLS regression can often provide a good variables without making assumptions about their multivariate structure. The
approximation to trends by employing the ladder of powers to “linearize” the nonparametric nature of copulas is in part why they are now so widely used
relationship (Helsel and Hirsch, 2002). Helsel and Hirsch (2002) provide a in hydrology and the geosciences, for the bivariate and multivariate frequency
detailed background on trend tests and how to improve their power, given analysis of extreme events (see Li et al., 2012 for wind events, and Corbella
the tremendous challenges associated with distinguishing between trends, et al., 2012 for ocean events and Chapter 30). Still, there is no unique or
seasonality, and persistence. Regression is an attractive approach to model- widely accepted definition of multivariate return periods in the scientific
ing trends because it provides a graphical summary display of the results, community. While the expected return period is univocally defined in a uni-
prediction intervals for trend extrapolations are easily derived, analytical variate environment and can be expressed as in Eqs. (78.2.2) and (78.2.13), for
expressions for the Type II error probabilities are available and rigorous stationary and nonstationary processes, respectively, there are several differ-
statistical tests are available for evaluating overall model integrity and resid- ent definitions of expected joint return periods (JRPs) for the multivariate
ual behavior. context (see Salvadori et al., 2011; Gräler et al., 2013; Requena et al., 2013;
Yue et al. (2002), Yue and Pilon (2004), Onoz and Bayazit (2003), and Serinaldi, 2015).
Morin (2011) have examined the power of the Mann–Kendall test and other
Here we describe the expected JRPs for the bivariate case. The primary
nonparametric trend tests, using Monte Carlo analysis. To our knowledge, no
published study has provided analytical expressions for the power of the return period TX∪,Y ( x *, y *) associated with the probability associated with the
Mann–Kendall test; however, Kendall reports analytical expressions for the exceedance of either one of the two variables, that is, the bivariate event
probability distribution of the Mann–Kendall statistic under both the null E ∪ ( x *, y *) = [ x > x * ∪ y > y *] (OR case), and the return period TX∩,Y ( x *, y *)
and alternative hypothesis, so it is surprising that none of the existing water related to the probability of exceedance for both variables, that is the bivariate
literature exploits that fact. It is also interesting to note that Morin (2011)
event E ∩ ( x *, y *) = [ x > x * ∩ y > y *] (AND case). TX ,Y ( x *, y *) and
∪
found that for his particular application, the power results of the linear regres-
sion and the Mann–Kendall were nearly identical, though results of only one TX ,Y ( x *, y *) can be expressed as follows for annual series of X and Y:
∩

technique was reported in the paper. 1 1 1

TX∪,Y ( x *, y *) = = =
78.3  MULTIVARIATE PROBABILISTIC AND RISK-BASED ∪
Pr [( X > x *) (Y > y *)] 1 − FX ,Y ( x *, y *) 1 − C FX(( x *), FY ( y *) )
APPROACHES TO HYDROLOGIC DESIGN (78.3.2a)
Above, our focus was on hydrologic design when a single hydrological vari- ∩ 1 1
X ,Y ( x *, y *) =
T =
able could be used to describe the return period or reliability associated with (
Pr [( X > x *) ∩ (Y > y *)] 1 − FX ( x *) − FY ( y *) + C FX ( x *), FY ( y *) )
system of interest. Such a univariate framework is useful for many practical
hydrologic applications which depend solely on peak flood discharges (e.g., (78.3.2b)
hydraulic design of bridge openings and culverts) or critical low flow dis- For these expected JRP’s the following inequalities relative to marginal
charges. Yet hydrological loads for structural design often depend on multiple expected return periods TX(x*) and TY(y*) hold:
hydrological variables that are significantly correlated to each other (see
Grimaldi et al., 2011). Common examples include flood risk mitigation prob- TX∪,Y ( x *, y *) ≤ minTX ( x *),TY ( y *) ≤ max TX ( x *),TY ( y *) ≤ TX∩,Y ( x *, y *)
lems in which hydraulic routing processes are important, such as in the design (78.3.2c)
of detention/retention basins, as well as dam sluice gates and spillways, where
there is a need to characterize the entire flood hydrograph both in terms of Salvadori and De Michele (2004) introduced a secondary return period, or
flood peak discharges and volumes (see Requena et al., 2014). In reference to Kendall’s return period, in order to identify in a multivariate context a uni-
such multivariate hydrologic processes, Yue and Rasmussen (2002) concluded variate critical threshold (see also Volpi and Fiori, 2014):
that “under a given return period, the flood peak/volume value given by the 1
single frequency analysis is greater than those by the joint distribution. This TK (t ) =, with K (t ) = Pr FX ,Y ( x , y ) ≤ t  (78.3.3)
implies that if one neglects the close correlation between flood peak and 1 − K (t )
volume, and carries out single-variable frequency analysis on flood peak or where, K(t) is the Kendall’s distribution function associated with FX,Y. TK(t)
volume only, the severity of a flood event may be overestimated. If a hydro- corresponds to the mean interarrival time of any event E t = FX ,Y ( x , y ) > t  ,
logic engineering design is based on the results from the single-variable fre- and therefore does not depend on a particular pair (x*,y*) and is defined for
quency analysis, then this overevaluation will lead to an increased cost. any value of t.
Hence, single-variable frequency analysis cannot provide a sufficient probabi- For a given pair (x,y), the following general inequality holds (see also
listic assessment of a correlated multivariate event.” Gräler et al., 2013):
In traditional multivariate frequency analyses, correlated variables are
modelled using standard bivariate or multivariate distributions (see Matalas TX∪,Y ≤ TK ≤ TX∩,Y (78.3.4)
and Langbein, 1962; Stedinger, 1983; Hosking and Wallis, 1988; Castellarin
et al., 2005). However, this approach assumes (1) the same marginal prob- The application of a traditional design event-based approach to hydrologi-
ability distribution for all variables involved in the analysis and (2) a linear cal design in a bivariate (or multivariate) context would require the selection
dependence between the PDFs; which have both been shown to be of lim- of a particular design return period (e.g. 100 years) and definition (i.e.,
∧
ited practical applicability (e.g., Favre et al., 2004; Klein et al., 2010). These TX∨,Y TX ,Y or TK ) of the JRP, which results in the identification of an infinite set
assumptions can be relaxed by resorting to the use of copula-based distribu- of events (x,y) in the xy space, that is, the curves illustrated in the examples of
tions (see Chapter 30) which are now widely used in hydrology as evidenced Fig. 78.3 (adapted from Requena et al., 2014). The hydrological design would
by the expanding literature on floods, precipitation, and droughts (see then refer to a single bivariate (or multivariate) event, or a set of events associ-
Genest and Favre, 2007; Zhang and Singh, 2007; Chen et al., 2013; Li et al., ated with the selected JRP, and, following this rationale, the return level of
2013; Sadri and Burn, 2014). With no loss of generality we focus on the structural failure would then be assumed to be equal to that of the selected
bivariate case, when the theory of copulas is based on Sklar’s theorem hydrological design events. The literature provides guidelines for selecting
(Sklar, 1959): multivariate design events (see Salvadori et al., 2011; Corbella and Stretch,
2012; Volpi and Fiori, 2012; Gräler et al., 2013).
FX ,Y ( x , y ) = Pr [( X ≤ x ) ∧ (Y ≤ y )] = C [ FX ( x ), FY ( y )] (78.3.1) However, this approach involves several degrees of subjectivity (e.g., ambi-
guities in the JRP definitions, infinite events associated with the same JRP,
where, FX,Y(x,y) is the joint cumulative distribution function of random vari- different probabilities of occurrence for these events) and, more importantly,
ables X and Y, with marginal PDFs FX(x) and FY(y), respectively, and the assumption of equivalence between the selected JRP and the expected
C:[0,1]2→[0,1] is the copula function. Equation (78.3.1) does not require any return period for the failure of the structure, which may not hold leading to
assumptions concerning the marginal PDFs. The copula function enables the under or overdesigning (see Salvadori and de Michele, 2004; Gräler et al.,
user to represent nonlinear dependencies between variables X and Y, for 2013). The most recent studies for hydrological design in a multivariate con-
example, higher extremes more correlated than lower extremes (for details on text apply a radically different perspective, switching from a design event-
copulas see Favre et al., 2004; Genest and Favre, 2007; Salvadori et al., 2007). based approach to a structure-based approach (see Volpi and Fiori, 2014).

78_Singh_ch78_p78.1-78.10.indd 7 01/07/16 4:29 PM

 78-8     Risk, Reliability, and Return Periods and Hydrologic Design

80 80
Simulated pairs Simulated pairs
Observed data 70 Observed data
70
TZ TZ
60 TX,Y 60 TX,Y

50 50
V [hm3]

V [hm3]
500 500
40 40
500 100 100 500
30 50 30 50
100 100
20 50 10 20 10 50
5 5
10 10 10
5 10
5
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
(a) Q [m3s–1] (b) Q [m3s–1]

80
Simulated pairs
70 Observed data
TZ
60 TK

50
V [hm3]

40 500
500 100
30 50
100
20 50 10
5
10 10
5

0
0 50 100 150 200 250 300
(c) Q [m3s–1]
Figure 78.3  Example of multivariate design of a dam spillway as a function of design hydrograph volume, V, and discharge peak, Q: regions in the QV space associated with expected
return period of 5, 10, 50, 100, and 500 years according to the definition of the structure-based return period TZ (Eq. 78.3.6, bold line) and TX∨,Y (OR case, Eq. 78.3.2a dashed line panel
a), TX∧,Y (AND case, Eq. 78.3.2b, dashed line panel b) and TK (Kemdall, Eq. 78.3.3, dashed line panel c) JRPs. [Source: Adapted from Fig. 10 in Requena et al., 2013, p. 3034.]

78.3.2  Hydrologic and Hydraulic Design Using According to Eq. (78.3.5) and (78.3.6), calculating TZ(z*) only requires
Multivarate Events
identification of the region DZ ( z *) =  g ( x , y ) ≤ z * and evaluation of the
Following Volpi and Fiori (2014), consider the design parameter Z (e.g., the above integral. For the sake of comparison, Fig. 78.3 (adapted from Requena
size of a spillway, the elevation of a levee, etc.) or a quantity related to it, which
et al., 2013) illustrates the infinite sets of bivariate events in the xy space asso-
measures the effects of the hydrological load on the structure; while X and Y
ciated with mean return periods of 5, 10, 50, 100, and 500 years, according to
are the pair of hydrological variables which identify the hydrological load on
definitions Eqs. (78.3.2a), (78.3.2b), (78.3.3) and (78.3.6).
the structure (e.g., the peak discharge and the volume of a hydrograph). Let
Unfortunately, in many cases of practical interest, the function g(x,y) can be
us suppose that Z depends on both hydrological variables X, Y through
very complex and its evaluation needs to be based on simulated experiments.
the structure function g(.), i.e., Z = g(X,Y), which mathematically expresses the
In those cases, the solution to the multivariate structure-based design can be
interactions among the structure and the hydrological loads acting on it. The
obtained through Monte Carlo procedures structured as follows (see Requena
PDF of Z can be derived as:
et al., 2013; Volpi and Fiori, 2014): (1) generate a long series of synthetic pairs
FZ ( z ) = ∫D f X ,Y ( x , y ) ⋅ dx ⋅ dy , where (78.3.5a) (x,y) with joint probability density function fX,Y(x,y); (2) compute the corre-
Z
sponding values of the design parameter z=g(x,y); (3) evaluate the univariate
∂2 FX ,Y ( x , y ) PDF of the design parameter FZ(z) based on the computed series; and (4)
DZ =  g ( x , y ) ≤ z  and f X ,Y ( x , y ) = , (78.3.5b)
select the design parameter zT with return level equal to T.
∂x ∂ y
provided that fX,Y(x,y) exists. From Eq. (78.3.5), if Z>z* determines a structure 78.3.3  Multivariate Risk-Based Design under
failure and X and Y are annual hydrological variables (i.e., annual maxima of Nonstationary Conditions
flood peak and volume), the mean time elapsing between two successive
structure failures can be expressed as: Multivariate risk-based design under nonstationary conditions is still an open
research topic. To our knowledge, the hydrological scientific literature pres-
1
TZ ( z *) = ents only one study addressing this very issue (see Bender et al., 2014). Bender
1 − Fz ( z *) (78.3.6) et al. (2014) present a multivariate flood frequency analysis focusing on flood
It can be easily shown that the structure-based and design event-based peaks and volumes, Q and V, for a streamgauge of the river Rhine and they
approaches coincide for the univariate case, i.e., when z=g(x). show the impacts in terms of AND JRP [i.e., Eq. (78.3.2b)], and associated

78_Singh_ch78_p78.1-78.10.indd 8 01/07/16 4:29 PM

 REFERENCES    78-9 

design events (q, v), of different hypotheses on the stationarity/nonstationar- OXFORD: Elsevier, United Kingdom, 2011, pp. 479–517, ISBN: 978-0-444-
ity of the dependence structure between Q and V, as well as of the position, 53199-5.
scale, and shape parameters of the Generalized Extreme Value distributions Gumbel, E. J., The calculated risk in flood control, Applied Science Research,
(see Chp. XXX in this volume) adopted for representing the Q and V The Hague, The Netherlands, Vol 5A, 1955, pp. 272–280.
marginals. Gumbel, E. J., The return period of flood flows, Annals of Mathematical
Statistics, 12 (2): 163–190, 1941.
Gumbel, E. J., “The return period of order statistics,” Annals of the Institute
REFERENCES
of Statistical Mathematics, 12 (3): 249–256, 1961.
Bender, J., T. Wahl, and J. Jensen, “Multivariate design in the presence of Harberg, R. J., Planning and Managing Reliable Urban Water Systems,
non-stationarity,” Journal of Hydrology, 514: 123–130, 2014, http://dx.doi. American Water Works Association, Denver, Co, 1997.
org/10.1016/j.jhydrol.2014.04.017. Hashimoto, T., J. R. Stedinger, and D. P. Loucks, “Reliability, resiliency, and
Bowling, L. C., P. Storck, and D. P. Lettenmaier, “Hydrologic effects of vulnerability criteria for water resource system performance evaluation,”
logging in western Washington, United States,” Water Resources Research, Water Resources Research, 18 (1): 114–20, 1982.
36 (11): 3223–3240, 2000. Helsel, D. R. and R. M. Hirsch, Statistical methods in water resources,
Bras, R. L., Hydrology: An Introduction to Hydrologic Science, Addison- Techniques of Water-Resources Investigations, Book 4, Chapter A3, U.S.
Wesley, Reading, MA., 1990, p. 643. Geological Survey, Reston, VA, 2002.
Castellarin, A., “Probabilistic envelope curves for design-flood estimation Hirsch, R. M., “Synthetic hydrology and water supply reliability,” Water
at ungaged sites,” Water Resources Research, 43: W04406, 2007. Resources Research, 15 (6): 1603–1615, 1979.
Castellarin, A., R. M. Vogel, and N. C. Matalas, “Probabilistic behavior of a Hobbs, B. F., P. T. Chao, and B. N. Venkatesh, “Using decision analysis to
regional envelope curve,” Water Resources Research, 41: W06018, 2005, include climate change in water resources decision making,” Climatic Change,
doi:10.1029/2004WR003042. 37: 177–202, 2007.
CEH, Flood Estimation Handbook, five printed volumes, Centre for Ecol- Holmes, R. R. Jr. and K. Dinicola, “100-Year flood–it’s all about chance,”
ogy & Hydrology, U.K., 1999, ISBN: 9781906698003. U.S. Geological Survey General Information Product, 106: 1, 2010.
Chao, P. T. and B. F. Hobbs, “Decision analysis of shoreline protection Hosking, J. R. M. and J. R. Wallis, “The effect of intersite dependence on
under climate change uncertainty,” Water Resources, 33 (4): 817–829, 1997. regional flood frequency-analysis,” Water Resources Research, 24 (4):
Chen, L., V. P. Singh, S. Guo, A. K. Mishra, and J. Guo, “Drought analysis 588–600, 1988.
using copulas,” Journal of Hydrologic Engineering, 18 (7): 797–808, 2013. IACWD (Interagency Committee on Water Data), Guidelines for deter-
Chow, V. T., Handbook of Applied Hydrology, McGraw Hill, New York, 1964. mining flood flow frequency, Bulletin 17B, Office of Water Data Coordina-
Cohen, J., “The earth is round (p < .05),” American Psychologist, 49 (12): tion, U.S. Geological Survey, Reston, VA, 1982.
997–1003, 1994. Katz, R. W., “Towards a Statistical Paradigm for Climate Change,” Climate
Cohn, T. A. and H. F. Lins, “Nature’s style: Naturally trendy,” Geophysical Research, 2: 167–175, 1993.
Research Letters, 32 (23): 1–5, 2005. Katz, R. W., “Statistics of extremes in climate change,” Climatic Change, 100:
Cooley, D., “ Extreme value analysis and the study of climate change: a 71–76, 2010.
commentary on Wigley 1988,” Climate Change, 97: 77–83, 2009. Klein, B., M. Pahlow, Y. Hundecha, and A. Schumann, “Probability analysis
Cooley, D., Return periods and return levels under climate change, Chapter of hydrological loads for the design of flood control systems using copulas,”
4, AghaKouchak, A., D. Easterling, K. Hsu, S. Schubert, and S. Sorooshian, Journal of Hydrologic Engineering (ASCE), 15: 360–369, 2010.
Extremes in a Changing Climate: Detection, Analysis and Uncertainty, edited Kottegoda, N. T. and R. Rosso, Statistics, Probability, and Reliability for Civil
by AghaKouchak, A., D. Easterling, K. Hsu, S. Schubert, and S. Sorooshian, and Environmental Engineers, McGraw-Hill, New York, 1997.
Springer Science + Business media, Dordrecht, 2013. H-H, Kwon, C. Brown, and U. Lall, “Climate informed flood frequency
Corbella, S. and D. Stretch, “Multivariate return periods of sea storms for analysis and prediction in Montana,” Geophysical Research Letters, 35, L05404,
coastal erosion risk assessment,” Natural Hazards and Earth System Sciences, doi:10.1029/2007GL032220, 2008.
12: 2699–2708, 2012, doi:10.5194/nhess-12-2699-2012. Lettenmaier, D. P., “Detection of trends in water quality data from records
Cudworth, A. G., Flood Hydrology Manual: A Water Resources Technical with dependent observations,” Water Resources Research, 12: 1037–1046, 1976.
Publication, Bureau of Reclamation, Denver, CO, 1989, p. 243. Li, C., V. P. Singh, and A. K. Mishra, “A bivariate mixed distribution with a
Douglas, E. M., R. M. Vogel, and C. N. Kroll, “Impact of streamflow heavy-tailed component and its application to single-site daily rainfall simula-
persistence on hydrologic design,” Journal of Hydrologic Engineering, ASCE, 7 tion,” Water Resources Research, 49 (2): 767–789, 2013.
(3): 220-227, 2002. Li, N., X. Liu, W. Xie, J. Wu, and P. Zhang, “The return period analysis of
Dupont, W. D. and W. D Plummer Jr., “Power and sample size calculations: natural disasters with statistical modeling of bivariate joint probability
a review and computer program,” Controlled Clinical Trials, 11: 116–128, 1990. distribution,” Risk Analysis, 33 (1): 134–145, 2013.
Dupont, W. D. and W. D Plummer, “Power and sample size calculations for Lloyd, E. H., “Occurrence interval in the presence of persistence,” Journal
studies involving linear regression,” Controlled Clinical Trials, 19: 589–601, 1998. of Hydrology, 10 (3): 291–298, 1970.
Favre, A-C., S. El Adlouni, L. Perreault, N. Thiemonge, and B. Bobee, “Mul- Loucks, D. P. and E. van Beek, Water Resource Systems Planning and
tivariate hydrological frequency analysis using copulas,” Water Resources Management: An Introduction to Methods, Models and Applications, UNESCO,
Research, 40: 1–12, 2004. WL – Delft Hydraulics, 2005, p 680, Paris, France.
Fernandez, B. and J. Salas, “Return period and risk of hydrologic events. I: Mandelbaum, M., M. Hlynka, and P. H. Brill, Nonhomogeneous geometric
Mathematical formulation, Journal of Hydrologic Engineering, 4 (4): distributions with relations to birth and death processes, TOP, , 15:281–296,
297–307, 1999. DOI 10.1007/s11750-007-0018-z, 2007.
Fiering, M. B. and N. C. Matalas, “Decision-making under uncertainty,” Manning, L. J., J. W. Hall, H. J. Fowler, C. G. Kilsby, and C. Tebaldi, “Using
Climate Change and U.S. Water Resources, edited by P. E. Waggoner, John probabilistic climate change information from a multimodel ensemble for
Wiley & Sons, New York, 1990. water resources assessment,” Water Resources Research, 45 (11): 1–13, 2009.
Fuller, W. E., “Flood flows,” Transactions of the American Society of Civil Matalas, N. C. and W. B. Langbein, “Information content of the mean,”
Engineers, 77: 564-617, 1914. Journal of Geophysical Research, 67 (9): 3441–3448, 1962.
Genest, C. and A. C. Favre, “Everything you always wanted to know about Matalas, N. C., “Comment on the announced death of stationarity,” Journal
copula modeling but were afraid to ask,” Journal of Hydrologic Engineering of Water Resources Planning and Management, 138: 311–312, 2012,
(ASCE), 12: 347–368, 2007. doi:10.1061/(ASCE)WR.1943-5452.0000215.
Gersonius, B., R. Ashley, A. Pathirana, and C. Zevenbergen, “Climate Mays, L. W., Water Resources Engineering, John Wiley & Sons, Second Edi-
change uncertainty: building flexibility into water and flood risk infrastruc- tion, 842. 2011.
ture,” Climatic Change, 116: 411–423, 2013. P. C. D. Milly, J. Betancourt, M. Falkenmark, R.M. Hirsch, Z. W. Kundze-
Gräler, B., M. J. van den Berg, S. Vandenberghe, A. Petroselli, S. Grimaldi, wicz, D. P. Lettenmaier, R. J. Stouffer “Climate change—stationarity is dead:
B. De Baets, and N. E. C. Verhoest, “Multivariate return periods in hydrology: Whither water management?” Science, 319 (5863): 573–574, 2008.
a critical and practical review focusing on synthetic design hydrograph Modarres, M., M. Kaminsky, and V. Krivtsov, Reliability Engineering and
estimation,” Hydrology and Earth System Sciences, 17: 1281–1296, 2013, Risk Analysis: A Practical Guide, CRC Press, Taylor and Francis, London, 2009.
doi:10.5194/hess-17-1281-2013. Morin, E., “To know what we cannot know: global mapping of minimal
Grimaldi, S., S. C. Kao, A. Castellarin, S. M. Papalexiou, A. Viglione, F. detectable absolute trends in annual precipitation,” Water Resources Research,
Laio, H. Aksoy, et al., Statistical hydrology, Treatise on Water Science, 47 (7): 1–9, 2011.

78_Singh_ch78_p78.1-78.10.indd 9 01/07/16 4:29 PM

 78-10     Risk, Reliability, and Return Periods and Hydrologic Design

Montanari, A. and D. Koutsoyiannis, “Modeling and mitigating natural tropical moisture exports,” Water Resources Research, 51: 1472–1492, 2015,
hazards: Stationarity is immortal!” Water Resources Research, 50: 9748–9756, doi:10.1002/2014WR016664.
2014, doi:10.1002/ 2014WR016092. Tasker, G. D., “Effective record length for the T-year event,” Journal of
National Earthquake Hazard Reduction Program (NEHRP) Recommended Hydrology, 64: 39–47, 1983.
Seismic Provisions for New Buildings and Other Structures, Federal Emergency Thomas, H. A., “Frequency of minor floods,” Journal of the Boston Society
Management Agency, Washington, D.C., 2010, pp. 1–406. of Civil Engineers, 35 (1): 425–442, 1948.
National Research Council (NRC), Estimating Probabilities of Extreme Trafimow, D. and M. Marks, Editorial, Basic and Applied Social Psychology,
Floods: Methods and Recommended Research, National Academy Press, 37: 1–2, 2015.
Washington, D.C., 1988, p. 139. Trenberth, K. E., “Attribution of climate variations and trends to human
National Research Council, Risk Analysis and Uncertainty in Flood Damage influences and natural variability,” WIREs Climate Change, 2: 925–930, 2011.
Reduction Studies, The National Academies Press, Washington, D.C., 2000. Tung, Y-K., “Flood defense systems design by risk-based approaches,”
Nicholls, N., “The insignificance of significance testing,” Bulletin of the Water International, 30 (1): 50–57, 2005.
American Meteorological Society, 81 (5): 981–986, 2000. Tung, Y-K., Risk/Reliability-Based Hydraulic Engineering Design, Chapter 7,
Olsen, J. R., J. H. Lambert, and Y. Y. Haimes, “Risk of extreme events under Hydraulic Design Handbook, edited by L. Mays, McGraw Hill, New York, 1999.
nonstationary conditions,” Risk Analysis, 18 (4): 497–510, 1998. Viessman, W. Jr. and G. L. Lewis, Introduction to Hydrology, 5th ed.,Prentice
Onoz, B. and M. Bayazit, “The power of statistical tests for trend detection,” Hall, Upper Saddle River, New Jersey 2003.
Turkish Journal of Environmental Engineering Science, 27: 247–251, 2003. Villarini, G., J. A. Smith, F. Serinaldi, J. Bales, P. D. Bates, and W. F. Krajew-
Parey, S., F. Malek, C. Laurent, and D. Dacunha-Castelle, “Trends and ski, “Flood frequency analysis for nonstationary annual peak records in an
climate evolution: statistical approach for very high temperatures in France,” urban basin,” Advances in Water Resources, 32: 1255–1266, 2009.
Climatic Change, 81: 331–352, 2007. Vogel, R. M., “Reliability indices for water supply systems,” Journal of Water
Parey, S., T. T. H. Hoang, and D. Dacunha-Castelle, “Different ways to Resources Planning and Management, 113 (4): 563–579, 1987.
compute temperature return levels in the climate change context,” Environ- Vogel, R. M. and C. N. Kroll, “The value of streamflow record augmenta-
metrics, 21: 698–718, 2010. tion procedures in low flow and flood-flow frequency analysis,” Journal of
Potter, K. W., “Estimating the probability of annual maximum water levels Hydrology, 125: 259–276, 1991.
on the Great Lakes,” Journal of Great Lakes Research, 18 (1): 229–235, 1992. Vogel, R. M., N. C. Matalas, J. F. England, and A. Castellarin, “An assess-
Prosdocimi, I., T. R. Kjeldsen, and C. Svensson, “Non-stationarity in annual ment of exceedance probabilities of envelope curves,” Water Resource
and seasonal series of peak flow and precipitation in the U.K.,” Natural Research, 43: W07403, 2007, doi:10.1029/2006WR005586.
Hazards and Earth System Sciences, 14: 1125–1144, 2014. Vogel, R. M., C. Yaindl, and M. Walter, “Nonstationarity: flood magnifica-
Read, L. and R. M. Vogel, “Risk, reliability, and return periods under non- tion and recurrence reduction factors in the United States,” Journal of Ameri-
stationarity,” Water Resources Research (under review), 2015. can Water Resources Association, 47 (3): 464–474, 2011.
Renard, B. and U. Lall, “Regional frequency analysis conditioned on large- Vogel, R. M., A. Rosner, and P. H. Kirshen, “Likelihood of societal
scale atmospheric or oceanic fields,” Water Resources Research, 50: 9536–9554, preparedness for global change—trend detection,” Natural Hazards and Earth
2014, doi:10.1002/ 2014WR016277. System Sciences, Brief Communication, 13: 1–6, 2013.
Requena, A. I., L. Mediero, and L. Garrote, “A bivariate return period based Vogel, R. M., U. Lall, X. Cai, B. Rajagopalan, P. Weiskel, R. P. Hooper, and
on copulas for hydrologic dam design: accounting for reservoir routing in risk N. C. Matalas, “Hydrology: the interdisciplinary science of water,” Water
estimation,” Hydrology and Earth System Sciences, 17: 3023–3038, 2013, Resources Research 51, 2015, doi:10.1002/2015WR017049.
doi:10.5194/hess-17-3023-2013. Volpi, E. and A. Fiori, “Hydraulic structures subject to bivariate hydrologi-
Rosner, A., R. M. Vogel, and P. H. Kirshen, “A risk based approach to flood cal loads: return period, design, and risk assessment,” Water Resources
management in a nonstationary world,” Water Resources Research, 50: 1928- Research, 50: 885–897, 2014, doi:10.1002/2013WR014214.
1942, 2014, doi:10.1002/2013WR014561. Volpi, E. and A. Fiori, “Design event selection in bivariate hydrological
Sadri, S. and D. H. Burn, “Copula-based pooled frequency analysis of frequency analysis,” Hydrological Sciences Journal, 57: 1506–1515, 2012, doi:1
droughts in the Canadian Prairies,” Journal of Hydrologic Engineering, 19 (2): 0.1080/02626667.2012.726357.
277–289, 2014. Wigley, T. M. L., “The effect of changing climate on the frequency of
Salas, J. D. and J. Obeysekera, “Revisiting the concepts of return period and absolute extreme events,” Climatic Change, 97: 67–76, 2009 (reprinted from
risk for nonstationary hydrologic extreme events,” Journal of Hydrologic Engi- his 1988 article in Climate Monitor).
neering, 19: 554–568, 2014. Wilks, S. S., “Recurrence of extreme observations,” Journal of the Australian
Salvadori, G. and C. De Michele, “Frequency analysis via copulas: theoreti- Mathematical Society, 1: 106–112, 1959.
cal aspects and applications to hydrological events,” Water Resources Research, Yen, B. C., “Risks in hydrologic engineering projects,” Journal of the
40: W12511, 2004, doi:10.1029/2004WR003133. Hydraulics Division, ASCE, 96 (4): 959–966, 1970.
Salvadori, G., C. De Michele, and F. Durante, “On the return period and Yue, S. and P. Rasmussen, “Bivariate frequency analysis: discussion of some
design in a multivariate framework,” Hydrology and Earth System Sciences, 15: useful concepts in hydrological application,” Hydrological Processes, 16:
3293–3305, 2011, doi:10.5194/hess-15-3293-2011. 2881–2898, 2002.
Salvadori, G., C. De Michele, N. T. Kottegoda, and R. Rosso, Extremes Yue, S. and P. Pilon, “A comparison of the power of the t test, Mann- Kendall
in Nature: An approach Using Copulas, Springer, Dordrecht, The Nether- and bootstrap tests for trend detection,” Hydrological Sciences Journal, 49 (1):
lands, 2007. 21–37, 2004.
Sen, Z., “Simple risk calculations in dependent hydrological series,” Hydro- Yue, S., P. Pilon, and G. Cavadias, “Power of the Mann-Kendall and Spear-
logical Sciences Journal, 44 (6): 871–878, 1999. man’s rho tests for detecting monotonic trends in hydrological series,” Journal
Serinaldi, F., “Dismissing return periods!” Stochastic Environmental of Hydrology, 259: 254–271, 2002.
Research and Risk Assessment, 29: 1179–1189, 2015, doi 10.1007/ Zhang, L. and V. P. Singh, “Trivariate flood frequency analysis using the
s00477-014-0916-1. Gumbel–Hougaard copula,” Journal of Hydrologic Engineering (ASCE), 12:
Sklar, A., “Fonctions de Répartition à n Dimensions et Leurs Marges,” Publi- 431–439, 2007.
cations de l’Institut de Statistique de L’Université de Paris, Vol. 8, 1959, Ziegler, A. D., J. Sheffield, E. P. Maurer, B. Nijssen, E. F. Wood, and D. P.
pp. 229–231. Lettenmaier, “Detection of intensification in global- and continental-scale
Stakhiv, E. Z., “Pragmatic approaches for water management under climate hydrological cycles: temporal scale of evaluation,” Journal of Climate, 16 (3):
change uncertainty,” Journal of American Water Resources Association, 47 (6): 535–547, 2003.
1183–1196, 2011, doi:10.1111/j.1752-1688.2011.00589.x. Ziegler, A. D., E. P. Maurer, J. Sheffield, B. Nijssen, E. F. Wood, and D. P.
Stedinger, J. R., “Estimating a regional flood frequency distribution,” Water Lettenmaier, “Detection time for plausible changes in annual precipitation,
Resources Research, 19: 503–510, 1983. evapotranspiration, and streamflow in three Mississippi river sub-basins,”
Stedinger, J. R., R. M. Vogel, and E. Foufoula-Georgiou, Frequency analysis Climatic Change, 72: 17–36, 2005.
of extreme events, Chapter 18, Handbook of Hydrology, edited by David Ziliak, S. T. and D. N. McCloskey, The Cult of Statistical Significance: How
R. Maidment, McGraw-Hill, New York, N.Y., 1993. the Standard Error Costs Us Jobs, Justice, and Lives, University of Michigan
Steinschneider, S. and U. Lall, “A hierarchical Bayesian regional model for Press: Ann Arbor, 2008.
nonstationary precipitation extremes in Northern California conditioned on

78_Singh_ch78_p78.1-78.10.indd 10 01/07/16 4:29 PM