CHAPTER 8

Low Flows and Droughts

Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Jose D. Salas
Charles N. Kroll
Antonino Cancelliere
Bonifacio Fernández
Jose A. Raynal
Dong R. Lee

8.0 INTRODUCTION

In other chapters, some features of the hydrologic cycle dealing with the occurrence of extreme
events such as extreme maximum precipitation and extreme maximum floods have been addressed.
This chapter addresses another feature of the hydrologic cycle that deals with the opposite extreme
or the other end of the spectrum, in other words, the occurrence of periods of low flows and drought.
Whereas a flood is quite visible and generally gets prompt attention from the public, the media, and
other institutions, a drought may not receive similar attention because it may take a while before
its impacts and consequences become noted. The effects of periods of low flows and drought can be
quite significant and sometimes devastating to the environment and to society. In the following
sections we describe some concepts and definitions associated with low flow and drought
phenomena and present some tools for analyzing them. Because of the random nature of such
phenomena, the methods are based on probabilistic and stochastic concepts. Several problems in
water resources and environmental engineering require the estimation of some low flow quantities
and the estimation of drought properties. For example, the design of the storage capacity of a
reservoir for water supply (conservation) is related to the occurrence of periods of low flows and
drought. Low-flow characteristics of streams are commonly used in the planning and management
of water resources systems, such as in designing water supply works, analyzing environmental
impacts of water resources development, modeling stream water quality, regulating instream water
uses, and improving the general understanding of natural regulated stream systems (Tasker 1987).
Also, certain water quality standards are based on specific low-flow quantiles.
Low flow at a site is often characterized by an index of low flow. For example, a widely used
index of low flow is the 7 day, 10 year low flow, which is the discharge having a 10 year return period
derived from a frequency curve of the lowest average flow for seven consecutive days in a year (Riggs
1980). In some cases, low-flow data may be serially uncorrelated. In those cases, one may apply the
usual procedures of frequency analysis and estimate useful statistics such as return period and risk.
However, when the time series of low flows are time dependent, the usual frequency analysis
assuming independence does not apply nor do the methods for estimating the return period and risk.
In this chapter we consider both cases. In addition, this chapter includes characterizing droughts by
stochastic methods. Drought is generally a complex phenomenon, which may involve a wide range

269

Statistical Analysis of Hydrologic Variables

 270 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

of temporal and spatial scales, and several hydrometeorological variables such as precipitation,
temperature, soil moisture, streamflow, groundwater levels, and water storages may be necessary for
drought identification and characterization depending on the particular problem at hand. In
addition, many other variables may be necessary for drought forecasting (e.g., Wong et al. 2013)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

and to quantify drought impacts and consequences on the environment and society. As the title of
the chapter implies, the emphasis here is based on streamflow as the key variable of analysis, and in
the case of drought characterization the focus is on multiyear droughts. A vast amount of literature is
available for analyzing various aspects of droughts such as definition, identification, prediction,
impacts, adaptation, and management (e.g., Alley 1984, Frick et al. 1990, Guttman 1998, Wilhite
2000, Svoboda 2000, Heim 2002, Panu and Sharma 2002, Salas et al. 2005, Bond et al. 2008, Mishra
and Singh 2010, Sheffield et al. 2012a).
Following this introductory section, the next section describes definitions of low flow and
drought. Then Section 8.2 introduces some elementary concepts and procedures for determining the
frequency analysis of low flows using empirical methods, and Section 8.3 describes in some detail
frequency analysis using traditional univariate distribution functions including the case of inter-
mittent low-flow data. Section 8.4 includes the case of regional analysis of low flows. Section 8.5
describes the analysis of autocorrelated low flows based on low-order discrete and continuous
stochastic models, and Section 8.6 discusses the characterization of multiyear droughts based on
analytical approximations. The sections include the estimation of return periods of low flows and
droughts, respectively. The case of regional droughts is briefly discussed in Section 8.7, and the
effects of hydraulic structures on low flows are briefly described in Section 8.8. The chapter ends with
a section of closing remarks, which includes some concepts and references on the effect of climate
variability on low flows and droughts.

8.1 LOW FLOW AND DROUGHT DEFINITIONS

Before statistically characterizing low flows and droughts, it is necessary to define them precisely.
Since both terms relate to those conditions where streamflow reaches some undesirable levels, there
are some similarities as well. In the following two sections, we include definitions that have
been suggested in literature, and indicate those cases where some similarities and differences arise
between them.

8.1.1 Definitions of Low Flows

Several definitions of low flows are found in the hydrologic literature. Two of the most popular ones
utilized in practice are called low flow and low-flow duration. Sometimes the terms low-flow volume,
low-flow discharge, and low-flow stage are utilized to emphasize that the underlying variable is either
volume, discharge, or stage, respectively. The term low flow refers to the minimum flow of a stream
over a consecutive number of days that may occur during a given time period (generally a year). The
term low-flow duration is defined as the number of consecutive days in which the flow (series) is
below a certain threshold value or crossing level. In either case, the referred quantities are random
variables so one must describe them in frequency terms.
The definition of low flow depends on the specified time duration d and the Tu, as shown in
Figure 8-1. Assuming a daily streamflow hydrograph, typical values of d are 1 day, 5 days, 7 days,
10 days, and higher, depending on the problem, while the unit time period Tu is generally a year or a
fraction of a year (e.g., the dry season). The year is the unit time period most commonly used for
most problems involving low-flow analysis (U.S. Army Corps of Engineers 1964). Figure 8-1 shows
schematically the definition of flow volume vi and low-flow discharge qi for the specified time
duration d. Thus, one can obtain m values of v1 , : : : , vm and q1 , : : : , qm from the daily hydrograph.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 271

Qt
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

v1 qi vm
vi
t
d d d
Unit time period
Figure 8-1. Definitions of time interval, d, unit time period Tu, flow volume v, and flow discharge q.
Source: Salas et al. (2018).

For example, assuming d = 10 days and Tu = 365 days, one can get m = 356 values of both vi and
qi (considering overlapping). Thus, the low-flow volume for the given year, say year 1, is
V 1 = minðv1 , : : : , vm Þ and the low-flow discharge is y1 = minðq1 , : : : , qm Þ. Considering N years of
record (i.e., total number of time periods Tu) we will have the sequences of low-flow volumes
V 1 , : : : , V N and low-flow discharges y1 , : : : , yN that can be used for frequency analysis. Example 8-1
illustrates some of these concepts.
Another type of low-flow index has been used in hydrologic practice where the duration of low
flows (and related quantities) is the variable of concern. Low-flow duration is defined as the number
of consecutive time intervals (e.g., days) in which the flow series is below a certain threshold or
crossing level. The low-flow spell has an associated low-flow volume (or cumulative flow deficit),
which is the accumulated flow below the crossing level. The low-flow intensity is defined as the ratio
of the low-flow volume and the duration of the low-flow spell. This type of analysis is called “low-
flow duration analysis” (US Army Corps of Engineers 1975, Viessman et al. 1989). This is also
referred to as “low-flow spells frequency analysis” (Shaw 1988).
Referring to Figure 8-2, let us assume that Qt is a daily flow hydrograph (although the time
interval may be hours, days, weeks, etc.) A low flow occurs when Qt is below the crossing level
(discharge) Qo for a consecutive number of days or duration d. The figure shows four low-flow
spells during the unit period Tu. In general, we may have m low-flow episodes during the unit time

Qt

d d d Q0
i3=v3 v

Unit time period

Figure 8-2. Definition of low-flow duration d, low-flow deficit v, and low-flow intensity I relative to
the threshold discharge Q0. The graph shows four episodes of low flows.
Source: Salas et al. (2018).

Statistical Analysis of Hydrologic Variables

 272 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

period Tu, with durations, d1, : : : , dm, low-flow deficits v1 , : : : , vm , and low-flow intensities i1 , : : : , im .
Note, however, that in some cases, low-flow spells may not occur in a particular year, and in those
cases, the low-flow variables equal zero. Generally for most water resources applications we are
interested in the maximum values, in other words., the longest dry period D, the largest deficit V, and
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

the largest intensity I, each of which is a random variable. Thus, for N years of data available and
Tu = 1 year, we will have the sequences of low-flow durations, D1, : : : , DN, low-flow deficits
V1, : : : , VN, and low-flow intensities I1, : : : , IN. The frequency analysis of each sequence will provide
the information needed to make probability statements about the variable of interest.
The various aforementioned low-flow definitions are particularly useful in perennial rivers.
However, in streams of semi-arid and arid regions where the flow regime may be intermittent and
zero values of flows may occur, appropriate adjustments may be needed for determining the
corresponding frequency distribution as outlined in Section 8.3.2. In addition, other characteristics
of low flows such as the duration of zero flows (e.g., considering Qo = 0 in Figure 8-2) may be of
interest, and determining the frequency distribution of the longest duration of zero flows (per year)
may be needed. Also another way of analyzing low flows in streams of intermittent flow regime may
be selecting an appropriate time scale of the hydrograph under consideration, for example, weekly
instead of daily, so as to avoid the zero flow values, then the low-flow analysis for such a scale can be
performed as described previously.

8.1.2 Definition of Drought

A drought is a complex phenomenon that evolves through time and space in a random fashion. It
may be characterized by its initiation, duration, magnitude (accumulated deficit), intensity,
termination, and spatial extent. These characteristics may be determined by comparing a water
supply time series with a water demand series. Because water supply quantities such as streamflows
are stochastic variables the corresponding drought characteristics are also random and must be
described using stochastic methods. Let us consider a stochastic water supply series denoted by
xt , t = 1, 2, : : : and a constant water demand threshold denoted by x0 . Following the drought
definition that Yevjevich (1967) suggests, a drought event is taken as a succession of consecutive
periods (run) in which the water supply xt remains below x0 where the run is preceded and
succeeded by water supply that is equal to or bigger than x0 . Thus, the drought length L (or negative
run length) is the number of consecutive time intervals (e.g., years) in which xt < x0 , preceded and
followed by (at least one period where) xt ≥ x0 . Figure 8-3 schematically shows this drought
definition. In addition, the drought magnitude P or accumulated deficit (run sum) is the total deficit
throughout the drought duration [i.e., D = tþL−1 j = t ðx 0 − x j Þ], and the drought intensity is the mean
deficit over the drought duration, (i.e., I = D/L). Furthermore, d0 ≥ 0 represents any given drought
magnitude, and i0 > 0 represents any given value of the intensity. To analyze the severity of droughts
and the associated risks, we need to specify the drought event under consideration. For instance, one
may consider only the duration of a drought regardless of the magnitude, or drought duration with a
certain degree of deficit, or duration and a given intensity, and so on.
Worth noting are the similarity and differences between the definitions of low-flow duration
(deficit and intensity) depicted in Figure 8-2 and the definition of drought duration (magnitude and
intensity) shown in Figure 8-3. The similarity of the definitions is fairly obvious, but the main
difference is in the ensuing analysis of the data arising from both definitions (figures). While the low-
flow duration (and magnitude and intensity) in Figure 8-2 yields a maximum value per year, and the
set of maximums is analyzed using frequency analysis methods (Sections 8.2 and 8.3), the drought
length/duration (and magnitude and intensity) in Figure 8-3 give information that can be analyzed
statistically but generally. Because of the small number of drought events that can be obtained from
the underlying flow series, the analysis involves analytical approximations and stochastic simulation
(Section 8.6). Furthermore, also note that in some cases where low flows are autocorrelated also,
stochastic modeling approaches may have to be applied (Section 8.5).

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 273

800

700
L=4 L=2
600
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

D
500
xt

400
I i0
300
x0
200
d0
100

0
Time t
Figure 8-3. Water supply xt, water demand x0, and drought properties: (a) duration L, (b) magnitude
(accumulated deficit) D, and (c) intensity I. Also specific drought magnitude and intensity thresholds
d0 and i0 are shown.
Source: Salas et al. (2005).

8.2 EMPIRICAL FREQUENCY ANALYSIS OF LOW FLOWS

The empirical frequency distribution of low flows may be determined using a certain plotting
position formula. The procedure is straight forward but will be summarized here for completeness.
Let us assume that the sample of the low-flow variable of interest is denoted as y1 , : : : , yN , where N is
the sample size. We will also assume that the sample has been arranged from the smallest value to
the largest one so that y1 represents the minimum and yN the maximum. Then, an estimate of the
cumulative probability corresponding to the ordered value yi is given by i∕ðN þ 1Þ, which is the
Weibull plotting position formula. The literature has suggested alternative plotting position formulas
(e.g., Stedinger et al. 1993), but for the purpose of this chapter we will use the Weibull formula.
Because we assumed that the sample is ordered from the smallest to the largest, one can write

FðiÞ = PðY ≤ yi Þ = i∕ðN þ 1Þ, i = 1, : : : , N (8-1)

in which F(i) is the nonexceeding probability. Thus Equation (8-1) gives the estimate of the
cumulative distribution function (CDF). This distribution is also referred to as the empirical CDF or
empirical frequency curve.
This formula is quite simple, but one must be aware of special cases where appropriate
interpretation and adjustments may be necessary. For instance, if the sample is a random sample in
the sense that the sequence of observations is independent (uncorrelated), then one can use the
nonexceedance probability F(i) = q to determine the return period of the value of interest. Per
illustration, if we are concerned with determining the return period of, say yi , one can calculate it by
T = 1∕FðiÞ = 1∕q. Note that sometimes F(i) = q has been called the exceedance probability,
particularly in relation to low flows. However, in this chapter we will stick to the usual terminology
and jargon commonly found in statistical literature, which is FðiÞ = PðY ≤ yi Þ = q = nonexceedance
probability. Also, in cases where the observations are autocorrelated, applying the methods outlined
in Section 8.5 instead may be useful, and where the sample has been censored, appropriate
adjustments may be necessary as described in Section 8.3.2. In addition, the foregoing plotting
position formula may provide a quick estimate of the probabilities of low-flow events and return

Statistical Analysis of Hydrologic Variables

 274 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

periods, but generally employing appropriate probabilistic or stochastic methods is better. This is
even more so in the analysis of multiyear droughts where even records of 100 years may not be
enough for estimating drought properties reliably. This is the case, for example, for determining the
return period of multiyear droughts or for characterizing critical droughts. In these cases, applying
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

approximate analytical techniques or stochastic simulations such as those described in Sections 8.5
and 8.6 may be necessary.

8.3 PROBABILITY DISTRIBUTION OF LOW FLOWS

When a historical streamflow record is available at a site of interest, low streamflow quantiles can be
estimated via frequency analysis. The standard procedure is to pick a probability distribution that
adequately describes the low streamflow series, to estimate the parameters of the probability
distribution, and then to estimate the relevant quantile from the distribution. The following sections
describe fitting univariate distributions, how to handle sites with intermittent (nonperennial)
streamflows, and the regionalization of low flows and droughts.

8.3.1 Fitting of Univariate Distributions

The most challenging issue with estimating low-flow quantiles at sites with historical records is the
choice of an appropriate probability distribution to describe the low-flow series. The few studies that
investigate fitting probability distributions to low-flow series do not reach a consensus as to the best
probability distribution to employ in practice. Tasker (1987) uses a bootstrap resampling experiment at
20 Virginia rivers to analyze how the three-parameter Weibull (W3), log–Pearson type III (LP3), log
Boughton, and Box–Cox distributions reproduce streamflow quantiles, recommending the W3 and
LP3 distributions. Condie and Nix (1975) recommend the W3 distribution for best describing the
lower bound of Canadian rivers. Vogel and Kroll (1989) employ a regional probability plot correlation
coefficient (PPCC) test at rivers in Massachusetts to compare various probability distributions,
recommending the two- and three-parameter log–normal (LN2 and LN3), LP3, and W3 distributions.
Onöz and Bayazit (2001) use a PPCC test to examine the fit of probability to low flows at a several
European rivers and recommend the use of the general extreme value (GEV) distribution.
Pearson (1995) employs an L-moment analysis of 1 day annual minimum flows at more than
500 New Zealand rivers, concluding that no single two- or three-parameter distribution provides an
adequate fit. Vogel and Wilson (1996) also employ L-moment diagrams to examine 1 day annual
minimum flows at more than 1,400 river sites across the United States and recommend the use of the
three-parameter Pearson (P3) distribution based on a visual interpretation of the L-moment
diagrams. Kroll and Vogel (2002) use an L-moment weighted distance measure to compare the
fit of probability distributions to 7 day annual minimum flows at more than 1,500 river sites across
the United States. They recommend the LN3 distribution at perennial sites and the P3 at intermittent
sites due to patterns in the shift of L-moment ratios.
In this section, we illustrate the applications of the log–Pearson type III, log–normal three-
parameter (log–normal 3), the Weibull, and the GEV models for fitting the distribution of low-flow
variables. The fitting procedure is essentially based on the method of moments (MOM), although the
literature proposes several other methods, such as the maximum likelihood and probability weighted
moments. Detailed examples are included and in most cases the fitted and empirical CDFs are
compared graphically.

8.3.1.1 Fitting the Log–Pearson Type III Distribution

Only a limited number of probability distributions are commonly fit to low-flow series, and in the
United States this is typically the LP3 distribution. The Committee on Surface-Water Hydrology of

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 275

the Hydraulics Division (ASCE 1980) recommends fitting low-flow series with the LP3. In addition,
in Bulletin 17, the US Water Resources Council recommends the LP3 distribution to describe annual
maximum flows (USIAC 1982), and its common use to describe low flows is by default. The LP3 has
also been employed in several USGS studies. The LP3 is a flexible distribution and converges to an
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

LN2 distribution when the log skew approaches zero. Typically, the LP3 is parameterized by the
method of moments. Whereas weighted skew maps and generalized least squares regression
estimators of the skew have been recommended for flood flow series, no such tools have been
recommended for low flows, and typically at-site skew estimators are employed, even though they
have been shown to have high bias and variance in small samples (Stedinger et al. 1993).
If one defines a d day low streamflow series as x1, x2, : : : , xN, and the logarithm of this series as
y1, y2, : : : , yN, i.e., yi = ln(xi), the log space moments of the series are

X
N
y
^y =
μ i
(8-2)
i=1
N

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN ffi
i=1 i ðy − ^
μ y Þ2
^y =
σ (8-3)
N −1

P
^y Þ3
N Ni−1 ð yi − μ
^γy = (8-4)
ðN − 1ÞðN − 2Þ^ σ3y

The quantile of interest can then be estimated as

^xq = expð^ ^y Þ
μy þ K q σ (8-5)

where Kq is the frequency factor that may be obtained from tables (e.g., USIAC 1982) or can be
estimated approximately using the Wilson–Hilferty transformation as
 
2 zq ^γy ^γ2y 3 2
Kq = 1þ − − (8-6)
^γy 6 36 ^γy

in which zq is the inverse standard normal variate such that P(Z ≤ zq) = q where Z ∼ N(0, 1). This
approximation is generally considered adequate when 0.01 ≤ q ≤ 0.99 and |γ| < 2 (Kirby 1972).
When ^γy → 0, Kq → zq, and the LP3 distribution turns into an LN2 distribution.

Example 8-1: Estimation of the 7 Day, 10 Year Low Streamflow, × 7,10 Based on the
Log–Pearson Type III Distribution
Table 8-1 contains 7 day annual minimum flows for Penns Creek at Penns Creek, PA (USGS
#01555000), for the 76 year period from 1930 to 2005. The 7 day annual minimum is the lowest
average daily streamflow over a 7 day period during the water year. The water year for low
streamflow series typically begins during the high flow period of the year so that a single low-flow
period is not included in two consecutive years. Here the water year was defined as April 1 to
March 31.
Using these data, the following statistics were calculated based on Equations (8-2), (8-3), and
^y = 4.13, σ
(8-4), respectively: μ ^y = 0.43, ^γy = 0.77. To estimate the 7 day low streamflow that has a

Statistical Analysis of Hydrologic Variables

 276 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Table 8-1. 7 Day Annual Minimum Flows for Penns Creek at Penns Creek, PA.

7 day min 7 day min 7 day min 7 day min

Year (ft3/s) Year (ft3/s) Year (ft3/s) Year (ft3/s)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

1930 30 1949 55 1968 56 1987 60

1931 46 1950 73 1969 69 1988 58
1932 39 1951 54 1970 63 1989 84
1933 82 1952 47 1971 62 1990 171
1934 70 1953 43 1972 79 1991 52
1935 44 1954 42 1973 84 1992 82
1936 35 1955 52 1974 88 1993 72
1937 69 1956 166 1975 78 1994 117
1938 56 1957 47 1976 85 1995 51
1939 37 1958 61 1977 103 1996 107
1940 61 1959 43 1978 72 1997 82
1941 46 1960 65 1979 151 1998 39
1942 67 1961 43 1980 50 1999 50
1943 48 1962 33 1981 65 2000 62
1944 56 1963 42 1982 57 2001 53
1945 89 1964 31 1983 57 2002 55
1946 78 1965 34 1984 81 2003 221
1947 70 1966 24 1985 49 2004 204
1948 47 1967 93 1986 66 2005 44

nonexceedance probability of 1/10 = 0.1, in other words, the 10th percentile of the distribution of
annual minimum flows, the appropriate frequency factor is the inverse of the standard normal
variate with a nonexceedance probability of 10%, in other words, P(Z ≤ z0.1) = 0.1. Employing a
table for a standard normal distribution, z0.1 = −1.282. Using this value, the frequency factor
obtained from Equation (8-6) is Kq = K0.1 = −1.17. From this the x7,10 can be estimated from
Equation (8-5) as

^x7,10 = ^x0.1 = exp½4.13 þ ð−1.17Þ × 0.43 = 37.6 ft3 ∕s

Note that if one instead fits a two-parameter log–normal distribution to the 7 day annual
minimum flows, then

^x7,10 = ^x0.10 = exp½^ ^y  = exp½4.13 þ ð−1.282Þ × 0.43 = 35.8 ft3 ∕s

μy þ z0.10 σ
^x7,10 = ^x0.10 = exp½^ ^y  = exp½4.13 þ ð−1.282Þ × 0.43 = 35.8 ft3 ∕s
μy þ z0.10 σ

8.3.1.2 Fitting the Three-Parameter Log–Normal Distribution (LN3)

The probability density function of the LN3 distribution is

 
1 ½lnðx − x0 Þ − μy 2
f ðxÞ = pffiffiffiffiffi exp − , x0 < x < ∞ (8-7)
ðx − x0 Þσy 2π 2σ2y

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 277

where x0, σy, and μy are the location, scale, and shape parameters, respectively. The CDF, defined as
FðxÞ = ∫xx0 f ðxÞ dx, cannot be integrated explicitly, thus, numerical procedures or tables for the
normal distribution must be used to calculate either probabilities for a given value of x or quantiles
for a given probability.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

The moment estimators of the parameters of the LN3 distribution are (Yevjevich 1972)

1X N
^y =
μ lnðxi − ^x0 Þ, (8-8)
N i=1

 X
N 1∕2
^y = ð1∕NÞ
σ ½lnðxi − ^x0 Þ − μ
^y 2
, (8-9)
i=1

 
^x ω1∕3
η
^x0 = μ
^x 1 − , (8-10)
ð1 − ω2∕3 Þ
in which

^x
σ
^x =
η (8-11)
^x
μ
and

1 
ω= −^γx þ ð^γ2x þ 4Þ1∕2 (8-12)
2
where μ^x , σ
^x , and ^γx are respectively the mean, standard deviation, and the skewness coefficient of
the x’s. Note that the moment solution requires that ^γx > 0.

Example 8-2: Estimation of the 1 Day, 10 Year Low Flow × 1,10 Based on the LN3 Distribution
Table 8-2 gives the 1 day low flows for the San Pedro River at the Villalba gauging station in Mexico
for 1939–1991. The results are obtained following the method of moments procedure described
previously. Using the data of Table 8-2 the following 1 day low flow statistics are obtained:

1XN
^x =
μ x = 0:3306 m3 ∕s
N i=1 i
 X 1∕2
1 N
^x =
σ ðx − μ^x Þ2
= 0:1465 m3 ∕s
N i=1 i
N X N
^γx = ^x Þ3 = 0.5853
ðx − μ
σ3x i = 1 i
ðN − 1ÞðN − 2Þ^

Equations (8-11) and (8-12) give η^x = 0.4431 and ω = 0.7493, respectively.
Then applying Equations (8-10), (8-8), and (8-9) gives the moment estimates of the parameters
of the LN3 distribution for the 1 day low flows for the San Pedro River, respectively,

Statistical Analysis of Hydrologic Variables

 278 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Table 8-2. 1 Day Low Flows for the San Pedro River at Villalba, Mexico, for 1939–1991.

Year x (m3/s) Year x (m3/s) Year x (m3/s)

1939 0.6220 1957 0.0790 1975 0.4090

Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

1940 0.6850 1958 0.1150 1976 0.5070

1941 0.4520 1959 0.4010 1977 0.4160
1942 0.3050 1960 0.2620 1978 0.2500
1943 0.6480 1961 0.2620 1979 0.4190
1944 0.4850 1962 0.1850 1980 0.3000
1945 0.4470 1963 0.2290 1981 0.7040
1946 0.4850 1964 0.3570 1982 0.4430
1947 0.4310 1965 0.1390 1983 0.3670
1948 0.3920 1966 0.3300 1984 0.1770
1949 0.2260 1967 0.3390 1985 0.3097
1950 0.2640 1968 0.3280 1986 0.3905
1951 0.1950 1969 0.3480 1987 0.4417
1952 0.1840 1970 0.2080 1988 0.3846
1953 0.1320 1971 0.3090 1989 0.2740
1954 0.1400 1972 0.3770 1990 0.1750
1955 0.2450 1973 0.3950 1991 0.1288
1956 0.1500 1974 0.2740
Source: CONAGUA (2016).

Figure 8-4. Empirical and fitted frequency curves for the 1 day low flows of the San Pedro River in
Mexico, based on MOM estimates of the parameters of the LN3 distribution.

^x0 = −0.4296 m3 ∕s, μ^y = −0.287, and σ

^y = 0.1945. Figure 8-4 shows the graphical fitting of the LN3
distribution using the MOM. Then the 10 year 1 day low flow is

^x1,10 = ^x0.1 = ^x0 þ expð^ ^y Þ = −0.4296 þ exp½−0.287

μy þ z0.1 σ
þ ð−1.2816Þð0.1945Þ = 0.155 m3 ∕s
where z0.1 = −1.282 is the 10% quantile of the standard normal distribution.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 279

8.3.1.3 Fitting the Extreme Value Type III (Weibull) Distribution

The probability density function of the Weibull distribution is (Gumbel 1958)
    
k x − ε k−1 x−ε k
f ðxÞ = exp − , ε ≤ x < ∞:
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

(8-13)
ðv − εÞ v − ε v−ε

where ε, k, and v are the location, scale, and shape parameters, respectively, such that k > 0 and
(v − ε) > 0. Likewise, the cumulative probability distribution function is
  
x−ε k
FðxÞ = 1 − exp − (8-14)
v−ε

Kite (1988) shows that based on the method of moments the parameters may be estimated by

^ε = μ ^x ½1 − Γð1 þ 1∕kÞBk
^x þ σ (8-15)

^v = ^ε − Bk σ
^x (8-16)

^k = 1
(8-17)
a0 þ a1 γx þ a2 γx þ a3 γ3x þ a4^γ4x þ a5 ^γ5x þ a6 ^γ6x
^ ^2 ^

where the approximation is valid in the range −1.04 ≤ γx ≤ 2, Bk = ½Γð1 þ 2∕kÞ − Γ2 ð1 þ 1∕kÞ−1∕2
Γ(.) = complete gamma function
a0 = 0.277597
a1 = 0.323127,
a2 = 0.061656,
a3 = −0.020235,
a4 = −0.007321,
a5 = 0.005578,
a6 = −0.001094, and
μ^x , σ
^2x , and ^γx = sample mean, variance, and skewness coefficient, respectively.

Example 8-3: Estimating the 1 Day, 10 Year Low Flow × 1,10 Based on the Weibull Distribution
We will use the 1 day low flow data for the San Pedro River at Villalba, Mexico, for 1939–1991
(Table 8-2). The results are obtained using the MOM procedures described previously. From
Example 8-2, the sample mean, standard deviation, and skewness coefficient are μ ^x = 0:3306 m3 ∕s,
^x = 0:1465 m ∕s, and ^γx = 0.5853, respectively. Then, applying the MOM procedure gives
σ 3
^k = 2.0693, Bk = ½Γð1 þ 2∕kÞ − Γ2 ð1 þ 1∕kÞ−1∕2 = ½0.9863 − ð0.8858Þ2 −1∕2 = 2.2269, ^ε = μ ^x þ
^x ½1 − Γð1 þ 1∕kÞBk = 0.3306 þ ð0.1465Þð1 − 0.8858Þð2.2269Þ = 0.0417, and ^v = ^ε − Bk σ
σ ^x = 0.0417 −
ð2.2269Þð0.1465Þ = 0.3678.
Thus, the moment estimates of the parameters of the Weibull distribution for the 1 day low-flow
sample data of the San Pedro River are ^ε = 0.0417 m3 ∕s, ^k = 2.0693, and ^v = 0.3678 m3 ∕s. Figure 8-5
shows the corresponding fitted distribution.
Then from Equation (8-14), the q-th quantile may be written as

xq = ε þ ðν − εÞ½− lnð1 − qÞ1∕k

Statistical Analysis of Hydrologic Variables

 280 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-5. Empirical and fitted frequency curves for the San Pedro River at Villalba (Mexico), based
on MOM estimates of the parameters of the Weibull distribution.

so that the 10 year, 1 day low flow is

^x1,10 = ^x0.1 = 0.0417 þ ð0.3678 − 0.0417Þ½− lnð1 − 0.1Þ1∕2.0693 = 0.152 m3 ∕s

8.3.1.4 Fitting the General Extreme Value Distribution

The GEV has been applied to many hydrologic events for a long time (e.g., NERC 1975). Various
forms and parameterizations of the GEV have been utilized in practice. Here we will follow that
utilized by Raynal-Villasenor (2013), where the probability density function is defined as

1
f ðxÞ = expf−½1 − βðω − xÞ∕α1∕β g½1 − βðω − xÞ∕αð1∕βÞ−1 (8-18)
α
in which ω is the location parameter, α is the scale parameter, and β is the shape parameter. Also
−∞ < x ≤ ω − α∕β for β < 0 and ω − α∕β ≤ x < ∞ for β > 0. Likewise, the cumulative probability
distribution function is

FðxÞ = 1 − expf−½1 − βðω − xÞ∕α1∕β g (8-19)

The moments estimators of the GEV distribution are

^
α ^
α
ω ^ − Γð1 þ ^βÞ þ ,
^ =μ (8-20)
^β ^β

j^βj^σ
^ = qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ,
α (8-21)
Γð1 þ 2^βÞ − Γ2 ð1 þ ^βÞ

and the shape parameter (β) may be estimated by Raynal-Villasenor (2013)

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 281

^β = 0.24662 þ 0.286678^γ þ 0.072454^γ2 þ 0.010176^γ3 þ 0.000816^γ4

þ 0.000037^γ5 (8-22a)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

for β < 0 and −19.0 < ^γ ≤ −1.1396 and

^β = 0.279434 − 0.333535^γ þ 0.048305^γ2 þ 0.024414^γ3 þ 0.003765^γ4

− 0.000263^γ5 (8-22b)

for β > 0 and −1.1396 ≤ ^γ < 11.35. Note that L-moment estimators are often used to parameterize
the GEV distribution (see Stedinger et al. 1993 for fitting procedures), because L-moment estimators
can have better properties than MOM estimators for real-space distribution with three or more
parameters.

Example 8-4: Estimation of the 1 Day, 10 Year Low Flow × 1,0.1 Based on the GEV Distribution
The 1 day low-flow data for the San Pedro River in Mexico for 1939–1991 are also used in this
example (Table 8-2). The estimation results have been obtained using the previously described
procedures. The following statistics have already been obtained in Examples 8-2 and 8-3 as
^x = 0.3306, σ
μ ^x = 0:1465 m3 ∕s, and ^γx = 0.5853. Then Equations (8-22b), (8-20), and (8-21) give,
respectively, ^β = 0.4965, ω
^ = 0.3668 m3 ∕s, and α
^ = 0.158 m3 ∕s.
Figure 8-6 shows the fitted distribution for the GEV. From Equation (8-19) one may find that
the 10 year, 1 day low flow is
  
α 1 β
^x1,10 = ^x0.1 = ω þ − ln 1 − −1
β T
  
0.158 1 0.4965
= 0.3668 þ − ln 1 − − 1 = 0.153 m3 ∕s
0.4965 10

Figure 8-6. Empirical and fitted frequency curves for the low flows of the San Pedro River in Mexico,
based on MOM estimates of the parameters of the GEV distribution.

Statistical Analysis of Hydrologic Variables

 282 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Note that the value obtained is practically the same as that obtained using the Weibull
distribution because it may be shown that the two distributions are equivalent.
Furthermore, as a matter of comparison we fitted the log–Pearson type III distribution, using the
procedure outlined in Section 8.3.1.1 and obtained ^x0.1 = 0.156 m3 ∕s. Thus, the estimates based on
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

the Weibull, GEV, LN3, and LP3 distributions are respectively 0.152, 0.153, 0.155, and 0.156—in
other words, they are very close. Note, however, that the sample size in this case is N = 53. The
results may have a wider variation for small sample sizes, smaller values of the nonexceedance
probability q, and for datasets with larger skews.

8.3.2 Case of Intermittent Flows

A common issue in hydrology is intermittent streamflows, where the hydrologic record contains
values recorded as zero. This typically occurs when the catchment area is small, the region is arid,
precipitation is stored for long periods as snow or ice, or the groundwater storage is depleted during
periods of little or no precipitation. When the streamflow is recorded as zero, it may actually be zero,
or may be too small to be measured by the recording instrumentation.
Typical hydrologic frequency analyses, which often rely on log-transformed distributions,
must be adapted to handle intermittent flows. Even when a real-spaced probability distribution,
such as the GEV, is employed, a truncation of the lower tail of the distribution at zero is still a
problem. While several techniques have been proposed to address this situation in regional
frequency analyses, the most common is the use of a conditional probability adjustment
(e.g., Haan 1992, USIAC 1982). This method, which represents the streamflow series as a mixed
distribution with a point mass at zero and a continuous distribution to describe the nonzero
observations, was originally recommended by Jennings and Benson (1969) for flood frequency
analysis. While an argument could be made to employ a truncated distribution to describe the
nonzero observations, with the truncation point at the measurement detection limit, typically a
nontruncated distribution is used.
Assume one is confronted with the typical problem of estimating some streamflow quantile,
xq, where q is the probability of the streamflow X to be less than or equal to xq can be written as

PðX ≤ xq Þ = Fðxq Þ = q (8-23)

If one has a streamflow series of length N, of which N0 of the values are reported as zero, an
estimator of the probability of a zero flow is

N0
^q0 = PðX = 0Þ = (8-24)
N

If q ≤ ^q0 , then ^xq = 0. If instead q > ^q0 , then one calculates an adjusted probability qa as

q − ^q0
qa = (8-25)
1 − ^q0

and then determines the qath percentile of the distribution of the nonzero observations as

PðX NZ ≤ xq Þ = qa (8-26)

where XNZ is the variable defining flows greater than zero, in other words, nonzeros.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 283

Table 8-3. 7 Day Annual Minimum Flows for Twelve Mile Creek near Waxhaw, NC (USGS #02146900).

7 day 7 day 7 day 7 day

Year min Year min Year min Year min
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

1961 0.400 1972 0.833 1983 0.000 1994 2.600

1962 0.500 1973 1.414 1984 2.000 1995 1.943
1963 0.529 1974 2.314 1985 0.553 1996 3.900
1964 6.071 1975 7.614 1986 0.094 1997 1.414
1965 3.371 1976 1.743 1987 0.096 1998 2.943
1966 0.543 1977 1.561 1988 1.031 1999 0.299
1967 2.471 1978 1.714 1989 6.371 2000 0.461
1968 0.014 1979 3.471 1990 1.287 2001 0.041
1969 1.686 1980 0.291 1991 3.600 2002 0.070
1970 0.000 1981 0.681 1992 2.843 — —
1971 6.686 1982 1.157 1993 1.130 — —

Example 8-5: Estimation of the 7-Day 10-Year Minimum Flows for the Case of Zero
Flood Years
Table 8-3 gives the 7 day average annual minimum streamflows for Twelve Mile Creek near
Waxhaw, North Carolina (USGS gauge #02146900), from 1961 to 2002. For two years (1970 and
1983), the 7 day annual minimum was recorded as zero. Assume one wishes to estimate x7,10, in
other words, the 7 day, 10 year annual minimum streamflow using a log–Pearson type III
distribution. The x7,10 is the 10th percentile of the distribution of 7 day annual minimums, in
other words, q = 0.1.
Using information from Table 8-3, we have N = record length = 42 and N0 = number of years
when 7 day annual minimum flows equals zero = 2. Then, Equations (8-24) and (8-25) give

N0 2 q − ^q0 0.1 − 0.0476

^q0 = = = 0.0476 and qa = = = 0.055:
N 42 1 − ^q0 1 − 0.0476

The x7,10 is thus estimated as the 5.5th percentile of the distribution of the nonzero 7 day annual
minimums. Using the nonzero 7 day annual minimum flows and Equation (8-2), (8-3), and (8-4) we
get μ ^y = 0.0022, σ^y = 1.439, and ^γy = −1.089, respectively. Also applying the Wilson–Hilferty
transformation [Equation (8-6)] where zq = −1.598 is the 5.5th percentile of the standard normal
distribution, we obtain the frequency factor Kq = −1.812. Thus, the ^x7,10 is estimated as
^x7,10 = ^x0.10 = exp½0.0022 − 1.812 × 1.439 = 0.074 ft3 ∕s.

8.4 REGIONAL ANALYSIS OF LOW FLOWS

Regional hydrologic analysis refers to the use of information from one or more gauged river sites
to improve the estimation of a hydrologic statistic or parameter at the site of interest. In these
situations, the site of interest typically has either no streamflow measurements (ungauged) or a
limited record (partially gauged). Regional analysis typically involves selecting a homogeneous
region of gauged sites that is similar to the site of interest and then employing that region of sites for
estimation at the ungauged or partially gauged site.

Statistical Analysis of Hydrologic Variables

 284 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

8.4.1 Methods for Selecting Homogeneous Regions

For many low streamflow studies, determining a homogeneous region within which low streamflow
processes are similar is beneficial. This is especially true when one has limited historical streamflow
data at the site of interest, and information from other sites in the region must be transferred to the
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

site of interest to improve low streamflow estimators at the site of interest. Low streamflow
estimation techniques such as regional regression and index flow are based on a homogeneous
region. Hosking and Wallis (1997) state that, “Of all the stages in a regional frequency analysis
involving many sites, the identification of homogeneous regions is usually the most difficult and
requires the greatest amount of subjective judgment.”
Although techniques have been developed to determine homogeneous regions when historical
streamflow records are available at the site of interest, such as L-moment techniques (Hosking and
Wallis 1997), streamflow data are not available at ungauged and partially gauged river sites. Thus,
even if homogeneous regions were developed, one may have a problem determining which region
the site of interest should include.
If contiguous regions are developed, then a site is placed into a region based on its geographic
location. These regions are often developed based on drainage basins. For instance, the USGS has
broken the conterminous United States into 18 regions containing the drainage areas of major basins
and then subdivided these regions into 221 subregions based on smaller drainage areas. This
classification is the basis of hydrologic unit codes (HUCs) by which watersheds are catalogued
(USGS 2010a), and often state-based hydrologic studies develop regions based on a combination of
HUCs and state boundaries.
For discontiguous regions, an allocation rule is required to place an ungauged or partially gauged
site within a specific region. These rules must be based on watershed characteristics that can be
measured at the ungauged or partially gauged site. The methods presented as follows identify several
techniques that have been applied to develop homogeneous regions for low streamflow estimation.
Hayes (1992) uses a residual pattern approach to develop homogeneous regions. In this technique
a global regression model between the flow characteristic of interest and watershed characteristics is
first developed for all sites in a region, and the residuals from this model are mapped and generalized
into homogeneous regions. One drawback of this approach is that the initial regression model, which
was developed using all potential sites, may be incorrect for the specific region developed, creating an
incorrect grouping of sites. Typically, with this technique, contiguous regions are developed, and thus
placing an ungauged site in a region is based on its geographic location.
Another technique is the use of multivariate statistics to determine homogeneous regions.
Nathan and McMahon (1990) compare several low-flow regionalization techniques, including
cluster analysis, multiple regression, and principal component analysis. Their analysis recommends a
weighted cluster analysis that creates groups that minimize the sum of squared errors between
observations and the group’s weighted mean (Ward’s method). This method creates discontiguous
regions, and thus a method for placing ungauged sites in a region is necessary. Nathan and
McMahon (1990) suggest using Andrews curves for a decision rule, though Laaha and Blöschl
(2006a) suggest that discriminant analyses and classification trees may also be used.
Still another approach is to use classification and regression tree (CART) models. These were
first applied to the problem of low streamflow regionalization by Laaha and Blöschl (2006a). With
this supervised classification technique, the initial heterogeneous domain is broken into several more
homogeneous groups by maximizing the homogeneity of low flows and catchment characteristics
simultaneously within each group. Usually the homogeneity is measured in terms of minimizing the
variance of the low-flow statistic of interest. Laaha and Blöschl (2006a) note that the benefits of this
technique include its nonparametric structure, the interpretability of small groupings, its ability to
handle nonlinear relationships, and the lack of sensitivity to unusual sites (outliers). One major
problem with this technique is that with large trees interpreting relationships is difficult. Once a
regression tree has been developed, it can be employed to place ungauged sites in a region.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 285

Young et al. (2000) and Laaha and Blöschl (2006a, b) examine a final approach, which is to use
seasonality to group sites. This method assumes that the time of occurrence of low flows indicates the
dominant hydrologic processes and thus can be used to determine low-flow regions. While Young et
al. (2000) find the spatial variability of low-flow seasons to be relatively small in the United Kingdom,
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Laaha and Blöschl (2006a, b) find the spatial variability of low-flow seasonality to be very high in
Austria, a humid, mountainous region. Ideally with this method regions are spatially contiguous, and
thus placing an ungauged site in a region is based on its location.
Laaha and Blöschl (2006a) compare all four of the aforementioned regionalization methods to
estimate the specific discharge exceeded 95% of the time at 325 watersheds in Austria. After
determining homogeneous regions, they develop regional regression models in each region and
perform a delete-one cross-validation to compare the regionalization techniques. They find that
regions based on seasonality create the best groupings in Austria, though note that this may be the
result of the wide variety of differences in seasonal low-flow processes within the study area. They
also note that all methods appear to underestimate specific discharge at wet catchments.

8.4.2 Methods for Regional Analysis and Estimation

Several common methods have been applied to estimate low streamflow statistics at a site of interest
using regional information. As mentioned previously, these techniques are commonly employed
where little or no streamflow data are available at the site of interest. This section discusses two such
techniques: low streamflow regional regression and baseflow correlation.

8.4.2.1 Regional Regression

When no historic streamflow record is available at the site of interest, a regional regression model
may be developed. In this technique, a relationship is developed between the streamflow statistic of
interest and watershed characteristics at gauged sites within the region, and then watershed
characteristics at the ungauged site are employed to determine a streamflow estimate at the site
of interest. Typically, these models have the following form:

Qd,T = eβ0 X β11 X β22 : : : eε (8-27)

where
Qd,T = d day,
T = year low streamflow statistic,
Xi = watershed characteristics,
βi = model parameters to be estimated, and
ε = model error term. The form of this model is consistent with a theoretically based low
streamflow watershed model derived from hillslope discharge models (Vogel and Kroll 1992).

When the logarithm of this equation is taken, one obtains a linear relationship:

lnðQd,T Þ = β0 þ β1 lnðX 1 Þ þ β2 lnðX 2 Þ þ : : : þ ε (8-28)

The parameters of this equation can then be estimated using ordinary, weighted, or generalized
least squares regression procedures (Stedinger and Tasker 1986). For ordinary least squares
parameter estimators to be efficient, the model residuals need to be independent and homoscedastic
(constant variance). Weighted least squares is employed to address heteroscedasticity (nonconstant
variance) of the model residuals, while generalized least squares can be employed to address both
heteroscedasticity and the lack of independence of the model residuals. Kroll and Stedinger (1998)
show that if the model error variance in the regional regression model is large, which is typical for

Statistical Analysis of Hydrologic Variables

 286 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

low-flow regional regression, the model error variance overwhelms the time sampling error, and
ordinary least squares produces similar parameter estimators as generalized least squares.
Low-flow regional regression models have been developed for many regions throughout the
world, including Europe (e.g., Gustard et al. 1989, Laaha and Blöschl 2007), Australia (Nathan and
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

McMahon 1992), and the United States (e.g., Thomas and Benson 1970, Kroll et al. 2004). One
challenging aspect of developing low-flow regional regression models is determining the appropriate
catchment characteristics to use as explanatory variables in the models. Whereas a theoretically
derived model is available (Vogel and Kroll 1992), a common approach is using stepwise regression
procedures (Tallaksen and van Lanen 2004). Such procedures are useful when the number of
potential explanatory variables is large to help determine the significant variables to include in the
final model (Kroll et al. 2004). Tallaksen and van Lanen (2004) and Kroll et al. (2004) list some
catchment characteristics commonly employed in low-flow models, including drainage area,
annual average rainfall, soil index, mean basin elevation, and summer precipitation. Hydro-
geologic indexes, such as the baseflow index (Institute of Hydrology 1980) or baseflow recession
constant (Vogel and Kroll 1996, Tallaksen 1995) have been shown to greatly improve low
streamflow regional regression models (Kroll et al. 2004), though often these indexes are difficult
to obtain at ungauged river sites.
In most instances, the standard errors associated with low-flow regression models have been
relatively high (Vogel and Kroll 1992, Smakhtin 2001). One reason may be low-flow processes are
too complex to be described with a linear or log–linear model. Another reason may be that important
explanatory variables have been excluded from these models, and/or the watershed characteristics
employed as explanatory variables have not been of high quality. Regardless of these issues, regional
regression is still a common technique applied in low streamflow estimation.
Typically, a statistical computing package is employed to develop regression models, especially
when ordinary least squares regression procedures are used to determine parameter estimators.
Hirsch et al. (1993) provide an excellent review of regression analysis, including the formulas used in
linear regression, a procedure for developing and testing a regression model, confidence and
prediction intervals, model diagnostics, and model selection. The US Geological Survey (USGS
2010a) has the publicly available software package GLSNet to perform generalized least squares
regression procedures (e.g., Stedinger and Tasker 1986, Kroll and Stedinger 1998).

8.4.2.2 Baseflow Correlation

The baseflow correlation (or baseflow regression) method is a data transfer technique where
information is transferred from a nearby long-record gauge to a short-record gauge. Baseflow
correlation can be performed with only a minimal number of streamflow measurements at the
partial record site (i.e., 5–15 measurements). One key assumption is the streamflow at the short- and
long-record sites is under baseflow conditions, meaning that all contributions to streamflow are from
groundwater discharge, which is typically considered to be at least three days after the peak of the
hydrograph.
The baseflow correlation method was proposed by Stedinger and Thomas (1985) and has several
assumptions. The first assumption of the baseflow correlation method is a linear relationship
between yi, the logarithm of the d day annual minimum flows at an ungauged site, and those at a
gauged site, xi:

yi = α þ βxi þ εi , (8-29)

where α and β are regression model parameters, and εi are independent normal error terms with a
mean of zero and a constant variance, σ2ε (i.e., εi ∼ Nð0, σ2ε Þ).
Second, because annual minimum flows are not available for the ungauged site, the relationship
between d day annual minimum flows is assumed to be similar to the relationship between

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 287

instantaneous baseflows. In a large simulation experiment, Zhang and Kroll (2007b) find that this is
generally a reasonable assumption, though in some cases wide variations occur. Thus, the linear
relationship between baseflow measurements at the ungauged site, ~yi , and corresponding baseflows
at the gauged site, ~xi , is given by
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

~yi = α þ β~xi þ εi , (8-30)

where εi ∼ Nð0, σ2ε Þ.
The third assumption is that the annual minimum streamflows are described by an LP3
distribution. By this assumption, the logarithm of Q7,10 at the ungauged site can be estimated by

^ 7,10 Þ = μ
lnðQ ^y
^y þ K y σ (8-31)
where
^y = estimator of the log-space mean,
μ
σ^y = estimator of the log-space variance, and
K y = associated frequency factor for the LP3 distribution.

The frequency factor is a function of the log-space skew of the 7 day annual minimum flows and
the percentile of interest. In the baseflow correlation method, the frequency factor for the ungauged
site, K y , is assumed equal to the frequency factor for the gauged site, Kx, an assumption Zhang and
Kroll (2007b) find has little impact on the performance of this method. Thus, only estimators of μ ^y
and σ ^y are required; Stedinger and Thomas (1985) suggest the unbiased estimators:

^y = a þ bmx
μ (8-32)

 
s2x
^2y = b2 s2x
σ þ s2e 1− (8-33)
ðL − 1Þs~2x
where
mx and s2x = log-space mean and variance of the 7 day annual minimum flows at the gauged site,
respectively,
s~2x = sample variance of the logarithms of the concurrent flows at the gauged site;
L = number of concurrent baseflow measurements, and
a, b, and s2e = ordinary least squares estimators of the parameters α, β, and σ2ε estimated using
baseflow measurements at the ungauged site as a function of the concurrent flows at the gauged site.

Stedinger and Thomas (1985) derive the variance of the Q7,10 estimator as
 
s2
^ 7,10 Þ ≅ þ
e ðm x − m~x Þ2 2
se b 2 2
s x K 2y 4b2 s4x s2e 2b4 s4x 2s4e
Var½lnðQ þ þ 2 þ þ
L ðL − 1Þs2~x n σy
4^ Ls~2x n L
2bs2x ðmx − m~x ÞK y s2e
þ (8-34)
σy s~2x
L^

where m~x is the sample mean of the logarithms of the flows at the gauged site. Stedinger and Thomas
(1985) examine the performance of baseflow correlation with 20 pairs of streamflow sites. Reilly and
Kroll (2003) expand this analysis to more than 1,300 streamflow sites in the United States and find this
method to perform well if nearly independent baseflow measurements were obtained. Reilly and Kroll
suggest choosing one baseflow measurement from at least 10 consecutive streamflow recessions and

Statistical Analysis of Hydrologic Variables

 288 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

choosing a long-record site within 200 km of the short-record site. Zhang and Kroll (2007b) examine
the tradeoffs between the number of required streamflows and the correlation coefficient between the
baseflows at the two sites. Zhang and Kroll (2007a) showed that when only five measurements are
available, employing multiple long-record sites can improve the performance of this technique.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

8.5 ANALYSIS OF AUTOCORRELATED LOW FLOWS

To quantify the return period and risk of low flows events it is convenient to analyze the time series
dependence structure of low flows. For this purpose, the mathematical models to be applied may
consider the low flow as a continuous or discrete variable, depending of the complexity of the
temporal dependence. In any case, some simple models are described in this section along with
illustrative examples. Likewise, the estimation of return period and risk are described for the case of
discrete variables.

8.5.1 Modeling of Autocorrelated Low Flows

Time series of low flows may be time dependent (autocorrelated) due to the effect of ground water,
lake regulations, wetland storage, and channel storage. Several models representing the dependence
structure of this type of process have been proposed in the hydrological literature (e.g., Salas 1993).
One of the simplest models is the first-order autoregressive, which is generally adequate for modeling
hydrologic processes having short-term time dependence. For processes with a more complex
autocorrelation structure, autoregressive moving average (ARMA) models may be more applicable.
Also, extensions or modifications thereof, for example, the gamma autoregressive (GAR) model
(Fernández and Salas 1990) may be useful. Furthermore, shifting mean (SM) models (Sveinsson et al.
2003) may be necessary where long-term climate variability affects hydrological processes. In
addition, where the hydrological process is a discrete autocorrelated variable defined by a finite
number of states, then Markov chain models and discrete autoregressive (DAR) and discrete ARMA
(DARMA) models may be useful (e.g., Salas 1993). In this section, we provide some details of the
simple models.

8.5.1.1 Simple Markov Chain

A simple model for representing the dependence structure of low flows is the Markov-dependent
process with two states and a homogeneous transition probability matrix, i.e., a simple Markov chain
(e.g., Fernández and Salas 1999a, 1999b). The two states may represent above or below a critical level.
Let Yt represent a time series of low flows and Y0 a certain threshold (e.g., the sample mean or a
particular quantile). Consider the states Yt < Y0 and Yt ≥ Y0 and for convenience we will denote
those states as zero and one, respectively. A simple Markov chain that represents the sequence of
zeros and ones is defined by its transition probability matrix: p00 = PðY t < Y 0 jY t−1 < Y 0 Þ,
p01 = PðY t ≥ Y 0 jY t−1 < Y 0 Þ, p10 = PðY t < Y 0 jY t−1 ≥ Y 0 Þ, and p11 = PðY t ≥ Y 0 jY t−1 ≥ Y 0 Þ. It fol-
lows that p00 þ p01 = 1 and p10 þ p11 = 1. Furthermore, p0 = PðY t < Y 0 Þ and p1 = PðY t ≥ Y 0 Þ are the
unconditional probabilities of states 0 and 1, respectively, and p0 þ p1 = 1.
Thus, to characterize the simple Markov chain estimating the transition probabilities is
necessary. Jackson (1975) shows that maximum likelihood estimates of such transition probabilities
can be obtained from a sample of the underlying process as follows. Let Nij be the number of
transitions from state i to state j. Then

N 01 N 01
^p01 = = (8-35)
N0 N 00 þ N 01

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 289

N 10 N 10
^p10 = = (8-36)
N1 N 10 þ N 11

And the remaining elements are ^p00 = 1 − ^p01 and ^p11 = 1 − ^p10 .
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Although ^p01 and ^p10 may be calculated by counting the occurrence of zeros and ones from the
available sample, the procedure is reliable only if a large number of occurrences of zeros and ones can
be obtained from the available record. However, for the typical lengths of hydrologic records and for
the cases where Y0 is small (such as for low-flow studies), zeros are rare with a small probability of
occurrence, consequently the foregoing procedure for estimating the probability matrix based on
Equations (8-35) and (8-36) is unreliable. An alternative procedure is to assume a bivariate
distribution function for (Yt, Yt−1), so that p00 can be determined from

PðY t < Y 0 and Y t−1 < Y 0 Þ

p00 = PðY t < Y 0 jY t−1 < Y 0 Þ =
PðY t−1 < Y 0 Þ

For instance, for a bivariate normal distribution, Cramer and Leadbetter (1967) give

Zρ
1 exp½−Y 20 ∕ð1 þ zÞ
p00 = p0 þ pffiffiffiffiffiffiffiffiffiffiffiffi dz (8-37)
2πp0 1 − z2
0

in which ρ = lag-one serial correlation coefficient of Yt. In addition, the following relationship,
proposed by Sen (1976), completes the transition probabilities:

p0
p10 = PðY t < Y 0 jY t−1 ≥ Y 0 Þ = ð1 − p00 Þ (8-38)
1 − p0

Figure 8-7 shows the relationship among p00 , p0 , and ρ obtained from Equation (8-37) by
numerical integration, which can be used as a first approximation for calculating the transition

Figure 8-7. Relationship among p00, p0, and ρ obtained by numerical integration for a bivariate
normal process.
Source: Fernández and Salas (1999a).

Statistical Analysis of Hydrologic Variables

 290 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

probabilities. Also, Example 8-7 illustrates the estimation of transition probabilities for the Parana
River 7 day low flows.
For the case of low flows, the focus is on computing the probability that Yt < Y0 occurs for the
first time at the nth time step, f n,Y 0 , where Y0 is the critical event. This represents a sequence of state
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

one during the first n−1 time steps and state zero at the nth time step, in other words,

States: ::: :::. 1, 1, 1, 1, ::: :::. 1, 1, 0

Time step: ::: :::. 1, 2, 3, 4, ::: :::. (n−2), (n−1), n

Then, the probability f n,Y 0 , can be calculated as

f 0,Y 0 = 0
f 1,Y 0 = p0
f 2,Y 0 = p1 p10 = ð1 − p0 Þp10
::::
f n,Y 0 = p1 pn−2
11 p10 = ð1 − p0 Þpn−2
11 p10 , n ≥ 2 (8-39)

The probability f n,Y 0 will be useful for calculating the return period and risk as we will see in
Section 8.6.2.

8.5.1.2 Autoregressive Moving Average Models

The family of ARMA models has been widely used for modeling hydroclimatic processes at various
time scales, such as annual streamflows (e.g., Salas et al. 1980, Loucks et al. 1981, Salas 1993). The
ARMA(p, q) model is defined as (Box and Jenkins 1976)

X
p X
q
Yt = μ þ ϕj ðY t−j − μÞ þ εt − θj εt−j
j=1 j=1

ϕðBÞðY t − μÞ = θðBÞεt (8-40)

where
μ, ϕ 0 s, θ 0 s, and σ2 ðεÞ = parameters of the model,
p = order of the autoregressive terms, and
q = order of the moving average terms, ϕðBÞ = − ϕ1 B1 − ϕ2 B2 − · · · −ϕp Bp , θðBÞ = − θ1 B1 −
θ2 B2 − · · · −θq Bq , and Bi Z t = Z t−i .

Particular models derived from Equation (8-40) are the ARMA(p, 0) or autoregressive AR(p)
and the ARMA(0, q) or moving average MA(q) models. These models assume that the variable Yt is
normally distributed, hence their applications to modeling hydroclimatic time series generally
require that the underlying data be converted to normal by some appropriate transformation (e.g.,
Box and Cox 1964). However, some models with ARMA-type dependence structure are applicable to
skewed marginal distributions. Some of these models can accommodate variables with gamma-
marginal distribution functions, such as GAR models. Also, for processes consisting of discrete-
valued random variables there are discrete DARMA models, which are more suitable for modeling

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 291

persistence characteristics with longer memory than simple Markov chain models. Low-order
models are most widely used for modeling streamflow time series (Salas et al. 2001). Some of them
are described in the following subsections.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

8.5.1.3 The ARMA(1, 1) Model

The ARMA(1,1) model is defined as (Box and Jenkins 1976, Salas and Pielke 2003)

Y t = μ þ ϕ1 ðY t−1 − μÞ þ εt − θ1 εt−1 (8-41)

where μ, ϕ1 , θ1 , and σ2ε are model parameters. It is a Gaussian stationary-dependent process with
a continuous valued variable Yt. It may be shown that the lag-k autocorrelation function of the
ARMA(1, 1) process is ρk ðYÞ = bϕk−1 1 , k ≥ 1 where b = ð1 − ϕ1 θ1 Þðϕ1 − θ1 Þ∕ð1 þ θ1 − 2ϕ1 θ1 Þ =
2

ρ1 ðYÞ. In the case that θ1 = 0, Model (8-41) becomes the AR(1) process, and its autocorrelation
function is ρk ðYÞ = ϕk1 , k ≥ 1. Note that the parameters of the ARMA(1, 1) and AR(1) models
are constrained, which in turn implies certain relationships between ρ2(Y) and ρ1(Y) (Box and
Jenkins 1976).

8.5.1.4 First-Order Gamma-Autoregressive Model

The GAR(1) model assumes that the underlying series is Markov dependent with a gamma marginal
distribution, thus the model does not require variable transformation. Lawrance and Lewis (1981)
develop the first-order gamma-autoregressive model, which is defined as

Y t = ϕY t−1 þ εt (8-42)
where
Y t = gamma-dependent variable at time t,
ϕ = autoregression coefficient, and
εt = independent variable.

The only difference with the well-known AR(1) model is that Yt has a marginal distribution
given by the three-parameter gamma density function,

αβ ð y − λÞβ−1 exp½−αð y − λÞ

f Y ðyÞ = , (8-43)
ΓðβÞ
in which λ, α, and β are the location, scale, and shape parameters, respectively. Thus, a noise term εt
must be found so that when incorporated into the autoregressive Equation (8-42), it will produce a
variable Yt that is gamma distributed as in Equation (8-43). Gaver and Lewis (1980) find that for
integer values of β, the noise ε of Equation (8-42) is given by

X
β
ε = λð1 − ϕÞ þ ηj (8-44)
j=1

where
ηj = 0 with probability ϕ,
ηj = exp(α) with probability (1 − ϕ), and
exp(α) = an exponentially distributed random variable with expected value 1/α.
This approach is valid for skewness coefficient less than or equal to 2. In addition, Lawrance
(1982) finds another solution for noninteger values of β based on the shot-noise process used by
Weiss (1977). In this case, ε can be obtained by

Statistical Analysis of Hydrologic Variables

 292 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

ε = λð1 − ϕÞ þ η (8-45)
and

η = 0 if M=0
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

X
M
η= E j ϕU j if M>0 (8-46)
j=1

in which M is a discrete random variable Poisson distributed with mean value equal to - β ln(ϕ). The
set (Uj) comprises independent identically distributed (iid) random variables with uniform (0, 1)
distribution and the set (Ej) comprises iid random variables exponentially distributed with mean 1/α.
The estimation of model parameters, in other words, λ, α, β, and ϕ, can be made by the method
of moments (Fernández and Salas 1990). The population moments of the underlying variable, Yt,
may be expressed as a function of the parameters of the GAR(1) model as

μ = λ þ ðβ∕αÞ (8-47a)

σ2 = β∕α2 (8-47b)
pffiffiffi
γ = 2∕ β (8-47c)

ρ1 = ϕ (8-47d)

where μ, σ2, γ, and ρ1 are the population mean, variance, skewness coefficient, and lag-one
autocorrelation coefficient of Yt, respectively. These moments can be estimated based on the
sample Y1, Y2, : : : , YN, using the well-known relationships:

1X N
m= Y (8-48a)
N i=1 i

1 X N
s2 = ðY − mÞ2 (8-48b)
N − 1 i=1 i

N X N
g1 = ðY − mÞ3 (8-48c)
ðN − 1ÞðN − 2Þs3 i = 1 i

1 X
N−1
r1 = ðY − mÞðY iþ1 − mÞ (8-48d)
ðN − 1Þs2 i = 1 i

where N is the sample size. However, for dependent and nonnormal variables these estimators are
biased (Fernández and Salas 1990). Hence, some corrections are needed before using them for
solving the system of Equations (8-47a–d) to estimate the parameters of the GAR(1) model (only the
estimator of the expected value can be used without a correction factor).

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 293

Wallis and O’Connell (1972) suggest the following correction to obtain an unbiased estimator of
ρ1 for an AR(1) model:

r1 N þ 1
^ρ1 = (8-49)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

N −4
in which r1 is given by Equation (8-48d). Also, Matalas (1966) and O’Connell (1977) show that if the
variables follow an AR(1) process, an unbiased estimator of the variance can be obtained by

N −1 2
^2 =
σ s (8-50)
N −K
where K = ½Nð1 − ^ρ21 Þ − 2^ρ1 ð1 − ^ρN1 Þ∕½Nð1 − ^ρ1 Þ2  and s2 and ^ρ1 are given by Equations (8-48b) and
(8-49), respectively. In addition, Fernández and Salas (1990) suggest the following correction to
obtain an unbiased estimator of γ for a GAR(1) process:

^γ0
^γ = −0.49 (8-51)
ð1 − 3.12ρ3.7
1 N Þ

where ^γ0 is the skewness coefficient suggested by Bobee and Robitaille (1975) for independent
gamma variables as

Lg 1 ½A þ BðL2 ∕NÞg 21 
^γ0 = pffiffiffiffi (8-52)
N
in which

N −2
L = pffiffiffiffiffiffiffiffiffiffiffiffi , A = 1 þ 6.51 N −1 þ 20.2 N −2 , and B = 1.48 N −1 þ 6.77 N −2 (8-53)
N −1

Therefore, the general procedure to estimate the parameters of a GAR(l) model based on
available sample series is as follows: (1) estimates of the mean m, variance s2, skewness coefficient g1,
and lag-one autocorrelation coefficient r1, are obtained from the sample using Equations (8-48a–d);
(2) the unbiased autocorrelation coefficient, ^ρ1 , is determined from Equation (8-49) as a function of
N and r1; (3) with this value of ^ρ1 , the unbiased variance σ ^2 is estimated from Equation (8-50);
(4) with the values of N, ^ρ1 , and g1, the unbiased estimate ^γ of γ is determined from Equation (8-51);
and (5) Equations (8-47a–d) are used to estimate the set of model parameters λ, α, β, and ϕ.

8.5.1.5 The DARMA(1, 1) Model

In Section 8.5.1.1 we discussed the simple Markov chain considering the state zero, in which Y t < Y 0 ,
and state one, where Y t ≥ Y 0 . For convenience, we can also use an additional symbol, say X, to
denote the states zero and one, i.e., X is a discrete random variable where X = 0 if Y t < Y 0 and X = 1
if Y t ≥ Y 0 . Saying this another way, a continuous valued process Yt was censored or clipped at Y0
leading to a discrete valued process Xt. If the original hydrological process Yt is autocorrelated, then
the clipped process Xt is also expected to be autocorrelated (Salas et al. 2001). What we did in
Section 8.5.1.1 was to model the sequence of zeros and ones using a simple Markov chain. We could
also say that we modeled the variable X with a simple Markov chain.
In some cases, modeling the X process with a simple Markov chain is adequate, but for cases
where the underlying process Y has longer dependence, in other words, an autocorrelation function

Statistical Analysis of Hydrologic Variables

 294 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

that persists for long time, a model with a more flexible correlation structure may be necessary to
represent Xt. Thus, we will assume that the autocorrelation of Xt can be represented by a DARMA
process. For example, the DARMA(1, 1) model is defined as (Jacobs and Lewis 1977a, b)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

X i = U i Z i þ ð1 − U i ÞW i−1 , i = 1, 2, : : : (8-54)
where
Ui = an independent Bernoulli (0,1) process with parameter P(Ui = 1) = β,
Zi = another independent Bernoulli (0,1) process with P[Zi = 0] = π0 and P[Zi = 1] = π1, and
Wi−1 is a discrete AR(1) process, in other words, a DAR(1) process with parameters λ and π0.

The autocorrelation function of the DARMA(1, 1) model is ρk ðXÞ = cλk−1 , k ≥ 1 with c = (1 − β)

(λ + β − 2 λ β). When β = 0, Xi becomes the DAR(1) process. In general, the DAR(1) process is
defined as

X i = V i X i−1 þ ð1 − V i ÞZ i i = 1, 2, : : : (8-55)

where Vi is an independent Bernoulli (0,1) process with parameter λ, and Zi is as defined previously.
The autocorrelation function of the DAR(1) model is ρk ðXÞ = λk , k ≥ 1.

Example 8-6: Estimation of the 7 Day, 10 Year Low Flow Based on the GAR(1) Model
Table 8-4 gives the 7 day low-flow data for the gauging station Mapocho River at Rinconada de
Maipú, Chile, for 1980–2007 (i.e., N = 28). Figure 8-8 shows the time series data. The following
statistics are obtained directly from historic data using Equations (8-48a–d): m = 14.30, s2 = 5.572,
g1 = 0.298, and r1 = 0.358. Also from Equation (8-49) the unbiased estimate of the lag-1
autocorrelation coefficient is ^ρ1 = 0.459, and Equation (8-50) gives the estimate of the standard deviation
^ = 5.74 [refer to Equation (8-50) for the value of K = 2.59]. In addition, from Equation (8-53),
σ
one obtains L = 5.004, A = 1.258, and B = 0.062 so that Equation (8-52) gives ^γ0 = 0.355, the unbiased
estimate of γ0 for independent values. Furthermore, from Equation (8-51) the unbiased estimate of the
skewness coefficient (for autocorrelated data) becomes ^γ = 0.368. Summarizing, the unbiased estimates
of the GAR(1) model statistics are μ ^ = 5.74, ^γ = 0.368, and ^ρ1 = 0.459. Using these values in
^ = 14.30, σ
Equations (8-47a–d), the estimates of the parameters of the GAR(1) model are obtained as ^λ = −16.9,
^ = 0.946, ^β = 29.5, and ϕ
α ^ = 0.459.

Table 8-4. 7 Day Low Flows (m3/s) of the Mapocho River at Rinconda de Maipú, Chile, for 1980–2007.

Year Q (m3/s) Year Q (m3/s) Year Q (m3/s)

1980 18.29 1990 11.26 2000 11.91

1981 8.53 1991 10.80 2001 19.73
1982 9.04 1992 15.90 2002 19.54
1983 11.70 1993 19.41 2003 24.90
1984 13.74 1994 13.87 2004 18.19
1985 12.76 1995 12.31 2005 26.30
1986 16.87 1996 7.11 2006 18.23
1987 22.79 1997 8.75 2007 14.30
1988 2.42 1998 9.91 — —
1989 12.46 1999 9.42 — —
Source: Banco Nacional de Aguas, Dirección General de Aguas, Ministerio de Obras Públicas, Chile.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 295
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-8. Time series data of 7 day low flows of the Mapocho River at Rinconada de Maipú for
1980–2007.

Using these parameters and the generating procedure of Equations (8-42) and (8-44),
1,000 values of 7 day low flows were obtained. Figure 8-9 shows the CDF of these values and
the CDF of the historical data. Thus, the 10 year low flow, in other words, the low flow corresponding
to 0.1 nonexceedance probability based on the simulated GAR(1) model, gives a value of 6.8 m3/s.
However, the 10 year low flow obtained from the historical data is 8.53 m3/s (because it corresponds
to the third value of the historic series using the Weibull plotting position). For comparison, if a
three-parameter gamma distribution function for independent values is also fitted (i.e., ρ1 = 0) using
the unbiased statistics m = 14.30, σ2 = 5.572, and ^γ0 = 0.355 (from the foregoing calculations) and
the parameters of the gamma distribution estimated from Equations (8-47a–c) (^λ = −17.1, α ^ = 1.012,
and ^β = 31.7), the value 7.4 m3/s is obtained for the 10 year, 7 day low flow. Figure 8-9 also shows the
CDF of the gamma independent series.

8.5.2 Return Period and Risk of Low Flows

In Section 8.5.1.1, simple Markov chains were introduced as one of the simplest models that can be
used to represent the dependence structure of two-state discrete processes. To estimate the return
period and risk of low-flow events some additional definitions and concepts are presented here
including an example.

Figure 8-9. CDF of the 7 day low flows of the Mapocho River at Rinconada de Maipú, Chile, obtained
from historical data, independent gamma, and the GAR(1) model based on 1,000 generated values.

Statistical Analysis of Hydrologic Variables

 296 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

The return period of a particular event has been generally defined as “the average number of
time steps (e.g., years) required to the first occurrence of the event.” For instance in relation to the
event Yt < Y0 the assumption is that such event has occurred in the past, a finite time τ has elapsed
since then, and the interest is in the remaining waiting time N for the next occurrence of Yt < Y0. An
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

alternative definition of return period is “the expected value of the number of time steps between any
two successive occurrences of the event.” In this case, the assumption is that an event Yt < Y0 has just
occurred and the interest is in the time of arrival of the next event Yt < Y0. This definition is
equivalent to the previous one when the time past τ, after the occurrence of Yt < Y0, is equal to zero.
In practice, both definitions have been accepted as being equivalent because in simple cases such as
those related to independent annual flood events they lead to the same result. But they lead to
different results when they are applied to complex hydrological events such as autocorrelated low
flows.
To estimate the return period of an event Yt < Y0, the following probability given in
Equation (8-39) is needed:

f n,Y 0 = PðY t < Y 0 occurs for the first time at the nth time stepÞ

where f 0,Y 0 = 0. Let N denote the random variable defining the number of time steps needed to the
first occurrence of such an event. Then, considering the first definition of return period given
previously, the return period T is the expected value of N:
X
∞
T = EðNÞ = nf n,Y 0
n=0

In addition, if a critical low-flow value Y0 is specified so that the occurrence of the event Yt < Y0
produces a failure, and L is the project life measured in the same time units as time steps (usually
years), then the risk of failure can be defined as

RL,Y 0 = PðY t < Y 0 occurs at least once in L time stepsÞ

Furthermore, the following probability is equivalent for evaluating the risk, i.e.,

Sn,Y 0 = PðY t < Y 0 has occurred at or before time step nÞ

Then the risk of failure of a project with a project life L is SL,Y 0 . Note that the foregoing
definitions assume that τ > 0 (the time past after the occurrence of a failure event) and the interest is
to estimate the probability of the time it will take for the failure event to occur for the first time after
the construction of the project. Such time has been denoted by N.
For estimating the return period of a low-flow event, we first need to determine f n,Y 0 , the
probability that Yt < Y0 occurs for the first time at the nth time step, in which Y0 is the critical value.
Under the assumption that the sequence of zeros and ones may be modeled by a first-order Markov
chain, the probability f n,Y 0 may be determined from Equation (8-39). Then, the return period
considering the first definition is given by (Fernández and Salas 1999a)

X
∞
1 − p0
T = EðNÞ = nf n,Y 0 = 1 þ (8-56)
n=0
p10

In addition, Sn,Y 0 , the probability that event Yt < Y0 occurred at or before time step n, can be
determined by

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 297

S0,Y 0 = 0
S1,Y 0 = p0
S2,Y 0 = S1,Y 0 þ f 2,Y 0
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

::::
Sn,Y 0 = Sn−1,Y 0 þ f n,Y 0 , n≥2 (8-57)

Therefore, for n = L, the risk of failure becomes (Fernández and Salas 1999a)

SL,Y 0 = 1 − ð1 − p0 Þð1 − p10 ÞL−1 (8-58)

Furthermore, for estimating the return period considering the second definition discussed
previously, one must calculate f w , the probability that a failure event occurs after another failure
event occurred w time steps earlier, in other words, we are interested in T = EðWÞ. Then considering
the sequence of zeros and ones in the following sketch,

States: 0, 1, 1, 1, 1, ::: :::. 1, 1, 0

Time step: 0, 1, 2, 3, 4, ::: :::. (w−2), (w−1), w

illustrates that after the occurrence of zero (i.e., the event Y t < Y 0 ), it takes w steps for the occurrence
of another zero. Then, in this case,

f0=0
f 1 = p00
f 2 = p01 p10
::::
f w = p01 p11
w−2
p10 , w≥2 (8-59)

and the return period is given by (Fernández and Salas 1999a)

X
∞
p01
T = EðWÞ = wf w = p00 þ ð1 þ p10 Þ (8-60)
w=0
p10

where p00 and p10 can be obtained from Equations (8-37) and (8-38), respectively.

Example 8-7: Estimating the Return Period and Risk of Low-Flow Events
This example illustrates the procedure for calculating the return period and risk of a low-flow event.
The estimation of return period and risk related to low-flow events in the Paraná River at Corrientes,
Argentina, is of special interest because of the importance of navigation on that river (Paoli et al.
1994). Figure 8-10 shows the 7 day low-flow time series for 1904–1992. The low-flow values of
interest are the minimum observed value that occurred in 1944, the 10 year minimum annual flow,
and the 50 year minimum annual flow. We would like to determine the following statistics: (a) return
period of the observed minimum 7 day low flow and the risk that a flow equal to or smaller than the
historical (observed) value will occur in the next 30 years; (b) 10 year low flow, Q10; and (c) 50 year
low flow, Q50.

Statistical Analysis of Hydrologic Variables

 298 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-10. Annual 7 day low flows of the Parana River at Corrientes, Argentina, for 1904–1992.
Source: Fernández and Salas (1999b).

(a) Return period and risk

Table 8-5 shows the basic statistical properties of the series of 7 day low flows. Because the
observed time series is skewed (skewness coefficient is 1.024) transforming the original series
into Gaussian series is necessary before applying Equation (8-37). Thus, the original data was
log-transformed, and Table 8-6 shows some basic properties of the log-transformed series.
Because the minimum low-flow value is X 0 = 4,070 m3 ∕s, the corresponding value in the
log-transformed (normal) domain is Y 0 = 3.610. Then one can calculate p0 using the normal
distribution function as
 
3.61 − 3.975
p0 = PðY < 3.61Þ = P Z < = 0.0041
0.138

With this value of p0 and ρ = 0.506 one can determine the value of p00 from Equation (8-37) as

Table 8-5. Statistical Properties of the 7 Day Low Flows of

the Paraná River at Corrientes, Argentina.

Statistical property Value

Period of record 1904–1992

Length of record (years) 89
Mean (m3/s) 9,915
Minimum (m3/s) 4,070
Standard deviation (m3/s) 3,208
Skew coefficient 1.024
Lag-1 correlation 0.491
Source: Fernández and Salas (1999b).

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 299

Table 8-6. Statistical Properties of the Log-Transformed Series of 7 Day

Low Flows of the Paraná River at Corrientes, Argentina.

Transformed value,
Y = log( X)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Statistical property

Mean 3.975
Minimum 3.610
Standard deviation 0.138
Skew coefficient ∼ 0.0
Lag-1 correlation 0.506
Source: Fernández and Salas (1999b).

Zρ
1 exp½−Y 20 ∕ð1 þ zÞ
p00 = p0 þ pffiffiffiffiffiffiffiffiffiffiffiffi dz = 0.00514
2πp0 1 − z2
0

Then from Equations (8-38) and (8-56) we get

p0
p10 = ð1 − p00 Þ = 0.0039954 and
ð1 − p0 Þ
1 − p0 1 − 0.0041
T =1 þ =1 þ ≈ 250 years
p10 0.0039954

Also, the risk of failure for a project life of 30 years can be computed from Equation (8-58) as

SL,Y 0 = 1 − ð1 − p0 Þð1 − p10 ÞL−1 = 1 − ð1 − 0.0041Þð1 − 0.0039954Þ30−1 = 0.113

(b) 10 year low flow

Solving Equations (8-37), (8-38), and (8-56) numerically for values of T = 10 and ρ = 0.506,
one obtains Y 10 = 3.7981. Then, Q10 = 103.798 = 6,282 m3 ∕s.
(c) 50 year low flow
Similarly, for T = 50 and ρ = 0.506 the 50 year low flow discharge is
Q50 = 103:6912 = 4,911 m3 ∕s.

8.6 STATISTICAL CHARACTERIZATION OF MULTIYEAR DROUGHTS

As indicated in Section 8.1, Yevjevich (1967) introduces the definition of a drought event as the
succession of consecutive intervals where the hydrological variable of interest remains below a
threshold level x0. Thus for a fixed threshold x0, the main drought characteristics are drought length
L (length of negative run or number of consecutive intervals where Xt < x0 followed and preceded by
at least one interval where Xt ≥ x0), drought magnitude or accumulated deficit D (sum of the
individual deficits St = x0 − xt over the drought duration L), and drought intensity, defined as the
ratio of drought magnitude to the drought length, I = D∕L.

Statistical Analysis of Hydrologic Variables

 300 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

The probabilistic characterization of drought events is more complex than for low flows. Indeed,
the need to consider at least two characteristics (e.g., length and magnitude), which are not mutually
independent, coupled with the relatively short hydrologic records that are generally available, makes the
traditional inferential approach (i.e., fitting a distribution function to the observed drought characteris-
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

tic) applicable only when long records are available. However, such difficulties can be circumvented by
deriving analytically the probability distributions of drought properties, assuming that a certain
stochastic model represents the underlying hydrological series. Then, instead of fitting a probability
distribution directly to the historical drought characteristics one may fit a stochastic model to the
underlying hydrological series and determine the distribution and the parameters of the drought
properties either by analytical methods or by Monte Carlo simulation. In this section we describe
alternative statistical methods for characterizing droughts, namely empirical methods based on
historical data or based on simulated series and analytical methods based on closed form or approximate
equations. Mishra and Singh (2011) is a recent review of drought modeling concepts and methods.

8.6.1 Probability Distributions and Moments of Drought Characteristics

Finding the distribution and moments of droughts can be a complex undertaking because of the
multiple variables that are needed to define them. For this reason, we approach the problem by
determining the statistical properties of individual variables first, then bringing them together in a
joint framework. This procedure is described in the following sections.

8.6.1.1 Probability Distribution and Moments of Drought Length

The properties of run length and its application to drought have been widely investigated in the
literature. For example, Downer et al. (1967) study the distribution and the statistical moments of
positive and negative run lengths for a sequence of independent identically distributed normal and
log–normal random variables. Also, Llamas and Siddiqui (1969) consider the case of a two-state
lag-1 Markov process and derive analytical expressions for determining the probabilities of runs of
wet and dry years of specified lengths. Since then numerous studies and developments have been
reported (e.g., Sen 1976; Chang et al. 1984a, b; Loaiciga and Leipnik 1996; Salas et al. 2001; Loaiciga
2005; Cancelliere and Salas 2010).
The simplest way to model drought length is assuming that the underlying water supply series is
iid. In this case, the probability mass function (PMF) of drought length is geometric:

f L ðℓÞ = PðL = ℓÞ = ð1 − p1 Þℓ−1 p1 (8-61)

where p1 = PðX t > x0 Þ is the parameter. The foregoing equation enables computing the probability
that a drought will last exactly ℓ time-steps. Then, the CDF, in other words, the probability that a
drought has a length equal to or smaller than ℓ time steps, can be computed as

X
ℓ
PðL ≤ ℓÞ = ð1 − p1 Þj−1 p1 = 1 − ð1 − p1 Þℓ (8-62)
j=1

Consequently, the expected value and the variance of drought length are given respectively by

EðLÞ = 1∕p1 (8-63)

VarðLÞ = ð1 − p1 Þ∕p21 (8-64)

The iid assumption may be acceptable in some cases, for instance when dealing with yearly
precipitation where the autocorrelation structure may be negligible. However, when dealing with

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 301

streamflows, the time-dependence structure is generally significant and therefore a model that is
capable of representing such temporal dependence may be needed. For example, if a Markov model
(Markov chain) is adopted for modeling the sequence of deficits and surpluses, the PMF of drought
length can be derived analytically. For a stationary simple Markov chain, the PMF of drought length
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

is also geometric:

f L ðℓÞ = P½L = ℓ = ð1 − p01 Þℓ−1 p01 (8-65)

where the parameter p01 is the transition probability as defined previously. Equation (8-65)
enables computing the probability that a drought will last exactly ℓ time steps. The CDF of L, in
other words, the probability that a drought has a length equal to or smaller than ℓ time steps, is
given by

X
ℓ
P½L ≤ ℓ = ð1 − p01 Þj−1 p01 = 1 − ð1 − p01 Þℓ (8-66)
j=1

Likewise, the probability that a drought has a length greater that ℓ time steps is

P½L > ℓ = 1 − P½L ≤ ℓ = ð1 − p01 Þℓ (8-67)

The expected value and the variance of drought length follow from Equation (8-65) as

1
EðLÞ = (8-68)
p01

1 − p01
VarðLÞ = (8-69)
p201

Given a sample x1, x2, : : : , xN, where N is the sample size, and the threshold x0, the transition
probability p01 can be estimated using maximum likelihood by counting the number of times a
deficit is followed by a surplus N01 and the number of times a deficit is followed by a deficit N00. Then

N 01
^p01 = (8-70)
N 00 þ N 01

Similarly, one can estimate p1 by counting the number of surpluses N1 as

^p1 = N 1 ∕N (8-71)

In practice these equations can be employed only if a sufficient number of transitions N00 and
N01 are observed. However, if the hydrological series is short or the threshold x0 low, the number of
observed transitions may be too small and the estimation of p01 by Equation (8-70) and p1 by
Equation (8-71) may be unreliable.
An alternative is to use a parametric approach similar to that shown in Section 8.5. By definition,
p01 can be written as

Statistical Analysis of Hydrologic Variables

 302 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

P½X t ≤ x0 , X t−1 ≤ x0 
p01 = 1 − p00 = 1 − P½X t ≤ x0 jX t−1 ≤ x0  = 1 − (8-72)
P½X t−1 ≤ x0 
where P½X t ≤ x0 , X t−1 ≤ x0  is the joint distribution of X t and X t−1 and P½X t−1 ≤ x0  is the CDF of
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

X t−1 . Assuming a bivariate normal distribution for (Xt, Xt−1), the following expression can be
adopted to estimate p01 (Cramer and Leadbetter 1967):

Zρ
1 exp½−x20 ∕ð1 þ zÞ
^p01 = 1 − ^p0 − pffiffiffiffiffiffiffiffiffiffiffiffi dz (8-73)
2π^p0 1 − z2
0

where ρ = lag-1 serial correlation coefficient of Xt, ^p0 = 1 − ^p1 , and ^p0 can be estimated by
 
x0 − μ^x
^p0 = Φ (8-74)
^
σx

where Φ(.) represents the standard normal CDF, and μ ^x and σ^x are the sample mean and sample
standard deviation of Xt, respectively.
Furthermore, when the underlying series exhibits a strong autocorrelation or serial dependence
a better alternative to the lag-1 Markov model for modeling the sequences of deficits and surpluses is
the DARMA(1, 1) model of Equation (8-54) in Section 8.5.1.5. Low-order DARMA models,
originally introduced by Jacobs and Lewis (1977, 1978), have been applied to daily precipitation
(Chang et al. 1984a) and monthly streamflows (Chebaane et al. 1995). Also Chung and Salas (2000)
use a DARMA(1, 1) model to derive the return period and risk of droughts. In particular, they show
that the DARMA(1, 1) model is better than the simple Markov chain for modeling drought length
probabilities where streamflows exhibit strong autocorrelation, as is the case of the Niger River’s
annual streamflows. The probability distribution of the length L of a run of state i = 0 (drought
length) for a DARMA(1, 1) model has been derived by Chang et al. (1984b). The expectation of L is

π0 ½1 − βλ þ βð1 − λ − β þ 2βλÞf1 − π0 g
E½L = (8-75)
f1 − π0 g½1 − λð1 − βÞf1 − βπ0 g − βπ0 f1 − βð1 − λÞg

Other properties and applications of DARMA models for modeling drought lengths can be
found in the literature (e.g., Chebaane et al. 1995, Salas et al. 2001, Cancelliere and Salas 2010). As an
alternative to DARMA models, higher-order Markov chains have been proposed (e.g., Akyuz et al.
2012, Tabari et al. 2015). Also, drought-length properties related to periodic processes such as
monthly streamflows have been reported by Cancelliere and Salas (2004).

Example 8-8: Fitting the Probability Distribution of Drought Lengths

Table 8-7 lists 119 years of records (1984–2002) of annual streamflows of the Poudre River at the
Mouth of the Canyon gauging station. Because diversions and storage facilities are upstream from
the gauging station, the measured flows have been adjusted (naturalized) so as to approximate the
natural flows that would have existed at the site. Figure 8-11 shows the time series of annual flows
and the threshold x0 = 299,000 ac-ft (equal to the long-term mean). The figure indicates that various
drought episodes have occurred on the Poudre River throughout the historical record, such as those
of the 1930s and 1950s. In particular, two 8 year droughts occurred in the years 1930–1937 and
1987–1994. The coefficients of variation, skewness, and lag-1 serial correlation are 0.36, 0.98, and
0.153, respectively, which are characteristic of streams in the semi-arid western United States.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 303

Table 8-7. Annual Streamflows (thousands acre-ft) of the Poudre River at Mouth of the Canyon
(1884–2002).

Streamflow Streamflow Streamflow Streamflow

ft3/s ft3/s ft3/s ft3/s
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Year Year Year Year

1884 695 1914 410 1944 236 1974 329

1885 514 1915 230 1945 249 1975 278
1886 338 1916 270 1946 203 1976 206
1887 332 1917 520 1947 336 1977 129
1888 202 1918 320 1948 226 1978 330
1889 224 1919 150 1949 379 1979 372
1890 264 1920 410 1950 205 1980 471
1891 298 1921 437 1951 330 1981 193
1892 236 1922 199 1952 316 1982 298
1893 252 1923 453 1953 202 1983 702
1894 341 1924 481 1954 122 1984 440
1895 392 1925 211 1955 167 1985 261
1896 255 1926 428 1956 242 1986 368
1897 377 1927 264 1957 441 1987 169
1898 221 1928 325 1958 289 1988 287
1899 420 1929 330 1959 251 1989 192
1900 516 1930 227 1960 238 1990 268
1901 368 1931 172 1961 350 1991 295
1902 206 1932 232 1962 312 1992 237
1903 353 1933 272 1963 155 1993 286
1904 395 1934 127 1964 192 1994 190
1905 378 1935 277 1965 340 1995 358
1906 300 1936 263 1966 150 1996 321
1907 410 1937 198 1967 247 1997 366
1908 290 1938 358 1968 264 1998 282
1909 500 1939 212 1969 256 1999 384
1910 290 1940 149 1970 351 2000 198
1911 230 1941 212 1971 367 2001 200
1912 350 1942 360 1972 238 2002 95
1913 230 1943 369 1973 377 — —
Source: Northern Colorado Water Conservancy District, Loveland, Colorado.

Because the lag-1 autocorrelation coefficient is small, a lag-1 Markov model has been adopted to
derive the PMF of drought length. Thus, by counting the number of transitions between states, the
following results are obtained: N00 = 39 and N01 = 27. Then, applying Equation (8-70), the estimate
of the parameter ^p01 of the geometric distribution is

N 01 27
^p01 = = = 0.41
N 00 þ N 01 39 þ 27

Then, from Equation (8-65) the PMF of drought length is

f L ðℓÞ = P½L = ℓ = ð1 − 0.41Þℓ−1 0.41, ℓ = 1, : : : :

Statistical Analysis of Hydrologic Variables

 304 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-11. Annual flow records of the Poudre River at Mouth of the Canyon (1884–2002) and
threshold level x0 equal to the long-term mean.
Source: Salas et al. (2005).

Figure 8-12. Sample PMF of drought length (histogram) and PMF obtained from the fitted simple
Markov chain model (continuous line) for the annual flows of the Poudre River at Mouth of the Canyon.

Figure 8-12 shows the derived geometric PMF and the sample frequency distribution of
observed drought lengths. The probability of a drought longer than the longest one observed on
record can be computed by Equation (8-67) as

P½L > 8 = 1 − P½L ≤ 8 = ð1 − 0.41Þ8 = 1.5%

In addition, one can calculate E(L) and σðLÞ from Equations (8-68) and (8-69), respectively.
They give E(L) = 2.44 and σðLÞ = 1.87 years, respectively.
If sufficiently long records of the underlying hydrological series are available (as is the case for the
data in this example), one may be able to estimate the parameters of the probability distribution of
drought length from the observed drought events. For example, assuming that a simple Markov chain
represents the underlying series of zeros and ones, Equation (8-68) can be used to estimate p01 as
1 1
^p01 = =
^
EðLÞ μ^L
^L can be estimated from the sample of observed droughts as
where μ

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 305

Table 8-8. Drought Length (L), Drought Magnitude (D), and Drought Intensity (I) Identified for the
Streamflows of the Poudre River at Mouth of the Canyon (1884–2002) Assuming a Threshold
x0 = μx = 299 × 103 acre-ft.

D (103 acre-ft) I (103 acre-ft/year)

Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Begin year End year L (years)

1888 1893 6 318.07 53.01

1896 1896 1 44.01 44.01
1898 1898 1 78.01 78.01
1902 1902 1 93.01 93.01
1908 1908 1 9.01 9.01
1910 1911 2 78.02 39.01
1913 1913 1 69.01 69.01
1915 1916 2 98.02 49.01
1919 1919 1 149.01 149.01
1922 1922 1 100.01 100.01
1925 1925 1 88.011 88.011
1927 1927 1 35.011 35.011
1930 1937 8 624.09 78.011
1939 1941 3 324.03 108.01
1944 1946 3 209.03 69.678
1948 1948 1 73.011 73.011
1950 1950 1 94.011 94.011
1953 1956 4 463.05 115.76
1958 1960 3 119.03 39.678
1963 1964 2 251.02 125.51
1966 1969 4 279.05 69.761
1972 1972 1 61.011 61.011
1975 1977 3 284.03 94.678
1981 1982 2 107.02 53.511
1985 1985 1 38.011 38.011
1987 1994 8 468.2 58.525
1998 1998 1 17.356 17.356
2000 2002 3 404.01 134.67
Mean 2.39 177.68 72.76
Variance 4.10 25,829.14 1,200.64

X
m
^L = ð1∕mÞ
μ Li
i=1

where m is the number of drought episodes. For example, Table 8-8 gives the durations of the various
drought lengths that occurred on the Poudre River. The sample mean of drought length is μ ^L = 2.39, so
that ^p01 = 0.418. This would be an alternative procedure for estimating the parameter of the geometric
distribution as referred to in the first part of this example.

8.6.1.2 Fitting Probability Distributions to Drought Magnitude and Drought Intensity

Long records of hydrologic data may enable fitting the distribution of drought magnitude and
intensity. For example, one may assume that the probability distribution of drought magnitude is a
two-parameter gamma as

Statistical Analysis of Hydrologic Variables

 306 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

 r −1
1 d D −βd
f D ðdÞ = e D (8-76)
βD Γðr D Þ βD

where βD and r D represent the scale and shape parameters, respectively. The moment estimators of
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

βD and rD are given by

^βD = σ
^2D ∕^
μD and ^rD = μ
^2D ∕^
σ2D (8-77)

^D and σ
where μ ^2D are the sample mean and variance, respectively, which may be determined from the
observations of drought magnitude Dj as

1X m
1 X m
^D =
μ ^2D =
Dj and σ ^D Þ2 :
ðD − μ
m j=1 m − 1 j=1 j

Similarly, one may assume a two-parameter gamma distribution for the drought intensity, fI(i),
and use the method of moments for estimating the parameters β1 and r 1 .

Example 8-9: Fitting the Probability Distribution of Drought Magnitude and Drought Intensity
for the Annual Flows of the Poudre River at the Mouth of the Canyon
Table 8-8 gives the drought length, magnitude, and intensity of the droughts identified for the annual
flows of the Poudre River at the Mouth of the Canyon (see Example 8-8), assuming a threshold x0
equal to the mean. One may observe that the drought with the largest magnitude (624.09 × 103 ac-ft)
occurred in 1930–1937 and lasted eight years. We would like to determine the probability that
droughts larger than the maximum observed drought magnitude will occur on the Poudre River.
Using the sample mean μ ^D and variance σ ^2D obtained in Table 8-8, we apply the moment
Equation (8-77) for estimating the parameters of the gamma distribution for drought magnitude as

^βD = σ
^2D ∕^
μD = 25829.14∕177.68 = 145.37 and
^2D ∕^
^rD = μ σ2D = ð177.68Þ2 ∕25829.14 = 1.22

Likewise, the corresponding parameter estimates for drought intensity are

^βI = σ
^2I ∕^
μI = 16.5 and ^rI = μ
^2I ∕^
σ2I = 4.41

Replacing the parameters in the corresponding probability distributions, the probability of

various drought events can be computed. Alternatively, one may be interested in determining
drought magnitude or intensity corresponding to a given probability.
For instance, the probability of a drought magnitude greater than the largest observed value
(i.e., D = 624.09 × 103 ac-ft) can be computed by integrating the corresponding PDF as

Z
624.09  1.22−1
1 d
e−145.37 dd = 0.02
d
P½D > 624.09 = 1 −
145.37Γð1.22Þ 145.37
0

The foregoing integral can be determined using standard routines for computing the gamma
integral. For instance, using Matlab, the command P = gamcdf(624.09, 1.22, 145.37) = 0.978 can be
used to estimate the required integral. In a similar fashion, one can compute the probability that a
drought with intensity greater than a fixed value, say 130 × 103 ac-ft/year, will occur as

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 307

Z
130  
1 d 4.41−1 − d
P½I > 130 = 1 − e 16.50 dd = 0.067
16.50Γð4.41Þ 16.5
0
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

In addition, one may want to determine the drought magnitude for a given exceedance
probability, for example, p = 5%, using the inverse of the gamma distribution. This can be achieved
numerically by using, for instance, the Matlab function gaminv(q, r, β), where q is the nonexceedance
probability, and β and r are the parameters. In our case, q = 1 − 0.05 = 0.95, and the desired
drought magnitude is obtained as D = gaminv(0.95, 1.22, 145.37) = 496.22 × 103 ac-ft.

8.6.1.3 Analytical Approximations of Probability Distributions of Drought Magnitude

Drought magnitude (accumulated deficit) may be considered as a random sum of random variables, in
other words., the sum of L deficits St = x0 − X t , where drought length L is a random variable. The
analytical derivation of the probability distribution of drought magnitude is generally cumbersome and
closed-form solutions are feasible only in a few cases (e.g., Sen 1976). However, analytical approx-
imations have been proposed that overcome the difficulties of applying the inferential approach when a
small number of droughts are observed. Such analytical approximations are generally based on
deriving the moments of drought magnitude as a function of the moments of the individual deficits
and the moments of drought length. Such moments can then be used to estimate the parameters of
probability distributions by the method of moments. To this end, several authors have adopted the
gamma distribution as the underlying distribution of drought magnitude (e.g., Guven 1983, Shiau and
Shen 2001, Bonaccorso et al. 2003, Gonzalez and Valdes 2003, Salas et al. 2005, Mishra et al. 2009).
More specifically, assuming serial independence for the underlying hydrological variable, the
expected value and variance of D are given by the following expressions (Sen 1977):

EðDÞ = EðLÞ EðSÞ = μL μS (8-78)

VarðDÞ = EðLÞVarðSÞ þ VarðLÞE2 ðSÞ = μL σ2S þ σ2L μ2S (8-79)

where
individual deficit S = truncated variable S = x0 − Xt such that x0 > Xt and t is any time interval,
μL and μS = means of L and S, respectively, and
σ2L and σ2S = corresponding variances.

The gamma distribution with PDF,

 r−1
1 d
e− β ,
d
f D ðdÞ = (8-80)
βΓðrÞ β

is assumed as the distribution of drought magnitude D where r and β are the parameters. Then using
the method of moments and Equations (8-78) and (8-79), the gamma parameters can be expressed
as a function of the moments of deficit S and drought length L as

μ2L μ2S
r= (8-81)
μL σ2S þ σ2L μ2S

μL σ2S þ σ2L μ2S

β= (8-82)
μL μS

Statistical Analysis of Hydrologic Variables

 308 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

The foregoing equations allow expressing the PDF of drought magnitude [Equation (8-80)] as a
function of the moments of drought length L and deficit S. The moments of L can be determined
based on the procedure outlined in Section 8.6.1.1, while the moments of S can be estimated from the
sample mean and sample variance of the observed deficits, in other words,
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

1X k
^S =
μ S (8-83)
k i=1 i

1 X k
^2S =
σ ^ S Þ2
ðS − μ (8-84)
k − 1 i=1 i

Alternatively, another method can be employed for estimating the moments of the individual
deficits S, noting that the expected value and the variance of S can be computed as a function of the
probability distribution of the underlying hydrological variable Xt. This method is particularly useful
in cases of short records or in cases where the threshold level x0 is small (lower than the sample
mean). In these cases, the number of drought events obtained from the historical records will be
small, and consequently the statistics derived from them will be unreliable. Bonaccorso et al. (2003)
derive the moments of S for the case of normal, log–normal, and gamma-distributed Xt and the
related expressions of the parameters r and β of the gamma distribution for drought magnitude,
under the assumption of negligible autocorrelation in the series. Table 8-9 gives the mean μS and
variance σ2S of the deficits, and Table 8-10 gives the parameters r and β for the three referred
distributions, where the threshold is parameterized as (Yevjevich 1967)

x0 = μx − ασx = μx ð1 − αCvx Þ (8-85)

where μx and Cvx are the mean and the coefficient of variation of the underlying hydrological
variable Xt, and α is a threshold coefficient. For instance, if α = 0, then x0 = μx.

Example 8-10: Fitting the Probability Distribution of Drought Magnitude for the Annual Flows of
the Salso River at Pozzillo Reservoir (Italy) Based on Analytical Approximations
Table 8-11 shows the annual flows of the Salso River at Pozzillo reservoir (Italy) for 1959–1998
(i.e., 40 years of records). The coefficients of variation and skewness and the lag-1 serial correlation are
0.64, 1.51, and −.02, respectively. Table 8-12 reports the main characteristics of the droughts identified
assuming a threshold x0 = 159.89 mm (the long-term mean). The table shows that only nine droughts
occurred during the historical period 1959–1998. We are interested in computing the probability of
observing a drought with magnitude larger than the maximum observed. However, the number of
identified droughts is inadequate to fit a probability distribution to the observed drought statistics.
Therefore, the previously outlined approach will be adopted for deriving the probability
distribution of drought magnitude. Such an approach capitalizes on the statistics of the whole
deficit series and therefore leads to more reliable estimation of the probability distribution of drought
characteristics, especially when the number of observed droughts is limited. We assume the gamma
PDF for drought magnitude D as in Equation (8-80), in which the parameters r and β will be
estimated using Equations (8-81) and (8-82), where the moments of the deficit S are replaced by their
sample moments, and the moments of L are computed from Equations (8-63) and (8-64). The
sample moments of the observed deficits S are

1X k
1 X k
^S =
μ ^2S =
Si = 51.65 and σ ^S Þ2 = 1,857.3
ðS − μ
k i=1 k − 1 i=1 i
in which k = 27 is the number of the observed deficits Si.

Statistical Analysis of Hydrologic Variables

 Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Table 8-9. Mean μS and Variance σ2S of the Deficits S for Different Distributions of Xt.
Distribution of Xt μS σ2S Other parameters
  n h i o
ϕð−αÞ
Normal (μx, σx) σx −α þ p10 ϕð−αÞ σ2x p0 α − ϕð−αÞ
p0 þ1 p0 = Φð−αÞ
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  i h i
hexpðσ2y Þ 2 σ
1 2 2 Δ xÞ
Log–normal (μy, σy) exp μy þ 2 σy 1 − α expðσy Þ − 1 − p0 exp 2σy þ σ2y p0 · Ψ − Δp2 p0 = Φ 2y þ lnð1−αCv
σy
0
h i
σy lnð1−αCv x Þ
Δ=Φ − 2 þ σy
h i
3σy lnð1−αCv x Þ
Ψ=Φ − 2 þ σy

h i h i h i
βx ðr x þ1Þ
Gamma (rx, βx) βx r x 1 − αCv x − pΘ0 βx r x p0 Ω − βpx 2r x Θ2 p0 = G r x , r x 1 − αCv x
0
h i
Θ = G r x þ 1, r x 1 − αCv x
h i
Ω = G r x þ 2, r x 1 − αCv x

Statistical Analysis of Hydrologic Variables

1
Note that the symbols ϕðzÞ and ΦðzÞ denote the PDF and CDF of the standard normal variable Z. Also Gða, xÞ = x a−1 e−t dt
ΓðaÞ ∫0 t is the regularized lower incomplete gamma function.
Source: Bonaccorso (2002).
LOW FLOWS AND DROUGHTS
309
 310 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Table 8-10. Parameters of the Gamma Distribution of Drought Magnitude for Different Distributions
of Xt.

Distribution of Xt R β
h i
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

 p0

ϕð−αÞ 2 ϕð−αÞ ϕ2 ð−αÞ
ϕð−αÞ
−αþ p σx p1 −αþ p0 − α
p0 þ p2 −1
Normal (μx, σx)  0
ϕð−αÞ
0 
p0 −αþ
ϕð−αÞ 2
−p1 α
ϕð−αÞ ϕ2 ð−αÞ −αþ p
p0 p0 þ p2 −1 0
0
h 2 i
2
−Δ2 þexpðσ2y Þpψ
p0 2
1−αCv x −pΔ μx p1 1−αCv x −pΔ
Log–normal (μy, σy) 2 0
0 p
0  0

p0 1−αCv x −pΔ −p1 Δ2

−expðσ2y ÞpΨ 1−αCv x −pΔ
0
0 p2 0
0
 
2
2 μx
p0
1−αCv x −pΘ
2
−Θ2 þpΩ ð Cv 2x þ1 Þ
1−αCv x −pΘ p1 0 p 0
Gamma (rx, βx) n 2 h 0
io
0

p0 1−αCv x −pΘ −p1 Θ2 Ω
− Cv 2x þ1 1−αCv x −pΘ
0
0 p2 p0
0

Source: Bonaccorso (2002), Cancelliere et al. (2003).

Table 8-11. Annual Flows (mm) of the Salso River at Pozzillo Reservoir (1959–1998).

Year Streamflows Year Streamflows Year Streamflows Year Streamflows

1959 149.72 1969 245.65 1979 158.23 1989 14.66

1960 147.58 1970 101.45 1980 157.10 1990 15.07
1961 138.91 1971 79.91 1981 172.96 1991 61.45
1962 128.81 1972 282.73 1982 138.54 1992 142.12
1963 209.52 1973 511.62 1983 91.21 1993 124.77
1964 334.31 1974 103.02 1984 144.02 1994 171.48
1965 175.08 1975 68.29 1985 202.29 1995 51.45
1966 198.02 1976 395.02 1986 108.79 1996 372.73
1967 201.72 1977 53.41 1987 153.67 1997 109.29
1968 120.16 1978 132.14 1988 157.31 1998 71.56

Table 8-12. Main Characteristics of the Droughts Identified for the Flows of the Salso River at Pozzillo
Reservoir (1959–1998), Assuming a Threshold x0 = μx = 159.89 mm.

Begin year End year L (years) D (mm) I (mm/year)

1959 1962 4 74.56 18.64

1968 1968 1 39.73 39.73
1970 1971 2 138.43 69.21
1974 1975 2 148.48 74.24
1977 1980 4 138.70 34.67
1982 1984 3 105.91 35.30
1986 1993 8 501.31 62.66
1995 1995 1 108.44 108.44
1997 1998 2 138.94 69.47

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 311

Conversely, the mean and the variance of L are determined from Equation (8-63) and (8-64),
respectively, as
1 1 − p1
EðLÞ = = 3.08 years and VarðLÞ = = 6.39 years2
p21
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

p1
where p1 = 0.325 has been computed following the same procedure as that outlined in Section 8.6.1.1.
Then, the parameters r and β of the gamma distribution are computed from Equations (8-81) and
(8-82), respectively, as

μ2L μ2S
r= = 1.109
μL σ2S þ σ2L μ2S
μL σ2S þ σ2L μ2S
β= = 143.23
μL μS

Note that for estimating the parameters r and β, the whole series of 27 deficits has been used,
which should lead in principle to more reliable estimates as compared with the case when r and β are
estimated directly from the observed nine droughts.
Then the probability of observing a drought with a magnitude greater than the largest observed
in the historical record, i.e., D = 501.31, is computed by

Z
501.31  1.109−1
1 d
e−143.23 dd = 0.037
d
P½D > 501.31 = 1 −
143.23Γð1.109Þ 143.23
0

The foregoing integral is computed using the Matlab gamma function command:

gamcdf ð501.31, 1.109, 143.23Þ = 0.963

Example 8-11: Deriving the Probability Distribution of Drought Magnitude for the Annual Flows
of the Salso River at Pozzillo Reservoir Using an Alternative Method
In this example, the probability distribution of drought magnitude for the Salso River at Pozzillo
Reservoir is derived, using the alternative approach outlined previously. More specifically, we are
interested in analyzing drought characteristics considering the threshold level x0 = 80 mm. This
threshold represents about 50% of the sample mean, and not many drought events will result from
the empirical flow data, as Table 8-13 shows, where only six droughts occurred in 1959–1998.
Furthermore, the number of deficits is only eight, which is too small to apply the procedure
illustrated in Example 8-10. Therefore, the alternative approach will be applied for deriving the
probability distributions of drought magnitude. Such an approach capitalizes on the statistics
computed from the whole hydrological series and therefore leads to more reliable estimation of the
probability distribution of drought characteristics, especially when the number of observed droughts
is small.
Assuming the gamma distribution is the underlying model for drought magnitude, its
parameters r and β can be determined from the equations in Table 8-10. We also assume that
the distribution of the streamflow series for the Salso River is gamma with parameters rx and βx.
Because the mean and variance of the historical streamflows are μ ^x = 159.89 and σ^2x = 10,397,
respectively, using the method of moments we estimate the parameters rx and βx as
^rx = μ σ2x = 2.46 and ^βx = σ
^2x ∕^ ^2x ∕^ ^v = σ
μx = 65.02. And the coefficient of variation is C ^x ∕^μx = 0.64.

Statistical Analysis of Hydrologic Variables

 312 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Table 8-13. Main Characteristics of the Droughts Identified for the Flows of the Salso River at Pozzillo
Reservoir (1959–1998) Assuming a Threshold x0 = 80 mm.

Begin year End year L (years) D (mm) I (mm/year)

Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

1971 1971 1 0.09 0.09

1975 1975 1 11.71 11.71
1977 1977 1 26.59 26.59
1989 1991 3 148.82 49.61
1995 1995 1 28.55 28.55
1998 1998 1 8.44 8.44

Also, because the assumed threshold is x0 = 80 mm, the threshold coefficient α can be computed
from Equation (8-85) as

μx − x0 159.89 − 80
α= = = :783
σx 101.96

And the parameters r and β of the gamma distribution of drought magnitude are computed
using the equations of the last row of Table 8-10. Thus, the following results are obtained:

p0 = G½rx , r x ð1 − αCv Þ = 0.23, p1 = 1 − p0 = 0.77

Θ = G½rx þ 1, rx ð1 − αC v Þ = 0.0736, Ω = G½r x þ 2, r x ð1 − αCv Þ = 0.0192,
2
1 − αC v − pΘ
^r = n h 0 io = 1.767, and
Θ 2
p0 1 − αCv − p − p1 Θp2 − pΩ C2v þ 1
2

0 0
h i
0

2
μx pp0 1 − αC v − pΘ − Θp2 þ pΩ C 2v þ 1
2

^β = 1 0 0 0
= 20.55:
1 − αC v − pΘ
0

Then, the probability that a drought with magnitude greater than the largest observed drought
magnitude is computed as

Z
148.82  1.767−1
1 d
e−20.55 dd = 0.004
d
P½D > 148.82 = 1 −
20.55Γð1.767Þ 20.55
0

Again, the previous integral can be computed using one of the numerical standard routines for
computing the gamma integral.

8.6.1.4 Analytical Approximations for Determining the Joint Probability Distributions of

Drought Characteristics
In some practical problems, considering two or more drought characteristics may be necessary. For
instance, one may be interested in computing the probability that a drought of magnitude D exceeds
a certain value given that the duration of the drought is equal to a given number of years. Because the
drought characteristics, L, D, and I are not mutually independent, one must use a multivariate

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 313

formulation to model their joint occurrence. For this purpose, one may use bivariate distributions for
any pair of drought characteristics or copulas (e.g., Salas et al. 2005, Shiau 2006, Serinaldi et al. 2009,
Songbai and Singh 2010). Such an approach, however, requires a rather long series of observations,
because it requires estimating the parameters of the distributions that are to be fitted to the observed
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

droughts.
An alternative approach is based on conditional distributions, which capitalizes on the links
between D and L and between I and L. For example, the bivariate PDF of D and L, f D,L ðd, ℓÞ can be
expressed as the product of the conditional distribution of the two variables as (e.g. Salas et al. 2005)

f D,L ðd, ℓÞ = f DjL = ℓ ðdÞf L ðℓÞ (8-86)

where f DjL = ℓ ðdÞ is the distribution of D conditional on a fixed value of L, and f L ðℓÞ is the marginal
PDF of L. A similar expression can be written for the bivariate distribution of I and L as

f I,L ði, ℓÞ = f IjL = ℓ ðiÞf L ðℓÞ (8-87)

Thus, the joint distributions f D,L ðd, ℓÞ and f I,L ði, ℓÞ can be determined once the conditional
distributions f DjL = ℓ ðdÞ and f IjL = ℓ ðiÞ and the marginal distribution f L ðℓÞ are known. Some authors
have assumed a parametric distribution for DjL, and have estimated the parameters from observed
droughts (e.g., Guven 1983, Sharma 1995, Shiau and Shen 2001). In some cases, due to the limited
number of droughts that can be observed from the available records, synthetic generation (Wang and
Salas 1989; Shiau and Shen 2001) or long hydrologic series reconstructed from tree ring records
(e.g., Gonzalez and Valdes 2003, Biondi et al. 2005) have been utilized.
Alternatively, an approach similar to that described in Section 8.6.1.3 can be employed for
estimating the parameters of the distribution of DjL that capitalizes on analytical expressions of the
moments of drought magnitude. More specifically, the approximate moments of drought magnitude
computed based on the distribution of the underlying process X t and on the threshold are used to
estimate the parameters by the method of moments. This enables exploiting the available
information from the observed series, thus allowing a reliable estimation of the distribution of
drought characteristics even for a relatively short sample series.
Assuming the underlying series X t is iid., the mean and variance of drought magnitude
conditional on a fixed drought length L = ℓ are given respectively by

EðDjL = ℓÞ = ℓEðSÞ (8-88)

VarðDjL = ℓÞ = ℓVarðSÞ (8-89)

However, the iid assumption is not valid because most hydrologic series exhibit significant
autocorrelation. In this case, the computation of the conditional moments of drought magnitude is
more involved, and no closed-form solution is available. Nevertheless, some approximations have
been suggested (e.g., Salas et al. 2005, Biondi et al. 2005, Cancelliere and Salas 2010). For example,
Cancelliere and Salas (2010) provide empirical approximations that enable computing the condi-
tional moments as a function of the skewness coefficient of the underlying variable X t , the threshold
parameter α, and the lag-1 autocorrelation coefficient ρ1 . The proposed approximations for the mean
and the variance of drought magnitude of fixed length ℓ are (Cancelliere and Salas 2010)

EðDℓ Þ = μS am ℓbm (8-90)

VarðDℓ Þ = σ2S av ℓbv (8-91)

Statistical Analysis of Hydrologic Variables

 314 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

where EðDℓ Þ and VarðDℓ Þ are respectively the expected value and variance of the drought magnitude
of length ℓ for series that are autocorrelated and skewed; μS and σ2S are respectively the expected value
and variance of a single-year deficit assuming that the series are skewed but uncorrelated, thus they
are functions of the given marginal distribution; and the threshold x0 , and the parameters am , bm , av ,
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

and bv are related to ρ1 and x0 through the following expressions:

am = 1.0 þ ð:6983α − :5592Þρ1 þ ð−:6634α − :3418Þρ21 (8-92a)

bm = 1.0 þ ð−0.1840α þ 0.5903Þρ1 þ ð0.1865α þ 0.0839Þρ21 (8-92b)

av = 1.0 þ ð0.7415α − 1.0325Þρ1 þ ð−0.7969α − 0.0928Þρ21 (8-93a)

bv = 1.0 þ ð−0.4414α þ 1.078Þρ1 þ ð0.4175α þ 0.5707Þρ21 (8-93b)

Note that for ρ1 = 0, EðDℓ Þ = μS ℓ and VarðDℓ Þ = σ2S ℓ because am = bm = av = bv = 1.

Cancelliere (2008) compares several distributions by simulation, including log–normal, gamma,
and beta, among others, for drought magnitude conditioned on a fixed drought length where the
underlying series is autocorrelated. In particular, the method of moments has been applied to
estimate the parameters of each distribution, using the approximate moments derived in Equa-
tions (8-90) and (8-91). The results of the comparisons based on statistical goodness-of-fit tests
reveal that the beta distribution is to be preferred, besides it is bounded, as is the case of the drought
magnitude.
The beta PDF takes the form (Johnson et al. 1994)

1 ðd − aÞp−1 ðb − dÞq−1
f DjL = ℓ ðdÞ = ða ≤ d ≤ bÞ (8-94)
Bðp, qÞ ðb − aÞpþq−1

where Bðp, qÞ is the complete beta function, and a and b are the lower and upper bounds,
respectively. In our case, a = 0 and b = ℓxo , because a drought of length ℓ cannot have magnitude
greater than ℓxo , and the parameters p and q can be estimated as a function of the first two moments
of the drought magnitude μD = EðDℓ Þ and σ2D = VarðDℓ Þ as (Cancelliere and Salas 2010)
 2  
μ μD μ
p= D 1− − D (8-95)
σD ℓxo ℓxo

μD ðℓxo − μD Þ
q= − ð1 þ pÞ (8-96)
σ2D

where μD and σ2D are determined from Equations (8-90) and (8-91), respectively. Then, the bivariate
PDF of drought magnitude and length takes the following form:

1 ðdÞp−1 ðℓxo − dÞq−1

f D,L ðd, ℓÞ = f L ðℓÞ, ð0 ≤ d ≤ ℓxo Þ (8-97)
Bðp, qÞ ðℓxo Þpþq−1

where f L ðℓÞ is the PDF of drought length (Section 8.6.1.1).

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 315

Furthermore, from the conditional distribution of drought magnitude given drought length, the
conditional distribution of drought intensity I given drought length L can be obtained. Indeed,
because drought intensity is the ratio of drought magnitude and drought length, in other words,
I = D∕L, the conditional PDF of drought intensity I given a fixed length L = ℓ can be derived from
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Equation (8-94) as

1 ðiÞp−1 ðx0 − iÞq−1

f IjL ðiÞ = , ð0 ≤ i ≤ x0 Þ (8-98)
Bðp, qÞ ðx0 Þpþq−1

where p and q are given by Equations (8-95) and (8-96), respectively. Thus, the bivariate PDF of
drought intensity and length can be found as (Cancelliere and Salas 2010)

1 ðiÞp−1 ðx0 − iÞq−1

f I,L ði, ℓÞ = f L ðℓÞ, ð0 ≤ i ≤ x0 Þ (8-99)
Bðp, qÞ ðx0 Þpþq−1

By integrating the bivariate PDFs, the occurrence probability of various drought events can be
found (Salas et al. 2005). For example, for specific drought events using the previously described
models we have
(1) for drought event E = [D > d0 and L = ℓ0 (ℓ0 = 1, 2, : : : )]:

ℓZo xo ℓZ
0 x0
1 ðzÞp−1 ðℓ0 x0 − zÞq−1
P½D > d 0 , L = ℓo  = f D,L ðz, ℓ0 Þdz = f L ðℓ0 Þ dz (8-100)
Bð p,qÞ ðℓ0 x0 Þpþq−1
d0 d0

(2) for drought event E = [D > d0 and L ≥ ℓ0 (ℓ0 = 1, 2, : : : )]:

P½D > d0 , L ≥ ℓ0 
∞ Z ∞  Zℓxo 
ℓ0 x 0
X X 1 ðzÞp−1 ðℓx0 − zÞq−1 (8-101)
= f D,L ðz, ℓÞdz = f L ðℓÞ dz
ℓ = ℓ0 ℓ = ℓ0
Bðp,qÞ ðℓx0 Þpþq−1
d0 d0

(3) for drought event E = [I > i0 and L = ℓ0 (ℓ0 = 1, 2, : : : )]:

Zx0 Zx0
1 ðzÞp−1 ðx0 − zÞq−1
P½I > i0 , L = ℓ0  = f I,L ðz, ℓ0 Þ dz = f L ðℓ0 Þ dz (8-102)
Bð p,qÞ ðx0 Þpþq−1
i0 i0

(4) for drought event E = [I > i0 and L ≥ ℓ0 (ℓ0 = 1, 2, : : : )]:

P½I > i0 , L ≥ ℓ0 
∞ Z ∞  Zx0 
x0
X X 1 ðzÞp−1 ðx0 − zÞq−1 (8-103)
= f I,L ðz, ℓÞ dz = f L ðℓÞ dz
ℓ = ℓ0 ℓ = ℓ0
Bð p,qÞ ðx0 Þpþq−1
i0 i0

Furthermore, the marginal probability of drought events E = {D > d0} or E = {I > i0} can be
obtained from Equations (8-101) or (8-103), respectively, by letting ℓ0 = 1. Note that despite the

Statistical Analysis of Hydrologic Variables

 316 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

apparent complexity of the aforementioned expressions, the integrations can be carried out
efficiently using numerical tools for the beta PDF that are available in most statistical software.

8.6.2 Return Period of Multiyear Droughts

Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

The return period of a drought event may be defined as the average elapsed time, or mean
interarrival time, between two such events (Lloyd 1970, Loaiciga and Mariño 1991, Fernández and
Salas 1999a). Shiau and Shen (2001) develop a formulation for determining the return period of
droughts with magnitude greater than or equal to a given value. Cancelliere and Salas (2004) extend
such a formulation for drought length in periodic series, while Cancelliere et al. (2003), Gonzalez and
Valdes (2003), Salas et al. (2005), and Biondi et al. (2005) include other characteristics such as
drought length and intensity, under the assumption of lag-1 Markov dependence.
The return period (or mean recurrence time) of droughts for any of the drought scenarios and
events E as specified previously, e.g., E = {L = ℓ0 and D > d0 } may be estimated as the average of the
recurrence times as

1 X E N −1
T= T ð jÞ (8-104)
ðN E − 1Þ j = 1 E

in which T E ð jÞ denotes the recurrence time of two successive drought events and N E is the number
of such drought events. Equation (8-104) is particularly useful for determining the return periods of
drought events based on a historical sample (particularly for a long record) or a sample generated
from a stochastic model. An alternative procedure based on analytical formulations is given in the
following.
To derive an analytical expression for estimating T, the formulation proposed by Shiau and Shen
(2001) for the case of drought events characterized only by the drought magnitude can be extended
to the more general case of drought events jointly defined in terms of drought magnitude and length
(or drought intensity and length). For this purpose, the interarrival time T(E) between two droughts
events E [e.g., E = (D > d0 and L = ℓ0 )] may be written as (Cancelliere and Salas 2002, Gonzalez and
Valdes 2003, Salas et al. 2005)

EðLÞ þ EðLn Þ
T= (8-105)
PðEÞ

where E(L) and EðLn Þ are the expected values of the duration of specific drought event and
nondrought event, respectively, and P(E) is the joint probability of the specific drought event that
may be determined from Equations (8-100)–(8-103) as the case may be. For the case of a lag-1
Markov process, E(L) can be computed from Equation (8-68) and EðLn Þ can also be obtained from
Equation (8-68) by replacing p01 with p10 . Alternatively, E(L) and EðLn Þ can be estimated from the
sample mean of the observed drought length and nondrought length, respectively. Note that
although for an autocorrelated process, the independence assumption between drought events is not
exactly met, yet Equation (8-105) provides an excellent approximation (e.g., Salas et al. 2005,
Cancelliere and Salas 2010).

Example 8-12: Computing the 50 Year Drought Magnitude for the Salso River at Pozzillo
Reservoir
In this example, the 50 year (return period) drought magnitude for the Salso River at Pozzillo
Reservoir (see Examples 8-10 and 8-11) will be computed assuming a threshold x0 equal to the mean.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 317

For the drought event E = {D > d}, i.e., a drought with a magnitude greater than d, the associated
return period will be computed by Equation (8-105), in other words,

EðLÞ þ EðLn Þ
T=
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

PðD > dÞ

In our case, E(L) and EðLn Þ will be estimated using the sample means of the observed drought
length L and nondrought length Ln . The desired T year magnitude d can be computed by solving for
d in the following equation:
L þ Ln
P½D > d =
T
where the probability can be computed by making use of the gamma CDF as in Example 8-10. In our
case, the following results are obtained:

L þ Ln
L = 3 years, Ln = 1.5 years, and P½D > d = = 0.09
T
Then, the drought magnitude d can be computed (as in Example 8-10) as the value with
exceedance probability 0.09. Recalling from Example 8-10 that the parameters of the gamma
distribution of drought magnitude for Salso River at Pozzillo Reservoir are r = 1.109 and β = 143.23
the following Matlab command is used to find the 50 year accumulated deficit d as

d = gaminvð1 − 0.09, 1.109, 143.23Þ = 372.12 mm

Example 8-13: Calculating the Return Period of Drought Magnitude and Length for the Annual
Flows of the Poudre River at the Mouth of the Canyon (Table 8-7)
In this example, we will analyze drought magnitude and length for the annual flows of the Poudre
River at the Mouth of the Canyon. In particular, the return period associated with a drought of
length L = 3 and magnitude greater than 300,000 ac-ft will be computed using the analytical
approximations for the joint distribution of drought length and magnitude illustrated in
Section 8.6.1.4. In what follows we will use the threshold x0 = 299,000 acre-ft (the sample mean)
and the threshold coefficient α = 0 [refer to Equation (8-85)].
The return period of such critical drought will be computed by Equation (8-105), in other
words,

EðLÞ þ EðLn Þ EðLÞ þ EðLn Þ

T= =
PðL = ℓ0 , D > d0 Þ PðL = 3, D > 300Þ

where E(L) and EðLn Þ are, respectively, the expected values of the durations of specific drought
events and nondrought events. These expected values can be estimated as the sample mean of
observed drought length L and nondrought length Ln , which gives L = 2.4 years and Ln = 1.9 years.
Furthermore, the probability P(L = 3, D > 300) can be computed by Equation (8-100). The
parameters p and q of the beta distribution can be estimated by Equations (8-95) and (8-96), in
which the mean and standard deviation of drought magnitude EðDℓ Þ and VarðDℓ Þ, are given by the
empirical expressions (8-90) and (8-91), respectively.
The coefficients in Equations (8-90) and (8-91) are computed considering that α = 0 and
ρ1 = 0.153 (Example 8-8) as

Statistical Analysis of Hydrologic Variables

 318 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

am = 1.0 þ ð:6983α − :5592Þρ1 þ ð−:6634α − :3418Þρ21 = 0.906

bm = 1.0 þ ð−0.1840α þ 0.5903Þρ1 þ ð0.1865α þ 0.0839Þρ21 = 1.092
av = 1.0 þ ð0.7415α − 1.0325Þρ1 þ ð−0.7969α − 0.0928Þρ21 = 0.840
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

bv = 1.0 þ ð−0.4414α þ 1.078Þρ1 þ ð0.4175α þ 0.5707Þρ21 = 1.178

and the moments μ^S and σ

^2S will be determined from the expressions reported in Table 8-9. For this
purpose, we assume that the marginal distribution of the Poudre River annual streamflow series is
log–normal, and applying the method of moments, the parameters are estimated as
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2 ffi
^x
σ ^2y
σ
^y = ln 2 þ 1 = 0.347 and μ
σ ^y = lnð^
μx Þ − = 5.64
^x
μ 2

in which μ ^ x = 0.36. Then the mean μ

^x = 299 and Cv ^S and the variance σ
^2S of the individual deficits are
computed as (Table 8-9)
  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Δ 
1 2
^s = exp μ
μ ^y þ σ ^ 1 − α expð^ σ2y Þ − 1 − = 72.401
2 y p0
 
σ2y Þ
expð^ Δ2
^s = expð2^
σ 2
μy þ σ^y Þ
2
· Ψ − 2 = 2,071.44
p0 p0
where
   
1 ^ xÞ
lnð1 − αCv 1 ^ xÞ
lnð1 − αCv
p0 = Φ σ ^ þ = 0.569, Δ = Φ − σ^ þ = 0.431, and
2 y ^y
σ 2 y ^y
σ
 
3 ^ xÞ
lnð1 − αCv
Ψ=Φ − σ ^ þ = 0.301:
2 y ^y
σ

For ℓ = 3, Equations (8-90) and (8-91) yield, respectively,

μ ^S am ℓbm = 217.89 and σ

^D = μ ^2D = σ
^2S av ℓbv = 6,348.40,

and the parameters p and q of the beta distribution are obtained from Equations (8-95) and (8-96),
respectively, as
 2  
^
μ ^D
μ ^
μ ^ ðℓx − μ
μ ^ Þ
p= D 1− − D = 5.419 and q = D o2 D − ð1 þ pÞ = 16.89
^D
σ ℓxo ℓxo ^D
σ

Therefore, the probability of the drought E = {D > d0 and L = 3} is obtained from Equation
(8-100) as

Z
3x0
1 ðzÞp−1 ð3x0 − zÞq−1
P½D > do , L = 3 = f L ð3Þ dz
Bðp,qÞ ð3x0 Þpþq−1
d0

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 319

From Example 8-8, f L ð3Þ = ð1 − 0.41Þ3−1 0.41 = 0.143 and the integral can be computed using,
for instance, the Matlab function betacdf as

Z
3x0
1 ðzÞp−1 ð3x0 − zÞq−1
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

dz = 1 − betacdf ½d0 ∕ð3 x0 Þ, p, q = 0.155

Bð p,qÞ ð3x0 Þpþq−1
d0

It follows that P½D > d0 , L = 3 = 0.0222, and the corresponding return period is

L þ Ln
T= = 194 years
P½D > d o , L = 3

The foregoing procedure can be repeated for different values of d0 and Land for different types
of drought events, such as those indicated in Equations (8-100)–(8-103). For instance, Figure 8-13
shows the return periods of droughts defined as E = fD > d0 and L = ℓ0 ðℓ0 = 1, 2, : : : Þg, which
were obtained from Equation (8-105) for various values of the deficit coefficient δ = d0 ∕x0 and
threshold x0 = μx .

8.7 REGIONAL ANALYSIS OF DROUGHTS

Drought identification and characterization in a region is an important component of drought

mitigation and management studies. While drought analysis at individual sites provides useful
information on drought occurrences in a watershed, regional analysis enables identifying droughts
that affect a region, considering the duration, magnitude (or intensity), and areal extent of the
drought based on data that are available at several sites of the study region (e.g., Tase 1976, Santos
1983, Rossi et al. 1992, Shin and Salas 2000, Rossi and Cancelliere 2003). As presented in this chapter
regional analysis of drought differs from regional analysis of low flows. In regional analysis of low
flows, the objective is determining low-flow characteristics at sites with short records or at ungauged

10000
Return period (years)

1000

100
Symbol δ =do/Xo
0
0,25
10 0,75
1,25
2
1
1 2 3 4 5 6 7 8 9 10
Drought length (years)
Figure 8-13. Return period of drought events defined by E = [D > d0 and L = ℓ0 (ℓ0 = 1, 2, : : : )] for
the Poudre River at Mouth of the Canyon annual flows obtained from Equation (8-105) for various
values of the deficit coefficient δ and the threshold x0 = μx.
Source: Cancelliere and Salas (2010).

Statistical Analysis of Hydrologic Variables

 320 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

sites, while in regional analysis of drought one would like to identify and characterize the variability
of drought as it varies through the region. Presenting regional drought analysis in detail would be too
lengthy for this chapter. We simply outline the underlying definitions following the method
originally developed by Tase (1976).
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Let us consider a region having a total area A and m precipitation stations located within and
nearby the area of interest. Station k has an area of influence denoted by s(k) (e.g., defined by the
Tiessen polygon), which may be expressed as a fraction a(k) of the total area as

sðkÞ
aðkÞ = Pm (8-106)
j = 1 sðjÞ

in which A = s(1) + s(2) + : : : + s(m). In addition, we will denote by X t ðkÞ the precipitation at
station k at time t, and x0 ðkÞ the threshold level for station k. Then the following indicator variable
can be defined for station k:

I t ðkÞ = 0 if X t ðkÞ ≥ x0 ðkÞ

I t ðkÞ = 1 if X t ðkÞ < x0 ðkÞ (8-107)

which indicates that if a deficit occurs at time t at site k, i.e., X t ðkÞ < x0 ðkÞ, then I t ðkÞ = 1. And the
fraction of the area affected by deficit in a time t is denoted as areal coverage of deficit and is given by
the index
X
m
Adt = aðkÞI t ðkÞ (8-108)
k=1

Then a regional deficit occurs at time t if the areal coverage of deficit Ad t exceeds a given
threshold ac (Tase 1976). Accordingly, the areal (regional) deficit Dt can be determined by

X
m
Dt = aðkÞ I t ðkÞ½x0 ðkÞ − X t ðkÞ if Adt ≥ ac (8-109)
k=1

The index Ad t is a measure of the area affected by deficit, expressed as a fraction of the total area
that ranges between 0 and 1, while the index Dt provides a measure of the total amount of deficit in
the area (region). It is basically the sum of the deficits at each site weighted by the corresponding
influence areas (computed for example by the Thiessen polygons).
Regional drought is defined as a consecutive sequence of regional deficits preceded and
succeeded by surpluses. Then the duration of the regional drought can be characterized by

L = tf − ti þ 1 (8-110)
where ti and tf are such that Dti −1 = 0, Dt > 0 for t i ≤ t ≤ t f , and Dtf þ1 = 0. In addition, regional
drought magnitude can be determined as
X
tf
DM = Dt (8-111)
t = ti

and the regional drought intensity by

DM
DI = (8-112)
L

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 321

Furthermore, the mean areal coverage of drought can be computed by

1X
tf
AD = Adt (8-113)
L t=t
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Once the time series of regional deficit Dt and the threshold ac have been determined, one can
apply the drought analysis methods described previously for determining the statistical properties of
droughts, such as expected values, variances, distribution, and return period of drought events. Note
that the foregoing definitions of regional drought have been illustrated using precipitation as the key
hydrological input representing water supply in the study region, and several developments and
applications thereof are available (e.g., Tase 1976, Santos 1983, Rossi et al. 1992, Rossi and
Cancelliere 2003). However, the application of regional drought as described previously for
hydrological inputs other than precipitation have not been made.

8.8 EFFECTS OF HYDRAULIC STRUCTURES ON LOW FLOWS

Human intervention directly and indirectly alters the natural flow regime of river systems (Wang
and Cai 2009). Indirect effects occur by changes in natural phenomena such as climate and changes
in vegetation, land use, soil, and topography. River works such as dams, river diversions, major
mining explorations and operational activities, and groundwater pumping for many water supply
purposes directly alter the natural flow regime of streams and river systems. For several decades,
many studies have been conducted to identify the effects of water development on the natural
streamflows. For example, Dynesius and Nilsson (1994) report that 77% of the total flow of the
largest rivers in the United States, Canada, Europe, and the former Soviet Union was affected by
dams, reservoirs, aqueducts, and irrigation. They estimate that hydrologic alterations have occurred
in 60% of rivers all over the world, including more than 85% of US rivers and 60% to 65% of
European rivers. In many of those rivers significant changes of the natural flow regime have been
observed, such as decreasing flows during the wet season while increasing flows during the dry
season. While control of natural flow regime has proper functions for regulating water supply and
flood control, it conflicts with other functions of the streams and rivers such as those related to their
natural environment and ecology.
The variability of streamflows through time at a given cross-section of a river is called flow
regime. Natural flow regime refers to the condition before the development of rivers, without
controlling the flow regime by hydraulic structures. Rivers have their own natural flow regime
characteristics based on the climate, morphology of the river basin, and land use, and these flow
regime characteristics maintain the health of the river environment. However, artificial river works
such as dams changes the natural flow regime and consequently their various statistical
characteristics.
Flow duration curves (FDC) and the Indicator of Hydrologic Alteration (IHA) have been useful
for identifying the degree of alteration in flow regime resulting from the effects of hydraulic
structures. Natural flows vary over the year, and FDC is one of the methods used to identify the flow
variability from low flows to floods. An FDC for a given site of a river is a plot of flow against the
percentage of time the flow has been equaled or exceeded. Flow records of 10 years or more
adequately define the shape of the curve (Hadley et al. 1987, Maheshwari et al. 1995, Subramanya
1994). In low-flow studies, the main interest is the low-flow section of a FDC. The section may be
determined as part of the curve with flows below median flow that corresponds to the discharge
equal to or exceeding 50% (Smakhtin 2001). The Indicators of Hydrologic Alteration (IHA),

Statistical Analysis of Hydrologic Variables

 322 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

developed by The Nature Conservancy (Richter et al. 1997), are based on 32 indexes that define a
series of biologically relevant hydrologic attributes and then quantify hydrologic alterations
associated with anticipated perturbations arising from the effect of hydraulic structures by
comparing the hydrologic regimes from pre-impact and post-impact time frames. Poff et al. (1997)
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

and Richter et al. (1997) classify flow regime in terms of magnitude of flow, frequency of its
occurrence, duration, timing of a specific flow, and rate of change. Zhang et al. (2015a, b) assess
temporal and spatial alteration of flow regimes in some reservoir-regulated rivers in China using
IHA. During the post-regulation period, they show a decrease in the high-flow magnitudes and an
increase in the low flows (relative to those during the pre-regulation period). Table 8-14 shows
several flow regime characteristics related to low flows.
Much has been written about the impacts on low flows by hydraulic structure such as dams
(e.g., Homa et al. 2013). The typical differences in the FDCs are that the flows, at the low-flow section
of the FDC during the post-impact period, become greater than those for the pre-impact period.
However, the opposite occurs for the flows at the high-flow section of the FDC (e.g., Gippel and
Stewardson 1995, Gustard et al. 1989, Hadley et al. 1987, McMahon and Finlayson 2003). Also, the
annual minimum flows affected by regulation by dams generally increase compared with the pre-
regulation period (Richter et al. 1997, Magilligan and Nislow 2005).
Figure 8-14 shows the Han River in Korea where several dams have been constructed. The
multipurpose dams impound water during the rainy season (June–September) and then the stored
waters in the reservoirs are released to be used for water supply in urban areas during the low-flow
period (November–March). Likewise, the Geum River has two multipurpose dams that also store the
rain waters that are then released to be used downstream during low-flow periods. For illustration,
Figure 8-15 shows the comparison of the average daily flows obtained using the flow data for the pre-
impact period (1921–1940) and the flows for the post-impact period (1988–2007). Clearly the flows
during the low-flow period are generally larger in the post-impact period than in the pre-impact
period. Also, Figure 8-16 shows the comparison of the FDCs obtained for the two periods. As
expected, the flows in the lower portion of the curve are bigger for the post-impact period than for
the pre-impact period. This effect also becomes evident for the 1 day and 30 day minimum flows as
shown in Figures 8-17 and 8-18, respectively.

Table 8-14. Low-Flow Hydrologic Parameters Used in the Indicators of Hydrologic Alteration (IHA).

Regime Low-flow hydrologic

IHA statistics group characteristics parameters

Group 2: Magnitude and duration Magnitude Annual minima. 1 day means

of annual extreme discharge Duration Annual minima. 3 day means
conditions Annual minima. 7 day means
Annual minima. 30 day means
Annual minima. 90 day means
Group 3: Timing of annual extreme Timing Julian date of each annual 1 day
discharge conditions minimum discharge
Group 4: Frequency and duration Frequency No. of low pulses each year
of high and low pulses Duration Mean duration of low pulses within
each year
Group 5: Rate and frequency of Rate of change Means of all negative differences
water condition changes Frequency between consecutive daily values
No. of falls
Source: Modified from Richter et al. (1997).

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 323
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-14. The Han and Geum river basins in Korea.

8.9 CLOSING REMARKS

In the statistical analysis of low flows and droughts discussed in this chapter we made the
implicit assumption that the underlying data and models thereof are stationary (i.e., the model
statistics and parameters remain constant through time). However, this assumption may not
be valid in some cases. The basic data used for deriving low flows and droughts may be
nonstationary for several reasons such as human intervention in the landscape and water cycle,
natural events such as volcanic explosions or forest fires, the effect of low-frequency components
of oceanic–atmospheric phenomena, and global warming (Salas et al. 2012). Numerous studies
and reports assess changes (e.g., increasing or decreasing trends) in low-flow conditions in
individual rivers, regions, and worldwide (e.g., Lins and Slack 1999, Douglas et al. 2000, Yue et al.
2003, Svensson et al. 2005, Hannaford and Marsh 2006), and the results of either significant
increases, decreases, or no changes vary depending on the country and region of study and the
tests utilized.

Statistical Analysis of Hydrologic Variables

 324 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-15. The daily average flows for the pre- and post-regulation periods in the Han river at Koan
station, Korea.

Figure 8-16. Daily flow duration curve in the Han River at Koan, Korea.

Figure 8-17. 1 day minimum flows for the Han River at the Koan station during the pre-impact and
post-impact periods.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 325
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Figure 8-18. 30 day minimum flows for the Han River at the Koan station during the pre-impact and
post-impact periods.

In most if not all cases, the justification of such studies has been the concern for the possible
effects of anthropogenic global climate change on low flows and droughts. However, determining the
cause of any change in low flows is not an easy task. The problem is even more difficult when trying
to assess with confidence the impacts of future climate change because of the uncertainties in climate
models and the complexity of the relationships of various components of the hydrologic cycle
(Hannaford and Marsh 2006). In some cases, such as in the Midwest in the United States, changes in
low flows may be related to similar changes in precipitation (e.g., Douglas et al. 2000). However, it
has been argued that in some places, such as Iowa (where agricultural lands comprise more than 70%
of the land area), increases in precipitation alone do not explain the observed increases in baseflow and
most likely result from improvements in land management practices (Schilling and Libra 2003). In
addition, studies in the United Kingdom based on a benchmark network of 120 relatively undisturbed
catchments indicate no compelling evidence of trends in low flows (Hannaford and Marsh 2006).
Furthermore, studies based on 21 daily river flows in many locations worldwide suggest increases in
low flows for 10 of the stations (Svensson et al. 2005), and the authors argue that such increases are not
consistent with an intensification of the hydrological cycle in a warming climate that would result in
more severe droughts and that the cause of such increases may be the increasing number of reservoirs
in the basins. The authors conclude that such “modifications to the river flow regime would likely
obscure any recent alteration in the hydrological cycle due to climate change.”
Despite the limitations and uncertainties involved in projecting the global climate into the
future, many studies aim to determine the impacts of climate change on low flows and droughts at
the continental, country, regional, and basin scales (e.g., Ryu et al. 2011, van Lanen et al. 2007). In
addition to the difficulty of quantifying the effects of land-use changes on low flows and droughts,
distinguishing between the effects of low-frequency components of the oceanic–atmospheric system
and those from global warming is not feasible at present. Studies addressing the effect of large-scale
low-frequency components of the climate system on the variability of runoff and droughts include,
for example, Ozger et al. (2009) and Gudmundsson et al. (2011). Furthermore, some studies
(e.g., Dai et al. 2004) have attributed the severity of droughts in the past decades to global warming.
However, Sheffield et al. (2012b) argue that several studies and reports, including the IV Assessment
Report of the Intergovernmental Panel on Climate Change (IPCC Climate Change 2007), have
overestimated the increase of global drought because the computations were made using the Palmer
Drought Severity Index (PDSI) based on a temperature-driven potential evapotranspiration (PE).
However, the drought estimates using the same PDSI, but where PE is determined based on a more
realistic model involving energy, humidity, and wind speed, suggested that minor change in drought
occurred in the past decades (Sheffield et al. 2012b).

Statistical Analysis of Hydrologic Variables

 326 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

As reported in Section 8.6, a problem in characterizing multiyear droughts is the limited sample
size commonly available from historical instrumental records. For example, based on the 1911–1998
record of the Meuse River, concluding whether drought has become more severe or frequent is not
possible (van Lanen et al. 2007). One technique that has been helpful to extend streamflow records is
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

based on tree ring records, which enable streamflow reconstruction back in time (e.g., Loaiciga et al.
1993, Meko et al. 2001, Woodhouse 2001, Biondi and Strachan 2011). Often the reconstructed flow
records show that periods of low flows and droughts more severe than those observed from
instrumental records have occurred in the past (e.g., Meko et al. 1995, Meko et al. 2001, Gonzalez
and Valdes 2003, Gedalof et al. 2004, Tarawneh and Salas 2008). A stochastic nonparametric
approach for streamflow generation combining observational and paleoreconstructed data has also
been suggested (Prairie et al. 2008). Furthermore, alternative stochastic models capable of generating
nonstationary-like changes in streamflow sequences have been developed (e.g., Montanari et al.
2000, Koutsoyiannis 2002, Sveinsson et al. 2003). Data generated based on these models can be
utilized for drought analysis using some of the empirical methods included in this chapter. Lastly,
recent advances have been made in developing statistical analysis of extreme events under
nonstationary conditions (e.g., Villarini et al. 2009, Vogel et al. 2011, Salas and Obeysekera
2014, Lopez and Frances 2013). Although the main applications of those techniques focus on
extreme flood and sea level events, some advances have been made to analyze low flows and droughts
as well (e.g., Garcia Galiano et al. 2011, Arpita and Mujumdar 2015, Kwon and Lall 2016, Cancelliere
and Bonaccorso 2016).

References
Akyuz, D. E., M. Bayazit, and B. Onoz. 2012. “Markov chain models for hydrological drought characteristics.”
J. Hydrometeorol. 13 (1): 298–309.
Alley, W. M. 1984. “The Palmer drought severity index: Limitations and assumptions.” J. Clim. Appl. Meteorol.
23 (7): 1100–1109.
Arpita, M., and P. P. Mujumdar. 2015. “Return levels of hydrologic droughts under climate change.” Adv. Water
Resour. 75: 67–79.
ASCE. 1980. “Characteristics of low flows.” J. Hydraul. Div. 106 (5): 717–731.
Biondi, F., T. J. Kozubowskib, and A. K. Panorska. 2005. “A new model for quantifying climate episodes.” Int. J.
Climatol. 25 (9): 1253–1264.
Biondi, F., and S. Strachan. 2011. “Dendrohydrology in 2050: Challenges and opportunities.” In Toward a
sustainable water future, visions for 2050, edited by W. M. Grayman, D. P. Loucks, and L. Saito. Reston, VA:
ASCE.
Bobee, B., and R. Robitaille. 1975. “Correction of bias in the estimation of the coefficient of skewness.” Water
Resour. Res. 11 (6): 851–854.
Bonaccorso, B. 2002. “Stochastic characterization of drought events.” Ph.D. thesis, Univ. of Catania.
Bonaccorso, B., A. Cancelliere, and G. Rossi. 2003. “An analytical formulation of return period of drought
severity.” Stochastic Environ. Res. Risk Assess. 17 (3): 157–174.
Bond, N. R., P. S. Lake, and A. H. Arthington. 2008. “The impact of drought on freshwater ecosystem: An
Australian perspective.” Hydrobiologia 600 (1): 3–16.
Box, G. E. P., and D. R. Cox. 1964. “An analysis of transformations.” J. R. Stat. Soc., Ser. B 26 (2): 211–252.
Box, G. E. P., and G. M. Jenkins. 1976. Time series analysis forecasting and control. San Francisco: Holden-Day.
Cancelliere, A. 2008. “Stochastic characterization of droughts in stationary and periodic series.” Ph.D.
dissertation, Colorado State Univ.
Cancelliere, A., and B. Bonaccorso. 2016. “A non-stationary analytical framework for the probabilistic
characterization of drought events.” In Proc., EWRI World Environmental and Water Resources Congress.
Reston, VA: ASCE.
Cancelliere, A., B. Bonaccorso, G. Rossi, and J. D. Salas. 2003. “On the probabilistic characterization of drought
events.” In Proc., 23rd AGU Hydrology Days. Fort Collins, CO: Colorado State Univ.
Cancelliere, A., and J. D. Salas. 2002. “Characterizing the recurrence of hydrologic droughts.” In Proc., AGU Fall
Meeting. San Francisco.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 327

Cancelliere, A., and J. D. Salas. 2004. “Drought length properties in periodic-stochastic hydrologic data.” Water
Resour. Res. 40 (2).
Cancelliere, A., and J. D. Salas. 2010. “Drought probabilities and return period for annual streamflows series.”
J. Hydrol. 391 (1–2): 77–89.
Chang, T., M. Kavvas, and J. Delleur. 1984a. “Daily precipitation modeling by discrete autoregressive moving
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

average processes.” Water Resour. Res. 20 (5): 565–580.

Chang, T., M. Kavvas, and J. Delleur. 1984b. “Modeling sequences of wet and dry days by binary discrete
autoregressive moving average processes.” J. Clim. Appl. Meteorol. 23 (9): 1367–1378.
Chebaane, M., J. Salas, and D. Boes. 1995. “Product periodic autoregressive processes for modeling intermittent
monthly streamflows.” Water Resour. Res. 31 (6): 1513–1518.
Chung, C. H., and J. D. Salas. 2000. “Return period and risk of droughts for dependent hydrologic processes.”
J. Hydrol. Eng. 5 (3): 259–268.
CONAGUA (Comision Nacional del Agua). 2016. “BANDAS.” Accessed July 9, 2016. http://www.conagua.-
gob.mx/CONAGUA07/Contenido/Documentos/Portada%20BANDAS.htm.
Condie, R., and G. A. Nix. 1975. “Modelling of low flow frequency distributions and parameter estimation.” In
Proc., Int. Water Resources Symp.on Water for Arid Lands. Madison, WI: International Water Resources
Association.
Cramer, H., and M. Leadbetter. 1967. Stationary and related stochastic processes. New York: Wiley.
Dai, A., K. E. Trenberth, and T. A. Qian. 2004. “A global data set of Palmer Drought Severity Index for 1870–2002:
Relationship with soil moisture and effects of surface warming.” J. Hydrometeorol. 5 (6): 1117–1130.
Douglas, E. M., R. M. Vogel, and C. N. Kroll. 2000. “Trends in floods and low flows in the United States: Impact
of spatial correlation.” J. Hydrol. 240 (1–2): 90–105.
Downer, R., M. Siddiqui, and V. Yevjevich. 1967. “Application of runs to hydrological droughts.” In Proc., Int.
Hydrology Symp. Fort Collins, CO: Colorado State Univ.
Dracup, J. A., K. L. Lee, and E. G. Paulson Jr. 1980. “On the statistical characteristics of drought events.” Water
Resour. Res. 16 (2): 289–296.
Dynesius, M., and C. Nilsson. 1994. “Fragmentation and flow regulation of river systems in the northern third of
the world.” Science 266 (5186): 753–762.
Fernández, B., and J. D. Salas. 1990. “Gamma-autoregressive models for streamflow simulation.” J. Hydrol. Eng.
116 (11): 1403–1414.
Fernández, B., and J. D. Salas. 1999a. “Return period and risk of hydrologic events. 1: Mathematical
formulation.” J. Hydrol. Eng. 4 (4): 297–307.
Fernández, B., and J. D. Salas. 1999b. “Return period and risk of hydrologic events. 2: Applications.” J. Hydrol.
Eng. 4 (4): 308–316.
Frick, D. M., D. Bode, and J. D. Salas. 1990. “Effect of drought on urban water supplies. I: Drought analysis.”
J. Hydrol. Eng. 116 (6): 733–753.
Garcia Galiano, S. G., J. D. Giraldo, M. A. Urrea, A. Merida, and C. N. Tetay. 2011. “Assessing drought hazard
under non-stationary conditions on Southeast Spain.” In Proc., Symp. Risk in Water Resources Management
H03, 85–91. Melbourne, Australia.
Gaver, D. P., and P. A. W. Lewis. 1980. “First order autoregressive gamma sequences and point process.” Adv.
Appl. Probab. 12 (3): 727–745.
Gedalof, Z., D. L. Peterson, and N. J. Mantua. 2004. “Columbia river flow and drought since 1750.” J. Am. Water
Resour. Assoc. 40 (6): 1579–1592.
Gonzalez, J., and J. Valdes. 2003. “Bivariate drought recurrence analysis using tree ring reconstructions.”
J. Hydrol. Eng. 8 (5): 247–258.
Gudmundsson, L., L. M. Tallaksen, K. Stahl, and A. K. Fleig. 2011. “Low-frequency variability of European
runoff.” Hydrol. Earth Syst. Sci. 15 (9): 2853–2869.
Gumbel, E. J. 1958. Statistics of extremes. New York: Columbia University Press.
Gustard, A., L. A. Roald, S. Demuth, H. S. Lumadjeng, and R. Gross. 1989. “Flow regimes from experimental and
network data (FREND).” In Hydrological studies, Vol. I. Wallingford, UK: Institute of Hydrology.
Guttman, N. B. 1998. “Comparing the Palmer drought index and the standardized precipitation index.” J. Am.
Water Resour. Assoc. 34 (1): 113–121.
Guven, O. 1983. “A simplified semiempirical approach to probabilities of extreme hydrologic droughts.” Water
Resour. Res. 19 (2): 441–453.

Statistical Analysis of Hydrologic Variables

 328 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Haan, C. T. 1992. Statistical methods in hydrology. Ames, IA: Iowa State Press.
Hadley, R. F., M. R. Kalinger, A. W. Burns, and T. R. Eschner. 1987. “Water development and associated
hydrologic changes in the Platte River, Nebraska, U.S.A.” Regul. Rivers: Res. Manage. 1 (4): 331–341.
Hannaford, J., and T. Marsh. 2006. “An assessment of trends in UK runoff and low flows using a network of
undisturbed catchments.” Int. J. Climatol. 26 (9): 1237–1253.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Hayes, D. C. 1992. Low flow characteristics of streams in Virginia: Water supply paper 2374. Denver: US
Geological Survey, Federal Center.
Heim, R. 2002. “A review of twentieth-century drought indices used in the United States.” Bull. Am. Meteorol.
Soc. 83 (8): 1149–1166.
Hirsch, R. M., D. R. Helsel, T. A. Cohn, and E. J. Gilroy, 1993. “Statistical analysis of hydrologic data (chapter 17).”
In Handbook of hydrology, edited by D. R. Maidment. New York: McGraw-Hill.
Homa, E. S., C. Brown, K. McGarigal, B. W. Compton, and S. D. Jackson. 2013. “Estimating hydrologic
alteration from basin characteristics in Massachusetts.” J. Hydrol. 503: 196–208.
Hosking, J. R. M., and J. R. Wallis. 1997. Regional frequency analysis. Cambridge, UK: Cambridge University
Press.
IFC (Instream Flow Council). (2002). “Instream flows for riverine resource stewardship.” Accessed November 8,
2018. https://www.instreamflowcouncil.org/.
Institute of Hydrology. 1980. Low-flow studies. Rep. No. 1. Wallingford, UK: Institute of Hydrology.
IPCC (Intergovernmental Panel on Climate Change) Climate Change. 2007. The physical science basis:
Contribution of working group i to the fourth assessment report of the intergovernmental panel on climate
change. Edited by S. Solomon, et al. Cambridge, MA: Cambridge University Press.
Jackson, B. B. 1975. “Markov mixture models for drought lengths.” Water Resour. Res. 11 (1): 64–74.
Jacobs, P. A., and P. A. W. Lewis. 1977. “A mixed autoregressive-moving average exponential sequence and
point process (EARMA1, 1).” Adv. Appl. Probab. 9 (1): 87–104.
Jacobs, P. A., and P. A. W. Lewis. 1978. “Discrete time series generated by mixtures. I: Correlational and runs
properties.” J. R. Stat. Soc. Oxford 40 (1): 94–105.
Jennings, M. E., and M. A. Benson. 1969. “Frequency curves for annual flood series with some zero events or
incomplete data.” Water Resour. Res. 5 (1): 276–280.
Johnson, N., S. Kotz, and N. Balakrishnan. 1994. Continuous univariate distributions. Vol. 1. New York: Wiley.
Kirby, W. 1972. “Computer-oriented Wilson-Hilferty transformation that preserves the first three moments and
the lower bound of the Pearson type 3 distribution.” Water Resour. Res. 8 (5): 1251–1254.
Kite, G. W. 1988. Frequency and risk analyses in hydrology. Littleton, CO: Water Resources Publishing.
Koutsoyiannis, D. 2002. “The Hurst phenomenon and fractional Gaussian noise made easy.” Hydrol. Sci. J.
47 (4): 573–595.
Kroll, C. N., J. G. Luz, T. B. Allen, and R. M. Vogel. 2004. “Developing a watershed characteristics database to
improve low streamflow prediction.” J. Hydrol. Eng. 9 (2): 116–125.
Kroll, C. N., and J. R. Stedinger. 1998. “Generalized least squares regression procedures revisited.” Water Resour.
Res. 34 (1): 121–128.
Kroll, C. N., and R. M. Vogel. 2002. “The probability distribution of low streamflow series in the United States.”
J. Hydrol. Eng. 7 (2): 137–146.
Kwon, H. H., and U. Lall. 2016. “A copula-based nonstationary frequency analysis for the 2012–2015 drought in
California.” Water Resour. Res. 52 (7): 5662–5675.
Laaha, G., and G. Blöschl. 2006a. “A comparison of low flow regionalisation methods—Catchment grouping.”
J. Hydrol. 323 (1–4): 193–214.
Laaha, G., and G. Blöschl. 2006b. “Seasonality indices for regionalizing low flows.” Hydrol. Processes 20 (18):
3851–3878.
Laaha, G., and G. Blöschl. 2007. “A national low flow estimation procedure for Austria.” Hydrol. Sci. J. 52 (4):
625–644.
Lawrance, A. J. 1982. “The innovation distribution of a gamma distributed autoregressive process.” Scand. J.
Stat. 9 (4): 234–236.
Lawrance, A. J., and P. A. W. Lewis. 1981. “A new autoregressive time series model in exponential variables
[NEAR(1)].” Adv. Appl. Probab. 13 (4): 826–845.
Lins, H. F., and J. R. Slack. 1999. “Streamflow trends in the United States.” Geophys. Res. Lett. 26 (2):
227–230.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 329

Llamas, J., and M. Siddiqui. 1969. Runs of precipitation series: Hydrology paper 33. Fort Collins, CO: Colorado
State Univ.
Lloyd, E. H. 1970. “Return period in the presence of persistence.” J. Hydrol. 10 (3): 202–215.
Loaiciga, H. A. 2005. “On the probability of droughts: The compound renewal model.” Water Resour. Res.
41 (1): W01009.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Loaiciga, H. A., L. Haston, and J. Michaelsen. 1993. “Dendrohydrology and long-term hydrological phenome-
na.” Rev. Geophys. 31 (2): 151–171.
Loaiciga, H. A., and R. B. Leipnik. 1996. “Stochastic renewal model of low-flow streamflow sequences.”
Stochastic Hydrol. Hydraul. 10 (1): 65–85.
Loaiciga, H. A., and M. A. Mariño. 1991. “Recurrence interval of geophysical events.” J. Water Resour. Plann.
Manage. 117 (3): 367–382.
Lopez, J., and F. Frances. 2013. “Non-stationary flood frequency analysis in continental Spanish rivers, using
climate and reservoir indices as external covariates.” Hydrol. Earth Sci. 17 (8): 3189–3203.
Loucks, D. P., J. R. Stedinger, and D. A. Haith. 1981. Water resource systems planning and analysis. Englewood
Cliffs, NJ: Prentice Hall.
Magilligan, F. J., and K. H. Nislow. 2005. “Changes in hydrologic regime by dam.” Geomorphology 71 (1–2):
61–78.
Maheshwari, B. L., K. F. Walker, and T. A. McMahon. 1995. “Effects of regulation on the flow regime of the river
Murray, Australia.” River Res. Applic. 10 (1): 15–38.
Matalas, N. C. 1966. “Time series analysis.” Water Resour. Res. 3 (3): 817–829.
McMahon, T. A., and B. L. Finlayson. 2003. “Droughts and anti-droughts: The hydrology of Australian rivers.”
Freshwater Biol. 48 (7): 1147–1160.
Meko, D., C. W. Stockton, and W. R. Boggess. 1995. “The tree-ring record of severe sustained drought.” Water
Resour. Bull. 31 (5): 789–801.
Meko, D., M. Therrell, C. Baisan, and M. Hughes. 2001. “Sacramento river flow reconstructed to A.D. 869 from
tree rings.” J. Am. Water Resour. Assoc. 37 (4): 1029–1039.
Mishra, A. K., and V. P. Singh. 2010. “A review of drought concepts.” J. Hydrol. 391 (1–2): 202–216.
Mishra, A. K., and V. P. Singh. 2011. “Drought modeling—A review.” J. Hydrol. 403 (1–2): 157–175.
Mishra, A. K., V. P. Singh, and V. R. Desai. 2009. “Drought characterization: A probabilistic approach.”
Stochastic Environ. Res. Risk Assess. 23 (1): 41–55.
Montanari, A., R. Rosso, and M. Taqqu. 2000. “A seasonal fractional ARIMA model applied to the Nile River
monthly flows at Aswan.” Water Resour. Res. 36 (5): 1249–1259.
Nathan, R. J., and T. A. McMahon. 1990. “Identification of homogeneous regions for the purpose of
regionalization.” J. Hydrol. 121 (1–4): 217–238.
Nathan, R. J., and T. A. McMahon. 1992. “Estimating low flow characteristics at ungauged catchments.” Water
Resour. Manage. 6 (2): 85–100.
NERC (National Environmental Research Council). 1975. Flood studies report. Vol. 1. London: NERC.
O’Connell, P. E. 1977. “ARIMA models in synthetic hydrology.” In Mathematical models for surface water
hydrology, edited by T. A. Ciriani, V. Maione, and J. R. Wallis, 51–68. New York: Wiley.
Onöz, B., and M. Bayazit. 2001. “Power distribution for low streamflow.” J. Hydrol. Eng. 6 (5): 429–435.
Ozger, M., A. K. Mishra, and V. P. Singh. 2009. “Low frequency drought variability associated with climate
indices.” J. Hydrol. 364 (1–2): 152–162.
Panu, U. S., and T. C. Sharma. 2002. “Challenges in drought research: Some perspectives and future directions.”
J. Hydrol. Sci. 47: S19–S30.
Paoli, C., J. Bolzicco, and P. Cacik. 1994. “Analysis of low flows in the Paraná River.” [In Spanish.] In Vol. 4 of
XVI IAHR Congreso Latinoamericano de Hidraulica, 59–70. Santiago, Chile.
Pearson, C. P. 1995. “Regional frequency analysis of low streamflows in New Zealand rivers.” J. Hydrol. 33 (2):
94–122.
Poff, N. L., et al. 1997. “The natural flow regime: A paradigm for river conservation and restoration.” BioScience
47 (11): 769–784.
Prairie, J., K. Nowak, B. Rajagopalan, and U. Lall. 2008. “A stochastic nonparametric approach for streamflow
generation combining observational and paleoreconstructed data.” Water Resour. Res. 44 (6): W06423.
Raynal-Villasenor, J. A. 2013. “Moment estimators of the GEV distribution for the minima.” J. Appl. Water Sci.
3 (1): 13–18.

Statistical Analysis of Hydrologic Variables

 330 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Reilly, C. F., and C. N. Kroll. 2003. “Estimation of low streamflow statistics using baseflow correlation.” Water
Resour. Res. 39 (9): 1236.
Richter, B. D., J. V. Baumgatner, R. Wigington, and D. P. Braun. 1997. “How much water does a river need?”
Freshwater Biol. 37 (1): 231–249.
Riggs, H. C. 1980. “Characteristics of low flows.” J. Hydraul. Div. 106 (5): 717–731.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Rossi, G., M. Benedini, G. Tsakiris, and S. Giakoumakis. 1992. “On regional drought estimation and analysis.”
Water Resour. Manage. 6 (4): 249–277.
Rossi, G., and A. Cancelliere. 2003. “At site and regional identification by REDIM model.” In Tools for drought
mitigation in Mediterranean regions, edited by G. Rossi, A. Cancelliere, T. Oweis, L. Pereira, M. Shatanawi,
and A. Zairi, 37–54. Dordrecht, Netherlands: Kluwer Academic.
Ryu, J. H., J. H. Lee, S. Jeong, S. K. Park, and K. Han. 2011. “The impacts of climate change on local hydrology
and low flow frequency in the Geum River basin, Korea.” Hydrol. Processes 25 (22): 3437–3447.
Salas, J. D. 1993. “Analysis and modeling of hydrologic time series.” In Handbook of hydrology, edited by
D. R. Maidment. New York: McGraw-Hill.
Salas, J. D., C. Chung, and B. Fernández. 2001. “Relating autocorrelations and crossing rates of continuous- and
discrete-valued hydrologic processes.” J. Hydrol. Eng. 6 (2): 109–118.
Salas, J. D., J. W. Delleur, V. M. Yevjevich, and W. L. Lane. 1980. Applied modeling of hydrologic time series.
Littleton, CO: Water Resources Publications.
Salas, J. D., C. Fu, A. Cancelliere, D. Dustin, D. Bode, A. Pineda, and E. Vincent. 2005. “Characterizing
the severity and risk of drought in the Poudre River, Colorado.” J. Water Resour. Plann. Manage. 131 (5):
383–393.
Salas, J. D., J. H. Heo, J. Obeysekera, M. Arabi, R. A. Smith, and G. Tabios. 2019. Probability and statistics in
water resources. Fort Collins, CO: Colorado State Univ.
Salas, J. D., and J. Obeysekera. 2014. “Revisiting the concepts of return period and risk for nonstationary
hydrologic events.” J. Hydrol. Eng. 19 (3): 554–568.
Salas, J. D., and R. A. Pielke. 2003. “Stochastic characteristics and modeling of hydroclimatic processes.”
Chap. 32 in Handbook of weather, climate, and water: Atmospheric chemistry, hydrology, and societal impacts,
edited by T. D. Potter and B. R. Colman. Hoboken, NJ: Wiley.
Salas, J. D., B. Rajagopalan, L. Saito, and C. Brown. 2012. “Climate nonstationarity and water resources
management: Introduction.” J. Water Resour. Plann. Manage. 138 (5): 385–388.
Saldarriaga, J., and V. Yevjevich. 1970. Applications of run lengths to hydrological series: Hydrology paper 40. Fort
Collins, CO: Colorado State Univ.
Santos, M. 1983. “Regional droughts: A stochastic characterization.” J. Hydrol. 66 (1–4): 183–211.
Schilling, K. E., and R. D. Libra. 2003. “Increased baseflow in Iowa over the second half of the 20th century.”
J. Am. Water Res. Assoc. 39 (4): 851–860.
Sen, Z. 1976. “Wet and dry periods of annual flow series.” J. Hydraul. Div. 102 (HY10): 1503–1514.
Sen, Z. 1977. “Run-sums of annual flow series.” J. Hydrol. 35 (3–4): 311–324.
Serinaldi, F., B. Bonaccorso, A. Cancelliere, and S. Grimaldi. 2009. “Probabilistic characterization of drought
properties through copulas.” Phys. Chem. Earth 34 (10–12): 596–605.
Sharma, T. 1995. “Estimation of drought severity on independent and dependent hydrologic series.” Water
Resour. Manage. 11 (1): 35–49.
Shaw, E. M. 1988. Hydrology in practice. London: Chapman and Hill.
Sheffield, J., B. Livneh, and E. F. Wood. 2012a. “Representation of terrestrial hydrology and large-scale drought
of the continental United States from the North American regional reanalysis.” J. Hydrometeorol. 13 (3):
856–876.
Sheffield, J., E. F. Wood, and M. L. Roderick. 2012b. “Little change in global drought over the past 60 years.”
Nature 491 (7424): 435–438.
Shiau, J. T. 2006. “Fitting drought duration and severity with two-dimensional copulas.” Water Resour. Manage.
20 (5): 795–815.
Shiau, J. T., and H. Shen. 2001. “Recurrence analysis of hydrologic droughts of differing severity.” J. Water
Resour. Plann. Manage. 127 (1): 30–40.
Shin, H. S., and J. D. Salas. 2000. “Regional drought analysis based on neural networks.” J. Hydraul. Eng. 5 (2):
145–155.
Smakhtin, V. U. 2001. “Low flow hydrology: A review.” J. Hydrol. 240 (3–4): 147–186.

Statistical Analysis of Hydrologic Variables

 LOW FLOWS AND DROUGHTS 331

Songbai, S., and V. Singh. 2010. “Meta-elliptical copulas for drought frequency analysis of periodic hydrologic
data.” Stochastic Environ. Res. Risk Assess. 24 (3): 425–444.
Stedinger, J. R., and G. D. Tasker. 1986. “Regional hydrologic analysis. 1: Ordinary, weighted, and generalized
least squares compared.” Water Resour. Res. 22 (5): 844.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Stedinger, J. R., and W. O. Thomas Jr. 1985. Low-flow frequency estimation using base-flow measurements.
Open-File Rep. No. 85-95. Reston, VA: USGS.
Stedinger, J. R., R. M. Vogel, and E. Foufoula-Georgiou. 1993. “Frequency analysis of extreme events.” Chap. 18
in Handbook of hydrology, edited by D. R. Maidment. New York: McGraw-Hill.
Subramanya, K. 1994. Engineering hydrology. 2nd ed. New Delhi: Tata McGraw-Hill.
Svensson, C., W. Z. Kundzewicz, and T. Maurer. 2005. “Trend detection in river flow series: 2. Flood and
low-flow index series.” Hydrol. Sci. J. 50 (5): 811–824.
Sveinsson, O., J. D. Salas, D. C. Boes, and R. A. Pielke Sr. 2003. “Modeling the dynamics of long term variability
of hydroclimatic processes.” J. Hydrometeorol. 4 (3): 489–505.
Svoboda, M. 2000. “An introduction to the drought monitor.” Drought Networks News 12 (1): 15–20.
Tabari, H., R. Zamani, H. Rahmati, and P. Willems. 2015. “Markov chains of different orders for streamflow
drought analysis.” Water Resour. Manage. 29 (9): 3441–3457.
Tallaksen, L. M. 1995. “A review of baseflow recession analysis.” J. Hydrol. 165 (1–4): 349–370.
Tallaksen, L. M., and H. A. J. van Lanen, eds. 2004. Hydrological drought— Processes and estimation methods
for streamflow and groundwater: Developments in water sciences 48, 580. Dordrecht, Netherlands: Elsevier.
Tarawneh, Z. S., and J. D. Salas. 2008. Extending the streamflows of the Colorado River using tree ring indices and
drought analysis. Fort Collins, CO: Colorado State Univ.
Tase, N. 1976. “Area-deficit-intensity characteristics of droughts.” Ph.D. thesis, Colorado State Univ. Fort
Collins, CO.
Tasker, G. D. 1987. “A comparison of methods for estimating low flow characteristics of streams.” Water Resour.
Bull. 23 (6): 1077–1083.
Thomas, D. M., and M. A. Benson. 1970. Generalization of streamflow characteristics from drainage-basin
characteristics: USGS Water Supply Paper 1975. Washington, DC: USGS.
USACE (US Army Corps of Engineers). 1964. Low flow frequency analysis. Technical Rep. No. N1. Tulsa, OK:
USACE.
USACE. 1975. “Hydrologic frequency analysis.” In Hydrologic engineering methods for water resources
development: IHD-3, Vol. 3. Davis, CA: USACE.
USGS (US Geological Survey). 1958. Handbook for hydrologists, Chapter VII. Washington, DC: US Dept. of
Interior, USGS.
USGS. 2010a. “GLSNet: Regional hydrologic regression and NETwork analysis using generalized least squares.”
Accessed September 10, 2010. https://water.usgs.gov/software/GLSNet.
USGS. 2010b. “Hydrologic unit maps.” Accessed September 10, 2010. https://water.usgs.gov/GIS/huc.html.
USIAC (US Interagency Advisory Committee on Water Data). 1982. Guidelines for determining flood flow
frequency. Bulletin 17B. Reston, VA: USGS.
van Lanen, H. A. J., L. M. Tallaksen, and G. Rees. 2007. “Droughts and climate change.” In Annex II in
Commission Staff Working Document Impact Assessment, Accompanying Document to Communication
Addressing the Challenge of Water Scarcity and Droughts in the European Union (COM 2007). Brussels,
Belgium: Commission of the European Communities.
Viessman, W., Jr., G. L. Lewis, and J. W. Knapp. 1989. Introduction to hydrology. 3rd ed. New York: Harper
and Row.
Villarini, G., J. A. Smith, F. Serinaldi, J. Bales, P. D. Bates, and W. F. Krajewski. 2009. “Flood frequency analysis
for nonstationary annual peak records in an urban drainage basin.” Adv. Water Resour. 32 (8): 1255–1266.
Vogel, R. M., and C. N. Kroll. 1989. “Low-flow frequency analysis using probability plot correlation coefficients.”
J. Water Resour. Plann. Manage. 115 (3): 338–357.
Vogel, R. M., and C. N. Kroll. 1991. “The value of streamflows record augmentation procedures in low-flow and
flood-flow frequency analysis.” J. Hydrol. 125 (3–4): 259–276.
Vogel, R. M., and C. N. Kroll. 1992. “Regional geohydrologic-geomorphic relationships for the estimation of
low-flow statistics.” Water Resour. Res. 28 (9): 2451–2458.
Vogel, R. M., and C. N. Kroll. 1996. “Estimation of baseflow recession constants.” Water Resour. Manage. 10 (4):
303–320.

Statistical Analysis of Hydrologic Variables

 332 STATISTICAL ANALYSIS OF HYDROLOGIC VARIABLES

Vogel, R. M., and I. Wilson. 1996. “Probability distribution of annual maximum, mean, and minimum
streamflows in the United States.” J. Hydrol. Eng. 1 (2): 69–76.
Vogel, R. M., C. Yaindl, and M. Walter. 2011. “Nonstationarity: Flood magnification and recurrence reduction
factors in the United States.” J. Am. Water Res. Assoc. 47 (3): 464–474.
Wallis, J. R., and P. E. O’Connell. 1972. “Small sample estimation of r1.” Water Resour. Res. 8 (3): 707–712.
Downloaded from ascelibrary.org by UNIVERSITY OF NEW SOUTH WALES on 08/24/19. Copyright ASCE. For personal use only; all rights reserved.

Wang, D., and X. Cai. 2009. “Detecting human interferences to low flows through base flow recession analysis.”
Water Resour. Res. 45 (7): W07426.
Wang, D. C., and J. D. Salas. 1989. “Stochastic modeling and generation of drought.” In Proc., ASCE National
Conf. on Hydraulic Modeling. Reston, VA: ASCE.
Weiss, G. 1977. “Shot noise models for the generation of synthetic streamflow data.” Water Resour. Res. 13 (1):
101–108.
Wilhite, D. A. 2000. “Drought as a natural hazard: Concepts and definitions.” In Drought: A global assessment.
Vol. 1, edited by D. A. Wilhite, 3–18. London: Routledge.
Wong, G., H. A. J. van Lanen, and P. J. J. F. Torfs. 2013. “Probabilistic analysis of hydrological drought
characteristics using meteorological drought.” Hydrol. Sci. J. 58 (2): 253–270.
Woodhouse, C. A. 2001. “A tree-ring reconstruction of streamflow for the Colorado front range.” J. Am. Water
Resour. Assoc. 37 (3): 561–569.
Yevjevich, V. M. 1967. An objective approach to definitions and investigations of continental hydrologic droughts:
Hydrology Paper No. 23. Fort Collins, CO: Colorado State Univ.
Yevjevich, V. M. 1972. Stochastic processes in hydrology. Fort Collins, CO: Water Resources Publications.
Young, A. R., C. E. Round, and A. Gustard. 2000. “Spatial and temporal variations in the occurrence of low flow
events in the UK.” Hydrol. Earth Syst. Sci. 4 (1): 35–45.
Yue, S., P. Pilon, and B. Phinney. 2003. “Canadian streamflow trend detection: Impacts of serial and cross-
correlation.” Hydrol. Sci. J. 48 (1): 51–63.
Zhang, Z., and C. N. Kroll. 2007a. “The baseflow correlation method with multiple gauged sites.” J. Hydrol. 347
(3–4): 371–380.
Zhang, Z., and C. N. Kroll. 2007b. “A closer look at baseflow correlation.” J. Hydrol. Eng. 12 (2): 190–196.
Zhang, Q., X. Gu, V. P. Singh, and X. Chen. 2015a. “Evaluation of ecological instream flow using multiple
ecological indicators with consideration of hydrological alterations.” J. Hydrol. 529 (3): 711–722.
Zhang, Y., X. Zhai, Q. Shao, and Z. Yan. 2015b. “Assessing temporal and spatial alterations of flow regimes in the
regulated Huai River Basin, China.” J. Hydrol. 529 (1): 382–397.

Statistical Analysis of Hydrologic Variables