Professional Documents
Culture Documents
1029/2004WR003075, 2005
Received 3 February 2004; revised 9 October 2004; accepted 26 October 2004; published 18 January 2005.
[1] Droughts influence the planning and design of water supply infrastructure.
Hydrologists ascertain drought duration, severity, and pattern of recurrence from
instrumental and reconstructed records (e.g., using tree rings) of streamflow and
precipitation. This work introduces a compound renewal model for the probabilistic
analysis of multiyear drought recurrence. The compound renewal process generalizes
the Poisson process. In the former the interarrival time between two consecutive events is
the duration of nondrought conditions, and the events (i.e., droughts) have a probabilistic
duration of at least q years. The sum of the interarrival time and its subsequent drought
duration is called the renewal time, which regenerates over time according to probabilistic
laws derived in this work. Drought severity is incorporated in the analysis by means of a
threshold quantile (e.g., the median or the average), so that low-streamflow conditions
become a drought whenever they last over q years. A case study dealing with a river basin
that has multiyear storage capacity, and in which droughts recurred frequently in the
twentieth century, demonstrates the analytical power of the compound renewal model.
Citation: Loáiciga, H. A. (2005), On the probability of droughts: The compound renewal model, Water Resour. Res., 41, W01009,
doi:10.1029/2004WR003075.
2 of 8
W01009 LOÁICIGA: ON THE PROBABILITY OF DROUGHTS W01009
return interval of droughts, and, for the probability distri- (cumulative) deficit defining overall drought severity. The
bution function of the number of droughts in an interval of modeling of drought duration and severity with a joint
arbitrary length. The pdf of the number of droughts, in probability distribution is appealing, but not devoid of
particular, is a very useful result from which many other simplifying assumptions. In contrast to the joint probability
variables of practical interest (e.g., moments), can be approach, this work treats drought severity parametrically.
derived. Specifically, drought severity enters the probability treat-
ment of drought recurrence through the choice of the
3.2. The pdf of the Renewal Time streamflow quantile qp, in which qp can be the median,
[9] Central to the theory of this work is the pdf of average, or some other suitable index of annual streamflow.
the renewal time, fR(t). It is shown in Appendix A that Annual streamflow less than the quantile qp is considered
fR(t) is low flow, and low-flow conditions become a drought
whenever they last over q years. It may be the case that as
a1 a2 h ðtqÞa2 i
the choice of qp changes, so does the duration threshold q. A
f R ðt Þ ¼ e eðtqÞa1 tq ð3Þ
a1 a2 reasonable choice of qp and q must be rooted on empirical
observations of historical droughts and their effects on water
The shape parameters must differ in equation (3), a1 6¼ a2, supply. That, in fact, was the rationale for choosing median
which is the situation of greatest practical interest. For annual streamflow and a drought duration threshold of
completeness, Appendix A presents the pdf fR(t) when a1 = 3 years in this study. Studies by the author and collaborators
equals the
a2. The expected value of the renewal time (R) in rivers basins and reservoir systems with multiyear carry-
sum of the expected values of D(= q + 1/a1) and T(= 1/a2), over capacity of the western United States [Loáiciga et al.,
or 1992a, 1993; Loáiciga and Leipnik, 1996; Loáiciga, 2002]
have shown that qp equal to the median annual flow and
¼qþ 1 þ 1
R ð4Þ
threshold q equal to 3 years are suitable choices in that
a1 a2 region. The compound renewal process parameterized in
terms of qp and q has a flexible structure, as shown in the
remainder of this article. In addition, it is mathematically
The expected value of the renewal time is a measure of the tractable and leads to sound descriptive equations, which are
probabilistic regularity with which droughts recur. The derived in section 3.5.
analyst chooses the threshold q, while the parameters a1 and
a2 are estimated from data. For example, one approach to 3.5. Probability of Multiple Droughts
estimate the parameter a2 is to use the fact that the expected
value of the interarrival time (T) is 1/a2. Therefore [12] The compound renewal process can be fully charac-
one calculates the sample average of T and then estimates terized by the probability of k droughts in a period (0, t),
the parameter a2 as the inverse of that sample average. P[N(t) = k]. To derive the latter probability distribution, the
The average drought duration estimated from data, D, ^ is time until the end of the kth drought, Rk, is most helpful:
used in conjunction with the threshold q to estimate a1 with X
k
^ q).
1/(D Rk ¼ Rr k ¼ 1; 2; 3; . . . ð6Þ
r¼1
3.3. Drought Risk
in which the renewal time R(=R1) is distributed according to
[10] The risk of drought, Ht, in a time interval [0, t] is equation (3). Note that N(t) k if and only if Rk+1 > t.
the probability that one or more droughts occur in that Therefore:
interval: Ht = P[N(t) 1] = 1 P[N(t) = 0] = 1 P[R t] =
P[R t], in which the last probability, after integrating the
right-hand side of equation (3), equals the following: P½NðtÞ ¼ k ¼ P½Rk t P½Rkþ1 t ð7Þ
k ¼ 1; 2; 3; . . . ; ½t=q
a1 a2 1 eðtqÞa2 1 eðtqÞa1
Ht ¼ tq ð5Þ
a1 a2 a2 a1
where [t/q] is the integral part of t/q. Clearly, P[N(t) = 0] =
Appendix A derives the risk for the case a1 = a2. P[R > t].
[13] The key to P[N(t) = k] is the pdf of Rk, fk(t).
3.4. On the Issue of Drought Severity Appendix A shows that fk(t) equals the following expres-
[11] Several authors [Yevjevich, 1967; Salas et al., 2004] sion (it was assumed, without loss of generality, that a1 > a2;
have proposed joint probability distributions for drought if a2 > a1 simply replace a1 with a2 and a2 with a1 in the
duration and severity, where the latter variable is modeled as following equation):
a deficit relative to baseline water availability. In the joint
probability approach, the marginal distributions of drought 1
duration and severity are obtained by integrating their joint pffiffiffi ða1 a2 Þk ðt q kÞk2 ða1 þ a2 Þðtq kÞ
f k ðt Þ ¼ p e 2
probability distribution. Furthermore, simplifying assump- ðk 1Þ! ða1 a2 Þk12
tions are introduced in the probabilistic characterization of
t qk
drought severity to render its marginal distribution Ik12 ða1 a2 Þ ð8Þ
2
mathematically manageable. A key assumption is the inde-
pendence of the interannual deficits that make up the k ¼ 1; 2; 3; . . . ; t q k
3 of 8
W01009 LOÁICIGA: ON THE PROBABILITY OF DROUGHTS W01009
½t=q
X
s2 ¼ k2 P½NðtÞ ¼ k ðE½NðtÞ Þ2 ð11Þ
k¼1
Figure 3. Probability density function (pdf) of the renewal [24] The above calculations and tests using the compound
time R, drought duration (D), and interarrival time between renewal model for drought analysis demonstrate several of
consecutive droughts (T). its several attractive features. First, it allows the treatment of
time series data in a completely probabilistic manner, thus
avoiding problems of autocorrelation and associated com-
Recall that the drought risk is the probability of at plications that arise in the statistical treatment of annual
least one drought in a time interval (0, t). The 100-year streamflow data. The author and collaborators [Loáiciga et
time upper bound was chosen because it is a typical service al., 1992a, 1993; Loáiciga and Leipnik, 1996; Loáiciga,
life of regional water storage facilities. It can be seen in 2002] have provided ample evidence that drought duration
Figure 4 that there is approximately a 50% risk of drought and the interarrival time are independently distributed, and
in any 8-year period, whereas the drought risk becomes this is all that is needed in constructing the compound
100% for any time interval in excess of 30 years. renewal model. Second, the compound renewal model
[22] Figure 5 shows the probability of observing k conforms effortlessly with the observed recurrence pattern
droughts (k = 0, 1, 2, 3, . . .) in any 100-year time interval of multiyear droughts. Third, the principles of probability
(P[N(t = 100) = k]). The mode (i.e., the most likely value) is and applied mathematical analysis used in conjunction with
k = 10 droughts. It is worth noting in Figure 5 that the range the compound renewal model produce elegant and useful
of the number of droughts k associated with nonnegligible equations that describe drought recurrence phenomena.
probabilities is relatively narrow. Specifically, nonnegligible Last, but not least, the compound renewal model has
probabilities were obtained for k in the range [4, 16]. built-in flexibility that permits handling drought severity
[23] The expected number of droughts (E[N(t)]) in any parametrically, through the specification of the threshold
period (0, t) in the upper Santa Ynez River is shown in quantile qp.
Figure 6, where t ranges between 10 to 100 years. It is [25] The prediction of the drought risk, the probability of
remarkable that, in spite of the mathematical complexity of droughts, the expected number of droughts, and like vari-
the equation for E[N(t)] (see equation (10)), the expected ables, is very useful in water resources research. These
number of droughts increases with increasing duration of probabilistic variables play a role in the sizing of reservoirs
the period in an approximately linear fashion. [Loáiciga, 2002], and, in general, in the risk analysis of
water resources infrastructure. Future research involving the
compound renewal model will focus on extending historical
streamflow records with tree-reconstructed streamflow and
assessing the changes in the probabilistic description of
drought risk brought about by hydrologic time series several in which Res 1 and Res 2 are the residues of the complex
centuries long. integral in equation (A6). Res 1 is given by
h iðtqkÞ i
Appendix A 1 dk1 ev þ ia2
A1. Derivation of the pdf of the Renewal Time Res 1 ¼
ðk 1Þ! dvk1
[26] One approach to obtaining the pdf of the renewal v¼ia1
time, fR(t), is to take the Fourier transform of the charac- ið1Þk Xk1
ð1Þs ðk þ s 1Þ!ðt qkÞks1 ea1 ðtqkÞ
¼
teristic function (cf) of R. Using the fact that R = D + T, in ðk 1Þ! s¼0 s! ðk s 1Þ!ða2 a1 Þkþs
which D and T are independent random variables, the cf of
R, CR(v), is ðA7Þ
a1 eivq a2 The expression for Res 2 is identical to that for Res 1,
CR ðvÞ ¼ ðA1Þ
ða1 ivÞ ða2 ivÞ with a1 replaced by a2 and a2 replaced by a1. Substitution of
pffiffiffiffiffiffiffi Res 1 and Res 2 into equation (A5) yields the pdf fk(t):
in which i = 1. The first and second terms in the right-
hand side of equation (A1) are the characteristic functions ða1 a2 Þk X k1
ðk þ s 1Þ!
of the drought duration (D) and the interarrival time (T), fk ðtÞ ¼ ð1Þs ðt qkÞks1
ðk 1Þ! s¼0 s!ðk s 1Þ!
respectively, whose pdf’s are given in equations (1) " #
and (2). Taking the Fourier transform of CR(v) yields ea1 ðtqkÞ ea2 ðtqkÞ
þ k ¼ 1; 2; 3; . . . ; t q k
fR(t) (for a1 6¼ a2) ða2 a1 Þkþs ða1 a2 Þkþs
Z1 ðA8Þ
1 a1 a2 h ðtqÞa2 i
fR ðtÞ ¼ eivt CR ðvÞdv ¼ e eðtqÞa1
2p a1 a2
1 [29] From a computational standpoint, it is advantageous
tq ðA2Þ to introduce modified Bessel functions in equation (A8).
Specifically, the modified Bessel functions used are
which is the same as equation (3). The integral in Kk(1/2)(z) [Gradshteyn and Ryzhik, 1994, equation 8.468]
equation (A2) was obtained with the method of residues and I±(k(1/2))(z) [Gradshteyn and Ryzhik, 1994, equation
[Churchill and Brown, 1990], which is presented in more 8.467], and the following relationship between them
general detail below (see equations (A6) and (A7)). The [Gradshteyn and Ryzhik, 1994, equation 8.485]:
case in which a1 = a2 = a can be treated in an analogous " #
manner. Let a1 = a2 = a in equation (A1) and then take the p Ikþ12 ðzÞ Ik12 ðzÞ
Fourier transform of the resulting cf to obtain Kk12 ðzÞ ¼ ðA9Þ
2 ð1Þkþ1
fR ðtÞ ¼ a2 ðt qÞeðtqÞa tq ðA3Þ
[30] The use of the modified Bessel functions and
expression (A9) in equation (A8) yields the desired form
A2. Drought Risk Equation of the pdf fk(t) (it was assumed, without loss of generality,
[27] When a1 = a2 = a, the risk Ht is obtained by that a1 > a2; if a2 > a1 simply replace a1 with a2 and a2
integrating the right-hand side of equation (A3) from q with a1 in the following equation):
to t, leading to 1
pffiffiffi ða1 a2 Þk ðt q kÞk2 ða1 þ a2 Þðtq kÞ
Ht ¼ 1 e aðtqÞ
½1 þ aðt qÞ tq ðA4Þ f k ðtÞ ¼ p e 2
ðk 1Þ! ða1 a2 Þk12
t qk
Ik12 ða1 a2 Þ ðA10Þ
A3. The pdf of Rk and fk(t) 2
tive. The modified Bessel function is positive. This is It is easily proven by induction that P[Rk 1] = 1 for k =
demonstrable from its alternative form [Gradshteyn and 1, 2, 3, . . ..
Ryzhik, 1994, equation 8.445]: [34] When a1 = a2 = a, the cdf of Rk is obtained by
integrating the right-hand side of equation (A15) to yield
1
z k2þ2s
X
1
2
Ik12 ðzÞ ¼ ðA12Þ g½2k; aðt q kÞ
s¼0 s!G k 12 þ s þ 1 P½Rk t ¼ k ¼ 1; 2; 3; . . . ; t q k ðA19Þ
ð2k 1Þ!
which is positive for positive z, with G( ) being the gamma Either equation (A16) or (A19), as the case might be, is
function. used in equation (7) to construct P[N(t) = k].
[31] If the shape parameters are equal, let a1 = a2 = a in
the cf (A5), take the Fourier transform to obtain
A5. Expected Value E[N(t)]
Z1 [35] By definition, the expected value is
1 ð1Þk ðaÞ2k
f k ðt Þ ¼ eivt Ck ðvÞdv ¼ f2pi Resg ðA13Þ
2p 2p ½t=q
1 X
E½NðtÞ ¼ s P½NðtÞ ¼ s ðA20Þ
s¼0
in which Res is the residue:
a2k From equation (A21) and the fact that N(t) k if and only if
f k ðt Þ ¼ ðt q kÞ2k1 eaðtq kÞ ðA15Þ
ð2k 1Þ! Rk t, the expected value E[N(t)] becomes
k ¼ 1; 2; 3; . . . ; t q k
½t=q
X
E½NðtÞ ¼ P½Rk t ðA22Þ
k¼1
A4. Probability P[N(t) = k]
[32] This probability is obtained from the cumulative dis- in which the probability P[Rk t] is that given by either
tribution function (cdf) of Rk, P[Rk t], see equation (7). equation (A16) or (A19). Equation (A22) may be the
The latter cdf is obtained integrating the right-hand side of computationally simplest formula to calculate the expected
equation (A8) term by term to produce the following: value E[N(t)]. Using equation (A16) in equation (A22)
produces E[N(t)] when a1 6¼ a2:
ða1 a2 Þk X
k1
ð1Þs ðk þ s 1Þ!
PðRk tÞ ¼ ½t=q
ðk 1Þ! s¼0 s!ðk s 1Þ! X ða1 a2 Þk X
k1
ð1Þs ðk þ s 1Þ!
" # E½NðtÞ ¼
g½k s; a1 ðt q kÞ g½k s; a2 ðt q kÞ k¼1
ðk 1Þ! s¼0 s!ðk s 1Þ!
þ " #
kþs kþs
aks
1 ða2 a1 Þ aks
2 ða1 a2 Þ g½k s; a1 ðt q kÞ g½k s; a2 ðt q kÞ
þ ks ðA23Þ
k ¼ 1; 2; 3; . . . ; t q k ðA16Þ a1ks ða2 a1 Þkþs a2 ða1 a2 Þkþs
which is the same as equation (9), where which is the same as equation (10). If a1 = a2 = a, the
substitution of the cdf (A19) in equation (A22) produces the
Z aj ðtq kÞ
desired expected value:
g k s; aj ðt q kÞ ¼ ez zks1 dz j ¼ 1; 2 ðA17Þ
0
½t=q
X g½2k; aðt q kÞ
is the incomplete gamma function. E½NðtÞ ¼ ðA24Þ
k¼1
ð2k 1Þ!
[33] Notice that from equation (A16),
k1
k X s References
a2 kþs1 a1
P½Rk 1 ¼ Bryant, E. A. (1991), Natural Hazards, Cambridge Univ. Press, New York.
a2 a1 s¼0 s a1 a2
Chung, C., and J. Salas (2000), Drought occurrence probabilities and risk of
k Xk1 s dependent hydrologic processes, J. Hydrol. Eng., 5, 259 – 268.
a1 kþs1 a2
þ ðA18Þ Churchill, R. V., and J. A. Brown (1990), Complex Variables and Applica-
a1 a2 s¼0 s a2 a1 tions, 5th ed., McGraw-Hill, New York.
7 of 8
W01009 LOÁICIGA: ON THE PROBABILITY OF DROUGHTS W01009
Dracup, J. A., K. S. Lee, and E. G. Paulson (1980), On the definition of Loáiciga, H. A., L. Haston, and J. Michaelsen (1993), Dendrohydrology
droughts, Water Resour. Res., 16(2), 297 – 302. and long-term hydrologic phenomena, Rev. Geophys., 31, 151 – 171.
Feller, W. (1957), An Introduction to Probability Theory and Its Applica- Loáiciga, H. A., J. B. Valdes, R. Vogel, J. Garvey, and H. H. Schwarz
tions, 2nd ed., John Wiley, Hoboken, N. J. (1996), Global warming and the hydrologic cycle, J. Hydrol., 174(1 – 2),
Forest, C. E., P. H. Stone, A. P. Sokolov, M. R. Allen, and M. D. Webster 83 – 128.
(2002), Quantifying uncertainties in climate system properties with the Mahlman, J. D. (1997), Uncertainties in projections of human-caused cli-
use of recent climate observations, Science, 295, 113 – 117. mate warming, Science, 278, 1416 – 1417.
Gradshteyn, I. S., and I. M. Ryzhik (1994), Table of Integrals, Series, and Parzen, E. (1964), Stochastic Processes, Holden-Day, Inc., Boca Raton, Fla.
Products, 5th ed., Spring, New York. Ross, S. M. (1993), Introduction to Probability Models, Springer, New
Grassi, H. (2000), Status and improvements of coupled general circulation York.
models, Science, 288, 1991 – 1997. Salas, J. D., C. Fu, A. Cancelliere, D. Dustin, D. Bode, A. Pineda, and
Gutman, N. B. (1998), Comparing the Palmer Drought Index and the E. Vincent (2004), Characterizing the severity and risk of droughts of
Standardized Precipitation Index, J. Am. Water Resour. Assoc., 34, the Poudre River, Colorado, J. Water Resour. Plann. Manage, in press.
1029 – 1039. Sen, Z. (1977), Rum-sums of annual streamflow series, J. Hydrol., 35,
Karl, T. R. (1983), Some spatial characteristics of drought duration in the 311 – 324.
United States, J. Clim. Appl. Meteorol., 22, 1356 – 1366. Sen, Z. (1980), Statistical analysis of hydrologic critical droughts, J. Hy-
Kerr, R. A. (2002), Dueling models: Future U.S. climate uncertain, Science, draul. Div. Am. Soc. Civ. Eng., 106, 99 – 104.
288, 2113. Shiau, J., and H. W. Shen (2001), Recurrence analysis of hydrologic
Loáiciga, H. A. (2002), Reservoir design and operation with variable lake drought of differing severity, J. Water Resour. Plann. Manage., 127(1),
hydrology, J. Water Resour. Plann. Manage., 128(6), 399 – 405. 30 – 40.
Loáiciga, H. A., and R. B. Leipnik (1996), Stochastic renewal model of Wolfram, S. (2000), The Mathematica Book, 4th ed., Cambridge Univ.
low-flow streamflow sequences, Stochastic Hydrol. Hydraul., 10, 65 – Press, New York.
85. Yevjevich, V. M. (1967), An objective approach to definitions and investi-
Loáiciga, H. A., and S. Renehan (1997), Municipal water use and water gations of continental hydrologic droughts, Hydrol. Droughts Pap. 23,
rates driven by severe drought: A case study, J. Am. Water Resour. Colo. State Univ., Fort Collins.
Assoc., 33, 1313 – 1326. Zelenhastic, E., and A. Salvai (1987), A method of stream drought analysis,
Loáiciga, H. A., J. Michaelsen, S. Garver, and L. Haston (1992a), Droughts Water Resour. Res., 23(1), 156 – 168.
in river basins of the western United States, Geophys. Res. Lett., 19,
2051 – 2053.
Loáiciga, H. A., J. Michaelsen, and P. F. Hudak (1992b), Truncated dis- H. A. Loáiciga, Department of Geography, University of California,
tributions in hydrologic analysis, Water Resour. Bull., 28, 853 – 863. Santa Barbara, Santa Barbara, CA 93106, USA. (hugo@geog.ucsb.edu)
8 of 8