American Society for Quality

Estimation of the Generalized Extreme-Value Distribution by the Method of Probability-

Weighted Moments
Author(s): J. R. M. Hosking, J. R. Wallis and E. F. Wood
Reviewed work(s):
Source: Technometrics, Vol. 27, No. 3 (Aug., 1985), pp. 251-261
Published by: American Statistical Association and American Society for Quality
Stable URL: http://www.jstor.org/stable/1269706 .
Accessed: 09/12/2012 10:52

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Statistical Association and American Society for Quality are collaborating with JSTOR to digitize,
preserve and extend access to Technometrics.

http://www.jstor.org

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
? 1985 AmericanStatisticalAssociation and TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3
the AmericanSociety forQualityControl

Estimationofthe GeneralizedExtreme-
Value Distribution
bythe Methodof
Probability-Weighted
Moments
J. R. M. Hosking J. R. Wallis E. F. Wood
Institute
of Hydrology IBM ResearchCenter of CivilEngineering
Department
Oxon OX10 8BB
Wallingford, YorktownHeights,NY 10598 PrincetonUniversity
England NJ08540
Princeton,

We use the methodof probability-weighted momentsto deriveestimatorsof the parameters

and quantilesof the generalizedextreme-value
distribution.We investigatethe propertiesof
theseestimatorsin large samples,via asymptotictheory,and in small and moderatesamples,
via computersimulation.Probability-weightedmomentestimatorshave low varianceand no
severebias, and theycompare favorablywithestimatorsobtained by the methodsof maxi-
mum likelihoodor sextiles.The methodof probability-weighted momentsalso yieldsa con-
venientand powerfultestof whetheran extreme-value distribution
is of Fisher-Tippett
Type I,
II, or III.

KEY WORDS: Generalizedextreme-value

distribution;Hypothesistesting;Order statistics;
moments.
Probability-weighted

1. INTRODUCTION case of the generalizedextreme-valuedistribution,is

The generalized extreme-valuedistributionof
Jenkinson(1955) is widely used for modeling ex-
tremesof naturalphenomena,and it is of consider-
able importancein hydrology,since it was recom-
mendedby the Flood StudiesReport[Natural Envi-
ronmentResearchCouncil (NERC) 1975a] formod-
eling the distributionof annual maxima of daily
streamflows of Britishrivers.Currentlyfavoredmeth-
ods of estimationof the parametersand quantilesof
thedistribution are Jenkinson's(1969) methodofsex-
tilesand themethodofmaximumlikelihood(Jenkin-
son 1969; Prescottand Walden 1980, 1983).Neither
method is completelysatisfactory: The justification
of the maximum-likelihood approach is based on
large-sampletheory,and therehas been littleassess-
mentof theperformance of themethodwhenapplied
to small or moderate samples; whereas the sextile
methodinvolvesan inherentarbitrariness (whysex-
tiles ratherthan,say, quartilesor octiles?),requires
interpolationin a table of values of a functionin
order to estimatethe shape parameterof the distri-
bution, and has statisticalpropertiesthat are not
knownevenforlargesamples.
Probability-weighted moments,a generalizationof
theusual momentsofa probabilitydistribution, were
introducedby Greenwood et al. (1979). There are
several distributions-for example, the Gumbel,
logistic,and Weibull-whose parameterscan be con-
venientlyestimatedfromtheirprobability-weighted
moments.The Gumbel distribution, being a special variable X with distribution function F(x) =
of particularinterest.Landwehret al. (1979) investi-
gated the small-samplepropertiesof probability-
weightedmoment(PWM) estimatorsof parameters
and quantilesfortheGumbeldistribution and found
themsuperiorin many respectsto the conventional
moment and maximum-likelihood estimators.The
estimatorsused by Landwehret al. (1979) are identi-
cal to Downton's(1966b) linearestimateswithlinear
coefficients,and thussharethe asymptoticproperties
of the latter;in particular,theasymptoticefficiencies
of the PWM estimatorsof the Gumbel scale and
locationparametersare .756and .996,respectively.
In this article we summarize some theory for
probability-weighted momentsand show that they
can be used to obtain estimatesof parametersand
quantiles of the generalized extreme-valuedistri-
bution. We derive the asymptoticdistributionsof
theseestimatorsand compare,via computersimula-
tion,the small-samplepropertiesof the PWM, sex-
tile, and maximum-likelihoodestimators. The
method of probability-weightedmoments out-
performsthe othermethodsin many cases and will
usually be preferred to them.We also derive,from
the PWM estimatorof the shape parameterof the
generalized extreme-valuedistribution,a test of
whetherthisshape parameteris zero, and we assess
theperformance ofthistestby computersimulation.
2. PROBABILITY-WEIGHTED MOMENTS
The probability-weighted
momentsof a random
251

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
 252 J. R. M. HOSKING,J. R. WALLIS,AND E. F. WOOD

P(X < x) are thequantities Gini's mean difference,has a historygoing back at

- F(X)}S],
least as faras van Andrae(1872), and ?x/lr U2 is a
Mp,r,s= E[XP{F(X)}'{1 (1)
98% efficient estimatorof the scale parameterof a
where p, r, and s are real numbers.Probability- Normaldistribution (Downton 1966a); and
weightedmomentsare likelyto be most usefulwhen
the inversedistribution
functionx(F) can be written 6b2-6b + bo =U3 = 1(3)l (xi-2xj +x)
i>j>k
in closed form,forthenwe maywrite
(6)
is a U-statisticfor estimatingskewness,which has
Mp,r,s = {x(F)}PFr(l - F)s dF, (2)
o been used as the basis of a test for Normalityby
and this is oftenthe most convenientway of evalu- Locke and Spurrier(1976). U-statisticsare widely
used in nonparametric statistics(e.g.,see Fraser 1957,
atingthesemoments.The quantitiesMpo0 (p = 1, 2,
...) are the usual noncentralmomentsof X. The mo- chap. 4, and Randles and Wolfe 1979,chap. 3), and
ments Ml,., may, however,be preferablefor esti- theirdesirablepropertiesof robustnessto outliersin
thesample,highefficiency, and asymptoticnormality
matingthe parametersof the distribution of X, since
the occurrenceof only the firstpower of X in the may be expected to extend to the probability-
expressionforMl.r,s means thatthe relationshipbe- weightedmomentestimatorsbrand otherquantities
tweenparametersand momentsoftentakesa simpler calculatedfromthem.
formin this case than when using the conventional 3. PWM ESTIMATORS FOR THE
moments.When r and s are integers,F'(1 - F)s may GENERALIZED EXTREME-VALUE
be expressedas a linearcombinationofeitherpowers DISTRIBUTION
of F or powersof (1 - F), so it is naturalto summa-
rize a distributioneither by the moments Ml.r0o The generalizedextreme-value (GEV) distribution,
introducedby Jenkinson(1955), combines into a
(r = 0, 1, 2, ...) or by Mlo,0 (s= 0, 1, 2, ...). Green-
singleformthe threepossibletypesof limitingdistri-
wood et al. (1979) generallyfavoredthe latterap- butionforextremevalues,as derivedby Fisher and
proach,but herewe willconsiderthemomentsfr= Tippett(1928).The distribution
functionis
Ml,,.o = E[X{F(X)}r] (r = 0, 1, 2, ...).
Given a random sample of size n fromthe distri- F(x)= exp [-1 - k(x -)/}k] k 0,
butionF, estimationof f, is mostconveniently
based = exp [- exp {-(x -/a}], k=0, (7)
on the orderedsample xl < x2 < < x". The sta-
tistic withx bounded by ? + a/kfromabove ifk > 0 and
frombelow if k < 0. Here ~ and a are location and
br= n-1, (j- 1)(j-2)... (-r) (3) scale parameters,respectively,and theshape parame-
=l (n- lXn-2) (n-r)xi ter k determineswhichextreme-value distribution is
is an unbiasedestimatorof /, (Landwehret al. 1979). represented:Fisher-Tippett TypesI, II, and III corre-
Insteadone mayestimatef/,by spond to k = 0, k < 0, and k > 0, respectively.When
k = 0 the GEV distributionreducesto the Gumbel
n
n- X distribution.
The inversedistribution function is
[Pn] ,n (4)
x(F) = ~ + a{1-(-log
j=l
F)k}/k, k #0
where pj,n is a plotting position-that is, a
distribution-freeestimate of F(xj). Reasonable = -a log (-log F), k = 0. (8)
choices of pj,,, such as pj, = (j - a)/n,0 < a < 1, or In practicethe shape parameterusually lies in the
pj, = ( j-a)/(n+ 1 -2a), - < a < 1, yield esti- range - < k < ?. For example,the data base used
mators l,[pj,p],whichare asymptotically equivalent in NERC (1975a) included 32 annual flood series
to brand, therefore, consistentestimatorsoffr. withsample sizes of 30 or more.GEV distributions
The estimatorsb, are closelyrelatedto U-statistics were fittedto these 32 samples by the method of
(Hoeffding1948),whichare averagesof statisticscal- maximumlikelihood:The estimatedshape parameter
culated fromall subsamplesof size j < n of a given rangedfrom-.32 to .48.
sample of size n. In particular,bo = n-l xj is a The probability-weighted momentsof the GEV
trivialexample of a U-statistic,and it is a natural distribution
fork - 0 are givenby
estimatoroflocationofa distribution;
fr = (r + 1)-1[ + a{l - (r + 1)-kr(l + k)}/k],
2bl -bo-U = i )-l E(i-x) (5)
i>j k > -1 (9)
is a U-statistic for estimating the scale of a (forproof,see AppendixA). When k < -1, flo(the
distribution-thestatisticU2, sometimesknown as mean ofthedistribution)
and therestofthefr,do not

TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
 ESTIMATIONOF GENERALIZEDEV DISTRIBUTION
exist.From (9) we have
+ a{l - r( + k)}/k,
io0=
2i1 -/ 0o= aF(1 + k)(l - 2-)/k,
and
(3#2 - fo)/(2# - fo) = (1 - 3 -k)/(1- 2 -).
The PWM estimatorsf, a, I of the parametersare
the solutionsof (10)-H12)for a,a, and k whenthe fr,
are replaced by theirestimatorsb, or f/r[Pj,n].To
obtaink we mustsolvetheequation
(3b2 - bo)/(2b1- bo) = (1 - 3 -')/(1 - 2-k). (13)
The exact solution requiresiterativemethods,but
because the function (1 - 3-k)/(1 - 2-k) is almost
linear over the range of values of k (- 1 < k < 2),
which is usually encounteredin practice,low-order
polynomialapproximationsfork are veryaccurate.
We proposetheapproximateestimator
2b -bo
k = 7.8590c+ 2.9554c2, =
3b2 - bo
theerrorin k due to using(14) ratherthan(13) is less
than .0009 throughoutthe range - < k < ?. Given
k,thescale and locationparameterscan be estimated
successivelyfromEquations(11) and (10) as
(2b, - bo)k
a = r(o1 k)( -{(1 + {(
Equations (13) and (15), or theirequivalentforms
with br replaced by /r[P,n],definethe PWM esti-
mators of the parametersof the GEV distribution.
Given the estimatedparameters,the quantilesof the
distributionare estimatedusing the inversedistri-
butionfunction(8).
When calculated using br as the estimatorof fir,
the PWM estimatesof theGEV distribution satisfya
feasibilitycriterion,namelythat k > -1 and a > 0
almost surely(for proof,see Appendix B). This is
clearlya desirableproperty,sinceone would likeesti-
mates calculated using a set of sample momentsto
yield an estimateddistributionforwhichthe corre-
sponding population momentsexist. We have not
been able to prove that this feasibilitycriterionis
satisfiedwhen plotting-position estimators/r[Pj,n]
are used, but no examplesof the criterionnot being
satisfiedhave been discoveredin practice.
4. ASYMPTOTIC DISTRIBUTION OF
PWM ESTIMATORS
When modeling the propertiesof extremesof
physicalprocesses,it rarelyoccurs thatthe available
This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
253
data set is large enough to ensure that asymptotic
large-sampletheorymay be directlyapplied to the
(10) problem.It is nonethelessvaluable to investigatethe
(11) asymptoticpropertiesof a new statisticaltechnique,
for two main reasons. First, one may seek to es-
tablish the integrityof the technique,in the sense
(12) that when a large sample is available, the new
methodshouldnot be grosslyinefficient comparedto
an established,asymptotically optimalmethodsuch
as maximumlikelihood.Second, asymptotictheory
may provide an adequate approximationto some
aspect of the distribution of a statisticeven in quite
small samples.In the presentcase we shall see that
the varianceof PWM estimatorsof parametersand
quantiles of the GEV distributionis well approxi-
matedby asymptotictheoryforsamplesizes of 50 or
larger.
We considerfirsttheasymptotic distribution ofthe
br. From (3), bris a linearcombinationof the order
statisticsxl, ..., xn,and theresultsof Chernoff et al.
log 2 (1967) may be used to prove that the vectorb = (bo
log 3' blb2)bhas asymptotically a multivariate Normaldis-
tribution with mean / = (O fllf2)T and covariance
matrixn- V. The elementsof V and details of the
proofare givenin AppendixC.
The asymptoticdistributionof the PWM esti-
matorsof the GEV parametersfollowsfromthe pre-
ceding result. Let 0 = (ock)T, 0 = (ak)T, and write
(13) and (15) as the vectorequation 0 =f(b). Define
the 3 x 3 matrix G = (gij) by gij = fil/bj. Then
+ k-
1
k)-}/k. asymptotically, 0 has a multivariateNormal distri-
bution with mean f(/f)= 0 and covariance matrix
(15) n-lGVGT (Rao 1973,p. 388). The covariancematrix
has theform
3
0C2Wil W12 (Xwl3
n- GVGT =n- 1 2W212 a2W22 caW23 . (16)
\ awI3 aw23 W33
The wij are functionsof k and have a complicated
algebraic form,but theycan be evaluated numeri-
callyand are givenin Table 1 forseveralvalues of k.
As k approaches - ?, thevarianceof theGEV distri-
bution becomes infiniteand the variancesof the br
and of the PWM parameterestimatorsare no longer
ofordern- asymptotically.
The asymptoticbiases of the estimatorsare of
ordern- 1 and can be evaluatedby methodssimilar
to thoseof Rao (1973,p. 388). The biases,graphedin
Figure 1, are negligiblein large samples provided
thatk > -.4.
For comparison, asymptotic biases of the
maximum-likelihood estimatorsof the parametersof
theGEV distribution are graphedin Figure2. These
biases were calculated using equation 3.12 of Shen-
ton and Bowman (1977) and are functionsof ex-
pected values of third derivatives of the log-
TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3
254 J. R. M. HOSKING,J. R. WALLIS,AND E. F. WOOD

Table 1. Elementsof theAsymptoticCovariance Matrixof the 10 -

PWM Estimatorsof the Parametersof the GEV Distribution 9-
8-
k w1l w12 wl3 w 22 w23 w 33 7-
6-
-.4 1.6637 1.3355 1.1405 1.8461 1.1628 2.9092
c 5-
-.3 1.4153 .8912 .5640 1.2574 .4442 1.4090 x
-.2 1.3322 .6727 .3926 1.0013 .2697 .9139 u1)
0
4-

-.1 1.2915 .5104 .3245 .8440 .2240 .6815 GD 3-

.0 1.2686 .3704 .2992 .7390 .2247 .5633 2-
.1 1.2551 .2411 .2966 .6708 .2447 .5103 1 -
.2 1.2474 .1177 .3081 .6330 .2728 .5021 0
.3 1.2438 -.0023 .3297 .6223 .3033 .5294 -1 -
.4 1.2433 -.1205 .3592 .6368 .3329 .5880
-2 II I I
I I
II
I
I I I
I
I

NOTE: PWM-Probability-weightedmoment;GEV-Generalized extremevalue. -0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5
Shape parameter k

likelihoodfunction.For k ?> theseexpectationsdo Figure2. AsymptoticBias of ML Estimatorsof Parametersof

not exist and the biases of the maximum-likelihood
estimatorsare not ofordern- asymptotically.
The asymptoticvariances of the estimatorsare
graphedin Figure3, and theirasymptoticefficiencies
in Figure 4. Asymptoticefficiency is definedas the
ratio
eff(0i) = lim (var Oi/varOi)
foreach elementOiof the parametervector0, where
0i is themaximum-likelihood estimatorof 0i. Overall particularlyin the extremeupper tail of the distri-
efficiency is the ratio of the determinantsof the
asymptoticcovariancematricesof 6 and 0. The over-
all efficiencyof the PWM estimatorstendsto zero at
k = +.5, but forvalues ofk not too farfromzero the
PWM method is reasonably efficient. Within the
range -.2 < k < .2, whichis valid formany hydro-
logicaldata sets,each PWM parameterestimatorhas
efficiencyofmorethan.7.
Correspondingresultsmay be obtainedforPWM
and maximum-likelihood estimatorsof quantiles of
theGEV distribution. These are not presentedin full,
but Tables 2 and 3 give resultsforvariousquantiles
when k = -.2 and for various values of k at the
F = .98 quantile. The tables illustratethe main
the GEV Distribution: ;- - - c; ----- .
characteristics of PWM quantile estimators,which
are (a) highpositivebias in extremeupper tail,aris-
ingfrompositivebias in k; (b) highvariancein upper
tail when k < 0; and (c) fairor highefficiency except
when k is close to +.5. The maximum-likelihood
quantile estimatorshave lower variance than the
PWM estimatorsbut have some verylarge biases,
bution.
The resultsof thissectionwere derivedforPWM
estimatorsthatuse br to estimate/r. If the plotting-
position estimates [r[pPj,n]are used instead, the
asymptoticvariances and efficienciesremain un-
changed,but the asymptoticbiases are different and
cannot be easily calculated,being affectedby the
biases in the/r.[pj,n]themselves.
5. SMALL-SAMPLE PROPERTIES OF
ESTIMATES OF THE GEV DISTRIBUTION
A computersimulationexperimentwas run to
comparethreemethodsof estimationof theparame-

10 - 3 -
9-
8-
7-
6- 2 -
c
c 5- x
x a)
U) 4- U
u c
c
m 3-
.C
a ,\,-------------------------
2- 1-
1 -
0
-1 -
-2 j1
I II II II II II . I. . I . . I I u
n
I I I I 1I I I
"
-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5
Shape parameter k Shape parameter k
Figure 1. AsymptoticBias of PWM Estimatorsof Parameters Figure3. AsymptoticVarianceof PWM Estimatorsof Parame-
of the GEV Distribution: k:- - - E; ----- e. tersof the GEV Distribution: ; --; ----- .

TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
 ESTIMATION OF GENERALIZED EV DISTRIBUTION 255

1.0 - Table 3. AsymptoticBias, Variance,and Efficiency

of the PWM and ML Estimatorsof the F = .98
Quantileof the GEV Distribution,and
0.8 -
the Efficiency
of the PWM Estimators
7 (parameters = 0 O, a = 1 )
> 0.6 - ,1/

c n x Bias n x Variance
._
' 0.4 - / k x (F) PWM ML PWM ML Efficiency
-.4 9.41 -64.8 31.5 1,170 574 .49
0.2 - -.3 7.41 -15.3 15.3 369 275 .75
-.2 5.91 -4.4 5.5 147 131 .89
-.1 4.77 -1.1 -.3 64.8 62.0 .96
0.0 .0
I I I I I I I I I I 3.90 -.1 -3.8 30.2 28.8 .95
-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 .1 3.23 .3 -6.2 14.7 13.0 .88
Shape parameter k
.2 2.71 .5 -8.9 7.53 5.62 .75
Figure 4. AsymptoticEfficiencyof PWM Estimatorsof Pa- .3 2.30 .5 -23.2 4.04 2.28 .56
rametersof the GEV Distribution: k; - - ----- .4 1.98 .6 - 2.28 .83 .36
f;
...... overall efficiency(i.e., ratio of determinantsof asymp-
NOTE: PWM-Probability-weighted moment; ML-Maximum likelihood; GEV-
toticcovariance matricesof ML and PWM estimators). Generalizedextremevalue.

tersand quantilesof the GEV distribution. Simula- sextiles(JS).The PWM methodrequiresa choiceofa
tionswereperformed forsample sizes n = 15, 25, 50, suitableestimatorof #,.Severalpossibilitieswerein-
100 with the shape parameterof the distribution vestigated,includingthe unbiasedestimatorb, and a
takingvalues k = -.4, -.2, .0, .2, .4. All themethods numberof plotting-position estimators/4,[Pj,n.The
of estimationare invariantunderlineartransforma- best overall results were given by the estimator
tions of the data, so withoutloss of generalitythe /,[Pj,n]withPj,n= (J - .35)/n,and the simulationre-
location and scale parameters~ = 0 and a = 1 were sults presentedfor the PWM method referto this
used throughout.For each combinationof values of versionofthePWM estimators.
n and k, 10,000randomsamplesweregeneratedfrom Maximum-likelihood estimationof the GEV dis-
the GEV distribution, and foreach sample the pa- tribution is not always satisfactory.The log-
rameters~, a, and k, and the quantilesx(F), where likelihoodfunctionfora sample{x,,..., x"} is
F = .001, .01, .1, .2, .5, .8, .9, .98, .99, .998, .999,were log L =-n log a - (1 - k) y, i- E e-yi,
estimatedby each of threemethods:(a) the method
of probability-weighted moments(PWM), described Yi= -k 1 log { 1 - k(xi- )/a},
before;(b) themethodof maximumlikelihood(ML), and log L can be made arbitrarily large by setting
using Newton-Raphson iterationto maximize the k > 1 and choosing~ and a so thatthe upperbound
likelihoodfunction, as recommended by Prescottand + ca/kof the distribution is sufficiently
close to the
Walden (1983); and (c) Jenkinson's (1969) methodof largestdata value. In practicemaximum-likelihood
estimatesof theparametersare obtainedby findinga
Table 2. AsymptoticBias and Varianceof PWM and
local maximumoflog L. For some samples,however,
ML Estimatorsof GEV Quantiles, and Efficiency it appearsthatlog L does not have a local maximum.
of the PWM EstimatorsCompared WithML In our simulationsof theGEV distribution thisnon-
(parameters =- 0, a = 1, k = -.2) regularityof the likelihood functioncaused oc-
casional nonconvergenceof the Newton-Raphson
n x Bias n x Variance
iteration that was used to maximize the log-
F x (F) PWM ML PWM ML Efficiency likelihood;such sampleswereomittedfromthe sim-
ulations.Such occurrenceswererareexceptwhenthe
.001 -1.60 -1.2 1.7 3.78 2.29 .61
.01 -1.32 -.2 1.6 2.06 1.35 .66 sample size was very small and when k = .4 (see
.1 -.77 .8 1.2 .86 .79 .92 Table 4), and in our opiniontheyhave no significant
.2 -.45 1.0 .9 .88 .88 1.00 effecton the conclusionsthatmay be drawnregard-
.5 .38 .6 -.1 1.92 1.79 .93 ing the relativemeritof the different estimationpro-
.8 1.75 -1.3 -1.2 6.10 6.00 .98
.9 2.84 -3.1 -1.1 16.1
cedures.Similardifficulties withmaximum-likelihood
15.9 .99
.98 5.91 -4.4 5.5 147 131 .89 estimationare encounteredwith otherdistributions
.99 7.55 -1.6 13.4 336 289 .86 whose range depends on theirparameters,such as
.998 12.33 23.9 54.8 1,760 1,430 .81 the three-parameter lognormal,Weibull,and gamma
.999 14.90 49.1 88.2 3,310 2,630 .80 distributions(e.g., see Griffiths 1980 or Cheng and
NOTE: PWM-Probability-weighted moment; ML-Maximum likelihood; GEV- Amin1983).
Generalizedextremevalue. The simulationresultsforestimationoftheparam-

TECHNOMETRICS, AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
256 J. R. M. HOSKING, J. R. WALLIS, AND E. F. WOOD

Table 4. FailureRate of Maximum- Likelihood and n = 25, and theirbias, thoughoftenlargerthan

Estimationforthe GEV Distribution
k
n -.4 -.2 .0
15 .1 .6 1.7
25 .0 .0 .0
50 .0 .0 .0
100 .0 .0 .0
NOTE: GEV-Generalized extremevalue. Tabulated values are numbersof failures
to converge of Newton-Raphson iterationsper 100 simulatedsamples.
eters of the GEV distributionare summarizedin
Tables 5 and 6. Resultsforthe estimatorof k are of
the greatestimportance,since this parameterdeter-
minesthe overallshape of the GEV distribution and
the rate of increaseof the upperquantilesx(F) as F
approaches 1. Apartfromthe case n = 100,whenall
the methods have comparable performance,the
PWM estimatorhas, consistently, the lowest stan-
dard deviationof thethreeestimatorsof k,its advan-
tage being particularlymarked in small samples,
n = 15 and n = 25. The PWM estimator has in gen-
eral a largerbias than the otherestimators,but its
bias is small near theimportantvalue k = 0 and is in
any case relativelyinsignificantcompared to the
standard deviationin its contributionto the mean
squared errorof k. The sextileestimatorof k has a
large positivebias in small sampleswhenk < 0, and
its standarddeviationis generallylargerthanthatof
thePWM estimator.
Similarresultscan be seen forestimatorsof ~ and
a, althoughthe differences betweenthe variancesof
the estimatorsare less markedthan is the case with
estimatorsof k. In general,PWM estimatorshave
smalleststandard deviation,particularlyfor n = 15
forthe ML or sextileestimators,is not severe.The
standard deviations of the PWM estimatorsfor
n > 50 are well approximatedby theirlarge-sample
.2 .4 values given by (16) and Table 1. Maximum-
likelihood estimatorsare the least biased but are
3.8 12.4
.3 1.9 more variable than PWM estimatorsin small sam-
.0 .0 ples. Even at sample size 100, the asymptoticinef-
.0 .0 ficiencyof the PWM methodcomparedto maximum
likelihoodis not apparentin the simulationresults.
Sextileestimatorsin generalhave largerstandardde-
viationsthan PWM estimatorsand have some sig-
nificantbiases in smallsampleswhenk < 0.
The statisticalpropertiesof estimatorsof quantiles
of the GEV distributionwere evaluated for many
combinationsof quantiles and values of the shape
parameterk, and only a few representative simula-
tion resultsare presentedin Table 7. The most im-
portantaspect of quantileestimationin hydrological
applicationsis estimationoftheextremeupperquan-
tiles,particularly forheavy-tailedGEV distributions
withk < 0. Table 7 givesthebias and standarddevi-
ation of the estimatedupperquantilesfortwo GEV
distributions,one withk < 0 and one withk > 0. Re-
sults are presentedfor the ratios x(F)/x(F) rather
than forthe x(F) themselves, since the formerquan-
titiesare moreeasilycomparedat different F values.
For sample size 100 the threemethodshave com-
parable performance.In small samples the upper
quantiles obtained by PWM estimationare rather
biased,but theyare stillpreferable to themaximum-
likelihood estimators,since these have very large
biases and standard deviations.The errorsin the
maximum-likelihood quantileestimatorsarisechiefly
froma small numberof simulatedseriesthat yield
large negativeestimatesof k and consequentlygive
verylargeestimatesofextremeupperquantiles.

Table 5. Bias of Estimatorsof GEV Parameters

Bias (e) Bias (d) Bias (k)

n Method -.4 -.2 .0 .2 .4 -.4 -.2 .0 .2 .4 -.4 -.2 .0 .2 .4

15 PWM .10 .05 -.02 .00 -.03 .00 -.06 -.10 -.11 -.12 .11 .03 -.03 -.08 -.12
ML .03 .03 .05 .05 .04 -.07 -.07 -.07 -.06 -.07 -.04 -.02 .02 .04 .03
JS .11 .08 .06 .04 .02 -.05 -.06 -.07 -.08 -.08 .10 .06 .03 .01 -.01
25 PWM .06 .03 .01 -.01 -.02 .00 -.04 -.06 -.07 -.07 .08 .02 -.02 -.05 -.07
ML .01 .02 .03 .03 .04 -.04 -.04 -.04 -.03 -.03 -.02 -.01 .02 .04 .05
JS .06 .04 .03 .02 .01 -.03 -.03 -.04 -.04 -.05 .07 .04 .02 .01 -.00
50 PWM .04 .02 .01 .00 -.01 .01 -.02 -.03 -.04 -.04 .05 .02 -.01 -.02 -.04
ML .01 .01 .02 .02 .02 -.02 -.02 -.02 -.02 -.01 -.01 .00 .01 .02 .03
JS .04 .02 .02 .01 .01 -.02 -.02 -.02 -.02 -.02 .04 .02 .01 .00 .00
100 PWM .02 .01 .00 .00 -.01 .00 -.01 -.02 -.02 -.02 .03 .01 .00 -.01 -.02
ML .00 .00 .01 .01 .01 -.01 -.01 -.01 -.01 .00 -.01 .00 .00 .01 .02
JS .02 .01 .01 .00 .00 -.01 -.01 -.01 -.01 -.01 .02 .01 .00 .00 .00

NOTE: GEV-Generalized extreme value; PWM-Probability-weighted moment; ML-Maximum likelihood; JS-Jenkinson's

(1969) sextiles.

TECHNOMETRICS, AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
 ESTIMATIONOF GENERALIZEDEV DISTRIBUTION 257

Table 6. Standard Deviation of Estimatorsof GEV Parameters

Standard Deviation (E,) Standard Deviation (a) Standard Deviation (k)

n Method -.4 -.2 .0 .2 .4 -.4 -.2 .0 .2 .4 -.4 -.2 .0 .2 .4

15 PWM .32 .30 .29 .28 .28 .33 .25 .21 .19 .19 .20 .19 .18 .18 .19
ML .32 .32 .31 .30 .28 .28 .25 .23 .22 .21 .36 .32 .29 .27 .23
JS .33 .31 .30 .29 .28 .32 .28 .24 .21 .21 .25 .24 .23 .23 .23
25 PWM .24 .23 .22 .22 .22 .24 .19 .17 .15 .16 .18 .16 .14 .14 .15
ML .24 .24 .23 .23 .22 .21 .19 .17 .16 .17 .24 .21 .20 .18 .17
JS .24 .23 .23 .22 .22 .24 .21 .18 .17 .16 .19 .18 .17 .17 .17
50 PWM .17 .16 .16 .16 .16 .17 .14 .12 .11 .11 .14 .12 .11 .10 .11
ML .17 .16 .16 .16 .16 .15 .13 .12 .11 .11 .15 .13 .12 .11 .11
JS .17 .16 .16 .16 .16 .17 .14 .13 .12 .11 .14 .13 .12 .11 .12
100 PWM .12 .12 .11 .11 .11 .12 .10 .09 .08 .08 .11 .09 .07 .07 .08
ML .12 .11 .11 .11 .11 .10 .09 .08 .08 .08 .10 .09 .08 .07 .07
JS .12 .12 .11 .11 .11 .12 .10 .09 .08 .08 .10 .09 .08 .08 .08

NOTE: GEV-Generalized extreme value; PWM-Probability-weighted moment; ML-Maximum likelihood; JS-Jenkinson's

(1969) sextiles.

Estimationof extremelower quantilestendsto be combininginformation froma numberof indepen-

less importantin practicethan estimationof upper
quantiles,so simulationresultsforthiscase are not
given in detail. All threemethodsgive comparable
resultswhenn > 50, but forsmall samplesthe PWM
estimators have smallest standard deviation and
small or moderatebias, and are generallyto be pre-
ferred.
All the methods of quantile estimationare very
inaccurate when estimatingextreme quantiles in
small samples with k < 0. It is of course to be ex-
pectedthat a quantilex(F) cannot be estimatedreli-
ably froma sampleof size n ifF > 1 - 1/n.Nonethe-
less it is sometimespossible to obtain usefulesti-
mates of extremequantilesfromshortdata sets,by
dent data sets. Such a "regionalization"procedure,
based on the PWM estimationmethodforthe GEV
is describedin Hoskinget al. (1985).
distribution,
6. TESTING WHETHER THE SHAPE
PARAMETER IS ZERO
The Type I extreme-value distribution, or Gumbel
is a particularlysimple special case of
distribution,
the GEV distribution, and it is oftenusefulto test
whethera givenset ofdata is generatedby a Gumbel
ratherthana GEV distribution. This is equivalentto
testingwhetherthe shape parameterk is zero in the
GEV distribution. A test of this hypothesismay be
based on the PWM estimatorof k. On the null hy-

Table 7. Bias and Standard Deviation of Estimatorsof GEV Quantiles

k = -.2 k =.2

F= .9, F = .99, F = .999, F =.9, F = .99, F = .999,

x(F) = 2.84 x(F) = 7.55 x(F) = 14.90 x(F) = 1.81 x(F) = 3.01 x(F) = 3.74

n Method Bias SD Bias SD Bias SD Bias SD Bias SD Bias SD

15 PWM -.06 .34 -.02 .55 .15 1.12 -.04 .23 .08 .32 .25 .56
ML .01 .50 * * * * -.06 .23 .02 .79 .44 8.10
JS -.08 .34 -.05 .56 .14 1.20 -.06 .22 -.01 .33 .11 .64

25 PWM -.04 .27 -.01 .45 .11 .88 -.02 .18 .05 .24 .14 .39
ML -.01 .32 .16 .97 .74 9.59 -.04 .17 -.04 .29 .02 .86
JS -.05 .27 -.02 .46 .09 .89 -.04 .17 -.01 .25 .05 .42

50 PWM -.02 .19 -.01 .33 .06 .61 -.01 .12 .02 .16 .07 .25
ML -.01 .22 .05 .40 .18 .86 -.03 .12 -.03 .16 -.02 .28
JS -.02 .19 -.01 .33 .04 .59 -.02 .12 -.01 .17 .02 .27

100 PWM -.01 .14 .00 .24 .04 .42 -.01 .09 .01 .12 .03 .17
ML .00 .15 .02 .25 .08 .43 -.01 .09 -.02 .11 -.01 .15
JS -.01 .14 .00 .24 .03 .40 -.01 .09 .00 .12 .01 .18

* Values thatvaried
widelybetween differentsets of 1,000 simulationsand consequentlycould not be estimatedaccurately.
NOTE: GEV-Generalized extremevalue; SD-Standard deviation; PWM-Probability-weighted moment; ML-Maximum likeli-
hood; JS-Jenkinson's (1969) sextiles. Tabulated values are bias and standarddeviation of the ratioR(F)/x(F) ratherthan of the
quantile estimator,x(F) itself.

TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
258 J. R. M. HOSKING,J. R. WALLIS,AND E. F. WOOD

pothesisHo: k = 0, the PWM estimatorkis asymp- Table 9. Powers of Two Testsof the Hypothesisk = 0
toticallydistributedas N(0, .5633/n),so the testmay Against One -Sided and Two -Sided Alternatives
be performedby comparing the statistic Z = (sample size 50, nominalsignificancelevel 5%)
k(n/.5633)1/2with the criticalvalues of a standard Applications
Normal distribution. positivevalues of Z
Significant
imply rejectionof Ho in favor of the alternative For Modified
k > 0, and significant For Z Test Likelihood-Ratio Test
negativevalues of Z implyre-
jectionin favorofk < 0. k k< 0 k>0 k O k<0 k>0 k O
The size of the testbased on Z forvarioussample
sizes is givenin Table 8. In Table 9 the powerof the -.5 .96 - .94 .97 - .96
-.4 .90 - .85 .93 - .88
Z test is compared with that of the modified
-.3 .77 - .68 .80 - .72
likelihood-ratiotest,which was found by Hosking -.2 .54 - .43 .57 - .46
(1984) to be the best test of the hypothesisk = 0. -.1 .25 - .17 .25 - .17
Tables 8 and 9 are based on computersimulationsof .0 .05 .05 .05 .04 .06 .05
50,000samplesforeach value of n and k. The Z test .1 - .18 .11 - .25 .17
has poweralmostas highas thelikelihood-ratio .2 - .50 .37 - .57 .43
test,
.3 - .83 .73 - .89 .80
and forsamplesof size 25 or moreits distribution on .4 - .96 .93 - .98 .96
Ho is adequately approximatedby the standard .5 - 1.00 .99 - 1.00 1.00
Normal significancelevels. Since the statisticZ is
verysimple to compute,the Z testcan be strongly are also plottedin Figure 5. The Z test of the hy-
recommendedas a convenientand powerfulindica- pothesisk = 0 in the GEV distribution yieldsa test
tor of the sign of the shape parameterof the GEV statisticZ = 1.00. This value is not significantand
distribution. suggeststhat the Nidd data may reasonablybe as-
7. sumedto come froma Gumbeldistribution.
EXAMPLE
As an example we fitextreme-value 8. CONCLUSIONS
distributions
to the 35 annual maximumfloodsof the riverNidd Estimators of parametersand quantiles of the
at Hunsingore,Yorkshire,England. The data are GEV distributionhave been derived using the
taken fromNERC (1975b, p. 235). In Figure 5 the methodof probability-weighted moments.These esti-
ordereddata values xl < ... < xnare plottedagainst matorshave severaladvantagesover existingmeth-
the corresponding Gumbel reduced variates ods of estimation.They are fastand straightforward
-log (- log Fi), i = 1, ..., n, whereFi = (i - .44)/(n to computeand always yieldfeasiblevalues forthe
+ .12) is the Gringorten(1963) plottingpositionfor estimatedparameters.The biases of the estimators
the ith smallestof n observationsfroma Gumbel are small, except when estimatingquantiles in the
distribution.The returnperiod of xi is 1 - 1/Fi. extremetails of the GEV distribution, and theyde-
Gumbel and GEV distributionswere fittedto the crease rapidlyas the samplesize increases.The stan-
data by the methodof PWM, and the estimatedpa- dard deviationsof the PWM estimatorsare com-
rameterswere as follows(figuresin parenthesesare parable withthose of the maximum-likelihood esti-
standarderrorsof estimatedparameters):Gumbel- matorsformoderatesamplessizes (n = 50, 100) and
= 108.6 (8.6), a = 48.5 (7.4); GEV- = 105.8 (8.2), are often substantiallyless than those of the
a = 42.5 (6.7), k = -.13(.14). The fitteddistributions maximum-likelihood estimatorsfor small samples

500
Table 8. EmpiricalSignificanceLevels of the StatisticZ for
Testingthe HypothesisH: k = 0 Against
One -Sided and Two -Sided Alternatives 4 400
u
Alternatives

k <0 k> 0 k 0

Sample Size 10%* 5%* 10%* 5%* 10%* 5%*

15 10.3 4.3 7.3 3.7 8.0 3.5 I

25 10.4 4.6 8.4 4.3 8.9 4.1 < - _- 2 5 10 20 50 100
Reburn period
50 10.5 4.9 8.9 4.6 9.6 4.7 0
100 I I I I I
10.4 5.1 9.4 4.9 10.0 5.1 -2 -1 0 1 2 3 4 5
Gumbel reduced vor lobe
200 10.4 5.0 9.7 5.1 10.2 5.2
500 10.5 5.3 9.6 4.9 10.2 5.1 Figure 5. DistributionsFittedby the Method of PWM to 35
Annual MaximumFloods of the RiverNidd: GEV distri-
*Nominallevel.
bution; ----- Gumbeldistribution.

TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
 ESTIMATIONOF GENERALIZEDEV DISTRIBUTION 259

(n = 15, 25). PWM estimatorsof GEV parameters are both positive, so (B.1) reduces to bo - 4b,
and quantileshave asymptoticNormal distributions, + 3b2 < 0. But we can write
and the large-sampleapproximationto the variance
2
oftheestimatorsis adequate forsamplesizes of 50 or bo -o 4b1 -+ 3b2 = n(n- 1)(n- 2) (-x
(i + xk);
k);
i>>k
more.AlthoughPWM estimatorsare asymptotically
inefficientcompared to maximum-likelihoodesti-
mators,no inefficiency is detectablein samplesofsize (B.4)
100 or less.The PWM estimatoroftheshape param- thusbo - 4b1 + 3b2 < 0 almostsurelyand, therefore,
eter k of the GEV distributionmay be used as the k > -1 almost surely.Results(B.2)-HB.4)are easily
basis of a testof the hypothesisHo: k = 0, and this provedby inductionon thesamplesize n.
testis simpleto perform, and accurate.
powerful, Furthermore, since
ACKNOWLEDGMENTS a = (2b, - b0)k/{F(1+ k)(1 - 2-)} (B.5)
This cooperativeworkwas supportedby theInsti- - -
and 2bI bo > 0 as notedbefore,k/(1 2-k) > 0 for
tute of Hydrology,Wallingford,England. We are all k, and r(1 + k) > 0 because > - 1, it follows
greatlyindebtedto J. S. G. McCulloch (Directorof thatwe musthave c > 0.
the Instituteof Hydrology),P. E. O'Connell, and R.
T. Clarke forinitiatingand sustainingthis research
effort.
Part ofJ.R. M. Hosking'sresearchwas carried APPENDIX C: ASYMPTOTIC DISTRIBUTION
out at the MathematicsResearchCenter,University OF THE b,
of Wisconsin-Madison,with the support of U.S.
ArmyContractDAAG29-80-C-0041. The statisticb, may be writtenas a linearcombi-
nationoftheorderstatisticsofa randomsample: We
APPENDIX A: PROBABILITY-WEIGHTED have
MOMENTS FOR THE GEV DISTRIBUTION n

For theGEV distribution

we have from(2) and (3) br = n- 1 ()Xj, (C.1)
j-=l

fr = Ml,r,O where c- = (j - 1) ' (j - r)/{(n - 1) ... (n - r)}

and x < x < ... < x is the ordered sample. As
= [ + a{l -(-log F)k}/k]F'dF n- oo and j- oo withj/n- 0 (0 < 0 < 1), c{r is
o
asymptoticallya functionof the plottingposition
= + a(1 - uk)/k}e-(r+ 1)udu, j/(n + 1): in fact, c(r
{j/(n + 1)}'. It is straightfor-
wardto verify thatbrsatisfiestheconditionsofTheo-
rem 1 of Chernoffet al. (1967), and fromthat theo-
u = - log F,
substituting
remit followsthatbris asymptotically Normallydis-
tributedwithmean fr, and variance
=( + a/k) e -(r+1)udu- (a/k) uke%-(r+ )u du

fJ'
Jo Jo
n- v = 2n-1 {F(x)}r{F(y)}r F(x)
= ( + oa/k)(r+ 1)- -(a/k)(r + 1)- 1 k(l + k)
x<y -
x {1 F(y)} dx dy. (C.2)
providedthatk > -1,
A similar argumentapplies to any linear combi-
= (r + 1)-`[L + a{l - (r + 1)-kF(1 + k)}/k]. (A.1) nation of the br (r = 0, 1, 2, ...), and it followsthat
APPENDIX B: FEASIBILITY OF PWM the br are asymptotically jointly Normallydistrib-
ESTIMATES OF THE GEV PARAMETERS utedwithcovariancegivenby
The PWM estimatork satisfies(13), and therefore vrs = lim n cov (b,, bs) = j (grs+ sr), (C.3)
> --1 providedthat
where

{J
(2b - bo)/(3b2- bo)> 2. (B.1)
Now grs = 2 {F(x)}r{F(y)}S F(x){l - F(y)} dx dy.
x<y (C.4)
2b - bo n(n-n( 1) (x - ) (B.2)
)i>j To evaluate the g,s for the GEV distributionwe
and considerfirstthecase k > 0 and let

Irs = 2 {F(x)}r{F(y)}s dx dy (C.5)

^3b2-bo (B.3)
n(n- 1)(n-2) i>j>k
x<y

TECHNOMETRICS,AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
260 J. R. M. HOSKING,J. R. WALLIS,AND E. F. WOOD

so that The resultsstated in this Appendixare valid for

- arbitrarypositiveintegersr and s, thoughonly the
grs = I,+ l.s r+ 1s+ 1 (C.6)
cases r, s = 0, 1, 2 are required for derivingthe
Substituting(7) in (C.5) and makingthe furthersub- asymptoticdistributionsofPWM estimators.
stitution u = {1 - k(x - )/a}1/k, V = [ReceivedJune1984.RevisedDecember1984.]
{1 - k(y- )/} 1/k,we have
REFERENCES
rs = 22 | Uk-
le-ru du . vk- le-sv dv Cheng,R. C. H., and Amin,N. A. K. (1983),"EstimatingParame-
ters in Continuous Univariate DistributionsWith a Shifted
Origin,"JournaloftheRoyal StatisticalSociety,Ser. B, 45, 394-
= 403.
2x2r-k vk- le-Sv(k, rv) dv
Chernoff, H., Gastwirth, J.L., and Johns,M. V. (1967),"Asymptotic
Distributionof Linear Combinationsof Functions of Order
= a2k-2(r + ) -2kFr( + 2k) StatisticsWithApplicationsto Estimation,"AnnalsofMathemat-
ical Statistics, 38,52-72.
x 2F1[1, 2k; 1 + k; s/(r+ s)] (C.7) Downton, F. (1966a), "Linear EstimatesWith PolynomialCoef-
ficients," Biometrika, 53, 129-141.
(Gradshteynand Ryzhik1980,pp. 317 and 663); here
(1966b), "Linear Estimatesof Parametersin the Extreme
Fr(, *) is the incompletegamma functionand 2F1 is Value Distribution," Technometrics,8, 3-17.
the hypergeometric function.It is convenientto Fisher,R. A., and Tippett,L. H. C. (1928),"LimitingForms of the
transform thehypergeometric
functionin (C.7), using FrequencyDistributionof the Largestor SmallestMemberof a
resultsfromGradshteynand Ryzhik(1980,p. 1043): Sample,"ProceedingsoftheCambridgePhilosophicalSociety,24,
180-190.
Fraser,D. A. S. (1957),Nonparameteric MethodsofStatistics,New
2F,[1, 2k; 1 + k; s/(r+ s)] York: JohnWiley.
= 22k{F(l + k)}2/Fr( + 2k) ifr = s, Gradshteyn,I. S., and Ryzhik,I. W. (1980), Table of Integrals,
Seriesand Products(4thed.),New York: AcademicPress.
= {r/(r+ s)}-2kG(s/r) ifr > s, Greenwood,J.A., Landwehr,J. M., Matalas, N. C., and Wallis,J.
R. (1979),"ProbabilityWeightedMoments:Definitionand Rela-
= -{s/(r + s)} -2kG(r/s) + 2r-s-k(r + s)2k tion to Parametersof Several DistributionsExpressablein In-
verseForm,"WaterResourcesResearch,15,1049-1054.
x {rF(l + k)}2/F(1 + 2k) ifr < s, (C.8) Griffiths,D. A. (1980), "Interval Estimation for the Three-
ParameterLognormal Distributionvia the Likelihood Func-
whereG denotesthehypergeometric functionG(x) = tion,"AppliedStatistics, 29,58-68.
2F,(k, 2k; 1 + k; -x); note that G(0) = 1. Substitut- Gringorten, I. I. (1963),"A PlottingRule forExtremeProbability
ing back into (C.6) and (C.3) we obtain thefollowing Paper,"JournalofGeophysical Research,68,813-814.
expressionsforthev,s': Hoeffding,W. (1948), "A Class of StatisticsWith Asymptotically
Normal Distribution,"Annals of MathematicalStatistics,19,
293-325.
Vrr= C2k-2(r + 1)-2k[r(1 + 2k)G{r/(r+ 1)} Hosking,J.R. M. (1984),"TestingWhethertheShape ParameterIs
Zero in the Generalized Extreme-ValueDistribution,"Bio-
- {r(1 + k)}2], (C.9) metrika, 71,367-374.
Hosking, J. R. M., Wallis, J. R., and Wood, E. F. (1985), "An
Vr,+1 = ?c2k-2 [(r + 2)-2kr(l + 2k)G{r/(r+ 2)}
AppraisaloftheRegionalFlood FrequencyProcedurein theUK
+ (r + 1)-k{(r + 1)-k - 2(r + 2)-k}{r(1 + k)}2], Flood Studies Report,"HydrologicalSciencesJournal,30, 85-
109.
(C.10) Jenkinson, A. F. (1955),"The FrequencyDistributionoftheAnnual
Maximum (or Minimum)of MeteorologicalElements,"Quar-
v,,r+s=ia2k-2[(r+s+ 1)-2kF(1 +2k)G{r/(r + s+ 1)} terlyJournaloftheRoyalMeteorological Society,81, 158-171.
- (r + s)-kF(1 + 2k)G{(r + 1)/(r+ s)} + 2(r + 1)-k (1969), "Statisticsof Extremes,"TechnicalNote 98, World
MeteorologicalOffice, Geneva.
x {(r + )-k- s > 2. Landwehr,J. M., Matalas, N. C., and Wallis,J. R. (1979), "Prob-
(r + s + 1)-k}{r(1 + k)}2],
abilityWeightedMomentsCompared With Some Traditional
(C.11) Techniquesin EstimatingGumbel Parametersand Quantiles,"
WaterResourcesResearch,15,1055-1064.
When k < 0 the foregoingargumentis not valid Locke, C., and Spurrier,J. D. (1976),"The Use of U-Statisticsfor
because theintegralin (C.5) does not converge.How- TestingNormalityAgainstNon-Symmetric Alternatives,"Bio-
metrika, 63, 143-147.
ever, the expressions(C.9)-(C.11) are analyticfunc- Natural EnvironmentResearch Council (1975a), Flood Studies
tionsof k forall k > - ?; henceby analyticcontinu- Report(Vol. 1),London: Author.
ation expressions(C.9HC. 11) are valid solutionsof (1975b),Flood StudiesReport(Vol. 4), London: Author.
theintegralrepresentation Prescott,P., and Walden,A. T. (1980),"Maximum-Likelihood Esti-
(C.3HC.4) throughoutthe
domain - ? < k < oo. At the value k = 0, the v7,are mation of the Parametersof the Generalized Extreme-Value
Distribution," Biometrika, 67,723-724.
given by the limitsof (C.9)-(C. 11) as k-0; these (1983), "Maximum-LikelihoodEstimationof the Parame-
limitsare welldefined. tersof the Three-Parameter GeneralizedExtreme-ValueDistri-

TECHNOMETRICS, AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions
 ESTIMATION OF GENERALIZED EV DISTRIBUTION 261

bution From Censored Samples," Journalof StatisticalCom- Shenton,L. R., and Bowman,K. 0. (1977), Maximum-Likelihood
putingand Simulation, 16,241-250. Estimation
inSmallSamples,London: CharlesW. Griffin.
Randles,R. H., and Wolfe,D. A. (1979),Introduction to theTheory Von Andrae(1872),"Uber die Bestimmung das Wahrscheinlichen
ofNonparametric New York: JohnWiley.
Statistics, FehlersDurch die GegebenenDifferenzenvon ein Gleich Ge-
Rao, C. R. (1973), Linear StatisticalInferenceand Its Applications nauen BeobachtungenEiner Unbekannten,"Astron.Nach., 79,
(2nded.),New York: JohnWiley. 257-272.

TECHNOMETRICS, AUGUST 1985, VOL. 27, NO. 3

This content downloaded by the authorized user from 192.168.52.72 on Sun, 9 Dec 2012 10:52:26 AM
All use subject to JSTOR Terms and Conditions