# ORIGINAL PAPER

D. L. Fitzgerald
Analysis of extreme rainfall using the log logistic distribution
Published online: 3 February 2005
Ó Springer-Verlag 2005
Abstract Computer-intensive methods are used to
examine the ﬁt of the log logistic distribution to annual
maxima of Irish rainfall. The characteristics of the
L-moment solutions are examined by using the con-
ventional bootstrap on the data and by random sam-
pling within the ellipse of concentration of the parameter
estimates. A statistical method of examining uncertainty
is provided by the maximum product of spacings solu-
tion. Factors derived from random division of an
interval are proposed for estimation of short-duration
falls for which no data are available.
Keywords Log logistic distribution Æ L-moment
estimate Æ Maximum product of spacings method Æ
Bootstrap Æ Ellipsoid of concentration Æ
Geometric max-stable distributions
Introduction
In the Flood Estimation Handbook (1999) the gener-
alised logistic distribution is preferred to the general-
ised extreme value distribution for ﬂood estimation in
the UK. The log logistic (LLG) distribution is a re-
parameterisation of the generalised logistic. Its three-
parameter version (3pLLG) was employed in a study
of river ﬂows in Scotland (Ahmed et al. 1988) but, in
spite of its theoretical attractiveness it has been little
used in more recent ﬂood or extreme rainfall studies.
The L-moment solution of its three parameters is
examined and then recast by ﬁxing the location
parameter; then the shape and scale parameters have a
clear interpretation in terms of the data and, more
importantly, their maximum likelihood (ML) and
L-moment estimates coincide. Pooling L-moment
solutions is treated in detail in Hosking and Wallis
(1997). In this paper the focus is on examining the
stability of the L-moment quantile estimates at indi-
vidual stations under variation of the data and of the
parameters. The statistical tools used are: (1) the
conventional bootstrap to sample the data with
replacement (2) the ellipsoid of concentration (Crame` r,
1946) which gives a method of randomly varying the
parameters within bounds set by their covariance
matrix. (3) The maximum product of spacings solution
(Cheng and Amin 1983); for a sample of size n (x
1
, x
2
,
..., x
n
) it is equivalent to the ML solution of a more
extreme sample of size (n+1) with x
1
¢
< x
1
, x
2
¢
< x
2
,
....x
n
¢
< x
n
, x
n+1
¢
> x
n
.
The log logistic distribution
This is of form
f ðxÞ ¼
b
a
ðx ÀgÞ=a ð Þ
bÀ1
1 þðx Àg=aÞ
b
_ _
2
; x > g a; b > 0 ð1Þ
and b ln [(x À g)/a] has a logistic distribution of mean 0.
Eq. 1 implies that the cumulative distribution function
is:
F ðxÞ ¼
ðx Àg=aÞ ð Þ
b
1 þ x Àg ð Þ=a ð Þ
b
_ _ ð2Þ
This shows it as a special case of a Burr XII distri-
bution (Johnson and Kotz 1970). Its inverse is given by
x ¼ g þa
F ðxÞ
ð1 ÀF ðxÞÞ
_ _
c
¼ g þaðT À1Þ
c
c ¼ 1=b > 0
ð3Þ
where T=1/[1 À F(x)] is the return period. It follows
from Eq. 1 to Eq. 3 that
D. L. Fitzgerald
Met E
´
ireann, Glasnevin Hill, Dublin 9, Ireland
E-mail: denis.ﬁtzgerald@met.ie
Fax: +353-1-8064247
Stoch Environ Res Risk Assess (2005) 19: 249–257
DOI 10.1007/s00477-004-0229-x
Median ¼ g þa; mean ¼ g þ
acp
sinðcpÞ
;
mode ¼ g þa
1 Àc
1 þc
_ _
c ð4Þ
To a good degree of approximation mode =
2(median)Àmean for values of c less than 0.4; this covers
all but the most extreme growth rates. Also from Eq. 2
c ¼ 1 À2F ðmodeÞ ð5Þ
The shape parameter c is thus the measure of skew-
ness proposed in Bickel(2002) and using a plotting po-
sition formula we can get a ﬁrst estimate of c. As Eq. 4
shows mode < median < mean and the skewness is
positive.
Since b ln((x À g)/a) has a logistic distribution of
mean 0 the scale parameter a can be estimated as the
geometric mean of (xÀg) for the sample.
L-moment estimates of the parameters
The theory of L-moments and of the closely related
probability weighted moments (PWM) is given in
Hosking and Wallis (1997). With
l
1
¼ EðxÞ; l
2
¼ 2EðxF Þ ÀEðxÞ;
l
3
¼ 6EðxF
2
Þ À6EðxF Þ þEðxÞ
we get from Eq. 3
c ¼
l
3
l
2
¼ L - moment measure of skewness ð6Þ
a ¼ l
2
sinðpcÞ
p c
2
ð7Þ
g ¼ l
1
À
l
2
2
l
3
ð8Þ
When g is known the two-parameter log logistic theory
applies to (x-g) giving
c ¼
l
2
l
1
¼
l
3
l
2
ð9Þ
i.e. the shape parameter = the L-moment coeﬃcient of
variation = the L-moment coeﬃcient of skewness
(Fitzgerald 1996).
Since x>g, the location parameter might, ideally, be
expected to be a realistic estimate of the minimum of the
annual rainfall maxima; in some cases it is but more
often it is appears low and on occasion quite acceptable
estimates of rainfall quantiles occur with a negative
value of g.
The maximum likelihood (ML) parameter estimates
For a sample of size n the ML equations may be written
c ¼
1
n

i¼n
i¼1
ð2F ðx
i
Þ À1Þ lnðx
i
ÀgÞ ð10Þ
0 ¼

i¼n
i¼1
ð2F ðx
i
Þ À1Þ ð11Þ
c ¼

i¼n
i¼1
1 À2F ðx
i
Þ= x
i
Àg ð Þ ð Þ

i¼n
i¼1
1=x
i
Àg ð Þ
ð12Þ
When g is known Eqs. 10 and 11 are the two-
parameter ML equations. In the three-parameter case
a value of the location parameter g was taken, e.g. the
L-moment estimate, and Eq. 10 and Eq. 11 used to
estimate the shape and scale parameters; in turn the
new values of c and a were used to get a new value of
the location parameter. While often successful, this
method did not always converge but occasional failure
Table 1 Parameter and quantile estimates for 2-day (0900-0900UCT) annual maximum rainfalls (mm)
Estimation method Parameters Return periods
c a g RP 2 RP 100 RP 500 RP 1000
L-moment 0.225 32.4 17.0 49.3 108.1 148.3 170.5
ML 0.288 25.5 23.5 49.0 119.2 176.0 209.7
BOOTSTRAP
Estimation method Statistic Return Periods
RP 2 RP 100 RP 500 RP 1000
L-moment Mean 49±2 107±11 146±23 168±31
Median 49.4 106.1 144.1 165.1
ML Mean 50±2 105±16 142±28 163±35
Median 49.7 103.7 139.1 158.1
Station: Ffoulkesmills
Years of record=62
Sample max=105.2; Sample min=35.5; Sample median=48.2
250
is to be expected from Eq. 10 to Eq. 12; the most
usual case of non-convergence was with g strongly
negative and c low.
For quantile estimation beyond the range of the
sample it is unwelcome that Eq. 10 gives high weight to
low as well as to high values of (x-g) in the estimation of
the crucial shape parameter.
The information matrix
I ¼ À
1
n
E ð@
2

_
@h
i
; @h
j
_ _
; #ðc; a; gÞ
is:
Here w is the digamma function.
Comparison of the L-moment and ML solutions
for the three-parameter LLG
Table 1 presents a typical case of a good ML ﬁt which
generally has a higher value of the crucial shape
parameter than the L-moment estimate. Bootstrapping
the data to generate 5,000 samples gave much-reduced
ML estimates at the higher return periods, closer to the
L-moment estimates which were largely unchanged.
However, nearly half the bootstrap samples did not
converge in a fully satisfactory manner and so all that
can safely be gleaned from this exercise is that the ML
solution is not stable under bootstrapping. Only with 12
of the 10,000 L-moment solutions was it considered
desirable to use the logistic instead of the LLG. Failures
tended to occur in samples having a mean less than the
median and hence unsuited to an LLG ﬁt. Typically, for
such samples the ML solution did not converge or
produced estimates which were distinctly low; the L-
moment solution produced low or even negative values
of the location parameter but usually acceptable esti-
mates of the rainfall quantiles.
Table 2 provides an example of an L-moment solu-
tion where the location parameter looked implausible
and ML solution converged smoothly.
Note that the sample mean is less than the median
and we may anticipate diﬃculty in ﬁtting the LLG dis-
tribution. The three parameter L-moment solution has a
very low value of the location parameter. The ML
solution converged with a distinctly more realistic value
of g. Bootstrapping provided L-moment estimates still
well in accord with the quantile estimates and problems
occurred with less than 12% of the 10,000 samples. As in
Table 1 the ML bootstrap estimates reduce at high re-
turn periods and have a higher variance than the
L-moment estimates. Again there is a strong element of
subjectivity in the ML bootstrap results since there were
convergence problems with nearly 30% of the 5,000
samples.
It seems clear that the L-moment estimates are the
more stable. From the bootstrap means and standard
errors it would also seem that they give fairly realistic
estimates of the 1,000-year rainfall quantile for an
individual station with about 50 years of record.
The 2p/3p solution
From (1) setting the location parameter g at some ﬁxed
value is an attractive option. For the two-parameter
LLG the requirement that the L-moment coeﬃcient of
Table 2 Parameter and Quantile estimates for 2-day (0900-0900UCT) Annual Maximum Rainfalls(mm)
Estimation Method Parameters Return Periods
c a g RP 2 RP 100 RP 500 RP 1000
L-moment 0.151 58.9 04.5 63.9 122.2 154.7 171.2
ML 0.226 39.6 23.3 62.9 135.3 184.7 212.2
BOOTSTRAP
Estimation Method Statistic Return Periods
RP 2 RP 100 RP 500 RP 1000
L-moment mean 64±3 121±12 153±24 169±32
median 63.9 120.5 151.7 167.6
ML mean 64±3 125±17 162±37 182±49
median 63.6 122.9 157.5 175.6
Station: Brosna (Mt Eagle)
Years of record=55
Sample max=118.2; min=38.4; median=66.3; mean=65.7
p
2
þ3
9c
0
Àpcð3cþð1Àc
2
Þðwð2ÀcÞÀwð2þcÞÞÞ
3ac
2
0
1
3a
2
c
2
ð1Àc
2
Þpc
3a
2
c
2
Àpcð3cþð1Àc
2
Þðwð2ÀcÞÀwð2þcÞÞÞ
3ac
2
ð1Àc
2
Þpc
3a
2
c
2
2pcð1Àc
2
Þ
3a
2
c
2
_
¸
¸
_
_
¸
¸
_
ð13Þ
251
variation (l-cv) and the L-moment coeﬃcient of skew-
ness (l-sk) be equal is a key test of ﬁt. A second test is
that the scale parameter should equal the geometric
mean of (x-g). From a plot, the mode of a typical 30- to
60-year sample can usually be seen to fall within a cer-
tain range; the estimates of the mode provided by
2(median) À mean and Eq. 4 should agree closely and
both should fall within the indicated range. For annual
maxima of Irish rainfall the ﬁt with g=0 is poor.
However, a non-zero value of g can be found which
gives a good ﬁt. In the course of systematically varying g
it was found that:
(a) The value of g obtained when l-cv = l-sk was never
appreciably diﬀerent from that of the three-param-
eter L-moment solution.
(b) The l-sk estimate of c was nearly constant as g
varied.
(c) The l-cv and ML estimates of c were broadly similar
and both varied systematically with g.
(d) A good ﬁt with a common value of g could be found
where the ML and the L-moment solutions for the
shape and scale parameters essentially coincided;
these values of g, a, c were in all cases so close to the
three-parameter L-moment solution that there was
no practical diﬀerence between their quantile esti-
mates.
Based on (d) the model for annual maxima adopted
was that the exceedances of a ﬁxed, if initially unknown,
threshold g follow a two-parameter LLG distribution; it
is convenient that this amounts to the three-parameter
L-moment solution with its location parameter g re-
interpreted. An advantage is that the two-parameter ML
theory now applies and Eq. 13 simpliﬁes to give for a
large sample of size n
varðcÞ ¼
9c
2
nðp
2
þ3Þ
; varðaÞ ¼
3a
2
c
2
n
;
covarianceða; cÞ ¼ 0
ð14Þ
Using Eqs. 3 and 14 the variance of the rainfall with
return period T is:
varðx
T
Þ ¼ ðx
T
ÀgÞ
2
varðcÞðlnðT À1ÞÞ
2
þ
varðaÞ
a
2
_ _
ð15Þ
Ellipsoids of concentration and random variation
Crame` r (1946) shows that for a random variable n (n
1
, n
2
,..., n
n
) of known covariance matrix an ellipsoid of
concentration can be deﬁned with the property that
randomly choosing a large number of points within the
ellipsoid yields a random variable with the same mean
and covariance matrix as n. The classical assumption is
of a parent population distribution with the parameters
estimated to a degree of accuracy determined by the
sample size n; the ellipsoid of concentration re-expresses
this in terms suited to computer simulation by deﬁning a
volume in parameter space within which all points can
be taken as equally likely. Belonging to the population
may then be taken as being within the ellipsoid of con-
centration and a random sample of n regarded as n
random choices within the ellipsoid.
In the case of the preferred solution for the LLG the
ellipsoid reduces to an ellipse having the parameter
estimates (a, c) as the mean of the random variable and
with covariance matrix given by Eq. 14. Since the ran-
dom variation is of the parameters, the eﬀect on the
mean, median and variability of the quantile estimates
Table 3 Comparison of rainfall estimates (mm)
RP Log logistic Log logistic Generalised extreme value
Years L M ML L-moment bootstrap Ellipse of concentration PWM ML PWM bootstrap
Mean SE Median Mean SE Median Mean SE Median
(Duration=1 clock hour Sample max=22.6 Sample min=7.2)
50 22.5 22.7 22.1 2.3 22.1 22.6 2.7 22.4 22.4 21.6 21.9 2.3 22.0
100 26.0 26.4 25.4 3.1 25.6 26.3 3.8 26.0 25.4 24.1 24.8 3.1 24.8
500 37.3 38.2 36.1 6.6 36.2 38.1 7.7 37.2 33.5 30.4 32.4 6.2 32.3
1000 43.9 45.2 42.4 9.0 42.3 45.1 10.2 43.7 37.7 33.4 36.4 8.2 36.1
(Duration=24 clock hours Sample max=94.0 Sample min=34.2)
50 94.1 101.6 92.8 7.3 92.9 94.4 8.2 94.0 92.4 91.4 91.0 7.4 91.2
100 104.5 116.1 102.9 9.8 103.0 104.9 10.5 104.4 99.7 98.7 98.2 9.7 98.3
500 133.0 159.6 130.7 18.9 130.0 133.9 17.3 132.8 116.0 115.0 114.5 16.8 113.6
1000 147.4 183.8 145.0 24.6 143.6 148.7 21.1 147.1 122.7 121.8 121.5 20.7 119.9
(Duration=96 clock hours Sample max=136.8 Sample min=53.8)
50 141.6 146.5 139.4 10.1 140.1 142.0 11.8 141.3 139.4 136.0 137.4 10.2 138.2
100 156.5 164.2 154.1 13.0 154.8 157.1 15.1 156.3 150.2 145.4 147.8 13.2 148.6
500 198.2 216.1 194.6 25.1 194.9 199.7 25.7 197.9 174.9 166.2 172.2 23.0 172.2
1000 219.7 244.2 215.8 32.8 215.2 221.8 31.7 219.2 185.3 174.8 182.9 28.4 181.9
Station: Cork Airport
Data: Annual maxima 1963–2002
L-M L-moment, PWM probability weighted moment, ML maximum likelihood, SE standard error, RP return period
Generalised extreme value: x=n+a(1 - (- ln F)
k
)/k
252
can be compared with the preferred solution estimates,
the three-parameter ML estimates and the values ob-
tained from bootstrapping the data and using the
L-moment solution. The results are presented in
Table 3. For purposes of comparison Table 3 includes
the probability weighted moment, the ML and the
bootstrap PWM estimates for the generalised extreme
value (GEV) distribution (Hosking and Wallis 1997).
The ellipse of concentration statistics are based on
20,000 sample points.
Remarks on Table 3
– The GEV displays a lower growth rate but even at the
1,000-year RP the estimates are within 1.5 standard
errors of the LLG estimates.
– Both the bootstrap and the ellipse of concentration
give broadly similar quantile estimates for the LLG
and both their medians and means agree well with the
L-moment values.
– The very close agreement between the median of the
ellipse of concentration estimates and the L-moment
values at all return periods allows the identiﬁcation of
the L-moment solution with a central value of the
well-deﬁned population based on the sample data.
Fit of the estimates to the top sample values
For some rough guidance on the return period T
m
associated with the higher order statistics of a sample of
size n note that the distribution of y=T
m
À 1 for the mth
order statistic of any distribution is given by
f ðyÞ ¼
1
Bðm; n Àm þ1Þ
y
mÀ1
ð1 þyÞ
Àðnþ1Þ
ð16Þ
where B is the beta function and f(y) an inverted beta
distribution.
From Eq. 16EðT
m
Þ ¼ n=ðn ÀmÞ ð Þ. The modal value
T
mode
¼ n þ1=ðn Àm þ2Þ ð Þ. The median can be ob-
tained from Eq. 16 using incomplete beta functions. A
restatement of Eq. 16 is that the CDF of the mth order
statistic F(x
m
) = y/ (1+y) has a beta distribution and
Jenkinson (1977) exploited this to show that its median
plotting position can be derived as
~
F ðx
m
Þ ¼
m À0:307
n þ0:386
.
In turn this gives the good approximation:
T
m
ðmedianÞ
~
T
m
¼
n þ0:386
n Àm þ0:693
ð17Þ
For the LLG the median value of x
m
is given by the
incomplete beta function expression
1
2
¼
B
y
m
=1þy
m
ðm; n Àm þ1Þ
Bðm; n Àm þ1Þ
where y
m
¼
ðx ÀgÞ
a
_ _
b
¼ T
m
À1
ð18Þ
and so, to the same approximation as in Eq. 17, we get
x
m
¼ g það
~
T
m
À1Þ
c
; c ¼
1
b
ð19Þ
As m approaches n Eq. 16 is a highly skewed distri-
bution. Hence the median seems by far the best central
value to select for the purpose of gauging the ﬁt of any
distribution by examining the quantile estimates for the
top sample values.
Remarks on Table 4.
– For both distributions the median plotting position
is used to make the estimates. The ﬁt to the sample
values is good for 1 h and 24 h durations; for 96 h
the ﬁt is greatly improved by taking the fourth
highest value (in brackets) into account. While it
might be observed that they agree more closely with
one another than with the data and that the GEV
Table 4 Comparison of top three sample values with median estimates
Duration Distribution
Clock hours Log logistic Generalised extreme value Sample
Parameter Estimates Parameter Estimates
01 c 0.272 23.3 k À0.153 23.0 22.6
a 5.9 19.6 a 2.2 19.5 21.3
g 5.6 17.7 n 10.6 17.9 18.5
24 c 0.148 96.5 k +0.035 94.2 94.0
a 53.4 84.1 a 12.2 84.4 87.3
g À0.80 78.1 n 8.0 79.1 78.7
96 c 0.163 145.4 k +0.011 141.8 136.8
a 65.4 127.7 a 16.2 127.7 130.2
g 18.8 119.4(113.9) n 77.5 120.2(115.0) 129.3(115.7)
Station: Cork Airport
Data: annual maxima 1963–2002
Generalised extreme value x=n+a (1 À (Àln F)
k
)/k
253
appears to ﬁt the data a little better than the LLG,
the notion of the LLG estimates as median values is
supported.
– For 1 clock hour we have an instance of the location
parameter g as a plausible minimum while the mini-
mum for the GEV is À3.7. In general we may rea-
sonably expect the growth rate to decrease or remain
steady with increasing duration greater than 1 h or
2 h but, in practice contradictions between return
periods arise e.g. here Table 3 gives 147.4 as the L-
moment estimate of the 1,000-year rainfall quantile
for 24 clock hours while the 18-h value is 149.3. Ta-
ble 4 shows that this is caused by the negative value of
the location parameter which forces a high value of
the scale parameter and a low value of the shape
parameter. The GEV estimates suﬀer from the same
lack of consistency as they give a 1,000-year rainfall
quantile of about 122 mm for 24 clock hours and an
18-h value of over 125 mm. This might fairly be taken
as an indication of the inadvisability of estimating
1,000-year quantiles from forty data points but it is
necessary to do so while imposing consistency over
durations and return periods. The ML solution for the
3-parameter LLG converges to give 0.226, 35.1 and
17.1 as the estimates for c, a, g respectively and gives
an 1,000-year quantile of 184 mm. However, its esti-
mate of the 50-year quantile as 102 mm appears a
little high as the median RP associated with the 40-
year sample maximum of 94 mm is over 58 years. The
location parameter may be considerably increased
without much change to the estimates in Table 3 and
in Table 4. The method, for a given value of the
location parameter, is to ﬁrst take the mean of the two
estimates of the shape parameter in Eq. 9 and to use
that in Eq. 7 to determine the scale parameter. Of
itself this gives RP estimates quite close to the pre-
ferred solution estimates. Taking g=15 we get esti-
mates of 0.177 and 37.6 for the shape and scale
parameters while the two-parameter ML solution
gives 0.212 and 37.3. Giving the two solutions equal
weight we get estimates of 95, 107, 141 and 159 for
Table 3 and 97.6, 84.1 and 77.7 for Table 4, thus
ironing out the contradiction between 18 h and 24 h
estimates and retaining a decent ﬁt to the top sample
used instead of or in conjunction with a depth-dura-
tion-frequency scheme [Flood Estimation Handbook
(1999), vol 2, Chap. 10].
Maximum product of spacings (MPS) solution
as a statistical safety factor
For parameter estimation of univariate distributions the
MPS method of maximising the geometric mean of the
sample spacings was developed by Cheng and Amin
(1983) as an alternative to maximum likelihood. It was
shown in Fitzgerald (1996) that the MPS method
applied to a sample of size n is equivalent to using ML
for a sample of size (n+1) where x
1
¢
< x
1
, x
2
¢
< x
2
, ..., x
n
¢
< x
n
, x
n+1
¢
> x
n
. It was further shown for the two-
parameter LLG that MPS gives a positively biased
estimate of the shape parameter while the scale param-
eter is essentially equal to the ML or L-moment esti-
mates. The location parameter of the preferred solution
of Sect. 2.4 is accepted for the MPS; then, the median
estimate of the two solutions is the same but for higher
return periods MPS gives more extreme quantile esti-
mates.
Maximum product of spacings chooses the sample
spacings closest to a perfectly regular pattern and
hence is a typical index of lack of clustering or
ignorance. Compared with the estimates of the pre-
ferred solution it produces quantiles that are in line
with climate change scenarios for mid latitudes, the
amount of the increase being determined by what
amounts to an application of the theorem of the mean
to the sample values. Table 5 shows the results of
using MPS estimates of the shape and scale parame-
ters and their variances in an ellipse of concentration
exercise as outlined in Sect. 3; the MPS results are
very much along the lines of those found for the L-
moment solution. The 100-year rainfall quantile in-
creases from 90 to 107±12 mm in what we may
choose to regard as the new, median-conserving cli-
mate. The 1,000-year quantile increases from 157±25
to about 210±35; if the period 1941–2002 is used
instead of the record for 1881–2002 the values become
173±36 and 230±54 respectively.
Table 5 Log logistic estimates of rainfall quantiles
MPS ellipse of concentration
RP
(years)
L-moment MPS Mean SE± Median
2 34.2 34.3 34.3 1.1 34.3
5 44.5 46.5 46.5 2.1 46.4
10 52.5 56.5 56.6 3.2 56.5
20 61.7 68.4 68.5 4.9 68.2
50 76.4 88.0 88.3 8.2 87.8
100 90.1 106.9 107.4 11.7 106.7
250 112.5 139.0 140.0 18.3 138.6
500 133.5 170.2 171.8 25.4 169.6
1,000 158.9 208.9 211.4 35.0 208.1
Station: Phoenix Park (Dublin)
Data: 1-day annual maxima 1881–2002
MPS Maximum product of spacings, RP return period, SE stan-
dard error
Table 6 Means and medians of maxima of the random division of
the unit interval
Number of intervals
2 3 4 5 6
Mean 0.75 0.61 0.52 0.46 0.41
Median 0.75 0.59 0.50 0.44 0.39
254
Why use the log logistic distribution?
In the Flood Studies Report (1975) the equation for the
rainfall growth curves is given by: ln
M
T
M
5
¼
cðy À1:5Þ; M
T
¼ T - year return period and y ¼
ÀlnðÀlnðF ðxÞÞ % lnðT À0:5Þ; T > 5
Here c= - k where k is the shape parameter of the
GEV distribution and c ranges from 0.23 to 0.07.
For the LLG the corresponding growth curve is of
strikingly similar form:
ln((x
T
À g)/(x
5
À g)) = c(ln(T À 1) À 1.4) where c is
the LLG shape parameter.
The LLG distribution also exhibits geometric-ex-
treme stability in the sense of Voorn (1987) i.e. if F(x) is
a continuous distribution independent of N, a distribu-
tion on the positive integers with generating function
Q(s), then F is maximum stable with respect to N if real
numbers a and b exist such that
QðF ðxÞÞ ¼

1
k¼1
p
k
F ðxÞ
k
¼ F
x Àa
b
_ _
; b > 0 ð20Þ
If the number of rainfall events capable of producing the
annual maximum is taken as geometric with generating
function Q(s) = (ps)/(1 À (1 À p)s)i.e. p
k
=p(1 À p)
k - 1
,
k=1,2, ... and the total rainfall from the individual
events is LLG(c, a, g) then
QðF ðxÞÞ ¼ F
x
u
_ _
; u ¼ p
Àc
; u > 1 ð21Þ
Hence over years with random numbers, N(p), of
events with an LLG distribution the maximum is an
LLG distribution of a rescaled variable.
However, the geometric distribution used here has
mean 1/p but its mode is 1 and this seems low for Irish
conditions as most years are nondescript, having several
events capable of yielding the maximum. If instead we
consider a series of annual maxima having an LLG
distribution and, after the manner of Mitnik and Rachev
(1993), take the number of years into the future that the
process is likely to remain free of radical change as a
geometric variable N(t), then the argument applied
above gives the CDF of the maximum of a sample of size
N(t) as LLG, most plausibly if p is small. Here t is the
rather ill-deﬁned and subjective probability of ‘major
change in any year’ and N(t) the waiting time to the
change. We might estimate t from our degree of belief in
the length of time the ellipse of concentration generated
from the LLG ﬁtted to the annual series will adequately
represent the process generating the maxima. A belief
that there was only a probability of 0.5 that the current
model would apply after 50 years would imply t=0.013
and for a shape parameter c=0.20 this yields a scaling
factor of 2.38 to allow for the uncertainty in the length
of the period. The MPS method deals with uncertainty
by increasing the shape parameter while the method of
(geometric) random sample size alters the scale param-
eter of the LLG distribution. However, the point of most
interest is that under both these randomisations the
LLG (and also the logistic distribution) is maximum
stable. If the negative binomial is used instead of the
geometric we get
QðF ðxÞÞ ¼
pF ðxÞ
1 Àð1 ÀpÞF ðxÞ
_ _
r
¼ F
x
u
_ _
r
,
a Burr III distribution (Johnson and Kotz 1970); this
may be viewed as LLG with a rescaling of the data and
of the time unit, especially if r is an integer since in this
case r bears the interpretation of the minimum number
of events in unit time.
Experience dictates that when choosing a probability
distribution for rainfall extremes it is wise to allow for
much higher values than appear even in a long-term
record. This leads to the choice of long-tailed distribu-
tions with shape and scale parameters. Among these the
LLG has the advantage that, as Eqs. 4 and 5 show, it
can be completely determined by three central values,
leading to the hope that population characteristics may
be accurately determined from a sample. Unfortunately
the sample sizes in this study are rather small and the
mode is not well-determined but may be inferred from
the mean and median. The ellipse of concentration
argument in Sect. 3 shows that the preferred solution
closely mimics the median of a plausible representation
of the population giving rise to the data. If, further, the
parameter estimates of the preferred solution yield close
approximations to the sample estimates of the median
and mean then its ﬁt to the LLG is regarded as good;
this solution is computationally inexpensive and the
simplicity of the underlying theory makes it easy to
investigate its properties either directly or by computer-
intensive methods.
A point to emphasise is that the ﬁt of the pure two-
parameter LLG to the rainfall data was mostly poor and
the move away from g=0 crucial to obtaining a good ﬁt.
Unfortunately for their applicability to this study, all the
recent papers on the LLG in hydrology that have come
to my attention have been on the pure two-parameter
case. An example of the latter, remarkable for the depth
of negative feeling expressed towards the 2pLLG, is the
paper of Rowin´ ski et al. (2002). They concentrate on the
analysis of the tail region and for this study the most
pertinent of their many reservations is the necessity to
restrict the ranges of the parameters if population mo-
ments are to be kept ﬁnite; the practical implications of
this are most easily discussed in terms of the following
equation derived from Eq. 3:
E
x Àg
a
_ _
r
_ _
¼ Bð1 þrc; 1 ÀrcÞ;
r ¼ 1; 2; 3; :::; 0\c\1,
B being the beta function.
In this form the fractional contribution to the pop-
ulation moment of the range 0–F
À 1
(u) is given by an
incomplete beta function. Also for a ﬁnite third moment
we must have a shape parameter less than one-third and
255
so on. This sort of restriction on the range of the shape
parameter constitutes a diﬃculty for full tail analysis.
For statistical practice, it is important to appreciate that
with c=0.33 about 90% of the value of the third mo-
ment is contributed by values with return periods of
10
5
years(time units) or more; with c=0.333 the ﬁgure is
99%. Thus, the approach of the population moment of
order three to inﬁnity is mediated by values very unlikely
to appear in samples. Clearly, in parameter estimation it
would be unwise to employ third order moments. The
restriction thus points up the advantages of the L-mo-
ment solution.
Values of the shape parameter c encountered in this
study range from 0.1 to 0.33, with values of 0.33 to 0.4
occasionally appearing for short durations where growth
rates are steeper. At c=0.4 the population mean receives
a 12% contribution from values with return periods of
50 years or more while for 100 years and 1,000 years the
ﬁgures are 8% and 2% respectively. For E(xF
r
) these
percentages rise only slowly with r . The indication is
that, while in no way a drastic problem even at c=0.4,
low values of the shape parameter can be more accu-
rately determined from a sample than high. In addition
Eqs. 14 and 15 show how the variance of the quantile
estimates increases with c. Values of c>0.33 were trea-
ted with caution and, insofar as possible, checked for
consistency against higher and lower durations, as in the
next section.
Short-duration falls and random division
of the unit interval
The sub-hourly falls available are predominantly from
Dines rainfall recorders (HMSO 1981). Durations from
15 min to 24 h have been extracted from 42 stations.
Fitting the LLG sometimes gives high growth rates, with
values of the shape parameter in excess of 0.35, espe-
cially for durations ranging from 15 min to 3 h; check-
ing for consistency of the quantile estimates raised the
question of the possibility of ﬁnding scaling factors be-
tween durations. More importantly, urban drainage re-
quires design values for durations less than15 min and
for such short durations no reliable long-term data are
available. When the sub-period can be considered part
of a single event the theory of the random division of an
interval can contribute plausible scaling factors. The
distribution of the maximum of n intervals on (0, 1)
(David 1981) gives the following.
That the factors are signiﬁcant can be seen by
checking them against the 5-year return period factors
in Table II of Logue (1975) or Tables 3.6 and 3.7, vol
2, Flood Studies Report (1975), especially the mid-
table values. However, their derivation as central val-
ues of the distribution of a maximum suggests that
they apply more widely. Since, for example, it must be
possible for more than 52% of the total to fall in one
quarter of the time period the factors should apply
better at low than at high return periods. Over Ireland
the 500-year quantile for 4-h falls is usually less than
80 mm; it is possible to get that amount in 1 h (Ro-
han 1986) and so, in practice, high may mean about
1,000 years for some durations. In the particular case
of the LLG using the factor to adjust the scale and
location parameters and assuming the shape parameter
is unchanged is equivalent to directly multiplying the
estimates by the scaling factors; this may be attributed
to the fact the parameters are determined by the
median, mode and mean and these three may rea-
sonably be taken to scale in the same way. If there is
a diﬀerence Dc between the shape parameters of the
two durations it appears as a multiplier (T À 1)
Dc
when estimating the T-year quantile; however, in the
absence of data for the shorter duration the assump-
tion that the shape parameter is constant seems a
reasonable ﬁrst approximation. Table 7 presents the
estimates of 5-min quantiles from 30- and 15-min
values.
Note that the 15-min estimates and 0.75 times the 30-
min estimates agree to within one standard deviation
even at the1,000-year return period, the scaled down 30-
min values being systematically higher. The consistency
between the 5-min estimates is heartening but their
accuracy is diﬃcult to gauge. It has been standard
practice at Irish synoptic stations to extract daily max-
imum rates for 5 min but, as Logue (1975) points out,
the Dines recorder is not accurate below about 15 min.
Maximum values for 5 min from Dines recorders range
from 12 mm to about 8 mm over a period of record
mostly of 40–50 years. Fairly near Rosslare there was a
remarkable fall of 38 mm in 12 min in August, 1949
(Rohan 1986); this strongly suggests that 5-min falls of
over 20 mm occur in the general area but attaching a
return period at a single point must be labelled a hopeful
exercise and it seems a good idea to employ a region-
alisation scheme to supplement single station analysis.
Table 7 Five-minute quantile estimates from 15 min and 30 min
values
LLG estimates Table 6 estimates
RP 15 min 30 min 5 from 15 5 from 30
Years mm mm mm mm
2 7.2±0.4 9.8±0.6 4.4 4.0
5 9.4±0.6 12.9±0.9 5.7 5.3
10 11.1±0.8 15.5±1.2 6.8 6.4
20 13.1±1.1 18.5±1.5 8.0 7.6
50 16.1±1.7 23.3±2.2 9.8 9.6
100 18.9±2.3 27.9±2.9 11.5 11.4
250 23.4±3.5 35.3±4.5 14.3 14.5
500 27.5±4.8 42.4±6.4 16.8 17.4
1000 32.5±6.6 51.1±9.0 19.8 20.9
Station: Rosslare
Data: 45 annual maxima for 15 min and 30 min (millimetres)
Distribution ﬁtted: log-logistic with shape, scale and location
parameters (c, a, g)
15-Min: c=0.254, a=5.3, g=1.9; 30-min: c=0.285, a=6.7, g=3.0
256
Discussion and conclusions
The parameters of the log logistic distribution are
readily interpretable in terms of the data. Its form and
L-moment solution are such that it is well suited to
computer-intensive investigations. Bootstrap and ellipse
of concentration exercises prove very informative about
the stability of the quantile estimates. Curiously, having
gone to some pains to obtain a solution with attractive
statistical properties, it was found that the parameters so
obtained can be varied fairly considerably without much
change to the quantile estimates; this is a useful property
that can obviate the need for depth-duration schemes to
iron out the occasional inconsistencies that arise at high
return periods when a single station record is analysed
over a range of durations.
Some idea of more extreme rainfall regimes may be
gleaned from the maximum product of spacings solution
or, even more vaguely, from the property of geometric-
maximum stability.
An interesting and important digression was that in
the examination of the ﬁt of the log logistic distribution
to falls of short duration D, scaling factors k based on
the theory of the random division of an interval were
found to give acceptable estimates for durations such as
D /2, D/3, D/4, D/5, .... In terms of the log logistic dis-
tribution the degree of approximation implied is that an
LLG(c, a, g) at duration D scales into an LLG(c, ka, kg)
at the shorter duration.
Over a wide range of durations ﬁtting the log logistic
distribution to annual maxima of Irish rainfalls using the
L-moment method of parameter estimation proved
successful. In comparisons with estimates based on the
generalised extreme value distribution the log-logistic
values agree closely for return periods up to about
100 years and tend to be higher at longer return periods.
References
Ahmad MH, Sinclair CD, Werrity A (1988) Log-logistic ﬂood
frequency analysis. J Hydrol 98:205–224
Bickel DR (2002) Robust estimators of the mode and skewness of
continuous data. Comput Stat Data Anal 39:153–163
Cheng RCH, Amin NAK (1983) Estimating parameters in con-
tinuous univariate distributions with shifted origin. J R Stat Soc
B 45(3):394–403
Crame´ r H (1946) Mathematical methods of statistics. Princeton
University Press
David HA (1981) Order statistics. Wiley, New York
Fitzgerald DL (1996) Maximum product of spacings estimators for
the generalised Pareto and log-logistic distributions, Stochastic
hydrology and hydraulics, vol 10. Springer, Berlin Heidelberg
New York, pp 1–15
Flood Estimation Handbook (1999) Institute of hydrology, vols 1
and 2. Wallingford, UK
Flood Studies Report (1975) Meteorological studies, vol 2. NERC,
London
HMSO (1981) Handbook of meteorological instruments. Her
Majesty’s Stationery Oﬃce, London
Hosking JRM, Wallis J R (1997) Regional frequency analysis (An
approach based on L-moments), CUP
Jenkinson AF (1977) The analysis of meteorological and other
geophysical extremes, MET O 13 Technical Note No 58, British
Meteorological Oﬃce
Johnson NL, Kotz S (1970) Continuous univariate distributions-1.
Wiley, New York
Logue JJ (1975) Extreme rainfalls in Ireland. Met E
´
ireann, Dublin
Mitnik S, Rachev ST (1993) Modelling asset returns with alterna-
tive stable distributions. Econometric Rev 12(3):261–230
Rohan PK (1986) The climate of Ireland. Met E
´
ireann, Dublin
Rowsin´ ski PM, Strupczewski WG, Singh VP (2002) A note on the
applicability of log-Gumbel and log-logistic probability
distributions in hydrological analyses: I known pdf. Hydrol Sci
J 47(1):107–122
Voorn WJ (1987) Characterization of the logistic and log-logistic
distributions by extreme value related stability with random
sample size. J Appl Prob 24:838–851
257

l2 ¼ 2EðxF Þ À EðxÞ. 3 c¼ l3 ¼ L . e. in some cases it is but more often it is appears low and on occasion quite acceptable estimates of rainfall quantiles occur with a negative value of g.5.4 25. 11 used to estimate the shape and scale parameters. Sample median=48.4 50±2 49.1 105±16 103. mean ¼ g þ sinðcpÞ  c 1Àc mode ¼ g þ a 1þc ð4Þ When g is known the two-parameter log logistic theory applies to (x-g) giving c¼ l 2 l3 ¼ l 1 l2 ð9Þ To a good degree of approximation mode = 2(median)Àmean for values of c less than 0.0 RP 1000 170.0 23. Since x>g.1 0.moment measure of skewness l2 sinðpcÞ p c2 l2 2 l3 ð6Þ ð7Þ ð8Þ i. 2 c ¼ 1 À 2F ðmodeÞ ð5Þ The shape parameter c is thus the measure of skewness proposed in Bickel(2002) and using a plotting position formula we can get a ﬁrst estimate of c. 10 and 11 are the twoparameter ML equations. and Eq.3 176.1 163±35 158. Since b ln((x À g)/a) has a logistic distribution of mean 0 the scale parameter a can be estimated as the geometric mean of (xÀg) for the sample.1 142±28 139. the L-moment estimate. While often successful.225 0. As Eq. Median ¼ g þ a.7 RP 500 146±23 144.e.5 209.5 g 17.1 RP 1000 168±31 165. Also from Eq. the location parameter might. With l1 ¼ EðxÞ.2 . In the three-parameter case a value of the location parameter g was taken.3 49.4. be expected to be a realistic estimate of the minimum of the annual rainfall maxima.g.2 RP 500 148. ideally.288 a 32.7 Station: Ffoulkesmills Years of record=62 Sample max=105. l3 ¼ 6EðxF 2 Þ À 6EðxF Þ þ EðxÞ we get from Eq.7 RP 100 107±11 106. the shape parameter = the L-moment coeﬃcient of variation = the L-moment coeﬃcient of skewness (Fitzgerald 1996). 10 and Eq. The maximum likelihood (ML) parameter estimates For a sample of size n the ML equations may be written c¼ i¼n 1X ð2F ðxi Þ À 1Þ lnðxi À gÞ n i¼1 i¼n X i¼1 ð10Þ 0¼ ð2F ðxi Þ À 1Þ ð1 À 2F ðxi Þ=ðxi À gÞÞ Pi¼n i¼1 ð1=xi À gÞ ð11Þ Pi¼n c¼ i¼1 ð12Þ a ¼ l2 g ¼ l1 À When g is known Eqs. Sample min=35.0 RP 100 108.2.1 119. this method did not always converge but occasional failure Table 1 Parameter and quantile estimates for 2-day (0900-0900UCT) annual maximum rainfalls (mm) Estimation method Parameters c L-moment ML BOOTSTRAP Estimation method Statistic Return Periods RP 2 L-moment ML Mean Median Mean Median 49±2 49. 4 shows mode < median < mean and the skewness is positive. in turn the new values of c and a were used to get a new value of the location parameter.5 Return periods RP 2 49.250 acp . this covers all but the most extreme growth rates. L-moment estimates of the parameters The theory of L-moments and of the closely related probability weighted moments (PWM) is given in Hosking and Wallis (1997).

mean=65. For quantile estimation beyond the range of the sample it is unwelcome that Eq.7 184.9 RP 100 122.2 135. the Lmoment solution produced low or even negative values of the location parameter but usually acceptable estimates of the rainfall quantiles. gÞ n is: 2 p2 þ3 6 9c produced estimates which were distinctly low.7 RP 1000 171. Only with 12 of the 10. However.9 64±3 63. Bootstrapping the data to generate 5.6 0.7 . Comparison of the L-moment and ML solutions for the three-parameter LLG Table 1 presents a typical case of a good ML ﬁt which generally has a higher value of the crucial shape parameter than the L-moment estimate.226 a 58. median=66. 10 to Eq.5 RP 1000 169±32 167.000 samples gave much-reduced ML estimates at the higher return periods.000 samples.9 39. closer to the L-moment estimates which were largely unchanged. Failures tended to occur in samples having a mean less than the median and hence unsuited to an LLG ﬁt.3 Return Periods RP 2 63. Typically.000-year rainfall quantile for an individual station with about 50 years of record. min=38.151 0.6 RP 100 121±12 120.6 182±49 175. The information matrix Á  1 À I ¼ À E ð@ 2 LÞ @hi . It seems clear that the L-moment estimates are the more stable.7 162±37 157.3 RP 500 154.2 Station: Brosna (Mt Eagle) Years of record=55 Sample max=118. nearly half the bootstrap samples did not converge in a fully satisfactory manner and so all that can safely be gleaned from this exercise is that the ML solution is not stable under bootstrapping. Note that the sample mean is less than the median and we may anticipate diﬃculty in ﬁtting the LLG distribution. The three parameter L-moment solution has a very low value of the location parameter. As in Table 1 the ML bootstrap estimates reduce at high return periods and have a higher variance than the L-moment estimates. Table 2 provides an example of an L-moment solution where the location parameter looked implausible and ML solution converged smoothly.9 62. Bootstrapping provided L-moment estimates still well in accord with the quantile estimates and problems occurred with less than 12% of the 10. The ML solution converged with a distinctly more realistic value 3 ð13Þ 0 1 3a2 c2 ð1Àc2 Þpc 3a2 c2 60 4 Àpcð3cþð1Àc2 Þðwð2ÀcÞÀwð2þcÞÞÞ 3ac2 Àpcð3cþð1Àc2 Þðwð2ÀcÞÀwð2þcÞÞÞ 3ac2 7 ð1Àc2 Þpc 7 2 c2 5 3a 2pcð1Àc2 Þ 3a2 c2 Here w is the digamma function. From the bootstrap means and standard errors it would also seem that they give fairly realistic estimates of the 1.4. 10 gives high weight to low as well as to high values of (x-g) in the estimation of the crucial shape parameter.5 23.000 samples.5 125±17 122.9 RP 500 153±24 151. 12.2.000 L-moment solutions was it considered desirable to use the logistic instead of the LLG. the most usual case of non-convergence was with g strongly negative and c low.3.6 g 04. #ðc. The 2p/3p solution From (1) setting the location parameter g at some ﬁxed value is an attractive option. a.251 is to be expected from Eq. @hj . for such samples the ML solution did not converge or of g. For the two-parameter LLG the requirement that the L-moment coeﬃcient of Table 2 Parameter and Quantile estimates for 2-day (0900-0900UCT) Annual Maximum Rainfalls(mm) Estimation Method Parameters c L-moment ML BOOTSTRAP Estimation Method Statistic Return Periods RP 2 L-moment ML mean median mean median 64±3 63. Again there is a strong element of subjectivity in the ML bootstrap results since there were convergence problems with nearly 30% of the 5.2 212.

3 38.2 43.8 21.8 15.2 Generalised extreme value PWM ML PWM bootstrap Mean 22..6 Sample min=7.3 36.4 100 104.3 22.6 143.7 31.1 22. Based on (d) the model for annual maxima adopted was that the exceedances of a ﬁxed. 13 simpliﬁes to give for a large sample of size n Table 3 Comparison of rainfall estimates (mm) RP Years Log logistic LM ML L-moment bootstrap Mean SE Median varðcÞ ¼ 9c2 3a2 c2 .7 18.8 Sample min=53. (b) The l-sk estimate of c was nearly constant as g varied.6 146.1 11.4 10.2 13.8 32.8 32.4 9.4 33. if initially unknown.0 104.ln F)k)/k .1 6.4 99. nðp2 þ 3Þ n covarianceða.3 3.8 136.3 45.7 10.4 33.1 13.2 38.0 98.1 141.7 139.0 121.9 199. it is convenient that this amounts to the three-parameter L-moment solution with its location parameter g reinterpreted. the eﬀect on the mean.2 114.1 194.7 (Duration=96 clock hours Sample max=136.4 91.8 32.2 98.4 91. n 2.9 94.6 36.4 98.2 8. 4 should agree closely and both should fall within the indicated range.2 181.3 21.9 9.8 7.2 42.0 24.2) 50 22.0 100 156.8 172.1 102.3 156.8 157.. In the course of systematically varying g it was found that: (a) The value of g obtained when l-cv = l-sk was never appreciably diﬀerent from that of the three-parameter L-moment solution.9 185.3 92.1 101.7 16.2 8.0 28. Since the random variation is of the parameters. In the case of the preferred solution for the LLG the ellipsoid reduces to an ellipse having the parameter estimates (a.8 7. varðaÞ ¼ .6 100 26. An advantage is that the two-parameter ML theory now applies and Eq.0 133.7 116.0 145.1 (Duration=24 clock hours Sample max=94. c were in all cases so close to the three-parameter L-moment solution that there was no practical diﬀerence between their quantile estimates. (c) The l-cv and ML estimates of c were broadly similar and both varied systematically with g. However.1 1000 43.1 25. The classical assumption is of a parent population distribution with the parameters estimated to a degree of accuracy determined by the sample size n.1 30..2 7.1 500 198. 14.3 113.. these values of g.0 Sample min=34.2 174.6 92.(. Belonging to the population may then be taken as being within the ellipsoid of concentration and a random sample of n regarded as n random choices within the ellipsoid. RP return period Generalised extreme value: x=n+a(1 .2 10.5 139.9 45.7 115.7 Median 22.0 159.2) 50 94.2 23.6 148.1 140.3 500 37. a non-zero value of g can be found which gives a good ﬁt. For annual maxima of Irish rainfall the ﬁt with g=0 is poor. ML maximum likelihood.6 24.4 25.1 91. SE standard error.2 221. A second test is that the scale parameter should equal the geometric mean of (x-g).0 104.4 25.8 103.0 24.7 10.4 132.2 36.5 37.7 92.8 Station: Cork Airport Data: Annual maxima 1963–2002 L-M L-moment.9 1000 147.6 119. PWM probability weighted moment. the ellipsoid of concentration re-expresses this in terms suited to computer simulation by deﬁning a volume in parameter space within which all points can be taken as equally likely. From a plot.4 Median 22.2 148.7 1000 219.5 121.4 9.3 197.1 6.5 164.4 150. threshold g follow a two-parameter LLG distribution. (d) A good ﬁt with a common value of g could be found where the ML and the L-moment solutions for the shape and scale parameters essentially coincided.1 2.9 130.to 60-year sample can usually be seen to fall within a certain range.7 244. the mode of a typical 30.5 22.4 36.8 145.2 216.3 21.6 172.5 137.8 215.6 130. the estimates of the mode provided by 2(median) À mean and Eq.9 219.4 183.4 166. c) as the mean of the random variable and with covariance matrix given by Eq.0 42.8) 50 141.9 500 133.7 22.2 174.7 3.4 26.0 122. 3 and 14 the variance of the rainfall with return period T is: ! varðaÞ 2 2 varðxT Þ ¼ ðxT À gÞ varðcÞðlnðT À 1ÞÞ þ ð15Þ a2 Ellipsoids of concentration and random variation ` Cramer (1946) shows that for a random variable n (n 1. n n) of known covariance matrix an ellipsoid of concentration can be deﬁned with the property that randomly choosing a large number of points within the ellipsoid yields a random variable with the same mean and covariance matrix as n.0 26.2 215.1 25.4 3.8 147.4 147.9 SE 2. cÞ ¼ 0 ð14Þ Using Eqs.5 17.9 24.6 26.8 20. median and variability of the quantile estimates Log logistic Ellipse of concentration Mean SE 2.7 94.1 142.2 154.9 138. a.0 37.1 194.5 116.9 (Duration=1 clock hour Sample max=22.6 25.0 154.2 182.252 variation (l-cv) and the L-moment coeﬃcient of skewness (l-sk) be equal is a key test of ﬁt.

2 10.0 19.8 127. A restatement of Eq.011 16.163 65.8 130.6 0.3(115.7 136.2 8. Remarks on Table 3 – The GEV displays a lower growth rate but even at the 1.253 can be compared with the preferred solution estimates.7 119.0) 22. 16EðTm Þ ¼ ðn=ðn À mÞÞ.7) Sample Station: Cork Airport Data: annual maxima 1963–2002 Generalised extreme value x=n+a (1 À (Àln F)k)/k .80 0.5 17.8 Estimates 23. The median can be ob- Table 4 Comparison of top three sample values with median estimates Duration Clock hours Distribution Log logistic Parameter 01 24 96 c a g c a g c a g 0. 16 using incomplete beta functions. Fit of the estimates to the top sample values For some rough guidance on the return period Tm associated with the higher order statistics of a sample of size n note that the distribution of y=Tm À 1 for the mth order statistic of any distribution is given by f ðyÞ ¼ 1 y mÀ1 ð1 þ yÞÀðnþ1Þ Bðm.1 141.1 145.153 2.5 84.7 120.9 94. 16 is a highly skewed distribution. – For both distributions the median plotting position is used to make the estimates. For purposes of comparison Table 3 includes the probability weighted moment.000 sample points. n À m þ 1Þ   ðx À gÞ b ¼ Tm À 1 where ym ¼ a ð18Þ and so.3 78. The ellipse of concentration statistics are based on 20. – The very close agreement between the median of the ellipse of concentration estimates and the L-moment values at all return periods allows the identiﬁcation of the L-moment solution with a central value of the well-deﬁned population based on the sample data.6 +0.272 5.2 84. While it might be observed that they agree more closely with one another than with the data and that the GEV where B is the beta function and f(y) an inverted beta distribution.2 129. Remarks on Table 4.7 96.035 12.3 19. n À m þ 1Þ ¼ 2 Bðm.6 17.4 18. the three-parameter ML estimates and the values obtained from bootstrapping the data and using the L-moment solution. – Both the bootstrap and the ellipse of concentration give broadly similar quantile estimates for the LLG and both their medians and means agree well with the L-moment values.5 Estimates 23.5 94.1 78. 16 is that the CDF of the mth order statistic F(xm) = y/ (1+y) has a beta distribution and Jenkinson (1977) exploited this to show that its median plotting position can be derived as m À 0:307 ~ . 17.3 18.9 5. to the same approximation as in Eq.5 standard errors of the LLG estimates. F ðxm Þ ¼ n þ 0:386 In turn this gives the good approximation: ~ Tm ðmedianÞTm ¼ n þ 0:386 n À m þ 0:693 ð17Þ For the LLG the median value of xm is given by the incomplete beta function expression 1 Bym =1þym ðm.6 21. for 96 h the ﬁt is greatly improved by taking the fourth highest value (in brackets) into account. c¼ 1 b ð19Þ As m approaches n Eq.4(113. From Eq.4 À0.9) Generalised extreme value Parameter k a n k a n k a n À0.2 77. we get ~ xm ¼ g þ aðTm À 1Þc .4 79. the ML and the bootstrap PWM estimates for the generalised extreme value (GEV) distribution (Hosking and Wallis 1997). The ﬁt to the sample values is good for 1 h and 24 h durations.148 53.0 87.0 +0.2(115.000-year RP the estimates are within 1. The results are presented in Table 3. Hence the median seems by far the best central value to select for the purpose of gauging the ﬁt of any distribution by examining the quantile estimates for the top sample values. n À m þ 1Þ ð16Þ tained from Eq. The modal value Tmode ¼ ðn þ 1=ðn À m þ 2ÞÞ.4 127.

9 and to use that in Eq. Compared with the estimates of the preferred solution it produces quantiles that are in line with climate change scenarios for mid latitudes. Maximum product of spacings chooses the sample spacings closest to a perfectly regular pattern and hence is a typical index of lack of clustering or ignorance. Taking g=15 we get estimates of 0.8 106.5 56.6 169. – For 1 clock hour we have an instance of the location parameter g as a plausible minimum while the minimum for the GEV is À3. g respectively and gives an 1.2 208.7 138.2 11.000-year rainfall quantile for 24 clock hours while the 18-h value is 149.9 8.44 6 0. x¢n+1 > xn.5 68.4 is accepted for the MPS. in practice contradictions between return periods arise e. is to ﬁrst take the mean of the two estimates of the shape parameter in Eq. Table 5 Log logistic estimates of rainfall quantiles MPS ellipse of concentration RP (years) 2 5 10 20 50 100 250 500 1. It was shown in Fitzgerald (1996) that the MPS method applied to a sample of size n is equivalent to using ML . The 100-year rainfall quantile increases from 90 to 107±12 mm in what we may choose to regard as the new. The location parameter may be considerably increased without much change to the estimates in Table 3 and in Table 4.3 46. thus ironing out the contradiction between 18 h and 24 h estimates and retaining a decent ﬁt to the top sample values.6 208.6 68. This might fairly be taken as an indication of the inadvisability of estimating 1. In general we may reasonably expect the growth rate to decrease or remain steady with increasing duration greater than 1 h or 2 h but. procedure can be used instead of or in conjunction with a depth-duration-frequency scheme [Flood Estimation Handbook (1999).59 4 0.1 Station: Phoenix Park (Dublin) Data: 1-day annual maxima 1881–2002 MPS Maximum product of spacings.254 appears to ﬁt the data a little better than the LLG..4 88.000-year rainfall quantile of about 122 mm for 24 clock hours and an 18-h value of over 125 mm.000-year quantiles from forty data points but it is necessary to do so while imposing consistency over durations and return periods.75 0.6.1 112.000-year quantile of 184 mm. It was further shown for the twoparameter LLG that MPS gives a positively biased estimate of the shape parameter while the scale parameter is essentially equal to the ML or L-moment estimates.1 3. Of itself this gives RP estimates quite close to the preferred solution estimates.5 158.7.7 for Table 4.000 L-moment 34. 3. This.3 46. if the period 1941–2002 is used instead of the record for 1881–2002 the values become 173±36 and 230±54 respectively.2 4.. Table 6 Means and medians of maxima of the random division of the unit interval Number of intervals 2 Mean Median 0. Chap. 35. here Table 3 gives 147.3. the median estimate of the two solutions is the same but for higher return periods MPS gives more extreme quantile estimates.1 as the estimates for c. The 1.5 68.4 90.g. 84.7 76. the MPS results are very much along the lines of those found for the Lmoment solution. vol 2.1 2.8 211.5 133. The method.39 Maximum product of spacings (MPS) solution as a statistical safety factor For parameter estimation of univariate distributions the MPS method of maximising the geometric mean of the sample spacings was developed by Cheng and Amin (1983) as an alternative to maximum likelihood.61 0.000-year quantile increases from 157±25 to about 210±35.2 87.9 139.0 171.6 for the shape and scale parameters while the two-parameter ML solution gives 0.212 and 37.9 MPS 34. 107.52 0.1 and 77. admittedly ad hoc. Giving the two solutions equal weight we get estimates of 95.5 88.3 107. The ML solution for the 3-parameter LLG converges to give 0. 2. 7 to determine the scale parameter. The GEV estimates suﬀer from the same lack of consistency as they give a 1. RP return period.5 61.2 44. SE standard error for a sample of size (n+1) where x¢1 < x1.226.5 52.0 Median 34. Table 4 shows that this is caused by the negative value of the location parameter which forces a high value of the scale parameter and a low value of the shape parameter. median-conserving climate.46 0.75 3 0.. 10]. The location parameter of the preferred solution of Sect.177 and 37.3.4 35.0 170.1 and 17.41 0.3 46. Table 5 shows the results of using MPS estimates of the shape and scale parameters and their variances in an ellipse of concentration exercise as outlined in Sect. for a given value of the location parameter. However.3 25.50 5 0.4 56. x¢n < xn. the amount of the increase being determined by what amounts to an application of the theorem of the mean to the sample values.5 56. 141 and 159 for Table 3 and 97.4 as the Lmoment estimate of the 1. the notion of the LLG estimates as median values is supported. . its estimate of the 50-year quantile as 102 mm appears a little high as the median RP associated with the 40year sample maximum of 94 mm is over 58 years.4 SE± 1.7 18.0 106. x¢2 < x2.4 140.9 Mean 34. a. then.

Unfortunately the sample sizes in this study are rather small and the mode is not well-determined but may be inferred from the mean and median. 2. leading to the hope that population characteristics may be accurately determined from a sample. this solution is computationally inexpensive and the simplicity of the underlying theory makes it easy to investigate its properties either directly or by computerintensive methods. it can be completely determined by three central values. and the total rainfall from the individual events is LLG(c. An example of the latter. For the LLG the corresponding growth curve is of strikingly similar form: ln((xT À g)/(x5 À g)) = c(ln(T À 1) À 1. a r ¼ 1. having several events capable of yielding the maximum. The MPS method deals with uncertainty by increasing the shape parameter while the method of (geometric) random sample size alters the scale parameter of the LLG distribution. take the number of years into the future that the process is likely to remain free of radical change as a geometric variable N(t).255 Why use the log logistic distribution? In the Flood Studies Report (1975) the equation for the rainfall growth curves is given by: ln MT ¼ M5 cðy À 1:5Þ. A point to emphasise is that the ﬁt of the pure twoparameter LLG to the rainfall data was mostly poor and the move away from g=0 crucial to obtaining a good ﬁt.4) where c is the LLG shape parameter. u > 1 QðF ðxÞÞ ¼ F ð21Þ u Hence over years with random numbers. If the negative binomial is used instead of the geometric we get  r  x r pF ðxÞ QðF ðxÞÞ ¼ ¼F . especially if r is an integer since in this case r bears the interpretation of the minimum number of events in unit time. this may be viewed as LLG with a rescaling of the data and of the time unit.e. 3. 1 À rcÞ.. 0\c\1. further. then the argument applied above gives the CDF of the maximum of a sample of size N(t) as LLG. (2002). the parameter estimates of the preferred solution yield close approximations to the sample estimates of the median and mean then its ﬁt to the LLG is regarded as good. after the manner of Mitnik and Rachev (1993). k=1. However. remarkable for the depth of negative feeling expressed towards the 2pLLG. 4 and 5 show. In this form the fractional contribution to the population moment of the range 0–FÀ 1 (u) is given by an incomplete beta function. all the recent papers on the LLG in hydrology that have come to my attention have been on the pure two-parameter case. Among these the LLG has the advantage that. if F(x) is a continuous distribution independent of N. T > 5 Here c= . a. 3 shows that the preferred solution closely mimics the median of a plausible representation of the population giving rise to the data.013 and for a shape parameter c=0. A belief that there was only a probability of 0.20 this yields a scaling factor of 2. Also for a ﬁnite third moment we must have a shape parameter less than one-third and . then F is maximum stable with respect to N if real numbers a and b exist such that 1 x À a X QðF ðxÞÞ ¼ . g) then x .23 to 0. MT ¼ T . Unfortunately for their applicability to this study. B being the beta function.5 that the current model would apply after 50 years would imply t=0. N(p).1. If. We might estimate t from our degree of belief in the length of time the ellipse of concentration generated from the LLG ﬁtted to the annual series will adequately represent the process generating the maxima.07.k where k is the shape parameter of the GEV distribution and c ranges from 0. as Eqs. b>0 ð20Þ pk F ðxÞk ¼ F b k¼1 If the number of rainfall events capable of producing the annual maximum is taken as geometric with generating function Q(s) = (ps)/(1 À (1 À p)s)i.2. They concentrate on the analysis of the tail region and for this study the most pertinent of their many reservations is the necessity to restrict the ranges of the parameters if population moments are to be kept ﬁnite. If instead we consider a series of annual maxima having an LLG distribution and. This leads to the choice of long-tailed distributions with shape and scale parameters. the practical implications of this are most easily discussed in terms of the following equation derived from Eq. u ¼ pÀc . .. The ellipse of concentration argument in Sect. Here t is the rather ill-deﬁned and subjective probability of ‘major change in any year’ and N(t) the waiting time to the change. :::. is the ´ paper of Rowinski et al. of events with an LLG distribution the maximum is an LLG distribution of a rescaled variable. most plausibly if p is small. 1 À ð1 À pÞF ðxÞ u a Burr III distribution (Johnson and Kotz 1970).38 to allow for the uncertainty in the length of the period. 3: x À gr  E ¼ Bð1 þ rc. the geometric distribution used here has mean 1/p but its mode is 1 and this seems low for Irish conditions as most years are nondescript.e. The LLG distribution also exhibits geometric-extreme stability in the sense of Voorn (1987) i. a distribution on the positive integers with generating function Q(s). Experience dictates that when choosing a probability distribution for rainfall extremes it is wise to allow for much higher values than appear even in a long-term record. pk=p(1 À p)k . However.year return period and y¼ À lnðÀ lnðF ðxÞÞ % lnðT À 0:5Þ. the point of most interest is that under both these randomisations the LLG (and also the logistic distribution) is maximum stable.

9 Station: Rosslare Data: 45 annual maxima for 15 min and 30 min (millimetres) Distribution ﬁtted: log-logistic with shape.4 occasionally appearing for short durations where growth rates are steeper.9±0. the Dines recorder is not accurate below about 15 min.7 18.6 11. The restriction thus points up the advantages of the L-moment solution. with c=0. g=3. especially for durations ranging from 15 min to 3 h.2 18.1 16. The distribution of the maximum of n intervals on (0.4 14.6 11.1 to 0. This sort of restriction on the range of the shape parameter constitutes a diﬃculty for full tail analysis. Maximum values for 5 min from Dines recorders range from 12 mm to about 8 mm over a period of record mostly of 40–50 years. this may be attributed to the fact the parameters are determined by the median.9 35. Over Ireland the 500-year quantile for 4-h falls is usually less than 80 mm. it must be possible for more than 52% of the total to fall in one quarter of the time period the factors should apply better at low than at high return periods. When the sub-period can be considered part of a single event the theory of the random division of an interval can contribute plausible scaling factors. Flood Studies Report (1975). 14 and 15 show how the variance of the quantile estimates increases with c. vol 2.3±2.33.4 20.4.8 5 from 30 mm 4.33 were treated with caution and.6 and 3.3 6. In the particular case of the LLG using the factor to adjust the scale and location parameters and assuming the shape parameter is unchanged is equivalent to directly multiplying the estimates by the scaling factors. 1) (David 1981) gives the following. a=5. Values of the shape parameter c encountered in this study range from 0. The indication is that. g) 15-Min: c=0.2±0.35. it is important to appreciate that with c=0.8 11.5 27.1±1.5±1. high may mean about 1.4±6.256 so on. scale and location parameters (c. 1949 (Rohan 1986). The consistency between the 5-min estimates is heartening but their accuracy is diﬃcult to gauge. At c=0.and 15-min values. this strongly suggests that 5-min falls of over 20 mm occur in the general area but attaching a return period at a single point must be labelled a hopeful exercise and it seems a good idea to employ a regionalisation scheme to supplement single station analysis. Durations from 15 min to 24 h have been extracted from 42 stations. Thus. That the factors are signiﬁcant can be seen by checking them against the 5-year return period factors in Table II of Logue (1975) or Tables 3.8±0. with values of 0. a. Table 7 presents the estimates of 5-min quantiles from 30. as in the next section.2 27.5 17.6 9. For statistical practice. while in no way a drastic problem even at c=0.4±3. in parameter estimation it would be unwise to employ third order moments.4 5. urban drainage requires design values for durations less than15 min and for such short durations no reliable long-term data are available.0 5. More importantly. in the absence of data for the shorter duration the assumption that the shape parameter is constant seems a reasonable ﬁrst approximation.8 8.333 the ﬁgure is 99%. In addition Eqs.9 15. insofar as possible.1±9. If there is a diﬀerence Dc between the shape parameters of the two durations it appears as a multiplier (T À 1)Dc when estimating the T-year quantile.7. especially the midtable values.000 years the ﬁgures are 8% and 2% respectively. 30-min: c=0. For E(xFr) these percentages rise only slowly with r .4 51.5 14. g=1.8 19. in practice. Values of c>0.7. checked for consistency against higher and lower durations.4±0.3 16. the approach of the population moment of order three to inﬁnity is mediated by values very unlikely to appear in samples. the scaled down 30min values being systematically higher.3 23.6 30 min mm 9.1±1. their derivation as central values of the distribution of a maximum suggests that they apply more widely.9±2. however.7 6.5±1.9.0 . Fitting the LLG sometimes gives high growth rates.5 23. Table 7 Five-minute quantile estimates from 15 min and 30 min values LLG estimates RP Years 2 5 10 20 50 100 250 500 1000 15 min mm 7.3.254. for example. Since.000-year return period.6 12.5 42. Fairly near Rosslare there was a remarkable fall of 38 mm in 12 min in August.9±2. Clearly.000 years for some durations.4 7.4 9.285. Short-duration falls and random division of the unit interval The sub-hourly falls available are predominantly from Dines rainfall recorders (HMSO 1981).8 13.3±4.33 to 0.75 times the 30min estimates agree to within one standard deviation even at the1.0 Table 6 estimates 5 from 15 mm 4. mode and mean and these three may reasonably be taken to scale in the same way. low values of the shape parameter can be more accurately determined from a sample than high.8 32.4 the population mean receives a 12% contribution from values with return periods of 50 years or more while for 100 years and 1.5±6. However. it is possible to get that amount in 1 h (Rohan 1986) and so.0 9. a=6. It has been standard practice at Irish synoptic stations to extract daily maximum rates for 5 min but. checking for consistency of the quantile estimates raised the question of the possibility of ﬁnding scaling factors between durations.33 about 90% of the value of the third moment is contributed by values with return periods of 105 years(time units) or more.1±0. with values of the shape parameter in excess of 0. Note that the 15-min estimates and 0.5±4. as Logue (1975) points out.

Wiley. Singh VP (2002) A note on the applicability of log-Gumbel and log-logistic probability distributions in hydrological analyses: I known pdf. New York Fitzgerald DL (1996) Maximum product of spacings estimators for the generalised Pareto and log-logistic distributions. Springer. Curiously. . References Ahmad MH. D/5. Kotz S (1970) Continuous univariate distributions-1. Econometric Rev 12(3):261–230 ´ Rohan PK (1986) The climate of Ireland. D/3. Met Eireann. Princeton University Press David HA (1981) Order statistics.257 Discussion and conclusions The parameters of the log logistic distribution are readily interpretable in terms of the data. Amin NAK (1983) Estimating parameters in continuous univariate distributions with shifted origin. J Hydrol 98:205–224 Bickel DR (2002) Robust estimators of the mode and skewness of continuous data. kg) at the shorter duration. Berlin Heidelberg New York. Met Eireann. this is a useful property that can obviate the need for depth-duration schemes to iron out the occasional inconsistencies that arise at high return periods when a single station record is analysed over a range of durations. In comparisons with estimates based on the generalised extreme value distribution the log-logistic values agree closely for return periods up to about 100 years and tend to be higher at longer return periods. London HMSO (1981) Handbook of meteorological instruments. J Appl Prob 24:838–851 . Wallis J R (1997) Regional frequency analysis (An approach based on L-moments). Hydrol Sci J 47(1):107–122 Voorn WJ (1987) Characterization of the logistic and log-logistic distributions by extreme value related stability with random sample size. g) at duration D scales into an LLG(c.. Comput Stat Data Anal 39:153–163 Cheng RCH. Dublin ´ Rowsinski PM. scaling factors k based on the theory of the random division of an interval were found to give acceptable estimates for durations such as D /2. vols 1 and 2. J R Stat Soc B 45(3):394–403 ´ Cramer H (1946) Mathematical methods of statistics. MET O 13 Technical Note No 58. Wallingford. Its form and L-moment solution are such that it is well suited to computer-intensive investigations. vol 2. having gone to some pains to obtain a solution with attractive statistical properties. CUP Jenkinson AF (1977) The analysis of meteorological and other geophysical extremes.. pp 1–15 Flood Estimation Handbook (1999) Institute of hydrology. An interesting and important digression was that in the examination of the ﬁt of the log logistic distribution to falls of short duration D. New York ´ Logue JJ (1975) Extreme rainfalls in Ireland. from the property of geometricmaximum stability. it was found that the parameters so obtained can be varied fairly considerably without much change to the quantile estimates. Werrity A (1988) Log-logistic ﬂood frequency analysis. British Meteorological Oﬃce Johnson NL. Stochastic hydrology and hydraulics. Over a wide range of durations ﬁtting the log logistic distribution to annual maxima of Irish rainfalls using the L-moment method of parameter estimation proved successful. In terms of the log logistic distribution the degree of approximation implied is that an LLG(c. NERC. Rachev ST (1993) Modelling asset returns with alternative stable distributions. Bootstrap and ellipse of concentration exercises prove very informative about the stability of the quantile estimates. even more vaguely.. Strupczewski WG. vol 10. UK Flood Studies Report (1975) Meteorological studies. London Hosking JRM. Her Majesty’s Stationery Oﬃce. D/4. Dublin Mitnik S. Wiley. ka. Sinclair CD. Some idea of more extreme rainfall regimes may be gleaned from the maximum product of spacings solution or. a.