See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337058827

The Frechet distribution: Estimation and Application an Overview

Article · May 2020

DOI: 10.1080/09720510.2019.1645400

CITATIONS READS

11 3,528

4 authors:

Pedro L. Ramos Francisco Louzada

University of São Paulo University of São Paulo
81 PUBLICATIONS 349 CITATIONS 414 PUBLICATIONS 2,902 CITATIONS

SEE PROFILE SEE PROFILE

Eduardo Ramos Sanku Dey

University of São Paulo St. Anthony's College
19 PUBLICATIONS 75 CITATIONS 218 PUBLICATIONS 1,666 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Machine learning approaches to credit risk modeling View project

Regression Analysis View project

All content following this page was uploaded by Pedro L. Ramos on 06 November 2019.

The user has requested enhancement of the downloaded file.

 Journal of Statistics and Management Systems

ISSN: 0972-0510 (Print) 2169-0014 (Online) Journal homepage: https://www.tandfonline.com/loi/tsms20

The Fréchet distribution: Estimation and

application - An overview

Pedro L. Ramos, Francisco Louzada, Eduardo Ramos & Sanku Dey

To cite this article: Pedro L. Ramos, Francisco Louzada, Eduardo Ramos & Sanku Dey (2019):
The Fréchet distribution: Estimation and application - An overview, Journal of Statistics and
Management Systems, DOI: 10.1080/09720510.2019.1645400

To link to this article: https://doi.org/10.1080/09720510.2019.1645400

Published online: 05 Nov 2019.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsms20
Journal of Statistics & Management Systems
ISSN 0972-0510 (Print), ISSN 2169-0014 (Online)
DOI : 10.1080/09720510.2019.1645400

The Fréchet distribution : Estimation and application - An overview

Pedro L. Ramos *
Francisco Louzada
Eduardo Ramos
Institute of Mathematical and Computer Sciences
University of São Paulo
Avenida Trabalhador São- Carlense 400
13566-590 Sao Carlos SP
Brazil

Sanku Dey
Department of Statistics
St. Anthony’s College
Shillong 793001
Meghalaya
India

Abstract
In this article we consider the problem of estimating the parameters of the Fréchet
distribution from both frequentist and Bayesian points of view. First we briefly describe
different frequentist approaches, namely, maximum likelihood, method of moments,
percentile estimators, L-moments, ordinary and weighted least squares, maximum product
of spacings, maximum goodness-of-fit estimators and compare them using extensive
numerical simulations. Next, we consider the Bayesian infer- ence approach using reference
priors. Five real data sets related to the minimum flow of water on Piracicaba river in Brazil
are used to illustrate the applicability of the discussed procedures.

Subject Classification: (2010) 62F10, 62P12

Keywords: Bayesian Inference, Fréchet distribution, Hydrological applications, Maximum product
of spacings, Reference prior.

*E-mail: [email protected]

©
 2 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

1. Introduction
Extreme value theory plays an important role in statistical analysis.
The most used distribution to describe extreme data is the generalized
extreme value (GEV) distribution [21]. Its cumulative density function
(CDF) is given by

F(t|σ , µ , ξ ) = 
}
exp{−[1 + ξ (t − µ ) / σ ]−1/ξ , for ξ ≠ 0
+
(1)
exp{− exp[−(t − µ ) / σ ]}, for ξ = 0

where s > 0, m, x ŒR and x+ = max(x, 0). Gumbel, Weibull and Fréchet

distributions are special cases of the so-called generalized extreme
value (GEV) distribution. The Fréchet distribution is named after French
mathematician Maurice René Fréchet, who developed it in the 1920s
as a maximum value distribution (which is also known as the extreme
value distribution of type II). Kotz and Nadarajah [25] describe this distri-
bution and discussed its wide applicability in different spheres such as
accelerated life testing, natural calamities, horse racing, rainfall, queues
in supermarkets, sea currents, wind speeds, track race records and so on.
Let the random variable T follows Fréchet distribution then its
probability density function (PDF) and cdf are given by
=f (t|λ , α ) λα
= t −(α +1)e − λt −α and F(t|λ , α ) e − λt −α , (2)

for all t > 0 and the quantities a > 0 and l > 0 are the shape and the scale
parameters respectively. This distribution is also referred as Inverse
Weibull distribution. The PDF can be unimodal or decreasing depending
on the choice of the shape parameter while its hazard function is always
unimodal.
Several researchers have studied different aspects of inferential
procedures for the Fréchet distribution. From the classical perspective,
Calabria and Pulcini [7] and Erto [14] discussed the properties of the
maximum likelihood estimators (MLE) and the ordinary least-square
estimators (LSE) respectively. Ramos et al. [37] presented the MLE for the
the Fréchet distribution in the presence of cure fraction. Loganathan and
Uma [28] compared the MLE, LSE, weighted LSE (WLSE) and the method
of mo- ments (MME) and outlined that the WLSE provided similar results.
Salman et al. [41] and Maswadah [31] studied the Fréchet distribution in the
context of order statistics and generalized order statistics respectively. The
Bayes estimators were discussed by Calabria and Pulcini [8] and Kundu
and Howlader [26] using informative or subjective priors such as Gamma
priors (also known as flat priors). However, Bernardo [5] argued that the
THE FRÉCHET DISTRIBUTION 3

use of simple proper flat priors presumed to be non-informative often

hide important unwarranted assumptions which may easily dominate, or
even invalidate the statistical analysis and should be strongly discouraged.
Recently, Abbas and Tang [1] studied Fréchet distribution based on Jeffreys
and reference priors.
Parameter estimation is significant for any probability distribution
and therefore various estimation methods are frequently studied in
the statistical literature. Traditional estimation methods such as the
MLE, MME, LSE and WLSE are often opted for parameter estimation.
Each has its own merits and demerits but the most popular method of
estimation is the maximum likelihood estimation method. Besides the
above cited methods, we consider five additional methods to estimate
the parameters of Fréchet distribution. These additional methods are the
maximum product spacing estimator (MPS), percentile estimator (PE),
Cramér-von-Mises estimator (CME), Anderson-Darling estimator (ADE)
and L-moment (LME) estimator. Several authors have compared these
different estimation methods for different distributions [2, 11– 13, 27, 32,
36, 40]. From the Bayesian point of view different Bayes estimators are
discussed using objective priors and different loss functions. Also, the
coverage probability with a confidence level equal to 95% for the estimates
are obtained. To evaluate the performance of the estimators, a simulation
study is carried out. Finally, five real life data sets have been analyzed for
illustrative purposes.
The objective of this paper is to estimate the parameters of the model
from both frequentist and Bayesian perspective and to develop a guideline
for choosing the best estimation method for the Fréchet distribution, which
we would be of profound interest to applied statisticians. The present
study is unique because of the fact that thus far, no attempt has been made
to compare all these aforementioned estimators for the two- parameter
Fréchet distribution. At the end, we present the better estimation procedure
for the Fréchet distribuion.
The paper is organized as follows. Section 2 describes nine frequentist
methods of estimation. The Bayes estimators are presented in section 3. In
Section 4, we present the Monte Carlo simulation results. In Section 5, the
usefulness of the Fréchet distribution is illustrated by using five real data
sets. Finally, concluding remarks are provided in Section 6.
 4 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

2. Classical parameter estimation methods

In this section, we describe nine methods for estimating the parameters
l and a of the Fréchet distribution.

2.1 Maximum Likelihood Estimation

Among the statistical inference methods, the maximum likelihood
method is widely used due its desirable properties including consistency,
asymptotic efficiency and in- variance. Under the maximum likelihood
method, the estimators are obtained by max- imizing the log-likelihood
function. Let T1, ..., Tn be a random sample such that T ~ Fréchet(l, a).
Then, the likelihood function from (2) is given by

n n − (α + 1 )   −α 
n n n

∏= f L(λ , α |t)
(
=t i
, λ , α ) λ α ∏ it  exp  ∑ ti  .
− λ ( 3)
=i 1 = i1   i =1 

The log-likelihood function (3) is given by

n n
l(λ , α=
|t) n log(λ ) + n log(α ) − (α + 1)∑ log(ti ) − λ ∑ ti−α .
=i 1 =i 1

From ∂l(l, a|t)/∂l = 0 and ∂l(l, a|t)/∂a = 0, we get the likelihood

equations
n n
− ∑ ti−α =
0 ( 4)
λ i =1

and
n n n
− ∑ log(ti ) + λ ∑ ti−α log(ti ) =
0 (5)
=i 1 =i 1 α
whose solutions provide λ̂ MLE and α̂ MLE. After some algebraic manipulations,
the estimate α̂ MLE can be obtained by solving the following non-linear
equation
n∑ i =1 ti−α log(ti )
n
n n
− ∑ log(ti ) + 0.
= (6)
α ∑ i=1 ti−α
n
i =1

The estimate λ̂ MLE can be obtained by substituting α̂ MLE in

n
λMLE = . (7)
∑
n
t–α
i =1 i
THE FRÉCHET DISTRIBUTION 5

The obtained ML estimates are asymptotically normally distributed

with a joint bivariate normal distribution given by
(λˆMLE , αˆ MLE ) ~ N 2 [(λ , α ), I −1 (λ , α ))] for n → ∞ ,

where I(l, a) is the Fisher information matrix given by

 n n(1 − γ − log(λ )) 
 
λ 2
λ
I (λ , α ) =  ,
α
(8)
 n(1 − γ − log(λ )) n  π 2  
 2 
+ (1 − γ − log(λ ))2  
 λα α  6  

and g ≈ 0.5772156649 is known as Euler-Mascheroni constant.

In the following we prove the existence and uniqueness of MLEs.

Theorem 2.1 : Let t1, ···, tn be not all equal. Then the MLEs of the parameters a
and l are unique and are given by α̂ and
n
λˆ = n
, (9)
∑t
i =1
−αˆ
i

where α̂ is the only solution of non-linear equation

n n n n
G(α ) =− ∑ log ti −
−
i ∑ t α log t .
i
n
α
=i 1 =

i =1
− i 1
i ∑t α

Proof : See Appendix A.                     

2.2 Moments Estimators

The method of moments is one of the oldest method used for
estimating the parameters of the statistical models. The raw moments of T
for the Fréchet distribution is
r
 r
E(T r |λ , α ) = λ α Γ  1 −  , (10)
 α 
where r ŒN and Γ(λ ) = ∫0 e x dx is the gamma function. Note that
∞
− x λ −1

E(T r|g, a) does not have a finite value for a > r. The moment estimators
(MEs) for the Fréchet distribution can be obtained by equating the first
two theoretical moments with the sample moments. However, instead of
equating the first two theoretical moments, we consider that
6 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

1
 1
E(T|λ , α ) = λ α Γ  1 −  and
 α
2
  2  1 
Var(T|λ ,=
α ) λ α  Γ  1 −  − Γ2  1 −   . (11)
  α  α 
Therefore, the population coefficient of variation is given by
Var(T|λ , α ) Γ(1 − 2α −1 )
CV
= (X|λ , α ) = − 1,
E(T|λ , α ) Γ 2 (1 − α −1 )

which is independent of the scale parameter l. So, the estimator for α̂ MLE
can be obtained by solving the following non-linear equation
Γ(1 − 2α −1 ) 8
− 1 − =0.
Γ (1 − α − 1)
2
t
Substituting α̂ MME in (11) the estimate λ̂ MME can be obtained by solving
tα
λˆME = . (12)
Γ (1 − α − 1)
α

However, this estimator can only be computed for a > 2 which is

undesirable.

2.3 Percentile Estimator

The percentile estimator is a statistical method used to estimate the
parameters by comparing the sample points with the theoretical points.
This method was originally suggested by Kao [22, 23] and has been
widely used for distributions that has the quantile function in closed
form expression, such as the Weibull and the Generalized Exponential
distribution. The quantile function of the Fréchet distribution has the
closed form and is given by
1
−
1 1  α
Q( p|λ , α ) =  log    . (13)
λ
  pi 
Therefore, the percentile estimates (PCEs), λ̂ PCE and α̂ PCE, can be
obtained by minimizing
2
 −
1

n
 1 1  α

∑  t( i ) −  λ log  p    ,
i =1
   i  
 
THE FRÉCHET DISTRIBUTION 7

with respect to l and a, where pi denotes an estimate of F(t(i); l, a) and t(i) is

the ith order statistics (we assume the same notation for the next sections).
The estimates of l and a can be also be obtained by solving the following
non-linear equations:
 − 
1
−
1
n
 1  1  α  1  1  α
∑ t( i ) −  λ log  p     λ log  p   = 0,
i =1
   i    i 
 
 − 
1
–
1
n
 1  1  α  1  1   1  1  α
∑ t( i ) −  λ log  p    log  λ log  p    log    = 0,
i =1
   i     i   λ  pi  
 
i
respectively. In this paper, we consider pi = . However, several estima-
n +1
tors of pi can be used instead [30].

2.4 L-Moments Estimators

Hosking [19] proposed an alternative method of estimation analogous
to conventional moments, namely L-moments estimators. L-moments
estimators can be obtained by equating the sample with the population
L-moments. Hosking [19] stated that the L-moment estimators are more
robust than the usual moment estimators, and also relatively robust to the
effects of outliers and reasonably efficient when compared to the MLE for
some distributions.
For the Fréchet distribution, the L-moments estimators can be obtained
by equating the first two sample L-moments with the corresponding
population L-moments. The first two sample L-moments are
1 n 2 n

=
∑ t( i ) , =
n i 1=
l1
= l2 ∑=
n(n − 1) i 1
(i 1) t( i ) − l1 ,

where t(1), t(2), ···, t(n) denotes the order statistics of a random sample from
a distribution function F(t|l, a). The first two population L-moments are
1
1  1
µ1 ( λ , α ) = ∫0 Q( P|λ , α ) dp = E( X |λ , α ) = λ α
Γ1− 
 α 
(14)

and
1
1
 α1   1
µ2 (λ ,=
α) ∫0 Q( p|λ , α )(2 p − 1)
=dp λ α
 2 − 1  Γ  1 − α  ,
 (15)
   
 8 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

where Q(p|l, a) is given in (13). After some algebraic manipulation, the

estimator for α̂ LME can be obtained as
log(2)
αˆ LME = . (16)
log(2) + log (∑ n
i =1
(i =
1)t( i ) ) − log(n(n − 1) t )
Note that, substituting α̂ LME in (14) the estimator for λ̂ LME can be
obtained by solving
tα
=λˆLME , α > 1. (17)
Γα (1 − α −1 )

It is worth noting that, among the chosen methods, the L-moments

estimator was the only that has closed-form solution for both parameters.

2.5 Ordinary and Weighted Least-Square Estimate

The least square estimators λ̂ LSE and α̂ LSE, can be obtained by
minimizing
2
 i 
n
S(λ , α ) ∑  F(t( i ) |θ , λ ) −
= ,
i =1  n + 1 

with respect to l and a. Similarly, they can also be obtained by solving the
following non-linear equations (see Erto [14] for more details):
n
 i 
∑  F(t (i )
|λ , α ) − η (t |λ , α ) =
n + 1  1 ( i )
0,
i =1
n
 i 
∑  F(t( i ) |λ , α ) − n + 1 η 2 (t( i ) |λ , α ) =
i =1 
0,

where
− λt − α − λt(−iα)
=η1 (t( i ) |λ , α ) t(=
−α
i)
e ( i ) and η 2 (t( i ) |λ , α ) λt(−iα) log(t( i ) ) e . (18)

The weighted least-squares estimators (WLSEs), λ̂ WLSE and α̂ WLSE, can

be obtained by minimizing
2
(n + 1)2 (n + 2)  n
i 
=W (λ , α ) ∑  F(t( i ) |λ , α ) − .
i =1 i(n − i + 1)  n + 1 

The estimators can also be obtained by solving the following non-

linear equations,
 THE FRÉCHET DISTRIBUTION 9

n
(n + 1)2 (n + 2)  i 
∑ i ( n − i + 1)  F(t( i ) |λ , α ) − n + 1 η1 (t( i ) |λ , α ) =
0,
i =1  
n
(n + 1)2 (n + 2)  i 
∑
i =1 i(n − i + 1)   F(t( i ) |λ , α ) − η (t |λ , α ) =
n + 1  2 ( i )
0,

2.6 Method of Maximum Product of Spacings

The maximum product of spacings (MPS) method is a powerful
alternative to ML method for estimating the unknown parameters of
continuous univariate distributions. This method was proposed by Cheng
and Amin [9, 10], and later independently developed by Ranneby [38] as
an approximation to the Kullback-Leibler measure of information.
Let= Di (λ , α ) F(t( i ) |λ , α ) − F(t( i −1) |λ , α ), for i = 1, 2, …, n + 1, be the
uniform spacings of a random sample from the Fréchet distribution, where
=F(t(0) |λ , α ) 0,= F(t( n+1) |λ , α ) 1 and ∑ in=+11 Di (λ , α ) = 1. The maximum
product of spacings estimators λ̂ MPS and α̂ MPS are obtained by maximizing
the geometric mean of the spacings
1
 n +1  n +1
G(λ , α ) = ∏ Di (λ , α ) ,
 i =1 

with respect to l and a, or, equivalently, by maximizing the logarithm of

the geometric mode of sample spacings
1 n +1
H (λ , α ) = ∑ log Di (λ ,α ).
n + 1 i =1

The estimators λ̂ PS and α̂ MPS of the parameters l and a can be obtained

by solving the following nonlinear equations
∂H (λ , α ) 1 n +1 1
=
∂λ
∑ [η (t |λ , α ) − η=
n + 1 i =1 Di (λ , α ) 1 ( i )
(t )|λ , α )] 0,
1 ( i −1
(19)
∂H (λ , α ) 1 n +1 1
=
∂α
∑ [η (t |λ , α ) − η=
n + 1 i =1 Di (λ , α ) 2 ( i )
(t )|λ , α )] 0,
2 ( i −1

Note that if t(i+k) = t(i+k−1) then Di+k(l, a) = Di+k−1(l, a) = 0 for some i.

Therefore, the MPS estimators are sensitive to closely spaced observations,
especially ties. When ties are due to multiple observations, Di(l, a) should
be replaced by the corresponding Cheng and Amin likelihood [10]
presented f (t(i), l, a) desirable since t(i) properties = t(i−1).
10 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

Cheng and Amin [10] presented desirable properties of the MPS

such as asymptotic efficiency and invariance. They also proved that the
consistency of maximum product of spacing estimators holds under
much more general conditions than for maximum likelihood estimators.
Therefore, the MPS estimators are asymptotically normally distributed
with a joint bivariate normal distribution given by
(λˆ , αˆ ) ~ N [(λ , α ), I −1 (λ , α ))], for n → ∞.
MPS MPS 2

2.7 The Cramér-von Mises maximum goodness-of-fit estimators

The Cramér-von Mises is a type of maximum goodness-of-fit
estimators (also called minimum distance estimators) and is based on the
difference between the estimate of the cumulative distribution function
and the empirical distribution function [6].
Macdonald [29] motivated the choice of the Cramér-von Mises
statistic and provided empirical evidence that the bias of the estimator is
smaller than the other goodness- of-fit estimators. The proposed estimator
is based on the Cramér-von Mises statistics given by
∞
=Wn2 n ∫ ( F(t( i ) − En (t( i ) ))2 dF(t( i ) ),
−∞

where En(·) is the empirical density function. Boos [6] discussed its
asymptotic properties and presented its computational form which is
given by
2
1 n
 2i − 1 
C(λ , α ) = + ∑  F(t( i ) |θ , λ ) − . (20)
12n i =1  2n 
The Cramér-von Mises estimators λ̂ CME and α̂ CME of the parameters l
and a are obtained by minimizing
n
 2i − 1 
∑  F(t (i )
|λ , α ) − η (t |λ , α ) =
2n  1 ( i )
0,
i =1 
n
 2i − 1 
∑  F(t( i ) |λ , α ) −
i =1 
η (t |λ , α ) =
2n  2 ( i )
0,

where h1 (·| l, a) and h2 (·| l, a) are given respectively in (18).

2.8 Method of Anderson-Darling

Other type of maximum goodness-of-fit estimators is based on
an Anderson-Darling statistic and is known as the Anderson-Darling
estimator. The Anderson-Darling statistic is given by
THE FRÉCHET DISTRIBUTION 11

∞ ( F(t( i ) ) − En (t( i ) ))2

ADSn2 = ∫–∞ F(t)(1 − F(t))
dF(t( i ) ).

Boos [6] also discussed the properties of the AD estimators and

presented its computational form which the Anderson-Darling estimators
λ̂ ADE and α̂ ADE are obtained by minimizing, with respect to l and a, the
function
1 n
(2 i − 1)  −lt(−iα) + log(1 − e ( n +1−i ) ) .
− λt − α
A(λ , α ) =− n − ∑
n i =1 
 

These estimators can also be obtained by solving the following non-

linear equations:
n η1 (t( i ) |λ , α ) η1 (tn+1−i : n |λ , α ) 
∑ (2 i − 1)  F(t − 0,
 =
i =1  ( i ) |λ , α ) 1 − F(tn+1−i : n |λ , α ) 
n η (t |λ , α ) η 2 (tn+1−i : n |λ , α ) 
∑ (2 i − 1)  1 ( i ) − 0,
 =
i =1  F(t( i ) |λ , α ) 1 − F(tn+1−i : n |λ , α ) 

where h1 (·| l, a) and h2 (·| l, a) are given respectively in (18).

3. Bayesian Analysis
In the previous sections, we have presented different estimation
procedures using the frequentist approach. In this section, we consider
the Bayesian inference approach for estimating the unknown parameters
of the Fréchet distribution. Bayesian analysis is an attractive framework
in practical problems and has grown popularity in recent years. The
prior distribution is a key part of the Bayesian inference and there are
different types of priors distribution available in the literature (see, for
instance, Ramos et al. [35]). Usually, non-informative priors are preferable
because if prior information on study parameters is unavailable or does
not exist for a device, then initial uncertainty about the parameters can be
quantified with a non-informative prior distribution. An important non-
informative reference prior was introduced by Bernardo [4], with further
developments (see Bernardo [5] and references therein). The proposed idea
was to maximize the expected Kullback-Leibler divergence between the
posterior distribution and the prior. The reference prior provides posterior
distribution with interesting properties, such as invariance under one-to-
one transformations, consistent marginalization and consistent sampling
 12 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

properties. Recently, Abbas and Tang [1] de- rived two reference priors as
well as the Jeffreys’ prior for the Fréchet distribution. Here, we derive such
priors by using different approach, the following proposition is used to
obtain the reference priors.

Proposition 3.1 : [Bernardo [5], pg 40, Theorem 14]. Let θ = (q1, q2) be
a vector of the ordered parameters of interest and I(q1, q2) is the Fisher
information matrix. If the parameter space of q1 does not depend of q2 and
Ij,j(θ ), j = 1, 2 are factorized in the form
1 1
−
=S1,12 (θ ) f=(θ ) g1 (θ 2 ) and I 2,2
1 1
2
(θ ) f2 (θ1 ) g2 (θ 2 ).

Then the reference prior for the ordered parameters q is given by

π θ (θ ) = f1(q1)g2(q2) and there is no need for compact approximations, even
if the conditional priors are not proper.

Theorem 3.2 : Let w1 = (l, a) and w2 = (a, l) be the vectors of the ordered
parameters then the reference priors for the ordered parameters are respectively
given by
1 1
π ω (λ , α ) ∝ and π ω (α , λ ) ∝ . (21)
1
λα 2
λα

Proof : For w1 = (l, a) we have f1(l) = l–1, g1(a) = 1,

2
π
f 2 (λ =
) + (1 − γ − log(λ ))2 , g2(a) = a–1. Therefore, π ω (λ , α ) ∝ f1 (λ ) g2 (α ) ∝ λ −1α −1 .
6 1

λ , α ) ∝ f1 (λ ) g2 (α ) ∝ λ α . Analogously, we obtain π ω (λ , α ).              

−1 −1
2

Moreover, Berger et al. [3] suggested by starting with a collection of

reference priors and then taking the arithmetic mean or the geometric
mean to obtain an overall reference prior. Therefore, an overall reference
prior is the same as (21). Another well-known non-informative prior was
introduced by Jeffreys [20] and can be obtained through the square root
of the determinant of Fisher information matrix (8), such prior is widely
used due to its invariance property under one-to-one transformations.
After some algebraic manipulations we have
1
π J (λ , α ) ∝ . (22)
λα
which is equal to (21). It is worth noting that such prior also arises
considering the Jeffreys rule (see, Kass and Wasserman [24]). Therefore,
THE FRÉCHET DISTRIBUTION 13

even considering different methods to obtain non-informative priors, we

have the same prior for the Fréchet distribution.
Combining the likelihood (3) and the prior (22), the posterior
distribution is
(λα )n−1 n −α  n

π (λ , α |t )
= ∏ ti exp −λ ∑ ti−α  , (23)
c (t ) i =1  i =1 
where
n
 n

=c (t ) ∫ ( λα ) n −1
∏ t −α
i
exp  − λ ∑ ti−α  dθ (24)
i =1  i =1 
and A = {(0, ∞) × (0, ∞)} is the parameter space of θ . To get reliable inference,
first we have to check whether the posterior distribution (23) is a proper
posterior, i.e., c (t) < ∞.

Theorem 3.3 : The posterior distribution (23) is proper if n ≥ 2.

Proof : Since (λα )n−1 ∏ in=1 ti−α exp{−λ ∑ in=1 ti−α } ≥ 0, by Tonelli theorem (see
Folland [15]) we have
n
 n

=C(t ) (λα )n−1 ∏ ti−α exp −λ ∑ ti−α  dθ
∫

i =1  i =1 
∞∞ n
 n

= ∫ ∫ (λα )n−1 ∏ ti−α exp −λ ∑ ti−α  dλ dα
0 0 i =1  i =1 
∞ n
 n

= Γ(n)∫ α n−1 ∏ ti−α ∑ ti−α  dα
0 i =1  i =1 
−α
∞ n 
ti 
≤ ∫ α n−1 ∏   dα < ∞ ,
0 i =1  min(t i
,..., t )
n 

the last inequality holds if n ≥ 2.                  

The marginal posterior distribution for a is given by
−n
n
 n

π (α |t ) ∝ α n−2 ∏ ti−α ∑ ti−α  . (25)
i =1  i =1 
The conditional posterior distribution for l is
 n 
p(λ|α , t ) ~ Gamman  n, ∑ ti−α  , (26)
 i =1 
14 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

where Gamma(a, c) is the Gamma distribution with PDF, f (x, a, c) = ca

xa−1 exp(−cx)/G(a). By considering marginal and conditional posterior
distributions (25) and (26), the convergence of Markov Chain Monte Carlo
(MCMC) method can be easily achieved.
Abbas and Tang [1] derived the same priors (21) and proved that the
obtained posterior is proper. However, the authors did not proved that the
obtained posterior means for a and l are finite. These proofs are important
in order to obtain reliable re- sults. The following theorem prove that the
posterior mean of l is improper depending on the data.
n  t 
Theorem 3.4 : The posterior mean for l is improper in ∏ i =1  min(t i,...,t )  ≤ min
 i n 
(t1, ..., tn) and n ≥ 2.

Proof : The posterior mean for l is given by

n
 n

=λˆ(t ) ∫ λ(λα ) ∏ ti exp −λ ∑ ti  dθ .
n −1 −α −α

 i =1  i =1 

Notice that, since ∏ in=1 ti / min(t1 ,...tn )n+1 ≤ 1 by hypothesis it follows

that (∏ in=1 ti / min(t1 ,...tn )n+1 )−α ≥ 1−α =
1.
Now, since λ (λα )n−1 ∏ in=1 ti−α exp{−λ ∑ in=1 ti−α } ≥ 0, by Tonelli theorem we
have
n
 n

=λˆ(t ) ∫ λ(λα ) ∏ t exp −λ ∑ t  dθ
n −1
α α −
i
−
i
 i =1 i =1
∞ ∞
n
 n

= ∫0 ∏ ti−α ∫ α n−1λ n exp −λ ∑ ti−α  dλ dα
i =1 0  i =1 
∞ −n
n
n 
=Γ(n + 1)∫ α n−1 ∏ ti–α ∑ ti–α  dα
0 i =1  i =1 
−α

∏ i=1 ti  dα
∞ n


≥ Γ(n + 1)n ∫α
−n n −1

 min(t1 ,..., tn )n+1 

0
 
∞
≥ Γ(n + 1)n− n ∫ α n−1dα = ∞ ,
0

where the last inequality holds for n ≥ 2.               

From the above theorem, we can see that the posterior mean of l
may be improper depending on the data, which is undesirable. The use of
Monte Carlo methods in improper posterior was discussed by Hobert and
Casella [18]. The authors argued that “one can not expect the Gibbs output
 THE FRÉCHET DISTRIBUTION 15

to provide a “red flag”, informing the user that the posterior is improper.
The user must demonstrate propriety before a Markov Chain Monte Carlo
technique is used.” In our case, as the properness of the posterior mean
of l depend on the data, other alternative is needed to be considered as
Bayes estimator. A discussion about other alternatives will be presented in
the next section.

4. Simulation Study
In this section, we present some experimental results to evaluate
the performance of the different methods of estimation discussed in the
previous sections.

4.1 Classical approach

In this subsection, we compare the efficiency of the different estimation
methods considering classical approach. The following procedures are
adopted:
(1) Generate N samples from the Fréchet(l, a) distribution with size n
and compute θˆ = (λ̂ , α̂ ) via MLE, ME, LME , LSE, WLSE, PCE, MPS,
CME and ADE.
(2) Using θˆ and θ , compute the mean relative estimates (MRE) =
∑∑inin==11((θθˆˆ11 −−θθjj)/
)/N N and the mean squared errors (MSE) = ∑ in=1 (θˆ1 − θ j )2 / N , for j =
1, 2.
n ˆ
∑ i =1 (θ1 − θ j ) / N , for j =
2
1, 2.
Considering the above approach, the most efficient estimator will be
the one whose MREs is closer to one with smaller MSEs. These results are
computed using the software R (see, R Core Team [34]). The seed used to
generate the pseudo-random samples from the Fréchet distribution was
2018. We have chosen the values to perform this procedure are N = 500,
000 and n = (15, 20, 25, ···, 140). We presented results only for θ = (2, 4) due
to space constraint. However, results are similar for other choices of l and
a. Figure 1 shows the MREs, MSEs for the estimates of θ . The horizontal
lines in the Figure correspond to MREs and MSEs being one and zero
respectively.
From Figure 1, we observe that the MSEs of all estimators of the
parameters tend to zero for large n and also, the values of MREs tend to
one, i.e. the estimators are asymptotically unbiased and consistent for the
parameters. For both parameters, we observe that the moment estimators
has the largest MREs and MSEs respectively among all the considered
estimators. Further, we also observe that the MPS, ADE, LSE and WLSE
16 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

Figure 1
MREs, MSEs for the estimates of l = 2 and a = 4 for N = 500, 000 simulated
samples, considering different values of n using the following estimation
method 1-MLE, 2-ME, 3-LME , 4-LSE, 5-WLSE, 6-PCE, 7-MPS, 8-CME, 9-ADE.

performs better than the MLEs for small and moderate sample sizes in
terms of MREs and MSEs. Moreover, the MPS estimators have the smallest
MSEs among all the considered estimators.
Combining all results with the good properties of the MPS method
such as consis- tency, asymptotic efficiency, normality and invariance, we
suggest to use MPS estima- tors of the parameters of Fréchet distribution
in all practical purposes.

4.2 Bayesian approach

In this subsection, we obtain the Bayes estimator under the same
assumptions of section 4.1. The 95% coverage probability of the asymptotic
confidence intervals under the also classical evaluated. set-up For and
large the number credible of experiments intervals (CI95%and ) considering
under the Bayesian confidence set-up level are of 95%, the frequencies of
intervals that covered the true values of q should be closer to 95%.
THE FRÉCHET DISTRIBUTION 17

4.2.1 Metropolis-Hastings (M-H) algorithm

Since the marginal posterior distribution of a does not have closed
form, the Metropolis-Hastings (M-H) algorithm is applied to generate
samples from this marginal distribution. In this case, we have used the
Gamma distribution as transition kernel q(a (j)|a (*), b) for sampling values
of a, where b is a known hyperparameter that controls the acceptance rate
of the algorithm. It is worth mentioning that, the choice of the transition
kernel is arbitrary, and other non-negative random variables could also be
used instead. The M-H algorithm operates as follows:
(1) Start with an initial value a (1) and set the iteration counter j = 1;
(2) Generate a random value a (∗) from the proposal Gamma(a (j), b);
(3) Evaluate the acceptance probability
 π (α (∗) |t ) q(α ( j ) , α (∗) |b) 
h(α ( j ), α (∗) ) = min  1,  ,
 π (α |t ) q(α , α |b) 
( j) (∗) ( j)

where p (·) is given in (25). Generate a random value u from an

independent uniform in (0.1);
(4) If h(a (j), a (*)) ≥ u(0, 1) then a (j+1) = a (*), otherwise a (j+1) = a (j);
(5)  Change the counter from j to j + 1 and return to step 2 until
convergence is reached.
In this case, we choose b to be equal to one. However, other values
can also be considered. To decrease the necessary time taken for M-H
method to reach the con- vergence, we can use (16) as a good initial value
for a (1). Considering the conditional posterior distribution (26), the Bayes
estimator for l can be obtained direcly from the Gamma distribution with
ˆ
(n, ∑ in=1 tiα Bayes ) as well as its respective credible interval is evaluated by the
quantile function. The decision rule used to obtain the Bayes estimators
will be presented in the following.

4.2.2 Bayes estimator

The selection of a decision rule to obtain posterior estimates is of
fundamental problem in Bayesian statistics. Usually, this problem is over
looked by many authors. The most common risk function used to obtain
the Bayes estimates is the mean squared error, by considering this risk
function, the obtained Bayes estimates are the posterior means. Other
alternative functions can be considered, for instance, the posterior mode,
also known as maximum a posteriori probability (MAP) is obtained by
18 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

assuming a 0-1 loss function, while the posterior median is obtained

considering a linear loss function.
For the Fréchet distribution, Abbas and Tang [1] presented two
Bayes estimators, the first is the MAP estimators obtained from Laplace’s
approximation. Despite of the fact that reference priors are invariant under-
one-to-one transformation, the MAP is not [33], and therefore, the obtained
Bayes estimator is not invariant. Although the authors considered other
Bayes estimator, but they did not specify how the proposed estimator was
obtained. This lack of information makes such results non-reproducible,
which is undesirable for any application. As the authors used Monte Carlo
methods, the most common Bayesian estimator is the posterior mean.
However, from Theorem 3.4 we have proved that the posterior mean may
be improper depending on the data, which is undesirable. In fact, consider
the example analyzed by Abbas and Tang [1] related to fatigue lifetime
data. The data is given by: 152.7, 172.0, 172.5, 173.3, 193.0, 204.7, 216.5,
234.9, 262.6, 422.6. From the proposed theorem, the posterior mean of l is
improper if
n  ti 
∏  min(t ,..., t )  ≤ min(t ,..., t ),
1 n
i =1  1 n 
since 25.4 ≤ 152.7, then the posterior mean of l is improper and can not be
used. As it was discussed earlier, for the proposed dataset we can easily
draw a sample for the marginal distribution and compute the posterior
mean without any “red flag”, but such estimate is meaningless. Therefore,
for the Fréchet distribution the posterior median is a reasonable choice as
it is invariant under-one-to-one transformation and finite for n ≥ 2 almost
surely.

4.2.3 Results
For each simulated data set under the Bayesian approach, 5, 500
iterations are per- formed using the MCMC methods. As a burning
sample, we discarded the first 500 initial values taking jumps of size 5 to
reduce the auto-correlation values among the chain, getting at the end one
chain of size 1, 000. To validate the convergence of the obtained chain, we
used the Geweke criterion [16] with a 95% confidence level. At the end, 10,
000 posterior medians for a and l were computed.
The results are presented only for θ = (2, 4) due to space constraint.
However, the following results are also similar for other choices of l and
a. Figure 2 presents the MREs, MSEs and the coverage probability with a
THE FRÉCHET DISTRIBUTION 19

confidence level equal to 95% for the estimates obtained for the MPS and
posterior medians. It is to be noted that, since MPS performs better than
their counterparts, we compare Bayes estimates with MPS.
From Figure 2 we observe that the Bayes estimates have the MRE
nearest to one while both approaches (MPS and Bayes estimates) have
approximately the same MSE. Moreover, considering the Bayesian
approach, we have obtained accurate coverage probability through
the highest posterior density intervals. Therefore, we conclude that the
Bayesian approach is preferred in order to make inference on unknown
parameters of the Fréchet distribution.

5. Applications
In this section, we analyze five real data sets related to minimum
monthly flows of water (m3/s) on the Piracicaba River, located in S ̃ao
Paulo state, Brazil. This study can be useful to protect and maintain aquatic

Figure 2
MREs, MSEs and the coverage probability with a 95% confidence level for the
estimates of l = 2 and a = 4 for N = 500, 000 simulated samples, considering
different values of n using the following estimation method,
7 - MPS, B - Bayes estimator.
20 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

resources for the state [39, 42]. The data sets (see in Supplemental material
for more details) are obtained from the Department of Water Resources
and Power agency manager of water resources of the State of S ̃ao Paulo
from 1960 to 2014.
Table 1 presents the posterior median (Bayes estimators) and 95%
credible intervals for l and a of the Fréchet distribution for the data sets
related to the total monthly rainfall during May, June, July, August and
September at Piracicaba River.
The results obtained using the Fréchet distribution are compared
with the Weibull, Gamma, Lognormal (LN), Gumbel and Generalized
Exponential (GE) distribution [17]. We consider certain discrimination
criteria such as BIC (Bayesian Information
Criteria), AIC (Akaike Information Criteria) and AICc (Corrected
Akaike information criterion) and computed respectively by BIC = −2l
(θˆ ; t) + k log(n), AIC = −2l(θˆ ; t) + 2k and AICc = AIC+[2k(k + 1)]/[(n − k −
1)], where k is the number of parameters to be fitted and θˆ is estimation
of θ . Given a set of candidate models for t, the preferred model is the
one which provides the minimum values of the aforementioned statistics.
Table 2 presents the results of BIC, AIC and AICc, for different probability
distributions. The goodness of fit can also be checked through the over
plot of the survival function adjusted by the proposed theoretical models
onto the empirical survival func- tion as shown in Figure 3.
Table 1
Bayes estimates, 95% credible intervals for l and a for the data sets related
to the total monthly rainfall during May, June, Month July, August and
September at Piracicaba River.

Month θ θˆ Bayes CI95%(θ )

l 309.890 (223.248; 16.505)

May
a 1.817 (1.401; 2.293)
l 89.758 (64.376; 121.075)
June
a 1.585 (1.194; 2.033)
l 204.493 (146.666; 275.840)
July
a 2.048 (1.549; 2.637)
l 401.656 (290.594; 537.969)
August
a 2.4585 (1.876; 3.119)
l 55.128 (39.539; 74.362)
September
a 1.529 (1.163; 1.939)
THE FRÉCHET DISTRIBUTION 21

Table 2
Results of the BIC, AIC and the AICc for different probability distributions
for the data sets related to the minimum flows of water (m3/s) during May,
June, July and August at Piracicaba River.

Month Test Fréchet Weibull Gamma LN Gumbel GE

BIC 361.43 390.91 386.90 369.93 394.23 384.04
May AIC 358.06 387.53 383.52 366.55 390.85 380.66
AICc 358.38 387.86 383.84 366.88 391.18 380.98
BIC 346.92 379.72 381.02 359.70 403.55 380.81
June AIC 343.60 376.39 377.69 356.37 400.22 377.48
AICc 343.93 376.73 378.03 356.70 400.55 377.81
BIC 302.86 336.50 332.78 316.75 341.31 330.30
July AIC 299.54 333.17 329.45 313.42 337.98 326.97
AICc 299.87 333.50 329.79 313.75 338.32 327.30
BIC 283.92 310.33 303.41 294.30 303.68 299.35
August AIC 280.49 306.90 299.98 290.87 300.25 295.92
AICc 280.81 307.22 300.30 291.19 300.57 296.24
BIC 329.06 344.21 341.77 332.96 351.45 340.68
September AIC 325.73 340.89 338.44 329.63 348.12 337.35
AICc 326.06 341.22 338.77 329.96 348.45 337.69

Comparing the empirical survival function with the adjusted

distributions, we observe that Fréchet distribution fits better than the
chosen models. These results are confirmed from the BIC, AIC and
AICc values, since Fréchet distribution has the minimum values for the
proposed data sets. Therefore, our proposed methodology can be used
successfully to analyze the minimum flow of water during May, June, July,
August and September at Piracicaba River using the Fréchet
distribution with the Bayesian approach.

6. Conclusions
In this paper, we considered different methods of estimation of the
unknown parameters both from frequentist and Bayesian viewpoint
of Fréchet distribution. We considered the MLEs, MMEs, LME, PCEs,
LSEs, WLSE, MPS, CME and ADE as frequentist estimators. As it is not
feasible to compare these methods of estimation theoretically, we have
performed an extensive simulation study to compare these nine methods
22 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

Figure 3
Survival function adjusted by the empirical and the different
distributions for the data sets related to the minimum flows of water
(m3/s) during May till September at Piracicaba River.

of estimation. We compared these frequentist estimators mainly with

respect to MREs and MSEs. The results show that among the classical
estimators, MPS performs bet- ter than their counterpart in terms of MSEs
in the simulation study. Additionally, we have obtained Bayes estimates
through posterior median and compared with respect to MREs and MSEs.
We have obtained accurate coverage probability through the credible
intervals. Therefore, combining all these results, we conclude that the
Bayesian approach is preferred in order to make inference on unknown
parameters of the Fréchet distribution. We also compared this model with
some existing competing models and the Fréchet distribution performed
reasonably well.

Acknowledgements
The authors are grateful to the principal editor and the anonymous
reviewers for their valuable suggestions and comments which have
THE FRÉCHET DISTRIBUTION 23

improved the paper substantially. Pedro L. Ramos is grateful to the S ̃ao

Paulo State Research Foundation (FAPESP Proc. 2017/25971-0). The data
used in the application and the codes are available in the supplemental
material.

Appendix A : Proof of Theorem 2.1.

Proof : It is important to notice that the solutions to the likelihood equation

(4) and (5) may fail to be the global maximum of l(l, a|t). However, the
reasoning that follows justify why in this case these solution will in fact be
the global maximum. Notice that

max l(λ , α |t) = max

λ∈ + ,α ∈ + +
max
+
l(λ , α |t)
α ∈
( λ∈
)
To find by max λ∈+ l(λ , α |t) we notice its derivative in relation
∂l( λ ,α|t ) n
to l, for a fixed, is given by = − ∑ in=1 ti−α which is positive for
∂λ λ
λ < λˆ(α ) =n
tα
and negative in case l > λ̂ (a). Therefore λ̂ (a) is the unique
∑in=1
−
i

value that provides max λ∈+ l(λ , α |t) for a fixed and
max
+
l(λ , α |t) = l(λ (α ), α )
λ∈

Now, to find max l(λ̂ (a), a|t) notice, by the chain rule, that l(λ̂ (a),
a|t) is a differentiable function in relation to a with derivative given by
dl( λ (α ),α|t ) ∂l( λ (α ),α|t ) ∂l( λ (α ),α|t ) ∂l( λ (α ),α|t ) ∂l( λ (α ),α|t )
=
dα ∂λ
λ ′(α ) + =
∂α ∂α
because
∂α
= 0.
dl( λ (α ),α|t )
Let G(α ) =
∂α

Let us prove that

n
n n
n
G(α=
i
) n
−α
i
α
= ∑ log t + ∑t log ti
=i 1 =

i =1
−α i 1
i ∑t
admits a unique solution. It is straightforward to see that lima → 0 + G(a) =
t
∞. We now prove that lima → 0 + G(a) exists and is negative. Let ui = n in
∏ i =1 ti
for i = 1..., n.
Then ti = ui n ∏ in=1 ti for i = 1, …, n and therefore

n n n n  n 
G(α ) =− ∑ log ti + ∑ u α log  u ∏ t
−
n 
n
α i i i

=i 1 =

i =1
−α i 1
i ∑u  i =1 
n n
n n n
= − ∑ log ti + log ∏ ti + ∑ u α log(u ) –
i i
n
α
=i 1 = i =1

i =1
–
i
i 1
∑u α

n
n n
n
–α
i i
=
α
− ∑u log(u ).
– α i =1
∑u
 n n n
n
n  
G(α ) =− ∑ log ti + −α
n
∑u log  u ∏t 
24 =i P. L. RAMOS,−F.
n
α i
α LOUZADA,
i
E. RAMOS
i
  AND S. DEY
1=

i =1
i
i 1
∑ui =1  
n
n n n
n
i
= − ∑ log t + log ∏ t +
i
–α
i i ∑u log(u )
α n
=i 1 = i =1

i =1
–α i 1
i ∑u
n
n n
n
= i
–α

α
− i ∑u log(u ).

i =1
– α i =1
i ∑u
Now let umin = min u1, ..., un, and r be the number of times umin appears
n
u− α n
u− α log( u )
in u1, ..., um. Then
= limα →∞ ∑i =1 i , limα →∞ ∑i =1 i −α
r= i
r log(umin ) and
umin umin
therefore
n
lim G(α ) =
0 + r log(umin ) =
n log(umin ) < 0,
α →∞ r
where the last inequality follows since umin < 1, which is true by consequence
of the hypothesis that not all ti are equal. It follows that, because of the
intermediate value theorem, G(a) has at least one root in the interval
[0, ∞).
We now prove that G¢(a) is negative, which in turn will imply that
G(a) cannot have more than one root in [0, ∞), where G¢(a) is given by
2
n
 n   n 
n
∑ ti−α (log)2  ∑ ti−α  −  ∑ ti−α log ti 
=i 1
G′(α ) =− 2 − n
=  i 1=  i 1  . (27)
a  n
−α 
2

 ∑ ti 
 i =1 

To show that G¢(a) is negative, it is enough to show that

2
n
 n   n 
∑ ti− β (log ti )2  ∑ ti− β  −  ∑ ti− β log ti  > 0. (28)
=i 1 =  i 1=  i 1 
To prove this, one can use the Cauchy-Schwartz inequality (a special
case of Holder’s inequality with p = q = 2) stated as follows
2
 n 2  n 2   n 
 ∑ ai  ∑ bi  ≥  ∑ ai bi  , (29)
= i 1=  i 1 = i 1 
ai
where equality holds if constant.
bi
THE FRÉCHET DISTRIBUTION 25

To prove the equation (28), take ai2 = ti–α (log ti )2 and bi2 = ti–α . Then
clearly ai bi = ti–α log ti , and hence the inequality (28) follows easily from
the application of the Cauchy-Schwartz inequality.
dl( λ (α ),α|t )
It follows that G(α ) = has only on root α̂ such that G(a) >
dα
0 for a < α̂ and G(a) < 0 for a > α̂ and therefore α̂ is the only value that
attains the maximum of l(l(a), a). Then
max = l(λ , α |t) l= (λˆ(αˆ ), αˆ ) l(λˆ , αˆ )
λ∈ + , α ∈ +

and (λ̂ , α̂ ) is the only pair that attains the maximum of l(l, a).     

Supplemental Material
Data set
The datasets used in Section 5 is given as follow:
• May: 29.19, 18.47, 12.86, 151.11, 19.46, 19.46, 84.30, 19.30, 18.47, 34.12,
374.54, 19.72, 25.58, 45.74, 68.53, 36.04, 15.92, 21.89, 40.00, 44.10, 33.35,
35.49, 56.25, 24.29, 23.56, 50.85, 24.53, 13.74, 27.99, 59.27, 13.31, 41.63,
10.00, 33.62, 32.90, 27.55, 16.76, 47.00, 106.33, 21.03.
• June: 13.64, 39.32, 10.66, 224.07, 40.90, 22.22, 14.44, 23.59, 47.02, 37.01,
432.11, 10.63, 28.51, 11.77, 25.35, 25.80, 39.73, 9.21, 22.36, 11.63, 33.35,
18.00, 18.62, 17.71, 100.10, 23.32, 11.63, 10.20, 12.04, 11.63, 50.57, 11.63,
33.72, 14.69, 12.30, 32.90, 179.75, 37.57, 7.95.
• July: 12.98, 15.66, 13.18, 174.94, 10.35, 47.52, 13.28, 24.03, 11.40, 22.71,
43.96, 9.38, 11.40, 13.28, 14.84, 14.44, 63.74, 12.04, 17.26, 28.74, 12.25,
10.22, 26.25, 13.31, 28.24, 12.88, 17.71, 8.82, 10.40, 7.67, 49.15, 17.93,
9.80, 105.88, 10.77, 13.49, 19.77, 34.22, 7.26.
• August: 16.00, 9.52, 9.43, 53.72, 17.10, 8.52, 10.00, 15.23, 8.78, 28.97,
28.06, 18.26, 9.69, 51.43, 10.96, 13.74, 20.01, 10.00, 12.46, 10.40, 26.99,
7.72, 11.84, 18.39, 11.22, 13.10, 16.58, 12.46, 58.98, 7.11, 11.63, 8.24,
9.80, 15.51, 37.86, 30.20, 8.93, 14.29, 12.98, 12.01, 6.80.
• September: 29.19, 8.49, 7.37, 82.93, 44.18, 13.82, 22.28, 28.06, 6.84,
12.14, 153.78, 17.04, 13.47, 15.43, 30.36, 6.91, 22.12, 35.45, 44.66, 95.81,
6.18, 10.00, 58.39, 24.05, 17.03, 38.65, 47.17, 27.99, 11.84, 9.60, 6.72,
13.74, 14.60, 9.65, 10.39, 60.14, 15.51, 14.69, 16.44
26 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

Code of the Metropolis-Hasting algorithm within Gibbs

library(coda)
###################################################################
########
### Gibbs with Metropolis-Hasting algorithm ###
### R: Iteration Number; burn: Burn in; ###
###jump: Jump size; b= Control generation values ###
### log_posteriori: logarithm of posteriori density ###
###################################################################
########
MCMC<-function(t, R, burn, jump, b=1) {
log_posteriori <- function (alfa) {
posterior= (n-2)*log(alfa)-(alfa*sum(log(t)))-n*log(sum(t^(-alfa)))
return(posterior) } ##logarithm of the marginal posterior of alpha
valpha<-length(R+1) ; n<-length(t)
valpha[1]<-max(log(2)/(log((((2/(n*(n-1))*sum(seq(0, n-1,
1)*sort(t)))-mean(t))/mean(t))+1)), 1)
##Set the initial value of alpha based on the L-moments
c1<-rep(0, times=R)
###Starts the M-H algorithm described in Section 4.2.1
for(i in 1:R){
prop1<-rgamma(1, shape=b*valpha[i], rate=b)
ratio1<-log_posteriori(prop1)+dgamma(valpha[i], shape=b*prop1,
rate=b, log=TRUE)-dgamma(
prop1, shape=b*valpha[i], rate=b, log=TRUE)-log_
posteriori(valpha[i])
h<-min(1, exp(ratio1)); u1<-runif(1)
if (u1<h & is.finite(h)) {valpha[i+1]<-prop1 ; c1[i]<-0} else
{valpha[i+1]<-valpha[i] ; c1[i]<-1} }
###Ends the M-H algorithm
valpha2<-valpha[seq(burn, R, jump)] ###Remove the burn-in and takes
jump
ge1<-abs(geweke.diag(valpha2)$z[1]) ### Compute the Geweke
diagnostic
alpha<-median(valpha2); ### Compute the median of alpha
### Compute the median of lambda
lambda<-qgamma(0.5, shape=n, rate = sum(t^(-alpha)), lower.tail =
TRUE)
prai<-quantile(valpha2, probs = 0.025) ## Compute the Lower
credibility interval of alpha
pras<-quantile(valpha2, probs = 0.975) ## Compute the Upper
credibility interval of alpha
## Compute the Lower credibility interval of lambda
prli<-qgamma(0.025, shape=n, rate = sum(t^(-alpha)), lower.tail =
TRUE)
## Compute the Upper credibility interval of lambda
prls<-qgamma(0.975, shape=n, rate = sum(t^(-alpha)), lower.tail =
TRUE)
return(list(acep=(1-sum(c1)/length(c1)), lambda=lambda,
alpha=alpha, LCI_alpha=prai,
UCI_alpha=pras, LCI_Lambda=prli, UCI_Lambda=prls, Geweke.
statistics=ge1)) }
THE FRÉCHET DISTRIBUTION 27

################################################################
## Example ### t: data vector ###
################################################################
rIW<-function(n, lambda, alpha) {
U<-runif(n, 0, 1)
t<-(((1/lambda)*(log(1/U)))^(-1/alpha))
return(t) }
set.seed(2018)
t<-rIW(n=30, lambda=4, alpha=2)
MCMC(t, R=15000, burn=500, jump=5)
$acep ##Aceptance rate
[1] 0.2490667
$lambda ##Posterior median of lambda
[1] 4.880441
$alpha ##Posterior median of alpha
[1] 2.236982
$LCI_alpha ##Lower credibility interval of alpha
2.5%
1.618934
$UCI_alpha ##Upper credibility interval of alpha
97.5%
2.917118
$LCI_Lambda ##Lower credibility interval of lambda
[1] 3.329736
$UCI_Lambda ##Upper credibility interval of lambda
[1] 6.851465
$Geweke.statistics ## Geweke Statitics
1.835165

References

[1] K. Abbas and Y. Tang, Analysis of frechet distribution using refer-
ence priors, Communications in Statistics-Theory and Methods 44,
pp. 2945–2956 (2015).
[2] M. R. Alkasasbeh and M. Z. Raqab, Estimation of the generalized
logistic distribution parameters: Comparative study, Statistical Meth-
odology 6, pp. 262–279 (2009).
[3] J. O. Berger, J. M. Bernardo, D. Sun, et al., Overall objective priors,
Bayesian Analysis 10, pp. 189–221 (2015).
[4] J. M. Bernardo, Reference posterior distributions for bayesian infer-
ence, Journal of the Royal Statistical Society. Series B (Methodological), pp.
113–147 (1979).
[5] J. M. Bernardo, Reference analysis, Handbook of statistics 25, pp.
17–90 (2005).
28 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

[6] D. D. Boos, Minimum distance estimators for location and good-
ness of fit, Journal of the American Statistical association 76, pp. 663–670
(1981).
[7] R. Calabria and G. Pulcini, Confidence limits for reliability and toler-
ance limits in the inverse weibull distribution, Reliability Engineer-
ing & System Safety 24, pp. 77–85 (1989).
[8] R. Calabria and G. Pulcini, Bayes 2-sample prediction for the inverse
weibull distribution, Communications in Statistics-Theory and Meth-
ods 23, pp. 1811–1824 (1994).
[9] R. Cheng and N. Amin, Maximum product of spacings estimation
with application to the lognormal distribution, Math. Report, pp.
79–1 (1979).
[10] R. Cheng and N. Amin, Estimating parameters in continuous univar-
iate distributions with a shifted origin, Journal of the Royal Statistical
Society. Series B (Methodological), pp. 394–403 (1983).
[11] S. Dey, T. Dey, S. Ali, and M. S. Mulekar, Two-parameter maxwell
distribution: Properties and different methods of estimation, Journal
of Statistical Theory and Practice 10, pp. 291–310 (2016).
[12] S. Dey, T. Dey, and D. Kundu, Two-parameter rayleigh distribution:
different methods of estimation, American Journal of Mathematical and
Management Sciences 33, pp. 55–74 (2014).
[13] S. Dey, D. Kumar, P. L. Ramos, and F. Louzada, Exponentiated chen
distribution: Properties and estimation, Communications in Statis-
tics-Simulation and Computation (2016).
[14] P. Erto, Genesis, properties and identification of the inverse weibull
lifetime model, Statistica Applicata 1, pp. 117–128.
[15] G.B. Folland, Real analysis: modern techniques and their applica-
tions, 2nd ed., Wiley, New York, (1999).
[16] J. Geweke, et al., Evaluating the accuracy of sampling-based ap-
proaches to the calculation of posterior moments, Vol. 196, Federal
Reserve Bank of Minneapolis, Research Depart- ment Minneapolis,
MN, USA, (1991).
[17] R. D. Gupta and D. Kundu, Generalized exponential distribution:
different method of estimations, Journal of Statistical Computation and
Simulation 69, pp. 315–337 (2001).
THE FRÉCHET DISTRIBUTION 29

[18] J. P. Hobert and G. Casella, The effect of improper priors on gibbs
sampling in hierarchical linear mixed models, Journal of the American
Statistical Association 91, pp. 1461– 1473 (1996).
[19] J. R. Hosking, L-moments: analysis and estimation of distributions
using linear combina- tions of order statistics, Journal of the Royal Sta-
tistical Society. Series B (Methodological), pp. 105–124 (1990).
[20] H. Jeffreys, An invariant form for the prior probability in estimation
problems, in Proceed- ings of the Royal Society of London A: Math-
ematical, Physical and Engineering Sciences, Vol. 186. The Royal So-
ciety, pp. 453–461(1946).
[21] A. F. Jenkinson, The frequency distribution of the annual maximum
(or minimum) val- ues of meteorological elements, Quarterly Journal
of the Royal Meteorological Society 81, pp. 158–171 (1955).
[22] J. H. Kao, Computer methods for estimating weibull parameters in
reliability studies, IRE Transactions on Reliability and Quality Con-
trol, pp. 15–22 (1958).
[23] J. H. Kao, A graphical estimation of mixed weibull parameters in life-
testing of electron tubes, Technometrics 1, pp. 389–407 (1959).
[24] R. E. Kass and L. Wasserman, The selection of prior distributions by
formal rules, Journal of the American Statistical Association 91, pp. 1343–
1370 (1996).
[25] S. Kotz and S. Nadarajah, Extreme value distributions: theory and
applications, World Scientific, (2000).
[26] D. Kundu and H. Howlader, Bayesian inference and prediction of the
inverse weibull distribution for type-ii censored data, Computational
Statistics & Data Analysis 54, pp. 1547–1558 (2010).
[27] D. Kundu and M. Z. Raqab, Generalized rayleigh distribution: differ-
ent methods of estimations, Computational statistics & data analysis
49, pp. 187–200 (2005).
[28] A. Loganathan and A. Uma, Comparison of estimation methods for
inverse weibull parameters, Global and Stochastic Analysis 4, pp.
83–93 (2017).
[29] P. Macdonald, An estimation procedure for mixtures of distribution,
J. Royal Statist. Soc. Ser 33, pp. 326–329 (1971).
[30] N. R. Mann, N. D. Singpurwalla, and R. E. Schafer, Methods for sta-
tistical analysis of reliability and life data (1974).
 30 P. L. RAMOS, F. LOUZADA, E. RAMOS AND S. DEY

[31] M. Maswadah, Conditional confidence interval estimation for the in-
verse weibull distri- bution based on censored generalized order sta-
tistics, Journal of Statistical Computation and Simulation 73, pp. 887–898
(2003).
[32] J. Mazucheli, F. Louzada, and M. Ghitany, Comparison of estimation
methods for the parameters of the weighted lindley distribution, Ap-
plied Mathematics and Computation 220, pp. 463–471 (2013).
[33] K. P. Murphy, Machine learning: a probabilistic perspective, MIT
press, (2012).
[34] R Core Team, R: A Language and Environment for Statistical Com-
puting, R Founda- tion for Statistical Computing, Vienna, Austria
(2017). Available at https://www.R- project.org/.
[35] P. L. Ramos, J. A. Achcar, F. A. Moala, E. Ramos, and F. Louzada,
Bayesian analysis of the generalized gamma distribution using non-
informative priors, Statistics 51, pp. 824–843 (2017).
[36] P. Ramos and F. Louzada, The generalized weighted lindley distribu-
tion: Properties, estimation and applications, Cogent Mathematics 3,
p. 1256022 (2016).
[37] P. Ramos, D. Nascimento, and F. Louzada, The long term fréchet dis-
tribution: Estimation, properties and its application, Biom Biostat Int
J 6, p. 00170 (2017).
[38] B. Ranneby, The maximum spacing method. an estimation method
related to the maximum likelihood method, Scandinavian Journal of
Statistics, pp. 93–112 (1984).
[39] D. W. Reiser, T. A. Wesche, and C. Estes, Status of instream flow leg-
islation and practices in north america, Fisheries 14, pp. 22–29 (1989).
[40] G. C. Rodrigues, F. Louzada, and P. L. Ramos, Poissonexponential
distribution: different methods of estimation, Journal of Applied Statis-
tics 45, pp. 128–144 (2018).
[41] M. A. W. M. K. SALMAN and S. S. M. AMER, Order statistics from
inverse weibull distribution and characterizations, Metron 61, pp.
389–401 (2003).
[42] D. L. Tennant, Instream flow regimens for fish, wildlife, recreation
and related environmental resources, Fisheries 1, pp. 6–10 (1976).

Received September, 2018

Revised May, 2019

View publication stats