Professional Documents
Culture Documents
Tablada & Cordeiro (2017) - The Modified Frechet Distribution and Its Properties
Tablada & Cordeiro (2017) - The Modified Frechet Distribution and Its Properties
To cite this article: Claudio J. Tablada & Gauss M. Cordeiro (2017) The modified Fréchet
distribution and its properties, Communications in Statistics - Theory and Methods, 46:21,
10617-10639, DOI: 10.1080/03610926.2016.1239115
tion, to extend the Fréchet distribution. By using the Lambert function, Extended distributions;
we obtain some properties of the new distribution. We provide a sim- Fréchet distribution; Lambert
ulation study to illustrate the performance of the maximum likelihood W function; lifetime analysis;
estimates. The flexibility of the introduced distribution is illustrated by modified Fréchet
means of a real data set. We use some goodness-of-fit statistics to ver- distribution.
ify the adequacy of the proposed model. We prove empirically that it is
MATHEMATICS SUBJECT
appropriate for lifetime applications.
CLASSIFICATION
E; N; C; B
1. Introduction
The type II extreme value distribution , also known as the Fréchet distribution, is a family
of continuous distributions developed within the general extreme value theory, which deals
with the stochastic behavior of the maximum and the minimum of independent and identi-
cally distributed (i.i.d.) random variables (Kotz and Nadarajah, 2000). This distribution was
introduced by Maurice Fréchet (1878–1973), who investigated it as one possible limit dis-
tribution for a sequence of maxima. The Fréchet distribution is widely used in applications
involving stochastic phenomena such as rainfall, floods, air pollution (Kotz and Nadarajah,
2000), material properties in engineering applications (Harlow, 2002), and advanced math-
ematical results on point processes and regularly varying functions (Resnick, 2013), among
others.
Given its applicability in extreme value theory, several extensions of the Fréchet distri-
bution were proposed in the literature. Some of them are as follows: the exponentiated
Fréchet (EF) (Nadarajah and Kotz, 2001), beta Fréchet (Nadarajah and Gupta, 2004; Barreto-
Souza et al., 2011), Marshall–Olkin Fréchet (MOF) (Krishna et al., 2013), transmuted Fréchet
(Mahmoud and Mandouh, 2013), gamma extended Fréchet (da Silva et al., 2013), transmuted
EF (Elbatal et al., 2014), Kumaraswamy Fréchet (Mead, 2014), transmuted Marshall–Olkin
Fréchet (Afify et al., 2015), and transmuted exponentiated generalized Fréchet (Yousof et al.,
2015) distributions.
In this paper, we propose a new three-parameter extended Fréchet model named the modi-
fied Fréchet (MF) distribution. Given its simple mathematical form, several of its mathematical
quantities and properties such as the quantile function (qf) and expansions for the ordinary
moments are obtained based on the Lambert W function (Corless et al., 1996; Carrasco et al.,
2008; Silva et al., 2010; Jodrá, 2010). Given that the new distribution has positive real support,
our aim is to define a flexible model for lifetime applications (Marshall and Olkin, 2007).
The paper unfolds as follows. In Sec. 2, we review some main issues of the Lambert W
function. In Sec. 3, we introduce the MF model and discuss the generation of random vari-
ates from this distribution by considering the Lambert function. In Sec. 4, we plot its density
and hazard rate functions for some parameter values. In Sec. 5, we express the qf in terms of
the Lambert function and obtain an expansion for it. In Secs. 6 and 7, we present expansions
for the ordinary and incomplete moments, generating function and Bonferroni and Lorenz
curves using an expansion for the Lambert function. In Sec. 8, we determine the order statis-
tics and their moments. Section 9 is devoted to the maximum likelihood estimates (MLEs) for
complete samples, and in Sec. 10 we carry out a simulation study to evaluate the performance
of these estimates. In Sec. 11, we provide an application of the MF distribution and compare
it with the Fréchet distribution and other three-parameter extended distributions including
the exponentiated Weibull (EW) distribution (Mudholkar and Srivastava, 1993), since it is a
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
satisfying W (x) ≥ −1 is denoted by W0 (x) and called the principal branch of W (·), and the
other branch satisfying W (x) ≤ −1 is denoted by W−1 (x) and called the negative branch of
W (·).
Next, we present some properties and expansions for W0 (·) (Corless et al., 1996). By defi-
nition of the Lambert function, the principal branch satisfies
x = W0 (x ex ), x ≥ −1 (1)
From Equation (1), we obtain (for x ≥ −1)
x
W0 (x ex ) eW0 (x e ) = x ex
and, therefore, letting y = x ex , we note that W0 (·) is the solution of the equation
W0 (y) eW0 (y) = y, y ≥ −e−1 (2)
Furthermore, for |x| ≤ e−1 and r ∈ N, the function W0r (x) has a power series around x0 = 0
given by
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
∞
r (−n)n−r n
W0r (x) = x (3)
n=r
n (n − r)!
Other useful expansion related to W0 (·) with the exponential function is given by
∞
c (c − n)n−1
ec W0 (x) = xn , c ∈ C, |x| ≤ e−1 (4)
n=0
n!
The properties of the function W0 (x) presented above will be used henceforth to obtain
several mathematical quantities and properties of the MF distribution, which is introduced
in the following section. Since the Lambert W function is implemented in various scientific
libraries, computations of quantities such as quantiles and random variates related to the MF
distribution can be implemented more efficiently.
Lai et al. (2003) have successfully defined the three-parameter modified Weibull distribu-
tion by taking appropriate limits on the beta integrated distributions. Also, it is obtained by
extending the Weibull distribution by including the additional term eλx in order to decrease
more rapidly the survival function. Following a similar approach in this paper, we study the
MF distribution by extending the cdf (5) to the form
α β −λx
F (x; α, β, λ) = exp − e , x > 0, α, β > 0, λ ≥ 0 (6)
x
where α, β, and λ are shape parameters. It is straightforward matter to observe that F (x) is
differentiable and strictly increasing in (0, ∞) and that limx→0 F (x) = 0 and limx→∞ F (x) =
1. Therefore, F (x) is, in fact, a legitimate absolutely continuous cdf. Thus, the corresponding
MF density is given by
α β α β
1 −λx
f (x; α, β, λ) = (β + λx) exp −λx − e (7)
x x x
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
Note that the Fréchet distribution is a special case of the MF distribution when λ = 0.
Hereafter, a random variable X with pdf (7) will be denoted by X ∼ MF(α, β, λ).
(β + λ x) ( αx )β e−λx
h(x; α, β, λ) = α
(8)
x exp ( x )β e−λx − 1
Figure . Plots of the exact MF densities and histograms of the simulated data for given parameters.
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
In Figure 2, we compare the exact MF densities and histograms from two simulated data
sets for specified parameters by showing the consistency of the simulated values from (10)
with the MF distribution. To simulate the data, we use the R software (version 3.0.2, lamW
package).
Indeed, we have the known result for all n ∈ N, limx→∞ xn e−x = 0. Letting y = ( αx )β , we have
α
x = y1/β . Thus, x → 0 if and only if y → ∞ and, therefore,
α β β+1
e− ( x ) y β e−y
lim β+1 = lim
x→0 x y→∞ α β+1
β+1
β+1
Let n ∈ N such that β
≤ n. Then, for y ≥ 1, we have y β ≤ yn and then
β+1
y β e−y yn e−y
0 ≤ lim ≤ lim =0
y→∞ α β+1 y→∞ α β+1
which proves (12). Finally, using (12) in the inequality (11), we obtain limx→0 f (x) = 0.
Since limx→0 f (x) = limx→∞ f (x) = 0, then f (x) must have at least one mode. The fol-
lowing result shows that, in fact, f (x) is unimodal.
Proposition 4.2. f (x) is unimodal, with mode x = x0 satisfying
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
β
α β
e−λx0 − −1=0 (13)
x0 (β + λx0 )2
Sketch of the Proof. For λ = 0, we obtain as particular case the Fréchet distribution, which
is well known to be unimodal. Suppose, then, λ > 0. Let s(x) = u(x) + v (x) be, where
α β
u(x) = (β + λx)2 , v (x) = −eλx β + (β + λx)2
x
It is easy to show that u(x) > 0, v (x) < 0, and v (x) is strictly decreasing, for all
x > 0 and α, β, λ > 0. Furthermore, limx→0 u(x) = ∞, limx→0 v (x) = −β(1 + β ), and
limx→∞ v (x) = −∞. So, limx→0 s(x) = ∞.
To calculate f (x), we obtain
1 α β α β −λx
f (x) = 2 exp − e − 2λx s(x)
x x x
and, therefore, f (x) = 0 if and only if s(x) = 0. Since f (x) must have at least one mode, there
exists x0 ∈ (0, ∞) such that s(x0 ) = 0.
We have
α β β(β + λx)
u (x) = (β + λx) 2λ −
x x
Consider two cases: β ≥ 2 and 0 < β < 2. An analysis of u (x) reveals that, in the first case,
u(x) is strictly decreasing in (0, ∞), with limx→∞ u(x) < ∞. So, s(x) is strictly decreasing in
(0, ∞) and limx→∞ s(x) = −∞. In the second case, we have that u(x) is strictly convex, with
β2
minimum point at x∗ = (2−β )λ
. Since u(x) is strictly decreasing in (0, x∗ ), then s(x) is also
strictly decreasing in this interval. To see that s(x) is strictly decreasing in (x∗ , ∞), consider
x = x∗ + z, with z > 0. Thus, x ∈ (x∗ , ∞) and
2 β 2
2β α(2 − β )λ β 2 /(2−β ) λz 2β
s(x) = s̃(z) = + λz −e e + λz +β
2−β β 2 + (2 − β )λ z 2−β
An analysis of s̃(z), for z > 0, reveals that s̃(z) is strictly decreasing and, because of the expo-
nential term, we have limz→∞ s̃(z) = −∞. We conclude then that, in the second case, s(x) is
strictly decreasing in (0, ∞) and limx→∞ s(x) = −∞, too.
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10623
Since s(x0 ) = 0, limx→0 s(x) = ∞, limx→∞ s(x) = −∞, and s(x) is strictly decreasing in
(0, ∞), we conclude that x0 is the unique point in (0, ∞) such that f (x0 ) = 0. Thus, f (x) is
unimodal, with mode x = x0 satisfying
β
e−λx0 α β
s(x0 ) = e−λx0 − −1=0
(β + λ x0 ) 2 x0 (β + λx0 )2
The shape parameters α, β, and λ allow extensive control on the right tail, providing
more light or heavy tails, accordingly β and λ decrease (α increases) or β and λ increase
(α decreases), respectively.
The following corollary of the Proposition 4.2 gives the behavior of the mode of f (x) in
function of the parameters. Hereafter, the symbol → means “tends to”.
b) β → 0+ implies that x0 → 0+ .
β 1/β
c) λ → 0+ implies that x0 → α( β+1 ) .
Proof.
a) Note that, from Equation (13), we have
1/β
β
α = x0 eλx0 /β +1
(β + λx0 )2
Thus, α → 0+ implies that x0 → 0+ and α → ∞ implies that x0 → ∞.
b) From (13), we have that β → 0+ implies that e−λx0 → 1, which implies that x0 → 0+ .
c) From (13), we have that λ → 0+ implies that ( xα0 )β − β1 − 1 → 0, which implies that
β 1/β
x0 → α( β+1 ) , the mode of the Fréchet distribution.
The plots in Figure 3 display the shapes of the pdf of X for some parameter values.
The plots of the hrf of X displayed in Figure 4 reveal the classical shapes such as decreasing,
unimodal, and bathtub. So, the new distribution can be appropriate for different applications
in lifetime analysis.
5. Quantile function
Since the cdf F (x) given in (6) is continuous and strictly increasing, the qf of X is Q(u) =
F −1 (u), for 0 < u < 1. From Sec. 3.2, we note that, for λ > 0, the qf of the MF distribution
can be given explicitly in terms of the Lambert function as in (10).
By using (3), we obtain the following result.
αλ )β eβ
Proposition 5.1. For 0 < u ≤ e−( β and λ > 0, the qf of the MF distribution can be
expanded as
∞
(−n λ)n−1 α n
Q(u) = n/β (14)
n=1 n! β n−1 log u1
Proof. Applying (3) (with r = 1) in (10), we obtain (14). Since the radius of convergence in (3)
is e−1 , then the expansion (14) converges only if β[log(α 1λ)]1/β ≤ e−1 , which is equivalent to u ≤
u
αλ β β αλ )β eβ
e−( β ) e . Finally, since α, β, λ > 0, then e−( β < 1 and therefore (14) holds only for 0 <
αλ β β
u ≤ e−( β ) e .
From the result above, we note that Equation (14) is particularly useful when αβλ ≈ 0, since
αλ β β
in this case e−( β ) e ≈ 1.
Useful skewness and kurtosis measures are given by α3 = μ3 /σ 3 and α4 = μ4 /σ 4 , respec-
tively, where μ j is the jth central moment and σ is the standard deviation. Since for some
parameter values the moments of the MF distribution cannot exist, alternative measures for
the skewness and kurtosis based on quantiles are sometimes more appropriate. The skewness
measure S of Bowley and the kurtosis measure K of Moors are, respectively, defined by
Q(6/8) + Q(2/8) − 2 Q(4/8)
S= (15)
Q(6/8) − Q(2/8)
and
Q(7/8) − Q(5/8) + Q(3/8) − Q(1/8)
K= (16)
Q(6/8) − Q(2/8)
These measures are more robust and they exist even for distributions without moments.
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10625
Figure 5 displays plots of the skewness (15) and kurtosis (16) as functions of λ for some
values of α and β. For evaluating the quantiles using the Lambert function given by (10),
we use the R software (version 3.0.2, lamW package). These plots reveal that, in general, the
skewness and kurtosis measures are decreasing functions of λ.
6. Moments
Moments are important in any statistical analysis. Some of the most important features of a
distribution can be studied through moments. For instance, the first four moments can be
used to describe some characteristics of a distribution. For r ∈ N, the rth ordinary moment
μr of X is given by
∞ ∞ α β α β
μr = E(X r ) = xr f (x) dx = xr−1 (β + λx) exp −λx − e−λx dx
0 0 x x
(17)
For practical purposes, μr can be evaluated numerically.
Also, for z ≥ 0, the rth incomplete moment of X is defined by
z z α β α β
mX(r) (z) = x f (x) dx =
r
x r−1
(β + λx) exp −λx − e −λx
dx (18)
0 0 x x
Clearly, mX(r) (0) = 0 and μr = limz→∞ mX(r) (z). The following result gives an expansion for
the rth incomplete moment of X in terms of generalized exponential integrals.
β
Proposition 6.1. For λ > 0 and 0 < z ≤ λ
W0 (e−1 ), the rth incomplete moment of X can be
expressed as
∞
α β −λz
mX(r) (z) = cn (z) En/β e
n=r
z
n
r (−n)n−r α n λn−r
∞
where cn (z) = an[( αz )β e−λz ]1− β , an = n (n−r)! β n−r
, and Eν (x) = 1
e−tx t −ν dt is the gener-
alized exponential-integral function.
10626 C. J. TABLADA AND G. M. CORDEIRO
∞
e−t
mX(r) (z) = an dt (19)
( αz )β e−λz n=r t n/β
where
r (−n)n−r α n λn−r
an = (20)
n (n − r)! β n−r
Since, for r ∈ N, there exists M > 1 such that W0r( β αt 1/β λ
) < t r in (M, ∞) and t r e−t is inte-
grable in (0, ∞), then W0r( β αt 1/β
λ
)e−t is integrable in (M, ∞). Therefore, by the continuity
α λ
of W0r( β t 1/β )e−t in [ε, M] for all 0 < ε < M, we conclude that this function is integrable in
−t
[( αz )β e−λz , ∞), since ( αz )β e−λz > 0. In addition, the functions ten/β , n = r, r + 1, . . . , are inte-
grable in [( αz )β e−λz , ∞). Thus, it is possible to exchange in (19) the infinite sum and the inte-
gral using the dominated convergence theorem. We obtain
∞ ∞
e−t
mX(r) (z) = an n/β
dt (21)
n=r ( αz )β e−λz t
which is equivalent to
∞
ν−1
Eν (x) = x e−t t −ν dt (22)
x
Thus, letting ν = n
β
and x = ( αz )β e−λz in (22) and replacing in (21), we obtain
∞
α β −λz
mX(r) (z) = cn (z) En/β e (23)
n=r
z
where
n
α β −λz 1− β
cn (z) = an e
z
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10627
We have that the radius of convergence in (3) is e−1 and therefore we must have that β αt 1/β
λ
≤
−1 αλ β β αλ β β
e , which is equivalent to t ≥ ( β ) e . Since the lower limit ( β ) e is attained in (19) for
z = βλ W0 (e−1 ) and the function ( αz )β e−λz is decreasing, expansion (3) can be applied only for
0 < z ≤ βλ W0 (e−1 ), which implies that (23) is only valid in this range.
The following result provides an expansion for μr by considering a suitable transformation
in (17).
Proposition 6.3. For λ > 0 and μr < ∞, an expression for μr is given by
β
∞ ∞ ∞
e αλ
μr = p1 (n) In (k) + p2 (n) En/β (24)
k=1 n=r n=r
β
where
n−r r
2β α n λ r(−n)n−r r(−n)n−r β e αλ β−n
p1 (n) = , p2 (n) =
β2 β n(n − r)! n(n − r)! λ β
and
β (1−β )β
1− βn 1− βn eαλ eαλ
In (k) = (ϑk+1 ) En/β (ϑk+1 ) − (ϑk ) En/β (ϑk ), ϑk =
2βkβ β
Proof. Following the proof in Proposition 6.1, we can write
r ∞
β αλ
μr = W0r e−t dt
λ 0 β t 1/β
Setting t1 = ( e βαλ )β and dividing the integration interval, we can write μr = ( βλ )r(I1 + I2 ),
where
t1 ∞
αλ −t αλ
I1 = W0r
e dt, I2 = W0 r
e−t dt
0 β t 1/β t1 β t 1/β
αλ
Since 0 < β t 1/β
< e−1 for t1 < t < ∞, applying the expansion (3), we obtain
∞
r(−n)n−r 1− βn
I2 = t En/β (t1 ) (25)
n=r
n(n − r)! 1
αλ β
Next, we aim to find an expression for I1 . Note that, since the set {tk = ( ekβ ) : k ∈ N}
is enumerable, it can be neglected of the integration interval of I1 . Further, since
μr < ∞, the function w(t ) = W0r( β αt 1/β λ
)e−t is integrable in (0, t1 ). Let wk (t ) =
w(t ) I(tk+1 ,tk ) (t ), where I(tk+1 ,tk ) (t ) is the indicator function of the interval (tk+1 , tk ). Thus,
10628 C. J. TABLADA AND G. M. CORDEIRO
we have w(t ) = ∞ k=1 wk (t ) almost everywhere t ∈ (0, t1 ). Applying the dominated conver-
gence theorem gives
∞ t1
I1 = wk (t ) dt (26)
k=1 0
Since | arctan(y)| < π/2, then | eπ2 arctan(y)| < e−1 and, using the expansion (3), we
obtain
t1 ∞ yk+1
r(−n)n−r 2n αλ β β−n
wk (t ) dt = (eπ )
0 n=r
n(n − r)! β β yk
β
eπαλ [arctan(y)]n−β−1
× exp − dy
2 β arctan(y) (1 + y2 )
where ϑk = ( eπαλ
2β zk
)β .
Replacing (28) in Equation (27) and considering the expression (26) gives
∞ ∞
r(−n)n−r 2β αλ n n n
I1 = (ϑk+1 )1− β En/β (ϑk+1 ) − (ϑk )1− β En/β (ϑk ) (29)
k=1 n=r
n(n − r)! β 2 β
where
n−r r n
2β α n λ r(−n)n−r r(−n)n−r β 1−
p1 (n) = , p2 (n) = t1 β
β2 β n(n − r)! n(n − r)! λ
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10629
and
n n
In (k) = (ϑk+1 )1− β En/β (ϑk+1 ) − (ϑk )1− β En/β (ϑk )
dxr
Using the expansion (4) we obtain the following result.
Proposition 6.4. For λ > 0 and 0 < z ≤ βλ W0 (e−1 ), the incomplete mgf of X can be expanded
as
∞
α β −λz
Mz (s) = dn (z, s) En/β e
n=0
z
n s(s− λβn )n−1 α n
where dn (z, s) = bn (s)[( αz )β e−λz ]1− β and bn (s) = n!
.
Proof. Letting t = ( αx )β e−λx and following the proof of Proposition 6.1, we obtain
∞
βs αλ
Mz (s) = exp W0 − t dt
( αz )β e−λz λ βt 1/β
β
Then, for 0 < z ≤ λ
W0 (e−1 ), we have from (4)
∞ ∞
e−t
Mz (s) = bn (s) n/β
dt (30)
n=0 ( αz )β e−λz t
where
n−1
λn
s s− β
αn
bn (s) =
n!
Letting ν = n
β
and x = ( αz )β e−λz
in Equation (22) and replacing in (30), we obtain
∞
α β −λz
Mz (s) = dn (z, s) En/β e (31)
n=0
z
where
1− βn
α β −λz
dn (z, s) = bn (s) e
z
We can obtain an expansion for M(s) using expansion (4) and following the insights of the
proof in Proposition 6.3. The result is given in the following proposition.
10630 C. J. TABLADA AND G. M. CORDEIRO
where
β−n
(2α)β s(βs − nλ)n−1 s(βs − nλ)n−1 α n eαλ
p1 (n, s) = , p2 (n, s) =
β n+2 n! β n−1 n! β
and In (k) is given in Proposition 6.3.
∞
Proof. From the proof in Proposition 6.1, we have M(s) = 0 ν(t ) dt, where
βs αλ
ν(t ) = exp W0 −t
λ β t 1/β
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
αλ β
Setting tk = ( ekβ ) , k ∈ N, and following the proof in Proposition 6.3, we have M(s) = I1 +
I2 , where
∞ t1 ∞
I1 = νk (t ) dt, I2 = ν(t ) dt
k=1 0 t1
n−1 n
α
where p2 (n, s) = s(βs−nλ)
β n−1 n!
t11−n/β .
Following the same sequence of transformations as in the proof of Proposition 6.3, we can
write
t1 ∞
(2α)β s(βs − nλ)n−1 ϑk −ϑ −n/β
νk (t ) dt = e ϑ dϑ
0 n=0
β n+2 n! ϑk+1
∞ ∞
(2α)β s(βs − nλ)n−1
I1 = In (k)
k=1 n=0
β n+2 n!
(2α)β s(βs−nλ)n−1
where p1 (n, s) = β n+2 n!
.
Equations (23), (24), (31), and (32) are the main results of this section.
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10631
From Proposition 6.1, we obtain the following expansion for the Bonferroni curve.
αλ )β eβ
Corollary 7.1. For λ > 0 and 0 < π ≤ e−( β , the Bonferroni curve can be expanded as
1
∞
1− n
B(π ) =
an − log π β En/β − log π
π μ1 n=1
8. Order statistics
Let X(1) ≤ X(2) ≤ · · · ≤ X(n) be the order statistics of a random sample of size n from the
distribution F (x). Then, for m = 1, 2, . . . , n, the pdf of the mth order statistic, X(m) , is given
by
f (m) (x) = K F (x)m−1 [1 − F (x)]n−m f (x)
10632 C. J. TABLADA AND G. M. CORDEIRO
Let Q(m) (u) be (for 0 < u < 1) the qf of X(m) . Then, we obtain from (34)
where Q(·) is the qf of X. Thus, from (10) and (35), we can write the qfs of X(n) and X(1) in
closed form in terms of the Lambert W function.
It is possible, when μr < ∞, to obtain an expression for the rth ordinary moment of the
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
order statistics using a result given by Barakat and Abdelkader (2004) for the case of i.i.d.
random variables.
Thus, if μr < ∞, we can express the rth moment of the mth order statistic X(m) as (Silva
et al., 2010)
(r)
r n
j−1 n
μ(m) = E X(m) = (−1) j−n+m−1
I j (r)
j=n−m+1
n−m j
∞
where I j (r) = r 0 xr−1 [1 − F (x)] j dx.
In particular, for the MF distribution, we obtain
n ∞
j−1 n α β −λx j
μ(r)
(m) =r (−1) j−n+m−1
xr−1 1 − e−( x ) e dx
j=n−m+1
n−m j 0
n
n
n
n
e−λxi
β
f (θ; x) = n β log(α) − (β + 1) log(xi ) + log(β + λ xi ) − λ xi − α
i=1 i=1 xiβ i=1 i=1
(36)
The MLE θ̂ of θ can be obtained by maximizing (36) directly by using the
SAS (PROC NLMIXED), R (optim and MaxLik functions), and Ox program (sub-
routine MaxBFGS). Details for fitting univariate distributions using maximum likeli-
hood in R for censored or non censored data can be obtained at http://www.inside-
r.org/packages/cran/fitdistrplus/docs/mledist.
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10633
Alternatively, we can obtain the components of the score vector Uθ = (Uα , Uβ , Uλ ) and
set them to zero. They are given by
∂ nβ e−λxi n
Uα = f (θ; x) = − β α β−1 β
(37)
∂α α i=1 xi
∂ n 1
n
Uβ = f (θ; x) = n log(α) − log(xi ) +
∂β i=1 i=1
β + λxi
n
e−λxi
n
e−λxi log(xi )
β β
− α log(α) +α
i=1 xiβ i=1 xiβ
∂ n
xi n
e−λxi
n
Uλ = f (θ; x) = + αβ β−1
− xi (38)
∂λ i=1
β + λxi i=1 xi i=1
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
The MLE θ̂ = (α̂, β̂, λ̂) can be determined by setting Uα = Uβ = Uλ = 0 and by solving these
equations simultaneously. Because they cannot be solved in closed form, numerical iterative
methods, such as Newton–Raphson type algorithms, can be applied.
a
Under general regularity conditions, we have (θ̂ − θ) ∼ N3 (0, K(θ)−1 ), where K(θ) is the
a
3 × 3 expected information matrix and ∼ denotes the asymptotic distribution. For n large,
K(θ) can be approximated by the observed information matrix. This normal approximation
for the MLE θ̂ can be used for constructing approximate confidence intervals and for testing
hypotheses on the parameters α, β, and λ.
In many cases, it is of interest to perform inference about some parameters of the model
by assuming that the remaining parameters are known.
Proposition 9.1. Let x = (x1 , . . . , xn ) be a sample of size n from X ∼ MF(α, β, λ), with log-
likelihood for θ given by (36). Then
a) If β and λ are known, then the MLE of α always exists and is unique, and is given by
−1/β
1 e−λxi
n
α̂ =
n i=1 xiβ
b) Assume α and β known and let x(n) = max{x1 , . . . , xn }. If α > x(n) , then the MLE of λ
exists and is unique.
Proof. Suppose β and λ are known. From (37) and considering that α, β > 0 and n > 0, we
have
−1/β
1 e−λxi
n
Uα = 0 ⇐⇒ α = (39)
n i=1 xiβ
nβ e−λxi n
Uαα = − 2 − β(β − 1)α β−2 β
α i=1 xi
10634 C. J. TABLADA AND G. M. CORDEIRO
To see that the critical point given in (39) is a point of maximum, it has to satisfy Uαα < 0.
But
1 e−λxi
n
−β
Uαα < 0 ⇐⇒ α > (1 − β )
n i=1 xiβ
Since the critical point clearly satisfies this condition, we conclude that the MLE of α always
exists and is unique, and is given by (39). It proves (a).
To prove (b), suppose α and β are known. Differentiating Uλ with respect to λ, we obtain
n
xi α β e−λxi
Uλλ = − + β−2
i=1
(β + λxi )2 xi
Since each term in the above sum is positive, then Uλλ < 0. Further, from (38), we have
n
lim Uλ = − xi < 0 (40)
Downloaded by [Claudio Tablada] at 07:43 14 November 2017
λ→∞
i=1
and
n α β − xiβ
n
lim Uλ = + (41)
λ→0 β i=1 xiβ−1
Thus, if α > x(n) , then limλ→0 Uλ > 0, which ensures the existence and uniqueness of the MLE
of λ.
Conditions for the existence of the MLE of β, given that α and λ are known, are much more
difficult to obtain and, therefore, are omitted from the above result.
Suppose that the parameter vector is partitioned as θ = (ψ
1 , ψ 2 ) , where dim(ψ 1 ) +
dim(ψ2 ) = dim(θ). The likelihood ratio (LR) statistic for testing the null hypothesis H0 :
ψ1 = ψ1(0) against the alternative hypothesis H1 : ψ1 = ψ1(0) is given by LR = 2 { f (θ̂) −
f (θ̃)}, where θ̂ = (ψ̂1 , ψ̂2 ) , θ̃ = (ψ1(0) , ψ̃2 ) , ψ̂i and ψ̃i are the MLEs under the alter-
native and null hypotheses, respectively, and ψ1(0) is a specified parameter vector. Based on
a
the first-order asymptotic theory, we know that LR ∼ χk2 , where k = dim(ψ1 ). Thus, we can
test submodels of the MF distribution and analyze how significant are the parameters tested
for modeling a particular data set.
Table . Relative bias and MSE values of the MLE θ̂ = (α̂, β̂, λ̂).
Relative bias MSE
α β λ n α̂ β̂ λ̂ α̂ β̂ λ̂
11. Application
In this section, the potentiality of the MF distribution is illustrated by means of one applica-
tion. We use a data set corresponding to 202 observations of plasma ferritin concentration in
athletes (Weisberg, 2005, Sec. 6.4) and fit the Fréchet, EF, MOF, and MF distributions to these
data. The data are also used by Alizadeh et al. (2015). All computations are performed using
the R software (version 3.0.2, AdequacyModel package).
For maximizing the log-likelihood function (36), we use the BFGS method with numerical
derivatives. Initial values of α0 , β0 , and λ0 for the BFGS method can be obtained by compar-
ing the histogram of the data with the shape of the density for different parameter values, to
using a package such as Mathematica or Maple. Alternatively, based on Proposition 9.1,
it is possible to obtain initial values for the MF distribution by taking α0 > x(n) and giv-
ing arbitrary positive values for β0 and λ0 . For purposes of comparison, we compute some
goodness-of-fit statistics: Akaike information criterion (AIC), Bayesian information criterion
(BIC), Hannan–Quinn information criterion (HQIC), Cramér–von Mises criterion (W*), and
Anderson–Darling criterion (A*) (Chen and Balakrishnan, 1995). In general, the smaller the
values of these statistics are, the better the fit is.
10636 C. J. TABLADA AND G. M. CORDEIRO
Distribution α̂ β̂ λ̂
For completeness purposes, we also fit and include in the comparison the EW distribu-
tion (Mudholkar and Srivastava, 1993), since it is a widely used lifetime model. Its cdf and
pdf are given, respectively, by
x β λ βλ x β−1
x β λ−1 − x β
R(x) = 1 − e−( α ) , r(x) = 1 − e− ( α ) e (α )
α α
The MLEs are given in Table 2 with their standard errors in parentheses and the goodness-
of-fit values for the fitted distributions are listed in Table 3.
Since the Fréchet distribution is a submodel of the MF distribution, a comparison between
them can be conducted by considering the AIC, BIC, and HQIC statistics. However, since
the EW, EF, MOF, and MF distributions are non nested models, a comparison among them is
more appropriate by considering the W* and A* statistics. The figures in Table 3 reveal that the
MF distribution has the smallest values of the W* and A* statistics among the fitted models
and the smaller values of all statistics comparatively with the Fréchet distribution. Therefore,
we can conclude that the MF distribution gives the best fit to the current data. The plots in
Figure 6 display the Fréchet, EW, MOF, EF, and MF estimated densities. Based on these plots,
it is possible to assess the overall best fit of the MF and EW distributions.
A graphical analysis of the quality of fit can be assessed by means of Q-Q plots. So, in
Figure 7 are given Q-Q plots for all current distributions. From these plots, we can observe
that the MF distribution gives the best fit to the current data among the Fréchet, EF, and MOF
distributions, and a fit similar to that given by the EW distribution. This graphical analysis is
consistent with the quantitative analysis given by the goodness-of-fit statistics listed in Table 3.
To analyze how significant is the additional parameter λ of the MF distribution for mod-
eling the current data, we use the LR statistic, as discussed in Sec. 9, for testing the MF model
versus the Fréchet model. The results are given in Table 4. We note that the rejection of the
Table . LR test.
Models Hypotheses LR statistic p-Value
Figure . Q-Q plots for the Fréchet, EF, MOF, and EW distributions (solid circles) vs. Q-Q plot for the MF dis-
tribution (open circles).
10638 C. J. TABLADA AND G. M. CORDEIRO
null hypothesis is significant. So, we have evidence of the potential need for the inclusion of
the parameter λ in the Fréchet distribution for modeling the current data.
using the classical goodness-of-fit statistics and Q-Q plots. The results confirm that the new
distribution is very appropriate for applications to real lifetime data.
Funding
This research was partially supported by CNPq and CAPES, Brazil.
References
Afify, A.Z., Hamedani, G., Ghosh, I., Mead, M. (2015). The transmuted Marshall-Olkin Fréchet distri-
bution: Properties and applications. Int. J. Stat. Probab. 4:132–148.
Alizadeh, M., Cordeiro, G.M., Brito, E., Demétrio, C.G.B. (2015). The beta Marshall–Olkin family of
distributions. J. Stat. Distrib. Appl. 2:1–18.
Alves, I.F., Neves, C. (2011). Extreme value distributions. In Lovric, M. ed., International Encyclopedia
of Statistical Science (pp.493–496). Berlin, Heidelberg: Springer.
Barakat, H.M., Abdelkader, Y.H. (2004). Computing the moments of order statistics from nonidentical
random variables. Stat. Methods Appl. 13:15–26.
Barreto-Souza, W., Cordeiro, G.M., Simas, A.B. (2011). Some results for beta Fréchet distribution. Com-
mun. Stat. – Theory Methods 40:798–811.
Carrasco, J.M.F. , Ortega, E.M.M. , Cordeiro, G.M. (2008). A generalized modified Weibull distribution
for lifetime modeling. Comput. Stat. Data Anal. 53:450–462.
Chen, G., Balakrishnan, N. (1995). A general purpose approximate goodness-of-fit test. J. Qual. Tech-
nol., (27: 154–161).
Chiccoli, C., Lorenzutta, S., Maino, G. (1990). Recent results for generalized exponential integrals. Com-
put. Math. Appl. 19:21–29.
Corless, R.M., Gonnet, G.H., Hare, D.E.G., Jeffrey, D.J., Knuth, D.E. (1996). On the Lambert W function.
Adv. Comput. Math. 5:329–359.
da Silva, R.V., de Andrade, T.A.N., Maciel, D., Campos, R.P.S., Cordeiro, G.M. (2013). A new lifetime
model: The gamma extended Fréchet distribution. J. Stat. Theory Appl. 12:39–54.
Elbatal, I., Asha, G., Raja, A.V. (2014). Transmuted exponentiated Fréchet distribution: Properties and
applications. J. Stat. Appl. Probab. 3:379–394.
Harlow, D.G. (2002). Applications of the Fréchet distribution function. Int. J. Mater. Prod. Technol.
17:482–495.
Jodrá, P. (2010). Computer generation of random variables with Lindley or Poisson–Lindley distribu-
tion via the Lambert W function. Math. Comput. Simul. 81:851–859.
Kotz, S., Nadarajah, S. (2000). Extreme Value Distributions: Theory and Applications. London: World
Scientific.
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 10639
Krishna, E., Jose, K.K., Alice, T., Ristić, M.M. (2013). The Marshall-Olkin Fréchet distribution. Com-
mun. Stat. - Theory Methods 42:4091–4107.
Lai, C., Xie, M., Murthy, D. (2003). A modified Weibull distribution. IEEE Trans. Reliab. 52:33–37.
Mahmoud, M.R., Mandouh, R.M. (2013). On the transmuted Fréchet distribution. J. Appl. Sci. Res.
9:5553–5561.
Marshall, A., Olkin, I. (2007). Life Distributions: Structure of Nonparametric, Semiparametric, and Para-
metric Families. New York: Springer.
Mead, M.E.A. (2014). A note on Kumaraswamy Fréchet distribution. Aust. J. Basic Appl. Sci. 8:294–300.
Mudholkar, G.S., Srivastava, D.K. (1993). Exponentiated Weibull family for analyzing bathtub failure-
rate data. IEEE Trans. Reliab. 42:299–302.
Nadarajah, S., Gupta, A.K. (2004). The beta Fréchet distribution. Far East J. Theor. Stat. 14:15–24.
Nadarajah, S., Kotz, S. (2001). The exponentiated Fréchet distribution. Interstat Electr. J. 31:1–7.
Resnick, S.I. (2013). Extreme Values, Regular Variation and Point Processes. New York: Springer.
Silva, G.O., Ortega, E.M.M., Cordeiro, G.M. (2010). The beta modified Weibull distribution. Lifetime
Data Anal. 16:409–430.
Weisberg, S. (2005). Applied Linear Regression. Hoboken, New Jersey: John Wiley & Sons.
Yousof, H.M., Afify, A.Z., Alizadeh, M., Butt, N.S., Hamedani, G.G., Ali, M.M. (2015). The transmuted
Downloaded by [Claudio Tablada] at 07:43 14 November 2017