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

net/publication/271073842

The Kummer Beta Normal: A New Useful-Skew Model

Article  in  Journal of data science: JDS · April 2015

CITATIONS READS

4 265

2 authors:

Rodrigo Rossetto Pescim Saralees Nadarajah

Universidade Estadual de Londrina The University of Manchester
34 PUBLICATIONS   684 CITATIONS    930 PUBLICATIONS   9,352 CITATIONS   

SEE PROFILE SEE PROFILE

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

Order Statistics View project

Nuclear Catastrophe Risk Bonds in a Markov Dependent Environment View project

All content following this page was uploaded by Rodrigo Rossetto Pescim on 19 January 2015.

The user has requested enhancement of the downloaded file.

Journal of Data Science Vol.(2015),

The Kummer Beta Normal: A New Useful-Skew Model

Rodrigo R. Pescim1 and Saralees Nadarajah2

1 ESALQ - Universidade de São Paulo and 2 University of Manchester

Abstract: The normal distribution is the most popular model in applica-

tions to real data. We propose a new extension of this distribution, called
the Kummer beta normal distribution, which presents greater flexibility to
model scenarios involving skewed data. The new probability density function
can be represented as a linear combination of exponentiated normal pdfs. We
also propose analytical expressions for some mathematical quantities: Or-
dinary and incomplete moments, mean deviations and order statistics. The
estimation of parameters is approached by the method of maximum likeli-
hood and Bayesian analysis. Likelihood ratio statistics and formal goodness-
of-fit tests are used to compare the proposed distribution with some of its
sub-models and non-nested models. A real data set is used to illustrate the
importance of the proposed model.

Key words: Bayesian analysis, Kummer beta generalized distribution, Maxi-

mum likelihood method, Moment, Normal distribution, Order statistic.

1. Introduction

The main motivation for statisticians to study new families of statistical dis-
tributions is to increase the flexibility to better model various data sets that
cannot be properly fitted by the existing distributions. In many applied areas
such as environmental and medical sciences, engineering, demography, biologi-
cal studies, lifetime analysis, actuarial, economics, finance and insurance there
is a clear need for extended forms of these distributions. Exponentiated gen-
eralized (EG), beta generalized (BG) (Eugene et al., 2002) and Kumaraswamy
generalized (KwG) (Cordeiro and de Castro, 2011) families of distributions are
very versatile to analyze different types of data. These families have been widely
studied in statistics and some authors have developed several special EG, BG
and KwG models. For EG models, Mudholkar and Srivastava (1993) and Mud-
holkar et al. (1995) defined the exponentiated Weibull (EW) distribution, Gupta
et al. (1998) defined the exponentiated Pareto (EPa) distribution, Gupta and
144 The Kummer Beta Normal: A New Useful-Skew Model

Kundu (2001) defined the exponentiated exponential (EE) distribution, Nadara-

jah and Gupta (2007) defined the exponentiated gamma (EGa) distribution and
Cordeiro et al. (2011) defined the exponentiated generalized gamma (EGG) dis-
tribution. For BG models, Eugene et al. (2002) defined the beta normal (BN)
distribution, Nadarajah and Kotz (2004) defined the beta Gumbel (BGu) distri-
bution, Nadarajah and Gupta (2004) defined the beta Fréchet (BF) distribution,
Nadarajah and Kotz (2006) defined the beta exponential (BE) distribution and
more recently, Pescim et al. (2010) and Paranaı́ba et al. (2011) studied important
mathematical properties of the beta generalized half-normal (BGHN) and beta
Burr XII (BBXII) distributions, respectively. For KwG models, Cordeiro and de
Castro (2011) defined the Kumaraswamy normal (KwN) distribution, Cordeiro
et al. (2010) defined the Kumaraswamy Weibull (KwW) distribution, Pascoa et
al. (2011) defined the Kumaraswamy generalized gamma distribution, Cordeiro
et al. (2012b) defined the Kumaraswamy Gumbel (KwGu) distribution, Cordeiro
et al. (2012c) defined the Kumaraswamy generalized half-normal (KwGHN) dis-
tribution and more recently, Paranaı́ba et al. (2013) defined the Kumaraswamy
Burr XII (KwBXII) distribution. However, the beta, Kumaraswamy and expo-
nentiated generators do not provide flexibility to the extremes (right and left) of
the probability density functions (pdfs). For this reason, they are not suitable
for analyzing real data with high levels of asymmetry.
For an arbitrary baseline distribution G(x; γ) with parameter vector γ and pdf
g(x; γ), Pescim et al. (2012) proposed the Kummer beta generalized (denoted
by the prefix “KB-G” for short) family of distributions that provides greater
flexibility to extremes. Its cumulative distribution function (cdf) is defined by
∫ G(x;γ )
FKBG (x) = K ta−1 (1 − t)b−1 e−ct dt, (1)
0
where a > 0 and b > 0 are shape parameters which induce skewness, and thereby
promote weight variation of the tails, whereas the parameter −∞ < c < ∞
“squeezes” the pdf to the left or right, i.e., it gives weights to the extremes of the
pdfs. Here,
Γ(a)Γ(b)
K −1 = 1 F1 (a; a + b; −c)
Γ(a + b)
and
∫ 1 ∞
∑
Γ(a + b) −ct (a)k (−c)k
1 F1 (a; a + b; −c) = (1 − t)
a−1 b−1
t e dt =
Γ(a)Γ(b) 0 (a + b)k k!
k=0

is the confluent hypergeometric function (Abramowitz and Stegun, 1968), Γ(·) is

the gamma function and (d)k = d(d + 1) . . . (d + k − 1) denotes the ascending
factorial.
 R. R. Pescim and S. Nadarajah 145

The pdf corresponding to (1) can be expressed as

fKBG (x) = K g(x; γ) G(x; γ)a−1 [1 − G(x; γ)]b−1 exp [−c G(x; γ)] . (2)

Equation (2) will be most tractable when both G(x; γ) and g(x; γ) have simple
analytic expressions. Its major benefit is to offer more flexibility to extremes
(right and/or left) of the pdfs and therefore it becomes suitable for analyzing
data with high degree of asymmetry.
The KB-G class of distributions includes two important special cases: the
beta-generalized (BG) distribution for c = 0 and the exponentiated genera-lized
(EG) distribution for c = 0 and b = 1. Pescim et al. (2012), Cordeiro et al.
(2014), Pescim et al. (2014) and Pescim and Nadarajah (2015) defined the Kum-
mer beta Weibull (KBW), Kummer beta generalized gamma (KBGG), Kummer
beta Birnbaun-Saunders (KBBS) and Kummer beta gamma (KBGa) distribu-
tions by taking G(x) and g(x) to be the cdf and the pdf of the Weibull, gener-
alized gamma, Birnbaum-Saunders and gamma distributions, respectively. They
studied several mathematical properties of these distributions and showed clear
evidence of the potential of the three skewness parameters when modeling real
data.
The normal distribution is the most popular model in applications to real
data. When the number of observations is large, it can serve as an approxima-
tion for other models. Over the past decades, several authors have proposed new
generalizations based on the normal distribution for modeling real data sets; see,
for example, Azzalini (1985), Eugene et al. (2002), Nadarajah (2005), Cordeiro
and de Castro (2011), Cordeiro et al. (2012a), Nadarajah et al. (2014) and
Alzaatreh et al. (2014) for skew-normal, beta normal, generalized normal, Ku-
maraswamy normal, McDonald normal, modified beta normal and gamma normal
distributions, respectively. The emergence of such distributions in the statistics
literature is only very recent.
The cdf and the pdf of the normal distribution with location parameter −∞ <
µ < ∞ and scale parameter σ > 0 are given by
( )
x−µ
G(x; µ, σ) = Φ
σ
and
[ ] ( )
1 (x − µ)2 1 x−µ
g(x; µ, σ) = √ exp − = ϕ , (3)
2 πσ 2σ 2 σ σ
respectively.
In this paper, we propose the Kummer beta normal (denoted with the pre-
fix “KBN”) distribution and provide a comprehensive description of some of its
146 The Kummer Beta Normal: A New Useful-Skew Model

mathematical properties with the hope that it will attract wider applications in
many areas of research. The main motivation for this extension is that the new
distribution is a highly flexible distribution which admits different degrees of kur-
tosis and asymmetry. The normal distribution represents only a special case of
the KBN distribution.
The cdf and the pdf of the KBN distribution are obtained from equations (1)
and (2) as
∫ Φ( x−µ )
σ
F (x) = K ta−1 (1 − t)b−1 exp(−c t)dt, (4)
0

and
( )[ ( )] [ ( )] [ ( )]
K x−µ x − µ a−1 x − µ b−1 x−µ
f (x) = ϕ Φ 1−Φ exp −c Φ , (5)
σ σ σ σ σ
where x ∈ R, µ ∈ R is a location parameter, σ > 0 is a scale parameter, a and b are
positive shape parameters, and c ∈ R is a real-valued shape parameter. Hereafter,
we denote by X a random variable following (5), and write X ∼ KBN(a, b, c, µ, σ).
For µ = 0 and σ = 1, we have the standard KBN distribution. This pdf has
three shape parameters a, b and c allowing for a high degree of flexibility. The
parameter c controls tail weights to the extremes of the distribution.
The study of the new distribution is important since it extends some distri-
butions previously considered in the literature. In fact, the normal distribution
(with parameters µ and σ) is clearly a basic exemplar for a = b = 1 and c = 0,
with a continuous crossover towards distributions with different shapes (e.g., a
specified combination of skewness and kurtosis). The KBN distribution contains
as sub-models the beta normal (BN) distribution for c = 0 and the exponentiated
normal (EN) distribution for c = 0, b = 1. Plots of the KBN pdf for selected
parameter values are displayed in Figure 1. It is evident that the shapes of the
new pdf are much more flexible than its sub-models.
The article is outlined as follows. In Section 2, we provide useful expansions
for the pdf of the KBN distribution. We obtain explicit expressions for the
ordinary and incomplete moments (Section 3) and order statistics (Section 4). In
Section 5, we discuss some statistical inference like maximum likelihood method
and Bayesian approach. A real data application given in Section 6 reveals the
usefulness of the new distribution for analyzing real data. Concluding remarks
are addressed in Section 7.

2. Useful expansions

Expansions for equations (4) and (5) can be derived using the concept of ex-
ponentiated distributions. Consider the exponentiated normal (EN) distribution
 R. R. Pescim and S. Nadarajah 147
(a) (b)

0.8
1.2

c = −10
1.0

c = −5
c=0

0.6
c = −20
c = −10 c=5
0.8

c=0 c = 10
c = 10
c = 15

f(x)
f(x)

0.4
0.6
0.4

0.2
0.2

0.0
0.0

−2 −1 0 1 2 3 −4 −2 0 2 4

x x

(c)
0.0020

c = −1
c=0
0.0015

c=1
0.0010
f(x)

0.0005
0.0000

−10 −5 0 5

Figure 1: Plots of the KBN pdf for some parameter values: (a) KBN(8, 2, c, 0, 1),
(b) KBN(1.5, 2, c, 0, 1) and (c) KBN(0.1, 0.1, c, 0, 1) pdfs (the red lines represent
the BN pdfs).

with power parameter a > 0 defined by Y ∼ EN(a, µ, σ), with the cdf and the pdf
( )a ( ) ( y−µ )a−1
given by H(y; a) = Φ y−µ
σ and h(y; a) = σa ϕ y−µ
σ Φ σ , respectively.
[ ( )]
By expanding the term exp −c Φ x−µ σ and using the binomial in equation
148 The Kummer Beta Normal: A New Useful-Skew Model

(5), we obtain the linear combination (for a > 0 integer)

∞
∑
f (x) = wj,k h(x; a + j + k, µ, σ), (6)
j,k=0

where h(x; a + j + k, µ, σ) denotes the EN(a + j + k, µ, σ) pdf and the coefficient

wj,k is given by
( )
K(−1)j+k cj b − 1
wj,k = .
j!(a + j + k) k

By integrating (6), we obtain

∞
∑ [ ( )]
x − µ a+j+k
F (x) = wj,k Φ . (7)
σ
j,k=0

[ ( )]a+j+k
If a is a positive non-integer, we can expand Φ x−µ
σ as
[ ( )] ∞ [ ( )]
x − µ a+j+k ∑ x−µ r
Φ = sr (a + j + k) Φ , (8)
σ σ
r=0

where
∞
∑ ( )( )
k+r m k
sr (m) = (−1) .
k r
k=r

Thus, from equations (3), (7) and (8), the KBN cdf can be expressed as
∞ ∑
∑ ∞ [ ( )]
x−µ r
F (x) = wj,k sr (a + j + k) Φ . (9)
σ
r=0 j,k=0

By differentiating (9) and changing indices, we can obtain

∞
∑
f (x) = br h(x; r, µ, σ), (10)
r=0

∞
∑
where br = wj,k sr (a + j + k). Equation (10) reveals that the KBN pdf is
j,k=0
a linear combination of EN pdfs. So, several properties of the KBN distribution
can be obtained by knowing those properties of the EN distribution.
 R. R. Pescim and S. Nadarajah 149

3. Ordinary and incomplete moments

Hereafter, let X denote the KBN(a, b, c, µ, σ) random variable. The sth mo-
ment of X for µ = 0 and σ = 1 can be expressed from (10) as
∞
∑ ∫ ∞
µ′s = E (X s ) = br xs h(x; r, 0, 1)dx
r=0 −∞

and then
∞
∑
µ′s = b∗r τs,r , (11)
r=0
∫ ∞
where b∗r = r br and τs,r = xs ϕ(x) Φ(x)r−1 dx is the (s, r)th probability
−∞
weighted moment (PWM) (for s and r positive integers) of the normal distribu-
tion.
Nadarajah (2008) demonstrated that the (s, r)th PWM of the normal distri-
bution can be expressed in terms of the Lauricella function of type A (Exton,
1978) as
s
∑
r−1 ( )( ) ( )
2 2 +1−r r−1 2 p p+s+1
τs,r = √ √ Γ
π p=0,p+s even p π 2
( )
(p) p + s + 1 1 1 3 3
× FA ; , . . . , ; , . . . , ; −1, . . . , −1 ,
2 2 2 2 2

where
(n)
FA (a; b1 , . . . , bn ; c1 , . . . , cn ; x1 , . . . , xn ) =
∞
∑ ∞
∑ (a)m1 +···+mn (b1 )m1 · · · (bn )mn xm1
1 . . . xn
mn
···
(c1 )m1 · · · (cn )mn m1 ! . . . mn!
m1 =0 mn =0

is the Lauricella function of type A and the Pochhammer symbol (a)k = a(a +
1) . . . (a + k − 1) indicates the kth rising factorial power of a with the convention
(a)0 = 1.
The skewness and kurtosis measures can be calculated from the ordinary
moments using well-known relationships. Plots of the skewness and kurtosis of
the KBN distribution as a function of c for selected values of a and b for µ = 0
and σ = 1.0 are displayed in Figures 2 and 3. Figures 2a and 2b immediately
indicate that the additional parameter c promotes high levels of asymmetry.
150 The Kummer Beta Normal: A New Useful-Skew Model
(a) (b)

0
0

−10
−10

−20
skewness

skewness
−20

−30
−30

b=1

−40
a = 1.0 b = 1.5
a = 1.2 b=2
a = 1.5 b = 2.5

−50
−40

a = 1.6

1 2 3 4 5 1 2 3 4 5

c c

Figure 2: Skewness of the KBN distribution as a function of c for some values of

a and b for µ = 0 and σ = 1.0. (a) b = 1.5 and (b) a = 1.5.

(a) (b)
0

a = 1.0
a = 1.2 b = 1.5
−1

a = 1.4 b=2
a = 1.5 b = 2.5
b = 2.8
−2
−2
kurtosis

kurtosis
−3

−4
−4
−5

−6
−6

1 2 3 4 5 1 2 3 4 5

c c

Figure 3: Kurtosis of the KBN distribution as a function of c for some values of

a and b for µ = 0 and σ = 1.0. (a) b = 1.5 and (b) a = 1.5.

The sth incomplete moment of X is defined by ms (y) = E (X s IX<y ) =

 R. R. Pescim and S. Nadarajah 151
∫ y
xs f (x)dx. Consider the case µ = 0 and σ = 1. Based on equation (10),
−∞
ms (y) reduces to
∞
∑ ∫ y
ms (y) = b∗r xs ϕ(x) Φ(x)r−1 dx, (12)
r=0 −∞

where b∗r is defined in (11).

∑ ∞ ( √ )−1
We can write Φ(x) as a power series Φ(x) = ηm xm , where η0 = 1 + 2/π /2,
[√ ] m=0
η2m+1 = (−1)m / 2π 2m (2m + 1) m! for m = 0, 1, 2, . . . and η2m = 0 for
m = 1, 2, . . .. Using an identity given by Gradshteyn and Ryzhik (2007) for a
power series raised to a positive integer j,
(∞ )j ∞
∑ ∑
ai xi = cj,i xi , (13)
i=0 i=0

where the coefficients cj,i (for i = 1, 2, . . .) are easily obtained from the recurrence
equation

−1
∑
i
cj,i = (ia0 ) [m(j + 1) − i] am cj,i−m (14)
m=1

and cj,0 = aj0 .

Further, using (13), we have
∞
∑
Φ(x)r−1 = ηr−1,m xm , (15)
m=0

where the coefficients ηr−1,m can be determined from the recurrence equation
(14). Thus, using (15) and changing variable in integral (12), the sth incomplete
moment of X is given by
 ∞ [ ( )]
 ∑ ( s+m ) 1 s+m−1 s + m + 1 y2

 ∗
−√ 2


br ηr−1,m E X 2 Γ
2
,
2
, if y > 0,
r,m=0
2π
ms (y) = ∞ ( ) (16)

 1 ∑ ∗ s + m + 1 y2

s+m−1

 √2π if y ≤ 0,
m+s
 br ηr−1,m (−1) 2 2 Γ , ,
2 2
r,m=0
∫ ∞
where Γ(a, x) = ta−1 e−t dt denotes the complementary incomplete gamma
x
function.
152 The Kummer Beta Normal: A New Useful-Skew Model

Consider the case µ = 0 and σ = 1. We can derive the mean deviations of X

about the mean µ′1 and about the median M (δ2 ) in terms of the first incomplete
∫ Φ(M )
moment. The median can be obtained by inverting F (M ) = K ta−1 (1 −
0
t)b−1 e−ct dt = 1/2 numerically. They can be expressed as
[ ( ) ( )]
δ1 = 2 µ′1 F µ′1 − m1 µ′1 , δ2 = µ′1 − 2m1 (M ) ,
where m1 (·) is the first incomplete moment of X given by (16) with s = 1. We
have
 ∞ [ ( )]
 ∑ ( 1+m ) 1 m m + 2 y2

 ∗
− √ 22Γ


br ηr−1,m E X
2
,
2
, if y > 0,
r,m=0
2π
m1 (y) = ∞ ( ) (17)

 1 ∑ ∗ m + 2 y2
 m+1 m
 √2π
 br ηr−1,m (−1) 2 Γ , , if y ≤ 0.
2
2 2
r,m=0

The measures δ1 and δ2 can be calculated from (17) by setting y = µ′1 and
y = M , respectively. An application of the mean deviations is to the Lorenz
and Bonferroni curves defined by L(π) = m1 (q)/µ′1 and B(π) = m1 (q)/πµ′1 ,
respectively, where q = F −1 (π) can be computed for a given probability π by
inverting (4) numerically. These curves have applications in several fields and
can be calculated from equation (17).

4. Order statistics

Order statistics have been used in a wide range of problems, including robust
statistical estimation and detection of outliers, characterization of probability
distributions and goodness-of-fit tests, entropy estimation, analysis of censored
samples, reliability analysis, quality control and strength of materials.
Suppose Z1 , . . . , Zn is a random sample from the standard KBN distribution
and let Z1:n < · · · < Zi:n denote the corresponding order statistics. The pdf
fi:n (z) of the ith order statistic can be written as
∑
n−i ( )
f (z) n! j n−i
fi:n (z) = (−1) F (z)i+j−1 .
(i − 1)!(n − i)! j
j=0

We now demonstrate that fi:n (z) can be written as a linear combination of stan-
dard EN pdfs. First, we provide an expansion for the cdf of the standard KBN
distribution. Using (9) and (10), the pdf of the ith order statistic, Zi:n , can be
expressed as
∑
n−i ( ) [∑
∞
][ ∞
∑
]i+j−1
n! n − i
fi:n (z) = (−1)j br r ϕ(z) Φ(z)r−1 br Φ(z)r .
(i − 1)!(n − i)! j
j=0 r=0 r=0
 R. R. Pescim and S. Nadarajah 153

From equation (13), we obtain

[∞ ]i+j−1 ∞
∑ ∑
r
br Φ(z) = bi+j−1,r Φ(z)r ,
r=0 r=0

where bi+j−1,0 = bi+j−1

0 and

∑
r
bi+j−1,r = (rb0 )−1 [m (i + j) − r] bm bi+j−1,r−m .
m=1

Hence, the pdf of the ith order statistic for the standard KBN distribution can
be expressed as

∑
n−i ( )∑ ∞
∞ ∑
n! j n−i r
fi:n (z) = (−1) bi+j−1,s br h(z; r + s).(18)
(i − 1)!(n − i)! j r+s
j=0 r=0 s=0

Equation (18) is the main result of this section. It gives the pdf of the standard
KBN order statistics as a linear combination of standard EN pdfs. So, several
mathematical quantities of standard KBN order statistics like ordinary moments,
generating function, and mean deviations follow immediately from those quanti-
ties of the standard EN distribution.

5. Inference and estimation

5.1 Maximum likelihood method

In this section, the estimation of the model parameters of the KBN distri-
bution will be investigated by maximum likelihood. Let X = (X1 , . . . , Xn )
be a random sample from this distribution with unknown parameter vector
θ = (a, b, c, µ, σ)T . The total log-likelihood function for θ is

[ ( )]
1 ∑ ∑
n n
n 2 xi − µ
ℓ(θ) = n log(K) − log(2π) − n log(σ) − 2 (xi − µ) + (a − 1) log Φ
2 2σ σ
i=1 i=1
∑
n [ ( )] ∑n ( )
xi − µ xi − µ
+(b − 1) log 1 − Φ −c Φ . (19)
σ σ
i=1 i=1
154 The Kummer Beta Normal: A New Useful-Skew Model

The elements of score vector are

 [ ( )2 ] 
∑n ∑n  exp − 1 xi −µ 
nµ 1 a−1 2 σ
Uµ (θ) = − 2 + 2 xi − √ ( xi −µ )
σ σ
i=1
σ 2π i=1  Φ σ 
 [ ( ) ]  [
1 xi −µ 2  ( ) ]
b − 1 ∑  exp − 2 c ∑
n n
σ 1 xi − µ 2
+ √ [ ( )] + √ exp − ,
σ 2π i=1  1 − Φ xiσ−µ  σ 2π
i=1
2 σ

 [ ( )2 ] 
∑n ∑n  (xi − µ) exp − 1 xi −µ 
n 1 a−1 2 σ
Uσ (θ) = − + 3 (xi − µ)2 − √ ( xi −µ )
σ σ
i=1
σ 2 2π i=1  Φ σ 
 [ ( ]
) 
1 xi −µ 2 
b − 1 ∑  (xi − µ) exp − 2
n
σ
+ √ [ ( )]
σ 2 2π i=1  1 − Φ xiσ−µ 
[ ( ) ]
c ∑n
1 xi − µ 2
+ √ (xi − µ) exp − ,
σ 2 2π 2 σ
i=1

[ ( )]
n ∂K ∑
n
xi − µ
Ua (θ) = + log Φ ,
K ∂a σ
i=1

[ ( )]
n ∂K ∑
n
xi − µ
Ub (θ) = + log 1 − Φ ,
K ∂b σ
i=1

and
( )
n ∂K ∑
n
xi − µ
Uc (θ) = − Φ ,
K ∂c σ
i=1

where the partial derivatives of K in relation to a, b and c are

{ }
∂ 1 F1 (a,a+b,−c)
∂K [ψ(a) − ψ(a + b)] F
1 1 (a, a + b, −c) + ∂a
=− 2 ,
∂a B(a, b) [1 F1 (a, a + b, −c)]

{ }
∂ 1 F1 (a,a+b,−c)
∂K [ψ(b) − ψ(a + b)] 1 F1 (a, a + b, −c) + ∂b
=− 2 ,
∂b B(a, b) [1 F1 (a, a + b, −c)]
 R. R. Pescim and S. Nadarajah 155
∂K a 1 F1 (a + 1, a + b + 1, −c)
= ,
∂c (a + b)B(a, b)1 F1 (a, a + b, −c)

∂ 1 F1 (a, a + b, −c)
= − [ψ(a) − ψ(a + b)] 1 F1 (a, a + b, −c)
∂a
∞
∑ (a)k (−c)k
− [ψ(a + b + k) − ψ(a + k)] ,
k!(a + b)k
k=0

and
∞
∑ (a)k (−c)k
∂ 1 F1 (a, a + b, −c)
= ψ(a + b) 1 F1 (a, a + b, −c) + ψ(a + b + k),
∂b k!(a + b)k
k=0

where ψ(x) = d log Γ(x)/dx denotes the digamma function.

Maximization of (19) can be performed by using well established routines like
the nlm routine or optimize in the R statistical package. Setting these equations
to zero, U(θ) = 0, and solving them simultaneously yields the maximum likeli-
hood estimate (MLE) θ b of θ. These equations cannot be solved analytically and
statistical software can be used to solve them numerically by means of iterative
techniques like the Newton-Raphson algorithm.
For interval estimation and hypothesis tests on the parameters in θ, we require
the 5 × 5 total observed information matrix J(θ) = − {Urs }, where the elements
Urs for r, s = µ, σ, a, b, c (can be obtained numerically. The estimated asymptotic
( )−1 )
multivariate normal N5 0, J θ b b can be used to construct
distribution of θ
approximate confidence regions for the parameters. An asymptotic confidence
interval with significance level γ for each parameter θr is given by
( √ √ )
ACI (θr , 100(1 − γ)%) = θbr − zγ/2 κ bθr ,θr , θbr + zγ/2 κbθr ,θr ,

where κ b for r = 1, 2, 4, 4, 5,
bθr ,θr is the rth diagonal element of J(θ)−1 estimated at θ
and zγ/2 is the 1 − γ/2 quantile of the standard normal distribution.
The likelihood ratio (LR) statistic is useful for comparing the new distribution
with some of its special models. For example, we may adopt the LR statistic to
check if the fit using the KBN distribution is statistically “superior” to a fit
using the normal distribution for a given data set. In any case, considering the
( )T (0)
partition θ = θ T1 , θ T2 , tests of hypotheses of the type H0 : {θ 1(=)θ 1 ( versus
)}
(0)
HA : θ 1 ̸= θ can be performed using the LR statistic w = 2 ℓ θ − ℓ θ , b e
1
where θb and θ
e are the estimates of θ under HA and H0 , respectively. Under the
d
null hypothesis H0 , w → χ2q , where q is the dimension of the vector θ 1 of interest.
156 The Kummer Beta Normal: A New Useful-Skew Model

The LR test rejects H0 if w > ξγ , where ξγ denotes the upper 100γ% point of
the χ2q distribution.

5.2 Bayesian inference

As it is well-known, the Bayesian approach allows for the incorporation of

previous knowledge of the parameters through informative prior pdfs. When this
information is not available, we can consider a non-informative prior. In the
Bayesian context, the information referring to the model parameters is obtained
through a posterior marginal distribution. Two difficulties usually arise. The first
one refers to obtaining the marginal posterior distribution, and the second to the
calculation of the moments of interest. Both cases require numerical integration
that, many times, do not present an analytical solution. To overcome these
problems, we use the simulation methods based on the Markov Chain Monte
Carlo (MCMC), like the Gibbs sampler and Metropolis-Hastings algorithms.
Since we have no prior information from historical data or from previous
experiments, we assign conjugate but weakly informative prior distributions to
the parameters. Since we assume an informative (but weakly) prior distribution,
the posterior distribution is a well-defined proper distribution. We assume that
the parameters (a, b, c, µ and σ) have independence priors and consider that the
joint prior distribution of all unknown parameters has a pdf given by
π(a, b, c, µ, σ) ∝ π(a) · π(b) · π(c) · π(µ) · π(σ), (20)
where a ∼ Ga((a1 , b1 )) with a1 and b1 known, b ∼ Ga((a2 , b2)) with a2 and b2
known, c ∼ N µ0 , σ02 with µ0 and σ02 known, µ ∼ N µ1 , σ12 with µ1 and σ12
known, and σ ∼ Ga (a3 , b3 ) with a3 and b3 known, where Ga (ai , bi ) denotes the
gamma
( distribution
) with mean ai /bi , variance ai /b2i for ai > 0 and bi > 0, and
N µi , σi represents the normal distribution with mean µi , variance σi2 for µi ∈ R
2

and σi2 > 0. We note that gamma and normal priors are the most commonly
used priors for positive and real-valued parameters.
Combining the likelihood function (19) and the prior distribution (20), the
joint posterior distribution for a, b, c, µ and σ reduces to
( )n [ ( ) n ( ) ]
K ∑n
xi − µ 1 ∑ xi − µ 2
π(a, b, c, µ, σ|x) ∝ exp −c Φ −
σ σ 2 σ
i=1 i=1
∏
n [ ( )] [ ( )]
xi − µ a−1 xi − µ b−1
· Φ 1−Φ
σ σ
i=1
· π(a, b, c, µ, σ). (21)
The joint posterior pdf (21) is analytically intractable because the integration
of the joint posterior pdf is not easy to perform. So, the inference can be based on
 R. R. Pescim and S. Nadarajah 157

MCMC simulation methods like the Gibbs sampler and Metropolis-Hastings al-
gorithm, which can be used to draw samples, from which features of the marginal
distributions of interest can be inferred. In this direction, we first obtain the full
conditional distributions of the unknown quantities given by
n [ (
∏ )]
xi − µ a−1
π(a|x, b, c, µ, σ) ∝ K n
Φ · π(a),
σ
i=1

n [
∏ ( )]
xi − µ b−1
π(b|x, a, c, µ, σ) ∝ K n
1−Φ · π(b),
σ
i=1

[ ( )]
∑
n
xi − µ
π(c|x, a, b, µ, σ) ∝ K n
exp −c Φ · π(c),
σ
i=1

[ ) ( n ( ) ]
∑
n
xi − µ 1 ∑ xi − µ 2
π(µ|x, a, b, c, σ) ∝ exp −c Φ −
σ 2 σ
i=1 i=1
n [ (
∏ )] [ ( )]
xi − µ a−1 xi − µ b−1
· Φ 1−Φ · π(µ),
σ σ
i=1

and
[ ( ) n ( ) ]
1 ∑n
xi − µ 1 ∑ xi − µ 2
π(σ|x, a, b, c, µ) ∝ exp −c Φ −
σn σ 2 σ
i=1 i=1
∏n [ ( )] [ ( )]
xi − µ a−1 xi − µ b−1
· Φ 1−Φ · π(σ).
σ σ
i=1

Since the full conditional distributions do not have explicit expressions, we

require the use of the Metropolis-Hastings algorithm to generate the variables a,
b, c, µ and σ for the KBN distribution.

6. Application - INPC data

This section contains an application of the Kummer beta normal (KBN) dis-
tribution to a real data referred to as INPC data. We shall compare the fits of
the KBN distribution with those of two sub-models (the beta normal (BN) and
normal distributions) and also to the following non-nested models: the McDonald
Normal (McN) (Cordeiro et al., 2012a), the gamma-normal (GN) (Alzaatreh et
158 The Kummer Beta Normal: A New Useful-Skew Model

al., 2014) and the modified beta normal (MBN) (Nadarajah et al., 2014) distri-
butions.
The INPC is a national index of consumer prices of Brazil, released by IBGE
(Brazilian Institute of Geography and Statistics). The period of collection goes
from day 01 to 30 of the reference month and the target population includes
families dwelling in the urban areas, whose head of the household is considered the
main employee. The survey was conducted in the metropolitan regions of Belém,
Belo Horizonte, Brası́lia, Curitiba, Fortaleza, Goiânia, Porto Alegre, Recife, Rio
de Janeiro, São Paulo and Salvador. The data set was extracted from the IBGE
database available at http : //www.ibge.gov.br and reported by De Morais (2009).
Table 1 presents a descriptive summary for the INPC data set and suggests a
skewed distribution with high degrees of skewness and kurtosis.

Table 1: Descriptive statistics for the INPC data set.

Mean Median SD Variance Skewness Kurtosis Min. Max.
0.64 0.50 0.60 0.36 1.56 6.59 -0.49 3.39

(i) Maximum likelihood estimation

Table 2 gives the MLEs and the corresponding SEs (given in parentheses) of
the model parameters and the values of the following statistics for some models:
Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC).
The computations were performed using the statistical software R. The AIC and
BIC values for the KBN model are the smallest values among those fitted sub-
models and non-nested models.
A formal test of the need for the third skewness parameter in KB-G distri-
butions can be based on the LR statistics. The results of this test are shown in
Table 3 for the INPC data set. We reject the null hypotheses of the LR test in
favor of the KBN distribution. The rejection is extremely highly significant and
it gives clear evidence of the potential need for the three skewness parameters
when modeling real data.
Figure 4 displays the histogram of the data and fitted pdfs of the KBN dis-
tribution, its sub-models and non-nested models. Further, Figure 5 plots the
empirical cdf and estimated cdfs of the KBN distribution, its sub-models and
non-nested models. We note that the KBN distribution produces better fit than
the other models.
We also apply formal goodness-of-fit tests in order to verify which distribution
fits the data better. We consider the Cramér-Von Mises (W ∗ ) and Anderson-
Darling (A∗ ) statistics. In general, the smaller the values of the statistics, W ∗
 R. R. Pescim and S. Nadarajah 159

Table 2: MLEs and the corresponding SEs (given in parentheses) of the model
parameters for the INPC data and the measures AIC and BIC.
Model µ σ a b c AIC BIC
KBN 0.4514 0.5806 4.3419 0.2658 9.4879 238.5 253.8
(0.4378) (0.1646) (1.8986) (0.0035) (0.0025)
BN -0.4391 0.4686 5.3041 0.2905 0 256.0 268.3
(0.1590) (0.0028) (2.4669) (0.0431) (-)
Normal 0.6442 0.5988 1 1 0 288.5 294.6
(0.0477) (0.0337) (-) (-) (-)
Gamma-normal µ σ α β
0.3098 0.3771 0.7881 3.3616 (-) 278.2 290.4
(0.2000) (0.0028) (0.2176) (0.4486) (-)
McN µ σ a1 b1 c1
-1.2530 0.5993 13.9336 0.2858 3.8102 251.1 266.4
(0.0205) (0.0178) (0.0631) (0.0307) (0.0412)
MBN µ σ a2 b2 β1
-0.3192 0.4578 5.9186 0.2954 1.70004 260.1 275.3
(0.0037) (0.0037) (4.0527) (0.0272) (1.3441)

Table 3: LR statistics for the INPC data.

Model Hypotheses Statistic w p-value
KBN vs BN H0 : c = 0 vs H1 : H0 is false 19.53 < 0.0001
KBN vs Normal H0 : a = b = 1 and c = 0 vs H1 : H0 is false 55.99 < 0.0001

( )
b be the cdf, where the form
and A∗ , the better the fit to the data. Let F x; θ
of F is known but θ b (a k-dimensional parameter vector, say) is unknown. To
obtain the statistics, W ∗ and A∗ , we proceed as follows:
( )
(i) compute vi = F xi ; θ b , where the xi ’s are in ascending order, yi = Φ−1 (vi )
is the standard normal quantile function and ui = Φ {(yi − y) /sy }, where
∑n ∑
n
y = n−1 yi and s2y = (n − 1)−1 (yi − y)2 ;
i=1 i=1

(ii) compute

∑
n
W2 = {ui − (2i − 1)/(2n)}2 + 1/(12n)
i=1
160 The Kummer Beta Normal: A New Useful-Skew Model
(a) (b)
1.2

1.2
1.0

1.0
KBN KBN
BN GN
0.8

0.8
Normal McN
MBN
f(x)

f(x)
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0
0 1 2 3 0 1 2 3

x x

Figure 4: (a) Estimated pdfs of the KBN distribution and its sub-models. (b)
Estimated pdfs of the KBN, GN, McN and MBN models.

(a) (b)
1.0

1.0
0.8

0.8
0.6

0.6
F(x)

F(x)

KBN KBN
0.4

0.4

BN GN
Normal McN
MBN
0.2

0.2
0.0

0.0

0 1 2 3 0 1 2 3

x x

Figure 5: (a) Empirical and estimated cdfs of the KBN distribution and its sub-
models. (b) Empirical and estimated cdfs of the KBN, GN, McB and MBN
models.
 R. R. Pescim and S. Nadarajah 161

and
∑
n
A2 = −n − n−1 {(2i − 1) log (ui ) + (2n + 1 − 2i) log (1 − ui )} ;
i=1
( )
(iii) modify W 2 into W ∗ = W 2 (1+0.5/n) and A∗ into A∗ = A2 1 + 0.75/n + 2.25/n2 .
For further details, the reader is referred to Chen and Balakrishnan (1995). The
values of the statistics, W ∗ and A∗ , for the distributions are given in Table
4. Overall, by comparing the measures of these formal goodness-of-fit tests in
Table 4, we conclude that the KBN distribution outperforms all the distributions
considered in this study. So, the proposed distribution can yield better fits than
the normal, BN, GN, McN and MBN distributions and therefore may be an
interesting alternative to these distributions for modeling skewed data sets. These
results illustrate the potentiality of the new distribution and the necessity for
additional shape parameters.

Table 4: Formal goodness-of-fit tests for the INPC data.

Model Statistic
W∗ A∗
KBN 0.0715 0.5513
BN 0.2465 1.5050
Normal 0.7635 4.4915
GN 0.5742 3.4008
McN 0.1401 0.8813
MBN 0.2699 1.6412

The QQ plots of the normalized quantile residuals was introduced by Dunn

and Smyth (1996) and more recently used by Cordeiro et al. (2013). Figures 6
and 7 indicate the improved fit achieved using the KBN distribution over other
distributions. We also emphasize the gain yielded by the KBN distribution in
relation to the normal, BN, GN, McN and MBN distributions.
(ii) Bayesian analysis
For the INPC data set, the following independent priors were considered to
perform the Metropolis-Hastings algorithm: a ∼ Ga(0.001, 0.001), b ∼ Ga(0.001, 0.001),
c ∼ N(0, 1000), µ ∼ N(0, 1000) and σ ∼ Ga(0.001, 0.001), so that we have a vague
prior distribution. Considering these prior pdfs, we generated two parallel inde-
pendent runs of the Metropolis-Hastings with size 300,000 for each parameter.
Disregarding the first 30,000 iterations to eliminate the effect of the initial values
and, to avoid correlation problems, we considered a spacing of size 10, obtain-
ing a sample of size 27,000 from each chain. To monitor the convergence of the
162 The Kummer Beta Normal: A New Useful-Skew Model

(a) (b)
3

4
2
Normalized quantile residuals

Normalized quantile residuals

3
1

2
0

1
−1

0
−2

−1
−3

−2

−2 −1 0 1 2 −2 −1 0 1 2

N(0,1) quantiles N(0,1) quantiles

(c)
4
3
Normalized quantile residuals
2
1
0
−1
−2
−3

−2 −1 0 1 2

N(0,1) quantiles

Figure 6: QQ plot of the normalized quantile residuals with an identity line for
the distributions: (a) KBN, (b) Normal and (c) BN.
 R. R. Pescim and S. Nadarajah 163

(a) (b)
4

3
Normalized quantile residuals
Normalized Quantile Residuals
3

2
2

1
0
1

−1
0

−2
−1

−3
−2

−2 −1 0 1 2 −2 −1 0 1 2

N(0,1) quantiles N(0,1) quantiles

(c)
4
3
Normalized Quantile Residuals
2
1
0
−1
−2
−3

−2 −1 0 1 2

N(0,1) quantiles

Figure 7: QQ plot of the normalized quantile residuals with an identity line for
the distributions: (a) GN, (b) McN, (c) MBN.
164 The Kummer Beta Normal: A New Useful-Skew Model

Metropolis-Hastings, we performed the methods suggested by Cowles and Carlin

(1996). To monitor the convergence of the samples, we used the between and
within sequence information, following the approach developed in Gelman and
b In all cases, these values
Rubin (1992) to obtain the potential scale reduction, R.
were close to one, indicating the convergence of the chain.
The approximate posterior marginal pdfs of the parameters are illustrated in
Figure 8. Table 5 reports the posterior summaries (posterior means, standard
deviation (SD) and the 95% highest posterior density (HPD) intervals) for all
parameters of the KBN distribution. We note that the values for posterior means
(Table 5) are in good agreement with the MLEs.
0 10 20 30 40

0 10 20 30 40
Density

Density

0.40 0.42 0.44 0.46 0.48 0.50 0.52 0.54 0.56 0.58

µ σ
0 10 20 30 40
0 10 20 30 40
Density

Density

4.30 4.32 4.34 4.36 0.24 0.26 0.28 0.30

a b
0 10 20 30 40
Density

9.40 9.42 9.44 9.46 9.48

Figure 8: Approximate posterior marginal pdfs for the parameters of the KBN
model for the INPC data.

7. Concluding remarks

We have introduced the Kummer beta normal (KBN) distribution with three
shape parameters. The new distribution has proved to be versatile and ana-
lytically tractable. The KBN pdf can be expressed as a linear combination of
exponentiated normal pdfs which allows us to derive some of its mathematical
properties like its ordinary and incomplete moments, mean deviations and order
 R. R. Pescim and S. Nadarajah 165

Table 5: Posterior summaries for the parameters of the KBN model for the INPC
data.
Parameter Mean SD HPD (95%) b
R
a 4.3298 0.0098 (4.3107; 4.3493) 0.9999
b 0.2699 0.0102 (0.2505; 0.2898) 1.0007
c 9.4499 0.0102 (9.4301; 9.4696) 0.9998
µ 0.4401 0.0099 (0.4199; 0.4589) 1.0004
σ 0.5515 0.0101 (0.5307; 0.5699) 1.0010

statistics. The estimation of parameters has been approached by the method

of maximum likelihood and Bayesian analysis. The usefulness of the KBN dis-
tribution has been illustrated by an application to a real data set. The new
distribution provides a rather flexible mechanism for fitting a real world data
and it may attract wider applications in many areas of research.

Acknowledgments

The authors would like to thank the Editor and the referee for careful reading
and for comments which greatly improved the paper. This research was granted
by the Brazilian institution CAPES.

References

Abramowitz, M. and Stegun, I.A. (1968). Handbook of Mathematical Functions.

Dover Publications, New York.

Alzaatreh, A., Famoye, F. and Lee, C. (2014). The gamma-normal distribution:

Properties and applications. Computational Statistics and Data Analysis,
69, 67-80.

Azzalini, A. (1985). A class of distributions which includes the normal ones.

Scandinavian Journal of Statistics, 31, 171-178.

Chen, G. and Balakrishnan, N. (1995). A general purpose approximate goodness-

of-fit test. Journal of Quality Technology, 27, 154-161.

Cordeiro, G.M., Castellares, F., Montenegro, L.C. and de Castro, M. (2013).

The beta generalized gamma distribution. Statistics, 47, 888-900.

Cordeiro, G.M., Cintra, R.J., Rêgo, L.C. and Ortega, E.M.M. (2012a). The Mc-
Donald normal distribution. Pakistan Journal of Statistics and Operation
166 The Kummer Beta Normal: A New Useful-Skew Model

Research, 8, 301-329.

Cordeiro, G.M. and de Castro, M. (2011). A new family of generalized distri-

butions. Journal of Statistical Computation and Simulation, 81, 883-898.

Cordeiro, G.M., Nadarajah, S. and Ortega, E.M.M. (2012b). The Kumaraswamy

Gumbel distribution. Statistical Methods and Applications, 21, 139-168.

Cordeiro, G.M., Ortega, E.M.M. and Nadarajah, S. (2010). The Kumaraswamy

Weibull distribution with application to failure data. Journal of the Franklin
Institute, 347, 1399-1429.

Cordeiro, G.M., Ortega, E.M.M. and Silva, G.O. (2011). The exponentiated
generalized gamma distribution with application to lifetime data. Journal
of Statistical Computation and Simulation, 81, 827-842.

Cordeiro, G.M., Pescim, R.R., Demétrio, C.G.B. and Ortega, E.M.M. (2014).
The Kummer beta generalized gamma distribution. Journal of Data Sci-
ence, Accepted for publication.

Cordeiro, G.M., Pescim, R.R. and Ortega, E.M.M. (2012c). The Kumaraswamy
generalized half-normal distribution for skewed positive data. Journal of
Data Science, 10, 195-224.

Cowles, M.K. and Carlin, B.P. (1996). Markov chain Monte Carlo convergence
diagnostics: A comparative review. Journal of the American Statistical
Association, 91, 133-169.

De Morais, A.L. (2009). A Class of Generalized Beta Distributions, Pareto

Power Series and Weibull Power Series. Dissertation - CCEN, Federal Uni-
versity of Pernambuco, Recife, Brazil.

Dunn, P.K. and Smyth, G.K. (1996). Randomized quantile residuals. Journal
of Computational and Graphical Statistics, 5, 236-244.

Eugene, N., Lee, C. and Famoye, F. (2002). Beta-normal distribution and its
applications. Communications in Statistics - Theory and Methods, 31, 497-
512.

Exton, H. (1978). Handbook of Hypergeometric Integrals: Theory, Applications,

Tables, Computer Programs. Halsted Press, New York.

Gelman, A. and Rubin, D.B. (1992). Inference from iterative simulation using
multiple sequences (with discussion). Statistical Science, 7, 457-472.
 R. R. Pescim and S. Nadarajah 167

Gradshteyn, I.S. and Ryzhik, I.M. (2007). Table of Integrals, Series, and Prod-
ucts. Academic Press, New York.

Gupta, R.C., Gupta, P.L. and Gupta, R.D. (1998). Modeling failure time data
by Lehman alternatives. Communications Statistics - Theory and Methods,
27, 887-904.

Gupta, R.D. and Kundu, D. (2001). Exponentiated exponential family: An

alternative to gamma and Weibull distributions. Biometrical Journal, 43,
117-130.

Mudholkar, G.S. and Srivastava, D.K. (1993). Exponentiated Weibull family

for analyzing bathtub failure-real data. IEEE Transactions on Reliability,
42, 299-302.

Mudholkar, G.S., Srivastava, D.K. and Friemer, M. (1995). The exponential

Weibull family: A reanalysis of the bus-motor failure data. Technometrics,
37, 436-445.

Nadarajah, S. (2005). A generalized normal distribution. Journal of Applied

Statistics, 32, 685-694.

Nadarajah, S. (2008). Explicit expressions for moments of order statistics.

Statistics and Probability Letters, 78, 196-205.

Nadarajah, S. and Gupta, A.K. (2004). The beta Fréchet distribution. Far East
Journal of Theoretical Statistics, 15, 15-24.

Nadarajah, S. and Gupta, A.K. (2007). A generalized gamma distribution with

application to drought data. Mathematics and Computer in Simulation, 74,
1-7.

Nadarajah, S. and Kotz, S. (2004). The beta Gumbel distribution. Mathemati-

cal Problems in Engineering, 10, 323-332.

Nadarajah, S. and Kotz, S. (2006). The beta exponential distribution. Reliabil-

ity Engineering and System Safety, 91, 689-697.

Nadarajah, S., Teimouri, M. and Shih, S.H. (2014). Modified beta distributions.
The Indian Journal of Statistics, B, 76, 19-48.

Paranaı́ba, P.F., Ortega, E.M.M., Cordeiro, G.M. and Pascoa, M.A.R. (2013).
The Kumaraswamy Burr XII distribution: Theory and practice. Journal
of Statistical Computation and Simulation, 83, 2117-2143.
 168 The Kummer Beta Normal: A New Useful-Skew Model

Paranaı́ba, P.F., Ortega, E.M.M., Cordeiro, G.M. and Pescim, R.R. (2011). The
beta Burr XII distribution with application to lifetime data. Computational
Statistics and Data Analysis, 55, 1118-1136.

Pascoa, M.A.R., Ortega, E.M.M. and Cordeiro, G.M. (2011). The Kumaraswamy
generalized gamma distribution with application in survival analysis. Sta-
tistical Methodology, 8, 411-433.

Pescim, R.R., Cordeiro, G.M., Demétrio, C.G.B., Ortega, E.M.M. and Nadara-
jah, S. (2012). The new class of Kummer beta generalized distributions.
Statistics and Operations Research Transactions, 36, 153-180.

Pescim, R.R., Cordeiro, G.M., Nadarajah, S., Demétrio, C.G.B. and Ortega,
E.M.M. (2014). The Kummer Birnbaum-Saunders: An alternative fatigue
life distribution. Hacettepe Journal of Mathematics and Statistics, 43, 473-
510.

Pescim, R.R., Demétrio, C.G.B., Cordeiro, G.M., Ortega, E.M.M. and Urbano,
M.R. (2010). The beta generalized half-normal distribution. Computational
Statistics and Data Analysis, 54, 945-957.

Pescim, R.R. and Nadarajah, S. (2015). An extension of the gamma distribu-

tion. Communications in Statistics - Theory and Methods, Accepted for
publication.

Received June 15, 2014; Accepted December 29, 2014.

Rodrigo R. Pescim
Departamento de Ciências Exatas
Universidade de São Paulo - ESALQ - USP
Av. Pádua Dias 11 - Caixa Postal 9, 13418-900, Piracicaba - São Paulo - Brazil
[email protected]

Saralees Nadarajah
School of Mathematics
University of Manchester
Manchester M13 9PL, UK
[email protected]

View publication stats