Chilean Journal of Statistics

Vol. 4, No. 2, September 2013, 3–26

DISTRIBUTION THEORY
RESEARCH PAPER

The beta Weibull Poisson distribution

Ana Percontini1∗ , Betsabé Blas2 , and Gauss M. Cordeiro3
1
Departamento de Ciências Exatas, Universidade Estadual de Feira de Santana, Feira de
Santana, Brazil,
2
Departamento de Estatística, Universidade Federal de Pernambuco, Recife, Brazil,
3
Departamento de Estatística, Universidade Federal de Pernambuco, Recife, Brazil
(Received: 20 November 2012 · Accepted in final form: 22 March 2013)

Abstract
Providing a wider distribution is always precious for statisticians. A new five-parameter
distribution called the beta Weibull Poisson is proposed, which is obtained by com-
pounding the Weibull Poisson and beta distributions. It generalizes several known life-
time models. We obtain some properties of the proposed distribution such as the survival
and hazard rate functions, quantile function, ordinary and incomplete moments, order
statistics and Rényi entropy. Estimation by maximum likelihood and inference for large
samples are addressed. The potentiality of the new model is shown by means of a real
data set. In fact, the proposed model can produce better fits than some well-known
distributions.

Keywords: Beta distribution · Generating function · Lifetime data · Mean Deviation ·

Moment · Quantile function.
Mathematics Subject Classification: Primary 60E05 · Secondary 62P99.

1. Introduction

The Weibull distribution is a very popular model in reliability and it has been widely used
for analyzing lifetime data. Several new models have been proposed that are either derived
from or in some way are related to the Weibull distribution. When modelling monotone
hazard rates, the Weibull distribution may be an initial choice because of its negatively
and positively skewed density shapes. However, it does not provide a reasonable parametric
fit for modelling phenomenon with non-monotone failure rates such as the bathtub shaped
and the unimodal failure rates that are common in reliability and biological studies. An
example of the bathtub-shaped failure rate is the human mortality experience with a high
infant mortality rate which reduces rapidly to reach a low level. It then remains at that
level for quite a few years before picking up again. Unimodal failure rates can be observed

∗ Corresponding author. Email: anappaixao@gmail.com

ISSN: 0718-7912 (print)/ISSN: 0718-7920 (online)

⃝c Chilean Statistical Society – Sociedad Chilena de Estadística
http://www.soche.cl/chjs
4 Percontini et al.

in course of a disease whose mortality reaches a peak after some finite period and then
declines gradually.
The statistics literature is filled with hundreds of continuous univariate distributions.
Recent developments focus on new techniques for building meaningful distributions. Several
methods of introducing one or more parameters to generate new distributions have been
studied in the statistical literature recently. Among these methods, the compounding of
some discrete and important lifetime distributions has been in the vanguard of lifetime
modeling. So, several families of distributions were proposed by compounding some useful
lifetime and truncated discrete distributions.
In recent years, there has been a great interest among statisticians and applied researchers
in constructing flexible distributions to facilitate better modeling of lifetime data. Several
authors introduced more flexible distributions to model monotone or unimodal failure rates
but they are not useful for modelling bathtub-shaped failure rates. Adamidis and Loukas
(1998) proposed the exponential geometric (EG) distribution to model lifetime data with
decreasing failure rate function and Gupta and Kundu (1999, 2001a,b) defined the ge-
neralized exponential (GE) (also called the exponentiated exponential) distribution. The
last distribution has only increasing or decreasing failure rate function. Following the key
idea of Adamidis and Loukas (1998), Kus (2007) introduced the exponential Poisson (EP)
distribution which has a monotone failure rate. Lee et al. (2007) proposed a generalization
of the Weibull distribution called the beta Weibull (BW) distribution. Barreto-Souza et al.
(2010) studied a Weibull geometric (WG) distribution which extends the EG and Weibull
distributions. In this paper, we propose a new compounding distribution, called the beta
Weibull Poisson (BWP) distribution, by compounding the beta and Weibull Poisson (WP)
distributions (Lu and Shi, 2012). The failure rate function of the WP distribution has
various shapes. In fact, it can be increasing, decreasing, upside-down bathtub-shaped or
unimodal.
The proposed generalization stems from a general class of distributions which is defined
by the following cumulative distribution function (cdf)

∫ G(x)
1
F (x) = IG(x) (a, b) = wa−1 (1 − w)b−1 dw, (1)
B(a, b) 0

where a > 0 and b > 0 are two additional shape parameters to the parameters of the
G-distribution, B(a, b) = Γ(a)Γ(b)/Γ(a + b) is the beta function and IG(x) (a, b) denotes the
incomplete beta function ratio evaluated at G(x). The parameters a and b govern both the
skewness and kurtosis of the generated distribution.
This class was proposed by Eugene et al. (2002) and has been widely used ever since. For
example, Eugene et al. (2002) introduced the beta normal (BN) distribution, Nadarajah and
Kotz (2004) defined the beta Gumbel (BGu) distribution and Nadarajah and Gupta (2004)
proposed the beta Fréchet (BF) distribution. Another example is the beta exponential (BE)
model studied by Nadarajah and Kotz (2006).
The probability density function (pdf) corresponding to (1) is given by

g(x)
f (x) = G(x)a−1 {1 − G(x)}b−1 , (2)
B(a, b)

where g(x) = dG(x)/dx is the baseline density function.

The paper is organized as follows. In Section 2, we define the BWP distribution and
highlight some special cases. In Section 3, we demonstrate that the new density function is
a linear combination of WP density functions. The proof is given in Appendix A. Also, we
 Chilean Journal of Statistics 5

derive the survival and hazard rate functions, moments and moment generating function
(mgf), order statistics and their moments and Rényi entropy. Maximum likelihood estima-
tion of the model parameters and the observed information matrix are discussed in Section
4. In Section 5, we provide an application of the BWP model to the maintenance data with
46 observations reported on active repair times (hours) for an airborne communication
transceiver. Concluding remarks are given in Section 6. Unless otherwise indicated, all re-
sults presented in the paper are new and original. It is expected that they could encourage
further research of the new model.

2. The BWP distribution

We assume that Z has a truncated Poisson distribution with parameter λ > 0 and proba-
bility mass function given by

p(z; λ) = e−λ λz Γ−1 (z + 1)(1 − e−λ )−1 , z = 1, 2, . . . ,

∫p
where Γ(p) = 0 xp−1 e−x dx (for p > 0) is the gamma function.
We define {Wi }Zi=1 to be independent and identically distributed random variable having
the Weibull density function defined by

π(w; α, β) = αβwα−1 exp(−βwα ), w > 0,

where α > 0 is the shape parameter and β > 0 is the scale parameter.
We define X = min{W1 , . . . , WZ }, where the random variables Z and W’s are assumed
independent. The WP distribution of X has density function given by

g(x; α, β, λ) = ċ u xα−1 eλu , x > 0, (3)

αβλe−λ
and u = e−βx .
α
where ċ = ċ(α, β, λ) =
1 − e−λ
The WP model is well-motivated for industrial applications and biological studies. As
a first example, consider the time to relapse of cancer under the first-activation scheme.
Suppose that the number, say Z, of carcinogenic cells for an individual left active after
the initial treatment follows a truncated Poisson distribution and let Wi be the time spent
for the ith carcinogenic cell to produce a detectable cancer mass, for i ≥ 1. If {Wi }i≥1
is a sequence of independent and identically distributed (iid) Weibull random variables
independent of Z, then the time to relapse of cancer of a susceptible individual can be
modeled by the WP distribution. Another example considers that the failure of a device
occurs due to the presence of an unknown number, say Z, of initial defects of the same
kind, which can be identifiable only after causing failure and are repaired perfectly. Define
by Wi the time to the failure of the device due to the ith defect, for i ≥ 1. If we assume
that the Wi ’s are iid Weibull random variables independent of Z, which is a truncated
Poisson random variable, then the time to the first failure is appropriately modeled by the
WP distribution. For reliability studies, the proposed models for X = min {Wi }Z i=1 and
T = max {Wi }Z i=1 can be used in serial and parallel systems with identical components,
which appear in many industrial applications and biological organisms. The first activation
scheme may be questioned by certain diseases. Consider that the number Z of latent factors
6 Percontini et al.

4
(2, 2) (2, 2)

12
(3, 3) (3, 3)
(4, 4) (4, 4)
(5, 5) (5, 5)
(6, 6) (6, 6)

10
3

8
Density

Density
2

6
4
1

2
0

0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

x x

(a) (b)
3.5

(2, 2) (2, 2)
(3, 3) (3, 3)
(4, 4) (4, 4)
3.0

2.5
(5, 5) (5, 5)
(6, 6) (6, 6)
2.5

2.0
2.0
Density

Density

1.5
1.5

1.0
1.0
0.5

0.5
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

x x

(c) (d)

Figure 1. Plots of the BWP density function for: (a) α = 1, β = 2 and λ = 1, (b) α = 0.5, β = 2 and λ = 1, (c)
α = 1.5, β = 2 and λ = 1, (d) α = 0.5, β = 0.5 and λ = 2.

that must all be activated by failure follows a truncated Poisson distribution and assume
that W represents the time of resistance to a disease manifestation due to the ith latent
factor has the Weibull distribution. In the last-activation scheme, the failure occurs after
all Z factors have been activated. So, the WP distribution is able for modeling the time to
the failure under last-activation scheme.
The cdf corresponding to (3) is

eλu − eλ
G(x) = , x > 0. (4)
1 − eλ

The BWP density function is obtained by inserting (3) and (4) in equation (2). It is
given by

f (x) = c u xα−1 eλu (eλ − eλu )a−1 (eλu − 1)b−1 , (5)

 Chilean Journal of Statistics 7

where

αβλ e−λ (eλ − 1)2−a−b

c= .
B(a, b)(1 − e−λ )

Hereafter, a random variable X having density function (5) is denoted by X ∼

BWP(α, β, λ, a, b).
The cumulative distribution of X is given by

F (x) = IG(x) (a, b) = I(eλu −eλ )/(1−eλ ) (a, b). (6)

We are motivated to study the BWP distributions because of the wide usage of the
Weibull and the fact that the current generalization provides means of its continuous exten-
sion to still more complex situations. A second positive point of the current generalization
is that the WP distribution is a basic exemplar of the proposed family. A third positive
point is the the role played by the two beta generator parameters to the WP model. They
can add more flexibility in the density function (5) by imposing more dispersion in the
skewness and kurtosis of X and to control the tail weights.

Figure 2. Relationships of the BWP sub-models.

The beta exponential Poisson (BEP) distribution is obtained from (5) when α = 1. For
b = 1, the exponentiated Weibull Poisson (EWP) distribution comes as a special model. In
addition, for α = 1, we obtain the exponentiated exponential Poisson (EEP) distribution.
8 Percontini et al.

On the other hand, if α = 2, the beta Rayleigh Poisson (BRP) distribution is obtained. In
addition, for b = 1, it follows the exponentiated Rayleigh Poisson (ERP) distribution. The
beta Weibull (BW) distribution comes as the limiting distribution of the BWP distribution
when λ → 0+ . For a = b = 1, equation (5) becomes the WP density function. In addition, if
α = 1, we obtain the exponential Poisson (EP) distribution. The following distributions are
new sub-models: the beta Rayleigh Poisson (BRP), exponentiated Weibull Poisson (EWP),
beta exponential Poisson (BEP), exponentiated Rayleigh Poisson (ERP), beta Rayleigh
(BR), Rayleigh Poisson (RP) and arc sine Weibull Poisson (ASWP) distributions (for more
details, see Appendix B). Other sub-models are the beta exponential (BE), beta Weibull
(BW), beta Rayleigh (BR), exponentiated Rayleigh (ER), exponentiated exponential (EE),
exponentiated Weibull (EW), Rayleigh (R), Weibull (W) and exponential (E) distributions.
Several special distributions of the BWP model are displayed in Figure 2.

3. Properties of the new distribution

3.1 Density function

We can derive a useful expansion for the BWP density function (see the proof in Appendix
A) given by

∞ ∑
∑ r
f (x) = vr,j g(x; α, β, λr,j ), (7)
r=0 j=0

where λr,j = λ(r − j + 1) > 0 and

( )
(−1)j (r + 1) vr ejλ (1 − e−λr,j ) r
vr,j = .
(r − j + 1) e−λr,j (1 − eλ )r (eλ − 1) j
∑∞ ∑r
Clearly, r=0 j=0 vr,j = 1. Equation (7) reveals that the BWP density function is a
linear combination of WP density functions. So, we can obtain some mathematical pro-
perties of the BWP distribution directly from those WP properties.

3.2 Cumulative function and quantiles

By integrating (7), the cdf F (x) becomes

∞ ∑
∑ r
F (x) = vr,j G(x; α, β, λr,j ). (8)
r=0 j=0

Quantile functions are in widespread use in general statistics and often find repre-
sentations in terms of lookup tables for key percentiles. For some baseline distributions
with closed-form cdf, it is possible to obtain the quantile function in closed-form. How-
ever, for some other distributions, the solution is not possible. The quantile function, say
x = Q(z; α, β, λ, a, b) = F −1 (z; α, β, λ, a, b), of the BWP distribution follows by inverting
(6) as
 Chilean Journal of Statistics 9

{ ( )− 1 } α1
1
log log[w + e (1 − w)] λ
λ β
x = Q(z; α, β, λ, a, b) = , (9)

where w = Qa,b (z) denotes the beta quantile function with parameters a and b.
Power series methods are at the heart of many aspects of applied mathematics and
statistics. We can obtain the moments of the beta G distribution using a power series
expansion for the quantile function x = QG (u) = G−1 (u) of the baseline cdf G(x) with
easily computed non-linear recurrence equation for its coefficients.
When the function Q(u) does not have a closed form expression, this function can usually
be written in terms of a power series expansion of a transformed variable v, which is usually
of the form v = p(qu − t)ρ for p, q, t and ρ known constants.
We can obtain a power series for Qa,b (z) in the Wolfram website given by

(b − 1) 2 (b − 1)(a2 + 3ba − a + 5b − 4) 3
Qa,b (z) = v + v + v
(a + 1) 2(a + 1)2 (a + 2)
v 4 (b − 1)
+ [a4 + (6b − 1)a3 + (b + 2)(8b − 5)a2 +
3(a + 1)3 (a + 2)(a + 3)
(33b2 − 30b + 4)a + b(31b − 47) + 18] + O(v 5 ), (10)

where v = [azB(a, b)]1/a for a > 0.

The simulation of the BWP distribution is easy. If T is a random variable having a beta
distribution with parameters a and b, then the random variable

{ ( )− 1 } α1
1
log log[W + e (1 − W )]
λ β
X= λ

follows the BWP distribution.

3.3 Survival and hazard rate functions

The BWP survival function is given by

S(x; θ) = 1 − F (x; θ) = 1 − I(eλu −eλ )/(1−eλ ) (a, b),

where θ = (α, β, λ, a, b) is the vector of the model parameters. The failure rate function
corresponding to (5) reduces to

f (x; θ) c u xα−1 eλu (eλ − eλu )a−1 (eλu − 1)b−1

h(x; θ) = = .
S(x; θ) {1 − I(eλu −eλ )/(1−eλ ) (a, b)}
10 Percontini et al.

12
(2, 2) (2, 2)
(3, 3) (3, 3)

14
(4, 4) (4, 4)
(5, 5) (5, 5)
10 (6, 6) (6, 6)

12
8

10
h(x)

h(x)
6

8
4

6
2

4
0

2
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

x x

(a) (b)

1.5
(2, 2) (2, 2)
(3, 3) (3, 3)
(4, 4) (4, 4)
(5, 5) (5, 5)
15

(6, 6) (6, 6)
1.0
10
h(x)

h(x)

0.5
5

0.0
0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

x x

(c) (d)

Figure 3. Plots of the BWP hazard rate function for (a) α = 1, β = 2 and λ = 1; (b) α = 0.5, β = 2 and λ = 1; (c)
α = 1.5, β = 2 and λ = 1; (d) α = 0.5, β = 0.5 and λ = 1.

3.4 Moments
We hardly need to emphasize the necessity and importance of moments in any statistical
analysis especially in applied work. Some of the most important features and characteristics
of a distribution can be studied through moments (e.g., tendency, dispersion, skewness and
kurtosis).
An expression for the mgf of X can be obtained from (7) using the WP generating
function. Setting y = λr,j e−βx in the definition of the mgf, we can express it as
α

∞ ∑
∑ r
MX (t) = vr,j (eλr,j − 1)−1
r=0 j=0
∫ λr,j
× exp{t(−β −1 [log(y) − log(λr,j )])1/α + y}dy.
0
 Chilean Journal of Statistics 11

Using the power series of the exponential function, after some simplification, we obtain

∞
∑ ∑
r
MX (t) = q(r, m, n, j) J(λr,j , m, n) tn , (11)
r,m,n=0 j=0

where

∫ λr,j ( )n
J(λr,j , m, n) = y m −β −1 [log(y) − log(λr,j )] α dy
0

and

vr,j
q(r, m, n, j) = .
(eλr,j − 1) m! n!

The last integral can be computed using the software Mathematica 8.0. Then,

∞
∑ ∑
r ( )
α+n
MX (t) = ϖ(r, m, n, j) Γ tn , (12)
α
r,m,n=0 j=0

where

ϖ(r, m, n, j) = β − α λm+1 (1 + m)−

n α+n

r,j
α q(r, m, n, j).

Equation (12) can be reduced to

∞
∑
MX (t) = δ n tn , (13)
n=0
) (
∑∞ ∑r α+n
where δn = m,r=0 j=0 ϖ(r, m, n, j) Γ , n = 0, 1, . . .
α
Hence, the nth ordinary moment of X, say µ′n = E(X n ), is simply given by µ′n = n! δn .
Further, the central moments (µn ) and cumulants (κn ) of X can be determined as

∑
n ( ) ∑(
n−1 )
n n−1
µn = (−1)s
µ′s ′
1 µn−s and κn = µ′n − κs µ′n−s ,
s s−1
s=0 s=1

respectively, where κ1 = µ′1 . Then, κ2 = µ′2 − µ′2 ′ ′ ′ ′3 ′

1 , κ3 = µ3 − 3µ2 µ1 + 2µ1 , κ4 = µ4 −
4µ′3 µ′1 − 3µ′2 ′ ′2 ′4 3/2
2 + 12µ2 µ1 − 6µ1 , etc. The skewness γ1 = κ3 /κ2 and kurtosis γ2 = κ4 /κ22
follow from the second, third and fourth cumulants.
The nth descending factorial moment of X is

∑
n
µ′(n) = E(X (n) ) = E [X(X − 1) × · · · × (X − n + 1)] = s(n, r) µ′r ,
r=0
12 Percontini et al.

where
[ ]
1 dr (n)
s(n, r) = x
r! dxr x=0

is the Stirling number of the first kind which counts the number of ways to permute a list of
n items into r cycles. So, we can obtain the factorial moments from the ordinary moments
given before.
The incomplete moments of X can be expressed in terms of the incomplete moments of
the WP distribution from equation (7). We obtain

∞ ∑
∑ r ∫ y
mn (y) = E(X |X < y) =
n
vr,j xn g(x; α, β, λr,j )dx
r=0 j=0 0

∞ ∑
∑ r ∫ y
= vr,j xn c u xα−1 eλu dx. (14)
r=0 j=0 0

Setting z = βxα and integrating by parts, we can write

∞ r
{ ∞
[ ∞
] }
e−λ y n ∑ ∑ ∑ λs ∑ (−1)m sm (βy α )m λe−βy
α
mn (y) = vr,j n −e .
1 − eλ s! n + mα
r=0 j=0 s=0 m=0

The sum in m converges to (n + mα)−1 e−sβy . Then, the nth incomplete moment of X
α

becomes

∞ ∑
{ ∞ [ ] }
∑ ∑
r α
λs e−sβy y n e−λ(1−e
α −βy
)
mn (y) = pr,j y n − , (15)
s!(n + mα) 1 − eλ
r=0 j=0 s=0

n vr,j e−λ
where pr,j = .
1 − eλ
We can derive the mean deviations of X about the mean µ′1 and about the median M
in terms of its first incomplete moment. They can be expressed as
[ ]
δ1 = 2 µ′1 F (µ′1 ) − m1 (µ′1 ) and δ2 = µ′1 − 2m1 (M ), (16)
∫q
where µ′1 = E(X) and m1 (q) = −∞ x f (x) dx. The quantity m1 (q) is obtained from
(15) with n = 1 and the measures δ1 and δ2 in (16) are immediately determined from
′
these formulae with n = 1 by setting q = µ1 and q = M , respectively. For a positive
random variable X, the Bonferroni and Lorenz curves are defined as B(π) = T1 (q)/[πµ′1 ]
and L(π) = T1 (q)/µ′1 , respectively, where q = F −1 (π) = Q(π) comes from the quantile
function (9) for a given probability π.
The formulae derived along the paper can be easily handled in most symbolic compu-
tation software platforms such as Maple, Mathematica and Matlab. These platforms have
currently the ability to deal with analytic expressions of formidable size and complexity.
Established explicit expressions to calculate statistical measures can be more efficient than
computing them directly by numerical integration. The infinity limit in the sums of these
 Chilean Journal of Statistics 13

expressions can be substituted by a large positive integer such as 20 or 30 for most practical
purposes.

3.5 Order statistics

Order statistics make their appearance in many areas of statistical theory and practice.
Moments of order statistics play an important role in quality control and reliability, where
some predictors are often based on moments of the order statistics. We derive an explicit
expression for the density function of the ith order statistic Xi:n , say fi:n (x) (see Appendix
C). For a beta-G model defined from the parent functions g(x) and G(x), fi:n (x) can be
expressed as an infinite linear combination of WP density functions

∞ ∑
∑ l
fi:n (x) = γi:n (l, s) g(x; α, β, λl,s ), (17)
l=0 s=0

where λl,s = λ(l − s + 1) and

∞
( l )(n−i)(r+a−1)(k+b−1)
∑
n−i ∑
(−1)l+s+j+k esλ s j k l (1 − e−λl,s ) ci+1−j,r
γi:n (l, s) = .
(l − s + 1)(1 − e−λ ) (1 − eλ )l B(a, b)i+j B(i, n − i + 1)
j=0 r,k=0

An expression for the mgf of Xi:n can be obtained from (17) using the WP generating
function. Setting y = λl,s e−βx in the definition of the generating function, we obtain
α

∞
∑ ∑
l ( )
α+n
MXi:n (t) = ϖi (l, m, n, s) Γ tn , (18)
α
l,m,n=0 s=0

where

β − α λm+1 (1 + m)−
n α+n

l,s
α γi:n (l, s)
ϖi (l, m, n, s) = .
m! n! (eλl,s − 1)
∑∞
Equation (18) can be reduced to MXi:n (t) = n=0 δi:n tn , where

∞ ∑
∑ l ( )
α+n
δi:n = ϖi (l, m, n, s) Γ , n = 0, 1, . . .
α
m,l=0 s=0

s ) = s! δ .
Hence, the sth ordinary moment of Xi:n becomes E(Xi:n i:n

3.6 Rényi entropy

The entropy of a random variable X with density function f (x) is a measure of the uncer-
tainty variation. The Rényi entropy is defined as
14 Percontini et al.

{∫ }
−1
IR (ρ) = (1 − ρ) log f (x) dx , ρ

where ρ > 0 and ρ ̸= 1. If a random variable X has the BWP distribution, we have

[ ]ρ
g(x; θ)
f (x) = ρ
G(x)(a−1)ρ [1 − G(x)](b−1)ρ . (19)
B(a, b)

By expanding the binomial term, the following expansion holds for any real a,

∞
∑
G(x) (a−1)ρ+j
= sr [(a − 1)ρ + j] G(x)r ,
r=0

∑∞ ( ) (i)
where sr [(a − 1)ρ + j] = i=r (−1)
r+i (a−1)ρ+j
j r . Equation (19) can be rewritten as

[ ]ρ ∑
∞ ∑
∞
g(x; θ)
ρ
f (x) = qj,r G(x)r ,
B(a, b)
j=0 r=0

( )
where qj,r = (−1)j (b−1)ρ
j sr [(a − 1)ρ + j].
From equations (3) and (4), we obtain

[ ]ρ ∑
∞ ∑
∞ ( )r
ρ c u xα−1 eλu eλu − eλ
f (x) = qj,r .
B(a, b) 1 − eλ
j=0 r=0

Then,

∞ ∑
∑ ∞ ∑
r
ρ
ρ
f (x) = pj,r,t uρ xα(ρ− α ) eλ(ρ+r−t)u , (20)
j=0 r=0 t=0

where u = e−βx and

α

()
qj,r (−1)t rt eλt c ρ
pj,r,t = .
[B(a, b)]ρ (1 − eλ )r

∑∞ [λ(ρ + r − t)]s −sβxα

Using the power series expansion eλ(ρ+r−t)u = s=0 e in (20) and
s!
setting y = βsxα , the Rényi entropy reduces to
 
∑∞ ( )
1−ρ
IR (ρ) = (1 − ρ)−1 log ϕj (ρ) Γ ρ + , (21)
 α 
j=0
 Chilean Journal of Statistics 15

where

∞ ∑
∑ r
pj,r,t λs (ρ + r − t)s
ϕj (ρ) = 1−ρ .
+ρ
r,s=0 t=0 α s! (βs) α

4. Maximum likelihood estimation

Let x1 , . . . , xn be a random sample of size n from the BWP(a, b, α, β, λ) distribution. The

log-likelihood function for the vector of parameters θ = (a, b, α, β, λ)T can be expressed as

l(θ) = n [log(αβλ) − λ − log[B(a, b)] − log(1 − e−λ ) − log(1 − eλ )a+b−2 ]

∑
n ∑
n ∑
n
+(α − 1) log(xi ) − β xαi + λ ui
i=1 i=1 i=1
∑
n ∑
n
+(a − 1) log(e λui
− e ) + (b − 1)
λ
log(1 − eλui ),
i=1 i=1

where ui = exp(−βxαi ) is a transformed observation. The components of the score vector

U (θ) are given by

n ∑ ∑ ∑
n n n
Uα (θ) = + log(xi ) − β xαi log(xi ) − λβ ui xαi log(xi )
α
i=1 i=1 i=1
∑
n ( )
1−a b−1
+λβ ui xαi eλui log(xi ) + ,
e −e
λui λ 1 − eλui
i=1

n ∑ α ∑ ∑
n n n
Uβ (θ) = − xi − λ ui xαi + λ ui xαi eλui
β
i=1 i=1 i=1
( )
1−a b−1
× λui + ,
e −e λ 1 − eλui
16 Percontini et al.

n ne−λ n(a + b − 2)eλ log(1 − eλ )a+b−2

Uλ (θ) = −n+ −
λ 1 − e−λ (1 − eλ ) log(1 − eλ )
∑
n ∑
n
ui eλui − eλ
+ ui + (a − 1) − (b − 1)
eλui − eλ
i=1 i=1
∑
n
ui eλui
× ,
i=1
1 − eλui

Ua (θ) = −n [ψ(a) − ψ(a + b)] + n log(eλ − 1)a+b−2 log[log(1 − eλ )]

∑
n
+ log(eλui − eλ ),
i=1

Ub (θ) = −n [ψ(b) − ψ(a + b)] + n log(eλ − 1)a+b−2 log[log(1 − eλ )]

∑
n
+ log(1 − eλui ),
i=1

where ψ(·) is the digamma function. The maximum likelihood estimates (MLEs) θb =
a, bb, α
(b b λ)
b, β, b T of θ = (a, b, α, β, λ)T are the simultaneous solutions of the non-linear equa-
tions: Ua (θ) = Ub (θ) = Uα (θ) = Uβ (θ) = Uλ (θ) = 0. They can be solved numerically
using iterative methods such as a Newton-Raphson type algorithm.
For interval estimation and hypothesis tests on the model parameters, we require the
5 × 5 observed information matrix J = J(θ) given in Appendix D. Under conditions
that are fulfilled for parameters in the interior √ b of the parameter −1 space but not on the
boundary, the asymptotic distribution of n(θ − θ) is N5 (0, I(θ) ), where I(θ) is the
expected information matrix. In practice, we can replace I(θ) by the observed information
matrix evaluated at θ, b say J(θ).b We can construct approximate confidence regions for the
parameters based on the multivariate normal N5 (0, J(θ) b −1 ) distribution.
Further, the likelihood ratio (LR) statistic can be used for comparing this distribution
with some of its sub-models. We can compute the maximum values of the unrestricted
and restricted log-likelihoods to construct the LR statistics for testing some sub-models of
the BWP distribution. For example, the test of H0 : a = b = 1 versus H1 : H0 is not true
is equivalent to compare the BWP and WP distributions and the LR statistic becomes
w = 2{l(â, b̂, α̂, β̂, λ̂) − l(1, 1, α̃, β̃, λ̃)}, where â, b̂, α̂, β̂ and λ̂ are the MLEs underH1 and
α̃, β̃ and λ̃ are the estimates under H0 .

5. Application

Here, we present an application regarding the BWP model to the maintenance data with
46 observations reported on active repair times (hours) for an airborne communication
transceiver discussed by Alven (1964), Chhikara and Folks (1977) and Dimitrakopoulou
et al. (2007). We also fit a five-parameter beta Weibull geometric (BWG) distribution
introduced by Cordeiro et al. (2011) to make a comparasion with the BWP model. The
 Chilean Journal of Statistics 17

BWG density function is given by

α(1 − p)b β α xα−1 e−b(βx) (1 − e−(βx) )a−1 (1 − pe−(βx) )−(a+b)

α α α

f (x; θ 1 ) = ,
B(a, b)

where θ 1 = (p, α, β, a, b) and x > 0.

The data are: 0.2, 0.3, 0.5, 0.5, 0.5, 0.5, 0.6, 0.6, 0.7, 0.7, 0.7, 0.8, 0.8, 1.0, 1.0, 1.0, 1.0,
1.1, 1.3, 1.5, 1.5, 1.5, 1.5, 2.0, 2.0, 2.2, 2.5, 2.7, 3.0, 3.0, 3.3, 3.3, 4.0, 4.0, 4.5, 4.7, 5.0, 5.4,
5.4, 7.0, 7.5, 8.8, 9.0, 10.3, 22.0 e 24.5.
In Table 1, we list the MLEs of the model parameters and the bias-corrected Akaike
information criterion (BAIC), Bayesian information criterion (BIC) and the Hannan-Quinn
information criterion (HQIC). We observe that the value of the BAIC criterion is smaller
for the BWP distribution as compared with those values of the other models. So, the new
distribution seems to be a very competitive model to these data.
Table 1. MLEs of the parameters and BAIC, BIC and HQIC statistics of the BWP, BWG, WP and Weibull models
for data of active repair times (hours) for an airborne communication transceiver.

Model â b̂ α̂ β̂ λ̂ BAIC BIC HQIC

BWP 21.969 0.320 0.722 1.439 5.342 207.838 216.981 211.263
(58.799) (0.256) (0.390) (1.418) (2.232)

WP 1.101 0.092 3.522 210.927 216.413 212.982

(0.120) (0.052) (1.917)

Weibull 0.899 0.334 212.939 216.597 214.309

(0.096) (0.075)
Model â b̂ α̂ β̂ p̂ BAIC BIC HQIC
BWG 3.269 0.587 1.417 0.212 0.988 208.205 217.348 211.630
(4.599) (0.323) (0.642) (0.076) (0.017)

The LR test statistic for testing H0 : a = b = 1 against H1 : H0 is not true is w = 7.08912

(p-value = 2.88 × 10−2 ), which is statistically significant. Figure 4 displays the histogram
of the data and the plots of the fitted BWP, WP, Weibull and BWG models.

6. Concluding remarks

The Weibull distribution is commonly used to model the lifetime of a system. However,
it does not exhibit a bathtub-shaped failure rate function and thus it can not be used to
model the complete lifetime of a system. We define a new lifetime model, called the beta
Weibull Poisson (BWP) distribution, which extends the Weibull Poisson (WP) distribution
proposed by Lu and Shi (2012), whose failure rate function can be increasing, decreasing
and upside-down bathtub. The BWP distribution is quite flexible to analyse positive data
instead of some other special models. Its density function can be expressed as a mixture
of WP densities. We provide a mathematical treatment of the distribution including ex-
plicit expressions for the density function, generating function, ordinary and incomplete
moments, Rényi entropy, order statistics and their moments. The estimation of the model
parameters is approached by the method of maximum likelihood and the observed informa-
tion matrix is determined. An application to real data reveals that the BWP distribution
could provide a better fit than other well-known lifetime models.
18 Percontini et al.

0.20
BWP
WP
Weibull
0.15 BWG
Density

0.10
0.05
0.00

0 5 10 15 20 25

repair times for an airborne communication transceiver

Figure 4. The density functions of the fitted BWP, WP, Weibull and BWG distributions.

Appendix A. The BWP density function

An expansion for the beta-G cumulative function is given by Cordeiro and Lemonte (2011)
and follows from equation (1) as

∑ ∞
1
F (x) = tr G(x)r , (A1)
B(a, b)
r=0

∑ ( )
where tr = ∞ wm sr (a+m) for any real a, wm = (−1)m (a+m)−1 b−1 m and sr (a+m) =
∑∞ r+j
(a+m)(j )
m=0
j=r (−1) j r . Differentiating equation (A1), we obtain an expansion for the BWP
density function

∞
∑
f (x) = vr hr+1 (x), (A2)
r=0

where vr = tr+1 /B(a, b). Note that hr+1 (x) = (r + 1)G(x)r g(x) is the density function of
the exponentiated
∑ G with power parameter r + 1, say exp-G(r + 1), distribution. We can
verify that ∞r=0 v r = 1. In fact,
 Chilean Journal of Statistics 19

∞
∑ ∑∑ ∞ ∞
1
vr = wm sr (a) = 1
B(a, b)
r=0 r=0 m=0

if and only if

∞ ∑
∑ ∞
wm sr (a) = B(a, b). (A3)
r=0 m=0

But

∫ ∞ (
∑ )
1
b − 1 (−1)j
B(a, b) = t a−1
(1 − t) b−1
dt = ,
0 j a+j
j=0

and, consequently,

∞ ∑
∑ ∞ ∞
∑ ( ) ∞ ∞ ( )( )
(−1)m b − 1 ∑ ∑ r+j a j
wm sr (a) = (−1) = B(a, b)
(a + m) m j r
r=0 m=0 m=0 r=0 j=r

Consider the expressions of g(x) and G(x) from equations (3) and (4), respectively.
Replacing them in (A2), we obtain an expansion for the BWP density function

∞
∑ ( )r
α−1 λu eλu − eλ
f (x) = c x ue vr (r + 1) . (A4)
1 − eλ
r=0

Hence, from this equation, the BWP density function can be expressed as a linear com-
bination of WP density functions.

Appendix B. Special cases of the BWP distribution

Setting b = 1 in equation (5), we obtain the EWP density function

( )a−1
α−1 λu eλu − eλ αβλ
f (x) = c u x e , c= .
1 − eλ B(a, 1)(1 − e−λ )
∑∞
Using equation G(x)α = k
k=0 sk (α) G(x) , we can write

∞
∑ ( )k
eλu − eλ
f (x) = c u x α−1 λu
e sk (a − 1)
1 − eλ
k=0
∞
∑ ( )
sk (a − 1) ∑
k
α−1 λu k λu(k−r) λr
r
=cux e (−1) e e . (B1)
(1 − eλ )k r
k=0 r=0
20 Percontini et al.

After some algebra, we obtain from (B1)

∞ ∑
∑ k
f (x) = vk,r g(x; α, β, λk,r ), (B2)
k=0 r=0

where λk,r = λ(k − r + 1) and

( )
(−1)r kr B(a, b) sk (a − 1) eλr (1 − e−λk,r )
vk,r = .
(k − r + 1) B(a, 1) (1 − eλ )k (1 − e−λ )

Equation (B2) reveals that the density function f (x) is a linear combination of the WP
densities.
From equation (5) with a = b = 1/2, we obtain

( )−1/2 ( )−1/2
c1 xα−1 ueλu eλu − eλ eλu − eλ
f (x; θ) = 1− ,
π 1 − eλ 1 − eλ

αβλe−λ (eλ − 1)
and u = e−βx . Thus,
α
where c1 =
(1 − e−λ )

c1 xα−1 ueλu
f (x; θ) = √( )( )
eλu − eλ 1 − eλu
π
1 − eλ 1 − eλ

If λ approaches to 0, then

c1 xα−1 ueλu αβxα−1 u

lim f (x; θ) = lim √( )( ) = √
λ→0 λ→0
eλu − eλ 1 − eλu π u(1 − u)
π
1 − eλ 1 − eλ

So, the BWP distribution reduces as a limiting case to a two-parameter arcsine Weibull-
Poisson distribution.

Appendix C. Expansion for the Density Function of the order statistics

The density function fi:n (x) of the ith order statistic, say Xi:n , for i = 1, 2, . . . , n, from
data values X1 , . . . , Xn having the beta-G distribution can be obtained from (2) as

( )
g(x) G(x)a−1 {1 − G(x)}b−1 ∑
n−i
j n−i
fi:n (x) = (−1) F (x)i+j−1 . (C1)
B(a, b) B(i, n − i + 1) j
j=0

By application of an equation in Section 0.314 of Gradshteyn and Ryzhik (2000) for a

 Chilean Journal of Statistics 21

power series raised to any j positive integer

( ∞
)j ∞
∑ ∑
ai ui = cj,i ui , (C2)
i=0 i=0

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

∑
i
−1
cj,i = (ia0 ) [m(j + 1) − i] am cj,i−m , (C3)
m=1

with cj,0 = aj0 . The coefficient cj,i comes from cj,0 , . . . , cj,i−1 and then from a0 , . . . , ai .
The coefficients cj,i can be given explicitly in terms of the quantities a′i s, although it is
not necessary for programming numerically our expansions in any algebraic or numerical
software.
For a > 0 real non-integer, we have
( ∞
)i+j−1
1 ∑
F (x)i+j−1 = tr (a, b)G(x)r
B(a, b)
r=0
( )i+j−1 (∑ ∞
)i+j−1
1 r
= tr G(x) .
B(a, b)
r=0

We now use equations (C2)-(C3)

∑ ( ) ∞
n−i ∑
n−i
g(x)G(x)a−1 [(1 − G(x)]b−1
fi:n (x) = (−1)j ci+j−1,r G(x)r
B(a, b)i+j B(i, n − i + 1) j
j=0 r=0

∑ ∞
n−i ∑ ( )
j n − i g(x)[(1 − G(x)]b−1 G(x)r+a−1
= (−1) ci+j−1,r , (C4)
j B(a, b)i+j B(i, n − i + 1)
j=0 r=0

where

∑
r
ci+j−1,r = (rt0 )−1 ((i + j)m − r)tm ci+j−1,r−m . (C5)
m=1

Equation (C4) can be written as

∞
n−i ∑
∑ ( )
n − i g(x)[(1 − G(x)]b−1 [1 − (1 − G(x))]r+a−1
fi:n (x) = (−1)j ci+j−1,r .
j B(a, b)i+j B(i, n − i + 1)
j=0 r=0

For any q > 0 real, we have

∞
∑ ( )
q
G(x) = [1 − {1 − G(x)}] =
q q
(−1) k
[1 − G(x)]k , (C6)
k
k=0
22 Percontini et al.

and then

∑ ∞
n−i ∑ ∞
∑ ( )
r+a−1
fi:n (x) = g(x) (−1) k
[1 − G(x)]k+b−1 .
k
j=0 r=0 k=0

In the same way, using equation (C6), it follows that

∞
(n−i)(r+a−1)(k+b−1)
∑
n−i ∑
(−1)j+k+l j k l ci+j−1,r
fi:n (x) = g(x) G(x)l .
B(a, b)i+j B(i, n − i + 1)
j=0 r,k,l=0

Replacing equations (3) and (4) in the above equation, fi:n (x) can be expressed as an
infinite linear combination of WP density functions

∞
(n−i)(r+a−1)(k+b−1) [ ][ ]l
∑
n−i ∑
(−1)j+k+l j k l ci+j−1,r αβλe−λ eλu − eλ
fi:n (x) = u xα−1 eλu
B(a, b)i+j B(i, n − i + 1) 1 − e−λ 1 − eλ
j=0 r,k,l=0

∞
(n−i)(r+a−1)(k+b−1) [ ]
∑
n−i ∑
(−1)j+k+l j k l ci+j−1,r αβλe−λ
= u xα−1 eλu
B(a, b)i+j B(i, n − i + 1)(1 − eλ )l 1 − e−λ
j=0 r,k,l=0

∑
l ( )
l
× (−1) s
(eλu )l−s esλ . (C7)
s
s=0

Equation (C7) reduces to

∞ ∑
∑ l
fi:n (x) = γi:n (l, s) g(x; α, β, λl,s ), (C8)
l=0 s=0

where λl,s = λ(l − s + 1) and

∞
( l )(n−i)(r+a−1)(k+b−1)
n−i ∑
∑ (−1)l+s+j+k esλ s j k l (1 − e−λl,s ) ci+1−j,r
γi:n (l, s) = .
(l − s + 1)(1 − e−λ )(1 − eλ )l B(a, b)i+j B(i, n − i + 1)
j=0 r,k=0

Appendix D. Information Matrix

Let ui = exp(−βxαi ). The elements of the observed information matrix J(θ) for the pa-
rameters (α, β, λ, a, b) are
 Chilean Journal of Statistics 23

n ∑ n ∑ n ∑ n
Jαα =− 2 −β xαi log2 (xi ) − λβ ui xαi log2 (xi ) + λβ 2 x2α 2
i ui log (xi )
α
i=1 i=1 i=1
n [
∑ ]
λβui xαi eλui log2 (xi )
+(a − 1) ψ(xi )
eλui − eλ
i=1
n [
∑ ]
λβui xα eλui log2 (xi )
+(b − 1) i
φ(xi ),
1 − eλui
i=1

∑
n ∑
n ∑
n
Jαβ = Jβα = − xαi log(xi ) − λ ui xαi log(xi ) + λβ ui x2α
i log(xi )
i=1 i=1 i=1
n [
∑ ]
λui xαi eλui log(xi )
+(a − 1) ψ(xi )
eλui − eλ
i=1
n [
∑ ]
λui xα eλui log(xi )
+(b − 1) i
φ(xi ),
1 − eλui
i=1

where
( )
λβui xα eλui
ψ(xi ) = −1 + βxαi + λβui xαi − λui i λ
e −e

and
( )
λβui xαi eλui
φ(xi ) = 1 − βxi − λβui xi −
α α
.
1 − eλui

Further,
[
∑
n ∑
n
Jαλ = Jλα = −β ui xαi log(xi ) + (a − 1) ρ(xi ) − 1 − λui
i=1 i=1
] ( )
λ(ui eλui − eλ ) ∑
n
λui eλui
+ + (b − 1) ϕ(xi ) 1 + λui + ,
eλui − eλ 1 − eλui
i=1

∑
n ∑
n
Jαa = Jaα = − λρ(xi ), Jαb = Jbα = λϕ(xi ),
i=1 i=1

βui xαi eλui log(xi ) βui xαi eλui log(xi )

where ρ(xi ) = and ϕ(xi ) = ,
eλui − eλ 1 − eλui
24 Percontini et al.

∑ n ∑ n ( )
n λui (xαi )2 eλui
Jββ =− 2 +λ i + (a − 1)
ui x2α
β eλui − eλ
i=1 i=1
( ) n (
∑ )
λui eλui λui x2α eλui
× 1 + λui − + (b − 1) i
eλui − eλ 1 − eλui
i=1
( )
λui eλui
× −1 − λui − ,
1 − eλui

∑
n ∑
n [ ]
λ(ui eλui − eλ )
Jβλ = Jλβ = − ui xαi + (a − 1) γ(xi ) −1 − λui +
eλui − eλ
i=1 i=1
∑
n ( )
λui eλui
+(b − 1) δ(xi ) 1 + λui + ,
1 − eλui
i=1

∑
n ∑
n
Jβa = Jaβ = − λγ(xi ), Jβb = Jbβ = λδ(xi ),
i=1 i=1

ui xαi eλui ui xαi eλui

γ(xi ) = and δ(xi ) = . Furthermore,
eλui − eλ 1 − eλui

n ne−λ ne−2λ n(a + b − 2)eλ log(1 − eλ )a+b−2

Jλλ = − + + −
λ2 1 − e−λ (1 − e−λ )2 1 − eλ log(1 − eλ )
[ ]
(a + b − 2)eλ eλ eλ
× 1− + +
(1 − eλ ) log(1 − eλ ) 1 − eλ (1 − eλ ) log(1 − eλ )
∑n [ ]
ui eλui − eλ (ui eλui − eλ )2
+(a − 1) −
eλui − eλ (eλui − eλ )2
i=1
n [ 2 λui
∑ ]
ui e u2i (eλui )2
−(b − 1) + ,
1 − eλui (1 − eλui )2
i=1

[ ]
neλ log(1 − eλ )a+b−2 { }
Jλa = Jaλ = 1 + (a + b − 2) log[log(1 − e λ
)]
(1 − eλ ) log(1 − eλ )
∑n ( )
ui eλui − eλ
+ ,
eλui − eλ
i=1

[ ]
neλ log(1 − eλ )a+b−2 { }
Jλb = Jbλ = 1 + (a + b − 2) log[log(1 − eλ
)]
(1 − eλ ) log(1 − eλ )
∑n ( )
ui eλui
+ ,
1 − eλui
i=1
 Chilean Journal of Statistics 25

[ ]2
¨ n B˙a (a, b)
nBa (a, b)
Jaa =− + − n log(1 − eλ )a+b−2
B(a, b) B(a, b)
× log2 [log(1 − eλ )],

nB̈(a, b) nB˙a (a, b)Ḃb (a, b)

Jab = Jba = − + − n log(1 − eλ )a+b−2
B(a, b) [B(a, b)]2
× log2 [log(1 − eλ )],

[ ]2
nB̈b (a, b) n Ḃb (a, b)
Jbb = − + − n log(1 − eλ )a+b−2
B(a, b) [B(a, b)]2
× log2 [log(1 − eλ )],

∂ ∂ ∂2
where B˙a (a, b) = B(a, b) , Ḃb (a, b) = B(a, b) and B̈(a, b) = B(a, b).
∂a ∂b ∂b∂a

References

Adamidis, K., Loukas, S., 1998. A lifetime distribution with decreasing failure rate, Statis-
tics & Probability Letters, 39, 35-42.
Alven, W.H., 1964. Reliability Engineering by ARINC. Prentice-hall, New Jersey.
Barreto-Souza, W., de Morais, A.L., Cordeiro, G.M, 2010. Weibull-geometric distribution,
Journal of Statistical Computation and Simulation, 81, 645-657.
Chhikara, R.S., Folks,J.L., 1977. The inverse Gaussian distribution as a lifetime model,
Technometrics, 19, 461-468.
Cordeiro, G.M., Lemonte, A.J., 2011. The β-Birnbaum-Saunders distribution: An improved
distribution for fatigue life modeling, Computational Statistics & Data Analysis, 55,
1445-1461.
Cordeiro, G.M., Silva, G.O., Ortega, E.M.M., 2011. The beta-Weibull geometric distribu-
tion, A Journal of Theoretical and Applied Statistics, 0, 1-18
Dimitrakopoulou, T., Adamidis, K., Loukas, S., 2007. A lifetime distribution with an
upside-down bathtub-shaped hazard function, IEEE Transactions on Reliability, 56, 308-
311.
Eugene, N., Lee, C., Famoye, F., 2002. Beta-normal distribution and its applications, Com-
munication in Statistics - Theory and Methods, 31, 497-512.
Gradshteyn, I.S., Ryzhik, I.M., 2000. Table of Integrals, Series and Products. Academic
Press, San Diego.
Gupta, R.D., Kundu, D., 1999. Generalized exponential distribution, Australian & New
Zealand Journal of Statistics, 41, 173-188.
Gupta, R.D., Kundu, D., 2001. Generalized exponential distribution: Different method of
estimations, Journal of Statistical Computation and Simulation, 69, 315-337.
Gupta, R.D., Kundu, D., 2001. Exponentiated exponential family: An alternative to gamma
and Weibull distributions, Biometrical Journal, 43, 117-130.
Kus, C., 2007. A new lifetime distribution. Computational Statistics & Data Analysis, 51,
4497-4509.
Lee, C., Famoye F., Olumolade, O., 2007. Beta-Weibull Distribution: Some properties and
26 Percontini et al.

applications to censored data. Journal of Modern Applied Statistical Methods, 6, 173-

186.
Lu, W., Shi, D., 2012. A new compounding life distribution: the Weibull-Poisson distribu-
tion, Journal of Applied Statistics, 39, 21-38.
Nadarajah, S., Kotz, S., 2004. The beta Gumbel distribution, Mathematical Problems in
Engineering, 10, 323-332.
Nadarajah, S., Gupta, A.K., 2004. The beta Fréchet distribution, Far East Journal of
Theoretical Statistics, 14, 15-24.
Nadarajah, S., Kotz, S., 2006. The beta exponential distribution, Reliability Engineering
and System Safety, 91, 689-697.