Mediterr. J. Math.

DOI 10.1007/s00009-015-0665-5
c Springer International Publishing 2015


A Two-Parameter Distribution Obtained by

Compounding the Generalized Exponential
and Exponential Distributions
Božidar V. Popović, Miroslav M. Ristić
and Gauss M. Cordeiro

Abstract. We introduce a new two-parameter lifetime distribution

obtained by compounding the generalized exponential and exponential
distributions. We assume that the shape parameter of the generalized
exponential distribution is a random variable having the exponential
distribution. The shapes of the density and hazard rate functions are
derived. The model parameters are estimated by maximum likelihood,
and an application of the proposed distribution is presented.
Mathematics Subject Classification. 62E15, 62P10, 62F10.
Keywords. Compounding model, Exponential distribution,
Generalized exponential distribution, Identifiability.

1. Introduction
There are many practical situations where well-known distributions do not
provide adequate fits to real data. Among the different methods for generating
new distributions, the compounding of some discrete and important lifetime
distributions has been very popular in lifetime modeling.
Adamidis and Loukas [1] introduced a two-parameter exponential–
geometric (EG) distribution by compounding the exponential and geometric
distributions. More general family of these distributions had been considered
by Marshall and Olkin [9]. Similarly, the exponential Poisson (EP) and expo-
nential logarithmic distributions were introduced by Kuş [6] and Tahmasbi
and Rezaei [15], respectively. Barreto-Souza et al. [3] proposed the Weibull
geometric (WG) model, while Lu and Shi [7] studied the Weibull–Poisson
(WP) distribution as a natural extension of the EG and EP distributions.
Further, Rodrigues et al. [13] defined the Weibull negative binomial distribu-
tion, which includes as sub-models the WG and WP distributions. Nadarajah
 B. V. Popović et al. MJOM

et al. [11] built in a Compounding R script for computing continuous distrib-

utions obtained by compounding a continuous and a discrete distribution.
Gupta and Kundu [5] introduced the generalized exponential (GE) dis-
tribution with cumulative distribution function (cdf) given by
G(x|α, λ) = (1 − e−λx )α , (1)
where x > 0, α > 0, λ > 0. They showed that the corresponding probability
density function (pdf) is log convex for 0 < α < 1 and log concave for α > 1.
Here, we propose a new distribution by assuming that the shape parameter
α > 0 of the GE distribution is a random variable itself. For a given α, let a
random variable Y have the GE cdf with parameters α > 0 and λ > 0 given
by (1). Let α be a random variable having the exponential distribution with
parameter θ > 0. Then, we define a new random variable X by the cdf (for
x > 0)

  −1
∞
1
F (x) = F (x; λ, θ) = G(x|α, λ) θ e−θα dα = 1 − log(1 − e−λx ) , λ, θ > 0.
0 θ
(2)

The pdf of X becomes

λe−λx
f (x) = f (x; λ, θ) =  2 . (3)
1
θ (1 − e−λx ) 1 − θ log(1 − e−λx )
Henceforth, we denote a random variable X having the density function (3)
by X ∼ GEE(λ, θ), where “GEE” stands for the generalized exponential
exponential distribution. The parameter λ is a scale parameter. It is easy to
prove that if the random variable X has the pdf f (x; λ, θ), then the random
variable Z = λX has the pdf f (x; 1, θ). The density function f (x; λ, θ) has
shapes given by Theorem 3.
We motivate the GEE distribution by the following facts:
1. Comparing with the GE and exponential distributions, one will see that
the pdf (3) provides more flexibility than these distributions. This fact
will make the GEE model more attract to the readers.
2. The next theorem links the GEE distribution with other well-known dis-
tributions.
Theorem 1. For the random variable X having pdf (3), we have:
(a) The random variable Y = − log(1−e−λX ) has the log-logistic distribution
with shape parameter β = 1 and scale parameter α = θ.
(b) The random variable Y = − log(1 − e−λX ) has the generalized Pareto
distribution with location parameter µ = 0, scale parameter σ = θ, and
shape parameter ξ = 1.
(c) The random variable Y = 1 − θ1 log(1 − e−λX ) has the Pareto distribution
with shape parameter one. 
(d) The random variable Y = − θ1 log(1 − e−λX ) has the Burr distribution
with shape parameters equal to two and one.
 Compounding the Generalized Exponential
 
(e) The random variable Y = log 1 − θ1 log(1 − e−λX ) has the exponential
distribution with unity scale parameter.
Proof. The proof is trivial and left as an exercise to the reader. 
Theorem 2. If the random variable V has the log-logistic distribution with
shape parameter β = 1 and scale parameter θ, then the random variable
X = − λ1 (1 − e−V ) has the distribution with pdf (3).
Proof. The proof is trivial and left as an exercise to the reader. 
Henceforth, we consider the following expansions:
∞
 q k
(1 − z)q = (−1)k z , if |z| < 1, (4)
k
k=0

∞
 zk
log(1 − z) = −z , if |z| < 1 (5)
k+1
k=0
and
−x (−1)n x(n)
= , (6)
n n!
where x(k) = x(x + 1) . . . (x + k − 1) denotes the rising factorial.
∞ −t
We shall use the exponential integral defined as Ei (x) = − −x e t dt.
The exponential integral can be easily computed in MATHEMATICA using
the function ExpIntegralEi[x].
In the sequel, we require the Whittaker’s function defined by
 +∞ k−1/2+m
e−z/2 z k −k−1/2+m t
Wk,m (z) = t 1 + e−t dt,
Γ 21 − k + m 0 z
where Γ(·) is the gamma function. The values of the Whittaker function can
be obtained using the MATHEMATICA function WhittakerW[k,m,z].
We use an equation of Gradshteyn and Ryzhik ([4], equation 0.314) for
a power series raised to a positive integer power p
⎛ ⎞p
 
⎝ an xn ⎠ = cp,n xn , (7)
n≥0 n≥0

where the coefficients cp,n can be determined using the following recurrence
relation (for m ≥ 1 and cp,0 = ap0 )
m

cp,m = (m a0 )−1 [k(p + 1) − m] ak cp,m−k .
k=1
The rest of the paper is organized as follows. We discuss the shapes of
the pdf and cdf of the new distribution in Sect. 2. We provide in Sect. 3
several mathematical properties of the GEE model such as the ordinary and
incomplete moments, probability-weighted moments (PWMs), mean devia-
tions, Bonferroni and Lorenz curves, generating function and order statistics.
The Rényi and Shannon entropies are derived in Sect. 4. The parameters
 B. V. Popović et al. MJOM

of the GEE distribution are estimated by maximum likelihood in Sect. 5.

Two applications to real data illustrate the potentiality of the new model in
Sect. 6. Finally, concluding remarks are addressed in Sect. 7.

2. Shapes of pdf and hrf

Here, we will study some shapes of the pdf and hazard rate function (hrf) of
the random variable X with the GEE distribution.
Theorem 3. If 0 < θ < 1−log 2, then there exist two positive real numbers a <
b such that f (x) is a decreasing function in (0, a) ∪ [b, ∞) and an increasing
function in [a, b). If θ > 1 − log 2, then the function f (x) is a decreasing
function. For both cases, we have that f (0+) = ∞.
Proof. The first derivative of the logarithm of the function f (x) can be
expressed as
λu(x)
[log f (x)]′ =  ,
θ(1 − e−λx ) 1 − θ1 log(1 − e−λx )
where the function u(x) is given by
u(x) = −θ + 2e−λx + log(1 − e−λx ).
The function u(x) is a unimodal function with u(0+) = −∞ and u(∞) =
−θ. The function u(x) has a maximum at x = λ−1 log 2, and it is equal to
1 − θ − log 2. Thus, if θ > 1 − log 2, the function u(x) is negative which
implies that f (x) is a decreasing function with f (0+) = ∞. In the second
case, 0 < θ < 1 − log 2, the function u(x) has two roots, say a < b. This
implies that u(x) is negative for x ∈ (0, a) ∪ [b, ∞) and positive for x ∈ [a, b).
So, we obtain that f (x) is a decreasing function for x ∈ (0, a) ∪ [b, ∞) and an
increasing function for x ∈ [a, b). Further, we have f (0+) = ∞. 

Based on Eqs. (2) and (3), the hrf h(x) of X is given by

λe−λx
h(x) =  1
.
(1 − e−λx ) [− log(1 − e−λx )] 1 − θ log(1 − e−λx )
Theorem 4. The hrf is bathtub-shaped for 0 < θ < 2 and decreasing for θ > 2.
For all parameter values, we have that h(∞) = λ.
Proof. The logarithm of h(x) is given by
s(x)
[log h(x)]′ =  1
,
θ(1 − e−λx ) [− log(1 − e−λx )] 1 − θ log(1 − e−λx )
where s(x) = λθe−λx + λ(θ − 2e−λx ) log(1 − e−λx ) − λ[log(1 − e−λx )]2 . If
0 < θ < 2, then there exists a positive real number a such that s(x) is
increasing in (0, a) and decreasing in (b, ∞) with s(0+) = −∞ and s(∞) = 0.
Thus, there exists a positive real number b such that, for x < b, the function
s(x) is negative and, for x > b, the function s(x) is positive. This fact implies
that h(x) is bathtub-shaped.
 Compounding the Generalized Exponential

a b

0.6
0.30

0.5
0.4
0.20

0.3
0.2
0.10

0.1
0.00

0.0
0 5 10 15 0 1 2 3 4 5

Figure 1. GEE pdf for λ = 0.25 (red), λ = 0.75 (blue),

λ = 1.25 (black) and: a θ = 0.15; b θ = 0.75 (color figure
online)

If θ > 2, then s(x) is an increasing function with s(0+) = −∞ and

s(∞) = 0. Thus, s(x) is a negative function, which implies that the hrf is a
decreasing function. 
Remark 1. The hazard rate function h(·) has a finite positive limit λ, which
implies according to Marshall and Olkin ([8], Proposition B.3 ) that residual
life function converges to the exponential distribution with parameter λ.
Figures 1 and 2 display, respectively, shapes of the pdf and hrf for
different parameter values.
The finite limit property of the hazard rate function allows to use to
prove the identifiability of the distribution with respect to parameters λ and θ.
Theorem 5. The distribution function F given by (2) is identifiable with
respect to parameters θ and λ.

a b
4.0
1.4

3.5
1.3

3.0
1.2

2.5
1.1

2.0
1.0

1.5
0.9

0 1 2 3 4 5 0 1 2 3 4 5

Figure 2. GEE hrf for λ = 0.25 (red), λ = 0.75 (blue),

λ = 1.25 (black) and: a θ = 0.75; b θ = 3.75 (color figure
online)
 B. V. Popović et al. MJOM

Proof. Let us suppose that F (x; λ1 , θ1 ) = F (x; λ2 , θ2 ) for all x > 0. We will
show that this condition implies that λ1 = λ2 and θ1 = θ2 . First, we note
that this condition implies that h(x; λ1 , θ1 ) = h(x; λ2 , θ2 ) for all x > 0. Now,
letting x tends to ∞ on both sides, and using the result from Theorem 4
that h(∞; λ, θ) = λ, we obtain that λ1 = λ2 . Next, condition F (x; λ1 , θ1 ) =
F (x; λ2 , θ2 ) implies that θ1 (1 − e−λ1 x ) = θ2 (1 − e−λ2 x ) and since λ1 = λ2 ,
we obtain finally that θ1 = θ2 . Thus, we have proved the identifiability of the
distribution function F . 

3. Mathematical Properties
Without loss of generality in order to simplify final expressions given in the
Sects. 3 and 4, we will set λ = 1.

3.1. Moments
By means of the Theorem 2, the nth moment of X can be expressed as
 +∞ n
n −V n n 1 [log(1 − e−v )]
E(X ) = E − log(1 − e ) = (−1) 2 dv .
θ 0 1 + θ1 v
By setting 1 + θ1 v = u, using both (5) and (7), the last equation reduces
to
  +∞
e−θ(n+k)v
E(X n ) = cn,k dv, (8)
0 (v + 1)2
k≥0

where (for i ≥ 1)
i
1  [j(n + 1) − i]
cn,i = cn,k−j , (9)
i j=1 (j + 1)

and cn, 0 = 1. From the result (3.353.2) in [4], Eq. (8) can be expressed as
  
E(X n ) = cn,k 1 − e−θ(n+k) Ei [−θ(n + k)] . (10)
k≥0

Incomplete moments play important role for inequality measurement,

for example mean deviations and the Lorenz curve, Pietra and Gini measures.
Using a similar approach to the one for ordinary moments and equation
(3.353.1) in [4], the nth incomplete moment of X is given by
 
(1 − e−z )n+k θ(n + k)
mn (z) = cn,k 1 +
k≥0
1 − θ log(1 − e −x ) (1 − e−z )−1/θ
 
Ei (log(1 − e−z ) − θ)(n + k) , (11)
where cn,k is defined by (9).
Probability-weighted moments (PWMs) cover the summarization and
description of theoretical distributions. The primary use of these moments is
in the estimation of parameters for a probability distribution whose inverse
cannot be expressed explicitly. One can use PWMs when maximum likelihood
 Compounding the Generalized Exponential

estimates are unavailable or difficult to compute. The (s, r)th PWM of X is

formally defined as
 ∞
s r
τs,r = E {X F (X) } = xs F (x)r f (x) dx.
0

By equation (3.353.1) in [4] and using similar arguments as above, we obtain

 r+1
  r+1−l
τs,r = cn,k (l − 1)! [−θ(s + k)]
k≥0 l=1

r+1 θ(s+k)
− [−θ(s + k)] e Ei [−θ(s + k)] .

3.2. Mean Deviations

The amount of scatter in a population is evidently measured to some extent
by the totality of deviations from the mean and median. The mean deviations
+∞
of X about the mean and the median are defined by δ1 = 0 |x − µ| f (x) dx
+∞
and δ2 = 0 |x − M | f (x) dx, where µ = E(X) and M is the median.
1
Hence, the quantile function (qf) of X is F −1 (p) = − log 1 − eθ(1− p ) , where
p ∈ [0, 1). So, the median equals to M = F −1 (1/2) = − log 1 − e−θ .
These measures can be calculated using the relationships δ1 = 2[µF (µ)−
m1 (µ)] and δ2 = E(X) − 2m1 (M ). Clearly, F (M ) and F (µ) are easily calcu-
lated from Eq. (2), whereas the first ordinary moment and the first incomplete
moment follow from (10) and (11), respectively.
Equation (11) with n = 1 is useful to derive the Bonferroni and Lorenz
curves and mean deviations, that is, the Bonferroni and Lorenz curves are
defined (for a given probability π) by B(π) = m1 (q)/(π µ′1 ) and L(π) =
m1 (q)/µ′1 , respectively, where q = F −1 (π) can be determined from the qf
above.

3.3. Generating Function

The moment generating function (mgf) M (t) = E(etX ) of X can be obtained
using the same approach for the moments. We have
 +∞
(1 − e−θv )−t
M (t) = dv .
0 (v + 1)2
Based on (4), equation (3.353.2) in [4] and (6), the last equation reduces to
 (t)(k)  
M (t) = 1 + θ k eθk Ei (−θk) .
k!
k≥0

3.4. Order Statistics

Order statistics make their appearance in many areas of statistical theory and
practice. Suppose X1 , . . . , Xn is a random sample from the GEE distribution.
 B. V. Popović et al. MJOM

Let Xi:n denote the ith order statistic. From Eqs. (2) and (3), the power series
expansion of the pdf of Xi:n can be expressed as
n−i
K e−x  (−1)j n−i
fi:n (x) =  i+j+1 ,
θ(1 − e−x ) j=0 1 − 1
log(1 − e−x ) j
θ

n!
where K = (i−1)!(n−i)! .
We obtain the moments of the order statistics using a similar approach
of Sect. 2. We have
n−i 

n (−1)i cn,k n−1
E(Xi:n )=K
j=0
(i + j)! j
k≥0
 i+j 
 (l − 1)! i+j θ(n+k)
× l
− [θ(n + k)] e Ei [−θ(n + k)] . (12)
l=1 [−θ(n + k)]

The importance of the moments of order statistics arises in the applied

statistics such as quality control testing and reliability. If the reliability of an
item is high, the duration of “all items fail” life test can be too expensive in
both time and money. This fact prevents a practitioner from knowing enough
about the product in a relatively short time. Therefore, a practitioner needs
to predict the failure of future items based on the times of a few early failures.
These predictions are often based on moments of order statistics.
The L-moments are summary statistics for distributions and data sam-
ples. They are analogous to ordinary moments but are computed from linear
functions of the ordered data values. The rth L-moment of X is defined by
r−1
 r−1 r−1+j
λr = (−1)r−1−j βj ,
j=0
j j

where βj = E{XF (X)j }. In particular, λ1 = β0 , λ2 = 2β1 − β0 , λ3 =

6β2 − 6β1 + β0 , and λ4 = 20β3 − 30β2 + 12β1 − β0 . In general, βr = (r +
1)−1 E(Xr+1:r+1 ), so it can be computed from (12). The L-moments have
several advantages over ordinary moments; for example, they apply for any
distribution having finite mean; no higher-order moments need be finite.

4. Entropy
An entropy is a measure of variation or uncertainty of a random variable.
Two popular entropy measures are the Rényi and Shannon entropies [12,14].
The Rényi entropy of a random variable with pdf f (·) is defined as
 ∞
1
IR (γ) = log f γ (x)dx
1−γ 0

for γ > 0 and γ = 1. The Shannon entropy of a random variable is defined

by E{− log[f (X)]}. It is the special case of the Rényi entropy when γ ↑ 1.
 Compounding the Generalized Exponential

Here, we derive explicit expressions for the Rényi and Shannon entropies of
X. We have
  
γ
1 1 ∞
e−γx
IR (γ) = log  2γ dx
1−γ θ 0 (1 − e−x )γ 1 − 1
log(1 − e−x )
θ
  
γ
1 1 ∞
e−(γ−1)x e−x
= log   dx .
1−γ θ 0 (1 − e−x )γ−1 (1 − e−x ) 1 − 1 log(1 − e−x ) 2γ
θ

Then, using the binomial expansion of (1 − e−x )1−γ and the same algebra
which leads to (10), the last equation becomes
⎧
⎨  1−γ γ−1+i
1 γ−1
1
IR (γ) = 1−γ log θ (−1)i+j
⎩ i j
i,j≥0
⎫
 +∞ ⎬
e−θjv
dx . (13)
0 (1 + v)2γ ⎭

By equation (3.382.3) from [4], Eq. (13) turns out to be

⎧ ⎫
1 ⎨ 1 γ−1  !1−γ " !γ −1+i" jθ
⎬
i+j γ−1
IR (γ) = log (−1) (jθ) e 2 W−γ; 1−2γ (jθ) .
1−γ ⎩ θ i j 2 ⎭
i,j≥0

(14)

Equation (14) is very complicated for limiting, and then, we can derive
an explicit expression for the Shannon entropy based on its definition. We
can write
 
E{− log[f (X)]} = log θ − E(X) + E log(1 − e−X )
  #
1 −X
+ 2E log 1 − log(1 − e ) . (15)
θ
The first expectation in (15) follows easily from (10) for n = 1. Using
Eq. (5) and the same approach of Sect. 2, we obtain
   (−1)j i
E log(1 − e−X ) = − [1 + jθ Ei (−jθ)]. (16)
i j
i,j≥0
1 −x
Setting 1 − θ log(1 − e ) = u, we easily obtain
  #
1
E log 1 − log(1 − e−X ) = 1. (17)
θ
By inserting (10) (for n = 1), (16), and (17) into (15), we obtain the Shannon
entropy.

5. Estimation
In this section, we consider the maximum likelihood estimation (MLE) of the
unknown parameters λ and θ. Suppose the observed sample x1 , . . . , xn of size
 B. V. Popović et al. MJOM

n from the distribution (3). The log-likelihood function for (λ, θ) is given by
n
 n

l(λ, θ) = n log λ − n log θ − λ xi − log(1 − e−λxi )
i=1 i=1
n
  
1
−2 log 1 − log(1 − e−λxi ) .
i=1
θ
So, the components of the score function satisfy equations
n n
∂l(λ, θ) n  xi 2 xi e−λxi
= − +   = 0,
∂λ λ i=1 1 − e−λxi θ i=1 (1 − e−λxi ) 1 − θ1 log(1 − e−λxi )
(18)
n
∂l(λ, θ) n 2  log(1 − e−λxi )
=− − 2 = 0. (19)
∂θ θ θ i=1 1 − θ1 log(1 − e−λxi )
Now, we will study the existence and uniqueness of the MLE estimates
when the other parameter is known (or given).
Theorem 6. If the parameter θ is known, then the equation (18) has at least
one root on the interval (0, +∞).
∂l(λ,θ) ∂l(λ,θ)
$nOne can readily verify that limλ→0 ∂λ = +∞ and limλ→+∞ ∂λ
Proof.
= − i=1 xi . So, there exists at least one solution on the interval (0, +∞).
This completes the proof. 
Theorem 7. Let us suppose that the parameter λ is known. Then, the root
of equation (19) lies in the interval − log 1 − e−λx(n) , − log 1 − e−λx(1)
and is unique for all λ > 0, where x(1) = min(x1 , x2 , . . . , xn ) and x(n) =
max(x1 , x2 , . . . , xn ).
$n yi
Proof. Let us define the function ψ(θ) = − n2 + i=1 θ+y i
, where yi =
− log(1 − e−λxi ). Then, the function ∂l(λ,θ)
∂θ can be represented as ∂l(λ,θ)
∂θ =
2
θ ψ(θ), which implies that we can derive its behavior through the behavior of
the function ψ(θ). We have that ψ(θ) is almost surely decreasing on (0, +∞),
and it holds lim ψ(θ) = n2 and lim ψ(θ) = − n2 . Thus, the function ψ(θ) has
θ↓0 θ↑0
the unique root, which implies that the Eq. (19) has the unique solution.
Also, since y(1) + yi ≤ 2yi for all i = 1, 2, . . . , n then
n n
n  yi n 1
ψ y(1) = − + ≥− + = 0,
2 i=1 y(1) + yi 2 i=1 2
and analogously
n n
n  yi n 1
ψ y(n) =− + ≤− + = 0.
2 i=1 y(n) + yi 2 i=1 2
From the last two equations, we found interval which is the solution of the
Eq. (19). 
 Compounding the Generalized Exponential

Table 1. Descriptive statistics for the Aarset data set

Mean Median SD Skewness Kurtosis Min Max

45.69 48.5 32.8352 −0.1378 1.4138 0.1 86

Table 2. Estimated parameters, AIC, KS, and p value for

Aarset data

Model %
λ α
% θ% β% AIC KS p
GE 0.0187 0.7798 – – 483.99 0.2042 0.0309
(0.0036) (0.1352)
GEE 0.0383 – 0.2158 – 473.57 0.1491 0.2162
(0.0058) (0.0883)
MOE 0.0326 2.6214 – – 483.10 0.1617 0.1464
(0.0070) (1.0927)
Weibull 0.9490 – – 44.9125 486 0.1928 0.0486
(0.1095) (6.9465)
0.06
0.05
0.04
Density

0.03
0.02
0.01
0.00

0 20 40 60 80

Figure 3. Histogram of the Aarset data and fitted density

functions of the Weibull (red), GE (yellow), GEE (brown),
and MOE (pink) distributions (color figure online)

6. Application
Here, we use two real data sets to compare the fits of the GEE distribution
with other fits from the Weibull, generalized exponential (GE), and Marshal–
Olkin exponential (MOE) distributions. The parameters are estimated using
maximum likelihood and reported jointly with standard errors in parentheses
in Tables 2 and 4.
 B. V. Popović et al. MJOM

1.0
0.8
0.6
Fn(x)

0.4
0.2
0.0

0 20 40 60 80
x

Figure 4. Empirical distribution function of the current data

and fitted cdfs of the Weibull (red), GE (yellow), GEE
(brown), and MOE (pink) distributions (color figure online)

Table 3. Descriptive statistics for the test stop data

Mean Median SD Skewness Kurtosis Min Max

0.0014 5.8460 3.4375 −0.2294 1.7753 0.0014 10.76
0.20
0.15
Density

0.10
0.05
0.00

0 2 4 6 8 10 12

Figure 5. Histogram of the test stopped data and fitted pdfs

of the Weibull (red), GE (yellow), GEE (brown), and MOE
(pink) distributions (color figure online)

6.1. Application 1: Aarset Data

This data set appears in [2]. The data represent the lifetimes of 50 devices.
Table 1 gives a data set descriptive summary.
 Compounding the Generalized Exponential

1.0
0.8
0.6
Fn(x)

0.4
0.2
0.0

0 2 4 6 8 10 12
x

Figure 6. Empirical distribution function of the current data

and fitted cdfs of the Weibull (red), GE (yellow), GEE
(brown), and MOE (pink) distributions (color figure online)

Table 4. Estimated parameter, AIC, KS, and p, values for

the test stopped data

Model %
λ α
% θ% β% AIC KS p
GE 0.1570 0.8377 – – 113.24 0.2493 0.1397
(0.0467) (0.2300)
GEE 0.4202 – 0.0974 – 100.41 0.1375 0.7958
(0.0939) (0.0707)
MOE 0.3984 8.0859 – – 107.78 0.1472 0.7254
(0.1046) (5.7289)
Weibull 1.0892 – – 5.8163 113.50 0.2205 0.2465
(0.2210) (1.2218)

Figure 3 displays the histogram of the current data and the fitted pdfs.
Figure 4 displays the empirical cdf for the current data superimposed with
the fitted cdfs. These two figures reinforce that the GEE distribution provides
the best fit to the blood cancer data.
We compute the MLEs of the model parameters and adopt the Akaike
information criteria (AIC) and p values corresponding to the Kolmogorov–
Smirnov (KS) test for comparing the fitted models. The results in Table 2
indicate that the GEE distribution has the lowest AIC and the largest p value
among all fitted distributions.

6.2. Application 2: Test Stopped Data

Here, we use data points representing failure times. The data are taken from
[10]. The descriptive statistics of the failure times are given in Table 3.
 B. V. Popović et al. MJOM

From the figures in Table 3, it is obvious that the GEE distribution

provides a better fit to the failure time data than the other models.
The histogram of the failure time data and the fitted pdfs is displayed in
Fig. 5, whereas the empirical cdf of these data and the fitted cdfs is displayed
in Fig. 6. These two figures reinforce our earlier observation that the GEE
distribution provides the best fit.

7. Concluding Remarks
We introduce a new two-parameter model, called the generalized exponential
exponential (GEE) distribution, and study some of its structural properties.
We provide explicit expressions for the density function, moments and incom-
plete moments, probability-weighted moments, generating function, mean
deviations, Bonferroni and Lorenz curves, and two measures of entropy. Our
formulas related with the GEE model are manageable and, with the use of
modern computer resources with analytic and numerical capabilities, may
turn into adequate tools comprising the arsenal of applied statisticians. The
model parameters are estimated by maximum likelihood, and the existence
of the ML estimates is proved. This distribution is a very competitive model
to other lifetime distributions. In fact, we prove that this can be superior to
some widely known lifetime distributions by means of two examples with real
data.
Acknowledgements
We are grateful to the Editor and anonymous referees for comments that
greatly improved the quality of the paper. The authors are very grateful to
one of the referee for providing them the proof of Theorem 7.

References
[1] Adamidis, K., Loukas, S.: A lifetime distribution with decreasing failure
rate. Stat. Probab. Lett. 39, 35–42 (1998)
[2] Aarset, M.V.: How to identify bath tub hazard rate. IEEE Trans. Reliab. 36(1),
106–108 (1987)
[3] Barreto-Souza, W., Morais, A.L., Cordeiro, G.M.: The Weibull-geometric dis-
tribution. J. Stat. Simul. Comput. 81, 645–657 (2010)
[4] Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series, and Products. 7th
edn. Academic Press, San Diego (2007)
[5] Gupta, R.D., Kundu, D.: Generalized exponential distribution. Aust. N. Z. J.
Stat. 41, 173–188 (1999)
[6] Kuş, C.: A new lifetime distribution. Comput. Stat. Data Anal. 51, 4497–
4509 (2007)
[7] Lu, W., Shi, D.: A new compounding life distribution: the Weibull-Poisson
distribution. J. Appl. Stat. 39, 21–38 (2012)
[8] Marshall, A.W., Olkin, I.: Life Distributions. Springer, New York (2007)
 Compounding the Generalized Exponential

[9] Marshall, A.W., Olkin, I.: A new method for adding a parameter to a family
of distributions with application to the exponential and Weibull families. Bio-
metrika. 84, 641–652 (1997)
[10] Murthy, D.N.P., Xie, M., Jiang, R.: Weibull Models. Wiley, New York (2005)
[11] Nadarajah, S., Popović, B.V., Ristić, M.M.: Compounding: an R package for
computing continuous distributions obtained by compounding a continuous
and a discrete distribution. Comput. Stat. 28, 977–992 (2013)
[12] Rényi, A.: On measures of entropy and information. In: Proceedings of the 4th
Berkeley Symposium on Mathematical Statistics and Probability, Volume I,
PP. 547–561. University of California Press, Berkeley (1961)
[13] Rodrigues, C., Cordeiro, G.M., Demetrio, C.G.B., Ortega, E.M.M.: The
Weibull negative binomial distribution. Adv. Appl. Stat. 22, 25–55 (2011)
[14] Shannon, C.E.: Prediction and entropy of printed English. The Bell System
Tech. J. 30, 50–64 (1951)
[15] Tahmasbi, R., Rezaei, S.: A two-parameter lifetime distribution with decreasing
failure rate. Comput. Stat. Data Anal. 52, 3889–3901 (2008)

Božidar V. Popović
Faculty of Philosophy
University of Montenegro
Nikšić, Montenegro
e-mail: [email protected]

Miroslav M. Ristić
Faculty of Sciences and Mathematics
University of Niš
Niš, Serbia
e-mail: [email protected]

Gauss M. Cordeiro
Departamento de Estatı́stica
Universidade Federal de Pernambuco
Recife, PE, Brazil
e-mail: [email protected]

Received: January 28, 2015.

Revised: November 9, 2015.
Accepted: November 26, 2015.