REVSTAT – Statistical Journal

Volume 15, Number 1, January 2017, 65–88

THE BETA GENERALIZED

INVERTED EXPONENTIAL DISTRIBUTION
WITH REAL DATA APPLICATIONS

Authors: R.A. Bakoban

– Department of Statistics, Faculty of Science AL-Faisaliah Campus,
King Abdulaziz University, Jeddah, Saudi Arabia
rbakoban@kau.edu.sa
Hanaa H. Abu-Zinadah
– Department of Statistics, Faculty of Science, AL-Faisaliah Campus,
King Abdulaziz University, P.O. Box 32691, Jeddah 21438, Saudi Arabia
habuzinadah@kau.edu.sa

Received: July 2014 Revised: April 2015 Accepted: May 2015

Abstract:
• The four-parameter beta generalized inverted exponential distribution is considered
in this article. Various properties of the model with graphs of the density function
are investigated. Moreover, the maximum likelihood method of estimation is used
for estimating the parameters of the model under complete samples. An asymptotic
Fisher information matrix of the estimators is found. Additionally, confidence interval
estimates of the parameters are obtained. The performances of findings of the arti-
cle are shown by demonstrating various numerical illustrations through Monte Carlo
simulation studies. Finally, applications on real data-sets are provided.

Key-Words:

• beta generalized inverted exponential distribution; Fisher information matrix;

goodness-of-fit test; maximum likelihood estimator; Monte Carlo simulation.

AMS Subject Classification:

• 62-07, 62E20, 62F10, 62F25, 62N02.

66 R.A. Bakoban and Hanaa H. Abu-Zinadah
The beta Generalized Inverted Exponential Distribution... 67

1. INTRODUCTION

The generalized inverted exponential distribution (GIED) was introduced

first by Abouammoh and Alshingiti (2009). It is a generalized form of the inverted
exponential distribution (IED). IED has been studied by Keller and Kamath
(1982) and Duran and Lewis (1989). GIED has good statistical and reliability
properties. It fits various shapes of failure rates.

The probability density function (pdf) of a two-parameter GIED is given

by
     α−1
αλ −λ −λ
(1.1) f (x) = exp 1 − exp , x > 0, α, λ > 0 ,
x2 x x
and the cumulative distribution function (cdf) is given by
−λ α
  
(1.2) F (x) = 1 − 1 − exp , x > 0, α, λ > 0 .
x
In the last few years, new classes of distributions have been found by extending
certain distributions such that these new classes will have more applications in
reliability, biology and other fields.

Let G(t) be a cdf of a random variable T , such that

Z G(t)
1
(1.3) F (t) = ̟a−1 (1 − ̟)b−1 d̟ ,
B(a, b) 0
R1
where a > 0, b > 0, and B(a, b) = 0 ̟a−1 (1 − ̟)b−1 d̟ is the beta function. The
skewness of the distribution is controlled by the two parameters a and b. The
cdf G(t) could be any arbitrary distribution, and, consequently, F is named the
beta G distribution. The previous formula in (1.3) was defined by Eugene et al.
(2002) as a class of generalized distributions.

The beta normal distribution (BND) was introduced by Eugene et al.

(2002). They used the cdf G(t) of the normal distribution in (1.3) and derived
some moments of the distribution. Expanding on this work, Gupta and Nadara-
jah (2004) established more general moments of BND. Based on the cdf G(t) of
the Gumbel distribution, Nadarajah and Kotz (2004) presented the beta Gum-
bel distribution and provided closed form expressions for the moments and the
asymptotic distribution of the extreme order statistics and obtained the maxi-
mum likelihood estimators (MLE) of the parameters. Further, by using the cdf
G(t) of the exponential distribution, Nadarajah and Kotz (2005) considered the
beta exponential distribution. They studied the first four cumulants, the moment
generating function, and the extreme order statistics and found the MLE. Fur-
thermore, Lee et al. (2007) considered the beta Weibull distribution and studied
applications based on censored data.
68 R.A. Bakoban and Hanaa H. Abu-Zinadah

Recently, Barreto-Souza et al. (2010) proposed the beta generalized expo-

nential distribution by taking G(t) in (1.3) to be the cdf of the exponentiated
exponential distribution and discussed the MLE of its parameters. Addition-
ally, Nassar and Nada (2011) presented several properties of the beta general-
ized Pareto distribution. They estimated the distribution’s parameters using the
MLE. An application on actual tax revenue data was investigated. Paranaiba
et al. (2011) discussed the beta Burr XII distribution. Mahmoudi (2011) pre-
sented the beta generalized Pareto distribution. Cordeiro and Lemonte (2011)
investigated the beta Laplace distribution. Zea et al. (2012) studied statistical
properties and inference of the beta exponentiated Pareto distribution (BEPD).
They provided an application of the BEPD to remission times of bladder cancer.
Leão et al. (2013) studied the beta inverse Rayleigh distribution. They provided
various properties, including the quantile function, moments, mean deviations,
Bonferroni and Lorenz curves, Rényi and Shannon entropies and order statistics,
as well as the MLE. Baharith et al. (2014) discussed properties, the MLE and the
Fisher information matrix for the beta generalized inverse Weibull distribution.

In this paper, a new beta distribution is introduced by taking G(·) to be the

GIED, and we refer to it as the beta generalized inverted exponential distribution
(BGIED). In Section 2, the BGIED is defined. Statistical properties of the model
are derived in Section 3. Maximum likelihood estimators of the parameters are
derived in Section 4. In Section 5, the asymptotic Fisher information matrix is
investigated. Additionally, interval estimates of the parameters are found using
the maximum likelihood method in Section 6. Section 7 explains the simula-
tion studies that illustrate the theoretical results. Finally, Section 8 provides
applications to real data-sets. Various conclusions are addressed in Section 9.

2. BETA GENERALIZED INVERTED EXPONENTIAL DISTRI-

BUTION

In this section, we introduce the four-parameter beta generalized inverted

exponential distribution (BGIED) by assuming G(x) to be the cdf of the gen-
eralized inverted exponential distribution (GIED). Substituting (1.2), the cdf of
GIED, into (1.3), the cdf of the BGIED is obtained in the following form
Z 1−[1−exp( −λ )]α
1 x
(2.1) F (x) = ̟a−1 (1 − ̟)b−1 d̟ ,
B(a, b) 0
x > 0, a, b, α and λ > 0 .
The pdf of the BGIED takes the form
 a−1 
−λ α −λ αb−1


αλ exp −λ
   
x
(2.2) f (x) = 1 − 1 − exp 1 − exp ,
x2 B(a, b) x x
x > 0, a, b, α and λ > 0 .
The beta Generalized Inverted Exponential Distribution... 69

For a positive real value a > 0, (2.2) can be rewritten as an infinite power
series in the form
−λ α(b+k)−1

∞
αλ exp −λ X (−1)k Γ(a)   
x
(2.3) f (x) = 1 − exp ,
x2 B(a, b) k!Γ(a − k) x
k=0
x > 0, a, b, α, and λ > 0 .
From (2.3), the corresponding cdf can be written as follows
−λ α(b+k)
∞
(−1)k+1
  
1 X
(2.4) F (x) = 1 − exp ,
B(a, b) a(b + k)B(a − k, k + 1) x
k=0
x > 0, a, b, α and λ > 0 .
The GIED is a special case of (2.2) when a = b = 1. Therefore, we can assume all
of the properties of the GIED that were investigated by Abouammoh and Alshin-
giti (2009) still hold. Additionally, when α = 1 in (2.2), the BIED is obtained,
which is related to the BGIWD when the shape parameters are equal to one and
has been discussed by Baharith et al. (2014).

Figure 1: The pdf curves of the BGIED with (a, b, α, λ).

3. STATISTICAL PROPERTIES

3.1. The reliability and hazard functions

The reliability function is the probability of no failure occurring before time t.

Alternately, the hazard function is the instantaneous rate of failure at a given
time. These two functions are very important properties of a lifetime distribution.
70 R.A. Bakoban and Hanaa H. Abu-Zinadah

The reliability function of the BGIED is given by

−λ α(b+k)
∞
(−1)k+1
  
1 X
(3.1) R(x) = 1 − 1 − exp ,
B(a, b) a(b + k)B(a − k, k + 1) x
k=0

x > 0, a, b, α and λ > 0 ,

and the corresponding hazard function of the BGIED can be written as

αλ exp( −λ ∞
x ) (−1)k Γ(a)  −λ α(b+k)−1
P

2
x B(a,b) k!Γ(a−k) 1 − exp x
k=0
(3.2) h(x) = ∞ ,
1 (−1)k+1 −λ α(b+k)
P 

1 − B(a,b) a(b+k)B(a−k,k+1) 1 − exp x
k=0

x > 0, a, b, α and λ > 0 .

Figure 2 shows different choices for the parameters of the BGIED. Additionally, it
is shown from Figure 3 that the hazard function of the BGIED has an upside down
bathtub shape. As is shown, the hazard function increases and then decreases.

Figure 2: The reliability curves of the BGIED with (a, b, α, λ).

The upside down bathtub hazard function indicates that the risk of failing de-
creases as soon as the item has passed a specific time, during which it may have
experienced some type of stress. Thus, the BGIED shows good statistical behav-
ior based on these two functions.
The beta Generalized Inverted Exponential Distribution... 71

Figure 3: The hazard curves of the BGIED with (a, b, α, λ).

3.2. Moments and various measures

The rth moment about the origin, µ′r = E(X r ) of a BGIED with pdf (2.2)
in the non-closed form is
 a−1 
−λ α −λ αb−1
−λ 


αλ exp
Z ∞    
′ r x
µr = x 1 − 1 − exp 1 − exp dx ,
0 x2 B(a, b) x x
r = 1, 2, ...
that is, for k ≥ r, µ′r takes the closed form
∞ X ∞
λr X (−1)k+j (j + 1)r−1
(3.3) µ′r =
B(a, b) a(b + k)B(a − k, k + 1)B(j + 1, α(b + k) − j)
k=0 j=0
(∞ )
X (−1)i
× + Er (1) ,
i!(i − r + 1)
i=0

where B(a, b) is the beta function, and En (z) is called the exponential integral
function (Abramowitz and Stegun (1972)), which is defined as
Z ∞
exp (−zt)
(3.4) En (z) = dt .
1 tn
Substituting r = 1 in (3.3), we obtain the mean of the BGIED as follows
∞ X ∞
λ X (−1)k+j
(3.5) µ =
B(a, b) a(b + k)B(a − k, k + 1)B(j + 1, α(b + k) − j)
k=1 j=0
(∞ )
X (−1)i
× + E1 (1) ,
i2 (i − 1)!
i=1

where E1 (1) = 0.577216 is Euler’s constant.

72 R.A. Bakoban and Hanaa H. Abu-Zinadah

Additionally, the variance of the BGIED can be found from

∞ ∞
λ2 X X (−1)(k+j) (j + 1)
(3.6) Var(x) =
B(a, b) a(b + k)B(a − k, k + 1)B(j + 1, α(b + k) − j)
k=2 j=0
(∞ )
X (−1)i
× + E2 (1) − µ2 .
i!(i − 1)
i=0

3.3. Quantile function and various related measures

The quantile function of the BGIED corresponding to (2.2) is

. n  1o
(3.7) q(u) = −λ log 1 − 1 − Iu−1 (a, b) α , 0<u<1,

where Iu−1 (a, b) is the inverse of the incomplete beta function with parameters a
and b, such that Z u
1
Iu (a, b) = ̟a−1 (1 − ̟)b−1 d̟ ,
B(a, b) 0
The above form of q(u) allows us to derive the following forms of statistical
measures for the BGIED:

1. The first quartile Q1, the second quartile Q2 (median), and the third
quartile Q3 of the BGIED correspond to the values u = 0.25, 0.50, and
0.75, respectively


2. The median (m), also, can be found using (2.4) such that 1− exp −λ
m < 1,
for a = 1, and then
−λ
(3.8) m = h 1
i.
log 1 − (−0.5) αb

3. The skewness and kurtosis can be calculated by using the following

relations, respectively:

Bowley’s skewness is based on quartiles; Kenney and Keeping (1962) cal-

culated it as follows
Q3 − 2 Q2 + Q1
(3.9) υ3 = ,
Q3 − Q1
Moors’ kurtosis (Moors (1988)) is based on octiles via the form
q(7/8) − q(5/8) − q(3/8) + q(1/8)
(3.10) υ4 = ,
q(6/8) − q(2/8)
where q(·) represents the quantile function defined in (3.7).
The beta Generalized Inverted Exponential Distribution... 73

When a = b = 1 in (2.3), (3.3) and (3.7) give the moments and the quantile
of GIED, and, when a = b = α = 1 in (2.3), (3.3) and (3.7) give the moments and
the quantile of IED. Therefore, all measures above are satisfied for GIED when
a = b = 1, and for IED when a = b = α = 1.

3.4. The mean deviation

Let X be a BGIED random variable with mean µ = E(X) and median m.

In this subsection, the mean deviation from the mean and the mean deviation
from the median are derived.

3.4.1. The mean deviation from the mean can be found from the following theorem:

Theorem 1. The mean deviation from the mean of the BGIED is in the
form
∞
∞ X
2 X (−1)k+1+j
E(|X − µ|) =
B(a, b) a(b + k)B(a − k, k + 1) [α(b + k) + 1]
k=0 j=0
µ exp (−jλ/µ) − jλΓ (0, jλ/µ)
× ,
B [j + 1, α(b + k) − j + 1]
R∞
where Γ (a, z) = z ta−1 exp (−t) dt.

Proof: The mean deviation from the mean can be defined as

Z ∞
E(|X − µ|) = |X − µ| f (x) dx
0
Z µ
= 2 (X − µ) f (x) dx
0
= 2µ F (µ) − 2I(µ) ,
Rz
where I(z) = 0 t dG(t), and d [t.dG(t)] = G(t) dt + t dG(t).
Rµ
Therefore, E(|X − µ|) = 2 0 F (x) dx.

Using (2.4), and expanding the term (1 − exp (−λ/x))α(b+k) we obtain

∞ X ∞
2 X (−1)(k+1+j)
E(|X − µ|) =
B(a, b) a(b + k)B(a − k, k + 1) [α(b + k) + 1]
k=0 j=0
Z µ
1
× exp (−jλ/x) dx ,
B(j + 1, α(b + k) − j + 1) 0
74 R.A. Bakoban and Hanaa H. Abu-Zinadah

where
Z c
(3.11) exp (−jλ/x) dx = c exp (−jλ/c) − jλΓ (0, jλ/c) .
0

Hence, the theorem is proved.

3.4.2. The mean deviation from the median can be found from the following theorem:

Theorem 2. The mean deviation from the median of the BGIED is in the
form
∞
∞ X
2 X (−1)(k+j) jλ
E(|X − m|) = µ +
B(a, b) a(b + k)B(a − k, k + 1) [α(b + k) + 1]
k=0 j=0
Γ (0, jλ/m)
× , jλ > 0, m > 0 .
B(j + 1, α(b + k) − j + 1)

Proof: The mean deviation from the median can be defined as

Z ∞
E(|X − m|) = |x − m|f (x)dx
0
Z m Z m Z ∞
= 2 (m − x) f (x)dx − (m − x) f (x)dx + (x − m) f (x)dx
Z0 m Z0 ∞ m

(3.12) = 2 (m − x) f (x)dx + (x − m) f (x)dx

0 0
 Z m 
= µ − 2 mF (m) − F (x)dx
0
Z m
= µ−m+2 F (x)dx .
0

Substituting (2.4) into (3.12) and using (3.11), we obtain

E(|X − m|) = µ − m
∞
∞ X
2 X (−1)k+j+1 [m exp (−jλ/m) − jλΓ (0, jλ/m)]
+
B(a, b) a(b + k)B(a − k, k + 1) [α(b + k) + 1]
k=0 j=0
1
×
B [j + 1, α(b + k) − j + 1]
= µ − m + 2mF (m)
∞
∞ X
2 X (−1)k+j jλΓ (0, jλ/m)
+
B(a, b) a(b + k)B(a − k, k + 1) [α(b + k) + 1]
k=0 j=0
1
× .
B [j + 1, α(b + k) − j + 1]
Hence, the theorem is proved.
The beta Generalized Inverted Exponential Distribution... 75

3.5. The mode

The mode for the BGIED can be found by differentiating f (x) with respect
to x; thus, (2.2) gives

−2 λ λ
f ′ (x) = f (x) + 2 − (αb − 1) [1 − exp (−λ/x)]−1 2 exp (−λ/x)
x x x
(3.13) + (a − 1) [1 − (1 − exp (−λ/x))α ]−1

αλ α−1
× 2 exp (−λ/x) (1 − exp (−λ/x)) .
x

By equating (3.13) with zero, we get

2x
(3.14) 1− + (exp (−λ/x) − 1)−1 ×
λ n o
−1
× α (a − 1) (1 − exp (−λ/x))−α − 1

− (αb − 1) = 0 .

Then, the mode of the BGIED can be found numerically by solving (3.14).

In Table 1, we present the values of the mean, standard deviation (SD),

mode, median, skewness and kurtosis for different values of a, b, α and λ.

Table 1: The mean, SD, mode, median, skewness and kurtosis

for different values of the parameters.

a b α λ mean SD mode median skewness kurtosis

1 1 4 2 1.35919 1.04298 0.76393 1.08802 0.23016 0.66022

1 1 4 4 2.71838 2.08595 1.52787 2.17604 0.23016 0.66022
1 2 4 4 1.79735 0.88156 1.30871 1.60709 0.16158 0.44815
1 3 4 4 1.50143 0.61555 1.19585 1.38881 0.13109 0.35990
1 2 4 2 0.89867 0.44078 0.65435 0.80354 0.16158 0.44815
2 1 4 2 1.81971 1.24786 1.12748 1.50317 0.22047 0.63372
2 2 2 2 1.98654 1.39808 1.19879 1.62873 0.22323 0.63985
2 2 2 4 3.97308 2.79615 2.39758 3.25747 0.22323 0.63985
3.5 2 1 0.5 1.97910 302.436 0.65618 1.17708 0.32893 1.01650

The results of studying the behaviour of the BGIED are shown in Table 1
and Figure 1. We note that the distribution is unimodal and positively skewed.
For fixed values of a, b and α, the kurtosis values remain constant; therefore, the
mode, median and mean increase with the increase of λ. As we increase the value
of α ≥ 1 and fix the other parameters, the kurtosis value increases and the mean
decreases. It is noted that the distribution has a long right tail for fixed values
of b, α and λ. Moreover, for fixed values of a, α and λ the kurtosis and the mean
76 R.A. Bakoban and Hanaa H. Abu-Zinadah

values decrease as we increase the value of b. Additionally, for different values of

α and λ and fixed values of a and b, the skewness and the kurtosis values remain
stable. Alternately, for fixed values of α and λ, the skewness and the kurtosis
values decrease as we increase a and b. Furthermore, we found that our results
for a = b = 1 are exactly the same as the results in Abouammoh and Alshingiti
(2009).

4. MAXIMUM LIKELIHOOD ESTIMATORS

In this section, we examine estimation by maximum likelihood and inference

for the BGIED. Let X1 , X2 , ..., Xn be a random sample from the BGIED with
pdf and cdf given, respectively, by (2.2) and (2.4). The likelihood function in this
case can be written as (Lawless (2003)):

n
Y
(4.1) L(θ|x) = f (xi ) ,
i=1

where f (·) is given by (2.2) and θ = (a, b, α, λ).

The natural logarithm of the likelihood function (4.1) is given by

n
X
(4.2) ℓ = log L(θ|x) = log f (xi ) .
i=1

For the BGEID, we have have

n  
X −λ
log L = n log (αλ) − n log B(a, b) + (αb − 1) log 1 − exp
xi
i=1
n n n    α 
X
−1
X X −λ
(4.3) −λ xi − 2 log (xi ) + (a − 1) log 1 − 1 − exp .
xi
i=1 i=1 i=1

Assuming that the parameters θ = (a, b, α, λ), are unknown, the likelihood equa-
tions are given for θ

∂ log L 1 ∂f (xi )
lj = = =0, j = 1, 2, 3, 4 .
∂θj f (xi ) ∂θj

From (2.2), we have

(4.4)  
n    "  −α #−1 
∂ log L n X −λ −λ
= + log 1− exp b − (a −1) 1− exp −1 ,
∂α α xi  xi 
i=1
The beta Generalized Inverted Exponential Distribution... 77

n    −1 X n
∂ log L n X
−1 λ
= + (αb − 1) xi exp −1 − x−1
i
∂λ λ xi
i=1 i=1
n
"  −α #−1    −1
X
−1 −λ λ
(4.5) −α(a − 1) xi 1 − exp −1 exp −1 ,
xi xi
i=1

n    α 
∂ log L −n X −λ
= φ1 + log 1 − 1 − exp ,
∂a B(a, b) xi
i=1

∂B(a, b) Γ(b) [Γ(a + b)Γ′ (a) − Γ(a)∂Γ(a + b)/∂a]

φ1 = =
∂a [Γ(a + b)]2
= B(a, b) [ψ(a) − ψ(a + b)] ,
1 ∂Γ(z) Γ′ (z)
where ψ(z) = Γ(z) ∂z = Γ(z) is called the Psi function (Abramowitz and Stegun
(1972)). Then,
n    α 
∂ log L X −λ
(4.6) = −n [ψ(a) − ψ(a + b)] + log 1 − 1 − exp ,
∂a xi
i=1
n   
∂ log L −n X −λ
= φ2 + α log 1 − exp ,
∂b B(a, b) xi
i=1

∂B(a, b) Γ(a) [Γ(a + b)Γ′ (b) − Γ(b)∂Γ(a + b)/∂b]

φ2 = =
∂b [Γ(a + b)]2
= B(a, b) [ψ(b) − ψ(a + b)] ,

n   
∂ log L X −λ
(4.7) = −n [ψ(b) − ψ(a + b)] + α log 1 − exp .
∂b xi
i=1

The solution of the four nonlinear likelihood equations via (4.4), (4.5), (4.6)
and (4.7) yields the maximum likelihood estimates (MLEs) θ̂ = (â, b̂, α̂, λ̂) of θ =
(a, b, α, λ). These equations are in implicit form, so they may be solved using
numerical iteration, such as the Newton–Raphson method via Mathematica 9.0.

5. ASYMPTOTIC VARIANCES AND COVARIANCES OF ESTI-

MATES

The asymptotic variances of maximum likelihood estimates are given

 2by the 
∂ ln L
elements of the inverse of the Fisher information matrix Iij (θ) = E − ∂θi ∂θj
.
Unfortunately, the exact mathematical expressions for the above expectation are
78 R.A. Bakoban and Hanaa H. Abu-Zinadah

very difficult to obtain. Therefore, the observed Fisher information matrix is

∂ 2 ln L
given by Iij = − ∂θ i ∂θj
which is obtained by dropping the expectation on operation
E (Cohen (1965)). The approximate (observed) asymptotic variance-covariance
matrix F for the maximum likelihood estimates of the BGIED can be written as
follows
F = [Iij (θ)] , i, j = 1, 2, 3, 4 and θ = (a, b, α, λ) .
The second partial derivatives of the maximum likelihood function for the BGIED
are given as the following
n    −1 "   −1 #
∂ 2 log L

−n X
−2 λ λ
2
= 2 − (αb − 1) xi exp −1 1 + exp −1
∂λ λ xi xi
i=1
n
"  −α #−1    −1
X
−2 −λ λ
− α(a − 1) xi 1− exp −1 exp −1
xi xi
i=1
  
   −1 "  −α #−1 
 λ λ
× −1 + exp −1 (α − 1) + α 1− exp − −1  ,
 xi xi 

n   2 "  −α #−1

∂ 2 log L

−n X −λ −λ
2
= 2 − (a −1) log 1− exp 1− exp −1
∂α α xi xi
i=1
 
"  −α #−1 
 −λ
× 1+ 1 − exp −1 ,
 xi 

∂ 2 log L  
2
= −n ψ ′ (a) − ψ ′ (a + b) ,
∂a
where
∂ψ(z)
ψ ′ (z) = ,
∂z
∂ 2 log L  ′ ′

= −n ψ (b) − ψ (a + b) ,
∂b2
∂ 2 log L
= n ψ ′ (a + b) ,
∂a ∂b
 #−1
n
"
∂ 2 log L
  −1 
−λ −α
  
X
−1 λ
= xi exp −1 b − (a − 1) 1− exp −1
∂λ ∂α xi  xi
i=1
    
α log 1 − exp −λ
xi

× 1 + h   α i  ,
1 − 1 − exp −λxi


n
" #−1 
∂ 2 log L −λ −α
    −1
X λ
= −α x−1
i 1 − exp −1 exp −1 ,
∂λ ∂a xi xi
i=1
The beta Generalized Inverted Exponential Distribution... 79

n
∂ 2 log L
   −1
X
−1 λ
= α xi exp −1 ,
∂λ ∂b xi
i=1

n  " #−1
∂ 2 log L −λ −α
   
X −λ
= − log 1 − exp 1 − exp −1 ,
∂α ∂a xi xi
i=1

n
∂ 2 log L
  
X −λ
= log 1 − exp .
∂α ∂b xi
i=1

Consequently, the maximum likelihood estimators of a, b, α and λ and have an

asymptotic variance-covariance matrix defined by inverting the Fisher informa-
tion matrix F and by substituting â for a, b̂ for b, α̂ for α and λ̂ for λ.

6. INTERVAL ESTIMATES

If Lθ = Lθ (y1 , ..., yn ) and Uθ = Uθ (y1 , ..., yn ) are functions of the sample

data y1 , ..., yn then a confidence interval for a population parameter θ is given by
 
P Lθ ≤ θ ≤ Uθ = γ ,

where Lθ and Uθ are the lower and upper confidence limits that enclose θ with
probability γ. The interval [Lθ , Uθ ] is called a 100γ % confidence interval for θ.

For large sample sizes (Bain and Engelhardt (1992)), the maximum like-
lihood estimates, under appropriate regularity conditions, are consistent and
asymptotically normally distributed. Therefore, the approximate 100γ % con-
fidence limits for the maximum likelihood estimate θ̂ of a population parameter θ
can be constructed, such that

 
θ̂ − θ
(6.1) P −z ≤ ≤z = γ,
σ(θ̂)

where z is the 100(1+γ)

2 standard normal percentile. Therefore, the approximate
100γ % confidence limits for a population parameter θ can be obtained such that
 
(6.2) P θ̂ − z σ(θ̂) ≤ θ ≤ θ̂ + z σ(θ̂) = γ .

Then, the approximate confidence limits for a, b, α and λ will be constructed

using (6.2) with a confidence level of 90%.
80 R.A. Bakoban and Hanaa H. Abu-Zinadah

7. SIMULATION STUDIES

Simulation studies have been performed using Mathematica 9.0 to illustrate

the theoretical results of the estimation problem. The performance of the result-
ing estimators of the parameters has been considered in terms of their absolute
relative bias (ARBias) and mean square error (MSE), where

  θ̂ − θ    2
ARBias θ̂ = and MSE θ̂ = E θ̂ − θ .
θ

Furthermore, the asymptotic variance, covariance matrix and confidence intervals

of the parameters are obtained. The algorithm for the simulation procedure is
described below:

Step 1. 1000 random samples of sizes n = 10(10)50, 100, 200 and 300 were
generated from the BGIED. The true parameter values are se-
lected as (a = 1, b = 2, α = 4, λ = 2).
Step 2. For each sample, the parameters of the distribution are estimated
under the complete sample.
Step 3. The Newton–Raphson method is used for solving the four non-
linear likelihoods for α, λ, a and b given in (4.4), (4.5), (4.6) and
(4.7), respectively.
Step 4. The ARBiase and MSE of the estimators for the four parameters
for all sample sizes are tabulated.
Step 5. For large sample sizes n = 100, 200 and 300, the Fisher informa-
tion matrix of the estimators are computed using the equations
presented in Section 5.
Step 6. By inverting the Fisher information matrix that was computed in
Step 5, the asymptotic variances and covariances of the estimators
are found.
Step 7. Based on the values of the asymptotic variances and covariances
matrix that were found in Step 6 and on Eq. (6.2), the approxi-
mate confidence limits at 90% for the parameters are computed.

Simulation results are summarized in Tables 2, 3 and 4. Table 2 gives

the ARBias and MSE of the estimators. The asymptotic variances and covari-
ances matrix of the estimators for complete samples of size n = 100, 200 and 300
and true parameter values (a = 1, b = 2, α = 4, λ = 2) are displayed in Table 3.
The approximate confidence limits at 90% for the parameters are presented in
Table 4.
The beta Generalized Inverted Exponential Distribution... 81

Table 2: The ARBias and MSE of the parameters θ = (a, b, α, λ).

â b̂ α̂ λ̂
n
ARBias MSE ARBias MSE ARBias MSE ARBias MSE

10 2.77997 57.22180 1.23642 39.93170 0.34929 57.85210 0.04445 4.49668

20 1.57487 18.92080 0.75400 18.13730 0.01542 23.75870 0.02126 3.08791
30 1.17703 8.04064 0.59348 10.72860 0.13843 12.21520 0.03641 2.48411
40 1.01435 7.53433 0.32958 4.15377 0.25344 7.47854 0.12458 1.52254
50 0.81716 5.63524 0.35502 5.58289 0.23793 8.44674 0.08708 1.65372
100 0.34507 2.01241 0.13037 1.57136 0.37237 5.26341 0.15154 0.73170
200 0.09587 0.78450 0.04185 0.88988 0.45129 4.69708 0.18296 0.46773
300 0.05260 0.56916 0.04620 0.88359 0.46429 4.67339 0.18716 0.31603

Table 3: Asymptotic variances and covariances of estimates

for complete samples.

n Parameters â b̂ α̂ λ̂

a 0.00860 0.00389 − 0.00021 − 0.00359

100 b 0.00389 0.05006 − 0.00513 0.00867
α − 0.00021 − 0.00513 0.04464 0.01414
λ − 0.00359 0.00867 0.01414 0.01308
a 0.00673 0.00327 − 0.00018 − 0.00411
200 b 0.00327 0.03238 − 0.00883 0.00346
α − 0.00018 − 0.00883 0.02928 0.00772
λ − 0.00411 0.00346 0.00772 0.00951
a 0.00630 0.00320 − 0.00075 − 0.00445
300 b 0.00320 0.02563 − 0.00722 0.00197
α − 0.00075 − 0.00722 0.02245 0.00664
λ − 0.00445 0.00197 0.00664 0.00888

Table 4: Confidence bounds of the estimates at a confidence level of 0.90.

n Parameters Estimated mean Lower bound Upper bound Width

a 1.34507 1.19294 1.49719 0.30424

100 b 2.26075 1.89380 2.62770 0.73390
α 2.51052 2.16399 2.85704 0.69304
λ 1.69693 1.50931 1.88455 0.375241
a 1.09587 0.96132 1.23041 0.26909
200 b 2.08370 1.78856 2.37883 0.59027
α 2.19482 1.91416 2.47549 0.56132
λ 1.63408 1.47409 1.79408 0.31999
a 1.05260 0.92241 1.18279 0.26038
300 b 2.09239 1.82981 2.35498 0.52517
α 2.14283 1.89709 2.38857 0.49147
λ 1.62569 1.47108 1.78030 0.30922
82 R.A. Bakoban and Hanaa H. Abu-Zinadah

From these tables, the following observations can be made on the perfor-
mance of the parameter estimation of the BGIED:

1. As the sample size increases, the MSEs of the estimated parameters

decrease. This indicates that the maximum likelihood estimates provide
asymptotically normally distributed and consistent estimators for the
parameters (see Table 2).
2. Although the estimators of a and b are consistent according to the
ARBias, it is noted that the estimators of α and λ are not consistent.
Table 2 shows that, for large sample sizes (n = 100, 200 and 300), the
ARBiases are increased, which indicates that the estimates of α and λ
are not consistent.
3. The asymptotic variances of the estimators decrease when the sample
size is increasing (see Table 3).
4. The interval of the estimators decreases when the sample size is in-
creasing (see Table 4).
5. The interval estimations of all parameters were reasonable except the
interval estimate of α. The estimated intervals at a confidence level of
0.90 for n = 100, 200 and 300 did not cover the real value of α.

We conclude from the previous points that the MLE of the parameters is
a good estimator for a, b and λ.

8. APPLICATIONS

In this section, two sets of data are presented to demonstrate the utility of
using the BGIED. These two sets were investigated by Abouammoh and Alshin-
giti (2009). The GIED was fitted to both of these sets.

8.1. The first data-set

The following data-set is presented in Lawless (2003). The data resulted

from a test on the endurance of deep groove ball bearings. The data are as
follows:

17.88, 28.92, 33.0, 41.52, 42.12, 45.60, 48.40, 51.84, 51.96, 54.12, 55.56, 67.80,
68.64, 68.64, 68.88, 84.12, 93.12, 98.64, 105.12, 105.84, 127.92, 128.04, 173.4 .
The beta Generalized Inverted Exponential Distribution... 83

Descriptive statistics of these data are tabulated in Table 5.

Table 5: Descriptive statistics for the ball bearing data.

Measure Value Measure Value

n 23 Minimum 17.880
Maximum 173.400 Mean 72.22
Q1 45.600 Q3 98.640
Median 67.800 Mean deviation 29.429
Variance 1405.580 SD 37.491
Skewness 0.941 Kurtosis 3.486

We apply the Kolmogorov–Smirnov (K-S) statistic to verify which distri-

bution better fits these data. The K-S test statistic is described in detail in
D’Agostino and Stephens (1986). In general, the smaller the value of K-S is, the
better the fit to the data is. All graphs and computations presented to analyse
the data were carried out by Mathematica 9.0. The model selection was carried
out using the AIC (Akaike information criterion), the BIC (Bayesian information
criterion) and the CAIC (consistent Akaike information criterion):

AIC = −2 l(θ̂) + 2p ,

BIC = −2 l(θ̂) + p log n

and
2pn
CAIC = −2 l(θ̂) + ,
n−p−1

where l(θ̂) denotes the log likelihood function evaluated at the maximum likeli-
hood estimates, p is the number of parameters and n is the sample size. Table 6
lists the values of the K-S statistic and of −2l(θ̂). The K-S goodness-of-fit test
of the BGIED as well as the GIED are the best among all models; accordingly,
the BGIED model can be used to analyse the ball bearing data. Table 7 provides
the MLEs with corresponding standard errors (SEs) of the model parameters.

Table 6: Goodness-of-fit measures and K-S statistics

for the ball bearing data.

Model BGIED GIED IED BIED

K-S statistics 0.097 0.091 0.306 0.104
−2 l(θ̂) 227.723 227.098 243.452 228.288
84 R.A. Bakoban and Hanaa H. Abu-Zinadah

Table 7: MLEs of the model parameters, the corresponding SEs and the
statistics of the AIC, BIC and CAIC for the ball bearing data.

Estimates Statistics
Model Method
â b̂ α̂ λ̂ AIC BIC CAIC
MLE 20.615 9.427 0.532 7.161
BGIED 235.723 240.265 237.946
SE 0.125 0.279 8.052 0.262
MLE 5.307 129.996
GIED 231.098 233.369 231.698
SE 0.188 0.015
MLE 55.055
IED 245.452 246.587 245.642
SE 0.018
MLE 15.858 3.692 11.858
BIED 234.288 237.695 235.552
SE 0.112 0.508 0.161

8.2. The second data-set

The data that is studied in this section was provided by Ed Fuller of the
NICT Ceramics Division in December 1993. It contains polished window strength
data. Fuller et al. (1994) described the use of this set to predict the lifetime for
a glass airplane window. The data are as follows:

18.83, 20.8, 21.657, 23.03, 23.23, 24.05, 24.321, 25.5, 25.52, 25.8, 26.96,
26.77, 26.78, 27.05, 27.67, 29.9, 31.11, 33.2, 33.73, 33.76, 33.89, 34.76,
35.75, 35.91, 36.98, 37.08, 37.09, 39.58, 44.045, 45.29, 45.381 .

Descriptive statistics of the window strength data are tabulated in Table 8.

Table 8: Descriptive statistics for the window strength data.

Measure Value Measure Value

n 31 Minimum 18.830
Maximum 45.381 Mean 30.820
Q1 25.500 Q3 35.910
Median 29.900 Mean deviation 6.145
Variance 52.539 SD 7.248
Skewness 0.403 Kurtosis 2.290

The K-S goodness-of-fit test of the BGIED as well as the BIED are the best
among all models; therefore, the BGIED model can be used to study the window
strength data. Table 10 presents the MLEs with corresponding SEs of the model
parameters.
The beta Generalized Inverted Exponential Distribution... 85

Table 9: Goodness-of-fit measures and K-S statistics

for the window strength data.

Model BGIED GIED IED BIED

K-S statistics 0.133 0.137 0.474 0.130
−2 l(θ̂) 208.207 208.454 274.523 208.105

Table 10: MLEs of the model parameters, the corresponding SEs and the statistics
of the AIC, BIC and CAIC for the window strength data.

Estimates Statistics
Model Method
â b̂ α̂ λ̂ AIC BIC CAIC
MLE 27.850 7.354 1.978 17.601
BGIED 216.207 221.943 217.745
SE 0.088 0.341 2.401 0.247
MLE 90.855 148.412
GIED 212.454 215.322 212.883
SE 0.011 0.029
MLE 29.215
IED 276.523 277.957 276.661
SE 0.034
MLE 14.506 20.169 26.053
BIED 214.105 218.407 214.994
SE 0.205 0.146 0.166

Finally, we conclude the following from studying the AIC, BIC and CAIC
statistics of the two previous data-sets.

Figure 4: The empirical distribution and estimated cdf

of the models for the ball bearing data.
86 R.A. Bakoban and Hanaa H. Abu-Zinadah

It is noted that the GIED has a smaller value compared with the values of
other models for the two data-sets. The BIED and BGIED follow next. That
indicates that the GIED seems to be a very competitive model for these data.
Because the values of the AIC, BIC and CAIC are approximately equivalent for
the GIED, BIED and BGIED, the BGIED can thus be a good alternative model
for these data, as can the GIED. Alternately, the IED presents the worst fit for
the second dataset. Figure 4 shows the empirical distribution and estimated cdf
of the models for the ball bearing data. Figure 5 shows the empirical distribution
and estimated cdf of the models for the window strength data.

Figure 5: The empirical distribution and estimated cdf

of the models for the window strength data.

9. CONCLUDING REMARKS

In this study, the four-parameter beta generalized inverted exponential dis-

tribution (BGIED) is proposed. BGIED generalizes the generalized inverted ex-
ponential distribution discussed by Abouammoh and Alshingiti (2009). Addition-
ally, the BGIED represents a generalization of the inverted exponential distribu-
tion (IED). IED has been considered by Keller and Kamath (1982) and Duran and
Lewis (1989). Statistical properties of the BGIED are studied. Maximum likeli-
hood estimators of the BGIED parameters are obtained. The information matrix
and the asymptotic confidence bounds of the parameters are derived. Monte
Carlo simulation studies are conducted under different sample sizes to study the
theoretical performance of the MLE of the parameters. Two real data-sets are
analysed, and the BGIED has provided a good fit for the data-sets.
The beta Generalized Inverted Exponential Distribution... 87

ACKNOWLEDGMENTS

The authors would like to thank the referees for their useful comments and
suggestions. Additionally, the authors acknowledge with thanks the Deanship
of Scientific Research, King Abdulaziz University, Jeddah, for the technical and
financial support.

REFERENCES

[1] Abouammoh, A. M. and Alshingiti, A. M. (2009). Reliability estimation of

generalized inverted exponential distribution, Journal of Statistical Computation
and Simulation, 79(11), 1301–1315.
[2] Abramowitz, M. and Stegun, I. A. (1972). Handbook of Mathematical Func-
tions with Formulas, Graphs and Mathematical Tables, 2nd Ed., Dover Publica-
tions, New York.
[3] Baharith, L. A.; Mousa, S. A.; Atallah, M. A. and Elgayar, S. H. (2014).
The beta generalized inverse Weibull distribution, British Journal of Mathematics
and Computer Science, 4(2), 252–270.
[4] Bain, L. J. and Engelhardt, M. (1992). Introduction to Probability and Math-
ematical Statistics, 2nd Ed., Duxbury Press, Boston.
[5] Barreto-Souza, W.; Santos, A. H. S. and Cordeiro, G. M. (2010). The
beta generalized exponential distribution, Journal of Statistical Computation and
Simulation, 80, 159–172.
[6] Cohen, A. C. (1965). Maximum likelihood estimation in the Weibull distribution
based on complete and censored samples, Technometrics, 5, 579–588.
[7] Cordeiro, G. M. and Lemonte, A. J. (2011). The beta Laplace distribution,
Statistics and Probability Letters, 81, 973–982.
[8] D’Agostino, R. B. and Stephens, M. A. (1986). Goodness-of-Fit Techniques,
Marcel Dekker, New York.
[9] Duran, B. S. and Lewis, T. O. (1989). Inverted gamma as life distribution,
Microelectron. Reliab., 29, 619–626.
[10] Eugene, N.; Lee, C. and Famoye, F. (2002). Beta-normal distribution and
its applications, Commun. Statist. – Theory and Methods, 31, 497–512.
[11] Fuller Jr, E. R.; Frieman, S. W.; Quinn, J. B.; Quinn, G. D. and Carter,
W. C. (1994). Fracture mechanics approach to the design of glass aircraft win-
dows: A case study, SPIE Proc., 2286, 419–430.
[12] Gupta, A. K. and Nadarajah, S. (2004). On the moments of the beta normal
distribution, Commun. Statist. – Theory and Methods, 33, 1–13.
[13] Keller, A. Z. and Kamath, A. R. (1982). Reliability analysis of CNC Machine
Tools, Reliab. Eng., 3, 449–473.
88 R.A. Bakoban and Hanaa H. Abu-Zinadah

[14] Kenney, J. F. and Keeping, E. S. (1962). Mathematics of Statistics, 3rd Ed.,

Part 1, Princeton, New Jersey.
[15] Lawless, J. F. (2003). Statistical Models and Methods for Lifetime Data, 2nd
Ed., Wiley, New York.
[16] Leão, J.; Saulo, H.; Bourguignon, M.; Cintra, R.; Rego, L. and
Cordeiro, G. (2013). On some properties of the beta inverse Rayleigh dis-
tribution, Chilean Journal of Statistics, 4(2), 111–131.
[17] Lee, C.; Famoye, F. and Olumolade, O. (2007). The beta-Weibull distribu-
tion: Some properties and applications censored data, J. Modern Applied Statis-
tical Method, 6(1), 173–186.
[18] Mahmoudi, E. (2011). The beta generalized Pareto distribution with application
to lifetime data, Mathematics and Computers in Simulation, 81, 2414–2430.
[19] Moors, J. J. (1988). A quantile alternative for kurtosis, Journal of the Royal
Statistical Society D, 37, 25–32.
[20] Nadarajah, S. and Kotz, S. (2005). The beta exponential distribution, Reli-
ability Engineering and System Safety, 91, 689–697.
[21] Nadarajah, S. and Kotz, S. (2004). The beta Gumbel distribution, Math.
Probab. Eng., 10, 323–332.
[22] Nassar, M. M. and Nada N. K. (2011). The beta generalized Pareto distribu-
tion, Journal of Statistics: Advances in Theory and Applications, 6(1/2), 1–17.
[23] Paranaiba, P. F.; Ortega, E. M. M.; Cordeiro, G. M. and Pescim, R. P.
(2011). The beta Burr XII distribution with application to lifetime data, Com-
putational Statistics and Data Analysis, 55(2), 1118–1136.
[24] Zea, L. M.; Silva, R. B.; Bourguignon, M.; Santos, A. M. and Cordeiro,
G.M. (2012). The Beta exponentiated Pareto distribution with application to
bladder cancer susceptibility, International Journal of Statistics and Probability,
1(2), 8–19.