arXiv:1603.03001v1 [math.ST] 9 Mar 2016
An Extension of the Generalized Linear Failure Rate
Distribution
M. R. Kazemi∗, A. A. Jafari† , S. Tahmasebi‡
∗
Department of Statistics, Fasa University, Fasa, Iran
Department of Statistics, Yazd University, Yazd, Iran
‡
Department of Statistics, Persian Gulf University, Bushehr, Iran
†
Abstract
In this paper, we introduce a new extension of the generalized linear failure rate distributions. It includes some well-known lifetime distributions such as extension of generalized
exponential and generalized linear failure rate distributions as special sub-models. In addition, it can have a constant, decreasing, increasing, upside-down bathtub (unimodal), and
bathtub-shaped hazard rate function depending on its parameters. We provide some of its
statistical properties such as moments, quantiles, skewness, kurtosis, hazard rate function,
and reversible hazard rate function. The maximum likelihood estimation of the parameters
is also discussed. At the end, a real data set is given to illustrate the usefulness of this new
distribution in analyzing lifetime data.
Keywords: Generalized exponential distribution; Hazard function; Maximum likelihood
estimation.
1
Introduction
The generalized linear failure rate (GLFR) distribution is defined by Sarhan and Kundu (2009)
and contains various well-known distributions: the generalized exponential (GE) distribution
introduced by Gupta and Kundu (1999), the generalized Rayleigh distribution introduced by
Surles and Padgett (2001, 2005), the exponential, Rayleigh, and linear failure rate (LFR) distributions are its special cases. The GLFR distribution has decreasing or unimodal probability
density function (pdf) and its hazard rate function (hrf) can have increasing, decreasing, and
bathtub-shaped. Unfortunately, the GLFR distribution cannot have unimodal hrf.
∗
Corresponding: kazemi@fasau.ac.ir
Recently, many studies have been done on GLFR distribution, and some authors have
extended it: the generalized linear exponential (Mahmoud and Alam, 2010), beta-linear failure rate (Jafari and Mahmoudi, 2015), Kumaraswamy-GLFR (Elbatal, 2013), modified-GLFR
(Jamkhaneh, 2014), McDonald-GLFR (Elbatal et al., 2014), Poisson-GLFR (Cordeiro et al.,
2015), GLFR-geometric (Nadarajah et al., 2014), and GLFR-power series (Alamatsaz and Shams,
2014) are some univariate extension of GLFR distribution.
Kundu and Gupta (2011) proposed an extension of GE distribution that is a very flexible
family of distribution. It is positively skewed, and has increasing, decreasing, unimodal and
bathtub shaped hrfs. It is included GE, exponential, generalized Pareto (Johnson et al., 1995),
and Pareto distributions.
Cordeiro et al. (2012) introduced a five-parameter called the McDonald extended exponential distribution as a generalization of extended generalized exponential (EGE). In this paper,
we introduce a new four-parameter distribution that contains the EGE distribution as its special case. In addition, this class of distribution extends the three-parameter GLFR distribution.
Therefore, it is called extended generalized linear failure rate (EGLFR) distribution. The hrf
of this new distribution is increasing, decreasing, bathtub, and unimodal. In addition, we
will show that the new distribution has been fit better than the EGE distribution and other
competing distributions to analyzing the lifetime data.
The paper is organized as follows. In Section 2, we introduce a new class of distributions.
Some statistical properties such as moments, quantiles, hrf, and reversible hazard rate are
provided in Section 3. The maximum likelihood estimation (MLE) of the parameters is obtained
in Section 4. An application of the EGLFR distribution using a real data set is presented in
Section 5.
2
A new class
In this section, we introduce a new class of distributions as extension of GLFR distribution.
Also, some properties of the pdf and hrf of this distribution are given here.
Theorem 2.1. For given α > 0, β ∈ R, a ≥ 0, and b ≥ 0 (with a + b > 0), consider the
function
α
1 − 1 − β(ax + b x2 ) 1/β
if
α2
F (x) =
b 2
−(ax+
x
)
1−e
2
if
β 6= 0
β = 0.
i. If β ≤ 0, then F (x) is a cumulative distribution function (cdf ) on (0, ∞).
(2.1)
ii. If β > 0 and b 6= 0, then F (x) is a cdf on (0, ψ) where ψ =
1
b
q
a2 +
2b
β
1
aβ .
iii. If β > 0 and b = 0, then F (x) is a cdf on (0, ψ) where ψ =
− ab .
Proof. When β = 0, proof is obvious. We consider β 6= 0. Without loss of generality, we
1
consider α = 1. Therefore, limx→0 F (x) = 0 and F ′ (x) = (a + bx)(1 − β(ax + 2b x2 )) β
−1
.
i. When β < 0, limx→∞ F (x) = 1 and F ′ (x) > 0 for 0 < x < ∞.
ii. When β > 0, limx→ψ F (x) = 1 and F ′ (x) > 0 for 0 < x < ψ.
iii. This part is similar to part ii.
Therefore, proof is completed.
When the function F in (2.1) is a cdf, it is said EGLFR distribution with parameters α, β,
a and b, and will be denoted by EGLFR(α, β, a, b). The pdf of this new class of distributions is
α−1
1
−1
α (a + bx) (1 − βz) β
1 − (1 − βz)1/β
if β 6= 0
f (x; α, β, a, b) =
(2.2)
α (a + bx) e−z (1 − e−z )α−1
if β = 0,
where z = ax + 2b x2 . The plots for pdf of EGLFR distribution are given in Figure 1, for some
different values of parameters.
0.20
α=0.8, a=0.1, b=0.1
β=−5, a=0.1, b=0.1
0.15
0.3
f(x)
0.4
0.5
β=−10
β=0.8
β=1.0
β=2.0
α=0.5
α=0.95
α=1
α=2
0.0
0.00
0.1
0.05
0.2
f(x)
Density
0.10
0.6
Density
0
1
2
x
3
4
0
2
4
6
8
x
Figure 1: The plot of pdf for some different values of parameters.
If β = 0, then this distribution reduces to the GLFR distribution which is introduced by
Sarhan and Kundu (2009). The GLFR includes the GE (if b = 0), exponential (if b = 0, α = 1),
LFR (if α = 1), generalized Rayleigh (if a = 0) and Rayleigh (if a = 0, α = 1) distributions as
its special sub-models.
If b = 0 then EGLFR distribution reduces to the EGE distribution introduced by Kundu and Gupta
(2011). The EGE distribution includes GE (if β = 0), exponential (if β = 0, α = 1), generalized
Pareto (if α = 1), and Pareto distributions (if α = 1, β < 0).
If a = 0, we have an extension of two-parameter Burr X distribution which is introduced by Surles and Padgett (2005) and also is known as generalized Rayleigh distribution
(Kundu and Raqab, 2005). Therefore, it is called extended generalized Rayleigh (EGR) distribution.
Proposition 1. If X has a EGLFR distribution with the pdf in (2.2), then G (X) = aX + 2b X 2
has a EGE distribution with parameters α, β, and 1 or equivalently, Y ∗ = X +
b
2
2a X
has a
EGE distribution with parameters α, β, and a.
Proposition 2. Suppose that X has the following cdf
(
1 − (1 − βz)1/β if
F1 (x) =
1 − e−z
if
β 6= 0
β = 0,
where z = ax + 2b x2 . We say that X has extended linear failure rate (ELFR) distribution. The
maximum of a random sample with size n from the ELFR distribution has EGLFR(n, β, a, b).
Therefore, the EGLFR(n, β, a, b) provides the cdf of a parallel system when each component has
the ELFR distribution.
Theorem 2.2. Let f (x) be the pdf of the EGLFR distribution. The limiting behavior of f (x)
for different values of its parameters is given bellow:
if α > 1
0
∞
lim f (x) =
and
lim f (x) =
a
if α = 1
+
−
0
x→0
x→c
∞
if α < 1
if β ≥ 1
if β < 1,
where c = ψ for β > 0, and c = ∞ for β < 0.
Proof. The proof is obvious.
Theorem 2.3. Let f (x) be the pdf of the EGLFR distribution. Then,
i. the mode of f (x) is obtained from the solution of the following nonlinear equation when
α ≥ 1, β < 1 and β 6= 0:
1 −1
(a + bx) 1 − β(ax + 2b x2 ) β
(β − 1)(a + bx)
b
+ (α − 1)
+
= 0,
1/β
a + bx 1 − β(ax + 2b x2 )
1 − 1 − β(ax + b x2 )
2
ii. the mode of f (x) is obtained from the solution of the following nonlinear equation when
α ≥ 1 and β = 0:
b
a + bx
= 0,
− (a + bx) + (α − 1)
b 2
a + bx
eax+ 2 x − 1
iii. the mode of f (x) is 0 when α < 1 and β < 1,
iv. f (x) has two modes at 0 and ψ when α < 1 and β ≥ 1.
Proof. Using Theorem 2.2, the proof is obvious.
Theorem 2.4. Let F be the pdf of the EGLFR distribution. Then, F is a heavy-tailed distribution when α ≥ 1 and β < 0.
Proof. Consider z = ax + 2b x2 and s = − β1 . Therefore,
z −s
F̄ (x) = 1 − F (x) = 1 − 1 − 1 +
s
Without loss of generality take α = 1. Clearly, we have F̄ (x) ∼
lim
x→∞
s s
z
eλx = ∞
for ∀λ > 0,
α
.
s s
z ,
as x → ∞. So
and F is a heavy tailed distribution (see Foss et al., 2011, Theorem 2.6).
The hrf of the EGLFR distribution is
1 −1
α−1
β
(1−(1−βz)1/β )
α(a+bx)(1−βz)
1/β α
1−(1−(1−βz)
)
h (x) =
−z 1−e−z α−1
α(a+bx)e
)
(
1−(1−e−z )α
if
β 6= 0
if
β = 0,
(2.3)
where z = ax + 2b x2 . For some cases of parameters, the plot for hazard of EGLFR distribution
are given in Figure 2.
Hazard
0.15
0.3
h(x)
0.4
0.20
0.5
α=0.5
α=0.95
α=1
α=2
0.10
α=0.8, a=0.1, b=0.1
0.05
0.00
0.1
0.2
β=−10
β=−0.5
β=0.1
β=0.5
0.0
h(x)
β=−5, a=0.1, b=0.1
0.25
0.6
0.30
Hazard
0
1
2
3
x
4
5
0
2
4
6
x
Figure 2: The plot of hrf for some different values of parameters.
8
10
Theorem 2.5. Let h(x) be the hrf of the EGLFR distribution. The limiting behavior of h(x)
for different values of its parameters is given bellow:
0
lim h (x) =
a
x→0+
∞
if α > 1
if α = 1
if α < 1
and
lim h (x) =
x→c−
∞
0
if β > 0
if β < 0,
where c = ψ for β > 0, and c = ∞ for β < 0.
Proof. The proof is obvious.
Theorem 2.6. Let h(x) be the hrf of the EGLFR distribution. Then h(x) is increasing function
for β ≥ 1 and α ≥ 1,
Proof. For b = 0, we have the EGE distribution and proof is given by Kundu and Gupta (2011).
√
2
2
Here, we consider b > 0. Let z = ax + 2b x2 = 2b x + ab − a2b . Then, x = 1b 2bz + a2 − ab , and
for β 6= 0
α−1
√
1
−1
α 2bz + a2 (1 − βz) β
1 − (1 − βz)1/β
α
.
h (z) =
1 − 1 − (1 − βz)1/β
With u (z) = log (h (z)), we have
∂
u (z) =
∂z
1
−1
b
(1 − βz) β
1
+ (α − 1)
+ (β − 1)
1
2
2bz + a
1 − βz
1 − (1 − βz) β
α−1
1/β
1
−
(1
−
βz)
(1 − βz)1/β−1
α
α
.
+
β
1 − 1 − (1 − βz)1/β
If β ≥ 1 and α ≥ 1, then the hrf is an increasing function.
The reversible hazard function of EGLFR distribution is
1 −1
α (a + bx) 1 − β(ax + 2b x2 ) β
f (x; α, β, a, b)
=
r (x; α, β, a, b) =
1/β .
F (x; α, β, a, b)
1 − 1 − β(ax + 2b x2 )
It is clear that r (x; α, β, a, b) = αr (x; 1, β, a, b) .
3
Properties of EGLFR distribution
We provide some of statistical properties of introduced distribution such as moments and quantiles. For more properties, reader can see Sarhan and Kundu (2009) and Kundu and Gupta
(2011) for β = 0 and b = 0, respectively.
3.1
Quantiles of EGLFR distribution
The quantile function of EGLFR distribution is
r
h
√ i
1
2 + 2b 1 − (1 − α u)β −
a
b
β
1p
√
a
α
2
Q (u) =
b a − 2blog (1 − u) − b
√
β
1
α
u)
βa 1 − (1 −
√
1
− a log (1 − α u)
a
b
if
β 6= 0, b > 0
if
β = 0, b > 0
if
β 6= 0, b = 0
if
β = 0, b = 0.
The median of the EGLFR distribution can be obtain by letting u = 0.5. We can use
this function for generating data form EGLFR distribution by generating data from a uniform
distribution. For checking, we have generated a random sample from EGLFR(0.8, 2, 0.5, 0.1).
The histogram of this random sample and pdf of EGLFR are plotted in Figure 3 (left). In
addition, the empirical cdf of this sample and cdf of EGLFR are plotted in Figure 3 (right).
Empirical Distribution
0.0
0.0
0.2
0.5
0.4
Fn(x)
Density
0.6
1.0
0.8
1.0
1.5
Histogram
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
Figure 3: The histogram of the simulated data set and the pdf of EGLFR distribution (left);
the empirical cdf of the simulated data set and the cdf of EGLFR distribution (right).
3.2
Skewness and kurtosis of EGLFR distribution
The Bowley skewness (see Kenney and Keeping, 1962) (based on quantiles can be calculated
by
B=
Q
3
4
− 2Q 12 + Q
Q 43 − Q 14
1
4
and the Moors kurtosis (see Moors, 1988) is defined as
Q 87 − Q 85 + Q 38 − Q
M=
Q 86 − Q 28
,
1
8
.
These measures are less sensitive to outliers and they exist even for distributions without
moments. For the standard normal and the classical standard t distributions with 10 degrees
of freedom, the Bowley measure is 0. The Moors measure for these distributions is 1.2331 and
1.27705, respectively. In Figure 4, we plot Bowley and Moors measures (with a = b = 0.1) as
a function of β for fixed values of α.
Moors kurtosis
1.0
10
Bowley skewness
8
α=2.0
α=1.0
α=0.5
α=0.2
0
−1.0
2
−0.5
4
M
0.0
B
6
0.5
α=2.0
α=1.0
α=0.5
α=0.2
−10
0
10
20
30
40
−10
−5
β
0
5
10
β
Figure 4: Plot of Bowley measure (left) and Moors measure for some parameters. (right).
3.3
Moments of EGLFR distribution
Using the following theorem, we can calculate the r-th non-central moment, µ(r) , of EGLFR
distribution in a series form.
Theorem 3.1. Let X has a EGLFR(α, β, a, b). Then, for β > 0 and b > 0,
µ
(r)
∞ X
∞ X
m
X
a
bψ
α
(−1)m+n Λ[n,m,k]β m am−k bk ψ j ( +
),
= E(X ) =
2k
j
j+1
r
n=0 m=0 k=0
where j = m + k + r + 1 and Λ[n,m,k] =
α−1
n
n+1
−1
β
m
m
k
.
Proof. Consider z = ax + 2b x2 . If β > 0, then support of X is (0, ψ). So, βz < 1 and
(1 − βz)1/β < 1. By using binomial series expansion, we have
∞
α−1 X
n
1/β
n α−1
(−1)
1 − (1 − βz)
=
(1 − βz) β ,
n
n=0
n+1
∞
X
n+1
−1
(−1)m β
(1 − βz) β −1 =
(βz)m ,
m
m=0
m
b m X m m−k bk k
x .
a
(a + x) =
2
2k
k
k=0
Therefore,
µ
(r)
=
=
=
=
=
=
Z
ψ
0
1
xr α (a + bx) (1 − βz) β
ψ
∞
X
−1
α−1
1 − (1 − βz)1/β
dx
n+1
α−1
(−1)
αx (a + bx)
(1 − βz) β −1 dx
n
0
n=0
n+1
Z ψ
∞ X
∞
X
−1 m
b
n+m α − 1
r
β
αx (a + bx)
(−1)
β (a + x)m dx
m
2
n
0
n=0 m=0
n+1
Z ψ
∞ X
∞ X
m
X
− 1 m m m−k bk m+k
n+m α − 1
r
β
β a
x
dx
αx (a + bx)
(−1)
m
k
2k
n
0
n=0 m=0 k=0
Z ψ
∞ X
∞ X
m
X
bk
αxr (a + bx)
(−1)n+m Λ[n,m,k] β m am−k k xm+k dx
2
0
n=0 m=0 k=0
Z ψ
∞ X
∞ X
m
X
α
m+n
m m−k k
(−1)
Λ[n,m,k]β a
b
xm+k+r (a + bx) dx
2k
0
Z
r
n
n=0 m=0 k=0
∞ X
∞ X
m
X
α
bψ
a
=
),
(−1)m+n Λ[n,m,k]β m am−k bk ψ j ( +
k
2
j
j
+1
n=0 m=0
k=0
and proof is completed.
Theorem 3.2. Let X has a EGLFR(α, β, a, b). Then, for β > 0 and b > 0,
∞ X
∞ X
m m+k
X
X (−1)n+m+i etψ (m+k)! α
Λ
β m am−k bk ψ m+k−i
ti+1 (m+k − i)!
2k [n,m,k]
n=0 m=0 k=0 i=0
X
∞ X
∞ X
m
(−1)n+2m+k a − bt α
b (m+k + 1) ψ
× a+
Λ
β m am−k bk .
−
m+k+1
k [n,m,k]
(m+k + 1 − i)
t
2
n=0 m=0
MX (t) = E(etx ) =
k=0
where Λ[n,m,k] =
α−1
n
n+1
−1
β
m
m
k
.
Proof. Since (see Gradshteyn and Ryzhik, 2007, Section 2.321)
Z
n tx
tx
x e dx = e n!
n
X
i=0
(−1)i
xn−i ,
ti+1 (n − i)!
we have
Z
ψ
0
xm+k (a + bx) etx dx = a
Z
ψ
xm+k etx dx + b
0
= etψ (m+k)!
m+k
X
i=0
Z
ψ
xm+k+1 etx dx
0
(−1)i
b (m+k + 1) ψ
m+k−i
ψ
a
+
ti+1 (m + k − i)!
(m + k + 1 − i)
−
(−1)m+k (a − bt )
.
tm+k+1
Similar to Theorem 3.1, we have
MX (t) =
Z ψ
∞ X
∞ X
m
X
m+k
α
n+m
m m−k k
(−1)
Λ
β
a
b
x
(a + bx)etx dx.
[n,m,k]
k
2
0
n=0 m=0
k=0
and proof is completed.
Corollary 3.1. For β > 0, b > 0 and a = 0, from Theorem 3.1 we have
n+1
r+1
∞ X
∞
X
αb
−1
2 2
n+m α − 1
(r)
β
µ =
(−1)
( ) .
(2m + r + 1)
bβ
n
m
n=0 m=0
Remark 3.1. Using the equation (6) of Kundu and Gupta (2011) and proposition 1 for Y =
aX + 2b X 2 , we have
Γ(α + 1)Γ(kβ + 1)
E (1 − βY )k =
:= gk ,
Γ(α + kβ + 1)
kβ + 1 > 0,
β 6= 0.
When a = 0,
2
E(X 2k ) = (−1)k ( )k
βb
gk − 1 −
k−1
X
r=1
!
k βb r
2r
(−1)
( ) E(X ) ,
r
2
r
k = 2, 3, . . . .
Therefore,
2
E X2 =
(1 − g1 ) ,
bβ
E X4 =
4
b2 β 2
(g2 − 2g1 + 1) .
Theorem 3.3. Let X has a EGLFR(α, β, a, b). Then, for β < 0 and α ≥ 1, the k-th moment
of X does not exist when k ≥ − β2 .
Proof. Consider α = 1. Then, the survival function of X is
z −s
,
F̄ (x) = 1 +
s
where z = ax + 2b x2 . Since X is a positive random variable, so
Z ∞
Z
h i Z ∞
1/k
k
1/k
dx =
P X>x
E X =
dx =
F x
0
0
0
∞
dx
1+
ax1/k + 2b x2/k
s
s ,
R∞
b 2/k
x , as x → ∞. The integral 0 x−2s/k dx
where s = − β1 . clearly, (s + ax1/k + 2b x2/k )/s ∼ 2s
diverges when k ≥ 2s, and therefore, E X k does not exists.
−s α
Consider Fα (x) = 1 − 1 + zs
. It can be easily seen that if α1 < α2 then Fα1 (x) >
Fα2 (x) and for the survival function F̄ (x) we have F̄α1 (x) < F̄α2 (x). So, (at least) the
divergence condition k ≥ 2s, satisfies for ∀α ≥ 1.
Table 1: Some moments of EGLFR distribution for β < 0.
β
α k -1.90 -1.00 -0.75 -0.50 -0.25 -0.10
1 1 20.195 2.221 1.829 1.571 1.388 1.303
2 —
— 8.000 4.000 2.667 2.222
3 —
—
— 18.849 6.664 4.598
4 —
—
—
—
21.333 11.111
5 —
—
—
—
88.857 30.656
2 1 38.886 3.332 2.617 2.159 1.846 1.702
2 —
— 14.400 6.667 4.190 3.391
3 —
—
— 35.342 11.610 7.732
4 —
—
—
—
39.619 19.883
5 —
—
—
— 171.467 57.127
Based on Remark 3.1, if k-th moment of EGLFR distribution exists, we have an expression
for it when k is even. But, we did not find an expression for other cases. In Table 1, we
calculated some moments of EGLFR distribution for β < 0 when they exist with considering
a = 0 and b = 1.
3.4
Entropy
The entropy of random variable is defined in terms of its probability distribution and can
be shown to be a good measure of randomness or uncertainty. The Shannon’s entropy of a
continuous random variable Y with pdf f (y) is defined by Shannon (1948) as
HSh (f ) = −Ef [log f (Y )] = −
Z
∞
f (y) log f (y)dy.
0
Since the pdf of EGLFR distribution can be written as
f (x) = αf1 (x)[F1 (x)]α−1
(3.1)
where f1 and F1 are pdf and cdf of ELFR distribution, respectively, given in Proposition 2, the
Shannon entropy for EGLFR distribution can be expressed in the form
α−1
− Ef [ln f1 (X)]
α
α−1
= − ln(α) +
− EW [ln f1 (F1−1 (W ))],
α
HSh (f ) = = − ln(α) +
(3.2)
where W has the beta distribution with parameters α and 1. The last term in (3.2) follows
immediately from the conditions in Lemma 1 of Zografos and Balakrishnan (2009). The Rényi
entropy is defined by
Z ∞
1
Hρ (f ) =
[f (y)]ρ dy),
log(
1−ρ
−∞
where ρ > 0 and ρ 6= 1. The Shannon entropy is derived from limρ→1 Hρ (f ). An explicit
expression of Rényi entropy for EGLFR distribution is obtained as
Hρ (f ) =
1
1
−ρ
ln(α) −
ln B(ρ(α − 1) + 1, 1) −
ln ET [f1ρ−1 (F1−1 (T ))],
ρ−1
ρ−1
ρ−1
where T has the beta distribution with parameters ρ(α − 1) + 1 and 1.
3.5
Characterization
Using the following theorem, we can characterize the EGLFR distribution.
Theorem 3.4. The random variable X follows EGLFR(α, β, a, b) if and only if for all real
t > 0, and for all non-negative integer n
δ(n) (t) = U n (t) +
n (n−1)
δ
(t) ,
α
where δ(n) (t) = E (U n (X) |X < t ) and U (t) = −log (redD(t)) = −log(1 − (1 − β(at + 2b t2 ))
1/β
Proof. Necessity condition:
Let F and f be the cdf and pdf of EGLFR(α, β, a, b), respectively. Also, let d(x) be the
derivative function of D(x). It can be shown that
Z t
α
δ(n) (t) =
[−log (D(x))]n d(x)[D(x)]α−1 dx
F (t) 0
Z D(t)
α
=
[−log (z)]n z α−1 dz
F (t) 0
Z D(t)
1
=
[−log (z)]n dz α
F (t) 0
Z D(t)
n
[−log (z)]n z α D(t)
+
[−log (z)]n−1 z α−1 dz
=
F (t)
F
(t)
0
0
n (n−1)
n
(t) .
= U (t) + δ
α
Sufficiency condition: Sarhan and Kundu (2009) showed that δ(n) (t) = U n (t) + αn δ(n−1) (t) ,
implies that
r (x) =
f (x; 1, β, a, b)
f (x)
=α
= αr (x; 1, β, a, b) .
F (x)
F (x; 1, β, a, b)
Therefore, from the uniqueness property of reversible hazard function, we can conclude that X
follows a EGLFR(α, β, a, b).
).
4
Estimation
In this section, we discuss the MLE of the parameters of the model when β 6= 0. Consider X1 , . . . , Xn is a random sample from EGLFR distribution with vector parameter θ =
(α, β, a, b)′ . The log-likelihood function based on this random sample is given as
ℓ(θ) = n log(α) +
n
X
n
n
log(a + bxi ) + (
i=1
X
X
1
1
log 1 − (1 − βzi ) β ,
log(1 − βzi ) + (α − 1)
− 1)
β
i=1
i=1
(4.1)
where zi = axi + 2b x2i . The log-likelihood can be maximized either directly or by solving the
nonlinear likelihood equations obtained by differentiating (4.1). The components of the score
vector U (θ) = (Uα (θ) , Uβ (θ) , Ua (θ) , Ub (θ))T are given by
n
Uα (θ) =
Uβ (θ) =
1
n X
∂ℓ (θ)
log 1 − (1 − βzi ) β ,
= +
∂α
α
i=1
X
n
n
1
1 X
zi
∂ℓ (θ)
log (1 − βzi ) −
=− 2
−1
∂β
β
β
1 − βzi
i=1
+
Ua (θ) =
α−1
β2
n
X
i=1
(1 − βzi )
1
−1
β
1 [βz i + (1 − βzi ) log (1 − βzi )],
β
1
−
(1
−
βz
)
i=1
i
1
X
n
n
n
−1
X
∂ℓ (θ) X
1
βxi
1
xi (1 − βzi ) β
=
−1
−
+ (α − 1)
1 ,
∂a
a + bxi
β
1 − βzi
1 − (1 − βz ) β
i=1
Ub (θ) =
∂ℓ (θ)
=
∂b
n
X
i=1
i=1
xi
−
a + bxi
i
i=1
1
X
n
n
X
βx2i
x2i (1 − βzi ) β −1
1
.
−1
+ (α − 1)
1
β
2 (1 − βzi )
2 1 − (1 − βz ) β
i=1
i=1
i
For given β, a, and b, the MLE of parameter α is
α̂ = − P
n
i=1 log
n
1
1 − (1 − βzi ) β
.
By maximizing the profile log-likelihood ℓ (α̂, β, a, b) with respect to β, a, b, the MLE of these
parameters can be obtained.
4.1
The regular case
When β ≤ 0, the support of the EGLFR distribution is in interval (0, ∞), and in this case the
asymptotic distribution of the MLE of vector parameter θ = (α, β, a, b)′ is multivariate normal
distribution as
√
d
n(θ̂ − θ) −→ N4 0, I −1 ,
where I is the Fisher information matrix.
4.2
The non-regular case
When β > 0, the support of the EGLFR distribution is in interval (0, ψ). So, in this case the
standard asymptotic normality distribution of the MLE’s of parameters does not hold. The
asymptotic distributions are proposed by Kundu and Gupta (2011) for b = 0. Here, similar
results are given when b > 0. The general approach is given by Smith (1985), and is used in
literature (for example Ng et al., 2012).
Take G(x) = ax + 2b x2 . Then G−1 ( β1 ) = ψ, and the pdf of EGLFR distribution can be
written as
α−1
,
f (x; α, a, b, ψ) = αg(x)(1 − t)G(ψ)−1 1 − (1 − t)G(ψ)
x ≤ ψ,
(4.2)
where t = G (x) /G (ψ), andg (x) is the derivative of G (x) with respect to x. So, the loglikelihood function correspond to the pdf in (4.2) is
n
X
ℓ(α, a, b, ψ) = nlog(α) +
+ (α − 1)
i=1
n
X
i=1
n
X
log 1 − t(i)
log g(x(i) ) + (G (ψ) − 1)
i=1
G(ψ)
,
log 1 − 1 − t(i)
where t(i) = G(x(i) )/G (ψ), and x(1) , x(2) , . . . , x(n) are the ordered statistics of random sample
x1 , x2 , . . . , xn . As mentioned in Smith (1985), at first, we find the MLE of the threshold
parameter of the model. For the EGLFR distribution, the MLE of the threshold parameter
ψis ψ̃ = x(n) . Then, the modified log-likelihood function based on the remaining (n − 1)
observations is
ℓ(α, a, b, ψ̃) = (n − 1)log(α) +
+ (α − 1)
n−1
X
i=1
n−1
X
i=1
n−1
X
log 1 − t̃(i)
log g(x(i) ) + (G(ψ̃) − 1)
i=1
G ψ̃
log 1 − (1 − t̃(i) ) ( ) ,
where t̃(i) = G(x(i) )/G ψ̃ . For more information about the modified likelihood function, refer
to Smith (1985). For fixed a and b the modified MLE of parameter α is obtained as
α̃ = − P
n−1
.
G(ψ̃ )
n−1
log
1
−
(1
−
t̃
)
(i)
i=1
By maximizing the modified log-likelihood ℓ α̃, a, b, ψ̃ with respect to a, and b, the modified
MLE of these parameters can be obtained.
Theorem 4.1. The asymptotic distribution of ψ̃ is
d
n1/G(ψ) (ψ̃ − ψ) −→ −
G (ψ)
V 1/G(ψ) ,
(a + bψ)
where the random variable V is distributed as exponential distribution with mean
1
α.
Proof. The corresponding cdf of the pdf in (4.2) is
α
F (x) = 1 − (1 − t)G(ψ) .
1/G(ψ)
where U has a standard uniform distriSo, G (X) /G (ψ) is distributed as 1 − 1 − U 1/α
G(X(n) )
1/α 1/G(ψ)
bution. Therefore, n1/G(ψ) ( G(ψ)
where U(n)
− 1) is distributed as −n1/G(ψ) (1 − U(n) )
has a beta distribution with parameters n and 1. Now, we have
1/α
lim P n(1 − U(n) ) ≤ x = 1 − e−αx ,
n→∞
and then n1/G(ψ) G(X(n) ) − G (ψ)
d
−→ − G (ψ) V 1/G(ψ) . By using the delta method and the
relation between the derivatives of the G (x) and G−1 (x), the proof is completed.
Theorem 4.2. a) Conditioning on X(n) , the asymptotic distribution of the modified MLE,
(α̃, ã, b̃) is multivariate normal distribution.
b)The asymptotic distribution of (α̃, ã, b̃) is (i) multivariate normal if G (ψ) < 12 , (ii) multivariate Weibull if G (ψ) > 12 , and (iii) a mixture of normal and Weibull if G (ψ) = 12 .
Proof. For the proof of this theorem, reader can be refer to Kundu and Gupta (2011).
5
A real example
The following data set is given by Aarset (1987) and represents the lifetimes of 50 devices.
Sarhan and Kundu (2009) and Silva et al. (2010) also analyzed this data.
0.1 0.2 1 1 1 1 1 2 3 6 7 11 12 18 18 18 18 1821 32 36 40 45 46 47 50 55
60 63 63 67 67 67 67 72 7579 82 82 83 84 84 84 85 85 85 85 85 86 86
Using this data, we obtained the MLE’s of parameters of six distributions: extended generalized
linear failure rate (EGLFR), extended generalized exponential (EGE), extended generalized
Rayleigh (EGR), generalized linear failure rate (GLFR), generalized exponential (GE), and
generalized Rayleigh (GR) distributions. The results are given in Table 2.
Based on the MLE’s of parameters, we calculated minus of log-likelihood function
(− log (L)), Kolmogorov-Smirnov (K-S) statistic with its p-value, Akaike information criterion
(AIC), Akaike information criterion corrected (AICC), Bayesian information criterion (BIC)
and likelihood ratio test (LRT) with its p-value. Table 2 indicates that the GLFR and GE
models are not suitable for this data set based on K-S statistic. Also, The EGLFR model has
the lowest − log (L), AIC, AICC and BIC values among all fitted models. A comparison of
the proposed distribution with some of its sub-models using p-value of LRT shows that the
EGLFR model yields a better fit than the other five distributions to this real data set.
The beta modified Weibull (BMW) distribution was introduced by Silva et al. (2010) and
has the following pdf:
fBM W (x) =
γ βx a−1
αxγ (γ + βx) βx−bαxγ eβx
1 − e−αx e
,
e
B (a, b)
x > 0,
where B (a, b) is the beta function with α, β, γ > 0 and 0 < a, b < 1. This distribution contains
several important distributions such as Weibull, modified Weibull (Lai et al., 2003), generalized
exponential (Gupta and Kundu, 1999), beta Weibull (Lee et al., 2007) and generalized modified Weibull (Carrasco et al., 2008) distributions. Silva et al. (2010) showed that the BMW
distribution produces a better fit than its sub-models for this data set. In Table 2, we present
the MLE’s, −log(L), and other criteria for the BMW distribution, and we can conclude that
the EGLFR, EGR and EGE have better fit than the BMW distribution for this data set.
In Figure 5, we plot the histogram of this data set and the estimated pdf of the seven
models. Moreover, the plots of empirical cdf of the data set and estimated cdf of the five
models are displayed in Figure 5. Again, we can conclude that EGLFR distribution is a very
satisfactory model for this data set.
There are various extensions of GLFR distribution in literature. The pdf of seven extensions
b 2
of this distribution are given as follows. Here, we consider g(x) = (a + bx) e−(ax+ 2 x ) and
b 2
G(x) = 1 − e−(ax+ 2 x ) , where a > 0 and b > 0.
Beta LFR (BLFR) distribution
This distribution is proposed by Jafari and Mahmoudi (2015) with the following pdf:
fBLFR (x) =
g (x)
(G(x))α−1 (1 − G (x))β−1 ,
B (α, β)
x > 0,
where α > 0 and β > 0.
Kumaraswamy GLFR (KGLFR) distribution
This distribution is proposed by Elbatal (2013) with the following pdf:
fKGLFR (x) = αβg (x) (G(x))α−1 (1 − (G(x))α )β−1 ,
x > 0,
where α = cθ, and θ > 0, c > 0 and β > 0.
McDonald GLFR (MCGLFR) distribution
This distribution is proposed by Elbatal et al. (2014) with the following pdf:
fMCGLFR (x) =
γg (x)
(G(x))αγ−1 (1 − (G(x))γ )β−1 ,
B (α, β)
x > 0,
where γ = cθ, θ > 0, c > 0, α > 0 and β > 0.
Modified GLFR (MGLFR) distribution
This distribution is proposed by Jamkhaneh (2014) with the following pdf:
α−1
−(ax+bxβ )
β
1 − e−(ax+bx )
,
fMGLFR (x) = α a + bβxβ−1 e
x > 0,
where a > 0, b > 0, α > 0 and β > 0.
Poisson GLFR (PGLFR) distribution
This distribution is proposed by Cordeiro et al. (2015) with the following pdf:
fPGLFR (x) =
α
βαg (x)
(G(x))α−1 e−β(G(x)) ,
−β
(1 − e )
x > 0,
where α > 0 and β > 0.
Geometric GLFR (GGLFR) distribution
This distribution is proposed by Nadarajah et al. (2014) with the following pdf:
fGGLFR (x) =
α (1 − β) g (x) (G(x))α−1
(1 − β (1 − (G(x))α ))2
,
x > 0,
where α > 0 and β < 1.
Generalized Linear Exponential (GLE) distribution
This distribution is proposed by Mahmoud and Alam (2010) with the following pdf:
α−1
α
b 2
b 2
e−(ax+ 2 x −β ) ,
x > 0,
fGGLFR (x) = α (a + bx) ax + x − β
2
where a > 0, b > 0, α > 0 and β > 0.
We also fitted these distributions to the data set, and obtained the MLE’s, −log(L) and
other criteria for them. The results are given in Table 3. It can be concluded that the EGLFR,
EGR and EGE have better fit than these seven extension of GLFR distribution for this data
set. The histogram of the data set with the estimated pdf’s and the empirical cdf of the data
set with estimated cdf’s for distributions are given in Figure 6.
Table 2: MLE’s of the model parameters, and the K-S, AIC, AICC, BIC, and LRT statistics.
Statistic
EGLFR
α̂
0.2620
4.5000
β̂
â
1.21 × 10−8
0.00006
b̂
γ̂
—
−log(L)
173.9487
K-S
0.0981
p-value (K-S)
0.7215
AIC
355.8974
AICC
356.7863
BIC
363.5455
LRT
—
p-value (LRT)
—
EGE
0.5368
1.8199
0.0064
—
—
189.1973
0.1558
0.1763
384.3945
384.9163
390.1306
30.4971
0.0000
Distribution
EGR
GLFR
0.2590
0.5327
4.2100
—
—
0.0038
0.00006 0.0003
—
—
180.5367 233.1447
0.0872
0.1832
0.8413
0.0696
367.0733 472.2895
367.5951 472.8110
372.8094 478.0256
13.1759 118.3921
0.0002
0.0000
GR
0.3520
—
—
0.0003
—
234.5655
0.2011
0.0350
473.1309
473.3862
476.9550
121.2335
0.0000
BMW
0.0002
0.0541
0.1975
0.1647
1.3771
220.6601
0.0846
0.3971
451.3201
452.6838
460.8802
—
—
Empirical Distribution
1.0
0.000 0.005 0.010 0.015 0.020 0.025 0.030
EGLFR
EGE
EGR
GLFR
GE
GR
BMW
0.6
0.8
EGLFR
EGE
EGR
GLFR
GE
GR
BMW
0.0
0.2
0.4
Fn(x)
Density
Histogram
GE
0.7798
—
0.0187
—
—
239.9951
0.2042
0.0309
483.9903
484.2456
487.8143
132.0929
0.0000
0
20
40
60
80
x
100
0
20
40
60
80
x
Figure 5: The histogram of the data set with the estimated pdf’s (left); the empirical cdf of
the data set and estimated cdf’s (right).
Table 3: MLEs of the model parameters, and the K-S (and its p-value), AIC, AICC, BIC for
other extensions of GLFR distribution.
Statistic
α̂
β̂
â
b̂
γ̂
−log(L)
K-S
p-value (K-S)
AIC
AICC
BIC
BLFR
0.3347
0.1243
0.0172
0.0035
—
230.3785
0.1554
0.1784
468.7570
469.6459
476.4051
KGLFR
0.6525
0.0622
0.2988
0.0007
—
238.0490
0.1666
0.1246
484.0980
484.9869
491.7461
Distribution
MCGLFR MGLFR
0.0295
19699.45
5.74 × 108 0.0164
0.0015
0.0246
−5
6.66 × 10
8.8393
1.8936
—
221.9929 235.3460
0.1949
0.1624
0.0448
0.1428
453.9858 478.6921
455.3494 479.5810
463.5459 486.3402
GLE
0.6262
0.0015
0.0149
0.0005
—
227.1663
0.2327
0.0088
462.3327
463.2216
469.9808
Empirical Distribution
0.6
0.8
BLFR
KGLFR
MCGLFR
MGLFR
PGLFR
GGLFR
GLE
0.4
Fn(x)
0.015
0.020
0.025
BLFR
KGLFR
MCGLFR
MGLFR
PGLFR
GGLFR
GLE
0.0
0.000
0.005
0.2
0.010
Density
GGLFR
0.2624
-5.5536
0.0086
0.0005
—
229.9373
0.1297
0.3694
467.8745
468.7634
475.5226
1.0
0.030
Histogram
PGLFR
0.5327
10−8
0.0038
0.0003
—
233.1447
0.1832
0.0696
474.2895
475.1784
481.9376
0
20
40
60
x
80
100
0
20
40
60
80
x
Figure 6: The histogram of the data set with the estimated pdf’s (left); the empirical cdf of
the data set and estimated cdf’s (right) for seven extensions of GLFR distribution.
6
Conclusion
In this paper, a four-parameter EGLFR distribution has been proposed. Although the GLFR
distribution cannot have a unimodal hazard rate function, but EGLFR distribution can have
a decreasing, increasing, unimodal, and bathtub-shaped hazard rate function depending on
its parameters. This new class of distributions includes some well-known distribution such as
EGE and GLFR distributions as special sub-models. Several properties of EGLFR distribution
such as quantiles function, Skewness and kurtosis measures, and the r-th non-central moment
have been given. the maximum likelihood estimators of the parameters are given and also, the
asymptotic distribution of estimates are investigated. By using a real data set, we have shown
that our new distribution fits to the lifetime data very well than the well-known distributions.
Recently Sarhan et al. (2011) introduced the bivariate GLFR distribution and its multivariate
extension. So, one can use the approach used in this paper to develop the multivariate EGLFR
distribution.
Acknowledgements
The authors would like to thank the anonymous referees for many helpful comments and
suggestions.
References
Aarset, M. V. (1987). How to identify a bathtub hazard rate. IEEE Transactions on Reliability,
R-36(1):106–108.
Alamatsaz, M. H. and Shams, S. (2014). Generalized linear failure rate power series distribution.
Communications in Statistics: Theory and Methods, In press.
Carrasco, J. M., Ortega, E. M., and Cordeiro, G. M. (2008). A generalized modified Weibull
distribution for lifetime modeling. Computational Statistics & Data Analysis, 53(2):450–462.
Cordeiro, G. M., Hashimoto, E. M., Ortega, E. M., and Pascoa, M. A. (2012). The McDonald
extended distribution: properties and applications. AStA Advances in Statistical Analysis,
96(3):409–433.
Cordeiro, G. M., Ortega, E. M. M., and Lemonte, A. J. (2015). The Poisson generalized linear
failure rate model. Communications in Statistics-Theory and Methods, 44(10):2037–2058.
Elbatal, I. (2013). Kumaraswamy generalized linear failure rate distribution. Indian Journal
of Computational & Applied Mathematics, 1(1):61–78.
Elbatal, I., Merovci, F., and Marzouk, W. (2014). McDonald generalized linear failure rate
distribution. Pakistan Journal of Statistics and Operation Research, 10(3):267–288.
Foss, S., Korshunov, D., and Zachary, S. (2011). An Introduction to Heavy-Tailed and Subexponential Distributions. Springer, New York.
Gradshteyn, I. S. and Ryzhik, I. M. (2007). Table of Integrals, Series, and Products, Edited by
Alan Jeffrey and Daniel Zwillinger. Academic Press, New York, 7th edition.
Gupta, R. D. and Kundu, D. (1999). Generalized exponential distributions. Australian & New
Zealand Journal of Statistics, 41(2):173–188.
Jafari, A. A. and Mahmoudi, E. (2015). Beta-linear failure rate distribution and its applications.
Journal of the Iranian Statistical Society, Accepted for publication.
Jamkhaneh, E. B. (2014). Modified generalized linear failure rate distribution: Properties and
reliability analysis. International Journal of Industrial Engineering Computations, 5(3):375–
386.
Johnson, N. L., Kotz, S., and Balakrishnan, N. (1995). Continuous Univariate Distributions,
volume 2. John Wiley & Sons, New York, second edition.
Kenney, J. F. and Keeping, E. (1962). Mathematics of Statistics. D. Van Nostrand Company.
Kundu, D. and Gupta, R. D. (2011). An extension of the generalized exponential distribution.
Statistical Methodology, 8(6):485–496.
Kundu, D. and Raqab, M. Z. (2005). Generalized Rayleigh distribution: different methods of
estimations. Computational Statistics & Data Analysis, 49(1):187–200.
Lai, C., Xie, M., and Murthy, D. (2003). A modified Weibull distribution. IEEE Transactions
on Reliability, 52(1):33–37.
Lee, C., Famoye, F., and Olumolade, O. (2007). Beta-Weibull distribution: some properties and
applications to censored data. Journal of Modern Applied Statistical Methods, 6(1):173–186.
Mahmoud, M. A. W. and Alam, F. M. A. (2010). The generalized linear exponential distribution. Statistics & Probability Letters, 80(1112):1005–1014.
Mahmoudi, E. and Jafari, A. A. (2015). The compound class of linear failure rate-power series
distributions: model, properties and applications. Communications in Statistics - Simulation
and Computation, 10.1080/03610918.2015.1005232.
Moors, J. J. A. (1988). A quantile alternative for kurtosis. Journal of the Royal Statistical
Society. Series D (The Statistician), 37(1):25–32.
Nadarajah, S., Shahsanaei, F., and Rezaei, S. (2014). A new four-parameter lifetime distribution. Journal of Statistical Computation and Simulation, 84(2):248–263.
Ng, H. K. T., Luo, L., Hu, Y., and Duan, F. (2012). Parameter estimation of three-parameter
Weibull distribution based on progressively type-II censored samples. Journal of Statistical
Computation and Simulation, 82(11):1661–1678.
Sarhan, A. M., Hamilton, D. C., Smith, B., and Kundu, D. (2011). The bivariate generalized
linear failure rate distribution and its multivariate extension. Computational Statistics and
Data Analysis, 55(1):644–654.
Sarhan, A. M. and Kundu, D. (2009). Generalized linear failure rate distribution. Communications in Statistics-Theory and Methods, 38(5):642–660.
Shannon, C. (1948). A mathematical theory of communication. Bell System Technical Journal,
27:379–432.
Silva, G. O., Ortega, E. M., and Cordeiro, G. M. (2010). The beta modified Weibull distribution. Lifetime Data Analysis, 16(3):409–430.
Smith, R. L. (1985). Maximum likelihood estimation in a class of nonregular cases. Biometrika,
72(1):67–90.
Surles, J. and Padgett, W. (2005). Some properties of a scaled Burr type X distribution.
Journal of Statistical Planning and Inference, 128(1):271–280.
Surles, J. G. and Padgett, W. J. (2001). Inference for reliability and stress-strength for a scaled
Burr type X distribution. Lifetime Data Analysis, 7(2):187–200.
Zografos, K. and Balakrishnan, N. (2009).
On families of beta-and generalized gamma-
generated distributions and associated inference. Statistical Methodology, 6(4):344–362.