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

net/publication/267673743

Marshall-Olkin extended Lindley distribution and its application

Article · January 2012

CITATIONS READS

9 437

4 authors, including:

M. E. Ghitany Maali Al-Buraies

Kuwait University Kuwait Institute for Scientific Research
63 PUBLICATIONS   1,463 CITATIONS    3 PUBLICATIONS   9 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Maali Al-Buraies on 13 June 2017.

The user has requested enhancement of the downloaded file.

International Journal of Applied Mathematics
————————————————————–
Volume 25 No. 5 2012, 709-721

MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION

AND ITS APPLICATION

M.E. Ghitany1 § , D.K. Al-Mutairi2 , F.A. Al-Awadhi3 , M.M. Al-Burais4

1,2,3 Departmentof Statistics and Operations Research
Faculty of Science
Kuwait University
P.O. Box 5969, Safat 13060, KUWAIT
4 Techno Economics Division

Quantitative Methods and Modeling Department

Kuwait Institute for Science and Research (KISR)
P.O. Box 24885, Safat 13109, KUWAIT

Abstract: A two-parameter Marshall-Olkin extended Lindley distribution

is introduced and its structural properties are investigated. These include the
compounding representation of the distribution, the shapes of the density and
hazard rate functions, the moments and quantiles as well as the limiting dis-
tributions of extreme order statistics. Maximum likelihood estimation of the
model’s parameters and their estimated asymptotic standard errors are derived
based on randomly right censored data. Application of the model to randomly
right censored real data set will be presented and compared with other com-
monly used two-parameter distributions.

AMS Subject Classification: 62F10, 60E15, 62N05

Key Words: lindley distribution, maximum likelihood estimation, hazard
rate, extreme order statistics, minimal and maximal domain of attraction

Received: September 4, 2012 c 2012 Academic Publications

§ Correspondence author
710 M.E. Ghitany, D.K. Al-Mutairi, F.A. Al-Awadhi, M.M. Al-Burais

1. Introduction

Marshall and Olkin [11] introduced a general method of extending a family of

distributions, with survival function F (t), by adding a tilt parameter α giving
rise to the survival function

αF 1 (x)
F (x) = , −∞ < x < ∞, α > 0, α = 1 − α. (1)
1 − αF 1 (x)

They have noted that the method has a stability property, i.e., if the method
is applied twice, nothing new is obtained in the second time around. The
Marshall-Olkin extended distributions give a wider range of behavior of the
probability density function (p.d.f.), hazard rate function (h.r.f.) than those of
the baseline distributions from which they are derived.
For different baseline distributions, the model given by (1) has been ex-
tensively studied by various authors including Ghitany [6], Ghitany et al. [7],
Ghitany et al. [4], and Ghitany and Kotz [5].
In this paper, we introduce a new variant of Marshall-Olkin model by con-
sidering the baseline distribution to be the Lindley distribution, denoted by
L(θ), with survival function
 θ 
F 1 (x) = 1 + x e−θx , x > 0, θ > 0, (2)
θ+1

see see Ghitany et al. [3] for more details about this distribution. The resulted
Marshall-Olkin extended Lindley distribution, denoted by MOEL(α, θ), with
survival function
 
θ
α 1 + θ+1 x e−θx
F (x) =   , x > 0, α, θ > 0, α = 1 − α. (3)
θ
1 − α 1 + θ+1 x e−θx

The aim of this paper is to reveal some statistical properties of the MOEL
distribution. These properties include: (i) expressing this proposed distribution
as a compounding process, (ii) shapes of the p.d.f. and h.r.f., (iii) existence of
the moments and quantiles, and (iv) the limiting distributions of order statis-
tics. Application of the proposed distribution to randomly right censored data
is presented and is compared with other commonly used two-parameter distri-
butions.
MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION... 711

2. Compounding

The following theorem shows that the MOEL distribution can be obtained by
the compounding argument.

Theorem 1. Let X|ψ have a (conditional) SF

( " # )
1
F (x|ψ) = exp − θ
− 1 ψ , x > 0, ψ, θ > 0,
(1 + θ+1 x)e−θx

and ψ have (unconditional) p.d.f.

m(ψ) = α e−α ψ , ψ > 0, α>0
Then the (unconditional) distribution of X is MOEL (α, θ).

Proof. The theorem holds, since

Z ∞
F (x) = F (x|ψ) m(ψ) dψ
0
∞ θ −θx
1 − α(1 + θ+1 x)e
Z n o
=α exp − θ
ψ dψ
(1 − −θx
0 θ+1 x)e
θ −θx
α(1 + θ+1 x)e
= θ
.
1 − α(1 + θ+1 x)e−θx

3. Probability Density Function

In this section we investigate the shape of the p.d.f. of MOEL(α, θ) distribution

given by
θ2
α θ+1 (1 + x) e−θx
f (x) = θ
, x > 0, α, θ > 0, α = 1 − α. (4)
[1 − α (1 + −θx ]2
θ+1 x)e

θ2
Note that f (0) = α(θ+1) and f (∞) = 0.

Theorem 2. The p.d.f. f (x) of MOEL(α, θ) distribution is decreasing

(unimodal) if

2θ 2 2θ 2
α≤ (α > ).
θ2 + 1 θ2 + 1
712 M.E. Ghitany, D.K. Al-Mutairi, F.A. Al-Awadhi, M.M. Al-Burais

Proof. The derivative of f (x) is given by

αθ 2 −θx
′ (1+θ)2 e
f (x) = h i3 ψ(θ(1 + x)),
θ
1 − α(1 + θ+1 x)e−θx

where
ψ(y) = (θ + 1)(1 − y) − α (1 + y 2 )e−(y−θ) , y > θ,
with ψ(θ) = α(θ 2 + 1) − 2θ 2 and ψ(∞) = −∞.
Since ψ(y) is negative (changes sign from positive to negative) if ψ(θ) ≤
0 (ψ(θ) > 0), it follows that f (x) decreases (unimodal) if ψ(θ) ≤ 0 (ψ(θ) > 0).
This completes the proof.

By Theorem 2, it can be seen that when α = 1, i.e. Lindley distribution,

f (x) is decreasing (unimodal) for θ ≥ 1 (0 < θ < 1), see Ghitany et al. [3].
Figure 1 shows the p.d.f. of the Marshall-Olkin Lindley distribution for
selected values of α and θ.

4. Hazard Rate Function

The h.r.f. of the MOE Lindley distribution is given by

θ 2 (1+x)
θ(1+x)+1
h(x) = θ
, x > 0, α, θ > 0 (5)
1−α (1 + θ+1 x)e−θx

θ2
Note that h(0) = f (0) = α(θ+1) and h(∞) = θ.

Theorem 3. The h.r.f. h(x) of MOEL(α, θ) distribution is:

(a) increasing if
(i) α ≥ 1 and θ > 0,
√
(ii) 0 < α ≤ 2(y(θ+1) e−θ
+1)2 e−y0
and θ < y0 , where y0 = 1 + 2,
0
1
(iii) 0 < α ≤ θ 2 +1
and θ ≥ y0 .
(b) decreasing-increasing if
(i) α ≥ θ21+1 and θ < y0 ,
(ii) α > θ21+1 and θ ≥ y0 .
MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION... 713

1.5

f HxL 1

0.5

0
0 1 2 3 4 5
x
(a)

0.2

0.15
f HxL

0.1

0.05

0
0 2 4 6 8 10 12
x
(b)

Figure 1: Density of MOE Lindley distribution: (a) (α, θ): (0.5, 1)

(–), (1, 2) (- -), (1.5, 3) (. . .) (b) (α, θ): (0.5, 0.25) (–), (1, 0.5) (- -),
(3, 0.75) (. . .).

(θ+1) e−θ 1
(c) increasing-decreasing-increasing if 2(y0 +1)2 e−y0
<α< θ 2 +1
and θ < y0 .

Proof. The derivative of h(x) is given by

θ2
′ (θ+1)2
h (x) =
2 2 η(θ(x + 1)),
θ θ


1+ θ+1 x 1−α 1+ θ+1 x e−θx
714 M.E. Ghitany, D.K. Al-Mutairi, F.A. Al-Awadhi, M.M. Al-Burais

where
eθ
η(y) = 1 − α (y + 1)(y 2 + 1)e−y , y > θ.
θ+1
Note that η(θ) = 1 − α (θ 2 + 1) and η(∞) = 1.
Clearly, η(y) ≥ 0, i.e. h(x) is increasing, if α ≥ 1 and θ > 0.
For α < 1, the sign of η(y) can be determined as follows.
Case I: θ < y0 : √
Here η(y) has an absolute minimum at the point y0 = 1+ 2. Consequently,
• η(y) ≥ 0 if η(y0 ) ≥ 0, proving (a) (ii).
• η(y) changes sign from - to + if η(θ) ≤ 0, proving (b) (i).
• η(y) changes sign from + to - to + if η(θ) > 0 and η(y0 ) < 0, proving (c).
Case II: θ ≥ y0 :
Here, η ′ (y) ≥ 0, i.e. η(y) is increasing. Consequently,
• η(y) ≥ 0 if η(θ) ≥ 0, proving (a) (iii).
• η(y) changes sign from - to + if η(θ) < 0, proving (b) (ii).

By Theorem 3, it can be seen that when α = 1, i.e. Lindley distribution,

h(x) is increasing with h(0) = θ 2 /(θ + 1) and h(∞) = θ, see Ghitany et al. [3].
Figure 2 shows the h.r.f. h(x) of the MOE Lindley distribution for selected
values of α and θ.

5. Moments and Quantiles

Theorem 4. The MOEL(α, θ) distribution has finite moments of all posi-

tive orders.

Proof. Since F (0) = 0 and

θ 2 x(x + 1)
lim inf x h(x) = lim inf = ∞,
x→∞ x→∞ θ(x + 1) + 1

it follows that the MOE Lindley distribution has finite moments of all positive
orders, see Barlow et al. (1963), page 386.

Remark. Although MOEL(α, θ) distribution has finite moments of all

positive orders, these moments cannot be expressed in simple closed form. This
is also the case for other MOE distributions such as MOE Weibull (Marshall
and Olkin [11]), MOE Lomax (Ghitany et al. [4]) and MOE linear-failure rate
(Ghitany and Kotz [5]). The MOE exponential (Marshall and Olkin [11]) has
MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION... 715

3 3

2 2
hHxL

hHxL
1 1

0 0
0 2 4 6 8 10 0 2 4 6 8 10
x x

(a) (b)

0.4
hHxL

3
hHxL

0.3
2

1 0.2
0 2 4 6 8 10 0 5 10 15 20 25
x x

(c) (d)

Figure 2: Hazard rate function of MOE Lindley distribution:

(a) (α, θ): (0.9, 2) (–), (0.9, 2.5) (- -), (0.9, 3) (. . .)
(b) (α, θ): (1, 1) (–), (2, 2) (- -), (3, 3) (. . .)
(c) (α, θ): (0.5, 1.5) (–), (0.5, 2) (- -), (0.5, 3) (. . .)
(d) (α, θ): (0.3, 0.3) (–), (0.4, 0.4) (- -), (0.5, 0.5) (. . .).

simple closed form expression for its mean but no closed form expression for
second or higher moments.
Recently, Jodra [8] showed that the quantile function of the Lindley distri-
bution is given by

1 1  θ+1 
F1−1 (v) = −1 − − W−1 − θ+1 (1 − v) , 0 < v < 1,
θ θ e

where W−1 (.) denotes the negative branch of the Lambert W function (i.e., the
solution of the equation W (z) eW (z) = z).
716 M.E. Ghitany, D.K. Al-Mutairi, F.A. Al-Awadhi, M.M. Al-Burais

The quantile function of the MOE Lindley distribution is given by

 
−1 −1 1−u
F (u) = F1 1−
1 − αu
 
1 1 θ+1 1−u
= −1 − − W−1 − θ+1 . , 0 < u < 1. (6)
θ θ e 1 − αu

The quartiles of the MOE Lindley distribution can be obtained by setting u =

1/4, 1/2, 3/4, respectively.

6. Extreme Order Statistics and Their Limiting Distributions

Let X1:n , . . . , Xn:n be the order statistics of a random sample of size n from
MOEL distribution with survival function F (x) given by (3).
The c.d.f. of the minimum order statistic X1:n is given by
   n
θ
α 1 + θ+1 x e−θx
F1:n (x) = 1 − [F (x)]n = 1 −     .
θ
1 − α 1 + θ+1 x e−θx

The c.d.f. of the maximum order statistic Xn:n is given by

   n
θ
α 1 + θ+1 x e−θx
Fn:n (x) = [1 − F (x)]n = 1 −    .
θ
1 − α 1 + θ+1 x e−θx

The minimum (maximum) order statistics represents the lifetime of a series

(parallel) system in reliability studies.

Theorem 5. Let X1:n , . . . , Xn:n be the order statistics of a random sample

of size n from MOEL(α, θ) distribution. Then
nX
1:n − a∗n o
(i) lim P ≤x = 1 − e−x , x>0
n→∞ b∗n
nX − an o
n:n
(ii) lim P ≤t = exp(−e−t ), −∞ < t < ∞,
n→∞ bn
where

a∗n = 0, b∗n = F −1 (1/n), an = F −1 (1 − 1/n), bn = 1,

MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION... 717

and F −1 (.) is given by (6).

Proof. (i) For the MOEL(α, θ), we have, by using L’Hospital rule,
F (ǫx) x f (ǫx) [1 − αF 1 (ǫ)]2 (ǫx + 1)e−θǫx
lim = lim = x lim = x.
ǫ→0+ F (ǫ) ǫ→0+ f (ǫ) ǫ→0+ [1 − αF 1 (ǫx)]2 (ǫ + 1)e−θǫ
Therefore, by Theorem 8.3.6 (ii) , page 214, of Arnold et al. [1], the minimal
domain of attraction of the MOEL(α, θ) distribution is the standard exponential
distribution.
(ii) For the MOEL(α, θ), we have
 
d 1 1 − αF 1 (x)
lim = lim αF 1 (x) − 2 = 0.
x→∞ dx h(x) x→∞ θ (x + 1)2
Therefore, by Theorem 8.3.3 , page 212, of Arnold et al. [1], the maximal
domain of attraction of the MOEL(α, θ) distribution is the standard Gumbel
distribution.

7. Estimation

In this section we use the maximum likelihood method to estimate the param-
eters α, and θ of the MOE Lindley distribution.
Let (t1 , δ1 ) . . . , (tn , δn ) be randomly right censored data of size n from
MOEL(α, θ), where δi = 0(1) if ti is right censored (failure) observation. For
non-informative censoring, i.e. the distribution of the censoring does not in-
volve the parameters of the failure distribution, the log-likelihood function is
given by
n
X n
X
ln L(α, θ) = δi ln[f (ti )] + (1 − δi ) ln[F (ti )].
i=1 i=1

The MLEs α̂, θ̂ of α, θ are the solutions of the non-linear equations:

n
∂ n X F 1 (ti )
ln L = − (1 + δi ) = 0,
∂α α 1 − αF 1 (ti )
i=1
n n
∂ X h θ+2 i X h 1 
ln L = δi − ti + (1 − δi ) ti 2
−1
∂θ θ(θ + 1) (θ + 1) + θ(θ + 1)ti
i=1 i=1
n h e−θti
X ti i
+α (1 + δi ) − F 1 (t i ) = 0,
i=1
1 − αF 1 (ti ) (θ + 1)2
718 M.E. Ghitany, D.K. Al-Mutairi, F.A. Al-Awadhi, M.M. Al-Burais

where F 1 (.) is the survival function of the Lindley distribution.

In practice, one has to use numerical methods to find the MLEs, such
methods are well implemented in MATHEMATICA, MAPLE and R packages.
The observed variance-covariance matrix of (α̂, θ̂) is given by
 −1
d11 (α, θ) d12 (α, θ)
Vn (α, θ) =   ,
d12 (α, θ) d22 (α, θ)

where
∂2 ∂2
d11 (α, θ) = − ln L(α, θ), d22 (α, θ) = − ln L(α, θ),
∂α2 ∂θ 2
∂2
d12 (α, θ) = − ln L(α, θ).
∂α∂θ
For right random censoring, under smoothness conditions, the MLE vector
(α̂, θ̂) is asymptotically distributed as a bivariate normal distribution, i.e.
    
α̂ d α
∼ N2 , Vn (α, θ) ,
θ̂ θ
see Miller [12], page 21.
The estimated standard error (S.E.) of α̂ and θ̂, respectively, are given by
s s
\ = d22 (α̂, θ̂) ,
S.E.(α̂) \θ̂) = d11 (α̂, θ̂) .
S.E.(
∆(α̂, θ̂) ∆(α̂, θ̂)
where
∆(α̂, θ̂) = d11 (α̂, θ̂)d22 (α̂, θ̂) − d212 (α̂, θ̂).

8. Application

In this section we fit the proposed MOE Lindley distribution to randomly right
censored data. Also, we fit the same data to the following four well known two-
parameter distributions, each of which is a generalization of the exponential
distribution.
(i) Gamma:
ba11
fG (x) = xa1 −1 e−b1 x , x > 0, a1 , b1 > 0.
Γ(a1 )
MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION... 719

Model MLEs S.E. ln L

Lindley θ̃ = 0.187 0.0117 -428.005
Gamma â1 = 0.983 0.1291 -421.708
b̂1 = 0.002 0.0165
Weibull â2 = 0.927 0.0681 -422.415
b̂2 = 0.004 0.0177
Generalized Exponential â3 = 1.215 0.1457 -421.444
b̂3 = 0.114 0.0129
MOE Exponential â4 = 0.346 0.3250 -422.713
b̂4 = 0.001 0.0192
MOE Lindley α̂ = 0.026 0.1006 -418.030
θ̂ = 0.001 0.0347

Table 1: Summary of the fitted distributions.

(ii) Weibull:
a2
fW (x) = a2 b2 xa2 −1 e−b2 x , x > 0, a2 , b2 > 0.

(iii) Generalized Exponential:

fGE (x) = a3 b3 e−b3 x (1 − e−b3 x )a3 −1 , x > 0, a3 , b3 > 0.

(iv) MOE Exponential:

a4 b4 e−b4 x
fEE (x) = , x > 0, a4 , b4 > 0, a4 = 1 − a4 .
(1 − a4 e−b4 x )2

The data considered here represents the remission times (in months) for
137 bladder cancer patients (Lee and Wang [10]). The data contains 9 right
censored observations.
Table 1 shows that the MOE Lindley model has the largest log-likelihood
among the competing distributions. Also, the log-likelihood ratio test for Lin-
dely model versus MOE Lindley model, i.e. for testing H0 : α = 1 versus
H1 : α 6= 1, gives a test statistic χ20 = 2[ln L(α̂, θ̂) − ln L(1, θ̃)] = 19.95 with
p-value of the test P (χ2(1) > 19.95) = 0.000008. Hence, the Lindley model
cannot be accepted for the given data. Finally, Figure 3 shows the probability-
probability (P-P) plot of the fitted models. The figure supports the claim hat
the MOE Lindley model is the most suitable among the fitted models.
720 M.E. Ghitany, D.K. Al-Mutairi, F.A. Al-Awadhi, M.M. Al-Burais

1.0

1.0
KME of Survival Function

KME of Survival Function

0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

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

Fitted Lindley Survival Function Fitted Gamma Survival Function

(a) (b)
1.0

1.0
KME of Survival Function

KME of Survival Function

0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0

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

Fitted Weibull Survival Function Fitted Generalized Exponential Survival Function

(c) (d)
1.0

1.0
KME of Survival Function

KME of Survival Function

0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0

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

Fitted MOE Exponential Survival Function Fitted MOE Lindley Survival Function

(e) (f)

Figure 3: P-P plots of the fitted distributions.

MARSHALL-OLKIN EXTENDED LINDLEY DISTRIBUTION... 721

References

[1] B.C. Arnold, N. Balakrishnan, H.N. Nagaraja, A First Course in Order

Statistics, Wiley, New York (1992).

[2] R.E. Barlow, A.W. Marshall, F. Proschan, Properties of probability distri-

butions with monotone hazard rate, The Annals of Mathematical Statistics,
34 (1963), 375-389.

[3] M.E. Ghitany, B. Atieh, S. Nadarajah, Lindley distribution and its appli-
cation, Mathematics and Computers in Simulation, 78 (2008), 493-506.

[4] M.E. Ghitany, F.A. Al-Awadhi, L.A. Alkhalfan, Marshall-Olkin extended

Lomax distribution and its application to censored data, Communications
in Statistics-Theory and Methods, 36 (2007), 1855-1866.

[5] M.E. Ghitany, S. Kotz, Reliability properties of extended linear failure-rate

distributions, Probability in the Engineering and Informational Sciences,
21 (2007), 441-450.

[6] M.E. Ghitany, Marshall-Olkin extended Pareto distribution and its appli-
cation, International Journal of Applied Mathematics, 18 (2005), 17-31.

[7] M.E. Ghitany, E.K. Al-Hussaini, R.A. Al-Jarallah, Marshall-Olkin ex-

tended Weibull distribution and its application to censored data, Journal
of Applied Statistics, 32 (2005), 1025-1034.

[8] P. Jodra, Computer generation of random variables with Lindley or

Poisson-Lindley distribution via the Lambert W function. Mathematics
and Computers in Simulation, 81 (2010), 851-859.

[9] E.L. Kaplan, P. Meier, Nonparametric estimation from incomplete observa-

tions, Journal of the American Statistical Association, 53 (1958), 457-481.

[10] E.T. Lee, J.W. Wang, Statistical Methods for Survival Data Analysis, Third
Edition, Wiley, N.Y. (2003).

[11] A.W. Marshall, I. Olkin, A new method for adding a parameter to a family
of distributions with applications to the exponential and Weibull families,
Biometrika, 84 (1997), 641-652.

[12] P.G. Miller Jr., Survival Analysis, Wiley, N.Y. (1981).

 722

View publication stats