REVSTAT – Statistical Journal

Volume 0, Number 0, Month 0000, 000-000

ONE PARAMETER POLYNOMIAL EXPONENTIAL

DISTRIBUTION WITH BINOMIAL MIXTURE

Authors: Molay Kumar Ruidas

– Department of Statistics, Triveni Devi Bhalotia College,
Kazi Nazrul University, Raniganj-713 347, India ([email protected])
Indrani Mukherjee
– Department of Statistics, Visva-Bharati University,
Santiniketan-731 235, West Bengal, India ([email protected])
Mriganka Mouli Choudhury
– Department of Statistics, Visva-Bharati University,
Santiniketan-731 235, West Bengal, India ([email protected])

Sudhansu S. Maiti *
– Department of Statistics, Visva-Bharati University,
Santiniketan-731 235, West Bengal, India ([email protected])
Sumanta Adhya
– Department of Statistics, West Bengal State University,
Barasat-700 126, India ([email protected])

Received: Month 0000 Revised: Month 0000 Accepted: Month 0000

Abstract:
ˆ A further generalized version of one parameter polynomial exponential distribution
with binomial probability mass as a mixture called a Binomial Mixture One Pa-
rameter Polynomial Exponential Distribution (BMOPPE) is proposed in the article.
The moments and stochastic orderings are studied. Maximum likelihood estimator
(MLE) and uniformly minimum variance unbiased estimator (UMVUE) of the prob-
ability density function and the cumulative distribution function have been derived
and compared in the mean squared error sense. Estimation issues (both MLE and
UMVUE) of reliability functions- mission time and stress-strength have been consid-
ered, and asymptotic variances of MLEs and variances of UMVUEs have been derived.
UMVUEs of the variance of UMVUE of reliability functions have also been derived.
Simulation study results have been reported to validate the theoretical findings. Few
data sets have been fitted and compared.

*
Corresponding author
2 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

Key-Words:
ˆ Akaike’s Information criterion; Asymptotic variance; Gamma distribution; Mixture
distribution; Reliability function.

AMS Subject Classification:

ˆ 60E05; 62E99.

1. INTRODUCTION

The probability density function (PDF) of the Lindley distribution [See,

Lindley ([6])] is specified by
θ2
(1.1) fX (x, θ) = (1 + x) exp (−θx), θ > 0, x > 0,
1+θ
and the corresponding cumulative density function (CDF) is given by
1 + θ + θx
(1.2) FX (x, θ) = 1 − exp (−θx), θ > 0, x > 0.
1+θ
A distribution that is close in form to (1.1) is the well-known exponential distri-
bution given by the PDF
fX (x, θ) = θ exp (−θx), θ > 0, x > 0.
Ghitany et al. ([5]) showed that in many ways the Lindley distribution is a
better model than one based on the exponential distribution. The distribu-
tion in (1.1) is a mixture of exponential and gamma distribution with shape
θ 1
parameter 2 and scale parameter θ with mixing proportions 1+θ and 1+θ , re-
spectively. So, it is a Bernoulli mixture of gamma distributions of the form

 1 k  θ 1−k
fX (x; θ) = 1k=0 k1 θ+1
P
θ+1 fGA (x; k + 1, θ), where, fGA (x; k + 1, θ) =
θk+1 k
Γ(k+1) xexp(−θx) is the gamma PDF with shape parameter k + 1 and scale pa-
rameter θ. We generalize this distribution to binomial mixing with parameter r

 1 k  θ r−k
1
of the form fX (x; θ) = rk=0 kr θ+1
P
and θ+1 θ+1 fGA (x; k + 1, θ).

It can be written in more generalized form of the PDF as

Pr
a p h (x; θ)
(1.3) fX (x, θ) = k=0Pr k k k , θ > 0, x > 0,
k=0 ak pk

 1 k  θ r−k θk+1
where, pk = kr θ+1 θ+1 , hk (x; θ)) = Γ(k+1) x(k+1)−1 exp(−θx) and ak ’s
are non-negative constants.
One Parameter Polynomial Exponential Distribution with Binomial Mixture 3

It can also be rewritten as

(1.4) fX (x, θ) = h(θ)p(x) exp (−θx), θ > 0, x > 0,

where, h(θ) = Pr a 1 r 1 , p(x) = rk=0 ak r

xk .
P

k! k
k=0 k (k) θ k+1

A random variable X is said to have a Binomial Mixture One Parameter

Polynomial Exponential (BMOPPE) with parameter θ, if its probability density
function (PDF) is given by

 k  r−k
Pr r 1 θk+1
θ (k+1)−1 exp(−θx)
k=0 ak k θ+1 Γ(k+1) x
θ+1
(1.5) fX (x, θ) =
 1 k  θ r−k ,
Pr r
a
k=0 k k θ+1 θ+1
θ > 0, x > 0.

The CDF of the random variable X is given by

Pr
 k  r−k
r 1
θ
k=0 ak k θ+1
θ+1 γ(k + 1, θx)
(1.6) F (x) =
 1 k  r−k , θ > 0, x > 0,
Pr r θ
k=0 ak k θ+1 θ+1

1
Rt
where γ(s, t) = Γs 0 exp(−x)xs−1 dx is the lower incomplete gamma function.

The CDF can also be written as

Pr r

1 !
k=0 ak k+1 Γ(k + 1, θx)
(1.7) F (x) = 1 − Pr k θ r
1 , x, θ > 0,
k=0 ak k θk+1
R∞
where Γ(m, x) = 1
Γ(m) x e−u um−1 du, the upper incomplete gamma function.

Bouchahed and Zeghdoudi ([4]) proposed a new and unified approach to

generalizing Lindley’s distribution and investigated its properties. Mukherjee et
al. ([10]) later called it the One Parameter Polynomial Exponential (OPPE)
family of distributions and studied the estimation aspect of the PDF and the
CDF of the distribution. The natural discrete version of the OPPE called the
Natural Discrete One Parameter Polynomial Exponential (NDOPPE) family of
distributions is studied by Maiti et al. ([9]), and the estimation aspect of the PMF
and the CDF is discussed by Mukherjee et al. ([11]). The OPPE is a mixture of
gamma distributions with some mixing probabilities. In contrast, the BMOPPE
is revisited with a different look, and it is also a mixture of gamma distributions
with binomial mixing probabilities, which is different from the previous one.
4 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

0.12

BMOPPE BMOPPE BMOPPE

0.10
OPPE OPPE OPPE

0.20
0.10

0.08
0.08

0.15
0.06
PDF

PDF

PDF
0.06

0.10
0.04
0.04

0.05
0.02
0.02
0.00

0.00

0.00
0 5 10 15 20 0 5 10 15 20 25 0 5 10 15

x x x

(a) For r = 3 and θ = 0.5 (b) For r = 4 and θ = 0.5 (c) For r = 4 and θ = 1

Figure 1: Plot of PDF of BMOPPE and OPPE for different combinations

The article is organised as follows. Section 2 discusses different order mo-

ments and stochastic orderings of the random variable. In section 3, The max-
imum likelihood estimator (MLE) and uniformly minimum variance unbiased
estimator (UMVUE) of the PDF and the CDF are discussed. The estimators
are compared in the mean squared error (MSE) sense. This section also consid-
ers the estimation of both mission time and stress-strength reliability functions.
Asymptotic variances of MLEs and variances of MVUEs are derived. UMVUEs
of variances of UMVUE of reliability functions have been derived. Simulation
study results have been reported to verify the theoretical findings in section 4.
Three data sets have been analysed for illustration purposes in section 5. Section
6 is for making some concluding remarks.

2. Moments and Stochastic Orderings

The sth raw moments for BMOPPE distribution is

Pr r

 1 k  θ r−k Γ(k+s+1)
a
k=0 k k θ+1 θ+1 θs Γ(k+1)
(2.1) µ′s =  k  r−k .
Pr r

1 θ
a
k=0 k k θ+1 θ+1

The coefficient of skewness and kurtosis measures have been shown in Figure 2.
These are shown for r = 1, 2, 3 and for different values of θ. The constants of
polynomial are taken as ai = 1, i = 0, 1, 2, 3. It is noticed that the distribution
is positively skewed, and skewness decreases with the increment of the degree of
the polynomial. Also, the distribution is leptokurtic, and it becomes long-tailed
with the increment of the degree of the polynomial.
The ordering relations between two BMOPPE random variables have been shown
in the following Theorem.
One Parameter Polynomial Exponential Distribution with Binomial Mixture 5

2.0

6
5
1.8

4
1.6
Skewness

Kurtosis

3
1.4

2
1.2

1
r=1 r=1
r=2 r=2
1.0

r=3 r=3

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

θ θ

(a) Skewness (b) Kurtosis

Figure 2: Plot of Skewness and Kurtosis of BMOPPE for different θ and

r

Theorem 2.1. Let Xi ∼ BM OP P E(θi ), i = 1, 2 be two random vari-

ables. If θ2 ≤ θ1 , thenX1 ≺lr X2 , X1 ≺hr X2 , X1 ≺st X2 and X1 ≺cx X2

Proof: Since, Likelihood ratio order ⇒ Hazard rate order ⇒ Stochastic

order and Convex order ⇐⇒ Stochastic order, it is sufficient to show that
Likelihood ratio order holds.
We have
 
fX1 (x)
L(x) = ln
fX2 (x)
 
θ1
= (k + 1) ln − (θ1 − θ2 )x
θ2

Now,   
d fX1 (x)
△L(x) = ln
dx fX2 (x)
Clearly, it is evident that △L(x) ≤ 0, ∀θ2 ≤ θ1

3. Estimation of PDF and CDF and their applications in Reliability

Estimation

The PDF and the CDF estimates have immense importance in estimating
reliability functions (both mission time and stress-strength), entropy functions,
6 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

Kullback-Leibler divergence measure, Fisher information, cumulative residual en-

tropy, quantile function, hazard rate function, etc. In this section, the MLE and
UMVUE of the reliability functions are to be attempted. The asymptotic vari-
ances/variances of the estimators and their estimators are to be discussed.

First, we will discuss the MLE and UMVUE of the PDF and the CDF of
BMOPPE family of distributions. Let X1 , X2 , ..., Xn be random sample of size n
drawn from the BMOPPE distribution in (1.5). The MLE of θ which is denoted
as θe is obtained by numerically solving the equation
Pr r k+1


k=0 ak k θk+2
(3.1) Pr r

1 − X̄ = 0.
k=0 ak k θk+1

Using the invariance property of MLE, one can obtain the MLEs of the PDF and
that of the CDF by substituting θe in (1.5) and (1.6) respectively. Theoretical
expressions for the MSE of the MLEs are not available. MSE is to be studied
through simulation.

To derive the UMVUE of the PDF and that of the CDF (stated in Theorem
3.2), we will use the following Theorem 3.1 and Lemma 3.1.

Theorem 3.1. Let X1 , X2 , ..., Xn independently follow BMOPPE(θ).

Then the distribution of T = X1 + X2 + .... + Xn is
XX X
f (t) = hn (θ) ... c(n, y0 , y1 , . . . , yr ) exp(−θt)
y0 y1 yr
Pr
×t k=0 (k+1)yk −1 , t > 0,

Qr r yk
n! k=0 [ak (k)]
with y0 + y1 + ... + yr = n and c(n, y0 , y1 , . . . , yr ) = y0 !y1 !...yr ! Γ( k=0 (k+1)yk ) .
P r

Proof: Since Xi ’s are independent and identically distributed, the mo-

ment generating function (mgf) of T is
" r #n
X r  1  t −(k+1)

n
MT (t) = h (θ) ak 1−
k θk+1 θ
k=0
"   
t −1
  
n r 1 r 1
= h (θ) a0 1− + .... + ar r+1
0 θ θ r θ
 −(r+1) #n
t
× 1−
θ
r   yk
n
XX X n! Y r
= h (θ) ... ak
y0 y1 yr
y0 !y1 ! . . . yr ! k
k=0
P r
t − k=0 (k+1)yk
Pr
 
×θ− k=0 (k+1)yk 1 − .
θ
One Parameter Polynomial Exponential Distribution with Binomial Mixture 7

Hence, the distribution of T is

r   yk
n
XX X n! Y r
f (t) = h (θ) ... ak
y0 y1 yr
y0 !y1 ! . . . yr ! k
k=0
Pr r
X
×θ− k=0 (k+1)yk f
GA (t, (k + 1)yk , θ)
k=0
XX X Pr
n
= h (θ) ... c(n, y0 , y1 ...., yr )t k=0 (k+1)yk −1 exp(−θt).
y0 y1 yr

Lemma 3.1. The conditional distribution of X1 given T = X1 + X2 +

.... + Xn is
p(x) X X X
fX1 |T (x|t) = ... c(n − 1, q0 , q1 , . . . , qr )
An (t) q q q
0 1 r
Pr
×(t − x) k=0 (k+1)qk −1 , 0 < x < t,

where
XX X Pr
An (t) = ... c(n, y0 , y1 , . . . , yr )t k=0 (k+1)yk −1 ,
y0 y1 yr

and
r   qk
(n − 1)! Y r 1
c(n − 1, q0 , q1 ...., qr ) = ak Pr ,
q0 !q1 ! . . . qr ! k Γ( k=0 (k + 1)qk )
k=0

with q0 + q1 + q2 + .... + qr = n − 1.

Proof: The proof is obviously conducted from

fX1 (x)f (t − x)
fX1 |T (x|t) =
f (t)
p(x) X X X
= ... c(n − 1, q0 , q1 , . . . , qr )
An (t) q q q
0 1 r
Pr
×(t − x) k=0 (k+1)qk −1 .

Theorem 3.2. Let T = t be given. Then

p(x) X X X
(3.2) fb(x) = ... c(n − 1, q0 , q1 , . . . , qr )
An (t) q q q 0 1 r
Pr
×(t − x) k=0 (k+1)qk −1 , 0 < x < t,
8 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

is UMVUE for f (x) and

1 XX X
(3.3) Fb(x) = 1 − ... c(n − 1, q0 , q1 , . . . , qr )
An (t) q q q
0 1 r
r  
X r 1 Pr
× ak t k=0 (k+1)qk +k
k Γ(k + 1)
k=0
r r
! !
X X
×B (k + 1), (k + 1)yk Ix/t (k + 1), (k + 1)qk ,
k=0 k=0
0 < x < t,
1
R1
is UMVUE for F (x), where Ix (α, β) = B(α,β) x uα−1 (1 − u)β−1 du is an incom-
Γ(α)Γ(β)
plete beta function and B(α, β) = Γ(α+β) .

Proof: The
P BMOPPE in (1.4) is a member of one parameter exponential
family with T = ni=1 Xi as complete sufficient statistic.Therefore, by the use of
Lehmann-Scheffe theorem, we get the UMVUE of the PDF from Lemma 3.1.
Z t
F (x) = 1 −
b fb(w)dw
x
Z t
p(w) X X X
= 1− ... c(n − 1, q0 , q1 , . . . , qr )
x An (t) q q 0 q1 r
Pr
×(t − w) k=0 (k+1)qk −1
dw
1 XX X
= 1− ... c(n − 1, q0 , q1 , . . . , qr )
An (t) q q qr
0 1
r  
X r 1 Pr
× ak t k=0 (k+1)qk +k
k Γ(k + 1)
k=0
r r
! !
X X
×B (k + 1), (k + 1)yk Ix/t (k + 1), (k + 1)qk .
k=0 k=0

3.1. Mission Time Reliability

Suppose the life length of a component X is distributed as BMOPPE (θ).

Then the reliability of that component for a fixed mission time t0 (> 0) is

F̄X (t0 ) = P (X ≥ t0 )
r
ak kr


X
= h(θ) Γ(k + 1, θt0 ).
θk+1
k=0
One Parameter Polynomial Exponential Distribution with Binomial Mixture 9

By using the relation between incomplete gamma function and Poisson probabil-
ity, F̄X (t0 ) can be written as
r r

k
X ak k
X e−θt0 (θt0 )j
F̄X (t0 ) = h(θ) .
θk+1 j!
k=0 j=0

3.1.1. The MLE

The MLE of F̄X (t0 ) based on a random sample of size n, is

r r

k
e−θt0 (θt
e e j
X ak k
X 0)
F
ē (t ) = h(θ)
X 0
e ,
θek+1 j!
k=0 j=0

where θe is the solution of the equation (3.1). Since M SE(F ē (t )) has no closed
X 0
form expression, we give the asymptotic distribution of F X (t0 ) by using delta
ē
theorem as follows
2 !
√  ē


d 1 dF̄X (t0 )
n F X (t0 ) − F̄X (t0 ) → N 0, ,
I(θ) dθ

where
    2
Pr ak (kr ) Pr ak (kr )(k+2)(k+1) Pr ak (kr )(k+1)
k=0 θk+1 k=0 θk+3
− k=0 θk+2
I(θ) =  2
Pr ak (kr )
k=0 θk+1

is the Fisher information about parameter θ for a single observation X.

3.1.2. The UMVUE

Application of incomplete beta function and binomial probability gives the

UMVUE of F̄X (t0 ) as
1 XX X
F
b̄ (t ) =
X 0 ... c(n − 1, x0 , x1 , . . . , xr ) ×
An (t) x x x Pr
0 1 r xi =n−1
i=0
r r
!
ak kr

Pr
(k+1)x +k
X X
k
t k=0 B (k + 1),× (k + 1)xk
Γ(k + 1) 0
k=0 k=0
k Pr   j  Prl=0 (l+1)xl +k−j
X
l=0 (l + 1)x l + k t 0 t0
1− .
j t t
j=0
10 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

For exponential family of distribution, Blight and Rao ([3]) considered

Bhattacharya bounds to calculate the variance of UMVUE of parametric function
ψ(θ). So, variance of F
b̄ (t ) can be expressed as
X 0
∞ (l)
X [F̄ (t0 )]2
X
V ar(F
b̄ (t )) =
X 0 ,
[Jl∗ (θ)]2
l=1

where
l X
l   
X l l
[Jl∗ (θ)]2 = [h(θ)] n
(−1) i+j
×
i j
j=0 i=0

[h(l−i) (θ)]n [h(l−j) (θ)]n

XX X
× ... c(n − 1, x0 , x1 , ..., xr )
x0 x1 xr Pr xi =n−1
i=0
Pr
Γ(i + j + l=0 (l + 1)xl )
× Pr
θi+j+ l=0 (l+1)xl
is the Bhattacharya function and h(i) (θ) denotes the ith derivative of h with re-
spect to θ.

Now, for the derivation of UMVUE of V ar(F

b̄ (t )), we consider the rep-
X 0
resentation

V ar(F b̄ 2 (t )) − [E(F
b̄ (t )) = E(F b̄ (t ))]2
X 0 X 0 X 0
2
b̄ (t )) − F̄ 2 (t ).
= E(F X 0 X 0

b 1 (t0 ) is the UMVUE of Q1 (t0 ) = F̄ 2 (t0 ), we get the UMVUE of V ar(F

If Q b̄ (t ))
X 0
X
2
as F (t0 ) − Q
b̄
X
b 1 (t0 ). We start from the fact that

I[X1 ≥ t0 , X2 ≥ t0 ] = I[X1 ≥ t0 ]I[X2 ≥ t0 ],

which implies that I[X1 ≥ t0 , X2 ≥ t0 ] is unbiased for Q1 (t0 ). Then, we get the
UMVUE by using Rao-Blackwell theorem as
r
1 XX X X
Q
b 1 (t0 ) = ... c(n − 2, x0 , x1 , . . . , xr )
An (t) x x x Pr
0 1 r xi =n−2 k=0
i=0
Pr
m=0 (m+1)xm −1
Pr
r
ak al kr rl (−1)i m=0 (m+1)xm −1




X X
i
×
Γ(k + 1)Γ(l + 1) l+i+1
l=0 i=0
Pr
m=0 (m+1)xm +l
Pr
(−1)u m=0 (m+1)x m +l


Pr X
×[t m=0 (m+1)xm +l u
tu (k + u + 1)
u=0
Pr
Pr m=0 (m+1)xm −i−1
X
×(t k+u+1
− t0k+u+1 ) − tl+i+1 t m=0 (m+1)xm −i−1 ×
0
v=0
Pr
(−1)v m=0 (m+1)xm −i−1


v
(tk+v+1 − tk+v+1
0 )].
tv (k + v + 1)
One Parameter Polynomial Exponential Distribution with Binomial Mixture 11

3.2. Stress Strength Reliability

For estimation of stress strength reliability, we assume the stress ran-

dom variable X follows BMOPPE (θ1 ) and strength random variable Y follows
BMOPPE (θ2 ) and they are independently distributed. Then the expression of
stress strength reliability becomes

R = P (X < Y )
Z ∞
= F̄Y (x)f (x)dx
0
r2 r2 Xr1 r1



X bk2 k2 ak1 k1
= h(θ1 )h(θ2 ) ×
θk2 +1 k1 =0
k2 =0 2
θ1k1 +1
k2   j  k1 +1
X k1 + j θ2 θ1
.
j θ1 + θ2 θ 1 + θ2
j=0

3.2.1. The MLE of R

Let (x1 , x2 , ..., xn1 ) and (y1 , y2 , ..., yn2 ) are independent samples drawn from
BMOPPE (θ1 ) and BMOPPE (θ2 ), respectively. Let the MLEs of θ1 and θ2 are
θe1 and θe2 , respectively. By putting the values of θe1 and θe2 in the expression of
R, we get R e by its invariance property as,
r2
bk2 kr22 X ak1 kr11

r1

X
Re = h(θe1 )h(θe2 ) ×
k2 =0 θ2
e k2 +1 k1 =0 θe1 k1 +1
k2   !j !k1 +1
X k1 + j θe2 θe1
.
j θe1 + θe2 θe1 + θe2
j=0

Similarlly as in mission time reliability, we derive the asymptotic distribution of

R
e in the following theorem.

Theorem 3.3. If the ratio nn12 converges to a positive number κ when

both n1 and n2 tends to infinity. Then
√ d
e − R) → N (0, σ 2 ),
n(R
1
where σ 2 = I1 (θ [ ∂R ]2 + I2 (θ
1 ) ∂θ1
κ
[ ∂R ]2 and I1 (θ1 ) is a Fisher information about
2 ) ∂θ2
parameter θ1 in a single observation X, I2 (θ2 ) is a fisher information about θ2 in
a single observation Y.
12 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

Proof: Since log-likelihood equations satisfies all regularity conditions

of asymptotic normality for MLE, then we have
√ d
nj (θej − θj ) → N (0, [Ij (θj )]−1 ),
where
    2
Pr1 ak (rk1 ) Pr1 ak (rk1 )(k+2)(k+1) Pr1 ak (rk1 )(k+1)
k=0 θk+1 k=0 − k=0
1 θ1k+3 θ1k+2
I1 (θ1 ) =  2 ,
Pr1 ak (rk1 )
k=0 θk+1
1

and
    2
Pr2 bk (rk2 ) Pr2 bk (rk2 )(k+2)(k+1) Pr2 bk (rk2 )(k+1)
k=0 θk+1 k=0 − k=0
2 θ2k+3 θ2k+2
I1 (θ2 ) =  2 .
Pr2 bk (rk2 )
k=0 θk+1
2

Again, from the fact of independence of θe1 and θe2 , we get

√ d
n1 (θe1 − θ1 , θe2 − θ2 ) → N2 (0, J(θ1 , θ2 )),
where
[I1 (θ1 )]−1
 
0
J(θ1 , θ2 ) = .
0 κ[I2 (θ2 )]−1
Now application of the transformation R = R(θ1 , θ2 ) together with Delta theorem
conclude the proof.

3.2.2. The UMVUE of R

By using the UMVUE of mission time reliability and the PDF of the
BMOPPE distribution, the UMVUE of R can be expressed as
Z M in(t1 ,t2 )
R =
b F
b̄ (x)fb (x)dx,
Y X
x=0

where, t1 = i=1 xi and t2 = ni=1

Pn1 P 2
yi , respectively. Evalution of the integral
gives the final form of the UMVUE of R as
1 XX X
Rb = ... c(n1 − 1, q0 , q1 , . . . , qr1 )
An1 (t1 )An2 (t2 ) q q q r
0 P 1 r1 1
i=0 qi =n1 −1
Pr1
(l+1)ql X X X
×t1 l=0
... c(n2 − 1, w0 , w1 , . . . , wr2 )
w0 w1 wr2 Pr2
i=0 wi =n2 −1
r1 r2
!
Pr2
m=0 (m+1)wm
X X
×t2 J1 (l + 1)ql , (m + 1)wm ,
l=0 m=0
One Parameter Polynomial Exponential Distribution with Binomial Mixture 13

where
k2 β+k2 −j
r1 r1
tk11 X
r2 r2
tk22 X


i+j
ak1 bk2

X k1 k2
X
i t1
J1 (α, β) = (−1) ×
Γ(k1 + 1) Γ(k2 + 1) t2
k1 =0 k2 =0 j=0 i=0
  
β + k2 β + k2 − j
B(k2 + 1, β)B(i + j + k1 + 1, α) ×
j i
α+i+j+k
X 1 α + i + j + k1   M in(t1 , t2 ) l
×
l t1
l=i+j+k1 +1

M in(t1 , t2 ) (α+i+j+k1 −l)

 
1− .
t1

Under certain regularity conditions, variance of R

b takes the form [See,
Bartoszewicz([1])] as follows

∞ X
k
" #2
X ∂kR 1
V ar(R)
b = ,
k=1 j=0 ∂θ1j ∂θ2k−j Ij (θ1 )Jk−j (θ2 )

where
l X
l   
2 n1
X
i+j l l
[Il (θ1 )] = [h(θ1 )] (−1) ×
i j
j=0 i=0

[h(l−i) (θ1 )]n1 [h(l−j) (θ1 )]n1

XX X
× ... c(n1 − 1, q0 , q1 , ..., qr1 )
q0 q1 qr1 Pr1
i=0 qi =n1 −1
Pr1
Γ(i + j + l=0 (l + 1)ql )
× Pr1 ,
i+j+ l=0 (l+1)ql
θ1

l X
l   
2 n2
X
i+j l l
[Jl (θ2 )] = [h(θ2 )] (−1) ×
i j
j=0 i=0

[h(l−i) (θ2 )]n2 [h(l−j) (θ2 )]n2

XX X
× ... c(n2 − 1, w0 , w1 , ..., wr2 )
w0 w1 wr2 Pr2
i=0 wi =n2 −1
Pr2
Γ(i + j + l=0 (l + 1)wl )
× Pr2 .
i+j+ l=0 (l+1)wl
θ2

As earlier, the following representation

V ar(R)
b = E(R b 2
c2 ) − [E(R)]
c2 ) − R2 .
= E(R
14 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

gives the UMVUE of V ar(R) c2 − Q

b as R b 2 , where Qb 2 is the UMVUE of Q2 = R2 .
Similarly, we give the final expression of Qb 2 in the following equation.

Q
b2
1 XX X
= ... c(n1 − 2, q0 , q1 , . . . , qr1 ) ×
An1 (t1 )An2 (t2 ) q q q Pr0 1 r1 1
i=0 qi =n1 −2
r1 X
r1
XX X X ak1 al1
... c(n2 − 1, w0 , w1 , . . . , wr2 )
w0 w1 wr2 Pr2
Γ(k1 + 1)
wi =n2 −1 k1 =0 l1 =0
i=0
Pr2
r1 r1 r2 X
r2 r2 r2 m=0 (m+1)wm −1





k1 l1
X bk2 bl2 k2 l2
X
× (−1)i ×
Γ(l1 + 1) Γ(k2 + 1) Γ(l2 + 1)
k2 =0 l2 =0 i=0
Pr2 Pr2 Pr2
m=0 (m+1)wm −1 m=0 (m+1)wm +l2
(−1)u m=0 (m+1)w m +l2

Pr2


m=0 (m+1)wm +l2
X
i u
[t2
l2 + i + 1 tu2 (k2 + u + 1)
u=0
r1
X
×[tu2 J2 (k1 , l1 , (m + 1)qm − 1) − J2 (k1 + k1 + u + 1, l1 ,
m=0
Pr2
r1 Pr2 m=0 (m+1)wm −i−1
m=0 (m+1)wm −i−1
X X
(m + 1)qm − 1)] − t2 (−1)v ×
m=0 v=0
Pr2
m=0 (m+1)wm −i−1

r1
X
v
[tk22 +v+1 J2 (k1 , l1 + l2 + i + 1, (m + 1)qm − 1)
tv2 (k2 + v + 1)
m=0
r1
X
−J2 (k1 + k2 + v + 1, l1 + l2 + i + 1, (m + 1)qm − 1)]],
m=0

where
β−1
β − 1 ta+b+β+1
X  
J2 (a, b, β) = (−1)i 1
B(a + 1, b + β + 1) ×
i b+β+1
i=0
a+b+β+1
X a + b + β + 1  M in(t1 , t2 ) j
×
j t1
j=a+1

M in(t1 , t2 ) a+b+β+1−j
 
1− .
t1

4. Simulation study

Monte Carlo Simulation technique will not be helpful in generating random

samples from the BMOPPE distribution, since the equation

F (x) = u, u ∈ (0, 1)
One Parameter Polynomial Exponential Distribution with Binomial Mixture 15

cannot explicitly be solved in x. However, since the distribution is a mixture of

gamma distributions given in (1.5), one can utilize this fact. For the BMOPPE
distribution, the generation of a random sample X1 , X2 , . . . , Xn is made using
the following algorithm:

1. Generate Ui ∼ U nif orm(0, 1), i = 1(1)n.

Pj−1 Pj
ak (kr )
1
ak (kr )
1
k=0 θk k=0 θk
2. If Pr r 1 < Ui ≤ , j = 1, 2, ..., r,then set Xi = Vi ,
Pr r 1
k=0 ()ak k k
θ
() k=0 ak k θ k
where Vi ∼ gamma(j + 1, θ) and if Ui ≤ Pr aa0 r 1 , then set Xi =
k=0 k (k) θ k
Vi , where Vi ∼ exp(θ).
16 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

0.0005 0.0010 0.0015 0.0020 0.0025 0.0030

0.10
mse_MLE mse_MLE
mse_UMVUE mse_UMVUE

0.08
0.06
MSE_F(x)
MSE_f(x)

0.04
0.02
0.00
5 10 15 20 5 10 15 20

Sample Size Sample Size

(a) For PDF at θ = 1.05, x = 2, a0 = 0, a1 = (b) For CDF at θ = 1.05, x = 2, a0 = 0, a1 =

0.1, a2 = 1 and r = 2 0.1, a2 = 1 and r = 2

0.10
mse_MLE mse_MLE
mse_UMVUE mse_UMVUE
0.0018

0.08
0.0014

0.06
MSE_F(x)
MSE_f(x)

0.04
0.0010

0.02
0.0006

0.00

5 10 15 20 5 10 15 20

Sample Size Sample Size

(c) For θ = 1.5, x = 2, a0 = 0.01, a1 = (d) For θ = 1.5, x = 2, a0 = 0.01, a1 =

0.02, a2 = 4 and r = 2 0.02, a2 = 4 and r = 2
1e−05
8e−06

mse_MLE mse_MLE
mse_UMVUE mse_UMVUE
8e−06
6e−06

6e−06
MSE_F(x)
MSE_f(x)

4e−06

4e−06
2e−06
2e−06

0e+00

5 10 15 20 5 10 15 20

Sample Size Sample Size

(e) For θ = 0.09, x = 2, a0 = 0.2, a1 = (f) For θ = 0.09, x = 2, a0 = 0.2, a1 =

0.1, a2 = 1 and r = 2 0.1, a2 = 1 and r = 2

Figure 3: Graph of simulated MSE of the MLE and UMVUE of the PDF
and the CDF of BMOPPE distribution

5. Data Analysis

In this section, we analyse three real data sets for comparing the perfor-
mances of the MLE and the UMVUE of the PDF and the CDF. The probability
model selection is made based on Akaike’s Information Criterion (AIC= −2 log-
One Parameter Polynomial Exponential Distribution with Binomial Mixture 17

likelihood+2k, where k is the number of parameters involved in the model) for

each data set. Minimum is the AIC; better is the model fit. For our BMOPPE
model, there is only one parameter θ; it is sufficient to select the model with a
negative log-likelihood. But since this model is compared with other models with
more than one parameter, the analyses are based on AIC. The AIC is calculated
first using the MLE of the parameter(s), and the best model is selected. Then
the AIC of the chosen model is compared with the AIC of this model calculated
using the UMVUE of the PDF. It is mentioned that a0 , a1 , ..., ar and r are
known constants in the model; in practice, these are chosen by trial and error
such that AIC is at its minimum.

Data set-I, which is cited from Gross and Clark ([7]) and is given in Table
1, represents the relief times (in minutes) of 20 patients receiving an analgesic.
We calculate the AIC of different standard distributions of One Parameter Poly-
nomial Exponential (OPPE) family and few others from literature (presented in
Table 4) and it is observed that BMOPPE family with r = 2, a0 = 0, a1 = 0.1
and a2 = 1 is a better fit. The negative log-likelihood values of the selected model
calculated using the MLE and the UMVUE of the PDF are presented in Table
7. Figures 3(a)- (b) show the histogram, the estimated PDF, and the CDF.

Data set-II represents the number of million revolutions before failure for
each of the 23 ball bearings in the life test. It is obtained from Lawless ([8]) and
shown in Table 2. For ease of calculation, we divide each observation into the
data set by 2. The calculated AIC of different standard distributions of OPPE
family and few others have been shown in Table 5 and it is noticed that BMOPPE
family with r = 2, a0 = 0.2, a1 = 0.1 and a2 = 1 is a better fit. The negative
log-likelihood values of the selected model calculated using the MLE and the
UMVUE of the PDF are also presented in Table 7. Figures 3(c) and 3(d) present
the corresponding histogram, the estimated PDF, and the CDF.

Data set-III is a collection from Bjerkedal ([2]), and Table 3 displays the
survival times (in days) of 72 guinea pigs infected with virulent tubercle bacilli.
This data set has been fitted with a distribution of BMOPPE family with r = 2,
a0 = 0.01, a1 = 0.02 and a2 = 4 and it is found to be a good fit. AIC for this
distribution and some other distributions available in the literature are listed in
Table 6 that supports our claim. The negative log-likelihood values of the selected
model calculated using the MLE and the UMVUE of the PDF are also presented
in Table 7. The histogram, the estimated PDF, and the CDF have been shown
in Figures 3(e) and 3(f).

6. Concluding Remarks

The article searches for a more generalised version of Lindley distribution.

Starting with the Lindley distribution as a Bernoulli mixture of gamma distribu-
tions, a generalised binomial mixture of gamma distributions called the BMOPPE
18 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

family of distributions has been derived. It is a revisit of the Lindley distribution

from a different angle. As a result, the generalised version of the Lindley, the
OPPE family of distributions, got mixed with the binomial probabilities, and
therefore, the BMOPPE is an improvement. Moments and stochastic orderings
are discussed. The process of generation of observations is pointed out, and the
results are summarised. Estimations of the PDF and the CDF are discussed. The
MLEs and UMVUEs are derived and compared. We have the estimators in biased
(i.e., MLE) and unbiased (i.e., UMVUE) classes. Estimators of reliability func-
tions are derived. Asymptotic variances of MLEs and variances of UMVUEs have
been derived. The UMVUEs of the variance of UMVUEs of reliability functions
have also been derived. These may be helpful in data analysis and comparison.
Few data sets have been fitted, and it is found that the proposed distribution
fits well in AIC sense. Even though the gain in AIC is minimal compared to the
OPPE family of distributions, the BMOPPE is an improvement and is preferred.

1.1 1.4 1.3 1.7 1.9 1.8 1.6 2.2 1.7 2.7
4.1 1.8 1.5 1.2 1.4 3 1.7 2.3 1.6 2

Table 1: Relief times (in minutes) of 20 patients receiving an analgesic

17.88 28.92 33.00 41.52 42.12 45.60 48.80 51.84

51.96 54.12 55.56 67.80 68.44 68.64 68.88 84.12
93.12 98.64 105.12 105.84 127.92 128.04 173.40

Table 2: The number of million revolutions before failure for each of the
23 ball bearings in the life tests

0.1 0.33 0.44 0.56 0.59 0.72 0.74 0.77 0.92

0.93 0.96 1 1 1.02 1.05 1.07 1.07 1.08
1.08 1.08 1.09 1.12 1.13 1.15 1.16 1.2 1.21
1.22 1.22 1.24 1.3 1.34 1.36 1.39 1.44 1.46
1.53 1.59 1.6 1.63 1.63 1.68 1.71 1.72 1.76
1.83 1.95 1.96 1.97 2.02 2.13 2.15 2.16 2.22
2.3 2.31 2.4 2.45 2.51 2.53 2.54 2.54 2.78
2.93 3.27 3.42 3.47 3.61 4.02 4.32 4.58 5.55

Table 3: Survival times (in days) of 72 guinea pigs

One Parameter Polynomial Exponential Distribution with Binomial Mixture 19

Model −2 log-likelihood value AIC

BMOPPE (a0 = 0, a1 = 0.1 and a2 = 1) 48.10 50.10
Length-biased Lindley 49.70 51.70
Akash 59.52 61.52
Shanker 59.78 61.78
Lindley 60.50 62.50
Moment Exponential 52.32 54.32
Exponential 65.67 67.67

Table 4: Model selection criterion for data set-I

Model −2 log-likelihood value AIC

BMOPPE (a0 = 0.2, a1 = 0.1 and a2 = 1) 195.26 197.26
Sujatha 195.38 197.38
Akash 227.06 229.06
Shanker 231.06 233.06
Lindley 231.47 233.47
Gamma 226.04 230.04
Weibull 232.27 236.27
Exponential 242.87 244.87

Table 5: Model selection criterion for data set-II

Model −2 log-likelihood value AIC

BMOPPE (a0 = 0.01, a1 = 0.02 and a2 = 4) 188.18 190.18
OPPE (a0 = 0.01, a1 = 0.02 and a2 = 4) 188.27 190.27
Lindley 213.85 215.85
New Generalized Lindley 188.36 194.36
Moment Exponential 208.40 210.40
Marshall-Olkin Exponential 206.36 210.36

Table 6: Model selection criterion for data set-III

Negative log-likelihood value

Data Set Model MLE UMVUE
I BMOPPE (a0 = 0, a1 = 0.1 and a2 = 1) 24.05 23.71
II BMOPPE (a0 = 0.2, a1 = 0.1 and a2 = 1) 97.63 97.52
III BMOPPE (a0 = 0.01, a1 = 0.02 and a2 = 4) 94.09 94.09

Table 7: Negative log-likelihood value using MLE and UMVUE fitted in

data set I-III
20 M. K. Ruidas, I. Mukherjee, M. M. Choudhury, S. S. Maiti and S. Adhya

1.0
f(x).MLE
f(x).UMVUE
0.8

0.8
0.6

0.6
Density

F(x)
0.4

0.4
0.2

0.2
F(x).MLE
0.0

0.0
F(x).UMVUE

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 0 1 2 3 4 5 6

x x
Histogram and estimated PDF fitted to the data set−I Estimated CDF fitted to the data set−I

(a) Fitted PDF at a0 = 0.2, a1 = 0.1 and (b) Fitted CDF at a0 = 0.2, a1 = 0.1 and
a2 = 1 to the data set-I a2 = 1 to the data set-I

1.0

f(x).MLE
f(x).UMVUE
0.030

0.8
0.6
0.020
Density

F(x)

0.4
0.010

0.2

F(x).MLE
0.000

0.0

F(x).UMVUE

0 20 40 60 80 0 20 40 60 80

x x
Histogram and estimated PDF fitted to the data set−II Estimated CDF fitted to the data set−II

(c) Fitted PDF at a0 = 0, a1 = 0.1 and a2 = (d) Fitted CDF at a0 = 0, a1 = 0.1 and a2 =
1 to the data set-II 1 to the data set-II
1e−05
8e−06

mse_MLE mse_MLE
mse_UMVUE mse_UMVUE
8e−06
6e−06

6e−06
MSE_F(x)
MSE_f(x)

4e−06

4e−06
2e−06
2e−06

0e+00

5 10 15 20 5 10 15 20

Sample Size Sample Size

(e) Fitted PDF at a0 = 0.01, a1 = 0.02 and (f) Fitted CDF at a0 = 0.01, a1 = 0.02 and
a2 = 4 to the data set-III a2 = 4 to the data set-III

Figure 4: Graph of the estimated PDF and CDF of BMOPPE distribution

for different data sets.
One Parameter Polynomial Exponential Distribution with Binomial Mixture 21

ACKNOWLEDGMENTS

The authors are thankful to the Editor, Associate Editor and two anony-
mous referees for their valuable comments that substantially improve the previous
version of the paper.

REFERENCES

[1] Bartoszewicz, J. (1980). On the convergence of the Bhattacharya bounds in the

multiparameter case. Zastosowania Mathematica, 16, 601–608.
[2] Bjerkedal. (1960). Acquisition of resistance in guinea pigs infected with different
doses of virulent tubercle bacilli. American Journal of Hygiene, 72, (1), 130–148.
[3] Blight, B. J. N. and Rao, P. V. (1974). The convergence of the Bhattacharya
bounds. Biometrika, 61, 137–142.
[4] Bouchahed, L. and Zeghdoudi, H. (2018). A new and unified approach in
generalizing the Lindley’s distribution with applications. Statistics in Transition,
19(1), 61–64.
[5] Ghitany, E. M., Al-Mutairi, D. K., Balakrishnan, N. and Al-Enezi,
J. L. (2013). Power Lindley distribution and associated inference. Computational
Statistics & Data Analysis, 64, 20–33.
[6] Lindley, D. V. (1958). Fiducial distributions and Bayes’ theorem. Journal of the
Royal Statistical Society. Series B (Methodological), 20, (1), 102–107.
[7] Gross, A.J. and Clark, V.A. (1975). Survival distributions: Reliability Appli-
cations in the Biometrical Sciences. John Wiley, New York, USA.
[8] Lawless, J.F. (1982). Statistical models and methods for lifetime data. John Wiley
& Sons, Inc., Hoboken, New jersey.
[9] Maiti, S. S., Ruidas, M. K. and Adhya, S. (2020). Natural Discrete One
Parameter Polynomial Exponential Family of Distributions and the Application.
Journal of Data Science, To appear.
[10] Mukherjee, I., Maiti, S. S. and Singh, V. V. (2021a). On estima-
tion of the PDF and the CDF of the one-parameter polynomial exponen-
tial family of distributions. Communications in Statistics—Theory and Methods,
https://doi.org/10.1080/03610926.2021.1910302.
[11] Mukherjee, I., Maiti, S. S. and Shanker, R. (2021b). On estimation of the
probability mass function and the cumulative distribution function of a natural
discrete one parameter polynomial exponential distribution. International Journal
of Applied Mathematics and Statistics, 60, (3), 49–61.