1

Statistics and Applications {ISSN 2454-7395 (online)}

Volume xx, No. x, yyyy (New Series), pp 1–22
http://www.ssca.org.in/journal

2 On the Bivariate Generalized Chen Distribution

1
3 Author1, 2 Author2, 3 Author3, 3 Author4 and 4 Author5
1
4 Department of Mathematical Statistics
5 Faculty of Graduate Studies for Statistical Research, Cairo University, Egypt
2
6 Formerly at Indian Agricultural Statistics Research Institute, New Delhi 110012
3
7 Department of Statistics, Pondicherry University, Puducherry
4
8 National Institute Medical Statistics, ICMR, New Delhi - 110029
9 Received: dd Month yyyy; Revised: dd Month yyyy; Accepted: dd Month yyyy
10

11 Abstract
12 A new model of bivariate distributions is presented in this paper. The model intro-
13 duced here is of the Marshall–Olkin type. The joint survival function, the joint probability
14 density function and the joint hazard function of the bivariate generalized Chen (BGCh)
15 distribution are obtained. The maximum likelihood and Bayesian methods are used to esti-
16 mate the unknown parameters of the BGCh distribution. Numerical methods are required
17 to calculate the desired estimates.
18 Key words: Marshall–Olkin type; Generalized Chen (BGCh) distribution; The maximum
19 likelihood and Bayesian methods; MCMC.
20 AMS Subject Classifications: 62K05, 05B05
21

22 1. Introduction
23 A suitable parametric model is often of interest in the analysis of survival data, as it
24 provides insight into the characteristics of the failure times and hazard functions that may
25 not be available with non-parametric methods. The Weibull distribution is one of the most
26 commonly used families for modeling such data. However, only monotonically increasing
27 and decreasing hazard functions can be generated from the classic two-parameter Weibull
28 distribution. As such this two-parameter model is inadequate when the true hazard shape is
29 of bathtub nature. Models with bathtub-shaped hazard rate are needed in reliability analysis
30 and decision making when the complete life cycle of the system is to be modeled. Many
31 authors have proposed models with bathtub-shaped failure rates. For example, Smith and
32 Bain (1975) proposed the exponential power distribution. Mudholkar and Srivastava (1993)
33 suggested the exponentiated Weibull distribution. Chen (2000) provided a two-parameter
34 lifetime distribution with bathtub shape or increasing failure function, now known as Chen
35 distribution. Xie et al. (2002) modified the Chen distribution to include a scale parameter
36 named modified Weibull extension and also referred to as the generalized Chen distribution.

Corresponding Author: Author Name

Email: author email
 2 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

37 They discussed the parameters’ estimation using maximum likelihood method. For more
38 generalizations and modifications of Weibull distribution, see Murthy et al. (2004) and Pham
39 and Lai (2007).
40 Bivariate lifetime data arise frequently in many practical problems and in these sit-
41 uations it is important to consider different bivariate models that could be used to model
42 such bivariate lifetime data. There are a number of papers dealing with bivariate models
43 of type of Marshal-Olkin. For example, Sarhan and Balakrishnan (2007) introduced a bi-
44 variate distribution using exponential and generalized exponential distributions, now known
45 as Sarhan-Balakrishnan bivariate (SBBV) distribution. Although, they derived several in-
46 teresting properties of this distribution, the marginal distributions of SBBV distribution
47 are not in known forms. Kundu (2012) modified the SBBV distribution to include a scale
48 parameter and discussed the estimation of parameters using maximum likelihood method.
49 Kundu and Gupta followed the idea using the generalized exponential to introduce the bi-
50 variate generalized exponential (BVGE) distribution so that the marginal distributions are
51 generalized exponential distributions. They derived several interesting properties of this
52 distribution and discussed the maximum likelihood estimation of the unknown parameters.
53 Also, they re-analyzed a real data set that was analyzed by Meintanis (2007) and concluded
54 that the BVGE distribution provides a better fit than the bivariate Marshall-Olkin distri-
55 bution. Sarhan (2019) noted that none of the marginal distributions of the SBBV and the
56 BVGE provide a bathtub shape of the hazard function and this lack of the bathtub property
57 limits the application of these distributions. Thus he introduced a new bivariate distribution
58 named the bivariate generalized Rayleigh (BVGR) distribution. The BVGR distribution has
59 generalized Rayleigh marginal distributions. The hazard rate functions of the marginals of
60 the BVGR can be either increasing or decreasing or bathtub shaped, and with this prop-
61 erty the BVGR distribution has wider applicability than other distributions. Sarhan (2019)
62 investigated several interesting properties of this distribution and estimated the unknown
63 parameters by using the maximum likelihood and Bayes methods. Many authors discussed
64 the Marshal-Olkin idea for different distributions; see for example; El-Gohary et al. (2016),
65 Kundu and Gupta (2017) Azizi et al. (2020), Muhammed (2019) and others.
66 Using the idea of Marshal-Olkin, we propose a new bivariate generalized Chen (BGCh)
67 distribution. The BGCh distribution has generalized Chen marginal distributions. The joint
68 survival function, the joint probability density function and the joint hazard function of the
69 BGCh distribution are obtained. The maximum likelihood and Bayesian methods are used
70 to estimate the unknown parameters of the BGCh distribution. Numerical methods are
71 required to calculate these estimates.

72 2. The bivariate generalized Chen distribution

73 In this section, we define a new bivariate distribution, shortly denoted by BGCh. We
74 start with the joint survival function of the distribution and then we derive the corresponding
75 joint probability density function.
 yyyy] SHORT RUNNING TITLE IN CAPITALS 3

76 2.1. The joint survival function

77 Chen (2000) introduced a two-parameter lifetime distribution with either bathtub-
78 shaped or increasing failure rate with the survival function
β
SCh (t) = exp(λ(1 − e(t) )), t ≥ 0, λ and β > 0.

79 and the corresponding probability density function

β
fCh (t) = λβ(t)β−1 exp((t)β + λ(1 − e(t) )), t ≥ 0, λ and β > 0.

80 Xie et al. (2002) modified the Chen distribution to include a scale parameter named the
81 generalized Chen distribution. The survival function of the univariate generalized Chen
82 (GCh) distribution is
β
SGCh (t) = exp(λα(1 − e(t/α) )), t ≥ 0, λ, α and β > 0. (1)

83 with probability density function (pdf)

β
fGCh (t) = λβ(t/α)β−1 exp((t/α)β + λα(1 − e(t/α) )), t ≥ 0, λ, α and β > 0. (2)

84 Now, suppose that Tj , j = 1, 2, 3 are independent random variables with Ti having GCh
85 distributions with scale parameters α, and λj , j = 1, 2, 3 and shape parameter β; i.e. Ti ∼
86 GCh(α, β, λj ), j = 1, 2, 3. Define Xi = min(Ti , T3 ), i = 1, 2. Then one can say that the vector
87 (X1 , X2 ) follows the bivariate generalized Chen distribution with scale parameters α,and
88 λj , j = 1, 2, 3 and shape parameter β. We will denote it by BGCh(α, β, λ1 , λ2 , λ3 ) and to
89 simplify we write λ123 = λ1 + λ2 + λ3 and λi3 = λi + λ3 , i = 1, 2.
90 Theorem 1: Let (X1 , X2 ) follows BGCh(α, β, λ1 , λ2 , λ3 ), then the joint survival function of
91 (X1 , X2 ) for x1 > 0, x2 > 0, is

SX1 ,X2 (x1 , x2 ) = P (X1 > x1 , X2 > X2 )

= P (T1 > x1 , T2 > x2 , T3 > x3 )
3
(3)
(xi /α)β
Y
= exp(λi α(1 − e )),
i=1

92 where x3 = max{x1 , x2 }.
93 Also, the joint survival function of (X1 , X2 ) can be written as

3
Y
SX1 ,X2 (x1 , x2 ) = SGCh (xi ; α, β, λi )
i=1

(4)
SGCh (x1 ; α, β, λ1 )SGCh (x2 ; α, β, λ23 ) if x1 < x2


= SGCh (x2 ; α, β, λ2 )SGCh (x1 ; α, β, λ13 ) if x2 < x1

S (x; α, β, λ ) if x1 = x2 = x.

GCh 123
 4 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

94 2.2. The joint probability density function

95 The following theorem gives the joint probability density function of the BGCh dis-
96 tribution.
97 Theorem 2: Let (X1 , X2 ) follows BGCh(α, β, λ1 , λ2 , λ3 ), then the joint pdf of (X1 , X2 )
98 takes the form

f1 (x1 , x2 )

 if 0 < x1 < x2 < ∞
fX1 ,X2 (x1 , x2 ) =

f (x , x2 )
2 1 if 0 < x2 < x1 < ∞ (5)
f (x) if 0 < x1 = x2 = x < ∞.

3

where
β +(x β (x1 /α)β ) +λ (x /α)β ) )
f1 (x1 , x2 ) = λ1 λ23 β 2 (x1 /α)(β−1) (x2 /α)(β−1) e(x1 /α) 2 /α)
eλ1 α(1−e 23 α(1−e 2

= fGCh (x1 ; α, β, λ1 )fGCh (x2 ; α, β, λ23 ),

β β (x1 /α)β )+λ α(1−e(x2 /α)β )
f2 (x1 , x2 ) = λ13 λ2 β 2 (x1 /α)(β−1) (x2 /α)(β−1) e(x1 /α) +(x2 /α) eλ13 α(1−e 2

= fGCh (x1 ; α, β, λ13 )fGCh (x2 ; α, β, λ2 ),

and
β (x/α)β )
f3 (x) = λ3 β(x/α)(β−1) e(x/α) eλ123 α(1−e
λ3
= fGCh (x; α, β, λ123 ).
λ123

Proof: The forms of f1 (., .) and f2 (., .) can be obtained simply by differentiating SX1 ,X2 (x1 , x2 )
in (4) with respect to x1 and x2 for x1 < x2 and x2 < x1 , respectively. The form of f3 (x)
can not obtained in the same way but it can be derived by using the following identity:
 ∞  x2  ∞  x1  ∞
f1 (x1 , x2 )dx1 dx2 + f2 (x1 , x2 )dx2 dx1 + f3 (x)dx = 1
0 0 0 0 0

99 which completes the proof of the theorem.

100 Proposition 1: Let (X1 , X2 ) follows BCh(β, λ1 , λ2 , λ3 ), then the joint pdf of (X1 , X2 ) takes
101 the form 
g1 (x1 , x2 )

 if 0 < x1 < x2 < ∞
gX1 ,X2 (x1 , x2 ) = g2 (x1 , x2 ) if 0 < x2 < x1 < ∞ (6)

g (x) if 0 < x1 = x2 = x < ∞.

3

where
β +(x )β (x1 )β )+λ (1−e(x2 )β )
g1 (x1 , x2 ) = λ1 λ23 β 2 (x1 )(β−1) (x2 )(β−1) e(x1 ) 2
eλ1 (1−e 23

= gCh (x1 ; β, λ1 )gCh (x2 ; β, λ23 ),

β +(x )β (x1 )β )+λ (1−e(x2 )β )
g2 (x1 , x2 ) = λ13 λ2 β 2 (x1 )(β−1) (x2 )(β−1) e(x1 ) 2
eλ13 (1−e 2

= gCh (x1 ; β, λ13 )gCh (x2 ; β, λ2 ),

and
β (x)β )
g3 (x) = λ3 β(x)(β−1) e(x) eλ123 (1−e
λ3
= gCh (x; β, λ123 ).
λ123
 yyyy] SHORT RUNNING TITLE IN CAPITALS 5

102 Proof: The result is obtained immediately from Theorem 2 upon setting α = 1.
103 The BGCh distribution has both a singular part and an absolutely continuous part
104 similar to Marshal-Olkin’s bivariate exponential distribution, Sarhan and Balakrishnan bi-
105 variate distribution, the bivariate generalized exponential introduced by Kundu and Gupta
106 (2009) and the bivariate generalized Rayleigh distribution provided by Sarhan (2019). The
107 function fX1 ,X2 (., .) may be considered to be a density function for the BGCh distribution
108 if it is understood that the first two terms are densities with respect to two-dimensional
109 Lebesgue measure and the third term is a density function with respect to one dimensional
110 Lebesgue measure, see Bemis et al. (1972). It is well known that although in one dimension
111 the practical use of a distribution with this property is unusual, but they do arise quite
112 naturally in higher dimensions, see Marshall and Olkin (1967).
113 In many practical situations it may happen that X1 and X2 both are continuous
114 random variables, but X1 = X2 has a positive probability. The BGCh distribution may be
115 used as a competing risk model or a shock model similar to the bivariate Marshall-Olkin
116 model. Marshall and Olkin (1967) has examples in this connection.. The following theorem
117 provides the explicit forms of the absolute continuous and the singular parts of the BGCh
118 distribution.
Theorem 3: If (X1 , X2 ) follows BGCh(α, β, λ1 , λ2 , λ3 ), then

λ3 λ12
SX1 ,X2 (x1 , x2 ) = Ss (x1 , x2 ) + Sa (x1 , x2 ).
λ123 λ123

For x = max(x1 , x2 ) we get,

(x)β )
Ss (x1 , x2 ) = eλ123 α(1−e ,
and
3
λ123 Y (x /α)β λ3 λ123 α(1−e(x)β )
Sa (x1 , x2 ) = eλi α(1−e i ) − e ,
λ12 i=1 λ1 2
119 here Ss (., .) and Sa (., .) are the singular and the absolutely continuous parts, respectively.

Proof: The joint survival function SX1 ,X2 (x1 , x2 ) can be written as

SX1 ,X2 (x1 , x2 ) = P (X1 > x1 , X2 > x2 |A)P (A) + P (X1 > x1 , X2 > x2 |Á)P (Á)

Let A = {T3 < T1 } ∩ {T3 < T2 } ≡ {X1 = X2 }, therefore

 ∞
β (x/α)β ) λ3
P (A) = λ3 β(x/α)(β−1) e(x/α) eλ123 α(1−e dx =
0 λ123

and
Ss (x1 , x2 ) = P (X1 > x1 , X2 > x2 |A)

λ123 ∞ β (x/α)β )
= λ3 β(x/α)(β−1) e(x/α) eλ123 α(1−e dx
λ3 0
(x/α)β )
= eλ123 α(1−e .
 6 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

120 Once P (A) and Ss (x1 , x2 ) are obtained, the function Sa (x1 , x2 ) can be obtained by
121 subtraction.
122 Different shapes of the joint pdf and corresponding contours for different sets of
parameters values are provided in Figure 1.

Set (1): (λ1 , λ2 , λ3 , β, α)=(2,2,2,2,1)

123

Set (2): (λ1 , λ2 , λ3 , β, α)=(0.5,0.5,1,1.5,1)

Figure 1: The joint probability density function of the BGCh distribution and
corresponding contour

124 2.3. The joint hazard rate function

125 Using the relation between the joint pdf of (X1 , X2 ) and the joint survival function of
126 (X1 , X2 ), one can obtain the joint hazard rate function of (X1 , X2 ) according to the relation

fX1 ,X2 (x1 , x2 )

hX1 ,X2 (x1 , x2 ) = .
SX1 ,X2 (x1 , x2 )

127 Here we use the forms (4) and (6) to obtain the joint hazard rate function. In Figure 2 we
128 provide the surface plots of the joint hazard rate function and corresponding contours for
129 different values of the parameters.
 yyyy] SHORT RUNNING TITLE IN CAPITALS 7

Set (1): (λ1 , λ2 , λ3 , β, α)=(2,2,2,2,1)

Set (2): (λ1 , λ2 , λ3 , β, α)=(0.2,0.2,0.8,0.8,1)

Figure 2: The joint hazard rate function of the BGCh distribution and corre-
sponding contour
130 3. Statistical properties
131 3.1. Marginal distributions
One can easily verify that the marginal distribution of Xi , i = 1, 2, follows GCh(β, α, λi ).
For this, we first derive the marginal survival function of Xi , say SXi (x), as follows
SXi (x) = P (Xi > x) = P (min(Ti , T3 ) > x) = P (Ti > x, T3 > x)
132 and since Ti , i = 1, 2 and T3 are independent random variables, then
(x/α)β )
SXi (x) = P (Ti > x)P (T3 > x) = SXi (x; β, α, λi3 ) = eλi3 α(1−e (7)
133 Using (7), the marginal pdf of Xi is
β (x/α)β ) )
fXi (x) = λi3 β(x/α)(β−1) e(x/α) eλi3 α(1−e , (8)
134 and the marginal hazard rate function (hrf) of Xi is
β
hXi (x) = λi3 β(x/α)(β−1) e(x/α) . (9)
135 Xie et al. (2002) noted that the hrf depends only on the shape parameter β and they observed
136 that: when β > 1, the hrf has an increasing shape and when β < 1, the hrf has a bathtub
137 shape. Shapes of the pdf and hrf of Xi for different values of β, α and λi3 are provided in
138 Figure 3. Also, Xie et al. (2002) showed that the GCh distribution can be used in modeling
139 bathtub-shaped failure rate univariate lifetime data. Hence, we expect the BGCh distribution
140 can be used in modeling bathtub-shaped failure rate bivariate lifetime data.
 8 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

Figure 3: The probability density and hazard rate functions of the marginal dis-
tribution of X1

141 3.2. Conditional distributions

142 Having obtained the marginal pdf of X1 and X2 , one can derive the conditional proba-
143 bility density function. The following theorem provides the conditional pdf of X1 given X2 =
144 x2 , say fX1 |X2 (x1 |x2 ).
145 Theorem 4: If (X1 , X2 ) follows BGCh(β, α, λ1 , λ2 , λ3 ), then the conditional pdf of X1 given
146 X2 = x2 is 
f1 (x1 |x2 ) if x1 < x2


fX1 |X2 (x1 , x2 ) = f2 (x1 |x2 ) if x2 < x1 (10)

f (x |x )

if x1 = x2 = x,
3 1 2

where
β ( β
f1 (x1 |x2 ) = λ1 β(x1 /α)(β−1) e(x1 /α) eλ1 α(1−e x1 /α) ) ,
β (x1 /α)β )−λ (x2 /α)β ))
f2 (x1 |x2 ) = (λ12 λ2 )/λ23 β(x1 /α)(β−1) e(x1 /α) eα(λ13 (1−e 3 (1−e
, and
β
α(1−e(x/α) )
f3 (x1 |x2 ) = λ3 /λ23 eλ1 .

147 Proof: The results of this theorem are easily derived using the definition of conditional
148 probability and the results of Theorem 2 and the form (8). Figure 4 shows some plots of the
149 conditional pdf’s of X1 given X2 = x2 for different values of x2 (x2 = 0.5, 1, 2) and different
150 values of parameters.
151 Similarly, the conditional pdf of X2 given X1 = x1 can be obtained in a similar
152 manner as above. Also, one can note that if α = 1, the conditional pdf in the case of BCh
153 distribution can be obtained.

154 4. Parameters’ estimation

155 Suppose that {(X11 , X21 ), (X12 , X22 ), . . . , (X1n , X2n )} is a random sample from BGCh
156 (λ1 , λ2 , λ3 , β, α). The likelihood function for this sample is
n
f1 (x1i , x2i )I(x1i <x2i ) f2 (x1i , x2i )I(x1i >x2i ) f3 (x1i , x2i )I(x1 i=x2i ) ,
Y
L(data; θ) = (11)
i=1
 yyyy] SHORT RUNNING TITLE IN CAPITALS 9

Figure 4: The conditional probability density function of X1 given X2 = x2 at

different sets of the parameters

157 where I(A) is an indicator function that is equal to 1 if A is true and 0 otherwise and
158 θ = (λ1 , λ2 , λ3 , β, α). Substituting (5) in (11) and taking the natural logarithm, we obtain
159 the log-likelihood function as

n
I(x1i < x2i ){ln(λ1 ) + ln(λ23 ) + 2ln(β) + (β − 1)ln(x1i /α) + ln(x2i /α) + (x1i /α)β
X
LL =
i=1
β β
+ (x2i /α)β + λ1 (1 − e(x1i /α) ) + (λ23 )(1 − e(x2i /α) ))}
+ I(x1i > x2i ){ln(λ2 ) + ln(λ13 ) + 2ln(β) + (β − 1)(ln(x1i /α) + ln(x2i /α) + (x1i /α)β
β β
+ (x2i /α)β + λ2 (1 − e(x2i /α) ) + (λ13 )(1 − e(x1i /α) )}
β
+ I(x2i = x1i ){ln(λ3 ) + ln(β) + (β − 1)ln(x1i /α) + (x1i /α)β + (λ123 )(1 − e(x1i /α) )}.
(12)
 10 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

160 4.1. Maximum likelihood estimation

161 Here we use maximum likelihood method to estimate the unknown parameters of the
162 BGCh distribution. For fixed α, the likelihood equations are
n
∂LL n1 n2 β
I(x1i < x2i )(1 − e(x1i /α) ) = 0,
X
= + +
∂λ1 λ1 λ13 i=1
n
∂LL n1 n2 X β
= + + I(x1i > x2i )(1 − e(x2i /α) ) = 0,
∂λ2 λ23 λ2 i=1
n
∂LL n1 n2 n3 X β
= + + + {I(x1i < x2i )(1 − e(x2i /α) )
∂λ3 λ23 λ13 λ3 i=1
β
+ {I(x1i > x2i ) + I(x1i = x2i )}(1 − e(x1i /α) )} = 0,
and
n
∂LL X β
= I(x1i < x2i ){2/β + ln(x1i /α) + ln(x2i /α) + (x1i /α)β ln(x1i /α)(1 − λ1 e(x1i /α) )
∂β i=1
β
+ (x2i /α)β ln(x2i /α)(1 − λ23 )e(x2i /α) }
β
+ I(x1i > x2i ){2/β + ln(x1i /α) + ln(x2i /α) + (x1i /α)β ln(x1i /α)(1 − λ13 e(x1i /α) )
β
+ (x2i /α)β ln(x2i /α)(1 − λ2 )e(x2i /α) )}
β
+ I(x2i = x1i ){1/β + ln(x1i /α) + (x1i /α)β ln(x1i /α)(1 − λ123 e(x1i /α) )} = 0,
Pn Pn Pn
(13)
163 where n1 = i=1 I(x1i < x2i ) , n2 = i=1 I(x1i > x2i ), and n3 = i=1 I(x1i = x2i ).
164 The likelihood equations (13) do not have a closed-form solution, so a numerical technique
165 must be used to find the maximum likelihood estimates (mles) of λ1 , λ2 , λ3 , and β. The
166 likelihood equations may have multiple roots, Small et al. (2000) discussed this problem
167 using the Hessian matrix. They showed that the likelihood equations have a unique root
168 when the Hessian matrix of the log-likelihood is negative definite for all value of θ. This
169 relies on maximizing the log-likelihood function. The Hessian matrix is written as
 
LLλ1 λ1 LLλ1 λ2 LLλ1 λ3 LLλ1 β
LL LLλ2 λ2 LLλ2 λ3 LLλ2 β 
T (θ) =  λ2 λ1
 
LLλ3 λ1 LLλ3 λ2 LLλ3 λ3 LLλ3 β 


LLβλ1 LLβλ2 LLβλ3 LLββ

∂ 2 LL
170 where LLθi θj = ∂θi ∂θj
is the second partial derivative of the log-likelihood function with re-
171 spect to the components θi and θj of θ and T (θ̂) is the Hessian matrix computed at θ = θ̂.
172

173 Large-sample confidence intervals: Under regularity conditions, the mles of the pa-
174 rameters λ1 , λ2 , λ3 , and β are asymptotically normally distributed with means equal to the
175 true values of these parameters and variances given by the inverse of the information matrix.
176 One can approximate the expected values of the second-order derivatives of logarithms of
177 likelihood function with the maximum likelihood estimates of the parameters as given in
178 Cohen (1965). That is, using normality property of mles, one can construct the asymptotic
179 confidence interval for each parameter.
 yyyy] SHORT RUNNING TITLE IN CAPITALS 11

180 4.2. Bayes estimation

181 Now, we discuss the Bayesian estimation of the unknown parameters of the BGCh
182 distribution. For fixed α, let the four parameters θ = (λ1 , λ2 , λ3 , β) are independent random
183 variables and follow the gamma prior distribution. That is, the joint prior pdf of θ is
(a −1) (a2 −1) (a3 −1) (a4 −1) (−b1 λ1 −b2 λ2 −b3 λ3 −b4 β)
g0 (θ) ∝ λ1 1 λ2 λ3 β e , λ1 , λ2 , λ3 , β > 0, (14)

184 where all the hyperparameters ai and bi , i = 1, 2, 3, 4 are assumed to be positive and known.
185 The log-prior density function is
3
X 3
X
g0 (θ) ∝ (ai − 1)ln(λi ) + (a4 − 1)ln(β) − bi λi − b4 β. (15)
i=1 i=1

186 Using (12) and (15) and applying Bayes theorem, the joint posterior probability density
187 function of θ, given data, is
1
g(θ|data) = exp(LL + g0 (θ)), (16)
K
188 where K is the normalizing constant. Bayes estimators of the unknown parameters and/or
189 of any function of the unknown parameters, say w(θ), can be obtained as follows

∞
∞
∞
∞
w(θ)exp(LL + g0 (θ))dλ1 dλ2 dλ3 dβ
ŵ(data) = 0
∞0
∞0
∞0
∞ . (17)
0 0 0 0
exp(LL + g0 (θ))dλ1 dλ2 dλ3 dβ

190 Formula (17) involves a ratio of two multidimentional integrals and does not have analytical
191 solution. Thus, some approximation methods were suggested to approximate these integrals
192 and calculate the ratio of the integrals such as the methods discussed by Lindley (1980)
193 and Tierney and Kadane (1986). These methods work well for low dimensions. In this
194 paper we will use Markov Chain Monte Carlo (MCMC) method that work well in the case
195 of high dimensions, see Gelman et al. (2013). MCMC method generates random draws from
196 the joint posterior distribution by generating draws from an arbitrary distribution (proposal
197 distribution) that easy to simulate from then apply an accept-reject method. Here, we
198 use multivariate normal as a proposal distribution. The following steps can be followed to
199 generate random draws from the joint posterior distribution (16):

200 1. Specify the size of the random draws we wish to generate, say m.
201 2. Choose an initial value of θ, say θ (0) .
202 3. For i = 1, 2, . . . , m, repeated the following steps:
203 (a) Generate θ ∗ from the multivariate normal with mean θ (i−1) and variance-covariance
204 Σ.
∗ |data)
205 (b) Compute the ratio κ = min{1, g(θg(θ
(i−1) |data)
}.
206 (c) Generate a random value from uniform distribution on (0, 1).
207 (d) If κ ≥ put θ (i) = θ ∗ , otherwise put θ (i) = θ (i−1) .
 12 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

Discarding the early m0 number of burn-in draws and using the remaining m − m0 , θ (m0 +1) ,
θ (m0 +2) , . . . , θ (m) , as the chosen draws from the joint posterior distribution, the Bayes esti-
mate of θj is
m−m (i)
X0 θj
θ̂j = , j = 1, 2, 3, 4.
i=m0 +1 m − m0

208 Furthermore, for 0 < ν < 1, one can obtain the lower and upper bounds of the 100(1 − ν)%
209 Bayesian probability interval of θj via (ν/2)100th and (1 − ν/2)100th percentiles of the
210 sequence of the m − m0 draws; θ (m0 +1) , θ (m0 +2) , . . . , θ (m) .

211 5. Simulation results and applications

212 In this section, some simulation results and the analysis of a data set are presented.

213 5.1. Simulation results

214 In this section, we provide the following steps to generate a random sample of the
215 BGCh distribution:

216 1. Generate u1 , u2 and u3 from uniform (0, 1).

217 2. Compute t1 = α(ln(1 − ln(1−u

λ1 α
1 ) 1/β
) , t1 = α(ln(1 − ln(1−u2 ) 1/β
λ2 α
) and
ln(1−u3 ) 1/β
218 t3 = α(ln(1 − λ3 α ) .

219 3. Obtain x1 = min(t1 , t3 ) and x2 = min(t2 , t3 ).

220 To obtain some simulation results for samples size (n=100) and for different parameter
221 values, we consider three different sets of parameter values namely: (i) λ1 = λ2 = λ3 = β =
222 1, (ii) λ1 = λ2 = λ3 = 2, β = 1, and (iii) λ1 = 0.5, λ2 = 0.5, λ3 = 1, β = 1.5. We replicate
223 the process 1000 times and report the average estimates and the root mean square errors
224 (RMSEs) in Table 1. Also, we compute the Bayes estimates of the unknown parameters as
225 mentioned in the previous section with assuming uniform priors. We simulate 10000 runs
226 and replicate the process 1000 times. The average estimates and the RMSEs are also listed
227 in Table 1 and one can note that results of Bayes estimates are better than mles.

228 5.2. Applications

229 In this section we present the analysis of a data set to discuss how the proposed
230 distribution can be used in practice. This data represents the UEFA Champion’s League
231 Data and it was analyzed in Meintanis (2007) using the Marshall-Olkin exponential model
232 (MO) and by Kundu and Gupta (2009) using the bivariate generalized exponential (BVGE)
233 model, then by Sarhan (2019) using the bivariate generalized Rayleigh (BVGR) model.
234 Kundu and Gupta (2009) reported that the BVGE model fits the data better than MO
235 model and Sarhan (2019) reported that the BVGR model fits the data better than both the
236 MO and the BVGE models. Here, we use the BCh model to reanalyze the same data and
237 compare it with the three models; the MO, the BVGE and the BVGR but first we have
238 fitted Ch(β, λ) model to the marginal and the minimum of the two marginals. The mles of
 yyyy] SHORT RUNNING TITLE IN CAPITALS 13

Table 1: The mles and the Bayes estimates and their RMSEs (in paren-
theses) of the parameters

Parameter value Method β λ1 λ2 λ3

MLE 1.3834 2.6776 1.3080 1.5323
(0.4243) (1.6961) (0.3732) (0.5492)
(1.0, 1.0, 1.0, 1.0) Bayes 1.3429 1.4783 1.2716 1.4297
(0.3609) (0.4788) (0.2993) (0.4316)
MLE 2.9372 5.0508 2.8083 1.5604
(1.9725) (3.0754) (0.8590) (0.4518)
(1.0, 2.0, 2.0, 2.0) Bayes 2.3606 2.4705 2.3693 1.3772
(1.3661) (0.4714) (0.3867) (0.6247)
MLE 1.7625 4.3562 1.5935 1.4858
(0.4276) (3.8752) (1.1251) (0.4916)
(1.5, 0.5, 0.5, 1.0) Bayes 0.9269 0.9891 1.3696 1.3696
(0.5764) (0.4892) (0.8753) (0.3736)

239 the unknown parameters, the Kolmogorov-Smirnov (K-S) distances between the empirical
240 distribution function (EDF) and the fitted distribution function and the associated p values
241 are reported in Table 2. Based on the p values, one can observe that Chen distribution may
242 be used to fit X1 , X2 and min(X1 , X2 ).
Table 2: The mles of the parameters, the K-S test statis-
tics and associated p-values

Variable mle K-S p-value

X1 β̂ = 0.403, λ̂ = 0.010 0.013 0.572
X2 β̂ = 0.379, λ̂ = 0.184 0.106 0.804
min(X1 , X2 ) β̂ = 0.389, λ̂ = 0.019 0.094 0.899

243 Now, to test whether BCh distribution fits the data or not, we use the two-dimensional
244 Kolomogorov-Sminrov test of goodness of fit as proposed by Peacock (1983). Using the
245 computational environmental R peacock package, we obtain the value of test statistic as
246 0.2712 with p value 0.6482. Based on the p value, we cannot reject the null hypothesiss that
247 the data came from the BCh distribution at 0.05 level of significance. For more details about
248 multivariate Kolomogorov-Sminrov test of goodness of fit see Justel et al. (1997).
249 Hence, we have used the BCh model to analyze the bivariate data set. We use R to
250 get mles of the unknown parameters. Table 3 shows the mles of the unknown parameters
251 of the proposed distribution together with the values of the log-likelihood values and the
252 Akaike information criterion (AIC=-2 LL+2k,k is the number of estimated parameters; see
253 Akaike (1974). The AIC suggests that the BCh distribution provides a better fit than the
254 three models; the MO, the BVGE and the BVGR.
255 To indicate that a unique root for the likelihood equations exist. We use the estimates
256 λ̂1 , λ̂2 , λ̂3 and β̂ obtained with respect to the given bivariate data set. These estimates
 14 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

Table 3: The mles of the parameters, the log-likelihood values and AIC values

Model mle L AIC

MO λ̂1 = 0.012, λ̂2 = 0.014, λ̂3 = 0.022, -339.006 684.012
BVGE α̂1 = 1.351, α̂2 = 0.465, α̂3 = 1.153, β̂ = 0.039 -296.935 601.870
BVGR α̂1 = 0.492, α̂2 = 0.166, α̂3 = 0.410, λ̂ = 0.020 -293.357 594.714
BCh ˆ 2 = 0.055, λ̂3 = 0.048, β̂ = 1.020
λ̂1 = 0.026, lambda 0.094 0.899

257 are obtained using nlm R package which minimize the negative of the log-likelihood function.
258 We obtain T (θ̂) as follows:

0.0424 −1.63E − 0710−7 9.8070 6.0055

 
 − 33.114 2.0183 8.427 
 
 − − 46.709 11.059 
 

− − − 104.0074
259 The eigen values of this matrix are -103.6646, -42.8617, -31.9991 and -2.2724. This indicates
260 that T (θ̂) is negative definite. Then according to Small et al. (2000), the likelihood equations
261 has a unique root. For more details see Thomas and Jose (2021).
262 For Bayesian computations, we obtain the Bayes estimates of the unknown parame-
263 ters based on the uniform priors and the gamma priors. In the case of the gamma priors,
264 we assume that all hyperparameters equal and equal to 0.5. For the two cases, the pro-
265 posal distribution is multinormal with variance covariance matrix and the choice of its value
266 depends on the acceptance rate which is assumed such that the acceptance rate (number
267 of accepted runs out of total runs) increases. Here, we simulate 10000 runs from the joint
268 posterior distribution of the four parameters and the early 20% of the runs were discarded.
269 The trace plots of the draws are plotted in Figures 5 and 6 after discarding the early 2000
270 draws (burn-in period). Tables 4-5 list the posterior descriptive summaries of interest such
271 as the posterior mean, median, standard deviation and the 95% Bayesian credible intervals.

Table 4: Summary results for the posterior parameters in the case of gamma
priors (the acceptance rate is 38.18%)

Parameter Mean Median Standard deviation 95% credible intervals

λ1 0.0837 0.0754 0.0408 (0.0533, 0.1053)
λ2 0.5379 0.5278 0.1352 (0.4347, 0.6332)
λ3 0.1707 0.1604 0.0620 (0.1274, 0.2031)
β 0.2114 0.2114 0.0168 (0.1995, 0.2227)

272

273 6. Conclusion
274 In this paper, the bivariate generalized Chen distribution (BGCh) is proposed as a
275 new bivariate lifetime distribution. The BGCh distribution is of Marshal-Olkin type whose
276 marginal are generalized Chen distributions. One can observe that the BGCh distribution is
 yyyy] SHORT RUNNING TITLE IN CAPITALS 15

Table 5: Summary results for the posterior parameters in the case of

uniform priors (the acceptance rate is 54.83%)

Parameter Mean Median Standard deviation 95% credible intervals

λ1 0.1137 0.1042 0.0533 (0.0756, 0.1401)
λ2 0.6255 0.6134 0.1575 (0.5175, 0.7273)
λ3 0.2227 0.2167 0.0779 (0.1657, 0.2697)
β 0.2019 0.2015 0.0171 (0.1900, 0.2127)

Figure 5: The trace plot of the random draws from the joint posterior distribu-
tion in the case of gamma priors

277 a singular distribution and has an absolute continuous and a singular part. Some statistical
278 properties are investigated. The estimation of the parameters has been approached by max-
279 imum likelihood and Bayesian methods. For Bayesian method, we used the MCMC method.
280 Numerical methods are required to calculate the desired estimates. One real data set is
281 analyzed using the BCh distribution which showed a better fit than the MO, the BVGE and
282 the BVGR distributions.
283
 16 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

Figure 6: The trace plot of the random draws from the joint posterior distribu-
tion in the case of uniform priors

284 Acknowledgements
285 I am indeed grateful to the Editors for their guidance and counsel. I am very grateful
286 to the reviewer for valuable comments and suggestions of generously listing many useful
287 references.

288 References
289 Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on
290 Automatic Control, 19, 716–723.
291 Azizi, A., Sayyareh, A., and Panahi, H. (2020). Inference about the bivariate new ex-
292 tended weibull distribution based on complete and censored data. Communications
293 in Statistics-Simulation and Computation, 51, 738–756.
294 Bemis, B. M., Bain, L. J., and Higgins, J. J. (1972). Estimation and hypothesis testing
295 for the parameters of a bivariate exponential distribution. Journal of the American
296 Statistical Association, 67, 927–929.
297 Chen, Z. (2000). A new two-parameter lifetime distribution with bathtub shape or increasing
298 failure rate function. Statistics & Probability Letters, 49, 155–161.
 yyyy] SHORT RUNNING TITLE IN CAPITALS 17

299 Cohen, A. C. (1965). Maximum likelihood estimation in the weibull distribution based on
300 complete and on censored samples. Technometrics, 7, 579–588.
301 El-Gohary, A., H El-Bassiouny, A., and El-Morshedy, M. (2016). Bivariate exponentiated
302 modified weibull extension distribution. Journal of Statistics Applications & Proba-
303 bility, 5, 67–78.
304 Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. (2013).
305 Bayesian Data Analysis. CRC press.
306 Justel, A., Peña, D., and Zamar, R. (1997). A multivariate kolmogorov-smirnov test of
307 goodness of fit. Statistics & Probability Letters, 35, 251–259.
308 Kundu, D. (2012). On sarhan-balakrishnan bivariate distribution. Journal of Statistics
309 Applications & Probability, 1, 163.
310 Kundu, D. and Gupta, A. K. (2017). On bivariate inverse weibull distribution. Brazilian
311 Journal of Probability and Statistics, 31, 275–302.
312 Kundu, D. and Gupta, R. D. (2009). Bivariate generalized exponential distribution. Journal
313 of Multivariate Analysis, 100, 581–593.
314 Lindley, D. V. (1980). Approximate bayesian methods. Trabajos de Estadı́stica Y de Inves-
315 tigación Operativa, 31, 223–245.
316 Marshall, A. W. and Olkin, I. (1967). A multivariate exponential distribution. Journal of
317 the American Statistical Association, 62, 30–44.
318 Meintanis, S. G. (2007). Test of fit for marshall–olkin distributions with applications. Journal
319 of Statistical Planning and inference, 137, 3954–3963.
320 Mudholkar, G. S. and Srivastava, D. K. (1993). Exponentiated weibull family for analyzing
321 bathtub failure-rate data. IEEE Transactions on Reliability, 42, 299–302.
322 Muhammed, H. Z. (2019). Bivariate generalized burr and related distributions: properties
323 and estimation. Journal of Data Science, 17, 535–550.
324 Murthy, D. P., Xie, M., and Jiang, R. (2004). Weibull Models. John Wiley & Sons.
325 Pham, H. and Lai, C.-D. (2007). On recent generalizations of the weibull distribution. IEEE
326 transactions on reliability, 56, 454–458.
327 Sarhan, A. (2019). The bivariate generalized rayleigh distribution. Journal of Mathematical
328 Sciences and Modelling, 2, 99–111.
329 Sarhan, A. M. and Balakrishnan, N. (2007). A new class of bivariate distributions and its
330 mixture. Journal of Multivariate Analysis, 98, 1508–1527.
331 Small, C. G., Wang, J., and Yang, Z. (2000). Eliminating multiple root problems in estima-
332 tion. Statistical Science, 15, 313–341.
333 Smith, R. M. and Bain, L. J. (1975). An exponential power life-testing distribution. Com-
334 munications in Statistics-Theory and Methods, 4, 469–481.
335 Thomas, P. Y. and Jose, J. (2021). On weibull–burr impounded bivariate distribution.
336 Japanese Journal of Statistics and Data Science, 4, 73–105.
337 Tierney, L. and Kadane, J. B. (1986). Accurate approximations for posterior moments and
338 marginal densities. Journal of the American Statistical Association, 81, 82–86.
339 Xie, M., Tang, Y., and Goh, T. N. (2002). A modified weibull extension with bathtub-shaped
340 failure rate function. Reliability Engineering & System Safety, 76, 279–285.
 18 NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL. [Vol. xx, No. x

341 ANNEXURE
Table 6: The AIL and CP for bootstrap CIs and HPD intervals

α n θ1 θ2 R Bootstrap HPD
AIL CP AIL CP
-0.75 6 5 1 0.67100 0.22830 85 0.11755 94
3 2 0.57589 0.21336 85 0.13603 92
2 4 0.34583 0.18995 87 0.12971 93
0.5 5 0.09334 0.18229 86 0.12178 92
8 5 1 0.67100 0.17457 87 0.14590 95
3 2 0.57589 0.16680 87 0.11195 94
2 4 0.34583 0.15912 88 0.12494 95
0.5 5 0.09334 0.15215 88 0.11086 94
10 5 1 0.67100 0.14993 88 0.11234 96
3 2 0.57589 0.13375 87 0.08454 96
2 4 0.34583 0.16985 89 0.12985 95
0.5 5 0.09334 0.21065 85 0.14707 95
-0.5 6 5 1 0.72511 0.23566 85 0.13176 94
3 2 0.58393 0.17582 84 0.11582 93
2 4 0.34167 0.17737 84 0.13245 93
0.5 5 0.09253 0.16639 86 0.11629 94
8 5 1 0.72511 0.18512 86 0.12621 95
3 2 0.58393 0.15397 86 0.13670 95
2 4 0.34167 0.14288 88 0.12190 94
0.5 5 0.09253 0.15278 87 0.13116 95
10 5 1 0.72511 0.14377 88 0.11098 94
3 2 0.58393 0.16415 87 0.18154 95
2 4 0.34167 0.16366 88 0.08478 96
0.5 5 0.09253 0.23831 85 0.12197 95
-0.25 6 5 1 0.77922 0.22563 85 0.13355 93
3 2 0.59196 0.14293 84 0.11844 94
2 4 0.33750 0.14189 85 0.12333 93
8 5 1 0.77922 0.15815 86 0.12882 95
3 2 0.59196 0.14140 87 0.11425 95
2 4 0.33750 0.13703 85 0.12929 95
10 5 1 0.77922 0.12808 87 0.12552 95
3 2 0.59196 0.13565 89 0.12982 96
2 4 0.33750 0.12893 88 0.11873 95
yyyy] SHORT RUNNING TITLE IN CAPITALS 19

Table 6: Continued

α n θ1 θ2 R Bootstrap HPD
AIL CP AIL CP
0.25 6 5 1 0.88745 0.15200 84 0.09624 93
3 2 0.60804 0.17320 84 0.11650 93
2 4 0.32917 0.18756 85 0.08888 92
0.5 5 0.09010 0.16752 85 0.09039 92
8 5 1 0.88745 0.13493 85 0.09131 95
3 2 0.60804 0.16607 85 0.11750 94
2 4 0.32917 0.16677 86 0.08159 95
0.5 5 0.09010 0.15563 86 0.08744 95
10 5 1 0.88745 0.12890 86 0.08205 93
3 2 0.60804 0.13831 87 0.10826 95
2 4 0.32917 0.12724 88 0.07175 95
0.5 5 0.09010 0.13398 88 0.07624 96
0.5 6 5 1 0.94156 0.16974 85 0.11738 94
3 2 0.61607 0.15253 86 0.12961 93
2 4 0.32500 0.14018 83 0.11854 94
0.5 5 0.08929 0.14557 84 0.12974 95
8 5 1 0.94156 0.13592 85 0.11546 94
3 2 0.61607 0.13528 85 0.10860 95
2 4 0.32500 0.12432 86 0.10255 96
0.5 5 0.08929 0.12819 88 0.11897 95
10 5 1 0.94156 0.12177 88 0.12423 94
3 2 0.61607 0.12387 86 0.12557 96
2 4 0.32500 0.11362 88 0.12771 96
0.5 5 0.08929 0.11695 89 0.08317 95
0.75 6 5 1 0.99567 0.18762 84 0.11731 93
3 2 0.62411 0.17517 85 0.12341 92
2 4 0.32083 0.16081 83 0.09304 94
0.5 5 0.08847 0.16947 84 0.09253 94
8 5 1 0.99567 0.14872 85 0.12235 93
3 2 0.62411 0.13736 87 0.13994 95
2 4 0.32083 0.12521 87 0.09638 95
0.5 5 0.08847 0.13757 88 0.09070 96
10 5 1 0.99567 0.12422 86 0.10111 95
3 2 0.62411 0.12523 88 0.11063 96
2 4 0.32083 0.12777 87 0.08842 96
0.5 5 0.08847 0.12388 89 0.09357 96
 20

Table 7: The bias and MSE for MLE and Bayes estimators for R

α n θ1 θ2 MLE SE LL EL
Bias MSE Bias MSE Bias MSE Bias MSE
-0.75 6 5 1 0.15841 0.12508 0.12893 0.02014 0.13763 0.02264 0.14868 0.02623
3 2 -0.12726 0.12173 0.02384 0.00752 0.00917 0.00907 0.01595 0.00935
2 4 0.12681 0.15718 -0.01435 0.00854 0.00669 0.00825 0.01677 0.00864
0.5 5 0.19078 0.16283 0.00522 0.00162 0.02135 0.00240 0.02290 0.00256
8 5 1 -0.14307 0.11006 0.12063 0.01724 0.12942 0.01956 0.14035 0.02282
3 2 0.11088 0.11727 -0.00596 0.00775 0.00451 0.00815 0.01270 0.00837
2 4 0.11998 0.09947 -0.01343 0.00837 0.00377 0.00653 0.01565 0.00675
0.5 5 0.16437 0.14880 0.00417 0.00158 0.01192 0.00195 0.01295 0.00216
10 5 1 0.12832 0.10176 0.10952 0.01594 0.11912 0.01818 0.13063 0.02139
3 2 -0.10925 0.10386 -0.00592 0.00612 0.00388 0.00769 0.00961 0.00783
2 4 0.11148 0.09322 0.01237 0.00576 0.00108 0.00549 0.00530 0.00564
0.5 5 0.13270 0.08616 0.00172 0.00149 0.00901 0.00178 0.00985 0.00185
-0.5 6 5 1 -0.14634 0.13198 0.07667 0.00944 0.08515 0.00995 0.09591 0.01322
3 2 0.17554 0.18434 0.01319 0.00853 0.01647 0.00841 0.02230 0.00877
2 4 0.19828 0.15297 -0.00146 0.00621 0.00271 0.00981 0.00674 0.00915
0.5 5 -0.12645 0.16493 0.00700 0.00122 0.01692 0.00178 0.01831 0.00190
8 5 1 0.13805 0.12617 0.06490 0.00858 0.07459 0.00898 0.08654 0.01228
3 2 0.15785 0.13086 -0.00762 0.00762 0.00343 0.00719 0.01002 0.00732
2 4 0.11615 0.14922 -0.00123 0.00580 -0.00863 0.00826 0.00412 0.00845
NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL.

0.5 5 0.10410 0.14556 0.00433 0.00095 0.01231 0.00163 0.01315 0.00128

10 5 1 -0.12854 0.11433 0.06238 0.00673 0.07095 0.00798 0.08148 0.00989
3 2 0.12208 0.12936 -0.00613 0.00641 0.00849 0.00665 0.00884 0.00678
2 4 0.12199 0.13811 0.00108 0.00476 -0.00417 0.00712 0.00353 0.00730
0.5 5 0.08247 0.11382 -0.00365 0.00079 0.00789 0.00109 0.00884 0.00104
[Vol. xx, No. x
 yyyy]

Table 7: Continued

α n θ1 θ2 MLE SE LL EL
Bias MSE Bias MSE Bias MSE Bias MSE
-0.25 6 5 1 0.12725 0.15281 0.01976 0.00428 0.02879 0.00486 0.04018 0.00604
3 2 -0.14357 0.12200 0.02778 0.00977 -0.01681 0.00880 -0.01049 0.00871
2 4 0.12894 0.14564 -0.00505 0.00807 0.01759 0.00801 0.02039 0.00841
0.5 5 0.12438 0.15906 0.00797 0.00146 0.01257 0.00193 0.01381 0.00204
8 5 1 0.11938 0.12006 0.01673 0.00374 0.02615 0.00417 0.03892 0.00518
3 2 -0.14834 0.10475 -0.01844 0.00855 0.01445 0.00796 0.01226 0.00819
2 4 -0.11049 0.13814 -0.00708 0.00782 0.01541 0.00785 0.02008 0.00818
0.5 5 0.11796 0.13364 0.00640 0.00118 0.01118 0.00149 0.01213 0.00155
10 5 1 0.10124 0.10853 0.01262 0.00255 0.02425 0.00304 0.03439 0.00396
3 2 0.11841 0.11926 -0.01207 0.00800 -0.01166 0.00735 -0.00628 0.00730
2 4 0.09921 0.12727 0.00337 0.00563 0.01496 0.00534 0.00435 0.00546
0.5 5 -0.09222 0.11548 -0.00393 0.00071 0.00190 0.00077 0.00565 0.00079
0.25 6 5 1 0.16839 0.12261 -0.07787 0.00996 -0.08808 0.01129 -0.09631 0.01265
3 2 -0.13160 0.11563 -0.02081 0.01138 -0.02764 0.01163 -0.03948 0.01314
2 4 0.12543 0.10017 0.01213 0.00565 0.01863 0.00848 0.01393 0.00552
0.5 5 0.12936 0.10947 0.01771 0.00260 0.01652 0.00477 0.00358 0.00189
8 5 1 -0.15803 0.11962 -0.07764 0.00917 -0.08717 0.01043 -0.09485 0.01169
SHORT RUNNING TITLE IN CAPITALS

3 2 -0.12533 0.10829 -0.01730 0.00680 -0.02298 0.00697 -0.03243 0.00783

2 4 0.11079 0.09383 0.01208 0.00795 0.00802 0.00771 0.00669 0.00786
0.5 5 0.12187 0.09976 0.01282 0.00180 0.01196 0.00145 0.00282 0.00114
10 5 1 0.13589 0.11125 -0.06289 0.00709 -0.07466 0.00835 -0.08383 0.00965
3 2 0.11101 0.10076 -0.01277 0.00537 -0.01760 0.00547 0.02536 0.00601
2 4 -0.10887 0.08118 0.00638 0.00149 0.00141 0.00470 -0.01720 0.00509
0.5 5 -0.11165 0.08092 0.01127 0.00091 0.01059 0.00088 0.00209 0.00069
21
 22

Table 7: Continued

α n θ1 θ2 MLE SE LL EL
Bias MSE Bias MSE Bias MSE Bias MSE
0.5 6 5 1 0.15162 0.11095 0.11730 0.01587 -0.12781 0.01819 -0.13604 0.02023
3 2 -0.13110 0.09081 -0.04190 0.00861 -0.04643 0.00895 0.05423 0.00998
2 4 0.11021 0.09824 0.02081 0.00658 0.01599 0.00629 -0.08451 0.00617
0.5 5 0.15137 0.10037 0.01325 0.00631 0.02208 0.00219 0.01924 0.00158
8 5 1 -0.13262 0.09341 -0.11504 0.01456 0.12539 0.01776 -0.12349 0.01969
3 2 0.12683 0.08046 -0.02679 0.00650 -0.03326 0.00686 0.04428 0.00810
2 4 -0.10137 0.07384 0.01085 0.00570 0.00701 0.00552 -0.07148 0.00564
0.5 5 0.13209 0.08589 0.00947 0.00538 0.00871 0.00133 0.00729 0.00110
10 5 1 0.12558 0.08917 -0.11237 0.01364 -0.12415 0.01608 0.11328 0.02029
3 2 -0.11186 0.06539 -0.01501 0.00628 -0.02041 0.00641 -0.02901 0.00709
2 4 -0.10113 0.06314 -0.00301 0.00439 -0.00617 0.00433 -0.01820 0.00465
0.5 5 0.11845 0.07321 0.00211 0.00089 0.00453 0.00086 0.00315 0.00076
0.75 6 5 1 0.17269 0.10671 -0.18041 0.03485 -0.19013 0.03819 -0.19789 0.05107
3 2 0.14495 0.11425 -0.02737 0.00675 -0.05236 0.00700 0.04044 0.00786
2 4 0.13727 0.10991 0.02759 0.00617 0.03377 0.00680 0.03682 0.00923
0.5 5 0.13290 0.11231 0.01795 0.00222 0.01814 0.00214 0.02786 0.00728
8 5 1 0.12592 0.09042 -0.15980 0.02955 0.16805 0.02916 0.18938 0.04232
3 2 -0.12520 0.07051 -0.02599 0.00568 -0.04199 0.00612 0.05215 0.00734
2 4 0.12947 0.09929 0.01740 0.00562 0.02961 0.00582 -0.02351 0.00736
NAME OF AUTHOR1, AUTHOR2, AUTHOR3 ET AL.

0.5 5 0.11302 0.08131 0.01433 0.00182 0.01335 0.00174 0.01885 0.00234

10 5 1 -0.12679 0.07175 -0.11297 0.02318 -0.13501 0.02640 -0.17271 0.03915
3 2 -0.11129 0.06732 -0.01954 0.00459 -0.03385 0.00583 -0.04074 0.00651
2 4 0.12162 0.06734 0.00728 0.00419 0.00826 0.00508 0.00897 0.00518
0.5 5 0.10832 0.06674 0.00415 0.00060 0.00363 0.00088 0.00858 0.00073
[Vol. xx, No. x