Georgia Southern University

Digital Commons@Georgia Southern

Mathematical Sciences Faculty Publications Mathematical Sciences, Department of

9-2016

The Log-Logistic Weibull Distribution with Applications to Lifetime

Data
Broderick O. Oluyede
Georgia Southern University, [email protected]

Susan Foya
Botswana International University of Science and Technology

Gayan Warahena-Liyanage
Central Michigan University

Shujiao Huang
University of Houston

Follow this and additional works at: https://digitalcommons.georgiasouthern.edu/math-sci-facpubs

Part of the Education Commons, and the Mathematics Commons

Recommended Citation
Oluyede, Broderick O., Susan Foya, Gayan Warahena-Liyanage, Shujiao Huang. 2016. "The Log-Logistic
Weibull Distribution with Applications to Lifetime Data." Austrian Journal of Statistics, 45 (3): 43-69. doi:
10.17713/ajs.v45i3.107
https://digitalcommons.georgiasouthern.edu/math-sci-facpubs/356

This article is brought to you for free and open access by the Mathematical Sciences, Department of at Digital
Commons@Georgia Southern. It has been accepted for inclusion in Mathematical Sciences Faculty Publications by
an authorized administrator of Digital Commons@Georgia Southern. For more information, please contact
[email protected].
Austrian Journal of Statistics
September 2016, Volume 45, 43–69.
http://www.ajs.or.at/
AJS
doi:10.17713/ajs.v45i3.107

The Log-logistic Weibull Distribution with

Applications to Lifetime Data

Broderick O. Oluyede, Susan Foya, Gayan Warahena-Liyanage, Shujiao Huang

Georgia Southern University, Botswana International University of Science and Technology,
Central Michigan University, University of Houston

Abstract
In this paper, a new generalized distribution called the log-logistic Weibull (LLoGW)
distribution is developed and presented. This distribution contain the log-logistic Rayleigh
(LLoGR), log-logistic exponential (LLoGE) and log-logistic (LLoG) distributions as special
cases. The structural properties of the distribution including the hazard function, reverse
hazard function, quantile function, probability weighted moments, moments, conditional
moments, mean deviations, Bonferroni and Lorenz curves, distribution of order statistics, L-
moments and Rényi entropy are derived. Method of maximum likelihood is used to estimate
the parameters of this new distribution. A simulation study to examine the bias, mean
square error of the maximum likelihood estimators and width of the confidence intervals
for each parameter is presented. Finally, real data examples are presented to illustrate the
usefulness and applicability of the model.

Keywords: generalized distribution, log-logistic distribution, Weibull distribution, log-logistic

Weibull distribution, probability weighted moments, L-moments, maximum likelihood estima-
tion.

1. Introduction
There are several generalizations of univariate distributions including those of (Eugene, Lee,
and Famoye 2002) dealing with the beta-normal distribution, as well general family of uni-
variate distributions generated from the Weibull distribution that was introduced by Gurvich,
Dibenedetto, and Ranade (1997). The cumulative distribution function (cdf) given by (Gurvich
et al. 1997) is
G(x; α, Θ) = 1 − exp[−αH(x; Θ)], x ∈ C, α > 0, (1)
where C is a subset of R, and H(x; Θ) is a non-negative monotonically increasing function that
depends on the vector of parameters Θ. The corresponding probability density function (pdf)
is given by
g(x; α, Θ) = α exp[−αH(x; Θ)]h(x; Θ), (2)
where h(x; Θ) is the derivative of H(x; Θ). The choice of the function H(x; Θ) can lead to
44 The Log-logistic Weibull Distribution with Applications to Lifetime Data

different models including for example, exponential distribution with H(x; Θ) = x, Rayleigh
distribution is obtained from H(x; Θ) = x2 and Pareto distribution from setting H(x; Θ) =
log(x/k).
There are several ways of generating new probability distributions from classic ones to relative
new distributions in the literature. Nelson mentioned in (Nelson 1982) that distributions with
bathtub-shaped failure rate are sufficiently complex and, therefore, difficult to model. The
distribution proposed by (Hjorth 1980) is such an example. Later on, (Rajarshi and Rajarshi
1988) presented a revision of these distributions, and (Haupt and Schäbe 1992) put forward
a new lifetime model with bathtub-shaped failure rates. Unfortunately, these models are not
sufficient to address various practical situations, so new classes of distributions were presented
based on the modifications of the Weibull distribution to satisfy non-monotonic failure rate. For
a review of these models, the reader can refer to (Mudholkar and Srivastava 1993), and (Pham
and Lai 2007), where the authors summarized some generalizations of Weibull distribution in
their papers. Other generalizations include the exponentiated Weibull (EW) (Gupta and Kundu
2001), the modified Weibull (MW) (Lai, Xie, and Murthy 2003), and the beta exponential (BE)
(Nadarajah and Kotz 2006). Some more recent extensions are the generalized modified Weibull
(GMW) (Carrasco, Ortega, and Cordeiro 2008), the beta modified Weibull (BMW) (Silva,
Ortega, and Cordeiro 2010), (Nadarajah, Cordeiro, and Ortega 2011), the Weibull-G family
(Bourguignon, Silva, and Cordeiro 2014) and the Gamma-exponentiated Weibull distributions
(GEW) (Pinho, Cordeiro, and Nobre 2012). (Gurvich et al. 1997) developed a new statistical
distribution for characterizing the random strength of brittle materials.
To motivate the model under study, consider a series system and assume that the lifetime of
the components follow the log-logistic and Weibull distributions with with reliability functions
β
R1 (t) = (1 + ( st )c )−1 and R2 (t) = e−αt , respectively. The reliability R(t) = P (T > t) of the
system is given by
Y2
R(t) = Ri (t). (3)
i=1

In some context, a series model is referred to as a competing risk model.

Also, a primary motivation for developing this model is the advantages presented by this gener-
alized distribution with respect to having a hazard function that exhibits increasing, decreasing
and bathtub shapes, as well as the versatility and flexibility of the log-logistic and Weibull
distributions in modeling lifetime data. We propose and study this new distribution called the
log-logistic Weibull distribution which inherits these desirable properties and also covers quite
a variety of shapes.
There is an added advantage to this model, in that it has an additional dispersion parameter,
depending on the overall form that accounts for the scale of the underlying random variable.
The distribution also has exponential dumping in the upper tail making the distribution suitable
for modeling samples that display power behavior for intermediate observations and decrease in
tail probability for large observations or beyond a certain threshold or specified value.
The proposed new distribution generalizes the log-logistic and Weibull distributions. Some
structural properties of this distribution are obtained and estimation the parameters via the
method of maximum likelihood presented.
This paper is organized as follows. In section 2, we present the generalized distribution includ-
ing the corresponding probability density functions (pdf), hazard and reverse hazard functions,
quantile function and various sub-models. In section 3, the probability weighted moments, mo-
ments and conditional moments are presented. Section 4 contain the derivation of the mean
deviations, Bonferroni and Lorenz curves. Section 5 is concerned with Rényi entropy, distribu-
tion of order statistics and L-moments. Estimation of model parameters is presented in section
6. Monte Carlo simulation study is conducted in section 7 to examine the bias and mean square
 Austrian Journal of Statistics 45

error of the maximum likelihood estimators and the width of the confidence intervals for each
parameter. Applications of the proposed model to real data are given in section 8, followed by
concluding remarks.

2. The log-logistic Weibull distribution

In this section, we present some statistical properties of the new log-logistic Weibull (LLoGW)
distribution, including pdf, cdf, quantile function, hazard and reverse hazard functions. Plots
of the hazard rate function for selected values of the model parameters are also given. We
first of all present the Burr-XII, log-logistic and Weibull distributions. The very popular Burr
Type III and Type XII distributions attract special attention because they include several
families of distributions (e.g., the gamma distribution) with varying degrees of skewness and
kurtosis. Further, these distributions have applications in a wide variety of areas in statistics and
applied mathematics including modeling events associated with fracture roughness, life testing,
operational risk, option market price distributions, forestry, meteorology, modeling crop prices,
software reliability growth, and reliability analysis. See (Burr 1942), (Burr 1973) for additional
details. The cdf and pdf of Burr XII distribution are given by
  c −k
x
FB (x) = 1 − 1 + , (4)
s

and  c −k−1
kc x c−1
  
x
fB (x) = 1+ , for s, c, k, and x ≥ 0,
s s s
respectively. The reliability and hazard rate functions are given by
 c −k  c −1
kc x c−1
   
x x
F B (x) = 1+ , and hFB (x) = 1+ ,
s s s s

respectively. Note that the pdf is unimodal with mode at x0 = ((c − 1)/(ck + 1))1/c when c > 1,
and L−shaped when c = 1. The rth non-central moment is given by

E(X r ) = ksr B(k − rc−1 , 1 + rc−1 ), for ck > r.

Note that k and c are shape parameters and s is a scale parameter. When k = 1 we obtain the
log-logistic distribution. The cdf of the well known Weibull (W) distribution is given by

FW (x; α, β) = 1 − exp(−αxβ ), x ≥ 0, α > 0, β > 0,

where α and β are scale and shape parameter, respectively.

Now, consider the log-logistic Weibull (LLoGW) distribution obtained by taking R1 (x) =
F 1 (x) = (1 + ( xs )c )−1 and R2 (x) = F 2 (x) = exp(−αxβ ) in equation (1) to obtain the new
LLoGW cdf G(x) = G(x; s, c, α, β) given by
  x c −1
G(x) = 1 − 1 + exp(−αxβ ), (5)
s
for s, c, α, β > 0 and x ≥ 0. If a random variable X has the LLoGW cdf, we write X ∼
LLoGW (s, c, α, β). The corresponding LLoGW pdf is given by
 c −1 
cxc−1
 
−αxβ x β−1
g(x) = e 1+ αβx + c , (6)
s (s + xc )
46 The Log-logistic Weibull Distribution with Applications to Lifetime Data

s, c, α, β > 0, and x ≥ 0. Plots of the pdf for selected values of the model parameters are given
in Figure 1. The plots suggests that the LLoGW pdf can be right skewed or decreasing for the
selected values of the parameters.

3.0
s=2.0,c=2.0,α=4.0,β=0.2
s=4.0,c=3.0,α=5.0,β=4.0
s=1.0,c=0.5,α=0.5,β=2.0
s=0.3,c=3.0,α=1.0,β=3.0
2.5

s=0.5,c=2.0,α=2.0,β=1.0
2.0
density

1.5
1.0
0.5
0.0

0.0 0.5 1.0 1.5 2.0

Figure 1: Plots of LLoGW pdf

2.1. Quantile function

The LLoGW quantile function can be obtained by inverting G(x) = 1 − u, where G(x) = u,
0 ≤ u ≤ 1, and
  x c −1 β
G(x) = 1+ e−αx . (7)
s
The quantile function of the LLoGW distribution is obtained by solving the equation
  x c 
αxβ + log 1 + + log(1 − u) = 0, (8)
s
using numerical methods. Consequently, random number can be generated based on equation
(8). Table 1 lists the quantile for selected values of the parameters of the LLoGW distribution.

Table 1: LLoGW quantile for selected values

(s, c, α, β)
u (1.5,1.5,0.5,0.5) (1.5,0.5,1.5,0.5) (0.5,0.5,1.5,1.5) (0.3,1.0,0.3,0.8) (1.0,1.0,2.0,2.0)
0.1 0.1388 0.0021 0.0061 0.0277 0.0923
0.2 0.2296 0.0095 0.0290 0.0629 0.1754
0.3 0.3148 0.0247 0.0744 0.1075 0.2547
0.4 0.4012 0.0516 0.1436 0.1656 0.3338
0.5 0.4936 0.0968 0.2351 0.2444 0.4157
0.6 0.5972 0.1730 0.3489 0.3574 0.5040
0.7 0.7202 0.3068 0.4907 0.5339 0.6046
0.8 0.8800 0.5675 0.6779 0.8520 0.7287
0.9 1.1292 1.2234 0.9688 1.6268 0.9098
 Austrian Journal of Statistics 47

2.2. Some new and known sub-models

There are several new as well as well known distributions that can be obtained from the LLoGW
distribution. Note that when s = m−1 , we have the log-logistic Weibull (LLoGW) distribution
β
with the survival function S(x; m, c, α, β) = [1 + (xm)c ]−1 e−αx . When c = 1 it reduces to the
generalized Pareto type II Weibull (GP-II-W) distribution.

• If β = 1, we obtain the log-logistic exponential (LLoGE) distribution.

• If β = 2, we have the log-logistic Rayleigh (LLoGR) distribution.
• When α → 0, we have the log-logistic (LLoG) distribution.
• If c = 1 and s → ∞, we obtain Weibull (W) distribution.
• If c = 1, s → ∞, and β = 2 we have Rayleigh (R) distribution.
• If c = 1, s → ∞, and β = 1, we have the exponential (E) distribution.
• If c = 1, then the LLoGW cdf reduces to the three parameter distribution with cdf given
by
  −1
x
G(x) = 1 − 1 + exp(−αxβ ), (9)
s
for s, α, β > 0, and x ≥ 0.
• If c = β = 1 then the LLoGW cdf reduces to to the two parameter distribution given by
  −1
x
G(x) = 1 − 1 + exp(−αx), (10)
s
for s, α > 0, and x ≥ 0.
• If c = 1 and β = 2, then the LLoGW cdf reduces to the two parameter model
  −1
x
G(x) = 1 − 1 + exp(−αx2 ), (11)
s
for s, α > 0, and x ≥ 0.

2.3. Hazard and reverse hazard functions

In general, if X is a continuous random variable with cdf F, and pdf f, then the hazard function,
reverse hazard function and mean residual life function are given by
R∞
f (x) f (x) F (u)du
λF (x) = , τF (x) = , and δF (x) = x ,
F (x) F (x) F (x)
respectively. The functions λF (x), δF (x), and F (x) are equivalent. See (Shaked and Shanthiku-
mar 1994) and references therein. In this subsection, the hazard and reverse hazard functions of
the LLoGW distribution are presented. The hazard and reverse hazard functions of the LLoGW
distribution are
cxc−1
 
β−1
hG (x) = αβx + c ,
(s + xc )
and
 c −1 −1  c −1 
cxc−1
   
x −αxβ −αxβ x β−1
τG (x) = 1 − 1 + e e 1+ αβx + c (12)
s s (s + xc )
for x ≥ 0, s, c, α, β > 0, respectively. The limiting behavior of the hazard function of the
LLoGW distribution, which can be readily established is as follows:
 48 The Log-logistic Weibull Distribution with Applications to Lifetime Data

• For β < 1 and c = 1, limx→0 hG (x) = ∞ and limx→∞ hG (x) = 0.

• For β = 1, (
∞ 0 < c < 1,
lim h (x) =
x→0 G α c > 1.

• For β = 1, and for each c > 0, limx→∞ hG (x) = α.

• For β = 1 and c = 1, limx→0 hG (x) = α + 1

s and limx→∞ hG (x) = α.

• For β < 1 and c < 1, limx→0 hG (x) = ∞ and limx→∞ hG (x) = 0.

• For β > 1 and c > 1, limx→0 hG (x) = 0 and limx→∞ hG (x) = ∞.

• For β > 1 and c = 1, limx→0 hG (x) = 1

s and limx→∞ hG (x) = ∞.

Plots of the hazard function are given in Figure 2. The graphs exhibit increasing, decreasing,
bathtub, upside down bathtub, and upside down bathtub followed by bathtub shapes for the
selected values of the model parameters. This very attractive flexibility makes the LLoGW
hazard function useful and suitable for non-monotonic empirical hazard behaviors which are
more likely to be encountered in practice or real life situations.
10

s=2.0,c=2.0,α=4.0,β=0.2
s=4.0,c=3.0,α=5.0,β=4.0
s=1.0,c=0.5,α=0.5,β=2.0
s=0.3,c=3.0,α=1.0,β=3.0
s=0.5,c=2.0,α=2.0,β=1.0
8
6
h(x)

4
2
0

0 1 2 3 4

Figure 2: LLoGW hazard functions

3. Probability weighted moments, moments and conditional moments

In this section, we obtain probability weighted moments (PWMs) (Greenwood, Landwehr, Mata-
las, and Wallis 1979), moments and conditional moments for the LLoGW distribution. The
probability weighted moments (PWMs) of the LLoGW distribution is given by
Z ∞
m m
E(X r Gl (X)G (X)) = xr Gl (x)G (x)g(x)dx
0
∞
X (−1)j Γ(l + 1)E(X r (G(X))j+m )
= .
Γ(l + 1 − j)Γ(j + 1)
j=0
 Austrian Journal of Statistics 49

Now, by setting y = (1 + (x/s)c )−1 , the PWMs of the LLoGW distribution can be written as:
∞
m X (−1)j+k Γ(l + 1)[α(j + m + 1)]k
E(X r Gl (X)G (X)) =
Γ(l + 1 − j)Γ(j + 1)k!
j,k=0
Z ∞   x c −(j+m+2) c  x c−1
× xr+kβ 1 + dx
0 s s s
Z ∞   x c −(j+m+1) 
r+kβ+β−1
+ αβ x 1+ dx
0 s
∞
X (−1)j+k Γ(l + 1)[α(j + m + 1)]k
=
Γ(l + 1 − j)Γ(j + 1)k!
j,k=0
 Z 1
r+kβ r+kβ
× sr+kβ y j+m−( c ) (1 − y) c dy
0
r+kβ+β Z 1 
αβs j+m−( r+kβ+β ) r+kβ+β
−1
+ y c (1 − y) c dy .
c 0
Consequently, the PWMs of the LLoGW distribution is given by
∞
m X (−1)j+k Γ(l + 1)[α(j + m + 1)]k sr+kβ
E(X r Gl (X)G (X)) =
Γ(l + 1 − j)Γ(j + 1)k!
j,k=0
    
r + kβ r + kβ + c
× B j+m+1− ,
c c
β
   
αβs r + kβ + β r + kβ + β
+ B j+m+1− , .
c c c
Remarks: special cases
• When l = m = 0, we obtain the rth non-central moment µ0r given by
∞
(−1)k αk skβ+r αβsβ
  
0 r
X kβ + β + r kβ + β + r
µr = E(X ) = B 1− ,
k! c c c
k=0
 
kβ + r kβ + r
+ B 1− ,1 + ,
c c
R1
where B(a, b) = 0 y a−1 (1 − y)b−1 dy is the beta function.
• When r = l = 0, we have
∞
m X (−1)k [α(m + 1)]k skβ
E(G (X)) =
k!
k=0
    
kβ kβ + c
× B m+1− ,
c c
β
   
αβs kβ + β kβ + β
+ B m+1− , .
c c c
• When l = 0, the LLoGW PWMs reduces to
∞
m X (−1)k [α(m + 1)]k sr+kβ
E(X r G (X)) =
k!
k=0
    
r + kβ r + kβ + c
× B m+1− ,
c c
β
   
αβs r + kβ + β r + kβ + β
+ B m+1− , .
c c c
50 The Log-logistic Weibull Distribution with Applications to Lifetime Data

• When m = 0, the LLoGW PWMs reduces to

∞
r l
X (−1)j+k Γ(l + 1)[α(j + 1)]k sr+kβ
E(X G (X)) =
Γ(l + 1 − j)Γ(j + 1)k!
j,k=0
    
r + kβ r + kβ + c
× B j+1− ,
c c
β
   
αβs r + kβ + β r + kβ + β
+ B j+1− , .
c c c

• When r = m = 0, we have
∞
X (−1)j+k Γ(l + 1)[α(j + 1)]k skβ
E(Gl (X)) =
Γ(l + 1 − j)Γ(j + 1)k!
j,k=0
    
kβ kβ + c
× B j+1− ,
c c
β
   
αβs kβ + β kβ + β
+ B j+1− , .
c c c

• When r = 0, we have
∞
l m
X (−1)j+k Γ(l + 1)[α(j + m + 1)]k skβ
E(G (X)G (X)) =
Γ(l + 1 − j)Γ(j + 1)k!
j,k=0
    
kβ kβ + c
× B j+m+1− ,
c c
β
   
αβs kβ + β kβ + β
+ B j+m+1− , .
c c c

Note that the rth raw moment µ0r can also be obtained as follows:
Z ∞
µ0r = E(X r ) = xr g(x; s, c, α, β)dx
Z0 ∞   x c −1 β
= αβxr+β−1 1 + e−αx dx
s
Z0 ∞  x c −2
c r+c−1  β
+ c
x 1+ e−αx dx.
0 s s
Let y = (x/s)c , and u = y β/c then
αβsr+β ∞ r+β −1
Z Z ∞
β y β/c r β β/c
−1
r
E(X ) = y c [1 + y] e αs
dy + s r
y c [1 + y]−2 eαs y dy
c
Z ∞0 0
r+β βu
r+β −1 c/β −1 −αs
= αs u β [1 + u ] e du
0
csr ∞ r+c
Z
−1 β
+ u β [1 + uc/β ]−2 e−αs u du.
β 0
We apply the following results which holds for m and k positive integers, w > −1 and p > 0
((Prudnikov, Brychkov, and Marichev 1992), page 21):
Z ∞
m m
I(p, w, , ν) = xw [1 + x k ]ν e−px dx
k 0
k −ν mw+1/2
 m 
k,k+m m  ∆(m, −w), ∆(k, ν + 1)
= m−1 Gk+m,k  ,
(2π) 2 Γ(−ν)pw+1 pm  ∆(k, 0)
 Austrian Journal of Statistics 51

where ∆(k, a) = ka , a+1 a+k

k , ...., k , and the Meijer G function is defined by
Qm
Γ(bj + t) nj=1 Γ(1 − aj − t)
   Z Q
a1 , .... , ap 1 j=1
m,n
Gp,q x  b1 , .... , bq = 2πi Qp Qp x−t dt,
L j=n+1 Γ(a j + t) j=m+1 Γ(1 − b j − t)
√
where i = −1 is the complex unit and L is the integration path (see (Gradshtein and Ryzhik
2000), sec. 9.3 for a description of this path).
Consequently, the rth moment of the LLoGW distribution is given by
r
+ 12
cc  ∆(c, − βr ), ∆(β, 0)
  
βc β
r
E(X ) = c−1 r Gβ,β+c
β+c,β

(2π) 2 α β (αsβ )c  ∆(β, 0)
r+c
+1
cc  ∆(c, 1 − r+c
  
+
βc β 2
G β,β+c  β ), ∆(β, −1) .
c−1 r+c β+c,β (αsβ )c  ∆(β, 0)
(2π) 2 α β sc

Alternatively, the following direct computation also gives the rth moment of the LLoGW dis-
(−1)k αk xkβ
tribution when we use the fact that e−αx = ∞
β P
k=0 k! .
Theorem 3.1. The rth raw moment µ0r = E(X r ) of the LLoGW distribution is
∞
(−1)k αk skβ+r αβsβ
  
r
X kβ + β + r kβ + β + r
E(X ) = B 1− ,
k! c c c
k=0
 
kβ + r kβ + r
+ B 1− ,1 + ,
c c
R1
where B(a, b) = 0 y a−1 (1 − y)b−1 dy is the beta function.

Proof: Note that

Z ∞
r
E(X ) = xr g(x; s, c, α, β)dx
Z0 ∞   x c −1 β
= αβxr+β−1 1 + e−αx dx
s
Z0 ∞  x c −2
c r+c−1  β
+ c
x 1 + e−αx dx
0 s s

β P∞ (−1)k αk xkβ
Let y = (1 + ( xs )c )−1 , and apply e−αx = k=0 k! , to obtain
∞ 1
(−1)k αk+1 βskβ+β+r
Z
X kβ+β+r kβ+β+r
r −1 −1
E(X ) = y 1− c (1 − y) c dy
k!c 0
k=0
∞ 1
(−1)k αk skβ+r
Z
X kβ+r kβ+r
−1
+ y 1− c (1 − y) c
+1−1
dy
k! 0
k=0
∞
(−1)k αk skβ+r
αβsβ
  
X kβ + β + r kβ + β + r
= B 1− ,
k! c c c
k=0
 
kβ + r kβ + r
+ B 1− ,1 + , (13)
c c
for c > kβ+β+r. To obtain the moment generating function (MGF) of the LLoGW distribution,
(tx)j
recall the Taylor’s series expansion of the function etx = ∞
P
j=0 j! , so that, we have MX (t) =
tn
E(etX ) = ∞ n n
P
n=0 n! E(X ), where E(X ) is given above.
52 The Log-logistic Weibull Distribution with Applications to Lifetime Data

Table 2 lists the first six moments together with the standard deviation (SD or σ), coefficient
of variation (CV), coefficient of skewness (CS) and coefficient of kurtosis (CK) of the LLoGW
distribution for selected values of the parameters, by fixing α = 1.5 and β = 1.5. And Table 3
lists the first six moments, SD, CV, CS and CK of the LLoGW distribution for selected values
of the parameters, by fixing s = 1.5 and c = 1.0. These values can be p determined numerically
using R and MATLAB. The SD, CV, CS and CK are given by σ = µ02 − µ2 ,
p 0 s
σ µ2 − µ2 µ02
CV = = = − 1,
µ µ µ2
E (X − µ)3
 
µ03 − 3µµ02 + 2µ3
CS = = ,
[E(X − µ)2 ]3/2 (µ02 − µ2 )3/2
and
E (X − µ)4
 
µ04 − 4µµ03 + 6µ2 µ02 − 3µ4
CK = = ,
[E(X − µ)2 ]2 (µ02 − µ2 )2
respectively.

Table 2: Moments of the LLoGW distribution for some parameter values; α = 1.5 and β = 1.5.

µ0s s = 0.5, c = 0.5 s = 0.5, c = 1.5 s = 1.5, c = 0.5 s = 1.5, c = 1.5

µ01 0.3784 0.4060 0.4627 0.5777
µ02 0.3181 0.2696 0.4099 0.4973
µ03 0.3680 0.2485 0.4871 0.5543
µ04 0.5244 0.2916 0.7062 0.7482
µ05 0.8740 0.4124 1.1918 1.1759
µ06 1.6522 0.6780 2.2751 2.0961
SD 0.4181 0.3237 0.4425 0.4044
CV 1.1049 0.7972 0.9565 0.7001
CS 1.5772 1.5918 1.3412 1.1800
CK 5.8595 6.6587 5.0516 4.8231

Table 3: Moments of the LLoGW distribution for some parameter values; s = 1.5 and c = 1.0.

µ0s α = 0.2, β = 1.8 α = 2.5, β = 0.8 α = 1.0, β = 1.0 α = 0.1, β = 3.5

µ01 1.2593 0.2890 0.6724 1.1314
µ02 2.7456 0.2143 0.9828 1.8173
µ03 7.9772 0.2744 2.2886 3.3807
µ04 28.0396 0.5190 7.4228 6.8809
µ05 113.5654 1.3309 31.0823 14.9509
µ06 514.8466 4.3789 160.0520 34.2132
SD 1.0770 0.3617 0.7285 0.7330
CV 0.8552 1.2517 1.0835 0.6479
CS 1.2797 2.8920 2.3639 0.2764
CK 4.7851 16.8416 11.7875 2.1587

3.1. Conditional moments

For lifetime models, it is also of interest to find the conditional moments and the mean residual
 Austrian Journal of Statistics 53

lifetime function. The rth conditional moments for LLoGW distribution is given by
Z ∞
r 1
E(X |X > t) = xr gLLoGW (x)dx
G(t) t
∞ Z ∞  c −1
(−1)k αk
X 
1 kβ+r+β−1 x
= αβ x 1+ dx
G(t) k=0 k! t s
∞  c −2 
(−1)k cαk ∞ kβ+r+c−1
Z 
X x
+ c
x 1+ dx . (14)
s k! t s
k=0

Let y = (1 + (x/s)c )−1 , then

∞
(−1)k+1 αk+1 β kβ+β+r
X
r 1
E(X |X > t) = s
G(t) k=0 k!
Z 1
kβ+β+r kβ+β+r
× y 1− c −1 (1 − y) c −1 dy
(1+(t/s)c )−1
∞
X (−1)k+1 αk β Z 1 
kβ+r kβ+r
−1 −1
+ skβ+r y 1− c (1 − y)1+ c dy
k! (1+(t/s)c )−1
k=0
∞
(−1)k+1 αk βskβ+r
X
1
=
G(t) k=0 k!
   
β kβ + β + r kβ + β + r
× αβs B 1− ,
c c
 
kβ + β + r kβ + β + r
− B(1+(t/s)c )−1 1 − ,
c c
  
kβ + r kβ + r
+ B 1− ,
c c
 
kβ + r kβ + r
− B(1+(t/s)c )−1 1 − ,1 + ,
c c
Rx
where Bx (a, b) = 0 y a−1 (1 − y)b−1 dy is the incomplete beta function, and c > kβ + β + r.
Alternatively, consider the following integral for q > 0 and b > 0, (Paranaı́ba, Ortega, Cordeiro,
and Pescim 2011),
Z ∞
J(q, a, b) = y a [1 + y]−b dy
q
a+1
 
2 F1 (b, a + 1; a + 2; −q)q Γ(b − a − 1)π(πa)
= − + , (15)
a+1 Γ(b)Γ(−a)
where 2 F1 is the hypergeometric function given by
∞
X (a)k (b)k xk
2 F1 (a, b; c; x) = , (16)
(c)k k!
k=0

Rand (a)k = a(a + 1)....(a + k − 1) is the ascending factorial. Now, consider the integral
∞ kβ+r+β−1
t x [1 + (x/s)c ]−1 dx and let y = (x/s)c , then
∞ ∞
skβ+r+β
Z Z
kβ+r+β
c −1 −1
x kβ+r+β−1
[1 + (x/s) ] dx = y c [1 + y]−1 dy
t c (t/s)c
 c
skβ+r+β

t kβ + r + β − c
= J , ,1 .
c s c
54 The Log-logistic Weibull Distribution with Applications to Lifetime Data

Consequently, the rth conditional moment of the LLoGW distribution is given by

∞  c
(−1)k αk skβ+r
X  
r 1 β t kβ + r + β − c
E(X |X > t) = αβs J , ,1
G(t) k=0 k! c s c
 c 
c t kβ + r
+ sJ , ,2 .
s c

The mean residual lifetime function of the LLoGW distribution is E(X|X > t) − t.

4. Mean deviations
The amount of scatter in a population is evidently measured to some extent by the totality of
deviations from the mean and median. These are known as the mean deviation about the mean
and the mean deviation about the median, and defined by

Z ∞ Z ∞
δ1 (x) = |x − µ|g(x)dx and δ2 (x) = |x − M |g(x)dx, (17)
0 0

respectively, where µ = E(X) is the mean and M =Median (X) is the median. The measures
δ1 (x) and δ2 (x) can be calculated using the relationships

Z ∞
δ1 (x) = 2µG(µ) − 2µ + 2 xg(x)dx, (18)
µ

and
Z ∞
δ2 (x) = −µ + 2 xg(x)dx, (19)
M

R∞
respectively. When r = 1, we get the mean µ = E(X). Note that T (µ) = µ xg(x)dx and
R∞
T (M ) = M xg(x)dx are given by

∞  c
(−1)k αk skβ+r
X  
β µ kβ + r + β − c
T (µ) = αβs J , ,1
k! c s c
k=0
 c 
c µ kβ + r
+ sJ , ,2
s c

and

∞  c
(−1)k αk skβ+r
X  
β M kβ + r + β − c
T (M ) = αβs J , ,1
k! c s c
k=0
 c 
c M kβ + r
+ sJ , ,2 ,
s c
 Austrian Journal of Statistics 55

respectively. Alternatively,
Z ∞
T (µ) = xg(x)dx
µ
∞
(−1)k+1 αk βskβ+r
X
=
k!
k=0
   
β kβ + β + r kβ + β + r
× αβs B 1− ,
c c
 
kβ + β + r kβ + β + r
− B(1+(µ/s)c )−1 1 − ,
c c
  
kβ + r kβ + r
+ B 1− ,
c c
 
kβ + r kβ + r
− B(1+(µ/s)c )−1 1 − ,1 + .
c c

Consequently, the mean deviation about the mean is

δ1 (x) = 2µG(µ) − 2µ + 2T (µ)

and the mean deviation about the median is

δ2 (x) = −µ + 2T (M ).

4.1. Bonferroni and Lorenz curves

In this subsection, we present Bonferroni and Lorenz Curves. Bonferroni and Lorenz curves
have applications not only in economics for the study income and poverty, but also in other
fields such as reliability, demography, insurance and medicine. Bonferroni and Lorenz curves
for the LLoGW distribution are given by
Z q
1 1
B(p) = xg(x)dx = [µ − T (q)],
pµ 0 pµ

and Z q
1 1
L(p) = xg(x)dx = [µ − T (q)],
µ 0 µ
R∞
respectively, where T (q) = q xg(x)dx, and q = G−1 (p), 0 ≤ p ≤ 1.

5. Order statistics, L-moments and Rényi entropy

The concept of entropy plays a vital role in information theory. The entropy of a random variable
is defined in terms of its probability distribution and can be shown to be a good measure of
randomness or uncertainty. In this section, we present the distribution of the order statistic,
L-moments and Rényi entropy for the LLoGW distribution.

5.1. Order statistics

Order statistics play an important role in probability and statistics. In this subsection, we
present the distribution of the ith order statistic from the LLoGW distribution. The pdf of the
56 The Log-logistic Weibull Distribution with Applications to Lifetime Data

ith order statistic from the LLoGW pdf g(x) is given by

n!g(x)
gi:n (x) = [G(x)]i−1 [1 − G(x)]n−i
(i − 1)!(n − i)!
n−i  
n!g(x) j n−i
X
= (−1) [G(x)]j+i−1
(i − 1)!(n − i)! j
j=0
Pn−i n−i
by using the binomial expansion [1 − G(x)]n−i = (−1)m [G(x)]m . Consequently,


m=0 m

n−i 
n − i (−1)m

1 X
gi:n (x) = (m + i)[G(x)]m+i−1 g(x)
B(i, n − i + 1) m m+i
m=0
n−i
X
= wi,m gm+i (x),
m=0

where gm+i (x) is the pdf of the exponentiated LLoGW distribution with parameters s, c, α, β
and (m + i), B(., .) is the beta function and the weights wi,m are given by
(−1)m n − i
    
1 m+i−1 n
wi,m = = (−1)m .
B(i, n − i + 1) m + i m m m+i

The tth moment of the ith order statistics from the LLoGW distribution can be derived via a
result of (Barakat and Abdelkader 2004) as follows:
n    Z ∞
p−n+i−1 p − 1 n
X
t
E(Xi:n ) = t (−1) xt−1 [1 − G(x)]p dx. (20)
n−i p 0
p=n−i+1

Note that
∞ ∞ ∞
(−1)k (pα)k
Z X Z
x t−1 p
[1 − G(x)] dx = xkβ+t−1 [1 + (x/s)c ]−p dx.
0 k! 0
k=0

Let y = [1 + (x/s)c ]−1 , then

Z ∞
skβ+t 1 p− kβ − t −1
Z
kβ+t
c −p
x kβ+t−1
[1 + (x/s) ] dx = y c c (1 − y) c −1 dy.
0 c 0

Now,
n ∞ k p−1
   kβ+t
p−n+i+k (pα) n s
XX
t
E(Xi:n ) = t (−1)
k! n−i p c
p=n−i+1 k=0
 
kβ + t kβ + t
× B p− , , (21)
c c
R1
where B(a, b) = 0 ta−1 (1 − t)b−1 dt is the beta function.

5.2. L-moments
The L−moments (Hoskings 1990) are expectations of some linear combinations of order statistics
and they exist whenever the mean of the distribution exits, even when some higher moments
may not exist. They are relatively robust to the effects of outliers and are given by
k  
1 X k
λk+1 = (−1)j E(Xk+1−j:k+1 ), k = 0, 1, 2, ....... (22)
k+1 j
j=0
 Austrian Journal of Statistics 57

The L−moments of the LLoGW distribution can be readily obtained from equation (21). The
first four L−moments are given by λ1 = E(X1:1 ), λ2 = 21 E(X2:2 −X1:2 ), λ3 = 13 E(X3:3 −2X2:3 +
X1:3 ) and λ4 = 14 E(X4:4 − 3X3:4 + 3X2:4 − X1:4 ), respectively.

5.3. Rényi entropy

In this subsection, Rényi entropy of the LLoGW distribution is derived. An entropy is a measure
of uncertainty or variation of a random variable. Rényi entropy is an extension of Shannon
entropy and is defined to be
Z ∞ 
1 v
IR (v) = log [g(x; s, c, α, β)] dx , v 6= 1, v > 0. (23)
1−v 0

Rényi entropy tends to Shannon entropy as v → 1. Note that [g(x; s, c, α, β)]v = g v (x) can be
written as
 c−1   c −2   c −1 v
cx x −αxβ −β−1 x −αxβ
g v (x) = 1 + e + αβx 1 + e
sc s s
v c −2k
cxc−1
X v      
x β
= 1+ e−kαx
k sc s
k=0
   c −1 v−k
β−1 x −αxβ
× αβx 1+ e .
s
Now,
∞ ∞  
v X
v (−1)m (vα)m ck (αβ)v−k
Z X
g v (x)dx =
0 k m!sck
k=0 m=0
Z ∞   c −v−k
mβ+vβ−kβ+ck−v x
× x 1+ dx
0 s
∞
v X
v (−1)m (vα)m ck−1 (αβ)v−k mβ+vβ−kβ−v+1
X  
= s
k m!
k=0 m=0
Z 1
v+ v+kβ−vβ−mβ−1 −1 mβ+vβ−kβ−v+1
× y c (1 − y) c dy
0
∞
v X
v (−1)m (vα)m ck−1 (αβ)v−k mβ+vβ−kβ−v+1
X  
= s
k m!
k=0 m=0
 
v + kβ − vβ − mβ − 1 mβ + vβ − kβ − v + 1
× B v+ ,1 + .
c c
Consequently, Rényi entropy is given by
v X ∞  
v (−1)m (vα)m ck−1 (αβ)v−k mβ+vβ−kβ−v+1
  X
1
IR (v) = log s
1−v k m!
k=0 m=0
 
v + kβ − vβ − mβ − 1 mβ + vβ − kβ − v + 1
× B v+ ,1 + , (24)
c c
for v 6= 1, and v > 0.

6. Maximum likelihood estimation

Let X ∼ LLoGW (s, c, α, β) and ∆ = (s, c, α, β)T be the parameter vector. The log-likelihood
58 The Log-logistic Weibull Distribution with Applications to Lifetime Data

` = `(∆) for a single observation x of X is given by

cxc−1
 
`(∆) = −αxβ − log(1 + (x/s)c ) + log αβxβ−1 + c . (25)
s + xc
The first derivative of the log-likelihood function with respect to ∆ = (s, c, α, β)T are given by
∂` βxβ−1
= −xβ + c−1 ,
∂α αβxβ−1 + [scxc +xc ]

∂` αxβ + αβxβ−1 log(x)

= −αxβ log(x) + ,
∂β αβxβ−1 + cxc−1 [sc + xc ]−1

∂` c(x/s)c−1 (sc + xc )−2

= −(1 + (x/s)c )−1 (c/s)(x/s)c − ,
∂s αβxβ−1 + cxc−1 (sc + xc )−1
and
∂` (1 + (x/s)c )−2 (x/s)c log(x/s)
= −
∂c (1 + (x/s)c )−1
(sc + xc )−1 (xc−1 + cxc−1 log(x) + cxc−1 − [sc + xc ]−2 (sc log(s) + xc log(x)))
+ .
αβxβ−1 + cxc−1 (sc + xc )−1
The total log-likelihood function based on a random sample of n observations: Pn x1 , x2 , ...., xn
drawn from the LLoGW distribution is given by `n = L(∆) = ∗
i=1 `i (∆), where `i (∆),
i = 1, 2, ....., n is given by equation (25). The equations obtained by setting the above partial
derivatives to zero are not in closed form and the values of the parameters s, c, α, β must be found
by using iterative methods. The maximum likelihood estimates of the parameters, denoted
by ∆ ˆ is obtained by solving the nonlinear equation ( ∂` , ∂` , ∂` , ∂` )T = 0, using a numerical
∂s ∂c ∂α ∂β
method such as Newton-Raphson procedure. The Fisher information matrix is given by I(∆) =
2`
[Iθi ,θj ]4X4 = E(− ∂θ∂i ∂θ j
), i, j = 1, 2, 3, 4, can be numerically obtained by MATLAB, SAS or R
software, where (θ1 , θ2 , θ3 , θ4 ) = (s, c, α, β). The total Fisher information matrix nI(∆) can be
approximated by
2`
 
ˆ ≈ − ∂
Jn (∆) , i, j = 1, 2, 3, 4. (26)
∂θi ∂θj ∆=∆ ˆ 4X4

For a given set of observations, the matrix given in equation (26) is obtained after the conver-
gence of the Newton-Raphson procedure.

6.1. Asymptotic confidence intervals

In this subsection, we present the asymptotic confidence intervals for the parameters of the
LLoGW distribution. The expectations in the Fisher Information Matrix (FIM) can be obtained
numerically. Let ∆ ˆ = (ŝ, ĉ, α̂, β̂) be the maximum likelihood estimate of ∆ = (s, c, α, β). Under
the usual regularity conditions and that the parameters are in the interior of the parameter
√ ˆ d
space, but not on the boundary, we have: n(∆ − ∆) −→ N4 (0, I −1 (∆)), where I(∆) is the
expected Fisher information matrix. The asymptotic behavior is still valid if I(∆) is replaced
by the observed information matrix evaluated at ∆, ˆ that is J(∆).
ˆ The multivariate normal
ˆ −1
distribution N4 (0, J(∆) ), where the mean vector 0 = (0, 0, 0, 0)T , can be used to construct
confidence intervals and confidence regions for the individual model parameters and for the
survival and hazard rate functions. That is, the approximate 100(1 − η)% two-sided confidence
intervals for s, c, α, and β are given by:
q q q q
−1 −1
sb ± Z η Iss (∆), b
b c ± Z η Icc (∆), α
b b ± Z η I−1
αα (∆) and β ± Z η
b b I−1
ββ (∆),
b
2 2 2 2
 Austrian Journal of Statistics 59

−1 b −1 ˆ
respectively, where I−1 −1 b −1 b
ss (∆), Icc (∆), Iαα (∆), and Iββ (∆), are the diagonal elements of In (∆) =
b
ˆ −1 , and Z η is the upper η th percentile of a standard normal distribution.
(nI(∆))
2 2

7. Simulation study
In this section, we study the performance and accuracy of maximum likelihood estimates of
the LLoGW model parameters by conducting various simulations for different sample sizes and
different parameter values. Equation (8) is used to generate random data from the LLoGW
distribution. The simulation study is repeated N = 5, 000 times, each with sample size n =
25, 50, 75, 100, 200, 400 and parameter values I : s = 5.5, c = 2.5, α = 0.8, β = 0.2 and II : s =
5.5, c = 8.5, α = 0.5, β = 0.5. Four quantities are computed in this simulation study.

(b) Average bias of the MLE ϑ̂ of the parameter ϑ = s, c, α, β :

N
1 X
(ϑ̂ − ϑ).
N
i=1

(b) Root mean squared error (RMSE) of the MLE ϑ̂ of the parameter ϑ = s, c, α, β :

v
u
u1 XN
t (ϑ̂ − ϑ)2 .
N
i=1

(c) Coverage probability (CP) of 95% confidence intervals of the parameter ϑ = s, c, α, β, i.e.,
the percentage of intervals that contain the true value of parameter ϑ.

(d) Average width (AW) of 95% confidence intervals of the parameter ϑ = s, c, α, β.

Table 4 presents the Average Bias, RMSE, CP and AW values of the parameters s, c, α and β
for different sample sizes. From the results, we can verify that as the sample size n increases,
the RMSEs decay toward zero. We also observe that for all the parametric values, the biases
decrease as the sample size n increases. Also, the table shows that the coverage probabilities
of the confidence intervals are quite close to the nominal level of 95% and that the average
confidence widths decrease as the sample size increases. Consequently, the MLE’s and their
asymptotic results can be used for estimating and constructing confidence intervals even for
reasonably small sample sizes.
60 The Log-logistic Weibull Distribution with Applications to Lifetime Data

Table 4: Monte Carlo simulation results: average bias, RMSE, CP and AW

I II

Parameter n Average Bias RMSE CP AW Average Bias RMSE CP AW

s 25 0.52811 2.64774 0.80320 7.31782 0.03521 0.63589 0.81620 1.84526

50 0.32822 1.61445 0.88920 5.31108 0.01472 0.41423 0.89860 1.41529
75 0.25455 1.25934 0.91700 4.27320 0.01039 0.32745 0.91660 1.17925
100 0.20528 1.05246 0.92200 3.72238 0.00437 0.27415 0.92720 1.00656
200 0.08874 0.66811 0.94000 2.55273 0.00214 0.18193 0.94120 0.70739
400 0.07761 0.47016 0.94740 1.79814 0.00097 0.12919 0.94140 0.49884

c 25 2.42437 11.61202 0.93080 8.93786 5.95590 17.41663 0.93140 24.84591

50 0.68223 5.33933 0.94160 4.00937 1.67338 5.23907 0.94320 12.11561
75 0.36400 1.06252 0.94480 2.81549 0.90145 4.07377 0.94500 8.94710
100 0.36400 1.06252 0.94480 2.81549 0.61040 2.21012 0.94460 7.36365
200 0.09824 0.40474 0.95220 1.47864 0.26576 1.30460 0.95380 4.89920
400 0.04759 0.26770 0.95320 1.01592 0.13375 0.88148 0.95400 3.38478

α 25 0.00953 0.23419 0.93000 0.87833 -0.00224 0.14260 0.92620 0.53624

50 0.00756 0.16517 0.93520 0.62066 -0.00165 0.09929 0.93560 0.37911
75 0.00320 0.13366 0.93240 0.50368 -0.00094 0.08211 0.93400 0.30948
100 0.00320 0.11391 0.94540 0.43755 -0.00048 0.06811 0.94740 0.26802
200 0.00146 0.07917 0.94460 0.30843 -0.00017 0.04886 0.94600 0.18899
400 0.00130 0.05606 0.95060 0.21773 -0.00002 0.03443 0.95100 0.13387

β 25 0.02114 0.08255 0.96400 0.22165 0.02681 0.13760 0.94160 0.48931

50 0.01420 0.03978 0.96100 0.14848 0.01408 0.08995 0.94700 0.33926
75 0.01397 0.03350 0.95480 0.12064 0.01086 0.07166 0.95500 0.27499
100 0.01181 0.02751 0.95900 0.10348 0.00608 0.06133 0.94760 0.23555
200 0.00590 0.01832 0.95920 0.07108 0.00382 0.04249 0.94620 0.16553
400 0.00503 0.01325 0.95080 0.04994 0.00159 0.02938 0.95140 0.11628

8. Applications
In this section, we present examples to illustrate the flexibility of the LLoGW distribution
and its sub-models for data modeling. Estimates of the parameters of LLoGW distribution
(standard error in parentheses), Akaike Information Criterion (AIC), Bayesian Information
Criterion (BIC), Cramer von Mises (W ∗ ), Andersen-Darling (A∗ ), and sum of squares (SS) from
the probability plots are presented for each data set. We also compare the LLoGW distribution
with the gamma-Dagum (GD) (Oluyede, Huang, and Pararai 2014) and beta Weibull (BW)
(Lee, Famoye, and Olumolade 2007), (Famoye, Lee, and Olumolade 2005) distributions. The
GD and BW pdfs are given by
λβδx−δ−1
gGD (x) = (1 + λx−δ )−β−1 (− log[1 − (1 + λx−δ )−β ])α−1 , x > 0,
Γ(α)
and
kλk k−1 −b(λx)k k
gBW (x) = x e (1 − e−(λx) )a−1 , x > 0,
B(a, b)
respectively.
The maximum likelihood estimates (MLEs) of the LLoGW parameters ∆ = (s, c, α, β) are
computed by maximizing the objective function via the subroutine NLMIXED in SAS as well as
the function nlm in R (R Development Core Team 2011). The estimated values of the parameters
(standard error in parenthesis), -2log-likelihood statistic, Akaike Information Criterion, AIC =
2p − 2 ln(L), and Bayesian Information Criterion, BIC = p ln(n) − 2 ln(L), where L = L(Θ̂)
is the value of the likelihood function evaluated at the parameter estimates, n is the number
of observations, and p is the number of estimated parameters are presented in Tables 5 and 6.
The LLoGW distribution is fitted to the data sets and these fits are compared to the fits using
LLoGR, LLoGE and LLoG distributions.
 Austrian Journal of Statistics 61

As stated earlier, we maximize the likelihood function using NLmixed in SAS as well as the
function nlm in R (R Development Core Team 2011). These functions were applied and executed
for wide range of initial values. This process often results or lead to more than one maximum,
however, in these cases, we take the MLEs corresponding to the largest value of the maxima. In a
few cases, no maximum was identified for the selected initial values. In these cases, a new initial
value was tried in order to obtain a maximum. The issues of existence and uniqueness of the
MLEs are theoretical interest and has been studied by several authors for different distributions
including (Seregin 2010), (Silva and Tenreyro 2010), (Zhou 2009), and (Xia, Mi, and Zhou
2009).
We can use the likelihood ratio (LR) test to compare the fit of the LLoGW distribution with
its sub-models for a given data set. For example, to test β = 1, the LR statistic is ω =
2[ln(L(α̂, β̂, ŝ, ĉ)) − ln(L(α̃, 1, s̃, c̃))], where α̂, β̂, ŝ, and ĉ are the unrestricted estimates, and α̃,
s̃, and c̃ are the restricted estimates. The LR test rejects the null hypothesis if ω > χ2 , where
χ2 denote the upper 100% point of the χ2 distribution with 1 degrees of freedom.
Plots of the fitted densities, the histogram of the data and probability plots (Chambers, Cleve-
land, Kleiner, and Tukey 1983) are given in Figures 3 and 4 for the first data set and Figures
5 and 6 for the second dataset. For the probability plot, we plotted G(x(j) ; ŝ, ĉ, α̂, β̂) against
j − 0.375
, j = 1, 2, · · · , n, where x(j) are the ordered values of the observed data. The measure
n + 0.25  
Pn  j − 0.375 2
of closeness given by the sum of squares SS = j=1 G(x(j) )− , was computed
n + 0.25
for each fitted model.
The goodness-of-fit statistics W ∗ and A∗ , described by (Chen and Balakrishnan 1995) are also
presented in the tables. These statistics can be used to verify which distribution fits better to
the data. In general, the smaller the values of W ∗ and A∗ , the better the fit. Let G(x; ∆)
be the cdf, where the form of G is known but the k-dimensional parameter vector, say ∆ is
unknown. We can obtain the statistics W ∗ and A∗ as follows: (i) Compute ui = G(xi ; ∆), ˆ
−1
where the xi ’s are in ascending order; (ii) Compute yi = Φ (ui ), where Φ(.) is the standardPn
normal cdf and Φ−1P (.) its inverse; (iii) Compute vi = Φ((y i − y)/sy ), where y = n
−1
i=1 yi
n n
and s2y = (n − 1)−1 P 2 ; (iv) Calculate W 2 = 2 + 1/(12n)
P
i=1 (yi − y) i=1 {v i − (2i − 1)/(2n)}
and A2 = −n − n−1 ni=1 {(2i − 1) log(vi ) + (2n + 1 − 2i) log(1 − vi )}; (v) Modify W 2 into
W ∗ = W 2 (1 + 0.5/n) and A2 into A∗ = A2 (1 + 0.75/n + 2.25/n2 ).

8.1. Fracture toughness of alumina (Al2 O3 )(in the units of MPa m1/2 )
This data set consists of 119 observations on fracture toughness of Alumina (Al2 O3 )(in the units
of MPa m1/2 . The data are taken from the web-site:
http://www.ceramics.nist.gov/ srd/summary/ftmain.htm. The same data set has also been
studied by (Nadarajah and Kotz 2007). Initial values for the LLoGW model in R code are
s = 2.34, c = 3.42, α = 0.155, β = 1. Estimates of the parameters of LLoGW distribution and
its related sub-models (standard error in parentheses), AIC, BIC, W∗ , A∗ and SS are give in
Table 5. Plots of the fitted densities and the histogram are given in Figure 3, and plots of the
observed probability vs predicted probability are given in Figure 4. The estimated variance-
covariance matrix for the LLoGW distribution is
 
0.09121 −0.09084 −0.00039 0.1929
−0.09084 2.9051 0.00136 −0.5062 
−0.00039 0.00136 4.21E − 06 −0.00166 ,
 

0.1929 −0.5062 −0.00166 0.6950

and the 95% confidence intervals for the model parameters are given by s ∈ (4.9735 ± 1.96 ×
0.3020), c ∈ (10.4933 ± 1.96 × 1.7044), α ∈ (0.00196 ± 1.96 × 0.0021) and β ∈ (3.6271 ± 1.96 ×
62 The Log-logistic Weibull Distribution with Applications to Lifetime Data

0.8337), respectively.

Table 5: Estimates of models for fracture toughness of alumina data

Estimates Statistics
Model ŝ ĉ α̂ β̂ −2 log L AIC BIC W∗ A∗ SS
LLoGW 4.9735 10.4933 0.00196 3.6271 334.7 342.7 353.8 0.0416 0.2545 0.0408
(0.3020) (1.7044) (0.0021) (0.8337)
LLoGE 4.5223 9.5011 0.02847 1 350.0 356.0 364.3 0.0659 0.5475 0.0395
(0.0988) (1.1499) (0.0122) -
LLoG 4.3234 7.0649 0 1 356.7 360.7 366.2 0.3549 2.2314 0.1803
(0.0959) (0.5535) - -
LLoGR 4.6670 10.6049 0.01245 2 339.6 345.6 354.0 0.0603 0.4009 0.0386
(0.1060) (1.2344) (0.0037) -
λ̂ β̂ δ̂ α̂
GD 0.2687 0.5433 17.2156 26.4212 387.6 395.6 406.7 0.9288 5.3449 1.0675
(0.8648) (0.0.1388) (2.4638) (0.5059)
λ̂ k̂ â b̂
BW 0.09969 5.6903 0.8010 55.2052 337.1 345.1 356.2 0.0829 0.5021 0.0785
(0.0102) (1.4654) (0.3214) (4.7545)
Note. Standard errors are in parentheses.

Figure 3: Fitted densities for fracture toughness of alumina data

The LR test statistic of the hypothesis H0 : LLoGE against Ha : LLoGW, H0 :LLoG against
Ha : LLoGW, and H0 :LLoGR against Ha : LLoGW are 15.3 (p-value < 0.0001), 22.0 (p-value <
0.0001), and 4.9 (p-value=0.02686 < 0.05). We can conclude that there are significant differences
between LLoGW and LLoGE, LLoG, LLoGR distributions. The values of the statistics: AIC
and BIC also shows that the LLoGW distribution is a better fit than the non-nested GD and
BW distributions for the fracture toughness of alumina data. There is also clear evidence based
on the goodness-of-fit statistics W ∗ and A∗ that the LLoGW distribution is by far the better
fit for the fracture toughness of alumina data.
 Austrian Journal of Statistics 63

Figure 4: Probability pots for fracture toughness of alumina data

8.2. Breaking stress of carbon fibres (in Gba)

This data set consists of 100 uncensored data on breaking stress of carbon fibres (in Gba),
(Nichols and Padgett 2006). Initial values for the LLoGW model in R code are s = 1, c =
1, α = 1, β = 3. Estimates of the parameters of LLoGW distribution and its related sub-models
(standard error in parentheses), AIC, BIC, W∗ , A∗ and SS are give in Table 6. Plots of the
fitted densities and the histogram, as well as observed probability vs predicted probability are
given in Figures 5 and 6, respectively. The estimated variance-covariance matrix for the LLoGW
distribution is
 
0.5771 −0.5628 0.0014 0.2189
−0.5628 4.1583 0.0415 −0.9957
 ,
 0.0014 0.0415 0.0008 −0.0128
0.2189 −0.9957 −0.0128 0.3387
and the 95% confidence intervals for the model parameters are given by s ∈ (3.4839 ± 1.96 ×
0.7597), c ∈ (4.3649±1.96×2.0392), α ∈ (0.0418±1.96×0.0274) and β ∈ (2.5196±1.96×0.5820),
respectively.
The LR test statistic of the hypothesis H0 : LLoGE against Ha : LLoGW and H0 :LLoG against
Ha : LLoGW are 7.5791 (p-value = 0.0059) and 10.1860 (p-value = 0.0014). We can conclude
that there are significant difference between LLoGW and LLoGE distributions as well between
LLoGW and LLoG distributions. There is also very clear and convincing evidence based on
the goodness-of-fit statistics W ∗ and A∗ that the LLoGW distribution is by far the better fit
than the sub-models. The SS value of 0.0584 for the LLoGW distribution is smaller than the
values for the non-nested GD and BW distributions. The values of AIC and BIC also shows
that the LLoGW distribution is a better fit than the non-nested GD and BW distributions for
the breaking stress of carbon fibres data.

9. Concluding remarks
We have presented a new distribution called the log-logistic Weibull (LLoGW) distribution that
is suitable for applications in various areas including reliability, survival analysis and actuarial
64 The Log-logistic Weibull Distribution with Applications to Lifetime Data

Table 6: Estimates of models for breaking stress of carbon fibres data

Estimates Statistics
Model ŝ ĉ α̂ β̂ −2 log L AIC BIC W∗ A∗ SS
LLoGW 3.4839 4.3649 0.0418 2.5196 282.37 290.37 300.79 22.2883 133.4936 0.0584
(0.7597) (2.0392) (0.0274) (0.5820)
LLoGE 2.6224 4.5962 0.0311 1 289.95 295.95 303.77 24.5245 137.8047 0.1268
(0.1482) (0.5469) (0.0298) -
LLoG 2.4984 4.1179 0 1 292.56 298.56 306.37 23.8320 139.0570 0.1664
(0.1054) (0.3441) - -
LLoGR 3.1737 5.3261 0.0595 2 283.48 289.48 297.30 23.6256 133.9022 0.0507
(0.3043) (1.1471) (0.0213) -
λ̂ β̂ δ̂ α̂
GD 36.8189 4.1911 3.7238 0.2278 289.06 297.06 307.48 23.7881 138.2885 0.1716
(16.3529) (1.5857) (0.3264) (0.1076)
λ̂ k̂ â b̂
BW 0.2002 2.3259 1.3707 4.8215 282.69 290.69 301.11 22.9775 133.4390 0.0680
(0.0179) (0.7563) (0.7529) (0.0031)
Note. Standard errors are in parentheses.

Fitted PDF

0.6 LLOGW
LLOGE
LLOG
LLOGR
0.5 GD
BW

0.4
f(x)

0.3

0.2

0.1

0.0

1 2 3 4 5

Figure 5: Fitted densities for breaking stress of carbon fibres data

sciences just to mention a few areas. The structural properties including hazard and reverse
hazard functions, quantile function, probability weighted moments (PWMs), moments, condi-
tional moments, mean deviations, Bonferroni and Lorenz curves, Rényi entropy, distribution of
order statistics, maximum likelihood estimates, asymptotic confidence intervals are presented.
Applications of the model to real data sets are given in order to illustrate the applicability and
usefulness of the proposed distribution.

Acknowledgements
The authors are very grateful to the referees for some useful comments on an earlier version of
this manuscript which led to this improved version.

References
 Austrian Journal of Statistics 65

The Graph of Observed vs Expected Probability

1.0
0.8
Expected Probability

0.6
0.4

LLOGW(SS=0.0584)
0.2

LLOGE(SS=0.1268)
LLOG(SS=0.1664)
LLOGR(SS=0.0507)
GD(SS=0.1716)
0.0

BW(SS=0.0680)

0.0 0.2 0.4 0.6 0.8 1.0

Observed Probability

Figure 6: Probability plots for breaking stress of carbon fibres data

Barakat HM, Abdelkader YH (2004). “Computing the Moments of Order Statistics from Non-
identical Random Variables.” Statistical Methods and Applications, 13(1), 15–26.
Bourguignon M, Silva RB, Cordeiro GM (2014). “The Weibull–G Family of Probability Distri-
butions.” Journal of Data Science, 12(1), 53–68.
Burr IW (1942). “Cumulative Frequency Functions.” The Annals of Mathematical Statistics,
13(2), 215–232.
Burr IW (1973). “Pameters for a General System of Distributions to Match a Grid of α3 and
α4.” Communications in Statistics-Theory and Methods, 2(1), 1–21.
Carrasco M, Ortega EM, Cordeiro GM (2008). “A Generalized Modified Weibull Distribution
for Lifetime Modeling.” Computational Statistics & Data Analysis, 53(2), 450–462.
Chambers JM, Cleveland WS, Kleiner B, Tukey PA (1983). Graphical Methods of Data Analysis.
Chapman and Hall.
Chen G, Balakrishnan N (1995). “A General Purpose Approximate Goodness-of-fit Test.” Jour-
nal of Quality Technology, 27(2), 154–161.
Durbin J, Koopman SJ (2012). Time Series Analysis by State Space Methods. 38. Oxford
University Press.
Eugene N, Lee C, Famoye F (2002). “Beta-normal Distribution and Its Applications.” Commu-
nications in Statistics-Theory and Methods, 31(4), 497–512.
Famoye F, Lee C, Olumolade O (2005). “The Beta-Weibull Distribution.” Journal of Statistical
Theory and Applications, 4(2), 121–136.
Gradshtein IS, Ryzhik IM (2000). Tables of Integrals, Series and Products. Academic Press,
San Diego.
Greenwood J, Landwehr JM, Matalas NC, Wallis JR (1979). “Probability Weighted Moments:
Definition and Relation to Parameters of Several Distributions Expressable in Inverse Form.”
Water Resources Research, 15(5), 1049–1054.
66 The Log-logistic Weibull Distribution with Applications to Lifetime Data

Gupta RD, Kundu D (2001). “Exponentiated Exponential Family: An Alternative to Gamma

and Weibull Distributions.” Biometrical Journal, 43(1), 117–130.

Gurvich MR, Dibenedetto AT, Ranade SV (1997). “A New Statistical Distribution for Char-
acterizing the Random Strength of Brittle Materials.” Journal of Materials Science, 32(10),
2559–2564.

Haupt E, Schäbe H (1992). “A New Model for a Lifetime Distribution with Bathtub Shaped
Failure Rate.” Microelectronics Reliability, 32(5), 633–639.

Hjorth U (1980). “A Reliability Distribution with Increasing, Decreasing, Constant and Bathtub-
Shaped Failure Rates.” Technometrics, 22(1), 99–107.

Hoskings JRM (1990). “L-moments: Analysis and Estimation of Distributions Using Linear
Combinations of Order Statistics.” Journal of the Royal Statistical Society, B52, 105–124.

Jones MC (2004). “Families of Distributions Arising from Distributions of Order Statistics.”

Test, 13(1), 1–43.

Lai CD, Xie M, Murthy DNP (2003). “A Modified Weibull Distribution.” IEEE Transactions
on Reliability, 52(1), 33–37.

Lee C, Famoye F, Olumolade O (2007). “Beta-Weibull Distribution: Some Properties and

Applications to Censored Data.” Journal of Modern Applied Statistical Methods, 6(1), 173–
186.

Mudholkar GS, Srivastava DK (1993). “Exponentiated Weibull Family for Analyzing Bathtub
Failure-Rate Data.” IEEE Transactions on Reliability, 42(2), 299–302.

Murthy DNP, Xie M, Jiang R (2004). Weibull Models, volume 505. John Wiley & Sons.

Nadarajah S, Cordeiro GM, Ortega EMM (2011). “General Results for the Beta-Modified
Weibull Distribution.” Journal of Statistical Computation and Simulation, 81(10), 1211–
1232.

Nadarajah S, Kotz S (2006). “The Beta Exponential Distribution.” Reliability Engineering &
System Safety, 91(6), 689–697.

Nadarajah S, Kotz S (2007). “On the Alternative to the Weibull Function.” Engineering Fracture
Mechanics, 74(3), 451–456.

Nassar MM, Eissa FH (2003). “On the Exponentiated Weibull Distribution.” Communications
in Statistics-Theory and Methods, 32(7), 1317–1336.

Nelson W (1982). Lifetime Data Analysis. Wiley, New York.

Nichols MD, Padgett WJ (2006). “A Bootstrap Control Chart for Weibull Percentiles.” Quality
and Reliability Engineering International, 22(2), 141–151.

Oluyede BO, Huang S, Pararai M (2014). “A New Class of Generalized Dagum Distribution
with Applications to Income and Lifetime Data.” Journal of Statistical and Econometric
Methods, 3(2), 125–151.

Paranaı́ba PF, Ortega EMM, Cordeiro GM, Pescim RR (2011). “The Beta Burr XII Distribution
with Application to Lifetime Data.” Computational Statistics & Data Analysis, 55(2), 1118–
1136.
 Austrian Journal of Statistics 67

Pham H, Lai CD (2007). “On Recent Generalizations of the Weibull Distribution.” IEEE
Transactions on Reliability, 56(3), 454–458.

Pinho LGB, Cordeiro GM, Nobre JS (2012). “The Gamma–Exponentiated Weibull Distribu-
tion.” Journal of Statistical Theory and Applications, 11(4), 379–395.

Prudnikov AP, Brychkov YA, Marichev O I (1992). Integrals and Series, volume 4. Gordon and
Breach Science Publishers, Amsterdam.

R Development Core Team (2011). “R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing, Vienna, Austria, 2011.”

Rajarshi S, Rajarshi MB (1988). “Bathtub Distributions: A Review.” Communications in

Statistics-Theory and Methods, 17(8), 2597–2621.

Renyi A (1960). “On Measures of Entropy and Information.” In Proceedings of the Fourth
Berkeley Symposium on Mathematical Statistics and Probability, volume 1, pp. 547–561.

Seregin A (2010). “Uniqueness of the Maximum Likelihood Estimator for k-Monotone Densities.”
Proceedings of the American Mathematical Society, 138(12), 4511–4515.

Shaked M, Shanthikumar JG (1994). Stochastic Orders and Their Applications. New York,
Academic Press.

Silva GO, Ortega EMM, Cordeiro GM (2010). “The Beta Modified Weibull Distribution.”
Lifetime Data Analysis, 16(3), 409–430.

Silva JMCS, Tenreyro S (2010). “On the Existence of the Maximum Likelihood Estimates in
Poisson Regression.” Economics Letters, 107(2), 310–312.

Xia J, Mi J, Zhou Y (2009). “On the Existence and Uniqueness of the Maximum Likelihood
Estimators of Normal and Log-normal Population Parameters with Grouped Data.” Journal
of Probability and Statistics, 2009.

Zhou C (2009). “Existence and Consistency of the Maximum Likelihood Estimator for the
Extreme Value Index.” Journal of Multivariate Analysis, 100(4), 794–815.

R code
In this section, the R codes to compute cdf, pdf, moments, Rényi entropy, mean deviations,
maximum likelihood estimates and variance-covariance matrix for the LLoGW distribution are
presented.
#D e f i n e t h e pdf o f LLoGW d i s t r i b u t i o n
f 1=f u n c t i o n ( x , s , c , alpha , b e t a ) {
y =(((1+( x/ s ) ∗ ∗ c )∗∗( −1))∗ exp(− a l p h a ∗ ( x ∗∗ b e t a ) ) )
∗ ( ( ( 1 + ( x/ s ) ∗ ∗ c ) ∗ ∗ ( − 1 ) ) ∗ ( c / s ) ∗ ( x/ s ) ∗ ∗ ( c −1)
+a l p h a ∗ b e t a ∗ ( x ∗ ∗ ( beta −1)))
return (y)
}

#D e f i n e t h e c d f o f LLoGW d i s t r i b u t i o n
F1=f u n c t i o n ( x , s , c , alpha , b e t a ) {
y=1−((1+(x/ s ) ∗ ∗ c )
∗∗( −1))∗ exp(− a l p h a ∗x ∗∗ b e t a )
return (y)
}
68 The Log-logistic Weibull Distribution with Applications to Lifetime Data

#D e f i n e t h e moments o f LLoGW d i s t r i b u t i o n
moment=f u n c t i o n ( s , c , alpha , beta , r ) {
f=f u n c t i o n ( x , s , c , alpha , beta , r )
{ ( xˆ r ) ∗ ( f 1 ( x , s , c , alpha , b e t a ) ) }
y=i n t e g r a t e ( f , l o w e r =0, upper=I n f ,
s u b d i v i s i o n s =100 , s=s , c=c , a l p h a=alpha , b e t a=beta , r=r )
return (y)
}

#D e f i n e t h e q u a n t i l e o f LLoGW d i s t r i b u t i o n
q u a n t i l e=f u n c t i o n ( s , c , alpha , beta , u ) {
f=f u n c t i o n ( x ) { a l p h a ∗xˆ b e t a+l o g (1+( x/ s ) ˆ c)+ l o g (1−u ) }
r c <− u n i r o o t ( f , l o w e r =0, upper =100 , t o l = 1 e −9)
r e s u l t=r c $ r o o t
#check
e r r o r=F1 ( r e s u l t , s , c , alpha , b e t a )−u
r e t u r n ( l i s t ( ” r e s u l t ”= r e s u l t , ” e r r o r ”= e r r o r ) )
}

#D e f i n e Mean D e v i a t i o n about t h e mean o f LLoGW d i s t r i b u t i o n

DU=f u n c t i o n ( s , c , alpha , b e t a ) {
mu=moment ( s , c , alpha , beta , 1 ) $ v a l u e
f=f u n c t i o n ( x , s , c , alpha , b e t a ) { ( abs ( x−mu) ∗ f 1 ( x , s , c , alpha , b e t a a ) }
y=i n t e g r a t e ( f , l o w e r =0, upper=I n f , s u b d i v i s i o n s =100
, s=s , c=c , a l p h a=alpha , b e t a=b e t a )
return (y)
}

#D e f i n e Mean D e v i a t i o n about t h e median o f LLoGW d i s t r i b u t i o n

DM=f u n c t i o n ( s , c , alpha , b e t a ) {
M=median ( c (X) ) #X i s t h e data s e t
f=f u n c t i o n ( x , s , c , alpha , b e t a ) { ( abs ( x−M) ∗ f 1 ( x , s , c , alpha , b e t a ) }
y=i n t e g r a t e ( f , l o w e r =0, upper=I n f , s u b d i v i s i o n s =100
, s=s , c=c , a l p h a=alpha , b e t a=b e t a )
return (y)
}

D e f i n e t h e Renyi e n t r o p y o f LLoGW d i s t r i b u t i o n
t=f u n c t i o n ( s , c , alpha , beta , v ) {
f=f u n c t i o n ( x , s , c , alpha , beta , v )
{ ( f 1 ( x , s , c , alpha , b e t a ) ) ˆ ( v ) }
y=i n t e g r a t e ( f , l o w e r =0, upper=I n f , s u b d i v i s i o n s =100
, s=s , c=c , a l p h a=alpha , b e t a=beta , v=v ) $ v a l u e
return (y)
}
Renyi=f u n c t i o n ( s , c , alpha , beta , v ) {
y=l o g ( t ( s , c , alpha , beta , v ))/(1 − v )
return (y)
}

#C a l c u l a t e t h e maximum l i k e l i h o o d e s t i m a t o r s
#o f LLoGW d i s t r i b u t i o n
l i b r a r y ( ' bbmle ' )
xvec<−c (X) #X i s t h e data s e t
ln<−f u n c t i o n ( s , c , alpha , b e t a ) {
−sum ( l o g ( ( ( ( 1 + ( x/ s ) ∗ ∗ c )∗∗( −1))∗ exp(− a l p h a ∗ ( x ∗∗ b e t a ) ) )
∗ ( ( ( 1 + ( x/ s ) ∗ ∗ c ) ∗ ∗ ( − 1 ) ) ∗ ( c / s ) ∗ ( x/ s ) ∗ ∗ ( c −1)
 Austrian Journal of Statistics 69

+a l p h a ∗ b e t a ∗ ( x ∗ ∗ ( beta − 1 ) ) ) ) )
}
mle . r e s u l t s 1 <−mle2 ( ln , s t a r t= l i s t ( s=s , c=c , a l p h a=a l p h a
, b e t a=b e t a ) , h e s s i a n . opt=TRUE)
summary ( mle . r e s u l t s 1 )

# Variance−c o v a r i a n c e matrx o f LLoGW d i s t r i b u t i o n

vcov ( mle . r e s u l t s 1 )

Affiliation:
Broderick O. Oluyede
Department of Mathematical Sciences
Georgia Southern University
Statesboro, GA, 30460, USA
E-mail: [email protected]

Susan Foya
Department of Mathematics and Computational Sciences
Botswana International University of Science and Technology
Palapye, BW

Gayan Warahena-Liyanage
Department of Mathematics
Central Michigan University
Mount Pleasant, MI, 48859, USA

Shujiao Huang
Department of Mathematics
University of Houston
Houston, TX, 77004, USA

Austrian Journal of Statistics http://www.ajs.or.at/

published by the Austrian Society of Statistics http://www.osg.or.at/
Volume 45 Submitted: 2015-04-20
September 2016 Accepted: 2015-06-15