Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Statistics

A Journal of Theoretical and Applied Statistics

ISSN: 0233-1888 (Print) 1029-4910 (Online) Journal homepage: https://www.tandfonline.com/loi/gsta20

An extension of the exponential distribution

Saralees Nadarajah & Firoozeh Haghighi

To cite this article: Saralees Nadarajah & Firoozeh Haghighi (2011) An extension of the
exponential distribution, Statistics, 45:6, 543-558, DOI: 10.1080/02331881003678678

To link to this article: https://doi.org/10.1080/02331881003678678

Published online: 16 Mar 2010.

Submit your article to this journal

Article views: 2278

View related articles

Citing articles: 45 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=gsta20
Statistics, Vol. 45, No. 6, December 2011, 543–558

An extension of the exponential distribution


Saralees Nadarajaha * and Firoozeh Haghighib
a School of Mathematics, University of Manchester, Manchester M13 9PL, UK;
b Department of Mathematics, Statistics and Computer Sciences, University of Tehran, Tehran, Iran

(Received 1 December 2008; final version received 25 January 2010 )

A generalization of the exponential distribution is presented. The generalization always has its mode at zero
and yet allows for increasing, decreasing and constant hazard rates. It can be used as an alternative to the
gamma, Weibull and exponentiated exponential distributions.A comprehensive account of the mathematical
properties of the generalization is presented. A real data example is discussed to illustrate its applicability.

Keywords: exponential distribution; exponentiated exponential distribution; gamma distribution; hazard


rate function; moments; Weibull distribution

1. Introduction

In lifetime data analysis, monotone hazard rate occurs commonly in practice. Such situations are
commonly modelled using the Weibull and gamma distributions. The Weibull distributions are
more popular than the gamma because the survival function (sf) of the latter cannot be expressed
in a closed form and one needs numerical integration. The Weibull distributions have closed form
survival and hazard rate functions (hrfs). We refer the readers to Murthy et al. [1] for details about
Weibull distributions.
Gupta and Kundu [2] introduced the exponentiated exponential (EE) distribution as an alter-
native to the gamma distribution. The EE distribution has many properties similar to that of the
gamma and yet has closed form survival and hrfs. See Gupta and Kundu [3] for a review and
some developments on the EE distribution.
The aim of this paper is to introduce a new extension of the exponential distribution as an
alternative to the gamma, Weibull and the EE distributions. It is most conveniently specified in
terms of its SF as follows:

S(t) = exp{1 − (1 + λt)α }, (1)

*Corresponding author. Email: mbbsssn2@manchester.ac.uk

ISSN 0233-1888 print/ISSN 1029-4910 online


© 2011 Taylor & Francis
http://dx.doi.org/10.1080/02331881003678678
http://www.tandfonline.com
544 S. Nadarajah and F. Haghighi

for α > 0, λ > 0 and t > 0. The corresponding cumulative distribution function (cdf), probability
density function (pdf) and the quantile function are as follows:

F (t) = 1 − exp{1 − (1 + λt)α }, (2)


f (t) = αλ(1 + λt)α−1 exp{1 − (1 + λt)α } (3)

and
1
Q(p) = {(1 − log(1 − p))1/α − 1}, 0 < p < 1,
λ
respectively. The median is given by

1
Median(T ) = {(1 − log(0.5))1/α − 1}.
λ

The hrf is given by


h(t) = αλ(1 + λt)α−1 . (4)

Note that Equation (3) has two parameters just like the gamma, Weibull and the EE distribu-
tions. Note also that Equation (3) has closed form sfs and hrfs just like the Weibull and the EE
distributions. For α = 1, Equation (3) reduces to the exponential distribution. As we shall see
later, Equation (3) has the attractive feature of always having the zero mode and yet allowing for
increasing, decreasing and constant hrfs.
Let T be a random variable having the pdf (3). Then the pdf of Y = λT is as follows:

f (y) = α(1 + y)α−1 exp{1 − (1 + y)α }, (5)

which is independent of λ. So, λ controls only the scale and we shall refer to it as the scale
parameter. It follows from Equation (5) that

f (y) ∼ exp(1)αy α−1 exp{−(1 + y)α },

as y → ∞. So, the upper tail behaviour of the pdf of Y is a product of a polynomial power and
an exponential decay, both of which depend only on α. For example, larger values of α will lead
to faster decay of the upper tail. Hence, we shall refer to α as the shape parameter.
We provide at least three possible motivations for introducing the new family. The first moti-
vation is based on the relationship between a pdf and its hrf. The gamma, Weibull and the EE
are flexible distributions. They allow for both monotonically decreasing and unimodal pdfs. They
also allow for increasing, decreasing and constant hrfs. However, they only allow for decreasing
or constant hrfs when their respective pdfs are monotonically decreasing. They do not allow for
an increasing hrf when their respective pdfs are monotonically decreasing. This feature can be a
serious limitation.
Consider the data set in Table 1. The data given are daily rainfall (in millimetres) on the first
of January for a location in Florida for the years from 1907 to 2000. It is clear from the table
that the mode for the data set is zero, meaning that the pdf is monotonically decreasing. The
empirical hrf for the data set is shown in Figure 1. This figure shows an approximately increasing
hrf. So, we have at least one example where the pdf is monotonically decreasing and yet the hrf
is increasing. The gamma, Weibull and the EE distributions are not suitable for examples of this
kind. Furthermore, these three distributions cannot be fitted to data sets of the kind in Table 1.
Statistics 545

Table 1. Daily rainfall (in mm) on the first of January at Gainsville


(Florida) for the years 1907–2000.

0.00 0.00 0.00 0.35 0.00 0.00 0.06 0.00 0.00 0.00 0.14 0.54 0.54 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.21 0.02 1.10 0.00 0.00 0.26 0.00 0.00
0.00 0.66 0.10 0.00 0.00 0.25 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.02
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.10 0.46 0.52
0.61 0.05 0.00 0.00 0.00 0.00 0.03 0.00 0.45 0.45 0.00 0.00 0.02 0.00 0.00
0.00 0.00 0.19 0.00 0.00 0.00 0.75 0.00 0.00 0.00 0.89 0.94 0.00 0.00 0.25
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
15
Empirical Hazard Rate Function
10
5

0.0 0.2 0.4 0.6 0.8 1.0


Daily Rainfall (in mm)

Figure 1. Empirical hrf for daily rainfall (in mm) on the first of January at Gainsville (Florida) for the years 1907–2000.

To see this, suppose t1 , . . . , tn is a random sample from a gamma distribution with shape parameter
α and scale parameter λ. Then the log-likelihood (LL) function of the two parameters is as follows:

n 
n
log L(α, λ) = nα log λ − n log (α) + (α − 1) log ti − λ ti . (6)
i=1 i=1

If the random sample is from a Weibull distribution with shape parameter α and scale parameter
λ, then
n 
n
log L(α, λ) = n log α + nα log λ + (α − 1) log ti − λα tiα . (7)
i=1 i=1
Finally, if the random sample is from an EE distribution with shape parameter α and scale
parameter λ, then

n 
n
log L(α, λ) = n log α + n log λ + (α − 1) log{1 − exp(−λti )} − λ ti . (8)
i=1 i=1

For the data set in Table 1, because of the presence of zeros, Equations (6)–(8) reduce to the
following:


⎪ ∞, for all α < 1 and λ > 0,
⎨ n
log L(α, λ) = n log λ − λ i=1 ti , for α = 1 and for all λ > 0,



−∞, for all α > 1 and λ > 0.
So, the maximum-likelihood estimate of α is non-unique. Hence, none of the three distributions
can be fitted to data sets of the kind in Table 1.
546 S. Nadarajah and F. Haghighi

The second motivation is based on the ability or the inability to model data that have their
mode fixed at zero. There are many physical processes which have their mode fixed at zero by
definition. For example, consider daily rainfall and daily snowfall. For most locations in the world,
as seen from Table 1, the mode for both these variables will be zero. However, if one tries to fit
the gamma, Weibull and the EE distributions to the data, the resulting fitted pdfs may not always
have their mode at zero. To illustrate this, we simulated a random sample of 100 observations
from the unit exponential distribution which always has its mode at zero. We fitted the gamma,
Weibull, EE and the new distributions to the simulated data. The fitted pdfs are shown in Figure 2.
It is clear that none of the gamma, Weibull or the EE distributions provides a valid fit. Only the
new distribution provides a plausible fit. This happened because the gamma, Weibull and the EE
distributions allow for a positive probability that their shape parameters will be greater than one
even if the data are not unimodal. The new distribution always has a zero probability that its shape
will be unimodal.
Our third and final motivation is based on a mathematical relationship. Suppose Y is a Weibull
random variable with shape parameter α and scale parameter λ. Let Z = Y − 1/λ. Then the
distribution in Equation (1) is the same as that of Z truncated at zero. So, the new distribution can
0.8
0.6
Fitted PDFs

0.4
0.2
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0


Unit Exponential Variate
Figure 2. Fitted pdfs of the new distribution (solve curve), the gamma distribution (curve of dashes), the Weibull
distribution (curve of dots) and the EE distribution (curve of dashes and dots) for simulated data from the unit exponential
distribution.
Statistics 547

be interpreted as a truncated Weibull distribution. We are most grateful to one of the reviewers
for pointing out this fact.
The fact that the new distribution is a truncated Weibull distribution does not mean that it does
not deserve a separate treatment. Truncated distributions are of special interest by themselves.
There are hundreds of papers published on truncated forms of known distributions. There are also
several papers comparing truncated versions with their original forms. For example, Vaughan [4,
p. 676] compares the fits of truncated normal and normal distributions for a problem involving
electrical insulation. More recently, Nadarajah [5, Sect. 4] has compared the fits of truncated
inverted beta and inverted beta distributions for an air pollution data.
The contents of this paper are organized as follows. A comprehensive account of mathematical
properties of the new distribution is provided in Sections 2–6. The properties studied include
shapes of pdf and hrf, raw moments, conditional moments, L-moments, mean deviation about the
mean, mean deviation about the median, Bonferroni curve, Lorenz curve, Rényi entropy, Shannon
entropy, order statistics and the asymptotic distribution of the extreme values. Estimation by the
methods of moments and maximum likelihood – including the case of censoring – is presented
in Section 7. Finally, Section 8 illustrates an application by using the real data set in Table 1.

2. Shapes

It follows from Equation (3) that

d log f (t) λ(α − 1)


= − αλ(1 + λt)α−1
dt 1 + λt

and
d2 log f (t) λ2 (α − 1)
=− − α(α − 1)λ2 (1 + λt)α−2 .
dt 2 (1 + λt)2
So, if α ≤ 1, then d log f (t)/d t < 0 for all t. If α > 1, then d log f (t0 )/d t0 = 0 and
d2 log f (t0 )/dt02 < 0, where t0 = (1/λ){(1 − 1/α)1/α − 1}. But t0 < 0. Hence, the only shape
possible is that f (t) monotonically decreases with f (0) = αλ and f (t) → 0 as t → ∞. Figure 3
illustrates this shape for selected parameter values.
The hrf in Equation (4) exhibits the following shapes:
1.5
Probability Density Function
1.0

α = 0.5
α=1
α=2
0.5
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0


t

Figure 3. The pdf, Equation (3), for λ = 1 and α = 0.5, 1, and 2.


548 S. Nadarajah and F. Haghighi

8
Hazard Rate Function

6
4
α = 0.5
α=1
α=2
2
0

0.0 0.5 1.0 1.5 2.0 2.5 3.0


t

Figure 4. The hrf, Equation (4), for λ = 1 and α = 0.5, 1, and 2.

• if α < 1, then h(t) monotonically decreases with h(0) = αλ and h(t) → 0 as t → ∞.


• if α > 1, then h(t) monotonically increases with h(0) = αλ and h(t) → ∞ as t → ∞.
• if α = 1, then h(t) = αλ for all t.

Figure 4 illustrates these shapes for selected parameter values.

3. Moment properties

Let T denote a random variable with the pdf (3). It follows from Lemma 1 in the appendix that

E(T k ) = αλeI (k, 0, 1), (9)

where
k  
1  k i
I (k, 0, 1) = (−1) k−i
 + 1, 1 .
αλk+1 i=0 i α

In particular, the first four moments of T are as follows:




1 1
E(T ) = −1 + e 1 + , 1 ,
λ α

 
1 1 2
E(T 2 ) = 2 1 − 2e 1 + , 1 + e 1 + , 1 ,
λ α α

  
1 1 2 3
E(T 3 ) = 3 −1 + 3e 1 + , 1 − 3e 1 + , 1 + e 1 + , 1
λ α α α

and

   
1 1 2 3 4
E(T 4 ) = 1 − 4e 1 + , 1 + 6e 1 + , 1 − 4e 1 + , 1 + e 1 + , 1 .
λ4 α α α α

The variance, skewness and kurtosis of T can be obtained using the following relationships:
Var(T ) = E(T 2 ) − (E(T ))2 , Skewness(T ) = E(T − E(T ))3 /(Var(T ))3/2 and Kurtosis(T ) =
E(T − E(T ))4 /(Var(T ))2 .
Statistics 549

For lifetime models, it is also of interest to know what E(T k | T > t) is. Using Lemma 3 in
the appendix, it is easily seen that

E(T k | T > t) = αλ exp{(1 + λt)α }K(t, k),

where
k  
1  k i
K(t, k) = k−i
(−1)  + 1, (1 + λt) .
α
αλk+1 i=0 i α
In particular,


1 1
E(T | T > t) = exp {(1 + λt)α } −  (1, (1 + λa)α ) +  + 1, (1 + λa)α ,
λ α


1 1
E(T | T > t) = 2 exp {(1 + λt) }  (1, (1 + λa) ) − 2
2 α α
+ 1, (1 + λa) α
λ α

2
+ + 1, (1 + λa) α
,
α


1 1
E(T | T > t) = 3 exp {(1 + λt) } −  (1, (1 + λa) ) + 3
3 α α
+ 1, (1 + λa) α
λ α
 
2 3
− 3 + 1, (1 + λa) + 
α
+ 1, (1 + λa) α
α α
and


1 1
E(T | T > t) = 4 exp {(1 + λt) }  (1, (1 + λa) ) − 4
4 α α
+ 1, (1 + λa) α
λ α
 
2 3
+ 6 + 1, (1 + λa)α − 4 + 1, (1 + λa)α
α α

4
+ + 1, (1 + λa)α .
α
The mean residual lifetime function is E(T | T > t) − t.
Some other important measures useful for lifetime models are the L-moments due to Hoskings
[6]. It can be shown using Lemma 1 in the appendix that the kth L-moment is given by


k−1  
k−1 k−1+j
λk = (−1)k−1−j βj ,
j =0
j j

where
k 

1 k 1 exp(i + 1) 1
βk = (−1)i − +  + 1, i + 1 .
λ i=0 i i + 1 (i + 1)1/α+1 α
In particular,

λ1 = β0 ,
λ2 = 2β1 − β0 ,
λ3 = 6β2 − 6β1 + β0
550 S. Nadarajah and F. Haghighi

and
λ4 = 20β3 − 30β2 + 12β1 − β0 ,
where


1 1
β0 = − 1 + e + 1, 1 ,
λ α

 
1 1 1 1
β1 = − + e + 1, 1 − 2−1/α−1 e2  + 1, 2 ,
λ 2 α α

  
1 1 1 −1/α 2 1 −1/α−1 3 1
β2 = − + e + 1, 1 − 2 e  + 1, 2 + 3 e  + 1, 3
λ 3 α α α
and

  
1 1 1 −1/α−1 2 1 −1/α 3 1
β3 = − + e + 1, 1 − 32 e  + 1, 2 + 3 e  + 1, 3
λ 4 α α α

−1/α−1 4 1
−4 e  + 1, 4 .
α
The L-moments have several advantages over ordinary moments: for example, they apply for any
distribution having finite mean; no higher-order moments need to be finite.
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 – defined by

δ1 (T ) = |t − μ|f (t) dt
0

and ∞
δ2 (T ) = |t − M|f (t) dt,
0
respectively, where μ = E(T ) and M = Median(T ) denotes the median. The measures δ1 (T )
and δ2 (T ) can be calculated using the following relationships:
μ ∞
δ1 (T ) = (μ − t)f (t) dt + (t − μ)f (t) dt,
0 μ
μ ∞
= μF (μ) − tf (t) dt − μ{1 − F (μ)} + tf (t) dt,
0 μ

= 2μF (μ) − 2μ + 2 tf (t) dt
μ

and
M ∞
δ2 (T ) = (M − t)f (t) dt + (t − M)f (t) dt,
0 M
M ∞
= MF (M) − tf (t) dt − M{1 − F (M)} + tf (t) dt
0 M

= 2MF (M) − M − μ + 2 tf (t) dt.
M
Statistics 551

By Lemma 3 in the appendix,



tf (t) dt = αλeK(μ, 1)
μ

and ∞
tf (t) dt = αλeK(M, 1),
M
so it follows that
δ1 (T ) = 2μF (μ) − 2μ + 2αλeK(μ, 1)
and
δ2 (T ) = 2MF (M) − M − μ + 2αλeK(M, 1).

4. Bonferroni and Lorenz curves

Bonferroni and Lorenz curves [7] have applications not only in economics to study income and
poverty, but also in other fields like reliability, demography, insurance and medicine. They are
defined by
q
1
B(p) = tf (t) dt (10)
pμ 0
and q
1
L(p) = tf (t) dt, (11)
μ 0

respectively, where μ = E(T ) and q = F −1 (p). By Lemma 3 in the appendix, one can reduce
Equations (10) and (11) to
μ − αλeK(q, 1)
B(p) =

and
μ − αλeK(q, 1)
L(p) = ,
μ
respectively.

5. Entropy

An entropy of a random variable T is a measure of variation of the uncertainty. Rényi entropy is


defined by

1
JR (γ ) = log f γ (t) dt , (12)
1−γ
where γ > 0 and γ = 1 [8]. Substituting y = (1 + λt)α , one can calculate the following:

f γ (t) dt = (αλ)γ −1 γ (γ −1−αγ )/α exp(γ ) y (αγ −γ +1)/α−1 exp(−y) dy,
γ

αγ − γ + 1
= (αλ)γ −1 γ (γ −1−αγ )/α exp(γ ) ,γ .
α
552 S. Nadarajah and F. Haghighi

So, one obtains the Rényi entropy as follows:




1 αγ − γ + 1 αγ − γ + 1
JR (γ ) = − log(αλ) + γ− log γ + log  ,γ . (13)
1−γ α α

Shannon entropy defined by E[− log f (T )] is the particular case of Equation (12) for γ ↑ 1. Lim-
iting γ ↑ 1 in Equation (13) and using L’Hospital’s rule, one obtains after considerable algebraic
manipulation that

(α − 1)e
E06[− log f (T )] = 1 − log(αλ) − {2 F2 (1, 1; 2, 2; −1) − C},
α
where C is Euler’s constant. Furthermore, 2 F2 (· · · ) is the hypergeometric function defined by

∞
(a)k (b)k x k
2 F2 (a, b; c, d; x) = ,
k=0
(c)k (d)k k!

where (z)k = z(z + 1) · · · (z + k − 1) denotes the ascending factorial.

6. Order statistics

Suppose T1 , T2 , . . . , Tn is a random sample from Equation (3). Let T1:n < T2:n < · · · < Tn:n denote
the corresponding order statistics. It is well known that the pdf and the cdf of the kth order statistic,
say Y = Tk:n , are given by

n!
fY (y) = F k−1 (y){1 − F (y)}n−k f (y),
(k − 1)!(n − k)!
n!
= {1 − S(y)}k−1 S n−k (y)f (y),
(k − 1)!(n − k)!
k−1 
n! k−1
= (−1)m S m+n−k (y)f (y)
(k − 1)!(n − k)! m=0 m

and
n 
 n
FY (y) = F j (y) {1 − F (y)}n−j ,
j =k
j

respectively, for k = 1, 2, . . . , n. It follows from Equations (1) and (3) that

k−1 

n!αλ k−1
fY (y) = (−1)m (1 + λt)α−1
(k − 1)!(n − k)! m=0 m

× exp{m + n − k + 1 − (m + n − k + 1)(1 + λt)α }

and
n 
 n  j
FY (y) = 1 − exp {1 − (1 + λt)α } exp {n − j − (n − j )(1 + λt)α }.
j =k
j
Statistics 553

Using Lemma 1 in the appendix, the qth moment of Y can be expressed as follows:

k−1 

n!αλ k−1
E(Y q ) = (−1)m exp(m + n − k + 1)I (q, 0, m + n − k + 1).
(k − 1)!(n − k)! m=0 m

√ If T = (T1 + ·√ · · + Tn )/n denotes the sample mean, then by the usual central limit theorem
n(T − E(T ))/ Var(T ) approaches the standard normal distribution as n → ∞. Sometimes
one would be interested in the asymptotics of the extreme values Mn = max(T1 , . . . , Tn ) and
mn = min(T1 , . . . , Tn ).
Let g(t) = (1/(αλ))(1 + λt)1−α , a strictly positive function. Take the cdf and the pdf as
specified by Equations (2) and (3), respectively. It can be seen that

1 − F (t + xg(t))
= exp{(1 + λt)α − (1 + λt + λxg(t))α },
1 − F (t)

 
λxg(t) α
= exp (1 + λt) 1 − 1 +
α
,
1 + λt

 
λxg(t)
= exp (1 + λt)α 1 − 1 + α + ··· ,
1 + λt
= exp {−x + o(1)},

as t → ∞. It can also be seen using L’Hospital’s rule that

F (tx) xf (tx)
lim = lim = x.
t→0 F (t) t→0 f (t)

Hence, it follows from Theorem 1.6.2 in Leadbetter et al. [9] that there must be norming constants
an > 0, bn , cn > 0 and dn such that

Pr {an (Mn − bn ) ≤ t} −→ exp {− exp(−t)}

and
Pr {cn (mn − dn ) ≤ t} → 1 − exp (−t),

as n → ∞. The form of the norming constants can also be determined. For instance, using
Corollary 1.6.3 in Leadbetter et al. [9], one can see that bn = F −1 (1 − 1/n) and an = 1/g(bn ),
where F −1 (·) denotes the inverse function of F (·).

7. Estimation

Here, we consider estimation by the methods of moments and maximum likelihood and provide
expressions for the associated Fisher information matrix. We also consider estimation issues for
censored data.
Suppose sample from Equation (3). For the moments estimation, let m1 =
 t1 , . . . , tn is a random
(1/n) nj=1 tj and m2 = (1/n) nj=1 tj2 . By equating the theoretical moments of Equation (3)
554 S. Nadarajah and F. Haghighi

with the sample moments, one obtains the following equations:



1 e 1
− +  1 + , 1 = m1
λ λ α
and  
1 2e 1 e 2
−  1 + , 1 +  1 + , 1 = m2 .
λ2 λ2 α λ2 α
These equations can be rearranged to give

1 λm1 + 1
 1 + ,1 =
α e
and 
2 λ2 m2 + 2λm1 + 1
 1 + ,1 = .
α e
The method of moments estimators are the simultaneous solutions of these two equations.
Now consider estimation by the method of maximum likelihood. The LL function of the two
parameters is as follows:

n 
n
log L(α, λ) = n + n log(αλ) + (α − 1) log(1 + λti ) − (1 + λti )α . (14)
i=1 i=1

It follows that the maximum-likelihood estimators (MLEs) are the simultaneous solutions of the
equations:
n  
n n
+ log(1 + λti ) − (1 + λti )α log(1 + λti ) = 0
α i=1 i=1
and
n n n
+ (α − 1) ti (1 + λti )−1 − α ti (1 + λti )α−1 = 0.
λ i=1 i=1

For interval estimation of (α, λ) and tests of hypothesis, one requires the Fisher information
matrix. The elements of this matrix for Equation (14) can be worked out as follows:
 2
∂ log L n  
E − = 2 + nE (1 + λT )α {log(1 + λT )}2 ,
∂α 2 α
 2
∂ log L n    
E − 2
= 2 + n(α − 1)E T 2 (1 + λT )−2 + nα(α − 1)E T 2 (1 + λT )α−2
∂λ λ
and

∂ 2 log L
E − = −nE[T (1 + λT )−1 ] + nαE[T (1 + λT )α−1 log(1 + λT )]
∂αλ
+ nE[T (1 + λT )α−1 ].

By using Lemmas 1 and 2 in the appendix, the expectations in these elements can be calculated
as follows:

E[(1 + λT )α {log(1 + λT )}2 ] = αλeJ (0, α, 2),


E[T 2 (1 + λT )−2 ] = αλeI (2, −2, 1),
Statistics 555

E[T 2 (1 + λT )α−2 ] = αλeI (2, α − 2, 1),


E[T (1 + λT )−1 ] = αλeI (1, −1, 1),
E[T (1 + λT )α−1 log(1 + λT )] = αλeJ (1, α − 1, 1)

and
E[T (1 + λT )α−1 ] = αλeI (1, α − 1, 1).
As shown in the appendix, each of the I and J terms can be expressed as a finite sum of the
complementary incomplete gamma function.
Often with lifetime data, one encounters censored data. There are different forms of censoring:
Type I censoring, Type II censoring, etc. Here, we consider the general case of multicensored
data: there are n subjects of which
• n0 is known to have failed at the times t1 , . . . , tn0 .
• n1 is known to have failed in the interval [si−1 , si ], i = 1, . . . , n1 .
• n2 survived to a time ri , i = 1, . . . , n2 , but not observed any longer.
Note that n = n0 + n1 + n2 . Note that Type I censoring and Type II censoring are contained as
particular cases of multicensoring. The LL function of the two parameters for this multicensoring
data is given by

n0 
n0
log L(α, λ) = n0 + n1 + n2 + n0 log(αλ) + (α − 1) log(1 + λti ) − (1 + λti )α
i=1 i=1


n2 
n1
 
− (1 + λri )α + log exp {−(1 + λsi−1 )α } − exp {−(1 + λsi )α } . (15)
i=1 i=1

It follows that the MLEs are the simultaneous solutions of the equations:

n0   
0n 0 n 2 n
+ log(1 + λti ) − (1 + λti )α log(1 + λti ) − (1 + λri )α log(1 + λri )
α i=1 i=1 i=1


n1
(1 + λsi−1 )α log(1 + λsi−1 ) exp{−(1 + λsi−1 )α }

i=1
exp{−(1 + λsi−1 )α } − exp{−(1 + λsi )α }

n1
(1 + λsi )α log(1 + λsi ) exp{−(1 + λsi )α }
+ =0
i=1
exp{−(1 + λsi−1 )α } − exp{−(1 + λsi )α }

and
n0 n0 n0 n2
+ (α − 1) ti (1 + λti )−1 − α ti (1 + λti )α−1 − α ri (1 + λri )α−1
λ i=1 i=1 i=1

n1
si−1 (1 + λsi−1 )α−1 exp{−(1 + λsi−1 )α }
−α
i=1
exp{−(1 + λsi−1 )α } − exp{−(1 + λsi )α }

n1
si (1 + λsi )α−1 exp{−(1 + λsi )α }
+α = 0.
i=1
exp{−(1 + λsi−1 )α } − exp{−(1 + λsi )α }

The Fisher information matrix corresponding to Equation (15) is too complicated to be


presented here.
556 S. Nadarajah and F. Haghighi

8. Data analysis

In this section, we return to the data set in Table 1. We consider four possible models for the data
set: the gamma, Weibull, EE and the new distributions. The first three distributions are unsuitable
for the following reasons explained in Section 1: (1) increasing hrfs are not allowed when the
mode is fixed at zero and (2) the method of maximum likelihood cannot be implemented.
On the other hand, the new distribution allows for increasing hrfs when the mode is fixed at
zero and guarantees that the fitted pdfs will have zero mode. So, the new distribution provides the
only plausible model for the data in Table 1. Fitting this distribution by the method of maximum
likelihood in Section 7, we obtain log L = 741.5, α̂ = 0.109(0.007), λ̂ = 6.829 × 105 (8.657 ×
104 ), Cov (α̂, λ̂) = −98.486 for the data set in Table 1. The standard errors given within brackets
and the covariances were computed by inverting the Fisher information matrix in Section 7. The
fitted pdfs superimposed with the empirical histograms are shown in Figure 5. The corresponding
probability plot is shown in Figure 6. Both figures suggest that the fit of the new distribution is
reasonable. As a further check, we performed a Kolmogorov Smirnov test for the fit. This test
yielded a p value of 0.125.
8
6
Fitted PDF
4
2
0

0.0 0.2 0.4 0.6 0.8 1.0


Daily Rainfall (in mm)

Figure 5. Fitted pdf of the new distribution (solid curve) and the empirical histogram for the data set in Table 1.
1.0
Expected Cumulative Probability
0.8
0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Observed Cumulative Probability

Figure 6. Probability plot for the fit of the new distribution to the data set in Table 1.
Statistics 557

Acknowledgements
The authors would like to thank the Editor-in-Chief, the Associate Editor and the two referees for carefully reading the
paper and for their comments which greatly improved the paper. The second author wishes to thank the University of
Tehran for providing support for this research.

References

[1] D.N.P. Murthy, M. Xie, and R. Jiang, Weibull Models, Wiley, New Jersey, 2004.
[2] R.D. Gupta and D. Kundu, Exponentiated exponential family: An alternative to gamma and Weibull distributions,
Biom. J. 43 (2001), pp. 117–130.
[3] R.D. Gupta and D. Kundu, Generalized exponential distribution: Existing results and some recent development,
J. Stat. Plan. Inf. 137 (2007), pp. 3537–3547.
[4] W.E. Vaughan, Digest of literature on dielectrics, Vol. 31, Conference on Electrical Insulation, Committee on Digest
of Literature, National Research Council, United States, 1973.
[5] S. Nadarajah, A truncated inverted beta distribution with application to air pollution data, Stoch. Environ. Res. Risk
Assess. 22 (2008), pp. 285–289.
[6] J.R.M. Hoskings, L-moments: Analysis and estimation of distribution using linear combinations of order statistics,
J. R. Stat. S. Ser. B 52 (1990), pp. 105–124.
[7] C.E. Bonferroni, Elementi di Statistica Generale, Seeber, Firenze, 1930.
[8] A. Rényi, On Measures of Entropy and Information, Proceedings of the Fourth Berkeley Symposium on Mathematical
Statistics and Probability Vol. I, University of California Press, Berkeley, 196, pp. 547–561.
[9] M.R. Leadbetter, G. Lindgren, and H. Rootzén, Extremes and Related Properties of Random Sequences and
Processes, Springer Verlag, New York, 1987.
[10] I.S. Gradshteyn and I.M. Ryzhik, Table of Integrals, Series, and Products, 6th ed., Academic Press, San Diego, 2000.

Appendix

We need the following lemmas.

Lemma A.1 Let


∞  
I (a, b, c) = t a (1 + λt)α−1+b exp −c(1 + λt)α dt.
0

We have
a  
1  a a−i −(b+i)/α−1 b+i
I (a, b, c) = (−1) c  + 1, c ,
αλa+1 i α
i=0
∞
where (a, x) = x t a−1 exp(−t)dt denotes the complementary incomplete gamma function.

Proof Substituting y = (1 + λt)α , one can write



1
I (a, b, c) = y b/α (y 1/α − 1)a exp(−cy) dy,
αλa+1 1
a  ∞
1  a
= (−1) a−i
y (b+i)/α exp(−cy) dy.
αλa+1 i 1
i=0

The result of the lemma follows by the definition of the complementary incomplete gamma function. 

Lemma A.2 Let


∞  
J (a, b, c) = t a (1 + λt)α−1+b (log(1 + λt))c exp −(1 + λt)α dt.
0

We have
a 
 
1 a ∂c 
J (a, b, c) = (−1)a−i (ν, 1) ,
α c+1 λa+1 i ∂ν c
ν=(b+i)/α+1
i=0

where (·, ·) denotes the complementary incomplete gamma function.


558 S. Nadarajah and F. Haghighi

Proof Substituting y = (1 + λt)α , one can write



1  a
J (a, b, c) = c+1 a+1 y b/α y 1/α − 1 (log y)c exp(−y) dy,
α λ 1

a  ∞
1 a
= c+1 a+1 (−1)a−i y (b+i)/α (log y)c exp(−y) dy.
α λ i 1
i=0

The result of the lemma follows by Equation (4.358.1) in Gradshteyn and Ryzhik [10]. 

Lemma A.3 Let


∞  
K(a, b) = t b (1 + λt)α−1 exp −(1 + λt)α dt.
a

We have
b  
1  b i
K(a, b) = (−1) b−i
 + 1, (1 + λa) α
,
αλb+1 i α
i=0

where (·, ·) denotes the complementary incomplete gamma function.

Proof Similar to the proof of Lemma A1. 

You might also like