ACCEPTED MANUSCRIPT

A New Count Model Generated from Mixed Poisson

Transmuted Exponential Family with an application to
Health Care Data
Deepesh Bhati†1 , Pooja Kumawat1 , and E. Gómez–Déniz∗2
1
Department of Statistics, Central University of Rajasthan
2
Department of Quantitative Methods in Economics and TiDES Institute.
University of Las Palmas de Gran Canaria, Spain.

Abstract

In this article, a new mixed Poisson distribution is introduced. This new distribution is ob-

tained by utilizing mixing process, with Poisson distribution as mixed distribution and Trans-

muted Exponential as mixing distribution. Distributional properties like unimodality, moments,

over-dispersion, infinite divisibility are studied. Three methods viz. Method of moment, Method

of moment and proportion, and Maximum likelihood method are used for parameter estimation.

Further, an actuarial application in context of aggregate claim distribution is presented. Finally,

to show the applicability and superiority of proposed model, we discuss count data and count

regression modeling and compare with some well established models.

Keywords Count regression, Health care data, Over-dispersion, mixed Poisson Distribution,

Transmuted exponential family.

Acknowledgements

*EGD was partially funded by grant ECO2013–47092 (Ministerio de Economı́a y Competitividad,

Spain). † deepesh.bhati@curaj.ac.in(Corresponding author).

ACCEPTED MANUSCRIPT
1
ACCEPTED MANUSCRIPT

1 Introduction

Recently, count data modelling have drawn attention of many researchers working in different ar-

eas like Insurance, economics, social sciences and biometrics. For this purpose, traditional models

like Poisson, Negative binomial, Geometric and their generalizations were used. However, often

it has been found that count data exhibits over-dispersion (variance > mean) and corresponding

distribution function shows long tail behaviour. Hence there is further scope to modify/generalize

these traditional models. In last two decades, many methods have been proposed by the researchers

to develop new models for count data, one such method is mixtures of distributions, which have

been widely used for modelling non-homogeneous populations(see Karlis and Xekalaki (2005)).

Count data model proposed by mixing Poisson parameter (λ) with a continuous distribution over-

comes the problems of under or equi-dispersion and possesses many interesting properties(see

Karlis and Xekalaki (2005)). For more detail, on Mixed Poisson distribution, readers’ may see the

following references

• Poisson-Gamma (Negative Binomial) -Greenwood and Yule (1920)

• Poisson Beta with specific parameter values (Yule) -Simon (1955)

• Poisson Beta Type-2 -Gurland (1958)

• Poisson-Exponential Beta -Pielou (1962)

• Poisson Truncated Poisson -Patil (1964)

• Poisson Beta Type 1 -Holla and Bhattacharya (1965)

• Poisson Truncated Gamma -Bhattacharya (1966)

• Poisson Linear Exponential -Sankaran (1969)

ACCEPTED MANUSCRIPT
2
ACCEPTED MANUSCRIPT

• Poisson Lindley -Sankaran (1970)

• Poisson Power Function -Rai (1971)

• Poisson Lognormal -Bulmer (1974)

• Poisson Generalized Inverse Gaussian -Sichel (1974)

• Poisson Inverse Gaussian -Sichel(1975)

• Poisson Gamma Product Ratio(Generalized Waring) -Irwin (1975)

• Poisson Generalized Pareto -Kempton (1975)

• Poisson-Poisson Distribution(Neyman) -Douglas (1980)

• Poisson Pearson’s Family of Distribution -Albrecht (1982)

• Poisson Generalized Gamma -Albrecht (1984)

• Poisson Truncated Beta Type 2 -Willmot (1986)

• Poisson Log-Student -Gover and O’ Muircheartigh (1987)

• Poisson Shifted Gamma -Ruhonen (1988)

• Poisson Exponential(Geometric) -Johnson et al. (2005)

• Poisson-Other Discrete Distribution(Neyman) -Johnson et al. (2005)

• Poisson Linear Exponential -Kling and Goovaerts (1993)

• Poisson Inverse Gamma - Willmot (1993)

• Poisson Truncated Gamma -Willmot (1993)

• Poisson Pareto -Willmot (1993)

• Poisson Shifted Pareto -Willmot (1993)

ACCEPTED MANUSCRIPT
3
ACCEPTED MANUSCRIPT

• Poisson Modified Bessel -Ong and Muthaloo (1995)

• Poisson-Power Variance -Hougaard et al. (1997)

• Poisson-Lomax -Al-Awadhi and Ghitany (2001)

• Poisson Lindley Distribution(Size Biased)- Al-Multairi (2008)

• Zero Truncated Poisson-Lindley -Ghitany et al. (2008)

• Poisson Generalized-Lindley -Mahmoudi and Zakerzadah (2010)

• Poisson-Lindley-Beta -Gómez-Déniz et al. (2014)

• Poisson-Generalized Gamma -Sastry et al. (2014)

• Poisson-Marshall-Olkin-Generalized Exponential -Gómez–Déniz et al. (2015)

The above literature present more flexible distribution which can be used in count data modelling.

In the present paper, a new discrete distribution is obtained by mixing Poisson distribution with

Transmuted-Exponential distribution which is proposed by Shaw and Buckley (2007). To the best

of our knowledge, we have not come across any literature on discrete distribution, particularly

on mixed Poisson distribution, where transmuted family is used as mixing distribution, except

Chakraborty and Bhati (2016) who proposed discretization of transmuted exponential distribution

called transmuted geometric distribution.

The rest of the article is structured as follows. Section 2 describes the theoretical development of

the new count distribution, including some properties. Estimation of parameters by three methods

are discussed in Section 4. An actuarial application of the proposed model is presented in Section

5. Section 6 comprises of applications of proposed distribution in count data and count regression

modelling. Finally some conclusions are drawn in Section 7.

ACCEPTED MANUSCRIPT
4
ACCEPTED MANUSCRIPT

2 Proposed model

Shaw and Buckley (2007) proposed a novel technique to introduce skewness and kurtosis into sym-

metric as well as to other family of distribution. In this technique they use “transmutation map”,

which is functional composition of cumulative distribution of one distribution with the quantile

function of another. One of the family discussed by Shaw and Buckley is “Transmuted Exponen-

tial family” denoted as (T ED(α, θ)) with density function given as






 x < 0,
0
 for
f (x) = 
 (1)



 ˉ −θx + 2αθe−2θx for
αθe x ≥ 0,

where αˉ = 1 − α with |α| < 1 and θ > 0.

Considering the fact that (T ED(α, θ)) possesses a wide range of statistical properties as compared

to exponential distribution, we introduce a new mixed Poisson distribution assuming (T ED(α, θ))

as prior distribution for Poisson parameter (λ). The definition of proposed model is as follows

Definition 1 A random variable(rv) X is said to follow Poisson-Transmuted Exponential distribu-

tion if it possesses the following stochastic representation

X|λ ∼ Po(λ)

λ|α, θ ∼ T ED(α, θ)

for λ > 0, |α| ≤ 1, αˉ = 1 − α and θ > 0. We call the unconditional distribution of X as Poisson-

Transmuted exponential distribution and denote it by PT E(α, θ) and its pmf is derived in Theorem

1.

Theorem 1 If X ∼ PT E(α, θ), then probability mass function (pmf) of X is

ACCEPTED MANUSCRIPT
5
ACCEPTED MANUSCRIPT
!
αˉ 2α
P(X = x; α, θ) = θ + , x = 0, 1, . . . ,
(1 + θ) x+1 (1 + 2θ) x+1
where αˉ = 1 − α with |α| ≤ 1 and θ > 0.

Proof: If X|λ ∼ Po(λ) and λ ∼ T ED(α, θ) then

Z∞
P (X = x) = θ P(X = x|λ) fΛ (λ; α, θ)dλ
0
Z∞
e−λ λ x −θλ  
=θ θe αˉ + 2αe−θλ dλ
x!
0
Z∞  
θ
= e−λ λ x αˉ + 2αe−θλ e−θλ dλ
x!
0
!
αˉ 2α
=θ + , (2)
(1 + θ) x+1 (1 + 2θ) x+1

where |α| ≤ 1 and θ > 0. 

Remarks

1. It is easy to see that, pmf in (2) can be rewritten as a finite mixture of two geometric distribu-

tion in the form,

! !
1 1
p x = αˉ Geo + α Geo . (3)
1+θ 1 + 2θ

where Geo(p) denotes the pmf of geometric random variable with parameter p.
 
2. As α → 0, (2) reduces to Geometric distribution Geo 1
1+θ
.
 
3. As α → 1, (2) reduces to Geometric distribution 1
1+2θ
.

ACCEPTED MANUSCRIPT
6
ACCEPTED MANUSCRIPT

Further, the cumulative distribution function(cdf) and the survival function(sf) of X ∼ PT E(α, θ)

can be given as
X
x !
αˉ 2α
F X (x) = P(X ≤ x) = θ +
n=0
(1 + θ) n+1 (1 + 2θ)n+1
!
αˉ α
=1 − + (4)
(1 + θ) x+1 (1 + 2θ) x+1

and
αˉ α
S X (x) = P(X ≥ x) = + (5)
(1 + θ) x (1 + 2θ) x
PMF of PT E(α, θ) for different values of α, θ, figure 1, confirms the unimodality of T GD rvs.

The probabilities defined in (2) can also be computed by recurrence relation

   
 1+θ x+1 
1  αˉ + 2α 1+2θ 
 p x
p x+1 =    (6)
1 + θ  αˉ + 2α 1+θ x+2 
1+2θ
with
θ (1 + α + 2θ)
p(0) = .
(1 + θ) (1 + 2θ)

Further,
2α(α)θ ˉ 2
p2x − p x+1 p x−1 = − .
(1 + θ) x+2 (1 + 2θ) x+2
For −1 < α < 0, the distribution is infinitely divisible since p2x − p x−1 p x+1 < 0 and p0 , 0, p1 ,

0, see Warde and Katti (1971) for details. In such case the distribution has mode at zero. For

{p x /p x−1 } , x = 1, 2, . . . , to form a monotone increasing sequence, it requires that {p x } to be a

decreasing sequence (see Johnson et al. 2005), which is congruent with the zero vertex of the new

distribution for −1 < α < 0. Moreover, any infinitely divisible distribution define on non-negative

integers is a compound Poisson distribution (see Proposition 9 in Karlis and Xekalaki, 2005), we

conclude that the new pmf presented in this paper is a compound Poisson distribution.

ACCEPTED MANUSCRIPT
7
ACCEPTED MANUSCRIPT

Furthermore, the infinitely divisible distributions play an important role in many areas of statis-

tics, viz, in stochastic processes and in actuarial statistics. When a distribution G is infinitely

divisible then for any integer x ≥ 2, there exists a distribution G x such that G is the x–fold convo-

lution of G x , that is, G = G∗x

x . Also, when a distribution is infinitely divisible an upper bound for

the variance can be obtained when −1 < α < 0 (see Johnson et al. (2005)), which is given by

p1
var(X) ≤ .
p0
Some implications of infinite divisibility are as follows:

1. For k, m = 1, 2, . . . ,
!
k+m
pk+m p0 ≥ pk pm . (7)
m

See Steutel and van Harn (2004), p. 51, Proposition 8.4. Expression (7) ensures that the

sequence ak = k!pk /p0 satisfies ak+l ≥ ak al , for all k and l (Steutel and van Harn (2004) say

that the sequence is therefore super-multiplicative).

2. For all k = 0, 1, . . . , pk ≤ exp(−1). See Steutel and van Harn (2004), p. 56, Proposition

9.2. This property says how small such a pk can be for large values of k and reveals some

information on the tail of the distribution which is infinitely divisible. In fact, this can have

tail larger than that of Poisson distribution (see Steutel and van Harn (2004)).

Furthermore, if the empirical data do not satisfy the inequalities above we can eliminate the

proposed distribution to fit them.

3. The cumulants of an infintely divisible distribution on the set of non–negative integers (as

far as they exist) are non negative, see Steutel and van Harn (2004), p. 47, Corollary 7.2.

This will imply that the skewness of the new distribution is positive, since the third cumulant

equals the third central moment.

4. The distribution is strictly log-concave and strongly unimodal, see theorem 3 in Keilson and

ACCEPTED MANUSCRIPT
8
ACCEPTED MANUSCRIPT

Gerber (1971).

Moreover, it can be observed that (6) possesses relation





 ∀ x ≤ x∗ ,
 p x+1 ≥ p x



 (8)




 p x+1 ≤ p x ∀ x ≥ x∗ ,

for x∗ (mode of PT E rv) depending on parameters α and θ, which indicate uni modality, see Keilson

and Gerber (1971).

As indicated in remark above, since (2) is log-concave and has zero vertex for some values of

parameters α and θ, we can determine the mode(x∗ ) of PT E(α, θ) as the value for which p x is

increasing on (0, 1, ∙ ∙ ∙ , x∗ ) and decreasing on (x∗ + 1, ∙ ∙ ∙ ). It can be easily verified that

2
1. For −1 < α < − (1+2θ)
3+4θ
, the pmf (2) is unimodal and the mode is at x0 = 0.
2 h   i
∗ ∗
2. For − (1+2θ)
3+4θ
< α < 1, the mode of pmf (2) will be at x + 1, with x = log θ ∗
α−1
4α
− 2 , where

θ∗ = 1+θ
1+2θ
. Here [.] denotes the integer part.
 
3. Further if logθ∗ α−1
4α
− 2 is an integer, then pmf (2) will be bimodal and the mode is at x∗ and

x∗ + 1 respectively. 

The new distribution being a mixed Poisson distribution, it has heavier tail than a Poisson distri-

bution with the same mean. Now, let Pr(X = x) be the probability mass function (2) and Pr(X|m)

be the probability function of a simple Poisson distribution with the same mean, say m. Then, as

shown by Feller (1943), Pr(X = 0) > Pr(0|m) and Pr(X = 1)/ Pr(X = 0) ≤ Pr(1|m)/ Pr(0|m) = m.

Reader’s can see asymptotic tail behavior of Poisson distributions in Willmot (1990).

Proposition 1 The probability generating function(pgf) of rv X defined in (2) is given by

θ(1 − t)(1 + α) + 2θ2
PX (t) = . (9)
(1 − t + θ)(1 − t + 2θ)

Proof: The proof is straight forward after using the definition PX (t) = E(t X ) and (2).

ACCEPTED MANUSCRIPT
9
ACCEPTED MANUSCRIPT

Proposition 2 The rth raw moment of PT E(α, θ) is given by

θ(α)Φ(
ˉ 1
1+θ
, −r, 0) 2θαΦ( 1+2θ 1
, −r, 0)
E(X r ) = + , (10)
1+θ 1 + 2θ
P∞
where Φ(z, s, a) = zk (k + a)−s is Hurwitz-Lerch transcendental function.
k=0

In particular, the first four raw moments of X can be obtained easily by putting r =1,2,3,4 in

(10) and are as follows

02−α
μ1 = ,
2θ
0 4 − 3α + 2θ − αθ
μ2 = ,
2θ2
0 24 + 24θ + 4θ2 − 21α − 18θα − 2θ2 α
μ3 = ,
4θ3
0 48 + 72θ + 28θ2 + 2θ3 − 45α − 63θα − 21θ2 − θ3 α
μ4 = ,
2θ4
   
other measures like variance(μ2 ), coefficient of skewness γ1 = μμ3/23 and kurtosis γ2 = μ4
μ22
are as
2

follows

4 + 2θ(2 − α) − α(2 + α)
μ2 = , (11)
4θ2
2(8 + 4θ(3 + θ) − 3α − 2θα(3 + θ) − 3α2 (1 + θ) − α3 )
γ1 = , (12)
[4 − 2θ(α − 2) − α(α + 2)]3/2
8(1 + θ)(2(9 + θ(9 + θ)) − (9 + θ(12 + θ))α) − 8(6 + θ(9 + 2θ))α2 − 12(1 + θ)α3 − 3α4
γ2 = . (13)
(4 − 2θ(α − 2) − α(α + 2))2

For fixed α(θ) both mean and variance decreases with increase in θ(α), whereas the skewness and

kurtosis show mixed behaviour. It could also be noted that skewness takes value more than 3 for

α > 0.5 whereas PT E(α, θ) have high kurtosis value for α > 0.5. Figure 2, gives more clear

picture of these observations.

Further, the coefficient of variation(C.V.) of the distribution comes out to be

√
σ 4 − 2θ(α − 2) − α(α + 2)
C.V. = = .
μ 2−α

ACCEPTED MANUSCRIPT
10
 ACCEPTED MANUSCRIPT

3 Steps to generate random observations from PT E(α, θ)

Since the proposed model is derived from mixed Poisson distribution, following algorithm can be

used to generate observations PT E(α, θ).

Algorithm

Step 1 Generate ui from U(0, 1), i = 1, 2, ∙ ∙ ∙

√ !
1 1−α+ 1+2α−4ui α+α2
Step 2 Corresponding to each ui , compute λi = log θ 2(1−ui )
.

Step 3 Generate xi from Po(λi ).

Step 4 Repeat Step 1,2 and 3 to get a sample from PT E(α, θ)

4 Estimation

In this section we discuss three methods to estimate the parameters α and θ. In the first subsection,

method of moments(MM) is presented. Method of proportion and moment(MPM) is discussed in

second section and finally Maximum likelihood(ML) method is shown as last subsection.

4.1 Method of moments

Given a random sample x1 , x2 , ∙ ∙ ∙ , xn of size n from (2), the moment estimates, α̃ and θ̃, of α and

θ can be obtained by solving the following equations

2−α 4 − 3α + 2θ − αθ
m1 = μ01 = and m2 = μ02 = . (14)
2θ 2θ2

ACCEPTED MANUSCRIPT
11
ACCEPTED MANUSCRIPT

where m1 and m2 are the first and second sample moments. Solving the above equations, we get
q
3m21 m1 4m1 + 9m21 − 4m2
α̃ = 2 + + ,
m1 − m2 m1 − m2
q
−3m1 − 4m1 + 9m21 − 4m2
θ̃ = . (15)
2(m1 − m2 )

Theorem 2 For fixed α, the estimator θ̃ of θ is positively biased, i.e. E(θ̃) > θ.

ˉ and g(t) =
Proof: Let θ̃ = g(X) 2−α
for t > 0. Then
2t

2−α
g00 (t) = > 0.
t3
   
ˉ > g E(X)
Therefore, g(t) is strictly convex. Thus, by Jensens inequality, we have E g(X) ˉ .
   
ˉ = g 2−α = θ, hence we obtain E(θ̃) > θ.
Finally, since g E(X) 2θ

4.2 Method of proportion and moment

Here, we compare the sample proportion of zero(p0 ) and sample mean ( xˉ) with the population

proportion of zero and population mean, i.e, estimated ᾰ and θ̆ will be obtained by solving the

following two equations

!
αˉ 2α 2−α
p0 = θ + and xˉ = . (16)
1 + θ 1 + 2θ 2θ
Thus, solving the above two equations, we obtain the estimates ᾰ and θ̆ as follows
 q 
 2
xˉ 9 − 10p0 − 8 xˉ p0 + p0 
1  3 xˉ 3 xˉ p0 
ᾰ = 4 −  + −  ,
2  xˉ + p0 − 1 xˉ + p0 − 1 xˉ + p0 − 1 
q
3 − 3p0 + 9 − 10p0 − 8 xˉ p + p20
θ̆ = . (17)
4( xˉ + p0 − 1)

ACCEPTED MANUSCRIPT
12
ACCEPTED MANUSCRIPT

4.3 Maximum likelihood estimation

Let x1 , x2 , ∙ ∙ ∙ , xn be n random observations from PT E(α, θ) distribution. The log-likelihood(LL)

function for the vector of parameter Θ = (α, θ)> is written as

X
n !
αˉ 2α
l(α, θ|x) = n log θ + log + .
i=1
(1 + θ) xi +1 (1 + 2θ) xi +1
By taking the first derivative with respect to both parameters. The normal equations are written as
 (α) ˉ

∂l n X n (x + 1)
(1+θ) x+2
+ 4α
(1+2θ) x+2
= −  (1−α)  = 0, (18)
∂θ θ x=0 x+1 + 2α
x+1
(1+θ) (1+2θ)
 
∂l X n − 1
(1+θ) x+1 + (1+2θ)
2
x+1
=  (α)  = 0. (19)
∂α x=0 ˉ
x+1 +
2α
x+1
(1+θ) (1+2θ)

As the above two equations cannot be solved explicitly, we obtain ML estimator of α and θ numeri-

cally by using nlm() function in R. Further, the second order derivatives of log-likelihood function

are
 2 
Xn  − θ
+ 2θ 
∂l2  (1+θ) i x +1 (1+2θ) ix +1

=−     , (20)
∂α 2  (1−α)θ 2αθ 2
i=1
(1+θ) xi +1
+ (1+2θ) xi +1


n − θ
 (1+θ)xi +1 + (1+2θ)xi +1 + (1+θ)xi +2 (xi + 1) − (1+2θ)xi +2 (xi + 1)
X 1 2 4θ
∂2 l
= 
∂θ∂α i=1  (1−α)θ
+ (1+2θ)2αθ
(1+θ) xi +1 xi +1
   (α)(1−θx
ˉ 2α(1−2θxi ) 

i)
θ − (1+θ)1 xi +1 + (1+2θ)2
xi +1 x +2 + x +2 
−  (α)θ
(1+θ) i

(1+2θ) i  , (21)
ˉ 2αθ 2 
(1+θ) xi +1
+ (1+2θ)xi +1

 2
n  (1−α)
n X  (1+θ)xi +2 (1 − θxi ) + (1+2θ)xi +2 (1 − 2θxi )
2α
∂2 l
=− 2 −   (1−α)θ 2
∂θ2 θ  2αθ
i=1 x
(1+θ) i +1 + x
(1+2θ) i +1
 (1−α) 
(xi + 1) (1+θ)xi +3 (2 − θxi ) + (1+2θ) − i 
8α
xi +3 (1 θx ) 
+  . (22)
(1−α)θ 2αθ
+ (1+2θ)1+xi 
(1+θ)1+xi

ACCEPTED MANUSCRIPT
13
ACCEPTED MANUSCRIPT

The Fisher information matrix can be computed by using the approximations

! !
∂2 l ∂2 l
E − 2 ≈− ,
∂α ∂α2 α̂,θ̂
! !
∂2 l ∂2 l
E − ≈− ,
∂α∂θ ∂α∂θ α̂,θ̂
! !
∂2 l ∂2 l
E − 2 ≈−
∂θ ∂θ2 α̂,θ̂

where α̂ and θ̂ be the maximum likelihood estimators of α and θ.

5 Collective risk model

In non-life insurance portfolio, for example in motor insurance, the aggregate loss (S ) is a rv

defined as the sum of claims incurred in a certain period of time. Let us consider now the following

actuarial model. Let X be the number of claims in a portfolio of policies in a time period. Let Yi ,

i = 1, 2, . . . be the amount of the i–th claim and S = Y1 +Y2 +∙ ∙ ∙+YX the aggregate claims generated

by the portfolio in the period under consideration. As usual, two fundamental assumptions are

made in risk theory: (i) The rvs Y1 , Y2 , . . . are independent and identically distributed and follow

a discrete (continuous) distribution with pmf (pdf) h(y) and (ii) The rvs X, Y1 , Y2 , . . . are mutually

independent. The distribution of the aggregate claims S is called the compound distribution and
P
assuming that Yi , i = 1, 2, . . . X are discrete rvs, the pdf of S is fS (y) = ∞ ∗x ∗x
x=0 p x h (y), where h (∙)

denotes the x–fold convolution of h(∙) and p x is given in (2). There exists an extensive literature

dealing with compound mixture Poisson distributions see Willmot (1986,1993) and Antzoulakos

and Chadjiconstantinidis (2004). An extensive review of the topic can be found in Sundt and Vernic

(2009). Based on the recursion provided by Panjer (1981) for the Poisson distribution, Sundt and

Vernic (2009, chapter 3, p.68) developed a simple algorithm to provide the probabilities of the rv

S when the amount of the single claim follows a discrete distribution with pmf h(.).

From Panjer (1981) recursion for the total claim amount when the pmf of the Poisson distribution

ACCEPTED MANUSCRIPT
14
ACCEPTED MANUSCRIPT

is assumed as the distribution of the number of claims is given by

λX
x
fS (x|λ) = yh(y) fS (x − y|λ), x = 1, 2, . . . (23)
x y=1

while fS (0|λ) = e−λ . Following Sundt and Vernic (2009, p.68), after multiplying both sides of (23)

by λi f (λ) and integrating we get,

1X
y
fSi (y)
= yh(x) fSi+1 (y − x), i = 0, 1, . . . ; y = 1, 2, . . .
y x=1
R∞
where fSi (y) = 0 λi fS (y|λ) f (λ)dλ. Now, starting with fS (0) = p0 , the probabilities fS (1), fS (2), . . .

can be evaluated by the algorithm described in Sundt and Vernic (2009, p.68) and taking into ac-

count (2).

For more on classic risk model, see Freifelder (1974), Rolski et al. (1999), Nadarajah and Kotz

(2006a and 2006b) and the references therein. Here, we consider two such situations: In first situa-

tion, the primary distribution is as defined in Section 2 and claim severity distribution as exponen-

tial distribution with parameter (α) and as we know, Erlang loss distribution may arise in insurance

settings when the individual claim amount is the sum of exponentially distributed claims. Hence

in the second situation, Erlang distribution with parameters r and α is considered as secondary

distribution.

Theorem 3 If we assume a PT E(α, θ) as primary distribution and an exponential distribution

P
X
with parameter (λ) as secondary distribution, then the pdf of aggregate loss rv S = Yi is given
i=0
by

 −θλy −2θλy 
 αλe
ˉ 1+θ 2αλe 1+2θ 

fS (y) = θ  2
+ 2
, for y > 0,
(1 + θ) (1 + 2θ)

ACCEPTED MANUSCRIPT
15
ACCEPTED MANUSCRIPT

whereas,
!
αˉ 2α
fS (0) = θ + .
(1 + θ) (1 + 2θ)
Proof: By assuming that the claim severity follows an exponential distribution with parameter

λ > 0, since the x-fold convolution of exponential distribution is gamma distribution with parame-

ters x and λ, the x−fold convolution is given by

∗x λx
f (y) = y x−1 e−λy , x = 1, 2, ...
(x − 1)!
Then the pdf of rv S is given by

∞
X ! x x−1 −λy
αˉ 2α λy e
fS (y) = θ +
x=1
(1 + θ) x+1 (1 + 2θ) x+1 (x − 1)!
X ∞ x−1 x−1
!
−λy αˉ x y 2α x y
= θe θ λ + λ
x=1
(1 + θ) x+1 (x − 1)! (1 + 2θ) x+1 (x − 1)!
 λy λy 
 αλe
−λy  ˉ 1+θ 2αλe 1+2θ 

= θe  2
+ 2

(1 + θ) (1 + 2θ)
 −θλy −2θλy 
 αλe
ˉ 1+θ 2αλe 1+2θ 
= θ  2
+ 2
.
(1 + θ) (1 + 2θ)

Theorem 4 If we assume PT E(α, θ) as primary distribution and an Erlang distribution (2, λ) with
PX
parameter (λ) > 0, as secondary distribution then the pdf of aggregate loss rv S = i=0 Yi is given

by

    yλ  
 αˉ sinh √ yλ 2α sinh √ 
−λy  (1+θ) (1+2θ) 
fS (y) = λθe   + 
(1 + θ)3/2 (1 + 2θ) 3/2

with
!
αˉ 2α
fS (0) = θ + .
(1 + θ) (1 + 2θ)
Proof: By assuming that the claim severity follows Erlang (2,λ) distribution, then the x− fold

convolution of Erlang distribution is gamma distribution with parameters (2x, λ) .The x− fold con-

ACCEPTED MANUSCRIPT
16
ACCEPTED MANUSCRIPT

volution is given by

λ2x
f ∗x (y) = y2x−1 e−λy , x = 1, 2, ...,
(2x − 1)!
Then, the pdf of the aggregate rv S is given by

∞
X ! 2x 2x−1 −λy
αˉ 2α λ y e
fS (y) = θ +
x=1
(1 + θ) x+1 (1 + 2θ) x+1 (2x − 1)!
X∞ 2x−1 2x−1
!
−λy αˉ 2x y 2α 2x y
=θe θ λ + λ
x=1
(1 + θ) x+1 (2x − 1)! (1 + 2θ) x+1 (2x − 1)!
  yλ   yλ  
 αˉ sinh √ 2α sinh √(1+2θ) 
=λθe−λy   .
(1+θ)
+
(1 + θ)3/2 (1 + 2θ)3/2 

6 Data analysis

6.1 Count data modelling

In this subsection, the applicability of PT E(α, θ) distribution is shown by considering a data set

representing epileptic seizure counts considered earlier by Albert(1991) and Hand et. al.(1994) p.

133 and is compared with the following five mass function namely

i Generalized Poisson-Lindley Distribution(GPL(α, θ)) (Mahmoudi and Zakerzadeh(2010)):

Γ(x + α) θα+1  x + α
f (x; α, θ) = α + , x = 0, 1, . . . , α, θ > 0.
x!Γ(α + 1) (1 + θ)1+x+α 1+θ

ii Weighted Generalized Poisson Distribution (WGPD(a, s, b)) (Chakraborty (2010)):

1 (a + xb) x+s e−xb

f (x; a, s, b) = , x = 0, 1, . . . , a > 0, |b| < 1
x! K(a, s, b)

ACCEPTED MANUSCRIPT
17
ACCEPTED MANUSCRIPT
∞
P (a+xb) x+s e−xb
where K(a, s, b) = x!
is a normalizing constant.
x=0

iii Poisson Distribution (Po(λ)):

e−λ λ x
f (x; λ) = , x = 0, 1, . . . , λ > 0.
x!

iv Negative Binomial Distribution (NB(r, p)):

!
x+r−1 x
f (x; r, p) = p (1 − p)r , x = 0, 1, . . . , r > 0, 0 < p < 1.
x

v A New Generalized Two Parameter Poisson-Lindley Distribution (NGPL(α, θ)) Bhati et

al.(2015)):
!
θ2 α(x + 1)
f (x; α, θ) = 1+ , x = 0, 1, . . .
(θ + α)(1 + θ) x+1 (1 + θ)

In each of these five distributions, the parameters are estimated by using the maximum likeli-

hood method and for comparability among these distributions log likelihood and chi-square values

are computed. Though, the log likelihood value of PT E(α, θ) falls close to NGPL and WGP but

minimum chi-square value for PT E(α, θ) distribution compared to other distributions confirms the

superiority of PT E(α, θ) over other models. Hence PT E(α, θ) could be a possible choice to model

over-dispersed and long tail data set.

6.2 Count regression including covariates

In last two decades researchers have contributed significantly in the area of counts regression mod-

elling, for example Duan et al.(1983), Christensen et al.(1987), Cameron et al.(1998), Cartwright

et al.(1992) and Deb and Trivedi(1997), Gómez–Déniz (2010). Here, in this section, we present

an application of the proposed distribution in count regression modelling, where we consider the

count variable (Yi ) (also called response variable) depends on one or more than one exogenous

variable (xi = x1 , x2 , ∙ ∙ ∙ , x s ).

ACCEPTED MANUSCRIPT
18
ACCEPTED MANUSCRIPT

In order to achieve this goal, let us consider the following re-parametrization

ν
ν=2−α and θ= , (24)
2μ
where, μi , the mean of the response variable (Yi ), and is related with set of independent variables

with log-link function log (μi ) = xi> β, for i = 1, 2, ..., s and β is a vector of unknown regression

coefficients.

The above link function will ensure the positivity of μi and after re-parametrization, the log-

likelihood of the new model including covariates will be written as

 
X  
(2 − ν)ν 
n
 ν(ν − 1)
ln (ν, β|yi , xi ) = log 
  yi +1 +  yi +1  . (25)
i=1 2μi 1 + 2μν i μi 1 + μνi
The parameters (ν, β1 , β2 , ..., β s ) in the above log-likelihood function can be estimated by maxi-

mizing the log-likelihood function for which we use an in-built R function called optim(), by

initializing the parameters chosen from Poisson regression model.

6.3 Illustrative example

The US National Medical Expenditure Survey 1987/88 (NMES) data were considered which can

be obtained from Journal of Applied Econometrics, 1997 Data Archive at http://qed.econ.queensu.ca/jae/1997-

v12.3/ deb-Trivedi/. This data were originally used by Deb and Trivedi(1997) in their analysis of

various measures of healthcare utilization. The data set consists of 4406 individuals covered by

Medicare, the USA public insurance programme. Here we consider, the number of stays after

hospital admission(HOSP) as the response variable because it possesses characteristics like over-

dispersion and large proportion of zeros. Detail summary statistics of response variable as well as

the set of explanatory variables are given in Table 2. Here the mean and variance of the response

variable indicate the over-dispersion as well as existence of large number of zeros. Hence it is

adequate to apply our model for this dataset.

ACCEPTED MANUSCRIPT
19
ACCEPTED MANUSCRIPT

The maximum likelihood estimates of Poisson Po, Negative Binomial (NB) and Poisson-

Transmuted Exponential(PT E) Regression Model, including the intercept β1 (the regression es-

timate when all variables in the model are evaluated at zero) were presented in Table 3. For

comparability of these models we use the value of the maximum log-likelihood function (lmax ) and

Akaike information criterion (AIC) defined as −2lmax +2k, where k being the number of parameters

in the regression model. These values are shown in Table 4.

Further, it is observed that the estimates of all parameters are found significant at 5% level of

significance as p-value for all the parameter estimates are less then 5%level of significance. It can

be clearly seen from Table 4 that the log-likelihood value of PT E regression model is highest,

whereas AIC also indicate the better fit of PT E regression model compared to other two models.

7 Comments

In this paper we have proposed a new mixed Poisson distribution and derived its distributional

properties. Actuarial application to model aggregate claim size were presented. Fitting of PT E

confirms its flexibility and potentiality in modelling of count data and count regression. Hence as

a closing remark, we believe that PT E can also be considered as a suitable model for count data

analysis.

Acknowledgements

The authors acknowledge profound thanks to anonymous referee for giving comments which have

immensely improved the presentation of the paper.

ACCEPTED MANUSCRIPT
20
ACCEPTED MANUSCRIPT

References

[1] Antzoulakos, D., Chadjiconstantinidis, S. (2004). On mixed and compound mixed Poisson

distributions. Scandinavian Actuarial Journal, 3, 161–188.

[2] Albrecht P. (1982). On Some Statistical Methods connected with the Mixed Poisson Process,

Scandinavian Actuarial Journal, 9, 1–14.

[3] Albrecht P. (1984). Laplace Transforms, Mellin Transforms and Mixed Poisson Process, Scan-

dinavian Actuarial Journal, 11, 58–64.

[4] Al-Awadhi, S.A. and Ghitany M.E. (2001). Statistical Properties of Poisson Lomax Distribu-

tion and its application to Repeated Accident Data, Journal of Applied Statistical Science, 10,

365–372.

[5] Bhattacharya S.K. (1966). Confluent Hypergeometric Distributions of Discrete and Continuous

type with Application to Accident Proneness, Bulletin of the Calcutta Statistical Association, 15,

20–31.

[6] Bhati D., Sastry, D.V.S., Qadri. P.Z. (2015). A New Generalized Poisson Lindley Distribution,

applications and Properties, Austrian Journal of Statistics, 44(4), 35–51.

[7] Bulmer M.G. (1974). On Fitting The Poisson Lognormal Distribution to species Abundance

Data, Biometrics, 30, 101–110.

[8] Cameron, A. C. and Trivedi, P. K. (1998). Regression Analysis of Count Data, Cambridge,

U.K., Cambridge University Press.

[9] Cartwright W.,Hu S.T. and Huang L-F. (1992). Impact of varying Medigap insurance coverage

on the use of medical services of the elderly, Applied Economics, 2(4), 529-539.

ACCEPTED MANUSCRIPT
21
ACCEPTED MANUSCRIPT

[10] Chakraborty S. (2010). On Some Distributional Properties of the Family of Weighted Gener-

alized Poisson Distribution. Communications in Statistics-Theory and Methods, 39(15), 2767–

2788.

[11] Chakraborty S. and Bhati D. (2016). Transmuted Geometric Distribution with application

in modelling and Regression Analysis, Statistics and Operation Research Transaction, 40(1),

153–176.

[12] Deb. P, Trivedi, P.K. (1997). Demand for Medical Care by the Elderly: A Finite Mixture

Approach. Journal of Applied Econometrics, 12, 313-336.

[13] Dean, C., Lawless J.F. and Willmot G.E. (1989). A mixed Poisson-Inverse Gaussian regres-

sion model, Canadian Journal of Statistics, 17(2), 171-181.

[14] Douglas J.B. (1980). Analysis with Standard Contagious Distributions, Statistical Distribu-

tions in Scientific Work Series 4, International Cooperative Publishing House.

[15] Duan N., Manning W., Morris C., and Newhouse J. (1983). A comparison of alternative

models for the demand for medical care, Journal of Business and Economic and Statistics 1,

115-126.

[16] Freifelder, F. (1974). Statistical decision theory and procedures, in Credibility theory and

applications, P.M. Kahn, ed., Academic Press, New York, 71-88.

[17] Gaver D. and O’Muircheartaigh I.G., (1987). Robust Empirical Bayes Analysis of Event

Rates, Technometrics, 29, 1–15.

[18] Ghitany M.E., Atieh B. and Nadarajah S. (2008). Zero-truncated Poisson-Lindley distribution

and its application, Mathematics and Computers in Simulation, 79, 279–287.

[19] Ghitany, M.E. and Al-Mutairi, D.K. (2009). Estimation methods for the discrete Poisson-

Lindley distribution, Journal of Statistical Computation and Simulation, 79(1), 1–9.

ACCEPTED MANUSCRIPT
22
ACCEPTED MANUSCRIPT

[20] Greene, W. (2008). Functional forms for the Negative Binomial model for count data. Eco-

nomics Letter, Vol(99), 585–590.

[21] Ghitany, M.E. and Al-Mutairi, D. K. (2008). Size-biased Poisson-Lindley Distribution and its

Application, Metron, LXVI(3), 299–31.

[22] Greenwood and Yule G. (1920). An enquiry into the nature of frequency distributions Repre-

sentative of Multiple happenings with particular reference to the occurrence of Multiple attacks

of disease or of repeated Accidents, Journal of Royal Statistical Society, 83, 255–279.

[23] Gómez–Déniz E. (2010). Another generalization of the geometric distribution, Test, 19, 399–

415.

[24] Gómez–Déniz E. and Calderı́n–Ojeda E. (2015). Parameters Estimation for a New Gener-

alized Geometric Distribution, Communications in Statistics - Simulation and Computation,

44(8).

[25] Gurland J. (1958). A Generalized Class of Contagious Distributions, Biometrics, 14, 229–

249.

[26] Holgate, P. (1970). The modality of some compound Poisson distribution. Biometrika, 57,

666–667.

[27] Holla M.S. Bhattacharya S.K. (1965). On a Discrete Compound Distribution, Annals of the

Institute of Statistical Mathematics, 15, 377–384.

[28] Hougard P. (1997). Analysis of Over Dispersed Count Data by Mixture of Poisson Variables

and Poisson Processes, Biometrics, 53, 1225–1238.

[29] Irwin J. (1975). The generalized Waring Distribution Parts I, II,III, Journal of the Royal Sta-

tistical Society A, 18-31(Part I), 204–227(Part II), 374-384(Part III).

ACCEPTED MANUSCRIPT
23
ACCEPTED MANUSCRIPT

[30] Johnson N.L., Kotz S. and Kemp A.W. (2005). Univariate Discrete Distributions, Second

Edition, John Wiley and Sons.

[31] Karlis, D. and Xekalaki E. (2005). Mixed Poisson distributions. International Statistical Re-

view, 73, 35-58.

[32] Keilson J. and Gerber H. (1971). Some results for Discrete Unimodality, Journal of the Amer-

ican Statistical Association, 66(334), 386-389.

[33] Kempton R.A. (1975). A generalized Form of Fisher’s Logarithmic Series, Biometrics, 62,

29-38.

[34] Kling B. and Goovaerts M. (1993). A Note on Compound Generalized Distributions, Scandi-

navian Actuarial Journal, 20, 60-72.

[35] Mahmoudi, E. and Zakerzadeh, H. (2010). Generalized Poisson-Lindley distribution. Com-

munications in Statistics-Theory and Methods, 39, 1785-1798.

[36] Nadarajah, S. and Kotz, S. (2006a). Compound mixed Poisson distributions I. Scandinavian

Actuarial Journal, 3, 141162.

[37] Nadarajah, S. and Kotz, S. (2006b). Compound mixed Poisson distributions II. Scandinavian

Actuarial Journal, 3, 163181.

[38] Ong S.H. and Muthaloo S. (1995). A Class of Discrete Distribution Suited to Fitting very

Long Tail Data, Communication in Statistics-Simulation and Computation, 24, 929-945.

[39] Ong, S.H. (1995). Computation of Probabilities of a generalised log-series and related distri-

butions, Communication is Statistics-Theory and Methods, 24, 253-271.

[40] Panjer, H.H. (1981). Recursive evaluation of a family of compound distributions. Astin Bul-

letin, 12, 22-26.

ACCEPTED MANUSCRIPT
24
ACCEPTED MANUSCRIPT

[41] Patil G.P. (1964). On Certain Compound Poisson and Compound Binomial Distributions,

Sankha A, 27, 929-945.

[42] Pielou E. (1962). Run of One Species with Respect to Another in Transects Through Plant

Population, Biometrics, 18, 579-593.

[43] Rai G. (1971). A Mathematical Model for Accdent Proneness, Trabajos Estadistica, 22, 207-

212.

[44] Rolski, T., Schmidli, H. Schmidt, V. and Teugel, J. (1999) Stochastic processes for insurance

and finance. John Wiley and Sons.

[45] Ruohonen M. (1988). A model for the Claim Number Process, ASTIN Bulletin, Vol(18),57-

68.

[46] S. Christensen, S. Long, and J. Rodgers, (1987). Acute health care costs for the aged Medicare

population: Overview and policy options, Millbank Q. 65 , 397425.

[47] Sankaran M. (1969). On certain Properties of a Class of Compound Poisson Distributions,

Sankha B, 32, 353-362.

[48] Sankaran M. (1970). The Discrete Poisson Lindley Distribution, Biometrics, 26:1, 145-149.

[49] Sastry D.V.S., Bhati D., Rattihalli R.N. and Gómez–Déniz E. (2016). Zero Distorted General-

ized Geometric Distribution. Communication in Statistics: Theory and Methods.45(18), 5427-

5442.

[50] Shaw, W. and Buckley I. (2007). The alchemy of probability distributions: beyond Gram-

Charlier expansions and a Skew-Kurtotic-normal distribution from a rank transmutation map.

Research report.

[51] Sichel H.S. (1974). On a Distribution Representing Sentence-Length in Written Prose, Jour-

nal of The Royal Statistical Society A, 137, 25-34.

ACCEPTED MANUSCRIPT
25
ACCEPTED MANUSCRIPT

[52] Sichel H.S. (1975). On a Distribution Law for Word Frequencies, Journal of the American

Statistical Association, 70, 542-547.

[53] Simon P. (1955). On a Class of Skew Distributions, Biometrika, 42, 425–440.

[54] Steutel, F.W. and Van Harn, K. (2004). Infinite Divisibility of Probability Distributions on the

Real Line. Marcel Dekker, New York.

[55] Sundt, B., Vernic, R. (2009) Recursions for Convolutions and Compound Distributions with

Insurance Applications. Springer–Verlag, New York.

[56] Warde, W.D. and Katti, S.K. (1971). Infinite divisibility of discrete distributions II. The An-

nals of Mathematical Statistics, 42, 3, 1088–1090.

[57] Willmot G.E. (1986) Mixed Poisson Distribution, ASTIN Bulletin Supplement, Vol(16), 59–

79.

[58] Willmot, G.E. (1990). Asymptotic tail behaviour of Poisson mixtures with applications. Ad-

vances in Applied Probability, 22, 147–159.

[59] Willmot G.E. (1993). On Recursive Evaluation of Mixed Poisson Probabilities and Related

Quantities, Scandinavian Actuarial Journal, 18, 114-133.

ACCEPTED MANUSCRIPT
26
ACCEPTED MANUSCRIPT

Table 1: Distribution of epileptic seizure counts

Observed Expected frequency

Count Frequency Po NB WGP GPL NGPL PT E
0 126 74.94 91 118.11 121.51 122 121.925
1 80 115.71 86.6 95.81 92 91 91.616
2 59 89.34 63.37 59.89 59 58.74 58.561
3 42 46 42.57 34.49 35.1 35.22 34.773
4 24 17.75 27.6 19.24 20.1 20.52 19.840
5 8 5.48 17.6 10.59 11.18 11.22 11.056
6 5 1.41 10.5 5.81 6.1 6.39 6.070
7 4 0.31 6.52 3.18 3.3 3.25 3.302
8 3 0.06 5 3.88 2.71 2.5 3.856
Total 351 351 351 351 351 351 351

ML λ̂ = 1.544 r̂ = 1.757 â = 1.089 θ̂=1.139 θ̂ = 1.116 α̂ = −0.701

Estimator p̂ = 0.463 b̂ = 0.295 α̂ = 1.292 α̂ = 2.906 θ̂ = 0.873
ŝ = −1
LL-value -636.05 -595.22 -595.83 -594.61 -594.48 -594.85
chi-square 256.54 22.53 7.12 5.94 5.75 5.36

ACCEPTED MANUSCRIPT
27
ACCEPTED MANUSCRIPT

Table 2: Summary Statistics of Response and Explanatory Variables

Variable Measurement Mean std dev

OFP Number of physician visit 5.774 6.759
HOSP Number of hospital stays 0.296 0.746
POORHLTH Self-perceived health status, 0.13 0.33
poor=1, else=0
EXCLHLTH Self-perceived health status, 0.08 0.27
excellent =1, else = 0
NUMCHRON Number of chronic conditions 1.54 1.35
MALE Gender; male = 1, else =0 0.404 0.491
SCHOOL Number of year of education 10.29 3.739
PRIVINS Private insurance indicator, yes =1, no = 0 0.776 0.417

ACCEPTED MANUSCRIPT
28
 Table 3: Maximum likelihood estimates their standard error, t-values and p-values of regression parameters
Po-Regression Model NB-Regression PT E-Regression Model

estimate s.e. t-value p-value estimate s.e. t-value p-value estimate s.e. t-value p-value
Intercept(β1 ) 1.0289 0.0238 43.2580 0.0000 0.9293 0.0546 17.0220 0.0000 –3.5740 0.0915 –39.0754 0.0000

HOSP(β2 ) 0.1648 0.0060 27.4780 0.0000 0.2178 0.0202 10.7930 0.0000 –0.0661 0.0298 –2.2142 0.0268

POORHLTH(β3 ) 0.2483 0.0178 13.9150 0.0000 0.3050 0.0485 6.2880 0.0000 1.0481 0.0705 14.8583 0.0000

EXCLHLTH(β4 ) –0.3620 0.0303 –11.9450 0.0000 –0.3418 0.0609 -5.6100 0.0000 –1.7797 0.5643 –3.1535 0.0016

NUMCHRON(β5 ) 0.1466 0.0046 32.0200 0.0000 0.1749 0.0121 14.4660 0.0000 0.0435 0.0139 3.1303 0.0017

MALE(β6 ) –0.1123 0.0129 –8.6770 0.0000 –0.1265 0.0312 -4.0520 0.0000 0.4583 0.0645 7.1032 0.0000

SCHOOL(β7 ) 0.0261 0.0018 14.1820 0.0000 0.0268 0.0044 6.1030 0.0000 0.1525 0.0112 13.5600 0.0000

PRIVINS(β8 ) 0.2017 0.0169 11.9630 0.0000 0.2244 0.0395 5.6860 0.0000 –0.1889 0.0333 –5.6730 0.0000

dispersion - - - - 1.2066 0.0336 35.5400 0.0000 0.0128 0.0002 67.0817 0.0000

29
ACCEPTED MANUSCRIPT

ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT

Table 4: log-likelihood (lmax ) and Akaike information criterion (AIC) Value

Regression Model
Criterion Pois NB PT E
lmax -17971.50 -12170.55 -10521.89
AIC 35959 24359 21061.78

ACCEPTED MANUSCRIPT
30
ACCEPTED MANUSCRIPT

Figure 1: PMF plot for different values of α and θ.

ACCEPTED MANUSCRIPT
31
ACCEPTED MANUSCRIPT

Figure 2: From left top-in clockwise direction we present contour plots of Mean, Variance, Kurtosis
and Skewness of PT E(α, θ)

ACCEPTED MANUSCRIPT
32