ADVANCED DISCRETE DISTRIBUTIONS
7.3 Mixed-Frequency Distributions
7.3.2 Mixed Poisson Distributions
If we letππ(π)in (7.13) have the Poisson distribution, this leads to a class of distributions with useful properties. A simple example of a Poisson mixture is the two-point mixture.
EXAMPLE 7.14
Suppose that drivers can be classified as βgood driversβ and βbad drivers,β each group with its own Poisson distribution. Determine the pf for this model and fit it to the data from Example 12.5. This model and its application to the data set are from TrΒ¨obliger [121].
From (7.13) the pf is
ππ=ππβπ1ππ1
π! + (1 βπ)πβπ2ππ2 π! .
The maximum likelihood estimates1were calculated by TrΒ¨obliger to be Μπ= 0.94, Μπ1= 0.11, and Μπ2 = 0.70. This means that about94%of drivers were βgoodβ with a risk ofπ1= 0.11expected accidents per year and6%were βbadβ with a risk ofπ2= 0.70 expected accidents per year. Note that it is not possible to return to the data set and
identify which were the bad drivers. β‘
This example illustrates two important points about finite mixtures. First, the model is probably oversimplified in the sense that risks (e.g. drivers) probably exhibit a continuum of risk levels rather than just two. The second point is that finite mixture models have a lot of parameters to be estimated. The simple two-point Poisson mixture has three parameters.
Increasing the number of distributions in the mixture toπwill then involveπβ 1mixing parameters in addition to the total number of parameters in theπcomponent distributions.
Consequently, continuous mixtures are frequently preferred.
The class of mixed Poisson distributions has some interesting properties that are developed here. Letπ(π§)be the pgf of a mixed Poisson distribution with arbitrary mixing
1Maximum likelihood estimation is discussed in Chapter 11.
distributionπ(π). Then (with formulas given for the continuous case), by introducing a scale parameterπ, we have
π(π§) =
β« πππ(π§β1)π’(π)ππ=
β«
[ππ(π§β1)]ππ’(π)ππ
=πΈ{[
ππ(π§β1)]π}
=πΞ[π(π§β 1)], (7.14) whereπΞ(π§)is the mgf of the mixing distribution.
Therefore,πβ²(π§) =ππΞβ²[π(π§β 1)]and withπ§= 1we obtain E(π) =πE(Ξ), where π has the mixed Poisson distribution. Also, πβ²β²(π§) = π2πΞβ²β²[π(π§β 1)], implying that E[π(πβ 1)] =π2E(Ξ2)and, therefore,
Var(π) =E[π(πβ 1)] +E(π) β [E(π)]2
=π2E(Ξ2) +E(π) βπ2[E(Ξ)]2
=π2Var(Ξ) +E(π)
>E(π),
and thus for mixed Poisson distributions the variance is always greater than the mean.
Most continuous distributions in this book involve a scale parameter. This means that scale changes to distributions do not cause a change in the form of the distribution, but only in the value of its scale parameter. For the mixed Poisson distribution, with pgf (7.14), any change inπis equivalent to a change in the scale parameter of the mixing distribution.
Hence, it may be convenient to simply setπ= 1where a mixing distribution with a scale parameter is used.
Douglas [29] proves that for any mixed Poisson distribution, the mixing distribution is unique. This means that two different mixing distributions cannot lead to the same mixed Poisson distribution and this allows us to identify the mixing distribution in some cases.
There is also an important connection between mixed Poisson distributions and compound Poisson distributions.
Definition 7.6 A distribution is said to be infinitely divisible if for all values of π = 1,2,3,β¦its characteristic functionπ(π§)can be written as
π(π§) = [ππ(π§)]π,
whereππ(π§)is the characteristic function of some random variable.
In other words, taking the (1βπ)th power of the characteristic function still results in a characteristic function. The characteristic function is defined as follows.
Definition 7.7 Thecharacteristic functionof a random variableπis ππ(π§) =E(πππ§π) =E(cosπ§π+πsinπ§π), whereπ=β
β1.
In Definition 7.6, βcharacteristic functionβ could have been replaced by βmoment generating functionβ or βprobability generating function,β or some other transform. That
is, if the definition is satisfied for one of these transforms, it will be satisfied for all others that exist for the particular random variable. We choose the characteristic function because it exists for all distributions, while the moment generating function does not exist for some distributions with heavy tails. Because many earlier results involved probability generating functions, it is useful to note the relationship between it and the characteristic function.
Theorem 7.8 If the probability generating function exists for a random variableπ, then ππ(π§) =π(βπlnπ§)andππ(π§) =π(πππ§).
Proof:
ππ(π§) =E(π§π) =E(ππlnπ§) =E[πβπ(πlnπ§)π] =ππ(βπlnπ§) and
ππ(π§) =E(πππ§π) =E[(πππ§)π] =ππ(πππ§). β‘ The following distributions, among others, are infinitely divisible: normal, gamma, Poisson, and negative binomial. The binomial distribution is not infinitely divisible because the exponentπin its pgf must take on integer values. Dividingπbyπ = 1,2,3,β¦will result in nonintegral values. In fact, no distributions with a finite range of support (the range over which positive probabilities exist) can be infinitely divisible. Now to the important result.
Theorem 7.9 Suppose thatπ(π§)is a mixed Poisson pgf with an infinitely divisible mixing distribution. Then,π(π§)is also a compound Poisson pgf and may be expressed as
π(π§) =ππ[π2(π§)β1],
whereπ2(π§)is a pgf. If we insist thatπ2(0) = 0, thenπ2(π§)is unique.
A proof can be found in Feller [37, Chapter 12]. If we choose any infinitely divisible mixing distribution, the corresponding mixed Poisson distribution can be equivalently described as a compound Poisson distribution. For some distributions, this is a distinct advantage when carrying out numerical work, because the recursive formula (7.5) can be used in evaluating the probabilities once the secondary distribution is identified. For most cases, this identification is easily carried out. A second advantage is that, because the same distribution can be motivated in two different ways, a specific explanation is not required in order to use it. Conversely, the fact that one of these models fits well does not imply that it is the result of mixing or compounding. For example, the fact that claims follow a negative binomial distribution does not necessarily imply that individuals have the Poisson distribution and the Poisson parameter has a gamma distribution.
To obtain further insight into these results, we remark that if a counting distribution with pgfπ(π§) =ββ
π=0πππ§πis known to be of compound Poisson form (or, equivalently, is an infinitely divisible pgf), then the quantitiesπandπ2(π§)in Theorem 7.9 may be expressed in terms ofπ(π§). Becauseπ2(0) = 0, it follows thatπ(0) =π0=πβπor, equivalently,
π= β lnπ(0). (7.15)
Thus, using (7.15),
π2(π§) = 1 +1
πlnπ(π§) = lnπ(0) β lnπ(π§)
lnπ(0) . (7.16)
The following examples illustrate the use of these ideas.
EXAMPLE 7.15
Use the preceding results and (7.14) to express the negative binomial distribution in both mixed Poisson and compound Poisson form.
The moment generating function of the gamma distribution with pdf denoted by π’(π)is (from Example 3.7 withπΌreplaced byπandπreplaced byπ½)
πΞ(π‘) = (1 βπ½π‘)βπ=
β«
β 0
ππ‘ππ’(π)ππ, π‘ <1βπ½.
This is clearly infinitely divisible because[
πΞ(π‘)]1βπis the mgf of another gamma distribution withπreplaced byπβπ. Thus, using (7.14) withπ= 1yields the negative binomial pgf
π(π§) =πΞ(π§β 1) =
β«
β 0
ππ(π§β1)π’(π)ππ= [1 βπ½(π§β 1)]βπ.
Because the gamma mixing distribution is infinitely divisible, Theorem 7.9 guar- antees that the negative binomial distribution is also of compound Poisson form, in agreement with Example 7.5. The identification of the Poisson parameterπand the secondary distribution in Example 7.5, although algebraically correct, does not provide as much insight as in the present discussion. In particular, from (7.15) we find directly that
π=πln(1 +π½) and, from (7.16),
π2(π§) = βπln(1 +π½) +πln[1 βπ½(π§β 1)]
βπln(1 +π½)
= ln
(1+π½βπ½π§ 1+π½
)
ln(1 +π½)β1
= ln
( 1 β π½
1+π½π§) ln
( 1 β π½
1+π½
) ,
the logarithmic series pdf as before. β‘
EXAMPLE 7.16
Show that a mixed Poisson with an inverse Gaussian mixing distribution is the same as a PoissonβETNB distribution withπ= β0.5.
The inverse Gaussian distribution is described in Appendix A. It has pdf π(π₯) =
( π 2ππ₯3
)1β2
exp [
β π 2π₯
(π₯βπ π
)2]
, π₯ >0,
and mgf
π(π‘) =
β«
β 0
ππ‘π₯π(π₯)ππ₯= exp [π
π (
1 β
β 1 β2π2
π π‘ )]
,
whereπ >0andπ >0are parameters. Note that [π(π‘)]1βπ= exp
[ π ππ
( 1 β
β 1 β2π2
π π‘ )]
= exp
β§βͺ
β¨βͺ
β© πβπ2
πβπ
β‘β’
β’β£ 1 β
β
1 β2(πβπ)2 (πβπ2) π‘β€
β₯β₯
β¦
β«βͺ
β¬βͺ
β .
This is the mgf of an inverse Gaussian distribution withπreplaced byπβπ2andπby πβπ, and thus the inverse Gaussian distribution is infinitely divisible.
Hence, by Theorem 7.9, the Poisson mixed over the inverse Gaussian distribution is also compound Poisson. Its pgf is then, from (7.14) withπ= 1,
π(π§) =π(π§β 1) = exp {π
π [
1 β
β 1 β2π2
π (π§β 1) ]}
,
which may be represented, using (7.15) and (7.16) in the compound Poisson form of Theorem 7.9 with
π= β lnπ(0) = π π
(β
1 +2π2 π β 1
) ,
and
π2(π§) =
π π
( 1 β
β 1 + 2π2
π
) + π
π
[β
1 β2π2
π (π§β 1) β 1 ]
ππ
( 1 β
β 1 +2ππ2
)
=
β 1 β2π2
π (π§β 1) β
β 1 +2π2
π
1 β
β 1 +2π2
π
.
We recognize thatπ2(π§)is the pgf of an extended truncated negative binomial distribution withπ= β1β2andπ½ = 2π2βπ. Unlike the negative binomial distribution, which is itself a member of the(π, π,0)class, the compound Poisson representation is of more use for computational purposes than the original mixed Poisson formulation.β‘ It is not difficult to see that, ifπ’(π)is the pf for any discrete random variable with pgfπΞ(π§), then the pgf of the mixed Poisson distribution isπΞ
[expπ(π§β1)]
, a compound distribution with a Poisson secondary distribution.
Table 7.2 Pairs of compound and mixed Poisson distributions.
Compound secondary Mixing
Name distribution distribution
Negative binomial Logarithmic Gamma
NeymanβType A Poisson Poisson
Poissonβinverse Gaussian ETNB (π= β0.5) Inverse Gaussian
EXAMPLE 7.17
Demonstrate that the Neyman Type A distribution can be obtained by mixing.
If in (7.14) the mixing distribution has pgf πΞ(π§) =ππ(π§β1), then the mixed Poisson distribution has pgf
π(π§) = exp{π[ππ(π§β1)β 1]},
the pgf of a compound Poisson with a Poisson secondary distribution, that is, the
Neyman Type A distribution. β‘
A further interesting result obtained by Holgate [57] is that, if a mixing distribution is absolutely continuous and unimodal, then the resulting mixed Poisson distribution is also unimodal. Multimodality can occur when discrete mixing functions are used. For example, the Neyman Type A distribution can have more than one mode. You should try this calculation for various combinations of the two parameters. The relationships between mixed and compound Poisson distributions are given in Table 7.2.
In this chapter, we focus on distributions that are easily handled computationally.
Although many other discrete distributions are available, we believe that those discussed form a sufficiently rich class for most problems.