Abstract
Two families of bivariate discrete Poisson–Lindley distributions are introduced. The first is derived by mixing the common parameter in a bivariate Poisson distribution by different models of univariate continuous Lindley distributions. The second is obtained by generalizing a bivariate binomial distribution with respect to its exponent when it follows any of five different univariate discrete Poisson–Lindley distributions with one or two parameters. The use of probability-generating functions is mainly employed to derive some general properties for both families and specific characteristics for each one of their members. We obtain expressions for probabilities, moments, conditional distributions, regression functions, as well as characterizations for certain bivariate models and their marginals. An attractive property of all bivariate individual models is that they contain only two or three parameters, and one of them is readily estimated by simple ratios of their sample means. This feature, and since all marginal distributions are over-dispersed, strongly suggests their potential use to describe bivariate dependent count data in many different areas.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Availability of count data, the number of occurrences of an event within a fixed period of time, is rapidly increasing in all areas of human activity. Modeling these data becomes an important issue in medicine, biology, ecology, economics, demography and other sciences.
In many real-life problems, the count represents two dependent variables. For example, in medical research we count coronavirus cases treated in a hospital and the number of deaths recorded among them. In actuarial science, we count the number of traffic accidents and the corresponding number of deaths or, in insurance claims, the count variable often represents damage and bodily injury.
Consequently, for the statistical analysis of such data, the use of bivariate discrete models may be appropriate.
Over several years, a large number of univariate discrete distributions were extended to the bivariate (multivariate) case and basic references are the books by Kocherlakota and Kocherlakota [1], Johnson, Kotz and Balakrishnan [2] and the review paper by Lai [3].
The one-parameter univariate Poisson–Lindley distribution was introduced by Sankaran [4]. Since it is a Poisson mixture constructed by mixing the Poisson parameter with a continuous Lindley model, this distribution is over-dispersed. Such a property is often useful in describing count data. In addition, since it has only one parameter, it has attracted the attention of several researchers. In the last decade, several univariate Poisson–Lindley distributions with two or more parameters were introduced by [5,6,7,8,9], among others. These distributions are Poisson mixtures derived by assuming that the mixing variable follows a continuous Lindley distribution with two or more parameters.
Bivariate (multivariate) extensions of the one-parameter Poisson–Lindley model were considered by Gómez-Déniz et al. [10], and their usefulness was demonstrated for bivariate count data sets. Their model was revisited by [11].
The main purpose of this paper is to introduce and study two families of bivariate Poisson–Lindley distributions, each with five members, by extending the univariate Poisson–Lindley models examined by [4,5,6,7,8] to the bivariate case. For this purpose, we adopted two widely used procedures, namely the mixing and the generalizing approach. The individual bivariate models examined can be useful in describing bivariate count data, since their basic characteristic is that the number of their parameters is limited to two or three. Furthermore, this assumption is further enhanced because one of the parameters is readily estimated by simple ratios of the sample means. In addition, in all marginal distributions the index of dispersion is greater than one. General properties of each class are derived mainly by employing the use of probability-generating functions (p.g.f.’s), and these are customized for each of their members.
The rest of the paper is organized as follows. In Sect. 2, we document certain properties of various univariate Poisson–Lindley distributions appeared in the literature. Additional characteristics are derived, including characterizations as special cases of a result due to Cacoullos and Papageorgiou [12]. In Sect. 3, bivariate Poisson–Lindley mixtures are derived by assuming that the common parameter in a bivariate Poisson distribution follows a univariate Lindley distribution with one or two parameters. This procedure was used by [4,5,6,7,8] in the univariate case to derive their corresponding Poisson–Lindley models. Furthermore, implementing a characterization theorem proved by Cacoullos and Papageorgiou [13] relative characterizations are given for the bivariate (X, Y) distribution. In Sect. 4, we generalize a bivariate binomial distribution with respect to its exponent, when it follows one of the five univariate Poisson–Lindley distributions derived by [4,5,6,7,8]. Finally, Sect. 5 concludes.
2 Univariate Poisson Mixtures and Poisson–Lindley Distributions
Let the parameter \(\lambda \) of a Poisson distribution be a continuous random variable (r.v.) with probability density function (d.f.) \(F'(\lambda )=f(\lambda )\) and moment-generating function (m.g.f.) \(M_\varLambda (\cdot )\). Then, if X is a nonnegative integer-valued r.v. with probability mass function (p.f.) \(p(x)=P(X=x)\), a Poisson mixture is defined by
Consequently, the p.g.f. of the r.v. X is
General properties of Poisson mixtures model were studied by Karlis and Xekalaki [14]. In particular, they pointed out that the index of dispersion, that is, the variance-to-the mean ratio for a mixed Poisson distribution is always greater than one. Consequently, Poisson mixture models are over-dispersed. A simple characterization of Poisson mixtures was indicated by Cacoullos and Papageorgiou [12], see also [15], using the one-parameter Poisson–Lindley distribution as an illustrative example. Their result is stated here as in Theorem 1, and a detailed proof is provided in the Appendix.
Theorem 1
Let X be a Poisson mixture defined by Eq. (1) and \(\varLambda >0\) a continuous r.v. with density function \(F'(\lambda )=f(\lambda )\). Then, the regression function of \(\varLambda \) on X
determines uniquely both the distributions of \(\varLambda \) and X.
Poisson–Lindley distributions are obtained by allowing the parameter \(\lambda \) of a Poisson distribution to follow a Lindley distribution with one or more parameters.
In this section, we document some basic properties of various univariate Poisson–Lindley distributions already introduced in the literature. Additional useful characteristics are also derived.
2.1 One-Parameter Poisson–Lindley Distribution
This distribution was introduced by Sankaran [4] by mixing (compounding) the Poisson parameter using a distribution given by Lindley [16] with m.g.f.
Then, the p.g.f. of the corresponding Poisson–Lindley r.v. X is
Since
we immediately derive the p.f.
and the factorial moments
where
2.2 A Two-Parameter Poisson–Lindley Distribution
A two-parameter Lindley distribution was introduced by [17] with m.g.f.
By assuming that the Poisson parameter \(\lambda \) in Eq. (2) follows a distribution with m.g.f. given by Eq. (9), Shanker and Mishra [5] obtained a two-parameter Poisson–Lindley distribution with p.g.f.
Since
we derive
and a simple relation for the factorial moments
Finally, from Theorem 1, the two-parameter Lindley and the two-parameter Poisson–Lindley distributions are determined uniquely, if
2.3 A New Generalized Poisson–Lindley Distribution
A slightly different two-parameter Lindley distribution from the one introduced by [17] was suggested by [18] with m.g.f.
Based on (14), a new generalized Poisson–Lindley distribution was proposed by Bhati et al. [6] with p.g.f.
Since
we have
and
Finally, from Theorem 1
2.4 A Generalized Poisson–Lindley Distribution
A generalized Lindley distribution with m.g.f.
was obtained by [19].
A generalized Poisson–Lindley distribution utilizing Eq. (19) was introduced by Mahmoudi and Zakerzadeh [7] with p.g.f.
From
the following properties are derived:
and
Another illustration of Theorem 1 is the following characterization. If
then the distributions of the generalized Lindley and the generalized Poisson–Lindley distributions are uniquely determined.
2.5 A New Two-Parameter Poisson-Generalized Lindley Distribution
This distribution was recently introduced by Altun [8] and is a Poisson mixture when the Poisson parameter \(\lambda \) follows the new generalized Lindley distribution studied by [20] with m.g.f.
Some properties of this distribution are:
2.6 Another Two-Parameter Poisson–Lindley Distribution
All previous models were based on the assumption that in a Poisson mixture model the Poisson parameter \(\lambda \) varies according to a Lindley distribution.
Let us now suppose that in a Poisson mixture the Poisson parameter is of the form \(\varphi \lambda \), where \(\varphi \) is a positive constant and \(\lambda \) is a continuous r.v. with m.g.f. \(M_\varLambda (\cdot )\).
Then, if Y is a nonnegative integer-valued r.v. with p.g.f. \(G_Y(t)\) the corresponding Poisson mixture model is defined as
A two-parameter Poisson–Lindley distribution was derived by [10], assuming that the m.g.f. of the r.v. \(\varLambda \) in Eq. (29) is given by expression (4) corresponding to the m.g.f. of the one-parameter Lindley distribution. The p.g.f. of this distribution is
and p.f.
In addition, they obtained
3 Mixed (Compounded) Bivariate Poisson Distributions
A general class of compounded bivariate Poisson distributions was extensively studied by [1], chapter 8] and Kocherlakota [21]. They considered the class of distributions (X, Y) with p.g.f.
where \(\varphi _1,\varphi _2,\varphi _{12}\) are constants and \(\lambda \) is a r.v. with m.g.f. \(M_\varLambda (\cdot )\). They proved that
and this representation enabled them to derive various general properties.
3.1 A Bivariate Poisson–Lindley Distribution
Let us now consider a simpler form of Eq. (33), that has p.g.f.
Gómez-Déniz et al. [10], by assuming that \(\lambda \) follows the one-parameter Poisson–Lindley distribution with m.g.f. given by Eq. (4), derived a bivariate Poisson–Lindley distribution with p.g.f.
The marginals are two-parameter Poisson–Lindley distributions with p.g.f.’s of the form given by expression (30).
Not only a detailed study of this bivariate distribution was presented by [10], but also they considered multivariate extensions.
Remark
Bivariate Poisson–Lindley distributions can also be derived by using an approach suggested by David and Papageorgiou [22]. They examined the general class of distributions with p.g.f.
where \(\varphi _1\) and \(\varphi _2\) are constants and \((\lambda _1,\lambda _2)\) r.v.’s of the discrete or the continuous type, with m.g.f. \(M_{\varLambda _1,\varLambda _2}(\cdot ,\cdot )\). Then, since
if \((\lambda _1,\lambda _2)\) follows a bivariate Lindley distribution, the corresponding bivariate Poisson–Lindley models can be constructed.
3.2 A Family of Mixed Bivariate Poisson–Lindley Distributions
A bivariate discrete model with the structure
where X follows a Poisson distribution with parameter \(\lambda \) and the \(Y_i\)’s are i.i.d. Bernoulli r.v.’s with parameter p were studied by Leiter and Hamdan [23] and Cacoullos and Papageorgiou [24] to express the joint distribution of the number of accidents and the number of fatal accidents.
Its p.g.f. is given by the relation
For other bivariate models with the structure given by relation (35), see, among others, [25,26,27]. By assuming that \(\lambda \) follows a distribution with m.g.f. \(M_\varLambda (\cdot )\), Eq. (36) becomes
where \(q=1-p\). Then,
which is Eq. (2) and
which is Eq. (29) with the parameter \(\varphi \) replaced by p.
Some general properties of this class of distributions can be easily derived. In particular, since
and
we have
To derive the conditional p.g.f. \(G_{Y|X=x}(z)\) of the r.v. Y given \(X=x\), we use the following result due to Subrahmaniam [28]:
For a bivariate discrete r.v. (X, Y) with p.g.f. \(G_{X,Y}(s,t)\), the conditional p.g.f. \(G_{Y|X=x}(z)\) of Y on X is
where
Hence, from Eqs. (37) and (41), we obtain
This result facilitates the calculation of the joint p.f. of X and Y, as
and, additionally, a characterization of the joint distribution of (X, Y) can be obtained by using the following theorem derived by Cacoullos and Papageorgiou [13].
Theorem 2
For a bivariate discrete r.v. (X, Y), let
and
Then, \(P(Y=y|X=x)\) and \(E[X|Y=y]\) together determine the distribution of (X, Y).
Furthermore, from Eq. (39) the parameter p can be immediately estimated by the ratio of the two marginal means, i.e.,
This property facilitates the applicability of this class of distributions, since the remaining parameters can be estimated by procedures suggested for univariate Poisson–Lindley models.
3.3 Examples
3.3.1 Bivariate Poisson–Lindley Distribution Defined by Relations (37) and (4)
The p.g.f. of this distribution is
with marginals
a univariate one-parameter Poisson–Lindley distribution discussed in Sect. 2.1, and
This distribution has p.f.
which is Eq. (31) with the parameter \(\varphi \) replaced by p.
Simple recurrences for probabilities can be obtained by using the ratios
In particular,
and
with
independent of the parameter p.
From expressions (40) and (8),
The conditional p.g.f. of \(G_{Y|X=x}(z)\) is given by Eq. (42).
Applying to Eq. (46) the general formula of [28] for the derivation of conditional p.g.f.’s as indicated by Eq. (41), \(G_{X|Y=y}(z)\) can be expressed as
where
i.e., it is a shifted Poisson–Lindley-type distribution.
From Eq. (49), we can derive the conditional expectation of X given \(Y=y\) as
The conditional expectation of X given \(Y=y\) can also be obtained from expression (45) given in Theorem 2 and Eq. (48).
Consequently, from Theorem 2, relations (44) and (51) characterize the joint distribution of (X, Y) with p.g.f. given by Eq. (46).
3.3.2 Bivariate Poisson–Lindley Distribution Defined by Relations (37) and (9)
The p.g.f. of this distribution is
with marginals
which is a two-parameter Poisson–Lindley distribution discussed in Sect. 2.2, and
The p.f. of this distribution is
and its moments can be derived from Eq. (38).
Furthermore, from expressions (43) and (12)
\(x=0,1,2,\ldots \), \(y=0,1,2,\ldots ,x\).
Also, from Eqs. (40) and (13) an expression for Cov(X, Y) is obtained.
Finally, from Eqs. (45) and (52)
Consequently, from Theorem 2 a characterization of (X, Y) can be obtained.
3.3.3 Bivariate Poisson–Lindley Distributions Defined by Relations (37) and (14)
This distribution has p.g.f. given by
with marginals
which is a new generalized Poisson–Lindley distribution examined in Sect. 2.3, and
The p.f. of r.v. Y is
and its moments can be obtained from Eq. (38). An expression for Cov(X, Y) is derived from Eqs. (40) and (18).
Utilizing expressions (43) and (17), we obtain
\(x=0,1,2,\ldots \), \(y=0,1,2,\ldots ,x\).
Finally,
3.3.4 Bivariate Poisson–Lindley Distributions Defined by Relations (37) and (19)
The p.g.f. of this distribution is
with marginals
given by Eq. (20), and
The p.f. of the r.v. Y is
In addition,
3.3.5 Bivariate Poisson–Lindley Distributions Defined by Relations (37) and (24)
Some basic characteristics of the distribution are
It should be noted that, as expected, for \(\alpha =1\) all relations in Sects. 3.3.2–3.3.4 become their corresponding relations in Sect. 3.3.1. This result also holds for the relations in Sect. 3.3.5 when \(\alpha =2\).
4 Generalized Bivariate Binomial Models
Generalized (or countable mixtures of) bivariate binomial models with respect to their index parameter(s) were studied by Papageorgiou and David [29], and illustrative examples were given.
A bivariate binomial distribution with p.g.f.
where N is a nonnegative integer-valued r.v. with p.g.f.
was introduced by Rao et al. [30] in their effort to study the correlation between the numbers of two types of children X and Y in a family where N is the family size (sibship size).
Consequently, the joint distribution of X and Y is given by the p.g.f.
Applications to actual set(s) of family size data were given by [29] and [30] when N follows a negative binomial or Neyman type A distributions. In addition, when N follows a “Short” distribution a corresponding bivariate model was fitted to accident data by [31].
4.1 Properties
For distributions with p.g.f. given by Eq. (53), we can obtain various properties of the marginal distributions of X and Y and the joint distribution of (X, Y), in terms of the corresponding properties of the distribution of the r.v. N.
In particular, since
we have
and
The joint p.f. is
and an expression for the factorial moments is
where
Using Eq. (41), we can prove that the conditional p.g.f. of Y given \(X=x\) is
Consequently, the conditional probability function of Y given \(X=x\) is
and the related conditional factorial moments are
Hence,
The corresponding expressions for
can be easily obtained.
4.2 Generalized Bivariate Poisson–Lindley Distributions
From Eq. (55) (see also [30]), we have
Consequently, the index of dispersion for the r.v. X denoted by \(D_X\) is
which is greater than one since in this section the r.v. N follows a Poisson–Lindley distribution. A similar property also holds for \(D_Y\).
In addition, an estimator of the parameter p can be easily obtained by using a simple ratio of the marginal means. That is
and the remaining parameters in the bivariate Poisson–Lindley models can be estimated by using procedures already employed in their univariate versions.
4.3 Examples
4.3.1 Bivariate Poisson–Lindley Distributions Defined by Relations (53) and (5)
The p.g.f. of this distribution is
Notice that Eq. (59) corresponds to Eq. (34) for \(\varphi _1=q\) and \(\varphi _2=p\). The p.g.f. of the X marginal is
while the p.g.f. of the Y marginal is given by Eq. (47).
Hence,
Simple recurrences for the probabilities are
with
independent of p. From Eqs. (56) and (6),
where \(A_x(\theta ,pz)\) can be obtained from Eq. (50).
Finally, from Eqs. (58) and (6)
4.3.2 Bivariate Poisson–Lindley Distributions Defined by Relations (53) and (10)
The p.g.f. of this distribution is
Some other properties of this distribution are
4.3.3 Bivariate Poisson–Lindley Distributions Defined by Relations (53) and (15)
This model has p.g.f. given by
Other basic properties are
4.3.4 Bivariate Poisson–Lindley Distributions Defined by Relations (53) and (20)
The p.g.f. of this distribution is
The p.g.f. of X is
Some other properties are
Expressions for \(P(X=x,Y=y)\) can be obtained from Eqs. (54) and (22) and for \(P(Y=y|X=x)\) from Eqs. (57), (22) and (21).
Finally,
4.3.5 Bivariate Poisson–Lindley Distributions Defined by Relations (53) and (25)
The p.g.f. of this distribution is
Some characteristic properties of this distribution are
Finally,
5 Conclusions
In this paper, two families of bivariate Poisson–Lindley distributions are introduced either by mixing or by generalizing. Each family extends to the bivariate case five univariate Poisson–Lindley models already appeared in the literature. We examined a number of characteristics both for the families and for their individual members. We also indicated that all bivariate models can be useful in analyzing count data because they contain only two or three parameters. The models derived by the generalization procedure also have the attractive property that \(Z=X+Y\) follows the same distribution with the generalizing variable a univariate Poisson–Lindley. As Kemp and Papageorgiou [32] pointed out “this property is often required for consistency in accident models where the split into two time periods is entirely arbitrary.” Obviously, more complicated bivariate Poisson–Lindley models can be derived, but their use may be restricted because of their increased number of parameters.
References
Kocherlakota S, Kocherlakota K (1992) Bivariate discrete distributions. Marcel Dekker, New York
Johnson NL, Kotz S, Balakrishnan N (1997) Discrete multivariate distributions. John Wiley and Sons, New York
Lai CD (2006) Constructions of discrete bivariate distributions. In: Balakrishnan N, Castillo E, Sarabia JM (eds) Advances on distribution theory, order statistics and inference. Birkhäuser, Boston, pp 29–58
Sankaran M (1970) The discrete Poisson-Lindley distribution. Biometrics 26:145–149
Shanker R, Mishra A (2014) A two-parameter Poisson-Lindley distribution. Int J Stat Syst 9(1):79–85
Bhati D, Sastry DVS, Qadri PZM (2015) A new generalized Poisson-Lindley distribution: applications and properties. Austrian J Stat 44:35–51
Mahmoudi E, Zakerzadeh H (2010) Generalized Poisson-Lindley distribution. Commun Stat Theory Methods 39(10):1785–1798
Altun E (2021) A new two-parameter discrete Poisson generalized Lindley distribution with properties and applications to health care data sets. Comput Stat 36(4):2841–2861
Das KK, Ahmed I, Bhattacharjee S (2018) A new three-parameter Poisson-Lindley distribution for modelling over-dispersed count data. Int J Appl Eng Res 13(23):16468–16477
Gómez-Déniz E, Sarabia JM, Balakrishnan N (2012) A multivariate discrete Poisson-Lindley distribution: Extensions and actuarial applications. Astin Bull 42(2):655–678
Zamani H, Faroughi P, Noriszura I (2015) Bivariate Poisson-Lindley distribution with application. J Math Stat 11(1):1–6
Cacoullos T, Papageorgiou H (1982) Characterizing the negative binomial distribution. J Appl Probab 19(3):742–743
Cacoullos T, Papageorgiou H (1983) Characterizations of discrete distributions by a conditional distribution and a regression function. Ann Inst Stat Math 35(1):95–103
Karlis D, Xekalaki E (2005) Mixed Poisson distributions. Int Stat Rev 73:35–58
Cacoullos T, Papageorgiou H (1984) Characterizations of mixtures of continuous distributions by their posterior means. Scandinavian Actuarial J 1984(1):23–30
Lindley DV (1958) Fiducial distributions and Bayes’s theorem. J Roy Stat Soc B 20:102–107
Shanker R, Mishra A (2013) A two-parameter Lindley distribution. Stat Trans New Ser 14(1):45–56
Shanker R, Sharma S, Shanker R (2013) A two-parameter Lindley distribution for modeling waiting and survival times data. Appl Math 4(2):363–368
Zakerzadeh H, Dolati A (2009) Generalized Lindley distribution. J Math Ext 3(2):13–25
Ekhosuehi N, Opone F, Odobaire F (2018) A new generalized two parameter Lindley distribution. J Data Sci 16(3):549–566
Kocherlakota S (1988) On the compounded bivariate Poisson distribution: A unified approach. Ann Inst Stat Math 40(1):61–76
David KM, Papageorgiou H (1994) On compounded bivariate Poisson distributions. Nav Res Logist 41(2):203–214
Leiter RE, Hamdan MA (1973) Some bivariate probability models applicable to traffic accidents and fatalities. Int Stat Rev 41:87–100
Cacoullos T, Papageorgiou H (1980) On some bivariate probability models applicable to traffic accidents and fatalities. Int Stat Rev 48:345–356
Cacoullos T, Papageorgiou H (1981) On bivariate discrete distributions generated by compounding. In: Taillie C, Patil GP, Baldessari BA (eds) Statistical distributions in scientific work, vol 4. Reidel, Dordrecht, pp 197–212
Cacoullos T, Papageorgiou H (1982) Bivariate negative binomial – Poisson and negative binomial – Bernoulli models with an application to accident data. In: Kalianpur G, Krishnaiah PR, Ghosh JK (Eds.). Statistics and probability: essays in honor of C. R. Rao, pp. 155–168. North-Holland Publishing Company, Amsterdam
Papageorgiou H (1985) On a bivariate Poisson-geometric distribution. Zastosowania Matematyki 18(4):541–547
Subrahmaniam K (1966) A test of “Intrinsic correlation’’ in the theory of accident proneness. J Royal Stat Soc Ser B 28:180–189
Papageorgiou H, David KM (1994) On countable mixtures of bivariate binomial distributions. Biom J 36(5):581–601
Rao BR, Mazumdar S, Waller JH, Li CC (1973) Correlation between the numbers of two types of children in a family. Biometrics 29:271–279
Papageorgiou H, Piperigou VE (1997) On bivariate “Short” and related distributions. In: Johnson NL, Balakrishnan N (eds) Advances in the theory and practice of statistics: A, vol in. Honor of Samuel Kotz. New York, John Wiley and Sons, pp 397–413
Kemp CD, Papageorgiou H (1982) Bivariate Hermite distributions. Sankyā Ser A 44(2):269–280
Teicher H (1961) Identifiability of mixtures. Ann Math Stat 32:244–248
Golberg S (1958) Introduction to difference equations. Wiley, New York
Acknowledgements
The authors would like to thank both referees for their comments and suggestions.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The article is dedicated to the memory of Professor T. Cacoullos.
Appendix
Appendix
Proof of Theorem 1
It is well known, Teicher [33], that a Poisson mixture as given by (1) identifies F, i.e., if
for \(x=0,1,2,\ldots \), then \(F_1=F_2\). Hence, it is sufficient to determine only the distribution of \(\varLambda \).
Denoting by
we have
Hence,
Since this is a linear first-order difference equation in p(x), a unique solution exists (see, for example, Goldberg [34], p. 61) and the assertion follows.
In fact, the explicit solution of (60) is
where p(0) is determined from the condition \(\sum \limits _xp(x)=1\). \(\square \)
Rights and permissions
About this article
Cite this article
Papageorgiou, H., Vardaki, M. Bivariate Discrete Poisson–Lindley Distributions. J Stat Theory Pract 16, 30 (2022). https://doi.org/10.1007/s42519-022-00261-z
Accepted:
Published:
DOI: https://doi.org/10.1007/s42519-022-00261-z