This article was downloaded by: [North Carolina State University]

On: 06 May 2013, At: 00:07

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK

Communications in
Statistics - Theory and
Methods
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/lsta20

On a bivariate poisson
distribution
a a
J. Lakshminarayana , S.N.N. Pandit & K.
b
Srinivasa Rao
a
Department of Statistics, Andhra University,
Visakhapatnam, A.P, 530 003, India
b
Department of Mathematics and Statistics,
Central University, Tejpur, Assam
Published online: 27 Jun 2007.

To cite this article: J. Lakshminarayana , S.N.N. Pandit & K. Srinivasa Rao

(1999): On a bivariate poisson distribution, Communications in Statistics -
Theory and Methods, 28:2, 267-276

To link to this article: http://dx.doi.org/10.1080/03610929908832297

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/

terms-and-conditions

This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make
any representation that the contents will be complete or accurate or
up to date. The accuracy of any instructions, formulae, and drug doses
should be independently verified with primary sources. The publisher
shall not be liable for any loss, actions, claims, proceedings, demand,
or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.
 COMMUN. STATIST.-THEORY METH , 28(2), 267-276 (1999)

ON A BIVARIATE POISSON DISTRIBUTION

J. Lakshminarayana*, S.N.N. Pandit** and K. Srinivasa Rao*

*n- -2 *.@A
Eiii~~i~
nl ~...d..'A~ . r ; w ,
u r pA~---Au m o Vi vi, & ~ ~- 530 m W3, A?., India
**Department of Mathematics and Statistics, Central University, Tejpur, Assam
Downloaded by [North Carolina State University] at 00:07 06 May 2013

Keywords : marginal distributions; negative correlation; multiplication factor

ABSTRACT

A bivariate distribution, whose marginals are Poisson is developed as a

product of Poisson marginals with a multiplicative factor. The correlation bemeen
the two variates can be either positive or negative, depending on the value chosen
for the parameter in the above multiplicative factor. The distributional properties
of this distribution are studied and this model is fitted to a bivariate discrete
distribution with negative correlation coefficient. In this context, it should be
worth noting that the bivariate Poisson distribution reported by Teicher (1954),
(Campbell (1934), Holgate (1964)) has the inherent limitation that the correlation
is necessarily positive, and hence is not usefbl in modelling the situations (e.g. in
biology) invohing competition for limited resources where a negative correlation is
expected.

Copyright Q 1999 by Marcel Dekker. Inc.

 LAKSHMINARAYANA, PANDIT, AND SRINIVASA RAO

Bivariate Poisson distniution plays a vital role in analysing many real life
situations. For example, in developing and analysing the bivariate population
growth models, interdependent computer communication systems, reliabrlity
analysis, etc., where the processes under study are correlated and follow a bivariate
Poisson distriiution. The bivariate Poisson distribution available m literature is
mainly due to Campbell (1934). Some work has been reported m the literature on
Multivariate Poisson distribution (cf. Pate1 and Bildikar, (1967), Banarjee, (1959).
Downloaded by [North Carolina State University] at 00:07 06 May 2013

Sibuya et al., (1964)). But mainly they deal with the case of 'independent
multivariate' distributions and hence, essentially reduce to consideration of
univariate Poisson distribution (cf Johnson and Kotz, (1969)). The only works
that consider nontrivial cases of multivariate Poisson distributions appear to be the
work of Holgate (1964) and Teicher (1954), based essentially on the work of
Campbell (1934). Even here, the model developed is itself such that the two
Poisson variables have to be only positively correlated. However, in biological
growth models, the variables under study wiU often have a negative correlation
while the marginal distributions will follow a Poisson distribution. Hence, there is
a need for developing a bivariate distribution with Poisson marginais and aliowing
negative correlation. In this paper we develop one such bivariate Poisson
distribution and study its distributional properties.

2. BlVARLATE POISSON DISTRIBUTION

Following heuristic arguments analogous to those in the context of Gram-

Charlier series expansion of near normal or near Poisson unit~ariatedistributions,
one can obtain the joint probability mass fhction of two variates fiom their
marginal probability h c t i o n s along with a suitable fhction containing a
dependency parameter a.

Let P (u, v) be the joint probability mass fimction of (u, v) having the
rnarginals PI (u) and Pz(v) respectively. Then it is easy to ver& that
 (where g,(u) g, (v) is a bounded fimction of (u, v) E R ~ is) a probability mass
function with Pl(u) and Pz(v) as marginals, for suitably chosen range of values of
the parameter a (Panditi et al., (1988)).

In particular, let Pl(x) and P2(y) be two Poisson probability mass fimctions
with parameters hl and h2 respecti-dy. hi g,(x) iiiiiiid g2(j..) be two ,bctircs
Downloaded by [North Carolina State University] at 00:07 06 May 2013

which are bounded in x and y respectively, with expectations g, and g,. Then,

has Poisson marginals for values of a such that

Any polynomials in x and y as g, and g2 are ruled out since the deviations from

themean i.e.I(g,(~)-~,)/and/(g~(y)-i,)~~~~~~inde~t

It is obvious that gl(u) and g4v) can be any arbitrary 'bounded' functions
of u and v, (with, of course, finite mean). Since polynomials as such are to be
ruled out: an obvious choice is 'damped polynomials', viz. polynomials Pk(x)
multiplied by a dampening h c t i o n W(x), such that g(x) = W(x) PL(x). An
obvious choice for the dampening function is e-' , since it tends to zero faster than
any Pk(x) tends to infinity as x increases indefinitely; thus g(x) = e-" Pk(x)w i l l be a
h c t i o n which tends to zero as x tends to infinity and hence one can choose a
nontrirvial range for a, over which the joint p.mE will be non-negative over all
allowed x, y. Of course, the more general the polynomial is, the more is the
number of parameters required for defining the joint p.mE and some simplification
is achievable by taking the polynomial Pk(x) as a linear combination of different
 270 LAKSHMINARAYANA, PANDIT, AND SRINIVASA RAO

fictorial powers of x : Pdx) = Ea,,x(') and one can work out the different
r

moments and other distributional properties in this situation.

However, though theoretically possl%le: one can expect that simple

polynomials, of degree not more than, say 2, should lead to a good fit to data fiom
situations where assumptions leading to the Poisson probabilitiies are only mildly
vidated. Hence, when co&onted :'i& data cf+&issox?, it is perhaps advkable to
start with simplest polynomials as Pk(x) and go for higher degree polynomials only
Downloaded by [North Carolina State University] at 00:07 06 May 2013

ifthe data so warrant. In fad, as is clear fiom the illustrative example given in the
sequel, even the case of k=0, viz. the trivial 'constant' polynomial viz.P,(x) = 1
itseKcan lead to a very good fit.

Hence,

,
Since (g (u) - ,) (g, (u) - g,) is finite for all x, y one can choose an n

such that P(x,y) will be non-negative for all x, y 2 0, i.e. a probability mass
bction.

The marginal probability mass functions of x and y are Poisson with

parameters hl and hz respectively.

n e mean vector E [r]I:[

=
 BNARIATE POISSON DISTRIBUTION

one gets the dispersion matrix of (x, y) as

Hence the correlation coefficient turns out to be

p = amC2e-(Al+;iz)c
Downloaded by [North Carolina State University] at 00:07 06 May 2013

Range of a : It can be shown that a should lie in the range

Range of correlation coefficient : In this distribution the correlation coefficient is

having the range

In magnitude the correlation is less than a. It is interesting to note that unlike in

the Campbell (1934) model the present model allows the correlation coefficient
between x and y to be negative or positive by a proper choice of the parameter a.
The moment generating h c t i o n of (x, y) is

The moment estimators of the parameters are

 LAKSHMINARAYANA, PANDIT, AND SRINIVASA RAO

1 " -
where c(x, y) = -- C ( x i - X)(Y - i)
n i-I

Variances of these estimators are

Downloaded by [North Carolina State University] at 00:07 06 May 2013

where p, is the second order joint central moment.

3. ILLUSTRATION

As an illustration of an application of the present model let us consider the

data given in Table - !, giving the number of seeds and plants gr~w
over a plot of
size five square feet (Subba Rao (1990)).

X2 = 18.62 with 16 d.c under the assumption of independence.

The Table- i also gives the expected frequencies of merent (x, y) values
under the assumptions of independence of x and y as the second entries in each
cell, As is obvious from this Table, the cells for x = 0, 2, 3 with y = 0 and for
y = 2, 3 and 4 for x = 0, show expected frequencies quite different from the
corresponding observed Erequencies.

It is true that the overall xZ value 18.62 is not significant even at 25% level
of significance, with 16 d.f after pooling the frequencies which are rather small
 BIVARIATE POISSON DISTRIBUTION 273

Table - 1

0 7* 41* 54* 40* 21*

21.14 45.30 47.61 33.22 17.41
10.64 44.86 53.76 38.16 19.57
1 36 79 83 59 30
36.88 79.02 83.05 57.95 30.37
39.61 81.09 82.04 55.10 27.71
Downloaded by [North Carolina State University] at 00:07 06 May 2013

2 39* 70 69 47 25
3 1.96 68.49 71.98 50.22 26.32
40.20 70.16 66.58 43.64 24.75
3 24* 41 39 26 14
18.44 39.51 41.52 28.97 15.18
24.04 39.86 36.94 23.98 11.91
4 10* 18 18 11 6
7.86 16.85 17.71 12.36 6.47
10.37 16.89 15.52 10.04 4.98
5 and above 3 6 6 4 2
2.70 5.79 6.09 4.25 2.22
3.53 5.72 5.23 3.38 1.68
"indicates cells with marked discrepancy between entries one and two in the cell.

(viz. Cells for x = 4 and 5 and cells for y = 4 and 5). But the discrepancies in some
of the cells, viz. (x, y) = (0, 0), (0, I), (0, 2), (0, 3), (2, 0) and (3, 0) are rather
striking. In particular the X2- component f?om cell (0,O) is 9.46 which is almost
half of the total x2 value. Hence, one is led to think of possibility of mild violations
of assumptions that lead one to expect independent Poisson frequencies. In other
words, one is led not to ignore the observed correlation of -0.09391 as if it were
effectively zero.

From the Table, we have x = 1.69, s: = 1.55, = 2.01, s: = 1.73 and

Cov (x, y) = -0.154. Since the mean and variance are approximately equal for
 274 LAKSHMINARAYANA, PANDIT, AND SRINIVASA RAO

each of the variables x and y, one can expect a Poisson fit for each marginal. Also,
the correlations is -0.09391, a negative vahe. However, since the correlation is
negative, one cannot consider the w i n g of Campbell distribution for this data; one
can only try the present model. Using moments-estimators, one gets as parameter
estimates,

K, = 1.69 x, = 2.01 and B = -1.176

The expected fiequencies, using these estimates, are given as third entries
Downloaded by [North Carolina State University] at 00:07 06 May 2013

in each cell of Table-I. The corresponding calculated X2-valuefor goodness of fit

is 5.92 with 21 d.E, which is less than the X2-value (32.7) at 5% level of
significance and, ofcourse much less than the x2-value 18.62 obtained under the
assunaption of independence.

A comparison of these expected fiequencies with the corresponding

observed fiequencies shows a marked improvement in the fit. The maximum
absolute difference between observed and expected fiequencies is now reduced to
about 5 and the largest x2- contribution fiom the cells is 1.24, all other
contributions being less than one, in contrast to the case of independence
assumption (and zero correlation), which have the largest contribution as 9.46,
with at least 5 other cells, even with observed fiequencies 18 and above,
contributing values more than 1 to the X2 value. Hence it is obvious that the
present model with a negative correlation gives a far superior fit to the observed
data than the model assuming independence of x and y (and hence zero
correlation).

4. CONCLUSION

The bivariate Poisson distribution given by Campbell (1934) admits only

positive correlation by the very nature of the genesis of the distribution as
envisaged by him. However, the present approach is purely theoretical and is
obtained as a special case of the situations where a class of joint distriiutions all
 BIVARIATE POISSON DISTRIBUTION 275

having the same marginals, is obtainable. As is obvious, the parameter a which

can be interpreted as a measure of dependence can take values in the interval

Hence by choosing the sign of a one get either a positive

or negative correlation in the theoretical bivariate distribution with Poisson

marghals.

ACKNOWLEDGEMENTS
Downloaded by [North Carolina State University] at 00:07 06 May 2013

The authors th-y acknowledge the role of illuminating comments and

observations of the anonymous referees and the editor which have greatly
enhanced the clarity of the paper.

BIBLIOGRAPHY

Banerjee, D.P. (1959). 'On some theorems on Poisson distribution', Proceedings

of the National Academy of Science; India, Section A, 28, 30-33.

Campbell, J.T (1934). The Poisson correlation hction', Proceedings of the

Edinburgh Mathematical Society, 4, 18-26.

Holgate, P. (1964). 'Estimation for the bivariate Poisson distribution', Biometrica,

5 1, 241-245.

Johnson, N.L. and Kotz, S. (1969). 'Discrete distributions', John Wiley and Sons,
New York.

Patil, G.P. and Bildikar Seela (1967). 'Multivariate logarithmic series distribution
as a probability model in population and community ecology and some of
its statistical properties', Journal of the American Statistical Association,
62, 655-674.

Pandit, S.N.N., Rao, C.R, Sarma, Y.V.S.S. (1988). 'On a bivariate binomial
distniution', Assam Statistical Review, 2(2), 67-72.

Sibuya, M., Yoshimura, I. and Shimizu, R (1964). 'Negative multinomial

distribution', Annals of the Institute of Statistical Mathematics', Tokyo, 16,
409-426.
 276 LAKSHMINARAYANA, PANDIT, AND SRINIVASA RAO

Subba Rao (1990). 'Experimental studies on the yield of groundnuts in coastal

region', Project report, Andhra University, Visakhapatnam.

Teicher (1954). 'On the multivariate Poisson distribution', Skandinavisk

Actuariedskift, 5.

Received September, 1996; Revised August, 1998.

Downloaded by [North Carolina State University] at 00:07 06 May 2013