Review Article Volume 4 Issue 6
1Department of Mathematics, Marquette University, USA
2Department of Mathematical Sciences, Ball State University, USA
Correspondence: Munni Begum, Department of Mathematical Sciences, Ball State University, Muncie, IN 47306, USA
Received: September 17, 2016 | Published: November 7, 2016
Citation: Hossain F, Begum M. Tests of hypotheses for the parameters of a bivariate geometric distribution. Biom Biostat Int J. 2016;4(6):244-249. DOI: 10.15406/bbij.2016.04.00112
A bivariate geometric distribution is an extension to a univariate geometric distribution where the occurrence of three different types of events is considered. Many statisticians have studied and given different forms of a bivariate geometric distribution. In this paper, we considered the form given by Phatak & Sreehari.1 We estimated the parameters of this distribution under three different models using maximum likelihood estimation (mle) and derived deviances as the goodness of fit statistics for testing the parameters and deviance difference for comparing two models. Using simulated data we found that the deviance measure works well to test a reduced model against a full model.
Keywords: bivariate geometric distribution, deviance, deviance difference
Many situations in real world cannot be described by a single variable. Simultaneous occurrence of multiple events warrants multivariate distributions. For instance, univariate geometric distribution can represent occurrence of failure of one component of a system. However, to study systems with several components that may have different types of failures, such as twin engines of an airplane or the paired organ in a human body, bivariate geometric distributions are suitable. Bivariate geometric distribution has increasingly important roles in various fields, including reliability and survival analysis. There are different forms of a bivariate geometric distribution. Phatak & Sreehari1 provided a form of the bivariate geometric distribution which is considered here. They introduced a form of probability mass function which take into consideration of three different types of events. There are other forms which can be seen in Nair & Nair,2 Hawkes,3 Arnold et al.4 and Sreehari & Vasudeva.5 Basu & Dhar6 proposed a bivariate geometric model which is analogous to bivariate exponential model developed by Marshal & Olkin.7 Characterization results are developed by Sun & Basu,8 Sreehari,9 and Sreehari & Vasudeva.5
Omey & Minkova10 considered the bivariate geometric distribution with negative correlation coefficient and analyzed some properties, probability generating function, probability mass function, moments and tail probabilities. Krishna & Pundir,11 studied the plausibility of a bivariate geometric distribution as a reliability model. They derived the maximum likelihood estimators and Bayes estimators of the parameters and various reliability characteristics. They also compared these estimators using Monte-Carlo simulation.
In this paper, the parameters of a saturated model, reduced model and generalized linear model (glm) for a bivariate geometric distribution are estimated using the maximum likelihood method. We also derived deviances as the goodness of fit statistics for testing parameters corresponding to these models and deviance difference to compare two related models in order to determine which model fits the data well. Rest of the paper is organized as follows: section 2 describes the univariate geometric distribution, section 3 presents the bivariate geometric distribution, section 4 presents hypothesis testing, section 5 discusses a numerical example with simulated data and section 6 has the conclusion.
The probability mass function (pmf) of a random variable Y which follows a geometric distribution with probability of success p can be written as,
P(Y=y)=p(1−p)y,y = 0,1,2,...; 0<p<1, 0<q=1−p<1P(Y=y)=p(1−p)y,y = 0,1,2,...; 0<p<1, 0<q=1−p<1 .
The moment generating function can be given by,
MY(t)=p1−qetMY(t)=p1−qet
the mean and the variance of this distribution are
E(Y)=μY=1−pp=qp and V ar(Y)=1−pp2=qp2E(Y)=μY=1−pp=qp and V ar(Y)=1−pp2=qp2
An extension to the univariate geometric distribution is the bivariate geometric distribution which is discussed in the next section.
The joint probability mass function of a bivariate geometric distribution can be obtained by the product of a marginal and a conditional distribution, introduced by Phatak & Sreehari.1 They considered a process from which the units could be classified as good, marginal and bad with probabilities q1q1 , q2q2 and q3=(1−q1−q2)q3=(1−q1−q2) respectively. They proposed that the probability mass function of observing the first bad unit after several good and marginal units are passed as follows:
P(Y1=y1,Y2=y2)=(y1+y2y1)q1y1q2y2(1−q1−q2),y1,y2=0,1,2,...;P(Y1=y1,Y2=y2)=(y1+y2y1)q1y1q2y2(1−q1−q2),y1,y2=0,1,2,...; (1)
0 < q1+q2< 10 < q1+q2< 1 .
Here Y1Y1 and Y2Y2 denote the number of good and marginal units respectively before the first bad unit is observed.
The marginal distribution of Y1Y1 is a geometric distributions with probability of success (1−q1−q21−q2)(1−q1−q21−q2) , and can be written as follows,
P(Y1=y1)=(1−q1−q21−q2)(q11−q2)y1, y1=0,1,2,...P(Y1=y1)=(1−q1−q21−q2)(q11−q2)y1, y1=0,1,2,...
The conditional distribution of Y2Y2 given Y1Y1 is
P(Y2=y2/Y1=y2)= (y1+y2y2)q2y2(1−q2)y1+1, y1,y2= 0,1,2,...P(Y2=y2/Y1=y2)= (y1+y2y2)q2y2(1−q2)y1+1, y1,y2= 0,1,2,... (3)
The product of the marginal distribution of Y1Y1 in equation (2) and the conditional distribution of Y2Y2 given Y1Y1 in equation (3) gives the mass function of bivariate geometric distribution in equation (1).
Maximum likelihood estimation
Estimation of parameters in the absence of regressors
In order to find the maximum likelihood estimators (mle)s from a saturated model (parameters are different for each pair of observations), it suffices to consider the likelihood functions based on the marginal and conditional mass functions. Let Y1,....,YnY1,....,Yn be independent random vectors each having bivariate geometric distribution with different pairs of parameters (q1i,q2i)(q1i,q2i) for i=1,2,....,n.i=1,2,....,n. .
The log likelihood function based on the conditional distribution of Y2Y2 given Y1Y1 can be written as follows using (3):
l= n∑i=1[ln(1−q1i−q2i)−ln(1−q2i)+y1ilnq1i−y1iln(1−q2i)]l= n∑i=1[ln(1−q1i−q2i)−ln(1−q2i)+y1ilnq1i−y1iln(1−q2i)] (4)
Differentiating (4) with respect to q2iq2i and setting it equal to zero, we get the mle of q2iq2i as,
ˆq2i=y2iy1i+y2i+1ˆq2i=y2iy1i+y2i+1 (5)
The log likelihood function based on the marginal distribution of Y1Y1 from (2) is,
l= n∑i=1[ln(1−q1i−q2i)−ln(1−q2i)+y1ilnq1i−y1iln(1−q2i)]l= n∑i=1[ln(1−q1i−q2i)−ln(1−q2i)+y1ilnq1i−y1iln(1−q2i)]
Differentiating (6) with respect to q1iq1i and setting it equal to zero, the mle of q1iq1i can be derived as,
ˆq1i=y1iy1i+y2i+1ˆq1i=y1iy1i+y2i+1 (7)
Here, ˆq1iˆq1i and q1iq1i are the maximum likelihood estimators of q1iq1i and q2iq2i , i=1,...,ni=1,...,n respectively under the saturated model.
Similarly the maximum likelihood estimators (mle)s from a reduced model (parameters are the same for each pair of observations) can be obtained as:
ˆq2=↼y2⇀y1+↼y2+1 (8)
ˆq1=↼y1⇀y1+↼y2+1 (9)
Where ˆq1 and ˆq2 are the maximum likelihood estimators of q1 and q2 respectively under the reduced model.
Estimation of parameters in the presence of regressors: In the presence of regressors, one can employ a generalized linear model and hence estimate the parameters in terms of the estimated model parameters. The conditional distribution of Y2 given Y1 in (3) can be set as exponential family representation as follows,
P(Y2=y2/Y1=y1) = exp[y2ln q2−{−(y1+1)ln(1−q2)}+ln(y1+y2)!y1!y2!]
Here the natural parameter and the function of the natural parameter respectively are,
θ = lnq2
b(θ) = − (y1+1)ln(1−q2)
Thus the mean of the conditional distribution of Y2 given Y1 is
μi= E[Y2/Y1=y1]/= b'(θ)=y1+11−q2
A generalized linear model based on the conditional distribution of Y2 given Y1 can be written as,
g(μi) = lnμi=/ n =p∑j=1x2ijβ2i; i = 1,2,....,n; n>p
Since, Y2 represents the number of trials before a certain event can occur it is considered as count response, the linear predictor can be written as the logarithm of the mean μi . Thus the conditional link function can be expressed as,
g(μi)=lny1i+11−q2i=p∑j=1x2ijβ2i
⇒ ˆq2i= 1− y1i+1exp{∑pj=1x2ijβ2j} (10)
Here, β is an element of the matrix β corresponding to the covariate x2ij which represents the effect of covariate to the mean responses through the link function g(μi)=ni .
Differentiating (6) again with respect to q1i , setting it to zero and using (10) we get,
ˆq1i=y1iexp{p∑j=1x2ijβ2j} (11)
In order to test the identical parameter assumption across each pair of observed data, we derived deviance as a goodness of fit statistics. Additional deviance statistics are derived for generalized linear model (glm) to compare two nested glms.
Deviance for reduced model with identical parameter assumption
The log likelihood function for the saturated model can be written using (1) and the maximum likelihood estimates of the parameters q1i and q2i from equations (5) and (7) respectively as follows,
l(bmax;y)=n∑i=1[y1ilny1i−y1iln(y1i+y2i+1)+y2ilny2i−y2iln(y1i+y2i+1)−ln(y1i+y2i+1)+ln(y1i+y2i)!−lny1i!−lny2i!] (12)
Similarly, the log likelihood function of the reduced model can be written using (1) and the maximum likelihood estimates of q1 and q2 from equations (8) and (9) respectively as follows,
l(b;y)=i=n∑i=1[y1ilnˆy−y1iln(ˆy1+ˆy2+1)+y2ilnˆy2−y2iln(ˆy1+ˆy2+1)−ln(ˆy1+ˆy2+1)+ln(y1i+y2i)!−lny1i!−lny2i!] (13)
Thus the deviance statistic for testing the identical parameter for each observed pair of data can be expressed as follows,
DI=2[l(bmax;y)−l(b;y)]=2i=n∑i=1[y1ilny1iˆy1−y1ilny1i+y2i+1ˆy1+ˆy2+1+y2ilny2iˆy2−y2ilny1i+y2i+1ˆy1+ˆy2+1−lny1i+y2i+1ˆy1+ˆy2+1] (14)
According to Dobson (2001), DI follows a X2 distribution with (2n−2) degrees of freedom.
Deviance for a GLM
The deviance statistic for the glm of interest can be written using (1) and the maximum likelihood estimates of q1i and q2i based on the glm from equations (10) and (11) respectively as follows,
l(b;y)=i=n∑i=1[y1ilny1iexp{∑j=pj=1x2ijβ2j}+y2iln(1−y1i+1exp{∑j=pj=1x2ijβ2j})+ln1exp{∑j=pj=1x2ijβ2j}+ln(y1i+y2i)!−lny1i!−lny2i!] (15)
Thus the deviance can be expressed as follows,
DII=2[l(bmax;y)−l(b;y)]=2i=n∑i=1[y2ilny2i−(y1i+y2i+1)ln(y1i+y2i+1)+{j=p∑j=1x2ijβ2j}(y1i+y2i+1)−y2iln(exp{j=p∑j=1x2ijβ2j}−y1i−1)] (16)
According to Dobson (2001), DII follows χ2 distribution with (2n−p) degrees of freedom.
Comparison between two GLMs
In order to compare two nested generalized linear models, we consider the following hypotheses. The null hypothesis corresponding to a smaller model (M0) in terms of number of regression parameters is
H0:β=β0=[β1β2...βq]
The alternative hypothesis corresponding to a bigger model (M1 with q < p < n) within which the smaller model is nested can be written as,
H1:β=β1=[β1β2...βp]
We can test H0 against H1 using the difference of the deviance statistics. Here, l(b0;y) is used to denote the likelihood function corresponding to the model M0 and l(b1;y) to denote the likelihood function corresponding to the model M1 . Hence the deviance difference can be written as,
△D=D0−D1=2[l(bmax;y)−l(b0;y)]−2[l(bmax;y)−l(b1;y)]=2[l(b1;y)−l(b0;y)]=2i=n∑i=1[{j=p∑j=1x2ijβ2j}(y1i+y2i+1)−y2iln(exp{j=p∑j=1x2ijβ2j}−y1i−1)−{j=q∑j=1x2ijβ2j}(y1i+y2i+1)+y2iln(exp{j=q∑j=1x2ijβ2j}−y1i−1)−{j=q∑j=1x2ijβ2j}(y1i+y2i+1)+y2iln(exp{j=q∑j=1x2ijβ2j}−y1i−1)]
According to Dobson12 this ΔD follows χ2 distribution with p−q degrees of freedom.
If the value of ΔD is consistent with the χ2(p−q) distribution we would generally choose the M0 corresponding to H0 because it is simpler. On the other hand, if the value of ΔD is in the critical region i.e., greater than the upper tail 100×α% point of the χ2(p−q) distribution then would reject H0 in favor of H1 on the grounds that model M1 provides a significantly better description of the data.
To determine the efficiency of our derived deviances we need to have data with known parameters. However, we cannot generate data directly from bivariate geometric distribution using the available computer software packages. Krishna and Pundir suggested an algorithm based on a theorem given by Hogg et al.13 to generate random numbers from bivariate geometric distribution. According to this, paired values can be generated from a bivariate geometric distribution using the following steps,
In this subsection, we use the following steps to check our derived deviance for the reduced model with identical values of parameters (q1,q2) for each observed pair of data.
We take the values of q1 and q2 ranging from 0.10 to 0.90 and satisfying the constraint q1+q2<1 . We considered several values for the pair (q1,q2) and generate random pairs to observe the efficiency of our derived deviance under different parametric values. For each specified pairs of parameters (q1,q2) , we ran this experiment twice to see whether there is a change in our decision due to randomness. The values of the pair of parameters and the corresponding deviance values are tabulated as follows.
The deviance we derived to test the parameters of the reduced model works well as we see that all, but four of the values of the deviances are smaller than X2198(0.95) . However, among these four values of the deviances three are greater than X2198(0.95) , but less than X2198(0.99) . So, it can be concluded that our derived deviance works well. On the other hand, if most of the values of the deviances had a larger value than our desired χ2 value, then we had to conclude that our derived deviance does not work in testing hypothesis regarding the parameters of the reduced model.
Parameters |
Deviance |
χ198 (0.95) |
χ198 (0.975) |
χ198 (0.99) |
q1=0.30,q2=0.30 |
177.4164 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.30 |
172.3071 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.40 |
185.3107 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.40 |
159.5293 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.50 |
193.8942 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.50 |
158.266 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.60 |
223.1697 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.60 |
193.667 |
231.8292 |
238.8612 |
247.2118 |
q1=0.40,q2=0.30 |
216.3456 |
231.8292 |
238.8612 |
247.2118 |
q1=0.40,q2=0.30 |
211.828 |
231.8292 |
238.8612 |
247.2118 |
q1=0.50,q2=0.30 |
148.1757 |
231.8292 |
238.8612 |
247.2118 |
q1=0.50,q2=0.30 |
254.3887 |
231.8292 |
238.8612 |
247.2118 |
q1=0.60,q2=0.30 |
239.3245 |
231.8292 |
238.8612 |
247.2118 |
q1=0.60,q2=0.30 |
215.4915 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.50 |
232.1984 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.50 |
191.7516 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.60 |
184.1803 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.60 |
236.0869 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.10 |
97.9206 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.10 |
85.10731 |
231.8292 |
238.8612 |
247.2118 |
Table 1 Estimation of deviance for different parameters under consideration
Parameters |
Deviance |
χ198
|
χ198
|
χ198
|
q1=0.10,q2=0.20 |
100.8624 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.20 |
155.157 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.30 |
155.157 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.30 |
123.3245 |
231.8292 |
238.8612 |
247.2118 |
q1=0.20,q2=0.20 |
113.3245 |
231.8292 |
238.8612 |
247.2118 |
q1=0.20,q2=0.20 |
147.3637 |
231.8292 |
238.8612 |
247.2118 |
q1=0.20,q2=0.30 |
166.6306 |
231.8292 |
238.8612 |
247.2118 |
q1=0.20,q2=0.30 |
157.8232 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.10 |
133.2772 |
231.8292 |
238.8612 |
247.2118 |
q1=0.30,q2=0.10 |
131.2191 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.80 |
183.8584 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.80 |
218.8224 |
231.8292 |
238.8612 |
247.2118 |
q1=0.80,q2=0.10 |
203.6515 |
231.8292 |
238.8612 |
247.2118 |
q1=0.80,q2=0.10 |
177.6116 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.40 |
144.1728 |
231.8292 |
238.8612 |
247.2118 |
q1=0.10,q2=0.40 |
168.524 |
231.8292 |
238.8612 |
247.2118 |
q1=0.70,q2=0.10 |
169.3248 |
231.8292 |
238.8612 |
247.2118 |
q1=0.70,q2=0.10 |
177.8397 |
231.8292 |
238.8612 |
247.2118 |
q1=0.60,q2=0.10 |
177.1335 |
231.8292 |
238.8612 |
247.2118 |
q1=0.70,q2=0.10 |
197.0526 |
231.8292 |
238.8612 |
247.2118 |
q1=0.50,q2=0.10 |
159.3473 |
231.8292 |
238.8612 |
247.2118 |
q1=0.50,q2=0.10 |
146.7018 |
231.8292 |
238.8612 |
247.2118 |
Table 2 Estimation of deviance for different parameters under consideration (contd)
In this paper, we addressed an important problem of inference regarding bivariate geometric distribution and developed testing procedure for the parameters of this distribution with and without covariate information. Our method depends on deriving the deviance statistics using maximum likelihood estimators (mle) of parameters. Our mles of the parameters of the bivariate geometric distribution are obtained using the conditional and the marginal distributions.
We conducted a numerical analysis based on simulated data for the testing the 11 identical parameter assumption for each pair of observed data. Our numerical example did not consider any covariate information. We found that without covariate information our derived deviance worked well in most cases.
None.
Author declares that there are no conflicts of interest.
©2016 Hossain, et al. This is an open access article distributed under the terms of the, which permits unrestricted use, distribution, and build upon your work non-commercially.
2 7