A New Multivariate Weibull Distribution

We have proposed a new multivariate Weibull distribution as a compound distribution of univariate Weibull distributions. We have studied some properties of the proposed distribution. The multivariate distribution of records for proposed distribution has also been studied. Estimation of the parameters has been done alongside application to the real data set.


Introduction
Weibull distribution, introduced by Weibull (1951), has been a popular model for modelling the lifetime data.The density and distribution function of Weibull distribution can be presented in various ways, see for example Johnson et al. (1995).A relatively simple form of the density function is The distribution function corresponding to (1) is Since its emergence the distribution has attracted so many scholars.Various authors have used the distribution to model lifetime data and in developing new distributions.Madholkar and Srivastava (1993)  Chandler (1952) introduced record statistics as a model for ordered random variables.The records are defined in a sequence of independent and identically distributed random variables as under.
Suppose a sequence of independently and identically distributed random variables 1 2 , , , n X X X  is available with same distribution function   F x .Suppose further that   From this definition it is clear that 1 X is an upper record value.We also associate the indices to each record value with which they occur.These indices are called the record time We can readily see that   The joint density function of mth and nth upper records is given by Chandler (1952) as Record statistics has been studied by several authors for various probability distributions.Ahsanullah (1992) has provided general results for distribution of records for continuous probability distributions.A comprehensive review of record statistics can be found in Ahsanullah (1995) and Nevzorov (2001).Often we have a sample from some bivariate distribution and the sample is arranged with respect to one of the variable.The automatically shuffled other variable is called concomitant of ordered variable.The concomitant of records appear when the bivariate sample is arranged with respect to records.The density function of nth concomitant of record is given by Ahsanullah (1995) as where x is defined in (3).The joint distribution of two concomitants of record values is given by Ahsanullah (1995) as where where   , | f y z x is the conditional distribution of Y and Z given X = x.Shahbaz where   is the conditional distribution of vector y given X and     X n f x is defined in (3).
In this paper we have proposed a new multivariate Weibull distribution and have studied its properties.The new multivariate distribution is proposed in section 2 with some graphical study.In section 3 some common properties of the proposed distribution has been studied.Multivariate concomitants of the distribution has been studied in section 4. Application of proposed distribution is given in section 5.

The New Multivariate Weibull Distribution
, 0 ; , , , , , 0, 9 , , , , , 0 x z z     .Using the idea given by Shahbaz et al. (2012), we have proposed the multivariate Weibull distribution in the following.Suppose X has a Weibull distribution with density function We write X has , , , W x y y   and so on, such that , , , ,

f y x x y x y f y x y x y y x y y f y x y y x y y y x y y y
and so on.Now the density function of   . Using above densities, the density function of   where , 0 i x y  .The distribution (12) can be studied for various choices of   In this paper we have studied the distribution (12) for following choices of   . Using these in (12), the density function of   x y y x y y x y y y x y y y x y y y x y y y y . The joint marginal distribution of y is readily obtained as The marginal distribution of ith component of y is f y f dy dy dy dy p y dy dy dy dy y The joint marginal distribution of any pair   , i k y y is given as ; , , , 0. 1 We now obtain the entries of mean vector and covariance matrix of y as under.
The qth moment of Yi is Using q = 1, the ith entry of mean vector of Yi is   Further, using q = 2, we have The variance can be easily obtained from (17) and (18).Further y y f y y dy dy y y dy dy y y y which after simplification becomes   The covariance can be obtained from ( 17) and (19).Following Further, the joint conditional distribution of y given X = x is obtained from ( 11) and ( 13) and is given as The conditional moments of order q for distribution (20) are given as and The conditional means and conditional variances can be computed from ( 21) and ( 22).Further, the joint conditional moments for distribution (20) are The conditional mean vector and covariance matrix can be computed for specific values of i  , k  ,  and x.
We now obtain the distribution of multivariate concomitants for the new multivariate Weibull distribution. where We now obtain the distribution of multivariate concomitants for the new multivariate Weibull distribution given in previous section.The density function of new multivariate Weibull distribution is given in (13).The marginal distribution of X is given in (11) and the conditional distribution of y given x is given in (20).Now for distribution (11) we have Now using ( 20) and ( 23) in ( 8), the distribution of multivariate concomitant of order statistics for new multivariate Weibull distribution is obtained as under The marginal distribution of ith concomitant is readily written from (24) as The joint marginal distribution of two concomitants is obtained from (24) as The qth moment of ith concomitant is readily obtained from (24) as The mean and variance can be easily obtained from (26).Further, the product moment between ith and kth concomitant is The covariance con be computed by using ( 26) and ( 27).The mean, variance, covariance and correlation coefficient can be computed for various choices of n, i  and k  .

Estimation
In this section we have discussed maximum likelihood estimation for parameters of new multivariate Weibull distribution.For this consider the density function of new multivariate Weibull distribution given in (13) as The log of density function is given as The log of likelihood function is immediately written as The derivatives of log of likelihood function wrt the parameter  is The maximum likelihood estimators can be obtained by solving (p + 1)-equations in (28) and ( 29).The solution is obviously done by using some numerical method.The entries of observed Fisher information matrix for parameter  are given as The entries of Fisher information matrix wrt the parameters 's  are The observed Fisher Information matrix can be computed for a given data.

Simulation Study and Application
In this section we have given simulation study and real data application of the proposed multivariate Weibull distribution.The simulation study has been conducted to see the performance of maximum likelihood estimates and the real data application has been conducted to see suitability of proposed multivariate distribution.

Simulation Study
In this sub-section we have presented the simulation results to see the performance of maximum likelihood estimates of proposed multivariate Weibull distribution.The simulation study has been carried out by using trivariate Weibull distribution.The algorithm for simulation is given below 1.
Draw sample of size n, for n = 50; 100; 300 and 500, from Weibull distribution with shape parameter  and scale parameter 1. Denote this sample with variable X.

2.
For each observation of X, draw sample of size 1 from Weibull distribution with shape parameter 1  and scale parameter x  .Repeat this procedure for all observations of X. Denote this sample with variable Y1.

3.
For each observation of X and Y1, draw sample of size 1 from Weibull distribution with shape parameter 2  and scale parameter 1 1 x y   .Repeat this procedure for all observations of X and Y1.Denote this sample with variable Y2.

4.
Using the trivariate sample   , , x y y obtain maximum likelihood estimates of 1 ,   and 2  .

6.
Compute mean and standard deviation of 5000 estimates obtained in step 5 to get the results.
The results of simulation study are given in Table-2 below.From the results we can see that the performance of maximum likelihood estimators increases with increase in the sample size.

Application
In this subsection we have applied the proposed multivariate Weibull distribution on two real data sets.The data set 1 contains information about body weight and brain weight of 84 mammals as reported by Ramsay and Schafer (1997).The second data set that we have used contains information about height and forced expiratory volume (FEV) of 655 smokers as reported by Rosner (1999).The bivariate Weibull distribution has been _tted on the data.The results are shown in the Table-3 below.The bivariate histogram for two datasets given below

Mammal's Data Smoker's Data
We can see that the surface plot is close to bivariate histograms.Hence the bivariate Weibull distribution fits data reasonably well.

Conclusions and Recommendation
In this paper we have proposed a multivariate Weibull distribution and have studied some of its common properties.The distribution of multivariate concomitants of records for the proposed distribution has also been obtained.The distribution can be used to obtain the distribution of concomitants for any number of variables.
have introduced the exponentiated Weibull distribution as an extension of Weibull distribution.Famoye et al. (2005) have used the Weibull distribution in context of Beta-G distribution of Eugene et al. (2002) to propose the Beta-Weibull distribution.Cordeiro et al. (2010) have proposed the Kumaraswamy-Weibull distribution as an alternative to Weibull distribution with much wider applicability.

1 1 U
 .The upper record values are denoted by   X n .The density function of nth upper record statistics is given by Chandler (1952) as

Figure 1 :
Figure 1: Bivariate Histograms of Two Data Sets Mammal's Data Smoker's Data

Figure 2 :
Figure 2: Surface Plots of Fitted Distributions ). Distribution of concomitants of records for various distributions has been studied by many authors.Shahbaz et al. (2011) have studied concomitants of record statistics for a new bivariate Weibull distribution.The distribution of concomitants has been extended to bivariate case by Shahbaz et al. (2012).The density function of nth bivariate record statistics is given as table contains means, variances, covariances and correlations between Yi and Yk for selected values of i  and k .From the table we can see that that the means, variances, covariances and correlations decrease with increase in i  and k  .

Distribution of Multivariate Concomitants The
concomitants of record values have been studied by several authors.Shahbaz et al. (2012) have studied the bivariate concomitants of record values for a trivariate Weibull distribution.Arnold et al. (2009) have provided the distribution of multivariate concomitants of record values which is given in (8) as