The bivariate generalized linear failure rate distribution and its multivariate extension

https://doi.org/10.1016/j.csda.2010.06.006Get rights and content

Abstract

The two-parameter linear failure rate distribution has been used quite successfully to analyze lifetime data. Recently, a new three-parameter distribution, known as the generalized linear failure rate distribution, has been introduced by exponentiating the linear failure rate distribution. The generalized linear failure rate distribution is a very flexible lifetime distribution, and the probability density function of the generalized linear failure rate distribution can take different shapes. Its hazard function also can be increasing, decreasing and bathtub shaped. The main aim of this paper is to introduce a bivariate generalized linear failure rate distribution, whose marginals are generalized linear failure rate distributions. It is obtained using the same approach as was adopted to obtain the Marshall–Olkin bivariate exponential distribution. Different properties of this new distribution are established. The bivariate generalized linear failure rate distribution has five parameters and the maximum likelihood estimators are obtained using the EM algorithm. A data set is analyzed for illustrative purposes. Finally, some generalizations to the multivariate case are proposed.

Introduction

The two-parameter linear failure rate (LFR) distribution, whose hazard function is monotonically increasing in a linear fashion, has been used quite successfully to analyze lifetime data. For some basic properties and for different procedures of estimation of the parameters of the LFR distribution, the readers are referred to Bain (1974), Pandey et al. (1993), Sen and Bhattacharyya (1995), Lin et al., 2003, Lin et al., 2006 and the references cited therein.

Recently, Sarhan and Kundu (2009) introduced a three-parameter generalized linear failure rate (GLFR) distribution by exponentiating the LFR distribution as was done for the exponentiated Weibull distribution by Mudholkar et al. (1995). The exponentiation introduces an extra shape parameter in the model, which may yield more flexibility in the shape of the probability density function (PDF) and hazard function. Several properties of this new distribution are established. It is observed that several known distributions like exponential, Rayleigh and LFR distributions can be obtained as special cases of the GLFR distribution.

The aim of this paper is to introduce a new bivariate generalized linear failure rate (BGLFR) distribution, whose marginals are GLFR distributions. This new five-parameter BGLFR distribution is obtained using a method similar to that used to obtain the Marshall–Olkin bivariate exponential model Marshall and Olkin (1967) and Sarhan and Balakrishnan’s bivariate distribution, Sarhan and Balakrishnan (2007). The proposed BGLFR distribution is constructed from three independent GLFR distributions using a maximization process. Creating a bivariate distribution with given marginals using this technique is nothing new. Alternatively, the same BGLFR distribution can be obtained by coupling the GLFR marginals with the Marshall–Olkin copula (Nelsen, 1999). This new distribution is a singular distribution, and it can be used quite conveniently if there are ties in the data. The joint cumulative distribution function (CDF) can be expressed as a mixture of an absolutely continuous distribution function and a singular distribution function. The joint probability density function (PDF) of the BGLFR distribution can take different shapes and the cumulative distribution function can be expressed in a compact form. The BGLFR distribution can be applied to a maintenance model or a stress model as introduced by Kundu and Gupta (2009).

Several dependency properties of this new distribution are investigated, which will be useful for data analysis purposes. The BGLFR copula has a total positivity of order 2 (TP2) property. Each component is stochastically increasing with respect to the other. This implies that the correlation is always non-negative and the two variables are positively quadrant dependent. Moreover, the correlation between the two variables varies between 0 and 1. Kendall’s tau index can be calculated using the copula property and can be positive. The population version of the medial correlation coefficient as defined by Blomqvist (1950) is always non-negative. The bivariate tail dependence is always positive.

The BGLFR distribution has five parameters, and their estimation is an important problem in practice. The usual maximum likelihood estimators can be obtained by solving five non-linear equations in five unknowns directly, which is not a trivial issue. To avoid difficult computation we treat this problem as a missing value problem and use the EM algorithm, which can be implemented more conveniently than the direct maximization process. Another advantage of the EM algorithm is that it can be used to obtain the observed Fisher information matrix, which is helpful for constructing the asymptotic confidence intervals for the parameters.

Alternatively, it is possible to obtain approximate maximum likelihood estimators by estimating the marginals first and then estimating the dependence parameter through a copula function, as suggested by Joe (1997, Chapter 10), which has the same rate of convergence as the maximum likelihood estimators. This is computationally less involved compared to the MLE calculations. This approach is not pursued here. Analysis of a data set is presented for illustrative purposes. The proposed model provides a better fit than the Marshall–Olkin bivariate exponential model or the recently proposed bivariate generalized exponential model (Kundu and Gupta, 2009).

Although in this paper we mainly discuss the BGLFR, many of our results can be easily extended to the multivariate case. Moreover, the LFR distribution is a proportional reversed hazard model, and our method may be used to introduce other bivariate proportional reversed hazard models.

The rest of the paper is organized as follows. We briefly introduce the GLFR distribution in Section 2. In Section 3 we introduce the BGLFR distribution and study its different properties. The EM algorithm is described in Section 4, and analysis of a data set is presented in Section 5. We discuss the multivariate generalization in Section 6, and finally conclude the paper in Section 7.

Section snippets

Generalized linear failure rate distribution

A random variable X has a linear failure rate distribution with parameters β0 and γ0 (such that β+γ>0), if X has the following distribution function: FLFR(x;β,γ)=1exp{βxγ2x2}, for x>0. The exponential distribution with mean 1/β (ED(β)) and the Rayleigh distribution with parameter γ (RD(γ)) can be obtained as special cases from the LFR distribution. The PDF of the LFR distribution can be decreasing or unimodal, but the failure rate function is either increasing or constant only (Sen and

Bivariate generalized failure rate distribution

In this section we introduce the BGLFR distribution using a method similar to that which was used by Marshall and Olkin (1967) to define the Marshall–Olkin bivariate exponential (MOBE) distribution.

Suppose U1,U2 and U3 are three independent random variables such that UiGLFR(αi,β,γ) for i=1,2 and 3. Define X1=max{U1,U3)andX2=max{U2,U3}. Then we say that the bivariate vector (X1,X2) has a bivariate GLFR (BGLFR) distribution, with parameters (α1,α2,α3,β,γ), and we denote it by BGLFR (α1,α2,α3,β,γ

Estimation

In this section we consider the estimation of the unknown parameters of the BGLFR model. It is assumed that we have a sample of size n, of the form {(x11,x12),,(xn1,xn2)} from BGLFR(α1,α2,α3,β,γ), and our problem is to estimate α1,α2,α3,β,γ from the given sample. First we obtain the MLEs of the unknown parameters. Since the computation of the MLEs is computationally quite involved, we propose alternative estimators, which can be obtained in a more convenient manner.

For further development we

Data analysis

In this section we present the analysis of a data set mainly to illustrate how the proposed model and the EM algorithm work in practice.

UEFA Champion’s League Data: The data set has been obtained from Meintanis (2007) and is presented in Table 2. It represents soccer data where at least one goal is scored by the home team and at least one goal is scored directly from a penalty kick, foul kick or any other direct kick (all of them will be called kick goals) by any team that has been considered.

Multivariate generalized linear failure rate distribution

In this section we are in a position to define the m-variate generalized linear failure rate distribution and provide some of its properties. It may be mentioned that recently Franco and Vivo (2009) provided a multivariate extension of the Sarhan–Balakrishnan bivariate distribution and studied its various properties.

Suppose U1,,Um+1 are m+1 independent random variables such that UiGLFR(αi,β,γ) for i=1,,m+1. Define Xj=max{Uj,Um+1},j=1,2,,m. Then we say that X=(X1,,Xm) is an m-variate GLFR

Conclusions

In this paper we have introduced the bivariate generalized linear failure rate distribution whose marginals are generalized linear failure rate distributions. The proposed bivariate distribution is a singular distribution, and it can be used quite effectively instead of the Marshall–Olkin bivariate exponential model, or the bivariate generalized exponential model when there are ties in the data. Several properties of this new distribution have been established, and also we proposed using the EM

References (22)

  • M. Franco et al.

    A multivariate extension of Sarhan and Balakrishnan’s bivariate distribution and its aging and dependence properties

    Journal of Multivariate Analysis

    (2009)
  • Cited by (45)

    • On a bivariate generalized inverted Kumaraswamy distribution

      2020, Physica A: Statistical Mechanics and its Applications
      Citation Excerpt :

      Contribution to this idea, Sarhan and Balakrishnan [5] introduced a bivariate distribution that is more flexible than the bivariate exponential distribution. Using the maximum instead of the minimum in the Marshall and Olkin scheme, Kundu and Gupta [6, 7] and Sarhan et al. [8] introduced the bivariate generalized exponential, Bivariate proportional reversed hazard and bivariate generalized linear failure rate distributions, respectively. And Muhammed [9–11] introduced the bivariate inverse Weibull distribution, bivariate Dagum Distribution and bivariate generalized Burr distributions respectively.

    • A new approach to measure systemic risk: A bivariate copula model for dependent censored data

      2019, European Journal of Operational Research
      Citation Excerpt :

      Based on our knowledge, there is a lack of theoretical justification in the literature to use the AIC for comparing absolutely continuous and non-absolutely continuous copulae. As El-Bassiouny, Shahen, and Bouhawwash (2018), Hanagal and Sharma (2015) and Sarhan, Hamilton, Smith, and Kundu (2011) used the AIC for a copula function with a singular component, we calculate this criterion for the MO copula using the pseudo likelihood function (3.6) in Eq. (4.1). We choose the MO copula based on the results in Table 2 and its characteristics described in Section 3.2.

    • On bivariate Teissier model using Copula: dependence properties, and case studies

      2024, International Journal of System Assurance Engineering and Management
    • A new bivariate lifetime distribution: properties, estimations and its extension

      2024, Communications in Statistics: Simulation and Computation
    • Bivariate Semi-Parametric Model: Bayesian Inference

      2023, Methodology and Computing in Applied Probability
    • Modelling bivariate failure time data via bivariate extended Chen distribution

      2023, Stochastic Environmental Research and Risk Assessment
    View all citing articles on Scopus
    View full text