# On a bivariate Birnbaum-Saunders distribution parameterized by its means: features, reliability analysis and application

Birnbaum-Saunders models have been widely used to model positively skewed data. In this paper, we introduce a bivariate Birnbaum-Saunders distribution which has the means as parameters. We present some properties of the univariate and bivariate Birnbaum-Saunders models. We discuss the maximum likelihood and modified moment estimation of the model parameters and associated inference. A simulation study is conducted to evaluate the performance of the maximum likelihood and modified moment estimators. The probability coverages of confidence intervals are also discussed. Finally, a real-world data analysis is carried out for illustrating the proposed model.

## Authors

• 18 publications
• 5 publications
• 18 publications
• 2 publications
• 5 publications
01/16/2018

### On a bimodal Birnbaum-Saunders distribution with applications to lifetime data

The Birnbaum-Saunders distribution is a flexible and useful model which ...
01/27/2019

### A computational method for estimating Burr XII parameters with complete and multiple censored data

Flexibility in shape and scale of Burr XII distribution can make close a...
06/08/2011

### Moment based estimation of stochastic Kronecker graph parameters

Stochastic Kronecker graphs supply a parsimonious model for large sparse...
12/14/2020

### The Bivariate Defective Gompertz Distribution Based on Clayton Copula with Applications to Medical Data

In medical studies, it is common the presence of a fraction of patients ...
01/10/2021

### Modelling wind speed with a univariate probability distribution depending on two baseline functions

Characterizing the wind speed distribution properly is essential for the...
06/03/2022

### Modified Bivariate Weibull Distribution Allowing Instantaneous and Early Failures

In reliability and life data analysis, the Weibull distribution is widel...
07/26/2021

### A Novel Bivariate Generalized Weibull Distribution with Properties and Applications

Univariate Weibull distribution is a well-known lifetime distribution an...
##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction

The Birnbaum-Saunders (BS) distribution was proposed by Birnbaum and Saunders (1969) motivated by problems of vibration in commercial aircrafts that caused fatigue in materials. Although, in principle, its origin is for modeling equipment lifetimes subjected to dynamic loads, the BS distribution has been widely studied and applied in many applied fields including, for example, engineering, business, economics, medicine, atmospheric contaminants, finance and quality control; see Jin and Kawczak (2003), Balakrishnan et al. (2007), Bhatti (2010), Villegas et al. (2011), Paula et al. (2012), Saulo et al. (2013), Marchant et al. (2013), Leiva et al. (2014), Leão et al. (2017a, b), and references therein. The interested reader on the BS distribution is refereed to Johnson et al. (1995); Leiva (2016)

. These works present a full review about this model. The BS distribution has been used quite effectively to model positively skewed data, especially lifetime data and crack growth data. This distribution is related to the normal distribution and has interesting properties.

Some works on reparameterized versions of the BS distribution were proposed by Ahmed et al. (2008), Lio et al. (2010) and Santos-Neto et al. (2012, 2016). In particular, the work of Santos-Neto et al. (2012) proposed several parameterizations of the BS distribution, which allow diverse features of data modeling to be considered. One of such parameterizations is indexed by the parameters and , where is a scale parameter and the mean of the distribution, whereas is a shape and precision parameter. The notation

is used when an random variable (RV) follows a reparameterized BS (RBS) distribution; more details about this model is found in Section

2.

The bivariate version of the BS distribution was proposed by Kundu et al. (2010), where the authors discussed maximum likelihood (ML) estimation and modified moment (MM) estimation of the model parameters. Recently, Khosravi et al. (2015) observed that the bivariate BS model proposed by Kundu et al. (2010)

can be written as the weighted mixture of bivariate inverse Gaussian distribution

(Kocherlakota, 1986) and its reciprocals. They also introduced a mixture of two bivariate BS distributions and discussed its various properties. Kundu et al. (2013) extended to the multivariate case the generalized BS distribution introduced by Díaz-García and Leiva (2005). Other bivariate and multivariate distributions related to the BS model can be found in Vilca et al. (2014a, b), Kundu (2015b, a) and Jamalizadeh and Kundu (2015).

In this context, the primary objective of this paper is to introduce a bivariate RBS (BRBS) distribution based on the RBS distribution proposed by Santos-Neto et al. (2012). The secondary objectives are: (i) to present the reliability measure for the BRBS model and obtain the monotonicity of the hazard rate (HR); (ii) to discuss some unimodality properties of the BRBS model; (iii) to derive the ML estimators and MM estimators of the unknown parameters as well as their asymptotic properties; (iv) to evaluate the performance of the ML and MM estimators via Monte Carlo (MC) simulations; and (v) to illustrate the potential applications by using real data.

The rest of the paper proceeds as follows. In Section 2, we describe briefly the BS and RBS distributions and their related properties. In Section 3, we introduce the BRBS distribution and discuss some of its properties. In Section 4, we present the ML and MM estimators of the unknown parameters and their corresponding asymptotic results. In Section 5, an evaluation of the ML and MM estimators using MC simulation is shown, as well as an illustrative example by using a real data set. Finally, in Section 5.2, we provide some concluding remarks and also point out some problems worthy of further study.

## 2 BS distributions

### 2.1 The BS distribution

The BS distribution is related to the normal distribution by means of the stochastic representation

 T=β4[αZ+√(αZ)2+4]2,

where and are shape and scale parameters, respectively, is a RV following a standard normal distribution , such that is BS distributed with notation

. The probability density function (PDF) of

is given by

 (1)

The mean and variance of

are given by and , respectively. The scale parameter is also the median of the distribution. The BS distribution holds the reciprocal property, that is, has the same distribution of with the parameter replaced by , , which implies

 E[1T]=β(1+α22), Var[1T]=α2β2(1+54α2).

### 2.2 The RBS distribution

The RBS distribution is indexed by the parameters and , where and are the original BS parameters of (1), and are the scale and shape (precision) parameters, respectively. If , then its PDF, for , is given by

 f(t;μ,δ)=exp(δ/2)√δ+14t3/2√πμ(t+δμδ+1)exp(−δ4[t(δ+1)δμ+δμt(δ+1)]), (2)

and its cumulative distribution function (CDF) is denoted by

. From (2), the survival function (SF) and hazard rate (HR) function are given by

 S(t;μ,δ) =12Φ(t+δ(t−μ)2√t(1+δ)μ), h(t;μ,δ) =exp(−(−δμ+δt+t)24(δ+1)μt)(δμ+δt+t)2√πμ(δ+1)√μt3/2Φ(t+δ(t−μ)2√t(1+δ)μ),

respectively, where is the CDF of the standard normal distribution. Figure 1 displays some shapes for the PDF and HR of .

Considering the function

 a(t;α,β)=1α[√tβ−√βt],α,β>0, (3)

and denoting , where and , expressions for the first, second and third derivatives of are given by

 a′(t)=12α [1√βt+1t√βt],a′′(t)=−14αt[1√βt+3t√βt],a′′′(t)=38αt2[1√βt+5t√βt]. (4)

Note that where denotes the PDF of the standard normal distribution. The function is a bijection of to and has inverse, denoted by , given by

 a−⊥(s)=β4[αs+√(αs)2+4]2,s∈R.

Some properties of the RBS distribution are the following. Let and . The RBS distribution satisfies the following properties:

1. The PDF of the RBS distribution has at most one mode, see Proposition 2.7 in Vila et al. (2017).

2. If , the moments about the origin of are given by

and in general

where . Then,

Since , the th moment always exists. As mentioned, can be interpreted as a precision parameter, that is, for fixed values of , when , the variance of tends to zero. Also, for fixed , if , then .

3. If , it is possible to show that , with , and . On the other hand, if , since the function is strictly increasing, it follows that

4. Let and be two RVs statistically independent such that represents “strength” and represents “stress”, then the reliability of a component is given by Since , with , and for each , by the independence of the RVs, we have that

 Z∗=α1a1(T1)−α2a2(T2)√α21+α22∼N(0,1).

Then, whenever .

5. Let , where

denotes the chi-squared distribution with

degrees of freedom, and be the PDF of the RV . Then,

 g(u;μ,δ)=2a2(u)f(u;μ,δ),u>0.

## 3 The bivariate RBS distribution

### 3.1 Density and shape analysis

The bivariate random vector

is said to follow a BRBS distribution with parameters , , , , , denoted by , if the joint CDF of and can be expressed as

 FT1,T2(t1,t2) =P(T1≤t1,T2≤t2) (5)

where

 ak=√(δk+1)tkδkμk,bk=√δkμk(δk+1)tk,k=1,2,

, , , , , , , and is the standard bivariate normal CDF with correlation coefficient . It follows that the joint PDF associated with (5) is given by

 fT1,T2(t1,t2) =ϕ2(√δ12(a1−b1),√δ22(a2−b2);ρ)2∏k=1√δk2√2tk(ak+bk), (6)

where is a normal joint PDF given by

 ϕ2(u,v;ρ)=12π√1−ρ2 exp(−12Q(u,v)),with Q(u,v)=1(1−ρ2)(u2−2ρuv+v2), u,v∈R.

Following the notation in (3) and considering , with and , for , note that

 FT1,T2(t1,t2)=Φ2(a1(t1),a2(t2);ρ),fT1,T2(t1,t2)=ϕ2(a1(t1),a2(t2);ρ)2∏k=1a′k(tk),

where , for

From now on, we will use the following notation

 cj,k(t,w;ρ)=1√1−ρ2[aj(t)−ρak(w)],t>0,w>0, (7)

for .

###### Lemma 3.1.

Some important properties of the function are:

1. is an increasing function for all .

If (), the function is increasing (decreasing) for all .

2. whenever , and .

3. whenever , and .

4. Let . For and , we have

 [1c1,2(t,v;ρ)−1c31,2(t,v;ρ)]c1,2(t,w;ρ)⩾1.
5. Let . For and , we have

 [1c1,2(t,w;ρ)−1c31,2(t,w;ρ)]c1,2(v,w;ρ)⩾1.

The joint PDFs and HRs of are unimodal and the surface plots for some values of the parameters are presented in Figure 2.

Next, some results on the unimodality properties of BRBS distribution are obtained. We will consider the following hypothesis:
Hypothesis 1. Let , and , for .

1. ,

2. ,  where  .

###### Proposition 3.1.

Under Hypothesis 1 there is an unique constant such that the point is critical for

###### Theorem 3.1 (Unimodality).

Under Hypothesis 1 there is an unique constant in the interval such that the point is a mode for That is, under Hypothesis 1 the BRBS distribution is unimodal.

### 3.2 Properties of the BRBS distribution

###### Proposition 3.2 (Marginal functions).

Let . The marginal PDFs, denoted , and the marginal CDFs, denoted , are given by

 fTk(tk)=f(tk;μk,δk),FTk(tk)=F(tk;μk,δk),k=1,2,

respectively, where and are the PDF and the CDF of the RBS distribution defined in (2). That is, .

###### Proposition 3.3 (Reliability function).

If  , the reliability can be expressed as

 R=E[Φ(c2,1(T1,T1;ρ))], T1∼RBS(μ1,δ1),

where is defined in (7).

###### Remark 3.1.

If and are identically distributed RVs following a RBS distribution, then

 R=E[Φ(√1−ρ1+ρ Z)],

where , and if in addition , then

###### Proposition 3.4.

Let . Then,

1. ;

where and are defined in (2) and (7), respectively.

Some authors like Basu (1971) or Puri and Rubin (1974) define the multivariate HR as a scalar quantity. In the bivariate case, Basu gives the HR as

 hT1,T2(t1,t2)=fT1,T2(t1,t2)S(t1,t2),whereS(t1,t2)=∫∞t1∫∞t2fT1,T2(w,t)dwdt

is the bivariate SF. An analogous procedure to Proposition 3.4 shows that

 (8)

Therefore, we get

 hT1,T2(t1,t2)=f(t1;μ1,δ1)ϕ(c2,1(t2,t1;ρ))∫∞t1f(w;μ1,δ1){1−Φ(c2,1(t2,w;ρ))}d% w. (9)
###### Proposition 3.5.

The function is decreasing in both and .

###### Proposition 3.6.

For the BRBS distribution,

1. if , and ; then the function is increasing in .

2. if , and ; then the function is increasing in .

3. if and ; then the function is increasing in .

###### Proposition 3.7.

Let and be the reliability. Then,

 ∫∞0hT1,T2(t1,t2=t1)dt1⩾R1−R.
###### Proposition 3.8.

Let , and . Then,

 E[a−⊥(X)]=βα22(σ2+b2)+β.

Let and be RVs, the covariance and correlation of and , as usual, are denoted by and respectively. The following result tells us that two RVs with BRBS distribution are associated and correlated positively.

###### Theorem 3.2.

Let , and , for . Then,

1. ,

Let be a bivariate positive RV with bivariate SF (defined in (8)) and with finite. If we assume that the sampling probability of is proportional to , the recurrence times are , where and , where and

have independent uniform distributions in

and are independent of . The vector is known as the size biased random vector and has PDF defined by

 fsb(x,y)=xyE[XY]f(x,y).

Under the previous hypothesis, the joint PDF of the bivariate equilibrium distribution is given by (see Navarro et al. (2006) or Navarro and Sarabia (2010))

 feq(x,y)=S(x,y)E[XY],x>0,y>0. (10)
###### Proposition 3.9 (Equilibrium distribution).

According to (10), the equilibrium PDF associated with (2) is

 feq(t1,t2)=∫∞t1f(w;μ1,δ1){1−Φ(c2,1(t2,w;ρ))}dwβ1β24[4+2(α21+α22)+α21α22(1+2ρ2)].
###### Proposition 3.10.

For the BRBS distribution, the function is decreasing in both and .

###### Theorem 3.3.

Let and , for . Then,

1. ;

2. ;

3. .

### 3.3 Failure rate with presence of dependence

A real function , which is defined on and , and are linearly ordered sets, is said to be totally positive of order two () and reverse rule of order two () in and if

 K(x,y)⩾0 (⩽0),x∈X, y∈Y,
 K(x,y)K(x′,y′)⩾K(x,y′)K(x′,y),for all  x

We also define a local dependence function (LDF)

 γK(x,y)=∂2∂x∂ylogK(x,y).

The LDF, , can be defined for any positive and mixed differentiable function , which does not need to be a density function. See Holland and Wang (1987) for definition and properties of the local dependence function.

###### Theorem 3.4.

(Holland and Wang, 1987) A function is if and only if .

It can be verified that

 γfT1,T2(t1,t2)=(ρ1−ρ2)2∏k=1a′k(tk)>0 (<0),t1>0,t2>0, (11)

whenever . Then, by Theorem 3.4 the joint PDF is an example of a function when . The LDF will be used for studying the monotonicity of certain HRs. If is a bivariate random vector distributed according to , then one can consider (the CDF of given that ). The conditional HR (CHR) of given and mean residual function (MRF), using obvious notation, are defined by

 h(x|Y∈A)=f(x|Y∈A)S(x|Y∈A),m(x|Y∈A)=∫∞xS(t|Y∈A)dtS(x|Y∈A),

respectively, where

 S(x|Y∈A)=1−F(x|Y∈A)

denotes the conditional SF (CSF).

We know a lot about the normal distribution. For example, if has a distribution we have (Feller (1968), Section 7.1)

 (1z−1z3)exp(−z2/2)√2π⩽P(Z>z)⩽1zexp(−z2/2)√2π,∀z>0. (12)

Clearly the inequalities are useful only for larger , because as decreases to zero the lower bound goes to and the upper bound goes to . The inequality above is known as the Gaussian tail inequality.

###### Theorem 3.5 (Monotonicity).

Let us assume that has a distribution and consider and for . Then,

1. The CHR: is decreasing for all and for all , whenever .

2. The MRF: is increasing for all and for all , whenever .

3. The CSF : is increasing for all and for all