TEXT SIZE

CrossRef (0)
How to Improve Classical Estimators via Linear Bayes Method?

Lichun Wanga

aDepartment of Mathematics, Beijing Jiaotong University, China
Correspondence to: Lichun Wang
Department of Mathematics, Beijing Jiaotong University, Beijing 100044, China. E-mail: lchwang@bjtu.edu.cn
Received October 17, 2015; Revised October 26, 2015; Accepted October 26, 2015.
Abstract

In this survey, we use the normal linear model to demonstrate the use of the linear Bayes method. The superiorities of linear Bayes estimator (LBE) over the classical UMVUE and MLE are established in terms of the mean squared error matrix (MSEM) criterion. Compared with the usual Bayes estimator (obtained by the MCMC method) the proposed LBE is simple and easy to use with numerical results presented to illustrate its performance. We also examine the applications of linear Bayes method to some other distributions including two-parameter exponential family, uniform distribution and inverse Gaussian distribution, and finally make some remarks.

Keywords : linear Bayes method, MCMC method, MSEM criterion, normal linear model, two-parameter exponential family, uniform distribution, inverse Gaussian distribution
1. Introduction

The linear Bayes method was originally proposed by Hartigan (1969), which suggests that in Bayesian statistics one can replace a completely specified prior distribution by an assumption with just a few moments of the distribution. It has been subsequently discussed by Rao (1973) from linear optimization viewpoint. Lamotte (1978) later develops a class of linear estimators, called Bayes linear estimators, by searching, among all linear estimators that have least average total mean squared error. Goldstein (1983) considers the problem of modifying the linear Bayes estimator for the mean of a distribution of unknown form using a sample variance estimate. Heiligers (1993) studies the relationship between linear Bayes estimation and minimax estimation in linear models with partial parameter restrictions. Hoffmann (1996) proposes a well-described subclass of Bayes linear estimators for the unknown parameter vector in the linear regression model with ellipsoidal parameter constraints and obtains a necessary and sufficient condition to ensure that the considered Bayes linear estimators improves the least squared estimator over the whole ellipsoid regardless of the selected generalized risk function. In the framework of empirical Bayes, Samaniego and Vestrup (1999) and Pensky and Ni (2000) respectively construct linear empirical Bayes estimators and establish their superiorities over standard and traditional estimators. In application fields, Busby et al. (2005) proposes the application of Bayes linear methodology to uncertainty evaluation in reservoir forecasting. Zhang and Wei (2005) also drive the unique Bayes linear unbiased estimator of estimable functions for the singular linear model. Wei and Zhang (2007) employs linear Bayes procedure to define Bayes linear minimum risk estimation in a linear model and discusses its superiorities. Recently, Zhang et al. (2011, 2012) extend the research on a linear Bayes estimator to the partitioned linear model and multivariate linear models, respectively.

In this paper, along the same line as in Wang and Singh (2014), we use the normal linear model as an example to demonstrate how to apply a linear Bayes method to simultaneously estimate all the parameters involved in the model and elaborate advantages and potential disadvantages.

Let W be a known p-dimension subspace of Rn. Suppose that we observe the random vector

Y~Nn(μ,σ2In),         μW,σ2>0.

This model is called the normal linear model as defined by Arnold (1980) and adopted by many other authors as well. Let X be a basis matrix for W. Then X is an n × p matrix of rank p, and there exists a unique βRp such that μ = . Hence, an equivalent version of the normal linear model can be presented, and we observe the random vector

Y~Nn(Xβ,σ2In),         βRp,σ2>0.

Define β̂ = (X′X)−1X′Y and σ̂2 = (||Y||2 – ||Xβ̂ ||2)/(np) and note that the fact that (β̂,σ̂2) is a complete sufficient statistic for the above linear model. Hence, the classical estimators for the parameters β and σ2 are β̂ and σ̂2, which are the uniformly minimum variance unbiased estimator (UMVUE) in the sense of minimizing mean squared error.

From the Bayesian viewpoint, note that in most cases past experience about the parameters β and σ2 are often available. Let f0(β,σ2) be the joint prior of β and σ2 and the loss function be

L(θ^,θ)=(θ^-θ)D(θ^-θ),

where D is a positive definite matrix and θ̂ denotes the estimate of the vector θ = (β′, σ2). Then, by virtute of the Bayes theorem, the usual Bayes estimators (UBE) for β and σ2, say β̂UB and σ^UB2, can be calculated by

β^UB=βg(β,σ2y)dβdσ2,σ^UB2=σ2g(β,σ2y)dβdσ2,

where g(β,σ2|y) denotes the conditional joint posterior density of β and σ2 given Y. However, it is difficult to handle complicated or non-standard integrations. Normally, in these cases approximate Bayes estimators are suggested such as Lindley’s approximation and Tierney and Kadane’s approximation, see Lindley (1980) and Tierney and Kadane (1986) for details. Simulation-based methods such as the Gibbs sampling procedure and Metropolis method also have emerged in the past twenty years, see Martinez and Martinez (2007). Traditional Bayes estimators (UBE) are somewhat complicated and inconvenient to use in these situations.

In the following, enlightened by Rao (1973), we employ the linear Bayes method to propose a linear Bayes estimator (LBE) for the parameters β and σ2 simultaneously as well as investigate superiorities. We also extend our discussions on the application of linear Bayes method to some other useful distributions.

The survey is organized as follows: In Section 2 we define the LBE for the parameter vector θ = (β′, σ2) and establish it superiorities over the classic UMVUE and MLE. Numerical comparisons between the LBE and the usual Bayes estimator (UBE) are presented in Section 3. Extended discussions and remarks are made in Section 4.

Throughout this paper, for two nonnegative definite matrices A1 and A2 of the same size, we say A1A2 if and only if A1A2 is a nonnegative definite matrix.

2. Linear Bayes Estimator and Its Superiorities

### 2.1. The proposed LBE

Denote θ = (β′, σ2). In what follows we assume that the prior G(θ) belongs to the distribution family:

G={G(θ):E[β2+(σ2)2]<}.

Put T = (β̂′, σ̂2) and define the linear Bayes estimator (LBE) of θ, say θ̂LB, be of the form θ̃ = BT + b satisfying

R(θ^LB,θ)=minB,bE(Y,θ)L(θ˜,θ)         and         E(Y,θ)(θ^LB-θ)=0,

where align and b are (p+1)×(p+1) and (p+1)×1 undetermined matrices respectively, E(Y,θ) denotes the expectation with respect to the joint distribution of Y and θ and the loss function is given by (1.3).

Thus, we have the following conclusion.

Theorem 1

Let θ̂LBbe defined by (2.2). If np + 1, then

θ^LB=T-W[W+Cov(θ)]-1(T-Eθ),

where W = E[Cov(T|θ)] = diag((X′X)−12, 24/(np)).

Proof

From the constraint E(Y,θ)(θ̃θ) = 0, we know b = BE(Y,θ)(T). Note that

E(Y,θ)(T)=E[E(Tθ)]=E(β,σ2)=Eθ.

Hence b = BEθ, and accordingly we have

R(θ˜,θ)=E(Y,θ)L(θ˜,θ)=E(Y,θ)[BT+Eθ-BEθ-θ]D[BT+Eθ-BEθ-θ]=E(Y,θ){tr (D[B(T-Eθ)-(θ-Eθ)][B(T-Eθ)-(θ-Eθ)])}=tr (DE(Y,θ)[B(T-Eθ)-(θ-Eθ)][B(T-Eθ)-(θ-Eθ)])=tr (DBE(Y,θ)[(T-Eθ)(T-Eθ)]B)-tr (DCov(θ)B)-tr(DBCov(θ))+tr(DCov(θ)).

For given θ, using the independence between β̂ and σ̂2, we have

E(Y,θ)[(T-Eθ)(T-Eθ)]=E[Cov(Tθ)]+Cov(E(Tθ))=W+Cov(θ),

where W = diag((X′X)−12, 24/(np)).

Substituting (2.5) into (2.4) and letting ∂R(θ̃, θ)/∂B be zero, we have

DB[W+Cov(θ)]-DCov(θ)=0,

which yields

B=Ip+1-W[W+Cov(θ)]-1.

Together with b = BEθ we come to the conclusion of Theorem 1.

Remark 1

In the definition of θ̂LB (2.2), if we discard the so-called unbiased constraint E(Y,θ)(θ̂LBθ) = 0, then by directly computing R(θ̃, θ) and denoting ∂R(θ̃, θ)/∂B = 0 and ∂R(θ̃, θ)/∂b = 0, we can obtain the same expression for the LBE θ̂LB, which means that θ̂LB satisfies the unbiased condition as well as performs best among linear Bayes estimators in the sense of minimizing E(Y,θ)L(BT + b, θ).

2.2. The superiorities of LBE

Note that

θ^U=(β^,σ^2)=(Ip001)(β^σ^2)=T,

where we use θ̂U to denote the UMVUE of θ = (β′, σ2).

### Theorem 2

Let θ̂LBand θ̂Ube given by (2.2) and (2.7) respectively. If np+1, then θ̂LBis superior to θ̂Uin terms of MSEM criterion, i.e. MSEM(θ̂LB)MSEM(θ̂U).

Proof

Since E(Y,θ)(θ̂LBθ) = 0, we have

MSEM (θ^LB)=E(Y,θ)[(θ^LB-θ)(θ^LB-θ)]=E[Cov (θ^LB-θθ)]+Cov (E[θ^LB-θθ]).

Denote M = [W + Cov(θ)]−1. Then by Theorem 1 we know

MSEM (θ^LB)=(I-WM)W(I-WM)+WMCov(θ)(WM)=(1-WM)W(I-MW)+WMCov(θ)MW=W-2WMW+WM[W+Cov(θ)]MW=W-WMW.

However,

MSEM (θ^U)=E(Y,θ)[(θ^U-θ)(θ^U-θ)]=E{E[(θ^U-θ)(θ^U-θ)θ]}=E{E[(T-θ)(T-θ)θ]}=W.

Comparing (2.9) with (2.10), we have

MSEM (θ^LB)MSEM (θ^U).

The proof of Theorem 2 is completed.

Moreover, note that the MLE of θ, denoted by θ̂ML, equals to B0T with

B0=(Ip00n-pn).

Thus,

MSEM (θ^ML)=B0WB0+(B0-Ip+1)E(θθ)(B0-Ip+1)=(XXEσ200n2(2n+p2-2p)Eσ4)-1.

### Theorem 3

Let θ̂LBand θ̂MLbe given by (2.2) and (2.12) respectively. If np + 1, then θ̂LBis superior to θ̂MLin terms of MSEM criterion, i.e. MSEM(θ̂LB)MSEM(θ̂ML).

Proof

We rewrite

MSEM (θ^LB)=W-WMW=[W-1+Cov-1(θ)]-1=[(XXEσ200n2(2n+p2-2p)Eσ4)+(000c0)+Cov-1(θ)]-1,

where c0 = (np2 – 4npp3 + 2p2)/{2(2n + p2 – 2p)4}.

Hence, in order to establish the MSEM superiority of θ̂LB over θ̂ML, it suffices to show that

(000c0)+Cov-1(θ)0,

for np + 1.

Denote Cov−1(θ) = S, where S = (Si j) is a 2 × 2 partition matrix and

S11=Cov-1(β)+Cov-1(β)E(σ2-Eσ2)(β-Eβ)S22E(σ2-Eσ2)(β-Eβ)Cov-1(β),S12=-Cov-1(β)E(σ2-Eσ2)(β-Eβ)S22,S21=-S22E(σ2-Eσ2)(β-Eβ)Cov-1(β),S22=[Var (σ2)-E(σ2-Eσ2)(β-Eβ)Cov-1(β)E(σ2-Eσ2)(β-Eβ)]-1.

Thus, to prove (2.15), it is adequate to show that

S11|c0+S22-S21(S11)-1S12|0,

or equivalently to show that

c0+S22-S21(S11)-1S120,

for np + 1, where we use the fact of S11 ≥ 0 and accordingly the value of the determinant |S11| is nonnegative.

Set Δ = Var(σ2) – E(σ22)(β)Cov−1(β)E(σ22)(β) and note that S11 = Cov−1(β)[Cov(β) + E(σ22)(β)(1/Δ)E(σ22)(β)]Cov−1(β), hence we have

c0+S22-S21(S11)-1S12=c0+1Δ-1ΔE(σ2-Eσ2)(β-Eβ)×[Cov(β)+E(σ2-Eσ2)(β-Eβ)1ΔE(σ2-Eσ2)(β-Eβ)]-1×E(σ2-Eσ2)(β-Eβ)1Δ.

Further, using [Σ+AΣ1A]-1=Σ-1-Σ-1A[AΣ-1A+Σ1-1]-1AΣ-1, we have

c0+S22-S21(S11)-1S12=c0+1Δ-aΔ2+a2(Δ+a)Δ2,

where a = E(σ22)(β)Cov−1(β)E(σ22)(β).

Note that Δ = Var(σ2) – a, hence

1Δ-aΔ2+a2(Δ+a)Δ2=1Var(σ2),

and accordingly

c0+S22-S21(S11)-1S12=c0+1Var (σ2)=[(n+4)p2-(4n+4)p-p3+4n]Eσ4+[4np-(n+2)p2+p3](Eσ2)22(2n+p2-2p)Var(σ2)Eσ4[2p2+4n-4p](Eσ2)22(2n+p2-2p)Var (σ2)Eσ4>0,         for   np+1,

where we use the facts that 4 ≥ (2)2 and (n + 4)p2 – (4n + 4)pp3 + 4n ≥ 0 for np + 1.

Hence, Theorem 3 has been proved.

Remark 2

For the two-parameter exponential family given by

f(x;μ,λ)=λ-1exp (x-μλ),

where x > μ, we assume that X(1)X(2) ≤ ··· ≤ X(r)(2 ≤ rn) denote the type II censored samples. Define Qi = [n – (i – 1)](X(i)X(i–1)), where X(0) = 0, then Q1 and P=i=2rQi are mutually independent and also (Q1, P) is sufficient for the parameter vector (μ, λ). Set T = (Q1, P), the classical UMVUE and MLE for θ = (μ, λ) can be defined as follows

θ^U=(1n-1n(r-1)01r-1)T,         θ^ML=(1n001r)T.

Under the assumption that the prior G(θ) satisfies the condition E||θ||2 < ∞, we can obtain the expression of LBE θ̂LB for the parameter vector θ = (μ, λ) in this case and establish its superiorities over the θ̂U and θ̂ML by virtue of MSEM criterion similarly. The interested readers are referred to Wang and Singh (2014) for more details.

Remark 3

Let X1, X2, . . . , Xn be independently drawn from the uniform distribution U(θ1, θ2) with density f (x; θ1, θ2) = (θ2θ1)−1, where θ1 < x < θ2. Note that X(1) = min1≤inXi and X(n) = max1≤inXi are sufficient and complete statistics, hence, set T = (X(1), X(n)), we obtain the classic UMVUE and MLE for θ = (θ1, θ2) in this case:

θ^U=(nn-1-1n-1-1n-1nn-1)T,         θ^ML=(1001)T.

Similarly, using the assumption that the prior G(θ) satisfies the condition E||θ||2 < ∞, the expression of LBE θ̂LB for the parameter vector θ = (θ1, θ2) can be easily obtained and its MSEM superiorities over the θ̂U and θ̂ML can also be proved.

Remark 4

Let X1, X2, . . . , Xn be a random sample from the two-parameter inverse Gaussian distribution IG(α1, α2) with pdf

f(x;α1,α2)=(α22πx3)12exp(-α2(x-α1)22α12x),

where x > 0. It is easily shown that the statistics X¯=(1/n)i=1nXi and S¯=n/{i=1n(1/Xi-1/X¯)} are sufficient and complete. Tweedie (1957) shows that and are independent, having an inverse Gaussian distribution with parameters α1 and 2, and 2/ having a χn-12 distribution. Schwarz and Samanta (1991) gives a proof of these facts using an inductive argument. Hence we obtain the classic UMVUE and MLE for the parameter θ = (α1, α2) as follows

θ^U=(100n-3n)T,         θ^ML=(1001)T,

where T = (, ). Assume that the prior G(θ) belongs to the prior family G={G(θ):E[α12+α22]<,E[α13α2-1]<}, we can obtain the expression of LBE θ̂LB for the parameter vector θ = (α1, α2) and prove that it prevails over the classic UMVUE and MLE under MSEM criterion.

2.3. An illustration example

To illustrate Theorem 2 and Theorem 3 we investigate the case of two-dimensional normal linear model, i.e.

Y~N(β0+β1x,σ2In),

where we assume that (β0, β1) ~ N((1, 2), Cov(β0, β1)) with Cov(β0, β1) having three alternative values and σ2 ~ U(a, b) with three different pairs a and b. We also assume that (β0, β1) and σ2 are uncorrelated, i.e. Cov(β0, β1, σ2) = diag(Cov(β0, β1), Var(σ2)), and x = (−4, −3, −2, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16).

Define the percentages of improvement of θ̂LB over θ̂U and θ̂ML, respectively, by

IP (θ^U)=tr (MSEM (θ^U)-MSEM (θ^LB))tr (MSEM (θ^U))         and         IP (θ^ML)=tr (MSEM (θ^ML)-MSEM (θ^LB))tr (MSEM (θ^ML)).

For different sample size n (= 5, 10, 20), the corresponding computation results for IP(θ̂U) and IP(θ̂ML) under the three different priors are presented in Table 1, where tr(Cov(β0, β1, σ2)) is used as an index of the variation of the prior information.

As stated in Theorems 2 and 3, since the above priors belong to the family (2.1), both MSEM(θ̂U)– MSEM(θ̂LB) and MSEM(θ̂ML) – MSEM(θ̂LB) are always nonnegative definite. From Table 1, firstly we can see that when the sample size n is fixed, as expected, both IP(θ̂U) and IP(θ̂ML) decrease as the variation of the prior increases (i.e, tr(Cov(β0, β1, σ2)) tends to be larger); secondly, for the same prior information, it is only natural that as the sample size n grows, which means that the sample information gets more, both IP(θ̂U) and IP(θ̂ML) decrease; finally, it seems that IP(θ̂U) is larger than IP(θ̂ML), the reason may be due to MSEM(θ̂U) ≥ MSEM(θ̂ML) for our case.

3. Numerical Comparisons between LBE and UBE

For the model (1.2), note that under the loss L(θ̂, θ) and the prior G(θ), the usual Bayes estimator (UBE) of θ, say θ̂UB, would be equal to E(θ|Y). In this Section, for given priors G(θ), we present some numerical comparisons between the LBE θ̂LB and the UBE θ̂UB, the latter is calculated by employing an MCMC sampling method.

Suppose p = 2 and let us consider the normal linear model

Y~N(γ0+γ1x,σ2In).

Denote β = (γ0, γ1). We assume that σ2 follows an inverse-Gamma distribution with density

π(σ2;λ0,t)=tλ0-1Γ(λ0-1)(1σ2)λ0exp (-tσ2)

and given σ2 the conditional distribution of β is N2(β̃0, σ20).

Note that the posterior density of (β,σ2) given Y is

f(β,σ2y)(σ-2)λ0+1+n2exp{-12σ2[2λ0+(n-2)σ^2+(β-β˜0)Σ0-1(β-β˜0)+(β-β^)XX(β-β^)]},

where X=(111x1x2xn) denotes the design matrix.

However, it is almost impossible to calculate θ^UB=(β^UB,σ^UB2) analytically, so we have to obtain it numerically.

Note that the posterior conditional densities of β given σ2 and σ2 given β are respectively proportional to

f1(βσ2,y)exp {-12σ2(β-β˜)Σ˜-1(β-β˜)},f2(σ2β,y)σ-2(λ0+1+n2)exp (-c12σ2),

where β˜=(Σ0-1+XX)-1(Σ0-1β˜0+XXβ^),Σ˜=σ2(Σ0-1+XX)-1 and c1=2λ0+(n-2)σ^2+(β-β˜0)Σ0-1(β-β˜0)+(β-β^)XX(β-β^).

The Gibbs sampler was originally developed by Geman and Geman (1984) as applied to image processing and the analysis of Gibbs distributions on a lattice. It is brought into mainstream statistics through the articles of Gelfand and Smith (1990) and Gelfand et al. (1990). The Gibbs sampler can also be shown to be a special case of the Metropolis-Hastings algorithm, see Gilks et al. (1996) and Robert and Casella (1999). In describing the Gibbs sampler, we follow the treatment in Casella and George (1992).

• Step 1. Choose the initial values of β and σ2 and denote the values of β and σ2 at the jth step by βj and σj2, respectively.

• Step 2. Generate βj+1 and σj+12 from f1(βσj2,y) and f2(σ2|βj, y), respectively.

• Step 3. Repeat Step 2 for N times.

• Step 4. Calculate the Bayes estimator of l(β,σ2) by 1/(N-m0)j=m0+1Nl(βj,σj2), where l(β,σ2) denotes any a function of β and σ2 and m0 is the burn-in period.

Note that under the above priors, it is readily seen that Eθ=(Eβ,Eσ2)=(β˜0,t/[λ0-2]) and Cov(θ) = diag(t/(λ0 – 2) · ∑0, t2/{(λ0 – 2)2(λ0 – 3)}) since Cov(β) = ECov(β|σ2) = ∑02 and Corr(β,σ2) = 0.

In the following table we first calculate the values of θ̂LB and the corresponding numerical results of θ̂UB for different prior parameters and then present θ^LB-θ^UB=β^LB-β^UB2+(σ^LB2-σ^UB2)2, which is defined as an index of degree of approximation between θ̂LB and θ̂UB.

The above numerical comparisons indicate two trends, one is that for the same prior, ||θ̂LBθ̂UB|| tends to be smaller as sample size gets larger, the other is that given sample size, ||θ̂LBθ̂UB|| increases as the prior variance becomes larger. In the process of simulation, we find that the value of ||θ̂LBθ̂UB|| is affected by the value of (σ^LB2-σUB2)2; however, the value of ||β̂LBβ̂UB||2 is always small, which means the LBE β̂LB is rather close to the UBE β̂UB and there could be a certain difference between the LBE σ^LB2 and the UBEσ^UB2 for our cases.

### Remark 5

Two cases are considered for the two-parameter exponential family. In case (I) we assume that the parameters μ and λ have independent prior distributions, where μ follows an exponential distribution and λ has an inverted Gamma prior. In case (II), we suppose that, given λ, the conditional prior of μ is an inverted Gamma density and λ follows an inverted Gamma prior. For the above two cases, numerical simulations show that ||θ̂LBθ̂UB||s are small, which means that as a linear approximation of θ̂UB, θ̂LB works better.

### Remark 6

In the case of the uniform distribution U(θ1, θ2), numerical computations show that θ̂LB works very good for both independent prior and non-independent prior. For example, for the single parameter uniform distribution U(0, θ2), we assume the prior π(θ2) has finite second-order moment and mimic the above discussions, then the LBE for the parameter θ2 is θ̂2,LB = a0X(n) + b0 with

a0=(n+1)(n+2)Var(θ2)(n+1)2Eθ22-n(n+2)(Eθ2)2         and         b0=[(1-a0)n+1]Eθ2n+1.

Specifically, let π(θ2)=t2t1θ2-t1-1exp(-t2/θ2)/Γ(t1) and together with f (x(n)|θ2) and the squared loss, we know that the UBE θ̂2,UB is

E(θ2x(n))=x(n)θ2-t1-nexp(-t2/θ2)dθ2x(n)θ2-t1-n-1exp(-t2/θ2)dθ2=t2t1+n-1P(χ2(2(t1+n-1))2t2/x(n))P(χ2(2(t1+n))2t2/x(n)),

where we utilize the relationship between the inverse Gamma and the χ2 distribution (Mao and Tang, 2012). Say, let n = 5, x(n) = 2 and t1 = 3 and t2 = 8, simple computations show that a0 = 1.1351, b0 = 0.2163 and P(χ2(14) ≤ 8) = 0.1107 and P(χ2(16) ≤ 8) = 0.0511. Hence, we have θ̂2,LB = 2.4865 and θ̂2,UB = 2.4758, which show that the LBE is very close to the UBE.

### Remark 7

For the two-parameter inverse Gaussian distribution IG(α1, α2), similarly, numerical studies show that LBE is adequate.

### Remark 8

In above simulation, it should be noted that the problem of deciding when to stop the chain is an important issue and is the topic of current research in MCMC methods. If the resulting sequence has not converged to the target distribution, then the estimators and inferences we get from it are suspect. Let γ represent the characteristic of the target distribution (mean, moments, quantiles, etc.) in which we are interested. An obvious method to monitor convergence to target distribution is to run multiple sequences of the chain and plot γ versus the iteration number.

4. Conclusions and Remarks

This paper uses the normal linear model Y ~ N(,σ2In) as an example to investigate the application of linear Bayes method, where we employ the linear Bayes method to simultaneously estimate regression parameter β and the variance parameter σ2 as well as propose a linear Bayes estimator for the parameter vector θ = (β′, σ2). The proposed linear Bayes estimator is shown superior to the classical estimators UMVUE and MLE, respectively, in terms of the mean squared error matrix criterion. Numerical simulations are presented to verify the validity of the linear Bayes estimator. The procedure used in this paper includes normal distribution as its special case and can be extended easily to other useful distributions (such as log-normal, inverse Gaussian distribution and two-parameter exponential family), which are frequently used parametric lifetime models in survival analysis and reliability theory. We also discuss and remark the applications of linear Bayes method to the two-parameter exponential family, uniform distribution and the inverse Gaussian distribution. Compared with the usual Bayes estimator, we find that

• The proposed linear Bayes estimator is simple and easy to calculate as well as a good approximation in many situations; the linear Bayes method works especially well for the case of uniform distribution.

• We can always define a linear Bayes estimator if there exists sufficient statistic for the parametric model; subsequently, the conclusions of Theorems 2 and 3 always hold.

However, an advantage of the usual Bayes estimator over the linear Bayes estimator is that the former allows for noninformative (improper) priors. Of note is that the linear Bayes estimator may be an inadequate approximation in some situations even for the cases of proper priors. Hence there is still scope for the linear Bayes method to be improved. For instance, for some cases a quadratic Bayes estimator would be a better alternative. However, for the case of normal linear model, one can consider to add more other statistics into the definition of T, for example, we can replace T = (β̂′,σ̂2) by T1 = (β̂′, ||β̂||2, σ̂2) or T2 = (β̂′, σ̂2, σ̂4) to redefine a new linear Bayes estimator. We note that the loss function often plays an important role in Bayesian analysis; consequently, some interesting loss functions such as the balanced loss and the linex loss can be integrated with the linear Bayes method in future studies.

TABLES

### Table 1

IP(θ̂U) and IP(θ̂ML) under different priors and sample sizes

PriorsnIP(θ̂U)IP (θ̂ML)tr (Cov (β0, β1, σ2))
(β0β1)~N((12),(113131))50.97120.9542
100.95160.94047/3
σ2 ~ U(7, 9)200.90660.8974

(β0β1)~N((12),(6226))50.92340.8779
100.88030.852440/3
σ2 ~ U(6, 10)200.77060.7480

(β0β1)~N((12),(186618))50.84580.7537
100.72650.6627124/3
σ2 ~ U(4, 12)200.53690.4961

### Table 2

||θ̂LBθ̂UB|| under different prior parameters and sample sizes

nThe prior parameters||θ̂LBθ̂UB||
λ0 = 2, t = 3, β˜0=(-23),Σ0=(10.70.71)1.3369
20λ0 = 2, t = 4, β˜0=(-23),Σ0=(53.53.55)1.4472
λ0 = 2, t = 8, β˜0=(-23),Σ0=(20141420)1.8390

λ0 = 2, t = 3, β˜0=(-23),Σ0=(10.70.71)1.0589
50λ0 = 2, t = 4, β˜0=(-23),Σ0=(53.53.55)1.1407
λ0 = 2, t = 8, β˜0=(-23),Σ0=(20141420)1.4855

Where (x1, x2, . . . , x20) is a subset of (x1, x2, . . . , x50) = (−4 3.4 2.4 0 1 2 3 4 3.5 0.6 7 8 9 10 11 12 −1.9 14 15 16 17 18 19 21 22 23 34 35 −13 17 18.5 19.9 24 28 32 33 37 39 −12 −16 −19 44 45 23.4 31.7 33.5 45.2 60.7 −14.3 −17).

References
1. Arnold SF (1980). The Theory of Linear Models and Multivariate Analysis, New York, John Wiley & Sons.
2. Busby D, Farmer CL, and Iske A (2005). Uncertainty evaluation in reservoir forecasting by Bayes linear methodology. Algorithms for approximation, proceedings of the 5th international conference Chester. , 187-196.
3. Casella G and George EI (1992). An introduction to Gibbs Sampling. The American Statistician, 46, 167-174.
4. Gelfand AE and Smith AFM (1990). Sampling-based approaches to calculating marginal densities. Journal of the American Statistical Association, 85, 398-409.
5. Gelfand AE, Hills SE, Racine-Poon A, and Smith AFM (1990). Illustration of Bayesian inference in normal data models using Gibbs sampling. Journal of the American Statistical Association, 85, 972-985.
6. Geman S and Geman D (1984). Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images. IEEE Transactions PAMI, 6, 721-741.
7. Gilks WR, Richardson S, and Spiegelhalter DJ (1996). Markov Chain Monte Carlo in Practice, London, Chapman and Hall.
8. Goldstein M (1983). General variance modifications for linear Bayes estimators. Journal of the American Statistical Association, 78, 616-618.
9. Hartigan JA (1969). Linear Bayesian methods. Journal of the Royal Statistical Society, Series B, 31, 440-454.
10. Heiligers B (1993). Linear Bayes and minimax estimation in linear models with partially restricted parameter space. Journal of Statistical Planning and Inference, 36, 175-183.
11. Hoffmann K (1996). A subclass of Bayes linear estimators that are minimax. Acta Applicandue Mathematicae, 43, 87-95.
12. Lamotte LR (1978). Bayes linear estimators. Technometrics, 3, 281-290.
13. Lindley DV (1980). Approximate Bayesian methods. Trabajos de Estadistica, 21, 223-237.
14. Mao SS and Tang YC (2012). Bayesian Statistics (Second Edition), Beijing, China Statistics Press.
15. Martinez WL and Martinez AR (2007). Computational Statistics Handbook with MATLAB (Second Edition), New York, Chapman & Hall/CRC.
16. Pensky M and Ni P (2000). Extended linear empirical Bayes estimation. Communications in Statistics - Theory and Methods, 29, 579-592.
17. Rao CR (1973). Linear Statistical Inference and Its Applications (Second Edition), New York, John Wiley & Sons.
18. Robert CP and Casella G (1999). Monte Carlo Statistical Methods, New York, Springer-Verlag.
19. Samaniego FJ and Vestrup E (1999). On improving standard estimators via linear empirical Bayes methods. Statistics & Probability Letters, 44, 309-318.
20. Schwarz CJ and Samanta M (1991). An inductive proof of the sampling distributions for the MLEs of the parameters in an inverse Gaussian distributions. The American Statistician, 45, 223-235.
21. Tierney L and Kadane JB (1986). Accurate approximations for posterior moments and marginal densities. Journal of the American Statistical Association, 81, 82-86.
22. Tweedie MCK (1957). Statistical properties of inverse Gaussian distributions. The Annals of Mathematical Statistics, 28, 362-377.
23. Wang LC and Singh RS (2014). Linear Bayes estimator for the two-parameter exponential family under type II censoring. Computational Statistics and Data Analysis, 71, 633-642.
24. Wei LS and Zhang WP (2007). The superiorities of Bayes linear minimum risk estimation in linear model. Communications in Statistics-Theory and Methods, 36, 917-926.
25. Zhang WP and Wei LS (2005). On Bayes linear unbiased estimation of estimable functions for the singular linear model. Science in China Series A Mathematics, 7, 898-903.
26. Zhang WP, Wei LS, and Chen Y (2011). The superiorities of Bayes linear unbiased estimation in partitioned linear model. Journal of System Science and Complexity, 5, 945-954.
27. Zhang WP, Wei LS, and Chen Y (2012). The superiorities of Bayes linear unbiased estimation in multivariate linear models. Acta Mathematicae Applicatae Sinica, 383-394.