Note on the EM Algorithm in Linear
Regression Model
Ji-Xia Wang and Yu Miao
College of Mathematics and Information Science Henan Normal University
Henan Province, 453007, China [email protected]
Abstract
Linear regression model has been used extensively in the fields of information processing and data analysis. In the present paper, we con-sider the linear model with missing data. Using the EM (Expectation and Maximization) algorithm, the asymptotic variances and the stan-dard errors for the MLE of the unknown parameters are established. Mathematics Subject Classification: 93C05; 93C41
Keywords: Conditional expectation; maximum likelihood estimator; EM algorithm; Newton-Raphson iteration
1
Introduction
As a typical statistical model, linear regression model has been widely used in the fields of information processing and data analysis. In fact, there have been several statistical methods for its learning or modeling (e.g., the expectation-maximization (EM) algorithm [2] for maximum likelihood and the self-organizing network with hyper-ellipsoidal clustering [5]). Generally, the parameters of lin-ear regressive model can be estimated via the EM algorithm under the maxi-mum likelihood framework, since the EM algorithm owns certain good conver-gence behaviors in certain situations. However, in some applications, there are many data sets including missing observations [9], which cause many problems if the missing data is related to the values of the missing item [8], for instance, in [4], Little and Rubin showed that this can cause bias and inefficiency for some estimations. So, an new algorithm for estimating unknown parameters is proposed based on the likelihood function. In [1], Baker and Laird used the
EM algorithm to obtain maximum likelihood estimates (MLE) of the unknown parameters in the model with the incomplete data. Ibrahim and Lipsitz [3] established Bayesian methods for estimation in generalized linear models.
In the present paper, we discuss the linear regression model with miss-ing data and propose a method for estimatmiss-ing parameters by usmiss-ing Newton-Raphson iteration to solve the score equation. Moreover, the standard errors of these estimators are calculated by the observed Fisher information matrix.
2
Linear regression model with missing data
Suppose thaty1, y2, . . . , ynare independent identically distributed normal ran-dom variables with unit variances. Let Xi = (X1i, X2i)T is a 2×1 random vec-tor of covariation, where X1i and X2i are independent observations and follow normal distributions with means μ1, μ2 and variances σ12, σ22, respectively. For notation convenience, letXi = (1, X1i, X2i)T and assume thatβT = (β0, β1, β2) are regression coefficients. It is also supposed that
p(yi|Xi, β) = √1 2πexp
⎧ ⎪ ⎨ ⎪
⎩−
yi−XiTβ
2 2
⎫ ⎪ ⎬ ⎪
⎭. (1)
We assume that X1i is completely observed, and X2i is partially missing for every i and our objective is to estimate β, μ1, μ2, σ12, σ22 and their standard errors from the known data with missing values.
Missing value indicators are introduced in [6] as
ri =
0, if yi is observed, 1, if yi is missing. si =
0, if x2i is observed,
1, if x2i is missing. (2) with probabilities p(ri) = ψi, p(si) = ϕi. Following the reference [8], for any
i= 1,2, . . . , n, the missing-data mechanism is defined as logit(ψi)log ψi
1−ψi =δ1X1i+δ2X2i +yiω (3) and
logit(ϕi)log ϕi
1−ϕi =α1X1i+α2X2i+yiτ, (4) where δ = (δ1, δ2)T, α = (α1, α2)T, ω and τ are parameters determining the missing mechanism. Then the conditional probability functions for ri and si are derived by Eqs. (2)-(4) as
p(ri|Xi, yi, δ, ω) = exp{ri(X T
i δ+yiω)} 1 + exp{XTδ+yiω},
p(si|Xi, yi, α, τ) = exp{si(X T
i α+yiτ)} 1 + exp{XiTα+yiτ}. Now we derive the joint probability function of yi, x2i, ri, si as
p(yi, x2i, ri, si|x1i)
=p(ri|Xi, yi, δ, ω)p(si|Xi, yi, α, τ)p(yi|Xi, β)p(x2i|X1i) ∝exp{ri(XiTδ+yiω)}
1 + exp{XiTδ+yiω}×
exp{si(XiTα+yiτ)}
1 + exp{XiTα+yiτ} ×(2π) −1
2
×exp
−(yi−Xi T
β)2 2
×(2πσ22)−12 ×exp
−(x
Therefore, we can write down the complete-data log-likelihood l(θ) by log L(θ|yi, Xi, ri, si)
= n
i=1 log
exp{ri(XiTδ+yiω)} 1 + exp{ri(XiTδ+yiω)}
+ n
i=1 log
exp{ri(XiTα+yiτ)} 1 + exp{si(XiTα+yiτ)}
+n
2 log(2π)− n
i=1
yi−XiTβ
2
2 −
n
2log(2πσ 2 2)−
n
i=1
(x2i−μ2)2 2σ22 ,
where θ = (β, δ, ω, α, τ, μ2, σ22) is the parameter related to developing EM al-gorithm. The complete-data log-likelihood specifies a model for the joint char-acterization of the observed data and the associated missing-data mechanism.
3
E-step of EM algorithm
The MLE ofθis a point which maximizes the observed-data likelihood function
L(θ|(y, X)obs, ri, si), where (y, X)obs is the observed components of (y, X). Let
θ(r) be ther-st iteration estimate ofθ and define the conditional expectation of
l(θ)-with respect to the conditional distribution of the missing data (y, X)mis given the observed data yi, Xi, ri, si and the value θ(r) as the following:
Q(θ|θ(r)) =E[l(θ)|(y, X)obs, r, s, θ(r)]. (5) The EM algorithm is composed of E-step and M-step iterations. Now for the expectation of the complete-data log-likelihood in the E-step of EM algorithm, we consider four possible-cases: response variable yi is missing, a covariance x2i is missing, both of them are missing, and no missing values. Then the expected log-likelihood function can be written by
(6) =
where x2i,mis denotes the missing components of x2i. Eqs.(3.1) and (3.2) lead to the conditional expectation of l(θ), which is our target quantity as
Q(θ|θ(r)) = n1
i=1
l(θ) + n2
i=n1+1
l(θ)pyi,mis|Xi, ri, si, θ(r)dyi,mis
+ n3
i=n2+1
l(θ)px2i,mis|Xi,obs, yi, ri, si, θ(r)dx2i,mis
+ n
i=n3+1
∞
yi=1
l(θ)pyi,mis, x2i,mis|Xi,obs, ri, si, θ(r)dyi,misdx2i,mis
where n1, n2, n3 are corresponding sample sizes, yi,mis is the missing compo-nents ofyi,Xi,obsis the observed component ofXi, andp(yi,mis, x2i,mis|Xi,obs, ri, si),
p(yi,mis|Xi, ri, si) and p(yi,mis, x2i,mis|Xi,obs, ri, si) are the conditional ities of the missing data given the observed data. These conditional probabil-ities are regarded as the weights in Q(θ|θ(r)). The weights have the following form:
pyi,mis, x2i,mis|Xi,obs, ri, si, θ(r)
= p
yi|Xi, θ(r)p(x2i|x1i)pri|yi, Xi, θ(r)psi|yi, Xi, θ(r)
∞
y1=1
p(yi|Xi, θ(r))p(x2i|x1i)p(ri|yi, Xi, θ(r))p(si|yi, Xi, θ(r)) ∝ pyi, x2i, ri, si|x1i, θ(r),
px2i,mis|Xi,obs, yi, ri, si, θ(r) = p
x2i|x1i, θ(r)psi|yi, Xi, θ(r)
p(x2i|x1i, θ(r))p(si|yi, Xi, θ(r)) ∝exp{ri(XiTα+yiτ)}
1 + exp{XiTα+yiτ}×(2πσ 2 2)−
1
2 ×exp
−(x2i−μ2)2 2σ22
,
and
pyi,mis|Xi, ri, si, θ(r)= p
yi|Xi, θ(r)pri|yi, Xi, θ(r)
∞
yi=1p(yi|Xi, θ(r))p(ri|yi, Xi, θ(r))
∝ pyi|Xi, θ(r)pri|yi, Xi, θ(r).
Metropolis-4
M-step of EM algorithm and convergence
Now we need to find a value of θ, saying θ(r), at which Q(θ|θ(r)) will attain the maximum. The Newton-Raphson method will be used to solve the score equation. The parameters θ(r+1) in the M-step at the (r+ 1)st EM iteration and the (r+ 1)stNewton-Raphson iteration take the following form (forβ for example):β(r+1)=β(r)+
−∂2Q(θ|θ(r))
∂β∂βT
−1
β=β(r) ×
∂Q(θ|θ(r))
∂β β=β(r).
The derivatives of the parameter β used in the iteration are given as follows:
∂Q(θ|θ(r))
∂β = n1 i=1 Xi
yi−XiTβ
+ n2
i=n1+1
E
Xi
yi−XiTβ
|Xi, θ(r)
+ n3
i=n2+1
E
Xi
yi−XiTβ
|Xobs, yi, θ(r)
+ n
i=n3+1
E
Xi
yi−XiTβ
|Xobs, θ(r)
,
and
∂2Q(θ|θ(r))
∂β∂βT =
n1
i=1
XiTXi+ n2
i=n1+1
E
XiTXi|Xi, θ(r)
+ n3
i=n2+1
E
XiTXi|Xobs, yi, θ(r)
+ n
i=n3+1
E
XiTXi|Xobs, θ(r)
.
The derivatives of other components of β used in the iteration are given in the reference [6].
The (r+1)stestimates ofμ2, σ22are obtained by solving the score equations:
∂Q(θ|θ(r))
∂μ2 = n
i=1
E(x2i|x1i, yi, ri, si)−nμ2 = 0,
∂Q(θ|θ(r))
∂σ22 =
n
i=1
E(x2i−μ2)2|x1i, yi, ri, si−nσ22 = 0.
Therefore, we can take μ(r+1)2 , σ22(r+1) by
μ(r+1)2 = 1
nE(x2i|x1i, yi, ri, si), σ
2(r+1)
2 =
1
nE
which are approximated by the sample averages of simulated and given obser-vations.
The sequence{Q(θ|θ(r))}often exhibits an increasing trend, and then fluc-tuate around the value of Q(θ|θ(r)) if r becomes large enough. The sequence {θ(r)}would also fluctuate the MLEθ(r)whenris sufficiently large. To monitor the convergence of the EM algorithm we can plot {Q(θ|θ(r))}as well as {θ(r)} against iteration number. We terminate the algorithm when the sequence of {Q(θ|θ(r))} become stationary. Otherwise, we continue by increasing the Monte Carlo precision in the E-step provided calculation is computationally feasible.
5
Standard errors of estimates
It is well know that the distribution of maximum likelihood estimates ˆθ asymp-totically tends to a normal distribution MV N(θ, V(θ)) under some regularity conditions. The expected Fisher information matrix I(ˆθ) which gives the in-verse of variance matrix of ˆθ is approximated by the observed information matrix Jθˆ(Y):
V(ˆθ)−1 =nE
−∂2logL(θ)
∂θ2 θ=ˆθ ∝n −
∂2logL(θ)
∂θ2
dx
≈ n
i=1
−∂2logL(θ)
∂θ2 θ=ˆθ ≈nJ(ˆθ).
By using the following relation which is obtained in [9]: observed
informa-tion=complete information-missing information, we have
I(ˆθ)≈Jθˆ(Y) =−∂
2logL(θ)
∂θ2 =
−∂2Q(θ|θ(r))
∂θ2 −V arθ
n
i=1
∂logL(θ)
∂θ
θ=ˆθ, where V ar(·) is the conditional variance given (y, X)obs, r, s, and θ(r). The details are to be provided in the reference [6].
ACKNOWLEDGEMENTS.
The authors acknowledge the financial support of the Foundation for Dis-tinguished Young Scholars of Henan Province (084100510013).
References
[1] S. G. Baker and N. M. Laird,Regression analysis for categorical variables with outcome subject to nonignorable nonresponse, J. Am. Stat. Assoc,
[2] A. P. Dempster, N. M. Laird and D. B. Rubin, Maximum likelihood from
incomplete data via the EM algorithm. J.Royal Stat. Soc. B, 1977, 39:
1-38.
[3] J. G. Ibrahim, S. R. Lipsitz, Missing covariates in generalized linear mod-els when the missing data mechanism is non-ignorable, J. Royal Stat. Soc. B, 1999, 61: 173-190.
[4] R. J. A. Little and D. B. Rubin, Statistical Analysis with Missing Data, New York, Wiley, 2002.
[5] J. Mao and A. K. Jain, A self-organizing network for hyperellipsoidal
clustering, IEEE Trans. Neural Networks, 1996, 7(1): 16-29.
[6] J. S. Park, G. Q. Qian and Y. Jun,Monte Carlo EM algorithm in logistic linear models involving non-ignorable missing data, Appl. Math. Comput., 2008,197: 440-450.
[7] C. P. Robert and G. Casella,Monte Carlo Statistical Methods, Berlin: Springer, 1999.
[8] M. M. Rueda,S. Gonzalez and A. Arcos,Indirect methods of imputation of
missing data based on available units, Appl. Math. Comput., 2005, 164:
249-261.
[9] Y. G. Smirlis and E. K. Despotis, Data envelopment analysis with missing
values: An interval DEA approach, Appl. Math. Comput., 2006, 177:
1-10.