Journal of Statistical Theory and Applications

Volume 19, Issue 2, June 2020, Pages 118 - 132

Robust Mixture Regression Based on the Mixture of Slash Distributions

Authors
Hadi Saboori1, *, Ghobad Barmalzan1, Mahdi Doostparast2
1Department of Statistics, University of Zabol, Sistan and Baluchestan, Iran
2Department of Statistics, Ferdowsi University of Mashhad, Khorasan Razavi, Iran
*Corresponding author. Email: h.saboori@uoz.ac.ir
Corresponding Author
Hadi Saboori
Received 17 June 2017, Accepted 18 June 2019, Available Online 9 April 2020.
DOI
10.2991/jsta.d.200304.001How to use a DOI?
Keywords
EM algorithm; Normal mixture regression; Outliers; Slash distribution
Abstract

The traditional estimation of Mixture regression models is based on the normal assumption of component errors and thus is sensitive to outliers and heavy-tailed errors. In this paper, we propose a robust Mixture regression models in which a mixture of slash distributions is assumed for the errors. Using the fact that the slash distribution can be written as a scale mixture of a normal and a latent distribution, we also estimate model parameters an expectation-maximization (EM) algorithm. The results of our simulation show that based on AIC and BIC criterion, the proposed robust regression model mixture on slash distribution dominates the robust regression based the normal and the t distribution. Finally, the proposed model is compared with other procedures, based on a real data set.

Copyright
© 2020 The Authors. Published by Atlantis Press SARL.
Open Access
This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

1. INTRODUCTION

Mixture models are applied in such diverse areas as biology, biometrics, genetics, medicine and marketing, among others. For a brief list of application of the MRM, see Table 1. There are various features of (finite) mixture distributions that make them useful in statistical modeling. For instance, statistical models on the base of finite mixture distributions can capture various properties of real data sets such as multimodality, skewness, kurtosis and unobserved heterogeneity. For more details, see [15].

Area of Application of MRM References
Introduced by [6]
Marketing [79]
Economics [10,11]
Agriculture [12]
Nutrition [13]
Psychometrics [14]
Table 1

Various application of the mixture regression model (MRM).

Mixture regression models (hereafter MRM), have been widely used to investigate the relationship between variables coming from several unknown latent homogeneous groups.

The model setting for mixtures of linear regression models can be stated as follows: Let Z be a latent class variable such that given Z=j, the (univariate) response Y depends on the p-dimensional predictor vector x=(x1,,xp) as

Y=xβj+ϵj,j=1,,m,(1)
where m is the number of groups (also called components in mixture models) in the population, the βj are unknown p-dimensional vectors of regression coefficients and ϵj~N(0,σj2) is independent of x. Suppose P(Z=j)=πj and Z is independent of x, then the conditional density of Y given x can be written as
ψ(y;x,δ)=j=1mπjϕ(y;xβj,σj2).(2)

The log-likelihood function (LLF) for observations (x1,y1),,(xn,yn) is

i=1nlog[j=1mπjϕ(yi;xiβj,σj2)],(3)
where ϕ(;μ,σ2) is the density function of the normal distribution with mean μ and variance σ2, hereafter denote by N(μ,σ2) and δ=(δ1,,δm) with δj=(πj,βj,σj2).

The computation can be further simplified based on the fact that the slash distribution can be considered as a scale mixture of normal distributions. Let U be the latent variable such that

ϵ|u~N(0,σ2u2)andU~Beta(q,1),
where corresponding density function to Beta(α,β) is
h(u;α,β)=b(α,β)uα1(1u)β1,0<u<1,
and b(α,β) is the Beta function. Then, marginally ϵ has a slash distribution with parameter q and scale parameter σ.

The unknown parameters in the model (2) can be estimated maximizing the LLF (3). Since the LLF (3) complicated, the maximum likelihood estimations (MLEs) are derived by maximizing (3) numerically. Among numerically methods for maximizing (3), the EM algorithm is commonly used; see [15].

The MLE δ̂ in (3) works well when the error distribution is normal. However, the normality based MLE is sensitive to outliers or heavy-tailed error distributions. To overcome this problem [16] and [17] proposed a weight factor for each data point to robustify the estimation procedure for MRM. Neykov et al. [18] proposed robust fitting of mixtures using the trimmed likelihood estimator (TLE). Bai et al. [19] proposed a modified EM algorithm to robustly estimate the mixture regression parameters by replacing the least squares criterion in M-step with a robust criterion. Through a simulation study, they show that their proposed estimate is robust when the data have outliers or the error distribution has heavy tails. Bashir and Carter [20] extended the idea of the S-estimator to mix linear regression models.

Yao et al. [21] proposed a robust estimation procedure for mixture linear regression models by assuming that the error terms follow the t Student distribution. Using the skew-normal distribution defined by [22,23] proposed a version of the mixture linear regression model. Dogru and Arslan [24] also proposed a new version of MRM modeling using the least trimmed squares (LTSs) estimation method.

Recently, researchers proposed the MRM based on skew and/or heavy tailed distributions. Lin et al. [25,4] respectively, using skew t and multivariate skew t-Normal distributions for this regression model. Song et al. [26] use the Laplace distribution. Zeller et al. [27] proposed a robust MRM based on the scale mixtures of skew-normal distributions. These authors also developed a simple EM-type algorithm to perform maximum likelihood inference of the parameters of the proposed model. In order to examine the robust aspect of this flexible model against outlying observations, some simulation studies have also been proposed. For example [28] proposed MRM using the mixture of different distributions. Dogru and Arslan [29,30] used other skew distribution for this regression model.

In recent years, attempts have been made by many researchers to propose flexible statistical models for fitting various data sets. Among the well-known distribution functions, the slash distribution, introduced by [31], provides more flexibility. See Section 2.

In this paper, we propose a robust MRM based on the slash distribution by assuming the mixture of slash distributions errors. Similar to the traditional M-estimate for linear regression (see [32]), the proposed estimate is expected to be sensitive to high leverage outliers. To overcome this problem, we also propose a modified version of the new method by fitting the new model to the data after adaptively trimming high leverage points. Compared to the TLE, the proportion of trimming of the new proposed model is data adaptive instead of a fixed value. Using a simulation study and real data application, we compare the new method to some existing methods, and demonstrate the effectiveness of the proposed method.

Therefore, the rest of this paper is organized as follows: In Section 2 the slash distribution and its properties are reviewed. In Section 3, we introduce our new robust mixture linear regression models based on the slash distribution and then develop the respective EM algorithm. In Section 4, we propose to further improve the robustness of the proposed method against high leverage outliers by adaptively trimming high leverage points. Also in this section, with simulation study, we compare the proposed method to the traditional MLE and some other robust methods by using a simulation study. Finally in Section 5, these results have been applied in real data.

The slash distribution with parameter q is defined as the ratio of two independent random variables Z and U1q, where Z is the standard normal random variable and U has uniform distribution in (0,1). The main feature of the slash distribution is to have greater flexibility in the degree of kurtosis with respect to the normal distributions. Other advantages of this distribution are the properties of symmetry, heavy tails and the inclusion of the normal family as a limit case when q tends to infinity. Also as this density is less peaked, in the center, than normal density, it is more realistic in representing data. For various properties of the slash distribution and its applications, we refer the interested reader to [33,34] and [35]. Also, Alcantara and Cysneiros [36] and [36] proposed a linear regression model with slash errors.

2. THE SLASH DISTRIBUTION

In this section, in order to more robustly estimate the mixture regression parameters, we assume that the error density function in (1) is a slash distribution with parameter qj>0 and scale parameter σj>0, as following:

f(ϵj;σj,qj)=qjσj01uqjϕuϵjσjdu,ϵj,j=1,,m.(4)

The MRM with slash distribution can be formulated in a similar way to the model defined in (2) as follows:

g(y;x,Θ)=j=1mπjf(yxβj;σj,qj),(5)
where f(;σ,q) is the density function of the slash distribution given by (4) and Θ=(θ1,,θm) with θj=(βj,σj,πj). The unknown parameter vector Θ in (5), given observations (x1,y1),,(xn,yn), is
Θ̂=argmaxΘi=1nlogj=1mπjfyixiβj;σj,qj.(6)

Note that the LLF (6) is unbounded if one of the observations yi is equal to xiβj and σj0. For more details, see [37,38] and [39]. Also notice that, the maximizer of (6) does not have an explicit solution, so we propose to use an EM-type algorithm (see [15]). For a gentle tutorial on the EM algorithm and its applications to parameter estimation for mixture models, see [40] and [41].

3. MAXIMUM LIKELIHOOD ESTIMATION

In this subsection, we present an EM algorithm for the ML estimation of the MRM with the slash distribution given by (5). For j=1,,m, and i=1,,n, denote zij as latent Bernoulli variables such that

zij=1,if the ith observation is from the jth component,0,otherwise.

If the complete data set T=(xi,yi,zij);i=1,,n,j=1,,m is observable, the complete LLF of Θ can be written as

c(Θ;T)=i=1nj=1mzijlogπjfyixiβj;σj,qj.(7)

Based on the theory of EM algorithm, in E-step, given current estimate Θ(k) at the kth step, we calculate Ec(Θ;T)|X,y,Θ(k) which simplifies to the calculation of EZij|X,y,Θ(k). Then in M-step, we find the maximizer of

Ec(Θ;T)|X,y,Θ(k)=i=1nj=1mEZij|X,y,Θ(k)logπjf(yixiβj;σj,qj).

Note that the above maximizer does not have explicit solutions for βj,σj and qj.

Since, we can simplify the computation of M-step of the proposed EM algorithm by introducing another latent variable U, therefore the complete likelihood for (y,u,z) given X is

c(Θ;X,y,z,u)=i=1nj=1mzijlogπjϕyixiβj;0,σj2/uij2huij;qj,1=i=1nj=1mzijlogπj+i=1nj=1mzijlogqj+i=1nj=1mzijqj1loguij+i=1nj=1mzij12log2πσj2+loguijuij22σj2yixiβj2,(8)
where u=(u11,,unm) is independent of z=(z11,,znm).

In the E-step on the (k+1)th iteration, we need to calculate the conditional expectation of the LLF of complete data, which is Ec(Θ;X,y,z,u)|X,y,Θ(k). Based on the above argument, the E-step requires the calculations of E(Zij|X,y,Θ(k)),E(Uij2|X,y,zij=1,Θ(k)) and ElogUij|X,y,zij=1,Θ(k). Thus, the EM algorithm can be written as

  1. Choose some initial value Θ(0)=(β1(0),q1(0),σ1(0),π1(0),,βm(0),qm(0),σm(0),πm(0)).

  2. E-step: On the (k+1)th iteration, according to Bayes theorem, we can compute conditional expectations as follows:

    pij(k+1)=EZij|X,y,Θ(k)=πj(k)fyixiβj(k);σj(k),qj(k)l=1mπl(k)fyixiβl(k);σl(k),ql(k),(9)
    lijk+1=ElogUij|X,y,zij=1,Θk=01loguqjkuqjkexpu2yixiβjk22σjk2σjk2πfyixiβjk;σjk,qjkdu,(10)
    and
    eij(k+1)=EUij2|X,y,zij=1,Θ(k)=qj(k)2qj(k)2(yixiβj(k))2(σj(k))2qj(k)+32Γqj(k)+32σj(k)πf(yixiβj(k);σj(k),qj(k))G(yixiβj(k))22(σj(k))2;qj(k)+32,11,yixiβj(k)0,qj(k)qj(k)+31σj(k)2πf(0;σj(k),qj(k)),yixiβj(k)=0,
    where Γ() and G(;r,1) are the complete gamma function and the cumulative distribution function of the gamma distribution with parameters shape r and scale 1, respectively.

  3. M-step: On the (k+1)th iteration, compute the estimator of parameters which maximize the expected complete log-likelihood. The estimators can be written as

    πj(k+1)=1ni=1npij(k+1),(11)
    βj(k+1)=i=1nxixipij(k+1)eij(k+1)1i=1nxiyipij(k+1)eij(k+1),(12)
    σj(k+1)=i=1npij(k+1)eij(k+1)(yixiβj(k+1))2i=1npij(k+1)1/2,(13)
    and
    qj(k+1)=i=1npij(k+1)i=1npij(k+1)lij(k+1).(14)

  4. Repeat the E-step and the M-step until the convergence is obtained. One stopping rule that we can choose is to stop the iteration when the change of the likelihood value is smaller than 106 or runs are more than 500.

Remarke:

If we further assume that all σj's are equal, then in the M-step, we can update σ by

σ(k+1)=i=1nj=1mpij(k+1)eij(k+1)(yixiβj(k+1))2n12.

3.1. Initial Point in the EM Algorithm

It is well known that mixture models may provide a multimodal LLF. In this sense, the method of MLE through EM algorithm may not give maximum global solutions if the starting values are far from the real parameter values. Thus, the choice of starting values for the EM algorithm in the mixture context plays an important role in parameters estimation. In our example and simulation studies, we consider the following procedure for the slash MRM.

  • Partition the sample into m groups using the K-means clustering algorithm (see [42]).

  • Compute the proportion of data points belonging to the same cluster j, say pj(0),j=1,,m. This is the initial value for pj.

  • For each group j, compute the initial values βj(0) using the method of least square. Then, use the residuals of each group to compute initial values σj(0) and qj(0).

3.2. Evaluation of the Standard Errors from the EM-Algorithm

In this subsection, we use the results of [43] to obtain the standard errors of the estimators from the EM-algorithm. For evaluation of the standard errors, one can also see [44]. Let v=X,y,z,u, then the elements of information matrix Ij(θj;v)=2c(Θ;v)θj2, for j=1,,m, are given in Appendix A.

Taking the conditional expectation of Ij(θj;v) given (X,y), we obtain the (p+3)×(p+3) matrix

cj(θj;X,y)=EIj(θj;v)|X,y=[cklj],(15)
that the elements of matrix [cklj] are presented in Appendix B.

Moving now to the computation of mj(θ̂j;X,y)) as

mj(θ̂j;X,y)=VarUcj(v;θ̂j)X,y=vklj(16)
where Ucj(v;θ̂j)=c(Θ;v)βj,c(Θ;v)qj,c(Θ;v)σj,c(Θ;v)πj and the elements of matrix [vklj] are given by Appendix C. Note that
Eg(Ui)|y,X,zij=1=01g(ui)qjuiqjexpui2(yixiTβj)22σj2σj2πf(yixiTβj;σj,qj)dui,(17)
where the g() is real function of ui from (15) and (16), we obtain the observed information matrix I^=[[Ij(θ̂j;X,y)]]
Ij(θ̂j;X,y)=cj(θ̂j;X,y)mj(θ̂j;X,y).

The standard errors of the MLEs of the EM-algorithm are the inverse of square root of the diagonal elements of the Ij(θ̂j;X,y).

4. SIMULATION STUDY

To assess the finite sample performance of the proposed model and the estimation procedure, we conducted an simulation study. In the EM algorithm, it is well khnown that label switching is always an issue when evaluating different estimation methods in mixture models, and there are no widely accepted labeling standards. In this simulation study, we choose the labels by minimizing the distance to the true parameter values. However, more research on choosing labels is needed (see [45]). Also, we considered equal variance for all components. The reason is that, if the variances are not same, the LLF (7) is unbounded and tends to infinity if one observation lies exactly on one component line and the corresponding variance vanishes, which makes the simulation very unstable.

4.1. Evaluate Robustness of the MRM

We use the method that proposed by [21] for evaluating robustness of the MRM. They proposed a trimmed version of the new method by fitting the new model to the data after adaptively trimming high leverage points. In addition, unlike TLE (see [18]), the proportion of trimming of the new method is data adaptive instead of a fixed value (For more explain, see [21]). Let X=(x1,,xn) and hii be the ith diagonal element of H, where H=X(XX)1X. The element hii is called the leverage for the ith predictor xi and xi is considered a high leverage point if hii is large. Note however

hii=n1+(n1)1MDi,(18)
where
MDi=(xix¯)S1(xix¯),(19)
is the Mahalanobis distance, x¯ is the sample mean of xi's, and S is the sample covariance of xi's (without the intercept 1). It is well known that x¯ and S are not resistant to outliers and might create the masking effect (see [46]), i.e., some high leverage points might not be identified due to the influence of other high-leverage points. In order to overcome this, a modified Mahalanobis distance is proposed
MDi=(xim(X))C(X)1(xim(X)),(20)
where m(X) and C(X) are robust estimates of location and scatter for X (after removing the first column 1s).

Suppose that (x1i,x2i,yi),i=1,,n is a random sampled from the following model

Y=0+X1+X2+ϵ1,ifZ=1;0X1X2+ϵ2,ifZ=2,
where Z is a component indicator of Y with P(Z=1)=1P(Z=2)=0.25, X1~N(0,1), X2~N(0,1), and ϵ1 and ϵ2 have the same distribution as ϵ. To see the effects of different distributions of ϵ and high leverage outliers in x-direction on various estimation methods, we considered the following six cases:
  • Case (I):ϵ~N(0,1), standard normal distribution.

  • Case (II):ϵ~slash(1,2), slash distribution with parameter q=2 and scale parameter σ=1.

  • Case (III):ϵ~t1, t-distribution with 1 degree of freedom.

  • Case (IV):ϵ~l(0,1), Laplace distribution with mean 0 and variance 1.

  • Case (V):ϵ~0.95N(0,1)+0.05N(0,25), a mixture of two normal distributions.

  • Case (VI):ϵ~N(0,1) with 5% high leverage outliers being X1=X2=20 and Y=100.

Case (I) is often used to evaluate the efficiency of different estimation methods compared to the traditional MLE when the error is normally distributed and there is no outliers. For Case (II), the estimation methods proposed in this paper will provide the MLE of unknown parameters, which, as in the first case, would serve as a baseline to evaluate the performance of other estimation procedures. Cases (III) and (IV) are heavy tailed distribution and are often used in the literature to mimic the outlier situations. Case (V) would produce 5% data likely to be low leverage outliers, and in Case (VI), 5% of the observations are replicated serving as the high leverage outliers, which will be used to check the robustness of estimation procedures against the high leverage outliers.

Five estimation methods were compared in the simulation study: (1) the maximum likelihood method based on the normality assumption (MLE); (2) the TLE proposed by [18]; (3) the robust modified EM algorithm based on bisquare (Bisquare) proposed by [19]; (4) the proposed robust EM mixture regression method based on the slash distribution (Mixregs); (5) the trimmed mixture regression method based on the slash distribution (Mixregs-MCD), with MCD trimming method. In the iterative numerical methods, the iteration is terminated when the change in the likelihood function is less than 106.

Tables 25 report the mean squared errors (MSE) and the absolute bias (Bias) of the parameter estimates for each estimation method for sample sizes n=50,100,200 and n=400, respectively. The number of replicates is 5000. As shown in Tables 25, in Case (I) through (IV), Mixregs and Mixregs-MCD performed at a level that is better or equal to the other three methods. In Case (V) where there are high leverage outliers, Mixregs-MCD also outperformed the other four methods.

True MLE TLE Bisquare Mixregt Mixregt-MCD Mixregs Mixregs-MCD
Case (I): ε~N(0,1)

β10:0 0.0035(0.4785) 0.0109(0.6015) 0.0040(0.4875) 0.0032(0.4561) 0.0071(0.5167) 0.0001(0.3494) 0.0056(0.3770)
β11:1 0.1528(0.5638) 0.8904(1.5952) 0.4678(0.9193) 0.2377(0.6354) 0.2862(0.7588) 0.0495(0.3492) 0.0722(0.4207)
β12:1 0.1366(0.5285) 0.8904(1.5965) 0.4586(0.8733) 0.2217(0.6059) 0.2696(0.7172) 0.0393(0.3625) 0.0612(0.4490)
β20:0 0.0006(0.0869) 0.0053(0.2304) 0.0038(0.1296) 0.0019(0.0981) 0.0059(0.1269) 0.0012(0.0455) 0.0024(0.0544)
β21:1 0.0058(0.0627) 0.0212(0.1404) 0.0277(0.0879) 0.0084(0.0702) 0.0068(0.0985) 0.0074(0.0472) 0.0027(0.0661)
β22:1 0.0160(0.0680) 0.0275(0.1278) 0.0432(0.0906) 0.0212(0.0663) 0.0246(0.0933) 0.0013(0.0483) 0.0007(0.0654)
π1:0.25 0.0326(0.0120) 0.1413(0.0374) 0.0891(0.0259) 0.0510(0.0149) 0.0626(0.0182) 0.0003(0.0130) 0.0068(0.0151)

Case (II): ε~slash(0,1)

β10:0 0.0592(52.9240) 0.0066(1.5574) 0.0065(1.5919) 0.0178(3.9062) 0.0531(4.1787) 0.0160(0.7033) 0.0200(0.8057)
β11:1 0.5294(65.3548) 0.7148(1.9506) 0.5117(1.6059) 0.3411(4.5422) 0.3415(4.9709) 0.0460(0.6372) 0.0646(0.8130)
β12:1 0.2843(39.1912) 0.7195(1.8359) 0.5054(1.6077) 0.4040(4.0941) 0.3734(4.7012) 0.0409(0.6399) 0.0537(0.8949)
β20:0 0.0018(38.2963) 0.0035(0.8519) 0.0112(0.7225) 0.0548(2.7391) 0.0200(2.6460) 0.0020(0.0799) 0.0028(0.0895)
β21:1 1.0935(36.3988) 0.0835(0.6123) 0.1200(0.6647) 0.2309(3.0690) 0.2967(4.3152) 0.0073(0.0898) 0.0082(0.1115)
β22:1 1.2521(41.2024) 0.0899(0.5964) 0.1553(0.5842) 0.2122(2.8183) 0.2575(3.6373) 0.0023(0.0891) 0.0042(0.1058)
π1:0.25 0.1820(0.1555) 0.1717(0.0547) 0.1528(0.0570) 0.1505(0.0643) 0.1592(0.0695) 0.0122(0.0278) 0.0122(0.0288)

Case (III): ε~t1

β10:0 11.6090(501705.6719) 0.1484(36.6703) 0.0082(2.7508) 0.0009(9.2711) 0.0010(10.8682) 0.0161(1.6247) 0.0444(1.7095)
β11:1 21.7989(49070.7569) 0.1158(37.1632) 0.5407(3.0793) 0.4597(12.4083) 0.2960(22.9286) 0.0536(1.6220) 0.0279(2.0650)
β12:1 44.2708(1982026.0620) 0.0477(19.2282) 0.5466(2.7890) 0.3747(17.2967) 0.3317(20.0879) 0.0122(1.9936) 0.0154(2.3890)
β20:0 6.5844(1410067.1062) 0.0457(13.4585) 0.0426(8.4725) 0.0329(13.4541) 0.0239(14.6770) 0.0048(0.1757) 0.0021(0.1658)
β21:1 48.3717(2540865.2077) 0.5829(15.6338) 0.1778(2.0443) 0.4748(13.1418) 0.5896(13.8616) 0.0002(0.1785) 0.0071(0.2151)
β22:1 46.2134(713106.5159) 0.6152(14.6803) 0.2375(22.8824) 0.4430(13.5257) 0.4694(18.4484) 0.0063(0.1780) 0.0109(0.2317)
π1:0.25 0.2414(0.2483) 0.1746(0.1025) 0.1844(0.0760) 0.1769(0.0677) 0.1829(0.0724) 0.0193(0.0436) 0.0226(0.0443)

Case (IV): ε~l(0,1)

β10:0 0.0097(0.7824) 0.0100(0.5476) 0.0222(0.4867) 0.0052(0.4622) 0.0059(0.5169) 0.0061(0.2894) 0.0090(0.3205)
β11:1 0.1221(0.8182) 0.6228(1.1672) 0.4430(0.9152) 0.1946(0.6037) 0.2138(0.6899) 0.0309(0.2877) 0.0716(0.4048)
β12:1 0.1172(0.7817) 0.6214(1.1486) 0.4314(0.8522) 0.1882(0.5904) 0.2190(0.7188) 0.0450(0.2821) 0.0514(0.4334)
β20:0 0.0009(0.1970) 0.0033(0.1314) 0.0020(0.1105) 0.0018(0.0652) 0.0013(0.0824) 0.0009(0.0217) 0.0005(0.0353)
β21:1 0.0229(0.1755) 0.0182(0.1083) 0.0241(0.0728) 0.0093(0.0687) 0.0149(0.0932) 0.0012(0.0243) 0.0036(0.0426)
β22:1 0.0413(0.1677) 0.0269(0.0955) 0.0457(0.0788) 0.0207(0.0669) 0.0242(0.0956) 0.0018(0.0246) 0.0070(0.0452)
π1:0.25 0.0355 (0.0181) 0.0953(0.0269) 0.0872(0.0249) 0.0450(0.0129) 0.0530(0.0156) 0.0086(0.0138) 0.0112(0.0147)

Case (V): ε~0.95N(0,1)+0.05N(0,25)

β10:0 0.0239(5.1865) 0.0070(0.6854) 0.0129(0.6047) 0.0077(1.1336) 0.0053(1.2544) 0.0000(0.2890) 0.0009(0.3453)
β11:1 0.1788(4.8738) 0.7779(1.4818) 0.4566(1.0430) 0.2930(1.4110) 0.3190(1.6596) 0.0299(0.3321) 0.0526(0.4502)
β12:1 0.0983(5.5528) 0.7600(1.3980) 0.4632(0.9917) 0.2787(1.3798) 0.3233(1.7069) 0.0445(0.4024) 0.0408(0.4929)
β20:0 0.0175(4.8827) 0.0012(0.2373) 0.0053(0.1499) 0.0052(0.4807) 0.0023(0.703) 0.0115(0.0469) 0.0114(0.0428)
β21:1 0.2626(3.9198) 0.0240(0.1881) 0.0415(0.1316) 0.0266(0.4103) 0.0515(0.7089) 0.0003(0.0413) 0.0002(0.0594)
β22:1 0.3090(3.8525) 0.0373(0.1679) 0.0492(0.1490) 0.0639(0.5072) 0.084(0.7683) 0.0054(0.0403) 0.0046(0.0579)
π1:0.25 0.1009(0.0886) 0.1257(0.0344) 0.0994(0.0308) 0.0749(0.0276) 0.0908(0.0343) 0.0037(0.0124) 0.0003(0.0158)

Case (VI): ε~N(0,1) with 5% high leverage outliers

β10:0 0.0763(5.3771) 0.0128(0.7567) 0.0438(1.0701) 0.1006(4.1444) 0.0019(0.5139) 0.0423(0.8197) 0.0046(0.3005)
β11:1 1.4844(6.1813) 0.4397(1.3832) 0.0892(1.4288) 1.4920(4.6566) 0.2778(0.7660) 0.5649(1.2906) 0.0066(0.3703)
β12:1 1.5091(6.2501) 0.4477(1.3542) 0.1050(1.4845) 1.4746(4.6102) 0.2815(0.7487) 0.5472(1.2253) 0.0150(0.3632)
β20:0 0.0024(0.0508) 0.0039(0.1774) 0.0029(0.0805) 0.0048(0.0568) 0.0037(0.1261) 0.0005(0.0425) 0.0013(0.0449)
β21:1 0.2027(0.1196) 0.0046(0.1412) 0.0082(0.0781) 0.1244(0.0913) 0.0303(0.1008) 0.0701(0.0583) 0.0003(0.0446)
β22:1 0.2099(0.1244) 0.0050(0.1365) 0.0030(0.0812) 0.1338(0.0964) 0.0237(0.1078) 0.0677(0.0586) 0.0004(0.0460)
π1:0.25 0.0886(0.0102) 0.0930(0.0247) 0.0461(0.0193) 0.0738(0.0089) 0.0709(0.0204) 0.0364(0.0129) 0.0048(0.0123)

MSE, mean squared error; MLE, maximum likelihood estimations; TLE, trimmed likelihood estimator.

Table 2

Bias and MSE (in parentheses) of point estimates for n=50.

True MLE TLE Bisquare Mixregt Mixregt-MCD Mixregs Mixregs-MCD
Case (I): ε~N(0,1)

β10:0 0.0097(0.1228) 0.0020(0.1002) 0.0006(0.1349) 0.0030(0.1340) 0.0008(0.1554) 0.0064(0.1342) 0.0054(0.1496)
β11:1 0.0149(0.1251) 0.0735(0.1139) 0.0240(0.1675) 0.0405(0.1513) 0.0546(0.1995) 0.0019(0.1312) 0.0007(0.1611)
β12:1 0.0030(0.1298) 0.0747(0.1147) 0.0328(0.1659) 0.0392(0.1501) 0.0580(0.1954) 0.0044(0.1327) 0.0098(0.1648)
β20:0 0.0009(0.0188) 0.0026(0.0168) 0.0000(0.0210) 0.0015(0.0218) 0.0029(0.0255) 0.0020(0.0204) 0.0019(0.0218)
β21:1 0.0019(0.0195) 0.0275(0.0173) 0.0025(0.0219) 0.0096(0.0209) 0.0107(0.0261) 0.0039(0.0207) 0.0028(0.0256)
β22:1 0.0008(0.0197) 0.0294(0.0165) 0.0061(0.0218) 0.0072(0.0202) 0.0090(0.0255) 0.0069(0.0202) 0.0051(0.0245)
π1:0.25 0.0006(0.0061) 0.0523(0.0096) 0.0025(0.0113) 0.0117(0.0049) 0.0157(0.0058) 0.0042(0.0064) 0.0026(0.0072)

Case (II): ε~slash(0,1)

β10:0 0.0883(67.0279) 0.0005(0.9510) 0.0018(0.9885) 0.0182(2.1725) 0.0104(2.5396) 0.0008(0.6941) 0.0049(0.6992)
β11:1 0.8606(72.2127) 0.5882(1.2027) 0.3194(0.9784) 0.1967(4.4901) 0.1811(4.3832) 0.0462(0.6795) 0.0424(0.7862)
β12:1 0.8628(49.1993) 0.5832(1.1459) 0.2903(0.9782) 0.1583(3.2396) 0.2100(3.1637) 0.0339(0.6743) 0.0489(0.7677)
β20:0 0.1339(84.6754) 0.0034 (0.3623) 0.0034 (0.3351) 0.0096 (1.5744) 0.0138 (1.4106) 0.0043(0.0629) 0.0049(0.0883)
β21:1 1.7241(64.2070) 0.0364(0.1853) 0.0829(0.2552) 0.0985(1.1663) 0.1289(1.2812) 0.0052(0.0665) 0.0054(0.1086)
β22:1 1.6774(64.7592) 0.0491(0.2135) 0.0832(0.2432) 0.1126(1.2056) 0.1139(1.1414) 0.0057(0.0688) 0.0010(0.1151)
π1:0.25 0.1798(0.1945) 0.1198(0.0328) 0.0988(0.0351) 0.0795(0.0355) 0.0924(0.0421) 0.0075(0.0260) 0.0113(0.0288)

Case (III): ε~t1

β10:0 10.2255(502197.693) 0.0724(7.9262) 0.0025(1.6847) 0.0193(2.6750) 0.0042(2.7491) 0.0165(0.8272) 0.0064(0.9149)
β11:1 53.9287(473663.046) 0.2454(6.7949) 0.4377(1.7240) 0.5182(4.3426) 0.5384(3.8060) 0.0625(0.8331) 0.0764(0.9773)
β12:1 57.7344(604387.717) 0.1955(6.1361) 0.4510(1.7334) 0.5412(3.3579) 0.5114(3.6211) 0.0624(0.8387) 0.0637(0.9625)
β20:0 13.9702(2522395.636) 0.0047(7.0569) 0.0081(0.7474) 0.0597(3.4518) 0.0420(5.3667) 0.0005(0.0544) 0.0047(0.0560)
β21:1 92.9882(2734642.730) 0.3511(5.3462) 0.0773(0.5873) 0.1203(3.2147) 0.1104(2.5059) 0.0136(0.0621) 0.0165(0.0712)
β22:1 81.6340(2400307.367) 0.3780(5.6155) 0.0908(0.5098) 0.1038(1.3682) 0.2075(6.1631) 0.0122(0.0628) 0.0148(0.0722)
π1:0.25 0.2405(0.277) 0.1350(0.0961) 0.1455(0.0543) 0.1159(0.0377) 0.1235(0.0416) 0.0090(0.0336) 0.0141(0.0363)

Case (IV): ε~l(0,1)

β10:0 0.0073(0.2388) 0.0011(0.3148) 0.0020(0.2660) 0.0020(0.1306) 0.0022(0.1452) 0.0059(0.0958) 0.0048(0.1045)
β11:1 0.0219(0.2414) 0.3869(0.6129) 0.1957(0.3853) 0.0382(0.1385) 0.0507(0.1716) 0.0268(0.1197) 0.0328(0.1484)
β12:1 0.0268(0.2044) 0.4095(0.6287) 0.1973(0.3881) 0.0406(0.1330) 0.0475(0.1663) 0.0363(0.1166) 0.0429(0.1409)
β20:0 0.0027(0.0847) 0.0033(0.0484) 0.0010 (0.0401) 0.0006(0.0147) 0.0002(0.0177) 0.0020(0.0127) 0.0022(0.0133)
β21:1 0.0219(0.0663) 0.0046(0.0271) 0.0202(0.0264) 0.0066(0.0129) 0.0065(0.0155) 0.0018(0.0125) 0.0017(0.0161)
β22:1 0.0212(0.0428) 0.0065(0.0274) 0.0228(0.0256) 0.0091(0.0139) 0.0111(0.0173) 0.0026(0.0115) 0.0029(0.0158)
π1:0.25 0.0096(0.0063) 0.0425(0.0111) 0.0461(0.0122) 0.0123(0.0043) 0.0144(0.0048) 0.0078(0.0069) 0.0091(0.0081)

Case (V): ε~0.95N(0,1)+0.05N(0,25)

β10:0 0.0363(5.8679) 0.0107(0.4219) 0.0034(0.2510) 0.0023(0.3526) 0.0008(0.3908) 0.0018(0.2013) 0.0027(0.2141)
β11:1 0.0887(4.9647) 0.6043(0.9888) 0.1826(0.4323) 0.0754(0.4468) 0.0709(0.5583) 0.0050(0.1991) 0.0102(0.2473)
β12:1 0.0461(4.9456) 0.6089(0.9898) 0.1928(0.4286) 0.0797(0.4676) 0.0964(0.5888) 0.0077(0.2032) 0.0096(0.2427)
β20:0 0.0404(7.0121) 0.0056(0.1193) 0.0022(0.0542) 0.0041(0.1568) 0.0008(0.2250) 0.0016(0.0243) 0.0023(0.0256)
β21:1 0.3090(4.3915) 0.0108(0.0553) 0.0165(0.0387) 0.0125(0.1123) 0.0222(0.1786) 0.0079(0.0240) 0.0078(0.0301)
β22:1 0.2940(4.3170) 0.0051(0.0524) 0.0136(0.0407) 0.0091(0.1268) 0.0174(0.3255) 0.0062(0.0236) 0.0062(0.0283)
π1:0.25 0.0871(0.0992) 0.0781(0.0196) 0.0460(0.0141) 0.0203(0.0095) 0.0238(0.0109) 0.0049(0.0090) 0.0040(0.0103)

Case (VI): ε~N(0,1) with 5% high leverage outliers

β10:0 0.1116(3.9889) 0.0143(0.5237) 0.0455(0.7972) 0.1189(2.8301) 0.0090(0.1585) 0.0655(0.6801) 0.0082(0.1237)
β11:1 1.4606(4.5974) 0.1432(0.8990) 0.3537(1.0996) 1.4982(3.6290) 0.0629(0.2004) 0.7505(1.3577) 0.0019(0.1302)
β12:1 1.5307(4.8079) 0.1652(0.8693) 0.3766(1.1550) 1.4945(3.6208) 0.0549(0.1984) 0.7475(1.3744) 0.0050(0.1253)
β20:0 0.0104(0.0262) 0.0130(0.0896) 0.0003(0.0426) 0.0105(0.0612) 0.0025(0.0269) 0.0061(0.0196) 0.0003(0.0192)
β21:1 0.2252(0.0901) 0.0131(0.0849) 0.0181(0.0337) 0.1373(0.0726) 0.0071(0.0265) 0.1090(0.0441) 0.0010(0.0199)
β22:1 0.2277(0.0932) 0.0204(0.0884) 0.0221(0.0371) 0.1352(0.0745) 0.0111(0.0267) 0.1040(0.0409) 0.0018(0.0207)
π1:0.25 0.0890(0.0092) 0.0353(0.0107) 0.0006(0.0087) 0.0762(0.0076) 0.0169(0.0062) 0.0624(0.0171) 0.0006(0.0061)

MSE, mean squared error; MLE, maximum likelihood estimations; TLE, trimmed likelihood estimator.

Table 3

Bias and MSE (in parentheses) of point estimates for n=100.

True MLE TLE Bisquare Mixregt Mixregt-MCD Mixregs Mixregs-MCD
Case (I): ε~N(0,1)

β10:0 0.0003(0.0451) 0.0001(0.0364) 0.0001(0.0504) 0.0031(0.0443) 0.0029(0.0474) 0.0035(0.0483) 0.0033(0.0517)
β11:1 0.0039(0.0441) 0.0407(0.0376) 0.0024(0.0547) 0.0051(0.0418) 0.0076(0.0493) 0.0037(0.0461) 0.0065(0.0581)
β12:1 0.0084(0.0451) 0.0417(0.0374) 0.0010(0.0559) 0.0065(0.0415) 0.0092(0.0497) 0.0084(0.0470) 0.0090(0.0576)
β20:0 0.0015(0.0093) 0.0000(0.0075) 0.0014(0.0100) 0.0005(0.0092) 0.0007(0.0094) 0.0029(0.0086) 0.0030(0.0091)
β21:1 0.0006(0.0093) 0.0305(0.0086) 0.0018(0.0101) 0.0044(0.0089) 0.0059(0.0104) 0.0012(0.0091) 0.0027(0.0108)
β22:1 0.0015(0.0089) 0.0299(0.0084) 0.0001(0.0099) 0.0021(0.0090) 0.0038(0.0105) 0.0003(0.0094) 0.0002(0.0112)
π1:0.25 0.0002(0.0026) 0.0581(0.0071) 0.0000(0.0038) 0.0031(0.0022) 0.0043(0.0024) 0.0020(0.0028) 0.0021(0.0032)

Case (II): ε~slash(0,1)

β10:0 0.2238(129.3761) 0.0066(0.4795) 0.0129(0.4322) 0.0154(0.5663) 0.0154(0.9366) 0.0035(0.2805) 0.0050(0.3098)
β11:1 1.6322(103.4913) 0.4177(0.6341) 0.0836(0.4221) 0.0852(0.6943) 0.0707(1.6737) 0.0240(0.3173) 0.0348(0.3931)
β12:1 1.7944(107.1121) 0.4148(0.6227) 0.0760(0.4140) 0.0886(0.7964) 0.0803(1.2111) 0.0225(0.3114) 0.0294(0.3955)
β20:0 0.0987(147.1717) 0.0004(0.1309) 0.0026(0.0980) 0.0058(0.4486) 0.0083(0.3612) 0.0028(0.0209) 0.0008(0.0341)
β21:1 2.4287(134.7825) 0.0230(0.0640) 0.0429(0.0570) 0.0314(0.4694) 0.0411(0.3362) 0.0044(0.0292) 0.0009(0.0499)
β22:1 2.4348(119.5413) 0.0221(0.0638) 0.0472(0.0604) 0.0321(0.2403) 0.0400(0.5037) 0.0046(0.0279) 0.0027(0.0494)
π1:0.25 0.1839 (0.2251) 0.0767(0.0158) 0.0540(0.0144) 0.0320(0.0117) 0.0397(0.0158) 0.0063(0.0140) 0.0121(0.0182)

Case (III): ε~t1

β10:0 25.0509(1.909716e+06) 0.0184(4.7935) 0.0115(0.7171) 0.0164(0.5915) 0.0047(0.6980) 0.0102(0.2730) 0.0078(0.2782)
β11:1 131.7579(2.203641e+06) 0.1625(3.3181) 0.2768(0.9563) 0.3669(1.1273) 0.3668(1.4800) 0.0552(0.3230) 0.0797(0.4144)
β12:1 101.4703(1.798630e+06) 0.1445(3.4201) 0.2997(0.9783) 0.3612(1.2457) 0.3758(1.2865) 0.0528(0.3290) 0.0604(0.3898)
β20:0 1584.5368(1.254946e+10) 0.0478(3.7613) 0.0008(0.3362) 0.0060(0.2418) 0.0056(0.2008) 0.0024(0.0213) 0.0025(0.0230)
β21:1 1150.8596(5.003399e+09) 0.1862(2.2135) 0.0526(0.2070) 0.0363(0.2245) 0.0363(0.4318) 0.0028(0.0241) 0.0040(0.0292)
β22:1 3101.5797(4.430502e+10) 0.1874(2.2794) 0.0424(0.2205) 0.0307(0.2821) 0.0440(0.5097) 0.0000(0.0248) 0.0016(0.0303)
π1:0.25 0.2413(2.912000e-01) 0.0894(0.0751) 0.1199(0.0353) 0.0654(0.0199) 0.0706(0.0218) 0.0120(0.0179) 0.0164(0.0211)

Case (IV): ε~l(0,1)

β10:0 0.0024(0.0480) 0.0028(0.1411) 0.0029(0.0749) 0.0008(0.0313) 0.0005(0.0339) 0.0000(0.0223) 0.0006(0.0346)
β11:1 0.0055(0.0481) 0.2040(0.2414) 0.0453(0.0997) 0.0060(0.0302) 0.0056(0.0356) 0.0107(0.0249) 0.0137(0.0423)
β12:1 0.0023(0.0551) 0.2173(0.2563) 0.0456(0.0974) 0.0073(0.0293) 0.0092(0.0346) 0.0146(0.0250) 0.0186(0.0441)
β20:0 0.0017(0.0485) 0.0029(0.0154) 0.0011(0.0078) 0.0007(0.0053) 0.0008(0.0055) 0.0000(0.0060) 0.0001(0.0062)
β21:1 0.0115(0.0077) 0.0049(0.0092) 0.0114(0.0070) 0.0019(0.0050) 0.0026(0.0059) 0.0034(0.0058) 0.0045(0.0068)
β22:1 0.0148(0.0200) 0.0014(0.0098) 0.0123(0.0072) 0.0032(0.0053) 0.0037(0.0061) 0.0033(0.0061) 0.0041(0.0070)
π1:0.25 0.0036(0.0023) 0.0207(0.0046) 0.0265(0.0036) 0.0057(0.0020) 0.0063(0.0021) 0.0054(0.0028) 0.0066(0.0032)

Case (V): ε~0.95N(0,1)+0.05N(0,25)

β10:0 0.0627(5.4191) 0.0003(0.2256) 0.0010(0.0908) 0.0002(0.0786) 0.0005(0.0754) 0.0042(0.0726) 0.0005(0.0767)
β11:1 0.0103(4.9720) 0.4269(0.5704) 0.0495(0.1354) 0.0100(0.0845) 0.0129(0.0931) 0.0006(0.0674) 0.0060(0.0811)
β12:1 0.0643(4.7922) 0.4246(0.5647) 0.0499(0.1379) 0.0075(0.0821) 0.0086(0.0900) 0.0098(0.0679) 0.0128(0.0859)
β20:0 0.0544(6.3131) 0.0029(0.0504) 0.0015(0.0136) 0.0014(0.0117) 0.0010(0.0129) 0.0001(0.0102) 0.0000(0.0108)
β21:1 0.2140(3.8070) 0.0014(0.0195) 0.0086(0.0135) 0.0008(0.0112) 0.0009(0.0144) 0.0017(0.0108) 0.0018(0.0129)
β22:1 0.2437(4.0257) 0.0015(0.0192) 0.0080(0.0137) 0.0019(0.0111) 0.0019(0.0147) 0.0011(0.0103) 0.0030(0.0125)
π1:0.25 0.0574(0.0849) 0.0365(0.0084) 0.0229(0.0063) 0.0013(0.0027) 0.0022(0.0031) 0.0014(0.0038) 0.0007(0.0044)

Case (VI): ε~N(0,1) with 5% high leverage outliers

β10:0 0.1759(2.6488) 0.0012(0.2241) 0.0576(0.4823) 0.1315(1.8995) 0.0015(0.0496) 0.0717(0.4401) 0.0015(0.0442)
β11:1 1.4949(3.5135) 0.0875(0.3978) 0.3566(0.7782) 1.5023(3.0228) 0.0089(0.0493) 0.6859(1.1268) 0.0033(0.0466)
β12:1 1.4925(3.5030) 0.0779(0.4179) 0.3686(0.8232) 1.4895(2.9826) 0.0040(0.0522) 0.6703(1.0896) 0.0048(0.0478)
β20:0 0.0080(0.0145) 0.0037(0.0366) 0.0006(0.0105) 0.0072(0.0271) 0.0011(0.0096) 0.0013(0.0098) 0.0015(0.0099)
β21:1 0.2441(0.0798) 0.0172(0.0465) 0.0259(0.0129) 0.1376(0.0679) 0.0048(0.0104) 0.1002(0.0312) 0.0002(0.0091)
β22:1 0.2445(0.0800) 0.0188(0.0460) 0.0266(0.0126) 0.1421(0.0652) 0.0004(0.0104) 0.1035(0.0314) 0.0031(0.0092)
π1:0.25 0.0954(0.0098) 0.0070(0.0041) 0.0138(0.0044) 0.0836(0.0080) 0.0042(0.0025) 0.0561(0.0159) 0.0008(0.0026)

MSE, mean squared error; MLE, maximum likelihood estimations; TLE, trimmed likelihood estimator.

Table 4

Bias and MSE (in parentheses) of point estimates for n=200.

True MLE TLE Bisquare Mixregt Mixregt-MCD Mixregs Mixregs-MCD
Case (I): ε~N(0,1)

β10:0 0.0001(0.0123) 0.0001(0.0115) 0.0000(0.0019) 0.0030(0.0098) 0.0028(0.0113) 0.0010(0.0203) 0.0011(0.0213)
β11:1 0.0012(0.0131) 0.0290(0.0191) 0.0013(0.0023) 0.0017(0.0083) 0.0037(0.0094) 0.0031(0.0185) 0.0058(0.0225)
β12:1 0.0086(0.0124) 0.0312(0.0189) 0.0008(0.0028) 0.0034(0.0070) 0.0039(0.0199) 0.0034(0.0191) 0.0049(0.0226)
β20:0 0.0019(0.0012) 0.0000(0.0011) 0.0012(0.0011) 0.0002(0.0013) 0.0000(0.0021) 0.0015(0.0043) 0.0032(0.0044)
β21:1 0.0002(0.0011) 0.0313(0.0009) 0.0009(0.0013) 0.0011(0.0010) 0.0023(0.0056) 0.0003(0.0043) 0.0008(0.0051)
β22:1 0.0016(0.0065) 0.0181(0.0011) 0.0001(0.0008) 0.0013(0.0011) 0.0016(0.0059) 0.0005(0.0044) 0.0003(0.0051)
π1:0.25 0.0001(0.0011) 0.0652(0.0020) 0.0000(0.0011) 0.0024(0.0009) 0.0023(0.0012) 0.0011(0.0012) 0.0017(0.0013)

Case (II): ε~slash(0,1)

β10:0 0.2343(254.24) 0.0070(0.1151) 0.0093(0.2230) 0.0152(0.0654) 0.0155(0.0366) 0.0071(0.0967) 0.0101(0.1025)
β11:1 1.7454(223.56) 0.4048(0.0958) 0.0795(0.2149) 0.0757(0.0492) 0.0691(0.9860) 0.0140(0.1103) 0.0177(0.1463)
β12:1 1.8329(213.27) 0.3959(0.2146) 0.0631(0.2017) 0.0832(0.0197) 0.0731(0.8741) 0.0050(0.1088) 0.0130(0.1411)
β20:0 0.0759(252.34) 0.0003(0.0957) 0.0021(0.0414) 0.0056(0.0178) 0.0083(0.0856) 0.0027(0.0137) 0.0006(0.0153)
β21:1 2.6730(257.54) 0.0211(0.0216) 0.0413(0.0372) 0.0311(0.0182) 0.0408(0.0811) 0.0052(0.0155) 0.0008(0.0197)
β22:1 2.6970(223.3501) 0.0103(0.0194) 0.0466(0.0419) 0.0309(0.0949) 0.0392(0.1131) 0.0043(0.0155) 0.0061(0.0195)
π1:0.25 0.1859(0.2324) 0.0524(0.0028) 0.0532(0.0097) 0.0319(0.0070) 0.0396(0.0084) 0.0063(0.0057) 0.0092(0.0073)

Case (III): ε~t1

β10:0 50.3757(6.4951e+8) 0.0102(2.5314) 0.0121(0.1130) 0.0101(0.0814) 0.0049(0.1160) 0.0084(0.0835) 0.0081(0.0904)
β11:1 265.7957(9.9166e+8) 0.0941(1.9187) 0.1923(0.9429) 0.2108(0.0830) 0.3234(1.1781) 0.0490(0.1022) 0.8163(0.1304)
β12:1 215.9238(10.6280e+8) 0.1153(2.0520) 0.2544(0.9559) 0.2134(0.0849) 0.3563(1.2229) 0.0483(0.1070) 0.0601(0.1299)
β20:0 2689.9093(6.4192e+12) 0.0253(2.0514) 0.0007(0.0129) 0.0031(0.0092) 0.0021(0.1932) 0.0019(0.0097) 0.0024(0.0102)
β21:1 2348.1087(9.86884)e+11 0.1153(1.5243) 0.0341(0.0148) 0.0268(0.0100) 0.0292(0.3334) 0.0021(0.0103) 0.0031(0.0120)
β22:1 10589.3970(10.6781e+12) 0.1184(1.6543) 0.0352(0.0149) 0.0281(0.0098) 0.0513(0.4334) 0.0000(0.0102) 0.0015(0.0121)
π1:0.25 0.2414(0.1872) 0.0223(0.0141) 0.1179(0.0374) 0.0633(0.0065) 0.0585(0.0173) 0.0118(0.0075) 0.0168(0.0091)

Case (IV): ε~l(0,1)

β10:0 0.0013(0.0914) 0.0031(0.1121) 0.0037(0.0543) 0.0036(0.0488) 0.0023(0.0887) 0.0000(0.0154) 0.0005(0.0254)
β11:1 0.0043(0.0962) 0.2037(0.2321) 0.0447(0.0817) 0.0394(0.0674) 0.0495(0.1073) 0.0101(0.0148) 0.0136(0.0325)
β12:1 0.0019(0.0026) 0.2169(0.2457) 0.0449(0.0821) 0.0405(0.0513) 0.0461(0.0928) 0.0137(0.0126) 0.0184(0.0329)
β20:0 0.0018(0.0087) 0.0017(0.0130) 0.0013(0.0011) 0.0006(0.0138) 0.0002(0.0081) 0.0000(0.0026) 0.0001(0.0036)
β21:1 0.0093(0.0046) 0.0035(0.0017) 0.0109(0.0010) 0.0065(0.0096) 0.0061(0.0073) 0.0033(0.0023) 0.0043(0.0031)
β22:1 0.0132(0.0089) 0.0009(0.0028) 0.0119(0.0051) 0.0087(0.0113) 0.0101(0.0078) 0.0031(0.0024) 0.0038(0.0047)
π1:0.25 0.0033(0.0021) 0.0173(0.0031) 0.0221(0.0025) 0.0117(0.0021) 0.0139(0.0019) 0.0054(0.0011) 0.0052(0.0019)

Case (V): ε~0.95N(0,1)+0.05N(0,25)

β10:0 0.1124(0.0204) 0.0003(0.1112) 0.0003(0.0147) 0.0002(0.0415) 0.0005(0.0345) 0.0041(0.0134) 0.0005(0.0141)
β11:1 0.1383(0.0405) 0.4023(0.3214) 0.0246(0.0751) 0.0095(0.0481) 0.0123(0.0489) 0.0006(0.0135) 0.0059(0.0164)
β12:1 0.1453(0.0444) 0.4018(0.3123) 0.0251(0.0813) 0.0071(0.0425) 0.0074(0.0461) 0.0099(0.0347) 0.0126(0.0174)
β20:0 0.0486(0.0037) 0.0021(0.0205) 0.0013(0.0070) 0.0011(0.0061) 0.0009(0.0068) 0.0001(0.0028) 0.0000(0.0029)
β21:1 0.0523(0.0044) 0.0010(0.0116) 0.0083(0.0086) 0.0008(0.0056) 0.0009(0.0071) 0.0017(0.0029) 0.0017(0.0032)
β22:1 0.0505(0.0040) 0.0013(0.0112) 0.0078(0.0088) 0.0017(0.0054) 0.0017(0.0073) 0.0010(0.0027) 0.0030(0.0031)
π1:0.25 0.0375(0.0026) 0.0315(0.0036) 0.0213(0.0051) 0.0009(0.0011) 0.0021(0.0016) 0.0013(0.0012) 0.0007(0.0014)

Case (VI): ε~N(0,1) with 5% high leverage outliers

β10:0 0.1873(4.0351) 0.0009(0.1219) 0.0632(0.1811) 0.1415(1.1217) 0.0011(0.0216) 0.0651(03112) 0.0064(0.0359)
β11:1 1.5082(5.8193) 0.0783(0.1734) 0.3422(0.3216) 1.5131(2.9812) 0.0081(0.0211) 0.6503(1.1170) 0.0030(0.0228)
β12:1 1.4889(6.6128) 0.0730(0.1964) 0.3541(0.3917) 1.4521(2.5438) 0.0017(0.0378) 0.6532(0.9896) 0.0026(0.0224)
β20:0 0.0182(0.0313) 0.0031(0.0113) 0.0006(0.0083) 0.0071(0.0118) 0.0006(0.0038) 0.0013(0.0087) 0.0006(0.0072)
β21:1 0.2495(0.0325) 0.0169(0.0247) 0.0231(0.0067) 0.1369(0.0321) 0.0041(0.0061) 0.1001(0.0213) 0.0010(0.0086)
β22:1 0.2512(0.0396) 0.0167(0.0239) 0.0252(0.0071) 0.0890(0.0326) 0.0003(0.0057) 0.1020(0.0207) 0.0015(0.0074)
π1:0.25 0.0986(0.0035) 0.0065(0.0017) 0.0132(0.0023) 0.0828(0.0035) 0.0041(0.0012) 0.0541(0.0113) 0.0006(0.0018)

MSE, mean squared error; MLE, maximum likelihood estimations; TLE, trimmed likelihood estimator.

Table 5

Bias and MSE (in parentheses) of point estimates for n=400.

5. A REAL DATA SET ANALYSIS

In this section, we apply the proposed robust procedure to tone perception data set [47]. In the experiment, a pure fundamental tone was played to a trained musician and electronically generated overtones were added, determined by a stretching ratio. A value of 2 for the stretch ratio corresponds to the harmonic pattern usually heard in traditional definite pitched instruments. The musician was asked to tune an adjustable tone to the octave above the fundamental tone, and a measurement called tuned gives the ratio of the adjusted tone to the fundamental. 150 pairs of (Actual tone ratio, Perceived tone ratio) values are obtained for the same musician. The variable Perceived tone ratio is treated as a response variable and Actual tone ratio as a predictor. Figure 1 shows the scater plot (in left plot) and histogram (in rigth) of tone data. The setup of the experiment indicates two mixture components in the model, and the scatter plot of the data collected from the experiment confirms this point. The right plot in Figure 2 shows that the robust regression based on slash distribution is better than robust regression based normal and t distributions. With the left plot in Figure 2, can be understand the reason of using MRMs, because the traditional models (MLE) is very affected by outlier. Many papers have analyzed this dataset using a mixture of linear regressions framework; see [48,49,21,27]. Now we revisit this data set with the aim of expanding the inferential results to the slash distributions. Table 5 reports various fitted models.

Figure 1

Left plot is the scatterplot of the tone data. Right plot is the histogram of the perceived tone ratio.

Figure 2

Mixture linear fitting with outlier (1.3,5). In left plot: black line-mixregs, green line-TLE, red line-MLE, in right plot: black line-mixregs, green line-bisquare, red line-mixregt.

From Table 6, we observe that the robust regression based on slash distribution is better than robust regression based normal and t distribution, according to the AIC [50] and BIC [51] criterion.

Model β̂01 β̂11 β̂02 β̂12 σ̂1 σ̂2 q̂1 q̂2 π̂ ̂ AIC BIC
Normal −0.039 1.008 1.892 0.056 0.084 0.084 0.325 107.257 −200.513 −179.439
t 0.006 0.998 1.978 0.017 0.011 0.011 1 1 0.485 202.804 −387.608 −360.513
Slash 0.003 0.999 1.954 0.029 0.002 0.020 0.569 1.455 0.443 229.436 −440.871 −413.776
Table 6

Fitted various models on the tone perception data set.

CONCLUSION

In this paper, we propose a robust MRM based on the slash distribution by extending the mixture of slash distributions to the regression setting. We suggest some guidelines to further improve the robustness of the proposed model against possible high leverage outliers by adaptively trimming high leverage points. The proposed method was compared to the traditional MLE and some other robust methods by using a simulation study and by using a real data set, it was shown that the results of the robust regression based on slash distribution is better than robust regression based normal and t distributions. Figure 2 also confirms this supremacy.

AUTHORS' CONTRIBUTIONS

The authors thank the Associate Editor and anonymous reviewers for their useful comments and suggestions on an earlier version of this manuscript which led to this improved one.

Funding Statement

This work has been conducted by University of Zabol.

APPENDIX A

2cqj2=i=1nzijqj2,2cqjβj=2cπjβj=0p×1,2cqjσj=2cqjπj=2cπjσj=0,
2cσj2=i=1nzij3ui2(yixiβj)2σj4+1σj2,2cβj2=1σj2i=1nzijui2xixi,
2cπj2=i=1nzijπj2,2cσjβj=2σj3i=1nzijui2xi(yixiβj).

APPENDIX B

c11j=i=1npijuijxixiσj2,c22j=i=1npijqj2,c33j=i=1npij3uij(yixiβj)2σj4+1σj2,
c44j=i=1npijπj2,c13j=2σj3i=1npijuijxi(yixiβj),c31j=(c13j),c21j=c41j=01×p,
c12j=c14j=0p×1,c23j=c32j=c24j=c42j=c34j=c43j=0,
with
pij=πjf(yixiβj;σj,qj)l=1mπlf(yixiβl;σl,ql)
and
uij=qj2qj2(yixiβj)2σj2qj+32σjπf(yixiβj;σj,qj)Γqj+32G(yixiβj)22σj2;qj+32,1,ifyixiβj0,qjqj+31σj2πf(0;σj,qj),ifyixiβj=0.

APPENDIX C

v11j=i=1nzij2xixi(yixiβj)2V3,ijσj4,v22j=i=1nzij2V1,ij,v33j=i=1nzij2(yixiβj)4V3,ijσj6,
v12j=i=1nzij2xi(yixiβj)V2,ijσj2,v21j=(v12j),v23j=(v32j)=i=1nzij2(yixiβj)2V2,ijσj3,
v13j=i=1nzij2xi(yixiβj)3V3,ijσj5,v31j=(v13j),v24j=v42j=v44j=v43j=v34j=0,
v14j=0p×1,v41j=01×p,
where
V1,ij=Varlog(Ui)|y,X,zij=1,V2,ij=Covlog(Ui),Ui2|y,X,zij=1
V3,ij=VarUi2|y,X,zij=1,
it is useful to note that V1,ij, V2,ij and V3,ij can be drive from (17).

REFERENCES

22.A. Azzalini, Scand. J. Stat., Vol. 12, 1985, pp. 171-178. https://www.jstor.org/stable/4615982
34.S. Morgenthaler and J.W. Tukey, Configural Polysampling: A Route to Practical Robustness, Wiley, NewYork, NY, USA, 1991. https://www.amazon.com/Configural-Polysampling-Robustness-Probability-Statistics/dp/0471523720
35.P. Bindu, Statistica, Vol. 72, 2012, pp. 235-249. https://rivista-statistica.unibo.it/article/view/3645/2996
39.J. Chen, X. Tan, and R. Zhang, Stat. Sin., Vol. 18, 2008, pp. 443-465. https://www.jstor.org/stable/24308490
42.D. MacKay, An Example Inference Task: Clustering, Information Theory, Inference and Learning Algorithms, Cambridge University Press, 2003, pp. 284-292. chap. 20 https://www.inference.org.uk/itprnn/book.pdf
45.Y. Wei, Robust MRM Using t-Distribution, Department of Statistics, Kansas State University, Manhattan, KS, USA, 2012. Master Report https://krex.k-state.edu/dspace/bitstream/handle/2097/14110/YanWei2012.pdf?sequence=1&isAllowed=y
50.H. Akaike, B.N. Petrov and B.F. Csaki (editors), Second International Symposium on Information Theory, Academiai Kiado, Budapest, Hungary, 1973, pp. 267-281. http://amosdevelopment.com/webhelp/index.html?akaike_1973.htm
Journal
Journal of Statistical Theory and Applications
Volume-Issue
19 - 2
Pages
118 - 132
Publication Date
2020/04/09
ISSN (Online)
2214-1766
ISSN (Print)
1538-7887
DOI
10.2991/jsta.d.200304.001How to use a DOI?
Copyright
© 2020 The Authors. Published by Atlantis Press SARL.
Open Access
This is an open access article distributed under the CC BY-NC 4.0 license (http://creativecommons.org/licenses/by-nc/4.0/).

Cite this article

TY  - JOUR
AU  - Hadi Saboori
AU  - Ghobad Barmalzan
AU  - Mahdi Doostparast
PY  - 2020
DA  - 2020/04/09
TI  - Robust Mixture Regression Based on the Mixture of Slash Distributions
JO  - Journal of Statistical Theory and Applications
SP  - 118
EP  - 132
VL  - 19
IS  - 2
SN  - 2214-1766
UR  - https://doi.org/10.2991/jsta.d.200304.001
DO  - 10.2991/jsta.d.200304.001
ID  - Saboori2020
ER  -