A comparison of some bayesian and classical procedures for simultaneous equation models with weak instruments
Abstract
EconStor is a publication server for scholarly economic literature, provided as a non-commercial public service by the ZBW.
Full text
Gao, Chuanming; Lahiri, Kajal Article A comparison of some bayesian and classical procedures for simultaneous equation models with weak instruments Econometrics Provided in Cooperation with: MDPI – Multidisciplinary Digital Publishing Institute, Basel Suggested Citation: Gao, Chuanming; Lahiri, Kajal (2019) : A comparison of some bayesian and classical procedures for simultaneous equation models with weak instruments, Econometrics, ISSN 2225-1146, MDPI, Basel, Vol. 7, Iss. 3, pp. 1-28, https://doi.org/10.3390/econometrics7030033 This Version is available at: https://hdl.handle.net/10419/247533 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/
econometrics Article A Comparison of Some Bayesian and Classical Procedures for Simultaneous Equation Models with Weak Instruments Chuanming Gao 1and Kajal Lahiri 2,* 1Fannie Mae, 1100 15th St NW, Washington, DC 20005, USA 2Department of Economics, University at Albany, SUNY, Albany, NY 12222, USA *Correspondence: [email protected] Received: 18 March 2019; Accepted: 23 July 2019; Published: 29 July 2019 Abstract: We compare the finite sample performance of a number of Bayesian and classical procedures for limited information simultaneous equations models with weak instruments by a Monte Carlo study. We consider Bayesian approaches developed by Chao and Phillips, Geweke, Kleibergen and van Dijk, and Zellner. Amongst the sampling theory methods, OLS, 2SLS, LIML, Fuller’s modified LIML, and the jackknife instrumental variable estimator (JIVE) due to Angrist et al. and Blomquist and Dahlberg are also considered. Since the posterior densities and their conditionals in Chao and Phillips and Kleibergen and van Dijk are nonstandard, we use a novel “Gibbs within Metropolis–Hastings” algorithm, which only requires the availability of the conditional densities from the candidate generating density. Our results show that with very weak instruments, there is no single estimator that is superior to others in all cases. When endogeneity is weak, Zellner’s MELO does the best. When the endogeneity is not weak and ρω12 > 0, where ρ is the correlation coefficient between the structural and reduced form errors, and ω12 is the covariance between the unrestricted reduced form errors, the Bayesian method of moments (BMOM) outperforms all other estimators by a wide margin. When the endogeneity is not weak and βρ <0 ( β being the structural parameter), the Kleibergen and van Dijk approach seems to work very well. Surprisingly, the performance of JIVE was disappointing in all our experiments. Keywords: limited information estimation; weak instruments; Metropolis–Hastings algorithm; Gibbs sampler; Monte Carlo method JEL Classification: C30; C11; C13; C15 1. Introduction Research on Bayesian analysis of the simultaneous equations models addresses a problem, raised initially by Maddala (1976), and now recognized as related to the problem of local nonidentification when diffuse/flat priors are used in traditional Bayesian analysis, e.g., Dr è ze (1976); Drèze and Morales (1976), and Dr è ze and Richard (1983). 1 In this paper, we examine the approaches developed by Chao and Phillips (1998, hereafter CP), Geweke (1996), Kleibergen and van Dijk (1998, hereafter KVD), and Zellner (1998). The idea in KVD was to treat an overidentified simultaneous equations 1 Zellner (1998) and Zellner et al. (2014) contain a comprehensive review of the finite sample properties of SEM estimators, and emphasize the need for finite sample optimal estimation procedure for such models. Andrews and Stock (2005) reviews recent developments in methods that deal with weak in IV regression models, and presents new testing results under “many weak-IV asymptotics”. Econometrics 2019,7, 33; doi:10.3390/econometrics7030033 www.mdpi.com/journal/econometrics
Econometrics 2019,7, 33 2 of 28 model (SEM) as a linear model with nonlinear parameter restrictions, and has been extended further in Kleibergen and Zivot (2003). While KVD focused mainly on resolving the problem of local nonidentification, CP explored further the consequences of using a Jeffreys prior. By deriving the exact and (asymptotically) approximate representations for the posterior density of the structural parameter, CP showed that the use of a Jeffreys prior brings Bayesian inference closer to classical inference in the sense that this prior choice leads to posterior distributions which exhibit Cauchy-like tail behavior akin to the LIML estimator. Geweke (1996), being aware of the potential problem of local nonidentification, suggests a shrinkage prior such that the posterior density is properly defined for each parameter. In another approach, Zellner (1998) suggested a finite sample Bayesian method of moments (BMOM) procedure based on given data without specifying a likelihood function or introducing any sampling assumptions. For the Bayesian approaches considered, while Geweke (1996) proposed Gibbs sampling (GS) to evaluate the posterior density with a shrinkage prior, the posterior densities as well as their conditional densities resulting from CP and KVD are nonstandard and cannot be readily simulated. In the category of “block-at-a-time” approach, we suggest a novel MCMC procedure, which we call a “Gibbs within M–H” algorithm. The advantage of this algorithm is that it only requires the availability of the conditional densities from the candidate generating density. These conditional densities are used in a Gibbs sampler to simulate the candidate generating density, whose drawings on convergence are then weighted to generate drawings from the target density in a Metropolis–Hastings (M–H) algorithm. In this study, we will focus on weak instruments, where the classical approach has been particularly problematic. 2 Ni and Sun (2003) have studied similar issues in the context of vector autoregressive models, see also Ni et al. (2007). Radchenko and Tsurumi (2006) used many of the procedures analyzed in this paper to estimate a gasoline demand model using an MCMC algorithm. The main objective of the present paper is to compare the small sample performance of some Bayesian and classical approaches using Monte Carlo simulations. For the purpose of comparison, a number of classical methods including OLS, 2SLS, LIML, Fuller’s modified LIML, and a jackknife instrumental variables estimator (JIVE) originally due to Angrist et al. (1999) and Blomquist and Dahlberg (1999) are also computed from the generated data. Our simulation results from repeated sampling experiments provide some unambiguous guidelines for empirical practitioners. The plan of the paper is as follows. In Section 2, we set up the model. Section 3reviews in limited details the recent Bayesian approaches and JIVE. Section 4suggests a new MCMC procedure for evaluating the posterior distributions for CP and KVD, and discusses the convergence diagnostics implemented. Section 5presents simulation results and some discussions. Section 6contains the main conclusions. 2. The Model Consider the following limited information formulation of the m-equation simultaneous equations model (LISEM): y1=Y2β+Z1γ+u(1) Y2=Z1Π1+Z2Π2+V2(2) where y 1 :(T × 1) and Y 2 :(T × (m − 1)) are the m included endogenous variables; Z 1 :(T × k 1 ) is an observation matrix of exogenous variables included in the structural Equation (1); Z 2 :(T × k 2 ) is an observation matrix of exogenous variables excluded from (1); and uand V 2 are, respectively, a T × 1 vector and a T × (m − 1) matrix of random disturbances to the system. We assume that (u,V 2 ) ∼ N 2 There has been a lot of interest in the estimation of LISEM with weak instruments. See Buse (1992); Bound et al. (1995); Staiger and Stock (1997); Angrist et al. (1999); Blomquist and Dahlberg (1999), among others. More recently, Andrews et al. (2019) review the literature on weak instruments in linear IV regression, and suggest that weak instruments remain an important issue in empirical practice.
Econometrics 2019,7, 33 3 of 28 (0, Σ⊗ I T ),where the m × mcovariance matrix is positive definite symmetric (pds) and is partitioned conformably with the rows of (u,V2) as follows Σ= σ11 σ0 21 σ21 Σ22 ! The likelihood function for the model described by (1) and (2) can be written as L(β,γ,Π1,Π2,Σ|Y,Z) =(2π)−Tm/2|Σ|−T/2expn−1 2trhΣ−1(u,V2)0(u,V2)io,(3) where Y=(y1,Y2)and Z=(Z1,Z2). The structural model described by (1) and (2) can alternatively be written in its reduced form y1Y2=Z1Z2 π1Π1 Π2βΠ2!+ξ1V2(4) where π1=γ+Π1β , ξ1=u+V2β , (ξ1,V2)∼N(0, Ω⊗IT) , and Σ=C0ΩC , C= 1 0 −βIm−1! . The likelihood function corresponding to this alternative representation is: L(β,γ,Π1,Π2,Ω|Y,Z)=(2π)−Tm/2|Ω|−T/2exp{−1 2trhΩ−1(ξ1,V2)0(ξ1,V2)i}(5) The likelihood functions (3) and (5) are equivalent since the Jacobian between Ωand Σis unity. Geweke (1996) considers the following reduced rank regression specification3 Y=Z1A+Z2Θ+E, (6) where A=(Π1,π1) , Θ=Π2Φ and Φ=(Im−1,β) , E=(V2,ξ1) ~ N( 0, Σ⊗IT with Σ−1 = Σ11 Σ12 Σ21 Σ22 ! partitioned conformably with the rows of (V2,ξ1) . Obviously, (6) is equivalent to (4) and the corresponding likelihood function is similar to (5). Note that in the absence of restrictions on the covariance structure, (1) is fully identified if and only if rank(Π1)=(m−1)≤k2. 3. Review of Some Bayesian Formulations Among the Bayesian approaches, Geweke (1996) used a shrinkage prior such that all parameters are identified (in the sense that a proper posterior distribution exists) even when Π2 has reduced rank. KVD treated overidentified SEMs as linear models with nonlinear parameter restrictions using the singular value decomposition. A diffuse or natural conjugate prior for the parameters of the embedding linear model results in the posterior for the parameters of the SEM having zero weight in the region of parameter space where Π2has reduced rank. This is a feature of the Jacobian of transformation from the multivariate linear model to the SEM. CP used a prior by applying Jeffreys principle on the model described by (1) and (2) and the assumptions regarding the disturbances. An important advantage of the Jeffreys prior in the present context is that it places no weight in the region of the parameter space where rank(Π2)<(m−1) and relatively low weight in close neighborhoods of this region where the model is nearly unidentified. 3 Geweke (1996) considered a more general specification. To facilitate comparison, for Geweke approach only, we have denoted Y=(Y2,y1).
Econometrics 2019,7, 33 4 of 28 3.1. Zellner’s Bayesian Method of Moments Approach (BMOM) Among the various Bayesian treatments of SEM proposed by Zellner (1971,1978,1986,1994,1998), the Bayesian method of moments approach applies the principle of maximum entropy and generates optimal estimates which can be evaluated by double K-class estimators. Given the unrestricted reduced form equation y1=Zπ1+ξ1,Zellner (1998) considered a balanced loss function, Lb=ωLg+(1−ω)Lp =ωy1−Xˆ δ0y1−Xˆ δ+(1−ω)Zπ1−Xˆ δ0Zπ1−Xˆ δ, for 0 ≤ω≤1 where X=(Y2,Z1) , δ=(β0,γ0) , and ˆ δ is an estimate of δ . The BMOM estimate that minimizes ELb , where the expectation is taken with respect to a probability density function of the π matrices of the unrestricted reduced form equations, is given by ˆ β ˆ γ!= Y0 2Y2−K1ˆ V0 2ˆ V2Y0 2Z1 Z0 1Y2Z0 1Z1!−1Y2−K2ˆ V20y1 Z0 1y1, (7) where K1= 1 −k/(T−k) , K2= 1 −(1−ω)k/(T−k) with 0 ≤ω≤ 1 and ˆ V2=I−Z(Z0Z)−1Z0Y2 . BMOM estimate will vary depending on the value of ω . When ω= 1, it is the optimal estimate resulting from a “goodness of fit” loss function L g . When ω =0, it is the optimal estimate given by a precision of estimation loss function L p . Meanwhile, the well-known minimum expected loss (MELO) estimator is derived using a precision of estimation loss function and may be evaluated as a K-class estimator with K1=K2=1−k/(T−k−m−1). Similar to the BMOM method, Conley et al. (2008) developed a Bayesian semiparametric approach to the instrumental variable problem assuming linear structural and reduced form equations, but with unspecified error distributions. 3.2. The Geweke Approach Geweke (1996) assumes the following reference prior Σ−(m+v+1)/2exp−1 2trSΣ−1exp"−τ2 2β0β+trΠ0 2Π2+tr A0A#, (8) which is the product of an independent inverted Wishart distribution for Σ with v degrees of freedom and scale matrix S, and an independent N(0, τ2 ) shrinkage priors for each element of β and Π2 . Geweke derived the respective conditional posterior distributions, which may be used to generate drawings through Gibbs sampling from the joint posterior distribution. Regarding the vector of parameters Σ−1, A, Π2,β, we obtain the full conditional densities as follows: (1) Conditional density of Σ−1 Σ−1(Π2,β, A, Z,Y)∼Wishart T+v,G−1, (9) where G=S+(Y−Z1A−Z2Θ)0(Y−Z1A−Z2Θ).
Econometrics 2019,7, 33 5 of 28 (2) Conditional density of A vec(A)Π2,β,Σ−1,Z,Y ∼N(hΣ−1⊗Z0 1Z1+τ2Imk1i−1hΣ−1⊗Z10Z1ivec(ˆ A), hΣ−1⊗Z0 1Z1+τ2Imk1i−1), (10) where ˆ A=Z0 1Z1−1Z0 1(Y−Z2Θ). (3) Conditional density of Π24 vec(Π2)β,Σ−1, A, Z,Y ∼N(e Σ11 ⊗Z0 2Z2+τ2Ik2(m−1)−1e Σ11 ⊗Z0 2Z2vec(ˆ Π2), e Σ11 ⊗Z0 2Z2+τ2Ik2(m−1)−1 ), (11) where ˆ Π2=ˆ Θ"Φ++Φ0e Σ21e Σ11−1# , ˆ Θ=Z0 2Z2−1Z0 2(Y−Z1A) . Φ+ is the Moore–Penrose generalized inverse of Φ and the columns of Φ+ and Φ0 are orthogonal, and C ≡ ( Φ+ , Φ0) is and m × m nonsingular matrix. Finally, e Σij denotes the partitioning of e Σ−1=(C0ΣC)−1 conformably with Y=(Y2,y1). (4) Conditional density of β βΠ2,Σ−1, A, Z,Y ∼NhΣ22 ⊗Π0 2Z0 2Z2Π2+τ2Im−1i−1hΣ22 ⊗Π0 2Z0 2Z2Π2iˆ β,hΣ22 ⊗Π0 2Z0 2Z2Π2+τ2Im−1i−1, (12) where ˆ β=Π0 2Z0 2Z2Π2−1Π0 2Z0 2(Y2−Z1Π1)Σ12Σ22−1−Σ12Σ22−1+Π0 2Z0 2Z2Π2−1Π0 2Z0 2(y1−Z1π1) 3.3. The Chao and Phillips Approach Using Jeffreys prior, CP obtains exact and approximate analytic expressions for the posterior density of the structural coefficient β in the LISEM (1) and (2). Their formulas are found to exhibit Cauchy-like tails analogous to comparable results in the classical literature on LIML estimation. For the model (1) and (2) under normality assumption for the disturbances, a Jeffreys prior on the parameters, θ=(β,γ,Π1,Π2,Σ),is of the form p(β,γ,Π1,Π2,Σ)∝−E(∂2 ∂θ∂θ0ln L(θ|Y,Z)) 1/2 ∝|σ11|(k2−m+1)/2|Σ|−(k+m+1)/2Π0 2Z0 2QZ1Z2Π21/2 (13) where ln L(θY,Z) is the log-likelihood function as specified in (3), and QX=IT−PX , PX= X(X0X)−1X0 . As first noted by Poirier (1996), the prior in (13) places no weight where rank(Π2)<(m−1) through the factor Π0 2Z0 2QZ1Z2Π21/2. 4 The expressions for the conditional densities of Π2 and β given in (Geweke 1996, expressions (11) and (13)) contain some typographical errors and are corrected here in (11) and (12).
Econometrics 2019,7, 33 6 of 28 The joint posterior of the parameters of LISEM (1) and (2) is constructed as proportional to the product of the prior (13) and the likelihood function (3), p(β,γ,Π1,Π2,ΣY,Z)∝p(β,γ,Π1,Π2,Σ)L(β,γ,Π1,Π2,Σ|Y,Z) ∝|σ11|(k2−m+1)/2|Σ|−(T+k+m+1)/2Π0 2Z0 2QZ1Z2Π21/2×exp {−1 2trhΣ−1(u,V2)0(u,V2)i}(14) where (u,V 2 ) is defined in (1) and (2). Note that (14) or its conditionals do not belong to any standard class of probability density functions. 3.4. The Kleibergen and van Dijk Approach To solve the problem of local nonidentification and to avoid the so-called Borel–Kolmogorov paradox, see Billingsley (1986) and Poirier (1995), KVD considered (4) as a multivariate linear model with nonlinear parameter restrictions: y1Y2=Z1Z2 π1Π1 φ2Φ2!+ξ1V2, (15) where φ1 is a k2× 1 vector, Φ2 is a k2×(m−1) matrix. Denote Φ=(φ1,Φ2) . The reduced form model (4) is obtained if a reduced rank restriction is imposed on the linear model (15) such that rank(Φ)=(m−1)instead of m. Using a singular value decomposition (SVD) of Φ , they show that (15) is identical to the so-called unrestricted reduced form (URF) model,5 y1Y2=Z1π1Π1+Z2Π2B+Z2Π2⊥λB⊥+ξ1V2, (16) where B=βIm−1 , λ is a (k2−m−1)× 1 vector. Π2⊥ and B⊥ are the orthogonal complements of Π2 and B, respectively, such that Π0 2Π2⊥≡ 0, BB0 ⊥≡ 0, and Π0 2⊥Π2⊥≡Ik2−m−1 , B⊥B0 ⊥≡ 1 (i.e., Π2⊥= −Π22Π−1 21 Ik2−m−10Ik2−m−1+Π22Π−1 21 Π−10 21 Π22−1/2 , where Π2=Π0 21 Π0 220 , Π21 : (m−1)×(m−1) , Π22 :(k2−m−1)×(m−1), and B⊥=(1+β0β)1/2(1−β0)). There is one-to-one correspondence between the parameters in (15) and (16). The SVD of Φis, Φ=USV0, (17) where U : k2×k2 , U0U=Ik2 ; V : m×m ; V0V=Im ; and S : k2×m is a rectangular matrix containing the (nonnegative) singular values (in decreasing order) on its main diagonal (i.e., (s11, s11,. .. ,smm)). Rewrite U= U11 U12 U21 U22 !,= S10 0 S2!, and V = v11 v12 v21 v22 !, (18) where U11 , S1 , v21 : (m−1)×(m−1) ; v12 : 1 × 1; v0 11 , v22 : (m−1)× 1; U12 : (m−1)×(k2−m+1)× (m−1) ; U21 : (k2−m−1)×(m−1) ; U22 : (k2−m−1)×(k2−m+1) ; S2 : (k2−m−1)× 1, then the following relationship between ( Π2 , β , λ ) and (U, S, V) results, Π2= U11 U12 !S1V0 21 , β=V0−1 21 v0 11 , and λ=U22U0 22−1/2U22S2v0 12v12v0 12−1/2. (19) 5 Note that this formulation or the singular value decomposition does not change the identification status of the LISEM specified by (1) and (2). If rank(Π2)<(m−1),βis locally nonidentified.
Econometrics 2019,7, 33 7 of 28 Note that λ is obtained through preand postmultiplication of s 2 by orthogonal matrices while s 2 contains the smallest singular values of Φ and is invariant with respect to the ordering of variables contained in Yand Z2. According to KVD, the above shows that the model described by (1) and (2) can be considered as equivalent to the linear model (16) with a nonlinear (reduced rank) restriction λ= 0 on the parameters. Therefore, the priors and posteriors of the parameters of the LISEM (1) and (2) may be constructed as proportional to the priors and posteriors of the parameters of the linear model (16) evaluated at λ= 0. A diffuse (Jeffreys) prior for the parameters (π1,Π1,Φ,Ω) of the linear model6 p(π1,Π1,Φ,Ω)∝|Ω|−(k+m+1)/2∝|Ω|−(m+1)/2Ω−1⊗Z0Z1/2(20) where k=k1+k2, implies the prior for the parameters (β,π1,Π1,Π2,Ω) of the LISEM (4) as p(β,π1,Π1,Π2,Ω)∝p(π1,Π1Φ(Π2,β,λ),Ω)|λ=0|J(Φ,(Π2,β,λ))λ=0 ∝|Ω|−(m+1)/2Ω−1⊗Z0Z1/2J(Φ,(Π2,β,λ))λ=0 ∝|Ω|−(m+1)/2Ω−1⊗Z0Z1/2 ×B0⊗Ik2e1⊗Π2B0 ⊥⊗Π2⊥, (21) where e1=(1, 0, 0, . . . , 0)0 . Note that the prior (21) is the Jeffreys prior of the unrestricted reduced form (16) evaluated at λ= 0. Most importantly, |J( Φ , ( Π2 , β , λ ))| λ=0 =0 when Π2 has reduced rank. This feature in KVD approach eliminates the potential impact of local nonidentification. The joint posterior of the parameters of the LISEM (4) is readily constructed as proportional to the product of the prior (21) and the likelihood function (5), p(β,π1,Π1,Π2,ΩY,Z) ∝p(β,π1,Π1,Π2,Ω)L∗(β,γ,Π1,Π2,ΩY,Z)∝|Ω|−(T+m+1)/2Ω−1⊗Z0Z1/2 ×B0⊗Ik2e1⊗Π2B0 ⊥⊗Π2⊥ ×exp{−1 2tr[Ω−1(y1Y2−Z1Z2 π1Π1 Π2βΠ2!)0 (y1Y2−Z1Z2 π1Π1 Π2βΠ2!)]} (22) Unfortunately, the above posterior or its conditional densities do not belong to a known class of probability density functions. 3.5. The Jackknife Instrumental Variable Estimator (JIVE) Motivated by split sample instrumental variables estimators, Angrist et al. (1999) and Blomquist and Dahlberg (1999) independently suggested a jackknife instrumental variable estimator (JIVE). For model (1) and (2), JIVE is given by ˆ δjive =ˆ X0 jiveX−1ˆ X0 jivey1(23) where ˆ Xjive is the T×(m−1+k1)matrix with t-th row defined by Ztˆ Π−t=ZtZ0 −tZ−t−1Z0 −tX−t=Ztˆ Π−htXt 1−ht, 6 This is the prior suggested in Dr è ze (1976). Zellner (1971) and Zellner et al. (1988) used a similar prior with − (m+1)/2 in the exponent.
Econometrics 2019,7, 33 8 of 28 Z−t and X−t are (T−1)×k and (T−1)×(m−1+k1) matrices obtained after eliminating the t-th rows of Z and X matrices respectively, ˆ Π=(Z0Z)−1(Z0X) and ht=Zt(Z0Z)−1Z0 t . In JIVE, the instrument is independent of the disturbances even in finite samples, which is achieved by using a ‘leave-one-out’ jackknife-type fitted value in place of the usual unrestricted reduced form predictions. Angrist et al. (1999) also proposed a second jackknife estimator that is a slight modification of (23). Similar to their study, we found that its performance is very similar to JIVE, and is not reported here. 4. Posterior Simulator: “Gibbs within M–H” Algorithm Given the full conditional densities in (9) through (12) for the four blocks of parameters, evaluating the joint posterior densities by Gibbs sampling is straightforward, see Geweke (1996) for a detailed description. Although Geweke’s (1996) shrinkage prior does not meet the argument in KVD that the implied prior/posterior on the parameters of an embedding linear model should be well-behaved, we found that the use of Geweke’s shrinkage prior does not lead to a reducible Markov Chain. With the specification of a shrinkage prior, when Π2 has reduced rank, the joint posterior density still depends on β and will not exhibit any asymptotic cusp. In the following, we only discuss the posterior simulation for CP and KVD. KVD suggested two simulation algorithms for the posterior (22): an Importance sampler and a Metropolis–Hastings algorithm. We found that their M–H algorithm performs unsatisfactorily with low acceptance rate even for reasonable parameter specifications. 7 As mentioned earlier, since the posteriors (14) and (22) as well as their conditional posteriors do not belong to any standard class of probability density functions, Gibbs sampling cannot be used. In this section, we suggest an alternative simulation algorithm which combines Gibbs sampling (see Casella and George (1992) and Chib and Greenberg (1996)) and the Metropolis–Hastings algorithm (see Metropolis et al. 1953; Hastings 1970;Smith and Roberts 1993;Tierney 1994;Chib and Greenberg 1995). Our algorithm is different from the “M–H within Gibbs” algorithm and can find its usefulness in other applications as well. To generate drawings from the target density p(x), we use a candidate-generating density r(x). An Independence sampler, which is a special case of the M–H sampler, in algorithmic form is as follows: 0. Choose starting values x0 1. Draw xifrom r(x) 2. Accept xiwith probability αxi−1,xi= min p(xi)r(xi−1) p(xi−1)r(xi), 1, if pxi−1rxi>0 1, if pxi−1rxi=0, (24) otherwise xi=xi−1 3. i=i+1. Go to 1. It is generally not feasible to draw all elements of the vector x simultaneously. A block-at-a-time possibility was first discussed in (Hastings 1970, sct. 2.4) and then in Chib and Greenberg (1995) along with an example. Chib and Greenberg (1995) considered applying the M–H algorithm in turn to sub-blocks of the vector x , which presumes that the target density p(x) may be manipulated to generate full conditional densities for each of the sub-blocks of x , conditioning on other elements of x. However, the full 7 Zellner et al. (2014) suggested a variant of this approach called Acceptance-Rejection within Direct Monte Carlo (ARDMC) to evaluate the posterior density, and report substantial gain in computational efficiency, particularly with weak instruments. They also studied the existence conditions for posterior moments of the parameters of interest in terms of the number of available instruments being greater than the number of endogenous variables plus the order of the moment.
Econometrics 2019,7, 33 15 of 28 Table 3. Cont. Mean Std RMSE MAD KVD_Mode 1.351 0.384 0.520 0.389 KVD_Median 1.381 0.367 0.529 0.405 Notes: Number of replications: 400. Geweke: nburn =100, n =2000. CP: nburn_GS =100, nburn_MH =100, n=10,000 , acceptance rate =0.475 (0.010). KVD: nburn_GS =100, nburn_MH =100, n =3000, acceptance rate =0.400 (0.217). Table 4. T=50, ρ=0.60, k2=9, R2=0.10. Mean Std RMSE MAD OLS 1.535 0.111 0.546 0.535 2SLS 1.363 0.221 0.425 0.371 MELO 1.463 0.139 0.483 0.463 BMOM 0.969 0.132 0.136 0.106 LIML 1.090 0.864 0.869 0.534 Fuller1 1.182 0.479 0.512 0.366 Fuller4 1.302 0.291 0.419 0.333 JIVE 0.706 0.933 0.978 0.728 Geweke_Mode 1.357 0.239 0.430 0.367 Geweke_Median 1.350 0.245 0.427 0.361 LIML_GS_Mode 1.375 0.218 0.328 0.380 LIML_GS_Median 1.367 0.228 0.432 0.374 CP_Mode 1.215 0.629 0.665 0.466 CP_Median 1.255 0.388 0.464 0.346 KVD_Mode 1.550 0.376 0.666 0.556 KVD_Median 1.573 0.322 0.657 0.576 Notes: Number of replications: 400. Geweke: nburn =100, n =1000. CP: nburn_GS =200, nburn_MH =200, n=10,000 , acceptance rate =0.242 (0.040). KVD: nburn_GS =200, nburn_MH =100, n =10,000, acceptance rate =0.267 (0.188). Table 5. T=100, ρ=0.60, k2=4, R2=0.10. Mean Std RMSE MAD OLS 1.538 0.077 0.543 0.538 2SLS 1.138 0.208 0.250 0.200 MELO 1.257 0.156 0.301 0.264 BMOM 0.954 0.156 0.163 0.127 LIML 1.023 0.280 0.281 0.210 Fuller1 1.069 0.250 0.259 0.197 Fuller4 1.171 0.195 0.259 0.209 JIVE 0.914 0.320 0.331 0.262 Geweke_Mode 1.149 0.215 0.262 0.208 Geweke_Median 1.111 0.228 0.254 0.198 LIML_GS_Mode 1.162 0.205 0.261 0.209 LIML_GS_Median 1.117 0.225 0.254 0.199 CP_Mode 1.155 0.207 0.259 0.206 CP_Median 1.107 0.228 0.252 0.196 KVD_Mode 1.233 0.205 0.310 0.258 KVD_Median 1.215 0.210 0.301 0.243 Notes: Number of replications: 400. Geweke: nburn =100, n =2000. CP: nburn_GS =200, nburn_MH =200, n=10,000 , acceptance rate =0.616 (0.008). KVD: nburn_GS =200, nburn_MH =100, n =10,000, acceptance rate =0.312 (0.175).
Econometrics 2019,7, 33 16 of 28 Table 6. T=100, ρ=0.60, k2=9, R2=0.10. Mean Std RMSE MAD OLS 1.542 0.078 0.548 0.542 2SLS 1.258 0.197 0.325 0.274 MELO 1.376 0.134 0.399 0.376 BMOM 0.972 0.132 0.135 0.110 LIML 1.003 0.437 0.437 0.291 Fuller1 1.071 0.311 0.319 0.243 Fuller4 1.180 0.233 0.294 0.232 JIVE 0.927 0.408 0.414 0.333 Geweke_Mode 1.253 0.201 0.323 0.269 Geweke_Median 1.238 0.206 0.315 0.261 LIML_GS_Mode 1.265 0.196 0.330 0.278 LIML_GS_Median 1.247 0.202 0.319 0.266 CP_Mode 1.196 0.264 0.329 0.266 CP_Median 1.192 0.232 0.301 0.240 KVD_Mode 1.371 0.278 0.464 0.382 KVD_Median 1.395 0.269 0.478 0.397 Notes: Number of replications: 400. Geweke: nburn =100, n =1000. CP: nburn_GS =200, nburn_MH =200, n=6000 , acceptance rate =0.434 (0.029). KVD: nburn_GS =200, nburn_MH =200, n =10,000, acceptance rate =0.210 (0.179). Table 7. T=100, ρ=0.60, k2=4, R2=0.05. Mean Std RMSE MAD OLS 1.565 0.080 0.571 0.565 2SLS 1.254 0.282 0.380 0.309 MELO 1.376 0.184 0.419 0.379 BMOM 0.953 0.183 0.189 0.150 LIML 1.052 0.584 0.586 0.392 Fuller1 1.158 0.377 0.409 0.307 Fuller4 1.296 0.244 0.384 0.317 JIVE 0.833 0.638 0.659 0.527 Geweke_Mode 1.264 0.285 0.388 0.314 Geweke_Median 1.224 0.316 0.387 0.305 LIML_GS_Mode 1.274 0.283 0.394 0.320 LIML_GS_Median 1.232 0.310 0.387 0.306 CP_Mode 1.263 0.295 0.395 0.318 CP_Median 1.223 0.316 0.387 0.304 KVD_Mode 1.388 0.389 0.549 0.418 KVD_Median 1.394 0.315 0.504 0.414 Notes: Number of replications: 400. Geweke: nburn =100, n =2000. CP: nburn_GS =100, nburn_MH =100, n=4000 , acceptance rate =0.611 (0.009). KVD: nburn_GS =200, nburn_MH =200, n =8000, acceptance rate =0.442 (0.224). Table 8. T=100, ρ=0.60, k2=9, R2=0.05. Mean Std RMSE MAD OLS 1.574 0.076 0.579 0.574 2SLS 1.386 0.219 0.444 0.394 MELO 1.478 0.131 0.496 0.478 BMOM 0.979 0.129 0.131 0.105 LIML 1.139 0.882 0.893 0.545 Fuller1 1.224 0.477 0.527 0.389 Fuller4 1.335 0.280 0.437 0.358 JIVE 0.844 0.823 0.838 0.663 Geweke_Mode 1.385 0.243 0.455 0.395 Geweke_Median 1.380 0.246 0.453 0.390
Econometrics 2019,7, 33 17 of 28 Table 8. Cont. Mean Std RMSE MAD LIML_GS_Mode 1.397 0.230 0.459 0.404 LIML_GS_Median 1.387 0.236 0.453 0.396 CP_Mode 1.338 0.465 0.575 0.433 CP_Median 1.337 0.311 0.459 0.376 KVD_Mode 1.584 0.462 0.745 0.592 KVD_Median 1.608 0.368 0.711 0.610 Notes: Number of replications: 400. Geweke: nburn =100, n =2000. CP: nburn_GS =200, nburn_MH =200, n=10,000 , acceptance rate =0.433 (0.035). KVD: nburn_GS =200, nburn_MH =200, n =10,000, acceptance rate =0.371 (0.221). Table 9. T=100, ρ=0.20, k2=4, R2=0.10. Mean Std RMSE MAD OLS 1.172 0.090 0.194 0.174 2SLS 1.046 0.253 0.257 0.206 MELO 1.083 0.189 0.206 0.164 BMOM 0.859 0.190 0.237 0.195 LIML 1.017 0.333 0.333 0.260 Fuller1 1.029 0.298 0.299 0.236 Fuller4 1.059 0.235 0.242 0.192 JIVE 0.957 0.417 0.419 0.340 Geweke_Mode 1.053 0.251 0.257 0.200 Geweke_Median 1.041 0.267 0.270 0.214 LIML_GS_Mode 1.058 0.244 0.251 0.197 LIML_GS_Median 1.044 0.265 0.269 0.212 CP_Mode 1.054 0.255 0.261 0.205 CP_Median 1.040 0.271 0.274 0.218 KVD_Mode 1.131 0.368 0.391 0.237 KVD_Median 1.161 0.328 0.365 0.245 Notes: Number of replications: 400. Geweke: nburn =100, n =1000. CP: nburn_GS =100, nburn_MH =100, n=5000 , acceptance rate =0.615 (0.011). KVD: nburn_GS =100, nburn_MH =100, n =1000, acceptance rate =0.548 (0.200). Table 10. T=100, ρ=0.20, k2=9, R2=0.10. Mean Std RMSE MAD OLS 1.179 0.096 0.203 0.181 2SLS 1.085 0.214 0.230 0.182 MELO 1.124 0.146 0.192 0.154 BMOM 0.823 0.143 0.228 0.193 LIML 0.992 0.397 0.397 0.301 Fuller1 1.015 0.347 0.347 0.270 Fuller4 1.055 0.267 0.273 0.216 JIVE 0.991 0.481 0.481 0.390 Geweke_Mode 1.084 0.218 0.234 0.184 Geweke_Median 1.079 0.223 0.237 0.187 LIML_GS_Mode 1.087 0.212 0.229 0.181 LIML_GS_Median 1.082 0.218 0.233 0.185 CP_Mode 1.054 0.308 0.313 0.223 CP_Median 1.063 0.254 0.262 0.207 KVD_Mode 1.249 0.234 0.342 0.283 KVD_Median 1.286 0.235 0.370 0.308 Notes: Number of replications: 400. Geweke: nburn =100, n =1000. CP: nburn_GS =100, nburn_MH =200, n=5000 , acceptance rate =0.456 (0.023). KVD: nburn_GS =100, nburn_MH =100, n =5000, acceptance rate =0.413 (0.202).
Econometrics 2019,7, 33 18 of 28 Table 11. T=50, ρ=0.95, k2=4, R2=0.10. Mean Std RMSE MAD OLS 1.846 0.052 0.848 0.846 2SLS 1.359 0.180 0.402 0.363 MELO 1.572 0.118 0.584 0.572 BMOM 1.057 0.118 0.131 0.102 LIML 0.988 0.404 0.404 0.255 Fuller1 1.169 0.196 0.259 0.221 Fuller4 1.417 0.120 0.434 0.417 JIVE 0.637 0.611 0.711 0.478 Geweke_Mode 1.347 0.302 0.460 0.358 Geweke_Median 1.277 0.377 0.468 0.305 LIML_GS_Mode 1.338 0.155 0.372 0.345 LIML_GS_Median 1.252 0.194 0.318 0.281 CP_Mode 1.314 0.162 0.353 0.325 CP_Median 1.234 0.194 0.304 0.266 KVD_Mode 1.411 0.379 0.559 0.428 KVD_Median 1.462 0.463 0.654 0.514 Notes: Number of replications: 400. Geweke: nburn =100, n =3000. CP: nburn_GS =200, nburn_MH =200, n=10,000 , acceptance rate =0.476 (0.010). KVD: nburn_GS =200, nburn_MH =200, n =10,000, acceptance rate =0.036 (0.038). Table 12. T=100, ρ=0.95, k2=4, R2=0.10. Mean Std RMSE MAD OLS 1.850 0.033 0.851 0.850 2SLS 1.230 0.126 0.262 0.234 MELO 1.414 0.094 0.425 0.414 BMOM 1.044 0.095 0.105 0.082 LIML 1.025 0.170 0.172 0.132 Fuller1 1.095 0.142 0.171 0.143 Fuller4 1.264 0.099 0.282 0.265 JIVE 0.873 0.199 0.236 0.191 Geweke_Mode 1.216 0.117 0.246 0.223 Geweke_Median 1.150 0.127 0.197 0.172 LIML_GS_Mode 1.227 0.118 0.256 0.235 LIML_GS_Median 1.158 0.128 0.203 0.180 CP_Mode 1.221 0.116 0.250 0.228 CP_Median 1.154 0.127 0.200 0.176 KVD_Mode 1.258 0.207 0.331 0.280 KVD_Median 1.252 0.294 0.387 0.260 Notes: Number of replications: 400. Geweke: nburn =100, n =3000. CP: nburn_GS =200, nburn_MH =200, n=10,000 , acceptance rate =0.626 (0.007). KVD: nburn_GS =200, nburn_MH =200, n =10,000, acceptance rate =0.022 (0.022). Table 13. Performance of BMOM and KVD when ρ<0. Mean Std RMSE MAD Remarks T=50, ρ=−0.60, k2=4, R2=0.40 BMOM 0.852 0.129 0.196 0.165 Compare Table 1. KVD_Mode 0.971 0.150 0.152 0.119 Acceptance rate for KVD_Median 0.999 0.153 0.152 0.119 KVD: 0.713 (0.130) T=50, ρ=−0.60, k2=4, R2=0.10 BMOM 0.551 0.191 0.488 0.453 Compare Table 3. KVD_Mode 0.851 0.327 0.359 0.271 Acceptance rate for KVD_Median 0.934 0.341 0.347 0.267 KVD: 0.680 (0.133)
Econometrics 2019,7, 33 19 of 28 Table 13. Cont. Mean Std RMSE MAD Remarks T=50, ρ=−0.60, k2=9, R2=0.10 BMOM 0.420 0.136 0.600 0.580 Compare Table 4. KVD_Mode 0.857 0.367 0.393 0.296 Acceptance rate for KVD_Median 0.927 0.399 0.406 0.291 KVD: 0.482 (0.155) T=100, ρ=−0.60, k2=4, R2=0.10 BMOM 0.676 0.160 0.362 0.326 Compare Table 5. KVD_Mode 0.901 0.213 0.235 0.186 Acceptance rate for KVD_Median 0.964 0.237 0.239 0.190 KVD: 0.772 (0.110) T=100, ρ=−0.60, k2=9, R2=0.10 BMOM 0.531 0.129 0.486 0.469 Compare Table 6. KVD_Mode 0.903 0.240 0.258 0.200 Acceptance rate for KVD_Median 0.952 0.247 0.252 0.198 KVD: 0.614 (0.138) T=100, ρ=−0.60, k2=4, R2=0.05 BMOM 0.514 0.181 0.519 0.486 Compare Table 7. KVD_Mode 0.813 0.306 0.358 0.285 Acceptance rate for KVD_Median 0.908 0.362 0.373 0.287 KVD: 0.720 (0.128) T=100, ρ=−0.60, k2=9, R2=0.05 BMOM 0.407 0.131 0.608 0.593 Compare Table 8. KVD_Mode 0.848 0.424 0.450 0.312 Acceptance rate for KVD_Median 0.907 0.349 0.361 0.275 KVD: 0.585 (0.144) T=100, ρ=−0.20, k2=4, R2=0.10 BMOM 0.753 0.195 0.314 0.266 Compare Table 9. KVD_Mode 1.002 0.267 0.267 0.208 Acceptance rate for KVD_Median 1.037 0.291 0.293 0.218 KVD: 0.699 (0.162) T=100, ρ=−0.20, k2=9, R2=0.10 BMOM 0.673 0.159 0.364 0.328 Compare Table 10. KVD_Mode 1.093 0.318 0.331 0.233 Acceptance rate for KVD_Median 1.129 0.279 0.307 0.241 KVD: 0.553 (0.181) T=50, ρ=−0.95, k2=4, R2=0.10 BMOM 0.427 0.120 0.585 0.573 Compare Table 11. KVD_Mode 0.737 0.244 0.359 0.312 Acceptance rate for KVD_Median 0.836 0.246 0.295 0.239 KVD: 0.173 (0.112) T=100, ρ=−0.95, k2=4, R2=0.10 BMOM 0.589 0.097 0.422 0.411 Compare Table 12. KVD_Mode 0.815 0.155 0.241 0.209 Acceptance rate for KVD_Median 0.889 0.153 0.189 0.156 KVD: 0.179 (0.103) Notes: Number of replications: 500. Table 13 summarizes the results for cases with ρ < 0 for BMOM and KVD for whom negative ρ made a surprising difference. As mentioned before, we focus on the estimates of the structural parameter β . Specifically, we analyze the sensitivity of the various estimates of β with respect to the strength of the instrumental variables Z, the degree of overidentification (k2−m+1) , the degree of endogeneity (ρ) , and the sample size (T). Also, we will examine whether the performance of
Econometrics 2019,7, 33 20 of 28 an estimator is symmetric with respect to the sign of parameter ρ , an issue generally overlooked in the literature.14 Note that the strength of the instrumental variables for the included endogenous variable Y 2 is measured in terms of the adjusted R2 by regressing Y 2 on Z=(Z 1 ,Z 2 ). In the data generating process, we controlled R2 to be within ± 2.5% of the specified value to reduce unnecessary variation. We did not experiment with extremely small R2 (say, 0.01 or less). In these cases, the mean values of all estimators approached the point of concentration ω12/ω22 , which is equal to (β+ρ) for our data generating process (DGP). For each specification, the number of replications is 400. The number of burn-ins (nburn_GS and nburn_MH), and subsequent number of iterations (n) determined at the convergence diagnosis step are reported in the footnotes to each table. The average acceptance rate and its standard deviation (in parentheses) across replications for each M–H routine are reported as well. To evaluate alternative estimators, we computed mean, standard deviation (Std), root of mean squared errors (RMSE), and mean absolute deviation (MAD) over repeated experiments for all the estimators considered. 15 Since LIML, posterior densities for CP and KVD, as well as 2SLS in the just-identified case do not have finite moments of positive order in finite samples, one should interpret the computed mean, standard deviation and RMSE across replications for these estimators with caution. In this sense, the MAD across replications is a preferred measure to consider. We will first look at cases reported in Tables 1–12 with ρ > 0. In Table 1, we consider a case (T=50, ρ =0.60, k 2 =4) with moderately strong instruments ( R2= 0.40). It is found that with reasonably strong instruments all estimators designed for simultaneous equations perform reasonably well. As expected, OLS is seriously biased. BMOM has a slight edge over others in terms of RMSE and MAD. For all Bayesian approaches and LIML-GS, the medians perform a little better than modes, and CP over KVD, in terms of bias, RMSE, and MAD. Notice that the classical LIML estimates are different from LIML-GS (mode or median). As noted by Dr è ze (1976), from a Bayesian viewpoint, LIML produces an estimate of β conditionally on the overidentifying restrictions, the modal values of all the remaining parameters, and a uniform prior. In other words, the concentrated likelihood function of β after concentrating out (i.e., maximizing with respect to) other reduced-form and nuisance parameters is a conditional density. However, LIML-GS is a marginal density with all other parameters being integrated out. Due to possible asymmetry in the distribution of the nuisance parameters, the modal/median values of LIML-GS may not coincide with classical LIML estimates. In all our experiments, we find that the median-unbiasedness property of (conditional) LIML does not carry over to the marginal LIML (i.e., LIML-GS); however, the former generally has a much larger standard deviation than the latter. In a way, LIML-GS brings the classical LIML estimator close to its Bayesian counterpart for the purpose of comparison. It is interesting to note that across all our tables, the difference between LIML-GS and CP can only be attributed to the importance of Jeffreys prior. Compared to LIML-GS, typically CP has a smaller bias, but slightly larger standard deviation, even though the differences are very small. In some cases, 14 Denote Ω="ω11 ω12 ω12 ω22 # . Using Σ=C0ΩC , we have σ12 =ω11 − 2 βω12 +β2ω22 , σ12 =ω12 −βω22 , and σ22 =ω22 . Letting ρ=σ12/√σ11σ22, the second relationship may be rewritten as: β−ω12 ω22 =−ρrσ11 ω22 If Σ is normalized as in (35) with σ11 =ω22 = 1, then ω12 = β + ρ . Therefore, in our context, given β =1, the sign and magnitude of ρ(or ω12) has a special significance. 15 Medians were also calculated. Since they were very close to the corresponding means in all our experiments, we did not report them in this paper.
Econometrics 2019,7, 33 21 of 28 however, the use of Jeffreys prior reduces the bias in CP quite substantially. For example, in Table 4 with T=50 and a high degree of overidentification, the bias is reduced from 0.36 to 0.25. A simple case when the structural model is just identified (k 2 =1) is reported in Table 2. For this case it is well known that classical LIML coincides with 2SLS. The KVD approach does not accommodate the case of just-identification since (15) requires k 2 >(m − 1). 16 In this case, we find that CP-Mode produces results closer to LIML-GS-Mode than to LIML. CP (1998) showed that for a two-equation just-identified SEM in orthonormal canonical form, the posterior density of β with Jeffreys prior has precisely the same functional form as the density of the finite sample distribution of the corresponding LIML estimator as obtained by Mariano and McDonald (1979). Our simulation results show that the assumption of orthonormal canonical form is crucial for their exact correspondence, which cannot be extended to a general SEM. 17 In general, the Bayesian marginal density is not the same as the classical conditional density. Interestingly, JIVE is considerably more biased and has larger standard deviation than 2SLS. Also, CP-Median and LIML-GS-Median perform significantly worse than their modes. This is because in an exactly identified model with weak instruments, the probability of local nonidentification is substantial, and the resulting nonstandard marginal density exhibits a very high variance. The same result holds true for Geweke-Median, but to a lesser extent. Thus, for exactly identified SEMs with very weak instruments, mode of the marginal density is a more dependable measure of β . We should point out that in all other cases in this study, the medians generally turned out to be more preferable than the modes in terms of bias, RMSE, and MAD (see Tables 11 and 12, for instance). Results reported in Tables 3–12 consider cases with general overidentification and weak instruments. As noted in the literature, OLS and 2SLS are median-biased in the direction of the correlation coefficient ρ , and the bias in 2SLS grows with the degree of over identification, and decreases as sample size increases. Results in Tables 3–10 confirm these results. Since MELO is a single K-class estimator with 0 <K<1, its performance is always between OLS and 2SLS estimates. The bias in MELO shows the same pattern as that of 2SLS. With moderate simultaneity, the median-bias in 2SLS can be as large as about 40% of the true value (see Table 8). We note that MELO, LIML-GS-Mode, and KVD-Mode or KVD-Median are also median-biased in the direction of ρ . However, the bias in JIVE is consistently in the opposite direction of ρ . Classical LIML is remarkably median-unbiased when the instrumental variables are not very weak, which is well documented in the literature. We find that LIML is median-biased in the direction of ρ when the instruments are very weak (Table 8), which is consistent with the finding in Staiger and Stock (1997) using local-to-zero asymptotic theory. Even in this situation, the bias of LIML is much smaller than that of any other estimator, except BMOM. The MAD of OLS is very close to its bias (i.e., relatively small Std) across all cases and it implies that OLS method is robust in the sense that it does not suffer from heavy tails or outlying estimates, see Zellner (1998). In this sense, MELO and BMOM are all robust with relatively small standard deviations across replications. However, OLS exhibits large bias in the presence of simultaneity and is not so appealing. It is known that for a degree of overidentification strictly less than seven, 2SLS would have a smaller asymptotic mean squared error (AMSE) than LIML, cf. Mariano and McDonald 16 When k2=(m−1), a diffuse prior in (20) for the linear model implies that the prior for the parameters of the LISEM (4) is p(β,π1,Π1,Π2,Ω)∝|Ω|−(k+m+1)/2|Π2|, and the prior for the parameters of the LISEM (1) and (2) is p(β,γ,Π1,Π2,Σ)∝|Σ|−(k+m+1)/2|Π2| which is identical to the Jeffreys prior; see also expressions (22) and (42) in CP. 17 Note that the relationship between the standardized parameter vector and the original parameter vector involves the nuisance parameters, cf. Chao and Phillips (1998). However, when a SEM is in orthonormal canonical form (i.e., the exogenous regressors are orthonormal and the disturbance covariance matrix Ω is an identity matrix), both the density of random parameter β from the CP approach and the probability density of the classical LIML estimator for β are conditional on these information.
Econometrics 2019,7, 33 22 of 28 (1979) and references therein. In cases with weak instruments the situation gets more complicated in finite samples. In our experiments, LIML has larger RMSE and MAD than 2SLS except in Tables 11 and 12 where ρwas 0.95. Note that the degree of overidentification is 8.0 in Tables 4,6,8and 10. Among classical estimators, JIVE turns out to be least appealing. Monte Carlo simulations in Angrist et al. (1999) showed that JIVE has slight median bias in the opposite direction of ρ (but less than 2SLS) and have heavier tails than LIML. Our Table 6is comparable to panel 2 of their Table 1, and the results are similar. Our other experiments show that JIVE may also have large absolute bias (larger than LIML) in the case with weak instruments, sometimes even greater than 2SLS (see Table 2). Generally, JIVE has slightly less bias than 2SLS, but this gain is overshadowed by enlarged standard deviation such that in finite samples it has no advantage over 2SLS in terms of MAD and RMSE. We also find that JIVE has greater RMSE and MAD than LIML. Blomquist and Dahlberg (1999) experimented with much larger sample sizes than ours. Comparing our Table 4with Table 6and with an unreported simulation with a sample size of 500, we found that the relative gain in JIVE is more than other estimators as sample size increases, even though its relative low standing remains valid. Examined from different angles, these results are very similar to those reported by Davidson and MacKinnon (2006a,2006b). 18 Fuller’s modified LIML estimators are included because Fuller1 is designed to minimize the median-bias, and Fuller4 to minimize the mean-squared error. It seems that this conclusion is also problematic in the presence of weak instruments. Between the two, Fuller1 has smaller median-bias, and Fuller4 has smaller standard deviations across replications. However, in terms of RMSE or MAD, Fuller4 shows no advantage over Fuller1 in most of the cases. Because all the estimators except OLS are consistent and their asymptotic distributions are also the same, results in Tables 3–6confirm that their bias and dispersion decrease as sample size increases. But if the instruments are very weak (see Tables 7and 8), their bias and dispersion may remain significant, a point emphasized forcefully by Zellner (1998). However, when the endogeneity is not strong (see Tables 9and 10), their bias and dispersion may not be a big concern for some of the estimators. Across all cases, we find that the bias in BMOM is small if ρ is not too small and the structural Equation (1) is overidentified. As sample size increases or degree of over-identification rises, the observed bias in BMOM decreases. The most striking feature of BMOM is that it exhibits the smallest MAD and Std when ρ is not too small. MELO shows slightly smaller MAD and Std than BMOM if ρ is small (see Tables 9and 10). In cases with very weak instruments and a high degree of overidentification, the MAD of BMOM is only one-fourth of that of other estimators (see Table 8). These are in accordance with Tsurumi (1990)’s finding that in many cases, ZEM has the least relative mean absolute deviation. Meanwhile, if ρ is very small and the structural equation is overidentified, the bias in BMOM can be large; 2SLS, LIML-GS, Geweke, and CP perform remarkably well in these situations. Next, we examine in more detail the performance of the Bayesian approaches. Overall, the median bias resulting from these approaches exhibits the same pattern as the bias of 2SLS, it increases with the degree of overidentification, and decreases as sample size rises. The Geweke (1996) approach used a shrinkage prior but its performance is comparable with LIML-GS and CP. The median-bias from PMOD-Geweke is the same or slightly less than that of LIML-GS-Mode, and the bias from Geweke-Median is always slightly less than that of LIML-GS-Median. Similar relationships are observed for MADs. These reflect the impact of the (informative) shrinkage prior on the posterior density. For each specification, the acceptance rate in the M–H algorithm using CP approach is stable while that using KVD approach shows huge variation across replications. The acceptance rate for CP is generally above 40%, except when sample size is small and the degree of overidentification is high. This shows that the posterior of CP is largely dominated by the likelihood function (3) and the Jeffreys 18 Ackerberg and Devereux (2006) and Blomquist and Dahlberg (2006) have suggested some ad hoc adjustments to the original JIVE formula to improve its performance.
Econometrics 2019,7, 33 23 of 28 prior generally carries little information. Second, in terms of the computed standard deviations (Stds) of the estimates across replications, CP-Mode has larger dispersion than LIML-GS-Mode, and CP-Median has larger dispersion than LIML-GS-Median. These also shed light on the notion that Jeffreys prior is less informative than a uniform prior. However, between the Jeffreys prior (13) used by CP and the implied prior (21) resulting from diffuse/Jeffreys prior on a linear model used by KVD, it is not clear which one is less informative. As for the KVD (1998) approach, we observe that it performs as well as any other estimator if the instruments are not weak (see Table 1). But when the instruments are weak, and ρ is positive, KVD shows more bias and higher MAD than those from CP. In Table 4with T=50 and high degree of overidentification, KVD performs as bad as OLS. Next, we consider cases with negative ρ , and the results are summarized in Table 13. We replicate each case in Tables 1–12 with the same specification except ρ being negative. Since the performance of all estimators except BMOM and KVD were basically the same with respect to the sign of ρ , we only report results on these two in Table 13. We find that when ρ changes sign, the bias of BMOM does not change sign and even increases in magnitude. Also note that the computed Stds for BMOM when ρ < 0 are close to the respective ones when ρ > 0. Therefore, for cases with ρ < 0, BMOM has large RMSEs/MADs and loses its attraction. Note that BMOM is the same as the double K-class estimator (DKC) with K values fixed. This asymmetry in the performance in DKC is not well recognized in the literature, and has been discussed in Gao and Lahiri (2001). The observed asymmetry in its bias with respect to ρ in our experiments is readily explained by examining an expression for the mean of double K-class estimator (DKC) in (Dwivedi and Srivastava (1984, Theorem (1)). We can express ˆ δDKC as: ˆ δDKC =ˆ δK1+ Y0 2Y2−K1ˆ V0 2ˆ V2Y0 2Z1 Z0 1Y2Z0 1Z1!−1 (K1−K2)ˆ V0 2y1 0!, (36) where ˆ δK1 is a single K-class estimator with characterizing scalar K1 . When Z0 1Z2= 0, which is satisfied in our experimental specifications, a double K-class estimator of βmay be written as ˆ βDKC =ˆ βK1+(K1−K2)Y0 2QZy1 Y0 2∆Y2 where ∆=(1−K1)Qz1+K1PZ2 . Observe that for 0 <K1< 1, βˆ K1 is biased in the direction of ρ , as noted in Mariano and McDonald (1979). Note also that Y0 2∆Y2> 0, and Y0 2QZy1 provides an estimate of ω12 . Although Dwivedi and Srivastava (1984) explored the dominance of double K-class over K-class using the exact MSE criterion, their guidelines for the selection of K 2 for a given K 1 are not entirely valid, because the conditions were derived from a small Monte Carlo simulation with positive ω12 and negative ρ only. Since K 1 <K 2 for BMOM, when ρ and ω12 have the opposite sign, the second term in ˆ βDKC will be of the same sign as the bias of ˆ βK1 , therefore ˆ βDKC (hence BMOM) will exhibit large bias. Otherwise, when ρ . ω12 > 0, the bias is mitigated. Based on our simulation results, we found that the sign of ρ has no effect on the standard deviation of BMOM. This finding shows that the greater RMSE of BMOM when ρω12 <0 is due to the aggravated bias. For the specification corresponding to Table 4in Table 13 (i.e., T=50, ρ = − 0.60, K 2 =4, R2 =0.10), we find that for given K 1 =0.947, RMSE is minimized if K 2 is chosen to be 0.829, which is much less than K 1 , and less than K 2 =0.987 used in BMOM. See Gao and Lahiri (2001) for further details. In Tables 3–12 we found that KVD with ρ >0 performs very poorly, often with substantial bias and relatively high RMSE and MAD. CP uniformly dominates KVD in these cases. However, with ρ<0 the picture turns around remarkably well in favor of KVD. As we see in Table 13, across all cases the bias tends to be negative and relatively small. With other parameter values being the same, KVD with ρ <0 has significantly less RMSE and MAD than cases when ρ >0, and performs unequivocally the best among all estimators when endogeneity is strong. However, since this observed asymmetry is essentially a finite sample problem with KVD, the improved performance when ρ <0 becomes less
Econometrics 2019,7, 33 24 of 28 significant when the sample size increases from 50 to 100. With ρ <0 the overall performance of KVD is very comparable to that of CP, if not slightly better in some cases. After experimenting with widely different negative and positive values of β and ρ , we found out that the performance of KVD is dependent on the sign of βρ , rather than on the sign of ρ . When βρ >0 , it performs very unsatisfactorily as documented in Tables 3–12.Kleibergen and Zivot (2003) have derived exact analytical expressions for the conditional densities of β given Ω for both the KVD and CP posteriors. They show that the difference between the two is in the Jacobian relating the unrestricted linear multivariate model to the restricted reduced form model. We expect that this additional term may account for the asymmetry in KVD with respect to βρ . In our experiments, we found that in finite samples, when βρ >0, the reduced rank restriction using singular value decomposition shifts the marginal posterior for KVD away from the marginal posterior of the linear multivariate model. However, when the sample size gets large, the problem seems to go away. 6. Conclusions This paper examines the relative merits of some contemporary developments in the Bayesian and classical analysis of limited information simultaneous equations models in situations where the instruments are very weak. Since the posterior densities and their conditionals in the Bayesian approaches developed by Chao and Phillips (1998) and Kleibergen and van Dijk (1998) are nonstandard, we proposed and implemented a “Gibbs within Metropolis–Hastings” algorithm, which only requires the availability of the conditional densities from the candidate-generating density. These conditional densities are used in a Gibbs sampler (GS) to simulate the candidate generating density, whose drawings, after convergence, are then weighted to generate drawings from the target density in a Metropolis–Hastings (M–H) algorithm. We rely on Raftery and Lewis (1992) to determine the number of burn-ins, and the subsequent number of required iterations in order to ensure convergence. Through a MCMC simulation study, our results provide useful guidelines for empirical practitioners. The first comforting result is that with reasonably strong instruments (marginal R2 in excess of 0.40), all estimators perform equally well in finite samples. In cases with very weak instruments (marginal R2 less than 0.10), there is no single estimator that is superior to others in all cases—a conclusion also reached by Andrews and Stock (2005). When endogeneity is weak ( ρ less than 0.20), Zellner’s MELO does the best. When the endogeneity is relatively strong ( ρ in excess of 0.60) and ρω12 >0, BMOM outperforms all other estimators by wide margins. When the endogeneity is strong but βρ <0, the KVD approach seems to get very appealing; but, otherwise, its performance is surprisingly poor. With βρ >0, as the sample size gets larger, the performance of KVD improves rapidly. Fortunately, the Geweke and CP approaches exhibit no such asymmetry and their performances based on bias, RMSE, and MAD are very similar. Based on the medians of marginal posteriors, their performance ranking is consistently a distant second. The record of JIVE is quite disappointing across all our experiments and is not recommended in practice. Even though JIVE is slightly less biased than 2SLS in most cases, its standard deviation is considerably higher, particularly in small samples. The most remarkable result in this study is that poor instruments can affect the performance of different estimators differently, depending on the signs and magnitudes of certain key parameters of the model. Given the finding that even in finite samples with very weak instruments BMOM and KVD perform remarkably well on certain parts of the parameter space, more research is needed to understand the reasons for the asymmetry and find ways to fix the problem. Another important caveat of our comparative study is that it was done in an iid setting. Heteroskedastic and autocorrelated errors, particularly in highly leveraged regressions, can affect inferences based on alternative instrument variable regressions differentially relative to ordinary least squares, see Young (2019). These issues remain unresolved. Author Contributions: The two authors contributed equally to this work. Funding: This research received no external funding.