Generalized Estimators of Stationary random-coefficients Panel Data models: Asymptotic and Small Sample Properties

Transcription

1 Generalized Estimators of Stationary random-coefficients Panel Data models: Asymptotic and Small Sample Properties Mohamed Reda Abonazel Department of Applied Statistics and Econometrics, Institute of Statistical Studies and Research, Cairo University, Egypt ABSTRACT This article provides generalized estimators for the random-coefficients panel data (RCPD) model where the errors are cross-sectional heteroskedastic and contemporaneously correlated as well as with the first-order autocorrelation of the time series errors. Of course, under the new assumptions of the error, the conventional estimators are not suitable for RCPD model. Therefore, the suitable estimator for this model and other alternative estimators have been provided and examined in this article. Furthermore, the efficiency comparisons for these estimators have been carried out in small samples and also we examine the asymptotic distributions of them. The Monte Carlo simulation study indicates that the new estimators are more efficient than the conventional estimators, especially in small samples. Keywords Classical pooling estimation; Contemporaneous covariance; First-order autocorrelation; Heteroskedasticity; Mean group estimation; Random coefficient regression.. Introduction The econometrics literature reveals a type of data called panel data, which refers to the pooling of observations on a cross-section of households, countries, and firms over several time periods. Pooling this data achieves a deep analysis of the data and gives a richer source of variation which allows for more efficient estimation of the parameters. With additional, more informative data, one can get more reliable estimates and test more sophisticated behavioral models with less restrictive assumptions. Also, Panel data sets are more effective in identifying and estimating effects that are simply not detectable in pure cross-sectional or pure time series data. In particular, panel data sets are more effective in studying complex issues of dynamic behavior. Some of the benefits and limitations of using panel data sets are listed in Baltagi (203) and Hsiao (204). The pooled least squares (classical pooling) estimator for pooled cross-sectional and time series data (panel data) models is the best linear unbiased estimator (BLUE) under the classical assumptions as in the general linear regression model. An important assumption for panel data models is that the individuals in our database are drawn from a population with a common regression coefficient vector. In other words, the coefficients of a panel data model must be fixed. In fact, this assumption is not satisfied in most economic models, see, e.g., Livingston et al. (200) and Alcacer et al. (203). In this Dielman (983, 989) discussed these assumptions. In the next section in this paper, we will discuss different types of classical pooling estimators under different assumptions.

2 article, the panel data models are studied when this assumption is relaxed. In this case, the model is called random-coefficients panel data (RCPD) model". The RCPD model has been examined by Swamy in several publications (Swamy 970, 973, and 974), Rao (982), Dielman (992a, b), Beck and Katz (2007), Youssef and Abonazel (2009), and Mousa et al. (20). Some statistical and econometric publications refer to this model as Swamy s model or as the random coefficient regression (RCR) model, see, e.g., Poi (2003), Abonazel (2009), and Elhorst (204, ch.3). In RCR model, Swamy assumes that the individuals in our panel data are drawn from a population with a common regression parameter, which is a fixed component, and a random component, that will allow the coefficients to differ from unit to unit. This model has been developed by many researchers, see, e.g., Beran and Millar (994), Chelliah (998), Anh and Chelliah (999), Murtazashvili and Wooldridge (2008), Cheng et al. (203), Fu and Fu (205), Elster and Wübbeler (206), and Horváth and Trapani (206). The random-coefficients models have been applied in different fields and they constitute a unifying setup for many statistical problems. Moreover, several applications of Swamy s model have appeared in the literature of finance and economics. 2 Boot and Frankfurter (972) used the RCR model to examine the optimal mix of short and long-term debt for firms. Feige and Swamy (974) applied this model to estimate demand equations for liquid assets, while Boness and Frankfurter (977) used it to examine the concept of risk-classes in finance. Recently, Westerlund and arayan (205) used the random-coefficients approach to predict the stock returns at the ew York Stock Exchange. Swamy et al. (205) applied a random-coefficient framework to deal with two problems frequently encountered in applied work; these problems are correcting for misspecifications in a small area level model and resolving Simpson's paradox. Dziechciarz (989) and Hsiao and Pesaran (2008) classified the random-coefficients models into two categories (stationary and non-stationary models), depending on the type of assumption about the coefficient variation. Stationary random-coefficients models regard the coefficients as having constant means and variance-covariances, like Swamy s (970) model. On the other hand, the coefficients in non-stationary random-coefficients models do not have a constant mean and/or variance and can vary systematically; these models are relevant mainly for modeling the systematic structural variation in time, like the Cooley-Prescott (973) model. 3 The main objective of this article is to provide the researchers with general and more efficient estimators for the stationary RCPD modes. To achieve this objective, we propose and examine alternative estimators of these models under an assumption that the errors are cross-sectional heteroskedastic and contemporaneously correlated as well as with the first-order autocorrelation of the time series errors. The rest of the article is organized as follows. Section 2 presents the classical pooling estimators of fixed-coefficients models. Section 3 provides generalized least squares (GLS) estimators of the different random-coefficients models. In section 4, we examine the efficiency of these estimators, theoretically. In section 5, we discuss alternative estimators for these models. The Monte Carlo comparisons between various estimators have been carried out in section 6. Finally, section 7 offers the concluding remarks. 2 The RCR model has been applied also in different sciences fields, see, e.g., Bodhlyera et al. (204). 3 Cooley and Prescott (973) suggested a model where coefficients vary from one time period to another on the basis of a non-stationary process. Similar models have been considered by Sant (977) and Rausser et al. (982). 2

3 2. Fixed-Coefficients Models Suppose the variable y for the ith cross-sectional unit at time period t is specified as a linear function of K strictly exogenous variables, x kit, in the following form: K y it = k= α ki x kit + u it = x it α i + u it, i =, 2,, ; t =, 2,, T, () where u it denotes the random error term, x it is a K vector of exogenous variables, and α i is the K vector of coefficients. Stacking equation () over time, we obtain: y i = X i α i + u i, (2) where y i = (y i,, y it ), X i = (x i,, x it ), α i = (α i,, α ik ), and u i = (u i,, u it ). When the performance of one individual from the database is of interest, separate equation regressions can be estimated for each individual unit using the ordinary least squares (OLS) method. The OLS estimator of α i, is given by: Under the following assumptions, α i is a BLUE of α i : α i = (X i X i ) X i y i. (3) Assumption : The errors have zero mean, i.e., E(u i ) = 0; i =, 2,,. Assumption 2: The errors have the same variance for each individual: E(u i u j ) = { σ u 2 I T if i = j 0 if i j i, j =,2,,. Assumption 3: The exogenous variables are non-stochastic, i.e., fixed in repeated samples, and hence, not correlated with the errors. Also, rank(x i ) = K < T; i =, 2,,. These conditions are sufficient but not necessary for the optimality of the OLS estimator. 4 When OLS is not optimal, estimation can still proceed equation by equation in many cases. For example, if variance of u i is not constant, the errors are either heteroskedastic and/or serially correlated, and the GLS method will provide relatively more efficient estimates than OLS, even if GLS was applied to each equation separately as in OLS. Another case, If the covariances between u i and u j (i, j =,2,, ) do not equal to zero as in assumption (2) above, then contemporaneous correlation is present, and we have what Zellner (962) termed as seemingly unrelated regression (SUR) equations, where the equations are related through cross-equation correlation of errors. If the X i (i =, 2,, ) matrices do not span the same column space and contemporaneous correlation exists, a relatively more efficient estimator of α i than equation by equation OLS is the GLS estimator applied to the entire equation system, as shown in Zellner (962). With either separate equation estimation or the SUR methodology, we obtain parameter estimates for each individual unit in the database. ow suppose it is necessary to summarize individual relationships and to draw inferences about certain population parameters. Alternatively, the process may be viewed as building a single model to describe the entire group of individuals rather than building a separate model for each. Again, assume that assumptions -3 are satisfied and add the following assumption: 4 For more information about the optimality of the OLS estimators, see, e.g., Rao and Mitra (97, ch. 8) and Srivastava and Giles (987, pp. 7-2). 3

4 Assumption 4: The individuals in the database are drawn from a population with a common regression parameter vector α, i.e., α = α 2 = = α = α. Under this assumption, the observations for each individual can be pooled, and a single regression performed to obtain an efficient estimator of α. ow, the equation system is written as: Y = Xα + u, (4) where Y = (y,, y ), X = (X,, X ), u = (u,, u ), and α = (α,, α K) is a vector of fixed coefficients which to be estimated. We will differentiate between two cases to estimate α in (4) based on the variance-covariance structure of u. In the first case, the errors have the same variance for each individual as given in assumption 2. In this case, the efficient and unbiased estimator of α under assumptions -4 is: α CP OLS = (X X) X Y. This estimator has been termed the classical pooling-ordinary least squares (CP-OLS) estimator. In the second case, which the errors have different variances along individuals and are contemporaneously correlated as in the SUR framework: Assumption 5: E(u i u j ) = { σ iii T σ ij I T where if i = j if i j i, j =,2,,. Under assumptions, 3, 4 and 5, the efficient and unbiased CP estimator of α is: α CP SUR = [X (Σ sur I T ) X] [X (Σ sur I T ) Y], σ σ 2 σ σ 2 σ 22 σ 2 Σ sur = ( ). σ σ 2 σ To make this estimator (α CP SUR ) a feasible, the σ ij can be replaced with the following unbiased and consistent estimator: σ ij = u i u j T K ; i, j =,2,,, (5) where u i = y i X i α i, is the residuals vector obtained from applying OLS to equation number i Random-Coefficients Models This section reviews the standard random-coefficients model proposed by Swamy (970), and presents the random-coefficients model in the general case, where the errors are allowed to be crosssectional heteroskedastic and contemporaneously correlated as well as with the first-order autocorrelation of the time series errors. 3.. RCR Model Suppose that each regression coefficient in equation (2) is now viewed as a random variable; that is the coefficients, α i, are viewed as invariant over time, but varying from one unit to another: 5 The σ ij in (5) are unbiased estimators because, as assumed, the number of exogenous variables of each equation is equal, i.e., K i = K for i =,2,,. However, in the general case, K i K j, the unbiased estimator is u i u j [T K i K j + tr(p xx )], where P xx = X i (X i X i ) X i X j (X j X j ) X j. See Srivastava and Giles (987, pp. 3-7) and Baltagi (20, pp ). 4

5 Assumption 6: (for the stationary random-coefficients approach): the coefficient vector α i is specified as: 6 α i = α + μ i, where α is a K vector of constants, and μ i is a K vector of stationary random variables with zero means and constant variance-covariances: E(μ i ) = 0, and E(μ i μ j ) = { Ψ if i = j i, j =, 2,,, 0 if i j where Ψ = diag{ψ 2 k }; for k =,2,, K, where K <. Furthermore, E(μ i x jt ) = 0 and E(μ i u jt ) = 0 i and j. Also, Swamy (970) assumed that the errors have different variances along individuals: Assumption 7: E(u i u j ) = { σ iii T if i = j 0 if i j i, j =,2,,. Under the assumption 6, the model in equation (2) can be rewritten as: Y = Xα + e; e = Dμ + u, (6) where Y, X, u, and α are defined in(4), while μ = (μ,, μ ), and D = diag{x i }; for i =,2,,. The model in (6), under assumptions, 3, 6 and 7, called the RCR model, which was examined by Swamy (970, 97, 973, and 974), Youssef and Abonazel (2009), and Mousa et al. (20). We will refer to assumptions, 3, 6 and 7 as RCR assumptions. Under these assumptions, the BLUE of α in equation (6) is: where Ω is the variance-covariance matrix of e: α RCR = (X Ω X) X Ω Y, Ω = (Σ rcr I T ) + D(I Ψ )D. where Σ rcr = diag{σ ii }; for i =,2,,. Swamy (970) showed that the α RCR estimator can be rewritten as: i=. α RCR = [ i= X i (X i ΨX i + σ ii I T ) X i ] X i (X i ΨX i + σ ii I T ) y i The variance-covariance matrix of α RCR under RCR assumptions is: var(α RCR ) = (X Ω X) = { i= [Ψ + σ ii (X i X i ) ] }. To make the α RCR estimator feasible, Swamy (97) suggested using the estimator in (5) as an unbiased and consistent estimator of σ ii, and the following unbiased estimator for Ψ: Ψ = [ ( α iα i i= α i i= α i i= )] [ σ ii(x i= i X i ) ]. (7) Swamy (973, 974) showed that the estimator α RCR is consistent as both, T and is asymptotically efficient as T. 7 It is worth noting that, just as in the error-components model, the estimator (7) is not necessarily non-negative definite. Mousa et al. (20) explained that it is possible to obtain negative estimates of Swamy s estimator in (7) in case of small samples and if some/all coefficients are fixed. 6 This means that the individuals in our database are drown from a population with a common regression parameter γ, which is a fixed component, and a random component μ i, allowed to differ from unit to unit. 7 The statistical properties of γ RCR have been examined by Swamy (97), of course, under RCR assumptions. 5

6 But in medium and large samples, the negative variance estimates does not appear even if all coefficients are fixed. To solve this problem, Swamy has suggested replacing (7) by: 8 Ψ + = ( α iα i i= α i i= i= α i ). This estimator, although biased, is non-negative definite and consistent when T. See Judge et al. (985, p. 542) Generalized RCR Model To generalize RCR model so that it would be more suitable for most economic models, we assume that the errors are cross-sectional heteroskedastic and contemporaneously correlated, as in assumption 5, as well as with the first-order autocorrelation of the time series errors. Therefore, we add the following assumption to assumption 5: Assumption 8: u it = ρ i u i,t + ε it ; ρ i <, where ρ i (i =,2,, ) are fixed first-order autocorrelation coefficients. Assume that: E(ε it ) = 0, E(u i,t ε jt ) = 0; i and j, and E(ε i ε j ) = { σ ε ii I T σ εij I T if i = j if i j i, j =,2,,. This means that the initial time period the errors have the same properties as in subsequent periods, i.e. E(u 2 i0 ) = σ εii ( ρ 2 i ) and E(u i0 u j0 ) = σ εij ( ρ i ρ j ) i and j. We will refer to assumptions, 3, 5, 6, and 8 as the general RCR assumptions. Under these assumptions, the BLUE of α is: where α GRCR = (X Ω X) X Ω Y, with Ω = X ΨX + σ ε ω σ ε2 ω 2 σ ε ω σ ε2 ω 2 X 2 ΨX 2 + σ ε22 ω 22 σ ε2 ω 2, (8) σ ( ε ω σ ε2 ω 2 X ΨX + σ ε ω ) ω ij = ρ i ρ j ( 2 T ρ i ρ i ρ i T 2 ρ j ρ i ρ i. (9) T 2 T 3 ρ j ρ j ) ρ j T Since the elements of Ω are usually unknown, we develop a feasible Aitken estimator of α based on consistent estimators of the elements of Ω : T ρ i = t=2 u itu i,t T 2, (0) t=2 u i,t where u i = (u i,, u it ) is the residuals vector obtained from applying OLS to equation number i. σ εij = ε i ε j T K, 8 This suggestion has been used by Stata program, specifically in xtrchh and xtrchh2 Stata s commands. See Poi (2003). 6

7 where ε i = (ε i, ε i2,, ε it) ; ε i = u i ρ i2, and ε it = u it ρ iu i,t for t = 2,, T. Replacing ρ i by ρ i in (9), yields consistent estimators of ω ij, say ω ij, which leads together with σ εij and ω ij to a consistent estimator of Ψ: 9 where Ψ = [ ( α i α i α i i= i= α i i= )] σ ε ii (X i ω ii X i ) + ( ) σ ε ij (X i ω ii X i ) X i ω ii ω ij ω jj X j (X j ω jj X j ), i j i,j= i= () α i = (X i ω ii X i ) X i ω ii y i. (2) By using the consistent estimators (σ εij, ω ij, and Ψ ) in (8), and proceed a consistent estimator of Ω is obtained, say Ω, that leads to get the generalized RCR (GRCR) estimator of α : α GRCR = (X Ω X) X Ω Y. The estimated variance-covariance matrix of α GRCR is: 4. Efficiency Gains var (α GRCR ) = (X Ω X). (3) In this section, we examine the efficiency gains from the use of GRCR estimator. Under the general RCR assumptions, It is easy to verify that the classical pooling estimators (α CP OLS and α CP SUR ) and Swamy s estimator (α RCR ) are unbiased for α and with variance-covariance matrices: var( α CP OLS ) = G Ω G ; var(α CP SUR ) = G 2 Ω G 2 ; var( α RCR ) = G 3 Ω G 3, (4) where G = (X X) X, G 2 = [X (Σ sur I T )X] X (Σ sur I T ), and G 3 = (X Ω X) X Ω. The efficiency gains, from the use of GRCR estimator, can be summarized in the following equation: EG γ = var(α γ) var(α GRCR ) = (G h G 0 )Ω (G h G 0 ), h =,,3, where the subscript γ indicates the estimator that is used (CP-OLS, CP-SUR, or RCR), G h (for h =,,3) matrices are defined in (4), and G 0 = (X Ω X) X Ω. Since Ω, Σ rcr, Σ sur, and Ω are positive definite matrices, then EG α matrices are positive semi-definite matrices. In other words, the GRCR estimator is more efficient than CP-OLS, CP-SUR, and RCR estimators. These efficiency gains increase when ρ i, σ εij, and ψ k 2 increase. However, it is not clear to what extent these efficiency gains hold in small samples. Therefore, this will be examined in a simulation study. 5. Alternative Estimators A consistent estimator of α can also be obtained under more general assumptions concerning α i and the regressors. One such possible estimator is the mean group (MG) estimator, proposed by 9 The estimator of ρ i in (0) is consistent, but it is not unbiased. See Srivastava and Giles (987, p. 2) for other suitable consistent estimators of ρ i that are often used in practice. 7

8 Pesaran and Smith (995) for estimation of dynamic panel data (DPD) models with random coefficients. 0 The MG estimator is defined as the simple average of the OLS estimators: α MG = i= α i. (5) Even though the MG estimator has been used in DPD models with random coefficients, it will be used here as one of alternative estimators of static panel data models with random coefficients. ote that the simple MG estimator in (5) is more suitable for the RCR Model. But to make it suitable for the GRCR model, we suggest a general mean group (GMG) estimator as: where α i is defined in(2). Lemma. α GMG = i= α i, (6) If the general RCR assumptions are satisfied, then α MG and α GMG are unbiased estimators of α, with the estimated variance-covariance matrices of α MG and α GMG are: var (α MG ) = Ψ + 2 σ ε ii (X i X i ) X i ω ii X i (X i X i ) i= + 2 σ ε ij (X i X i ) X i ω ij X j (X j X j ), (7) i j i,j= var (α GMG ) = ( ) [ ( α i α i i= α i i= α i i= + σ εij (X i ω ii X i ) X i ω ii ω ij ω jj X j (X j ω jj X j ). (8) i j i,j= ] It is noted from lemma that the variance of the GMG estimator is less than the variance of the MG estimator when the general RCR assumptions are satisfied. In other words, the GMG estimator is more efficient than the MG estimator. But under RCR assumptions, we have: var(α MG ) = var(α GMG ) = ( α ( ) i= iα i α i= i i= α i ) = Ψ+. The next lemma explains the asymptotic variances (as T with fixed) properties of GRCR, RCR, GMG, and MG estimators. In order to justify the derivation of the asymptotic variances, we must assume the following: Assumption 9: plimt X i X i and plimt X i ω ii X i are finite and positive definite for all i and for ρ i <. ) 0 For more information about the estimation methods for DPD models, see, e.g., Baltagi (203), Abonazel (204), Youssef et al. (204a,b), and Youssef and Abonazel (207). 8

9 Lemma 2. If the general RCR assumptions and assumption 9 are satisfied, then the estimated asymptotic variance-covariance matrices of GRCR, RCR, GMG, and MG estimators are equal: plimvar (α GRCR ) = plimvar (α RCR ) = plimvar (α GMG ) = plimvar (α MG ) = Ψ+. From lemma 2, we can conclude that the means and the variance-covariance matrices of the limiting distributions of α GRCR, α RCR, α GMG, and α MG are the same and are equal to α and Ψ respectively even if the errors are correlated as in assumption 8. it is not expected to increase the asymptotic efficiency of α GRCR about α RCR, α GMG, and α MG. This does not mean that the GRCR estimator cannot be more efficient than RCR, GMG, and MG in small samples when the errors are correlated as in assumption 8. This will be examined in our simulation study. 6. Monte Carlo Simulation In this section, the Monte Carlo simulation has been used for making comparisons between the behavior of the classical pooling estimators (CP-OLS and CP-SUR), random-coefficients estimators (RCR and GRCR), and mean group estimators (MG and GMG) in small and moderate samples. The program to set up the Monte Carlo simulation, written in the R language, is available upon request. Monte Carlo experiments were carried out based on the following data generating process: 3 y it = k= α ki x kit + u it, i =, 2,, ; t =, 2,, T. (9) To perform the simulation under the general RCR assumptions, the model in (9) was generated as follows:. The independent variables, (x kit ; k =, 2, 3), were generated as independent standard normally distributed random variables. The values of x kit were allowed to differ for each cross-sectional unit. However, once generated for all cross-sectional units the values were held fixed over all Monte Carlo trials. 2. The errors, u it, were generated as in assumption 8: u it = ρu i,t + ε it, where the values of ε i = (ε i,, ε it ) i =, 2,, were generated as multivariate normally distributed with means zeros and variance-covariance matrix: σ εii σ εij σ εij σ εij σ εii, σ εij σ εij σ εij σ εii ( ) where the values of σ εii, σ εij, and ρ were chosen to be: σ εii = or 00; σ εij = 0, 0.75, or 0.95; and ρ = 0, 0.55, or 0.85, where the values of σ εii, σ εij, and ρ are constants for all i, j =, 2,, in each Monte Carlo trial. The initial values of u it are generated as u i = ε i ρ 2 i =, 2,,. The values of errors were allowed to differ for each cross-sectional unit on a given Monte Carlo trial and were allowed to differ between trials. The errors are independent with all independent variables. 3. The coefficients, α ki, were generated as in assumption 6: α i = α + μ i, where α = (,,), and μ i were generated from two distributions. First, multivariate normal distribution with means zeros and variance-covariance matrix Ψ = diag{ψ k 2 }; k =,2,3. The values of ψ k 2 were chosen to 9

10 be fixed for all k and equal to 5 or 25. Second, multivariate student s t distribution with degree of freedom (df): df = or 5. To include the case of fixed-coefficients models in our simulation study, we assume that μ i = The values of and T were chosen to be 5, 8, 0, 2, 5, and 20 to represent small and moderate samples for the number of individuals and the time dimension. To compare the small and moderate samples performance for the different estimators, three different samplings schemes have been designed in our simulation, where each design contains four pairs of and T. The first two represent small samples while the moderate samples are represented by the second two pairs. These designs have been created as follows: First, case of < T, the pairs of and T were chosen to be (, T) = (5, 8), (5, 2), (0, 5), or (0, 20). Second, case of = T, the pairs are (, T) = (5, 5), (0, 0), (5, 5), or (20, 20). Third, case of > T, the pairs are (, T) = (8, 5), (2, 5), (5, 0), or (20, 0). 5. All Monte Carlo experiments involved 000 replications and all the results of all separate experiments are obtained by precisely the same series of random numbers. To raise the efficiency of the comparison between these estimators, we calculate the average of total standard errors (ATSE) for each estimator by: ATSE = 000 {trace [var(α l)] 0.5 } 000 l=, where α l is the estimated vector of the true vector of coefficients mean (α ) in (9), and var(α l) is the estimated variance-covariance matrix of the estimator. The Monte Carlo results are given in Tables -6. Specifically, Tables -3 present the ATSE values of the estimators when σ εii =, and in cases of < T, = T, and > T, respectively. While case of σ εii = 00 is presented in Tables 4-6 in the same cases of and T. In our simulation study, the main factors that have an effect on the ATSE values of the estimators are, T, σ εii, σ εij, ρ, ψ k 2 (for normal distribution), and df (for student s t distribution). From Tables -6, we can summarize some effects for all estimators in the following points: When the values of and T are increased, the values of ATSE are decreasing for all simulation situations. When the value of σ εii is increased, the values of ATSE are increasing in most situations. When the values of (σ εij, ρ) are increased, the values of ATSE are increasing in most situations. When the value of ψ k 2 is increased, the values of ATSE are increasing for all situations. When the value of df is increased, the values of ATSE are decreasing for all situations. For more deeps in simulation results, we can conclude the following results:. Generally, the performance of all estimators in cases of T is better than their performance in case of > T. Similarly, Their performance in cases of σ εii = is better than the performance in case of σ εii = 00, but not as significantly better as in and T. 2. When σ εij = ρ = μ i = 0, the ATSE values of the classical pooling estimators (CP-OLS and CP-SUR) are approximately equivalent, especially when the sample size is moderate and/or T. However, the ATSE values of GMG and GRCR estimators are smaller than those of the classical pooling estimators in this situation (σ εij = ρ = μ i = 0) and other simulation situations (case of 0

11 σ εii, σ εij, ρ, ψ k 2 are increasing, and df is decreasing). In other words, GMG and GRCR are more efficient than CP-OLS and CP-SUR whether the regression coefficients are fixed or random. 3. If T 5, the values of ATSE for the MG and GMG estimators are approximately equivalent. This result is consistent with Lemma 2. According to our study, this case (T 5) is achieved when the sample size is moderate in Tables, 2, 4 and 5. Moreover, convergence slows down if σ εii, σ εij, and ρ are increased. But the situation for the RCR and GRCR estimators is different; the convergence between them is very slow even if T = 20. So the MG and GMG estimators are more efficient than RCR in all simulation situations. 4. When the coefficients are random (whether they are distributed as normal or student s t), the values of ATSE for GMG and GRCR are smaller than those of MG and RCR in all simulation situations (for any, T, σ εii, σ εij, and ρ). However, the ATSE values of GRCR are smaller than those of GMG estimator in most situations, especially when the sample size is moderate. In other words, the GRCR estimator performs better than all other estimators as long as the sample size is moderate regardless of other simulation factors. 7. Conclusion In this article, the classical pooling (CP-OLS and CP-SUR), random-coefficients (RCR and GRCR), and mean group (MG and GMG) estimators of stationary RCPD models were examined in different sample sizes for the case where the errors are cross-sectionally and serially correlated. Analytical efficiency comparisons for these estimators indicate that the mean group and random-coefficients estimators are equivalent when T is sufficiently large. Furthermore, The Monte Carlo simulation results show that the classical pooling estimators are absolutely not suitable for random-coefficients models. And, the MG and GMG estimators are more efficient than the RCR estimator for random- and fixed-coefficients models, especially when T is small (T 2). But when T 20, the MG, GMG, and GRCR estimators are approximately equivalent. However, the GRCR estimator performs better than the MG and GMG estimators in most situations, especially in moderate samples. Therefore, we conclude that the GRCR estimator is suitable to stationary RCPD models whether the coefficients are random or fixed.

12 Appendix A. Proof of Lemma a. Show that GMG and MG are unbiased estimator for α : By substituting (2) and (2) into (6): α GMG = (X i ω ii X i ) X i= i ω ii (X i α i + u i ) = α i= i + (X i ω ii X i ) X i ω ii u i, (A.) Similarly, we can rewrite α MG in (5) as: α MG = α i= i + (X i X i ) X i u i. (A.2) Taking the expectation for (A.) and (A.2), and using assumption : E(α GMG ) = E(α MG ) = i= α i = α. b. Derive the variance-covariance matrix of GMG: ote first that under assumption 6, α i = α + μ i. Add α i to the both sides: α i + α i = α + μ i + α i, α i = α + μ i + α i α i = α + μ i + τ i, where τ i = α i α i = (X i ω ii X i ) X i ω ii u i. From (A.3): i= α i = α + μ i= i + i which means that α GMG = α + μ + τ, i= τ i, (A.3) (A.4) where μ = μ i= i and τ = τ i= i. From (A.4) and using the general RCR assumptions: var(α GMG ) = var(μ ) + var(τ ) = Ψ + 2 σ ε ii (X i ω ii X i ) i= + 2 σ ε ij (X i ω ii X i ) X i ω ii ω ij ω jj X j (X j ω jj X j ). i j i,j= Using the consistent estimators of Ψ, σ εij, and ω ij defined above: var (α GMG ) = ( ) [ c. Derive the variance-covariance matrix of MG: ( α i α i α i i= As above, equation (3) can be rewritten as follows: where λ i = α i α i = (X i X i ) X i u i. From (A.5): i= α i i= + σ εij (X i ω ii X i ) X i ω ii ω ij ω jj X j (X j ω jj X j ). i j i,j= ] α i = α + μ i + λ i, ) (A.5) 2

13 which means that i= α i = α + μ i= i + i α MG = α + μ + λ, i= λ i, (A.6) where μ = μ i= i, and λ, = λ i= i. From (A.6) and using the general RCR assumptions: var(α MG ) = var(μ ) + var(λ ) = Ψ + σ ε ii (X i X i ) X i ω ii X i (X i X i ) σ ε ij (X i X i ) X i ω ij X j (X j X j ). 2 i j i,j= 2 i= + As in the GMG estimator, and by using the consistent estimators of Ψ, σ εij, and ω ij : var (α MG ) = Ψ + A.2 Proof of Lemma 2 and then, 2 i= σ ε ii (X i X i ) X i ω ii X i (X i X i ) + 2 i j i,j= σ ε ij (X i X i ) X i ω ij X j (X j X j ). Following the same argument as in Parks (967) and utilizing assumption 9, we can show that: plim α i = plim α i = α i, plim ρ ij = ρ ij, plim σ εij = σ εij, and plim ω ij = ω ij, plim T σ ε ii T(X i ω ii X i ) = plim T σ ε ii T(X i X i ) X i ω ii X i (X i X i ) = plim T σ ε ij T(X i X i ) X i ω ij X j (X j X j ) = plim T σ ε ij T(X i ω ii X i ) X i ω ii ω ij ω jj X j (X j ω jj X j ) = 0. Substituting (A.7) and (A.8) in (): (A.7) (A.8) plim Ψ = ( α i= iα i α i= i i= α i ) = Ψ +. (A.9) By substituting (A.7)-(A.9) into (7), (8), and (3): plimvar (α MG ) = plim Ψ + plim 2 i j plimvar (α GMG ) = plim ( ) ( ) i j i,j= i,j= ( α i α i i= α i 2 i= T σ ε ii T(X i X i ) X i ω ii X i (X i X i ) + plim T σ ε ij T(X i X i ) X i ω ij X j (X j X j ) = Ψ+, (A.0) i= α i i= ) + [plim T σ ε ij T(X i ω ii X i ) X i ω ii ω ij ω jj X j (X j ω jj X j ) ] = Ψ+, (A.) plimvar (α GRCR ) = plim(x Ω X) = [ i= Ψ + ] = Ψ+. (A.2) Similarly, we will use the results in (A.7)-(A.9) in case of RCR estimator: plimvar (α RCR ) = plim [(X Ω X) X Ω Ω Ω X(X Ω X) ] = Ψ+. From (A.0)-(A.3), we can conclude that: (A.3) 3

14 plimvar (α GRCR ) = plimvar (α RCR ) = plimvar (α GMG ) = plimvar (α MG ) = Ψ+. References Abonazel, M. R. (2009). Some Properties of Random Coefficients Regression Estimators. MSc thesis. Institute of Statistical Studies and Research. Cairo University. Abonazel, M. R. (204). Some estimation methods for dynamic panel data models. PhD thesis. Institute of Statistical Studies and Research. Cairo University. Alcacer, J., Chung, W., Hawk, A., Pacheco-de-Almeida, G. (203). Applying random coefficient models to strategy research: testing for firm heterogeneity, predicting firm-specific coefficients, and estimating Strategy Trade-Offs. Working Paper, o Harvard Business School Strategy Unit. Anh, V. V., Chelliah, T. (999). Estimated generalized least squares for random coefficient regression models. Scandinavian journal of statistics 26():3-46. Baltagi, B. H. (20). Econometrics. 5th ed. Berlin: Springer-Verlag Berlin Heidelberg. Baltagi, B. H. (203). Econometric Analysis of Panel Data. 5th ed. Chichester: John Wiley and Sons. Beck,., Katz, J.. (2007). Random coefficient models for time-series cross-section data: Monte Carlo experiments. Political Analysis 5(2): Beran, R., Millar, P. W. (994). Minimum distance estimation in random coefficient regression models. The Annals of Statistics 22(4): Bodhlyera, O., Zewotir, T., Ramroop, S. (204). Random coefficient model for changes in viscosity in dissolving. 59(4): pulp. Wood Research Boness, A. J., Frankfurter, G. M. (977). Evidence of on-homogeneity of capital costs within risk-classes. The Journal of Finance 32(3): Boot, J. C., Frankfurter, G. M. (972). The dynamics of corporate debt management, decision rules, and some empirical evidence. Journal of Financial and Quantitative Analysis 7(04): Chelliah,. (998). A new covariance estimator in random coefficient regression model. The Indian Journal of Statistics, Series B 60(3): Cheng, J., Yue, R. X., Liu, X. (203). Optimal Designs for Random Coefficient Regression Models with Heteroscedastic Errors. Communications in Statistics-Theory and Methods 42(5): Cooley, T. F., Prescott, E. C. (973). Systematic (non-random) variation models: varying parameter regression: a theory and some applications. Annals of Economic and Social Measurement 2(4): Dielman, T. E. (983). Pooled cross-sectional and time series data: a survey of current statistical methodology. The American Statistician 37(2):-22. Dielman, T. E. (989). Pooled Cross-Sectional and Time Series Data Analysis. ew York: Marcel Dekker. Dielman, T. E. (992a). Misspecification in random coefficient regression models: a Monte Carlo simulation. Statistical Papers 33(): Dielman, T. E. (992b). Small sample properties of random coefficient regression estimators: a Monte Carlo simulation. Communications in Statistics-Simulation and Computation 2(): Dziechciarz, J. (989). Changing and random coefficient models. A survey. In: Hackl, P., ed. Statistical Analysis and Forecasting of Economic Structural Change. Berlin: Springer Berlin Heidelberg. Elhorst, J. P. (204). Spatial Econometrics: From Cross-Sectional Data to Spatial Panels. Heidelberg, ew York, Dordrecht, London: springer. Elster, C., Wübbeler, G. (206). Bayesian inference using a noninformative prior for linear Gaussian random coefficient regression with inhomogeneous within-class variances. Computational Statistics (in press). DOI: 0.007/s Feige, E. L., Swamy, P. A. V. B. (974). A random coefficient model of the demand for liquid assets. Journal of Money, Credit and Banking, 6(2): Fu, K. A., Fu, X. (205). Asymptotics for the random coefficient first-order autoregressive model with possibly heavy-tailed innovations. Journal of Computational and Applied Mathematics 285:

15 Horváth, L., Trapani, L. (206). Statistical inference in a random coefficient panel model. Journal of Econometrics 93(): Hsiao, C. (204). Analysis of Panel Data. 3rd ed. Cambridge: Cambridge University Press. Hsiao, C., Pesaran, M. H. (2008). Random coefficient models. In: Matyas, L., Sevestre, P., eds. The Econometrics of Panel Data. Vol. 46. Berlin: Springer Berlin Heidelberg. Judge, G. G., Griffiths, W. E., Hill, R. C., Lütkepohl, H., Lee, T. C. (985). The Theory and Practice of Econometrics, 2nd ed. ew York: Wiley. Livingston, M., Erickson, K., Mishra, A. (200). Standard and Bayesian random coefficient model estimation of US Corn Soybean farmer risk attitudes. In Ball, V. E., Fanfani, R., Gutierrez, L., eds. The Economic Impact of Public Support to Agriculture. Springer ew York. Mousa, A., Youssef, A. H., Abonazel, M. R. (20). A Monte Carlo study for Swamy s estimate of random coefficient panel data model. Working paper, o University Library of Munich, Germany. Murtazashvili, I., Wooldridge, J. M. (2008). Fixed effects instrumental variables estimation in correlated random coefficient panel data models. Journal of Econometrics 42: Parks, R. W. (967). Efficient Estimation of a System of regression equations when disturbances are both serially and contemporaneously correlated. Journal of the American Statistical Association 62: Pesaran, M.H., Smith, R. (995). Estimation of long-run relationships from dynamic heterogeneous panels. Journal of Econometrics 68:79-4. Poi, B. P. (2003). From the help desk: Swamy s random-coefficients model. The Stata Journal 3(3): Rao, C. R., Mitra, S. (97). Generalized Inverse of Matrices and Its Applications. John Wiley and Sons Ltd. Rao, U. G. (982). A note on the unbiasedness of Swamy's estimator for the random coefficient regression model. Journal of econometrics 8(3): Rausser, G.C., Mundlak, Y., Johnson, S.R. (982). Structural change, updating, and forecasting. In: Rausser, G.C., ed. ew Directions in Econometric Modeling and Forecasting US Agriculture. Amsterdam: orth-holland. Sant, D. (977). Generalized least squares applied to time-varying parameter models. Annals of Economic and Social Measurement 6(3): Srivastava, V. K., Giles, D. E. A. (987). Seemingly Unrelated Regression Equations Models: Estimation and Inference. ew York: Marcel Dekker. Swamy, P. A. V. B. (970). Efficient inference in a random coefficient regression model. Econometrica 38: Swamy, P. A. V. B. (97). Statistical Inference in Random Coefficient Regression Models. ew York: Springer- Verlag. Swamy, P. A. V. B. (973). Criteria, constraints, and multicollinearity in random coefficient regression model. Annals of Economic and Social Measurement 2(4): Swamy, P. A. V. B. (974). Linear models with random coefficients. In: Zarembka, P., ed. Frontiers in Econometrics. ew York: Academic Press. Swamy, P. A. V. B., Mehta, J. S., Tavlas, G. S., Hall, S. G. (205). Two applications of the random coefficient procedure: Correcting for misspecifications in a small area level model and resolving Simpson's paradox. Economic Modelling 45: Westerlund, J., arayan, P. (205). A random coefficient approach to the predictability of stock returns in panels. Journal of Financial Econometrics 3(3): Youssef, A. H., Abonazel, M. R. (2009). A comparative study for estimation parameters in panel data model. Working paper, o University Library of Munich, Germany. Youssef, A. H., Abonazel, M. R. (207). Alternative GMM estimators for first-order autoregressive panel model: an improving efficiency approach. Communications in Statistics-Simulation and Computation 46: Youssef, A. H., El-sheikh, A. A., Abonazel, M. R. (204a). Improving the efficiency of GMM estimators for dynamic panel models. Far East Journal of Theoretical Statistics 47:7 89. Youssef, A. H., El-sheikh, A. A., Abonazel, M. R. (204b). ew GMM estimators for dynamic panel data models. International Journal of Innovative Research in Science, Engineering and Technology 3: Zellner, A. (962). An efficient method of estimating seemingly unrelated regressions and tests of aggregation bias. Journal of the American Statistical Association 57:

16 Table : ATSE for various estimators when σ εii = and < T (ρ, σ εij ) (0, 0) (0.55, 0.75) (0.85, 0.95) (, T) (5, 8) (5, 2) (0, 5) (0, 20) (5, 8) (5, 2) (0, 5) (0, 20) (5, 8) (5, 2) (0, 5) (0, 20) μ i = 0 CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 25) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t(5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t() CP-OLS CP-SUR MG GMG RCR GRCR

17 Table 2: ATSE for various estimators when σ εii = and = T (ρ, σ εij ) (0, 0) (0.55, 0.75) (0.85, 0.95) (, T) (5, 5) (0, 0) (5, 5) (20, 20) (5, 5) (0, 0) (5, 5) (20, 20) (5, 5) (0, 0) (5, 5) (20, 20) μ i = 0 CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 25) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t(5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t() CP-OLS CP-SUR MG GMG RCR GRCR

18 Table 3: ATSE for various estimators when σ εii = and > T (ρ, σ εij ) (0, 0) (0.55, 0.75) (0.85, 0.95) (, T) (8, 5) (2, 5) (5, 0) (20, 0) (8, 5) (2, 5) (5, 0) (20, 0) (8, 5) (2, 5) (5, 0) (20, 0) μ i = 0 CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 25) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t(5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t() CP-OLS CP-SUR MG GMG RCR GRCR

19 Table 4: ATSE for various estimators when σ εii = 00 and < T (ρ, σ εij ) (0, 0) (0.55, 0.75) (0.85, 0.95) (, T) (5, 8) (5, 2) (0, 5) (0, 20) (5, 8) (5, 2) (0, 5) (0, 20) (5, 8) (5, 2) (0, 5) (0, 20) μ i = 0 CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 25) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t(5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t() CP-OLS CP-SUR MG GMG RCR GRCR

20 Table 5: ATSE for various estimators when σ εii = 00 and = T (ρ, σ εij ) (0, 0) (0.55, 0.75) (0.85, 0.95) (, T) (5, 5) (0, 0) (5, 5) (20, 20) (5, 5) (0, 0) (5, 5) (20, 20) (5, 5) (0, 0) (5, 5) (20, 20) μ i = 0 CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ (0, 25) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t(5) CP-OLS CP-SUR MG GMG RCR GRCR μ i ~ t() CP-OLS CP-SUR MG GMG RCR GRCR