A NOTE ON THE DERIVATION OF THEIL'S BLUS RESIDUALS Gregory ¢. Chow ECONOMETRIC RESEARCH PROGRAM Research Memorandum No. 176 March 1975 ‘The research described in this paper was supported by NSF Grant Gs43747X. Paper typed by Alice Furth. Econometric Research Program PRINCETON UNIVERSITY 207 Dickinson Hall Princeton, New JerseyA NOTE ON THE DERIVATION OF ‘THEIL'S BLUS RESIDUALS Gregory C. Chow* In two papers the results of which are summarized in his Principles of Econometrics (1, Chapter 5], H. Theil has proposed a best linear unbiased (BLU) estimator of the residuals ina standard linear regression which has a scalar (S) covariance matrix (of the form ot). Theil's proof in the form of seven Lemma's [1, pp. 209-213] and other related theorens appears complicated and can be further motivated. This note provides a simple con~ structive proof of the BLUS residuals. To set up the problem, consider n observations of a regression model with k explanatory variables mM yrxBte where Eee’ = Ic”. For a linear estimator cy to be unbiased, we have @ E(Cy) = BC(KB + €) = CXB= 0, which implies @ K=O. would like to thank Donald T.Sant and T. James Trussell for useful comments and to acknowledge financial support from the National Science Foundation (NSF Grant 6843747x) .Tt follows from (3) (1, p. 207, Theorem 5.5] that the linear estimator can be written in three alternative forns «a cy = ce = ce where e is the vector of residuals by the method of least squares ) es It xxx Tee Im order for Cy = Ce to have a scalar covariance matrix 6) B(cee'c') = ecto” = 107, we require ” eclar. The restriction (3), with X having k columns, is a set of k linear restrictions on the columns of C. Therefore, the rank of C cannot exceed nck. We thus let ¢ bean (n-k) xn matrix and Cy be an estimator of only n-k components of ¢ . Accordingly Theil partitions model (1) as e) - + e ewith X, being a nonsingular kxk matrix and x, being an (nck)%k matrix. 1 ‘The linear estimator can be written as (9) ce= Ce, + ce), ing (nok) x(n-) ising X3e wee, = -(xx5))' ©) being (nek)x(n-k) . Using xXJe, + Xje) = 0, we have e, = -(%)¥>")'e) . 2 Z'e, , where 2=X,x)". Further using CX, +¢,x,= 0, we have Cc, = 1 -C,X)X5" = -C,2 . (9) thus becomes 0) Ge = (-0,2) (-Z'e,) + Ce) = CIT + Z2"Je, . The constraint (7) for a scalar covariance matrix yields an Coch + CC} = CyzZtCH + CyCy 1 cfr + agtict =r. ‘The linear estimator is alternatively written as (12) Coeo + Oy8y = -CyZe, + Che, ‘The vector of errors in using (12) to estimate ¢, has a covariance matrix 2 ‘yer eee aa) cov[-cyze, + (C,-Te,] = ofc, (I + 22")VCy + 2 on. To find the matrix Cc, in the estimator (10) which minimizes the trace of (13), following Theil's definition of being "best," subject to the constraint (11) for a scalar covariance matrix, I would propose to ‘form the Lagrangean expression(4) L = tric, (t422")cf + I - c, - Ci] - tr Mic, (x4Zz")Ct - 1) Where M is a symmetric (n-k)x(n-k) matrix of Lagrange multipliers, and Gifferentiate with respect to C, . Using the differentiation rule dtr (AB) /3tr(BA)/3A = BY , we have as) 3 = 2c, re22") - 21 - amc, (r422") = 0 act To solve equations (15) and (11) for the two unknowns C, and M , we post- multiply (15) by Cj and use (11) to obtain ae) M=I- Substituting (16) for M in (15), one gets a7 Cje, (re2z") = 2. since ¢, is symmetric because in (16) is symmetric, (27) implies c2 = (re224) 72 or 2 (1g) (x42g") 7/7 which is the solution. Theil has written the solution in the form of Cy POP! where P is a matrix consisting of the characteristics vectors of (a#22") corresponding to the roots dj? , D having diagonal elements 4, and (I+ZZ') = pop", The reader may consult Theil [1] for further treatment of this topic.REFERENCE, (1] ‘Theil, H: Principles of Econometrics. New York, John Wiley & Sons, Inc., igs71.