Kobe University Repository : Kernel Title PMSE performance of the biased estimators in a linear regression model when relevant regressors are omitted Author(s) Namba, Akio Citation Econometric Theory, 18( 5): 1086-1098 Issue date 2002-10 Resource Type Journal Article / 学術雑誌論文 Resource Version publisher URL http://www.lib.kobe-u.ac.jp/handle_kernel/90000343 Create Date: 2014-10-21 Econometric Theory, 18, 2002, 1086–1098+ Printed in the United States of America+ DOI: 10+1017+S0266466602185033 PMSE PERFORMANCE OF THE BIASED ESTIMATORS IN A LINEAR REGRESSION MODEL WHEN RELEVANT REGRESSORS ARE OMITTED AK I O NA M B A Kobe University In this paper, we consider a linear regression model when relevant regressors are omitted+ We derive the explicit formulae for the predictive mean squared errors ~PMSEs! of the Stein-rule ~SR! estimator, the positive-part Stein-rule ~PSR! estimator, the minimum mean squared error ~MMSE! estimator, and the adjusted minimum mean squared error ~AMMSE! estimator+ It is shown analytically that the PSR estimator dominates the SR estimator in terms of PMSE even when there are omitted relevant regressors+ Also, our numerical results show that the PSR estimator and the AMMSE estimator have much smaller PMSEs than the ordinary least squares estimator even when the relevant regressors are omitted+ 1. INTRODUCTION In the context of linear regression, the Stein-rule ~SR! estimator proposed by Stein ~1956! and James and Stein ~1961! dominates the ordinary least squares ~OLS! estimator in terms of predictive mean squared error ~PMSE! if the model is specified correctly+ Further, as is shown in Baranchik ~1970!, the SR estimator is further dominated by the positive-part Stein-rule ~PSR! estimator when the specified model is correct+ As an improved estimator, Theil ~1971! proposes the minimum mean squared error ~MMSE! estimator+ Because Theil’s ~1971! MMSE estimator includes unknown parameters, Farebrother ~1975! suggests an operational variant of the MMSE estimator that is obtained by replacing the unknown parameters of the MMSE estimator by the OLS estimators+ Hereafter, we use the term MMSE estimator to denote the operational variant of the MMSE estimator+ Because the MMSE estimator satisfies Baranchik’s ~1970! condition, it dominates the OLS estimator if the number of regressors is larger than or equal to three and if The author is grateful to Kazuhiro Ohtani for his helpful guidance and to Peter C+B+ Phillips and anonymous referees for their useful comments and suggestions+ Address correspondence to: Akio Namba, Graduate School of Economics, Kobe University, Rokko, Nada-ku, Kobe 657-8501, Japan; e-mail: namba@kobe-u+ac+jp+ 1086 © 2002 Cambridge University Press 0266-4666002 $9+50 PERFORMANCE OF THE BIASED ESTIMATORS 1087 the model is specified correctly+ As an extension of the MMSE estimator, Ohtani ~1996! considers the adjusted minimum mean squared error ~AMMSE! estimator that is obtained by adjusting the degrees of the freedom of the component of the MMSE estimator+ These estimators are called biased estimators because they are not unbiased even if the model is specified correctly+ However, in most of the practical situations, it is hard to determine which regressors should be included in the model+ Thus, a researcher may exclude relevant regressors mistakenly+ If relevant regressors are omitted, even the OLS estimator is not unbiased+ In such situations, there may be a strong incentive to use the biased estimators if we consider that they are superior to the OLS estimator in some criterion+ However, there is little research on the properties of biased estimators when relevant regressors are omitted in the specified model+ Some exceptions are Mittelhammer ~1984! and Ohtani ~1993, 1998!+ Mittelhammer ~1984! shows that the SR estimator no longer dominates the OLS estimator when relevant regressors are omitted+ Ohtani ~1993! derives the formulae for the PMSEs of the SR estimator and the PSR estimator in the misspecified model and derives a sufficient condition for the PSR estimator to dominate the SR estimator+ Also, Ohtani ~1998! derives the formulae for the PMSEs of the MMSE estimator and the AMMSE estimator+ His numerical results show that the PMSE performance of the AMMSE estimator is better than that of the PSR estimator when the number of regressors included in the specified model is less than or equal to five and the model misspecification is severe+ However, in Ohtani ~1993, 1998!, the formulae for the PMSEs of the SR, PSR, MMSE, and AMMSE estimators are not correct+ Thus, in this paper, we limit our attention to PMSE and compare the sampling performances of the OLS, SR, PSR, MMSE, and AMMSE estimators though there are several other criteria+ ~As to other criteria, see, e+g+, Gourieroux and Monfort, 1995+! The plan of this paper is as follows+ In Section 2 the model and the estimators are presented, and in Section 3 we derive the explicit formulae for the PMSEs of the SR, PSR, MMSE, and AMMSE estimators+ It is shown analytically that the PSR estimator dominates the SR estimator in terms of PMSE even when relevant regressors are omitted+ In Section 4, using the formulae derived in Section 3, we compare the PMSEs of the estimators by numerical evaluations+ 2. THE MODEL AND THE ESTIMATORS Consider a linear regression model, y 5 X 1 b1 1 X 2 b2 1 e, e ; N~0, s 2 In !, (1) where y is an n 3 1 vector of observations on a dependent variable, X 1 and X 2 are n 3 k 1 and n 3 k 2 matrices of observations on nonstochastic independent variables, b1 and b2 are k 1 3 1 and k 2 3 1 vectors of regression coefficients, 1088 AKIO NAMBA and e is an n 3 1 vector of normal error terms+ We assume that X 1 and @X 1 , X 2 # are of full column rank+ Suppose that the matrix of regressors X 2 is omitted mistakenly and the model is specified as y 5 X 1 b1 1 h, h 5 X 2 b2 1 e+ (2) Then, based on the misspecified model, the ordinary least squares ~OLS! estimator of b1 is 21 ' b1 5 S11 X 1 y, (3) where S11 5 X 1' X 1 + Also, the SR, PSR, MMSE, and AMMSE estimators based on the misspecified model are, respectively, S bSR1 5 1 2 F D ae1' e1 b1 , b1' S11 b1 bPS1 5 max 0,1 2 bM1 5 bAM1 5 S S (4) G D ae1' e1 b1 , b1' S11 b1 (5) b1' S11 b1 b1 , b1' S11 b1 1 e1' e1 0n1 (6) b1' S11 b1 0k 1 b1 , b1' S11 b1 0k 1 1 e1' e1 0n1 (7) D where e1 5 y 2 X 1 b1 , n1 5 n 2 k 1 , and a is a constant such that 0 # a # 2~k 1 2 2!0~n1 1 2!+ If the model is specified correctly, the PMSE of the SR estimator is minimized when a 5 ~k 1 2 2!0~n1 1 2!+ Thus, we use this value of a hereafter+ 3. PMSE OF THE ESTIMATORS To derive the explicit formulae for the PMSEs of the SR, PSR, MMSE, and AMMSE estimators, we consider the general pretest estimator defined as S bZ 1 5 I ~F $ c! 1 1 a e1' e1 b1' S11 b1 D r b1 , (8) where I ~A! is an indicator function such that I ~A! 5 1 if an event A occurs and I ~A! 5 0 otherwise, F 5 ~b1' S11 b1 0k 1 !0~e1' e1 0n1 ! is the test statistic for the null hypothesis H0 : b1 5 0 against the alternative H1 : b1 Þ 0 based on the misspecified model, c is the critical value of the pretest, and r is an arbitrary integer+ The term bZ 1 reduces to the SR estimator when c 5 0, a 5 2a, and r 5 1, and it reduces to the PSR estimator when c 5 an1 0k 1 , a 5 2a, and r 5 1+ Furthermore, bZ 1 reduces to the MMSE estimator when c 5 0, a 5 10n1 , and r 5 21, PERFORMANCE OF THE BIASED ESTIMATORS 1089 and it reduces to the AMMSE estimator when c 5 0, a 5 k 1 0n1 , and r 5 21, respectively+ The PMSE of bZ 1 is defined as PMSE @ bZ 1 # 5 E @~X 1 bZ 1 2 Xb! ' ~X 1 bZ 1 2 Xb!# 5 E @ bZ 1' S11 bZ 1 # 2 2E @ b ' X ' X 1 bZ 1 # 1 b ' Sb, (9) where X 5 @X 1 , X 2 # , b ' 5 @ b1' , b2' # , and S 5 X ' X+ The meaning and some discussions of the PMSE are given in Mittelhammer ~1984! and Ohtani ~1993!+ Denoting u 1 5 b1' S11 b1 0s 2 and u 2 5 e1' e1 0s 2 , ~9! reduces to FS FS DS DS n1 u 1 $c k1 u2 PMSE @ bZ 1 # 5 s 2 E I 2 2E I n1 u 1 $c k1 u2 u 1 1 au 2 u1 D G D G 2r u1 u 1 1 au 2 u1 r b ' X ' X 1 b1 1 b ' Sb+ (10) We define the functions H~ p, q; a, c! and J ~ p, q; a, c! as FS FS DS DS H~ p, q; a, c! 5 E I n1 u 1 $c k1 u2 u 1 1 au 2 u1 J ~ p, q,; a, c! 5 E I n1 u 1 $c k1 u2 u 1 1 au 2 u1 D G D p q u1 , p (11) , q G u 1 ~ b ' X ' X 1 b1 0s 2 ! , (12) where p and q are arbitrary integers+ Then, we obtain PMSE @ bZ 1 #0s 2 5 H~2r,1; a, c! 2 2J ~r,0; a, c! 1 l 1 1 l 2 , (13) 21 ' 21 ' where l 1 5 b ' X ' X 1 S11 X 1 Xb0s 2, l 2 5 b ' X ' M1 Xb0s 2, M1 5 In 2 X 1 S11 X1 , ' 2 and l 1 1 l 2 5 b Sb0s + As is shown in the Appendix, the explicit formulae of H~ p, q; a, c! and J ~ p, q; a, c! are ` H~ p, q; a, c! 5 ` ( ( wi ~l 1 !wj ~l 2 !Gij ~ p, q; a, c!, i50 j50 ` J ~ p, q; a, c! 5 l 1 ( (14) ` ( wi ~l 1 !wj ~l 2 !Gi11, j ~ p, q; a, c!, (15) i50 j50 where Gij ~ p, q; a, c! 5 2 q G~~k 1 1 n1 !02 1 q 1 i 1 j ! G~k 1 02 1 i !G~n1 02 1 j ! 3 E 1 t k1 022p1q1i21 ~1 2 t ! n1 021j21 @a 1 ~1 2 a!t # p dt, c* wi ~l! 5 exp~2l02!~l02! i0i!, and c * 5 k 1 c0~n1 1 k 1 c!+ (16) 1090 AKIO NAMBA The formula for J ~ p, q; a, c! is substantially different from the formula derived in Theorem 1 of Ohtani ~1993!+ In Ohtani ~1993!, there are several unnecessary terms in the formula for the PMSEs of the SR estimator and the PSR estimator+ When c * [ ~0,1!, the integral in ~16! can be written as E 1 t k1 022p1q1i21 ~1 2 t ! n1 021j21 @a 1 ~1 2 a!t # p dt c* ` 5 ( l50 3 S k 1 02 2 p 1 q 1 i 2 1 l E ` ( l50 S ~21! l 1 ~1 2 t ! n1 021l1j21 @a 1 ~1 2 a!t # p dt c* 5 D k 1 02 2 p 1 q 1 i 2 1 l D ~21! l ~1 2 c * ! n1 021j1l n1 02 1 j 1 l 3 2 F1 ~2p, n1 02 1 j 1 l; n1 02 1 j 1 l 1 1; ~1 2 a!~1 2 c * !!, (17) where 2 F1~{,{;{;{! is the hypergeometric function+ ~For the definition and properties of the hypergeometric function, see, e+g+, Luke, 1969; Abramowitz and Stegun, 1972; Abadir, 1999+! In the case of the MMSE and AMMSE estimators, the hypergeometric function in ~17! converges because 6~1 2 a! 3 ~1 2 c * !6 , 1+ In the case of the SR and PSR estimators, it is a finite series with p 1 1 terms because p is a positive integer+ Thus, ~17! converges for all the estimators considered in this paper+ Also, when 6~1 2 a!~1 2 c * !6 . 1, analytic continuation formulae of the hypergeometric function make ~17! convergent+ ~The author is deeply grateful to one of the referees for suggesting that ~17! be derived and instructing the properties of the hypergeometric function+! Let a 5 2a and r 5 1+ We have PMSE~ bZ 1 !0s 2 5 H~2,1; 2 a, c! 2 2J ~1,0; 2 a, c! 1 l 1 1 l 2 ` ` 52( ( wi ~l 1 !wj ~l 2 ! i50 j50 3 E G~~k 1 1 n1 !02 1 i 1 j 1 1! G~k 1 02 1 i !G~n1 02 1 j ! 1 t k1 021i22 ~1 2 t ! n1 021j21 @2a 1 ~1 1 a!t # 2 dt c* ` 2 2l 1 ( ` G~~k 1 1 n1 !02 1 i 1 j 1 1! ( wi ~l 1 !wj ~l 2 ! G~k1 02 1 i 1 1!G~n1 1 j ! i50 j50 3 E c 1 t k1 021i21 ~1 2 t ! n1 021j21 @2a 1 ~1 1 a!t # dt 1 l 1 1 l 2 + * (18) PERFORMANCE OF THE BIASED ESTIMATORS 1091 Differentiating ~18! with respect to c and performing some manipulations, we obtain ` ` 2G~~k 1 1 n1 !02 1 i 1 j 1 1! ]PMSE~ bZ 1 !0s 2 5 ( ( wi ~l 1 !wj ~l 2 ! ]c G~k 1 02 1 i !G~n1 02 1 j ! i50 j50 3 k 1k1 021i21 n1n1 021j c k1 021i22 ~n1 1 k 1 c! ~n11k1 !021i1j21 F HF 3 2a 1 ~1 1 a! k1 c n1 1 k 1 c 3 2 2a 1 ~1 1 a! G G J k1 c k1 c 1 1 l1 + n1 1 k 1 c k 1 02 1 i n1 1 k 1 c (19) From ~19!, when a 5 2a and r 5 1, a condition for PMSE @ bZ 1 # to be monotonically decreasing is 2a 1 ~1 1 a! k1 c # 0+ n1 1 k 1 c (20) Thus, PMSE @ bZ 1 # is monotonically decreasing on c [ @0, an1 0k 1 # if a 5 2a and r 5 1+ Because bZ 1 reduces to the SR estimator when a 5 2a, r 5 1, and c 5 0, and it reduces to the PSR estimator when a 5 2a, r 5 1, and c 5 an1 0 k 1 , we obtain the following theorem+ THEOREM 1+ The PSR estimator dominates the SR estimator in terms of PMSE even when relevant regressors are omitted in the specified model. Theorem 2 in Ohtani ~1993! requires the condition l 1 $ l 2 for the PMSE of the PSR estimator to be smaller than that of the SR estimator+ This condition is caused by the fact that his formula of PMSE is mistakenly derived+ However, our result shows that such a condition is not required when the formula is derived correctly+ Because further theoretical analysis of the PMSEs of the SR, PSR, MMSE, and AMMSE estimators is difficult, we compare them numerically in the next section+ 4. NUMERICAL ANALYSIS In this section, we compare the PMSE performances of the SR, PSR, MMSE, and AMMSE estimators by numerical evaluations+ As is shown in Ohtani ~1993!, the noncentrality parameters are expressed as l 1 5 Fb R 12 and l 2 5 Fb ~1 2 R 12 !, 21 ' where Fb 5 b ' Sb0s 2 and R 12 5 b ' X ' X 1 S11 X 1 Xb0b ' Sb+ Here Fb is the noncentrality parameter that appeared in the test for the null hypothesis that all the regression coefficients are zeros in the correctly specified model+ Also, R 12 is 1092 AKIO NAMBA Table 1. Relative PMSEs of the SR, PSR, MMSE, and AMMSE estimators for k 1 5 5 and n 5 20 R 12 Fb 0+1 0+3 0+5 +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +4706 +5616 +6306 +7299 +7995 +8519 +8935 +9692 1+0216 1+0605 1+0908 1+1344 1+1640 1+2274 1+2471 +4706 +5632 +6349 +7381 +8080 +8579 +8948 +9538 +9873 1+0079 1+0215 1+0377 1+0467 1+0621 1+0662 +4706 +5640 +6372 +7421 +8116 +8594 +8935 +9447 +9715 +9872 +9971 1+0086 1+0147 1+0249 1+0275 +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +3533 +4457 +5148 +6121 +6783 +7267 +7640 +8292 +8720 +9028 +9262 +9598 +9829 1+0390 1+0618 +3533 +4522 +5294 +6419 +7197 +7764 +8196 +8920 +9363 +9658 +9866 1+0133 1+0294 1+0582 1+0651 +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +7254 +7614 +7883 +8260 +8514 +8699 +8842 +9089 +9251 +9367 +9455 +9582 +9669 +9880 +9963 +7254 +7631 +7925 +8354 +8651 +8868 +9033 +9310 +9479 +9591 +9671 +9774 +9838 +9963 1+0003 0+7 0+9 1+0 +4706 +5641 +6374 +7415 +8093 +8552 +8873 +9347 +9592 +9734 +9825 +9930 +9987 1+0081 1+0104 +4706 +5635 +6352 +7355 +7997 +8427 +8727 +9176 +9416 +9562 +9659 +9777 +9844 +9965 +9997 +4706 +5629 +6332 +7300 +7909 +8314 +8596 +9018 +9248 +9391 +9489 +9613 +9689 +9843 +9895 +3533 +4628 +5506 +6792 +7653 +8244 +8659 +9262 +9558 +9721 +9820 +9929 +9987 1+0081 1+0104 +3533 +4666 +5566 +6858 +7693 +8245 +8621 +9149 +9410 +9561 +9659 +9777 +9844 +9965 +9997 +3533 +4682 +5584 +6856 +7658 +8175 +8521 +9003 +9245 +9391 +9489 +9613 +9689 +9843 +9895 +7254 +7654 +7977 +8460 +8792 +9029 +9203 +9475 +9625 +9716 +9777 +9849 +9891 +9964 +9985 +7254 +7660 +7985 +8466 +8794 +9025 +9192 +9454 +9599 +9690 +9750 +9825 +9869 +9949 +9972 +7254 +7661 +7984 +8456 +8772 +8993 +9152 +9400 +9539 +9626 +9686 +9762 +9809 +9904 +9935 SR PSR +3533 +4579 +5414 +6643 +7486 +8084 +8522 +9204 +9571 +9785 +9919 1+0066 1+0139 1+0249 1+0275 MMSE +7254 +7644 +7957 +8421 +8743 +8976 +9149 +9427 +9586 +9685 +9752 +9834 +9882 +9969 +9995 continued PERFORMANCE OF THE BIASED ESTIMATORS 1093 Table 1. continued R 12 Fb 0+1 0+3 +3357 +4328 +5050 +6057 +6731 +7216 +7585 +8215 +8616 +8897 +9107 +9399 +9595 1+0049 1+0224 +3357 +4316 +5064 +6154 +6910 +7463 +7885 +8600 +9044 +9346 +9563 +9853 1+0037 1+0424 1+0556 0+5 0+7 0+9 1+0 +3357 +4267 +5023 +6193 +7037 +7661 +8133 +8902 +9342 +9614 +9793 1+0003 1+0117 1+0295 1+0333 +3357 +4229 +4963 +6110 +6946 +7566 +8036 +8797 +9228 +9490 +9659 +9855 +9957 1+0103 1+0127 +3357 +4207 +4921 +6033 +6841 +7439 +7891 +8620 +9032 +9284 +9448 +9641 +9745 +9913 +9953 AMMSE +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +3357 +4295 +5056 +6202 +7016 +7615 +8070 +8824 +9272 +9562 +9761 1+0011 1+0157 1+0426 1+0503 interpreted as the coefficient of determination in regression of Xb on X 1 + Thus, if R 12 is close to unity, the magnitude of model misspecification is regarded as small and vice versa+ The parameter values used in the numerical evaluations were k 1 5 3, 4, 5, 8, n 5 20, 30, 40, R 12 5 0+1, 0+3, 0+5, 0+7, 0+9, 1+0, and Fb 5 various values+ To compare the PMSEs of the estimators, we evaluated the values of relative PMSE defined as PMSE @ bN 1 #0PMSE @b1 # , where bN 1 is any estimator of b1 + Thus, the estimator bN 1 has smaller PMSE than the OLS estimator when the value of relative PMSE is smaller than unity+ The numerical evaluations were executed on a personal computer using the FORTRAN code+ In evaluating the integral in Gij ~ p, q; a, c! given in ~16!, we used Simpson’s rule with 200 equal subdivisions+ The double infinite series in H~ p, q; a, c! and J ~ p, q; a, c! were judged to converge when the increment of series became smaller than 10212+ Because the results for the cases of k 1 5 5, 8, and n 5 20 are qualitatively typical, we do not show the results for the other cases+ Table 1 shows the relative PMSEs of the SR, PSR, MMSE, and AMMSE estimators for k 1 5 5 and n 5 20+ From Table 1, we can make sure of Theorem 1; that is, the PMSE of the PSR estimator is uniformly smaller than or equal to that of the SR estimator even if relevant regressors are omitted+ Also, we can see that the AMMSE estimator has the smallest PMSE over a wide region of the parameter space ~e+g+, Fb # 25+0!+ In particular, the PMSE of the AMMSE estimator is smaller than the other estimators when the model misspecification is severe ~e+g+, R 12 5 0+1 and Fb # 50+0!+ Though the maximum of the relative PMSE of the AMMSE estimator is slightly larger than unity, the minimum is much smaller than unity+ This indicates that the gain in PMSE 1094 AKIO NAMBA Table 2. Relative PMSEs of the SR, PSR, MMSE, and AMMSE estimators for k 1 5 8 and n 5 20 R 12 Fb 0+1 0+3 0+5 +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +3571 +4321 +4980 +6105 +7053 +7877 +8611 1+0170 1+1456 1+2553 1+3506 1+5089 1+6355 2+0101 2+1888 +3571 +4320 +4975 +6067 +6944 +7663 +8265 +9408 1+0212 1+0804 1+1255 1+1891 1+2314 1+3252 1+3589 +3571 +4318 +4964 +6020 +6839 +7485 +8005 +8931 +9529 +9938 1+0233 1+0624 1+0868 1+1364 1+1527 +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +2606 +3218 +3739 +4585 +5251 +5792 +6244 +7106 +7722 +8184 +8542 +9057 +9407 1+0200 1+0489 +2606 +3279 +3864 +4834 +5609 +6245 +6778 +7801 +8539 +9099 +9542 1+0200 1+0670 1+1872 1+2400 +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +8030 +8145 +8242 +8400 +8525 +8627 +8714 +8885 +9014 +9117 +9202 +9334 +9435 +9713 +9844 +8030 +8165 +8283 +8479 +8636 +8766 +8875 +9083 +9232 +9345 +9432 +9559 +9647 +9855 +9935 0+7 0+9 1+0 +3571 +4313 +4946 +5960 +6724 +7312 +7774 +8571 +9066 +9396 +9630 +9932 1+0116 1+0477 1+0590 +3571 +4306 +4920 +5880 +6584 +7116 +7526 +8225 +8657 +8946 +9152 +9422 +9589 +9926 1+0033 +3571 +4301 +4904 +5831 +6500 +6998 +7379 +8023 +8419 +8684 +8874 +9127 +9288 +9629 +9750 +2606 +3389 +4072 +5198 +6076 +6769 +7322 +8293 +8898 +9295 +9568 +9909 1+0108 1+0477 1+0590 +2606 +3437 +4153 +5304 +6165 +6818 +7318 +8144 +8626 +8935 +9147 +9421 +9589 +9926 1+0033 +2606 +3460 +4188 +5335 +6171 +6786 +7247 +7984 +8408 +8682 +8874 +9127 +9288 +9629 +9750 +8030 +8203 +8353 +8600 +8793 +8945 +9069 +9290 +9435 +9535 +9608 +9706 +9768 +9898 +9942 +8030 +8220 +8382 +8643 +8841 +8994 +9116 +9329 +9465 +9559 +9626 +9716 +9774 +9893 +9934 +8030 +8228 +8395 +8660 +8857 +9008 +9126 +9331 +9460 +9549 +9612 +9698 +9753 +9871 +9912 SR PSR +2606 +3336 +3976 +5040 +5888 +6576 +7144 +8198 +8917 +9432 +9815 1+0337 1+0669 1+1326 1+1518 MMSE +8030 +8184 +8320 +8546 +8725 +8870 +8990 +9211 +9361 +9468 +9549 +9659 +9731 +9889 +9944 continued PERFORMANCE OF THE BIASED ESTIMATORS 1095 Table 2. continued R 12 Fb 0+1 0+3 +3201 +3783 +4271 +5047 +5642 +6116 +6504 +7229 +7737 +8115 +8409 +8837 +9135 +9855 1+0143 +3201 +3783 +4284 +5107 +5756 +6282 +6718 +7542 +8123 +8556 +8892 +9380 +9717 1+0526 1+0847 0+5 0+7 0+9 1+0 +3201 +3772 +4279 +5139 +5835 +6406 +6882 +7772 +8384 +8825 +9154 +9609 +9904 1+0534 1+0747 +3201 +3762 +4260 +5103 +5782 +6338 +6797 +7648 +8222 +8628 +8926 +9327 +9578 1+0077 1+0226 +3201 +3755 +4245 +5068 +5727 +6259 +6696 +7493 +8021 +8388 +8653 +9005 +9223 +9650 +9781 AMMSE +0 1+0 2+0 4+0 6+0 8+0 10+0 15+0 20+0 25+0 30+0 40+0 50+0 100+0 150+0 +3201 +3779 +4287 +5139 +5822 +6382 +6847 +7725 +8339 +8789 +9133 +9622 +9951 1+0705 1+0988 from using the AMMSE estimator instead of the OLS estimator is much larger than the loss even when there are omitted regressors+ Also, the loss from using the MMSE estimator instead of the OLS estimator is very small, though the gain is not so large+ The relative PMSEs of the SR, PSR, MMSE, and AMMSE estimators for k 1 5 8 and n 5 20 are shown in Table 2+ Similar to the case of k 1 5 5, we can make sure of Theorem 1+ When k 1 5 8, the PSR estimator has the smallest PMSE over a wide region of the parameter space+ Also, the AMMSE estimator has much smaller PMSE than the OLS estimator over a wide region of the parameter space+ Though the maximums of the relative PMSEs of the PSR and the AMMSE estimators are slightly larger than unity, their minimums are much smaller than unity+ Also, the MMSE estimator has smaller PMSE than the OLS estimator for all the values of Fb and R 12 considered here, though the minimum of its PMSE is not so small+ The results stated previously are almost similar to the numerical results in Ohtani ~1998!+ In other words, the effect of his calculative error on the numerical results is not so large+ REFERENCES Abadir, K+M+ ~1999! An introduction to hypergeometric functions for economists+ Econometric Reviews 18, 287–330+ Abramowitz, M+ & I+A+ Stegun ~1972! Handbook of Mathematical Functions+ New York: Dover Publications+ Baranchik, A+J+ ~1970! A family of minimax estimators of the mean of a multivariate normal distribution+ Annals of Mathematical Statistics 41, 642– 645+ 1096 AKIO NAMBA Farebrother, R+W+ ~1975! The minimum mean square error linear estimator and ridge regression+ Technometrics 17, 127–128+ Gourieroux, C+ & A+ Monfort ~1995! Statistics and Econometric Models, vol+ 1+ New York: Cambridge University Press+ James, W+ & C+ Stein ~1961! Estimation with quadratic loss+ In Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, vol+ 1, pp+ 361–379+ Berkeley: University of California Press+ Luke, Y+L+ ~1969! The Special Functions and Their Approximations, vol+ 1, New York: Academic Press+ Mittelhammer, R+C+ ~1984! Restricted least squares, pre-test, OLS, and Stein rule estimators: Risk comparison under model misspecification+ Journal of Econometrics 25, 151–164+ Ohtani, K+ ~1993! A comparison of the Stein-rule and positive-part Stein-rule estimators in a misspecified linear regression model+ Econometric Theory 9, 668– 679+ Ohtani, K+ ~1996! On an adjustment of degrees of freedom in the minimum mean squared error estimator+ Communications in Statistics—Theory and Methods 25, 3049–3058+ Ohtani, K+ ~1998! MSE performance of the minimum mean squared error estimators in a linear regression model when relevant regressors are omitted+ Journal of Statistical Computation and Simulation 61, 61–75+ Stein, C+ ~1956! Inadmissibility of the usual estimator for the mean of a multivariate normal distribution+ In Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, pp+ 197–206+ Berkeley: University of California Press+ Theil, H+ ~1971! Principles of Econometrics+ New York: Wiley+ APPENDIX In this Appendix, we derive the formulae for H~ p, q; a, c! and J ~ p, q; a, c!+ First, we derive the formula for H ~ p, q; a, c!+ The OLS estimator b 1 is distributed as 21 21 21 ' 21 N~ b1 1 S11 S12 b2 , s 2 S11 ! 5 N~S11 X 1 Xb, s 2 S11 !, where X 5 @X 1 , X 2 # , b ' 5 ' ' ' ' 2 @ b1 , b2 # , and S12 5 X 1 X 2 + Thus, u 1 5 b1 S11 b1 0s is distributed as the noncentral chi-square distribution with k 1 degrees of freedom and noncentrality parameter l 1 5 21 ' b ' X ' X 1 S11 X 1 Xb0s 2 + Also, the residual vector e1 is distributed as N~M1 Xb, s 2 M1 !, 21 ' where M1 5 ~In 2 X 1 S11 X 1 !+ Thus, u 2 5 e1' e1 0s 2 is distributed as the noncentral chi-square distribution with n1 degrees of freedom and noncentrality parameter l 2 5 b ' X ' M1 Xb0s 2 + Furthermore, u 1 and u 2 are mutually independent+ Using u 1 and u 2 , H~ p, q; a, c! is expressed as ` H~ p, q; a, c! 5 ` ( ( Kij i50 j50 EE u 1k1 022p1q1i21 u 2n1 021j21 ~u 1 1 au 2 ! p R 3 exp@2~u 1 1 u 2 !02# du 2 du 1 , (A.1) where Kij 5 wi ~l 1 !wj ~l 2 ! 2 ~k11n1 !021i1j G~k 1 02 1 i !G~n1 02 1 j ! , (A.2) wi ~l! 5 exp~2l02!~l02! i0i!, and R is the region such that $u 1 , u 2 6u 1 0u 2 $ k 1 c0n1 %+ PERFORMANCE OF THE BIASED ESTIMATORS 1097 Making use of the change of variables, v1 5 u 1 0u 2 and v2 5 u 2 , ~A+1! reduces to ` E E ` ` ( ( Kij i50 j50 k 1 c0n1 ` ~k 11n1 !021q1i1j21 v1k1 022p1q1i21 v2 ~a 1 v1 ! p 0 3 exp@2v2 ~1 1 v1 !02# dv2 dv1 + (A.3) Again, making use of the change of variable, z 5 ~1 1 v1 !v2 02, ~A+3! reduces to ` ` ( ( Kij 2 ~k 1n !021q1i1j G~~k1 1 n1 !02 1 q 1 i 1 j ! i50 j50 1 3 E 1 v1k1 022p1q1i21 ~a 1 v1 ! p ` ~1 1 v1 ! ~k11n1 !021q1i1j k 1 c0n1 dv1 + (A.4) Finally, making use of the change of variable, t 5 v1 0~1 1 v1 !, and performing some manipulations, we obtain ~14! in the text+ Next, we derive the formula for J ~ p, q; a, c!+ Noting that ]l 1 0]b1 5 2~S11 b1 1 S12 b2 !0s 2 and ]l 2 0]b1 5 0, and differentiating H~ p, q; a, c! given in ~14! with respect to b1 , we obtain ]H~ p, q; a, c! ]b1 ` 5 ` (( i50 j50 ` 5 ` (( i50 j50 F F 52 1 F ]wi ~l 1 ! F S11 b1 1 S12 b2 ]b1 G wj ~l 2 !Gij ~ p, q; a, c! s2 S11 b1 1 S12 b2 s2 @2wi ~l 1 ! 1 wi21 ~l 1 !#wj ~l 2 !Gij ~ p, q; a, c! G G( ( H~ p, q; a, c! S11 b1 1 S12 b2 s2 G ` ` wi ~l 1 !wj ~l 2 !Gi11, j ~ p, q; a, c!+ (A.5) i50 j50 21 ' 21 X 1 Xb, s 2 S11 !, H~ p, q; a+c! can be exBecause u 1 5 b1' X 11 b1 0s 2 and b1 ; N~S11 pressed as H~ p, q; a, c! 5 EE S R u 1 1 au 2 u1 D p q u 1 fN ~b1 ! f2 ~u 2 ! du 2 db1 , (A.6) where f2 ~u 2 ! is the density function of u 2 and fN ~b1 ! 5 1 ~2p! k 1 02 21 102 6s 2 S11 6 F exp 2 21 ' 21 ' ~b1 2 S11 X 1 Xb! ' S11 ~b1 2 S11 X 1 Xb! 2s 2 G + (A.7) 1098 AKIO NAMBA Differentiating ~A+6! with respect to b1 , we obtain ]H~ p, q; a, c! ]b1 5 EE S R 5 1 s 2 2 u 1 1 au 2 FS E I F u1 D n1 u 1 k1 u2 p q u 1 fN ~b1 ! f2 ~u 2 ! S11 b1 1 S12 b2 s2 DS $c G u 1 1 au 2 u1 F D H~ p, q; a, c!+ S11 b1 2 S11 b1 2 S12 b2 p s2 q u 1 S11 b1 G G du 2 db1 (A.8) 21 Equating ~A+5! and ~A+8!, and multiplying b ' X ' X 1 S11 from the left, we obtain ~15! in the text+
© Copyright 2024