Prncples of Econometrcs (3e) 013 년 1 학기 윤성민
8.1. 이분산의본질 ( 예 ) 식료품지출 / 식료품지출과소득에관한 40 개표본
8.1 이분산의본질 <Plot of sample data> 3
8.1 이분산의본질 <Plot of sample data> 4
8.1 이분산의본질 동분산가정 5
8.1 이분산의본질 이분산가정 6
8.1 이분산의본질 저소득가계의식료품지출액과고소득가계의식료품지출액중어느것을추정하는것이더쉬운가? 저소득가계는식료품에대한낭비적인선택을못함 - 상대적으로선택의폭이좁음, 소득의특정부분을식료품지출에사용, 소득을알면식료품지출액을추정하기쉬움 고소득가계는선택의폭이넓음 ( 가격보다취향을중시 ), 소득은설명변수로서덜중요하게되고, 소득을알더라도식료품지출액을추정하기어려움 7
8.1 이분산의본질 OLS 추정 가정 : 추정결과 : Ee ( ) = 0 y =β +β x + e 1 var( ) yˆ = 83.4 + 10.1x e e = e = σ cov(, j) 0 식료품지출액의퍼진정도가소득수준에따라다르다면, 동분산가정 var( y ) = var( e) = σ 다음과같은이분산가정이타당할것임 var( y ) = var( e) = σ 은부적절함 횡단면자료의경우이분산이존재하는경우가흔히있음 시계열자료에서도가끔나타남 ( 예 : 외환위기이후변동성증가 ) 8
8. 이분산이 OLS 추정량에미치는영향 OLS 추정량은선형불편추정량이기는하지만최소분산을가지지는않음 (BLUE 아님 ) 컴퓨터 SW가계산해주는추정치의표준오차는잘못된것 이것에근거한가설검정및신뢰구간도올바르지않게됨 9
8. 이분산이 OLS 추정량에미치는영향 OLS 추정량의표준오차계산식 < 동분산가정 > y =β 1+β x + e var( e) =σ var( b ) = N = 1 σ ( x x) < 이분산가정 > y =β 1+β x + e var( e) =σ var( b ) N ( ) N x x σ = 1 = wσ = N = 1 ( x x) = 1 10
8. 이분산이 OLS 추정량에미치는영향 표준오차를정확하게계산하기위한 Whte 표준오차 H. Whte 는 OLS 잔차를이용하여, 모수추정량에대한정확한표준오차를계산하는방법을제시 var( b ) N ( ) N x x σ = 1 = wσ = N = 1 ( x x) = 1 var( b ) N ( ) ˆ N x x e = 1 ˆ = we = N = 1 ( x x) = 1 11
8. 이분산이 OLS 추정량에미치는영향 Whte 표준오차계산결과 yˆ = 83.4 + 10.1x (7.46) (1.81) (Whte se) (43.41) (.09) (ncorrect se) (OLS se) β 의 95% 신뢰구간 Whte: b ± t se( b ) = 10.1±.04 1.81 = [6.55, 13.87] c Incorrect: b ± tcse( b) = 10.1±.04.09 = [5.97, 14.45] (OLS) - Whte 표준오차를사용하면신뢰구간의폭이더좁게계산됨 ( 더유용한구간추정결과를알려줌 ) 1
8. 이분산이 OLS 추정량에미치는영향 OLS ( 동분산가정 ) Dependent Varable: FOOD_EXP Method: Least Squares Date: 05/11/11 Tme: 14:47 Sample: 1 40 Included observatons: 40 Varable Coeffcent Std. Error t-statstc Prob. C 83.41600 43.41016 1.91578 0.06 INCOME 10.0964.09364 4.877381 0.0000 R-squared 0.38500 Mean dependent var 83.5735 Adjusted R-squared 0.368818 S.D. dependent var 11.675 S.E. of regresson 89.51700 Akake nfo crteron 11.87544 Sum squared resd 304505. Schwarz crteron 11.95988 Log lkelhood -35.5088 Hannan-Qunn crter. 11.90597 F-statstc 3.78884 Durbn-Watson stat 1.893880 Prob(F-statstc) 0.000019 13 13
OLS ( 이분산가정, Whte 표준오차 ) Dependent Varable: FOOD_EXP Method: Least Squares Date: 05/11/11 Tme: 14:48 Sample: 1 40 Included observatons: 40 Whte Heteroskedastcty-Consstent Standard Errors & Covarance 8. 이분산이 OLS 추정량에미치는영향 Varable Coeffcent Std. Error t-statstc Prob. C 83.41600 7.46375 3.037313 0.0043 INCOME 10.0964 1.809077 5.643565 0.0000 R-squared 0.38500 Mean dependent var 83.5735 Adjusted R-squared 0.368818 S.D. dependent var 11.675 S.E. of regresson 89.51700 Akake nfo crteron 11.87544 Sum squared resd 304505. Schwarz crteron 11.95988 Log lkelhood -35.5088 Hannan-Qunn crter. 11.90597 F-statstc 3.78884 Durbn-Watson stat 1.893880 Prob(F-statstc) 0.000019 14 14
8.3 일반최소제곱추정량 y =β 1 +β x + e, var( e ) =σ 이분산이존재하는경우에적용할수있는최선의추정방법 GLS (Generalzed Least Squares) (BLUE 임 ) 가정을 var( y ) = var( e) = σ 으로수정하는것만으로는 회귀식의모수를추정할수없음 - N개의표본만으로 N개의상이한분산과 모수들 β 1, β ) 을추정하는것은불가능 ( 자유도부족 ) ( 이문제를극복하기위해서는 σ 에대한추가적인가정이필요 (,,, ) σ σ σ 1 N 15
8.3 GLS 16
8.3.1 비례적이분산 식료품지출액의경우오차분산은소득수준에비례 비례적이분산을가정하는것이적절함, 다음은한사례 ( e ) =σ =σ var x 17
이분산이존재하는경우 OLS 추정량은 BLUE가아님 이문제를해결하는한방법은회귀모형을변형시켜동분산오차를갖도록한다음, OLS를적용하는것 y =β 1+β x + e 회귀모형의양변을 x 로나누어보자 y 1 x e =β 1 +β + x x x x 8.3.1 비례적이분산 다음과같이변형된변수를정의하면, y y * = x * 1 x = 1 x x * = x x e = e * x 변형된모형은다음과같음 ( 상수항이없는식이라는점에유의 ) y =β x +β x + e 1 1 18
8.3.1 비례적이분산 19
변형된모형의장점은오차항 ( 증명 ) 오차항의평균도 0, 독립성가정도충족됨 Ee ( ) = 0, cov( e, e ) = 0 가동분산을가진다는것 변형된변수를이용하여 OLS 로추정하면, 모수 y =β x +β x + e 1 1 * β 1, β ) 에대한 BLUE를구할수있음 이렇게구한추정량을 generalzed least squares (GLS) estmator 혹은 weghted least squares (WLS) estmator 라고함 e e var( e ) = var e 1 1 = var( e) = σ x =σ ( x x x ( y, x, x ) 1 j = e x 8.3.1 비례적이분산 ( e ) =σ =σ var x 0
8.3.1 비례적이분산 SAS program data food ; nfle 'C:\tmp\table3-1.prn' ; nput y x ; * create dataset; * read n data=food; * nput varables; w = 1/x ; * create weght varable; proc reg ; food_gls : model y=x ; weght w ; run ; * estmate regresson; * use orgnal data; * specfy weght for weghted LS = GLS; 1
8.3.1 비례적이분산 < 식료품지출액사례 > OLS 추정결과 yˆ = 83.4 + 10.1x (7.46) (1.81) (Whte se) (43.41) (.09) (ncorrect se) (OLS se) WLS 추정결과 yˆ = 78.68 + 10.45x (se) (3.79) (1.39) WLS 추정치의 95% 신뢰구간 β ˆ ± t ˆ c se( β ) = 10.451 ±.04 1.386 = [7.65,13.6]
8. 이분산이 OLS 추정량에미치는영향 WLS Dependent Varable: FOOD_EXP Method: Least Squares Date: 05/1/11 Tme: 14:57 Sample: 1 40 Included observatons: 40 Weghtng seres: 1/INCOME^(0.5) yˆ = 78.68 + 10.45x (se) (3.79) (1.39) Varable Coeffcent Std. Error t-statstc Prob. C 78.68408 3.7887 3.30761 0.001 INCOME 10.45101 1.385891 7.54100 0.0000 Quck-Estmate Equaton-Optons- WLS 클릭 -Weght 부분에 1/ncome^(0.5) 입력 Weghted Statstcs R-squared 0.599438 Mean dependent var 63.3689 Adjusted R-squared 0.588897 S.D. dependent var 76.43899 S.E. of regresson 76.30741 Akake nfo crteron 11.5561 Sum squared resd 167. Schwarz crteron 11.64057 Log lkelhood -9.15 Hannan-Qunn crter. 11.58666 F-statstc 56.8667 Durbn-Watson stat 1.905701 Prob(F-statstc) 0.000000 Unweghted Statstcs yˆ = 83.4 + 10.1x (43.41) (.09) R-squared 0.384787 Mean dependent var 83.5735 Adjusted R-squared 0.368597 S.D. dependent var 11.675 S.E. of regresson 89.5366 Sum squared resd 304611.7 Durbn-Watson stat 1.89377 3 3
8.3. 분산함수의추정 식료품지출액모형에서이분산을나타낼수있는방법들 1/ ( e ) =σ =σ x ( e ) =σ =σ x ( e ) =σ =σ x var var var 위의가정들을일반화하면아래와같이나타낼수있음 ( e ) =σ =σ var x γ ( γ=1 로가정할필요없음 ) 이식을약간변형하면, 위가정을아래와같이쓸수있음 ln( σ ) = ln( σ ) +γln( x) ( x ) σ = exp ln( σ ) +γln( ) = exp( α +α z ) 1 α = 1 ln( σ ) α =γ z = ln( x ) 4
8.3. 분산함수의추정 일반적인분산함수 σ = exp( α +α z ) 보다더일반화된분산함수 1 σ = exp( α +α z + +α z ) 1 s S 여기서 z k 는분산과관련된설명변수들 다음과같이표현해도됨 ln( σ ) =α 1 +α z + +αszs 5
8.3. 분산함수의추정 분산함수의추정방법 아래와같은간단한경우로설명해보자. y = E( y ) + e =β +β x + e ln( σ ) =α +α 1 (1) OLS 추정하여잔차를구함 z () 를선택하여아래분산함수를추정 식료품지출액사례 : (3) 분산의추정 ˆ 1 eˆ eˆ ˆ = y y = y b1 bx ln( e ) = ln( σ ) + v =α 1 +α z + v σ = α ˆ +αˆ ˆ ln( σ ) = 0.9378 +.39ln( x) ˆ exp( 1 z) z γ=.39 1 6
8. 이분산이 OLS 추정량에미치는영향 분산함수의추정 Dependent Varable: LOG(EHAT) Method: Least Squares Date: 05/13/11 Tme: 13:54 Sample: 1 40 Included observatons: 40 LOG(EHAT)=C(1)+C()*LOG(INCOME) ˆ ln( σ ) = 0.9378 +.39ln( x) Coeffcent Std. Error t-statstc Prob. C(1) 0.937796 1.583106 0.59377 0.5571 C().3939 0.541336 4.30761 0.0001 R-squared 0.37597 Mean dependent var 7.648159 Adjusted R-squared 0.309903 S.D. dependent var.071519 S.E. of regresson 1.70855 Akake nfo crteron 3.976 Sum squared resd 11.5310 Schwarz crteron 4.056670 Log lkelhood -77.4445 Hannan-Qunn crter. 4.00758 F-statstc 18.51375 Durbn-Watson stat.175575 Prob(F-statstc) 0.000114 7 7
8.3. 분산함수의추정 모수의 GLS 추정방법 y =β +β x + e 1 앞에서구한분산 σ 를이용하여 GLS 추정량구하면됨 (1) 회귀모형의양변을로나눔 변형된오차는동분산임을확인할수있음 ˆ σ () 을이용하여변형된변수를계산함 y 1 x y =, x1 =, x = σˆ ˆ ˆ σ σ (3) 아래식을 OLS로추정함 (BLUE) σ y 1 x e =β 1 +β + σ σ σ σ e 1 1 var = var( ) 1 e = σ = σ σ σ y =β x +β x + e 1 1 8
8.3. 분산함수의추정 GLS 추정절차요약 y =β +β x +β x + +β x + e 1 3 3 K K (1) OLS 로추정하여잔차을계산함 () eˆ =α +α z + +α z + v 에 OLS 적용하여 e ˆ ln 1 S S α1, α,, αs 추정 (3) 분산추정값계산 σ = α ˆ +α ˆ + +αˆ ˆ exp( 1 z S zs ) (4) 회귀모형의양변을로나누어, 변형된자료계산 y 1 x k y =, x1 =, xk = σˆ ˆ ˆ σ σ (5) 변형된아래모형을 OLS로추정 ˆ σ y =β x +β x +β x + +β x + e 1 1 3 3 K K 9
8.3. 분산함수의추정 < 식료품지출액사례 > OLS 추정결과 γ=0 yˆ = 83.4 + 10.1x ( e ) =σ =σ var x γ (7.46) (1.81) (Whte se) (43.41) (.09) (OLS se) WLS 추정결과 γ=1 yˆ = 78.68 + 10.45x (se) (3.79) (1.39) GLS 추정결과 γ=.39 yˆ = 76.05 + 10.63x (se) (9.71) (0.97) 30
GLS Dependent Varable: FOOD_EXP Method: Least Squares Date: 05/1/11 Tme: 17:14 Sample: 1 40 Included observatons: 40 Weghtng seres: 1/(INCOME^.39)^(0.5) 8. 이분산이 OLS 추정량에미치는영향 yˆ = 76.05 + 10.63x (se) (9.71) (0.97) Varable Coeffcent Std. Error t-statstc Prob. C 76.05387 9.71491 7.88563 0.0000 INCOME 10.63348 0.971543 10.94494 0.0000 Weghted Statstcs R-squared 0.759176 Mean dependent var 5.3697 Adjusted R-squared 0.75839 S.D. dependent var 105.968 S.E. of regresson 55.59375 Akake nfo crteron 10.973 Sum squared resd 117445.3 Schwarz crteron 11.00717 Log lkelhood -16.4545 Hannan-Qunn crter. 10.9536 F-statstc 119.7918 Durbn-Watson stat 1.905316 Prob(F-statstc) 0.000000 Unweghted Statstcs R-squared 0.38466 Mean dependent var 83.5735 Adjusted R-squared 0.368063 S.D. dependent var 11.675 S.E. of regresson 89.57055 Sum squared resd 304869.6 Durbn-Watson stat 1.890159 31 31
8.3.3 이분산적분할 (Heteroskedastc Partton) ( 예 ) 임금함수임금 = f ( 교육수준, 경험, 인종, 성별, 거주지역,.) 단순화를위해임금 (WAGE) 이다음세가지에만의존한다고하자 - 교육받은연수 (EDUC) - 경험의연수 ( 숙련도, 생산성의대리변수 )(EXPER) - 거주지역 (METRO, 더미변수 : 대도시거주면 1, 아니면 0) OLS 추정결과 WAGE = 9.914 + 1.34 EDUC + 0.133 EXPER + 1.54 METRO (se) (1.08) (0.070) (0.015) (0.431) - 대도시지역평균임금이시골지역보다시간당 $1.54 더높음 3
임금함수 (OLS) Dependent Varable: WAGE Method: Least Squares Date: 05/1/11 Tme: 17:3 Sample: 1 1000 Included observatons: 1000 WAGE=C(1)+C()*EDUC+C(3)*EXPER+C(4)*METRO 8. 이분산이 OLS 추정량에미치는영향 WAGE = 9.914 + 1.34 EDUC + 0.133 EXPER + 1.54 METRO (se) (1.08) (0.070) (0.015) (0.431) Coeffcent Std. Error t-statstc Prob. C(1) -9.913984 1.075663-9.16631 0.0000 C() 1.33964 0.069961 17.6378 0.0000 C(3) 0.13344 0.0153 8.747835 0.0000 C(4) 1.54104 0.431091 3.535459 0.0004 R-squared 0.66903 Mean dependent var 10.130 Adjusted R-squared 0.64695 S.D. dependent var 6.46641 S.E. of regresson 5.356490 Akake nfo crteron 6.198487 Sum squared resd 8577.1 Schwarz crteron 6.18118 Log lkelhood -3095.43 Hannan-Qunn crter. 6.05948 F-statstc 10.8733 Durbn-Watson stat 0.50560 Prob(F-statstc) 0.000000 33 33
8.3.3 이분산적분할 대도시지역 (M) 과시골지역 (R) 의임금분산은다를수있음 - 대도시에는다양한형태의직업이존재, 임금분산이더클가능성 두지역에서교육수준과경험이임금에미치는영향은동일하지만임금분산은다르다고가정해보자 WAGEM =β M 1 +β EDUCM +β 3EXPERM + em = 1,,, NM (808) WAGER =β R1 +β EDUCR +β 3EXPERR + er = 1,,, NR (19) 두식의오차분산이동일하다면 OLS 로각각분리추정하였을때두식의 ( β, β ) 은동일하고 두식의상수항은다음의관계에있을것임 ( 실제로는다름. Why?) b M (var( M ) var( R ) ) 3 1 = br 1 + 1.54 = 9.914 + 1.54 = 8.39 e = e = σ 34
8.3.3 이분산적분할 대도시지역과시골지역의임금분산이다르다고가정하면 ( 즉, 이분산적분할이존재한다면 ), 다음과같이나타낼수있음 var( e ) =σ, var( e ) =σ M M R R 대도시 808개, 시골 19개표본을이용하여앞의두식을각각분리추정한결과는다음과같음 σ ˆ = 31.84, σ ˆ = 15.43 M R b = 9.05 b = 1.8 b = 0.1346 M1 M M3 b = 6.166 b = 0.956 b = 0.160 R1 R R3 교육과경험이임금에미치는영향이어느정도인지알수없음 전체표본을이용하여 GLS로추정하는것이바람직함 35
임금함수 ( 도시 808 명, metro=1) 8. 이분산이 OLS 추정량에미치는영향 Dependent Varable: WAGE Method: Least Squares Date: 05/13/11 Tme: 17:5 Sample: 193 1000 Included observatons: 808 eˆ 5618.1 var( e ) =σ = M M 31.84 N K = (808 3) = Varable Coeffcent Std. Error t-statstc Prob. C -9.05478 1.189456-7.610603 0.0000 EDUC 1.81714 0.079763 16.06910 0.0000 EXPER 0.134560 0.017948 7.497370 0.0000 R-squared 0.58183 Mean dependent var 10.5780 Adjusted R-squared 0.56340 S.D. dependent var 6.541667 S.E. of regresson 5.64153 Akake nfo crteron 6.301795 Sum squared resd 5618.10 Schwarz crteron 6.3196 Log lkelhood -54.95 Hannan-Qunn crter. 6.308488 F-statstc 140.0868 Durbn-Watson stat 0.47766 Prob(F-statstc) 0.000000 36 36
임금함수 ( 농촌 19 명, metro=0) 8. 이분산이 OLS 추정량에미치는영향 Dependent Varable: WAGE Method: Least Squares Date: 05/13/11 Tme: 17:4 Sample: 1 19 Included observatons: 19 eˆ 880.94 var( e ) =σ = R R 15.43 N K = (19 3) = Varable Coeffcent Std. Error t-statstc Prob. C -6.165855 1.898511-3.4773 0.0014 EDUC 0.955585 0.133190 7.174608 0.0000 EXPER 0.15974 0.04771 5.085538 0.0000 R-squared 0.58748 Mean dependent var 8.676979 Adjusted R-squared 0.50904 S.D. dependent var 4.510933 S.E. of regresson 3.9047 Akake nfo crteron 5.577498 Sum squared resd 880.94 Schwarz crteron 5.68397 Log lkelhood -53.4398 Hannan-Qunn crter. 5.59811 F-statstc 3.98706 Durbn-Watson stat 0.516503 Prob(F-statstc) 0.000000 37 37
8.3.3 이분산적분할 이분산적분할이존재하는모형의추정방법 (1) 두하위표본 ( 대도시 / 시골 ) 에대해각각추정하여분산계산 () 전체표본을각그룹의오차항표준오차로나누어변수변형 WAGE 1 EDUC EXPER METRO e =β R1 +β +β 3 +δ + σˆ ˆ ˆ ˆ ˆ ˆ σ σ σ σ σ σ ˆ M when METRO = 1 σ ˆ = σ ˆ R when METRO = 0 (3) 변형된모형에 OLS 적용 (BLUE) 38
8.3.3 이분산적분할 임금함수추정결과비교 < 동분산가정하여 OLS 적용한결과 > WAGE = 9.914 + 1.34 EDUC + 0.133 EXPER + 1.54 METRO (se) (1.08) (0.070) (0.015) (0.431) < 이분산가정하여 GLS 적용한결과 > WAGE = 9.398 + 1.196 EDUC + 0.13 EXPER + 1.539 METRO (se) (1.0) (0.069) (0.015) (0.346) 모수추정치는비슷함, 이는이분산이존재하는경우에도 OLS 추정량은불편추정량이기때문임 (GLS 추정치가더정확함 ) GLS 경우의표준오차가별로줄지않은이유는표본이작기때문임 39
임금함수 (OLS) Dependent Varable: WAGE Method: Least Squares Date: 05/1/11 Tme: 17:3 Sample: 1 1000 Included observatons: 1000 WAGE=C(1)+C()*EDUC+C(3)*EXPER+C(4)*METRO 8. 이분산이 OLS 추정량에미치는영향 WAGE = 9.914 + 1.34 EDUC + 0.133 EXPER + 1.54 METRO (se) (1.08) (0.070) (0.015) (0.431) Coeffcent Std. Error t-statstc Prob. C(1) -9.913984 1.075663-9.16631 0.0000 C() 1.33964 0.069961 17.6378 0.0000 C(3) 0.13344 0.0153 8.747835 0.0000 C(4) 1.54104 0.431091 3.535459 0.0004 R-squared 0.66903 Mean dependent var 10.130 Adjusted R-squared 0.64695 S.D. dependent var 6.46641 S.E. of regresson 5.356490 Akake nfo crteron 6.198487 Sum squared resd 8577.1 Schwarz crteron 6.18118 Log lkelhood -3095.43 Hannan-Qunn crter. 6.05948 F-statstc 10.8733 Durbn-Watson stat 0.50560 Prob(F-statstc) 0.000000 40 40
임금함수 (GLS) Dependent Varable: WA_S Method: Least Squares Date: 05/14/11 Tme: 16:43 Sample: 1 1000 Included observatons: 1000 WA_S=C(1)*CONS_S+C()*ED_S+C(3)*EX_S+C(4)*ME_S 8. 이분산이 OLS 추정량에미치는영향 WAGE = 9.398 + 1.196 EDUC + 0.13 EXPER + 1.539 METRO (se) (1.0) (0.069) (0.015) (0.346) Coeffcent Std. Error t-statstc Prob. C(1) -9.398355 1.01967-9.17037 0.0000 C() 1.19570 0.068508 17.45375 0.0000 C(3) 0.1309 0.014548 9.087453 0.0000 C(4) 1.538803 0.34685 4.443749 0.0000 R-squared 0.6551 Mean dependent var 1.941801 Adjusted R-squared 0.63038 S.D. dependent var 1.16683 S.E. of regresson 1.00113 Akake nfo crteron.84494 Sum squared resd 998.4178 Schwarz crteron.86395 Log lkelhood -1418.147 Hannan-Qunn crter..851755 Durbn-Watson stat 0.510038 41 41
8.4 이분산의탐지 GLS를적용해야할정도로이분산이문제되는지를탐지하는방법 8.4.1 잔차의도표화 단순회귀경우 plot을그려보는방법 ( 예 ) 식료품지출액 다중회귀경우 각설명변수와잔차의 plot을그려보는방법 그래프만으로판단하기곤란할수도있음 일반적인방법 골드펠드-콴트검정 (Goldfeld-Quandt test, GQ test) 4
8.4 이분산의탐지 8.4. Goldfeld-Quandt test 전체표본을상이한분산을갖고있을것으로예상되는 개의그룹 ( 하위표본 ) 으로나눔 - 비례적이분산경우 잔차의크기순으로두그룹으로구분 ( 다음슬라이드에서설명 ) - 이분산적분할의경우 쉽게구분가능 ( 대도시 / 시골 ) 다음의가설을검정 H : σ =σ aganst H : σ σ 0 1 1 1 43
8.4 이분산의탐지 44
임금함수 ( 도시 808 명, metro=1) 8. 이분산이 OLS 추정량에미치는영향 Dependent Varable: WAGE Method: Least Squares Date: 05/13/11 Tme: 17:5 Sample: 193 1000 Included observatons: 808 eˆ 5618.1 var( e ) =σ = M M 31.84 N K = (808 3) = Varable Coeffcent Std. Error t-statstc Prob. C -9.05478 1.189456-7.610603 0.0000 EDUC 1.81714 0.079763 16.06910 0.0000 EXPER 0.134560 0.017948 7.497370 0.0000 R-squared 0.58183 Mean dependent var 10.5780 Adjusted R-squared 0.56340 S.D. dependent var 6.541667 S.E. of regresson 5.64153 Akake nfo crteron 6.301795 Sum squared resd 5618.10 Schwarz crteron 6.3196 Log lkelhood -54.95 Hannan-Qunn crter. 6.308488 F-statstc 140.0868 Durbn-Watson stat 0.47766 Prob(F-statstc) 0.000000 45 45
임금함수 ( 농촌 19 명, metro=0) 8. 이분산이 OLS 추정량에미치는영향 Dependent Varable: WAGE Method: Least Squares Date: 05/13/11 Tme: 17:4 Sample: 1 19 Included observatons: 19 eˆ 880.94 var( e ) =σ = R R 15.43 N K = (19 3) = Varable Coeffcent Std. Error t-statstc Prob. C -6.165855 1.898511-3.4773 0.0014 EDUC 0.955585 0.133190 7.174608 0.0000 EXPER 0.15974 0.04771 5.085538 0.0000 R-squared 0.58748 Mean dependent var 8.676979 Adjusted R-squared 0.50904 S.D. dependent var 4.510933 S.E. of regresson 3.9047 Akake nfo crteron 5.577498 Sum squared resd 880.94 Schwarz crteron 5.68397 Log lkelhood -53.4398 Hannan-Qunn crter. 5.59811 F-statstc 3.98706 Durbn-Watson stat 0.516503 Prob(F-statstc) 0.000000 46 46
8.4 이분산의탐지 임금함수경우의 Goldfeld-Quandt Test 가설 : H : σ =σ aganst H : σ >σ 0 M R 1 M R 검정통계량 : σˆ F = F σ M R ˆ M M R R ( N K, N K ) 임계치 : 유의수준 5% 경우 F F F C = ( α; N K, N K ) = (0.95;808 3,19 3) = 1. M M R R 검정결과 : F ˆ M 31.84 σˆ R 15.43 σ = = =.09 > F = 1. C 귀무가설기각, 대도시의임금분산이시골보다더큼 ( 이분산 ) 47
8.4 이분산의탐지 식료품지출액경우의 Goldfeld-Quandt Test (1) 표본을비슷한크기의두그룹으로분리함 () 큰분산을가질것같은그룹 (1) 의분산추정값 σˆ1 작은분산을가질것같은그룹 () 의분산추정값 σˆ (3) 가설 : H : σ =σ aganst H : σ >σ 0 1 1 1 ˆ σ1 (4) 검정통계량 : GQ = ~ F ( 이분산이심할수록 GQ는큰값 ) ( N1 K), ( N K) ˆ σ (5) 자유도 (18,18) 인 F- 분포에서 5% 임계값은 F c =. (6) ˆ 191.9 1 σ = ˆ 3574.8 σ = ˆ 191.9 1 GQ σ σˆ 3574.8 = = = 3.61 귀무가설기각 이분산존재, 오차항분산은소득수준에의존함 48
8.4 이분산의탐지 8.4.3 분산함수의검정 앞에서분산함수를다음과같이나타내었음 var( e ) =σ = exp( α 1+α z + +αszs) 이보다더일반적인경우는다음과같음 var( y ) = var( e ) = Ee ( ) =σ = h( α 1 +α z + +αszs) 앞에서사용한분산함수들은이것의특정한형태로볼수있음 h( α +α z + +α z ) = exp( α +α z + +α z ) 1 S S 1 S S 1 ( ) h( α +α z ) = exp ln( σ ) +γln( x ) 49
8.4 이분산의탐지 일반적인분산함수 =σ = = α +α + +α var( y ) Ee ( ) h( 1 z S zs ) 이일반적인분산함수의다른특별한예는선형함수형태임 σ = Ee ( ) =α +α z + +α z 1 s S 이분산여부를알아보기위해서는다음의가설검정을수행하면됨 H H : α =α = =α = 0 0 3 : not all the α n H are zero 1 s 0 만약귀무가설이옳다면, 오차분산은다음과같은동분산형태가됨 σ = Ee ( ) =α 귀무가설이기각되면, 오차분산은이분산형태가될것임 1 S 50
8.4 이분산의탐지 분산함수형태에대한가설검정 ( 이분산검정 ) = Lagrange multpler test 혹은 Breusch-Pagan test (1) 다음의회귀식을 OLS로추정 e =α +α z + +α z + v ˆ 1 S S () 가설 H H : α =α = =α = 0 0 3 : not all the α n H are zero 1 s 0 S (3) 아래의검정통계량을이용해검정 χ = N R χ ( S 1) 51
8.4 이분산의탐지 이분산에관한 Whte test 앞의 Lagrange multpler test ( 혹은 Breusch-Pagan test) 를사용하기쉽도록단순화시킨검정방법 z를설명변수 x, x의제곱항, x들의교차곱항 ( 상호작용항 ) 으로정의함 E( y ) =β +β x +β x 1 3 3 z = x, z = x, z = x, z = x, ( z = x x ) 3 3 4 5 3 6 3 5
8.4 이분산의탐지 ( 예 ) 식료품지출액모형의이분산검정 (1) y =β +β x + e 1 < Lagrange multpler test > 분산함수 : σ = Ee ( ) =α +α 1 z 가설 : H : α = 0 H : α 0 0 1 OLS 이용 e =α +α x + v ˆ 1 를추정 SST = 4,610,749, 441 SSE = 3,759,556,169 R SSE = 1 = 0.1846 SST χ = N R = 40 0.1846 = 7.38 >χ (1,0.95) = 3.84 유의수준 5% 에서귀무가설기각, 따라서이분산존재함 53
8. 이분산이 OLS 추정량에미치는영향 식료품이분산 LM 검정 (1 차항이분산과관련되는경우 ) Dependent Varable: EHAT Method: Least Squares Date: 05/14/11 Tme: 16:53 Sample: 1 40 Included observatons: 40 EHAT=C(1)+C()*INCOME e =α +α x + v ˆ 1 χ = N R = 40 0.1846 = 7.38 Coeffcent Std. Error t-statstc Prob. C(1) -576.370 483.501-1.194645 0.396 C() 68.36 3.590.93317 0.0057 R-squared 0.184611 Mean dependent var 761.69 Adjusted R-squared 0.163153 S.D. dependent var 10873.10 S.E. of regresson 9946.64 Akake nfo crteron 1.9656 Sum squared resd 3.76E+09 Schwarz crteron 1.38101 Log lkelhood -43.9313 Hannan-Qunn crter. 1.3710 F-statstc 8.603501 Durbn-Watson stat.34453 Prob(F-statstc) 0.005659 54 54
8. 이분산이 OLS 추정량에미치는영향 식료품이분산 Breusch-Pagan test 검정 / Evews output (1 차항이분산과관련되는경우 ) χ = N R = 40 0.1846 = 7.38 Heteroskedastcty Test: Breusch-Pagan-Godfrey F-statstc 8.603501 Prob. F(1,38) 0.0057 Obs*R-squared 7.38444 Prob. Ch-Square(1) 0.0066 Scaled explaned SS 6.67901 Prob. Ch-Square(1) 0.0100 55 55
8.4 이분산의탐지 ( 예 ) 식료품지출액모형의이분산검정 () y =β +β x + e 1 < Whte test > 분산함수 : σ = Ee ( ) =α +α x+α x 1 3 가설 : H : α =α = 0, H : α 0 or α 0 0 3 1 3 OLS 이용 e =α +α x +α x + v ˆ 1 3 를추정 χ = N R = 40 0.18888 = 7.555 > χ (,0.95) = 5.99 p-value = 0.03 유의수준 5% 에서귀무가설기각, 따라서이분산존재함 56
8. 이분산이 OLS 추정량에미치는영향 식료품이분산 Whte 검정 ( 차항까지이분산과관련되는경우 ) Heteroskedastcty Test: Breusch-Pagan-Godfrey F-statstc 4.307884 Prob. F(,37) 0.008 Obs*R-squared 7.555079 Prob. Ch-Square() 0.09 Scaled explaned SS 6.78107 Prob. Ch-Square() 0.0337 Test Equaton: Dependent Varable: RESID^ Method: Least Squares Date: 05/14/11 Tme: 17:37 Sample: 1 40 Included observatons: 40 e =α +α x +α x + v ˆ 1 3 χ = N R = 40 0.18888 = 7.555 Varable Coeffcent Std. Error t-statstc Prob. C -908.783 8100.109-0.359104 0.716 INCOME 91.7457 915.846 0.318553 0.7519 INCOME^ 11.1659 5.30953 0.441150 0.6617 R-squared 0.188877 Mean dependent var 761.69 Adjusted R-squared 0.14503 S.D. dependent var 10873.10 S.E. of regresson 10053.75 Akake nfo crteron 1.3413 57 57
< 과제 > 8.8 8.16 Evews output을출력하고, 출력물의빈여백에간단하게답을적으시오. 참고 : 필요한 data 는 WILEY 교과서홈페이지에있음 http://prncplesofeconometrcs.com/ 58