- Curs 10
Ipoteze statistice:
I1 - Media erorilor egal cu zero
I2 Homoscedasticitatea erorilor
1
~
N
(
0
,
)
c. Ipoteza de normalitate:
i
d. Ipoteza de necorelare sau de independen a
erorilor: cov(i, j)=0.
a) M(i) = = constant
Considerm Y = 0 + 1xi + .
Fie Y = 0 + + 1xi + = 0 *+ 1xi + *, unde:
0 * = 0 + , iar * = .
*
Parametrul 0 este estimat deplasat de estimatorul 0 : M ( 0 ) 0
*
b) M(i) = i
Considerm Yi = 0 + i+ 1xi + i = 0 *+ 1xi + *,
Parametrul 1 este estimat deplasat de estimatorul 1 :
n xi i xi i
M ( 1 ) 1
2
2
n xi xi
Etape:
- Se estimeaz modelul de regresie yi = 0 + 1xi + i
- Se determin erorile estimate ei = yi b0 b1xi
- Se verific dac media erorilor este zero.
tcalc
M ( ei )
sM ( )
Minimum
58508,12
-98125,2
-,778
-1,290
Maximum
6084511
131202,7
2,300
1,725
Mean
1582428
,00000
,000
,000
St d. Dev iation
1957596,554
73271,63549
1,000
,964
N
15
15
15
15
One-Sample Statistics
N
Unstandardized Residual
15
Mean
,0000000
St d. Dev iation
73271,63549
St d. Error
Mean
18918,65
One-Sample Test
Test Value = 0
t
Unstandardized Residual
df
,000
14
Sig. (2-tailed)
1,000
Mean
Dif f erence
,00000000
*
yi
unde:
*
yi
0 i xi ui
yi M ( i )
10
Definiie
V( i ) 2
11
Erori homoscedastice
Erori heteroscedastice
12
13
14
15
1.
2.
3.
Y 0 1 X
i 0 1 xi ui
17
Model
1
(Constant)
PI B
Unstandardized
Coef f icients
B
St d. Error
50921,663 12000,771
,016
,012
St andardized
Coef f icients
Beta
,348
t
4,243
1,337
Sig.
,001
,204
1. Ipoteze statistice:
H0: ipoteza de homoscedasticitate ( = 0)
H1: ipoteza de heteroscedasticitate ( 0)
19
2. Statistica test:
Se folosete statistica test t Student:
tcalc
r n2
1 r 2
unde: r este estimaia coeficientului Spearman .
3. Regula de decizie:
Dac tcalc >tteor sau valoarea lui Sig. asociat statisticii test t
Student calculate < 0,05 se respinge ipoteza H0.
20
Spearman's rho
Unstandardized Residual
X
1,000
.
5
,000
1,000
5
Unstandardiz
ed Residual
,000
1,000
5
1,000
.
5
2. Statistica test :
t
n 2
1 2
0
2
10
1 r
3. Decizia:
yi 0 1 xi i
Corectarea heteroscedasticitii presupune ponderarea
modelului iniial cu variabila 1 .
i
23
0
xi i
1
i i
i i
yi
24
i2 2 xi2
Corectarea heteroscedasticitii presupune ponderarea
modelului iniial cu variabila 1/xi.
25
1
xi
xi
xi
Coefficientsa
Model
1
(Constant)
Gross domestic
product / capita
Unstandardized
Coeff icients
B
Std. Error
64.365
3.802
-.004
Standardized
Coeff icients
Beta
.000
-.640
t
16.927
Sig.
.000
-8.625
.000
a. Dependent Variable: Inf ant mortality (deaths per 1000 liv e births)
Coefficientsa,b
Model
1
(Constant)
Gross domestic
product / capita
Unstandardized
Coeff icients
B
Std. Error
36.414
2.989
-.002
Standardized
Coeff icients
Beta
.000
-.609
t
12.184
Sig.
.000
-7.933
.000
a. Dependent Variable: Inf ant mortality (deaths per 1000 liv e births)
b. Weighted Least Squares Regression - Weighted by inv
26
27
Testul Breusch-Pagan-Godfrey
Pentru modelul de regresie multipl Y=0+1X1+2X2+,
testarea homoscedasticitii presupune parcurgerea
urmtorilor pai:
1. Estimarea parametrilor modelului de regresie liniar multipl
(0 , 1 i 2)
2. Pe baza modelului estimat se obin valorile erorii de modelare
(ei)
3. Se construiete modelul auxiliar de regresie:
28
Testul Breusch-Pagan-Godfrey
4. Se estimeaz raportul de determinaie pentru modelul auxiliar
(R2).
5. Pe baza raportului de determinaie se caluleaz statistica test
2 calc = n*R2
6. Se afl valoare teoretic din tabelul repartiiei Hi ptrat:
2, k-1
unde k reprezinta numrul parametrilor din modelul auxiliar
7. Se compar valoarea teoretic cu valoarea calculat i se decide
acceptarea/ respingerea ipotezei de homoscedasticitate a
erorilor
Dac 2 calc < 2, k-1 AH0
Dac 2 calc 2, k-1 RH0
SAU
Dac Sig. > AH0
Dac Sig. < RH0
29
Modelul estimat:
SAL = -7808,1+1,67*SAL0 + 1020,39*ED+e
Coefficient
-1.31E+08
6150.668
6452228
R-sq.
0.103839
Adjusted R-sq. 0.100033
S.E. of reg.
1.82E+08
Sum sq. resid 1.56E+19
Log likelihood -9686.895
F-statistic
27.28751
Prob(F-statistic) 0.000000
Prob. F(2,471)
0.0000
Prob. Chi-Square(2)
0.0000
Prob. Chi-Square(2) 0.0000
t-Statistic
-3.203380
4.472026
1.719509
P
0.0015
0.0000
0.0862
60401068
1.92E+08
40.88563
40.91197
40.89599
1.796592
31
F-statistic
27.28751
Obs*R-squared
49.21954
Scaled explained SS 245.2163
Var.
0
SAL0
ED
Coefficient
-1.31E+08
6150.668
6452228
R-sq.
0.103839
Adjusted R-sq. 0.100033
Prob. F(2,471)
0.0000
Prob. Chi-Square(2)
0.0000
Prob. Chi-Square(2) 0.0000
Std. Error
40991028
1375.365
3752366
Mean dependent var
S.D. dependent var
t-Statistic
-3.203380
4.472026
1.719509
P
0.0015
0.0000
0.0862
60401068
1.92E+08
32
Testul White
Testul White urmeaz acelai algoritm ca i n cazul testului
Breusch-Pagan-Godfrey.
Singura diferen const n faptul c se utilizeaz o form mult
mai complex a modelului auxiliar:
ei2=0 +1X1+ 2X2+3X1X2+ 4X1 2+ 5X2 2+u
Calculul statisticii 2 i luarea deciziei se realizeaz ca i n
cazul testului Breusch-Pagan-Godfrey.
33
Modelul estimat:
SAL = -7808,1+1,67*SAL0 + 1020,39*ED+e
Test Equation:
Dependent Variable: ^2
Raportul de determinaie
Method: Least Squares
(R2 = 0,1134).
Date: 01/15/13 Time: 23:19
Sample: 1 474
Included observations: 474
Variable Coefficient
Std. Error
t-Statistic
0
-2.13E+08
1.67E+08
-1.275593
SAL0
13826.12
9771.187
1.414989
SAL0^2
-0.138886
0.082003
-1.693663
SAL0*ED 72.27410
765.1561
0.094457
ED
9214356.
26870194
0.342921
ED^2
-291287.7
1381016.
-0.210923
R-sq.
0.113415
Mean dependent va 60401068
Adj R-sq
0.103943
S.D. dependent var 1.92E+08
S.E. of reg
1.82E+08
Akaike info criterion 40.88755
Sum sq. res
1.55E+19
Schwarz criterion 40.94022
Log likelihood
-9684.348
Hannan-Quinn criter 40.90826
F-statistic
11.97360
Durbin-Watson stat 1.799331
Prob(F-statistic)
0.000000
Prob.
0.2027
0.1577
0.0910
0.9248
0.7318
0.8330
35
11.97360
Prob. F(5,468)
Prob. Chi-Square(5) 0.0000
267.8303
Prob. Chi-Square(5)
Variable Coefficient
0
-2.13E+08
SAL0
13826.12
SAL0^2
-0.138886
SAL0*ED 72.27410
ED
9214356.
ED^2
-291287.7
R-sq.
0.113415
Std. Error
1.67E+08
9771.187
0.082003
765.1561
26870194
1381016.
0.0000
0.0000
t-Statistic
-1.275593
1.414989
-1.693663
0.094457
0.342921
-0.210923
Prob.
0.2027
0.1577
0.0910
0.9248
0.7318
0.8330
36