Добавил:
Upload Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Скачиваний:
668
Добавлен:
03.06.2015
Размер:
8.25 Mб
Скачать

32—Chapter 19. Additional Regression Tools

Dependent Variable: LPRICE

Method: Least Squares

Date: 08/08/09 Time: 22:15

Sample: 1 506

Included observations: 506

Variable

Coefficient

Std. Error

t-Statistic

Prob.

 

 

 

 

 

 

 

 

 

 

C

8.811812

0.217787

40.46069

0.0000

LNOX

-0.487579

0.084998

-5.736396

0.0000

ROOMS

0.284844

0.018790

15.15945

0.0000

RADIAL=1

0.118444

0.072129

1.642117

0.1012

RADIAL=2

0.219063

0.066055

3.316398

0.0010

RADIAL=3

0.274176

0.059458

4.611253

0.0000

RADIAL=4

0.149156

0.042649

3.497285

0.0005

RADIAL=5

0.298730

0.037827

7.897337

0.0000

RADIAL=6

0.189901

0.062190

3.053568

0.0024

RADIAL=7

0.201679

0.077635

2.597794

0.0097

RADIAL=8

0.258814

0.066166

3.911591

0.0001

 

 

 

 

 

 

 

 

 

 

R-squared

0.573871

Adjusted R-squared

0.565262

S.E. of regression

0.269841

Sum squared resid

36.04295

Log likelihood

-49.60111

F-statistic

66.66195

Prob(F-statistic)

0.000000

Mean dependent var

9.941057

S.D. dependent var

0.409255

Akaike info criterion

0.239530

Schwarz criterion

0.331411

Hannan-Quinn criter.

0.275566

Durbin-W atson stat

0.671010

Robust Standard Errors

In the standard least squares model, the coefficient variance-covariance matrix may be derived as:

S =

E(b b)(b b)¢

 

 

ˆ

ˆ

 

=

(X¢X)–1E(X¢ee¢X)(X¢X)–1

(19.8)

=

(X¢X)–1T Q (X¢X)–1

 

=

j2(X¢X)–1

 

A key part of this derivation is the assumption that the error terms, e , are conditionally homoskedastic, which implies that Q = E(X¢ee¢X § T) = j2(X¢X § T) . A sufficient, but not necessary, condition for this restriction is that the errors are i.i.d. In cases where this assumption is relaxed to allow for heteroskedasticity or autocorrelation, the expression for the covariance matrix will be different.

EViews provides built-in tools for estimating the coefficient covariance under the assumption that the residuals are conditionally heteroskedastic, and under the assumption of heteroskedasticity and autocorrelation. The coefficient covariance estimator under the first assumption is termed a Heteroskedasticity Consistent Covariance (White) estimator, and the

Robust Standard Errors—33

estimator under the latter is a Heteroskedasticity and Autocorrelation Consistent Covariance (HAC) or Newey-West estimator. Note that both of these approaches will change the coefficient standard errors of an equation, but not their point estimates.

Heteroskedasticity Consistent Covariances (White)

White (1980) derived a heteroskedasticity consistent covariance matrix estimator which provides consistent estimates of the coefficient covariances in the presence of conditional

heteroskedasticity of unknown form. Under the White specification we estimate

Q using:

ˆ

 

T

T

 

2

 

 

=

Â

ˆ

XtXt¢ § T

 

Q

------------

 

(19.9)

T k

e

t

 

 

 

t = 1

 

 

 

 

ˆ

are the estimated residuals, T is the number of observations, k is the number of

where et

regressors, and T § (T k) is an optional degree-of-freedom correction. The degree-of-free-

dom White heteroskedasticity consistent covariance matrix estimator is given by

 

 

ˆ

=

T

(X¢X)

–1

T

e

2

 

–1

(19.10)

 

SW

------------

 

 

 

X X ¢ (X¢X)

 

 

 

 

 

 

tÂ= 1

ˆ

t

t t

 

 

 

 

 

T k

 

 

 

 

 

To illustrate the use of White covariance estimates, we use an example from Wooldridge (2000, p. 251) of an estimate of a wage equation for college professors. The equation uses dummy variables to examine wage differences between four groups of individuals: married men (MARRMALE), married women (MARRFEM), single women (SINGLEFEM), and the base group of single men. The explanatory variables include levels of education (EDUC), experience (EXPER) and tenure (TENURE). The data are in the workfile “Wooldridge.WF1”.

To select the White covariance estimator, specify the equation as before, then select the Options tab and select White in the Coefficient covariance matrix drop-down. You may, if desired, use the checkbox to remove the default d.f. Adjustment, but in this example, we will use the default setting.

The output for the robust covariances for this regression are shown below:

34—Chapter 19. Additional Regression Tools

Dependent Variable: LOG(WAGE)

Method: Least Squares

Date: 04/13/09 Time: 16:56

Sample: 1 526

Included observations: 526

White heteroskedasticity-consistent standard errors & covariance

Variable

Coefficient

Std. Error

t-Statistic

Prob.

 

 

 

 

 

 

 

 

 

 

C

0.321378

0.109469

2.935791

0.0035

MARRMALE

0.212676

0.057142

3.721886

0.0002

MARRFEM

-0.198268

0.058770

-3.373619

0.0008

SINGFEM

-0.110350

0.057116

-1.932028

0.0539

EDUC

0.078910

0.007415

10.64246

0.0000

EXPER

0.026801

0.005139

5.215010

0.0000

EXPER^2

-0.000535

0.000106

-5.033361

0.0000

TENURE

0.029088

0.006941

4.190731

0.0000

TENURE^2

-0.000533

0.000244

-2.187835

0.0291

 

 

 

 

 

 

 

 

 

 

R-squared

0.460877

Mean dependent var

1.623268

Adjusted R-squared

0.452535

S.D. dependent var

0.531538

S.E. of regression

0.393290

Akaike info criterion

0.988423

Sum squared resid

79.96799

Schwarz criterion

1.061403

Log likelihood

-250.9552

Hannan-Quinn criter.

1.016998

F-statistic

55.24559

Durbin-Watson stat

1.784785

Prob(F-statistic)

0.000000

 

 

 

 

 

 

 

 

As Wooldridge notes, the heteroskedasticity robust standard errors for this specification are not very different from the non-robust forms, and the test statistics for statistical significance of coefficients are generally unchanged. While robust standard errors are often larger than their usual counterparts, this is not necessarily the case, and indeed this equation has some robust standard errors that are smaller than the conventional estimates.

HAC Consistent Covariances (Newey-West)

The White covariance matrix described above assumes that the residuals of the estimated equation are serially uncorrelated. Newey and West (1987b) have proposed a more general covariance estimator that is consistent in the presence of both heteroskedasticity and autocorrelation of unknown form. They propose using HAC methods to form an estimate of

E(X¢ee¢X § T). Then the HAC coefficient covariance estimator is given by:

ˆ

= (X'X)

–1

ˆ

–1

(19.11)

SNW

 

T Q (X'X)

 

where ˆ is any of the LRCOV estimators described in Appendix E. “Long-run Covariance

Q

Estimation,” on page 775.

To use the Newey-West HAC method, select the Options tab and select HAC (Newey-West) in the Coefficient covariance matrix drop-down. As before, you may use the checkbox to remove the default d.f. Adjustment.

Robust Standard Errors—35

Press the HAC options button to change the options for the LRCOV estimate.

We illustrate the computation of HAC covariances using an example from Stock and Watson (2007, p. 620). In this example, the percentage change of the price of orange juice is regressed upon a constant and the number of days the temperature in Florida reached zero for the current and previous 18 months, using monthly data from 1950 to 2000 The data are in the workfile “Stock_wat.WF1”.

Stock and Watson report Newey-West standard errors computed using a non pre-whitened Bartlett Kernel with a user-specified bandwidth of 8 (note that the bandwidth is equal to one plus what Stock and Watson term the “truncation parameter” m ).

The results of this estimation are shown below:

Соседние файлы в папке EViews Guides BITCH