Estimation with eteroscedaticity

Prof. Rizzi Laura

laura.rizzi@uniud.it

November 2023
DIES

1
Outline
Brief Overview on:

• Weighted Least Squares

• Example 1

2
Weighted Least Squares
Weighted Least Squares

If the test procedure points out the presence of eteroscedastic errors, it is necessary to verify if
there are specification errors first.
If there are not specification errors we may:

• use Weighted least squares (WLS) and not OLS. The WLS are a special form of GLS
(generalized least squares);
• define the variables and transform them on the basis of the underlying theory;

WLS
Usually we don’t know the form of eteroscedasticity. We may suppose, however, that:

2 α
σt = α0 + α1 zt 2

With t = 1, 2, . . . , n, where the r.v. z is a model regressor.

3
Weighted Least Squares

• If α1 = 0 there is homoscedasticity and σt2 = α0 (> 0);

• if α0 = 0 and α2 = 1 the errors variance is proportional to z;
• if α0 = 0 and α2 = 2 the errors variance is proportional to z 2

Often we assume one of these forms of eteroscedasticity and then the GLS take a particular form
called WLS (weighted least squares).
~ , then it follows:
If we suppose to have σt2 = α1 zt , defining the E(uuT ) = V

σ12
   
0 ... 0 z1 0 ... 0
 0 σ22 ... 0   0 z2 ... 0 
~ =
   
V . . .  = α1  . . .  = α1 Ω
 . . .   . . . 
 . 0 . .   . 0 . . 
2
0 0 ... σn 0 0 ... zn

4
Weighted Least Squares

Considering the constituents of the GLS estimator we derive:

 
. . 1
··· 0
 
 .. . z1 y1
.  n
T
~ Ω ~ −1   . . .

.  X 1
X ~
y=
 ~
x1 ··· ~
xn . . . .  = ( )~xt yt
 
. . . . 

 .
  zt
. 
1
t=1
. . 0 ··· zn
yn
. .

In the same way we may verify that:

n
~ −1 X
~TΩ ~ =
X 1 T
X ( )~xt ~
xt
t=1
zt

Then the GLS estimates are derived from the:

n n
−1
X T
X
bGLS = [ ~
xt ~
xt /zt ] [ ~
xt yt /zt ]
t=1 t=1

5
Weighted Least Squares

If yt and each element of xt are multiplied by the square root of the reciprocal of zt the
application of the OLS estimator to the transformed data will give the bGLS estimator.
If we assume that σt2 = α1 zt2 the appropriate weighting factor is the reciprocal of zt .

An alternative approach would imply to not derive the transformed data but to estimate the
xT
structure of eteroscedasticity through the OLS residuals, et = yt − ~ ~
t β, which are consistent
estimators of ut .
We may estimate by nonlinear regression:

2 αˆ
et = αˆ0 + αˆ1 zt 2 + v̂t

Then the estimates of the errors variances are:

σˆt2 = αˆ0 + αˆ1 zt 2

αˆ

With t = 1, 2, . . . , n.

6
Weighted Least Squares

Problems related to GLS estimation:

• there may be problems related to the choice of the proportionality factor, z;
• there may be problems related to the specification of the functional form of the relationship
between errors variances and the proportionality factor;
• it is important to avoid GLS application when eteroscedasticity is due to model specification
errors;

Example - 1:
We consider a random sample of n = 100 observations from the Current Population Survey, 1988.
A conventional earnings equation has been estimated on these data. The variables considered are:

• LNWAGE - log of wage, the dependent variable;

• GRADE - years of education;
• POTEXP and EXP2 - years of experience and its square;
• UNION - dummy variable for the membership in a union.

7
Example - 1

The results of estimated equation are reported below and confirm that education has a positive
and significant effect, experience has a quadratic effect and the union dummy variable has a
positive but not very significant coefficient.

Variable Coefficient S.E. t-statistic Prob.

Constant 0.595 0.283 2.099 0.038
GRADE 0.083 0.020 4.158 0.000
POTEXP 0.050 0.014 3.556 0.000
EXP2 -0.000 0.000 1.951 0.054
UNION 0.166 0.124 1.333 0.185
R2 = 0.371 R̄2 = 0.345 F-statistic = 14.05 Prob(F)= 0.000

To apply the White test for eteroscedasticity we need first to square the regression residuals
(RESSQ) and regress these versus the original covariates, their squares and cross-products. In the
auxliary regression derived has twelve covariates (not the square of UNION).The results of the
estimated auxiliary regression are below.

8
Example - 1

Variable Coefficient S.E. t-statistic Prob.

Constant -0.077 0.985 -0.078 .937
GRADE -0.012 0.125 -0.097 0.922
POTEXP 0.077 0.071 -0.097 0.281
EXP2 -0.003 0.004 -0.974 0.332
UNION 0.648 0.861 0.753 0.453
GRADE2 0.002 0.004 0.516 0.606
EXP4 -3.34E-07 1.51E-06 -0.220 0.825
EXP3 6.17E-05 0.000 0.434 0.664
GRADEXP -0.003 0.004 -0.759 0.449
GRADEXP2 0.000 0.000 1.052 0.295
GRADUNION -0.051 0.044 -1.159 0.249
POTEXPUNI 0.001 0.060 0.031 0.974
EXP2UNI -0.000 0.001 -0.176 0.860
R2 = 0.107

The test statistic is nR2 = 10.79 and χ20.05 = 21.03 then the null hypothesis of omoscedasticity
is not rejected.

9
Example - 1

To apply the Breush - Pagan test we have to define the variables that influence eteroscedasticity.
We select: GRADE, POTEXP and UNION. Then we regress RESSQ = e2t versus these variables
and correcting for the scale factor σ̃ 2 = 0.209.
We obtain the ESS of this auxiliary regression, then:

1 1 R2
ESS = RSS = 5.35
2 2 1 − R2

The relevant critical value is χ20.05 (3) = 7.815, then omoscedasticity is not rejected.

10