Professional Documents
Culture Documents
Generalized Least Squares and Heteroskedasticity: Walter Sosa-Escudero
Generalized Least Squares and Heteroskedasticity: Walter Sosa-Escudero
Generalized Least Squares and Heteroskedasticity: Walter Sosa-Escudero
Heteroskedasticity
February 4, 2009
Walter Sosa-Escudero
Linearity: Y = X + u.
Walter Sosa-Escudero
Walter Sosa-Escudero
V (|X)
will now be 2 (X 0 X)1 (X 0 X)1 (check it).
V (|X)
is no longer 2 (X 0 X)1 , and S 2 (X 0 X)1 will be a
of and V (|X)
= S 2 (X 0 X)1 , keeps estimation of unbiased
though ineficient, and invalidates all standard inference.
Walter Sosa-Escudero
Walter Sosa-Escudero
Now check:
1
(X ) = (CX) = K, w.p.1.
(u |X)
Walter Sosa-Escudero
Note that
gls = (X X )1 X Y
= (X 0 C 0 CX)1 X 0 C 0 CY
= (X 0 1 X)X 0 1 Y
When = In , gls = .
The practical implementation of gls requires that we know
(though not 2 .)
It is easy to check that V (gls ) = 2 (X 0 X )1 .
Walter Sosa-Escudero
Feasible GLS
Then, replacing
Suppose there is an estimate for , label it .
by :
1 X)X 0
1 Y
f gls = (X 0
This is the feasible GLS estimator.
Is it linear and unbiased? Efficient?
Walter Sosa-Escudero
Heteroskedasticity
Suppose we keep the no-serial correlation assumption, but allow for
heteroskedasticity, that is:
V (u|X) = 2 diag(12 , 22 , . . . , n2 )
It is easy to see that in this case
C = diag(11 , 21 , . . . , n1 )
Hence
X = CX
is a matrix with typical element Xik /i1 , i = 1, . . . , n,
k = 1, . . . , K. Y = CY is constructed in the same fashion.
gls for this particular case is called the weighted least squares
estimator.
Walter Sosa-Escudero
Xik
=
Xik
Xij
ui
Xij
Walter Sosa-Escudero
Walter Sosa-Escudero
1, . . . , n
Walter Sosa-Escudero
Walter Sosa-Escudero
Walter Sosa-Escudero
Comments:
Intuition is as in the White test (a model for the variance).
By focusing on a particular group, if we reject the null we
have a better idea of what causes heterocedasticity.
Accepting the null does not mean there isnt heterocedasticity
(why?).
Also a large sample test.
2 as a test, which is
Koenker (1980) has proposed to use nRA
still valid if errors are non-normal.
Walter Sosa-Escudero