Ch. 8 Nonspherical Disturbance
This chapter will assume that the full ideal conditions hold except that the covari-
ance matrix of the disturbance , i.e. E(""0) = 2 , where is not the identity
matrix. In particular, may be nondiagonal and/or have unequal diagonal ele-
ments.
Two cases we shall consider in details are heteroscedasticity and auto-
correlation. Disturbance are heteroscedastic when they have di erent variance.
Heteroscedasticity usually arise in cross-section data where the scale of the de-
pendent variable and the explanatory power of the model tend to vary across
observations. The disturbance are still assumed to be uncorrelated across obser-
vation, so 2 would be
2 =
2
66
66
66
4
21 0 : : : 0
0 22 : : : 0
: : : : : :
: : : : : :
: : : : : :
0 0 : : : 2T
3
77
77
77
5
:
Autocorrelation is usually found in time-series data. Economic time-series
often display a "memory" in that variation around the regression function is
not independent from one period to the next. Time series data are usually ho-
moscedasticity, so 2 would be
2 = 2
2
66
66
66
4
1 1 : : : T 1
1 1 : : : T 2
: : : : : :
: : : : : :
: : : : : :
T 1 T 2 : : : 1
3
77
77
77
5
:
In recent studies, panel data sets, constituting of cross sections observed at
several points in time, have exhibited both characteristics. The next three chap-
ter examines in details speci c types of generalized regression models.
Our earlier results for the classical mode; will have to be modi ed. We rst
consider the consequence of the more general model for the least squares estima-
tors.
1
1 Properties of the Least Squares Estimators
Theorem:
The OLS estimator ^ is unbiased. Furthermore, if limT!1(X0 X=T) is nite, ^
is consistent.
Proof:
E(^ ) = + E(X0X) 1X0" = , which proves unbiasedness.
Also
plim ^ = + lim
T!1
X0X
T
1
plimX
0"
T :
But X0"T has zero mean and covariance matrix
2X
0 X
T2 :
If limT!1(X0 X=T) is nite, then 2T X0 XT = 0. Hence X0"T has zero mean and
its covariance matrix vanishes asymptotically, which implies plim X0"T = 0, and
therefore, plim ^ = .
Theorem:
The covariance matrix of ^ is 2(X0X) 1X0 X(X0X) 1.
Proof:
E(^ )(^ )0 = E(X0X) 1X0""0X(X0X) 1
= 2(X0X) 1X0 X(X0X) 1:
Note that the covariance matrix of ^ is no longer equal to 2(X0X) 1. It
may be either "larger" or "smaller", in the sense that (X0X) 1X0 X(X0X) 1
(X0X) 1 can be either positive semide nite, negative semide nite, or neither.
Theorem:
s2 = e0e=(T k) is (in general) a biased and inconsistent estimator of 2.
2
Proof:
E(e0e) = E("0M")
= trace E(M""0)
= 2 trace M
6= 2(T k):
Also since E(s2) 6= 2, it is hard to see that it is a consistent estimator of 2
from convergence in mean square error.
2 E cient Estimators
To begin, it is useful to consider cases in which is a known, symmetric, positive
de nite matrix. This assumption will occasionally be true, but in most models,
will contains unknown parameters that must also be estimated.
Example:
Assume that 2t = 2x2t, then
2 =
2
66
66
66
4
2x21 0 : : : 0
0 2x22 : : : 0
: : : : : :
: : : : : :
: : : : : :
0 0 : : : 2x2T
3
77
77
77
5
= 2
2
66
66
66
4
x21 0 : : : 0
0 x22 : : : 0
: : : : : :
: : : : : :
: : : : : :
0 0 : : : x2T
3
77
77
77
5
;
therefore, we have a "known" .
2.1 Generalized Least Square (GLS) Estimators
Since is a positive symmetric matrix, it can be factored into
1 = C 1C0 = C 1=2 1=2C0 = P0P;
, where the column of C are the eigenvectors of and the eigenvalues of are
arrayed in the diagonal matrix and P0 = C 1=2.
3
Theorem:
Suppose that the regression model Y = X +" satisfy the ideal conditions except
that is not the identity matrix. Suppose that
lim
T!1
X0 1X
T
is nite and nonsingular. Then the transformed equation
PY = PX + P"
satis es the full ideal condition.
Proof:
Since P is nonsingular and nonstochastic, PXis nonstochastic and of full rank if
X is. (Condition 2 and 5). Also, for the consistency of OLS estimators
lim
T!1
(PX)0(PX)
T = limT!1
X0 1X
T
is nite and nonsingular by assumption. Therefore the transformed regressors ma-
trix satis es the required conditions, and we need consider only the transformed
disturbance P".
Clearly, E(P") = 0 (Condition 3). Also
E(P")(P")0 = 2P P0
= 2( 1=2C0)(C C0)(C 1=2)
= 2 1=2 1=2
= 2I (Condition 4):
Finally, the normality (Condition 6) of P" follows immediately from the nor-
mality of ".
Theorem:
The BLUE of is just
~ = (X0 1X) 1X0 1Y:
Proof:
Since the transformed equation satis es the full ideal conditions, the BLUE of
4
is just
~ = [(PX)0(PX)] 1(PX)0(PY)
= (X0 1X) 1X0 1Y:
Indeed, since ~ is the OLS estimator of in the transformed equation, and since
the transformed equation satis es the ideal conditions, ~ has all the usual de-
sirable properties{it is unbiased, BLUE, e cient, consistent, and asymptotically
e cient.
~ is the OLS of the transformed equation, but it is a generalized least square
(GLS) estimator of the original regression model which take the OLS as a sub-
cases when = I.
Theorem:
The variance -covariance of the GLS estimator ~ is 2(X0 1X) 1.
Proof:
Viewing ~ as the OLS estimator in the transformed equation, it is clearly has
covariance matrix
2[(PX)0(PX)] 1 = 2(X0 1X) 1:
Theorem:
An unbiased, consistent, e cient, and asymptotically e cient estimator of 2 is
~s2 = ~e
0 1~e
T k ;
where ~e = Y X~ .
Proof:
Since the transformed equation satis es the ideal conditions, the desired estimator
of 2 is
1
T k(PY PX
~ )0(PY PX~ ) = 1
T k(Y X
~ )0 1(Y X~ ):
5
Finally, for testing hypothesis we can apply the full set of results in Chapter
6 to the transformed equation. For the testing the m restrictions R = q, the
appropriate (one of) statistics is
(R~ q)0[~s2R(PX)0(PX) 1R0] 1(R~ q)
m
= (R
~ q)0[~s2R(X0 1X) 1R0] 1(R~ q)
m Fm;T k:
Exercise:
Derive the other three test statistics (in Chapter 6) of the F Ratio test statistics
to test the hypothesis R = q when 6= I.
2.2 Maximum Likelihood Estimators
Assume that " N(0; 2 ), if X are not stochastic, then by results from "func-
tions of random variables" (n ) n transformation) we have Y N(X ; 2 ).
That is, the log-likelihood function
ln f( ; Y ) = T2 ln(2 ) 12 lnj 2 j 12(Y X )0( 2 ) 1(Y X )
= T2 ln(2 ) T2 ln 2 12 lnj j 12 2(Y X )0 1(Y X )
where = ( 1; 2; :::; k; 2)0 since by assumption is known.
The necessary condition for maximizing L are
@L
@ =
1
2X
0 1(Y X ) = 0
@L
@ 2 =
T
2 2 +
1
2 4(Y X )
0 1(Y X ) = 0
The solution are
^ ML = (X0 1X) 1X0 1Y;
^ 2ML = 1
T (Y X
^ ML)0 1(Y X^ ML);
6
which implies that with normally distributed disturbance, generalized l;east squares
are also MLE. As is the classical regression model, the MLE of 2 is biased. An
unbiased estimator is
^ 2 = 1
T k(Y X
^ ML)0 1(Y X^ ML):
3 Estimation When is Unknown
If contains unknown parameters that must be estimated, then GLS is not
feasible. But with an unrestricted , there are T(T +1)=2 additional parameters
in 2 . This number is far too many to estimate with T observations. Obviously,
some structures must be imposed on the model if we are to proceed.
3.1 Feasible Generalized Least Squares
The typical problem involves a small set parameter such that = ( ). For
example, we may assume autocorrelated disturbance in the beginning of this
chapter as
2 = 2
2
66
66
66
4
1 1 : : : T 1
1 1 : : : T 2
: : : : : :
: : : : : :
: : : : : :
T 1 T 2 : : : 1
3
77
77
77
5
= 2
2
66
66
66
4
1 1 : : : T 1
1 1 : : : T 2
: : : : : :
: : : : : :
: : : : : :
T 1 T 2 : : : 1
3
77
77
77
5
;
then has only one additional unknown parameters, . A model of heteroscedas-
ticity that also has only one new parameters, , is
2t = 2x 2t:
De nition:
If depends on a nite number of parameters, 1; 2; :::; p, and if ^ depends on
consistent estimator, ^ 1; ^ 2; :::; ^ p, the ^ is called a consistent estimator of .
De nition:
Let ^ be a consistent estimator of . Then the feasible generalized least square
estimator (FGLS) of is
= (X0^ 1X) 1X0^ 1Y:
7
Conditions that imply that is asymptotically equivalent to ~ are
lim
T!1
1
T X
0^ 1X
1
T X
0 1X
= 0
and
lim
T!1
1
pT X0^ 1"
1
pT X0 1"
= 0:
Theorem:
An asymptotically e cient FGLS does not require that we have an e cient es-
timator of ; only a consistent one is required to achieve full e ciency for the
FGLS estimator.
8