GLS (Generalized Least Squares) - Paul Johnson Homepage
GLS (Generalized Least Squares) - Paul Johnson Homepage
GLS (Generalized Least Squares) - Paul Johnson Homepage
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
⎡<br />
V ar(e) =<br />
⎢<br />
⎣<br />
In weighted least squares, we use estimates of σ 2 e i<br />
σe 2 ⎤<br />
1<br />
0 0 0<br />
0 σe 2 2<br />
0 0<br />
. .. 0<br />
.<br />
0<br />
..<br />
0 0 σe 2 ⎥<br />
N−1<br />
0 ⎦<br />
0 0 0 0 σe 2 N<br />
as weights.<br />
(6)<br />
S(ˆb) =<br />
N∑<br />
i=0<br />
1<br />
̂σ 2 e i<br />
(y i − ŷ i ) 2 =<br />
Consider the WLS problem. If the heteroskedastic case occurs<br />
N∑<br />
w i (y i − ŷ i ) 2 (7)<br />
i=0<br />
N∑<br />
N∑<br />
S(ˆb) = w i (y i − ŷ i ) 2 = ( √ w i y i − √ w i ŷ i ) 2 (8)<br />
i=1<br />
i=1<br />
= w 1 (y 1 − ŷ 1 ) 2 + w 2 (y 2 − ŷ 2 ) 2 + · · · + w N (y N − ŷ N ) 2 (9)<br />
3 <strong>Generalized</strong> <strong>Least</strong> <strong>Squares</strong> (<strong>GLS</strong>)<br />
Suppose, generally, the Variance/Covariance matrix of residuals is<br />
⎡<br />
σ1 2 Cov(e 1 , e 2 ) Cov(e 1 , e 2 ) · · · Cov(e 1 , e N )<br />
Cov(e 1 , e 2 ) σ2<br />
2 σ3<br />
2<br />
V =<br />
. .<br />
..<br />
⎢<br />
⎣<br />
σN−1 2 Cov(e 1 , e N−1 )<br />
Cov(e 1 , e N ) Cov(e 1 , e N−1 σN<br />
2<br />
⎤<br />
⎥<br />
⎦<br />
(10)<br />
You can factor out a constant σ 2 if you care to (MM&V do, p. 51).<br />
If all the off diagonal elements of V are set to 0, then this degenerates to a problem of heteroskedasticity,<br />
know as Weighted <strong>Least</strong> <strong>Squares</strong> (WLS).<br />
If the off diagonal elements of V are not 0, then there can be correlation across cases. In a<br />
time series problem, that amounts to “autocorrelation.” In a cross-sectional problem, it means<br />
that various observations are interrelated.<br />
The “sum of squared errors” approach uses this variance matrix to adjust the data so that the<br />
residuals are homoskedastic or that the cross-unit correlations are taken into account.<br />
Let<br />
W = V −1 (11)<br />
The idea behind WLS/<strong>GLS</strong> is that there is a way to transform y and X so that the residuals<br />
are homogeneous, some kind of weight is applied.<br />
2