SAS/STAT 9.2 User's Guide: The MIXED Procedure (Book Excerpt)
SAS/STAT 9.2 User's Guide: The MIXED Procedure (Book Excerpt)
SAS/STAT 9.2 User's Guide: The MIXED Procedure (Book Excerpt)
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
3986 ✦ Chapter 56: <strong>The</strong> <strong>MIXED</strong> <strong>Procedure</strong><br />
<strong>The</strong> <strong>MIXED</strong> procedure computes DFFITS when the EFFECT= or SIZE= modifier of the<br />
INFLUENCE option is not in effect. In general, an external estimate of the estimated standard<br />
error is used. When ITER > 0, the estimate is<br />
ese.byi/ D<br />
q<br />
x 0 i .X0 V.b .u// X/ 1 xi<br />
When ITER=0 and 2 is profiled, then<br />
q<br />
ese.byi/ D b .u/<br />
x 0 i .X0 V.b / 1 X/ xi<br />
When the EFFECT=, SIZE=, or KEEP= modifier is specified, the <strong>MIXED</strong> procedure computes a<br />
multivariate version suitable for the deletion of multiple data points. <strong>The</strong> statistic, termed MDFFITS<br />
after the MDFFIT statistic of Belsley, Kuh, and Welsch (1980, p. 32), is closely related to Cook’s<br />
D. Consider the case V D 2 V. / so that<br />
VarŒbˇ D 2 .X 0 V. / 1 X/<br />
and let fVarŒbˇ .U / be an estimate of VarŒbˇ .U / that does not use the observations in U . <strong>The</strong> MDF-<br />
FITS statistic is then computed as<br />
MDFFITS.ˇ/ D ı 0<br />
.U / fVarŒbˇ .U / ı .U /=rank.X/<br />
If ITER=0 and 2 is profiled, then fVarŒbˇ .U / is obtained by sweeping<br />
b 2<br />
.U / .X0<br />
.U / V .U /.b / X .U //<br />
<strong>The</strong> underlying idea is that if were known, then<br />
.X 0<br />
.U / V .U /. / 1 X .U //<br />
would be VarŒbˇ= 2 in a generalized least squares regression with all but the data in U .<br />
In the case of iterative influence analysis, fVarŒbˇ .U / is evaluated at b .U /. Furthermore, a MDFFITStype<br />
statistic is then computed for the covariance parameters:<br />
MDFFITS. / D .b b .U // 0 cVarŒb .U / 1 .b b .U //<br />
Covariance Ratio and Trace<br />
<strong>The</strong>se statistics depend on the availability of an external estimate of V, or at least of 2 . Whereas<br />
Cook’s D and MDFFITS measure the impact of data points on a vector of parameter estimates, the<br />
covariance-based statistics measure impact on their precision. Following Christensen, Pearson, and<br />
Johnson (1992), the <strong>MIXED</strong> procedure computes<br />
CovTrace.ˇ/ D jtrace. cVarŒbˇ fVarŒbˇ .U // rank.X/j<br />
CovRatio.ˇ/ D detns. fVarŒbˇ .U //<br />
detns. cVarŒbˇ/<br />
where detns.M/ denotes the determinant of the nonsingular part of matrix M.