01.06.2013 Views

Statistical Methods in Medical Research 4ed

Statistical Methods in Medical Research 4ed

Statistical Methods in Medical Research 4ed

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

ti ˆ<br />

ei<br />

s… i† …<br />

p 1 hi†<br />

gives the externally studentized residual or jackknife residual. From (11.67),<br />

ti ˆ ri<br />

s<br />

s … i†<br />

11.9 Check<strong>in</strong>g the model 367<br />

…11:68†<br />

…11:69†<br />

and it can be shown mathematically that<br />

n p 2<br />

ti ˆ ri<br />

n p 1 r2 r<br />

: …11:70†<br />

i<br />

so that the jackknife residuals are easily computed for every po<strong>in</strong>t without<br />

hav<strong>in</strong>g to fit the n separate regressions, each exclud<strong>in</strong>g one po<strong>in</strong>t. The numerator<br />

and denom<strong>in</strong>ator of (11.68) are uncorrelated and the jackknife residuals are<br />

distributed exactly as the t distribution on n p 2DF.<br />

The def<strong>in</strong>ition of ti might seem to be no more than an adjustment us<strong>in</strong>g a<br />

more appropriate estimate of standard deviation, but it is <strong>in</strong> fact much more than<br />

this. Suppose the residual were redef<strong>in</strong>ed as the difference between the observed<br />

yi and the value, Y… i†, that would be predicted for the ith po<strong>in</strong>t us<strong>in</strong>g the<br />

regression based on all the po<strong>in</strong>ts except that po<strong>in</strong>t. The standard error of this<br />

residual could be calculated, us<strong>in</strong>g the multivariate extension of (7.22). Then<br />

def<strong>in</strong>e as a standardized residual<br />

ti ˆ yi Y… i†<br />

: …11:71†<br />

SE…yi Y… i††<br />

The residual def<strong>in</strong>ed <strong>in</strong> this way, as an externally studentized residual, is mathematically<br />

equivalent to the jackknife residual as def<strong>in</strong>ed <strong>in</strong> (11.70). Therefore<br />

each jackknife residual is a standardized residual about the regression fitted<br />

us<strong>in</strong>g all the data except the po<strong>in</strong>t under consideration. This rather remarkable<br />

mathematical property h<strong>in</strong>ges on the leverage. From (11.70) the studentized<br />

residual, ri, cannot exceed …<br />

p n p 1†, and this limit is only reached when<br />

deletion of the ith po<strong>in</strong>t results <strong>in</strong> a regression that is a perfect fit to all the other<br />

po<strong>in</strong>ts (Gray & Woodall, 1994). In this extreme case the jackknife residual has an<br />

<strong>in</strong>f<strong>in</strong>itely high value. Mathematically <strong>in</strong>cl<strong>in</strong>ed readers are referred to the books<br />

by Cook and Weisberg (1982) and Belsley (1991) for a full mathematical treatment.<br />

These residuals are given <strong>in</strong> Table 11.7 for the data shown <strong>in</strong> Fig. 11.10. The<br />

column headed D10 will be referred to later. S<strong>in</strong>ce all the residuals are standardized<br />

and <strong>in</strong>dependent of the scale on which y and x are measured, these scales<br />

have been chosen so that the slope of the l<strong>in</strong>e and the standard deviation <strong>in</strong><br />

Fig. 11.10(a) are both unity. For Fig. 11.10(b to e) the residual diagnostic statistics<br />

are given for the extra 10th po<strong>in</strong>t. When judged aga<strong>in</strong>st the t distribution with 8

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!