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.

202 Regression and correlation<br />

y , Increase <strong>in</strong> weight between 70 and 100 days, as % of x<br />

120<br />

100<br />

80<br />

60<br />

40<br />

20<br />

0<br />

80 90 100 110 120 130 140 150<br />

x , Birth weight (oz)<br />

Fig. 7.7 Confidence limits (95%) for the predicted mean value of y for specified values, x0, ofx (data<br />

as <strong>in</strong> Fig. 7.4).<br />

In (7.21) the width of the confidence <strong>in</strong>terval <strong>in</strong>creases with …x0 x† 2 , and is<br />

therefore a m<strong>in</strong>imum when x0 ˆ x. Figure 7.7 shows the limits for various values<br />

of x0 <strong>in</strong> Example 7.1. The reason for the <strong>in</strong>crease <strong>in</strong> the width of the <strong>in</strong>terval is<br />

that slight sampl<strong>in</strong>g errors <strong>in</strong> b will have a greater effect for values of x0 distant<br />

from x than for those near x. The regression l<strong>in</strong>e can be thought of as a rod<br />

which is free to move up and down (correspond<strong>in</strong>g to the error <strong>in</strong> y) and to<br />

pivot about a central po<strong>in</strong>t (correspond<strong>in</strong>g to the error <strong>in</strong> b). Po<strong>in</strong>ts near the<br />

ends of the rod will then be subject to greater oscillations than those near<br />

the centre.<br />

A different prediction problem is that of estimat<strong>in</strong>g an <strong>in</strong>dividual value y0 of<br />

y correspond<strong>in</strong>g to a given x0. The best s<strong>in</strong>gle estimate is aga<strong>in</strong> the value given by<br />

the regression equation, but the limits of error are different from those <strong>in</strong> the<br />

previous case. Two slightly different problems can be dist<strong>in</strong>guished.<br />

1 A s<strong>in</strong>gle prediction is required. The appropriate limits are not strictly confidence<br />

limits because the quantity to be estimated is a value taken by a<br />

random variable, not a parameter of a distribution. However, writ<strong>in</strong>g<br />

y0 ˆ Y ‡ e,<br />

where e is the deviation of y0 from the predicted value Y, we have

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

Saved successfully!

Ooh no, something went wrong!