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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Conduction velocity<br />

30<br />

25<br />

20<br />

15<br />

10<br />

5<br />

–1 . 0<br />

0 0 . 0 1 . 0 2 . 0 3 . 0 4 . 0<br />

log(age + 0 . (a) 75)<br />

(b)<br />

Conduction velocity<br />

25<br />

20<br />

15<br />

10<br />

5<br />

Conduction velocity<br />

Conduction velocity<br />

0<br />

–1<br />

(c) (d)<br />

. 0 0 . 0 1 . 0 2 . 0 3 . 0 4 . 0<br />

log(age + 0 . 75)<br />

25<br />

20<br />

15<br />

10<br />

5<br />

–1 . 0<br />

0<br />

25<br />

20<br />

15<br />

10<br />

5<br />

–1 . 0<br />

0<br />

0 . 0 1 . 0 2 . 0 3 . 0<br />

log(age + 0 . 75)<br />

0 . 0 1 . 0 2 . 0 3 . 0<br />

log(age + 0 . 75)<br />

Fig. 12.4 Nerve conduction velocity: (a) top left, raw data, vertical l<strong>in</strong>es at full term, age 1, 10, 30 and 50<br />

years; (b) top right, runn<strong>in</strong>g mean (- - -), runn<strong>in</strong>g l<strong>in</strong>e ( ), lowess (ÐÐ); (c) bottom left, Gaussian<br />

kernel smooth<strong>in</strong>g with smooth<strong>in</strong>g parameter 0 06 (± ± ±) and 0 37 (ÐÐ); (d) bottom right, penalized<br />

least squares a ˆ 1 0 (cont<strong>in</strong>uous), GCV ( ) and fractional polynomial degree 2, p ˆ (1, 3) (± ± ±).<br />

weight, so the effect on s(x) is far less abrupt. This expla<strong>in</strong>s why the lowess smoother is the<br />

smoothest of the three estimates of s(x) <strong>in</strong> this figure.<br />

Figure 12.4(c) shows the effect of fitt<strong>in</strong>g two Gaussian kernel smoothers to the data.<br />

The parameter h <strong>in</strong> (12 8) is set at 0 06 and 0 37. It seems implausible that mean conduction<br />

velocities should go up and down with age <strong>in</strong> the rather `wiggly' way predicted by the<br />

estimate of s…x† for the smaller h, suggest<strong>in</strong>g that a larger value of h, such as <strong>in</strong> the other<br />

estimate shown <strong>in</strong> Fig. 12.4(c), is to be preferred.<br />

Penalized least squares and smooth<strong>in</strong>g spl<strong>in</strong>es<br />

12.2 Smooth<strong>in</strong>g and non-parametric regression 391<br />

The local nature of the smooth<strong>in</strong>g methods just described is achieved quite<br />

directlyÐthe methods deliberately construct estimates of s(x) that rely only, or<br />

largely, on those observations that are with<strong>in</strong> a certa<strong>in</strong> distance of x. Regression<br />

l<strong>in</strong>es, on the other hand, are obta<strong>in</strong>ed by optimiz<strong>in</strong>g a suitable criterion that<br />

expresses a desirable property for the estimator. So, for example, <strong>in</strong> the case of<br />

normal errors, regression l<strong>in</strong>es are obta<strong>in</strong>ed by m<strong>in</strong>imiz<strong>in</strong>g<br />

4.0<br />

4 . 0

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

Saved successfully!

Ooh no, something went wrong!