15.01.2013 Views

an introduction to generalized linear models - GDM@FUDAN ...

an introduction to generalized linear models - GDM@FUDAN ...

an introduction to generalized linear models - GDM@FUDAN ...

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.

Central F<br />

Non-central F<br />

Figure 2.5 Central <strong>an</strong>d non-central F distributions.<br />

tion. However, as σ 2 is unknown, we c<strong>an</strong>not compare ( � S0 − � S1)/σ 2 directly<br />

with the χ 2 (J − 1) distribution. Instead we eliminate σ 2 by using the ratio<br />

of( � S0 − � S1)/σ 2 <strong>an</strong>d the r<strong>an</strong>dom variable � S1/σ 2 with a central chi-squared<br />

distribution, each divided by the relev<strong>an</strong>t degrees offreedom,<br />

F = ( � S0 − � S1)/σ 2<br />

(J − 1)<br />

/<br />

�S1/σ 2<br />

(JK − 2J) = ( � S0 − � S1)/(J − 1)<br />

�S1/(JK − 2J)<br />

IfH0 is correct, from Section 1.4.4, F has the central distribution F (J −<br />

1,JK − 2J). IfH0 is not correct, F has a non-central F -distribution <strong>an</strong>d<br />

the calculated value of F will be larger th<strong>an</strong> expected from the central F -<br />

distribution (see Figure 2.5).<br />

For the example on birthweight <strong>an</strong>d gestational age, the value of F is<br />

(658770.8 − 652424.5)/1<br />

652424.5/20<br />

=0.19<br />

This value is certainly not statistically signific<strong>an</strong>t when compared with the<br />

F (1, 20) distribution. Thus the data do not provide evidence against the hypothesis<br />

H0 : β0 = β1, <strong>an</strong>d on the grounds ofsimplicity model (2.6), which<br />

specifies the same slopes but different intercepts, is preferable.<br />

These two examples illustrate the main ideas <strong>an</strong>d methods ofstatistical<br />

modelling which are now discussed more generally.<br />

2.3 Some principles of statistical modelling<br />

2.3.1 Explora<strong>to</strong>ry data <strong>an</strong>alysis<br />

Any <strong>an</strong>alysis ofdata should begin with a consideration ofeach variable separately,<br />

both <strong>to</strong> check on data quality (for example, are the values plausible?)<br />

<strong>an</strong>d <strong>to</strong> help with model formulation.<br />

1. What is the scale ofmeasurement? Is it continuous or categorical? Ifit<br />

© 2002 by Chapm<strong>an</strong> & Hall/CRC<br />

.

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

Saved successfully!

Ooh no, something went wrong!