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.

492 Modell<strong>in</strong>g categorical data<br />

DF. Add<strong>in</strong>g the ma<strong>in</strong> effect of A to the model reduces the deviance to 13 59 with 11 DF, so<br />

that the deviance test for the effect of A, after allow<strong>in</strong>g for B, C and D, is 45 47 13 59 ˆ<br />

31 88 as an approximate x2 p<br />

…1† . This test is numerically similar to Wald's test, s<strong>in</strong>ce<br />

31 88 ˆ 5 65,<br />

but <strong>in</strong> general such close agreement would not be expected. Although the<br />

deviance tests of ma<strong>in</strong> effects are not necessary here, <strong>in</strong> general they are needed. For<br />

example, if a factor with more than two levels were fitted, us<strong>in</strong>g dummy variables (§11.7),<br />

a deviance test with the appropriate degrees of freedom would be required.<br />

The deviance associated with the model <strong>in</strong>clud<strong>in</strong>g all the ma<strong>in</strong> effects is 13 59 with 11<br />

DF, and this represents the 11 <strong>in</strong>teractions not <strong>in</strong>cluded <strong>in</strong> the model. Tak<strong>in</strong>g the deviance<br />

as a x2 …11† , there is no evidence that the <strong>in</strong>teractions are significant and the model with just<br />

ma<strong>in</strong> effects is a good fit. However, there is still scope for one of the two-factor <strong>in</strong>teractions<br />

to be significant and it is prudent to try <strong>in</strong>clud<strong>in</strong>g each of the six two-factor<br />

<strong>in</strong>teractions <strong>in</strong> turn to the model. As an example, when the <strong>in</strong>teraction of the two k<strong>in</strong>ds<br />

of read<strong>in</strong>g, AC, is <strong>in</strong>cluded, the deviance reduces to 10 72 with 10 DF. Thus, this <strong>in</strong>teraction<br />

has an approximate x2 …1† of 2 87, which is not significant (P ˆ 0 091). Similarly,<br />

none of the other <strong>in</strong>teractions is significant.<br />

The adequacy of the fit can be visualized by compar<strong>in</strong>g the observed and fitted<br />

proportions over the 16 cells. The fitted proportions are shown <strong>in</strong> column (4) of Table<br />

14.1 and seem <strong>in</strong> reasonable agreement with the observed values <strong>in</strong> column (3). A formal<br />

test may be constructed by calculat<strong>in</strong>g the expected frequencies, E(r) and E…n r†, for<br />

each factor comb<strong>in</strong>ation and calculat<strong>in</strong>g the Pearson's x2 statistic (8.28). This has the<br />

value 13 61 with 11 DF (16 5, s<strong>in</strong>ce five parameters have been fitted). This test statistic is<br />

very similar to the deviance <strong>in</strong> this example, and the model with just the ma<strong>in</strong> effects is<br />

evidently a good fit.<br />

The data of Example 13.2 could be analysed us<strong>in</strong>g logistic regression. In this<br />

case the observed proportions are each based on one observation only. As a<br />

model we could suppose that the logit of the population probability of survival,<br />

Y, was related to haemoglob<strong>in</strong>, x1, and bilirub<strong>in</strong>, x2 by the l<strong>in</strong>ear logistic<br />

regression formula (14.6)<br />

Y ˆ b 0 ‡ b 1x 1 ‡ b 2x 2:<br />

Application of the maximum likelihood method gave the follow<strong>in</strong>g estimates of<br />

b 0, b 1, and b 2 with their standard errors:<br />

^b 0 ˆ 2 354 2 416<br />

^b 1 ˆ 0 5324 0 1487<br />

^b 2 ˆ 0 4892 0 3448:<br />

…14:7†<br />

The picture is similar to that presented by the discrim<strong>in</strong>ant analysis of Example<br />

13.2. Haemoglob<strong>in</strong> is an important predictor; bilirub<strong>in</strong> is not. An <strong>in</strong>terest<strong>in</strong>g<br />

po<strong>in</strong>t is that, if the distributions of the xs are multivariate normal, with the same<br />

variances and covariances for both successes and failures (the basic model for<br />

discrim<strong>in</strong>ant analysis), the discrim<strong>in</strong>ant function (13.4) can also be used to<br />

predict Y. The formula is:

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

Saved successfully!

Ooh no, something went wrong!