14.11.2012 Views

Faecal occult blood testing for population health screening May 2004

Faecal occult blood testing for population health screening May 2004

Faecal occult blood testing for population health screening May 2004

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.

3. The positive predictive value (PPV) <strong>for</strong> each FOBT was also determined; these data<br />

are reported in Appendix F, <strong>for</strong> completeness.<br />

The PPV is the probability that if a person tests positive that they actually have the<br />

condition:<br />

True positive/(True positive + False positive).<br />

Meta-analyses<br />

This project required separate meta-analyses of various outcomes measuring the<br />

efficacy of the test being examined. The analyses were also carried out <strong>for</strong> various<br />

<strong>population</strong>s.<br />

Typically, each set of trial results consisted of two or more observed proportions, one <strong>for</strong><br />

each test used. For each trial, the tests were usually repeated on the same set of patients.<br />

These circumstances, in general, preclude the use of classical meta-analyses methods <strong>for</strong><br />

summary data, as these methods were developed <strong>for</strong> two-arm trials with independent sets<br />

of subjects in each arm.<br />

In addition, there was usually a small number of trials involved in each meta-analysis.<br />

Classical random-effects meta-analysis (Der Simonian and Laird 1986) implicitly assumes<br />

that a between-studies variance estimate used in the calculation of the weights <strong>for</strong> the<br />

method is known with precision, and this is not likely to be true <strong>for</strong> a small number of<br />

trials.<br />

A further source of variability may be that diagnostic test results are sensitive to the<br />

threshold set <strong>for</strong> the test, which may vary between studies (see Lijmer et al 2002 <strong>for</strong><br />

details).<br />

The circumstances outlined above suggest that a more sophisticated random-effects<br />

analysis is required, and a Bayesian approach was deemed to provide a more appropriate,<br />

conservative approach (Higgins and Whitehead 1996; Smith et al 1995). The analyses<br />

were carried out using Markov Chain Monte Carlo (MCMC) simulation as implemented<br />

in the statistical package WinBUGS (version 1.3, Spiegelhalter et al 2000).<br />

The method is based on fitting logistic regression models incorporating random effects.<br />

A random intercept model of the <strong>for</strong>m was adopted:<br />

logit(p ij) = α + b i + β j × x ij,<br />

where,<br />

• p ij = prob(Y ij = 1) is the probability that the event occurs<br />

• α is an intercept, interpreted as the log-odds <strong>for</strong> the event occurring in a<br />

reference category, in this case the guaiac test group<br />

• b i is a random effect <strong>for</strong> the i th study, which is common across all categories of<br />

interest<br />

24 <strong>Faecal</strong> <strong>occult</strong> <strong>blood</strong> <strong>testing</strong>

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

Saved successfully!

Ooh no, something went wrong!