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.

642 Cl<strong>in</strong>ical trials<br />

orig<strong>in</strong>al data. Pool<strong>in</strong>g the data at the outset would unnecessarily reduce<br />

precision if there were systematic differences <strong>in</strong> response between trials, and<br />

bias might be caused if some trials departed from equal assignment proportions.<br />

2 Although trials will never be exactly identical <strong>in</strong> treatments used, patient<br />

characteristics, outcome variables, etc., they should be broadly similar <strong>in</strong> all<br />

these respects. This is a vague recommendation, and differences of view as to<br />

the choice of trials for <strong>in</strong>clusion are <strong>in</strong>evitable.<br />

3 The trials should all <strong>in</strong>volve random assignment and adequately masked<br />

assessment. Protocol departures should be treated <strong>in</strong> the same way, preferably<br />

by ITT.<br />

4 All relevant and available data should be used.<br />

Precept 4 is clearly open to different <strong>in</strong>terpretations and practices. A widely<br />

held view is that efforts should be made to retrieve unpublished data, partly to<br />

avoid delays <strong>in</strong> publication, but more importantly to avoid publication bias<br />

(§18.8), discussed further below. Completeness of coverage may be relatively<br />

easy <strong>in</strong> some major areas of research, such as the acute treatment of myocardial<br />

<strong>in</strong>farction, where all current research is known to senior practitioners, but will be<br />

more difficult <strong>in</strong> less highly publicized areas. As noted <strong>in</strong> §18.8, a general system<br />

for registration of trials, at their outset, would be beneficial (Simes, 1986b;<br />

Dickers<strong>in</strong> et al., 1995).<br />

Publication bias may sometimes be detected by a funnel plot, <strong>in</strong> which<br />

an estimate of a parameter represent<strong>in</strong>g an efficacy difference for each trial is<br />

plotted aga<strong>in</strong>st the trial size (or some similar measure, such as the reciprocal of<br />

the variance or standard error of the estimate). The estimates should cluster<br />

round a constant value, with variability decreas<strong>in</strong>g with trial size (and hence<br />

produc<strong>in</strong>g the `funnel' shape). For small trial sizes, however, where non-significant<br />

results are more likely, publication bias may cause the absence of po<strong>in</strong>ts near<br />

zero effect, and hence a skewness <strong>in</strong> the shape of the funnel. Examples, ma<strong>in</strong>ly<br />

from epidemiology, are provided by Givens et al. (1997), who discuss methods of<br />

adjust<strong>in</strong>g for the bias. Other relevant work is by Dear and Begg (1992) and<br />

Hedges (1992).<br />

Many early meta-analyses worked merely with the summary statistics provided<br />

<strong>in</strong> the published reports of each trial. Current practice favours the use of<br />

<strong>in</strong>dividual patient data, enabl<strong>in</strong>g data check<strong>in</strong>g and standardization (Stewart &<br />

Parmar, 1993; Stewart & Clarke, 1995).<br />

Different meta-analysis <strong>in</strong>vestigators are likely to differ <strong>in</strong> their<br />

approach to some of the choices <strong>in</strong>dicated <strong>in</strong> 1±4 above. It is, therefore,<br />

not too surpris<strong>in</strong>g if different meta-analysis on ostensibly the same question<br />

occasionally reach different conclusions (Chalmers et al., 1987). An example is<br />

provided by two meta-analyses of trials compar<strong>in</strong>g the anti-caries efficacy<br />

of sodium fluoride and sodium monofluorophosphate dentifrices (Johnson,

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

Saved successfully!

Ooh no, something went wrong!