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.

quite acceptable as be<strong>in</strong>g with<strong>in</strong> the tolerated range of equivalence, but also<br />

because a non-significant result gives no <strong>in</strong>dication as to whether or not the true<br />

difference lies with<strong>in</strong> the tolerated <strong>in</strong>terval.<br />

Suppose the parameter u measures the difference <strong>in</strong> efficacy between T and S,<br />

with high values favour<strong>in</strong>g T. In a two-sided equivalence trial, the limits of<br />

equivalence may be denoted by uL < 0 and uU > 0. The <strong>in</strong>vestigator will be <strong>in</strong><br />

a position to claim equivalence if it can be asserted that u lies with<strong>in</strong> the range<br />

(uL, uU). A simple approach is to assert equivalence if an appropriate confidence<br />

<strong>in</strong>terval for u lies wholly with<strong>in</strong> this range.<br />

Suppose that, for this purpose, we use 100(1 2a)% confidence limits; e.g.<br />

95% limits would have a ˆ 0 025. An equivalent approach would be to carry out<br />

significance tests us<strong>in</strong>g each of the equivalence limits for the null hypothesis:<br />

1 Test the null hypothesis HL that u ˆ uL aga<strong>in</strong>st the one-sided alternative that<br />

u > uL, with one-sided significance level a.<br />

2 Test the null hypothesis HU that u ˆ uU aga<strong>in</strong>st the one-sided alternative that<br />

u < uU, with one-sided significance level a.<br />

Equivalence is asserted if and only if both these null hypotheses are rejected. For<br />

a non-<strong>in</strong>feriority trial, only test 1 is necessary, and this corresponds to the<br />

requirement that a one-sided 100(1 a)% confidence <strong>in</strong>terval, extend<strong>in</strong>g to<br />

plus <strong>in</strong>f<strong>in</strong>ity, falls wholly to the right of uL.<br />

A Bayesian <strong>in</strong>terpretation<br />

18.9 Special designs 637<br />

A Bayesian formulation for equivalence might require that the posterior<br />

probabilities that u < uL and that u > uU are both less than some small value<br />

a. With the usual assumptions of a non-<strong>in</strong>formative prior and a normally<br />

distributed estimator, this is equivalent to the frequentist approach given<br />

above. For a non-<strong>in</strong>feriority trial, only the first of these two requirements is<br />

needed.<br />

A slightly less rigorous Bayesian requirement would be that the posterior<br />

probability that u lies with<strong>in</strong> the equivalence range …uL, uU† is 1 2a. This is<br />

slightly more lenient towards equivalence than the previous approach, s<strong>in</strong>ce an<br />

estimate near the upper limit of uU might give a posterior probability for<br />

exceed<strong>in</strong>g that bound which was rather greater than a, and yet equivalence<br />

would be conceded because the probability that u < uL was very small and the<br />

total probability outside the range was still less than 2a. The frequentist analogue<br />

of this approach is a proposal by Westlake (1979) that a 100(1 2a)%<br />

confidence <strong>in</strong>terval should be centred around the mid-po<strong>in</strong>t of the range, rather<br />

than about the parameter estimate. This would be equivalent to do<strong>in</strong>g separate<br />

tests of HL and HU at the one-sided a level when the estimate is near the centre of<br />

the range, but a test of the nearest limit at the one-sided 2a level when the<br />

estimate is near that limit.

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

Saved successfully!

Ooh no, something went wrong!