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.

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

Treatment period <strong>in</strong>teraction, and carry-over<br />

The analysis illustrated <strong>in</strong> Example 18.4 assumes that the response is affected<br />

additively by treatment and period effects, i.e. that there is no treatment period<br />

(TP) <strong>in</strong>teraction. The terms g AB and g BA <strong>in</strong> (18.4) represent such an <strong>in</strong>teraction.<br />

Note that it is only their difference that matters: if these two terms were equal<br />

they could be absorbed <strong>in</strong>to the period parameter p2 and we should be back with<br />

the no-<strong>in</strong>teraction model. There are various possible reasons for an <strong>in</strong>teraction:<br />

1 If the wash-out period is too short, the response <strong>in</strong> period 2 may be affected<br />

by the treatment received <strong>in</strong> period 1 as well as that <strong>in</strong> period 2.<br />

2 Even if the wash-out period is sufficiently long for a period 1 drug to be<br />

elim<strong>in</strong>ated, its psychological or physiological effect might persist <strong>in</strong>to period<br />

2.<br />

3 If there is a strong period effect, chang<strong>in</strong>g the general level of response from<br />

period 1 to period 2, the treatment effect might be changed merely because it<br />

varies with different portions of the scale of measurement.<br />

A test for the existence of a TP <strong>in</strong>teraction may be obta<strong>in</strong>ed as follows.<br />

Denote the sum of the two responses for subject j <strong>in</strong> group i by<br />

sij ˆ yij1 ‡ yij2:<br />

If there is no TP <strong>in</strong>teraction, the expectation of sij for this subject is, from (18.3)<br />

and (18.4), 2m ‡ jij ‡…tA ‡ tB†‡…p1 ‡ p2†. S<strong>in</strong>ce the subjects were assigned<br />

randomly, the jij can be taken to be randomly distributed with the same mean<br />

(which can be taken as zero, s<strong>in</strong>ce the overall mean is accounted for by the term<br />

m <strong>in</strong> (18.4)), and a variance s2 0 , say. The expectations of s1j and s2j are thus equal.<br />

If, on the other hand, there is a TP <strong>in</strong>teraction, the expectations will differ: if<br />

gAB > gBA, for <strong>in</strong>stance, as might happen if the two treatments had different<br />

carry-over effects, the expectation of s1j will exceed that of s2j. We can therefore<br />

test the difference between the two mean values of sij, s1 and s2, aga<strong>in</strong> us<strong>in</strong>g a<br />

two-sample t test. Here, however, the with<strong>in</strong>-group variance estimate s2 s is that of<br />

the sij, not (as before) the dij. S<strong>in</strong>ce the sij vary <strong>in</strong> part because of the variability<br />

between subjects, we should expect s2 s > s2 d .<br />

Example 18.4, cont<strong>in</strong>ued<br />

The values of sij are given <strong>in</strong> Table 18.5. The two means are s1 ˆ 13 412, s2 ˆ 16 583. The<br />

pooled variance estimate is s2 s ˆ 41 890 (substantially larger than s2 d , as expected) and<br />

SE…s1 s2† ˆ …41 890† 1<br />

r<br />

1<br />

‡<br />

17 12<br />

ˆ 2 440:<br />

Thus, t ˆ…13 412 16 583†=2 440 ˆ 1 30 on 27DF (P ˆ 0 20). There is no clear<br />

evidence of a TP <strong>in</strong>teraction, and the earlier analysis therefore seems reasonable.

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

Saved successfully!

Ooh no, something went wrong!