Dear Sue,
The problem is this. Imagine two sets of data.
Set 1) 100 subjects each measured one day only, 50 assigned randomly
to treatment A 50 to treatment B.
45 of the A group have the symptom, 5 of the B group have it.
Set 2) 2 subjects each measured for 50 days 1 assigned randomly to
treatment A 1 to treatment B.
On 45 of the days A shows the symptom, on 5 days B shows the symptom.
If you analyse both sets of data using r/n and ignoring the repeated
measures aspects, the apparent conclusion is that treatment B reduces the
prevalence of the symptom with identical statistics in both cases.
In the case of set 2 however the observations could quite easily be the
result of one patient being sicker than the other irrespective of treatment.
So I would be wary of doing it. There are sophisticated procedures
around that let you do repeated measures for binary data and loglinear
models. Failing that, converting your r/n for each subject to a proportion
p, applying a transformation to make the distribution of p more normal
and then doing a conventional Anova would probably be safer.
Philip
|