>One more point. This sample is really too small for serious analysis and
>interpretation. Most sources suggest that the chi-square statistic be
>applied only when there is a frequency of at least 10 in each cell. If the
>results of this analysis turn out to be interesting, your friend should
>collect a larger sample of data and attempt to replicate the findings.
Shouldn't that be _expected_ frequency? The general rule of thumb is 80% of
expected frequencies >5. Low expected frequencies can distort the results
because one observation in such a cell will overwhelm lots of observations
in other cells. Low observed frequencies are not a problem if the expected
frequencies are high.
I'd use the binomial test. Under equiprobability P < .001 in this case.
(This comes out significant even if not returning is a priori thought to be
3 times as likely as the returning for the final dose).
Thom
|