Jean-Louis Tu <[log in to unmask]>:
>Do you think that, in some situations, these assumptions are obviously
>met, like when the random variable only takes two values?
>
>Example: take the 3-year survival rate in a double-blind study, some
>patients being administered drug A and others drug B. The random
>variable takes 2 values, 0 if the patient dies, 1 if the patient
>survives. If calculations are done carefully, P-values should be
>correct?
Tom:
Yes, in some cases the assumptions are easily met, because:
1) the assumptions are very general, e.g. what are called "non-
parametric" procedures,
2) the assumptions are very limited or easy to meet - the example you give.
However, if one is talking about a Bayesian procedure that assumes
the prior distribution is, say, the Cauchy distribution, then that
will be very difficult to test. Similarly, assuming something has
a normal distribution is hard to test because it is unimodal and
symmetric - such distributions are hard to test for, without huge amounts
of raw data.
Note also that the example you give above would probably not be analyzed
as a purely dead/not dead situation; instead a "survival curve" would be
fitted - a more complicated analysis that tries to fit a distribution
to the observed mortality data. Some kinds of survival analysis
assume the form/type of the distribution, others do not.
P.S. While on the topic, let me make one clarifying note on my post: the
multiple test procedures that control overall P-value are specialized
and limited in scope. The Bonferroni correction will work for any
tests, but it requires that one make individual tests at very low P-values, so
little is "significant". That is, the solutions to the problem of
controlling overall P-values in multiple tests, are limited and not completely
satisfying. That is a point I should have made clear, but did not,
in my post.
Tom Billings
|