The word "exact" in statistics is, I assert, always a misnomer.
Even when the result so labelled comes out of a complete combinatorial
enumeration there is usually a hidden independence assumption somewhere
that is often dubious given clustering, time or space structure or the
nature of sampling. And usually the argument is calculus-based, with yet
more assumptions behind it, or something like a bootstrap, with its own
different flavour of approximation.
Stata's implementation of Kolmogorov-Smirnov [NB spelling] is no
exception to that. At best, it uses an approximation.
A bigger point (in my experience) is that this kind of test is not
really useful in research. I can sense a genuine substantive and
scientific interest in whether males and females differ in some way. If
the K-S test doesn't confirm a difference, the main signal is that your
data are inadequate to detect it. If it does, the big question is to say
what that difference is. Why not short-circuit the process and look
directly at a quantile-quantile plot or cumulative distribution plot
(official -qqplot-, or -qplot- and -distplot- from SJ)?
Nick
[email protected]
[email protected]
So, are my conclusions exact? I do not know if I have to consider the
first and the second p-values too..
Garry
Why not use the exact p-value, as reported by
ksmirnov x, by(group) exact
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
