[snip, previous]
Post by c***@gmail.comHello, Dr. Ulrich.
First, the reason that I said "not random samples" is that I am not
likely to use parametric statistical methods.
We seem to have a disjunction or discordance of
vocabulary -- or something -- that is rather extreme.
There are random samples, there are observational samples,
there are selective samples, and there are odd observations.
ANY of those might be approached with "parametric statistical
methods," with assumptions that vary according to the case, if
there are any statistical methods to be used at all.
The usual contrast to "parametric" is "non-parametric" -- which
most often implies rank-transformations, followed by tests
that borrow the "parametric" methods for large Ns.
My preliminary guess is that you are fairly new to
statistics, and your use of terms does not follow anyone's
conventions. But I'm willing to consider otherwise.
What you (below) seem to be interested in,
"re-randomization" methods, are typically parametric,
though not the standard set. Perhaps you were seeking
the jargon-term, "robust"?
Post by c***@gmail.comSecond, there are two kinds of randomness: random selection (sampling)
and random assignment. Conventional parametric statistics are based
on the notion of random sampling.
- Yes, randomness is the philosophical underpinning. Most
people do not have trouble extending the statistics to
observational studies, with proper cautions about limits of inference.
Post by c***@gmail.comHowever, when we conduct a
randomized experiment based on random assignment, we usually do not
have a
population for statistical generalization but will get a reference
after the experiment. The reference is made by my samples.
Huh? Nonsense? If we don't have a population for generalization,
what do we have?
Post by c***@gmail.comThird, Fisher's permutation test is also based on a notion of random
sampling; so my samples which are not based on random selection cannot
be
analyzed with parametric statistical methods.
Fisher's permutation is often discussed with "nonparametric"
tests, but it does not fit well under the heading.
Post by c***@gmail.comIn sum, my samples are not selected at random,
? No? It sounded like they were.
Post by c***@gmail.comso I am not likely to
use conventional parametric statistical methods, which are based on
random sampling.
If they are not selected at *random*, you will be best
served, I suspect, by using conservative, conventional
methods to try to account for the non-randomness.
If you can't account for the non-randomness through
*parameters*, you are probably stuck with anecdotes.
If your problem is an awkward variance structure, based
on *dependencies* in the design, then you might want
something non-conventional. (SAS proc ph has 'sandwich
estimators' that are bootstrapped. I don't know if SPSS
offers the same.)
Post by c***@gmail.comThus, I would like to do a re-sampling method,
especially a randomization test.
However, I cannot find out an appropriate function for the test in
SPSS.
There are examples available for constructing bootstraps,
etc., (Google groups, < group:comp.soft-sys.stat bootstrap >
but these are not necessary or useful for most procedures,
for most data, in SPSS.
--
Rich Ulrich, ***@pitt.edu
http://www.pitt.edu/~wpilib/index.html