## small sample multivariate analyses

### small sample multivariate analyses

> I am doing a small sample (N=7) crossover design project for a client
> where the influence of two drug treatments on 5 continous dv's are being
> assessed. I normally do large sample work and would use a repeated
> measures MANOVA or GLM for this type of problem, with bonferroni multiple
> comparisons. I am
> concerned about 1) violations of assumptions 2) skewness of variables with
> this sample size and 3) power if I use MANOVA/GLM for this analysis. Most
> of my nonparametric texts are designed for a single dv.

a) for N=7, a univariate non-par test is barely reaching for a
maximum achievement of a 5% test, referring to rank-order statistics.
That does not leave much scope for multiple variables, or for
Bonferroni adjustment for multiple tests.
b) There *exist* hardly any multivariate non-par tests, even if you
had a huge N.

Quote:>                                                                                                I've looked
> through the published data in the area and multiple ANOVAs seem to be used
> a lot or several non-parametric tests on each dv. I would like to avoid
> inflating type 1 error by the latter choices. Is there an alternative to
> MANOVA or the set of several univariate analyses?

PICK  a single analysis.
ESTABLISH  one hypothesis, and one variable to measure the hypothesis.
Maybe you want to construct a composite variable to measure it.

If there were really huge effects, like the ones that will show up
with N=7, you really ought to have a pretty good idea what they are
from the uncontrolled observations that (ordinarily) precede a
designed study.

Of course, if you really don't know anything about the area you are
collecting data on,  then you have to do the multiple tests and
proclaim the whole exercise to be Exploratory.

--

http://www.pitt.edu/~wpilib/index.html

What is the best method for analysing a study in which 7 subjects are
exposed to 3 treatments and for which 5 continous dependent variables are
measured? The design is a crossover type for which my text recommends
several t-tests. Unfortunately the text procedures are for single
dependent variables. I would do a MANOVA if the sample size was larger but
due to the small N I am pretty sure the assumptions underlying MANOVA will
be violated. Plus power becomes a problem with this size N. I am not that
familiar with nonparametric stats but they seem to be directed at
experiments with one dependent variable. Can someone point me in the right
direction for a test where I can look at the 3 conditions effect of 5 dv's
with 7 subjects where I can have a hope of controlling type 1 error?

Thanks,
Tony