Re: Don's comments.
Much of what Don says is true, but I disagree with his portrayel of cluster analysis
as being used with only "quantitative" measures. Clustering can be done not only with
distance measures but with similarity measures. There are a number of coefficents
that can be used to measure similarity or dissimilarity of cases on a set of
dichotomous variables, and any categorical variable can be transformed to a set of
dichotomous dummy variables. SPSS has about 20 similarity measures for binary data
(some measures more sensible than others, IMO).
rick
--- "Donald F. Burrill" wrote:
On Wed, 31 May 2000 Claire <[EMAIL PROTECTED]> wrote:
> I am currently attempting to look at the effects of several nominal
> independent variables (all categorical
Well, yes: if they're ("only") nominal they must be categorical.
(What, by the way, is "Sun"? Time and day seem more or less
self-explanatory.)
> e.g Sun 1, 2 ,3 - Time 1,2,3 day 1,2,3) upon one nominal dependent
> variable (behavior coded eg. 55=upright sit).
Your dependent variable is also categorical, then.
> I was hoping to examine relationships with PCA or CA
Now I begin to get very puzzled. If you mean principal components
analysis (PCA) and cluster analysis (CA), both these techniques require
some sort of distance metric in the dependent variable(s), which cannot
exist with categorical variables, and you apparently don't have any
"quantitative" (not even ordinal) variables. (If you mean something
else by these acronyms, my comments will probably not be helpful.)
> and then to relate the strongest structures to environmental variables
> using CCA.
By which do you mean canonical correlation analysis?
Same objection as above -- can't imagine it working with categories.
> Unfortunately my data has a kurtosis value of > 5 (majority of
> frequencies at extremes) and therefore is not normally distributed.
Of course it isn't. Categories (as opposed to counts, or measured
variables of various kinds) cannot follow a normal distribution, in the
nature of things. What led you to think you wanted (or needed) normally
distributed data?
> Could anyone give me some advice regarding which ordination techniques
> are best for non-normal data.
This surely depends on the actual shape
of the data distribution(s). You can't really want to know "for non-
normal data" -- you want to know what to do with YOUR data, and so far
you haven't described it in enough detail (or clearly enough) for any
sensible suggestions to emerge.
> I have run spearmans rank correlations
Aqain, puzzling. All the variables you have mentioned are alleged to be
nominal (i.e., categorical), and to run rank correlations requires
imposing an order on the categories (or assuming that the codes assigned
to the categories actually reflect some underlying order). Else the
output is meaningless: a classic instance of GIGO.
> and there are definite trends there. I have also performed cluster
> analysis to group individual animals according to behavior and hoped to
> use MANOVA to evaluate how well differentiated clusters are and to use
> DFA to find which variables contribute most strongly to clustering.
ALL of these techniques require variables that are in some degree
"quantitative" -- that is, at least ordinal, interval for preference,
ratio would be an unexpected bonus -- and you don't seem to have any.
I conclude that either I have egregiously misread what you have told us
of your problem, or that you have not told us a LOT of pertinent
information about your problem and your variables, or you're generating
quantities of misleading and/or meaningless output.
> I know that these techniques generally rely on the assumption of
> normality
Actually, they don't. How could one possibly find
clusters, for example, in data that is multivariate normal?
> however how strict is this? There seems to be barely anything I can
> use to analyse my non-normal data, surely this cant be right???
Score one for your intuition, I think.
> I thought about transforming the data but as I am monitoring behaviours
> (105 in all) approximately half of which barely happen (e.g frequency 1
> or 2) and the rest which happen with great frequency (e.g frequency
> 560) I cannot see how this would help.
It begins to look as though your data may not actually be categories, but
rather frequencies with which the categories occur. This still leaves
ambiguous what you're actually doing with the frequencies -- I do not see
clearly what it was you were calculating kurtosis of, for example.
> Any advice would be greatly appreciated!
>
> Claire
Not sure I have any _advice_ -- what I seem to be offering mainly is
puzzlement. But perhaps even that will help, a little.
-- Don.
------------------------------------------------------------------------
Donald F. Burrill [EMAIL PROTECTED]
348 Hyde Hall, Plymouth State College, [EMAIL PROTECTED]
MSC #29, Plymouth, NH 03264 603-535-2597
184 Nashua Road, Bedford, NH 03110 603-471-7128
===========================================================================
This list is open to everyone. Occasionally, less thoughtful
people send inappropriate messages. Please DO NOT COMPLAIN TO
THE POSTMASTER about these messages because the postmaster has no
way of controlling them, and excessive complaints will result in
termination of the list.
For information about this list, including information about the
problem of inappropriate messages and information about how to
unsubscribe, please see the web page at
http://jse.stat.ncsu.edu/
===========================================================================
--- end of quote ---
===========================================================================
This list is open to everyone. Occasionally, less thoughtful
people send inappropriate messages. Please DO NOT COMPLAIN TO
THE POSTMASTER about these messages because the postmaster has no
way of controlling them, and excessive complaints will result in
termination of the list.
For information about this list, including information about the
problem of inappropriate messages and information about how to
unsubscribe, please see the web page at
http://jse.stat.ncsu.edu/
===========================================================================
