Posting a query here reminds me of calling the Car Guys on National Public
Radio for help. I doubt that the world is ready for a radio call in program
for morphometric problems, but... some of us would listen!

I have been doing some old-fashioned morphometrics using PCAs and
experimenting with a way to weight my characters by a coefficient
representing "within-plant consistency". The effect of this would hopefully
be to get rid of some non-genetic noise/variation. I was evaluating this
weighting scheme by whether the between-plant within-population distances
from PRIN1, PRIN2 coordinates were smaller in the weighted analysis than in
an unweighted analysis. Well, as it turns out all pairs of specimens, both
within and between populations, have smaller between specimen distances in
the weighted set, so I transformed each set of between specimen distances
into Z-scores and found that there was no significant differences between
the unweighted and weighted analyses.

BUT (and here is the real issue) the EigenValue percentages on the two
analyses are radically different. I am using JMP statistical software. The
unweighted analysis uses the raw data scores and is a PCA using the
correlation matrix. The weighted analysis uses the data as scaled Z-scores
and does an "uncentered/unscaled" PCA. The weighted analysis EigenValue
percentage for PRIN1 is 84% and reaches 90% by PRIN3. The unweighted
EigenValue percentage for PRIN1 is 24% and on PRIN9 it is only 81%. How
should I interpret this difference in EigenValue percentages?

A bit of additional data: I am using 19 linear distance characters and have
154 specimens.

Your advice will be most appreciated,
Patricia Barlow-Irick
______________________________________
Patricia Barlow-Irick
Largo Canyon School
Route 80 Box 5
Counselor, NM 87018-9701
(505)568-9131 voice
(505)568-9412 fax
http://www.largocanyon.org