Dear Brett, this is just a fast opinion, and I am sure of that more experienced people in the list will be able to give you better advice. If I got it right, do not find discriminant analysis to be the best technique for your purposes. You will get as many groups as you enter into the analysis. Besides, discriminant analysis is quite exigent on respect to sample size as the number of variables and groups increase.
If what you are considering are other more exploratory techniques, such as PCA, to go afterwards on factor and cluster analysis, things will be easier. "Additive" (sensu "highly correlated") measurements will not be a problem, as the amount of variability explained by the variables already into the analysis is taken into account to enter each new one, solving the problem of colinearity. Sexual size differences shouldn't be either a problem, as the first component would be an allometric size variable (allometric size is the usual suspect to be the main source of variability), being size effects thus controlled in the remaining components. If you find sexual differences not to depend just on size, I think that sexes should be entered as different groups in any further analysis... well, they are there. This way, through the interpretation of the variables entering into each component, you would also be able to select the ones for a discriminant analysis, if you want to be able to predict group pertenence later on. Theoretically, you would not be losing any relevant information, as the variables not entering into the PCA would contain mostly redundant one. I wish this helps, and would really like to know from other more experienced members of the list if this reasoning is right. Cheers, Luis Cabo Forensic Anthropology Laboratory, Mercyhurst Archaeological Institute, Erie, PA ----- Mensaje Original ----- De: [EMAIL PROTECTED] Fecha: Lunes, Septiembre 22, 2003 3:48 pm Asunto: linearity of variables > Dear Morphometricians, > > I am preparing to do a discriminant analysis on a genus of sharks > to see > if I get species clustering. > The first consideration is that most sharks are sexually > dimorphic, and > extremely so for the genus I am investigating, so I am expecting > to see > more groupings than there are species. There is also a considerable > amount of ontogenetic change. > I have a considerable data set in terms of variables (about 100 > measurements for each individual), however for most species (which are > extremely rare) I have a sample size of only 4 individuals and for > a few > species n=20+. Quite a number of the measurements are additive, > howeverthe majority of the measurements are non-additive. > > First question: will the non-additive measurements 'swamp' the effects > of the additive measurements? > > Secondly, if the additive measurements are not 'swamped' by the > non-additive measurements, what is the best approach to removing > measurements from the analysis? > For example, if I have: > > total length (TL) = head length (HL) + body length (BL) + tail length > (tL) > > then is removal of TL sufficient, or do I need to remove HL or BL > or tL, > or do I need to remove TL plus one of the other measurements, or > can I > retain TL and just remove one of the additive measurements?? > > Thirdly, how much information am I losing in removing a measurement > besides TL, because any one of HL, BL or tL may be a major diagnostic > feature between species, between sexes, or even between ontogenetic > stages?? > > Any comments will be appreciated. > > Thanks, > > Brett > > *************************************************** > Mr Brett Human (PhD candidate) > Shark Research Center > South African Museum > PO Box 61 > Cape Town 8000 > South Africa > ph: 27 21 481 3856 > fax: 27 21 481 3993 > > or > > email: [EMAIL PROTECTED] > ph: 27 21 406 6430 > fax: 27 21 448 8350 > **************************************************** == Replies will be sent to list. For more information see http://life.bio.sunysb.edu/morph/morphmet.html.
