-------- Original Message -------- Subject: Size effect in PCA Date: Tue, 17 May 2011 11:18:34 -0400 From: [email protected] To: [email protected] Hi, I am doing a PCA analysis on a network of 26 truss measurements in my fish species, capelin, and I am having some difficulties understanding my results. These measurements have been previously used to determine stock structure in this species (Roby et al. 1991). I have 3 collection sites (n~100 individuals/sex/site) and I performed separate analyses for each sex due to the dimorphism in mature adults (used in this study). The data are normally distributed within each group and length is not correlated with the 26 truss measurements for either sex. As such, I did not perform a size correction (i.e. Meng and Stocker 1984, Lleonart et al. 2000) on the dataset. There was a small range of lengths in the samples so log transformation was not necessary. Length did not correlate with the PC1 scores so I assume that there is no "size variation" factor in the analyses. The resulting PCAs (based on a correlation matrix) return only one PC with an eigenvalue >1 (14.06 and 13.54 for males and females, respectively) but all variables loadings on PC1 are similar in sign and magnitude (0.19-0.25 in females, 0.17-0.24 in males). Based on my reading, this would indicate a size component. So, I am inclined to interpret these results as: the first component is size, and all significant subsequent components are shape. Therefore, there are no shape differences in males and females. But, if there are not length correlations, why is PC1 a size vector? If anyone could provide any insight into my problem, it would be greatly appreciated!! Thank you!
