Yes, the PC axes are “comparable”. I think the best way to think about what a PCA does is to interpret it as a projection of a multidimensional space down to a low dimensional space that captures as much of the overall variation as possible. The first axis is somewhat special because it represents the best 1-dimensional space. Past that one should think of 1 and 2 giving the best 2-dimensional space and 1, 2, and 3 giving the best 3-dimensional space, etc. The axes themselves are not of a priori interest in an application – it is the space that is of interest. A consequence is that plots showing projections of points relative to PC1, PC2,etc. must be plotted to the same scale (i.e., consistent with the fact that the eigenvalues give the variances along each axis). If, as unfortunately often the case, the axes are plotted using different scales then the space has been distorted and is no longer the space that best accounts for the overall variation in the data. That also distorts the impressions one gets in looking at the plot as using different scales changes the relative distances between points.
Within that reduced space one may find that particular axes can seem to be interpretable but one should really look at the space and decide which directions within the space are most interesting based on the patterns of the data. That is, the data need to suggest interesting direction unless one has some a priori groups one wishes to compares. Often the first PC is of special interest but that is often due to allometry and the relatively large impact of size variation. That is, by now, a rather boring result! The individual PC axes are defined based on convenient mathematical properties – not based on any biological models so each one should not be considered separately as things of special interest. The above also means that one need not just visualize variation along each axis separately. One can, as in tpsRelw software, visualize any specified point within the PC space or in any direction of interest within the PC space. _ _ _ _ _ _ _ _ _ F. James Rohlf, Distinguished Prof. Emeritus Depts. of Anthropology and of Ecology & Evolution From: Yinan Hu <yinanhu...@gmail.com> Sent: Friday, May 18, 2018 2:19 PM To: MORPHMET <morphmet@morphometrics.org> Subject: Re: [MORPHMET] How to project shape difference onto different PC Dear James, Thanks for the reply. Yes I have completed a PCA on a GM dataset with 11 landmarks, and you got it exactly right that I'm trying to decompose shape differences onto individual PCs. The reason I was hesitating to do the vector projection is that I'm not sure if PC scores on different PCs are directly comparable to each other. For simplicity, let's say I'm only considering PC1 and PC2, which explains 80% of shape variation in total (60% + 20%). Group A has a mean PC1 score of 0.5, and PC2 score of 0.1; where as Group B has a mean PC1 score of 0.4 and PC2 score of 0.3. Then I'm looking at a 0.1 difference along PC1 and a 0.2 difference along PC2 between these two groups. Would this mean they differ twice as much along PC2 than PC1, such that in the 80% of shape variation explained by these two PCs, 1/3 is along PC1 and 2/3 is along PC2? But considering that PC1 explains three times more variation than PC2 (60% vs 20%), would this mean I should weigh the PC score difference (distance along each PC)? i.e. although the absolute difference in mean PC1 score is 0.1, it should be weighed three times more than the difference along PC2 so in the 80% of shape variation explained by these two PCs, 3/5 is along PC1 and 2/5 is along PC2? On the other hand, I agree visualizing the shape difference along each PC can be helpful, and I'm pretty sure the plotRefToTarget function from the R package geomorph can achieve this. Thanks again. Best, Yinan On Friday, May 18, 2018 at 12:53:32 PM UTC-4, K. James Soda wrote: Dear Dr. Hu, Let me begin by restating how I understand the question: You have completed a PCA on a morphological data set in which there are two subsets of interest. Now you would like to decompose the difference between the two subsets into differences along individual PCs. Here is my two cents on the issue: I would say that the literal solution to this problem would probably be something along the lines of what you proposed. For simplicity, say that you summarized each subset using its mean position in the PC space. This would be expressed as a vector where each element is a position along a single PC. The difference between these two vectors would then be a decomposition of how far you would need to move along each PC axis to move from one mean to the other. You could then standardize the elements so that their absolute values sum to one. This would be an expression of what percentage of the distance is along each PC. What I perceive as the subtext of your question, though, is whether this sort of decomposition has a reasonable interpretation, and the answer to this question is somewhat trickier. Assuming this is a GM data set, the more relevant point might be how you convert the difference into visualizations. A nice feature of GM data is that each PC will correspond to a "type" of deformation. This feature can be used to decompose the difference between two shapes in a shape-PC space as well. For example, imagine you moved from one mean shape in the PC space to the other by only moving parallel to PC axes. If you are interested in two PCs, this could be accomplished in two ways. You could then visualize the shape at the points where you make a turn; that is, you would visualize how mean shape 1 would need to be deformed to have the same PC1 or PC2 score as mean shape 2 if all other PCs were held constant. The degree of deformation would then provide a qualitative measure of how radical each PC's contribution is to the shape difference. Of course, this is not a quantitative measure, as you requested, but I would argue it is a more helpful assessment b/c it directly corresponds to observable phenomena. How helpful, though, will depend on your research question. Hope something in there helps a little, James On Thu, May 17, 2018 at 10:15 AM, Yinan Hu <yinan...@gmail.com <javascript:> > wrote: Dear colleagues, I'm trying to figure out how to break down shape differences onto individual PC axes. I have a morphospace where PC1 explains 60% of shape variation and PC2 explains 20% of variation. Two subsets of samples of particular interest do not differ much along PC1, but differs significantly along PC2. How should I project the shape difference between these subsets onto seperate PC axes, such that I can quantitatively show X% of shape difference between them are along PC1 and Y% is along PC2? A simple vector projection (i.e. using the mean difference of PC1 score and PC2 score) doesn't feel right to me as I don't think PC scores are directly comparable between different PCs. Or am I wrong? Any suggestions would be greatly appreciated. Many thanks for your time. -- MORPHMET may be accessed via its webpage at http://www.morphometrics.org --- You received this message because you are subscribed to the Google Groups "MORPHMET" group. To unsubscribe from this group and stop receiving emails from it, send an email to morphmet+u...@morphometrics.org <javascript:> . -- MORPHMET may be accessed via its webpage at http://www.morphometrics.org --- You received this message because you are subscribed to the Google Groups "MORPHMET" group. To unsubscribe from this group and stop receiving emails from it, send an email to morphmet+unsubscr...@morphometrics.org <mailto:morphmet+unsubscr...@morphometrics.org> . -- MORPHMET may be accessed via its webpage at http://www.morphometrics.org --- You received this message because you are subscribed to the Google Groups "MORPHMET" group. To unsubscribe from this group and stop receiving emails from it, send an email to morphmet+unsubscr...@morphometrics.org.
