Hello all, I was hoping someone could give me guidance as to the correct course of action, when using GPA followed by PCA to describe difference in form between 2 subgroups. Any assistance would be greatly appreciated!
I have a dataset of 200 surface 3-D bone models, from which I have taken seventeen representative 3-D landmarks. The dataset contains Asian/ Caucasian, Female/ Male, and healthy/ osteo-arthritic specimens. I wish to describe the differences in form caused by these factors. I have decided to analyse them in the following way. I have registered all 200 specimens simultaneously, using GPA. I have outputted the corresponding PC scores for PC1-PC6 (90% of variance) for each specimen. I have calculated the average PC scores for each subgroup described above, for PC1-PC6. I have used ANOVA on the PC-scores to identify which PCs (if any) distinguish significantly between subgroups. Alternatively, I could register all (say) Caucasians with one GPA, then register all Asians with a separate GPA, arriving at the mean configuration for each group. Procrustes Analysis, followed by PCA could then be applied to these 2 mean forms, in order to describe the modes of variation between them. ANOVA obviously could not be carried out on the output in this instance, given PC scores for just 2 specimens. Any difference described by PCs could not be verified statistically, using this method, I dont think? Which of these 2 approaches would you recommend for describing form difference, in terms of its PCs? Additionally, in an effort to generate a surface 3-D model representing each subgroup for further analysis (eg. average Asian and average Caucasian), the surface model of the most normal bone was identified (ie. the specimen whose configuration is closest to the GPA mean configuration/ specimen whose PC scores are closest to zero), then this model was warped to the average PC1-PC6 scores for each subgroup. I am able to write out these average surface models for each subgroup, along with their corresponding mean landmark configuration. For this analysis, I have decided that, for instance, the Asian and Caucasian average models should be generated by using the same base model (be it Asian or Caucasian) and warping it to the respective PC means of each ethnic subgroup. The alternative option is to model the average Asian using a close-to-normal Asian base model and to model the average Caucasian using a close-to-normal Caucasian base model. Although each mean model thus generated would be true to its origins as it were, I fear that to do it this way would be to introduce the difference between the 2 base models, additional to the essential difference in ethnic subgroups, as described by the difference in PC-score averages. Which of these 2 approaches would you recommend? Many many thanks! Niall Rooney -- Replies will be sent to the list. For more information visit http://www.morphometrics.org
