Well, that took longer than I'd hoped... Our multivariate mixed model approach to 3DGM is now available in early view. Using a global human sample, we treat population history and structure as a random effect in order to quantify the (fixed) effect of the transition to agriculture on skull shape and form.
http://www.pnas.org/content/early/2017/07/18/1702586114.full I hope you find it interesting. David On Fri, Jan 20, 2017 at 8:50 AM, David Katz <dck...@ucdavis.edu> wrote: > Elahe and Ari, > > If your dependent observations have more than a few dimensions, such as is > typical with landmark data or even a collections of linear measurements, > then I think the common covariance matrix recommendation from Dr. Rohlf is > the more standard approach. A typical strategy for removing group structure > is to perform your analysis on the pooled within group covariance matrix. > Chs.11 and 12 of Zelditch et al's *Geometric Morphometrics for Biologists* > provides an overview and citations, as well as the equation for > partitioning overall covariance into within and between group components. > > However, Dr. Belk is correct that, in principle, a mixed model suits your > needs. The mixed model provides a way to partition random deviances > attributable to population history and structure from common (fixed) > effects of developmental stage and sex. You will have to think about how > you want to model conditionality of sexual dimorphism with respect to > stage. For example, if dimorphism does not arise until later stages of > development in your subject species, it does not make much sense to model a > single dimorphism coefficient across the entire sample. > > To make the model go, you will need data with which to estimate pairwise > evolutionary relationships among the groups in your sample. The computation > of random effects depends upon it. Stone et al ( > http://rstb.royalsocietypublishing.org/content/366/1569/1410) describe > some of the challenges of estimating relatedness where there is gene flow > among the sample clusters (i.e., where your groups are populations rather > than non-reticulating species). A paper from me+colleagues, linked to > below, provides an overly simple but (we think) reasonable solution to this. > > This leaves one more potential problem: dimensionality. If the dependent > observation is univariate (centroid size, the first shape principal > component, etc.) or very low dimensional, the MCMCglmm package (see > http://onlinelibrary.wiley.com/doi/10.1111/j.1420-9101.2009.01915.x/full) > could suit your mixed model needs. However, if your data is truly > multivariate (GM or anything more than a handful of linear measurements), I > would recommend looking at the BSFG package. We used it to estimate sex and > climate effects on the human cranium, using a subset of populations from > the Howells linear craniometric collection. You can find that paper on my > ResearchGate page (ResearchGate profile > <https://www.researchgate.net/profile/David_Katz29>). You will also find > a poster where we extend the application to cranial and mandibular 3D > landmark data. In the next few months, we should have a paper out that > combines high-dimensional mixed model analysis with 3DGM in a more > satisfying way than was managed in the poster. > > Unfortunately, unless the developers have finished updating the BSFG > package, it is not quite plug and play. It takes some time to figure out. > You need familiarity with Bayesian analysis, and with Matlab. However, the > outputs are worth the effort. For example, with GM data, your fixed effect > posterior will estimate shape contrasts for sex and developmental stage > effects for the whole shape configuration, rather than on synthetic, > orthogonal subsets (PCs). > > Conclusion: if your data is multidimensional and you don't think you can > get going with BSFG, for the time being, the common covariance matrix > approach is probably your best option. > > David > > On Thu, Jan 19, 2017 at 5:24 PM, Mark Belk <mark_b...@byu.edu> wrote: > > What you describe – samples from multiple populations – is best considered > as a random effect in a typical generalized linear model format. You have > randomly sampled some populations from all of those that might be > available. If I understand your data correctly, to evaluate allometry, use > a mixed model approach where some trait measurement is the response > variable and some measure of body size would be the predictor variable, > then population would be included as a random effect in the model. This > structure has the advantage of accounting for and adjusting for covariation > among populations before the fixed effect is evaluated. Appropriately > crafted mixed models can rigorously account for a range of complicated > covariance structures within the context of one model. Several examples of > the use of mixed models in ecology and evolution can be found in the > literature. > > Hope that helps, > > > > Mark > > > > Mark C. Belk, Professor of Biology > > Brigham Young University > > Editor, *Western North American Naturalist* > > 801-422-4154 <(801)%20422-4154> > > > > *From:* Ariadne Schulz [mailto:ariadne.sch...@gmail.com] > *Sent:* Wednesday, January 18, 2017 1:27 PM > *To:* Elahe > *Cc:* MORPHMET > *Subject:* Re: [MORPHMET] eliminating the effect of population differences > > > > I would like to hear any responses to this as well. I did something > similar and I wasn't sure how to approach this question. In future studies > I would like to address precisely this issue. My inclination would be that > first you would want to determine how much morphological variation you're > getting between sites. You could then look at sexual dimorphism within each > site and/or you could look at variation of only females and only males over > all sites. But this is all rather clunky and does not eliminate any > interpopulation variation. If anyone has already proposed or can propose a > better methodology I'd be interested in it as well. > > Best, > > Ari > > > > On Wed, Jan 18, 2017 at 5:29 PM, Elahe <ellie.parv...@gmail.com> wrote: > > Dear all, > > > > I have pooled samples from 7 different populations of one species in order > to study the allometric growth and sexual dimorphism in that species. As > different populations may have subtle differences in terms of body > dimensions with each other, I want to remove their effects. > > Can anyone suggest a way to eliminate population effects and maybe finding > some residuals that are homogeneous and can be used for further analyses? > > > > I would appreciate any helps :) > > -- > 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. > > > > -- > 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. > > -- > 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. > > > > > -- > David C. Katz, Ph.D. > Evolutionary Anthropology > University of California, Davis > Young Hall 204 > ResearchGate profile <https://www.researchgate.net/profile/David_Katz29> > Personal webpage <https://davidckatz.wordpress.com/> > -- David C. Katz, Ph.D. Postdoctoral Fellow Benedikt Hallgrimsson Lab University of Calgary Research Associate Department of Anthropology University of California, Davis ResearchGate profile <https://www.researchgate.net/profile/David_Katz29> Personal webpage <https://davidckatz.wordpress.com/> -- 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.