I have been working on a study trying to estimate body mass in mammals using a skeletal proxy in R. My dataset has high phylogenetic signal (Pagel's Lambda = 0.9) and I am interested in removing the effects of phylogenetic signal to produce more accurate estimates of body mass.
I had tried to do this using PGLS, but PGLS ended up producing higher error rates than the same data under OLS despite the high phylogenetic signal. Part of this appears to be because there are several particular clades with long branches (e.g., Monotremata) that form regression lines above or below the line of best fit, and their phylogenetic position distorts the resulting best-fit line. I had thought PGLS took phylogenetic covariance into account when estimating values, but it looks as though it only uses phylogenetic signal into account to minimize the residuals of the best fit line. It does not put the phylogenetic covariance back into the estimation of the y-value.
I.e., the function does not go: "Taxon X is positioned as sister to Monotremata. Therefore it should deviate above or below the regression line to a degree similar to its phylogenetic distance and the mass estimate should be adjusted accordingly?"
Given this, is there any way to use a topology to estimate a value that considers phylogenetic covariance and the phylogenetic postion of the unknown taxon when estimating this value, rather than just using it to minimize the residuals when calculating the best fit line?
Hi Russell
Do you have branch lengths in your tree? If so, you could calculate patristic distance and then run Phylogenetic Eigenvector Regression.
See this paper and references therein:
https://www.sbpbrasil.org/assets/uploads/files/rbp17-2/01_Diniz_Filho_et_al_pg_105a122_online.pdf
I note the following from the discussion of the above paper, when comparing to PGLS: "rather than assuming an evolutionary model, such as Brownian motion or Ornstein-Uhlenbeck, and adding the expected covariance among taxa into model residuals (so that results are conditional to this model, see Martins & Hansen, 1997), the PVR uses the phylogenetic eigenvectors to empirically model the trait variation"
No specific R functions that I know of but can be done in R by doing the separate steps of calculating patristic distance, then PCO decomposition into eigenvectors, then selection of eigenvectors, and then regression model with selected eigenvectors as explanatory variables.
David
Hi, first - make sure you use robust branch legths. Secondly - why not use proper mixed model to correctly separate phylogenetic and residual sources of variation (via any mixed model method like ASReml-R, MCMCglmm etc)? Also, I would argue that what you call "distortion" is actual biology - the behaviour you are seeing is expected, and it is the OLS estimator that has biased error estimates, not the reverse... If you're tree has very different branch lengths in different portions you may also try to relax the assumption of uniform rate of evolution across the tree, there is a number of methods able to estimate rate shifts in the tree.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R