Unclear output from MCMCglmm with categorical predictors
Dear Jarrod and Ben, Thank you both for the advice. Apologies if I accidentally caused anyone to receive an unexpected email. Running the model specified above returns the error: "non-reduced nodes do not appear first". The error persists when using a fully-bifurcating tree so the issue is not tree polytomies. I tried to go through the source code line-by-line to locate the problem but got lost in the many loops and conditional statements. This error message is conditioned on a match() output that also returns a warning that the two arguments matched are unequal in length: one corresponds to the tips while the other includes the tree's internal nodes as well. As always, any help is much appreciated. Best, Roi
On Wed, 21 Nov 2018 at 15:05, HADFIELD Jarrod <j.hadfield at ed.ac.uk> wrote:
Hi, You could upload the tarball to winbuilder ( https://win-builder.r-project.org/) and build a Windows source package. Cheers, Jarrod On 21/11/2018 14:09, roee maor wrote: Hi Jarrod, Many thanks for your reply. I couldn't install the tarball on R v3.4.3 or v3.5.1 so sourced the files directly to the workspace. I tried to run this model as you suggested:
T1 <- MCMCglmm(Activity ~ -1 + log(Mass) + Max.Temp * Annual.Precip,
random = ~ animal,
prior = list(R = list(fix=1, V=1e-15), G = list(G1 =
list(V=1, nu=0.002))),
pedigree = datatree,
reduced = TRUE,
burnin = 50000, nitt = 750001, thin = 700,
family = "threshold",
data = Rdata,
pl = TRUE, saveX = TRUE, saveZ = TRUE,
verbose = TRUE)
It returns some errors about missing functions "is.positive.definite" and
"Matrix", which I addressed with:
library("corpcor", lib.loc="~/R/win-library/3.5")
library("MatrixModels", lib.loc="~/R/win-library/3.5")
but I can't figure this one out:
'Error in .C("MCMCglmm", as.double(data$MCMC_y),
as.double(data$MCMC_y.additional), :
C symbol name "MCMCglmm" not in load table'
Detaching these packages doesn't necessarily cause that same error to
appear although I execute the exact same code.
Also, several attempts (same code again) caused a fatal error and
automatic session termination (info for a similar session below if
interesting).
I tried to use the fully bifurcating tree as an experiment but that made
no difference.
Any ideas what this last error means?
Thanks!
Roi
sessionInfo()
R version 3.5.1 (2018-07-02) Platform: x86_64-w64-mingw32/x64 (64-bit) Running under: Windows >= 8 x64 (build 9200) Matrix products: default locale: [1] LC_COLLATE=English_United Kingdom.1252 LC_CTYPE=English_United Kingdom.1252 LC_MONETARY=English_United Kingdom.1252 [4] LC_NUMERIC=C LC_TIME=English_United Kingdom.1252 attached base packages: [1] stats graphics grDevices utils datasets methods base other attached packages: [1] corpcor_1.6.9 Matrix_1.2-14 phytools_0.6-60 maps_3.3.0 ape_5.2 loaded via a namespace (and not attached): [1] igraph_1.2.2 Rcpp_0.12.19 magrittr_1.5 MASS_7.3-50 mnormt_1.5-5 [6] scatterplot3d_0.3-41 lattice_0.20-35 quadprog_1.5-5 fastmatch_1.1-0 tools_3.5.1 [11] parallel_3.5.1 grid_3.5.1 nlme_3.1-137 clusterGeneration_1.3.4 phangorn_2.4.0 [16] plotrix_3.7-4 coda_0.19-2 yaml_2.2.0 numDeriv_2016.8-1 animation_2.5 [21] compiler_3.5.1 combinat_0.0-8 expm_0.999-3 pkgconfig_2.0.2 On Tue, 20 Nov 2018 at 20:01, HADFIELD Jarrod <j.hadfield at ed.ac.uk> wrote:
Hi, Most likely the phylogenetic heritability (the phylogenetic variance / the phylogenetic +residual variance) is approaching one resulting in numerical difficulties. Probably the best thing is to assume that the phylogenetic heritability equals 1 and use the reduced phylogenetic mixed model implementation. This allows the phylogenetic heritability to be equal to 1 without causing numerical issues. At some point I will integrate these models into the main MCMCglmm package, but for now you can download it from here: http://jarrod.bio.ed.ac.uk/MCMCglmmRAM_2.24.tar.gz. Change the name of the ??Binomial? column to ?animal? and fit: T1 <- MCMCglmm(Activity ~ -1 + log(Mass) + Max.Temp * Annual.Precip, random = ~ animal prior = list(R = list(fix=1, V=1e-15), G = list(G1 = list(V=1, nu=0.002))), pedigree = tree, reduced=TRUE, burnin = 150000, nitt = 2650001, thin = 2500, family = "threshold", data = Tdata, pl = TRUE, pr = TRUE, saveX = TRUE, saveZ = TRUE, verbose = FALSE) You should need fewer iterations. Cheers, Jarrod On 20 Nov 2018, at 18:08, roee maor <roeemaor at gmail.com> wrote: Dear Jarrod (and list), Following your previous comment I added "random = ~ Binomial" to my model to allow for a phylogenetic analysis. This causes convergence problems: the trace plots show increasing oscillations along each chain (although no directional trends, so it's not a burn-in issue). Also, the posterior samples are highly correlated, residual variance estimates are >10^3 and threshold estimates are high (>20 on the latent scale). Surprisingly (to me), predictors that are strongly significant in the non-phylogenetic model lose their effect in the phylogenetic model (I tried several alternative parameter configurations). It seems that this model attributes the explained variance to phylogeny alone. Can anyone explain what is going on here? Am I specifying the model poorly or just asking my data more than it can answer? I tried to overcome this issue by using a fully resolved variant of the phylogeny, which only improved things slightly. I also changed the random effect to "random=~Family" or "random=~Order", which reduced the variance and threshold estimates to more acceptable levels (<10), but still no significant predictors (and I'm not sure how the algorithm calculates covariance between higher taxa in the phylogeny). Separately I tried parameter expanded prior: "prior = list(R = list(V=1, fix=1), G = list(G1 = list(V=1, nu=1, alpha.mu=0, alpha.V=1000)))". That didn't help, and messing with priors for this reason feels like poor practice. This is the model: T1 <- MCMCglmm(Activity ~ -1 + log(Mass) + Max.Temp * Annual.Precip, random = ~ Binomial, prior = list(R = list(fix=1, V=1), G = list(G1 = list(V=1, nu=0.002))), ginverse = list(Binomial=INphylo$Ainv), burnin = 150000, nitt = 2650001, thin = 2500, family = "threshold", data = Tdata, pl = TRUE, pr = TRUE, saveX = TRUE, saveZ = TRUE, verbose = FALSE) The data I use looks like this (not all variables appear in each model): str(Tdata) 'data.frame': 1389 obs. of 10 variables: $ Binomial : Factor w/ 1421 levels "Abrocoma_bennettii",..: 1 2 3 4 5 6 7 8 9 10 ... $ Order : Factor w/ 27 levels "Afrosoricida",..: 24 24 24 24 24 3 24 24 2 2 ... $ Family : Factor w/ 126 levels "Abrocomidae",..: 1 26 26 26 26 46 74 87 10 10 ... $ Activity : Factor w/ 3 levels "1","2","3": 1 3 2 2 2 3 2 3 1 3 ... $ Habitat : Factor w/ 6 levels "Aqua","Arbo",..: 5 5 5 5 5 5 5 5 5 5 ... $ Diet : Factor w/ 3 levels "Faun","Herb",..: 2 3 3 3 3 1 3 2 2 2 ... $ Mass : num 250.5 24.9 34.5 38.9 24.5 ... $ Max.Temp : num 22 16.6 19.1 19.8 17.2 ... $ Annual.Precip : num 166 645 558 903 1665 ... Any advice would be much appreciated! Many thanks, -- Roi Maor PhD candidate School of Zoology, Tel Aviv University Centre for Biodiversity and Environment Research, UCL The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
-- Roi Maor PhD candidate School of Zoology, Tel Aviv University Centre for Biodiversity and Environment Research, UCL The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
Roi Maor PhD candidate School of Zoology, Tel Aviv University Centre for Biodiversity and Environment Research, UCL [[alternative HTML version deleted]]