RELL Test for Phylogenetic Inference
Performs the RELL test for finding the largest item. This calculates AU p-values for each item via the multiscale bootstrap resampling. This is particularly useful for testing tree topologies in phylogenetic analysis.
relltest(dat,nb=10000,sa=9^seq(-1,1,length=13),ass=NULL, cluster=NULL,nofit=FALSE,models=NULL,seed=100)
dat: a matrix. Row vectors are to be resampled. Each column vector gives score values to be evaluated for an item. For the phylogenetic analysis, dat[i,j] is the site-wise log-likelihood value at site-i for tree-j, and we are to find the tree with the largest expected value of sum(dat[,j]).nb: Number of replicates for each scale.sa: Scales in sigma squared ().ass: A list of association vectors for testing edges as well as trees. If ass=NULL, then only the results for trees are returned.cluster: parallel cluster object which may be generated by function makeCluster.nofit: logical. Passed to sbfit.models: character vectors. Passed to sbfit.seed: If non NULL, then a random seed is set. Specifying a seed is particularly important when cluster is non NULL, in this case seed + seq(along=cluster) are set to cluster nodes.relltest performs the resampling of estimated log-likelihoods (RELL) method of Kishino et al. (1990). For resampling indices stored in a vector i, the resampled log-likelihood for a tree-j is approximately calculated by sum(dat[i,j]). This approximation avoids time-consuming recalculation of the maximum likelihood estimates of tree parameters, which are to be calculated by an external phylogenetic software such as PAML as described in mam15. In the implementation of relltest, the resampled log-likelihood is calculated by sum(dat[i,j])``*nrow(dat)/length(i) so that the statistic is comparable to the case when .
relltest first calls scaleboot internally for multiscale bootstrap resampling, and then scaleboot
calls sbfit for fitting models to the bootstrap probabilities. The AU p-values (named "k.3") produced by the summary method are improvements of the third-order p-values calculated by CONSEL software (Shimodaira and Hasegawa 2001). In addition, relltest calls scaleboot with sa=1 for calculating p-values via the Shimodaira-Hasegawa test (SH-test) of Shimodaira and Hasegawa (1999).
See mam15 for details through an example.
relltest returns an object of class "relltest" that is inherited from the class "scalebootv" by adding two extra components called "stat" and "shtest". "stat" is a vector of the test statistics from the SH-test (i.e., the log-likelihood differences), and "shtest" is a list with two components: "pv", a vector of SH-test p-values, and "pe", a vector of standard errors of the p-values. The results of multiscale bootstrap resampling are stored in the "scalebootv" components returned by a call to sbfit.
Kishino, H., Miyata, T. and Hasegawa, M. (1990). Maximum likelihood inference of protein phylogeny and the origin of chloroplasts., J. Mol. Evol., 30, 151-160.
Shimodaira, H. and Hasegawa, M. (1999). Multiple comparisons of log-likelihoods with applications to phylogenetic inference, Molecular Biology and Evolution, 16, 1114-1116.
Shimodaira, H. and Hasegawa, M. (2001). CONSEL: for assessing the confidence of phylogenetic tree selection, Bioinformatics, 17, 1246-1247 (software is available from http://stat.sys.i.kyoto-u.ac.jp/prog/consel/).
Luke Tierney, A. J. Rossini, Na Li and H. Sevcikova. snow: Simple Network of Workstations. R package version 0.2-1.
Hidetoshi Shimodaira
sbfit, scaleboot, mam15.
## Not run: ## a quick example data(mam15) # loading mam15.mt mam15.trees <- relltest(mam15.mt,nb=1000) # nb=10000 is default mam15.trees # SH-test p-values and result of fitting summary(mam15.trees) # AU p-values ## End(Not run) ## Not run: ## An example from data(mam15). ## It may take 20 mins to run relltest below. mam15.mt <- read.mt("mam15.mt") # site-wise log-likelihoods mam15.trees <- relltest(mam15.mt) # resampling and fitting summary(mam15.trees) # AU p-values ## End(Not run)