Heard back from Graham Friday on manuscript revisions. Notes on reviewing edits:
Reconsider title? (Currently: “is your phylogeny informative? estimating power in phylogenetic trees”). Suggestions from Graham/Brian?
1000 taxa, 500 taxa seem to break the ouch package inferences.
Add table for AIC comparisons etc
Does AIC work better/worse on the lambda = 0 trees?
n[1:3][1] 10 50 100 aic_errors_size <- sapply(1:(length(n)-2), function(i) sum(size[[i]]\(null_dist > aic)/size[[i]]\)nboot ) aic_errors_size [1] 0.1505 0.1285 0.0845 lambda [1] 0.01 0.10 0.40 0.70 1.00 aic_errors_shape <- sapply(1:length(n), function(i) sum(shape[[i]]\(null_dist > aic)/shape[[i]]\)nboot ) aic_errors_shape [1] 0.0000 0.2580 0.0305 0.0415 0.0685
Improves as tree size increases, impact of lambda somewhat unclear – AIC seems to perform better on heavily transformed trees, except for lambda = 0.1 seems anomalous? Worth a closer look
Rerrunning with a few other examples, now on zero nice 1:
alpha <- c(seq(.1, 1, length=10), 2:10, seq(20,50, by=10))
n <- c(10, 20, 40, 60, 80, 100, 150, 200)
lambda <- c(.01, .05, .1, .2, .4, .6, .8, 1)
Recall this results only from the null distribution tightening up, as it is Type I error (area of null right of the AIC threshold, independent of test distribution).
edit/rewrite introductory paragraphs (less passive voice). Intro is set up as addressing two key issues, framed as: model problems/model selection/model adequacy, etc, vs data problems. Perhaps better to phrase more classically in terms of Type I /Type II error explicitly? emphasize power effectively.
Rest of edits have been relatively straight forward changes.