The RAD partitioning works and is much cleaner, and tells us something meaningful

I was beginning to think this was a waste of time, but in fact the analysis now is so much cleaner and nicer. The workflow is explained in the last few posts. The way I'm deciding who votes for vs. against a locus now is by a likelihood threshold. In the plots below, I've filtered out all loci that have < 5 log-likelihood point spread, and I only consider trees voted for if they are in the top 2 lnL point range, and voted against if they are in the bottom 2 lnL point range. Narrowing the spread doesn't change the story, but makes it noisier.

The result is striking: there is no tree favored by more loci than the best tree. There should be an outlier test associated with this, but just eyeballing it, there seems no obvious second story lurking among the suboptimal trees that is supported by a lot of loci. What is needed now is some simulation of what the data look like from a clean, single-story dataset. It's hard to imagine it looking much cleaner than what we get here.

Loci that favor the optimal topology and
each of 200 permuted topologies

Loci that disfavor the optimal topology and
each of 200 permuted topologies