Reasonable way to estimate an appropriate alpha?

[E Guevara]

Dear BUCKy group,

I am analyzing a number of datasets for which I have little prior knowledge about how much discordance among loci to expect. So, to guide my estimation of alpha, I have been setting mrbayes to estimate contrees for each locus during the mrbayes portion of the analysis. I then count how many different topologies I see in the contrees to get a sense of discordance.

For several of my datasets, all or nearly all of the mb contrees have a different topology, so I have been using a very high alpha or the alpha independence prior. I've also done some runs using the rate multiplier alpha prior for comparison. I've noticed the CFs vary a lot with how high alpha is (being, as expected, much lower when alpha increases toward infinity). I've also noticed, though, that in one of my datasets, the topology of the concordance trees and population trees actually changes with whether the rate multiplier or independence prior is used. 

In this dataset, all loci have a different mb contree topology. The results of the run using the rate multiplier prior indicate MCMC acceptance was entirely concentrated in the highest alpha class (1000-10000). Interestingly, the taxa that "switch" positions between the trees estimating by the rate multiplier and independence runs are part of a trio of taxa that I am interested in examining ILS in.

Do you think using the degree of discordance among mb contrees is a reasonable approach to estimating alpha? For the dataset where the topologies changes depending on alpha, should I "trust" the topology estimated in the run using the independence prior, as available evidence suggests that there is very high discordance in this dataset?

Thanks very much for considering my questions!

Elaine