Post date: Aug 21, 2013 11:1:5 PM
I returned on the allele frequency change analysis for the single generation Timema field experiment. This analysis uses software that I developed (bgsr) to estimate selection differentials and effective population size from allele frequency change. A key problem is that I want to model selection differentials hierarchically while inducing shrinkage on these estimates (point estimates are too high given a general expectation of genetic drift). I modified the software to model components of the selection coefficient as t-distributed across loci with a fixed standard deviation. Results from an initial test with sd = 0.01 or 0.05 and a subset of 10k loci are in projects/timema_reproduction_experiment/mcmc. I need to look at these more. I considered taking an empirical Bayes approach and setting the sd based on point estimates of the allele frequency change and the selection differential that would lead to that change, but again these are way too high to be reasonable (i.e., they don't adequately penalize the selection hypothesis). It is probably best to stick with 'reasonable' prior values for the sd.