Post date: Aug 29, 2013 6:4:48 PM
After working for the last two days I think the bgsr code is bug-free. The main (only) issues were numerical (i.e., probability calculations returning Inf, nan, 0, etc.). The fixed version is in cvs. I simulated a data set with 1000 neutral and 20 fitness-affecting loci (selection is quite strong, s = 0.5 for each locus). The R code I used is here. 10 of the selected loci have fitness effects that vary with a normally distributed co-variate. I am analysing this data with the full, constrained, and null model in bgsr (2 chains each, 5000 steps, burnin of 1000 and thinning intervals of 4), e.g.,
bgsr -g simoffspring.mpgl -f simfounder.mpgl -o ~/Scratch/bgsrtestFullC0.hdf5 -m 1 -e 1 -s 0.01 -p 1 -n 5000 -b 1000 -t 4
bgsr -g simoffspring.mpgl -f simfounder.mpgl -o ~/Scratch/bgsrtestNullC0.hdf5 -m 0 -e 1 -s 0.01 -p 1 -n 5000 -b 1000 -t 4
bgsr -g simoffspring.mpgl -f simfounder.mpgl -o ~/Scratch/bgsrtestConsC0.hdf5 -m 2 -e 1 -s 0.01 -p 1 -n 5000 -b 1000 -t 4