Post date: Jan 27, 2015 11:45:44 PM
Cross-validation did not appear to be that informative using the approach of dropping some individuals. Perhaps this isn't surprising, and dropping loci would be more informative. I modified the code to try this and a few quick tests suggest that this will likely be the case. The nice thing is that I have simulations with known parameters to make sure. So, I have committed the revised code to cvs, and I am re-running the cross-validation procedures (scale 0.1, 0.5, 1, 5, 10) to see if the scale selected by cross-validation better matches the scale that gives the lowest RMSE and best CI coverage of true parameter values. I am running one chain per data set with a 4000 iteration run, 2000 iteration burnin, and thinning interval of 4. The results will be in /labs/evolution/projects/popanc_sims/mcmc.cv/. Here is an example command:
cd /local/scratch/
popanc -o outcvF_r9f0.3gen_demog_gens10.hdf5 -m 4000 -b 2000 -t 4 -f 1 -w 0 -v 1 -c 0.1,0.5,1,5,10 /labs/evolution/projects/popanc_sims/sims/genoP0F_r9f0.3gen_demog_gens10.txt /labs/evolution/projects/popanc_sims/sims/genoP1F_r9f0.3gen_demog_gens10.txt /labs/evolution/projects/popanc_sims/sims/genoAdmxF_r9f0.3gen_demog_gens10.txt > /labs/evolution/projects/popanc_sims/mcmc.cv/cvF_r9f0.3gen_demog_gens10
scp outF_r9f0.3gen_demog_gens10.hdf5 /labs/evolution/projects/popanc_sims/mcmc.cv/