Post date: Nov 03, 2014 3:23:56 PM
PVE estimates from pimass looked more reasonable (median around 0.15 with 95% CIs spanning most of the distribution), but there were still severe MCMC problems. Effective sample sizes were on the order of 10 for parameters with 30,000 samples!, and posteriors varies among replicates. The SNP posterior inclusion probs. were also barely correlated across runs. So, while I think pimass is our best bet, we are not there yet. My solution is to run the chains longer (10 million) with a longer burn-in (5 million) and increased thinning (1000). And to run many chains (30). I have this running now, and the results will be in /home/A01963476/projects/timema_wgexperiment/gemma/. Here is an example command:
pimass-lin -g fix_geno_timemaC.txt -p pheno_samplesC.txt -w 5000000 -s 10000000 -num 1000 -o pimass_timemaCrep23 -cc