Post date: Mar 23, 2015 4:17:51 PM
The first test runs exhibited good mixing except for the scale parameter. I think the problem was with the update, which was based on the probability of the population ancestry frequency. The current value was nearly always better as the scale parameter directly affected the probability calculation. I decided instead to calculate the probability of the sample ancestry frequency integrating over the current value of the population ancestry frequency using the beta-binomial distribution. Note, I had to modify my code for this calculation so that everything is done on the log-scale to avoid underflow/overflow. Things appear to be working now and it looks like I get different estimates for the posterior of the scale parameter for different data sets (this is good). I am re-running my tests with two of the data sets. Here is an example:
cd /local/scratch/
sleep 5
popanc -o estpanc_demogGens50Rep3Chain1.hdf5 -m 30000 -b 10000 -t 5 -n 4 -d 10 -s 1 -c 1.3 -u 10 -l 0.1 -a 0.1 -z 1 /labs/evolution/projects/popanc_sims/sims/genoP0F_r3gen_demog_gens50.txt /labs/evolution/projects/popanc_sims/sims/genoP1F_r3gen_demog_gens50.txt /labs/evolution/projects/popanc_sims/sims/genoAdmxF_r3gen_demog_gens50.txt
scp estpanc_demogGens50Rep3Chain1.hdf5 /labs/evolution/projects/popanc_sims/mcmc/
One other note. The whole analysis is pretty fast (3 or 4 hours).