Posterior Distributions: M looks great, θ does not

[George zaragoza]

I'm revisiting an old dataset with a small number of SNPs (25) but large number of individuals (>700). I am assuming 3 subpopulations. I am using the following parameters to estimate M and θ:

updatefreq=0.200000 0.200000 0.200000 0.200000 #tree, parameter haplotype, timeparam updates

bayes-posteriorbins=1500 1500 1500 1500 1500

bayes-posteriormaxtype=TOTAL

bayes-file=NO

bayes-allfile=NO

bayes-all-posteriors=NO

bayes-proposals= THETA METROPOLIS-HASTINGS Sampler

bayes-proposals= MIG METROPOLIS-HASTINGS Sampler

bayes-proposals= DIVERGENCE METROPOLIS-HASTINGS Sampler

bayes-proposals= DIVERGENCESTD METROPOLIS-HASTINGS Sampler

bayes-proposals= GROWTH METROPOLIS-HASTINGS Sampler

bayes-priors= THETA * * UNIFORMPRIOR: 0.000000 10.000000 1.000000

bayes-priors= MIG * * UNIFORMPRIOR: 0.000000 1000.000000 100.000000

bayes-priors= RATE * * UNIFORMPRIOR: 0.000000 10000000000.000000 1000000000.000000

While M has looked great and given nice normal posterior distributions, θ has been either susceptible to what seems like local optimums (many small humps near a peak) or has been at the maximum end of the posterior distribution. What I've resorted to now is slightly increasing my upper bound of θ priors with the hope that my run time does not increase substantially. I was hoping to hear feedback on whether I should stick to this approach or if there's something I may not be considering when setting my parameters.

[George zaragoza]

Was revisiting this and realized a few more details would be good:

• I'm running migrate v5.0.3
• I'm running Migrate on a computer cluster using a single node with 16 CPUs, each with 16,000 MB of RAM
  • I am not using the parallel version of migrate as suggested by my cluster admin as I am only using one node
• My estimated time of completion for a job with a SNP HapMap with 25 alleles and (corrected from above) 449 animals (i.e., 898 individual alleles/gametes) 15 days (though my actual time of completion was 15 days)
  • My actual time of completion was 8 days
• Migrate usually runs until about ~0.5 "done" as below (maybe something to due with convergence? Most of my data is usually OK by this point):
  • 20:34:20   Sampling: prognosed end of run is 18:38 August 22 2023 [0.545455 done]

Hopefully this helps. With 16 CPUs I was hoping such runs would be quicker but I may just need to be patient.