Markov chain parameters for BSP
J
Hello BEAST2 users,
I have several basic questions about the Markov chain parameters for BSP.
Under the MarkovChainPopsizes.t prior, there are three options: Jeffreys, Reverse, and UseLog.
I was figuring out the impact of these options on the analysis results. As far as I understand, “Jeffreys checkbox” means that Jeffrey’s uninformative prior is assigned on the most recent population size and “Reverse checkbox” defines the direction of Markov chain (if checked, Jeffrey’s prior is assigned on the last population size). My questions are as follows:
1. What is for the checkbox “UseLog”?
Is it just for transforming estimated parameter values to log-scale to facilitate computation?
When I checked this option box, ESS related to Bayesian Skyline, bPopSizes, and bGroupSizes
increased dramatically, resulting in much faster convergence.
Should I always use this option (logarithm of parameter values)?
2. When I experimentally checked the “Reverse” option, I could not find any notable changes in the parameter values estimated (sometime ESS are much increased). Is it for enabling reversible jump Markov chain or for the special case of data?
3. Are there any advantage to using Gamma Markov priors (by setting shape parameter [alpha] other than 1)? If yes, what criteria can be used to set the alpha value? Or is it perfectly OK to use the default value (alpha=1, exponential with mean=1)?
4. Final question. If I turn off “Jeffreys,” what prior is assigned by BEAST on the MarkovChainPopSizes.t? Uniform?
It is very hard to find documents related to above questions. Any comments are very welcome.
J.
J
Sam Zimmerman
Sam