Priors and single-locus estimations iin Extended Bayesian Skyline Plot

[Andrey Yurchenko]

 Dear Beast-users!! I'm new in the satisfied field of the Bayesian Skyline Plotting.  I learn to use this methods, but although I read almost all manuals I cannot find some crucial moments of the estimating. Help me with two issues please:

 

Q1. How I understand , Extended Bayesian Skyline Plot is more new and powered method then BSP and GMRF. But there is one challenge - EBSP has been positioned as  multi-locus estimator, but my data on birds mostly consist of large samples with single locus - Is it reasonably and properly to use EBSP  on my data (because it doesn't need a special insertion of the change points number and give more rapid ESS)  ??

 

Q2. When I use EBSP with my data, I get very good results, but I am not confident of using priors in BEAUTY.  Should I change the priors: 

demographic.populationSizeChange (default sittings is Poisson [0.693147]); 

treeModel.rootHeight (Using tree prior in [0: infinit.]), 

demographic.populationMean (1/x, initial = 0.016)    

or could leave default settings?

 

Q3. What is the distribution I have to choose in the section of priors from the Model Test??  For example, I have got in the Model Test such values: kappa = 18.4990, p-inv = 0.7820, gamma shape = 0.4970.  Could I choose normal distribution with extremely low value of Stdev (i.e. 0,00000000000001) to truncate and make shape of this parameters very strict or should I use other and more flexible distributions (uniform, exponential) ???

 

Thank you very much!

sorry for my English :)

 

[Alexei Drummond]

On Thursday, November 15, 2012 2:32:45 AM UTC+13, Andrey Yurchenko wrote:

 Dear Beast-users!! I'm new in the satisfied field of the Bayesian Skyline Plotting.  I learn to use this methods, but although I read almost all manuals I cannot find some crucial moments of the estimating. Help me with two issues please:

 

Q1. How I understand , Extended Bayesian Skyline Plot is more new and powered method then BSP and GMRF. But there is one challenge - EBSP has been positioned as  multi-locus estimator, but my data on birds mostly consist of large samples with single locus - Is it reasonably and properly to use EBSP  on my data (because it doesn't need a special insertion of the change points number and give more rapid ESS)  ??

EBSP can be used with a single locus. All three models (EBSP, BSP, GMRF) have their pro's and con's. GMRF and BSP can actually be regarded as two instances of a superfamily of methods, where in GMRF-language the BSP (with # groups set to # coalescent events) has a non-time-aware exponential smoothing function whereas GMRF uses a log-normal smoothing function (i.e. gaussian in log space) and has both time-aware and non-time-aware models. I would recommend either GMRF or EBSP depending on which model you prefer.

 

Q2. When I use EBSP with my data, I get very good results, but I am not confident of using priors in BEAUTY.  Should I change the priors: 

demographic.populationSizeChange (default sittings is Poisson [0.693147]); 

treeModel.rootHeight (Using tree prior in [0: infinit.]), 

demographic.populationMean (1/x, initial = 0.016)    

or could leave default settings?

There has been a tendency among BEAST users to move away from improper priors (like 1/x) as in many cases they lead to improper posteriors. Even when they produce a proper posterior, they cause problems for methods like path-sampling and stepping-stone sampling that aim to estimate marginal likelihoods by sampling posterior-prior mixtures. So you may want to change the demographic.populationMean to some very broad proper distribution like a log-normal with suitable M and large S. The treeModel.rootHeight prior should be left alone because the coalescent density already provides a prior for the root height.

 

Q3. What is the distribution I have to choose in the section of priors from the Model Test??  For example, I have got in the Model Test such values: kappa = 18.4990, p-inv = 0.7820, gamma shape = 0.4970.  Could I choose normal distribution with extremely low value of Stdev (i.e. 0,00000000000001) to truncate and make shape of this parameters very strict or should I use other and more flexible distributions (uniform, exponential) ???

I would not recommend fixing the substitution model parameters. BEAST can easily estimate them for your, and in doing so will correctly incorporate any uncertainty in those parameters into the estimates of your parameters of interest.

[Zheng Hou]

Dear Alexei,

 

What do time-aware models and non-time-aware models mean respectively? Thank you!

2012/11/16 Alexei Drummond <alexei....@gmail.com>

--
You received this message because you are subscribed to the Google Groups "beast-users" group.

To view this discussion on the web visit https://groups.google.com/d/msg/beast-users/-/IEDhBFaQuo0J.

To post to this group, send email to beast...@googlegroups.com.
To unsubscribe from this group, send email to beast-users...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/beast-users?hl=en.

--
School of Life Sciences ,  Fudan University
220# Handan Road, Shanghai(200433)P.R.China.

[Alexei Drummond]

Both BSP and GMRF use the times of coalescent events as the boundaries for changes in population size. "Time-aware" smoothing takes account that coalescent events are separated by different amounts of time so that population size may have changed more in a longer inter-coalescent time interval. Non-time-aware smoothing only takes account of the number of coalescent events separating two coalescent intervals in determining the amount of population size change expected in the prior between those two intervals. That is, the prior expectation of population size change between two subsequent intervals is the same regardless of how much time is spanned by the intervals.