sample from prior only

[Peter]

G'day folks

I'm hoping for some advice on what the results from sample from prior
only mean. I couldn't really find anything that explained how to
interpret the output and what the implications are in the manual,
website or from previous posts to the google group.

I am using a single calibration point using a lognormal prior with a
mean and standard deviation of one with an offset of 14. This is for
a small dataset with 9 species and three genes (7,235 bp).

When I run beast with no data for 50 million generations I get the
following from my mcra values

Summary Statistic
mean 16.9396
stderr of mean 1.4585E-2
median 16.0343
geometric mean 16.7313
95% HPD lower 14.0776
95% HPD upper 22.5217
auto-correlation time (ACT) 1046.664
effective sample size (ESS) 42994.6954

When I run it with data I get the following values

Summary Statistic
mean 15.9065
stderr of mean 2.0541E-2
median 15.441
geometric mean 15.832
95% HPD lower 14.0719
95% HPD upper 19.0662
auto-correlation time (ACT) 6969.902
effective sample size (ESS) 6456.4752

First of all I presume that this is the key value that I should be
comparing between the runs?

The two results for mcra look very similar to me. I'm presuming that
this indicates that the prior is driving the results and that the data
may not be very informative?

So, what if anything can one to? Am I screwed as far as my results
go?

For what it's worth r8s spits out very similar dates for the same
dataset, which gives me some confidence that the results seem
reasonable. But I'm not sure what to do relative to what I say /
present in the paper.

Any help would be greatly appreciated.

Thanks
Peter

[pepster]

On Aug 19, 2:21 am, Peter <peter.goo...@unmack.net> wrote:
> G'day folks
>
> I'm hoping for some advice on what the results from sample from prior
> only mean.  I couldn't really find anything that explained how to
> interpret the output and what the implications are in the manual,
> website or from previous posts to the google group.
>
> I am using a single calibration point using a lognormal prior with a
> mean and standard deviation of one with an offset of 14.  This is for
> a small dataset with 9 species and three genes (7,235 bp).
>
> When I run beast with no data for 50 million generations I get the
> following from my mcra values
>
> Summary Statistic
> mean    16.9396
> stderr of mean  1.4585E-2
> median  16.0343
> geometric mean  16.7313
> 95% HPD lower   14.0776
> 95% HPD upper   22.5217
> auto-correlation time (ACT)     1046.664
> effective sample size (ESS)     42994.6954

What you want to do first is plot the distribution from the run vs.
the log-normal prior. This will give you an indication how far the
"induced" prior (the one you actually turn out to use) is from your
intended prior.

>
> When I run it with data I get the following values
>
> Summary Statistic
> mean    15.9065
> stderr of mean  2.0541E-2
> median  15.441
> geometric mean  15.832
> 95% HPD lower   14.0719
> 95% HPD upper   19.0662
> auto-correlation time (ACT)     6969.902
> effective sample size (ESS)     6456.4752
>
> First of all I presume that this is the key value that I should be
> comparing between the runs?
>
> The two results for mcra look very similar to me.  I'm presuming that
> this indicates that the prior is driving the results and that the data
> may not be very informative?

The prior may be driving the results, or perhaps the data "agrees"
with the prior. I would try other runs where the mean of the prior
varies (away from 16) and see if the results follow the prior or not.

But at the end this is a Bayesian analysis. It may be interesting to
know if the data is consistent with the prior, but your result
reflects the best estimate given your data and your prior knowledge.
If the data can change your mind about the prior - is it prior
knowledge anymore?

-Joseph

[Peter Unmack]

Thanks for the response Joseph.

> The prior may be driving the results, or perhaps the data "agrees"
> with the prior. I would try other runs where the mean of the prior
> varies (away from 16) and see if the results follow the prior or not.

Right, so I ran it again with the mean at 2 which gave a broader
confidence interval and a higher mean, thus it seems like the data and the
original prior are both agreeing with one another fairly well, and that
after setting the prior to a broader range, the data were not matching the
prior as well as it did with a mean of 1.

Perhaps someone else can offer some insight though as it is still not
clear to me, besides the prior that I am defining my calibration point,
what else should one be looking at in a run that samples from the prior
only? The other values seem like they are all over the place, and many of
them have low ESS scores.

Thanks
Peter

[Reena]

Dear Beast and Bayesian statistics Gurus,

I too am confused about the "sample from prior only" option.

I have trees from 3 independent runs with data that have converged well with very good mixing on all the statistics in the log file (as viewed in tracer). I am trying to determine the prior probability on the MRCA so that I can calculate a bayes factor. When I choose the "sample only from prior" option in BEAUTi, the resulting runs  (100 million chain length) have extremely horrible looking trace files and ESS
values of 4 on all statisitcs except for the MRCA and the clock rates. Can I still use the MRCA statistic from this prior only run to calculate a bayes factor if the overall posterior and prior values do not converge?

Thank you

[mumm...@gmail.com]

Hey 

 I have problem about "Sample from prior",when i run Beauti without data in MCMC,I can not save,it shows "there is no data to save to file ",i do not know why.

thanks

Jin

在 2010年8月18日星期三 UTC+8下午10:21:42，Peter写道：

[Remco Bouckaert]

Hi Jin,

Before being able to sample from the prior, you need to specify a model, that is, set up a tree prior, clock model, site model, etc. Importing data (via the File/Import alignment menu) is the easiest way to specify which taxa are required for the tree, and to set up a model, which you then can edit in BEAUti. After you set up the model you will be able to save.

Cheers,

Remco

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at https://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.

[mumm...@gmail.com]

Hi Remco

    Thanks for your reply.Another problem about sample time used in the tip dates ,i see some sequences tip dates with "2007.34246575" in Beauti,i can not understand why the year is not the integer.

     I also do not know the ”Gamma category“  in Beauti site model ,why it is better to choose it from 4 to 6.

     Best

Jin

在 2018年11月20日星期二 UTC+8上午7:41:44，Remco Bouckaert写道：

[Artem B]

Categories above 6 and more make computing more complicated without having significant difference between results.

[mumm...@gmail.com]

Hi

  Thanks for your advice.

   The MCMC option have the "Pre Burnin",no matter how i change the number of it the result after run BEAST the number of trees do not change,i am very confused about it .Maybe i misunderstanding the meaning .

   Best

   Jin

在 2018年12月6日星期四 UTC+8下午7:31:18，Artem B写道：