gamma value and substitution rate
cec
I'm a new user on Beast and I'm more a biologist than a statistician.
I'm working with mtDNA and I have two simple questions that I need to
be sure before starting analysing:
1) I run modeltest to know the best model for my sample set and so I
obtained a gamma value. I don't understand well how to fix the gamma
value in the Priors in Beast.
2) Previous publication estimate substitution rate around 2.3 10-7 and
4.1 10-7. I don't know how is the best solution to implement this
information in prior settings.
Thanks a lot four your help
Simon Ho
> 1) I run modeltest to know the best model for my sample set and so I
> obtained a gamma value. I don't understand well how to fix the gamma
> value in the Priors in Beast.
It's better to let BEAST estimate the value of the gamma shape
parameter (alpha) rather than fixing it to the maximum-likelihood
estimate obtained in Modeltest (because the estimate has been obtained
from the same data set that you're analysing). To do this, choose
"Gamma" from the "Site Heterogeneity Model" option in the "Site
Models" tab in BEAUti, and leave the prior for "alpha" unchanged.
> 2) Previous publication estimate substitution rate around 2.3 10-7 and
> 4.1 10-7. I don't know how is the best solution to implement this
> information in prior settings.
You can put a prior on the clock.rate parameter (if you are using a
strict clock) or the ucld.mean parameter (if you are using an
uncorrelated lognormal relaxed clock). With the two values you have
given, perhaps you could use a normal prior, with a mean of 3.2x10-7
and a standard deviation of 0.45918x10-7. This gives a normal
distribution centred on the midpoint of the interval that you have
given, with the central 95% of the probability lying within the
interval. Note that the values for this interval can be found in the
same way as you would for a 95% confidence interval (95% probability
lying within +/- 1.96 standard deviations from the mean).
Cheers,
Simon
JonBK
On Jan 27, 1:14 am, Simon Ho <drsimo...@gmail.com> wrote:
> Hi cec,
> >
> It's better to let BEAST estimate the value of the gamma shape
> parameter (alpha) rather than fixing it to the maximum-likelihood
> estimate obtained in Modeltest (because the estimate has been obtained
> from the same data set that you're analysing). To do this, choose
> "Gamma" from the "Site Heterogeneity Model" option in the "Site
> Models" tab in BEAUti, and leave the prior for "alpha" unchanged.
> Simon
In my data set (two mtDNA loci) the alpha estimates I get from BEAST
seldom seem correct. If I leave the default prior for "alpha" (uniform
between 0 and 1000) I often get estimates of alpha in the hundreds,
indicating little rate variation among sites, whereas the truth is
that there is significant rate variation (confirmed by manual
inspection of the data). Modeltest indicates gamma between 0.5 and
1.5. Inspection of the trace shows that the alpha value does not
converge at any particular value, but continues to vary between 0 and
1000. This behaviour happens repeatedly in different runs and in
different clades. If I use a normal prior with the mean equal to the
Modeltest estimate I get an alpha estimate that is close the Modeltest
value. I also tried increasing the number of gamma categories from 4
to 8 but that did not improve the behaviour.
Jon
alexei
The prior for the shape parameter of the gamma distribution in BEAST
you are using is a uniform distribution between 0 and 1000. That means
that the *prior* probability of the shape being >1 (and smaller than
1000) is 99.9% and the *prior* probability of 10<shape<1000 is 99%. As
you can appreciate this is a fairly strong prior. Even if you data has
"significant" (in a likelihood ratio sense) support for a shape of
(say) 1.0, this may still be overcome by the prior distribution. A
more sensible prior may assist you. All priors in BEAST are merely
defaults, and its impossible ahead of time to pick sensible defaults
for all possible analyses. So its extremely important in a Bayesian
analysis to assess your priors. To answer the original question: its
better to pick a good prior, then to fix the parameter value to the ML
estimate. Doing the latter artificially reduces the variance of the
posterior distribution by using the data twice.
Cheers
Alexei
JonBK
that makes sense. If I limit the uniform prior to a maximum of e.g.
10, the estimated alpha becomes far more reasonable.
Jon