"Improper priors" on posterior statistics

[GoogleCodeExporter]

Hey Andrew,

Why does BEAST 1.6 complain that the user should specify a prior on covariance, 
meanRate and coefficientOfVariation? They are not even parameters of the model. 
I propose to remove all recommendations to place proper priors on arbitrary 
statistics of the posterior.  
Otherwise we will have to demand that the user puts a prior on the logarithm of 
the left child of the root and any other statistic we come up with.

There is no sense in which not having an explicit density on the covariance 
statistic is "improper". The posterior distribution is defined on the 
parameters that are sampled and the prior distribution should integrate to 1 on 
that same set of parameters. A proper prior on ucld.stdev and ucld.mean will 
lead to a well-defined posterior distribution on the parameters without any 
further explicit densities or bounds being placed on the meanRate or the 
covariance or the coefficientOfVariation.

I am going to get Walter to remove all of these types of warnings from 
non-parameters and in the prior panel they will say "none" or "indirectly 
specified through parameter x" or something like that -- with the option of 
placing a prior on them but no warning about leaving them as is.

Is this okay?

Cheers
Alexei

Original issue reported on code.google.com by dong.w.xie@gmail.com on 4 Oct 2010 at 12:34

[GoogleCodeExporter]

It has being doing this since day one (v1.5) and we should probably have 
noticed. I agree it is inappropriate to flag them as problematic. I think these 
all return true for isStatistic() so it is easy to omit them.

A.

Original comment by dong.w.xie@gmail.com on 4 Oct 2010 at 9:49

[GoogleCodeExporter]

Original comment by dong.w.xie@gmail.com on 5 Oct 2010 at 2:47

• Changed state: Fixed