pymc-user

[Pymc-user] Correlation between parameters

[David H. <dav...@et...>]

I have a model with two parameters A and B. When I compute their
distribution using PyMC, I get a 'cigar' distribution, in other words,
parameters A and B are strongly correlated. I figure this must put a
strain on the routine, forcing it to decrease the jump radius to
maintain an acceptable acceptance ratio.  Is there a way to indicate
that those parameters are correlated and to use a multivariate
distribution for the jumps ?


Thanks, 


David


P.S. Adding the generalized extreme distribution to flib would sure be
appreciated. 

Re: [Pymc-user] Correlation between parameters

[Chris F. <fon...@gm...>]

On 2/15/06, David Huard <dav...@et...> wrote:
>  I have a model with two parameters A and B. When I compute their
> distribution using PyMC, I get a 'cigar' distribution, in other words,
> parameters A and B are strongly correlated. I figure this must put a stra=
in
> on the routine, forcing it to decrease the jump radius to maintain an
> acceptable acceptance ratio.  Is there a way to indicate that those
> parameters are correlated and to use a multivariate distribution for the
> jumps ?
>

Sorry for the late reply. The best way to deal with highly correlated
parameters is to block update -- i.e. sample them together. Thus,
rather than have A and B, you can just specify a vector-valued
parameter A with init_val =3D [x, y]. Of course, this assumes A and B
both use the same likelihood function. Is this the case? I should
probably add the ability to block update among arbitrary parameters.

>  P.S. Adding the generalized extreme distribution to flib would sure be
> appreciated.

Look for it in the next release.

C.

--
Chris Fonnesbeck + Atlanta, GA + http://trichech.us

Re: [Pymc-user] Correlation between parameters

[David H. <dav...@gm...>]

Thanks Chris,

Just so I understand correctly, does specyfing p =3D [A,B] means that
A,B are updated using bivariate normal jumps with a correlation
coefficient that is tuned along with the covariances ?


Also, I'm planning to use PyMC to do mixed Gibbs-Metropolis sampling.=20
That is, some parameters would be sampled directly from a distribution
and others would be chosen via MH jumping rules. If I ever get there,
I could write a little how-to for inclusion in the PyMC tutorial.
Would that be of any interest to you ?


Thanks,

David



2006/2/24, Chris Fonnesbeck <fon...@gm...>:
> On 2/15/06, David Huard <dav...@et...> wrote:
> >  I have a model with two parameters A and B. When I compute their
> > distribution using PyMC, I get a 'cigar' distribution, in other words,
> > parameters A and B are strongly correlated. I figure this must put a st=
rain
> > on the routine, forcing it to decrease the jump radius to maintain an
> > acceptable acceptance ratio.  Is there a way to indicate that those
> > parameters are correlated and to use a multivariate distribution for th=
e
> > jumps ?
> >
>
> Sorry for the late reply. The best way to deal with highly correlated
> parameters is to block update -- i.e. sample them together. Thus,
> rather than have A and B, you can just specify a vector-valued
> parameter A with init_val =3D [x, y]. Of course, this assumes A and B
> both use the same likelihood function. Is this the case? I should
> probably add the ability to block update among arbitrary parameters.
>
> >  P.S. Adding the generalized extreme distribution to flib would sure be
> > appreciated.
>
> Look for it in the next release.
>
> C.
>
> --
> Chris Fonnesbeck + Atlanta, GA + http://trichech.us
>
>
> -------------------------------------------------------
> This SF.Net email is sponsored by xPML, a groundbreaking scripting langua=
ge
> that extends applications into web and mobile media. Attend the live webc=
ast
> and join the prime developer group breaking into this new coding territor=
y!
> http://sel.as-us.falkag.net/sel?cmdlnk&kid=110944&bid$1720&dat=121642
> _______________________________________________
> Pymc-user mailing list
> Pym...@li...
> https://lists.sourceforge.net/lists/listinfo/pymc-user
>