[R--gR] Modelformulae

Steffen Lauritzen steffen at stats.ox.ac.uk
Fri Aug 20 11:14:47 CEST 2004


In a slightly longer perspective, I think the "CoCoCg"-models are far from
being sufficient, so the language should preferably be extendable to express
many other models, and their combinations, to achieve a long-term goal of
"Programming with models".

This may not be easy to do in a single step, but there is no reason to
exclude anything from the language which can be given a clear and
unambiguous meaning.

If terms are included like log(X) or X*Y*Z, the CoCoCg-parser can just say
that this model is not available as a CoCoCg-model.

But some crisp conventions should be formulated in any case.

Steffen


----- Original Message ----- 
From: "Poul Svante Eriksen" <svante at math.aau.dk>
To: "Steffen Lauritzen" <steffen at stats.ox.ac.uk>
Cc: "gRlist" <R-sig-gR at stat.math.ethz.ch>
Sent: Friday, August 20, 2004 8:33 AM
Subject: Re: [R--gR] Modelformulae


> A few comments - just to clear my mind.
>
> Steffen Lauritzen wrote:
>
> >Dear gR-folks
> >
> >The Danish gR-gang have been talking about describing a model language
for
> >graphical models that
> >
> >1) could specify at least chain graph models, based on the most general
> >hierarchical mixed models as
> >described in Lauritzen (1996) [my book], section 6.4, pages 199-216.
(More
> >general than MIM-models).
> >
> >
> Are we limiting the discussion to the Conditionally Gaussian(CG) case
> with independent and identically distributed observations?
> That would make life much more easy and clearly rule out X*Y*Z between
> continuous variables.
> And I have no problem with interpreting "~" as "log f ~" as long as we
> are careful to include quadratic terms of  continuos variables.
> But it also rules out things like ~Z+log(X)  and is far from having the
> same flexibility as the BUGS syntax.
> We might allow ~Z*(Z+log(X))|X  and thereby leaving CG, but sticking to
> gaussian error.
>
> If we want a more general BUGS/glim-like syntax, we need to consider,
> how to include information on link and errordistribution. And it would
> be nice
> to facilitate correlation across unit, e.g. by allowing ~X*B(X)|B(X),
> where B is the backshift operator.
>
> Regards, svante
>
> -- 
> ----------------------------------------------------
> Poul Svante Eriksen, office G1-113
> Department of Mathematical Sciences
> Aalborg University
> Fredrik Bajers Vej 7G, DK-9220 Aalborg East, Denmark
> tel.: (+45) 96358868
> ----------------------------------------------------
>
>




More information about the R-sig-gR mailing list