Constraints in INLA
Gwénaël Leday
Is it possible in INLA to specify linear equality or inequality constraints on regression parameters??
Any help would be appreciated,
Gwenael
INLA help
inqualities constraints: no. equality constraits: yes.
each f() model has two arguments
constr = TRUE will impose a sum-to-zero constraint
extraconstr (see ?f) is a general one. additinal to 'constr'.
in case you have constraints 'between' different 'f()'s, then there is
an option in the linear predictor, see ?control.predictor and argument
'cross', which essential, allow you do any anything if you define
formula accordingly.
if you have a spesific problem, please post some detals.
--
INLA help <he...@r-inla.org>
R-INLA
INLA help
> On Fri, 2012-03-16 at 02:39 -0700, Gwénaël Leday wrote:
> > Dear all,
> > Is it possible in INLA to specify linear equality or inequality
> > constraints on regression parameters??
oops, you were spesific.
if you have to effects, v and w, and you want beta.v + beta.w = 0, say,
then
y ~ f(i, v, model="iid", constr=TRUE, values = c(1,2)) +
f(j, w, copy="i")
where
i = rep(1,n)
j = rep(2,n)
would do the trick.
an alterntive solution is to extend the model so that (eta = linear
predictor,) eta[n+1] = beta.v, and eta[n+2] = beta.w, and then use
argument
control.predictor=list(cross = c(rep(NA,n),1,1))
see ?control.predictor
A third option is to use the 'A' matrix in the control.predictor and
observe it with 0 and high precision.