Difficulties estimating a model with user.expert
Philip Mostert
Dear INLA help,
I’m trying to implement a model using pre-defined integration points using the `user.expert` argument in `control.inla`, however the model seems to keep failing to estimate right at the end. The last hyperparameter value (rho for the ar1 model) appears to be going everywhere, and so it’s rejecting the integration given that the diff.logdens is incredibly large. From what I believe, you can control the cutoff for integration using the `diff.logdens` argument, but changing this value doesn’t seem to help (if I set `int.strategy = ‘grid’`, changing `diff.logdens` doesn’t change anything unless `inla.mode = “classic”`). I was wondering if there was anything I could do to help with the estimation for a model such as this? Attached is the verbose output from one of the models that crashed.
Helpdesk (Haavard Rue)
I see now what is happening.
if the log density at theta, during the integration, is to different from the
value at the mode, then its contribution is ignored. this stops the computations
at a early stage and '(interupt)' is recorded in the output. this saves time.
in your case, since you give the integration points yourself, then you end up in
a situation where ALL integration points are rejected,
> number of early stops = 79
and there is nothing to summarize, and the program fails.
This should not happen. the correct behaviour, should be: with user defined
points (that are not standarized), then there should be no rejection, and it
should just compute whatever is requested.
I'll add this fix and try to make a new testing-version tomorrow
best
H
--
Håvard Rue
he...@r-inla.org