vb.correction issues

Eke Arua

unread,

Sep 19, 2023, 10:14:57 AM9/19/23

to R-inla discussion group

Hello,

I am currently doing my PhD and i'm trying to map the diabetes risk in Cape Town. I'm reatively new to INLA and spatial statistics as a whole but its been good so far.

I am trying to get a poisson disease risk model with BYM2 specifications to run. I have covariates and i've attached a reproducible example. The problem I have is I get the following error:

max_correction = 25.01 >= 25.00, so 'vb.correction' is aborted
Please (re-)consider your model, priors, confounding, etc.
You can change the emergency value (current value=25.00) by
'control.inla=list(control.vb=list(emergency=...))'

I still get model output and the results of the model seem valid and are inline with literature but I have doubts on the validity of the results. Additionally, I wanted to test the effects of varying the prior for \phi on the marginal posterior of its hyper-parameters (a sort of sensitivity analysis).

I have browsed this group for similar queries with the same error message. Although I've seen some error messages, they seem to be for different models and have different underlying causes. I tried the suggestions from the package (varying the initial values and changing the emergency value) and from those posts but nothing seems to work. My belief is it could be something wrong with the data but the error message does not reveal what exactly is wrong.

Thank you for your time.

Reproducable example.zip

Helpdesk (Haavard Rue)

unread,

Sep 19, 2023, 1:05:16 PM9/19/23

to Eke Arua, R-inla discussion group

Hi Eke,

it seems like there are some issues in your data with NA.

INLA replace NA's in covariates with 0, and I think this is not what you
want for the covariates.

Also inla(..., E=0, family="poisson") *is* allowed, but for a different
purpose (see the documentation for poisson likelihood), hence cases with
E=0, I set casest to NA.

with 58 regions (- 2 that are NA), then there is not that much to learn
from, so you could consider to put a more strict prior for the precision
in bym2

I have revised your code as below, and with this there are no 'warning'
you asked for (which is a sign a to vague model...)

Best
H

> --
> You received this message because you are subscribed to the Google
> Groups "R-inla discussion group" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to r-inla-discussion...@googlegroups.com.
> To view this discussion on the web, visit
> https://groups.google.com/d/msgid/r-inla-discussion-group/536f37c4-c289-42cf-8c7a-3cb0cb86cd0en%40googlegroups.com
> .

--
Håvard Rue
he...@r-inla.org

Helpdesk (Haavard Rue)

unread,

Sep 19, 2023, 1:06:50 PM9/19/23

to Eke Arua, R-inla discussion group

Now with the revised code

INLA code.R

Eke Arua

unread,

Sep 20, 2023, 3:42:32 AM9/20/23

to R-inla discussion group

Thank you so much for your prompt response. The model works for the first time with no warning or errors messages. Regarding the NA in the covariates, you are right, it would be ideal if they were not viewed as NA in the model but I will try and figure that out. Regardless, you have solved a big issue for me and I am greatful.