Meta-analysis style model in INLA

[Will Fincham]

Hello All,

I'm new to INLA and I'm trying to use INLA to implement some models that I have previously been running in MCMCglmm. Specifically, I'm trying to run a meta-analysis type model where my response variables are weighted by another variable. My explanatory variables are coefficients from separate models, so the weighting is their standard error. The goal is to weight the response variables so that values with lower standard errors have a greater weighting in the model over those with larger standard errors. My models have so far been based on the MCMCglmm tutorial here.

Are these type of models currently possible in INLA?

My current (basic) INLA code is below as a sort of minimum working example/template.

Thanks in advance! 

priorpc <- list(prec = list(prior = "pc.prec", param = c(3, 0.05)))

 mod <- inla(beta ~ var1 + f(ran_eff,  model = 'iid', hyper = priorpc),

          data = datset1,

          quantiles = c(0.025, 0.975),

          control.compute = list(dic = TRUE, waic = TRUE),

          family = "lognormal")

[Helpdesk]

I guess what you're asking for if the reponse can have different
variances, like variance[i] = w[i] * common.variance
where 'common.variance' could f.ex be fixed to be 1, in the case you
know this.

this is possible through the argument 'scale', see inla.doc("gaussian")

inla(..., scale=1/w, ...)

make a small simulated example to verify that you get it right...

Best
H

On Thu, 2021-09-09 at 08:37 -0700, 'Will Fincham' via R-inla discussion

> --
> 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/17d3d9f6-1a88-42f0-907e-014a0c795256n%40googlegroups.com
> .

--
Håvard Rue
he...@r-inla.org

[Will Fincham]

Hi Håvard,

Thanks so much for your help and such a helpful answer!

Using the "...scale=1/w" variable does indeed result in very similar results being obtained by INLA and MCMCglmm. I'd imagine the slight differences are likely due to variation in priors and the INLA vs full MCMC methods. I'll be sure to check my correct implementation of this method further.

Thanks again,

Will 

[Helpdesk]

yes, you would need the same priors to get the same results...

On Fri, 2021-09-17 at 04:45 -0700, 'Will Fincham' via R-inla discussion