Variance/precisions of structured random effects

[Maria]

Hi,

I am fitting a model with split random effects and separate intercepts (taking account the sampling weights) of a country separated by the sea into 2 separate regions. I was wondering if there were any recent changes in INLA as the output for the variance/ precisions of the structured random effects in the models fitted in April were different from the output in July (using the same codes and data). Below are the codes:

props.w$unstruct <- props.w$struct <- 1:n.area
formula1 = logit.p ~ -1 + factor(Region) +
  f(struct1, model = 'besag', graph = mat1, scale.model = TRUE, param = c(0.5,0.005)) +  
  f(unstruct1, model = 'iid',  param = c(0.5,0.005)) +
  f(struct2, model = 'besag', graph = mat2, scale.model = TRUE, param = c(0.5,0.005)) +  
  f(unstruct2, model = 'iid',  param = c(0.5,0.005))

fit1 <- inla(formula1,
                     family="gaussian", data=props.w,
                     control.predictor = list(compute = TRUE),
                     control.compute = list(dic = TRUE, cpo = TRUE, waic=TRUE),
                     control.family = list(hyper = list(prec = list(
                       initial = log(1), fixed=TRUE))),
                     scale=props.w$logit.prec)
summary(fit1)

# variance region1
m <- fit1$internal.marginals.hyperpar$"Log precision for struct1"
m.var <- inla.tmarginal(function(x) 1/exp(x), m)
inla.zmarginal(m.var)

m1 <- fit1$internal.marginals.hyperpar$"Log precision for unstruct1"
m1.var <- inla.tmarginal(function(x) 1/exp(x), m1)
inla.zmarginal(m1.var)

# variance region2 
m2 <- fit1$internal.marginals.hyperpar$"Log precision for struct2"
m2.var <- inla.tmarginal(function(x) 1/exp(x), m2)
inla.zmarginal(m2.var)

m3 <- fit1$internal.marginals.hyperpar$"Log precision for unstruct2"
m3.var <- inla.tmarginal(function(x) 1/exp(x), m3)
inla.zmarginal(m3.var) 

Thank you in advance for your help.

BR,

Maria

[Helpdesk]

even if you run the same model twice you'll not get exact the same
results unless you turn off parallelisation. the code is developing all
the time, and for inla.mode="experimental", then different
approximations are used.

if you think there is an error, then I would need code & data to rerun
here

> --
> 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/d114f43c-9356-41a8-b01d-fa54f691362bn%40googlegroups.com
> .

--
Håvard Rue
he...@r-inla.org

[Maria]

Hi,

Thank you very much for your feedback. 

I am new in INLA and Bayesian inference. I understand that I will not get the same results if I run the same model twice. I found the recent outputs were more stable compared to the previous output. I was wondering if there were any changes on the way INLA handle the analysis. Please find below the output when running these codes:

props.w$unstruct <- props.w$struct <- 1:n.area

formula.weighted1 = logit.p ~ -1 + factor(Region) +

  f(struct1, model = 'besag', graph = mat1, scale.model = TRUE, param = c(0.5,0.005)) +  
  f(unstruct1, model = 'iid',  param = c(0.5,0.005)) +
  f(struct2, model = 'besag', graph = mat2, scale.model = TRUE, param = c(0.5,0.005)) +  
  f(unstruct2, model = 'iid',  param = c(0.5,0.005))

fit.weighted1 <- inla(formula.weighted1,

                     family="gaussian", data=props.w,
                     control.predictor = list(compute = TRUE),
                     control.compute = list(dic = TRUE, cpo = TRUE, waic=TRUE),
                     control.family = list(hyper = list(prec = list(
                       initial = log(1), fixed=TRUE))),
                     scale=props.w$logit.prec)

summary(fit.weighted1)

Thank you in advance for your kind help. I really appreciate it.

BR,

Maria