Sampling warning errors

[rmcelreath]

Sampling warning errors

[rmcelreath]

That final sample is to compute DIC, is my guess. If so, there's no harm in it.

That said, I get no divergent iterations for the example you provide.

[statwonk]

however when I directly check for divergences I see 0,

> divergent <- get_sampler_params(test@stanfit, inc_warmup=FALSE)[[1]][,'divergent__']
> sum(divergent)
[1] 0

I was lead here by another case where 1 divergence was reported. When I check directly, no divergence. When I look for any sign of the divergence in the stanfit object I can find none.

[abendorf]

I've run into the same issue. Below is code that triggers a warning that there was 1 divergent transition after warmup, but the "divergent" function says that there were none.

library(rethinking)
set.seed(123)
# create some fake data for a few different groups
result <- data.frame()
for (i in 1:20)
{
  rand_temp <- sample.int(10,1)
  temp <- data.frame(group=rep(i, rand_temp*5), sample=rnorm(rand_temp*5, ifelse(i %in% c(1,2,3),3,0), 1))
  print(temp)
  result <- rbind(result, temp)
}

# model specification
model_spec <- alist(
  sample ~ dnorm(mu[group], sigma),
  mu[group] ~ dnorm(a, b),
  sigma ~ dcauchy(0, 2.5),
  a ~ dnorm(0, 5),
  b ~ dcauchy(0, 50)
)

# feed the model to stan
mod_obj <- map2stan(model_spec, data = result, iter=10000, control = list(adapt_delta = 0.99))
divergent(mod_obj)

[rmcelreath]

This is just the DIC sample taken after real sampling finishes. The way DIC is calculated is to compute the posterior mean and then pass it back into the model to get the deviance.

In the Experimental branch, DIC is disabled by default, I think. I have also phased it out in the new ulam() tool.

[abendorf]

Oh, I see. Thank you very much for clearing that up! And thanks for this great package.

(By the way, DIC must not be disabled for the Experimental branch, since I got the same message after installing that version.)