LOO error with binomial mixture model

Interfaces brms
1
  • Operating System: Ubuntu 18.04
  • brms Version: 2.11.6

I am trying to calculate loo for a binomial mixture model but it loo throws the error:

Error in validate_ll(x) : All input values must be finite.

The model is specified as follows:

brm(u | trials(t) ~ 1
            , data = data
            , family =
                  mixture(binomial(link = "identity")
                        , binomial(link = "identity")
                          )
            , prior =
                  c(prior(beta(1, 20), class = Intercept, dpar = mu1)
                  , prior(beta(20, 1), class = Intercept, dpar = mu2))
            , chains = 4
            , cores = 4)

Is this an issue for these mixture models in general? is there any way around it?

Thanks!

2

What are your data u and t? And why link = “identity”? Do you get something sensible for posterior?

3

u are just counts of something happening, t is the total number of observations.

u    t
0    100
2    23
1000 1003

And why link = “identity”?

I though it was easier to specify the concrete priors I wanted that way (to make the model identifiable). But I get similar results without link = “identity”.

4

The posterior estimates are what I’d expect them to be.

5

Similar? Do you mean the error message is not the same as with the identity?

With the identity I would expect that you get probabilities which are less than 0 or larger than 1 which would cause non-finite log likelihood. This is even more likely with just three rows of observations, as then in LOO you have only two rows of observations and two mixture components.

6

Similar? Do you mean the error message is not the same as with the identity?

I mean it’s the same error message if I fit the model as:

model <- brm(bf(u | trials(t) ~ 1)
      , data = df_train_small
      , family =
            mixture(binomial, binomial)
      , prior =
            c(prior(normal(-4, 0.0001),
                    class = Intercept, dpar = mu1)
            , prior(normal(4, 0.2),
                    class = Intercept, dpar = mu2))
      , chains = 4
      , cores = 4)

Here’s the dataset.data-test.csv (5.1 KB)

7

Thanks for the clarification and the data. I hope I can test this some day next week.