Posterior_predict() in brms

Interfaces brms
1

I have the fitted brms model below:

brms_example = brm(formula = prev_gpa ~ maternal_edu + gender + (1|race_eth) + (1|school),
                   prior = c(
                     prior(normal(0, .25), class=b),
                     prior(normal(2.7,.25), class=Intercept),
                     prior(normal(0, .25), class = sd),
                     prior(normal(0, .25), class = sigma)
                   ),
                   data = dat,
                   chains = 1,
                   core = 1,
                   control = list(adapt_delta = 0.99)
)

The covariates in the dataset look like this:

> head(dat)
  prev_gpa maternal_edu gender race_eth school
1 3.661914            0      1        5  11101
2 2.789677            1      0        2  01329
3 3.221341            0      1        4  10625
4 2.644177            0      1        3  07923
5 4.094903            1      0        3  10442
6 3.653299            1      1        2  06654

The dataset dat used to fit brms_example only has 68 unique schools. Suppose the school level 99978 is not in the sample dat, and I want to do:

posterior_predict(brms_example, newdata = data.frame(maternal_edu=1,gender=1,race_eth=5,school="99978"), allow_new_levels=TRUE)

What does this actually return? My intuition says that
data.frame(maternal_edu=1,gender=1,race_eth=5,school="x") is generated each time where x is a sample from the 68 schools in dat, then a sample from the posterior predictive distribution is taken for this new datapoint. This is repeated 1000 times.

  • Operating System: 10.13.6
  • brms Version: 2.12.0
2

The argument that controls that is called sample_new_levels and is documented in prepare_predictions (formely extract_draws).

1 Like