Log_lik fails for bernoulli intialized with one data point

Interfaces brms
1

The log_lik function does not seem to work if the Bernoulli model is created with data that only has one data point. This is the case even if I only sample the prior. Is this expected?

The use case is I would actually like to fit the model with a single data point, but then get the posterior likelihood for new theoretical data. Is there a way to specify the levels of the Bernoulli response without having all present in the data?

library(brms)

# this works
dat <- data.frame(x = 0:1, y = 0:1)

priors = c(
  prior(normal(0,1), class = "Intercept"),
  prior(normal(0,1), class = "b")
)
model <- brm(
  y ~ x,
  family = bernoulli(),
  data = dat,
  prior = priors,
  sample_prior = "only"
)
model

new_dat <- data.frame(x = 0:1, y = 0:1)

ll <- log_lik(model, newdata = new_dat)

# this doesn't
dat <- data.frame(x = 1, y = 1)

model <- brm(
  y ~ x,
  family = bernoulli(),
  data = dat,
  prior = priors,
  sample_prior = "only"
)
model

new_dat <- data.frame(x = 0:1, y = 0:1)

ll <- log_lik(model, newdata = new_dat)
  • Operating System: Ubuntu 20.04.4
  • brms Version: 2.16.3
1 Like
2

I am able to work around this by including dummy data points and setting their weights to 0.

1 Like