Refusals for the refit-based and marginal-likelihood brmsfit methods

These brmsfit methods have frmtmb spellings that do not exist yet. They are defined so that a ported script fails with the reason and the alternative rather than with "could not find function", and they are documented so the reason is findable.

Usage

# S3 method for class 'frmtmb_draws'
loo_moment_match(x, ...)

# S3 method for class 'frmtmb_draws'
loo_subsample(x, ...)

# S3 method for class 'frmtmb_draws'
reloo(x, ...)

# S3 method for class 'frmtmb_draws'
kfold(x, ...)

bridge_sampler(samples, ...)

# S3 method for class 'frmtmb_draws'
bridge_sampler(samples, ...)

bayes_factor(x1, x2, log = FALSE, ...)

# S3 method for class 'frmtmb_draws'
bayes_factor(x1, x2, log = FALSE, ...)

post_prob(x, ..., prior_prob = NULL, model_names = NULL)

# S3 method for class 'frmtmb_draws'
post_prob(x, ..., prior_prob = NULL, model_names = NULL)

Arguments

x, x1, x2, samples, log, prior_prob, model_names, ...

Ignored; these methods always stop. The names are the formals of each generic's owner, loo or bridgesampling, which a method must carry.

Value

These functions never return; they signal an error.

Details

• loo_moment_match(), loo_subsample(), reloo() and kfold() all need to refit the model on modified data. frm_bootstrap() is the resampling machinery frmtmb does have, and AIC() on the maximum-likelihood fits answers the comparison question directly. The generics of these four and their refusals on a fit are frmtmb's (frmtmb::kfold()); these are the methods for draws.

• bridge_sampler(), bayes_factor() and post_prob() are marginal-likelihood quantities. A marginal likelihood is an integral against the PRIOR, so it is undefined under prior = "flat", and even under the default priors the bridge-sampling estimator needs a normalized log-posterior evaluator that the RTMB tape does not expose.

Examples

# \donttest{
if (requireNamespace("tmbstan", quietly = TRUE) &&
    requireNamespace("rstan", quietly = TRUE) &&
    !frmtmb.sample:::tmbstan_build_broken()) {
  set.seed(1)
  dd <- data.frame(x = rnorm(40))
  dd$y <- rnorm(40, 1 + 0.5 * dd$x, 1)
  fit <- frm(bf(y ~ x), family = gaussian(), data = dd)
  ds <- frm_sample(fit, chains = 1, iter = 400, refresh = 0)
  # each refusal names its reason and the replacement
  try(reloo(ds))
  try(bayes_factor(ds, ds))
}
#> frm_sample(): default priors (brms 2.23 defaults; prior = "flat" puts brms's (flat) in every slot)
#>   Intercept          student_t(3, 1.1, 2.5)
#>   sigma              student_t(3, 0, 2.5)  [natural scale]
#>   b                  (flat), as brms leaves slopes
#> Error : reloo() is not implemented for frmtmb draws: it re-runs the sampler once per observation with a high Pareto k, and frm_sample() has no stored program to re-run on modified data. Read loo()'s Pareto k table and treat a bad k as the diagnostic it is (usually many group-level parameters left to the data alone; see the prior section of ?loo), or compare the maximum-likelihood fits with AIC()
#> Error : bayes_factor() is not available for frmtmb draws: it is a ratio of the marginal likelihoods bridge_sampler() would have to estimate, and those are undefined under prior = "flat" and unavailable from the tape. hypothesis() gives the posterior probability of a directional claim, and loo() the predictive comparison
# }