Compares observed data to replicated datasets drawn from either the
posterior predictive distribution (standard use) or the prior predictive
distribution (when the model was fitted with sample_prior = TRUE).
Posterior predictive check (sample_prior = FALSE, the default): draws
y_rep from p(y_rep | y). Agreement between y and y_rep indicates the
fitted model can reproduce the observed data.
Prior predictive check (sample_prior = TRUE): draws y_rep from
p(y_rep), integrating over the prior without conditioning on the data.
This is a tool for prior elicitation and sanity-checking — verifying that
the prior does not place substantial mass on data values that are impossible
or implausible given domain knowledge, before the data have been seen.
Arguments
- object
A
bayesmaobject. If fitted withsample_prior = TRUE, a prior predictive check is produced; otherwise a posterior predictive check.- type
Character. Plot type passed to
bayesplot::ppc_*. One of"dens_overlay"(default),"hist","stat","scatter","bars"(discrete only),"ecdf_overlay","ribbon".- ndraws
Integer. Number of draws to display. Default
100.- ...
Additional arguments passed to the underlying
bayesplot::ppc_*function.
Prior predictive checks and double-dipping
It is legitimate to use a prior predictive check to verify that a prior is
coherent on the observable scale (e.g. that it does not imply impossible
event counts). It is not legitimate to iteratively adjust priors until
y_rep matches the observed y: doing so smuggles the data into the prior,
inflating posterior confidence. Priors should be specified from external
knowledge, expert elicitation, or independent reference data — not from the
analysis dataset itself.
Examples
if (FALSE) { # \dontrun{
# Posterior predictive check
fit <- bayesma(data, likelihood = "binomial", ...)
pp_check(fit)
pp_check(fit, type = "stat", stat = "mean")
# Prior predictive check — use to verify priors are sensible,
# not to calibrate them against the analysis data
prior_fit <- bayesma(data, likelihood = "binomial", ..., sample_prior = TRUE)
pp_check(prior_fit)
} # }