Prior predictive check for a fitted SCD model

Resamples the model from the prior alone and compares the implied outcome distribution to the observed data range, the prior-predictive step of the Bayesian workflow. A prior-predictive interval that fails to span the observed data signals a mis-scaled prior.

Usage

prior_predictive_check(model, data = NULL, ndraws = 100)

Arguments

model

A fitted brmsfit.

data

Data holding the outcome column; defaults to the model's data. A different frame must carry every model term, since the replicates are drawn on it so that both sides use the same rows.

ndraws

Prior-predictive draws summarized.

Value

A list with the prior-predictive median and central 99\ the observed outcome range, covers_observed, and note (NULL, or how an ar() term conditions the predictions); NULL if the model cannot be sampled from the prior (for example under improper priors).

Details

With a Gaussian ar() term, brms predicts each occasion conditional on lagged residuals computed from the observed outcomes, so prior predictions are pulled toward the observed series and the coverage check is optimistic; the returned note says so.

Examples

if (interactive()) {
fit <- run_scd_analysis(simulate_scd_data())$model
prior_predictive_check(fit)
}