Construct a fitter-agnostic prior-draws data generator

prior_draws_generator() is the fitter-agnostic analogue of prior_predictive_generator(). It works through the S7 Fitter interface rather than brms-specific functions, so it can be used with LinearRegressionFitter, BrmsFitter, CmdStanFitter, or any custom Fitter.

Usage

prior_draws_generator(
  fitter,
  fit_spec,
  pilot_bundle,
  predictor_generator,
  response = NULL,
  n_draws = NULL
)

Arguments

fitter

An S7 Fitter object (e.g. LinearRegressionFitter).

fit_spec

A list (single-row fit_grid entry) carrying at least formula (a base R formula). For LinearRegressionFitter a formula like y ~ x is expected.

pilot_bundle

A data_bundle list (train, response, etc.) used for the one-time preconditioning fit. The caller is responsible for providing a representative pilot dataset. Must contain a train data.frame whose column names match the design implied by fit_spec$formula.

predictor_generator

Function (data_spec, task_ctx) -> data.frame producing the design matrix of predictors (everything except the response). Must consume the ambient RNG state.

response

Name of the response column. Defaults to pilot_bundle$response, falling back to the LHS of fit_spec$formula, then to "y".

n_draws

Optional integer override for the number of draws to store; if NULL (default), uses the number of draws returned by extract_draws().

Value

A generator function (data_spec, task_ctx) -> data_bundle.

Details

The factory fits the model ONCE on pilot_bundle (provided by the caller), extracts parameter draws via extract_draws(), and stores them. The returned closure, on each call, picks a draw deterministically indexed by task_ctx$rep_idx (wrapped modulo the number of stored draws), uses it as true_params, and forward-simulates y from the supplied predictors.

Forward simulation: the response is drawn from a Gaussian with mean equal to the linear predictor X theta (using the coefficient columns of the draw) and standard deviation equal to the sigma column of the draw (if present, else 1). This is the natural data-generating process for Gaussian linear models — the common case for LinearRegressionFitter.

Limitations (non-brms)

A true model prior is directly accessible only for brms fits (brms::prior_draws()). For other fitters there is no generic "prior_draws" S7 method, so this factory falls back to the draws stored on the pilot fit. For LinearRegressionFitter those are NIG-prior-conditioned posterior draws (the prior is weak by default, prior_precision = 1e-6), which makes this an approximate prior-predictive path concentrated on the pilot's posterior region. Brms users who need full prior-predictive coverage should prefer the brms-specific prior_predictive_generator().

See also

prior_predictive_generator(), forward_sim_generator(), Fitter, LinearRegressionFitter