Predict Grid Response from Posterior Draws (Bayesian)

For each grid point, evaluates the forward model at every posterior draw, adds observation noise, then back-calculates concentration to produce a precision profile.

Usage

predict_grid_bayes(
  grid,
  bayes_fit,
  curve_idx = 1L,
  n_draws = NULL,
  cv_x_max = 150,
  pcov_threshold = 20,
  is_log_x = TRUE,
  is_log_response = TRUE
)

Arguments

grid

Data frame from curveRcore::generate_prediction_grid().

bayes_fit

Output of fit_bayes_single().

curve_idx

Integer. Which curve (1-based Stan index).

n_draws

Integer or NULL. Subsample this many draws.

cv_x_max

Numeric. Cap for pcov/pcov_rmse. Default 150.

pcov_threshold

Numeric. Percent CV threshold for pcov_pass. Default 20.

is_log_x

Logical. Default TRUE.

is_log_response

Logical. Whether the response is log10-transformed. Passed to curveRcore::enrich_grid_with_d2y() for second-derivative enrichment. Default TRUE.

Value

grid with added columns: predicted_response, ci_lower, ci_upper, predicted_concentration, se_concentration, pcov, pcov_rmse, pcov_pass, noise_mode.

Details

When the model was fitted with use_heteroscedastic_noise = TRUE, the noise injected at Step 2 scales with the predicted response magnitude (sigma_i = exp(log_sigma0 + log_sigma_slope * log(|mu_i|))), giving the O'Malley (2008) CDAN precision profile. When use_heteroscedastic_noise = FALSE, a constant sigma_obs is used and the profile reflects posterior-predictive uncertainty.