Fit a Bayesian Location-Scale Model with Saturated Location and Shared Scale

Fits a brms Gaussian location-scale model where the location uses saturated cell means (one coefficient per level of cell_col) to absorb ALL systematic variation in the response, and the scale estimates the power-law relationship between a per-observation precision index and residual variance, shared across all cells.

Usage

fit_saturated_weight(
  df,
  cell_col = "cell",
  concentration_col = "predicted_concentration",
  se_col = "se_concentration",
  pcov_col = "pcov",
  plate_col = "plate",
  predictor = c("auto", "se", "pcov", "se_over_conc"),
  response_is_log10 = FALSE,
  prior_gamma0 = brms::set_prior("normal(0, 1)", class = "Intercept", dpar = "sigma"),
  prior_gamma1 = brms::set_prior("normal(1, 0.5)", class = "b", dpar = "sigma"),
  prior_location = brms::set_prior("normal(0, 2)", class = "b"),
  prior_plate_sd = brms::set_prior("normal(0, 0.5)", class = "sd"),
  iter = 4000,
  warmup = 1000,
  chains = 4,
  cores = 4,
  adapt_delta = 0.95,
  seed = 42
)

Arguments

df

Data frame with observation-level data.

cell_col

Character: name of the saturated cell-means factor column. Create this externally, e.g. df$cell <- interaction(df$Arm, df$Timeperiod, drop = TRUE), or let as_weight_data() build it for you.

concentration_col

Character: predicted concentration column name.

se_col

Character: SE of concentration column name.

pcov_col

Character: posterior CV column name. NULL disables it.

plate_col

Character: plate column name for random intercept. NULL = no plate random effect.

predictor

Scale predictor, passed to prepare_cv(). One of "auto" (default; legacy behaviour), "se" (recommended for curveR ecosystem input), "pcov", "se_over_conc".

response_is_log10

Logical, passed to prepare_cv(): is concentration_col already on the log10 scale (as in a curveR calibration_result)? Default FALSE.

prior_gamma0, prior_gamma1, prior_location, prior_plate_sd

brms::prior objects.

iter, warmup, chains, cores, adapt_delta, seed

MCMC controls.

Value

Named list (see original documentation): fit, phi, beta1, gamma_0, gamma_1, credible intervals, interpretation, effective_se_power, data, diagnostics, cv_diagnostics, formula, priors_used.

Details

The location model uses yi ~ 0 + cell [+ (1|plate)] (saturated cell means). The scale model uses log(sigma) = gamma_0 + gamma_1 * log(cv).

Default prior on gamma_0 is Normal(0, 1), centering phi = exp(gamma_0) at 1. Default prior on gamma_1 is Normal(1, 0.5), centering beta1 at the delta-method prediction of 1, with a 95 percent interval from 0 to 2.

See also

fit_precision_weights() for the curveR-ecosystem entry point that wraps this estimator, as_weight_data(), prepare_cv(), diagnose_cv(), weight_diagnostics().