Predict category probabilities for observed or replacement raters

Evaluate a shared-rater RSM at explicitly supplied ability values.

Usage

# S3 method for class 'mfrm_random_rater'
predict(
  object,
  newdata,
  ability,
  rater = c("observed", "new"),
  quad_points = 41L,
  ...
)

Arguments

object: A result from fit_mfrm_random_rater().
newdata: Data frame with the fitted rater column and every fixed-facet column. Observed-rater labels must be known; replacement-rater labels must be new. Fixed-facet levels must always be known.
ability: A finite numeric value for all rows, one value per row, or the name of a numeric column of newdata. These are specified abilities on the fitted logit scale (mean ability zero, Rasch slope one), not estimated person scores. A value of one is one logit, not one population SD.
rater: "observed" uses the approximate conditional distribution of an observed rater given the calibration responses. "new" integrates a replacement rater from the estimated population distribution. The target is explicit; an unfamiliar label is never silently treated as observed.
quad_points: Normal quadrature points for the prediction integral, from 7 to 121; default 41. Compare orders when needed.
...: Unused.

Value

A list with probabilities (rows by categories), expected_scores, newdata, ability, settings and matching source metadata for mfrm_results(). No model is refitted. Save this list together with the fitted model and prediction inputs. The settings record the effect-sharing unit and whether integration uses the observed rater's conditional distribution or the replacement-rater population. These are preserved in the prediction settings table of mfrm_results().

Details

Predictions condition on estimated calibration and supplied ability. They do not propagate calibration-estimation uncertainty or uncertainty about ability. Observed-rater integration uses the conditional Laplace mean/mode and covariance, not the calibration-adjusted prediction SE. New-rater integration uses the population SD: substituting severity zero generally produces different probabilities.

Rows are marginal probabilities for individual future ratings, not a joint distribution or an interval for their average. Ratings with the same new rater ID share one random severity in the model; multiplying these row probabilities would discard that dependence. This function does not score latent abilities from responses or impute missing assigned scores. Use score_mfrm_random_rater() for conditional Person scoring from a complete joint response roster.