Explicit normal-approximation intervals for testlet calibration

Usage

# S3 method for class 'mfrm_testlet'
confint(object, parm = "calibration", level = 0.95, ...)

Arguments

object: A result from fit_mfrm_testlet().
parm: "calibration" returns all fixed-facet and step intervals. Variance components and Person abilities are not included.
level: Nominal confidence level between zero and one; default 0.95.
...: Unused.

Value

A matrix with Lower and Upper, one row per fixed-facet level or step, and attributes level, method, target, note, checks and diagnostics. attr(ci, "diagnostics") retains each estimate, SE, IntervalStatus, CIEligible and plain-language InferenceReview; checks retains the source checks without recomputing them. Save the whole object with saveRDS(); for CSV, export the diagnostics table to retain reasons.

Details

The explicit request computes estimate plus/minus a normal quantile times the saved observed-information SE. It needs no fitting, scoring or live optimizer and does not change the source fit. Nominal finite-sample coverage is not established. Estimated variance boundaries, unresolved numerical/information checks and invalid SEs retain missing bounds. No regular variance interval, simultaneous comparison or automatic rater classification is supplied. Default fit tables, summaries and plots omit these bounds. Use summary(fit, calibration_intervals = "normal"), plot(fit, intervals = "normal"), or mfrm_results(fit, calibration_intervals = "normal") to select the same approximation in those outputs; their level arguments retain its nominal interpretation. predict.mfrm_testlet() supplies the separate conditional Person-scoring intervals.

Reading missing bounds

estimated_variance_boundary means an estimated variance is at zero; regular normal inference is withheld even when a point is retained. numerical_review_failed means numerical readiness or positive information has not been established; InferenceReview retains all detected reasons, including any simultaneous boundary. Increasing iterations alone does not repair a variance-boundary restriction. Inspect attr(ci, "checks") first. invalid_estimate_or_se and nonfinite_bounds identify row-specific calculation failures. Missing bounds are neither zero uncertainty nor infinite endpoints. Retaining a point does not certify its reliability. Default calibration tables also distinguish not_requested from unsupported_target (population SDs/variances). CIEligible is true only for a returned normal interval; it is not a coverage guarantee.