This vignette outlines a reproducible workflow for:
For a plot-first companion guide, see the separate
mfrmr-visual-diagnostics vignette.
For a faster preliminary run without changing the final analysis target:
quad_points = 7 grid; restore the
prespecified final MML grid (31 points by default) and check quadrature
sensitivity before reportingmethod = "JML" only when its person-parameter
treatment is methodologically appropriate, not merely as a faster
substitute for MMLdiagnose_mfrm(..., residual_pca = "none", diagnostic_mode = "both", fit_df_method = "both")
when the diagnostics will feed a comprehensive summarymfrmr treats MML and JML
differently on purpose.
MML integrates over the person distribution with
Gauss-Hermite quadrature.mml_engine = "direct" (the default) optimizes the
quadrature-based marginal log-likelihood directly.
mml_engine = "em" and "hybrid" provide the
documented EM and EM-warm-start routes for supported RSM/PCM fits.JML is useful for JMLE-oriented comparisons and
analyses that avoid a parametric person distribution. MML
is the package default and supports marginal and fixed-calibration
follow-up when its response-model and population-distribution
assumptions are defensible.For RSM and PCM, diagnostics now expose two
distinct evidence paths:
diagnostic_mode = "legacy" keeps the residual/EAP-based
stack.diagnostic_mode = "marginal_fit" adds the strict
latent-integrated screen.diagnostic_mode = "both" is the safest default when you
want to inspect both views side by side.Strict marginal diagnostics are screening-oriented. Use
summary(diag)$diagnostic_basis to separate the legacy
residual evidence from the strict marginal evidence rather than pooling
them into one decision.
Start with the packaged example_operational dataset. It
is intentionally compact but uses a connected two-rater assignment,
unequal rater workloads, and six planned omissions represented by absent
long-format rows rather than NA or sentinel scores. This
makes the main tutorial closer to an applied rating design without using
empirical records. The same object is also available via
data("mfrmr_example_operational", package = "mfrmr"). A
separate score-free assignment roster lets the pre-fit review identify
those six omissions without guessing which cells should have existed.
example_core remains available as an idealized
complete-crossing example for fast help-page checks.
data("mfrmr_example_operational", package = "mfrmr")
data("mfrmr_example_operational_design", package = "mfrmr")
toy <- mfrmr_example_operational
data_review_toy <- describe_mfrm_data(
data = toy,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
rating_min = 1,
rating_max = 4,
expected_design = mfrmr_example_operational_design
)
data_summary_toy <- summary(data_review_toy)
data_summary_toy$structural_missingness
data_summary_toy$design_connectivity
fit_toy <- fit_mfrm(
data = toy,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
method = "MML",
model = "RSM"
)
diag_toy <- diagnose_mfrm(
fit_toy,
residual_pca = "none",
diagnostic_mode = "both",
fit_df_method = "both"
)
# Fast fit-only summary: this does not compute diagnostics.
fit_summary_toy <- summary(fit_toy, profile = "fit", detail = "brief")
# Comprehensive review, reusing diagnostics already computed above.
facets_summary_toy <- summary(
fit_toy,
profile = "facets",
detail = "brief",
diagnostics = diag_toy
)
res_toy <- facets_summary_toy$results
fit_summary_toy$overview
fit_summary_toy$readiness
fit_summary_toy$data_review
summary(diag_toy)$overview
facets_summary_toy
# Required first fitted-scale figure.
plot(res_toy, type = "wright", preset = "publication", show_ci = TRUE, top_n = Inf)
# Optional FACETS-style ruler. Replace these examples with the study rubric.
rubric_labels <- c(
"1" = "Level 1",
"2" = "Level 2",
"3" = "Level 3",
"4" = "Level 4"
)
plot(
res_toy,
type = "wright",
renderer = "facets",
category_labels = rubric_labels,
show_ci = FALSE,
preset = "publication"
)
# Setting show_ci = TRUE on this FACETS-style ruler is available as a
# deliberate hybrid: FACETS ruler grammar plus mfrmr uncertainty intervals.
# Optional follow-up: Infit on x, measure on y; persons are explicit opt-in.
plot(
res_toy,
type = "fit_pathway",
fit_stat = "Infit",
include_person = TRUE,
top_n_person = 12,
person_labels = "none",
facet_labels = "flagged",
preset = "publication"
)Optimizer code zero is only one numerical signal.
fit_mfrm() also checks the terminal gradient; when the
initial direct or hybrid solution stops with a larger gradient and
reltol <= 1e-9, it runs a bounded sequence of
warm-started polishing stages and retains the best non-worsening
candidate under the recorded selection rule. Inspect
fit_toy$opt$optimizer_polish$Stages when
Numerical is not pass.
maxit is only a computational ceiling. It must not be
increased until a preferred coefficient, fit statistic, or expected
conclusion appears. For a final analysis, prespecify the estimator and
controls and start from the package default maxit = 400
unless the analysis protocol states otherwise. If a run reaches
ConvergenceStatus = "iteration_limit", keep it review-only
and repeat the same data, model, method, anchors, optimizer, tolerance,
and quadrature rule with the next ceiling in a prespecified sequence.
Interpret only a run with Converged = TRUE,
InferenceReady = TRUE, and Numerical = pass.
If two separately ready runs differ materially, investigate numerical
stability instead of selecting the preferred result. Report the
requested ceiling, actual evaluations, convergence reason, and terminal
gradient.
InferenceReady describes this numerical gate only: Data,
Design, Stability, Diagnostics, and Reporting remain separate rows in
fit_summary_toy$readiness. For example, a disconnected
design remains a reporting hold even if its optimizer converges.
Conversely, ready_for_diagnostics_and_reporting_follow_up
means that fitting completed and diagnostic review is the next stage; it
is neither an optimization failure nor a claim that the analysis is
already manuscript-ready.
expected_design is a declared roster, not an inferred
complete crossing. When no roster is available, the
structural-missingness status is "not_declared"; the
package does not label unassigned cells as missing.
The facets profile is FACETS-style organization, not
evidence that FACETS was run and not a numerical-equivalence claim. Its
brief print is selective and does not print person identifiers; full
tables remain in facets_summary_toy$results$tables. For a
report-oriented result set, use profile = "reporting". To
inspect availability without allowing diagnostic computation, use
compute = "never"; requested dependent sections are then
recorded as not_computed:
reporting_summary_toy <- summary(
fit_toy,
profile = "reporting",
diagnostics = diag_toy
)
availability_only <- summary(
fit_toy,
profile = "facets",
compute = "never"
)
availability_only$section_statusAvailability and interpretability are intentionally separate. Review
res_toy$readiness and the InterpretationStatus
columns in res_toy$plot_map before treating an available
plot as a final result. Plots remain available during a numerical, data,
design, or stability review, but they warn and carry
REVIEW ONLY in the returned subtitle and drawn title until
the readiness issue is resolved.
Bias/DIF, residual PCA, and anchor-drift/linking analyses are deliberately not auto-run by any summary profile. They require explicit contrasts, diagnostic settings, or multi-fit designs.
#> Model Method MethodUsed N Persons Facets FacetInteractions
#> 1 RSM MML MML 282 48 2 0
#> InteractionParameters InteractionCells InteractionSparseCells Categories
#> 1 0 0 0 4
#> LogLik AIC BIC Converged InferenceReady
#> 1 -347.239803707285 712.479607414569 745.256771053012 TRUE TRUE
#> Iterations IterationsBasis MMLEngineRequested MMLEngineUsed
#> 1 28 function_evaluations direct direct
#> MMLEngineDetail EMIterations EMConverged
#> 1 Direct marginal likelihood optimization.
#> EMRelativeChange OptimizerMethod OptimizerInitialMethod OptimizerPolished
#> 1 L-BFGS-B L-BFGS-B TRUE
#> OptimizerPolishSucceeded OptimizerPolishStages RequestedReltol
#> 1 TRUE 2 1e-09
#> EffectiveReltol OptimizerFactr OptimizerPgtol
#> 1 1e-13 450.35996273705 1.49011611938477e-08
#> InitialTerminalGradientSupNorm ConvergenceCode ConvergenceBasis
#> 1 0.000693846262430853 0 optimizer_gradient
#> ConvergenceStatus ConvergenceReason ConvergenceSeverity
#> 1 converged tolerance_met pass
#> ConvergenceMessage
#> 1 CONVERGENCE: REL_REDUCTION_OF_F <= FACTR*EPSMCH
#> ConvergenceDetail ReviewableWarning
#> 1 Optimizer returned convergence code 0. FALSE
#> GradientReviewTolerance FunctionEvaluations GradientEvaluations
#> 1 1e-04 28 28
#> TerminalGradientSupNorm TerminalGradientRMS FacetSampleSizeFlag
#> 1 1.58249274520017e-05 8.46761078292423e-06 standard
#> FacetMinLevelN FacetSparseCount ExtremeHighN ExtremeLowN
#> 1 38 0 0 0
#> Observations Persons Facets Categories Subsets ResidualPCA DiagnosticMode
#> 1 282 48 2 4 1 none both
#> Method PrecisionTier MarginalFit
#> 1 MML model_based available
#> Component
#> 1 name
#> 2 dataThe same fit can then move through the recommended first reporting workflow:
report_toy <- mfrm_report(res_toy, style = "qc")
summary(res_toy)$next_actions
summary(report_toy)$overview
# This is a controlled analysis archive, not a deidentified shareable export.
export_dir <- file.path(tempdir(), "mfrmr-workflow-export")
export_toy <- export_mfrm_results(
res_toy,
output_dir = export_dir,
include = c("default", "report"),
overwrite = TRUE,
acknowledge_sensitive = TRUE
)
head(export_toy$written_files)The acknowledgement suppresses the warning only; it does not redact person identifiers, person-level results, local paths, or the complete RDS object. Review every exported file under the study’s data-handling policy before sharing.
#> Priority Area
#> 1 1 Overview
#> 2 2 Triage
#> 3 2 Wright map
#> 4 3 Diagnostics
#> 5 4 Visual diagnostics
#> 6 5 Fit pathway
#> 7 5 Precision
#> 8 6 Reporting
#> 9 11 Tables
#> Action
#> 1 Read the compact results summary.
#> 2 Read the first-screen triage before branching.
#> 3 Create and inspect the required shared-logit scale map.
#> 4 Review diagnostic key warnings before report drafting.
#> 5 Open the QC dashboard after reviewing the Wright map.
#> 6 Review Infit against measure, including selected person rows when useful.
#> 7 Inspect fit, separation, reliability, and ZSTD wording boundaries.
#> 8 Use the reporting checklist as a guide for preparing a manuscript.
#> 9 Create an appendix-ready summary-table bundle.
#> Route
#> 1 summary(res)
#> 2 summary(res)$triage
#> 3 plot(res, type = "wright", preset = "publication", show_ci = TRUE, top_n = Inf)
#> 4 summary(res$diagnostics)$key_warnings
#> 5 plot(res, type = "qc", preset = "publication")
#> 6 plot(res, type = "fit_pathway", fit_stat = "Infit", include_person = TRUE, top_n_person = 12, person_labels = "none", facet_labels = "flagged", preset = "publication")
#> 7 summary(res$components$precision_review)
#> 8 summary(res$components$reporting_checklist)
#> 9 build_summary_table_bundle(res)
#> Reason
#> 1 Confirms input mode, model, method, section status, table coverage, and available figures.
#> 2 Triage orders unavailable, review, information, and OK signals across diagnostics, tables, plots, and reporting outputs.
#> 3 The Wright map is the primary fitted-scale figure: compare person targeting with facet levels and step thresholds before branching into diagnostics.
#> 4 Diagnostic warnings identify the highest-priority fit, precision, residual, or category follow-up checks.
#> 5 The QC dashboard gives a focused follow-up view of fit, residual, and category summaries.
#> 6 This follow-up separates measure uncertainty from fit displacement while keeping person inclusion explicit.
#> 7 Precision review keeps fit-size, standardized fit, and separation evidence in separate reporting categories.
#> 8 Checklist rows identify report-ready, missing, and caveated sections.
#> 9 The bundle exposes table roles, plot readiness, and conservative appendix presets.
#> Style OverallStatus FirstAction ReviewAreas NotComputedAreas CaveatAreas
#> 1 qc review Start with Fit. 1 0 0
#> OptionalAreas UnavailableAreas OkAreas
#> 1 3 0 1
#> SourceInclude
#> 1 fit, diagnostics, tables, precision, reporting, categories, plots
#> Component Format Path Note
#> 1 summary_overview csv mfrmr_results_summary_overview.csv
#> 2 summary_status csv mfrmr_results_summary_status.csv
#> 3 summary_component_index csv mfrmr_results_summary_component_index.csv
#> 4 summary_table_index csv mfrmr_results_summary_table_index.csv
#> 5 summary_plot_map csv mfrmr_results_summary_plot_map.csv
#> 6 summary_triage csv mfrmr_results_summary_triage.csv
#> DataHandling
#> 1 review_before_sharing
#> 2 review_before_sharing
#> 3 review_before_sharing
#> 4 review_before_sharing
#> 5 review_before_sharing
#> 6 review_before_sharingt4_toy <- unexpected_response_table(
fit_toy,
diagnostics = diag_toy,
abs_z_min = 1.5,
prob_max = 0.4,
top_n = 10
)
t12_toy <- fair_average_table(fit_toy, diagnostics = diag_toy)
t13_toy <- bias_interaction_report(
estimate_bias(fit_toy, diag_toy,
facet_a = "Rater", facet_b = "Criterion",
max_iter = 2),
top_n = 10
)
class(summary(t4_toy))
class(summary(t12_toy))
class(summary(t13_toy))
names(plot(t4_toy, draw = FALSE))
names(plot(t12_toy, draw = FALSE))
names(plot(t13_toy, draw = FALSE))
chk_toy <- reporting_checklist(fit_toy, diagnostics = diag_toy)
subset(
chk_toy$checklist,
Section == "Visual Displays",
c("Item", "DraftReady", "NextAction")
)#> Object SummaryClass
#> 1 unexpected_response_table summary.mfrm_bundle
#> 2 fair_average_table summary.mfrm_bundle
#> 3 bias_interaction_report summary.mfrm_bundle
#> Object Components
#> 1 unexpected_response_table name, data
#> 2 fair_average_table name, data
#> 3 bias_interaction_report name, data
#> Item DraftReady
#> 1 Wright map TRUE
#> 2 QC / facet dashboard TRUE
#> 3 Residual PCA visuals FALSE
#> 4 Connectivity / design-matrix visual TRUE
#> 5 Inter-rater / displacement visuals TRUE
#> 6 Strict marginal visuals FALSE
#> 7 Bias / DIF visuals FALSE
#> 8 Precision / information curves TRUE
#> 9 Fit/category visuals TRUE
#> NextAction
#> 1 Include a Wright map when the manuscript benefits from a shared-scale targeting display.
#> 2 Use the dashboard as a first-pass triage view, then move to the specific follow-up plot behind each flag.
#> 3 Run residual PCA if you want scree/loadings visuals for residual-structure follow-up.
#> 4 Use the design-matrix view to support linkage and comparability claims.
#> 5 Use displacement and inter-rater views to localize QC issues after dashboard screening.
#> 6 Treat strict marginal plots as exploratory corroboration screens, then corroborate with design review and legacy diagnostics.
#> 7 Run bias or DIF screening before discussing interaction-level visuals.
#> 8 Use information curves to describe precision across theta when that is the reporting question.
#> 9 Use category curves and fit visuals as local descriptive follow-up after QC screening.For a larger sparse synthetic illustration, use the packaged Study 1 dataset:
fit <- fit_mfrm(
data = ej2021_study1,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
method = "MML",
model = "RSM",
quad_points = 7
)
diag <- diagnose_mfrm(
fit,
residual_pca = "none",
diagnostic_mode = "both",
fit_df_method = "both"
)
summary(fit, profile = "fit", detail = "brief")
summary(diag)
# Keep the final figure flow explicit: fit -> Wright map -> follow-up plots.
s <- summary(fit, profile = "facets", diagnostics = diag)
res <- s$results
s
plot(res, type = "wright", preset = "publication", show_ci = TRUE, top_n = Inf)
# Optional closest FACETS-style asterisk ruler (without mfrmr CI overlays).
plot(res, type = "wright", renderer = "facets",
category_labels = rubric_labels, show_ci = FALSE,
preset = "publication")
plot(
res,
type = "fit_pathway",
fit_stat = "Infit",
include_person = TRUE,
top_n_person = 12,
person_labels = "none",
facet_labels = "flagged",
preset = "publication"
)This full-data figure caps the displayed person layer at 12 and
suppresses routine point labels to keep the first screen legible. The
selected person IDs and every retained facet row remain in
plot(..., draw = FALSE)$data$table; use
person_labels = "all" or facet_labels = "all"
for a point-identification figure.
If you need residual-structure evidence for a final report, you can add residual PCA after the initial diagnostic pass. Treat this as an exploratory screen, not as a standalone unidimensionality test or as a DIMTEST/UNIDIM substitute. In MFRM reporting, a cautious claim should combine global residual fit, element-level fit, residual PCA, and local-dependence screens, for example: “evidence consistent with essential unidimensionality under the specified facet structure.”
For RSM and PCM, the package can now keep
the legacy residual path and the strict marginal path side by side:
fit_rsm_strict <- fit_mfrm(
data = toy,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
method = "MML",
model = "RSM",
quad_points = 7,
maxit = 30
)
diag_rsm_strict <- diagnose_mfrm(
fit_rsm_strict,
diagnostic_mode = "both",
residual_pca = "none"
)
fit_pcm_strict <- fit_mfrm(
data = toy,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
method = "MML",
model = "PCM",
step_facet = "Criterion",
quad_points = 7,
maxit = 30
)
diag_pcm_strict <- diagnose_mfrm(
fit_pcm_strict,
diagnostic_mode = "both",
residual_pca = "none"
)
summary(diag_rsm_strict)$diagnostic_basis[, c("DiagnosticPath", "Status", "Basis")]
summary(diag_pcm_strict)$diagnostic_basis[, c("DiagnosticPath", "Status", "Basis")]When you want a compact simulation-based screening check for the
strict branch, use evaluate_mfrm_diagnostic_screening() on
a small design:
screen_rsm <- evaluate_mfrm_diagnostic_screening(
design = list(person = 18, rater = 3, criterion = 3, assignment = 3),
reps = 1,
scenarios = c("well_specified", "local_dependence"),
model = "RSM",
maxit = 30,
quad_points = 7,
seed = 123
)
screen_pcm <- evaluate_mfrm_diagnostic_screening(
design = list(person = 18, rater = 3, criterion = 3, assignment = 3),
reps = 1,
scenarios = c("well_specified", "step_structure_misspecification"),
model = "PCM",
maxit = 30,
quad_points = 7,
seed = 123
)
screen_rsm$performance_summary[, c("Scenario", "EvaluationUse", "LegacyAnyFlagRate", "StrictAnyFlagRate")]
screen_pcm$performance_summary[, c("Scenario", "EvaluationUse", "LegacySensitivityProxy", "StrictSensitivityProxy", "DeltaStrictMinusLegacyFlagRate")]The same strict branch is now reflected in the reporting router:
chk_rsm_strict <- reporting_checklist(fit_rsm_strict, diagnostics = diag_rsm_strict)
subset(
chk_rsm_strict$checklist,
Section == "Visual Displays" &
Item %in% c("QC / facet dashboard", "Strict marginal visuals", "Precision / information curves"),
c("Item", "Available", "DraftReady", "NextAction")
)pca <- analyze_residual_pca(diag_pca, mode = "both")
plot_residual_pca(pca, mode = "overall", plot_type = "scree")data("mfrmr_example_bias", package = "mfrmr")
bias_df <- mfrmr_example_bias
fit_bias <- fit_mfrm(
bias_df,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
method = "MML",
model = "RSM",
quad_points = 7
)
diag_bias <- diagnose_mfrm(fit_bias, residual_pca = "none")
bias <- estimate_bias(fit_bias, diag_bias, facet_a = "Rater", facet_b = "Criterion")
fixed <- build_fixed_reports(bias)
apa <- build_apa_outputs(fit_bias, diag_bias, bias_results = bias)
mfrm_threshold_profiles()
vis <- build_visual_summaries(fit_bias, diag_bias, threshold_profile = "standard")
vis$warning_map$residual_pca_overallThe same example_bias dataset also carries a
Group variable so DIF-oriented examples can show a non-null
pattern instead of a fully clean result. It can be loaded either with
load_mfrmr_data("example_bias") or
data("mfrmr_example_bias", package = "mfrmr").
spec <- specifications_report(fit, title = "Study run")
data_qc <- data_quality_report(
fit,
data = ej2021_study1,
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score"
)
iter <- estimation_iteration_report(fit, max_iter = 8)
subset_rep <- subset_connectivity_report(fit, diagnostics = diag)
facet_stats <- facet_statistics_report(fit, diagnostics = diag)
cat_structure <- category_structure_report(fit, diagnostics = diag)
cat_curves <- category_curves_report(fit, theta_points = 101)
bias_rep <- bias_interaction_report(bias, top_n = 20)
plot_bias_interaction(bias_rep, plot = "scatter")The package also supports a separate simulation/prediction layer. The key distinction is:
evaluate_mfrm_recovery() checks whether known
generating parameters are recovered under a stated simulation design. It
is the first simulation check to run when you are validating a model
specification or a planned design.evaluate_mfrm_design() and
predict_mfrm_population() are design-level helpers that
summarize expected operating characteristics under an explicit
simulation specification.mfrm_generalizability() and mfrm_d_study()
summarize observed univariate G-study components and analytic D-study
projections. Read IdentificationStatus,
GStatus, and PhiStatus before reporting
projected coefficients; boundary or singular mixed-model fits are
design-identification warnings rather than high-stakes-ready reliability
evidence.predict_mfrm_units() and
sample_mfrm_plausible_values() score future or partially
observed persons under a fixed MML calibration.if (requireNamespace("lme4", quietly = TRUE)) {
gt <- mfrm_generalizability(fit)
gt$coefficients[, c("G", "Phi", "GStatus", "PhiStatus",
"IdentificationStatus")]
ds <- mfrm_d_study(
gt,
data.frame(Rater = c(2, 3, 4), Criterion = 4),
residual_scaling = "sensitivity"
)
ds[, c("n_Rater", "n_Criterion", "ResidualScaling",
"G", "Phi", "GStatus", "PhiStatus", "IdentificationStatus")]
}sim_spec <- build_mfrm_sim_spec(
n_person = 30,
n_rater = 4,
n_criterion = 4,
raters_per_person = 2,
assignment = "rotating"
)
recovery <- suppressWarnings(
evaluate_mfrm_recovery(
sim_spec = sim_spec,
reps = 2,
maxit = 30,
include_diagnostics = TRUE,
diagnostic_fit_df_method = "both",
seed = 2
)
)
summary(recovery)$recovery_summary[, c("ParameterType", "Facet", "RMSE", "Bias")]
plot(recovery, type = "summary", metric = "rmse", draw = FALSE)$data$plot_table
recovery_review <- assess_mfrm_recovery(
recovery,
min_reps = 2,
min_se_available = NULL,
max_mcse_rmse_ratio = NULL,
max_rmse = c(facet = 1, step = 1, default = 1.5),
max_abs_bias = c(default = 0.75)
)
summary(recovery_review)$checklist[, c("Section", "Item", "Status")]
summary(recovery_review)$reading_order
summary(recovery_review)$condition_reporting_notes
summary(recovery_review)$condition_review
summary(recovery_review)$diagnostic_reporting_notes
summary(recovery_review)$diagnostic_review
status_plot <- plot(recovery_review, type = "status", draw = FALSE)
status_plot$data$section_status
status_plot$data$reading_order
metric_plot <- plot(recovery_review, type = "metrics", metric = "rmse", draw = FALSE)
metric_plot$data$plot_table
metric_plot$data$guidance
recovery_bundle <- build_summary_table_bundle(
recovery_review,
appendix_preset = "recommended"
)
recovery_bundle$table_index[, c("Table", "Rows", "Role")]
pred_pop <- predict_mfrm_population(
sim_spec = sim_spec,
reps = 2,
maxit = 30,
seed = 1
)
summary(pred_pop)$forecast[, c("Facet", "MeanSeparation", "McseSeparation")]
keep_people <- unique(toy$Person)[1:18]
toy_mml <- suppressWarnings(
fit_mfrm(
toy[toy$Person %in% keep_people, , drop = FALSE],
person = "Person",
facets = c("Rater", "Criterion"),
score = "Score",
method = "MML",
quad_points = 5,
maxit = 30
)
)
new_units <- data.frame(
Person = c("NEW01", "NEW01"),
Rater = unique(toy$Rater)[1],
Criterion = unique(toy$Criterion)[1:2],
Score = c(2, 3)
)
pred_units <- predict_mfrm_units(toy_mml, new_units, n_draws = 0)
pv_units <- sample_mfrm_plausible_values(toy_mml, new_units, n_draws = 2, seed = 1)
summary(pred_units)$estimates[, c("Person", "Estimate", "Lower", "Upper")]
summary(pv_units)$draw_summary[, c("Person", "Draws", "MeanValue")]For a report or appendix handoff, pass the recovery objects through the same summary-table export route used by the rest of the package:
export_summary_appendix(
list(recovery = recovery, recovery_review = recovery_review),
output_dir = tempdir(),
prefix = "mfrmr_recovery_appendix",
preset = "recommended",
include_html = FALSE,
overwrite = TRUE
)For an initial exploratory run, reps = 2 or another very
small value is useful only to check the data-generating setup and refit
path. For a study report, increase reps, keep the
ADEMP-style metadata in the exported tables, and set substantive
RMSE/Bias thresholds so that assess_mfrm_recovery() can
mark those rows as ok, review, or
concern rather than not_assessed.