Extract coefficients, covariance, likelihood and observation counts
Source:R/api-fit-extractors.R, R/core-cml.R
mfrm_fit_extractors.RdStandard R accessors for native many-facet fits. They retain the estimator's coordinates and limitations; extracting a number does not qualify inference.
Usage
# S3 method for class 'mfrm_fit'
coef(object, type = c("free", "locations", "steps", "slopes", "persons"), ...)
# S3 method for class 'mfrm_cml_fit'
coef(object, type = c("free", "locations", "steps", "slopes", "persons"), ...)
# S3 method for class 'mfrm_comparison_fit'
coef(object, ...)
# S3 method for class 'mfrm_fit'
vcov(object, ...)
# S3 method for class 'mfrm_comparison_fit'
vcov(object, ...)
# S3 method for class 'mfrm_fit'
nobs(object, type = c("likelihood", "persons", "ratings"), ...)
# S3 method for class 'mfrm_cml_fit'
nobs(object, type = c("likelihood", "persons", "ratings"), ...)
# S3 method for class 'mfrm_comparison_fit'
nobs(object, type = c("likelihood", "persons", "ratings"), ...)
# S3 method for class 'mfrm_fit'
logLik(object, ...)
# S3 method for class 'mfrm_comparison_fit'
logLik(object, ...)
# S3 method for class 'mfrm_fit'
AIC(object, ..., k = 2)
# S3 method for class 'mfrm_fit'
BIC(object, ...)
# S3 method for class 'mfrm_cml_fit'
AIC(object, ..., k = 2)
# S3 method for class 'mfrm_cml_fit'
BIC(object, ...)
# S3 method for class 'mfrm_cml_fit'
vcov(object, ...)
# S3 method for class 'mfrm_cml_fit'
logLik(object, ...)Arguments
- object
A native result from
fit_mfrm(). Imported and extended-model results use their documented summaries and model-specific methods.- type
For
coef(),"free"(default) returns the independent fitted coordinates;"locations","steps","slopes"or"persons"returns the corresponding saved primary estimates. Fornobs(),"likelihood"(default) counts Persons for MML/CML and retained response rows for JML;"persons"and"ratings"request those counts explicitly.- ...
Unused for coefficient, covariance, likelihood and count extraction; misspelled arguments are rejected. Multiple-model calls to
AIC()/BIC()are refused: usecompare_mfrm()for supported comparisons with matched observations, estimation and integration settings.- k
Penalty per free parameter for single-fit
AIC(); default 2.
Value
coef() returns a named numeric vector; vcov() a matrix in the
matching default coordinates. MML covariance retains checks,
numerical_checks, information_review and cautions attributes when
supplied by the interval route. logLik() returns a logLik object; nobs() an
integer count; single-fit AIC()/BIC() a numeric value or NA when
unsupported. Existing CML covariance coordinates and numeric values are
unchanged. No accessor silently computes new Person scores.
Details
Coordinates. The default coef(fit) is a named numeric vector aligned
with both dimensions of vcov(fit). These are free coordinates, not every
displayed facet level: anchors and dependent centered coordinates are omitted.
MML/JML names such as Rater[1] identify positions within parameter blocks,
not literal level names. Steps are free step offsets, slopes are log slopes,
and estimated population variance is on the log-variance scale. Population
regression coefficients remain on their fitted design scale. CML retains its
existing conditional-coordinate names. Use type = "locations" for all
saved facet locations with their level labels, and summary(fit) for tables,
readiness, constraints and interpretation. Location signs are the fitted
model's signs; this accessor does not reorient them as severity.
Default confint() intervals use these same coordinates;
GPCM positive-slope intervals require explicit parm = "slopes".
type = "persons" returns saved primary Person estimates: MML EAPs and
JML point estimates are different targets. Person posterior SDs are not
structural standard errors and do not occur in vcov(fit). CML has no
Person estimates. Free-coordinate extraction refuses corrected or penalized
JML and any fit whose primary estimates are nonfinite: finite optimizer
proxies must not replace primary boundary values. Explicit component
extraction preserves the saved infinite or missing primary estimates.
Covariance. MML vcov() reuses the same qualified full joint covariance
as mfrm_facet_intervals(), including its model/design restrictions and
numerical checks. It may reevaluate information and integration, but does
not refit. An ineligible source remains an error. For an otherwise eligible
ordinary fixed-population RSM/PCM fit, information failure returns an NA
matrix with unavailable_reason, checks and cautions; default confint()
retains coefficient estimates and reasons in its diagnostics attribute
while bounds are missing. No regularized inverse is substituted. If even
one dense NA matrix exceeds mfrmr.max_information_bytes, vcov() errors
with that reason; use the smaller target-level mfrm_facet_intervals() result.
Other MML routes retain their separate qualification checks and error rules.
RSM/PCM singleton-category fits can pass the interval-specific information
review while retaining InferenceReady = FALSE; returned cautions and
checks explain the small-sample limitation. No count label alone certifies
or invalidates this covariance.
No JML structural-truth covariance is supplied. CML vcov() reuses saved
conditional information. If CML's finite numerical point passed its
identification, finite-MLE and stationarity checks but information failed,
vcov() returns an NA matrix with unavailable_reason, checks and
cautions attributes. coef() and logLik() still expose the retained
values; default confint() supplies missing bounds with diagnostic reasons.
The warning also limits confidence in point accuracy along weak directions;
no regularized covariance or formal inference is substituted.
Free-coordinate covariance is not the covariance
of a selected component vector or of all displayed levels; use
mfrm_facet_intervals() or confint.mfrm_fit() for supported transformations.
Comparison-only coef() returns the declared contrast estimates, retaining
unavailable rows as NA; its vcov() is available only for native CML
conditional comparisons, with the same row/column order.
Likelihood and counts. logLik() returns the retained MML marginal or
ordinary JML joint log likelihood, with likelihood_basis and ICReason
attributes. Nonunit weights give a weighted objective, not an ordinary
sampling likelihood. The df attribute is the free parameter count only
when the current information-criterion contract and solution checks permit
its use; otherwise it is NA, so AIC() and BIC() are also NA.
parameter_count retains the actual dimension separately. JML and two-family
GPCM do not gain information-criterion eligibility through these accessors.
Corrected JML, penalized objectives and comparison-only fits do not supply
a generic logLik() here. For estimated-population/one-family GPCM MML,
likelihood qualification may reevaluate solution information.
CML retains its conditional log likelihood and free dimension. Its
nobs attribute counts all retained Persons, including extreme totals
contributing zero conditional information; informative_persons records
nonextreme totals separately. This preserves the existing CML method and
is not a new validation of conditional information criteria. Conditional
and marginal likelihoods must not be compared. For MML, BIC uses Persons,
not rating rows or the sum of rating weights. Counts describe the retained
analysis, not omitted input records or an effective sample size under
dependence. Use nobs(fit, type = "ratings") to count retained ratings.
Examples
ratings <- load_mfrmr_data("example_core")
fit <- fit_mfrm(ratings, "Person", c("Rater", "Criterion"), "Score",
method = "MML", quad_points = 31, maxit = 500)
coef(fit, type = "locations") # Labelled facet locations, including constraints
#> "Rater":"R01" "Rater":"R02"
#> -0.18381527 -0.30884777
#> "Rater":"R03" "Rater":"R04"
#> 0.17950268 0.31316037
#> "Criterion":"Accuracy" "Criterion":"Content"
#> 0.23363879 -0.38997303
#> "Criterion":"Language" "Criterion":"Organization"
#> 0.09146368 0.06487055
coef(fit) # Independent coordinates, including steps
#> Rater[1] Rater[2] Rater[3] Criterion[1] Criterion[2] Criterion[3]
#> -0.18381527 -0.30884777 0.17950268 0.23363879 -0.38997303 0.09146368
#> steps[1] steps[2]
#> -1.21981374 -0.05745669
nobs(fit) # Persons contributing to marginal likelihood
#> [1] 48
nobs(fit, type = "ratings") # Retained response rows
#> [1] 768
ll <- logLik(fit)
attr(ll, "ICReason") # Read before interpreting AIC/BIC
#> [1] "Current MML likelihood and solution checks permit information criteria; model comparisons require matched sources."
AIC(fit)
#> [1] 1816.025
# When covariance is qualified, its names match the free coefficients:
# V <- vcov(fit); identical(rownames(V), names(coef(fit)))