Package 'magmaanlab'

Description

Composes exact empirical first-stage influence with the local correlation ML residual metric for global tests, or observed sensitivity and exact parameter restrictions for nested tests. Reports All, SB and PEBA4 references. Ordinary inference remains unavailable.

Usage

association_ml_global_test(fit, ordinal_stats)
association_ml_nested_test(fit_alt, fit_null, ordinal_stats)

Arguments

Value

A list with availability and a typed reason, statistic, degrees of freedom, spectrum, three references and reconstruction matrices. Zero degrees of freedom has reason zero_df and undefined p-values. Moment nesting, penalties, boundaries and unsupported Stage-1 channels are refused.

Description

Lab-only evaluation-point inference for the Gaussian discrepancy on estimated polychoric correlations. Uses empirical-Jacobian first-stage rows, observed sensitivity including correlation-map curvature, and fixed-allocation group sampling. Ordinary inference remains unavailable.

Usage

association_ml_ij(fit, ordinal_stats)

Arguments

Details

Requires complete all-ordinal rows, saturated thresholds, fixed response scales, linear association constraints, an identified model, positive observed sensitivity and interior primitive covariance matrices. Missing rows, mixed indicators, nonlinear constraints, covariance boundaries and penalties are unsupported. Sensitivity is computed by central differences of the analytic score; no global, nested or modification-index reference law is supplied.

Value

A list with available; unavailable results include reason and optionally detail. Available results contain the objective value, active score, observed H, meat B, coordinate map K, group target derivatives D, vcov_active, full vcov, and group lists influence_active and influence. Influence rows are unscaled: summing their cross-products and dividing by total sample size squared reconstructs the covariance. Full rows include saturated thresholds and their cross-covariance with association parameters.

Description

One-dimensional score tests using observed sensitivity for nuisance projection and exact empirical Stage-1 influence for the score variance.

Usage

association_ml_modification_indices(fit, ordinal_stats, absent = TRUE)
association_ml_score_tests(fit, ordinal_stats)

Arguments

Details

The robust statistic is N times the squared efficient score divided by its empirical meat, referenced to chi-square with one degree of freedom. EPC uses the observed Schur sensitivity. The unscaled mi uses that same sensitivity. Equality rows index the original affine constraint system; EPC is the coefficient along its normalized release direction, with epc_full giving the full parameter displacement. Fixed-row EPC is in the candidate parameter's units. No joint multi-direction release is exposed. Threshold, mean and response-scale releases, singular information, covariance boundaries, penalties, missing data and mixed indicators remain unavailable. Ordinary association inference remains guarded pending policy calibration.

Value

A list with available, metric = "observed" and rows. Each row retains availability or a typed reason. Available rows contain statistic, df, p_value, epc, epc_full, mi and diagnostic score, sensitivity, meat and coordinate matrices.

Description

Constructs parameter bounds in model free-parameter order. Sample-based bounds use the supplied data summaries.

Usage

bounds_loading(model, data)

Arguments

Value

A list of lower and upper parameter-bound vectors.

Description

Constructs parameter bounds in model free-parameter order. Sample-based bounds use the supplied data summaries.

Usage

bounds_pos_var(model)

Arguments

Value

A list of lower and upper parameter-bound vectors.

Description

Constructs parameter bounds in model free-parameter order. Sample-based bounds use the supplied data summaries.

Usage

bounds_standard(model, data)

Arguments

Value

A list of lower and upper parameter-bound vectors.

Description

Constructs parameter bounds in model free-parameter order. Sample-based bounds use the supplied data summaries.

Usage

bounds_variance(model)

Arguments

Value

A list of lower and upper parameter-bound vectors.

Description

Constructs parameter bounds in model free-parameter order. Sample-based bounds use the supplied data summaries.

Usage

bounds_wide(model, data)

Arguments

Value

A list of lower and upper parameter-bound vectors.

Description

The exact leave-one-out engine: for each selected case, drop its row, rebuild the sample statistics, and refit (warm-started from the full-sample solution). Mirror of semfindr::lavaan_rerun(). Continuous ML/ULS/GLS.

Usage

case_rerun(fit, data = NULL, to_rerun = NULL, warm_start = TRUE)

Arguments

Details

Single-group fits work directly from the raw data the fit carries. For multiple-group fits, pass the original data frame (the per-group raw blocks on the fit have lost the original row order needed to map a global case index to its group); case ids are then the data-frame rows, as in semfindr.

Value

A magmaan_case_rerun list: rerun (per-case refit fits, NULL when a refit fails or is inadmissible), fit, selected, case_id, converged.

Description

Computes composite weights and uncertainty from a fit and an explicitly supplied parameter covariance.

Usage

composite_weights(fit, vcov)

Arguments

Value

Native composite-weight results.

Description

Evaluates := expressions and their delta-method uncertainty using an explicit parameter covariance.

Usage

compute_defined(model, fit, vcov)

Arguments

Value

A table of defined parameters and uncertainty summaries.

Description

The scalar rung explicitly frees every non-reference-group latent mean. This is required when observed intercepts are constrained equal: the scalar model is not an affine parameter subset of the metric model, so its robust difference test uses the delta restriction map.

Usage

continuous_invariance(
  model,
  data,
  group = NULL,
  steps = c("configural", "metric", "scalar"),
  estimator = c("ML", "FIML"),
  group_labels = NULL,
  optimizer = "nlopt-lbfgs",
  control = NULL,
  bounds = NULL
)

Arguments

Value

A magmaan_continuous_invariance list with fitted models, robust nested tests, and the constructed specifications.

Description

Composes standard, scaled and robust fit-index families in C++, including a matching independence baseline, noncentral chi-square intervals and close-fit p-values. Estimates are retained.

Usage

convention_fit_measures(fit, convention)

Arguments

Value

A data frame with index, estimate and reason columns and a lavaan_compat attribute. Incompatible bundles error. FIML ML/MLR use saturated H1 EM parameter tolerance 1e-10.

Description

Thin adapters to C++ compatibility composers. A convention changes the reported inference without changing fitting. Components outside the checked model/data slice return an unavailable reason. Complete mixed DWLS supports "WLSMV": expected-bread NACOV covariance, scaled-shifted global and Satorra-2000 delta nested tests. Missing mixed observations and other mixed bundles are unsupported. Delta/theta reporting is point-gated; grouped/theta retained endpoints have limited validation due to optimizer stopping differences.

Usage

convention_inference(fit, convention)

convention_nested(fit_H1, fit_H0, convention)

Arguments

Value

A list with covariance and test results, component availability and convergence metadata. Nested inference returns one difference test.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim(design = convergence_sim_catalog()$design, ...)

Arguments

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim_catalog()

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim_dejonckere_crossloading(n = 20L, variant = c("2025", "2022"), fit_model = c("correct", "omit_lambda8",
    "omit_lambda8_lambda9", "omit_all_crossloadings"), beta = NULL, seed = NULL)

Arguments

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim_dejonckere_msst(n = 10L, state_loadings = c(1, 0.5, 1.3), trait_loadings = c(1, 0.5, 1.3),
    measurement_error = c(0.6, 0.25, 0.7), trait_var = 0.6, state_residual_var = c(0.6,
        0.25, 0.7), seed = NULL)

Arguments

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim_dejonckere_shrinkage(n = 20L, study = c("study1", "study2"), indicators = NULL, beta = NULL,
    seed = NULL)

Arguments

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim_dejonckere_simple(n = 20L, variant = c("2025", "2022"), beta = NULL, seed = NULL)

Arguments

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Provides a small reproducible convergence-study design. The catalog lists published design names and defaults; generators return data and model information without running a calibration study.

Usage

convergence_sim_ludtke_cfa(n = 30L, rho = 0.3, loading = 0.5, seed = NULL)

Arguments

Value

The design catalog, or generated data and model metadata for the selected design.

Description

Converts observations into native sample moments or raw-data input for the corresponding fitting route.

Usage

df_to_data(x, model, group = NULL, missing = c("listwise", "error"), scaling = c("n",
    "n-1"))

Arguments

Value

A prepared data list with observed-variable and group ordering.

Description

Converts observations into native sample moments or raw-data input for the corresponding fitting route.

Usage

df_to_fcsem_data(x, model, missing = c("listwise", "error"), scaling = c("n", "n-1"))

Arguments

Value

A prepared data list with observed-variable and group ordering.

Description

Converts observations into native sample moments or raw-data input for the corresponding fitting route.

Usage

df_to_fiml_data(x, model, group = NULL)

Arguments

Value

A prepared data list with observed-variable and group ordering.

Description

Translate explicit EQS covariance-model sections into the shared magmaan model. Estimation and inference use magmaan, not the EQS application.

Usage

eqs_model(syntax, observed_names = NULL)

Arguments

Details

This initial subset supports single-group continuous covariance models. Every equation has one distinct error variable with a fixed unit coefficient: E for observed variables and D for factors. Every factor must have an observed indicator. Error numbering need not match equation numbering. A coefficient followed by * is free; a preceding number supplies its start. A coefficient without * is fixed (implicit coefficient one). Missing independent variances are free and missing covariances are zero. Identification follows explicit EQS specifications; no marker loading or single-indicator residual fix is added. Free parameters without supplied starts use magmaan's start policy, not EQS's internal start algorithm.

Job files, /MODEL, /LABELS, constraints, means (V999), multiple groups, shared errors, nonunit error paths and error/predictor or E/D covariances are rejected. Observed indicators participating in structural regressions are also outside this subset; separate observed predictors and purely observed path models are supported. Supply data and estimator separately. Scientific exponents must have a sign, e.g. 1e-3, to distinguish 1E1, which is a unit coefficient on error E1.

Value

A magmaan_model_spec for fit_model() and prepare_model(). eqs_source retains the original input; syntax is its explicit lavaan projection for rebuilding and refitting.

Examples

model <- eqs_model("
/EQUATIONS
V1 = *F1 + E1;
V2 = .7*F1 + E2;
V3 = .8*F1 + E3;
/VARIANCES
F1 = 1;
E1-E3 = .5*;
", observed_names = c("x1", "x2", "x3"))

Description

Summarizes parameter changes from case_rerun under the selected standard-error convention.

Usage

est_change(rerun_out, parameters = NULL, se = c("standard", "robust.sem", "robust.huber.white"))

Arguments

Value

A parameter-change results table.

Note

The SE/type defaults standardize SAM or case-change diagnostics. They do not assert a generic misspecification-robust inference law; explicit robust or estimated-weight choices remain available within their supported scope.

Description

The one-step approximation of est_change() without refitting: the approximate raw change ⁠Δ ≈ (N/(N-1))·V·s_i⁠ standardized by the full-sample SE, plus an approximate generalized Cook's distance ⁠gcd_approx = Δ' V_sel^{-1} Δ⁠.

Usage

est_change_approx(
  fit,
  parameters = NULL,
  type = c("standard", "estimated.weight")
)

Arguments

Details

This deliberately uses the correct finite-sample scaling, which differs from semfindr::est_change_approx() (v0.2.0) by two known constant factors: semfindr applies N/(N-1) twice to DFTHETAS (one too many — the factor is already in the raw one-step change) and applies it only once inside gCD (one too few). With matched columns, est_change_approx(magmaan) relates to semfindr by ⁠dftheta_magmaan = dftheta_semfindr · (N-1)/N⁠ and ⁠gcd_magmaan = gcd_semfindr · N/(N-1)⁠. Both factors are O(1/N) and immaterial relative to the one-step error, but they have no first-principles basis (see the influence-function derivation; tracked as an upstream-PR item). The raw est_change_raw_approx() is unaffected (semfindr is correct there).

Value

A matrix with one column per selected parameter plus gcd_approx, one row per case.

Note

The SE/type defaults standardize SAM or case-change diagnostics. They do not assert a generic misspecification-robust inference law; explicit robust or estimated-weight choices remain available within their supported scope.

Description

For each case, ⁠(estimate with all cases) - (estimate without this case)⁠. With standardized = FALSE (default) these are raw coefficient changes; with standardized = TRUE they are changes in the standardized (std.all) solution. The change is NOT divided by a standard error (see est_change()). Mirror of semfindr::est_change_raw().

Usage

est_change_raw(rerun_out, parameters = NULL, standardized = FALSE)

Arguments

Value

A matrix, one row per case, one column per selected parameter, rownames = case ids.

Note

The SE/type defaults standardize SAM or case-change diagnostics. They do not assert a generic misspecification-robust inference law; explicit robust or estimated-weight choices remain available within their supported scope.

Description

⁠(N/(N-1))·V·s_i⁠ per case, the one-step approximation of est_change_raw() without refitting. Mirror of semfindr::est_change_raw_approx().

Usage

est_change_raw_approx(
  fit,
  parameters = NULL,
  type = c("standard", "estimated.weight")
)

Arguments

Details

With type = "estimated.weight" this is the misspecification-robust ("complete-sandwich") one-step change for a moment-quadratic fit whose weight is estimated from the data (continuous GLS/WLS, ordinal DWLS/WLSMV; ULS is fixed-weight and so unchanged): the per-case influence carries the data-dependent-weight term that semfindr / Pek & MacCallum drop by treating the estimator weight as fixed. That term is O_p(N^{-1}) under a correct model but O_p(N^{-1/2}) under misspecification (Hall-Inoue order promotion), so the two regimes agree at the null and diverge under misfit; it is the casewise dual of the estimated-weight standard error. The result then carries two attributes: "naive" (the fixed-weight one-step change) and "weight_diagnostic" (their difference, the per-case ⁠Δ'W'_d⁠ term). There is no semfindr counterpart; this is a frontier extension.

Both regimes support single- and multiple-group fits; multiple-group rows are block-stacked (⁠g{b}_{row}⁠ ids, group-suffixed columns).

Value

A matrix, one row per case, one column per selected parameter (plus the "naive" / "weight_diagnostic" attributes when type = "estimated.weight").

Note

The SE/type defaults standardize SAM or case-change diagnostics. They do not assert a generic misspecification-robust inference law; explicit robust or estimated-weight choices remain available within their supported scope.

Description

Computes precision diagnostics for factor scores from a fit and complete raw observations.

Usage

factor_score_precision(fit, data)

Arguments

Value

Native factor-score precision diagnostics.

Description

Computes factor scores from raw observations. The default is regression for continuous fits and EBM for ordinal fits; explicit choices include regression, bartlett, EBM, ML and EAP.

Usage

factor_scores(fit, data, method = NULL)

Arguments

Value

Factor-score estimates returned by the native scoring operation.

Description

Constructs, fits or summarizes the native factor-component SEM representation. These objects use their own fitting route.

Usage

fcsem_fit_measures(fit)

Arguments

Value

A factor-component model specification, fitted model or named post-fit summary, according to the operation.

Description

Constructs, fits or summarizes the native factor-component SEM representation. These objects use their own fitting route.

Usage

fcsem_model_spec(syntax, ...)

Arguments

Value

A factor-component model specification, fitted model or named post-fit summary, according to the operation.

Description

Constructs, fits or summarizes the native factor-component SEM representation. These objects use their own fitting route.

Usage

fcsem_standard_errors(fit)

Arguments

Value

A factor-component model specification, fitted model or named post-fit summary, according to the operation.

Description

Constructs, fits or summarizes the native factor-component SEM representation. These objects use their own fitting route.

Usage

fcsem_standardized_rows(fit, vcov = NULL)

Arguments

Value

A factor-component model specification, fitted model or named post-fit summary, according to the operation.

Description

Lab-only Wald intervals for continuous complete-data ML and FIML, including independent groups, random x and affine equality constraints. Points agree with policy_fit_measures(). The variance candidate has not been adopted or coverage-calibrated.

Usage

fit_index_intervals_misspec(fit, level = 0.90)

Arguments

Details

The casewise likelihood contrast gives the first-order user discrepancy variance V1. The second-order V2 is $2\mathrm{tr}(M^2)/N^2$, where $M=(J_s^{-1}-\Delta J_0^{-1}\Delta^\top)K_s$ uses saturated observed information, constraint-reduced model observed information, the implied-moment Jacobian and the saturated score outer product. Covariance-only ML includes its profiled group-mean nuisance parameters.

wald_f maps a Wald interval for the corrected discrepancy to RMSEA. wald_sqrt uses its square-root coordinate and requires a positive corrected discrepancy. CFI and TLI use the joint user/baseline first-order delta method plus V2 for the user discrepancy; their intervals coincide for both variants. CFI endpoints are clipped to [0,1]; TLI endpoints are unbounded. RMSEA lower bounds are clipped at zero. Ordinal, ML2S, other estimators, clustered, penalized and failed fits return typed unavailable intervals.

Value

A data frame with index, variant, estimate, lower, upper, V1, V2 and reason. V1/V2 are always on the user discrepancy scale, before mapping to an index. The details attribute includes casewise and fit discrepancies, their corrected point, Takeuchi and quadratic traces, sample size and level. Unavailable intervals have NA bounds and explicit reasons.

See Also

policy_fit_measures

Description

Convenience wrapper over the existing fit-measure primitives. By default it reports the standard ML chi-square, df, p-value, baseline chi-square, and ordinary fit measures. Categorical models report the native n F statistic; lavaan's (n - G) F standard reporting is available through convention_inference(). Passing fmg = TRUE or a character vector of FMG test labels adds an ⁠$fmg⁠ diagnostic table from fmg_tests().

Usage

fit_measures(fit, baseline = NULL, fmg = NULL, robust = NULL, data = NULL)

Arguments

Description

For each case, ⁠(fit measure with all cases) - (fit measure without this case)⁠ for each requested measure. Mirror of semfindr::fit_measures_change().

Usage

fit_measures_change(
  rerun_out,
  fit_measures = c("chisq", "cfi", "rmsea", "tli")
)

Arguments

Value

A matrix, one row per case, one column per measure, rownames = case ids.

Description

Estimated-weight (categorical DWLS) misspecification inference for the absolute and incremental fit indices RMSEA, CRMR, SRMR, CFI and TLI, each with a confidence interval that propagates the sampling variability of the estimated polychoric weight (the gamma channel). The single consolidated surface over the per-index entry points. Single- and multi-group all-ordinal DWLS fits only. Exact residual inference includes pooled weights and cross-group parameter influence; the OPG default keeps its legacy blockwise approximation. Residual point fields in this consolidated table remain uncorrected.

Usage

fit_measures_misspec(
  fit,
  ordinal_stats = NULL,
  estimated_weight = TRUE,
  conf_level = 0.9,
  eig_tol = 1e-10,
  first_stage = c("opg", "exact")
)

Arguments

Details

CFI carries a trustworthy interval and is largely robust to weight estimation; TLI's point is calibrated but its interval is conservative and unreliable at strong misfit (the generalized-df ratio is ill-conditioned). See cfi_tli_misspec_inference.tex.

Value

A named list: rmsea/crmr/srmr/cfi/tli with .ci.lower and .ci.upper, the exact-fit rmsea.pvalue and crmr.pvalue, the user and baseline chisq/df, conf.level, estimated.weight, and warnings.

Description

Exact discrepancy points also include the known N-divisor bias of the continuous covariance moments. Polychoric/polyserial bias remains omitted. Exact residual inference uses the full pooled influence map; OPG retains its legacy blockwise approximation and defaults. The consolidated residual point fields remain uncorrected.

Usage

fit_measures_misspec_mixed_ordinal(
  fit,
  mixed_stats = NULL,
  estimated_weight = TRUE,
  conf_level = 0.9,
  eig_tol = 1e-10,
  first_stage = c("opg", "exact")
)

Arguments

Value

A named list with the same fields as fit_measures_misspec().

Description

The user statistic is the residual goodness-of-fit T on df degrees of freedom under the chosen discrepancy, referred to an independence-model baseline under the SAME discrepancy. Naive indices treat T as central chi-square; robust indices use the Satorra-Bentler scaling scale_c (user) and the baseline scaling (independence model) to correct the mixture reference. logl/unrestricted.logl/aic/bic/bic2 are reported as NA: these are likelihood information criteria and are not defined for the closed-form non-ML estimator.

Usage

fit_measures_noniterative(
  fit,
  discrepancy = c("ntml", "uls"),
  gamma = c("nt", "empirical"),
  data = NULL,
  baseline = NULL,
  fmg = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Constructs, fits or summarizes the native factor-component SEM representation. These objects use their own fitting route.

Usage

fit_ml_fcsem(model, data, missing = c("listwise", "error"), control = NULL)

Arguments

Value

A factor-component model specification, fitted model or named post-fit summary, according to the operation.

Description

Fits the requested objective and returns estimates and diagnostics. Inference is composed explicitly after fitting.

Usage

fit_model(model, data, estimator = "ML", groups = NULL, ..., cluster = NULL, ordered = NULL,
    parameterization = "delta", missing = c("listwise", "error", "pairwise"), pd_gamma = c("overlap",
        "nominal"), se = "none", test = "none", W = NULL, optimizer = NULL, control = NULL,
    bounds = NULL, stage2_weight = "nt", dls_a = 0.5, stage1_regularization = NULL,
    psd = FALSE, covariance = NULL, barrier = NULL, weight = NULL, options = NULL)

Arguments

Value

A magmaan_fit list with parameter estimates, fitting options, convergence and admissibility diagnostics.

Deletion provenance

Dataframe helpers, prepared data and fitted lab objects retain a read-only deletion record. Its policy identifies listwise, error, FIML or pairwise row handling; known indicates whether supplied-row counts are available. per_group contains group, rows_supplied, rows_used, rows_deleted, and exclusive reason counts missing_model_values, missing_group_label, unknown_group_label, and missing_cluster_id. Reasons take precedence in that order: group labels, cluster IDs, model values. Rows without an assignable group appear in a separate NA group row. Single-level routes still refuse missing or unknown group labels; two-level listwise preparation deletes missing group or cluster IDs. FIML and ML2S retain incomplete rows without listwise deletion; existing native refusals remain. Summary-only inputs without retained provenance report unknown supplied/deleted counts, rather than claiming that no earlier deletion occurred.

Description

Fit a non-iterative CFA estimator.

Usage

fit_noniterative_cfa(
  partable,
  sample_stats,
  estimator = "guttman_lavaan",
  composite = "auto",
  admissibility = "raw",
  margin = 1e-04,
  beta0 = 1,
  rate = 0.5,
  score_conditioning = "raw",
  score_floor0 = 1,
  score_rate = 0.5,
  h_conditioning = "raw",
  h_floor0 = 1,
  h_rate = 0.5
)

Arguments

Value

A magmaan fit object (same shape as fit_model()), usable by the inference helpers below and by partable inspection.

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

This is not the inference-side Omega projection returned by noniterative_cfa_constrained(). It imposes the metric loading-shape constraint inside the Sigma/H reconstruction itself: a common standardized loading shape is estimated across groups and then converted back to the marker chart represented by partable.

Usage

fit_noniterative_cfa_metric(
  partable,
  sample_stats,
  estimator = "guttman_aligned",
  composite = "auto",
  admissibility = "raw",
  margin = 1e-04,
  beta0 = 1,
  rate = 0.5,
  score_conditioning = "raw",
  score_floor0 = 1,
  score_rate = 0.5,
  h_conditioning = "raw",
  h_floor0 = 1,
  h_rate = 0.5
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Residual-variance equality constraints are imposed in the Guttman H communality step before the loadings are reconstructed. Loading-only linear equalities are then projected with a Sigma-only composite metric. Rows that mix residual and loading parameters, or rows involving factor variances, intercepts, or latent means, are rejected.

Usage

fit_noniterative_cfa_restricted(
  partable,
  sample_stats,
  estimator = "guttman_aligned",
  communality = "triad_wls",
  composite = "auto",
  admissibility = "raw",
  margin = 1e-04,
  beta0 = 1,
  rate = 0.5,
  score_conditioning = "raw",
  score_floor0 = 1,
  score_rate = 0.5,
  h_conditioning = "raw",
  h_floor0 = 1,
  h_rate = 0.5
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Fits the native two-level model using cluster identifiers and optional group identifiers.

Usage

fit_twolevel(model, data, cluster, group = NULL, optimizer = "nlopt-lbfgs", control = NULL,
    bounds = NULL, missing = c("listwise", "error"))

Arguments

Value

A two-level fitted-model result.

Description

Applies the FMG eigenvalue-tail transforms to the Satorra-2000 restriction-map difference spectrum returned by robust_nested_lrt(). For complete-data ML pairs an explicit ⁠_rls⁠ suffix uses the difference between the two Browne RLS residual statistics, evaluated at the ML estimates, as the asymptotically equivalent base statistic. An ⁠_ug⁠ suffix selects the Browne/Du-Bentler finite-sample correction to the nested Gamma spectrum. Unsuffixed tests and ⁠_ml⁠ retain the likelihood-ratio difference. FIML and two-stage ML (ML2S) support only the biased Gamma and ML base.

Usage

fmg_nested(
  fit_H1,
  fit_H0,
  data = NULL,
  tests = NULL,
  A.method = c("exact", "delta"),
  weight = NULL,
  gamma = NULL
)

Arguments

Value

A magmaan_fmg_tests data frame. Numerical diagnostics from the nested-test backend are available in attr(x, "warnings").

Note

This is an FMG/semTests replication-compatibility route. Its NT/fixed-weight conventions reproduce the documented comparator and are distinct from the generic nested policy in robust_nested_lrt().

Description

The nested-model analogue of fmg_tests_ordinal(). The Satorra-2000 reduction of the two polychoric least-squares fits yields a difference triple — the unscaled LS difference statistic $T_{H0} - T_{H1}$, the restriction rank (df_diff), and the difference spectrum (the nested UGamma eigenvalues) — and the same FMG eigenvalue-tail transforms are applied to it, exactly as fmg_tests_ordinal() applies them to a single-model UGamma spectrum.

Usage

fmg_nested_ordinal(
  fit_H1,
  fit_H0,
  ordinal_stats,
  tests = NULL,
  weight = "",
  A.method = c("delta", "exact")
)

fmg_nested_mixed_ordinal(
  fit_H1,
  fit_H0,
  mixed_stats,
  tests = NULL,
  weight = "",
  A.method = c("delta", "exact")
)

Arguments

Details

A.method defaults to "delta": the configural-to-metric ordinal invariance step is non-nested (Wu-Estabrook metric frees the group-2+ scale/intercepts that configural fixes), so the exact parameter-nesting restriction does not apply — the moment-Jacobian column-space (delta) restriction is what lavaan's lavTestLRT(method = "satorra.2000") uses there. (robust_nested_lrt() keeps the "exact" default for nested complete-data/FIML pairs.)

Value

A magmaan_fmg_tests data frame, identical in shape to fmg_tests_ordinal(). Numerical diagnostics from the nested-test backend are available in attr(x, "warnings").

Description

Back-compatible named-vector view over fmg_tests(). New code that needs diagnostics should call fmg_tests() or fit_measures(..., fmg = tests).

Usage

fmg_pvalues(
  fit,
  data = NULL,
  tests = NULL,
  weight = NULL,
  gamma = c("empirical", "normal")
)

Arguments

Value

A named numeric vector of p-values; names are the canonical test labels (e.g. peba4_rls, sb_ug_rls), matching semTests::pvalues().

Description

Computes FMG p-values and diagnostics for a complete-data ML or continuous ULS/GLS/WLS fit, or a FIML (missing-data) fit. Complete-data fits must carry complete raw data, as fits from fit_model(..., data.frame) do, or callers can pass complete raw data explicitly.

Usage

fmg_tests(
  fit,
  tests = NULL,
  data = NULL,
  weight = NULL,
  gamma = c("empirical", "normal")
)

Arguments

Details

FIML fits (fit_fiml() / fit_model(..., estimator = "FIML"), single- or multi-group) are supported: the missing-data UGamma spectrum is computed first-principles from the saturated-model EM ACOV and either saturated or model-implied H1 information, with the FIML LRT as the base statistic. Under FIML only the biased Gamma-hat and the ML base are defined, so ⁠_ug⁠ and ⁠_rls⁠ are rejected (an unsuffixed base resolves to ML). This is a principled construction, not a port of semTests' (unsound) FIML handling.

Two-stage ML (ML2S) fits (fit_ml2s() / fit_model(..., estimator = "ML2S")) are supported the same way: the df-dimensional UGamma spectrum and the Stage-2 ML base chi-square are taken from the two-stage inference already attached to the fit (fit$ml2s). Unsuffixed tests use that ML base; an explicit ⁠_rls⁠ suffix instead uses the full fitted-moment RLS residual quadratic, including restricted means, with the same spectrum. As under FIML, ⁠_ug⁠ is rejected.

Value

A data frame with scalar diagnostics plus list-columns for the raw UGamma spectrum and the method-specific lambda vectors.

Description

Applies the FMG eigenvalue-tail transforms to an ordinal (all-categorical) or mixed continuous/ordinal least-squares fit (DWLS/WLS/ULS over polychoric and polyserial moments). The UGamma spectrum, base chi-square, and df come from the same polychoric-NACOV sandwich as robust_ordinal() / robust_mixed_ordinal(), so the categorical stats are supplied explicitly the same way. Single- and multi-group fits are supported.

Usage

fmg_tests_ordinal(fit, ordinal_stats, tests = NULL, weight = "")

fmg_tests_mixed_ordinal(fit, mixed_stats, tests = NULL, weight = "")

Arguments

Details

An ordinal LS fit has one base statistic, the LS chi-square $N\,F_\mathrm{min}$; there is no ML/RLS split and no Du-Bentler unbiased Gamma, so test names suffixed ⁠_ml⁠ or ⁠_ug⁠ are rejected. As on the complete-data and FIML paths, the pEBA/pOLS/PALL transforms are magmaan constructions with no external oracle.

Value

A magmaan_fmg_tests data frame, identical in shape to fmg_tests().

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_fiml_multiinfo(model, data, eta = 1.25, weight = NULL, optimizer = "port", control = NULL,
    bounds = NULL, target = c("joint", "determinacy"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_fiml_psd(model, data, optimizer = "nlopt-slsqp", control = NULL, start_eigen_floor = 1e-06,
    feasibility_tol = 1e-06, psd_factor = c("triangular", "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_gls_psd(model, data, optimizer = "nlopt-slsqp", control = NULL, start_eigen_floor = 1e-06,
    feasibility_tol = 1e-06, missing = c("listwise", "error"), psd_factor = c("triangular",
        "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_mixed_ordinal_psd(model, data, estimator = c("DWLS", "WLS", "ULS"), optimizer = "nlopt-slsqp", control = NULL,
    bounds = NULL, start_eigen_floor = 1e-06, feasibility_tol = 1e-06, psd_factor = c("triangular",
        "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_ml_multiinfo(model, data, eta = 1.25, weight = NULL, optimizer = "port", control = NULL,
    bounds = NULL, missing = c("listwise", "error"), target = c("joint", "determinacy"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_ml_psd(model, data, optimizer = "nlopt-slsqp", control = NULL, start_eigen_floor = 1e-06,
    feasibility_tol = 1e-06, missing = c("listwise", "error"), preconditioning = c("diagonal",
        "none"), psd_factor = c("triangular", "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits ordinary complete-data normal-theory ML first, accepting it only when both the common fit verdict and covariance admissibility pass. Otherwise tries PSD-constrained ML once. The selection policy is implemented in C++.

Usage

frontier_fit_ml_psd_fallback(model, data, ordinary_optimizer = "nlopt-lbfgs", psd_optimizer = "nlopt-slsqp", ordinary_control = NULL,
    psd_control = NULL, start_eigen_floor = 1e-06, feasibility_tol = 1e-06, missing = c("listwise",
        "error"), preconditioning = c("diagonal", "none"), psd_factor = c("triangular", "square"))

Arguments

Details

A warm start transfers ordinary parameter estimates only, not optimizer memory. The ordinary return must have finite parameters and objective, a positive-definite implied covariance, and satisfied model equalities to be used as a start. It need not pass accuracy or primitive-covariance admissibility. The PSD initializer projects covariance-factor starts; link equalities may initially be violated and are enforced by the solver. After an ordinary error or unusable return, PSD uses the original start.

Both stages must pass their common numerical verdict and covariance admissibility before their result can be selected. This does not certify global optimality or add boundary-aware inference. Ordinary ML defaults are unchanged. Parsing, data preparation, and initial start-construction errors occur before the fitting policy and raise R errors as usual.

Value

A list with fit (the accepted ordinary or PSD fit, otherwise NULL), converged, fallback_used, fallback_reason, and warm_start_used. The reason is "none", "ordinary-error", "ordinary-rejected", or "ordinary-inadmissible"; a failed accuracy verdict takes precedence over inadmissibility when both fail.

ordinary and psd retain each attempt as a list with fit and error, one of which is NULL. An error records kind, detail, iterations, and f_value from the core error value. The entire psd entry is NULL when skipped. Rejected returned estimates remain accessible even when fit is NULL.

Examples

s <- list(S = list(matrix(c(1, .7, .7, .7, 1, .3, .7, .3, 1), 3)),
          nobs = 300L)
result <- frontier_fit_ml_psd_fallback("f =~ x1 + x2 + x3", s)
result$fallback_reason
result$warm_start_used

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_ml2s_psd(model, data, optimizer = "nlopt-slsqp", control = NULL, h_step = 1e-04, stage1 = NULL,
    stage1_regularization = NULL, stage2_weight = "nt", dls_a = 0.5, start_eigen_floor = 1e-06,
    feasibility_tol = 1e-06, psd_factor = c("triangular", "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_ordinal_psd(model, data, estimator = c("DWLS", "WLS", "ULS"), optimizer = "nlopt-slsqp", control = NULL,
    bounds = NULL, start_eigen_floor = 1e-06, feasibility_tol = 1e-06, psd_factor = c("triangular",
        "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits in sphere coordinates with an explicit metric and pinning penalty, then optionally polishes in ordinary parameter coordinates. Inference remains a separate operation.

Usage

frontier_fit_sphere(model, data, estimator = "ML", groups = NULL, ..., cluster = NULL, ordered = NULL,
    parameterization = "delta", psd = FALSE, missing = c("listwise", "error"), W = NULL, optimizer = NULL,
    control = NULL, bounds = NULL, metric = c("unit_free", "raw"), pin_weight = 1, pole_tol = 1e-06,
    polish = TRUE, start = c("canonical", "user"), start_eigen_floor = 1e-06, feasibility_tol = 1e-06,
    preconditioning = c("none", "diagonal"), psd_factor = c("triangular", "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_uls_psd(model, data, optimizer = "nlopt-slsqp", control = NULL, start_eigen_floor = 1e-06,
    feasibility_tol = 1e-06, missing = c("listwise", "error"), psd_factor = c("triangular",
        "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Fits the named objective through an explicit experimental covariance-domain route. PSD variants constrain covariance blocks; multiinfo variants apply the selected barrier target. Inference remains a separate operation.

Usage

frontier_fit_wls_psd(model, data, W, optimizer = "nlopt-slsqp", control = NULL, start_eigen_floor = 1e-06,
    feasibility_tol = 1e-06, missing = c("listwise", "error"), psd_factor = c("triangular",
        "square"))

Arguments

Value

A fitted-model list with parameter estimates and route diagnostics.

Description

Computes the local Newton distance for complete-data ML in units of information-based standard errors. It is a quadratic approximation, not a global-optimality certificate. PSD fits can restrict the step to the active covariance face.

Usage

frontier_newton_accuracy(fit, budget = 0.01, psd = NULL)

Arguments

Value

A diagnostic list with status, distance, predicted gain and whether the requested budget passed.

Description

Projects a sphere-coordinate fitted model into the requested model identification.

Usage

frontier_reidentify(fit, model, ..., pole_tol = 1e-06)

Arguments

Value

A fitted model in the requested identification with projection diagnostics.

Description

Tests a fitted continuous ML, FIML, or normal-theory ML2S model against its local saturated mean/covariance alternative. ML/FIML use direct saturated likelihood scores. ML2S uses the Stage-2 saturated score for the observed gate and propagates Stage-1 EM casewise moment influence through the fitted normal-theory metric for multiplier calibration. In each case df is the saturated moment dimension minus the tangent rank. This is the general goodness-of-fit counterpart to score_flip_test(), which instead releases specified affine restrictions.

Usage

global_score_flip_test(
  fit,
  data = NULL,
  n_flips = 999L,
  seed = 1,
  multiplier = c("rademacher", "mammen", "two-point", "gaussian", "centered-exponential"),
  two_point_skewness = 1,
  center_multiplier_scores = FALSE,
  multiplier_studentization = c("none", "weighted-meat"),
  sensitivity = c("expected", "observed", "observed-shrink-light",
    "observed-shrink-sqrt", "observed-h1"),
  metric = c("expected", "observed", "observed-h1")
)

Arguments

Details

The default sensitivity is observed for FIML under MAR and expected for complete-data global goodness-of-fit under the correct-model null. Explicit sensitivity and metric choices remain available.

Value

A magmaan_global_score_flip_test list containing the effective multiplier result, asymptotic comparators, and saturated/tangent geometry diagnostics.

Description

Estimate the diagonal of the Guttman H matrix.

Usage

guttman_h(
  S,
  blocks,
  method = c("triad_ls", "extended_triad_ls", "triad_mean", "triad_pooled", "triad_wls",
    "triad_wls_joint", "ar", "rs", "gmm_block", "gmm_full", "anchor_triad_ls", "ilm",
    "anchor_ilm")
)

Arguments

Value

A list with correlation-scale communalities h2, covariance-scale diagonal h_diag, and H, equal to S with the diagonal replaced by h_diag.

Description

Extracts casewise contributions in group-block order from a prepared inference quadratic.

Usage

inference_rows(quadratic)

Arguments

Value

A matrix of casewise contributions.

Description

Methods-development primitives exposed through magmaan_core. Their ingredients must describe the component's null and sampling law.

Usage

infer_build_u_factor(fit, bread, moments = "structured")
infer_build_u_factor_parts(partable, sample_stats, theta, bread,
  moments = "structured")
infer_build_u_factor_pairwise(fit, X, bread, mask = NULL)
infer_build_u_factor_fit(fit, bread, moments = "structured")
infer_continuous_ls_robust(fit, raw_data = NULL, weight = NULL,
  bread = "observed", gamma = NULL, fixed_weight = FALSE)
infer_continuous_ls_profile_lrt(fit_H1, fit_H0, X_per_group = NULL,
  weight = NULL, eig_tol = 1e-10, gamma = NULL)
infer_ml_profile_lrt(fit_H1, fit_H0, X_per_group = NULL,
  eig_tol = 1e-10, gamma = NULL)
measures_reliability_cov(S, gamma = NULL, n = 0L, raw_data = NULL)
measures_reliability_omega_multidim(S, block, target = "total",
  weights = NULL, gamma = NULL, n = 0L, raw_data = NULL)

Arguments

Details

Scalar frontier_profile_lrt_* and interval primitives default to the misspecification-scaled reference (MisspecScaled) when reference = NULL, robust = FALSE. Supply raw data where required. The explicit reference = "ordinary" and reference = "robust_scaled" comparators remain available. Explicit robust = TRUE with NULL reference retains RobustScaled. Named Satorra, lavaan bundle and FMG/semTests replication APIs retain their documented compatibility conventions. SAM and case-change SE/type options standardize diagnostics and do not certify a generic inferential covariance.

Description

Returns standardized residual matrices, residual SE and z statistics, SRMR and block summaries. Estimated-weight inference for continuous least squares requires raw observations.

Usage

lav_residuals(fit, estimated_weight = TRUE, data = NULL)

Arguments

Value

A list of residual matrices and block-level diagnostic summaries.

Description

A named list of low-level native operations, organized by domain. Use the exported preparation, fitting and inference wrappers when they cover the required operation.

Usage

magmaan_core

Value

A list of callable native primitives.

Description

Constructs, fits or summarizes the native factor-component SEM representation. These objects use their own fitting route.

Usage

magmaan_fcsem(model, data, ..., missing = c("listwise", "error"), control = NULL)

Arguments

Value

A factor-component model specification, fitted model or named post-fit summary, according to the operation.

Description

Model-free leverage measure: stats::mahalanobis(X, colMeans(X), stats::cov(X)) (N-1 covariance), needing no refit. Mirror of semfindr::mahalanobis_rerun(). For multiple-group fits the distance is computed within each group (its own mean/cov); pass the original data frame to place the per-group distances back at the original rows (semfindr order).

Usage

mahalanobis_rerun(fit, data = NULL)

Arguments

Value

A one-column matrix of Mahalanobis distances, rownames = case ids.

Description

Adds multiplicity-adjusted p-values and a verdict based on significance and absolute effect size. Missing inputs produce missing verdicts.

Usage

mi_screen(x, p = NULL, effect = NULL, alpha = 0.05, effect_min = 0.1, adjust = "BH")

Arguments

Value

The input table with adjusted p-values and verdict (free, trivial, underpowered or ok).

Description

Constructs the model structure, names and start hints used by the lab fitting interfaces.

Usage

model_spec(syntax, ..., auto_var = TRUE, auto_cov_lv_x = TRUE, auto_cov_y = FALSE,
    orthogonal = FALSE, auto_fix_first = TRUE, marker = NULL, auto_fix_single = TRUE, std_lv = FALSE,
    effect_coding = FALSE, fixed_x = TRUE, meanstructure = "default", model_type = c("sem",
        "growth"), ordered = NULL, parameterization = "delta", group = NULL, group_labels = NULL,
    group_equal = NULL, group_partial = NULL)

Arguments

Value

A model specification containing the projected parameter table and construction metadata.

Description

Computes explicit release or restriction diagnostics. LR-type variants use the supplied observations for robust spectral inference. Unsupported candidate regimes are reported explicitly.

Usage

modification_indices(fit, data = NULL, ..., candidates = "all")

Arguments

Value

A table of candidate parameter diagnostics.

Description

Computes explicit release or restriction diagnostics. LR-type variants use the supplied observations for robust spectral inference. Unsupported candidate regimes are reported explicitly.

Usage

modification_indices_lrt(fit, data, candidates = c("all", "loadings", "covariances"), robust = TRUE,
    weight = NULL)

Arguments

Value

A table of candidate parameter diagnostics.

Description

Compute one-parameter modification indices or one-at-a-time equality-release tests, with a direction-specific sandwich correction.

Usage

modification_indices_robust(fit, data = NULL, weight = NULL,
  bread = "observed", moments = "auto", cov = "empirical",
  candidates = "all", include_loadings = TRUE, include_covariances = TRUE,
  information = "expected", estimated_weight = NULL, gamma = NULL)
score_tests_robust(fit, data = NULL, weight = NULL,
  bread = "observed", moments = "auto", cov = "empirical",
  estimated_weight = TRUE, gamma = NULL)

Arguments

Details

ML/FIML use expected information as the metric and bread, observed nuisance sensitivity when bread="observed", and empirical score meat (FIML uses observed-pattern scores). EPCs use the expected-information projection. Negative observed release curvature does not prevent a robust statistic. It needs no saturated-model EM computation for these tests. Explicit data rebuild their missingness pack; omitted data reuse the retained pack when available. Alternative FIML moment or covariance recipes and supplied second-stage weights are rejected.

Continuous ULS/GLS/WLS support empirical covariance from fitting observations, or model-implied normal-theory covariance from structured or unstructured moments. WLS-computed fits (WLS, DWLS, DLS, supplied W) use the recorded fitting weight for either source; the correction reduces to the ordinary statistic only when that weight is the inverse of the chosen covariance. Browne-unbiased covariance is not implemented for continuous LS. Estimated-weight mode requires raw observations and empirical covariance; other covariance choices are rejected.

Two-stage (ML2S) fits report mi as the naive Stage-2 statistic on the Stage-1 EM moments (attribute mi_type = "naive_stage2") and mi.scaled with the Stage-1 moment covariance as meat. They use the recorded Stage-2 weight, expected information and the default bread, moments and cov; a DWLS, ADF or DLS Stage-2 weight adds its data influence with estimated_weight = TRUE. The naive NT row uses the structured information; lavaan's two-stage modindices() uses the unstructured one and differs.

Identification-only releases are excluded. Ordinal and mixed-ordinal methods use their retained categorical moment covariance and weight conventions; the all-ordinal bread selects observed or expected nuisance sensitivity, and mixed methods now honor the same selection. Complete mixed estimated-weight MI/releases use exact empirical first-stage rows and mixed Gamma influence. Expected mixed projection and caller-Gamma comparators retain NACOV. All-ordinal moments and cov choices use the retained NACOV.

Generic modification_indices() and score_tests() compose empirical robust inference. To request the historical unscaled comparator explicitly, use cov = "model_implied", estimated_weight = FALSE, and the desired bread/information. Continuous ULS/GLS/DWLS/WLS/DLS and all-ordinal ULS/DWLS/WLS and complete mixed DWLS/WLS support observed-Hessian nuisance sensitivity with estimated-weight influence. Their quadratic metric stays expected and their reference law uses the matching projected meat. Selecting expected bread or fixed weights requests a comparator. At exact fit the observed and expected projections coincide and the weight-influence correction vanishes. Observed-sensitivity ML2S score laws remain unavailable.

Value

Tables carry reason and detail; identified numerical failures retain their candidate rows with NA values. A data frame with candidate names and groups, unscaled quadratic mi, EPCs, mi.scaled, scaling.factor, and the robust pvalue.

Description

Reads numeric individual or summary ASCII data using the C++ data plan. Relative paths are tried in the working directory, then beside the input file. Free format accepts blanks, tabs and commas, wrapped observations, and discards extra fields at the end of a completed observation's record. Fixed FORMAT supports Fw.d, Fw or w.d, X skips, Tn positions, record breaks and repeated groups. Implied decimals apply only without an explicit point. Missing flags compare after that scaling. A terminal DOS Ctrl-Z, optionally followed only by whitespace, denotes EOF. Embedded EOF markers and other control bytes are rejected; remove the marker or the trailing records.

Usage

mplus_data(model, file = NULL)

Arguments

Details

Separate-file groups use a reserved .mplus_group column containing FILE labels in declaration order. NGROUPS summary groups are named g1, g2, ... . GROUPING drops unlisted codes and reports counts. NOBSERVATIONS limits individual observations per file. LISTWISE and analysis-sample deletion rules are only reported: use fit_model(..., missing = "listwise") to remove incomplete rows or missing = "fiml" for FIML; explicitly filter missing covariates to reproduce Mplus's conditional analysis sample.

Summary matrices use divisor N-1 on input. As in Mplus ML and lavaan's default sample.cov.rescale, they are multiplied by (N-1)/N for fit_model(). Correlations are multiplied by supplied SDs; without SDs the input has unit variances. The report retains input covariances and conversion factors.

Accepted summary TYPE values are COVARIANCE, CORRELATION, FULLCOV and FULLCORR, optionally with MEANS (and STDEVIATIONS for correlations). Supply NOBSERVATIONS for every summary group; summary input uses free format. MISSING accepts one global ., * or fixed-format BLANK symbol, or numeric flags per NAMES list/ALL. Fixed blanks without BLANK missing read as zero. Empty comma fields are errors; numeric records are bounded to 10000 characters. File and matrix shape errors identify the data rule. Perform DEFINE/qualified DATA transformations and USEOBSERVATIONS or SUBPOPULATION selection in R before reading/fitting. Multiple imputation, Monte Carlo datasets and two-level weight matrices are outside this reader; supply one supported individual or summary dataset per declared group.

Value

Individual data as a data.frame, or summary moments as a list with S, mean and nobs accepted by fit_model(). Attribute mplus_data_report records paths, counts, dropped codes and unapplied sample rules.

Description

Imports single- and multiple-group linear SEM: BY, ON, WITH, PON, PWITH, means, variances, starts, fixes, labels and equality numbers, NOCOVARIANCES, and NOMEANSTRUCTURE with INFORMATION = EXPECTED. Names resolve without regard to case and retain the spelling in NAMES; factors retain their first BY spelling. Estimator and execution settings are reported, not imported.

Usage

mplus_model(input = NULL, file = NULL)

Arguments

Details

CATEGORICAL lists binary and ordinal outcomes (two through ten categories). Thresholds use ⁠[u$k]⁠, DELTA scales use {u}, and THETA residuals use variance statements. Categories and thresholds are completed from data; all groups must contain the same categories. All-ordinal models fit with estimator = "DWLS"; mixed, conditional and ML categorical routes report their unsupported route explicitly. Retained-estimate lavaan WLSMV reporting is available through convention_inference().

Fixed-time polynomial and piecewise growth (⁠i s | y1@0 y2@1 ...⁠) imports outcome intercept, growth mean, threshold and scale/residual defaults. Growth accepts one through four factors; free time scores are accepted when enough scores remain fixed to identify the polynomial loadings. MODEL CONSTRAINT imports NEW (starts default to 0.5), explicit and implicit equalities, derived quantities, and nested DO loops. Functions include EXP, LOG, SQRT, PHI and LOG10; ⁠**⁠ denotes power. Continuous ML, LS and FIML use the existing equality backends. All-ordinal nonlinear equalities use constrained LS fitting and expected-information lavaan reporting. Native policy and observed/IJ inference require unimplemented Lagrangian curvature. Inequalities are deliberately refused: active-bound inference requires boundary asymptotics; for variance positivity drop the constraint and use covariance = "psd" or "barrier" in fit_model(). MODEL INDIRECT imports total IND, specific IND and VIA products, including paths through factors. Derived names are ⁠ind_g<group>_<outcome>_ind_<names>⁠ or ⁠ind_g<group>_<outcome>_via_<mediator>_<predictor>⁠; specific names retain the written mediator order. Defined estimates and delta-method SEs are available through compute_defined(). A free NEW coordinate appears as a new partable row, with no observed variable or moment cell; rebuilding always uses the original Mplus source. INFORMATION defaults, MODEL TEST, LOOP and PLOT are reported in notes.

NAMES accepts bounded numeric- or letter-suffix ranges; MODEL and USEVARIABLES ranges follow schema order. Latent ranges follow factor-definition order. Labels apply on their physical line; subsequent mentions override earlier parameter specifications. PON and PWITH require equally sized lists. Model/option content must fit within 90 columns; wrap longer statements. TITLE and comments may be longer. Line and block comments are accepted, but a block opened after code must close on its opening line. TITLE, OUTPUT, SAVEDATA and PLOT are reported.

DEFINE and DATA transformations must be performed in R; USEOBSERVATIONS and SUBPOPULATION must be applied as case selection in R before fitting. Counts, nominal/censored/survival outcomes, survey weights, mixtures, multilevel models, Bayes and full-information categorical links are outside the linear SEM family: retain those analyses in Mplus. ESEM rotation is replaced only by explicitly specified ordinary BY factors. Bare @ uses data-dependent Mplus starts; write an explicit ⁠@value⁠ instead. Remove all observed-independent means/variances/WITH mentions to preserve fixed-x conditioning, or bring every observed independent variable into the joint random-X model (e.g. ⁠x1 x2;⁠). Complete mentions free X means, variances and default covariances; partial mentions are rejected with the exact variance statement that completes the joint model. Recode data-dependent GROUPING forms to explicit integer code/label pairs in R. Categorical summary inputs require individual observations for magmaan's moment preparation.

DEFINE, mixtures, multilevel models, ESEM, unsupported name ranges and mixed conditioning on observed covariates are rejected with rule IDs and an explanation of what to write instead. Data descriptions are parsed into ⁠$mplus_data_plan⁠; mplus_data() reads them. This constructor does not read the data file or run Mplus. GROUPING imports explicit integer code = label pairs, ordered by numeric code. Group MODEL sections override the overall model; repeated sections apply cumulatively. CONFIGURAL, METRIC and SCALAR accept one setting at a time. Categorical shortcuts accept CONFIGURAL and SCALAR; METRIC is rejected for binary and ordinal outcomes, following Mplus 9.1. Group-only changes to variable roles are rejected; put the relation in the overall model and fix it in the other groups instead. Variance identification in a shortcut uses free first loadings and explicit factor variances at one. mplus_data() drops unlisted GROUPING codes with counts. Separate FILE groups retain source labels, use .mplus_group, and follow FILE order. Summary NGROUPS labels are g1, g2, ... . Summary data without MEANS have no mean structure and reject explicit means/intercepts.

Value

A magmaan_model_spec with original mplus_source, a data.frame of mplus_notes (class, rule, line, col, message), and a lavaan syntax projection. Grouped specs also carry group_var, ordered integer-code group_labels and mplus_groups (a data.frame with label and code columns). Rebuilding uses mplus_source; construction options cannot be overridden.

Description

Constructs and fits the requested ordinal invariance sequence using the Mplus-style WLSMV conventions.

Usage

mplus_wlsmv_invariance(model, data, ordered, group, steps = c("configural", "metric", "scalar"),
    estimator = c("WLSMV", "DWLS", "WLS"), missing = c("listwise", "pairwise", "error"),
    group_labels = NULL, optimizer = "nlopt-lbfgs", control = NULL, bounds = NULL,
    pd_gamma = c("overlap", "nominal"), validate = TRUE)

Arguments

Value

An invariance result with the constructed models, fits and comparisons.

Description

Computes the same nuisance-orthogonalized observed score and robust restriction spectrum used by score_flip_test(), but skips every random sign draw and the flip-specific covariance calculations. The recommended p_value is the pEBA4 mixture approximation. Scalar SB, exact-mixture, and direct sandwich comparators remain available for diagnostics.

Usage

nested_score_test(
  fit_H1,
  fit_H0,
  data = NULL,
  sensitivity = c("observed", "expected")
)

Arguments

Value

A list of class magmaan_nested_score_test containing the effective score statistic, restriction eigenvalues, pEBA4/SB/exact-mixture/sandwich p-values, conditioning diagnostics, and setup/asymptotic timings.

Description

Minimum-distance projection of the configural (per-group) fit onto the linear equality constraints ⁠A theta = b⁠ carried by the partable's group.equal family (metric / tau-equivalence / strict invariance, fixed values, ...). The Omega-metric projection is closed-form: theta_tilde = theta_hat - Omega A'(A Omega A')^-1 (A theta_hat - b), and the constraint statistic W is an exact chi-square on k (Wald = minimum-distance duality). Returns the projected estimate (theta_tilde), its rank-deficient covariance (vcov_constrained, se_constrained), and the constraint test (W, k, p_wald).

Usage

noniterative_cfa_constrained(
  fit,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

T_d = T(H0) - T(H1) with a weighted-chi-square reference (H1-anchored). Secondary to the Wald test: T_d can be negative for a non-minimizing estimator (flagged in warnings).

Usage

noniterative_cfa_difference_test(
  fit0,
  fit1,
  df_d,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data0 = NULL,
  data1 = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Fits every group's Guttman block independently and stacks them: a block-diagonal delta-method covariance (vcov, se), a joint goodness-of-fit T on df degrees of freedom with the union weighted-chi-square reference, and block_of_param (0-based group index per free parameter). Reduces to noniterative_cfa_inference() for a single group with no mean structure. Required for mean-structure models (the single-group path omits the intercept block).

Usage

noniterative_cfa_grouped_inference(
  fit,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Returns delta-method standard errors (se, vcov), the goodness-of-fit statistic T on df degrees of freedom, its weighted-chi-square p-value family (p_scaled Satorra-Bentler, p_meanvar, p_scaled_shifted, p_mixture), the reference eigenvalues, and the Satorra-Bentler scale_c. For discrepancy = "ntml", rls_check is the reweighted-least-squares cross-check of T.

Usage

noniterative_cfa_inference(
  fit,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Closed-form Guttman fits are not stationary minimizers of the ML/ULS discrepancy, so the usual score-test modification index is not defined. This helper reports residual-space analogues for fixed-zero and absent one-parameter candidates: raw residual scores, tangent-residualized residual scores, and local one-step discrepancy-drop/EPC diagnostics.

Usage

noniterative_cfa_modification_indices(
  fit,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL,
  candidates = c("all", "fixed"),
  include_loadings = TRUE,
  include_covariances = TRUE
)

Arguments

Value

A data frame with candidate columns and .raw, .resid, and drop diagnostics.

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

fit_H1 is the less restricted model and fit_H0 is the more restricted estimator-side fit, typically from fit_noniterative_cfa_restricted(). The statistic is T_diff = T(H0) - T(H1) with an H1-anchored weighted-chi-square reference. This is an LRT-shaped discrepancy comparison; unlike ML, it can be negative for a non-minimizing estimator.

Usage

noniterative_cfa_pseudo_lrt(
  fit_H1,
  fit_H0,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Frees the latent means. Reads the common metric off the reference group and solves the latent means by least squares through the reference loadings: ⁠alpha_g = (Lr'Lr)^-1 Lr'(m_g - m_r)⁠, nu = m_r. Scalar invariance is tested on the mean residual d_g = (I - P_Lr)(m_g - m_r) with a linearized pseudo-inverse Wald W on ⁠df = (G-1)(p - #factors)⁠ degrees of freedom. The fit must carry mean structure (fit a configural meanstructure model first) and at least two groups.

Usage

noniterative_cfa_scalar_invariance(
  fit,
  ref_group = 1,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Details

Returns the latent means alpha (a list, one per non-reference group) with delta-method alpha_se / alpha_cov, the common intercept nu, the stacked residual d, and the test (W, df, rank, p_value).

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Returns only se and vcov, skipping residual goodness-of-fit spectra.

Usage

noniterative_cfa_se(
  fit,
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Exact chi-square on nrow(R) degrees of freedom, using the delta-method covariance. The primary nested test for these estimators.

Usage

noniterative_cfa_wald(
  fit,
  R,
  q,
  discrepancy = c("uls", "ntml"),
  gamma = c("nt", "empirical"),
  data = NULL,
  estimator = NULL
)

Arguments

Note

Noniterative development is postponed. The NT default is a normal-theory comparator and does not provide structural-misspecification robustness.

Description

Estimator-agnostic: works for any magmaan fit given a vcov on the free parameters. Rows are the free-parameter block of the fitted partable, in theta order (by the free index), so est, se, z, and pvalue are aligned.

Usage

parameter_table(fit, vcov, level = 0.95)

Arguments

Value

A data frame with lhs, op, rhs, group, est, se, z, pvalue, ci.lower, ci.upper.

Description

Computes indices on request, including an independence baseline under the same estimator and data treatment. ML and FIML use the raw-likelihood Takeuchi trace difference; ordinal and mixed DWLS use exact first-stage influence with the estimated-weight channel. Continuous ULS uses its profile trace plus covariance moment bias; ML2S-NT uses Stage-1 Gamma. No intervals are returned. RMSEA includes the multigroup factor sqrt(G). TLI uses nominal degrees of freedom and truncated corrected discrepancies; the lab misspecification family's generalized-df TLI remains a separate comparator. Residual indices pool squared residuals across groups and subtract their influence trace. DWLS indices measure misfit in the DWLS metric; ML cutoffs do not transfer.

Usage

policy_fit_measures(fit)

Arguments

Value

A data frame with index, estimate and reason, and a details attribute carrying discrepancies, traces, corrected discrepancies and degrees of freedom. Unsupported components have NA estimates and explicit reasons.

Description

Requests the C++ ordinary-user inference policy for the fitted estimator and data regime. Unsupported regimes remain explicit.

Usage

policy_inference(fit, data = NULL)

Arguments

Details

ML/FIML use observed-bread empirical-score sandwich covariance and score and likelihood-ratio tests with PEBA4 as the default reference. All-ordinal and complete mixed DWLS use exact empirical first-stage sampling influence and estimated-weight IJ covariance. Global fit-function tests and observed-Hessian/IJ nested fit-function differences use the exact weighted chi-square All reference; SB/PEBA4 comparators remain available. Fitting retains OPG NACOV weights. Reference-law studies support the DWLS default across ordinal, mixed, threshold-invariance and textbook designs; finite-sample size varies across settings. Nested DWLS score tests are unavailable.

Value

An inference bundle with availability status, covariance and test summaries where supported.

Description

Frees one candidate (or releases one equality) and refits from the fitted estimate embedded in the augmented model, replaying the fit's options. The template is an evaluation point for the native embedding map; it is never optimized.

Usage

policy_mi_refit(fit, row, candidate_row)

Arguments

Value

The fitted augmented model, for policy_nested.

Description

Thin adapter to the C++ ordinary-policy modification-index composer.

Usage

policy_modification_indices(fit, data = NULL, releases = TRUE)

Arguments

Details

ML/FIML use observed nuisance sensitivity with expected metric/EPCs and empirical likelihood-score meat (pattern scores for FIML), and all-ordinal/mixed DWLS use exact first-stage sampling rows and estimated-weight influence with observed sensitivity. Each available robust score statistic has a chi-square(1) limiting reference. Singular nuisance sensitivity or nonpositive score variance retains numeric_failure rows. Fits outside the policy return typed unavailable rows. No alternative inference convention is substituted.

Value

A data frame with kind, lhs, op, rhs, group, test, statistic, df, pvalue, epc, sepc.lv, sepc.all and reason. Internal candidate-row metadata supports native one-restriction augmentation and embedded-start refits.

Description

Requests the C++ ordinary-user nested policy for a restricted and unrestricted fit.

Usage

policy_nested(fit_H1, fit_H0, data = NULL)

Arguments

Details

ML/FIML use observed-bread empirical-score sandwich covariance and score and likelihood-ratio tests with PEBA4 as the default reference. All-ordinal and complete mixed DWLS use exact empirical first-stage sampling influence and estimated-weight IJ covariance. Global fit-function tests and observed-Hessian/IJ nested fit-function differences use the exact weighted chi-square All reference; SB/PEBA4 comparators remain available. Fitting retains OPG NACOV weights. Reference-law studies support the DWLS default across ordinal, mixed, threshold-invariance and textbook designs; finite-sample size varies across settings. Nested DWLS score tests are unavailable.

Value

A nested inference bundle with availability status and supported test summaries.

Description

Own sample contributions and fitted NTML geometry across score, likelihood-ratio, goodness-of-fit and covariance calculations. These are methods-development interfaces for interior, random-X complete-data ML fits with affine equality constraints.

Usage

prepare_inference_data(fit, data = NULL,
  storage = c("auto", "casewise", "tiled"))
prepare_hypothesis(null, alternative)
inference_quadratic(object, test = c("score", "lr"),
                    geometry = NULL)
inference_covariance(context, robust = TRUE)
inference_reuse(context)

Arguments

Details

The core reuses the existing NTML U-factor construction and casewise covariance reducers. It retains structured expected-information geometry and each test's reduction separately. Global LR uses the centered empirical GOF spectrum; global score uses likelihood contributions at the fitted null point. Restricted means require additional linear and constant terms, so these spectra are not silently equated. Nested LR uses the exact H1-anchored Satorra–2000 reduction; nested score evaluates at H0.

No empirical Gamma or full U matrix is required. Casewise storage reuses one contribution matrix across fits. Tiled storage reduces global score and GOF together when the sample size is at least df, retaining reduced matrices; otherwise the smaller row-space spectrum is used. Tiled storage trades repeated raw-data passes for bounded contribution memory when different hypotheses need different projections. Explicit Browne-unbiased GOF requests use the existing casewise correction and may materialize contributions.

score_spectrum and calibrate_quadratic accept these quadratics. Spectrum calculations are lazy and retained; SB alone uses the trace. inference_covariance returns a snapshot-tagged matrix usable with wald_test. Existing GOF, expected-information, expected sandwich-SE and exact empirical streaming nested-LR wrappers also reuse compatible prepared snapshots. Other bread/Gamma/nesting conventions retain their existing paths.

Shared NTML geometry does not yet cover FIML, ML2S, ordinal or fixed-X fits. Their existing score interfaces remain available. Native handles are local to the current R process: prepare again after serialization. Locked R environments preserve ownership; edited extracted fit lists invalidate the native caches.

Value

Prepared data/hypothesis and quadratic calls return locked environments owning native objects. Covariance returns a matrix; reuse returns construction counts, including data contributions, fit geometry, U, information and global spectra. Counts are diagnostics, not a timing guarantee.

Examples

## Not run: 
d <- prepare_inference_data(fit1, data)
g0 <- prepare_inference(fit0, d)
g1 <- prepare_inference(fit1, d)
h <- prepare_hypothesis(g0, g1)
qs <- inference_quadratic(h, "score")
ql <- inference_quadratic(h, "lr")
calibrate_quadratic(qs, c("sb", "peba4"))
calibrate_quadratic(ql, c("sb", "peba4"))
V <- inference_covariance(g0, robust = TRUE)
inference_reuse(g0)

## End(Not run)

Description

Separate structural model preparation, dataset preparation, estimation weights and estimation. Inference is requested separately.

Usage

prepare_model(model, ..., prototype = NULL)
augment_model_spec(model, categories)
prepare_data(model, data, kind = NULL, missing = c("error", "listwise"))
prepare_weight(data, method = c("DWLS", "WLS", "ULS", "GLS", "DLS"),
               W = NULL, full = TRUE, dls_a = 0.5)
estimate(model, data, estimator = NULL, weight = NULL,
         optimizer = NULL, control = NULL, bounds = NULL,
         covariance = NULL, psd = FALSE, barrier = NULL, dls_a = 0.5,
         options = NULL)

Arguments

Details

Rebuild handles after serialization and in each worker. A new dataset requires new data and weight preparation, but reuses the model. Category levels, group order, observable order and mean-structure convention must remain compatible. Group labels must be declared explicitly. Covariances supplied as sample moments use divisor N.

The covariance policy changes final model fitting without repairing supplied moments. A barrier minimizes f - weight / N * P, with fmin retaining the unpenalized discrepancy and penalty$penalized_fmin recording the fitted objective. Positive-weight barriers replace additional bounds. Their general sampling/inference contract is not yet validated. Ordinal ML fits associations with Stage-1 thresholds held fixed; it does not provide Gaussian likelihood or generic post-fit sampling inference.

Continuous diagonal weight blocks use owning numeric vectors of entries of W, not square roots; full blocks remain matrices. Supplied W accepts either form, or a list of blocks. Prepared native handles are process-local: serialize W and prepare a new handle after loading it.

FIML retains its missingness pack but does not compute saturated H1 during estimation unless versioned fitting options require it. Categorical inference needs full Gamma; obtain it separately with prepare_weight(data, "DWLS", full = TRUE)$stats and pass it to an existing explicit inference function, without refitting.

Versioned fitting options use the same configured engines as fit_model() for continuous ML/FIML and ordinal/complete mixed DWLS with unrestricted covariance, including options = list(preset = "lavaan-0.7.2").

The interface covers the estimators listed above. ML2S raises a typed magmaan_unsupported_estimator error. Two-level, FC-SEM, SAM and frontier methods continue using their existing entry points, which remain supported without deprecation warnings. Categorical weight preparation currently recomputes stage-one moments in the existing core builder; fitting itself reuses the prepared objects.

Value

augment_model_spec returns a model specification with threshold and scale rows added once. Preparation returns immutable, process-local handles. A prepared model exposes its specification; a weight exposes W for continuous data and stats for categorical inference. Estimation returns the existing magmaan_fit list with convergence and admissibility diagnostics.

Deletion provenance

Dataframe helpers, prepared data and fitted lab objects retain a read-only deletion record. Its policy identifies listwise, error, FIML or pairwise row handling; known indicates whether supplied-row counts are available. per_group contains group, rows_supplied, rows_used, rows_deleted, and exclusive reason counts missing_model_values, missing_group_label, unknown_group_label, and missing_cluster_id. Reasons take precedence in that order: group labels, cluster IDs, model values. Rows without an assignable group appear in a separate NA group row. Single-level routes still refuse missing or unknown group labels; two-level listwise preparation deletes missing group or cluster IDs. FIML and ML2S retain incomplete rows without listwise deletion; existing native refusals remain. Summary-only inputs without retained provenance report unknown supplied/deleted counts, rather than claiming that no earlier deletion occurred.

Description

Embeds the restricted estimate using the native, moment-verified nesting map, then replays the larger model's estimator and fitting options with that explicit start. Neither input is modified.

Usage

refit_from_null(fit_H1, fit_H0)

Arguments

Details

Supports single-level ML, FIML and all-ordinal DWLS, ULS and WLS fits from fit_model() or estimate(). FIML requires retained observations; complete-data and ordinal routes reuse retained sufficient statistics. This explicit lab operation returns the new endpoint even when its verdict fails; callers must inspect convergence diagnostics. Fit warnings are emitted as for any explicit user-requested fit.

Value

A new estimate-only magmaan_fit, with its own convergence diagnostics and native verdict.

Description

Compares a restricted model fit_H0 against the less-restricted superset fit_H1. method = "restriction_map" uses Satorra's streaming low-rank reduction of the asymptotic weighted-chi-square distribution. The "lavaan_sb2001" and "lavaan_sb2010" options mirror lavaan's compatibility difference-test approximations.

Usage

robust_nested_lrt(
  fit_H1,
  fit_H0,
  data = NULL,
  gamma = c("empirical", "unbiased", "NT", "both"),
  method = c("restriction_map", "lavaan_sb2001", "lavaan_sb2010"),
  A.method = c("exact", "delta"),
  computation = c("streaming", "materialized", "dense"),
  convention = c("magmaan", "lavaan"),
  ud_method = c("2000", "2001"),
  h1_reference_regularization = NULL,
  weight = NULL
)

nestedTest(
  fit_H1,
  fit_H0,
  data = NULL,
  gamma = c("empirical", "unbiased", "NT", "both"),
  method = c("satorra.2000", "satorra.bentler.2001", "satorra.bentler.2010",
    "restriction_map", "lavaan_sb2001", "lavaan_sb2010"),
  A.method = c("exact", "delta"),
  computation = c("streaming", "materialized", "dense"),
  convention = c("magmaan", "lavaan"),
  ud_method = c("2000", "2001"),
  h1_reference_regularization = NULL,
  weight = NULL
)