Changelog
Source:NEWS.md
idiographic 0.4.0.9006
Unified idiographic statistics
New
failures()accessor returns the failure table of any result as a tidy data frame, withmodel,scopeandsubjectfilters. Print methods and error messages now point tofailures()instead of$failures, and rolling network results return their windows withas.data.frame(x, what = "windows").fit_gimme()checks its arguments before it checks forlavaan, so an unsupported GIMME mode raises its own error whenlavaanis absent.Changed results: pooled and subgroup models from
fit_lm()andfit_glm()now report standard errors clustered on the person (CR1). They previously used model-based standard errors that treated every occasion as an independent observation; on the bundledsrldata these were about five times too small. Person-specific models are unchanged.Predictors with no variance for some persons now raise a warning of class
idiographic_constant_predictornaming those persons; their person-specific models previously omitted the predictor silently.Folded the
idiostatsanalytical engine intoidiographic. The package now covers person-level descriptions and correlations, LM/GLM and expanded ML, rolling-origin validation, within-between effects, coefficient pooling and shrinkage, subgroup existence/discovery/modelling, treatment effects, and heterogeneity analysis alongside all existing dynamic-network estimators.Added
describe_persons(),correlate_persons(),variance_components(),fit_lm(),fit_glm(),fit_rolling(),fit_within_between(),pool_coefs(),shrink_coefs(),test_subgroups(),find_subgroups(),fit_subgroups(),fit_effects(), andfit_heterogeneity()with their tidy accessors, views, diagnostics, and plots.Added
preprocess_panel()for person/grand centring and scaling, detrending, within-between decomposition, and lags that never cross person boundaries. The establishedpreprocess()network audit remains unchanged.fit_ml()now supports both interfaces. Existingoutcome/predictorscalls retain the originalidioml_result; consolidatedy/xcalls support pooled, subgroup, and individual scopes, richer native and optional model registries, tuning, diagnostics, and rolling validation.Consolidated statistical results use an
idiographic_*primary class while retaining theiridiostats_*compatibility class for downstream migration.Code written for
idiostatsruns withlibrary(idiographic). The formeridiostats::preprocess()transforms arepreprocess_panel(), andpreprocess()remains the network audit. Namedfit_ml(y = , x = )calls use the scoped engine,fit_ml_panel(data, y, x, id)keeps the former positionalidiostatscall, and positionalfit_ml(data, outcome, predictors, id)keeps the originalidiographiccall.Added a statistical workflow index and six topic vignettes covering the 16 major non-network functions: describing persons, panel preparation, regression models, prediction and temporal validation, pooling and shrinkage, and subgroups, effects, and heterogeneity. Each section uses bundled data, prints tidy results, renders a decision-focused figure, interprets the observed output, and states assumptions and failure conditions; each vignette ends with guidance on choosing among its functions. The prediction vignette includes temporal validation, baselines, tuning, scope comparison, person-level errors, prediction trajectories, and permutation importance. The subgroup vignette includes a worked mixture of regressions and shows why its own BIC is not used to choose the number of groups.
as.data.frame()forfind_subgroups()results returns the assignment table (default), the mixture component regressions (what = "components"), or the mixture’s BIC sweep (what = "bic"), so these no longer have to be read from the object’s internals. A missing table raises an error of classidiographic_missing_table.Redesigned statistical plotting around ranked dot plots, interval displays, binned calibration, direct labels, temporal trajectories, and consistent colour-accessible styling. Plots now foreground uncertainty, reference values, and person-level failures instead of relying on default bar charts.
Documentation
Documentation revision.
fit_ml()now documents both interfaces, the consolidated controls passed through...and the structure of each result, with runnable examples. Every exported accessor and view (metrics(),predictions(),coefs(),importance(),diagnostics(),tuning(),best_model(),compare(),person(),people(),individuals(),pooled(),overall(),plot_diagnostics(),plot_importance(),plot_tuning()) has an example and a return value that names its columns. Thecoefs()help page no longer claims that only multilevel VAR fits are supported.Examples and print methods no longer reach into results with
$or[[. They useas.data.frame(),summary(),edges(),matrices(x, fit = ),matrices(x, subject = )andplot(x, fit = )instead. The multilevel VAR examples use the bundledsrldata rather than white noise, which removes 28 singular-fit warnings, and the Bayesian VAR andedges()examples now run underR CMD checkinstead of sitting entirely in\donttest{}.The DESCRIPTION gives author and year for both method references, and the graphical VAR, subgroup-test and heterogeneity references carry DOIs.
idiographic 0.3.4
CRAN release: 2026-08-24
New features
The pure-R graphical-lasso kernel that powers
fit_graphical_var()is now a supported public API:glasso_fit(),glasso_path(), andglasso_kkt().glasso_fit()returns aglasso_resultwhose$wiand$welements are named exactly asglasso::glasso()’s, so existing call sites port unchanged, and it supports element-wise penalty matrices and hard zero constraints in addition to a scalar penalty.glasso_kkt()certifies a solution from the graphical-lasso stationarity conditions rather than against another solver. Both results haveprint()and tidyas.data.frame()methods. This exists so sibling packages can drop their own copies of the same kernel and depend on this one; seeOFFLOAD.mdin the repository.matrices()gains aprintargument. The default (print = TRUE) is unchanged: it prints each matrix and returns the list invisibly. Withprint = FALSEthe function prints nothing and returns the list visibly, which is the form dependent packages and resampling loops need. Threaded through everymatrices()method, including those that delegate to another method.printfollows...in every method, so it must be named in full and can never be matched positionally or by partial name.
Bug fixes
glasso_kkt()no longer reports optimal zero-constrained fits as non-optimal. At an entry hard-constrained byzero, the inactive-edge condition|W_ij - S_ij| <= rho_ijdoes not apply: the equality constraint carries its own Lagrange multiplier, which absorbs the residual. Constrained pairs are now excluded from the check. A fit matchingglasso’s Fortran kernel to 2e-12 previously certified as violating optimality by 0.028.glasso_fit()now rejects a covariance matrix that is not positive semi-definite (classed conditionidiographic_not_psd). A negative eigenvalue produced a precision matrix with a negative diagonal, from which every partial correlation wasNaN. Singular but positive semi-definite covariances remain valid input. The tidy accessor also refuses a precision matrix with a non-positive diagonal (idiographic_bad_precision) rather than lettingstats::cov2cor()warn and emitNaNweights.An asymmetric element-wise penalty matrix is now rejected by
glasso_fit()and byfit_graphical_var()’sregularize_mat_kappa. The penalty on edge(i, j)is a single scalar over a symmetricTheta, so an asymmetric penalty is ill-posed rather than stricter; previously it was accepted and produced a fit that failed the package’s own optimality check.glasso_kkt()validatespenalize_diagonalinstead of lettingisTRUE()silently mapNAand other invalid values toFALSE, and warns (idiographic_glasso_kkt_override) when a suppliedrhoorpenalize_diagonaldiffers from the one the fit was made with, since the returned violation then certifies a different problem.The internal graphical-lasso optimality checker used to measure the unpenalised diagonal stationarity condition (
W_ii = S_ii) even for fits made withpenalize.diagonal = TRUE, whose condition isW_ii - S_ii = rho. It therefore reported a spurious violation of exactlyrhofor every such fit — includingglasso’s own Fortran output, which is how this was found. It now takes the diagonal-penalty flag into account. No estimator result and no previously published number changes; only the diagnostic was wrong.
Documentation
matrices()now documents that it is a display verb by default, and points atprint = FALSEfor programmatic extraction.The three most computationally expensive vignettes – Graphical VAR, Bayesian VAR/DSEM, and GIMME – are now pkgdown articles rather than installed vignettes. Their content is unchanged and they remain published at https://pak.dynasite.org/idiographic/, but they are no longer rebuilt during
R CMD check. Rebuilding all ten vignettes took 574 seconds on Windows R-devel (68% of the total check time, against CRAN’s 10-minute guideline); these three accounted for the large majority of it. The remaining seven vignettes are renumbered 1-7.
idiographic 0.3.2
CRAN release: 2026-08-03
- New package Title — “Idiographic Person-Specific and Heterogeneous Complex Networks” — and a rewritten Description with method references doi:10.1007/978-3-031-95365-1_20 and doi:10.1080/00273171.2018.1454823.
- Added Sonsoles López-Pernas as package author.
- Documentation language standardized to British English (
Language: en-GB), with dialect fixes across the documentation prose and a newinst/WORDLISTso the package spell check runs clean. - Slimmer CRAN footprint: all competitor-equivalence tests now live only in the repository’s opt-in validation lane and are excluded from the CRAN tarball. Suggests trimmed from 16 to 8 packages — removed
gimme,graphicalVAR,glasso,corpcor,data.table,qgraph,rio, andjsonlite, none of which the shipped package uses.
idiographic 0.2.0
Made the CRAN package offline-first: the only mandatory imports are standard R packages, while
lme4,lavaan, plotting, and external backends are optional. Competitor-oracle tests and the real-panel corpus now run in a separate opt-invalidation/lane and are excluded from the CRAN tarball.Added a registry-backed
fit_idiographic()front door, estimator discovery, method-specificequivalence()declarations, package-wideequivalence_table()and argument-by-argumentargument_coverage()ledgers, and common tidy accessors. All 17 registered methods and 315 current public formals now have an executable evidence classification; new unassessed arguments fail the closure test.Expanded direct-oracle testing across graphicalVAR option combinations, mlVAR multi-lag/preprocessing/unique-model configurations, and bivariate plus three-variable GIMME standard, hybrid, and VAR searches. GIMME evidence now also covers fit statistics, uneven panels, exogenous-variable dimensions, and interacting correction/standardization controls. Tightened public argument validation so engine-specific controls cannot be silently ignored.
Closed the remaining executable evidence cells: all 12 supported lag-1 lmer mlVAR structure combinations, per-subject/missing-data graphicalVAR fits, GIMME 10.0 correction/stopping/standardization/cutoff/forced-path controls, standardized ML/MLR uSEM fits, Mplus wrapper forwarding/conversion, Bayesian burn-in/thinning, positive random-residual recovery, parallel mlVAR, and base-R linear/logistic idiographic-ML engine equality.
Migrated the 20-panel real ESM mlVAR validation corpus from the Dynalytics/psychaj work into the CRAN-excluded
validation/lane, with self-contained raw inputs, mlVAR 0.7.3 frozen oracles, provenance hashes, and explicit regression coverage for missing IDs, irregular occasion gaps, and degenerate between-person networks. Duplicate observation keys now fail clearly instead of producing row-order-dependent preprocessing.Uniform
fit_*naming for all estimators (breaking). Every model-fitting verb now uses a singlefit_prefix:fit_var(),fit_graphical_var(),fit_mlvar(),fit_rolling_var(), and so on for all estimators. Short model nicknames passed tocompare_idiographic(),estimate_stability(), andvalidate_forecast()(for example,"var"and"graphical_var") are unchanged.-
New native Bayesian estimators that statistically reproduce Mplus DSEM output without requiring Mplus:
-
fit_mlvar_bayes()— two-level Bayesian VAR(1) with latent mean centring.temporal = "fixed"is statistically validated against frozen Mplus DSEM fixed-temporal + random-intercept fixtures;temporal = "random"fits the full DSEM with person-specific temporal matrices and a random-effect covariance (reports random-slope SDs). -
fit_var_bayes()— single-level Bayesian VAR(1), the unregularized Bayesian analogue offit_graphical_var().
-
Pure-R conjugate Gibbs sampler (hand-rolled inverse-Wishart draws; no new dependencies). Posterior median / SD / 95% CI / one-tailed p, three networks (temporal, contemporaneous, between), and a Gelman-Rubin PSR diagnostic.
Validated to statistical (Monte-Carlo-error) equivalence against real Mplus 9 output with frozen ground-truth fixtures and parity tests.
Added
fit_ml()for idiographic supervised machine-learning: ordered within-person train/test splits, person-specific models, pooled baselines on the same held-out rows, regression/classification metrics, row-level predictions, and coefficient extraction viacoefs().modelnames the statistical/ML model (for example,"ridge"), whileestimatornames the implementation/backend (default"native"). No new dependencies: native models include mean/majority baselines, OLS/logistic, ridge, lasso, elastic net, PCR, LDA, Gaussian naive Bayes, kNN, and one-split trees.fit_idiographic_ml()andfit_individualized_ml()remain aliases.
idiographic 0.1.0
- Initial CRAN submission.
- Provides idiographic network estimators for intensive longitudinal data, including ordinary VAR, graphical VAR, mlVAR, uSEM, and GIMME-style models.
- Includes preprocessing audits, rolling-window estimation, forecast validation, edge stability diagnostics, model comparison, tidy accessors, and cograph plotting support.