Skip to contents

idiographic

Person-specific statistics and dynamic networks for intensive longitudinal data — from raw repeated observations to within-person conclusions.

idiographic describes, prepares, models, validates, compares, and explains person-specific processes in ESM, EMA, diary, and other panel data. It combines person-level descriptions, LM/GLM and machine-learning prediction, within-between decomposition, coefficient pooling and shrinkage, subgroup discovery, treatment effects, and heterogeneity analysis with the package’s established VAR, graphical VAR, mlVAR, Bayesian DSEM, uSEM, and GIMME methods.

People and time order are first-class throughout: lags and validation splits do not cross person boundaries, future observations are kept out of training, and within-person quantities are not silently substituted for between-person ones. Pooled and subgroup models are available for comparison and stabilization of individual results.

Clean-room by design

The core estimators are native R implementations of the published modelling targets, with a consistent interface and validation against reference outputs where a reference implementation is available:

Estimator Method Validated against Agreement
fit_graphical_var() Regularized graphical VAR (graphical lasso + EBIC) graphicalVAR committed tolerance 1e-6 across the supported lag-1 beta/kappa option matrix
fit_mlvar() Multilevel and person-specific VAR mlVAR 0.7.3 committed tolerance 1e-8 across 20 real ESM panels plus fixed lmer lag 1/1+2, preprocessing, and lag-1 lm/unique oracle slices
fit_gimme() Group and individual uSEM path search gimme 10.0 exact search/matrix agreement on bivariate and three-variable standard/hybrid/VAR panels, including exogenous and uneven-panel structures; fit tables within 5e-5
fit_mlvar_bayes() Native Bayesian multilevel VAR / DSEM real Mplus DSEM + Stan/JAGS Monte-Carlo error
fit_var_bayes() Native Bayesian VAR(1) real Mplus ESTIMATOR = BAYES committed statistical bounds 0.02-0.03

The CRAN package is offline-first: its imports are standard R packages that ship with R. It has no mandatory third-party package dependency. lme4 and lavaan are optional engines for multilevel frequentist VAR and SEM/GIMME respectively; plotting and the licensed Mplus bridge are optional too. Competitor packages and the 20-panel oracle corpus live in the repository’s separate validation/ lane and are not shipped in the CRAN tarball.

The Bayesian DSEM sampler is a particular highlight: fit_mlvar_bayes() targets the output of mlVAR::mlVAR(estimator = "Mplus") — Mplus’s two-level Bayesian VAR with latent mean centring — without Mplus installed, using a pure-R conjugate Gibbs sampler with hand-rolled inverse-Wishart draws (no MCMCpack/rstan). The committed evidence consists of fixed bivariate Mplus fixtures, one univariate random-AR fixture, and parameter-recovery tests; use equivalence(fit) to inspect the precise scope rather than assuming blanket DSEM equivalence.

Installation

The core can be installed from a downloaded source tarball without network access; optional engines are only checked when their corresponding methods are called.

From CRAN:

install.packages("idiographic")

From the author’s r-universe (recommended — no compilation, binaries included):

install.packages("idiographic",
                 repos = c("https://mohsaqr.r-universe.dev",
                           "https://cloud.r-project.org"))

Or from GitHub:

# install.packages("pak")
pak::pak("mohsaqr/idiographic")

Plotting uses the cograph package; it stays optional and is offered for on-demand install the first time you call plot().

Quick start: one idiographic workflow 

library(idiographic)

## simulate an ESM panel: 30 people, 40 occasions, 3 predictors
set.seed(1)
panel <- do.call(rbind, lapply(1:30, function(id) {
  x <- matrix(rnorm(120), 40, 3)
  data.frame(id = id, time = 1:40, A = x[, 1], B = x[, 2], C = x[, 3])
}))
panel$Y <- 0.7 * panel$A - 0.3 * panel$B +
  rep(rnorm(30, sd = 0.5), each = 40) + rnorm(nrow(panel), sd = 0.4)

## 1. inspect variation and dependence person by person
describe_persons(panel, id = "id", vars = c("A", "B", "Y"), time = "time")
correlate_persons(panel, id = "id", vars = c("A", "B", "Y"))
variance_components(panel, id = "id", vars = c("A", "B", "Y"))

## 2. prepare within-person predictors and honest lags
prepared <- preprocess_panel(
  panel, id = "id", time = "time", vars = c("A", "B"),
  decompose = TRUE, lag = 1
)

## 3. compare pooled and person-specific regression on later observations
reg <- fit_lm(prepared, y = "Y", x = c("A", "B"), id = "id",
              time = "time", scope = "both")
metrics(reg, overall = TRUE)
coefs(reg, scope = "individual")

## 4. separate within-person and between-person effects directly
wb <- fit_within_between(panel, y = "Y", x = c("A", "B"), id = "id",
                         time = "time")
contextual(wb)

## 5. stabilize noisy individual coefficients
pool_coefs(reg)
shrink_coefs(reg)

## 6. fit and tune scoped machine-learning models
ml <- fit_ml(panel, y = "Y", x = c("A", "B", "C"), id = "id",
             time = "time", scope = "both",
             model = c("ridge", "knn"), tune = TRUE)
best_model(ml)
predictions(ml, scope = "individual")

Dynamic-network workflow

The statistical workflow and network estimators live in the same package and operate on the same person-by-time panels.

## multilevel VAR: temporal, contemporaneous, and between networks
net <- fit_mlvar(panel, vars = c("A", "B", "C"),
                 id = "id", beep = "time")

net                 # tidy printout of all three networks
edges(net)          # one row per edge (network, from, to, weight)
coefs(net)          # fixed-effect estimates with SE / p / CI
plot(net, layer = "temporal")

## the same call through the registry-driven front door
fit2 <- fit_idiographic(
  panel, method = "mlvar",
  params = list(vars = c("A", "B", "C"), id = "id", beep = "time")
)
equivalence(fit2)  # exact validation scope and tolerance declaration

## inspect the complete package and argument-by-argument evidence ledgers
equivalence_table()
argument_coverage("mlvar")

All fitting functions use named, readable arguments. list_estimators(), estimator_info(), and get_estimator() expose the dynamic-network and legacy ML registry; custom methods can be added with register_estimator(). models() is the separate algorithm/backend registry used by the consolidated ML engine. Classical scoped fitters (fit_lm(), fit_glm(), effects, within-between, and subgroup methods) remain explicit verbs because their results and inferential contracts are not interchangeable with network estimators. equivalence_table() and argument_coverage() report evidence for the methods in the dynamic estimator registry.

Together these ledgers provide complete evidence closure for registered methods: there are no unassessed registered methods or arguments. Numerical equivalence remains method- and configuration-specific rather than a blanket package claim.

Native Bayesian DSEM (no Mplus needed) 

bayes <- fit_mlvar_bayes(panel, vars = c("A", "B", "C"),
                         id = "id", beep = "time",
                         n_iter = 4000, n_chains = 2)
bayes               # posterior medians, SDs, 95% CIs, convergence (max PSR)
coefs(bayes)

## full DSEM with person-specific slopes, random residuals, and
## within-model imputation of missing observations (needs enough subjects to
## identify the random-effect covariance: at least 2 * (p + p^2) + 1):
fit_mlvar_bayes(panel, vars = c("A", "B", "C"), id = "id", beep = "time",
                temporal = "random", residual = "random", impute = TRUE)

What’s included

Descriptions, regression, and explanation

Dynamic-network estimators

  • fit_var() / fit_var_each() — ordinary VAR(1) (OLS), pooled or per subject
  • fit_graphical_var() / fit_graphical_var_each() — regularized graphical VAR (GLASSO + EBIC), including explicit multi-lag layers
  • fit_mlvar() — frequentist multilevel VAR with fixed, correlated, orthogonal, or unique person-specific temporal/contemporaneous structures
  • fit_mlvar_bayes() — native Bayesian multilevel VAR / DSEM (fixed or random slopes, fixed or random residual covariance, optional within-model imputation)
  • fit_var_bayes() — native Bayesian VAR(1)
  • fit_mlvar_mplus() — true-Mplus backend (wraps mlVAR(estimator = "Mplus"))
  • fit_usem() — unified Structural Equation Modeling (lavaan)
  • fit_gimme() — Group Iterative Multiple Model Estimation with explicit Bonferroni/FDR corrections, alpha, and stopping criteria

Workflow & diagnostics

Tidy contract

Scoped statistical results: metrics() · predictions() · coefs() · diagnostics() · importance() · tuning() · person() · individuals() · pooled() · subgroups() · overall()

Every result: as.data.frame() · summary() · print() where meaningful

Network results: edges() · nodes() · coefs() · matrices() · plot() / plot_gimme() · as_netobject()

Idiographic machine learning 

ml <- fit_ml(
  panel,
  y = "Y",
  x = c("A", "B", "C"),
  id = "id",
  time = "time",
  scope = "both",
  model = c("linear", "ridge", "knn"),
  tune = TRUE
)

ml                 # per-person and pooled held-out performance
metrics(ml)        # MAE / RMSE / bias / R-squared by subject and overall
coefs(ml)          # coefficients for each individualized and pooled model
predictions(ml)    # row-level held-out predictions

Use model = "all" to run all native models for the selected task. For regression this includes mean baseline, OLS (linear), ridge, lasso, elastic net, PCR, kNN, and a one-split tree. For binary classification this includes majority baseline, logistic regression, ridge/lasso/elastic-net logistic, LDA, Gaussian naive Bayes, kNN, and a one-split tree. Use estimator = "native" explicitly only when you want to pin the implementation; optional package backends live behind the same model name. Historical calls using outcome, predictors, day, and beep remain supported.

Migrating from idiostats

Use library(idiographic) in place of library(idiostats). Almost all public analysis verbs retain their names. Two collision bridges are explicit:

  • Use preprocess_panel() for the former idiostats::preprocess() transforms; preprocess() remains the established network-readiness audit.
  • Named fit_ml(y = ..., x = ...) calls use the consolidated scoped engine. Use fit_ml_panel(data, y, x, id, ...) when retaining the former positional calling style. Positional fit_ml(data, outcome, predictors, id) remains the historical idiographic interface for backward compatibility.

Consolidated results report idiographic_* as their primary class and retain the former idiostats_* class as a compatibility bridge.

Bundled data

  • srl — a self-regulated-learning ESM dataset (data(srl))
  • inst/extdata/esm_demo.tsv — a small synthetic demo panel

Documentation

Package page and binaries: https://mohsaqr.r-universe.dev/idiographic.

Start with the statistical workflow guide, which maps research questions to the 16 major non-network functions. Each function then has a detailed worked vignette:

Analysis Function-specific vignette
Person-level description describe_persons()
Within-person correlation correlate_persons()
Variance decomposition variance_components()
Panel preparation preprocess_panel()
Linear modelling fit_lm()
Generalized linear modelling fit_glm()
Machine learning fit_ml()
Rolling-origin validation fit_rolling()
Within-between modelling fit_within_between()
Coefficient pooling pool_coefs()
Coefficient shrinkage shrink_coefs()
Subgroup-existence testing test_subgroups()
Subgroup discovery find_subgroups()
Subgroup modelling fit_subgroups()
Treatment-effect estimation fit_effects()
General heterogeneity analysis fit_heterogeneity()

Citation

Saqr, M., & López-Pernas, S. (2026). idiographic: Person-Specific Statistics and Heterogeneous Dynamic Networks. R package. https://github.com/mohsaqr/idiographic

License

GPL-3.