Tidy Intensive Longitudinal Data Analysis.

tidyILD

A reproducible, tidyverse-style framework for intensive longitudinal data (ILD) analysis in R, with built-in methodological safeguards, provenance tracking, and reporting tools.
Author: Alex Litovchenko.

Install

remotes::install_github("alitovchenko/tidyILD")

Backend validation benchmarks (developers / CI)

From a source checkout of the package (directory containing DESCRIPTION), optional cross-backend simulation benchmarks can be run locally:

Rscript scripts/run-backend-validation-benchmarks.R --tier smoke --out-dir /tmp/bench
Rscript scripts/check-backend-validation-thresholds.R \
  --summary /tmp/bench/benchmark_summary.csv \
  --thresholds inst/benchmarks/thresholds-smoke.json \
  --out /tmp/bench/benchmark_checks.csv

This uses pkgload::load_all() when available. Tiers:smoke (fast), nightly, full — see inst/dev/BACKEND_VALIDATION_BENCHMARK_CONTRACT.md. GitHub Actions workflow: .github/workflows/backend-validation-benchmarks.yml (scheduled + manual dispatch; artifacts uploaded).

Quick start

library(tidyILD)

# Prepare: validate time structure, add .ild_* columns and metadata
d <- data.frame(
  id = rep(1:3, each = 5),
  time = rep(as.POSIXct(0:4 * 3600, origin = "1970-01-01"), 3),
  mood = rnorm(15)
)
x <- ild_prepare(d, id = "id", time = "time", gap_threshold = 7200)

# Inspect (summary tibble + list)
ild_summary(x)

# Within-between decomposition
x <- ild_center(x, mood)

# Spacing-aware lags (max_gap from metadata if omitted)
x <- ild_lag(x, mood, mode = "gap_aware", max_gap = 7200)

# Missingness (summary tibble + plot + by_id)
ild_missing_pattern(x, vars = "mood")

# Fit and report: tidy fixed effects, fitted vs observed, residual ACF + QQ
fit <- ild_lme(mood ~ 1 + (1 | id), data = x, ar1 = FALSE, warn_no_ar1 = FALSE)
tidy_ild_model(fit)
ild_plot(fit, type = "fitted")
ild_plot_predicted_trajectory(fit, time_var = ".ild_seq") # observed + fitted vs time
diag <- ild_diagnostics(fit); diag; plot_ild_diagnostics(diag)
bundle <- ild_diagnose(fit) # diagnostics bundle; use ild_autoplot(bundle, ...)

Pipeline

• ild_prepare() — encode longitudinal structure, spacing, gaps
• ild_summary() — one-shot summary
• ild_center() — person-mean centering (WP/BP)
• ild_lag() — index, gap-aware, or time-window lags (supports lubridate::hours(2) etc.)
• ild_decomposition() — WP/BP variance and ratio; optional WP vs BP density plot
• ild_check_lags() — lag validity (valid/invalid, pct_invalid, lag order)
• ild_panel_lag_prepare() — several ild_lag() columns + one lag audit table
• ild_crosslag() — one-call cross-lag: lag predictor, check lags, fit outcome ~ lag
• ild_compare_fits() — AIC/BIC/nobs table for a list of models (optional guardrail counts)
• ild_brms_dynamics_formula() — template brms formula for random lag slopes (does not fit)
• ild_acf() — ACF on a variable or on residuals (pre-model check for AR1)
• ild_spacing_class() — regular-ish vs irregular-ish
• ild_spacing() — spacing diagnostics (median/IQR, large gaps %, CV) and AR1/CAR1 recommendation
• ild_design_check() — aggregate spacing, WP/BP, missingness, and recommendations
• ild_center_plot() — standalone WP vs BP density plot
• ild_missing_pattern() — missingness by person/variable
• ild_missing_bias() — test if missingness is associated with a predictor (informative missingness)
• ild_align() — align secondary stream (e.g. wearables) to primary ILD within a time window
• ild_lme() — mixed-effects model (lmer or nlme with AR1/CAR1)
• ild_robust_se() — cluster-robust variance (clubSandwich); tidy_ild_model(fit, se = "robust") for robust SE/CI/p
• ild_missing_model() — model missingness from covariates; ild_ipw_weights() and ild_ipw_refit() for IPW sensitivity
• ild_iptw_weights(), ild_iptw_msm_weights(), ild_ipcw_weights(), ild_joint_msm_weights() — treatment/censoring/joint MSM weight pipelines
• ild_msm_estimand() + ild_msm_fit() — estimand-first MSM runner with explicit inference capability status and strict_inference
• ild_msm_bootstrap() / tidy_ild_msm_bootstrap() — cluster bootstrap inference for weighted lmer (fixed or re-estimated weights)
• ild_msm_balance(), ild_ipw_ess(), ild_msm_overlap_plot() — weighted balance, effective sample size, and overlap diagnostics
• ild_msm_diagnose() + ild_msm_contrast_over_time() — one-call diagnostics bridge and time-indexed post-fit contrasts
• ild_msm_simulate_scenario() + ild_msm_recovery() — causal simulation and recovery harness with scenario-grid validation
• ild_ctsem() — continuous-time latent-dynamics backend with ild_tidy(), ild_augment(), ild_diagnose(), and ild_autoplot()
• ild_tvem() — time-varying effects (GAM); ild_tvem_plot() for the coefficient curve
• Temporal model choice:vignette("temporal-dynamics-model-choice", package = "tidyILD") (lags vs AR vs TVEM vs state-space); vignette("brms-dynamics-recipes", package = "tidyILD") for Bayesian templates
• ild_person_model() — fit model per person (N-of-1); ild_person_distribution() — plot distribution of estimates
• ild_diagnostics() — residual ACF, residuals vs fitted/time (use print() for summary)
• plot_ild_diagnostics() — build diagnostic plots from an ild_diagnostics object
• ild_plot() — trajectory, heatmap, gaps, missingness, fitted vs observed, predicted trajectory (observed + fitted vs time), residual ACF; optional facet_by for panels (e.g. cluster)
• ild_plot_predicted_trajectory() — convenience wrapper for predicted trajectory vs time_var
• ild_heatmap(), ild_spaghetti() — aliases; pass facet_by through to ggplot2::facet_wrap()
• ild_diagnose() — assemble ild_diagnostics_bundle (data, design, fit, residual, …); ild_autoplot(bundle, ...) for sectioned plots
• ild_circadian() — variable by hour of day (when time is POSIXct)
• augment_ild_model() / ild_augment() — tibble per ild_augment_schema(): .outcome, .fitted, .resid, .resid_std (Pearson when available), engine, model_class, etc.
• tidy_ild_model() — fixed-effect table (estimate, SE, CI, p); use se = "robust" for cluster-robust inference
• ild_simulate() — simulated ILD (n_id, n_time/n_obs_per, ar1, wp_effect, bp_effect, irregular)
• ild_power() — simulation-based power for a fixed effect (ild_simulate → ild_lme → effect recovery)
• ema_example — built-in dataset (data(ema_example))
• broom: Load broom.mixed for tidy(fit) and augment(fit) on ild_lme fits.

Vignettes

• Specialist backends (vignette("ild-specialist-backends", package = "tidyILD")): when to move beyond lme4 / default tidyILD fits; handoffs to dynamite, PGEE, DSEM, and multivariate workflows; export patterns after ild_prepare() / ild_center() / ild_lag().
• Visualization in tidyILD: index of plots, bundle sections, facet_by, predicted trajectories, and partial-effects templates (marginaleffects / ggeffects on _wp / _bp).
• From raw data to model: full pipeline with ild_prepare() through ild_lme() and ild_plot().
• MSM identification and recovery: assumptions, estimand-first MSM workflow, strict/degraded inference behavior, and scenario-grid recovery checks.
• Continuous-time dynamics with ctsem: ild_ctsem() workflow, diagnostics, and guardrails for ctsem fits.
• Short analysis report: fit, tidy fixed effects, fitted vs observed, residual ACF and Q-Q.
• Within-between decomposition and irregular spacing: centering and gap-aware lags.
• Glossary and quick-start: function glossary and checklist.

pkgdown site

Documentation and vignettes are built with pkgdown. From the package root: pkgdown::build_site(). Config: _pkgdown.yml.

Version tags

Release tags follow vMAJOR.MINOR.PATCH and are listed in version order with:

git tag -l 'v*' --sort=version:refname

Current tags: v0.0.1, v0.2.0, v0.3.0 (each annotated message is tidyILD <version>).

License

MIT.