Antedependence Models for Longitudinal Data

Fitting, simulation, and inference for antedependence models for longitudinal data, as described in Zimmerman and Nunez-Anton (2009, ISBN:9781420011074). Supports integer-valued antedependence (INAD) models for count data with thinning operators (binomial, Poisson, negative binomial) and flexible innovation distributions (Poisson, Bell, negative binomial), categorical antedependence models for discrete-state longitudinal outcomes, and Gaussian antedependence (AD) models for continuous data. Implements maximum likelihood estimation via time-separable optimization and block coordinate descent, with confidence intervals based on Louis' identity and profile likelihood.

antedep

antedep fits antedependence models for longitudinal data:

• Gaussian AD models for continuous outcomes
• INAD models for count outcomes
• Categorical AD (CAT) models for discrete state outcomes

Installation

# install.packages("remotes")
remotes::install_github("TanchyKing/antedep")

Production-Readiness Matrix

Quick Start

Included datasets:

• bolus_inad (morphine bolus counts)
• cattle_growth (book companion cattle growth data; continuous)
• cochlear_implant (book companion speech recognition/cochlear data; continuous)
• labor_force_cat (labor-force categorical table expanded to subject sequences)
• race_100km (100km race split times; continuous)

AD example (continuous)

library(antedep)
set.seed(1)

y <- simulate_gau(n_subjects = 80, n_time = 5, order = 1)
fit <- fit_gau(y, order = 1)
fit$log_l

Missing-data workflow

library(antedep)
set.seed(1)

y <- simulate_inad(n_subjects = 60, n_time = 5, order = 1)
y[sample(length(y), 20)] <- NA

# Fit observed-data likelihood under MAR
fit_miss <- fit_inad(y, order = 1, na_action = "marginalize")
fit_miss$log_l

CAT missing-data workflow

library(antedep)
set.seed(1)

y_cat <- simulate_cat(n_subjects = 80, n_time = 5, order = 1, n_categories = 3)
y_cat[sample(length(y_cat), 30)] <- NA

# Observed-data likelihood (orders 0/1/2)
fit_cat_marg <- fit_cat(y_cat, order = 1, na_action = "marginalize")

# EM (orders 0/1) via explicit entry point or fit_cat dispatch
fit_cat_em1 <- em_cat(y_cat, order = 1, max_iter = 80)
fit_cat_em2 <- fit_cat(y_cat, order = 1, na_action = "em", em_max_iter = 80)

If EM becomes unstable or converges slowly, try increasing epsilon, increasing max_iter, and using safeguard = TRUE.

Known Limitations

• EM entry points: em_gau (Gaussian), em_inad (INAD), and em_cat (CAT, orders 0/1) are available; for CAT order 2 with missing data, use fit_cat(na_action = "marginalize").
• For CAT models, fit_cat() supports both observed-data likelihood (na_action = "marginalize") and EM (na_action = "em" for orders 0/1).
• Missing-data confidence intervals are not yet implemented (ci_gau, ci_inad, ci_cat require complete-data fits).
• AD missing-data EM for order 2 is currently available with simplified updates and should be treated as provisional.
• AD missing-data LRT/mean/covariance tests remain complete-data only.
• CAT missing-data stationarity/time-invariance tests (lrt_stationarity_cat, lrt_timeinvariance_cat, run_stationarity_tests_cat) remain complete-data only.
• INAD missing-data LRTs and CAT missing-data order/homogeneity LRTs are supported via observed-data likelihood (na_action = "marginalize").
• Current model order support is up to order 2 for AD, INAD, and CAT.
• CAT missing-data marginalization currently supports orders 0, 1, and 2.

Vignette

Source vignette: vignettes/antedep-intro.Rmd. Rendered site article: docs/articles/antedep-intro.html. Integrated function reference: docs/reference/index.html.

Function Reference Site (pkgdown)

• Local build: Rscript -e "pkgdown::build_site(preview = FALSE)"
• CI deployment: .github/workflows/pkgdown.yml (GitHub Pages via Actions)

Local Check Guide

• See LOCAL_CHECKS.md for tarball-mode vs directory-mode checks and expected NOTE messages.