JSS reviewer start here • gamlss.longitudinal

This article is the reviewer-facing entry point for the JSS submission version of gamlss.longitudinal. It connects the paper, package interface, diagnostics, examples, and replication workflow.

What the package does

gamlss.longitudinal fits longitudinal GAMLSS models with first-order copula dependence between adjacent repeated measurements. The main interface is gamlss_longitudinal(). Each supported marginal parameter and copula parameter can have its own formula, including fixed and smooth covariate effects.

The package is intended for analyses where the marginal distributional shape or within-subject dependence is part of the scientific question. It is not an automatic imputation tool, an arbitrary higher-order vine fitting package, or a time-series forecasting engine beyond prediction on supplied new panels.

Quick install check

After installing the package, run the smoke test:

source(system.file("smoke-tests", "new-user-smoke.R",
  package = "gamlss.longitudinal"
))

Recommended workflow

The main user workflow is:

Inspect the longitudinal response and missingness pattern.
Screen candidate marginal distributions.
Screen candidate copula families.
Fit the longitudinal GAMLSS-copula model.
Inspect convergence, coefficients, and uncertainty.
Check marginal and dependence diagnostics.
Predict distributional summaries or simulate from the fitted model.

The minimal workflow article demonstrates this path:

vignette("standard-workflow", package = "gamlss.longitudinal")

For a longer known-truth example, use:

vignette("native-simulation-workflow", package = "gamlss.longitudinal")

Main interface

A typical model fit has separate formulas for marginal and dependence parameters:

fit <- gamlss_longitudinal(
  mu.formula = response ~ time + group + s(baseline),
  sigma.formula = ~ time + group,
  theta.formula = ~ time + group,
  dataset = dat,
  time_var = "time",
  subject_var = "id",
  margin_dist = GA(),
  copula_dist = "t"
)
summary(fit)

Use the final JSS paper example for the exact family, copula, and covariates.

Standard methods to inspect

Reviewers should be able to inspect fitted objects through standard R methods:

coef(fit)
vcov(fit)
confint(fit)
logLik(fit)
nobs(fit)
formula(fit)
terms(fit)
fitted(fit)
residuals(fit)
model.frame(fit)

Prediction and simulation use standard generics:

predict(fit, newdata = newdat, type = "response")
simulate(fit, nsim = 10)

Diagnostics

Marginal diagnostics and dependence diagnostics are both required for this model class:

plot(fit)
plot_terms(fit)
plot_margin_fit(fit)
plot_copula_diagnostics(fit)
check_model(fit)
check_missingness(dat, subject_var = "id", time_var = "time")

See the diagnostics guide for interpretation:

vignette("diagnostics-decisions", package = "gamlss.longitudinal")

Inference and uncertainty

The inference guide covers Wald intervals, likelihood comparisons, bootstrap helpers, and uncertainty language:

vignette("inference-uncertainty", package = "gamlss.longitudinal")

Paper replication

The repository version of the paper uses paper/replicate.R as the authoritative replication entry point. From the repository root:

source("paper/replicate.R")

The smoke profile is intended as a fast reviewer check. The expanded profile is the full paper regeneration:

Sys.setenv(GAMLSS_LONGITUDINAL_JSS_PROFILE = "expanded")
source("paper/replicate.R")

Generated outputs are written under results/jss-replication/<profile>/. The manifest maps paper result IDs to files, and the logs include session information and output hashes.

Private application data are not committed to the repository. The main JSS paper should therefore use a public, package-shipped, or simulated primary example unless accepted external data access instructions are provided.