gp3bayes

gp3bayes is an independent R package for transparent, contract-first Bayesian workflows for repeated-measures and hierarchical behavioural data, with a governed dynamic-pupillometry layer and explicit validation, sensitivity, prediction, and reporting infrastructure.

Current GitHub release: 0.6.0 (v0.6.0, 20 September 2026)

Website · Reference · Articles · Release

What gp3bayes provides

• explicit model contracts, readiness audits, transformation replay, prior specifications, and prior-predictive checks;
• restricted hierarchical Bernoulli-logit and positive uncensored lognormal-duration workflows;
• backend-portable full-MCMC fitting through brms with rstan or cmdstanr where explicitly requested;
• posterior diagnostics, prediction, calibration, PSIS-LOO/K-fold validation, recovery, SBC, sensitivity, and publication evidence;
• governed dynamic pupillometry covering temporal dependence, robust distributions, Gaussian-process trajectories, measurement uncertainty, missingness, binocular modelling, response-shape models, and predictive calibration;
• structured model cards, evidence inventories, figures, tables, dashboards, and reproducibility manifests without automatic adequacy, robustness, exclusion, or model-selection claims.

September 2026: retention-first Bayesian SCR responsivity

gp3bayes 0.6.0 adds estimate_scr_responsivity_bayes() for participant-level SCR responsivity accountability.

The function keeps the conventional amplitude-threshold non-responder flag for provenance, estimates a Beta-Binomial posterior response probability, and sets retain_for_modeling = TRUE for every participant with finite trial data. Low-reactive participants are therefore retained in the primary modelling dataset, while hard exclusion can be evaluated explicitly as a sensitivity specification rather than being imposed silently during preprocessing.

Posterior responsivity is a graded modelling quantity, not a diagnostic or psychological label. The implementation is dependency-light and does not claim to reproduce a Dirichlet-process mixture model.

See the Retention-first Bayesian SCR responsivity article and estimate_scr_responsivity_bayes() reference.

September 2026: trial-level multilevel gaze mediation

gp3bayes 0.6.0 adds governed repeated-measures mediation for trial-level gaze/process mechanisms. The workflow consumes within/between components prepared upstream by eyeprocess; it does not silently reconstruct or aggregate the mediator inside the Bayesian package.

The new surface supports Gaussian, lognormal, Gamma, beta, Bernoulli, and count mediator families; Gaussian, Bernoulli, count, and ordinal outcomes where supported; explicit weakly informative priors; participant random effects; within- and between-participant indirect effects; convergence gating; prior/posterior predictive checks; participant-specific effects when both random path slopes are estimated; and conservative serial/moderated extensions.

For nonlinear outcomes, coefficient-product indirect effects are reported on the model’s linear-predictor scale and are not mislabeled as probability-scale natural indirect effects. Missing gaze is never interpreted automatically as zero, and non-estimable between-person paths are omitted rather than filled by priors.

See the Bayesian multilevel gaze mediation article and the mediation reference section.

Bayesian dynamic pupillometry

gp3bayes 0.5.0 extends the contract-first workflow with governed vendor-neutral pupil time-course analysis. Pupil data are explicitly mapped and audited; the package does not silently interpolate missing samples, infer cognitive states, or select a preferred model automatically.

The 0.5 layer includes:

• Gaussian and Student-t observation models with governed residual-scale structures;
• AR/ARMA temporal dependence and governed Gaussian-process trajectories;
• measurement-error declarations and MAR-oriented missing-value models;
• binocular preparation and modelling without requiring upstream eye averaging;
• experimental nonlinear pupil response-shape models;
• temporal derivatives, dynamic contrasts, declared threshold-duration estimands, and posterior trajectories;
• PSIS-LOO, exact K-fold, leave-future-out planning, predictive calibration, model comparison, model cards, and sensitivity suites.

Start with the Bayesian dynamic pupillometry article or browse the advanced pupillometry 0.5 articles.

Model-family scope

The original contract-first core remains deliberately restricted to:

1. hierarchical Bernoulli-logit models for binary trial-level outcomes; and
2. hierarchical lognormal models for strictly positive uncensored durations.

Advanced pupil models are separately governed by their explicit pupil specifications and evidence gates. No interface accepts unrestricted formulas or arbitrary model families under an existing approved function name.

Installation

Install the exact GitHub 0.6.0 release:

install.packages("remotes")
remotes::install_github("stefanosbalaskas/gp3bayes", ref = "v0.6.0")

To install the current master branch:

remotes::install_github("stefanosbalaskas/gp3bayes")

Optional Bayesian backends are only required for workflows that actually fit models.

Minimal contract example

library(gp3bayes)

contract <- create_model_contract(
  family = "binary",
  outcome_col = "selected",
  participant_col = "participant_id",
  item_col = "stimulus_id",
  trial_col = "trial_id",
  condition_col = "condition"
)

audit_model_readiness(data, contract)

Creating a contract does not validate a substantive hypothesis, fit a model, or establish model adequacy.

Release validation

The v0.6.0 release record includes:

• 484 public exports;
• 230 S3 registrations;
• frozen pre-0.6 public API compatibility contracts remain enforced;
• 467 Rd files and 62 vignette sources;
• full backend-free test suite passed; and
• release-preparation R CMD check --as-cran: 0 errors, 0 warnings, 0 notes.

The immutable v0.5.0 archive remains unchanged; the September SCR responsivity and multilevel mediation additions form part of the 0.6.0 release scope.

Interpretation boundaries

Behavioural, gaze, pupil, EDA/SCR, cardiovascular, and other physiological measurements do not directly reveal emotion, stress, cognition, comprehension, personality, diagnosis, deception, intention, or another latent psychological state.

A fitted model, favorable diagnostic, posterior probability, predictive score, or model weight does not by itself establish causal identification, model adequacy, robustness, or substantive validity. Automatic participant exclusion and automatic model selection remain outside the package’s governed decision boundary.

Citation

Use the citation metadata supplied with the installed package:

citation("gp3bayes")
packageVersion("gp3bayes")

The software concept DOI is 10.5281/zenodo.21518698.

Licence

MIT License.