NEWS

---

BayesRTMB 0.4.0 (2026-09-10)

• Added MCMC_Fit$continue_sampling() to extend existing NUTS chains without another warmup. It resumes from each chain's final unconstrained state, reuses the adapted step size and mass matrix, and can update the fit in place or return an independently extended copy.
• Corrected fixed-effect standard errors for Gaussian rtmb_lmer() models fitted with classic(). REML fits now use the generalized least-squares covariance at the estimated variance components, matching lmerTest/lme4.
• Corrected rtmb_corr(prior_flat()) marginal-likelihood normalization so Bayes factors for correlations use the normalized LKJ(1)/uniform correlation prior without changing the printed flat-prior model code.
• Made posterior_predict() and pp_check() report a clear error for classic() results, which do not store posterior draws.
• Fixed simulation-based standard errors for one-parameter models, including optimize(se_method = "sampling") on scalar binomial models.
• Changed plot_mdu(distance = "auto") to use Euclidean distance when no distance metadata is available. Stored model distance settings continue to take precedence.
• Clarified posterior-prediction examples in the Japanese wrapper vignette.

BayesRTMB 0.3.0 (2026-08-20)

• Added posterior predictive simulation and checks through posterior_predict() and pp_check(). Regression wrappers automatically support continuous density checks, discrete binned-bar checks, scalar test statistics, predictor-based and fitted-value calibration scatter checks, and conditional, population-level, or newly simulated random effects. Custom models can provide replicated outcomes in a generate block.
• Added random-intercept mediation models to rtmb_mediation(). A random intercept can be included in one equation or in multiple equations; random intercepts sharing a grouping variable are modeled jointly with an estimated correlation matrix.
• Added grand-mean centering (gmc/centering) and centering within cluster (cwc) to rtmb_mediation(). Centering is applied to predictor uses while preserving response variables on their original scale; cluster means remain user-specified model terms.
• Corrected degrees of freedom for classical mediation models. Fixed Gaussian equations now use the rank of their own design matrix, derived effects inherit degrees of freedom from their contributing coefficients, and Gaussian random-intercept mediation models use Satterthwaite degrees of freedom by default.
• Added the setup-only .data binding for accessing the original object passed to rtmb_model(). Wrapper-generated setup code now reads matrices and data frames directly, or reads related inputs such as responses, IDs, covariates, and choice sets from one named data list. Structural options remain visible as fixed assignments in the generated code.
• Added std = TRUE to the regression wrappers. It reports post-hoc standardized fixed-effect coefficients as b_std, using every column of the fixed-effect design matrix, including factor and interaction columns.
• Added summary_mcmc() for summarizing numeric MCMC arrays arranged as iterations by chains by variables. It provides the same summary columns and print format as MCMC_Fit$summary(), with parameter and chain selection.
• Added exported center_grand_mean() and center_within_cluster() helpers for generated wrapper code and hand-written models.
• Made regression-wrapper print_code() expose data-frame columns instead of a nested data/formula payload, with explicit formula preprocessing and readable grand-mean and within-cluster centering helpers.
• Fixed rtmb_table(x, y, data = ...) so unquoted column names are resolved from the supplied data frame while preserving reproducible generated code.
• Simplified factor-analysis and multidimensional-unfolding print_code() output by assigning nfactors and ndim values directly to the internal constants K and D.
• Updated rtmb_corr() normal priors to use the model-specific mean_sd and sd_rate aliases and to add an LKJ prior with lkj_eta = 1 by default.
• Corrected Jacobian adjustments for constrained parameters supplied through fixed, so elements removed from the free-parameter map no longer contribute to optimization or sampling target densities.
• Updated the English and Japanese vignettes to document .data, unified named data inputs, structural constants in setup, and reproducible wrapper-generated code.

BayesRTMB 0.2.4 (2026-07-24)

• Allowed rtmb_lm(), rtmb_glm(), rtmb_lmer(), and rtmb_glmer() to resolve bare variable names from the formula environment when data is omitted. Formulas using $, [[, or . continue to require explicit data.
• Corrected JZS t-test priors to include the Jeffreys scale prior (p(\sigma) \propto 1/\sigma), including paired and unequal-variance models.
• Corrected exponential-prior scaling across wrappers. The default sigma_rate and tau_rate in prior_normal() are now 1 / 5, giving exponential priors with mean 5, and wrapper-specific aliases are applied consistently.
• Added the default discrimination prior a ~ exponential(1 / 2) when prior_normal() is used with IRT models.
• Improved weak-prior calibration from y_range in mixture and latent-rank models, including response-specific ranges for multivariate outcomes.
• Improved multivariate normal log-density performance and removed an unintended blank line from optimization progress output.
• Updated English and Japanese vignettes for the new defaults and wrapper behavior.

BayesRTMB 0.2.3 (2026-07-13)

• Fixed factor-analysis model construction on platforms where RTMB is sensitive to AD matrix containers that include structural zeros from lower-triangular parameters. The rtmb_fa() wrapper now avoids reading upper-triangular structural-zero entries of lower_tri loading matrices and constructs constrained AD matrices with rtmb_array().
• Simplified the runnable rtmb_fa() example to a one-factor model. Advanced factor-analysis workflows remain covered by documentation and CI regression checks.

BayesRTMB 0.2.2

• Added response-time distributions for model code: exp_mod_normal_lpdf() and diffusion_lpdf(), with sampling syntax support via exp_mod_normal(...) and obs(RT, Choice) ~ diffusion(...).
• Added obs(...) sampling syntax for multivariate observed values on the left side of ~.
• Improved setup helper capture so functions referenced in setup are more reliably available when building models and running parallel workers.
• Improved factor-analysis AD robustness and rotation output naming.

BayesRTMB 0.2.1 (2026-06-23)

• Added upgrade_fit() to rebuild saved MCMC, VB, MAP, and classic fit objects with the currently loaded class definitions, optionally upgrading their embedded model objects as well.
• Improved rtmb_vector() and rtmb_array() tape construction time by automatically reusing an AD seed from model parameters when available.
• Made log_sum_exp(), softmax(), and log_softmax() work more reliably with RTMB automatic-differentiation values, including baseline-category patterns such as softmax(c(0, eta)) inside rtmb_code().
• Improved wrapper-generated model code to use AD-compatible rtmb_vector() and rtmb_array() containers in loop-filled generated quantities and generated likelihood contributions where needed.
• Improved report() handling in transformed and generated quantities, including namespaced BayesRTMB::report() calls and wrapper-generated print_code() output.
• Changed VB point estimates to use only the best variational estimate by default, aligning EAP(), MAP(), and rotation references with the selected best ELBO run while still allowing explicit chains or best_chains selection.
• Made EAP() and MAP() drop their list wrapper by default when a single parameter is requested, matching the behavior of estimate().
• Added optional taped evaluation for transformed parameters and generated quantities, with automatic fallback to R evaluation when taping is not possible.
• Improved parallel worker robustness by reducing exported globals and preserving wrapper setup environments needed by generated model code.
• Improved MCMC runtime behavior by caching metric calculations, speeding up retained draw conversion, and refining progress checks.
• Changed bootstrap progress reporting to use percentage-style progress output, consistent with other long-running workflows.
• Added diagnostic recommendations to help interpret common fitting warnings.
• Improved matrix-valued Gaussian process log-density evaluation.
• Updated MDU defaults and internals, including Euclidean distance as the default MDU distance and more explicit use of namespace-qualified factor rotations.

BayesRTMB 0.2.0

• Reworked NUTS sampling internals with Stan-style multinomial tree expansion, warmup diagnostics, Stan-window metric adaptation, and support for diagonal, dense, hybrid, and automatic metric selection.
• Improved MCMC diagnostics by reporting divergence counts and percentages, metric auto-selection details, warmup summaries, metric condition numbers, and positive-definite fallback counts.
• Added configurable progress output for MCMC and VB workflows, including streamed message-style progress for parallel workers and percentage reporting.
• Added delta-method standard errors and confidence intervals for conditional_effects() and simple_effects() with optimized and classic fits; simple_effects() for classic fits now also reports df, t value, and Pr.
• Added sd_slice and sd_multiplier controls for conditional and simple effects, including automatic SD slicing for moderators with many observed values.
• Added rhat_summary() for MCMC fits, returning a numeric R-hat vector with a compact printed summary.
• Expanded data-reshaping helpers: to_long() now supports multiple value columns, list-based column groups, and preserves input row order by default while still allowing sorted output with sort = TRUE.
• Added AD-compatible helper constructors rtmb_vector() and rtmb_array() for model code that needs mutable RTMB-compatible containers.
• Improved RTMB model setup error messages for common AD and NA/NaN failures.
• Improved wrapper behavior, including rtmb_glmer(cwc = list(ID, "all")), hierarchical lambda in rtmb_mdu(), stronger prior validation, and more robust handling of non-finite VB optimization attempts.
• Improved MDU plotting and initialization, including principal-axis reference rotation and clearer radius display controls.
• Fixed several model-specific issues, including AD-compatible negative-binomial log densities and unequal-variance JZS t-test examples.

BayesRTMB 0.1.1 (2026-06-01)

• Addressed CRAN resubmission feedback by quoting 'RTMB' in the package title, adding method references to DESCRIPTION, and replacing \dontrun{} examples with \donttest{} where appropriate.
• Trimmed long-running examples for CRAN checks while retaining representative MCMC examples for correlation, t-test, and mixed-model workflows.
• Updated IRT post-estimation examples to use ordered response data explicitly.
• Fixed an AD-compatible negative-binomial log-density issue.
• Corrected the unequal-variance JZS t-test example and documentation.