Every function keeps its name, its arguments and the names of its results, so existing scripts run unchanged. Results print, summarise and plot (print(), summary(), plot()), and any component converts to a table with data.table::as.data.table(x, component = "...").
bayes_ammi(): new arguments .k (number of bilinear terms, default 2 as in Crossa et al. 2011; earlier versions fitted all min(G, E) - 1), .burnin, .thin, .chains, .priors (bam_priors()), .level, .seed (the random number state of the session is restored) and .align (sign of every term); .nIter now has a default (1000). New components summary (posterior mean, SD, Monte Carlo error, HPD interval, effective sample size, potential scale reduction factor of every parameter), scores, draws (coda::mcmc.list), diagnostics (Gelman-Rubin, Raftery-Lewis, Heidelberger-Welch), ls, settings.
tau1 now holds the draws of the error variance sigma^2 (it held a constant precision).
biplots(): HPD regions of the biplot coordinates lambda_k^(1/2) alpha_ik of the paper, at one or more levels (pb = c(0.05, 0.01)); new arguments .which, .grid; biplot_data gains axis1_lower, axis1_upper, axis1_significant and group.
Genotypes and environments keep their labels in all tables.
Beyond the article
Unbalanced trials: missing plots and whole genotype-environment cells are handled by data augmentation inside the sampler (earlier versions required balanced data).
bayes_ammi(): new arguments .model ("AMMI" or "GGE", the sites regression model of Perez-Elizalde et al. 2012), .het (an error variance for every environment; the environment scores are then drawn from their Bingham-von Mises-Fisher conditional) and .waic; new component fit (DIC, WAIC and the contribution of every plot to WAIC).
Decisions: bam_predict() (posterior of every cell and genotype mean, including empty cells), bam_winners() (probability that each genotype wins in each environment, groups of environments), bam_superiority() (probabilities of superior performance and stability, Dias et al. 2022), bam_ranks() and bam_stability() (Bayesian AMMI stability value, interaction variance, yield stability index).
Model choice: bam_compare() (DIC, WAIC and the standard error of WAIC differences) and bam_ppc() (posterior predictive checks).
as_draws_df(), as_draws() and as.mcmc.list() methods, for the posterior, bayesplot and coda packages.
bam_app() and bam_app_object(): a built-in shiny application with tabs for the interpretation, the plot, the table and the printed output, each with copy buttons, and the R code of the fit.
All plots are ggplot2 with scales::pretty_breaks(). bam_save() writes lossless WebP; the HTML vignettes and the application use lossless WebP images (about half the size of PNG).
Bayesian stability analysis: bayes_stability() and bam_stab_priors() fit the Finlay-Wilkinson (1963), Eberhart-Russell (1966) and Shukla (1972) models by a partially collapsed Gibbs sampler (compiled, with an R reference giving the same draws), with the environmental index estimated jointly, unbalanced data, slopes, deviation and stability variances, the safety-first index of Eskridge (1990), DIC and WAIC. The decision tools, bam_compare() and bam_ppc() accept these fits.
bam_indices(): posterior distributions of Wricke’s ecovalence, Shukla’s stability variance, Lin and Binns’ superiority, Annicchiarico’s reliability and Eskridge’s safety-first index, from any fit. bam_ammi_indices(): ASV, SIPC, EV, Za, WAAS and WAASY from AMMI or GGE fits. Each index has HPD intervals, probabilities of being best and expected ranks.
Citations: every printed result and interpretation ends with the sources of its method; bam_cite() gives them as text or BibTeX; citation("bayesammi") lists the package and every method; inst/REFERENCES.bib has all references. The application shows the sources under each interpretation.
The application gains a Stability menu (regression models, stability indices, AMMI indices).
Seven new vignettes: unbalanced trials and environment-specific variances, the GGE model, decision tools, the Bayesian stability models, posterior stability indices, the application, and the references of the methods.
Corrections
bayes_ammi(): the singular values were computed once before the loop and never sampled (lambdas1 was constant); they are now drawn every iteration from their truncated normal conditional.
bayes_ammi(): the residuals for sigma^2 compared every cell with the environment of a data row instead of its own environment (with Maiz, always site 1); the within-cell variances were summed in place of the sums of squares; the covariance matrices of the main effects were square roots of variances; the environment effects used B E in place of B G.
bayes_ammi(): the scores were drawn on the unconstrained Stiefel manifold and then orthogonalised; they are now von Mises-Fisher draws on the constrained spheres (Crossa et al. 2011, eq. 8-9).
bayes_ammi(): the prior of sigma^2 was improper (0 degrees of freedom); it is now proper.
biplots(): the contours were of the singular vectors with a fixed bandwidth of 0.2, and “stable” was decided by the bounding box of a contour.
Maiz help page: it described another data set; the data have 720 rows (9 genotypes, 20 sites, 4 replicates).
Internal
The sampler is written in C++ (Rcpp, RcppArmadillo) and draws its random numbers from R; a plain R version is kept as a reference, and both give the same draws from the same seed.
The tidyverse, rlang, rstiefel, tmvtnorm, mvtnorm and ks are no longer needed; lme4 is suggested (only ge_model() uses it; ge_ammi() computes the error variance in closed form). The diagnostics use coda.
Tests against Tables 1 and B1 of Crossa et al. (2011), simulated data, the building blocks of the sampler and the results of version 0.3.0 for the tables.