diagnostics() is replaced by validity(), which also reports chain convergence and can compute the Gelman-Rubin statistic on request.
Changed results
The 0.1.0 sampler had been ported from a superseded revision of the reference code, and it differed from the current reference implementation in all four sampler blocks. The coefficient step used a plain symmetric random walk instead of the Hastings corrected IWLS proposal. The variance step used an inverse gamma independence proposal instead of the Laplace proposal. Exogeneity was imposed by zeroing entries of an unrestricted inverse Wishart draw and rejecting, instead of the block recursive decomposition that now holds the zeros exactly. The marginal transform used the cumulative right endpoint instead of the cell midpoint. On the same data and seed the two samplers produce different posteriors, and the 0.2.0 sampler is the one that reproduces the published reference to 1e-12.
Bug fixes
A regressor with tied values used to raise a hard error, because the 0.1.0 code added a vector of length N to a table of shorter length. The marginal probability masses now sit on the unique observed values, so ties are the normal case. A count, a dummy, a Likert item or a rounded price all fit without error.
Differences from the reference implementation
The package deliberately differs from the reference implementation in three places. All three sit inside validity(). None of them changes any estimate: the sampler itself, and so every coefficient, draw and posterior summary, is unchanged.
The effective sample size function .bayes_ess() truncates its autocorrelation sum at the first negative lag unconditionally. The reference implementation only truncates when that lag is not the first one, which let it report an effective sample size larger than the number of retained draws, up to 390 times the number of draws, on 45% of chains in a simulation check. The package instead always truncates at the first negative lag, so the reported value never exceeds the number of retained draws.
validity.copregbayes() captures the calling frame with where <- parent.frame() in its own body, before the lapply() that runs the extra chains for the Gelman-Rubin statistic, and evaluates the original call there. The reference implementation calls parent.frame() inside the lapply() callback instead, which resolves to the lapply() call’s own frame rather than the frame validity() was called from. That made validity(fit, chains = 2) fail with “object ‘d’ not found” whenever it was called from inside a function whose data was not in the global environment. The package instead finds the data reliably regardless of where validity() is called from.
.bayes_ess() returns NA_real_ for a chain that never moves, since stats::acf() of a constant series is NaN and the truncation above never fires for it. The reference implementation returns NaN in that case. The package instead reports NA, matching what .bayes_geweke() already returns for the same input, so the convergence table reads as “not available” rather than showing NaN.
Dependencies
The package now imports endogCopula for the two shared core helpers. The sampler itself is base R only, so copula, LaplacesDemon, invgamma and mvtnorm are dropped.
The test suite is replaced, since the old tests targeted the removed API.