Prepared the first CRAN release. The detailed guide is now a PDF vignette
with an executable introductory example. Revised documentation wording and
added citation("GIR") entries for the software and its underlying working
paper. The paper has not yet been published on arXiv.
Added full release checks that build the vignette and PDF function reference, plus Linux checks on GitHub Actions. These changes do not alter estimation or inference calculations.
Made the polynomial design-grouping test portable across numerical libraries. It now checks exact reconstruction of the design and its actual distinct rows, in addition to agreement with the independent coefficient calculation.
Empty estimation samples after na.omit, na.exclude or a custom action
now give an informative error before matrix construction or the routine
removal warning. Positive weights are checked in the retained sample,
and entirely missing weights follow the requested missing-value policy.
First-stage support checks now use retained observations. If omission leaves
an undeclared numeric covariate with at most two distinct values, an explicit
discrete or continuous declaration is required. Custom actions report
missing versus complete removals accurately, preserve their own error
context and reject invalid row identifiers. Estimator formulas are unchanged.
Added regression tests for sample validation, row alignment and empty prediction/interval shapes. Replaced all broad warning suppression in the point-estimator and bootstrap/prediction tests with checks of the complete expected warning set or an explicit absence of warnings. User-facing numerical and inference warnings remain enabled.
Added deterministic mathematical validation of both estimation stages and
bootstrap refits against independent QR-based threshold regressions for
full-rank Legendre and spline tensor bases. Added pairwise group-Gini,
positive outcome-scale and row-order invariance, and GIR-R-squared identity
tests. Existing expected bootstrap warnings are now asserted explicitly.
The new GIR-validation help topic explains the tests and their limits;
estimator calculations, defaults and returned results are unchanged.
First-stage ridge penalties now scale with the average diagonal of the
weighted Gram matrix, without the former absolute floor of 1. The initial
and retry penalties both respect common weight rescaling. Unusable scales,
penalties and adjusted matrices give informative errors, and successful
ridge retries are identified in the diagnostics. The ridge = 1e-10
default is unchanged and documented as a numerical heuristic. Ridge-based
point estimates and inference can change where the old floor was active;
affected analyses should refit and rerun their bootstrap.
Added cluster_order = "auto", "original", or "estimation" to
bboot.gir(), also available through automatic bootstraps in predict().
Automatic ordering now rejects cluster vectors when the original data and
estimation sample have equal lengths but different row mappings. Explicit
ordering prevents silent cluster misassignment after rows are reordered or
repeated. Retained-column formulas are already aligned. Affected analyses
should rerun their bootstrap; original point estimates are unchanged.
Fixed multi-column cluster formulas accidentally merging distinct groups
when their combined text labels collided (for example, ("a.b", "c") and
("a", "b.c")). Grouping now uses exact value combinations, including
numeric identifiers that print alike, and retains first-occurrence order.
Affected analyses should rerun their bootstrap: draws and inference can
change; original point estimates are unaffected. Added regression tests
and clarified that multi-column formulas specify intersection clusters.
Added bboot.gir(): common normalized exponential multipliers in both stages, basic/type-7 intervals, bootstrap covariance and SEs, centered marginal tests, optional joint bootstrap Wald tests, smooth coefficient transformations, supplied multipliers and a one-draw mode for Monte Carlo studies.
Added predict.gir(): training/newdata linear Gini indices, pointwise bootstrap confidence intervals and standard errors, reusable coefficient draws, missing-data handling and estimability checks for aliased point predictions.
Fits retain a resolved estimation snapshot; bootstrap bases and ordering are cached without changing the sums used in estimation.
Added help files and edge-case tests.
bboot.gir() gains a cluster argument: one Exp(1) multiplier per cluster,
shared by all its observations in both stages (a Bayesian cluster
bootstrap), given as a one-sided formula over retained columns or as a
vector aligned to the estimation rows or to the original data. The result
reports the number of clusters and whether the first-stage cells are nested
in the clusters, in which case only the second stage varies across draws.
The second stage accumulates its sums over the distinct regressor rows, in
gir() and in every bootstrap draw. Estimates are unchanged. Fits and
bootstrap draws with cell-level regressors (industry and year effects in
firm data, for example) are stored and refitted at the size of the distinct
rows instead of the sample; x = TRUE still returns the full design.
The automatic sieve rule now targets the total tensor-product dimension (the paper's K) for splines as well as Legendre polynomials; a knots-per-margin alternative remains available as sieve_size_type = "knots". Every Legendre margin has at least two columns, and the diagnostics report when that floor, or the spline minimum, exceeded the requested size.