vecmatch

• match_gps() now works with any reference when method = "nnm". The columns kept for matching were selected with a logical mask built on the full csmatrix, but the mask was applied after the treatment column had already been dropped, so every position shifted by one. The gps column of the reference was therefore only retained when the reference happened to be the first gps column of the csmatrix; any other reference failed with undefined columns selected. When it did not fail, the shifted mask kept the neighbouring gps column as well, so [Matching::Matchby()] silently matched on two covariates instead of one and returned a different set of matches. Both branches now index the gps column of the first group of the current combination by name, which also fixes the same error for a user-supplied combos whose group1 differs from reference.
• match_gps() now actually honours the max_controls argument for method = "fullopt". It was passed to [optmatch::fullmatch()] under a name that function does not accept and carried the value of caliper, so it was silently discarded and had no effect on the matching. Calls that set max_controls to a non-default value will now produce different (correct) results.
• csregion() now stops with an informative message when the common support region turns out to be empty, or when a treatment group loses all of its observations. Previously such a call failed with non-numeric argument to binary operator, which came from the fallback warning of the internal refitting step rather than from the empty region itself, and gave no hint about the actual cause. With refit = FALSE the same situation returned a zero-row csr object silently, and the problem only surfaced later in match_gps(). Emptying a group is now reported as well, since dropping it would silently change the estimand. The new messages name the groups involved and suggest a larger sample, fewer treatment groups, a different borders value or a different model specification in estimate_gps().

Minor changes and bug fixes

• match_gps() no longer warns when a method-specific tuning parameter is left at its default. Previously, a call such as match_gps(csr, method = "fullopt") emitted a warning for every unset parameter (order, min_controls, max_controls), which made an ordinary default call look like it had gone wrong. Defaults are now filled in silently. Arguments that do not apply to the selected method are still reported, but only when they were explicitly supplied by the user.
• The descriptive metrics of balqual() describe the covariates as they are given in the formula, while the balance metrics are computed on the model matrix. A factor is therefore described once by its levels rather than once per dummy-coded column, and interaction terms are covered by the balance metrics only.
• The statistic argument of balqual() is not used by the descriptive metrics, and a warning is now issued when it is supplied together with type = "desc_full" or type = "desc_reduced" alone.
• Fixed a bug in balqual(), where the values of the cutoffs argument were matched to the metrics by position across all three balance metrics rather than by name. Passing a subset of the metrics, e.g. type = c("smd", "var_ratio"), raised a recycling warning and could evaluate a metric against the cutoff of another one.
• print.quality() now returns its argument invisibly, as documented for print() methods, instead of returning it visibly. The visible return value made the quality tables be printed twice whenever the result of print() was itself auto-printed, e.g. inside capture.output() or a knitted document.
• Fixed the fallback warning of csregion(), which was emitted when refitting the GPS model on the CSR-restricted data fails. Its message was split into three fragments that strwrap() received as the width and indent arguments instead of as text, so the warning aborted instead of being shown.
• optimize_gps() no longer aborts with $ operator is invalid for atomic vectors when the GPS estimation fails for the first combination of the search space. Failed combinations are passed through as the message of the underlying error, but the treatment levels were read from the first result unconditionally, so a single failure at the head of the estimation space brought down the whole sweep. The first result that actually holds a GPS matrix is now used instead. If every combination fails, the function stops with a message that reports how many were tried and the reason given by the first failure.

Documentation

• The package documentation is now also available as a pkgdown website at <https://polymerase3.github.io/vecmatch/>. The site collects the reference pages of all exported functions, grouped by the five steps of the vecmatch workflow, together with the vignettes and this changelog, and is rebuilt automatically by a GitHub Actions workflow on every push to main.