ggpubr

• geom_bracket() gains an orientation argument. orientation = "vertical" draws a native vertical bracket (the bar spans the y axis, the tips point left and the label is rotated), specified with ymin/ymax/x.position — useful to annotate plots where the comparison runs along the y axis, e.g. Kaplan-Meier curves (it is designed for a continuous x axis). The default orientation = "horizontal" is unchanged, and it cannot be combined with coord.flip = TRUE (#456).
• ggpaired() gains an opt-in jitter argument that spreads the paired points horizontally to reduce overlap. Each subject (id) gets a single offset, so its two points move together and the connecting line stays intact, and only the horizontal positions change (the values are never moved). Default jitter = 0 leaves the plot unchanged (#407).
• ggboxplot(), ggviolin() and ggstripchart() gain an opt-in show.n argument. When show.n = TRUE, the number of observations ("n = <count>") is displayed at the top of each group. When the groups are dodged (a color/fill grouping with a dodging position), one count is shown per group; otherwise a single count is shown per x-axis tick. Counts respect select/remove and are computed per facet. Default is show.n = FALSE, so existing plots are unchanged (#598, #627).
• ggarrange() gains a spacing argument to increase the gap between the arranged plots (a long-requested option). It adds a uniform margin, in text-line units, around each plot. Default is 0 (no extra space; each plot keeps its own margins, so existing arrangements are unchanged) (#151).
• compare_means() gains an id argument for paired comparisons. Without it, a paired test (paired = TRUE) pairs observations by row order, so the p-value is wrong when the data are not sorted so that the compared groups align by subject. Passing id (the name of a subject-identifier column) pairs the observations by id instead — row-order independent, using only the complete pairs (via rstatix, so the result matches a direct rstatix::t_test()/wilcox_test() call). It works for a two-group, a pairwise (more than two groups) and a ref.group comparison with t.test/wilcox.test. Default (id = NULL) is unchanged (#560).
• stat_compare_means() also gains the id argument, so a paired test can be aligned by subject id directly on a plot (e.g. ggpaired(...) + stat_compare_means(paired = TRUE, id = "subject")), displaying the correct p-value even when the data are not sorted by subject. Plots without id are unchanged (#560).
• geom_pwc() and stat_pwc() gain a p.adjust.n argument giving the number of comparisons to use for the p-value adjustment (passed as n to stats::p.adjust()). Default is NULL (adjust by the number of computed p-values, unchanged behavior); set it when the displayed comparisons are a subset of a larger family so the adjustment reflects that larger size. It must be at least the number of p-values being adjusted, and applies to the panel-level adjustment (p.adjust.by = "panel") and the single-comparison case (#612).
• annotate_figure() gains column.titles and row.titles arguments to add a title above each column and to the left of each row of an arranged grid (useful for publication panels). Each accepts a character vector (one title per column / per row, rendered in bold) or a list of grobs (e.g. built with text_grob()) for full styling control; row.titles are rotated 90 degrees. The titles are placed in reserved space (they do not overlap the plots) and are evenly spaced, so they assume equal-sized columns / rows (the ggarrange() default). Default is NULL (no titles; existing figures are unchanged) (#573).
• ggbarplot() gains a numeric.x.axis argument (already available in ggline() and ggerrorplot()). With numeric.x.axis = TRUE the x variable is kept numeric instead of being coerced to a discrete factor, so bars are drawn at their numeric x positions (e.g. a time axis). Default is FALSE (unchanged behavior) (#463).
• geom_bracket() and stat_pvalue_manual() gain a tip.length.ref argument controlling what tip.length is a fraction of. With the default tip.length.ref = "data" the tips are a fraction of the data range (unchanged behavior). With tip.length.ref = "axis" the tips are a fraction of the y-axis range (ylim/scale_y_*), which renders at the same physical fraction across plots and therefore gives visually constant tip lengths regardless of the data range - useful to keep tips consistent across facets or across separate plots with different scales (#362). stat_compare_means(comparisons =) draws its brackets via ggsignif::geom_signif() and does not support tip.length.ref; passing it there now emits an informative message pointing to stat_pvalue_manual(..., tip.length.ref = "axis") instead of being silently ignored (#362).
• stat_cor() and stat_regline_equation() gain a label.y.step argument giving the vertical spacing (in text-line units) between the labels of successive groups. Default is 1.4 (unchanged behavior). Setting label.y.step = 0 stops the per-group vertical shift so labels align across facet panels when a factor is mapped to an aesthetic that also defines the facets (the long-standing "labels climbing stairs across facets" behavior; #248). This control was inspired by the vstep argument in the ggpmisc package (by Pedro J. Aphalo), from which ggpubr's correlation/regression label-positioning logic was originally adapted; this is acknowledged via @seealso in ?stat_cor.
• ggviolin() gains a drop argument (passed to ggplot2::geom_violin()), default TRUE (unchanged behavior). Previously the argument was silently dropped, so grouped violins where a sub-sample has too few points to draw a density (which geom_violin() removes, including from the dodge position) could not be kept aligned with added boxplots / dot plots. Setting drop = FALSE together with position = position_dodge(0.8, preserve = "single") now reserves the empty dodge lane so all geoms stay aligned (#381).
• ggarrange() gained the ability to choose which plot(s) supply the shared legend: common.legend now also accepts one or several plot indices. common.legend = 2 uses the second plot's legend (handy when the first plot's legend is not representative, e.g. a group is missing in the first plot), and common.legend = c(1, 2) keeps and combines the legends of the listed plots into a single shared block (side by side for legend = "top"/"bottom", stacked for "left"/"right") - useful when the arranged plots genuinely need different legends. The documentation of common.legend was also clarified: common.legend = TRUE keeps the first plot's legend (it does not merge or validate legends), so plots should share a consistent scale for a single shared legend to be correct. Logical TRUE/FALSE behave exactly as before (#347).
• ggdonutchart() gained a label.repel argument (default FALSE). When TRUE, slice labels are placed with ggrepel::geom_text_repel() and connected to their slice with leader lines, so the labels of many small slices no longer overlap (or get dropped). The default output is unchanged (#655).
• ggboxplot() now accepts a position argument (e.g. position = position_dodge(0.9)), like ggviolin()/ggdotplot() already do. Previously passing position errored (formal argument "position" matched by multiple actual arguments) because the dodge was hardcoded; the default is unchanged (position_dodge(0.8)), so grouped box plots look the same unless you set it (#615).
• ggsummarytable() gains an angle argument to rotate the summary-table text; default (angle = 0) is unchanged (#595).
• ggmaplot() gains a line.color argument to set the threshold line color; default ("black") is unchanged (#322).
• ggarrange() gains a byrow argument (default TRUE) to fill the plot grid by column (byrow = FALSE) instead of by row; forwarded to cowplot::plot_grid() (#225).
• stat_regline_equation() gains coef.digits and rr.digits arguments to control the number of significant digits shown for the regression-equation coefficients and R2; defaults (2) reproduce the previous output (#312).
• Added formatting parameters to statistical helpers:
• p.format.style
• p.digits
• p.leading.zero
• p.min.threshold
• p.decimal.mark
• Added p.format.signif support and related label handling paths.
• stat_pvalue_manual() gains a p.digits argument (default 3) that formats **numeric** p-value label columns for display (e.g. label = "p" or label = "p.adj"), using the same format_p_value() engine as stat_anova_test(). This restores clean labels (e.g. 0.0156 instead of 0.015625) with rstatix >= 1.0.0, which now returns full-precision pairwise p-values. Character labels (significance symbols, pre-formatted strings, glue expressions) are unaffected. Set p.digits = NULL to print the raw value.
• stat_pvalue_manual() additionally gains p.format.style, p.leading.zero, p.min.threshold and p.decimal.mark, matching the p-value formatting arguments already available in stat_compare_means(), geom_pwc() and stat_anova_test(). These are opt-in and default so that the rendered labels are unchanged; for example p.min.threshold = 0.001 displays very small p-values as < 0.001, and p.format.style = "nejm" applies a journal style.
• stat_cor() gains two label-formatting arguments:
• r.leading.zero — set to FALSE to drop the leading zero of the correlation coefficient (e.g. .73 instead of 0.73), completing APA-style reporting together with p.leading.zero (#540). The dropped leading zero is preserved through plotmath rendering (the value is quoted so it is not silently re-normalized back to 0.73 in the default expression output).
• p.coef.name — symbol for the p-value label; use "P" for an uppercase p-value (#541).
• stat_cor() exposes two new computed variables, rmse and rmse.label, for the root mean square deviation (RMSE/RMSD) between x and y — useful for reporting agreement between paired measurements on the same scale (e.g. predicted vs. reference values). Display it with aes(label = after_stat(rmse.label)), or combine it with the correlation coefficient using paste(). The default label is unchanged (#458).
• stat_cor() can now display the confidence interval of the correlation coefficient, via the new computed variables conf.int.low, conf.int.high and conf.int.label (e.g. "95% CI [0.21, 0.75]") and a conf.level argument (default 0.95). Show it with aes(label = after_stat(conf.int.label)), or combine it with the coefficient using paste(). The confidence interval is available for method = "pearson" only; it is NA for Spearman/Kendall. The default label is unchanged (#418).

Main changes

Compatibility

• Added compatibility updates for modern ggplot2, dplyr, and tidyr.
• Updated legacy size usage to linewidth where required by recent ggplot2.
• Completed the size -> linewidth migration for the remaining line layers that still passed the deprecated size argument: the mean/median reference line added by gghistogram(add = ...) and ggdensity(add = ...), and the connector segments of ggdotchart(add = "segments"). These no longer emit ggplot2's "size aesthetic for lines was deprecated" or "Ignoring empty aesthetic: size" warnings, and the requested line width is now applied via linewidth.
• ggmaplot() no longer emits an "Ignoring empty aesthetic: size" warning on its default call; the point layer sets size only when the user supplies a value.
• Replaced deprecated tidyverse APIs in affected helper functions.
• Added package startup lock-file checks and clean_lock_files() helper.
• Relaxed minimum ggrepel dependency to >= 0.9.2 to keep Ubuntu oldrel (R 4.4.x) CI dependency resolution working.
• geom_pwc() now detects an absent ref.group in a grouped subset via the rstatix_missing_ref_group condition class raised by recent rstatix (walking the parent chain with rlang::cnd_inherits()), in addition to matching the error message. This keeps the "skip ref-less subsets" behaviour working after rstatix made that error message clearer (rstatix #153); older rstatix versions are still handled via the message fallback.
• The minimum required rstatix version is now >= 1.0.0 (the version that introduced the id argument used by compare_means(id = ), and the full-precision pairwise p-values relied on by the p-value formatting).

Validation and messaging

• Hardened sparse-group handling in stat_compare_means() and geom_pwc().
• Non-comparable grouped subsets (insufficient levels/observations) are now skipped while preserving valid inferential comparisons in the same layer.
• Added per-group skip diagnostics in geom_pwc() so users can see which grouped subsets were skipped and why (e.g., missing ref.group or insufficient levels).
• Added regression tests for mixed comparable/non-comparable subsets and fully sparse subsets.
• Passing ggtheme = NULL to the plotting functions (e.g. ggscatter(), ggboxplot(), ggline(), gghistogram(), ...) now skips applying a ggpubr theme, so the plot keeps ggplot2's default theme or the theme set globally with theme_set(). Previously an explicit ggtheme = NULL was treated like an unset argument and theme_pubr() was still applied. Calls that omit ggtheme or pass a specific theme are unchanged (#561).
• compare_means() now adjusts p-values **within** each group.by level (and response) rather than pooling all groups together, so a grouped adjustment matches filtering to one group and adjusting there. This changes p.adj values for calls that use group.by; ungrouped calls are unchanged (#200).
• stat_compare_means(comparisons = ) now emits a one-time message (and the docs note) clarifying that the displayed pairwise p-values are **not** adjusted for multiple comparisons, pointing to geom_pwc() / stat_pvalue_manual() + compare_means(p.adjust.method = ) for corrected p-values (#293).

Documentation

• Documented a recipe for labelling groups with significance letters (compact letter display): compute the pairwise comparisons and derive the letters with rstatix::add_cld(), then place them with geom_text() (see the "Significance letters" section in ?compare_means). Also covers the per-control letters case (a = differs from control 1, b = differs from control 2) (#464, #434).
• Documented that a summarized ggbarplot() (e.g. add = "mean_se") must be faceted with the facet.by= argument, not by appending + facet_wrap()/+ facet_grid(): the summaries are pre-computed over facet.by, so a manually added facet pools the bars (and, for stacked bars, the error bars) across all panels (#739).
• Clarified document