rstatix 1.0.0

rstatix 1.0.0

New features

• anova_test() gains a ci argument to add a confidence interval for partial eta squared. With effect.size = "pes" and e.g. ci = 0.95, two columns conf.low/conf.high are returned, computed in base R from the noncentral F distribution (Steiger, 2004) — no new dependency. The intervals match effectsize::eta_squared(partial = TRUE, ci = , alternative = "two.sided"). Works for one-way, factorial, repeated-measures and mixed designs (using the uncorrected degrees of freedom, consistent with the effect-size point estimate). No interval is provided for generalized eta squared ("ges", the default), which has no standard closed-form interval. The default ci = NULL leaves the output unchanged (#18).
• t_test() and wilcox_test() gain an error.as.na argument: when TRUE, a comparison that cannot be computed (a group with fewer than two observations, or essentially constant data) returns an NA result row with a warning naming the comparison, instead of stopping with an error — so the remaining comparisons (or groups, in a grouped analysis) are still computed. The default error.as.na = FALSE keeps the previous behavior (the comparison errors) (#208, #158).
• t_test() and wilcox_test() gain an id argument for paired tests: it names the subject/sample identifier so the two compared groups are matched by subject (instead of by row order), using only subjects present in both groups (complete pairs). For more than two groups the matching is done independently per pairwise comparison, so comparisons can be based on different numbers of pairs (per-comparison pairwise deletion, like SPSS). This makes paired tests work when observations are missing or the groups have unequal sizes, and avoids silent mis-pairing when rows are not in subject order. The default (id = NULL) is unchanged (groups paired in row order) (#136, #175, #192).
• cor_test() now returns the degrees of freedom (df) for the Pearson method (e.g. to report r(df) = …, p); the column sits next to statistic. Spearman/Kendall (which have no df) and cor_mat()/cor_pmat() are unchanged (#107).
• get_summary_stats() can now report skewness and kurtosis (bias-corrected, type-2 estimator, matching e1071's type = 2). They are opt-in via show, e.g. get_summary_stats(data, x, show = c("mean", "sd", "skewness", "kurtosis")), and are not added to any default type, so existing output is unchanged (#99).
• get_summary_stats() gains a digits argument to control the number of decimal places (default 3); useful when summarizing very small values that would otherwise round to 0 (#145, #186, #218).
• pairwise_mcnemar_test() now returns the McNemar chi-squared statistic and df columns for type = "mcnemar" (the default), as the documentation already described; they were previously dropped. The exact-binomial path (type = "exact"), mcnemar_test(), and the p/p.adj values are unchanged (#122).
• dunn_test() gains a ref.group argument to compare each group only to a reference (control) level. When specified, the multiple-comparison adjustment is computed over only the k - 1 comparisons against the reference (matching e.g. GraphPad Prism), which is not equivalent to filtering the full pairwise result afterwards. The default ref.group = NULL performs all pairwise comparisons exactly as before (#101).
• chisq_test() now accepts a pipe-friendly data-frame interface for the test of independence between two categorical variables: data %>% chisq_test(var1, var2) (positional) or data %>% chisq_test(vars = c("var1", "var2")). The contingency table is built internally. The existing table/matrix/vector interfaces (including passing a data frame that already is a contingency table) are unchanged (#43).
• wilcox_effsize() gains a detailed argument: when detailed = TRUE, the output additionally includes the Z statistic (used to compute r = Z/sqrt(N)), the p-value and the test method, so the effect size and the underlying Z are reported in a single data frame. The default (detailed = FALSE) output is unchanged (#122).
• New function dunnett_test() for Dunnett's many-to-one comparisons: each treatment group is compared against a single control (ref.group), with the family-wise error rate controlled over only the k - 1 comparisons using the exact multivariate-t distribution. Results match DescTools::DunnettTest() / multcomp::glht(). Built on emmeans (Suggests) (#129).
• New function ks_test() — a pipe-friendly wrapper around stats::ks.test() for the two-sample Kolmogorov-Smirnov test, with automatic pairwise comparisons (and ref.group) for grouping variables with more than two levels (#92, #168).
• New function conover_test() for Conover's (Conover-Iman) all-pairs rank comparison test, a post-hoc procedure following a significant Kruskal-Wallis test. It mirrors dunn_test() (including ref.group and grouped-data support) but uses the pooled within-group rank variance and a t-distribution, making it generally more powerful. Implemented in base R (no new dependency); results match PMCMRplus::kwAllPairsConoverTest() (#222, #17).
• New function friedman_conover_test() for Conover's all-pairs comparisons (also known as the Durbin-Conover test), a post-hoc procedure following a significant Friedman rank sum test (the repeated-measures analogue of Conover's Kruskal-Wallis post-hoc). It supports ref.group and grouped data. Implemented in base R (no new dependency); results match PMCMRplus::frdAllPairsConoverTest() (#8).
• New function friedman_nemenyi_test() for the Nemenyi all-pairs post-hoc test following a significant Friedman rank sum test (the rank-based, repeated-measures analogue of Tukey's HSD). The statistic is referred to the studentized range distribution, which already accounts for the multiplicity of the comparisons (so the reported p.adj needs no further adjustment). Implemented in base R (no new dependency); results match PMCMRplus::frdAllPairsNemenyiTest() (#141).
• New function add_cld() adds the compact letter display to an all-pairwise comparison result (e.g. from tukey_hsd(), dunn_test(), games_howell_test(), conover_test() or a pairwise t_test()/wilcox_test()): groups that do not share a letter are significantly different — convenient for annotating plots. The letters are computed in base R via the insert-and-absorb algorithm (Piepho, 2004), matching multcompView::multcompLetters() but without adding a dependency. Works on grouped tests (one display per group) (#110).
• New function fligner_test() — a pipe-friendly wrapper around stats::fligner.test() for the Fligner-Killeen test, a non-parametric (rank-based) test of the homogeneity of group variances and a robust alternative to levene_test(). Supports grouped data and reports the complete-case n (#179).

Main changes

• P-values are no longer rounded to 3 significant figures before being returned. Test functions (t_test(), wilcox_test(), kruskal_test(), cor_test(), tukey_hsd(), games_howell_test(), welch_anova_test(), the pairwise_*/*_test proportion/chi-square/Fisher/binomial/McNemar families, etc.) now return full-precision p and p.adj, consistent with dunn_test() which already did. Adjusted p-values are now computed from the full-precision p-values, so re-running adjust_pvalue() (or p.adjust()) downstream is exact. This changes the stored/printed p/p.adj (more digits), and pairwise p.adj values can shift slightly because adjustment is no longer applied to pre-rounded inputs. To round for display, use p_format() / p_round() or options(pillar.sigfig=). Test statistics, effect sizes, the correlation coefficient, and the omnibus ANOVA tests (anova_test(), welch_anova_test(), whose summary p stays at 3 significant figures) are unaffected (#108, #135, #219).
• The method column of detailed test results (detailed = TRUE) now reports the specific test variant instead of a generic label: t_test() gives "Welch t-test" / "T-test" / "Paired t-test" / "One-sample t-test", and wilcox_test() gives "Wilcoxon rank sum test" / "Wilcoxon signed rank test" (#124). Non-detailed output and the p-values/statistics are unchanged.

Minor changes

• freq_table() now supports grouped data: data %>% group_by(g) %>% freq_table(x) computes the frequency table of x within each group (the grouping columns are kept and the proportions sum to 100% within each group), returning a single tidy data frame. Previously a grouped call errored. Ungrouped behavior is unchanged. Based on the contribution by @jakub-jedrusiak (#191).
• Documented how to use emmeans_test()'s model argument for two common cases: (i) computing estimated marginal means averaged over another factor in a factorial design (fit lm(y ~ a * b) and compare a via formula = y ~ a, model = ), and (ii) repeated-measures / mixed designs (pass a within-subject model such as aov(... + Error()) or nlme::lme()). Both already worked via model = — this clarifies the parameter and adds examples (#139, #131).
• Documented the columns added by add_y_position() / add_x_position() / add_xy_position(), and clarified that the groups column is a reserved internal column (used to position dodged grouped comparisons) that is overwritten — users should not rely on it or use groups as one of their own column names in the test object. A future major version may rename it to a dotted, less collision-prone name (#16).
• Documented that anova_test()'s default sums-of-squares type can differ between its two interfaces on unbalanced designs: the formula interface defaults to type II for between-subjects designs, while the dv=/between= interface uses type III for unbalanced between-subjects designs with more than one factor (both use type III for repeated measures). For balanced designs the types coincide; for unbalanced designs, pass type= explicitly for reproducible, interface-independent results. A Note and example were added to ?anova_test. No change to computed results (#190).
• New documentation topic ?\rstatix-programming`showing how to userstatixprogrammatically (variable names held in strings, wrapper functions). The selection interface (e.g.cor_test(), get_summary_stats()) supports tidy evaluation — !!/!!!, embracing with {{ }}, vars=, and all_of()/any_of()— and the formula tests (e.g.t_test(), anova_test()) take a formula built with reformulate()/as.formula(paste())`. These patterns already work; they are now documented and cross-referenced from the relevant help pages (#142).
• cor_test() now errors with a clear message when a weights argument is supplied, instead of silently returning the unweighted correlation (it wraps stats::cor.test(), which has no weights argument). The message points to base R stats::cov.wt(..., cor = TRUE) for a weighted Pearson correlation. Unweighted calls are unaffected (#47).
• The "reference group not present" error (see below) now carries the S3 condition class rstatix_missing_ref_group, so downstream callers (e.g. ggpubr::geom_pwc(), which skips ref-less grouped subsets) can detect it reliably by class — via rlang::cnd_inherits() to also catch it through the grouped (doo()/purrr) wrapping — instead of matching the translatable message text (#153).
• t_test() / wilcox_test() (and the other tests that accept ref.group) now give a clear, actionable error when the specified ref.group is not present in the data, instead of a cryptic '<ref>' must be an existing level from relevel(). This most often happens with grouped data where some group does not contain the reference level; the message lists the available levels and shows the group_by() %>% filter(any(... == ref.group)) idiom to keep only the groups that have it. Valid ref.group values (including ref.group = "all") are unaffected (#153).
• Documented the sign convention of t_test() / wilcox_test() when a ref.group is used: the reference is taken as group1, so estimate = mean(group1) - mean(group2) = mean(ref.group) - mean(other) (the base R t.test()/wilcox.test() convention) and a positive statistic/estimate means the value is higher in the reference group. The note also shows how to flip the sign if you prefer a positive sign to mean "higher in the non-reference group". Documentation only — computed values are unchanged (#153).
• add_xy_position() / add_x_position() with scales = "free" now compute the bracket x positions (xmin/xmax) per facet, so they align correctly when a faceted plot (facet_*(scales = "free")) shows a different set of x-axis levels in each panel. Previously the global x positions were used, placing brackets off-panel in facets that didn't start from the first level. scales = "fixed" (default) and "free_y" are unchanged (#203).
• add_y_position() / add_xy_position() now handle repeated-measures test results whose stashed formula has the outcome ~ within | subject form (e.g. friedman_conover_test()): the | subject block is dropped when locating the x/grouping variable, so the brackets position correctly. Ordinary outcome ~ group results are unaffected (#8).
• anova_test() now gives clear, actionable error messages for two common repeated-measures data problems instead of cryptic ones: when the data has more than one observation per subject per within-subject cell ("duplicated cells"), and when no subject has a complete set of within-cell observations (previously a 0 (non-NA) cases error). Valid designs are unaffected (#216, #146, #116, #134, #102).
• Updated the CRAN checks badge in the README to the current badges.cranchecks.info endpoint (the old cranchecks.info badge no longer resolves) (#174).
• Clarified the ?anova_test documentation about contrasts: the formula / dv + between/within interface fits the model internally with contr.sum (matching SPSS / type-III), whereas a pre-fitted lm()/aov() keeps its own contrasts (R's default contr.treatment unless set otherwise) (#225).
• Clarified the ?wilcox_effsize documentation of the sample size N used in r = Z/sqrt(N). For the one-sample and paired tests, N is the number of pairs (difference scores), because the paired test reduces to a one-sample signed-rank test on the differences; this matches the default of rcompanion::wilcoxonPairedR(). The alternative convention N = 2 x pairs (Field, 2012; Tomczak & Tomczak, 2014) is now documented, with how to obtain it. Documentation only — computed effect sizes are unchanged (#213).
• Removed all internal tidyselect deprecation warnings (Use of .data in tidyselect expressions..., Using an external vector in selections..., and mutate_at()/pull() external-vector usage) across the package by selecting columns with all_of()/any_of(). This is an internal change only — results are unchanged — but it future-proofs rstatix against upcoming tidyselect versions and cleans up the test output (#202).

Bug fixes

• kruskal_test() (and therefore kruskal_effsize()) now reports the number of observations actually used by the test as n, rather than nrow(data). When the outcome or group contained missing values, kruskal.test() dropped them but the reported n was the inflated raw row count; because kruskal_effsize() (eta²[H]) uses n in its denominator, the effect size could be wrong too. n is now the complete-case count (matching the test's own NA handling); the test statistic and p-value are unchanged, and data without NAs report the same n as before (#224).

• Grouped repeated-measures anova_test() followed by get_anova_table() no longer errors (Can't combine <data.frame> and <list>) when the per-group results have different shapes — for example when one group needs a sphericity correction (so its result carries the Mauchly/sphericity-correction tables) while another does not. The internal doo() helper now decides whether to unnest based on all groups rather than only the first, keeping heterogeneous results in a list-column so get_anova_table() can extract each group. Homogeneous grouped results (the usual case) are unchanged (#83).

• wilcox_test() no longer hides the warning when the underlying wilcox.test() silently lowers the confidence interval's confidence level (e.g. to 60% instead of the requested 95%) because it cannot be achieved with tied or zero data. Previously this was suppressed, so the returned interval looked like a 95% CI and could contradict the p-value (e.g. p > 0.05 while the CI excludes 0). A clear warning is now emitted in that case (only when detailed = TRUE, i.e. when a CI is requested). The returned values are unchanged; t_test() and clean-data calls are unaffected (#127).

• games_howell_test() no longer crashes (`df` must be size 15 or 1, not 12) when groups have zero variance (constant values) or undefined variance (a single observation). The Welch correction is undefined for the affected pairs (degrees of freedom 0/0, or NA variance), so those comparisons are now returned as NA with a warning while all other comparisons are computed as usual. Previously such data either errored or silently returned recycled/incorrect values. Results for data without zero/undefined-variance groups are unchanged (#183).

• add_xy_position()/add_y_position() now keep significance brackets evenly spaced after the test results have been filtered (e.g. to keep only significant comparisons). Previously the y positions were computed for the full comparison set and then joined onto the filtered rows, producing uneven spacing; they are now computed for exactly the comparisons present. Unfiltered results, ref.group = "all", one-sample and grouped tests are unchanged (#197).

• cor_test() (and cor_mat()) no longer emit the tidyselect "Using an external vector in selections was deprecated" warning when vars/vars2 are passed as character vectors (e.g. cor_test(data, vars = my_vars)); the columns are now selected via all_of(). Bare names and tidyselect helpers are unaffected (#202).

• p_format() no longer strips a trailing 0 from the exponent of scientific-notation p-values (e.g. p_format(5.1e-10) returned "5.1e-1"; it now correctly returns "5.1e-10"). Ordinary decimal formatting, including trailing.zero padding, is unchanged (#112).

• p_mark_significant() no longer produces "e-NA" (with a coercion warning) when given a scientific-notation p-value string such as the output of p_format() on 1e-4; it now correctly returns e.g. "1e-04****" (#148).

• get_y_position()/add_y_position() now space significance brackets by exactly step.increase. Previously the gap between consecutive brackets was step.increase * n/(n-1) (wider than requested) for n >= 2 comparisons. This changes the computed y.position values for plots with 3 or more groups (brackets are slightly closer together); single-comparison (two-group) plots and ref.group = "all" are unchanged (#201).

• get_test_label() now includes the sample size n for ANOVA results (e.g. Anova, F(1,58) = 105.06, p = <0.0001, eta2[g] = 0.644, n = 60), consistent with the other tests; it was previously dropped because the slicing step stripped the attributes get_n() needs (#150).

• The comparison tests (t_test(), wilcox_test(), dunn_test(), emmeans_test(), etc.) now drop unused levels of the grouping factor, so a filtered factor that still carries empty levels no longer triggers "not enough observations" errors — matching stats::t.test() (#133).

• wilcox_test() and pairwise_wilcox_test() no longer error on degenerate (all-tied / constant) data. The location confidence interval (which can fail to compute on such data) is now requested only when detailed = TRUE, so the default call returns gracefully; statistic/p are unchanged (#79, #167).

• anova_test() results (grouped and ungrouped) and get_anova_table() output are now compatible with dplyr verbs again (e.g. filter(), mutate(), add_xy_position()): the class order is corrected so rstatix_test precedes data.frame, which recent vctrs/dplyr require (#106, #111).

• cohens_d() now honours the mu argument for two-sample tests (independent and paired): mu is the hypothesized mean difference and is subtracted before standardizing, consistent with the one-sample case. Previously mu was silently ignored for two-sample tests. Calls using the default mu = 0 are unchanged (#200).

• emmeans_test() no longer fails with "Nonconforming number of contrast coefficients" when the covariate is a numeric variable with only two distinct values (e.g. a 0/1 indicator). The covariate is now correctly averaged over (ANCOVA) instead of being kept as a grid factor (#206, #86).

• kruskal_effsize() now clamps the eta-squared effect size to its valid [0, 1] range, so a near-null effect no longer returns a negative value (the formula could yield a small negative for a tiny H); consequently it is correctly labelled "small" rather than "large" (the magnitude had used the absolute value) (#217).

• anova_test() / anova_summary() now return both effect sizes when effect.size = c("pes", "ges") is requested; previously only partial eta squared was kept (#180).

• tukey_hsd() now respects conf.level (and other TukeyHSD() arguments) for the ungrouped data-frame interface; previously ... was dropped so the confidence interval was always 95% (#188).

• pairwise_binom_test_against_p() no longer errors on an unnamed x; groups are auto-labelled grp1, grp2, … (named/table input is unchanged) (#44).

• Fixed a missing space in the anova_test() error message for single-level factors (now reads "Variable x has only one level") (#137).

• anova_test() results now carry rstatix_test as the first class, with the specific test class (anova_test / grouped_anova_test) in second position — matching every other rstatix test. The #106 dplyr/vctrs invariant (rstatix_test before data.frame) is preserved, but the earlier c("anova_test", "rstatix_test", "data.frame") order had shifted the specific class out of position 2, breaking reverse dependencies that dispatch on class(x)[2] (e.g. GimmeMyStats, GimmeMyPlot). inherits(), print()/plot() dispatch, dplyr verbs and ggpubr are unaffected.