LoginSign Up
← Back to all sparks
R

rstatix

ANALYTICS
Velocity2.5

Pipe-friendly Framework for Basic Statistical Tests in R

github.com/kassambara/rstatix ↗

rstatix hit 1.0 by unrounding every p-value it has ever returned

statisticsreffect sizesconfidence intervalsreproducibilitytidyverse
◆Current state
rstatix is the pipe-friendly test wrapper behind most ggpubr annotation workflows — t-tests, Wilcoxon, ANOVA, post-hoc comparisons, effect sizes, all returning tidy data frames. After three quiet years of CRAN-compat patching, it shipped 1.0.0 and 1.1.0 three weeks apart in mid-2026. Both releases push in the same direction: interval estimates and full-precision output for numbers the package previously rounded or omitted.
◆Where it's heading
The work is about matching what dedicated effect-size packages give you without taking on their dependencies. Confidence intervals for partial eta squared and for Cohen's d are both computed in base R from noncentral distributions and both check against effectsize; compact letter displays are computed in base R against multcompView. The pattern is deliberate — reproduce the reference implementation, add no imports. Alongside that, the package has started correcting statistical hygiene it got wrong for years, most visibly by no longer rounding p-values before adjusting them.
◆Prediction
The analytic-interval machinery now exists for eta squared and Cohen's d; the untouched effect sizes in the package — eta squared for nonparametric tests, Cramer's V, rank-biserial correlation — are the obvious next targets for the same base-R noncentral treatment.

◆Recent moves

  1. 22d ago

    rstatix 1.1.0 gives Cohen's d a deterministic confidence interval

    The interval work from 1.0.0 extended to Cohen's d: ci.method = "analytic" inverts the noncentral t distribution instead of bootstrapping, so the interval needs no seed and reproduces exactly across runs. It scales by the Hedges factor when requested and matches effectsize::cohens_d(). Incremental against the arc, but it removes a genuine reproducibility trap.

    View source ↗
  2. 1mo ago

    rstatix 1.0.0 stops rounding p-values before adjusting them

    ⚡ SPARK

    The 1.0.0 milestone changes numbers users have already published. Full-precision p and p.adj replace the old 3-significant-figure rounding across the whole test family, and adjustment now runs on unrounded inputs — so pairwise p.adj values can shift. Confidence intervals for partial eta squared, compact letter displays and fligner_test() land alongside.

    View source ↗
  3. 9mo ago

    rstatix 0.7.3

    CRAN survival work during the quiet years — tests loosened to accept both legacy and R-devel Wilcoxon p-values under ties, plus roxygen formatting fixes. No user-facing change.

    View source ↗
  4. 3y ago

    rstatix 0.7.2

    A tidyselect 1.2.0 floor and two emmeans_test() fixes, one restoring the factor class of the grouping variable. Dependency upkeep rather than product work.

    View source ↗
  5. 3y ago

    rstatix 0.7.1

    A cluster of small correctness fixes across the annotation path: shapiro_test() name collisions, get_comparisons() dropping unused levels, get_summary_stats() preserving user column order, and group sizes counted from non-NA values. The same category of quiet inaccuracy 1.0.0 later addressed at scale.

    View source ↗
  6. 5y ago

    rstatix 0.7.0

    Accessors for test results — get_n(), get_description(), remove_ns() — and a rewritten add_x_position(). This is rstatix building the plumbing that lets ggpubr annotate plots from test objects, the role that makes its later p-value change consequential.

    View source ↗