Where the methods in rstatix come from, and which other packages
implement them.
Most of the statistical methods in rstatix are implemented in base R
from the published formulas. A few call another package to compute the
result, and one is adapted from another package's code; Adapted code
below says which. Where a function records the source of its method, it does
so in its own References section; cite those authors, not this
package, when you report a result.
Cramer's V — Cramer, H. (1946) Mathematical Methods of
Statistics. See cramer_v().
Effect-size confidence intervals (partial eta squared,
Cramer's V) — obtained by inverting a noncentral distribution: Smithson, M.
(2003) Confidence Intervals; Steiger, J. H. (2004) Beyond the F test.
Psychological Methods, 9, 164-182. See anova_test() and
cramer_v().
Conover's all-pairs test — Conover, W. J. (1999)
Practical Nonparametric Statistics, 3rd edition. See
conover_test() and friedman_conover_test().
Nemenyi's all-pairs test — Nemenyi, P. (1963)
Distribution-free Multiple Comparisons. See
friedman_nemenyi_test().
Compact letter display — Piepho, H.-P. (2004) An algorithm
for a letter-based representation of all-pairwise comparisons. Journal
of Computational and Graphical Statistics, 13, 456-466. See
add_cld().
Dunnett's many-to-one comparisons — Dunnett, C. W. (1955) A
multiple comparison procedure for comparing several treatments with a
control. Journal of the American Statistical Association, 50,
1096-1121. See dunnett_test().
The following packages implement some of the same methods. rstatix
compares its results against them while developing, and they offer
functionality that rstatix does not:
effectsize — a broad effect-size toolkit;
effectsize::cramers_v() and effectsize::eta_squared() produce
the same intervals as cramer_v(ci = TRUE),
anova_test(ci = ), and eta_squared(ci = ) /
partial_eta_squared(ci = ) when called with
alternative = "two.sided", which is not their default.
DescTools — DescTools::CramerV() and
DescTools::DunnettTest().
multcomp — multcomp::glht() for general linear
hypotheses, including Dunnett contrasts.
multcompView — multcompView::multcompLetters() for
compact letter displays.
PMCMRplus — a large collection of all-pairs and many-to-one
nonparametric procedures, including the Conover and Nemenyi tests.
These packages are not dependencies of rstatix. The values they return
are recorded in the rstatix test suite as fixed numbers, checked
against the package they came from at the time they were written; a recorded
number cannot detect a later change in the package that produced it.
Of the functions documented in this package, sign_test() is the
exception to the description above. Its one- and two-sample test code, and
the confidence interval it reports for the median, are adapted with
modifications from DescTools::SignTest() and
DescTools::MedianCI(), written by Andri Signorell. DescTools is
distributed under GPL (>= 2); rstatix is distributed under GPL-2.
No other source file in rstatix carries a statistical method adapted
from a contributed package. Argument-validation idioms are shared with base R
— check_two_samples_test_args() follows the preamble of
stats::wilcox.test(), as DescTools and PMCMRplus do — and
those are not methods anyone cites. The test suite scans the sources for the
phrases that declare an adaptation and fails when one appears in a file this
section does not name; code copied without a word about it would not be caught
that way.
Copying code is not the same as calling it. Several functions compute their
result by calling another package, as each of them states in its own
documentation: anova_test() uses
car::Anova() for type II and type III sums of squares, and
stats::aov() for type I; dunnett_test() and
emmeans_test() use emmeans::emmeans(); and
wilcox_effsize() takes its default statistic (r = Z /
sqrt(N)) from coin::wilcoxsign_test() or coin::wilcox_test(),
while its method = "rank_biserial" alternative is computed in base R.
cramer_v(), anova_test(),
conover_test(), add_cld(),
dunnett_test().