anovakit

anovakit has eight analysis-of-variance functions: Welch's ANOVA, Kruskal-Wallis, ANCOVA, repeated measures, MANOVA, and analyses of deviance for binary, count and other GLM outcomes. They take the same arguments and return the same kind of object. Each one checks the assumptions its method needs, reports effect sizes, marginal means and pairwise comparisons, builds ggplot2 figures without drawing them, and records in $notes every decision it made for you.

Installation

# install.packages("remotes")
remotes::install_github("elkronos/anovakit", build_vignettes = TRUE)

install_github() skips vignettes unless you pass build_vignettes = TRUE, so without it vignette("anovakit") is not available.

Which function

Function	Response	Use it when
anova_welch()	one numeric	Comparing means. The default choice: it does not assume equal variances.
anova_kw()	one numeric	No distributional assumption. Kruskal-Wallis with Dunn post-hoc. Convert an ordered factor with as.integer() first.
anova_ancova()	one numeric	Adjusting for one or more numeric covariates.
anova_rm()	one numeric	The same subjects measured repeatedly.
anova_manova()	two or more numeric	Several outcomes analysed jointly, with or without covariates.
anova_bin()	binary	Logistic regression, odds ratios, marginal probabilities.
anova_count()	counts	Poisson, quasi-Poisson or negative binomial, with exposure offsets.
anova_glm()	anything	A classical factorial ANOVA with main effects and interactions, Type II or III (the default Gaussian family), or any other GLM family.

Six take data, response, groups, with columns named as character strings. anova_manova() takes responses (plural), and anova_rm() takes response with subject, within and between. Choose an analysis walks through the choice.

A first example

Fuel economy of 32 cars, by number of cylinders:

library(anovakit)

fit <- anova_welch(mtcars, "mpg", "cyl")
fit
#> Welch's analysis of variance 
#> ----------------------------
#> Call: anova_welch(data = mtcars, response = "mpg", groups = "cyl")
#> Observations used: 32
#> 
#> Omnibus test (Welch F)
#>   term statistic num_df den_df   p_value
#> 1  cyl     31.62      2  18.03 1.271e-06
#> 
#> Notes
#>   - Largest group variance is 9.63 times the smallest. Welch's test handles this; the classical F test would not.
#> 
#> Plots available: means, box, qq
#>   (use plot(x, which = "means"))

fit$posthoc[, c("group1", "group2", "difference", "conf_low", "conf_high", "p_adjusted")]
#>   group1 group2 difference conf_low conf_high   p_adjusted
#> 1      4      6   6.920779 3.751376 10.090182 4.048495e-04
#> 2      4      8  11.563636 8.318518 14.808755 4.924043e-06
#> 3      6      8   4.642857 2.802925  6.482789 9.080709e-05

summary(fit)       # adds the assumption checks, effect sizes and group summaries
plot(fit, "means") # a ggplot2 object

The comparisons are Welch t tests with Holm-adjusted p-values. cyl is numeric in mtcars; grouping columns are converted to factors.

What you get

• The omnibus test in $anova, with the sums of squares, statistic and any correction named in the printed heading.
• Effect sizes in $effect_sizes, with intervals for standardised mean differences, odds ratios and rate ratios.
• Marginal means in $emmeans and pairwise comparisons in $posthoc, with both the unadjusted and the adjusted p-value.
• The assumption checks that fit the method in $assumptions.
• ggplot2 objects in $plots, returned and never drawn.
• One conf_level for every interval, and no console output from any analysis function.
• $notes: dropped rows, centred covariates, a switch from Poisson to negative binomial, a perfectly separated cell.

Documentation

The website, https://elkronos.github.io/anovakit/, has:

• Get started: one analysis from start to finish, and the arguments every function shares.
• Choose an analysis: four questions that lead to one of the eight functions.
• Walkthroughs and guides: one worked analysis per function, and guides to results, diagnostics, plots, weights and the design choices behind the defaults.
• The function reference.

From R:

vignette("anovakit")      # Get started
?anova_welch              # any function
?anovakit_fit             # what the returned object holds
citation("anovakit")

Report bugs and request features at https://github.com/elkronos/anovakit/issues.

Dependencies

Imports car, emmeans, ggplot2 (3.4.0 or later) and base R.

afex (repeated measures), MASS (negative binomial) and sandwich (robust standard errors) are Suggests. Without afex, anova_rm() stops with a message, as does anova_count(model = "negbin") without MASS. Where a fallback exists, the function uses it and says so in $notes: model = "auto" without MASS fits quasi-Poisson where it would have chosen the negative binomial, and vcov_type without sandwich uses model-based standard errors. bit64 is needed only to read integer64 columns.

Dunn's test, Mardia's tests, Box's M and the canonical discriminant analysis are implemented in the package rather than imported.

Licence

GPL-3.