Visualizations with statistical details: The 'ggstatsplot' approach

Indrajeet Patil

doi:10.21105/joss.03167

A combination of box and violin plots along with jittered data points for between-subjects designs with statistical details included in the plot as a subtitle.

Usage

ggbetweenstats(
  data,
  x,
  y,
  type = "parametric",
  pairwise.display = "significant",
  pairwise.alpha = 0.05,
  p.adjust.method = "holm",
  bf.prior = 0.707,
  bf.message = TRUE,
  results.subtitle = TRUE,
  xlab = NULL,
  ylab = NULL,
  caption = NULL,
  title = NULL,
  subtitle = NULL,
  digits = 2L,
  conf.level = 0.95,
  tr = 0.2,
  alternative = "two.sided",
  centrality.plotting = TRUE,
  centrality.type = type,
  centrality.point.args = list(size = 5, color = "darkred"),
  centrality.label.args = list(size = 3, nudge_x = 0.4, segment.linetype = 4,
    min.segment.length = 0),
  point.args = list(position = ggplot2::position_jitterdodge(dodge.width = 0.6), alpha =
    0.4, size = 3, stroke = 0, na.rm = TRUE),
  boxplot.args = list(width = 0.3, alpha = 0.2, na.rm = TRUE),
  violin.args = list(width = 0.5, alpha = 0.2, na.rm = TRUE),
  ggsignif.args = list(textsize = 3, tip_length = 0.01, na.rm = TRUE),
  ggtheme = ggstatsplot::theme_ggstatsplot(),
  palette = "ggthemes::gdoc",
  ggplot.component = NULL,
  ...
)

Arguments

data

A data frame (or a tibble) from which variables specified are to be taken. Other data types (e.g., matrix, table, array, etc.) will not be accepted. Additionally, grouped data frames from {dplyr} should be ungrouped before they are entered as data.

x

The grouping (or independent) variable from data.

y

The response (or outcome or dependent) variable from data.

type

A character specifying the type of statistical approach:

"parametric"
"nonparametric"
"robust"
"bayes"

You can specify just the initial letter (e.g. "np" or "bf"). Matching is on the initial lowercase letter only, so any other value (including capitalized values such as "Bayes") falls back to "parametric" without a warning.

pairwise.display

Decides which pairwise comparisons to display. Available options are:

"significant" (abbreviation accepted: "s")
"non-significant" (abbreviation accepted: "ns")
"all"
"none"

Use this to keep the plot uncluttered when many groups are compared. Pairwise comparisons are carried out (and displayed) only when x has three or more levels. For type = "bayes", there are no p-values, so all comparisons are displayed unless this is "none".

pairwise.alpha

Numeric alpha threshold used to decide which pairwise comparisons are displayed when pairwise.display = "significant" or pairwise.display = "non-significant" (Default: 0.05). It is compared against p-values adjusted using p.adjust.method.

p.adjust.method

Adjustment method for p-values for multiple comparisons. Possible methods are: "holm" (default), "hochberg", "hommel", "bonferroni", "BH", "BY", "fdr", "none".

bf.prior

A number between 0.5 and 2 (default 0.707), the prior width to use in calculating Bayes factors and posterior estimates. In addition to numeric arguments, several named values are also recognized: "medium", "wide", and "ultrawide", corresponding to r scale values of 1/2, sqrt(2)/2, and 1, respectively. In case of an ANOVA, this value corresponds to scale for fixed effects.

bf.message

Logical that decides whether to display the Bayes Factor in favor of the null hypothesis as a caption. Relevant only for parametric tests (Default: TRUE).

results.subtitle

Logical that decides whether the results of statistical tests are displayed as a subtitle (Default: TRUE). Setting this to FALSE also skips the automatic Bayes Factor caption. Other plot details have their own arguments.

xlab

Label for x axis variable. If NULL (default), variable name for x will be used.

ylab

Label for y axis variable. If NULL (default), variable name for y will be used.

caption

The text for the plot caption. It will be replaced by the Bayes Factor caption whenever that is displayed (see bf.message).

title

The text for the plot title.

subtitle

The text for the plot subtitle. Used only if results.subtitle = FALSE.

digits

Number of digits for rounding or significant figures. May also be "signif" to return significant figures or "scientific" to return scientific notation. Control the number of digits by adding the value as suffix, e.g. digits = "scientific4" to have scientific notation with 4 decimal places, or digits = "signif5" for 5 significant figures (see also signif()).

conf.level

Scalar between 0 and 1 (default: 95% confidence/credible intervals, 0.95).

tr

Trim level for the mean when carrying out robust tests. In case of an error, try reducing the value of tr, which is by default set to 0.2. Lowering the value might help.

alternative

a character string specifying the alternative hypothesis, must be one of "two.sided" (default), "greater" or "less". You can specify just the initial letter.

centrality.plotting

Logical that decides whether central tendency measure is to be displayed as a point with a label (Default: TRUE). Function decides which central tendency measure to show depending on the type argument.

mean for parametric statistics
median for non-parametric statistics
trimmed mean for robust statistics
MAP estimator for Bayesian statistics

Use centrality.type to choose a different centrality measure.

centrality.type

Decides which centrality measure is displayed. Defaults to the same value as type. Can be one of:

"parametric" (for mean)
"nonparametric" (for median)
"robust" (for trimmed mean)
"bayes" (for MAP estimator)

As with type, abbreviations are accepted.

centrality.point.args, centrality.label.args

A list of additional aesthetic arguments passed to ggplot2::geom_point() and ggrepel::geom_label_repel(), respectively, which display the centrality measure.

point.args

A list of additional aesthetic arguments passed to ggplot2::geom_point().

boxplot.args

A list of additional aesthetic arguments passed to ggplot2::geom_boxplot(). By default, the whiskers extend to 1.5 times the interquartile range (IQR) from the box (Tukey-style). Use coef to change this, e.g., boxplot.args = list(coef = 3) extends the whiskers to the most extreme observations within 3 * IQR of the hinges.

violin.args

A list of additional aesthetic arguments passed to ggplot2::geom_violin().

ggsignif.args

A list of additional aesthetic arguments passed to ggsignif::geom_signif().

ggtheme

A {ggplot2} theme (Default: theme_ggstatsplot()). Any {ggplot2} theme (e.g., ggplot2::theme_bw()) or theme from an extension package (e.g., ggthemes::theme_fivethirtyeight()) can be used. Note that some themes remove plot details. For example, ggthemes::theme_fivethirtyeight() removes the secondary Y-axis, which ggbetweenstats() uses to show details about pairwise comparisons.

palette

Name of the discrete palette in "package::palette" format to be used for coloring (via {paletteer}). Run View(paletteer::palettes_d_names) to see all available options.

ggplot.component

A {ggplot2} component, or a list of components, to add to the plot (Default: NULL). This is mostly helpful for the grouped_ variants, where it is applied to each individual plot.

...

Currently ignored.

Value

A ggplot object, which can be further modified with {ggplot2} functions. Use extract_stats() to get the statistical details shown in the plot as data frames.

Details

Rows with a missing value in either x or y are removed before plotting and analysis, and unused levels of x are dropped.

For details, see: https://www.indrapatil.com/ggstatsplot/articles/web_only/ggbetweenstats.html

Summary of graphics

graphical element	`geom` used	argument for further modification
raw data	`ggplot2::geom_point()`	`point.args`
box plot	`ggplot2::geom_boxplot()`	`boxplot.args`
density plot	`ggplot2::geom_violin()`	`violin.args`
centrality measure point	`ggplot2::geom_point()`	`centrality.point.args`
centrality measure label	`ggrepel::geom_label_repel()`	`centrality.label.args`
pairwise comparisons	`ggsignif::geom_signif()`	`ggsignif.args`

Statistical defaults

This function uses statistically justified defaults that are not user-configurable:

For parametric tests, effect sizes use Hedges' g for two groups and omega-squared for three or more groups. These estimators correct for positive bias, which can be especially relevant in small samples.
Welch's t-test and one-way test are used instead of Student's versions, so equal variances are not assumed.

Users who need non-default values for these settings can call {statsExpressions} directly.

Centrality measures

Type	Measure	Function used
Parametric	mean	`datawizard::describe_distribution()`
Non-parametric	median	`datawizard::describe_distribution()`
Robust	trimmed mean	`datawizard::describe_distribution()`
Bayesian	MAP	`datawizard::describe_distribution()`

Two-sample tests

Below is a summary of:

the statistical test carried out for inferential statistics
the type of effect size estimate and a measure of uncertainty for this estimate
the functions used internally to compute these details

Between-subjects

Hypothesis testing

Type	No. of groups	Test	Function used
Parametric	2	Student's or Welch's t-test	`stats::t.test()`
Non-parametric	2	Mann-Whitney U test	`stats::wilcox.test()`
Robust	2	Yuen's test for trimmed means	`WRS2::yuen()`
Bayesian	2	Student's t-test	`BayesFactor::ttestBF()`

Effect size estimation

Type	No. of groups	Effect size	CI available?	Function used
Parametric	2	Cohen's d, Hedges' g	Yes	`effectsize::cohens_d()`, `effectsize::hedges_g()`
Non-parametric	2	r (rank-biserial correlation)	Yes	`effectsize::rank_biserial()`
Robust	2	Algina-Keselman-Penfield robust standardized difference	Yes	`WRS2::akp.effect()`
Bayesian	2	difference	Yes	`bayestestR::describe_posterior()`

Within-subjects

Paired data: See complete_subjects() for subject matching, duplicate observations, and removal of incomplete subjects.

Hypothesis testing

Type	No. of groups	Test	Function used
Parametric	2	Student's t-test	`stats::t.test()`
Non-parametric	2	Wilcoxon signed-rank test	`stats::wilcox.test()`
Robust	2	Yuen's test on trimmed means for dependent samples	`WRS2::yuend()`
Bayesian	2	Student's t-test	`BayesFactor::ttestBF()`

Effect size estimation

Type	No. of groups	Effect size	CI available?	Function used
Parametric	2	Cohen's d, Hedges' g	Yes	`effectsize::cohens_d()`, `effectsize::hedges_g()`
Non-parametric	2	r (rank-biserial correlation)	Yes	`effectsize::rank_biserial()`
Robust	2	Algina-Keselman-Penfield robust standardized difference	Yes	`WRS2::wmcpAKP()`
Bayesian	2	difference	Yes	`bayestestR::describe_posterior()`

One-way ANOVA

Below is a summary of:

the statistical test carried out for inferential statistics
the type of effect size estimate and a measure of uncertainty for this estimate
the functions used internally to compute these details

Between-subjects

Hypothesis testing

Type	No. of groups	Test	Function used
Parametric	> 2	Fisher's or Welch's one-way ANOVA	`stats::oneway.test()`
Non-parametric	> 2	Kruskal-Wallis one-way ANOVA	`stats::kruskal.test()`
Robust	> 2	Heteroscedastic one-way ANOVA for trimmed means	`WRS2::t1way()`
Bayesian	> 2	Fisher's ANOVA	`BayesFactor::anovaBF()`

Effect size estimation

Type	No. of groups	Effect size	CI available?	Function used
Parametric	> 2	partial eta-squared, partial omega-squared	Yes	`effectsize::omega_squared()`, `effectsize::eta_squared()`
Non-parametric	> 2	rank epsilon squared	Yes	`effectsize::rank_epsilon_squared()`
Robust	> 2	Explanatory measure of effect size	Yes	`WRS2::t1way()`
Bayesian	> 2	Bayesian R-squared	Yes	`performance::r2_bayes()`

Within-subjects

Paired data: See complete_subjects() for subject matching, duplicate observations, and removal of incomplete subjects.

Hypothesis testing

Type	No. of groups	Test	Function used
Parametric	> 2	One-way repeated measures ANOVA	`afex::aov_ez()`
Non-parametric	> 2	Friedman rank sum test	`stats::friedman.test()`
Robust	> 2	Heteroscedastic one-way repeated measures ANOVA for trimmed means	`WRS2::rmanova()`
Bayesian	> 2	One-way repeated measures ANOVA	`BayesFactor::anovaBF()`

Effect size estimation

Type	No. of groups	Effect size	CI available?	Function used
Parametric	> 2	partial eta-squared, partial omega-squared	Yes	`effectsize::omega_squared()`, `effectsize::eta_squared()`
Non-parametric	> 2	Kendall's coefficient of concordance	Yes	`effectsize::kendalls_w()`
Robust	> 2	Algina-Keselman-Penfield robust standardized difference average	Yes	`WRS2::wmcpAKP()`
Bayesian	> 2	Bayesian R-squared	Yes	`performance::r2_bayes()`

Pairwise comparison tests

Below is a summary of:

the statistical test carried out for inferential statistics
the type of effect size estimate and a measure of uncertainty for this estimate
the functions used internally to compute these details

Between-subjects

Hypothesis testing

Type	Equal variance?	Test	p-value adjustment?	Function used
Parametric	No	Games-Howell test	Yes	`PMCMRplus::gamesHowellTest()`
Parametric	Yes	Student's t-test	Yes	`stats::pairwise.t.test()`
Non-parametric	No	Dunn test	Yes	`PMCMRplus::kwAllPairsDunnTest()`
Robust	No	Yuen's trimmed means test	Yes	`WRS2::lincon()`
Bayesian	`NA`	Student's t-test	`NA`	`BayesFactor::ttestBF()`

Effect size estimation

Not supported.

Within-subjects

Paired data: See complete_subjects() for subject matching, duplicate observations, and removal of incomplete subjects.

Hypothesis testing

Type	Test	p-value adjustment?	Function used
Parametric	Student's t-test	Yes	`stats::pairwise.t.test()`
Non-parametric	Durbin-Conover test	Yes	`PMCMRplus::durbinAllPairsTest()`
Robust	Yuen's trimmed means test	Yes	`WRS2::rmmcp()`
Bayesian	Student's t-test	`NA`	`BayesFactor::ttestBF()`

Effect size estimation

Not supported.

Examples

# for reproducibility
set.seed(123)

p <- ggbetweenstats(mtcars, am, mpg)
p


# extracting details from statistical tests
extract_stats(p)
#> $subtitle_data
#> # A tibble: 1 × 18
#>   parameter1 parameter2 mean.parameter1 mean.parameter2 statistic df.error
#>   <chr>      <chr>                <dbl>           <dbl>     <dbl>    <dbl>
#> 1 mpg        am                    17.1            24.4     -3.77     18.3
#>   p.value method                  alternative effectsize estimate conf.level
#>     <dbl> <chr>                   <chr>       <chr>         <dbl>      <dbl>
#> 1 0.00137 Welch Two Sample t-test two.sided   Hedges' g     -1.35       0.95
#>   conf.low conf.high conf.method conf.distribution n.obs expression
#>      <dbl>     <dbl> <chr>       <chr>             <int> <list>    
#> 1    -2.17    -0.512 ncp         t                    32 <language>
#> 
#> $caption_data
#> # A tibble: 1 × 16
#>   term       effectsize      estimate conf.level conf.low conf.high    pd
#>   <chr>      <chr>              <dbl>      <dbl>    <dbl>     <dbl> <dbl>
#> 1 Difference Bayesian t-test    -6.44       0.95    -10.1     -2.74 0.999
#>   prior.distribution prior.location prior.scale  bf10 method         
#>   <chr>                       <dbl>       <dbl> <dbl> <chr>          
#> 1 cauchy                          0       0.707  86.6 Bayesian t-test
#>   conf.method log_e_bf10 n.obs expression
#>   <chr>            <dbl> <int> <list>    
#> 1 ETI               4.46    32 <language>
#> 
#> $pairwise_comparisons_data
#> NULL
#> 
#> $descriptive_data
#> NULL
#> 
#> $one_sample_data
#> NULL
#> 
#> $tidy_data
#> NULL
#> 
#> $glance_data
#> NULL
#> 
#> attr(,"class")
#> [1] "ggstatsplot_stats" "list"             

# show non-significant pairwise comparisons (needs 3+ groups for ggsignif)
ggbetweenstats(mtcars, cyl, mpg, pairwise.display = "non-significant")


# show all pairwise comparisons
ggbetweenstats(mtcars, cyl, mpg, pairwise.display = "all")


# use a stricter alpha threshold for significant pairwise comparisons
ggbetweenstats(mtcars, cyl, mpg, pairwise.alpha = 0.001)


# modifying defaults
ggbetweenstats(
  morley,
  x    = Expt,
  y    = Speed,
  type = "robust",
  xlab = "The experiment number",
  ylab = "Speed-of-light measurement"
)


# you can remove a specific geom to reduce complexity of the plot
ggbetweenstats(
  mtcars,
  am,
  wt,
  # to remove violin plot
  violin.args = list(width = 0, linewidth = 0, colour = NA),
  # to remove boxplot
  boxplot.args = list(width = 0),
  # to remove points
  point.args = list(alpha = 0)
)