N.B. All statistical analysis in {ggstatsplot} is
carried out in {statsExpressions}. Thus, to see changes
related to statistical expressions, read the NEWS for that
package: https://indrajeetpatil.github.io/statsExpressions/news/index.html
extract_stats() returns a list of class
ggstatsplot_stats which contains all the statistical
summaries and expressions for a given plot.
extract_stats(), extract_subtitle(),
extract_caption() now works out of the box for the grouped
plots as well.
ggpiestats() and ggbarstats() now respect
ratio() argument for proportion tests run in case of
two-way contingency tables (#818).bugs_wide.gghistostats()) has been removed. This feature always felt
like an ad hoc addition to the plot, and has nothing to do with the key
statistical analysis in question (which is not about checking the
normality of the distribution).ggpiestats() and ggbarstats() (#935).The cryptic but very useful parameter k has been
renamed to digits to improve its discoverability.
To be consistent with other functions, ggpiestats()
and ggbarstats() now default to two-sided alternative
hypothesis.
Maintenance updates for changes in upstream dependencies.
ggbarstats() gains
sample.size.label.args parameter to pass additional
arguments to ggplot2::geom_text().
To be internally consistent, the plot.type argument
has been removed from ggbetweenstats(), since no such
argument exists for ggwithinstats(). This argument was also
redundant. Since removing a specific geom is straightforward using
*.args arguments. Examples for these two functions
illustrate how.
ggbetweenstats() and ggwithinstats()
retire pairwise.comparisons argument since it was
redundant. In order to turn off showing pairwise comparisons, you can
now use pairwise.display = "none".
ggbetweenstats() gets boxplot.args
argument to pass additional arguments to the underlying geom function.
This also fixes regression introduced in 0.11.1 release
where outlier points were displayed along with box plot.ggbetweenstats()
and ggwithinstats() has been removed. It was too crude to
be useful or reliable, and users should instead prefer more informative
methods (e.g. performance::check_outliers()).{parameters}.R 4.1
because a crucial dependency ({pbkrtest}) requires this R
version.{ggplot2} and
{easystats} updates.output parameter for all functions has been
removed. All functions now return only the plot, which itself contains
all necessary details that were previously extracted using the
output argument. You can extract all the necessary details
(including expressions containing statistical details) from a plot using
extract_stats() function. There are two additional helpers
to get expressions: extract_subtitle() and
extract_caption().xfill and yfill arguments for
ggscatterstats() have been removed. You can specify all
aesthetic modifications for side histograms in scatter plot using
xsidehistogram.args and ysidehistogram.args
arguments.
Updates to changes made in the latest {ggplot2}
release (3.4.0).
Due to changes to the underlying API of
{parameters}, the effsize argument has been
renamed to effectsize.type.
Removes unnecessary re-exports of {tidyverse}
operators.
The pairwise_comparions() function implementation
now lives in {statsExpressions} package, although it will
continue to be exported from {ggstatsplot}
package.
The details about pairwise test for ggbetweenstats()
and ggwithinstats() functions are now displayed as a label
for the secondary axis. Previously, this information was displayed in
the caption. Given that caption already contained Bayesian test details,
it was becoming difficult to stack different expressions on top of each
other. To avoid unnecessary code complexity and also to avoid crowded
caption, this decision was made. Additionally, the pairwise test label
has been slightly abbreviated, and so is the label for significance
bars. This is done to not let the text overwhelm the numeric values, the
latter being more important.
{PMCMRplus} package from Imports to Suggests. So,
if, as a user, you wish to use pairwise comparisons in
ggbetweenstats() and ggwithinstats(), you will
need to download this package.pairwise_comparisons() function for carrying out
one-way pairwise comparisons has now moved in {ggstatsplot}
from {pairwiseComparisons} package.A number of effect size estimates and their confidence intervals
have changed due to respective changes made in {effectsize}
package version 0.5 release. For full details of these
changes, see: https://easystats.github.io/effectsize/news/index.html
For the same reason, the effect size for one-way contingency table has changed from Cramer’s V to Pearson’s C.
For plotting marginal distributions in
ggscatterstats, {ggstatsplot} now relies on
{ggside} package instead of {ggExtra}. This
was done to remove a glaring inconsistency in the API. All functions in
{ggstatsplot} produced ggplot objects and
could be further modified with ggplot2 functions, except
ggscatterstats, which led to a lot of confusion among users
(e.g. #28). This change gets rid of this inconsistency. But it comes at
a cost: there is no more marginal.type argument that lets
you change the type of marginal distribution graphic and histogram is
the only possible option. Note that this is not a
breaking change. Your past code will continue to work but it will now
always produce a histogram instead of other marginal graphic you might
have chosen.
Minimum needed R version is now 4.0.
Online vignette about combine_plots has been
removed. In case you want to create a grid of plots, it is highly
recommended that you use patchwork package directly and not
this wrapper around it which is mostly useful with
{ggstatsplot} plots.
ggscatterstats labeling arguments accept only
unquoted inputs now, and not quoted or string inputs. Allowing this was
a bad design choice in the past since most functions in
{ggstatsplot}, inspired by tidyverse, expect
unquoted (x) - and not quoted ("x") -
arguments. So this function was the odd one out.
Gets rid of ipmisc dependency.
Removes movies_wide dataset, which was virtually
identical to movies_long dataset and was not used anywhere
in the package. Also removes the unused VR_dilemma
dataset.
extract_stats function to extract dataframes
containing statistical details.There is finally a publication for {ggstatsplot}
package! https://joss.theoj.org/papers/10.21105/joss.03167
The ggcoefstats function defaults to
NULL for xlab and ylab arguments,
which lets users change these labels if they wish to do so.
Additionally, the x-axis label, if not specified, now defaults to
"estimate". Whether this estimate corresponds to regression
coefficient or effect size like partial eta-squared should be clear from
the label itself.
To reduce the dependency load, ggcorrplot moves from
Imports to Suggests.
The bar.fill argument in gghistostats
is retired in favor of the new bin.args argument that can
be used to pass aesthetic arguments to
ggplot2::stat_bin.
ggstatsplot.layer argument has been retired. If the
user chooses a certain ggplot2 theme, it means
they want that theme, and not {ggstatsplot}’s
varnish on it. So the previous behavior was undesirable. This is a
backward compatible change, so the plots should not look
different.
The pch size for ggcorrmat has been
increased to 14 (#579) to increase its visibility compared to the
correlation value text.
ggwithinstats gains point.args to
change geom_point.
Minor change to ggcorrmat legend title - content in
parentheses is now shown outside of it.
ggcoefstats didn’t work when statistic for the given
model was chi-squared. This has been fixed.To reduce the dependency load, ggExtra moves from
Imports to Suggests.
All functions are more robust in the sense that when
statistical analysis fails, they will return only the plots with no
subtitles/captions. This helps avoid difficult-to-diagnose edge case
failures when the primary functions are used in grouped_
functions (e.g., #559). The ggpiestats and
ggbarstats functions always behaved this way, but the rest
of the functions now also mimic this behavior.
ggcoefstats labels do not contain degrees of
freedom when they are not available instead of displaying
Inf.title.prefix is now removed. It led to redundant title
prefixes across different facets of the plot. Given that
grouped_ functions require users to set
grouping.var, it is fair to assume what variable the levels
in the title correspond to.Adapts to changes made in
statsExpressions 1.0.0.
sample.size.label argument is retired for
ggbetweenstats, ggwithinstats, and
ggbarstats. I do not think it is ever a good idea to not do
this. If the users wish to not display sample sizes, they can easily do
this using scale_* functions from
ggplot2.
In ggpiestats and ggbarstats,
parametric proportion tests are now turned off when
type = "bayes".
combine_plots has been completely revised to rely
not on patchwork, but on patchwork, to combine
a list of ggplot together. This was done to have a leaner
syntax. With this revision, its vestigial twin
combine_plots is no longer needed and has been removed.
This should not break any of the existing instances of
grouped_ functions, although it will lead to changed
graphical layouts. The only instance in which this change will lead to a
breakage is when you specified labels argument. So, if you
used plotgrid.args = list(labels = "auto"), you will now
have to replace it with
plotgrid.args = list(tag_level = "keep"). You can also use
annotation.args (e.g.,
annotation.args = list(tag_levels = "a") to customize
labels (this will create labels with pattern a,
b, c, etc.). Another instance of breakage is
if you had used combine_plots function and provided
individual plots to ... instead as a
list.
To avoid confusion among users, the default trimming level for
all functions is now changed from tr = 0.1 to
tr = 0.2 (which is what WRS2 defaults
to).
All robust tests in this package were based on trimmed means,
except for correlation test. This has been changed: the robust
correlation measure is now Winsorized correlation, which is based on
trimming. Therefore, the beta argument has been replaced by
tr argument. This should result only in minor changes in
correlation coefficient estimates.
Using annotate instead of geom_label
had significantly slowed down gghistostats and
ggdotplotstats functions. This has been fixed.
Removes the vestigial notch and
notchwidth arguments for ggbetweenstats and
ggwithinstats.
All Bayesian expression templates are now explicit about the type of estimate being displayed.
For gghistostats and ggdotplotstats,
the centrality measure labels used to be attached to the vertical line,
but this occluded the underlying data. Now this label is instead shown
on the top x-axis. Note that this means that if you make
any further changes to the resulting plot using the
ggplot2::scale_x_continuous function, this label will
likely disappear. The centrality.k argument is retired in
favor of k.
More models supported in ggcoefstats:
crr, eglm, elm,
varest.
ggbetweenstats, ggwithinstats,
gghistostats, ggdotplotstats gain argument
centrality.type that can be used to specify which
centrality parameter is to be displayed. So one can have
type = "robust" and still show median as centrality
parameter by choosing
centrality.type = "nonparametric".
gghistostats removes bar.measure
argument. The function now defaults to showing the count
information on the x-axis and the proportion
information on the duplicated x-axis.
ggscatterstats removes method and
method.args arguments. It will no longer be possible to use
this function to visualize data for when the model is not linear. It
also retires margins argument.
For ggbetweenstats and ggwithinstats
functions, the arguments of type mean.* have all been
replaced by centrality.*. This is because now these
functions decide which central tendency measure to show depending on the
type argument (mean for parametric,
median for non-parametric, trimmed
mean for robust, and MAP estimator for
Bayes).
Similarly, gghistostats and
ggdotplotstats functions also decide which central tendency
measure to show depending on the type argument
(mean for parametric, median for
non-parametric, trimmed mean for robust, and
MAP estimator for Bayes). Therefore,
centrality.parameter argument has been removed. If you want
to turn off displaying centrality measure, set
centrality.plotting = FALSE.
gghistostats and ggdotplotstats
functions remove the functionality to display a vertical line
corresponding to test.value. This feature was turned off by
default in prior releases. Accordingly, all related arguments from these
two functions have been removed.
ggscatterstats defaults to densigram as
the marginal distribution visualization.
ggbetweenstats and ggwithinstats now
display the centrality tendency measure in such a way that the label
doesn’t occlude any of the raw data points (#429).
mean.ci argument is retired for
ggbetweenstats and ggwithinstats. Future
{ggstatsplot} releases will be providing different
centrality measures depending on the type argument and it
is not guaranteed that all of them will have CIs available. So, for the
sake of consistency, this argument is just going to be retired.
ggcorrmat uses pretty formatting to display sample
size information.
ggcoefstats now also displays degrees of freedom for
chi-squared tests.
Expects minor changes in some of the effect sizes and their
confidence intervals due to changes in
{statsExpressions}.
More models supported in ggcoefstats:
fixest, ivFixed, ivprobit,
riskRegression.
ggcorrmat supports partial correlations.
ggcoefstats no longer supports
exponentiate argument. If it is specified, the user will
have to themselves adjust the scales appropriately.
ggcorrmat defaults have changed significantly:
As a matter of good practice, the p-values are adjusted by default for multiple comparisons.
The default matrix is upper type, and not the full matrix, which features many redundant comparisons and self-correlations diagonally.
Default text size for legend has been increased to 15 and background grid has been removed.
In the prior release, when the GitHub version of
BayesFactor wasn’t present, ggwithinstats just
outright failed to run for ANOVA designs. This has been fixed.
Setting mean.path = FALSE in
ggwithinstats produced incorrect colors for points (#470).
This bug was introduced in 0.6.5 and is now fixed.
If user had set options(scipen = 999) in their
session, the p-value formatting for ggpiestats and
ggcoefstats looked super-ugly (#478). This has been
fixed.
Drops broomExtra from dependencies. All regression
modeling-related analysis now relies on easystats
ecosystem.
ggpiestats and ggbarstats don’t support
returning dataframes. See FAQ vignette on how to get these dataframes:
https://indrajeetpatil.github.io/ggstatsplot/articles/web_only/faq.html#faq-1
ggpiestats and ggbarstats were not
supposed to support returning Bayes Factor for paired contingency table
analysis, which is not supported in BayesFactor
itself.
ggcoefstats defaults to displaying the intercept
term. Also, when the degrees of freedom are not available for
t-statistic, they are displayed to be Inf, in
keeping with easystats conventions.
Instead of showing significance of p-values with APA’s
asterisks conventions, ggbarstats now instead shows the
actual p-values from one-sample proportion tests.
ggcoefstats:
Glm.ggpiestats and ggbarstats no longer have
the vestigial arguments main and condition,
which are superseded by x and y,
respectively.For consistency and to reduce confusion, all Bayes Factor
(irrespective of whether in the subtitle or caption) are always in favor
of null over alternative (BF01).
Retires centrality parameter tagging functionality of
ggscatterstats. Although it was not the default, when
turned on, it definitely created a cluttered plot.
ggbetweenstats and ggwithinstats functions
now default to pairwise.comparisons = TRUE.Plot borders are now removed from the default theme.
Small p-values (< 0.001) are now displayed in scientific notation.
pairwiseComparisons re-exports are deprecated.More models supported in ggcoefstats:
BFBayesFactor, betamfx, crq,
coxph.penal, geeglm, glht,
glmm, lm_robust, lqm,
lqmm, manova, maov,
margins, negbinmfx, logitmfx,
logitsf, margins, poissonmfx,
betaor, negbinirr, logitor,
metafor, metaplus, orm,
poissonirr, semLm, semLme,
vgam.
ggpiestats gains label.repel argument
to cover contexts in which the labels might overlap. Setting it to
TRUE will minimize such an overlap.
ggbetweenstats and ggwithinstats gain
ggsignif.args argument to make it easy to change aesthetics
of the pairwise comparison geom.
The subtitle and caption for Bayes Factor tests now also provide information about posterior estimates, when relevant.
Removed unused intent_morality dataset.
ggcoefstats retires caption.summary
argument. So, by default, the caption is going to contain as much
information as it can and the users can then choose to modify the
default caption using ggplot2 functions.
The argument method for ggcorrmat has
been renamed to matrix.method, since it was confusing
whether this method referred to correlation method.
For both ggpiestats and ggbarstats, the
count labels no longer include n = in them as this was
confusing since all labels had n = in them with no further
explanation about how this n differed from n
in the proportion test.
No longer relies on groupedstats package.
The pairwise.annotation argument for
ggbetweenstats and ggwithinstats is
deprecated. This was done because-
Different fields have different schema for what significance levels asterisks represent.
The p-value labels also contain information about whether they are adjusted for multiple comparisons.
The normality_message and
bartlett_message helper functions have been removed. This
is because model assumption checks don’t really fall under the purview
of this package. There are excellent visualization tools out there for
model assumption checks (ggResidpanel,
performance, DHARMa, olsrr,
etc.), which should be preferred over unhelpful messages with only
p-values that these functions were printing. For what it’s
worth, the functions where these messages were displayed
(ggbetweenstats or ggwithinstats) feature
visualizations rich enough and defaults sensible enough that most of the
time one can either assess these assumptions from the plots or need not
worry about them.
ggcoefstats has been refactored to reflect that
broomExtra::tidy_parameters now defaults to
parameters package instead of broom. It also
loses the following vestigial arguments: p.adjust.method
and coefficient.type.
Reverts aligning title and subtitle with the plot and not the
axes, since it looked pretty ugly (esp., ggcoefstats) and
was causing problems for labels.
factor.levels (for ggpiestats) and
labels.legend (for ggbarstats) are deprecated.
If users would like to changes the names for factor levels, this should
be done outside of {ggstatsplot}.
The non-parametric post hoc test for between-subjects design has been changed from Dwass-Steel-Crichtlow-Fligner test to Dunn test.
ggcoefstats:
bayesGARCH, clm2, clmm2,
mcmc.list, robmixglm.ggcorrmat no longer returns matrices of correlation
coefficients or other details. It now returns either a plot or a data
frame and this can data frame can then be used to create
matrices.
ggbarstats loses x.axis.orientation
argument. This argument was supposed to help avoid overlapping
x-axis label, but now ggplot2 3.3.0 has a better
way to handle this: https://www.tidyverse.org/blog/2020/03/ggplot2-3-3-0/#rewrite-of-axis-code
More models supported in ggcoefstats:
bayesx, BBmm, brmultinom,
lmerModLmerTest, lrm.
Specifying output = "proptest" for
ggpiestats and ggbarstats functions will now
return a data frame containing results from proportion test.
ggbetweenstats and ggwithinstats will
display pairwise comparisons even if results.subtitle is
set to FALSE.
ggcorrmat supports computing Bayes Factors for
Pearson’s r correlation.
ggbetweenstats and ggwithinstats now
support pairwise comparisons for Bayes Factor test.
For changes related to subtitle details, see changes made in new
version of statsExpressions 4.0.0: https://CRAN.R-project.org/package=statsExpressions/news/news.html
ggbetweenstats and ggwithinstats no
longer print dataframes containing results from pairwise comparisons
tests because this is too cluttering for the user’s console. The users
are now instead advised to either extract this data frame using
ggplot2::ggplot_build() function or use the
pairwiseComparisons::pairwise_comparisons() function used
in the background by {ggstatsplot} to carry out this
analysis.
Due to changes in one of the downstream dependencies,
{ggstatsplot} now expects the minimum R version to be
3.6.0.
ggcorrmat now internally relies on
correlation for correlation analyses.
ggbarstats no longer displays "percent"
for Y-axis label as this was redundant information.
Continuing the argument cleanup that began in 0.3.0,
ggcoefstats gains point.args argument instead
of individuals point.* arguments.
The subtitles are more explicit about the details of the test.
For the same reason stat.title argument from all relevant
functions is retired since this argument was supposed to be for entering
some additional details about the test. Additionally, the plot titles
and subtitles for some of the plots are aligned with the plot.
ggcorrmat legend, in case of missing values, shows
mode - instead of median - for the distribution of sample
pairs.
The following vestigial arguments are retired:
caption.default in ggcorrmat
k.caption.summary in
ggcoefstats
This is a hotfix release to correct some of the failing tests and
other minor breakages resulting from the new release of
ggplot2 3.3.0.
ggpiestats loses sample.size.label
argument since this information is included in the goodness of fit test
results itself. So setting proportion.test to
FALSE will suppress this information.To give users more flexibility in terms of modifying the aesthetic
defaults for all geoms included in the
{ggstatsplot} plots (each plot typically has multiple
geoms), the package now uses a new form of syntax. Previously, each
geom had a separate argument to specify each aesthetic
(e.g., geom_point would get arguments like
point.size, point.color, etc.), which resulted
in functions with a massive number of arguments and was unsustainable in
the long run. Instead, {ggstatsplot} functions now expect a
list of such arguments for the respective geom (e.g.,
geom_point will have point.args argument where
a list of arguments
list(size = 5, color = "darkgreen", alpha = 0.8) can be
supplied).
All grouped_ functions have been refactored to
reduce the number of arguments. These functions now internally use the
new combine_plots instead of combine_plots.
The additional arguments to primary functions can be provided through
.... These changes will not necessarily break the existing
code but will lead to some minor graphical changes (e.g., if you were
providing labels argument explicitly, it will be
ignored).
All functions lose the return argument, which was
supposed to be alternative to enter output. But this was
just leading to more confusion on the user’s part. The biggest
user-visible impact this is going to have is that ggcorrmat
will no longer be backward-compatible. The older scripts will still work
but if the return argument was anything except
"plot", it will just be ignored.
ggcorrmat no longer has corr.method
argument. To be consistent with rest of the functions in this package,
the type of statistics should be specified using type
argument. Additional, it gains a new argument
ggcorrplot.args, which can be used to pass additional
arguments to the underlying plotting function
(ggcorrplot::ggcorrplot).
Both gghistostats and ggdotplotstats
now use the following arguments to modify geoms
corresponding to the lines and labels:
test.value.line.args, test.value.label.args,
centrality.line.args, centrality.label.args.
This helps avoid specifying millions of arguments.
Removes the vestigial ggplot_converter
function.
ggpiestats and ggbarstats remove the
following vestigial arguments: facet.wrap.name,
bias.correct, bar.outline.color. The
bar.proptest and facet.proptest arguments were
difficult to remember and confusing and are replaced by a common
proportion.test argument. Additionally, the following
arguments have all been removed and replaced by label
argument: slice.label, bar.label,
data.label. These plethora of options was a headache to
remember.
gghistostats loses the following arguments:
fill.gradient, low.color,
high.color. It made no sense to add a color gradient to
this plot when the Y-axis already displayed the information about what
the bar represented.
ggscatterstats loses the following arguments:
palette and package. Since this function
requires only two colors, it didn’t make much sense to use color
palettes to specify this. They can be instead specified using
xfill and yfill. You can always use
paletteer::paletteer_d to get a vector of color values and
then provide values of your choosing to xfill and
yfill.
Removes sorting options in ggbetweenstats and
ggwithinstats functions. This is something the users can
easily do before entering the data in these functions.
ggcorrmat was never supposed to work with Kendall’s
correlation coefficient but it accidentally did. This is no longer the
case.
{ggstatsplot} now has a logo, thanks to Sarah!
:)
The default theme_ggstatsplot changes slightly. The
biggest change is that the title and the subtitle for plots are now
aligned to the left of the plot. This change also forced the legend for
ggpiestats to be displayed on the right side of the plot
rather than at the bottom.
More models supported in ggcoefstats:
BBreg, bcplm, bife,
cglm, crch, DirichReg,
LORgee, zcpglm,
zeroinfl.
Following functions are now re-exported from ipmisc:
bartlett_message, normality_message. A few
other internal data wrangling functions now reside in
ipmisc.
To have a more manageable length of function arguments,
additional aesthetic specifications for any given geom can be provided
via a dedicated *.args argument. For example, all aesthetic
arguments for geom_vline can be provided via
vline.args, for geom_errorbarh via
errorbar.args, etc.
{ggstatsplot} continues with its conscious
uncoupling that started in 0.1.0 release: The following
functions have now been moved to {statsExpressions}
package: subtitle_meta_parametric and
bf_meta_message and follow a more logical nomenclature. For
the same reason, lm_effsize_ci function is also no longer
exported and lives in the groupedstats package.
The summary caption no longer displays log-likelihood value because it tends to be not available for a number of regression model objects and so the caption was unnecessarily being skipped.
Supports robust and Bayes Factors for random-effects meta-analysis.
New dataset included: bugs_wide
More models supported in ggcoefstats:
cgam, cgamm, coxme,
cpglm, cpglmm, complmrob,
feis, flexsurvreg, glmx,
hurdle, iv_robust, mixor,
rqss, truncreg, vgam.
Removed vestigial arguments from ggcorrmat (e.g.,
exact, continuity, etc.) and
ggpiestats (bf.prior,
simulate.p.value, B, etc.).
ggbetweenstats and ggwithinstats no longer
produce error with variables with pattern mean (#336).pairwise_p has been reintroduced as a number of users
found it useful to call the function from {ggstatsplot}
itself rather than using pairwiseComparisons package.ggbetweenstats and ggwithinstats use
[ instead of ( to display confidence
intervals. Additionally, \[\mu\]
denoted sample mean, but was confused with population mean by some
users. So these functions instead display \[\hat{\mu}\].
More models supported in ggcoefstats:
bmlm, coeftest
Adapts to the new syntax provided in paletteer
package.
ggrepel in ggcoefstats function have been
removed. The users can instead provide all such arguments in a list to
stats.labels.args argument.ggbetweenstats and ggwithinstats no
longer produce incorrect label if the data frame already contains a
variable named n (#317) or variables with pattern
mean (#322).
ggbetweenstats and ggwithinstats mean
labels respect k argument (#331).
MINOR
ggcoefstats now uses
parameters::p_value instead of
sjstats::p_value, as requested by the maintainer of that
package. This might lead to differences in p-values for
lmer models.
More models supported in ggcoefstats:
blavaan, bracl, brglm2,
glmc, lavaan, nlreg,
slm, wbgee.
ggcoefstats gains only.significant
argument to only display display stats labels for significant effects.
This can be helpful when a large number of regression coefficients are
to be displayed in a single plot.
MINOR
magrittr, ellipsis,
purrrlyr.MAJOR
ggbetweenstats and
ggwithinstats pairwise comparisons were adjusted or not for
multiple comparisons.ANNOUNCEMENTS
{ggstatsplot} is undergoing conscious
uncoupling whereby all the statistical processing functions that
make stats subtitles are being moved to a new package called
{statsExpressions}. This new package will act as a backend
that handles all things statistical processing. This will
not affect the end users of {ggstatsplot} unless
you have been using the helper functions.
Additionally, multiple pairwise comparison tests are being moved to
an independent package called pairwiseComparisons.
This uncoupling is designed to achieve two things:
Make the code base of more manageable size in
{ggstatsplot}, which will make package development a bit
easier.
Make the workflow more customizable since now you can prepare
your own plots and then use {statsExpressions} to display
results in the plot rather than relying on {ggstatsplot}
default plots which are heavily opinionated and not appealing to
everyone.
All helper functions subtitle_* and
bf_* have been moved to the new
{statsExpressions} package.
To be consistent with all the other subtitle_ and
bf_ functions, subtitle_contingency_tab and
bf_contingency_tab now use the arguments x and
y instead of main and
condition.
Major refactoring to reduce the codesize and to rely fully on
rlang.
There was confusion about what did the red point in
ggbetweenstats and ggbetweenstats plots
represents. Now the label also contains \(\mu\) to highlight that what is being
displayed is a mean value.
To be consistent with the rest of the functions,
ggpiestats and ggbarstats now uses the
following aliases for arguments: x for main
and y for condition. This change is
backward-compatible and should not pose any problems for scripts that
used main and condition arguments in these
functions.
Most subtitle expressions now report details about the design. In case of between-subjects design, this will be \(n\_{obs}\), while in case of repeated measures design, this will be \(n\_{pairs}\).
pairwise.annotation now defaults to
"p.value" rather than "asterisk" for
ggbetweenstats and ggwithinstats (and their
grouped_ variants) functions. This was done because the
asterisk conventions are not consistent across various scientific
disciplines.
New dataset included: bugs_long, for repeated
measures designs with NAs present in the data.
{ggstatsplot} now uses rcompanion to
compute Spearman’s rho and Kendall’s W. Therefore,
DescTools is removed from dependencies.
ggcoefstats supports following objects:
bglmerMod, blmerMod, lme,
mclogit, mmclogit, tobit,
wblm.
ggcoefstats now respects conf.int. It
internally always defaulted to conf.int = TRUE in
broom::tidy irrespective of what was specified by the
user.
It was painfully confusing for a lot of users what exactly the
asterisks in each facet of ggpiestats signified. So instead
now ggpiestats displays more detailed results from a
goodness of fit (gof) test. No such change is made for
ggbarstats because there is no space to include more
details above the bar.
Removed conf.method and conf.type
arguments for ggcoefstats. Also, p.kr argument
removed because ggcoefstats will begin to rely on
parameters instead of sjstats package to
compute p-values for some regression models.
ggwithinstats caption, displayed by
default, was incorrect. This has been fixed. This stemmed from a line of
code which should have been paired = TRUE, but was instead
paired = FALSE.ggcoefstats defaults to
bf.message = TRUE to be consistent with the rest of the
functions in the package.
ggcoefstats supports the following class of objects:
epi.2by2, negbin, emmGrid,
lmrob, glmrob, glmmPQL,
data.table.
bf_ttest is introduced as a general function. The
previously exported bf_one_sample_ttest and
bf_two_sample_ttest become its aliases.
bf_meta_message syntax changes to adapt to updates
made to metaBMA package (thanks to #259).
The vestigial arguments axis.text.x.margin.t,
axis.text.x.margin.r, axis.text.x.margin.b,
axis.text.x.margin.l for ggcorrmat have been
removed. The margins can be adjusted using
ggplot2::margin().
gghistostats no longer allows data
argument to be NULL. This is to make this function’s syntax
consistent with rest of the functions in this package (none of which
allow data to be NULL). This also removes
confusion that arose for some users when data couldn’t be
NULL for its grouped_ cousin
(grouped_gghistostats).
outlier_df function is no longer exported since it
was always meant to be an internal function and was accidently exported
during initial release and was retained for a while for backward
compatibility.
Instead of having two separate functions that dealt with repeated
measures (subtitle_friedman_nonparametric) and
between-subjects (subtitle_kw_nonparametric), a single
function subtitle_anova_nonparametric handles both of these
designs with the paired argument determining which test is
run.
All functions that supported Bayes Factor analysis
(type = "bf") will only return BF value and the scale used.
Previously, this was a mix of parametric statistics and BF, which was
confusing and often times misleading since these two types of analyses
relied on different tests.
The default for bf.message has been changed from
FALSE to TRUE. This is to make the Bayes
Factor analysis more visible to the user.
ggscatterstats returns only plot (without any
statistical details) when the specified model is not linear (i.e.,
either when method argument is not "lm" or
when formula is not y ~ x).New functions ggwithinstats (and its
grouped_ variant) are introduced as a counterpart to
ggbetweenstats to handle repeated measures
designs.
For repeated measures ANOVA,
subtitle_anova_nonparametric now returns confidence
intervals for Kendall’s W.
All functions get return argument that can be used
to return either "plot", "subtitle", or
"caption". This makes it unnecessary to remember which
subtitle function is to be used where. As a result, in the next release,
all subtitle making functions will not be exported and are encouraged
not be used either by other developers or by users.
Both subtitle_anova_robust and
subtitle_anova_parametric gain a new argument
paired to support repeated measures designs.
ggcoefstats can support following new model objects:
drc, mlm.
ggcoefstats gains bf.message argument
to display a caption containing results from Bayesian random-effects
meta-analysis. It therefore gains a new dependency:
metaBMA.
ggpiestats and ggcatstats will now
display Cramer’s V as effect size for one-sample proportion
tests.
All functions gain stat.title argument
(NULL by default) that can be used to prefix the subtitle
with a string of interest. This is possibly useful for specifying the
details of the statistical test.
pairwise_p() function no longer outputs
conf.low and conf.high columns when parametric
post hoc tests are run. This is because these values were
accurate only when no p-value adjustment was carried
out.
Instead of using the internal function cor_test_ci,
ggscatterstats instead used SpearmanRho
function from DescTools package. This was done to reduce
number of custom internal functions used to compute CIs for various
effect sizes. {ggstatsplot} therefore gains
DescTools as a dependency.
The sampling.plan argument default for
ggbarstats function has been changed from
"indepMulti" to "jointMulti" to be consistent
with its sister function ggpiestats.
ggcoefstats can support following new model objects:
rjags.
New VR_dilemma dataset for toying around with
within-subjects design.
subtitle_t_onesample supports both Cohen’s
d and Hedge’s g as effect sizes and also produces
their confidence intervals. Additionally, non-central variants of these
effect sizes are also supported. Thus, gghistostats and its
grouped_ variant gets two new arguments:
effsize.type, effsize.noncentral.
ggpiestats used to display odds ratio as effect size
for paired designs (McNemar test). But this was only working when the
analysis was a 2 x 2 contingency table. It now instead displays Cohen’s
G as effect size, which generalizes to any kind of
design.
The internal function outlier_df to add a column
specifying outlier status of any given data point is now
exported.
{ggstatsplot} previously relied on an internal
function chisq_v_ci to compute confidence intervals for
Cramer’s V using bootstrapping but it was pretty slow. It now
instead relies on rcompanion package to compute confidence
intervals for V. {ggstatsplot}, therefore, gains a
new dependency.
subtitle_mann_nonparametric and
subtitle_t_onesample now computes effect size r
and its confidence intervals as \[Z/\sqrt{N}\] (with the help of
rcompanion package), instead of using Spearman
correlation.
subtitle_t_onesample no longer has data as
the optional argument. This was done to be consistent with other
subtitle helper functions.New function ggbarstats (and its
grouped_ variant) introduced for making bar charts (thanks
to #78).
ggcoefstats also displays a caption with model
summary when meta-analysis is required.
gghistostats and its grouped_ variant
has a new argument normal.curve to superpose a normal
distribution curve on top of the histogram (#138).
ggcoefstats can support following new regression
model objects: brmsfit, gam, Gam,
gamlss, mcmc, mjoint,
stanreg.
New function to convert plots which are not of
gg/ggplot class to ggplot class
objects.
Instead of using effsize to compute Cohen’s
d and Hedge’s g, {ggstatsplot} now relies
on a new (#159) internal function effect_t_parametric to
compute them. This removes effsize from
dependencies.
To be consistent with other functions in this package, both
ggbarstats and ggpiestats gain
results.subtitle which can be set to FALSE if
statistical analysis is not required, in which case
subtitle argument can be used to provide alternative
subtitle.
ggbetweenstats now defaults to using
noncentral-t distribution for computing Cohen’s d and
Hedge’s g. To get variants with central-t
distribution, use effsize.noncentral = FALSE.All grouped_ functions had argument
title.prefix that defaulted to "Group". It now
instead defaults to NULL, in which case the prefix will
variable name for grouping.var argument.
To accommodate non-parametric tests,
subtitle_template function can now work with
parameter = NULL.
For ggbetweenstats, details contained in the
subtitle for non-parametric test are modified. It now uses Spearman’s
rho-based effect size estimates. This removes coin
from dependencies.
ggbetweenstats and its grouped_ variant
gain a new argument axes.range.restrict (which defaults to
FALSE). This restricts y-axes limits to
minimum and maximum of y variable. This is what these
functions were doing by default in the past versions, which created
issues for additional ggplot components using the
ggplot.component argument.
All bayes factor related subtitle and captions replace
prior.width with r_{Cauchy}.
ggcoefstats passes dots (...) to
augment method from broom.
The helper function bf_extractor no longer provides
option to extract information about posterior distribution because these
details were incorrect. The posterior = TRUE details were
not used anywhere in the package so nothing about the results
changes.
ggcorrmat didn’t output pair names when
output == "ci" was used. This is fixed.
ggcoefstats gains meta.analytic.effect
that can be used to carry out meta-analysis on regression estimates.
This especially useful when a data frame with regression estimates and
standard error is available from prior analyses. The
subtitle is prepared with the new function
subtitle_meta_ggcoefstats which is also exported.
ggbetweenstats, ggscatterstats,
gghistostats, and ggdotplotstats (and their
grouped_ variants) all gain a new
ggplot.component argument. This argument will primarily be
helpful to change the individual plots in a grouped_
plot.
ggcoefstats can support following new regression
model objects: polr, survreg,
cch, Arima, biglm,
glmmTMB, coxph, ridgelm,
aareg, plm, nlrq,
ivreg, ergm, btergm,
garch, gmm, lmodel2,
svyolr, confusionMatrix,
multinom, nlmerMod, svyglm,
MCMCglmm, lm.beta, speedlm,
fitdistr, mle2, orcutt,
glmmadmb.
ggcoefstats didn’t work when statistic
argument was set to NULL. This was not expected behavior.
This has been fixed. Now, if statistic is not specified,
only the dot-and-whiskers will be shown without any labels.
subtitle_t_parametric was producing incorrect sample
size information when paired = TRUE and the data contained
NAs. This has been fixed.
ggscatterstats and its grouped_ variant
accept both character and bare exressions as input to arguments
label.var and labe.expression (#110).
To be consistent with rest of the functions in the package, both Pearson’s r, Spearman’s rho, and robust percentage bend correlations also display information about statistic associated with these tests.
ggscatterstats, by default, showed jittered data
points (because it relied on position_jitter defaults).
This could be visually inaccurate and, therefore,
ggscatterstats now displays points without any jitter. The
user can introduce jitter if they wish to using
point.width.jitter and point.height.jitter
arguments. For similar reasons, for ggbetweenstats and its
grouped_ variant, point.jitter.height default
has been changed from 0.1 to 0 (no vertical
jitter, i.e.).
Confidence interval for Kendall’s W is now computed
using stats::kruskal.test. As a result,
PMCMRplus removed from dependencies.
ggcoefstats gains a caption argument.
If caption.summary is set to TRUE, the
specified caption will be added on top of the
caption.summary.
ggcoefstats was showing wrong confidence intervals
for merMod class objects due to a bug in the
broom.mixed package (https://github.com/bbolker/broom.mixed/issues/30#issuecomment-428385005).
This was fixed in broom.mixed and so
ggcoefstats should no longer have any issues.
specify_decimal_p has been modified because it
produced incorrect results when k < 3 and
p.value = TRUE (e.g., 0.002 was printed as
< 0.001).
ggpiestats produced incorrect results if some levels
of the factor had been filtered out prior to using this function. It now
drops unused levels and produces correct results.
gghistostats wasn’t filtering out NAs
properly. This has been fixed.
New function ggdotplotstats for creating a dot
plot/chart for labelled numeric data.
All primary functions gain conf.level argument to
control confidence level for effect size measures.
As per APA guidelines, all results show results with two decimal
places. That is, the default value for k argument for all
functions has been changed from 3 to
2.
All helper functions for the ggbetweenstats
subtitles have been renamed to remove _ggbetween_ from
their names as this was becoming confusing for the user. Some of these
functions work both with the between- and within-subjects designs, so
having _ggbetween_ in their names made users suspect if
they could use these functions for within-subjects designs.
{ggstatsplot} now depends on R 3.5.0.
This is because some of its dependencies require 3.5.0 to work (e.g.,
broom.mixed).
All theme_ functions are now exported
(theme_pie(), theme_corrmat()).
ggbetweenstats now supports multiple pairwise
comparison tests (parametric, nonparametric, and robust variants). It
gains a new dependency ggsignif.
ggbetweenstats now supports eta-squared and
omega-squared effect sizes for anova models. This function gains a new
argument partial.
Following functions are now reexported from the
groupedstats package to avoid repeating the same code in
two packages: specify_decimal_p,
signif_column, lm_effsize_ci, and
set_cwd. Therefore, groupedstats is now added
as a dependency.
gghistostats can now show both counts and
proportions information on the same plot when bar.measure
argument is set to "mix".
ggcoefstats works with tidy dataframes.
The helper function untable has been deprecated in
light of tidyr::uncount, which does exactly what
untable was doing. The author wasn’t aware of this function
when untable was written.
All vignettes have been removed from CRAN to reduce
the size of the package. They are now available on the package website:
https://indrajeetpatil.github.io/ggstatsplot/articles/.
subtitle_t_robust function can now handle dependent
samples and gains paired argument.
A number of tidyverse operators are now reexported by
{ggstatsplot}: %>%,
%<>%, %$%.
ggscatterstats, ggpiestats, and their
grouped_ variant support bayes factor tests and gain new
arguments relevant to this test.
Effect size and their confidence intervals now available for Kruskal-Wallis test.
Minor stylistic changes to how symbols for partial-eta-/omega-squared were being displayed in subtitles.
ggbetweenstats supports bayes factor tests for anova
designs.
ggpiestats (and its grouped_ version)
gain slice.label argument that decides what information
needs to be displayed as a label on the slices of the pie chart:
"percentage" (which has been the default thus far),
"counts", or "both".
ggcorrmat can work with
cor.vars = NULL. In such case, all numeric
variables from the provided data frame will be used for computing the
correlation matrix.
Given the constant changes to the default behavior of functions,
the lifecycle badge has been changed from stable to
maturing.
When the number of colors needed by a function exceeds the number
of colors contained in a given palette, informative message is displayed
to the user (with the new internal function
palette_message()).
Several users had requested an easier way to turn off subtitles
with results from tests (which was already implemented in
ggscatterstats and gghistostats with the
argument results.subtitle), so ggbetweenstats
also gains two new arguments to do this: results.subtitle
and subtitle.
New dataset added: iris_long.
More tests added and the code coverage has now jumped to over 75%.
To avoid code repetition, there is a now a function that produces a generic message any time confidence intervals for effect size estimate are computed using bootstrapping.
The package now exports all functions used to create text
expressions with results. This makes it easy for people to use these
results in their own plots at any location they want (and not just in
subtitle, the current default for
{ggstatsplot}).
ggcorrmat gains p.adjust.method
argument which allows p-values for correlations to be corrected
for multiple comparisons.
ggscatterstats gains label.var and
label.expression arguments to attach labels to
points.
gghistostats now defaults to not showing (redundant)
color gradient (fill.gradient = FALSE) and shows both
"count" and "proportion" data. It also gains a
new argument bar.fill that can be used to fill bars with a
uniform color.
ggbetweenstats, ggcoefstats,
ggcorrmat, ggscatterstats, and
ggpiestats now support all palettes contained in the
paletteer package. This helps avoid situations where people
had large number of groups (> 12) and there were not enough colors in
any of the RColorBrewer palettes.
ggbetweenstats gains bf.message
argument to display bayes factors in favor of the null (currently works
only for parametric t-test).
gghistostats function no longer has
line.labeller.y argument; this position is automatically
determined now.
legend.title.margin function has been deprecated
since ggplot2 3.0.0 has improved on the margin issues from
previous versions. All functions that wrapped around this function now
lose the relevant arguments (legend.title.margin,
t.margin, b.margin).
The argument ggstatsplot.theme has been changed to
ggstatsplot.layer for ggcorrmat function to be
consistent across functions.
For consistency, conf.level and
conf.type arguments for ggbetweenstats have
been deprecated. No other function in the package allowed changing
confidence interval or their type for effect size estimation. These
arguments were relevant only for robust tests
anyway.
ggocorrmat argument type has been
changed to matrix.type because for all other functions
type argument specifies the type of the test, while for
this function it specified the display of the visualization matrix. This
will make the syntax more consistent across functions.
ggscatterstats gains new arguments to specify
aesthetics for geom point (point.color,
point.size, point.alpha). To be consistent
with this naming schema, the width.jitter and
height.jitter arguments have been renamed to
point.width.jitter and point.height.jitter,
resp.
gghistostats: To be compatible with
JASP, natural logarithm of Bayes Factors is displayed, and
not base 10 logarithm.
ggscatterstats gains method and
formula arguments to modify smoothing functions.
ggcorrmat can now show robust
correlation coefficients in the matrix plot.
For gghistostats, binwidth value, if
not specified, is computed with (max-min)/sqrt(n). This is
basically to get rid of the warnings ggplot2 produces. Thanks to Chuck
Powell’s PR (#43).
ggcoefstats gains a new argument
partial and can display eta-squared and omega-squared
effect sizes for anovas, in addition to the prior partial variants of
these effect sizes.
ggpiestats gains digits.perc argument
to show desired number of decimal places in percentage labels.
grouped_ggpiestats wasn’t working when only
main variable was provided with counts data.
Fixed that.For the sake of consistency, theme_mprl is now
called theme_ggstatsplot. The theme_mprl
function will still be around and will not be
deprecated, so feel free to use either or both of them since they are
identical.
ggcoefstats no longer has arguments
effects and ran_params because only fixed
effects are shown for mixed-effects models.
ggpiestats can now handle within-subjects designs
(McNemar test results will be displayed).
ggbetweenstats was producing wrong axes labels when
sample.size.label was set to TRUE and user had
reordered factor levels before using this function. The new version
fixes this.
ggcoefstats wasn’t producing partial omega-squared
for aovlist objects. Fixed that with new version of
sjstats.
Removed the trailing comma from the robust correlation analyses.
gghistostats has a new argument to remove color fill
gradient.
ggbetweenstats takes new argument
mean.ci to show confidence intervals for the mean
values.
For lmer models, p-values are now computed
using sjstats::p_value. This removes lmerTest
package from dependencies.
sjstats no longer suggests apaTables
package to compute confidence intervals for partial eta- and
omega-squared. Therefore, apaTables and MBESS
are removed from dependencies.
ggscatterstats supports densigram with
the development version of ggExtra. It additionally gains
few extra arguments to change aesthetics of marginals (alpha, size,
etc.).
New function: ggcoefstats for displaying model
coefficients.
All functions now have ggtheme argument that can be
used to change the default theme, which has now been changed from
theme_grey() to theme_bw().
The robust correlation is no longer MASS::rlm, but
percentage bend correlation, as implemented in WRS2::pbcor.
This was done to be consistent across different functions.
ggcorrmat also uses percentage bend correlation as the
robust correlation measure. This also means that
{ggstatsplot} no longer imports MASS and
sfsmisc.
The data argument is no longer NULL for
all functions, except gghistostats. In other words, the
user must provide a data frame from which variables or
formulas should be selected.
All subtitles containing results now also show sample size
information (n). To adjust for the inflated length of the
subtitle, the default subtitle text size has been changed from
12 to 11.
Switched back to Shapiro-Wilk test of normality to remove
nortest from imports.
ggbetweenstats and ggpiestats now
display sample sizes for each level of the groping factor by default.
This behavior can be turned off by setting
sample.size.label to FALSE.
Three new datasets added: Titanic_full,
movies_wide, movies_long.
Added confidence interval for effect size for robust ANOVA.
The 95% CI for Cramer’V computed using boot::boot.
Therefore, the package no longer imports
DescTools.
To be consistent across correlations covered, all correlations now show estimates for correlation coefficients, confidence intervals for the estimate, and p-values. Therefore, t-values and regression coefficients are no longer displayed for Pearson’s r.
The legend.title.margin arguments for
gghistostats and ggcorrmat now default to
FALSE, since ggplot2 3.0.0 has better legend
title margins.
ggpiestats now sorts the summary dataframes not by
percentages but by the levels of main variable. This was
done to have the same legends across different levels of a grouping
variable in grouped_ggpiestats.
To remove cluttered display of results in the subtitle,
ggpiestats no longer shows titles for the tests run (these
were “Proportion test” and “Chi-Square test”). From the pie charts, it
should be obvious to the user or reader what test was run.
gghistostats also allows running robust version of
one-sample test now (One-sample percentile bootstrap).
The ggbetweenstats function can now show notched box
plots. Two new arguments notch and notchwidth
control its behavior. The defaults are still standard box
plots.
Removed warnings that were appearing when
outlier.label argument was of character
type.
The default color palette used for all plots is colorblind friendly.
gghistostats supports proportion and
density as a value measure for bar heights to show
proportions and density. New argument bar.measure controls
this behavior.
grouped_ variants of functions
ggcorrmat, ggscatterstats,
ggbetweenstats, and ggpiestats introduced to
create multiple plots for different levels of a grouping
variable.
To be internally consistent, all functions in
{ggstatsplot} use the spelling color, rather
than colour in some functions, while color in
others.
Removed the redundant argument binwidth.adjust from
gghistostats function. This argument was relevant for the
first avatar of this function, but is no longer playing any
role.
To be internally consistent, the argument lab_col
and lab_size in ggcorrmat have been changed to
lab.col and lab.size, respectively.
Added a new argument to ggstatsplot.theme function
to control if ggstatsplot::theme_mprl is to be overlaid on
top of the selected ggtheme (ggplot2 theme, i.e.).
Two new arguments added to gghistostats to allow
user to change colorbar gradient. Defaults are colorblind
friendly.
Both gghistostats and ggcorrmat have a
new argument legend.title.margin to control margin
adjustment between the title and the colorbar.
The vertical lines denoting test values and centrality parameters
can be tagged with text labels with a new argument
line.labeller in gghistostats
function.
centrality.para argument for
ggscatterstats was not working properly. Choosing
"median" didn’t show median, but the mean. This is fixed
now.Bayesian test added to gghistostats and two new
arguments to also display a vertical line for test.value
argument.
Vignette added for gghistostats.
Added new function grouped_gghistostats to
facilitate applying gghistostats for multiple levels of a
grouping factor.
ggbetweenstats has a new argument
outlier.coef to adjust threshold used to detect outliers.
Removed bug from the same function when outlier.label
argument is of factor/character type.
Functions signif_column and
grouped_proptest are now deprecated. They were exported in
the first release by mistake.
Function gghistostats no longer displays both
density and count since the density information was redundant. The
density.plot argument has also been deprecated.
ggscatterstats argument intercept has
now been changed to centrality.para. This was due to
possible confusion about interpretation of these lines; they show
central tendency measures and not intercept for the linear model. Thus
the change.
The default for effsize.type = "biased" effect size
for ggbetweenstats in case of ANOVA is
partial omega-squared, and not omega-squared.
Additionally, both partial eta- and omega-squared are not computed using
bootstrapping with (default) 100 bootstrap samples.
More examples added to the README document.
95% confidence intervals for Spearman’s rho are now computed
using broom package. RVAideMemoire package is
thus removed from dependencies.
95% confidence intervals for partial eta- and omega-squared for
ggbetweenstats function are now computed using
sjstats package, which allows bootstrapping.
apaTables and userfriendlyscience packages are
thus removed from dependencies.