Packages ‘ggpmisc’, ‘ggpp’ and ‘gginnards’

Data labels, annotations and insets for ‘ggplot2’

R packages

Pedro J. Aphalo






R, ggplot2, labels, annotations, regression, anova, correlation, fitted models, data labels, plot annotations, plot insets

How do these packages extend ‘ggplot2’?

‘ggpp’ extends ‘ggplot2’ and the grammar of graphics to more consistently and powerfully handle data labels, annotations and insets. New geometries extend the grammar so that whole plots, tables and graphical objects (‘grid’ grobs) can be used as data labels using an almost identical syntax as used for text labels in ‘ggplot2’. Another group of new geoms and scales, implements an enhanced grammar for annotations, by adding two pseudo aesthetics using normalized plot coordinated (NPC). Except for the use of npcx and npcy instead of x and y the grammar remains unchanged, allowing full support of grouping and facets for annotations. Position functions as implemented in ‘ggplot2’ do not preserve the original position, so it is difficult to draw connecting arrows and segments. This problem was first solved in ‘ggrepel’ for position_nudge() and the segment drawing for repulsed labels and text. ‘ggpp’ implements the “keeping” of the original position with new position functions matching all ‘ggplot2’ position functions. This did not fully solve the problem, as the positioning of data labels was constrained by the fact that ggplot2 position functions can not be combined. So, ‘ggpp’ defines combined position functions that implement the usual displacements like stacking plus nudging. Another recent addition to ‘ggpp’ is support for new types of nudging, including computed nudging based on the local data 1D or 2D density, based on fitted lines, or away or towards a computed centroid or arbitrary point or line. When nudging is applied both along x and y, even radial nudging is supported. Thanks to a fruitful collaboration with Kamil Slowikowski, the author of ‘ggrepel’, these new approaches to nudging are compatible with and extremely effective when combined with repulsive geoms. A few convenience and utility functions are also included. Perhaps surprisingly, given the good design of ‘ggplot2’ and its support for extensions, all these features were implemented without any overwriting of ‘ggplot2’ code except for a wrapper on annotate() to add support for NPC.

Package ‘ggpmisc’ makes use of ‘ggpp’ to add specific annotations and insets to plots. ‘ggpmisc’ mainly defines stats, that help annotate the plots based on the results of model fitting, although it also provides stats for adding fitted and predicted curves and for highlighting and or plotting residuals. These stats either complement or enhance stat_smooth() and stat_quantile() from ‘ggplot2’. Annotations supported included fitted equations and other estimates like \(R^2\) and P-values for continuous x and y and ANOVA tables when x or y is a factor. Additional stats make it possible to automatically annotate whole plots or quadrants in plots with the number of observations, and to locate peaks or valleys, and label them with their x and/or y coordinates. A few convenience and utility functions are also included. I have tracked changes to ‘ggplot2’, and ‘ggpp’ stats implement the orientation formal parameter if meaningful. (‘ggpmisc’ does not support annotations based on multiple comparisons, including pairwise comparisons, as these seem to be effectively by other pakages.)

gginnards is mainly useful for debugging and learning about ‘ggplot2’. It does implement the manipulation of ggplot layers (insertion, deletion and moving up or down) which can be useful not only for learning, but also for tweaking some ggplot objects returned by “canned” functions.

What is their history?

It all started in 2016 from an innocent question from my colleague, Titta Kotilainen, that went something like this: “I see in Stackoverflow some answers to the question of how to add a regression line equation to a ggplot, but they are so complex… Isn’t there any simpler way of doing this?

I looked at the answers and they were not only not straightforward to code, but were case specific. So after some thinking and “googling”, a primitive version of stat_poly_eq() was born. Lacking a good idea of what the package would develop into, following the trend set by ‘Gmisc’ and a few other packages I decided to use ‘ggpmisc’ (ggplot miscellanea).

Over the seven years since then ‘ggpmisc’ grew both because of my own needs and thanks from suggestions and questions from users. Rather soon it became clear that ‘ggpmisc’ needed to be split into more homogeneous “units”. The first spin-off was ‘gginnards’ in June 2018, which contains mostly functions I wrote to help myself maintain my extensions to ‘ggplot2’ and help me understand how ‘ggplot2’ works.

The second spin-off took place in 2021. The reason was to make the geometries and some other functions available on their own so that they could be more easily depended upon by other packages. Because of this history, ‘ggpmisc’ loads and attaches ‘ggpp’ when it is loaded and attached. So, the aim was akin to providing a subset of ‘ggpmisc’ to some users while keeping the behaviour of ‘ggpmisc’ unchanged.

In early 2023 I checked CRAN landing pages for the packages: ‘ggpmisc’ has 10 reverse dependencies and two reverse suggests, ‘ggpp’ has five reverse dependencies and one reverse suggest, and to my surprise, even ‘gginnards’ has two reverse dependencies.

The current chapter in this history relates to quality control, unit-test coverage and possible accreditation/certification of ‘ggpp’. By the initiative of Daniel Sabanes Bove and his team, and with a lot of help from them, progress is being made towards achieving high enough test coverage for accreditation. I am trying in parallel to enhance testing for validity of alues passed as arguments and improving the corresponding error messages. I am also checking the documentation and revamping some of the contents at this web site.

Naturally, I intend next to apply what I am learning to ‘ggpmisc’ and later to other packages.

What does their design aim at?

The underlying aim behind the design of ‘ggpmisc’ has been to make it easy to add data labels, annotations and insets to ggplots, using a grammar consistent with that implemented in ‘ggplot2’ and without imposing arbitrary restrictions on the use of the layered grammar of graphics.

How I approached and still approach this aim, is by trying to imaging how to remain conceptually consistent with the existing grammar. In other words, finding ways of reusing as much as possible the existing grammar to solve new problems. For example, statistics that return character labels from model fits, also return the corresponding numerical values. New functions for adding graphical elements as data labels, are consistent with ‘ggplot2’ stats used to add text-based data labels to plots.

The approach I use in ‘ggpmisc’ is different to that of popular extensions like ‘ggpubr’ which attempt to simplify plot creation by packaging the code for several plot layers into a single function. Using such an approach, much of the grammar of graphics is not accessible to users, and even if easier to use, such functions are much less flexible. Clearly, the two approaches target somehow different audiences.

How is the code tested?

The release of package ‘testthat’ made testing R code producing numerical or textual output rather easy, but testing graphical output remained very difficult until vdiffr was released. Developing and maintaining ‘ggpmisc’ and publishing it through CRAN would not have been manageable without using unit tests implemented with these two packages ‘testthat’ and ‘vdiffr’.

Unit tests for ‘ggplot2’ extensions had been quite tricky to implement, causing in the past trouble for CRAN and breaking frequently due to inconsequential changes in ‘ggplot2’ or its dependencies. For this reason, even though I implemented the first unit tests for ‘ggpmisc’ in 2017 and kept adding more since then, I have kept these tests local and not included them in the package releases. As I have been until recently the only developer and maintainer of the package, this approach was good anough and managed to keep the package nearly bug free and made development of enhancements and fixing bugs comparatively easy. However, local testing became insufficient once other developers started contributing pull-requests to ‘ggpp’.

I set up continuous integration actions for running CRAN checks in GitHub rather recently, as I was previously using CRAN winbuilder and ‘rhub’ to run tests on demand. However, ‘testthat’ unit tests were only run locally in my computer as they had not been included in the package build. This has now changed for ‘ggpp’ and ‘ggpmisc’, and from the next version tests will be included in the builds. As for reporting of test coverage its implementation is already working in ‘ggpp’ at GitHub thanks to a pull request from Danikar.

My goal is to follow the recommendations of ROpenScience and once requirements are met, submit ‘ggpp’ and ‘ggpmisc’ to their peer review.

More information

The documentation, as websites, including the output from examples and all vignettes is available for ‘ggpp’, ‘ggpmisc’ and ‘gginnards’.

At this web site there are also galleries of plot examples with the corresponding R code, organized by type of plot or plot features.