User Guide: 1 Debugging ggplots"

library(knitr)
opts_chunk$set(fig.align = 'center', 
               fig.show = 'hold', fig.width = 7, fig.height = 4)
options(warnPartialMatchArgs = FALSE)

Preliminaries

library(gginnards)
library(tibble)

We generate some artificial data.

set.seed(4321)
# generate artificial data
x <- 1:100
y <- (x + x^2 + x^3) + rnorm(length(x), mean = 0, sd = mean(x^3) / 4)
my.data <- data.frame(x, 
                      y, 
                      group = c("A", "B"), 
                      y2 = y * c(0.5, 2),
                      block = c("a", "a", "b", "b"))

We change the default theme to an uncluttered one.

old_theme <- theme_set(theme_bw())

ggplot construction

Package 'ggplot2' defines its own class system, and function ggplot() can be considered as a constructor.

class(ggplot())

These objects contain all the information needed to render a plot into graphical output, but not the rendered plot itself. They are list-like objects with heterogeneous named members.

The structure of objects of class "ggplot" can be explored with R's method str() as is the case for any structured R object. Package 'gginnards' defines a a specialization of str() for class "ggplot". Our str() allows us to see the different slots of these special type of lists. The difference with the default str() method is in the values of default arguments, and in the ability to control which components or members are displayed.

We will use the str() to follow the step by step construction of a "ggplot" object.

If we pass no arguments to the ggplot() constructor an empty plot will be rendered if we print it.

p0 <- ggplot()
p0

Object p contains members, but "data", "layers", "mapping", "theme" and "labels" are empty lists.

str(p0)

If we pass an argument to parameter data the data is copied into the list slot with name data. As we also map the data to aesthetics, this mapping is stored in slot maaping.

p1 <- ggplot(data = my.data, aes(x, y, colour = group))
str(p1)
str(p1, max.level = 2, components = "data")

A geometry adds a layer.

p2 <- p1 + geom_point()
str(p2)

A summary() method that produces a more compact output is available in recent versions of 'ggplot2'. However, it does not reveal the internal structure of the objects.

summary(p2)
str(p2, max.level = 2, components = "mapping")
p3 <- p2 + theme_classic()
str(p3)

Themes are stored as nested lists. To keep the output short we use max.level = 2 although using max.level = 3 would be needed to see all nested members.

str(p3, max.level = 2, components = "theme")

Data mappings in ggplots

How does mapping work? Geometries (geoms) and statistics (stats) do not "see" the original variable names, instead the data passed to them is named according to the aesthetics user variables are mapped to. Geoms and stats work in tandem, with geoms doing the actual plotting and stats summarizing or transforming the data. It can be instructive to be able to see what data is received as input by a geom or stat, and what data is returned by a stat.

Both geoms and stats can have either panel- or group functions. Panel functions receive as input the subset of the data that corresponds to a whole panel, mapped to the aesthetics and with factors indicating the grouping (set by the user by mapping to a discrete scale). Group functions receive as input the subset of data corresponding to a single group based on the mapping, and called once for each group present in a panel.

The motivation for writing the "debug" stats and geoms included in package 'gginnards' is that at the moment it is in many cases not possible to set breakpoints inside the code of stats and geoms, because frequently nameless panel and group functions are stored within list-like "ggplot" objects as seen above.

This can make it tedious to analyse how these functions work, as one may need to add print statements to their definitions to see the data. I wrote the "debug" stats and geoms as tools to help in the development of my packages 'ggpmisc' and 'ggspectra', and as a way of learning myself how data are passed around within the different components of a ggplot object when it is printed.

Data input to geometries

Data pass through a statistics before being received by a geometry. However, many geometries, like geom_point() and geom_line() use by default stat_identity() which simply relays the unmodified data to the geometries.

The debug geometries and statistics in package 'gginnards' by default do not add any graphical element to the plot but instead they make visible the data as received as their input.

The geometry geom_debug_panel() uses stat_identity() by default. Here the same data as rendered by geom_point() is printed as a tibble to the R console. We can see that the columns are named according to the aesthetics the variables in the user-supplied data have been mapped. In the case of colour, the levels of the factor have been replaced by colour definitions. Columns PANEL and group have been also added.

ggplot(mpg, aes(cyl, hwy, colour = factor(cyl))) + 
  geom_point() +
  geom_debug_panel()

Below we show how geom_debug_panel() can be used together with functions that take a data frame as input and return a value that can be printed. We use here head() but other functions such summary(), nrow() and colnames() as well as user defined functions can be useful when data is large. As shown here, additional arguments can be passed by name to the function.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  geom_debug_panel(dbgfun.data = head, dbgfun.data.args = list(n = 3))

When using a statistic that modifies the data, we can pass geom_debug_panel() as argument in the call to this statistic. In this way the data printed to the console will be those returned by the statistics and received by the geometry.

ggplot(mpg, aes(cyl, hwy)) +
  stat_summary(fun.data = "mean_se") +
  stat_summary(fun.data = "mean_se", geom = "debug_panel") 

As shown above an important use of geom_debug_panel() it to display the data returned by a statistic and received as input by geometries. Not all extensions to 'ggplot2' document all the computed variables returned by statistics. In other cases like in the next example, the values returned will depend on the arguments passed. While in the previous example the statistic returned a data frame with one row per group, here the returned data frame has 160 rows. The data are by default plotted as a line with a confidence band.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  stat_smooth(method = "lm", formula = y ~ poly(x, 2)) +
  stat_smooth(method = "lm", formula = y ~ poly(x, 2), 
              geom = "debug_panel", dbgfun.data = head)

Data input to statistics

Statistics can be defined to operate on data corresponding to a whole panel or separately on data corresponding to each individual group, as created by mapping aesthetics to factors. The statistics described below print a summary of their data input by default to the console. These statistics, in addition return a data frame containing summary information including labels suitable for "plotting" with geom = "text" or geom = "label". However, package 'gginnards' defines a "null" geom, geom_null(), which is used as default by the debug statistics. This geom is similar to the more recently added ggplot2::geom_blank().

ggplot(my.data, aes(x, y)) + 
  geom_null()

Using geom "null" allows to add the debug stats for the side effect of console output without altering the rendering of the plot when there is at least one other plot layer. The default geom "null" does not alter the rendering of the plot or print to the console the data output by the debug stats.

Because of the way 'ggplot2' works, the values are listed to the console at the time when the ggplot object is printed. As shown here, no other geom or stat is required, however in the remaining examples we add geom_point() to make the data also visible in the plot.

ggplot(my.data, aes(x, y)) + 
  stat_debug_group()

In the absence of facets or groups we get the printout of a single data frame, which is similar to that returned by geom_debug_panel(). Without grouping, group is set to -1 for all observations. As the we override the default geom with geom_debug_panel() a summary computed by the stat is also printed to the console.

ggplot(my.data, aes(x, y)) + 
  geom_point() + 
  stat_debug_group(geom = "debug_panel")

In a plot with no grouping, there is no difference in the data input for compute_panel() and compute_group() functions except for the order of the variables or columns in the data frame (this applies in general to ggplot statistics).

ggplot(my.data, aes(x, y)) + 
  geom_point() + 
  stat_debug_panel()

By mapping the colour aesthetic we create a grouping. In the case, compute_group() is called with the data subset by group, and a separate data frame is displayed for each call compute_group(), corresponding each to a level in the mapped factor. In this case group takes as values positive consecutive integers. As a factor was mapped to colour, colour is encoded as a factor.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  stat_debug_group()

Without facets, we still have only one panel.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  stat_debug_panel()

When we map the same factor to a different aesthetic the data remain similar, except for the column named after the aesthetic, in this case shape.

ggplot(my.data, aes(x, y, shape = group)) + 
  geom_point() + 
  stat_debug_group()

Facets based on factors create panels within a plot. Here we create a plot with both facets and grouping. In this case, for each panel the compute_panel() function is called once with a subset of the data that corresponds to one panel, but not split by groups. For our example, it is called twice.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  stat_debug_panel(dbgfun.data = "nrow") +
  facet_wrap(~block)

with grouping and facets, within each panel the compute_group() function is called for each group, in total four times.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  stat_debug_group(dbgfun.data = "nrow") +
  facet_wrap(~block)

Controlling the debug output

In the examples above we have demonstrated the use of the statistics and geometries using default arguments. Here we show examples of generation of other types of debug output.

stat_debug_group() and stat_debug_panel() return summary data that can be inspected using a geometry in addition to printing the data received as argument. If we use geom_debug_panel() a summary is printed to the console. With two groups, we get two summaries when we use stat_debug_group().

ggplot(my.data, aes(x, y, shape = group)) + 
  geom_point() + 
  stat_debug_group(geom = "debug_panel")

If we use stat_debug_panel() we get a single summary.

ggplot(my.data, aes(x, y, shape = group)) + 
  geom_point() + 
  stat_debug_panel(geom = "debug_panel")

In principle one can use other geoms to annotate the plot with the debug summary. In this case we silence all output to the R console and use the stat as any other ggplot stat.

ggplot(my.data, aes(x, y, colour = group)) + 
  geom_point() + 
  stat_debug_group(geom = "text",
                   mapping = aes(label = sprintf("group = %i", 
                                                 after_stat(group))),
                   dbgfun.data = function(x) {NULL})


Try the gginnards package in your browser

Any scripts or data that you put into this service are public.

gginnards documentation built on May 29, 2024, 2:08 a.m.