Skip to contents

Visualizing data

Source: vignettes/articles/Visualizing-data.Rmd
Visualizing-data.Rmd

In this article, we will explore the visualization of different types of datasets using tidyplots. We will cover the plotting of raw data points, amounts, and heatmaps, as well as measures for central tendency, dispersion, and uncertainty. We will conclude by visualizing distributions and proportions, and adding statistical comparisons and annotations.

Data points

Plotting the raw data points is probably the most bare bone way to visualize a dataset. The corresponding function in tidyplots is called add_data_points().

library(tidyplots)

animals %>% 
  tidyplot(x = weight, y = size) %>% 
  add_data_points()

In the above example some data points appear to overlap other points. To account for this so called overplotting, you might want to add a thin white border around the points. This is achieved by setting the argument white_border = TRUE.

animals %>% 
  tidyplot(x = weight, y = size) %>% 
  add_data_points(white_border = TRUE)

Another way is to make the points transparent using the alpha argument.

animals %>% 
  tidyplot(x = weight, y = size) %>% 
  add_data_points(alpha = 0.4)

Or to change the plotting symbol to an open shape.

animals %>% 
  tidyplot(x = weight, y = size) %>% 
  add_data_points(shape = 1)

However, data points can also be used to plot a discrete variable against a continuous variable.

study %>% 
  tidyplot(x = treatment, y = score) %>% 
  add_data_points()

To avoid overplotting in this scenario, there are two additional options. You can add some random noise to the y position, also known as jitter.

study %>% 
  tidyplot(x = treatment, y = score) %>% 
  add_data_points_jitter()

Alternatively, you can use an algorithm that keeps the points centered and just moves potentially overlapping points to the sides.

study %>% 
  tidyplot(x = treatment, y = score) %>% 
  add_data_points_beeswarm()

Amounts

For some datasets, it makes sense to count or sum up data points in order to arrive to conclusions. As one example, let’s have a look at the spendings dataset.

spendings
#> # A tibble: 19 × 4
#>    date       title            amount category      
#>    <date>     <chr>             <dbl> <chr>         
#>  1 2023-10-01 Groceries           100 Food          
#>  2 2023-10-01 Gasoline             40 Transportation
#>  3 2023-10-01 Rent               1200 Housing       
#>  4 2023-10-02 Electricity          80 Utilities     
#>  5 2023-10-03 School Supplies      75 Education     
#>  6 2023-10-03 Health Insurance    200 Insurance     
#>  7 2023-10-04 Dining Out           60 Food          
#>  8 2023-10-04 Cell Phone Bill      50 Utilities     
#>  9 2023-10-05 Groceries            90 Food          
#> 10 2023-10-06 Gasoline             40 Transportation
#> 11 2023-10-07 Medical Checkup     150 Health        
#> 12 2023-10-07 Dining Out           70 Food          
#> 13 2023-10-08 Groceries           110 Food          
#> 14 2023-10-08 Internet Bill        60 Utilities     
#> 15 2023-10-09 Entertainment        30 Entertainment 
#> 16 2023-10-10 Groceries            50 Food          
#> 17 2023-10-12 Public Transport     70 Transportation
#> 18 2023-10-13 Dentist              90 Health        
#> 19 2023-10-15 Car Insurance        40 Insurance

As you can see, this dataset contains family spendings over a time period of 15 days in October. Here, it might be informative to see which spending categories are reoccurring and which are just one time spendings.

spendings %>% 
  tidyplot(x = category) %>% 
  add_count_bar()

One thing to note here is that the x-axis labels are overlapping and are thus unreadable. There are at least two possible solutions for this. One is to swap the x and y-axis.

spendings %>% 
  tidyplot(y = category) %>% 
  add_count_bar()

The other one is to rotate the x-axis labels.

spendings %>% 
  tidyplot(x = category) %>% 
  add_count_bar() %>% 
  adjust_x_axis(rotate_labels = TRUE)

Now we can appreciate that this family had reoccurring spendings for Food but just one spending for Housing.

Next, we ask the question how much was spend on each of the categories by plotting the sum amount.

spendings %>% 
  tidyplot(x = category, y = amount, color = category) %>% 
  add_sum_bar() %>% 
  adjust_x_axis(rotate_labels = TRUE)

Note that we had to introduce the argument y = amount in the tidyplot() function to make it clear which variable should be summed up.

I also added color = category in the tidyplot() function to have the variable category encoded by different colors.

Since the labels for the variable category are now duplicated in the plot, one could argue that it would be justified to remove the duplicated information on the x-axis.

spendings %>% 
  tidyplot(x = category, y = amount, color = category) %>% 
  add_sum_bar() %>% 
  adjust_x_axis(rotate_labels = TRUE) %>% 
  remove_x_axis_labels() %>% 
  remove_x_axis_title() %>% 
  remove_x_axis_ticks()

Note that besides the x-axis labels, I also removed the x-axis ticks and x-axis title to achieve a cleaner look.

Of course you are free to play around with different graphical representations of the sum values. Here is an example of a lollipop plot constructed from a thin bar and a dot.

spendings %>% 
  tidyplot(x = category, y = amount, color = category) %>% 
  add_sum_bar(width = 0.03) %>% 
  add_sum_dot() %>% 
  add_sum_value(accuracy = 1) %>% 
  adjust_x_axis(rotate_labels = TRUE) %>% 
  remove_x_axis_labels() %>% 
  remove_x_axis_title() %>% 
  remove_x_axis_ticks()

I also added the sum value as text label using the add_sum_value() function.

Heatmaps

Heatmaps are a great way to plot a continuous variable across two additional variables. To exemplify this, we will have a look at the gene_expression dataset.

gene_expression %>% 
  dplyr::glimpse()
#> Rows: 800
#> Columns: 11
#> $ ensembl_gene_id    <chr> "ENSMUSG00000033576", "ENSMUSG00000033576", "ENSMUS…
#> $ external_gene_name <chr> "Apol6", "Apol6", "Apol6", "Apol6", "Apol6", "Apol6…
#> $ sample             <chr> "Hin_1", "Hin_2", "Hin_3", "Hin_4", "Hin_5", "Ein_1…
#> $ expression         <dbl> 2.203755, 2.203755, 2.660558, 2.649534, 3.442740, 5…
#> $ group              <chr> "Hin", "Hin", "Hin", "Hin", "Hin", "Ein", "Ein", "E…
#> $ sample_type        <chr> "input", "input", "input", "input", "input", "input…
#> $ condition          <chr> "healthy", "healthy", "healthy", "healthy", "health…
#> $ is_immune_gene     <chr> "no", "no", "no", "no", "no", "no", "no", "no", "no…
#> $ direction          <chr> "up", "up", "up", "up", "up", "up", "up", "up", "up…
#> $ log2_foldchange    <dbl> 9.395505, 9.395505, 9.395505, 9.395505, 9.395505, 9…
#> $ padj               <dbl> 3.793735e-28, 3.793735e-28, 3.793735e-28, 3.793735e…

We will start by plotting the expression values of each external_gene_name across the sample variable.

gene_expression %>% 
  tidyplot(x = sample, y = external_gene_name, color = expression) %>% 
  add_heatmap()

One thing to note here is that the y-axis labels are overlapping. So let’s increase the height of the plot area from 50 to 100 mm.

gene_expression %>% 
  tidyplot(x = sample, y = external_gene_name, color = expression) %>% 
  add_heatmap() %>%
  adjust_size(height = 100)

The next thing to note is that some of the rows like Map1a and Kif1a show very high values while other rows show much lower values. Let’s apply a classical technique to reserve the color range for differences within each row. This is done by calculating row z scores for each row individually. Luckily, tidyplots does this for us when setting the argument scale = "row" within the add_heatmap() function call.

gene_expression %>% 
  tidyplot(x = sample, y = external_gene_name, color = expression) %>% 
  add_heatmap(scale = "row") %>%
  adjust_size(height = 100)

Now it much easier to appreciate the dynamics of individual genes across the samples on the x-axis.

However, the rows appear to be mixed. Some having rather high expression in the “Eip” samples while others have high value in the “Hip” samples. Conveniently, there is a variable called direction in the dataset, which classifies genes as being either “up” or “down” regulated. Let’s use this variable to sort our y-axis.

gene_expression %>% 
  tidyplot(x = sample, y = external_gene_name, color = expression) %>% 
  add_heatmap(scale = "row") %>%
  adjust_size(height = 100) %>% 
  sort_y_axis_labels(direction)

Central tendency

In cases with multiple data points per experimental group, the mean and the median are a great way to compute a typical center value for the group, also known as central tendency measure. In tidyplots, these function start with add_mean_ or add_median_.

study %>% 
  tidyplot(x = treatment, y = score) %>% 
  add_data_points_beeswarm() %>% 
  add_mean_dash()

The second part of the function name is dedicated to the graphical representation. These include the representation as bar, dash, dot, value, line or area. Of course, these different representations can also be combined. Like in this case line and dot.

time_course %>% 
  tidyplot(x = day, y = score, color = treatment) %>% 
  add_mean_line() %>% 
  add_mean_dot()

Or in this case line and area.

time_course %>% 
  tidyplot(x = day, y = score, color = treatment) %>% 
  add_mean_line(linewidth = 1) %>% 
  add_mean_area(alpha = 0.2)

Here is one more example using bar and value.

study %>% 
  tidyplot(x = treatment, y = score) %>% 
  add_mean_bar() %>% 
  add_mean_value()

You could also plot the mean and the median together to explore in which cases they diverge. In the example below the mean is shown in orange and the median in purple.

study %>% 
  tidyplot(x = treatment, y = score) %>% 
  add_data_points_beeswarm(color = "grey") %>% 
  add_mean_dash(color = "orange") %>% 
  add_median_dash(color = "purple")

Dispersion & uncertainty

To complement the central tendency measure, it is often helpful to provide information about the variability or dispersion of the data points. Such measures include the standard error of the mean sem, the standard deviation sd, the range from the highest to the lowest data point and the 95% confidence interval ci95.

A classical representation of dispersion is an errorbar.

time_course %>% 
  tidyplot(x = day, y = score, color = treatment) %>% 
  add_mean_line() %>% 
  add_mean_dot() %>% 
  add_sem_errorbar(width = 2)

Or the use of a semitransparent ribbon.

time_course %>% 
  tidyplot(x = day, y = score, color = treatment) %>% 
  add_mean_line() %>% 
  add_mean_dot() %>% 
  add_sem_ribbon()

Another widely used alternative, especially for not normally distributed data is the use of violin or boxplot. Starting with the violin, the shape of these plots resembles the underlying distribution of the data points.

study %>% 
  tidyplot(x = treatment, y = score, color = treatment) %>% 
  add_violin()

These can be further augmented by adding, for example, the 0.5 quantile and the underlying data points.

study %>% 
  tidyplot(x = treatment, y = score, color = treatment) %>% 
  add_violin(draw_quantiles = 0.5) %>% 
  add_data_points_beeswarm()

The boxplot is the more classical approach, in which the quantiles are visualized by a box and whiskers.

study %>% 
  tidyplot(x = treatment, y = score, color = treatment) %>% 
  add_boxplot()

Finally, although it is not strictly a measure of central tendency, you can fit a curve through your data to derive an abstracted representation.

time_course %>% 
  tidyplot(x = day, y = score, color = treatment) %>% 
  add_curve_fit()

Distribution

When looking at a single distribution of values, a classical approach for visualization is a histogram.

energy %>% 
  tidyplot(x = power) %>% 
  add_histogram()

If you want to compare multiple distributions, violin or boxplot are two potential solutions.

distributions %>% 
  tidyplot(x = name, y = value) %>% 
  add_violin() %>% 
  add_data_points_beeswarm()

Proportion

Proportional data provides insights into the proportion or percentage that each individual category contributes to the total. To explore the visualization of proportional data in tidyplots, let’s introduce the energy dataset.

energy %>% 
  dplyr::glimpse()
#> Rows: 242
#> Columns: 5
#> $ year          <dbl> 2002, 2002, 2002, 2002, 2002, 2002, 2002, 2002, 2002, 20…
#> $ energy_source <fct> "Nuclear", "Fossil brown coal / lignite", "Fossil hard c…
#> $ energy_type   <fct> Nuclear, Fossil, Fossil, Fossil, Fossil, Other, Renewabl…
#> $ power         <dbl> 23.538, 20.304, 28.294, 20.300, 5.300, 0.000, 4.937, 0.9…
#> $ power_unit    <chr> "GW", "GW", "GW", "GW", "GW", "GW", "GW", "GW", "GW", "G…

As you might appreciate, this dataset contains the power in gigawatt produced per energy_source in Germany between year 2002 and 2023. Let’s start with a pie plot.

energy %>% 
  tidyplot(color = energy_type) %>% 
  add_pie()

The above plot represents the count of values across the different energy_type categories.

However, we might be more interested, in the sum contribution of each energy_type to the total power production. Therefore, we have to provide the variable power as a y argument to the tidyplots() function.

energy %>% 
  tidyplot(y = power, color = energy_type) %>% 
  add_donut()

Now we can appreciate the contribution of each energy type. Note that I also changed the pie for a donut plot, which is basically a pie chart with a white hole in the center.

The main criticism of pie and donut plots is that the human brain struggles to accurately interpret the proportions represented. For example, it’s challenging to determine whether fossil or renewable power has a larger share in the plot.

Another graphical representation is a barstack plot.

energy %>% 
  tidyplot(y = power, color = energy_type) %>% 
  add_barstack_absolute()

Here it might be slightly easier to compare the energy types. However, to really pinpoint this, we probably need to go back to a classical bar plot.

energy %>% 
  tidyplot(x = energy_type, y = power, color = energy_type) %>% 
  add_sum_bar() %>% 
  add_sum_value() %>% 
  reorder_x_axis_labels("Renewable")

Now we can see that renewable power production had a slightly larger overall share than fossil power production in the period between 2002 and 2023.

If we want to visualize proportional data across time or another variable, barstack plots are a great way to.

energy %>% 
  tidyplot(x = year, y = power, color = energy_type) %>% 
  add_barstack_absolute()

Moreover, to see the relative instead of the absolute contribution, we can use the add_barstack_relative() function.

energy %>% 
  tidyplot(x = year, y = power, color = energy_type) %>% 
  add_barstack_relative()

A similar plot can be achieved using an areastack.

energy %>% 
  tidyplot(x = year, y = power, color = energy_type) %>% 
  add_areastack_relative()

In both plots, the increasing contribution of renewable energy to the total energy production over time becomes apparent.

This can also be shown using donut plots. However, we need to downsample the dataset to 4 representative years.

energy %>% 
  # downsample to 4 representative years
  dplyr::filter(year %in% c(2005, 2010, 2015, 2020)) %>% 
  # start plotting
  tidyplot(y = power, color = energy_type) %>% 
  add_donut() %>% 
  adjust_colors(new_colors = c("Fossil" = "grey",
                               "Nuclear" = "#F6C54D",
                               "Renewable" = "#4FAE62",
                               "Other" = "#C02D45")) %>%  
  split_plot(by = year)

Now, let’s examine a related dataset that presents one week of energy data with higher time resolution.

energy_week %>% 
  tidyplot(date, power, color = energy_source) %>% 
  add_areastack_absolute()

In this plot, one can appreciate the higher contribution of solar power during day time in comparison to night time.

Also this plot can be shown as a relative areastack.

energy_week %>% 
  tidyplot(date, power, color = energy_source) %>% 
  add_areastack_relative()

This illustrates nicely how wind energy compensates for the lack of solar power during the night. However, when wind is weak, as on September 10, fossil energy sources need to step in to fill the gap.

Statistical comparison

To test for differences between experimental groups, tidyplots offers the functions add_test_asterisks() and add_test_pvalue(). While the first one includes asterisks for symbolizing significance.

study %>% 
  tidyplot(x = dose, y = score, color = group) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_asterisks()

add_test_pvalue() provides the computed p value.

study %>% 
  tidyplot(x = dose, y = score, color = group) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_pvalue()

As you might have noted, when using these functions, a caption is automatically included that provides details about the statistical testing performed. The default is a Student’s t test without multiple comparison adjustment.

Both can be changed by providing the method and p.adjust.method arguments.

For example, let’s perform a Wilcoxon signed-rank test with Benjamini–Hochberg adjustment.

study %>% 
  tidyplot(x = dose, y = score, color = group) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_pvalue(method = "wilcoxon", p.adjust.method = "BH")

It often makes sense to compare all experimental conditions to a control condition. For example, let’s say treatment “A” is our control.

study %>% 
  tidyplot(x = treatment, y = score, color = treatment) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_pvalue(ref.group = "A")

In some scenarios you have a mixture of significant and non-significant p values.

gene_expression %>% 
  # filter to one gene
  dplyr::filter(external_gene_name == "Apol6") %>% 
  # start plotting
  tidyplot(x = condition, y = expression, color = sample_type) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_pvalue()

Here you can choose to hide the non-significant p value using hide.ns = TRUE.

gene_expression %>% 
  # filter to one gene
  dplyr::filter(external_gene_name == "Apol6") %>% 
  # start plotting
  tidyplot(x = condition, y = expression, color = sample_type) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_pvalue(hide.ns = TRUE)

Finally, if you want to hide the caption with statistical information you can do this by providing hide_info = TRUE.

gene_expression %>% 
  # filter to one gene
  dplyr::filter(external_gene_name == "Apol6") %>% 
  # start plotting
  tidyplot(x = condition, y = expression, color = sample_type) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_test_pvalue(hide.ns = TRUE, hide_info = TRUE)

There are many more things you can do with statistical comparisons. Just check out the documentation of add_test_pvalue() and the underlying function ggpubr::geom_pwc().

Annotation

Sometimes you wish to add annotations to provide the reader with important additional information. For example, tidyplots let’s you add a title and a caption.

study %>% 
  tidyplot(x = treatment, y = score, color = treatment) %>% 
  add_mean_dash() %>% 
  add_sem_errorbar() %>% 
  add_data_points() %>% 
  add_title("Interesting study") %>% 
  add_caption("Here is some more detail how the study was performed")

In other cases you might want to highlight specific data points or reference values in the plot. Let’s take the animals dataset and plot speed versus weight.

animals %>% 
  tidyplot(x = weight, y = speed) %>% 
  add_data_points()

Here it might be interesting to have closer at the extreme values. First, let’s highlight the heaviest and the fastest animal.

animals %>% 
  tidyplot(x = weight, y = speed) %>% 
  add_data_points() %>% 
  add_data_points(data = max_rows(weight, 1), color = "red", shape = 1, size = 2) %>% 
  add_data_points(data = max_rows(speed, 1), color = "red", shape = 1, size = 2)

Now it would interesting to know the names of these animals. We can plot the names of all animals.

animals %>% 
  tidyplot(x = weight, y = speed) %>% 
  add_data_points() %>% 
  add_data_labels(label = animal)

Note that I provided the label argument to the add_data_labels() function to indicate the variable in the dataset that should be used for the text labels.

Another thing to note is that there is quite some overlap of labels in the lower left of the plot. Let’s try to separate the data labels using the add_data_labels_repel() function.

animals %>% 
  tidyplot(x = weight, y = speed) %>% 
  add_data_points() %>% 
  add_data_labels_repel(label = animal)

While the general idea might have been good, there are still too many labels to be plotted. So let’s restrict the labels to the 3 heaviest and the 3 fastest animals.

animals %>% 
  tidyplot(x = weight, y = speed) %>% 
  add_data_points() %>% 
  add_data_labels_repel(data = max_rows(weight, 3), animal) %>% 
  add_data_labels_repel(data = max_rows(speed, 3), animal)

There is lot tweaking that can be done with repelling data labels. For more details have a look at the documentation of add_data_labels_repel(), the underlying function ggrepel::geom_text_repel() and ggrepel examples.

As one last thing, let’s add some reference lines, to highlight specific values on the x and y-axis.

animals %>% 
  tidyplot(x = weight, y = speed) %>% 
  add_reference_lines(x = 4000, y = c(100, 200)) %>% 
  add_data_points() %>% 
  add_data_labels_repel(data = max_rows(weight, 3), animal) %>% 
  add_data_labels_repel(data = max_rows(speed, 3), animal)

What’s more?

To dive deeper into code-based plotting, here a couple of resources.

tidyplots documentation

Other resources