Friends don't let friends make certain types of data visualization - What are they and why are they bad.
Author: Chenxin Li, postdoctoral associate at Center for Applied Genetic Technologies, University of Georgia.
Contact: Chenxin.Li@uga.edu
This is an opinionated essay about good and bad practices in data visualization. Examples and explanations are below.
The Scripts/ directory contains .Rmd files that generate the graphics shown below.
It requires R, RStudio, and the rmarkdown package.
- R: R Download
- RStudio: RStudio Download
- rmarkdown can be installed using the intall packages interface in RStudio
- Friends Don't Let Friends Make Bar Plots For Mean Separation
- Friends Don't Let Friends Make Violin Plots for Small Sample Sizes
- Friends Don't Let Friends Use Bidirectional Color Scales for Unidirectional Data
- Friends Don't Let Friends Make Bar Plot Meadow
- Friends Don't Let Friends Make Heatmap without Reordering Rows & Columns
- Friends Don't Let Friends Make Heatmap without Checking Outliers
- Friends Don't Let Friends Forget to Check Data Range at Each Factor Level
- Friends Don't Let Friends Make Network Graphs without Trying Different Layouts
- Friends Don't Let Friends Confuse Position and Length Based Visualizations
- Friends Don't Let Friends Make Pie Charts
- Firends Don't Let Friends Make Concentric Donuts
This has to be the first one. Means separation plots are some of the most common in scientific publications. We have two or more groups, which contains multiple observations; they may have different means, variances, and distributions. The task of the visualization is to show the means and the spread (dispersion) of the data.
In this example, two groups have similar means and standard deviations, but quite different distributions. Are they really "the same"? Just don't use bar plot for means separation, or at least check a couple things before settling down on a bar plot.
It's worth mentioning that I was inspired by many researchers who have tweeted on the limitation of bar graphs. Here is a pulication: Weissgerber et al., 2015, PLOS Biology.
This is quite common in the literature as well, but unfortunately, violin plots (or any sort of smoothed distribution curves) make no sense for small n.
Distributions and quartiles can vary widely with small n, even if the underlying observations are similar. Distribution and quartiles are only meaningful with large n. I did an experiment before, where I sampled the same normal distribution several times and computed the quartiles for each sample. The quartiles only stablize when n gets larger than 50.
Excuse my language, but this is a truly data visualization sin, and again quite common. I can understand why this error is common, because it appears that many of us have not spent a lot of thoughts on this issue.
Color scales are pretty, but we have to be extra careful. When color scales (or color gradients) are used to represent numerical data, the darkest and lightest colors should have special meanings. You can decide what those special meanings are: e.g., max, min, mean, zero. But they should represent something meaningful. A data visualization sin for heat maps/color gradients is when the lightest or darkers colors are some arbitrary numbers. This is as bad as the longest bar in a bar chart not being the largest value. Can you imagine that?
We talked about no bar charts for means separation, but this is a different issue. It has to do with presenting results of a multi-factorial experiment. Bar plot meadows are very common in scientific publications and unfortunately also ineffective in communicating the results.
Data from: Matand et al., 2020, BMC Plant Biology
Bar plot meadows are common because multi-factorial experiments are common. However, bar plot meadows are poorly designed for its purpose. To communicate results of a multi-factorial experiment, it requires thoughtful designs regarding grouping/faceting by factors of interest.
In this example, I focus on comparing the effect of Treatment & Explant on Response at the level of each Variety.
However, if the focus is the effect of Treatment & Variety on Response at the level of each Exaplant, then it will require a different layout.
Heatmaps are very common in scientific publications, and very very common in omics papers. However, for heatmaps to be effective, we have to consider the ordering of rows & columns.
In this example, I have cells as columns and features as rows. Each grid is showing z scores. It is impossible to get anything useful out of the heatmap without reordering rows and columns. We can reorder rows and columns using clustering, but that is not the only way. Of course, if the rows and columns are mapping to physical entities (rows and columns of a 96-well plate), then you can't reorder them. But it is a very good idea to at least consider reordering rows and columns.
Data from: Li et al., 2022, BioRxiv
Outliers in heatmap can really change how we perceive and interpret the visualization. This generalizes to all sort of visualizations that use colors to represent numeric data. Let me show you an example:
In this example, I have 2 observations. For each observations, I measured 20 features. Without checking for outliers, it may appear that the 2 observations are overall similar, except at 2 features. However, after maxing out the color scale around 95th percentile of the data, it reveals that the two observations are distinct across all features.
This is a common issue that many of us have encountered. In a multifactor experiment, sometimes the range of the response variable changes widely between different factor levels.
This hypothetical experiment measured 3 compounds across 2 groups (control vs. treatment). Without checking data range for each compound, you will likely have missed that the treatment had a strong effect on compound 1. This is because the concentration of compound 1 has a much narrower range than the other compounds in this experiment.
Network graphs are common in scientific publications. They are super useful in presenting relationship data. However, the apparence (not the topology) of the network can make a huge difference in determing if a network graph is effective.
Layouts can drastically change the appearance of networks, making them easier or harder to interpret. Here are 3 network graphs from the same data. They look very different from each other. Data from: Li et al., 2022, BioRxiv
This is always an elephant in the room and the essence of many misleading visualizations. In this example, I measured a response variable across 3 time points. Two of the following graphs are fine, but one of them is a data visualization crime. Can you see why?
In dot and line plots, values are represented by positions along the x and y axis. The same idea applies to other position based visualizations, such as box plots. In bar plots, values are represented by the distance from the x axis, and thus the length of the bar.
The 3rd graph is not 0-based, which makes the bar length at time point 2 about 3x longer than that at time point 1. In fact, the true difference in means is closer to 1.6x. I hope you can see how confusing length and position based visualizations can lead to misleading graphs.
Pie chart is a common type of visualization for fractional data, where fractions add up to 100%. This is achieved by dividing a circle into sectors, and the sectors add up to a full circle. Pie charts have been criticized, because human are much worse in reading angles and area than reading lengths. Here is a blog post that explores that.
In this example, we have two groups, each contains 4 sub-category. In classic pie charts, the angles (and thus arc lengths & sector area) represent the data. The problem is that it is very difficult to compare between groups. We can visually simplify the pie chart into donut charts, where the data are now represented by arc lengths. However, if we want to use lengths to represent the data, why don't we just unwrap the donut and make stacked bars? In stacked bar graphs, bars are shown side-by-side and thus easier to compare across groups.
Fun fact: the scripts underlying stacked bars are much simpler than those underlying the pie charts and donut charts. If you want to produce sub-optimal graph types with ggplot, you actually have to work extra hard.
In this example, we have 3 groups, each of which contains two sub-categories (Type I or Type II).
In concentric donuts, you might be tempted to say the data is represented by the arc length, which is in fact inaccurate. The arc lengths on the outer rings are much longer than those in the inner rings. Group 2 and Group 3 have the same exact values, but the arc lengths of Group 3 are much longer. In fact the data is represented by the arc angle, which we are bad at reading.
Since outer rings are longer, the ordering of the groups (which group goes to which ring) has a big impact on the impression of the plot. It can lead to the apparent paradox where larger values have shorter arcs. The better (and simplier!) alternative is just unwrap the donuts and make a good old stacked bar plot. BTW, this is also my main issue with circos plots and other circular plot layouts.
That's it for now. I will update this when I have the time (and inspirations) to produce more examples. Not sure what the next one will be, but stay tuned!