Beyond the 1D scatterplot
When we discussed the fundamental charts, we saw the most basic way to visualize one dimensional quantitative data. In this post we will discuss alternatives ways to perform this task.
1-D Scatterplot
As we already discussed, the one dimensional scatterplot is the easiest way to visualize one dimensional quantitative data.
This visualization can however become very crowded as the number of points grows, and in this case it is hard to count the number of points within a certain range of values. This is known as the curse of dimensionality, and it’s one of the most central problems in dataviz.
In order to allow for a larger number of points, you can jitter your points, namely put them randomly on your y axis. This operation may however confuse your audience if they are not used to it.
Histogram
Histograms are very common ways to show how a certain quantity is distributed. In a first step, a range of values is divided in $n$ equally spaced intervals. We then visualize the number of objects belonging to each interval.
While this visualization scales much better than the scatterplot, it has as main drawback the fact that the results (and so the conclusions) depend on the choice of the interval, and the discretization may hide some important feature of the data.
Boxplot
Boxplot is another popular way to show one dimensional distributions. In this visualization, rather than showing the data, we show the main features of the underlying distribution.
In order to understand this visualization, your audience should be trained, so it’s not suited for general audience. There is no agreement on the exact values one should visualize. As an example, we used the 10th, 25th, 50th, 75th and 90th percentile for the five values shown (the leftmost point, the left border of the box, the central value, the right border of the box and the rightmost point respectively). Moreover, this visualization is only meaningful for unimodal data, so you should always make sure that these condition hold.
Kernel Density Estimate
Kernel Density Estimates, KDE for brevity, are a set of methods to estimate the probability distribution function of a set of data.
The result strongly depends on the kernel choice as well as on its free parameters, which are both subjective choices. Before using this method you should make sure that your kernel is appropriate for your data. Due to the strong dependence of the result on the kernel choice, many authors both show the KDE plot and the scatterplot.
In case you want to know more about KDE, there are many resources online, as they have been very popular in the Machine Learning community. You can take a look at the Wikipedia page or at the scikit-learn documentation.
Conclusions
We have seen few methods to visualize the distribution of a single quantitative variable. If your dataset is small, then the best way to show your data is simply to show them on a scatterplot, as this is the most transparent way to visualize a single quantitative variable.
If your dataset is large, however, choosing how to visualize your data is a trade off between transparency and readability, and there is no single answer to the question of how to choose the most appropriate visualization.