Visualising uncertainty

When and how to show uncertainty in charts

Dr Nicola Rennie

About me

Data visualisation specialist developing charts, standards, and training.

Background in statistics, operational research, data science consultancy, and academia.

Author of several R packages (mainly for visualisation).

Co-author of Royal Statistical Society’s Best Practices for Data Visualisation guidance.

Sometimes seeing uncertainty is necessary to get an accurate understanding of the data and its limitations.
However, it can add complexity that makes charts harder to read and understand.
Not all chart types can show uncertainty easily.
Showing uncertainty does not make your work less trustworthy.

Ask yourself: Does seeing the uncertainty change the conclusion?

You do not need to show uncertainty in charts if:

the uncertainty is consistently very small and it does not change the interpretation;
there is little or no overlap between ranges for values being compared;
there is some overlap between ranges but the overall trends are clear.

Estimate + Range(s)
- Probably the most common
- e.g. mean and confidence intervals
Distribution
- Show many (possible) values
- e.g. all estimates from 1,000 simulations or different scenarios
Summary statistics
- One number representing uncertainty
- e.g. variance
- Unlikely to be useful in a chart on its own

You have point estimates for:

And you want to show the range of uncertainty around those estimates.

Show how uncertainty in estimate(s) changes, rather than implying a single precise value throughout.
Might need to show changes in main trend line e.g. observed vs forecast.
Might only show uncertainty ranges for part of a time period e.g. forecasts.
Might show different metrics of uncertainty in estimate and forecast.

Your choice!

Go to nrennie.gitlab.io/dl4sg-uncertainty/exercises/visualising-uncertainty.html . You’ll see some R code to load a dataset and create a line chart with three lines.
Inspect the data. You will see additional columns containing different levels of confidence intervals for different estimates in the data.
Decide if, and how, the uncertainty could be presented.
Edit the code to show uncertainty on the chart.
Hint: you may find geom_ribbon() in ggplot2 useful.