Data Visualization under Uncertainty

Talk for the Boston Data Viz meetup. Covering visual display of uncertainty, modeling and bringing statistics into data vis. Notes, links and source code available here: https://gist.github.com/Protonk/4961378

Transcript

1. Data Visualization Under Uncertainty Adam Hyland, Data Viz intern at Bocoup
2. Some date!
3. Why are we here? • Data visualization is a tool • Develop a narrative • Inform a decision • Make clear a message • Data is complex enough • But...
5. Certainty and Narrative The...decision we made was to present data that fell within documented ranges, rather than reflect the findings of a particular report, because of the inherent challenge in collecting data on this issue. -Sarah Beaulieu, Enliven
6. What happened? • Enliven presented a point estimate • They felt this was a responsible choice • Actual data is animated by uncertainty
7. How do we solve this? • Mix of high and low level topics • Some examples from the web, some in R • Strategies
8. How do we do it now? • Where do we handle uncertainty in visualization well? • Sciences • Forecasting • Poorly? • Pretty much everywhere else
9. Dark Sky Weather data
10. But sometimes...
11. Taxonomy • Stochastic uncertainty • Model uncertainty (error) • Uncertainty about the future
12. But I repeat myself • What do we mean by stochastic uncertainty? • Dispersion • Sampled or not • A canonical example
13. Polls • Not a strawman • Standard presentation • Cognitive load
14. It gets better
15. What’s better, exactly? • The Good: • Conveys dispersion in results • Helps reader to infer estimate • The Bad: • Let’s see the site itself... • The Ugly: • Overplotting, etc.
16. Nate Silver
17. Visual language • How do we think about stochastic uncertainty? • How to show it? • Boxplots • Density estimates • Bootstrap • Models: in our department?
18. Boxplot
19. What is a boxplot?
20. Density Estimates
21. Think back to Pollster
22. Is this really uncertainty • Think about enliven example again • We are always picking an estimator
23. Anscombe's quartet
24. Some other solutions • Contour plots • Margins • Rug plot • Others
25. Lovely web visualization
26. Model Uncertainty • Model error is a term of art • Can refer to errors in statistical models, or simpler models
27. Analytics
28. Messy data
30. Forecasting • Distinct only because our data will usually embed the prediction • Uncertainty comes from obvious and non- obvious places • Already have a strong visual language
31. New York Times
32. Practical examples • Confidence regions • The bootstrap (again)
33. Confidence Intervals
34. Bootstrap
35. Good news • Many ways to show forecast data • Most of them not terrible
36. • Almost (all) visualization leans on estimators • Alternative to good statistics... • Expose more of the model
37. Let’s do it! • Push new things into our visual language • People trained by bad visualization • Retraining is hard work
38. That was fun, right?