Regression by Eye: Estimating Trends in Bivariate Visualizations
summary
Crowdsourced experiments (n=144 on Mechanical Turk) use an interactive wiggle test slider to let viewers fit slopes, quadratics and trig curves; slope estimates show a log error around 2.4 for well-behaved zero-centered Gaussian residuals. Introducing outliers shifts visual fits toward a robust line (not ordinary least squares nor equal-weighted), and area charts induce a within-the-area bias that underestimates y-intercepts by about 5–10%. Study materials and code are available on GitHub.
AI-written summary · about
chapters (11)
- 0:00Introduction to regression by eye
- 3:37Why not always show trendlines
- 6:10Assessing accuracy of visual estimation
- 9:02Results of slope estimation study
- 10:26Impact of outliers on visual regression
- 11:49Chart type biases in trend estimation
- 13:57Prescriptive advice on visual asymmetry
- 14:18Conveying uncertainty in trends
- 15:21Nudges and their implications
- 16:04Conclusion and future work
- 16:29Q&A session
related talks
01h 10mStorytelling through data
31mFundamentals of Data Visualization
22mVisualizing uncertainty with hypothetical outcomes plots
02h 23mEffective Data Visualization
09mAn intelligent assistant for mediation analysis in visual analytics
23mEffective Visualizations
01h 42mRecreating, Understanding, and Visualizing FiveThirtyEight's...
35mVisual Diagnostics for More Effective Machine Learning