Statistician Francis Anscombe constructed four small datasets in 1973, each with an identical mean, an identical variance, and an identical correlation coefficient between their two variables — and when actually plotted, the four datasets look nothing alike: one shows a genuine, roughly linear relationship; one shows a clear, smooth curve that a linear correlation coefficient badly mischaracterizes; one shows a relationship driven almost entirely by a single influential outlier; and one shows essentially no real relationship among most of the points, except for one extreme value that alone is generating the entire calculated correlation.
Why this demonstration remains genuinely important decades later
Anscombe's quartet isn't a historical curiosity — it's a direct, concrete illustration of a genuinely important and still frequently overlooked point: identical summary statistics, including something as commonly relied upon as a correlation coefficient, can correspond to fundamentally different underlying data structures, meaning a report or analysis that presents only the summary statistic, without ever actually visualizing the underlying data, can be concealing a structural pattern the statistic alone gives no hint of.
What each of the four specific datasets actually reveals when plotted
The first dataset shows a genuinely appropriate linear relationship, where the calculated correlation and a fitted linear regression line meaningfully describe the actual underlying pattern. The second shows a clear, smooth, nonlinear curve, where fitting the same linear regression badly misrepresents the actual relationship, understating how strong the true, non-linear relationship actually is. The third shows a relationship that would be nearly perfect except for one clear outlier, which is dragging the calculated correlation and the fitted line away from what nearly all the other data points actually show. The fourth shows almost no genuine relationship among most of the data points, clustered at a single x-value, with the entire calculated correlation being generated by one single extreme point positioned far from that cluster.
Why this specifically argues for visualizing data, not simply computing more statistics
Each of these four very different underlying patterns produces exactly the same reported correlation coefficient, meaning no amount of additional numerical summarization using that same statistic alone would reveal which of the four very different situations you're actually looking at — only actually plotting the data directly reveals the difference, which is precisely Anscombe's original point: numerical summaries are useful and necessary, and they are not a substitute for actually looking at the data itself.
Why this matters directly for everyday business and research data analysis
A business analysis reporting a strong correlation between two variables, without accompanying visualization, could be describing any of Anscombe's four very different underlying patterns — a genuine, generalizable linear relationship, a mischaracterized nonlinear one, an outlier-driven artifact, or a near-total absence of relationship outside a single extreme data point — and a decision made based on the correlation coefficient alone, without checking which of these situations actually applies, risks acting on a badly mischaracterized understanding of what the data actually shows.
What this means for how data analysis should actually be conducted
- Always visualize the underlying data directly before relying on a summary statistic like a correlation coefficient to characterize a relationship
- Check specifically for outlier-driven relationships and nonlinear patterns that a linear correlation coefficient would badly mischaracterize
- Treat a reported correlation without accompanying visualization as an incomplete analysis, regardless of how confidently or precisely the number itself is reported
- Use Anscombe's quartet as a standing reminder that identical statistics can correspond to fundamentally different underlying realities
Anscombe's quartet remains one of the most efficient, concrete demonstrations available of a genuinely important principle — a summary statistic tells you something true and useful about data, and it never tells you the whole story on its own, which is exactly why looking at the actual data remains a non-negotiable step, however sophisticated the statistical summary might otherwise seem.