Why a Scatter Plot Belongs Next to Every Correlation Number

The single best argument for always plotting your data

Anscombe's quartet is four small datasets, constructed deliberately in 1973, that share nearly identical summary statistics — the same mean, variance, and Pearson correlation coefficient (~0.816) across all four — while looking completely different when actually plotted:

1. A roughly linear relationship with normal scatter — the pattern the correlation coefficient is actually describing accurately.
2. A clear, smooth curve (non-linear) — the correlation number is misleadingly summarizing a real but non-linear relationship as if it were linear.
3. A near-perfect linear relationship, except for a single outlier that pulls the calculated correlation down from what it would otherwise be.
4. Nearly all identical x-values, plus one extreme point that entirely drives the calculated correlation on its own.

All four report the same correlation coefficient. Only one of them actually looks like what most people picture when they hear "correlation of 0.8."

What this demonstrates about relying on summary numbers alone

Any single summary statistic — correlation, mean, standard deviation — compresses an entire dataset's shape into one number, and that compression necessarily discards information. Anscombe's quartet is the canonical illustration that datasets with identical summary statistics can have entirely different actual structures: a real linear relationship, a real non-linear relationship, a relationship dominated by one outlier, or a relationship that's really just an artifact of a single extreme point. No amount of staring at the correlation number alone distinguishes between these — only plotting the actual data does.

What to specifically look for in a scatter plot

• Does the pattern actually look linear, or does it curve — suggesting the Pearson correlation (which only measures linear relationship) is summarizing a non-linear pattern inaccurately?
• Is one or a few points doing most of the work? A scatter plot immediately reveals whether a correlation is driven by the bulk of the data or by one or two extreme, influential points.
• Are there distinct clusters rather than one continuous relationship — a pattern a single correlation coefficient can't reveal, since it treats the whole dataset as one relationship.

Why this matters beyond a famous historical example

This isn't just a textbook curiosity — any time you compute a correlation matrix (or any summary statistic) across real data, the same risk applies: a "strong" or "weak" correlation number alone doesn't tell you why it's strong or weak, or whether the relationship is actually what a linear correlation implies. A quick scatter plot check, especially for any correlation you're about to act on, catches exactly the kind of misleading summary Anscombe's quartet was built to demonstrate.

Common mistakes

• Reporting a correlation coefficient without ever visualizing the underlying scatter. The number alone can't distinguish a genuine linear relationship from a non-linear one, an outlier-driven one, or a coincidental one.
• Assuming a low correlation means "no relationship to see." As Anscombe's quartet's curved dataset shows, a real, strong relationship can exist that a linear correlation coefficient doesn't capture at all.
• Not checking whether a single extreme point is driving an otherwise unremarkable correlation. This is exactly the pattern in two of Anscombe's four datasets, and it's invisible without an actual plot.

FAQ

What is Anscombe's quartet, in one sentence?
Four datasets with nearly identical summary statistics (mean, variance, correlation) that look completely different when actually plotted — a deliberate demonstration that summary numbers alone can be misleading.

Does a scatter plot replace the need for a correlation coefficient?
No — they're complementary; the coefficient gives a precise number for comparison, while the scatter plot reveals the actual shape and whether that number is a fair summary of what's really there.

Why does this matter for real, non-textbook data analysis?
Because the same risk Anscombe's quartet illustrates applies to any real correlation calculation — a reported number alone can't tell you whether it reflects a genuine linear relationship, a non-linear one, or an artifact of a few extreme points.

Plot any two numeric columns and see the actual relationship alongside the Pearson coefficient with the Scatter Plot Generator — entirely in your browser.