Akshay DineshNumbers can sound more trustworthy than pictures. A mean, a variance, a correlation coefficient, and a regression line all feel precise, while a graph can seem like a quick sketch. Anscombe’s quartet turns that instinct upside down. It shows that four datasets can share nearly the same summary statistics and still tell four very different stories once the points are graphed.
The example comes from Francis J. Anscombe’s 1973 paper Graphs in Statistical Analysis, published in The American Statistician. Anscombe built four small datasets, each with eleven paired points. On paper, the four sets look almost identical: the x-values and y-values have the same means, the same variances, almost the same correlation, and the same straight-line regression equation. In a scatterplot, they stop looking alike.
The Trick Behind the Quartet
Each dataset in Anscombe’s quartet has an average x-value of 9 and an average y-value of about 7.5. Each has a similar spread. Each has a correlation of about 0.816, which suggests a fairly strong positive linear relationship. Each produces the same regression line, commonly written as y = 3 + 0.5x.
If those numbers were all a reader saw, the four datasets would appear to support the same conclusion. A student might say, As x increases, y tends to increase, and the line of best fit explains the pattern fairly well. That statement is not exactly false, but it is badly incomplete. The missing piece is the shape of the data.
When the points are graphed, the first dataset looks like an ordinary linear trend with some natural scatter. The second follows a curved pattern, so a straight line misses the real shape. The third is mostly linear, but one outlier pulls attention away from the main cluster. The fourth has most points stacked vertically, with one high-leverage point creating the appearance of a relationship. Same summaries, different stories.
Why Averages Can Hide the Shape
An average is useful because it compresses a group of values into one reference point. That compression is also its weakness. Two classes can have the same average quiz score even if one class has most students near the middle and the other has half very high scores and half very low scores. The mean helps answer one question, but it cannot show the whole distribution.
Variance and standard deviation add more information because they describe spread. Yet spread still does not reveal arrangement. A dataset shaped like a loose line, a curve, a cluster with one outlier, or a nearly vertical stack can share similar spread values. The numbers report how far values tend to sit from the average, not how the points are arranged in the plane.
Correlation has the same limitation. It measures the strength and direction of a linear relationship, but it does not know whether the pattern is truly linear. A curved relationship can produce a misleading correlation. A single unusual point can make a relationship look stronger or weaker than it really is. Correlation is a useful clue, not a full description of the data.
What the Four Graphs Teach
The first graph in Anscombe’s quartet is the one people often expect from a positive correlation. The points roughly follow a straight upward trend. A regression line is not perfect, but it gives a reasonable short summary. In that case, the summary statistics and the graph support each other.
The second graph is different because the points curve. A straight regression line can still be calculated, but it misses the pattern that the eye notices immediately. This is a common lesson in real data: a formula may fit the assignment, but the graph may show that the chosen model is the wrong shape. For curved data, another model may explain the relationship better.
The third graph warns readers about outliers. Most points follow a clean line, but one unusual point changes the calculation. Without the graph, the outlier is easy to miss. With the graph, the next question becomes obvious: Is that point a measurement error, a special case, or a real observation that deserves attention?
The fourth graph shows the power of a high-leverage point. Most x-values are the same, so they cannot show much relationship by themselves. One point sits far to the right and high enough to pull the regression line upward. The correlation and line may look impressive numerically, but the graph reveals that one point is doing most of the work.
Graphs Are Not Decoration
Anscombe’s quartet matters because it changes how readers should treat graphs. A graph is not a decorative picture added after the serious work is finished. It is part of the serious work. It helps reveal whether a calculation is appropriate, whether a model is plausible, and whether unusual points are shaping the conclusion.
This is why good data analysis usually begins with exploratory graphs. A scatterplot can show clusters, gaps, curves, outliers, changing spread, and strange measurement patterns. Those details guide better questions. Should the data be split into groups? Is a line of best fit reasonable? Is one point so influential that it needs investigation? Are the axes and units being interpreted correctly?
Graphs can also protect readers from overconfident claims. A report may state that two variables have a strong correlation, but the scatterplot may show that the relationship depends on one extreme value. Another report may give the same mean for two groups, while a graph shows that one group is far more uneven. The picture does not replace the numbers. It keeps the numbers honest.
The Modern Version: Same Stats, Different Graphs
Anscombe’s quartet became a classic because it made a small, sharp point. Decades later, researchers at Autodesk Research extended the idea with a project often called the Datasaurus Dozen. Those datasets share nearly the same summary statistics too, but their graphs form very different shapes, including lines, circles, stars, and a small dinosaur-shaped cloud of points.
The Datasaurus example is playful, but the lesson is serious. Summary statistics can be identical even when the visual structure is completely different. This does not mean summary statistics are useless. It means they must be interpreted alongside the data’s shape, context, and collection method.
Modern tools make it easy to calculate statistics quickly. That speed is helpful, but it can also encourage readers to accept an answer before understanding the data. A quick graph slows the process in the right way. It asks the reader to notice whether the calculation is describing the pattern or hiding it.
How to Use the Lesson Well
The practical lesson is simple: graph the data before trusting the summary. For paired numerical data, a scatterplot is often the best starting point. Look for direction, strength, shape, clusters, and outliers. Then decide whether a correlation coefficient or regression line is a reasonable next step.
It also helps to compare the graph with the summary statistics. If the correlation is strong but the graph depends on one unusual point, say so. If the line of best fit exists but the points curve, say that too. If two datasets share a mean but one is far more uneven in the graph, the average alone is not enough.
- Start visually: make a graph before settling on a numerical summary.
- Check the model: use a straight line only when the pattern is reasonably linear.
- Investigate outliers: ask what unusual points represent before ignoring or emphasizing them.
- Match the claim to the evidence: a calculation can support a claim only if the graph and context support it too.
Anscombe’s quartet is memorable because it is a small example with a large warning. Data can look tidy in a table and still behave very differently in a graph. The careful reader uses both. Numbers give precision, graphs give shape, and the strongest conclusions come from letting each one test the other.
