A spurious correlation is a statistical relationship between two variables that appears meaningful but isn’t caused by one affecting the other. The two things move together in the data, but only because of coincidence or because a hidden third factor is driving both of them. Recognizing spurious correlations is one of the most important skills in interpreting data, whether you’re reading a news headline, evaluating a business decision, or understanding a scientific study.
How Spurious Correlations Work
Two variables are correlated when they tend to rise or fall together. Sometimes that relationship reflects a genuine cause and effect. But in a spurious correlation, the link is an illusion. The classic formulation goes like this: you observe a correlation between variable X and variable Y, then introduce a third variable Z and “hold it constant” statistically. If the correlation between X and Y drops to nearly zero once Z is accounted for, the original relationship was spurious. It existed only because Z was independently pushing both X and Y in the same direction.
That third variable is called a confounding variable (or sometimes a “lurking” variable). It affects both of the things you’re measuring, creating a pattern that looks like a direct connection between them. The correlation between X and Y is real in the data. You can calculate it, graph it, and get a statistically significant result. But it doesn’t mean what it seems to mean, because the true cause sits elsewhere.
The Ice Cream and Shark Attack Example
The most commonly cited spurious correlation is the relationship between ice cream sales and shark attacks. Both increase during the same months, producing a tidy upward trend if you plot them on a graph. But buying ice cream doesn’t attract sharks. The confounding variable is warm weather: higher temperatures lead people to buy more ice cream and also to spend more time swimming in the ocean. Once you account for temperature and beach activity, the apparent connection between ice cream and sharks disappears.
Other well-known examples follow the same pattern. The number of college degrees awarded correlates with movie ticket sales over time, but both are driven by population growth. Swimming pool drownings track closely with nuclear energy production, again because a growing population independently increases both. These examples sound absurd, and that’s the point. They make it obvious that correlation alone tells you nothing about cause and effect. The trickier cases are the ones where the connection seems plausible enough that people accept it without checking for a hidden third factor.
Why Plausible-Sounding Correlations Are Dangerous
Absurd pairings like ice cream and sharks are easy to dismiss. The real damage from spurious correlations happens when the relationship sounds reasonable. Microsoft once studied whether users who tried advanced features in Office software were less likely to cancel their subscriptions. The data showed a clear correlation: advanced feature use predicted retention. But the relationship was spurious. Heavy users were independently more likely to explore advanced features and more likely to stick around. The features weren’t causing loyalty; a third factor, how engaged someone already was, explained both behaviors.
This kind of mistake drives bad decisions constantly. A company might invest millions in promoting advanced features to casual users, expecting it to reduce cancellations, when the real lever is overall engagement. In medicine, observational studies can show that people who take a certain supplement have lower rates of a disease, when the actual explanation is that health-conscious people are more likely to both take supplements and exercise regularly. The supplement gets the credit; the exercise was the real factor.
How Scientists Test for Causation
Scientists use several tools to distinguish genuine causal relationships from spurious ones. The most straightforward is the randomized controlled trial, where participants are randomly assigned to different groups. Randomization is powerful because it distributes confounding variables evenly across groups, making it much harder for a hidden third factor to create a false pattern.
When experiments aren’t possible, researchers rely on a set of guidelines originally proposed by the epidemiologist Austin Bradford Hill. These nine viewpoints help evaluate whether a correlation likely reflects true causation:
- Strength: A strong association is harder to explain away as confounding than a weak one.
- Consistency: If the same relationship appears across different populations and settings, it’s less likely to be a fluke of one dataset.
- Temporality: The supposed cause must come before the effect. This sounds obvious, but in many datasets the timing is unclear.
- Dose-response: If more exposure leads to more of the outcome, that pattern is difficult to produce through confounding alone.
- Plausibility: There should be a reasonable biological or mechanical explanation for how the cause produces the effect.
- Coherence: The relationship should fit with what’s already known from other lines of evidence.
- Experiment: Associations observed in controlled experiments provide the strongest evidence against spuriousness.
No single criterion proves causation on its own. Researchers look at the full picture. A relationship that is strong, consistent across populations, follows a dose-response pattern, and has a plausible mechanism is far more likely to be real than one that simply shows up in a single dataset.
Modern researchers also use tools called directed acyclic graphs (essentially causal diagrams) to map out which variables could be confounders and design studies that account for them. These diagrams force researchers to be explicit about their assumptions, making it easier to spot where a spurious relationship might be hiding.
P-Hacking and Manufactured Correlations
Spurious correlations don’t always arise naturally. They can be manufactured through a practice called p-hacking, where someone tests dozens or hundreds of variable combinations until they find a pair that happens to correlate. With enough variables, you’re virtually guaranteed to find patterns that look statistically significant but are pure noise.
This happens in both research and business settings. A data analyst under pressure to find “insights” might test every possible combination of marketing spend, weather data, website clicks, and sales figures until something lines up. The resulting correlation is technically real in the dataset but meaningless as a predictor of anything. It’s the statistical equivalent of dealing enough poker hands until you get a royal flush, then claiming you’ve found a system.
The website “Spurious Correlations,” created by Tyler Vigen, illustrates this perfectly by pairing unrelated datasets that happen to track together over time. Per capita cheese consumption correlates with the number of people who die tangled in their bedsheets. U.S. spending on science correlates with suicides by hanging. The numbers line up beautifully, and every one of them is meaningless.
Spurious Correlations in Artificial Intelligence
This problem has taken on new urgency with the rise of machine learning. AI models learn by finding patterns in training data, and they are highly susceptible to latching onto spurious correlations as shortcuts. If a model is trained on a dataset where most photos of waterbirds happen to have water in the background (and most photos of land birds have land backgrounds), the model may learn to classify birds by their background rather than by the bird itself. It performs well on similar data but fails badly when shown a waterbird standing on land.
This kind of shortcut learning can have serious consequences. A medical imaging model might learn to associate a hospital’s specific label format with a diagnosis rather than learning to read the actual scan. A hiring algorithm might pick up on zip codes or names as proxies for demographics rather than evaluating qualifications. In each case, the model has found a genuine pattern in the training data, but the pattern doesn’t generalize to the real world because the correlation was spurious.
Researchers are actively developing techniques to make models less reliant on these shortcuts, including training approaches that deliberately remove spurious features or penalize models for using them. The core challenge remains the same one that statisticians have wrestled with for over a century: separating the patterns that reflect reality from the ones that just happen to appear in a particular dataset.