When two things seem linked (but aren’t): Understanding different types of correlations
- Sarah H. Woodhouse
- Jul 2
- 6 min read
Updated: Jul 3
Key Takeaways
Not all correlations mean that one thing causes the other. Some correlations are real but misleading. Others are simply due to chance. Understanding which is which can help you avoid common reasoning errors—and make better decisions.
There are five common types of correlations. Only one involves a direct causal link. The others include spurious patterns (flukes), reverse causation, confounding variables, and feedback loops, each with different implications for how we interpret data.
Random patterns can look convincing. The more variables you compare, the more likely it is you’ll find strong but meaningless correlations. This is known as “data dredging,” and it’s a major cause of misinformation in science and media.
Before assuming causality, consider alternatives. Could B be causing A? Could both be caused by C? Could the link be coincidental? By questioning the source of a correlation, you can avoid flawed conclusions and spot more reliable patterns.
Imagine if a friend tells you that the per capita consumption of margarine is correlated with the number of divorces in Maine or that the cost of sending a letter using the United States Postal Service (USPS) increases at the same rate as people searching for “I am dizzy” on Google.
What would you make of these strange correlations? Do they mean that one thing causes the other?
The answer is: it depends.
You’ve probably heard the common maxim that correlation doesn’t imply causation – that’s because there are actually several different types of correlation, only one of which is explained by one thing directly causing another.
To illustrate the differences between types and help you avoid common mistakes when interpreting correlations, we’ve explained five different types of correlation – and provided examples of some funny correlations you may not have heard of!
1) Causal Correlation (A causes B)
Let’s start with the most obvious type of correlation: causal. This is when two things are associated because one is causing the other to happen. There are plenty of well-established causal relationships (exercise and cardiovascular health, seatbelt wearing and the chances of surviving a car crash).
Actually demonstrating that two things are causally related is challenging. We may have strong intuitions in that direction, or even be able to make compelling arguments for the presence of a casual relationship – but proof requires experimentation. For example, scientists had suspected that smoking may cause lung cancer since the 1930s, but conclusively proving this link (and getting the government to acknowledge it) necessitated a study with a sample size of over 40,000.
This chart comparing cigarette sales in the US with lung cancer deaths shows a strong positive correlation. But it also demonstrates how concrete evidence of causal relationships can take time to materialize! If we calculate the correlation in this graph relying solely on the extent to which cigarette sales and lung cancer deaths rise simultaneously, we get a fairly modest positive correlation of 0.43. But repeating this calculation accounting for the lag between smoking behavior and ultimate death from lung cancer produces a very strong correlation of 0.96.
2) Random fluke (A and B are correlated by chance)
Sometimes, two things can be correlated – even strongly so – for no reason at all. This is surprisingly common, especially when datasets are relatively small (meaning that correlations with no underlying cause appear particularly pronounced). An online database of these spurious correlations contains thousands of examples. Here are a few of our favorites:
The correlation between per capita margarine consumption in the US and the divorce rate in Maine is 0.9993 (that’s almost as high as it can be!):
Almost as high at 0.980 is the correlation between the cost of sending a letter via the US Postal Service and the number of people Googling “i am dizzy”:
These spookily high correlations don’t just occur when two things are on a general upward or downward trajectory! Take this 0.912 correlation between rain in San Francisco and the number of printing press operators in Rhode Island:
These strange correlations can be amusing – but they are actually unsurprising when you consider just how many variables there are out there. As Tyler Vigen, creator of the spurious correlations database, explains, he uncovers these patterns using the practice of ‘data dredging’, which involves comparing vast numbers of variables against each other and seeing what emerges. In his case, that’s over 636 million correlation calculations!
In cases where researchers use it to claim causal connections or statistical significance, data dredging can be a form of scientific malpractice.
3) Reverse Correlation (B causes A)
It’s possible for two things to be causally correlated, but not in the way they initially appear to be. One example is when A appears to cause B, but further analysis reveals that B actually causes A. This is called a reverse correlation.
One example is the relationship between smoking and mental illness. Studies had consistently shown higher rates of mental health problems such as depression among smokers. One 2019 study found evidence for the claim that although health professionals had historically assumed that people suffering from poor mental health were more likely to smoke as a form of self-medication, it may actually be the other way around – that smoking itself increases the risk of mental illness.
The study acknowledges that the correlation between smoking and mental health is likely caused by both adverse effects of smoking itself and the use of smoking as a coping mechanism. So this is not a straightforward case of reverse causality, but it does show how assuming that a causal link runs in one direction can obscure the possibility of it also running in the other.
A nice thing about reverse correlations is that they are often easy to eliminate. For example, it’s implausible that increased umbrella use causes it to rain! But as the smoking example demonstrates, there are cases where they can be trickier to detect.
4) Correlations caused by confounding variables (C causes both A and B)
A confounding variable or ‘confounder’ is one that causes two variables, making it appear as if those two variables are causally linked to each other.
One humorous and oft-cited example is the correlation between the stork population of a given country and its birth rate. As one researcher demonstrated, these two variables are highly correlated, which could be used as evidence to validate the popular urban myth that storks deliver babies:
This conclusion is obviously ridiculous, but we still need to explain the correlation in order to confidently dismiss it. The explanation is that land area is the confounding variable, which causes the apparent correlation between birth rate and stork population. The birth rate is measured in absolute terms rather than per capita, so it’s simply the case that larger areas of land have more people and more storks!
5) Cyclical Correlations (A causes B cause A causes B…)
Cyclical correlations occur when two variables mutually reinforce each other, creating a feedback loop. When people identify a “chicken and egg situation”, they’re often pointing to a cyclical correlation.
The Pygmalion effect is one example of a cyclical correlation. This is a psychological phenomena where people tend to meet the expectations set for them by others. One landmark 1968 study found that the Pygmalion effect is common in school classrooms. When a teacher has high expectations of a student, this leads them to perform better, which in turn heightens the teacher’s expectations further (and vice versa).
For example, a study investigating the impact of racial bias on student attainment found a strong positive correlation between a teacher’s confidence that a student will complete college and the likelihood that they actually will:
There are many reasons why two things may appear correlated! Next time you identify a correlation, before jumping to the conclusion that A is causing B, consider whether:
The correlation is a pure coincidence
B is causing A
Both A and B are actually caused by C
A and B are mutually reinforcing each other.
This will make you more likely to avoid the correlation-causation trap and understand how variables truly influence each other.
