Causation vs Correlation | The Most Confused Concept in Data Science

Key Takeaways

Did you know that as ice cream sales increase, shark attacks also tend to increase? While studying machine learning, you have probably seen correlation heat maps showing relationships between variables. In this case, ice cream sales and shark attacks show a positive correlation. But can we say that if you eat a lot of ice cream, you will be attacked by a shark? Probably not. And there are some important reasons behind this which we will discuss in this video. Correlation is a statistical term that describes a relationship between two variables. When two variables are correlated, it means they tend to change together. This defines association. It means two variables are somehow connected to each other, but we cannot determine whether one causes the other or whether they cause each other. Causation, on the other hand, means that one event directly causes another. For example, if you flip a light switch and the light turns on, flipping the switch causes the light to turn on. Often, what looks like a direct relationship is actually driven by hidden factors. And if you think deeply about how complex different things interact in the world, this will start to make sense. For example, let's take temperature as our hidden factor. During hot summer days, people buy more ice cream to cool down and more people go swimming in the ocean to beat the heat. Logically, the increased number of swimmers raises the chance of shark incidents. If we assume that only these three factors interact with each other without influence from any other variables, then in this simplified scenario, temperature causes both ice cream sales and shark attacks to increase, creating the appearance that these two variables are directly related. This example highlights what is called a confounding variable, a hidden factor that influences both variables you're studying. Without accounting for these, we risk making false conclusions. In data science, confusing correlation with causation can lead to costly mistakes. Imagine you're working for an e-commerce company and you discover that customers who browse the premium watches section tend to spend more overall. If you jump straight to the conclusion that visiting this section caused higher spending, you might redesign the site to push more people there. But what if the real reason is that wealthier customers both like browsing premium watches and tend to spend more in general? Here, customer income is the confounding variable, not the watches. So, you probably want to uncover causation more than association. But in data science, demonstrating causation requires stronger evidence. As a content creator, I might want to optimize my thumbnails to get more clicks on my videos. If I simply post the video with thumbnail A for the first week and then switch to thumbnail B for the second week, I might see that B got more clicks. But can I really say that B caused the increase? Not necessarily, because other factors might have changed. Maybe the second week was a holiday or YouTube recommended my video more or maybe my topic just started trending. A better way to measure causation is to run an AB test. That means I would randomly show thumbnail A to half of my audience and thumbnail B to the other half at the same time. By randomizing who sees which thumbnail, we make sure that other factors like time of day, seasonality, or viewer demographics are evenly balanced between groups. Now, if I see that thumbnail B consistently gets a higher click-through rate across both groups, I have much stronger evidence that the design of the thumbnail itself is what's causing the difference, not some hidden variable. But here's the thing. Even with a perfectly designed AB test, we can rarely claim pure 100% certain causation. There's always the possibility that some unknown factor we didn't measure is still influencing the results. The goal is to reduce the uncertainty as much as possible and make well-informed decisions based on the best available evidence. Subscribe and like the video if you enjoyed it. There are great topics yet to cover.