acquired into information science, there was a phrase that we’d all heard; everybody is aware of it, younger and outdated:
“Correlation doesn’t suggest causation.”
It’s a catchy phrase, and also you’ve undoubtedly stated it a couple of times, and may even have nodded confidently when another person stated it. Particularly for datasets that don’t relate to one another, however the place it’s humorous and intriguing to suggest causation!
Listed here are two very attention-grabbing info:
- International locations that eat extra pizza are inclined to have larger math scores.
- The extra sun shades bought, the extra shark assaults happen.
Now, if that have been all the data you’ve… what do you have to conclude?
Does consuming pizza make you higher at math? Will shopping for a brand new pair of sun shades trigger a shark assault?
Although it’s humorous to consider, the reply to these questions is “most likely not”.
And but, these are examples of one thing very actual: Correlation.
The query price asking now’s: if correlation doesn’t equal causation, then what does it imply?
That’s the place issues get fuzzy.
As a result of we are inclined to deal with correlation like a imprecise concept, we consider it as if it means “They’re form of associated”, or “They transfer collectively someway”. However correlation isn’t only a feeling; it’s a exact mathematical measurement of how two variables transfer collectively.
As a substitute of simply repeating the warning, let’s really perceive the idea. When you do, these bizarre examples cease being stunning and begin making sense.
So, let’s get into it!
What’s correlation?
When folks say two issues are “correlated,” they normally imply one in all three issues:
- “These two issues appear associated.”
- “These two issues transfer collectively.”
- “There’s some connection between these two issues.”
On a floor stage, all three of those should not unsuitable, however they’re lacking some nuances.
Correlation shouldn’t be a vibe. It’s a measurement! And like several measurement, it solutions a really particular query.
Taking a step again, think about you acquire the information on what number of hours college students studied and their examination scores.
You plot it, and also you see one thing like this:
Every level represents one pupil. The x-axis is how lengthy they studied, and the y-axis is their rating.
While you take a look at this plot, you discover that the factors have a tendency to maneuver upward. So that you conclude, “As research time will increase, scores have a tendency to extend too”, which is what we name a constructive correlation.
However, is that only a development or is the information telling you one thing extra?
On this instance, the connection you simply plotted is: when one variable is above its common, the opposite tends to be above its common too.
That’s the important thing concept most individuals miss: correlation isn’t about uncooked values, it’s about how variables transfer relative to their averages.
So, the query correlation solutions is:
Do two variables transfer collectively in a constant manner?
That query has one in all three solutions:
- Up + up → constructive correlation
- Up + down → damaging correlation
- No constant sample → no correlation
The Math Behind Correlation
Let’s attempt to make interested by correlation extra intuitive. We’ll do this utilizing the Pearson correlation coefficient, which we will outline as:
Okay, I do know that equation isn’t what anybody thinks of once I say “intuitive”… However keep on with me and let’s unpack it with out turning it right into a lecture.
Step 1: Covariance (AKA Do They Transfer Collectively?)
Covariance appears to be like at how two variables transfer relative to their averages. For instance, if each variables are above their averages, we get constructive covariance; if one is above and the opposite under, we get damaging covariance.
Mainly, covariance solutions: “Are these variables aligned in how they deviate from their averages?”
Step 2: Normalize It
Covariance alone is difficult to interpret as a result of it depends upon scale. To beat that, we divide by the usual deviations: and . This rescales the whole lot right into a clear vary: -1 to 1. That provides us frequent floor for evaluating variable values.
After these two steps, we will now calculate the Pearson coefficient! If we get:
- +1 → good constructive relationship.
- 0 → no linear relationship.
- -1 → good damaging relationship.
This code merely measures how constantly these two variables transfer collectively—not how huge they’re, however how nicely aligned they’re.
What Completely different Correlations Look Like
- Left: robust constructive correlation → clear upward sample
- Center: no correlation → random scatter
- Proper: robust damaging correlation → downward sample
Correlation measures consistency of motion, not simply whether or not two variables are associated.
What Correlation Really Tells You
Correlation tells you: these variables transfer collectively in a structured manner. It tells us that there’s a sample right here to concentrate to.
However, it does NOT let you know why or how they do, or whether or not one causes the opposite.
The traditional instance of correlation is that ice cream gross sales and drowning incidents are correlated.
In truth, we will plot the variety of ice cream gross sales and drowning incidents to get:
We are able to see a transparent upward relationship between these two variables… extra ice cream gross sales result in extra drownings?…
However that’s deceptive. As a result of the true driver is temperature: sizzling climate means extra ice cream gross sales, extra folks going to the seaside, and extra swimming.
So, although we will clearly see that correlation is actual, the reason is hidden.
Correlation and Nonlinearity
Now contemplate this relationship:
y = x²
That is clearly a powerful relationship, as x will increase or decreases, y will increase! However for those who compute correlation:
np.corrcoef(x, y)[0,1]
You’ll get one thing near 0.
That’s as a result of correlation solely measures: How nicely a straight line matches the connection. This can be a essential limitation. If the connection is curved, correlation could fail, even when a powerful relationship exists.
So, as a substitute of pondering: “Correlation = relationship”, it’s higher to suppose: “Correlation = how nicely a straight line explains the connection.”
The Misunderstanding
The vagueness of the idea of correlation, and the way in which we’re taught it, results in some misunderstandings. Three quite common ones are:
- Assuming causation: Simply because two variables transfer collectively doesn’t imply one causes the opposite.
- Ignoring hidden variables: There could also be a 3rd issue driving each.
- Lacking nonlinear relationships: Correlation solely sees straight-line patterns.
You be questioning now, if correlation is a quite simple time period that doesn’t inform us a lot, why is it vital nonetheless?
As a result of it’s extremely helpful as a primary sign. It tells you:
“One thing attention-grabbing may be taking place right here.”
From there, you examine additional. Correlation measures alignment; additional investigation gives an evidence.
Last Takeaway
“Correlation doesn’t suggest causation.” That’s true. However right here’s the issue: folks hear this and suppose: “Correlation is meaningless.” That isn’t true!
Correlation measures how variables transfer collectively; it ranges from -1 to 1, captures linear relationships, but it surely does NOT suggest causation.
Correlation isn’t deceptive. We simply count on an excessive amount of from it when it isn’t making an attempt to elucidate the world. It’s only a sign indicating:
“Hey… this appears to be like attention-grabbing.”
Now, the true work begins, as we examine why that is actually attention-grabbing.
