Introduction
Correlation describes a statistical tendency for two variables to move together, either in the same direction or in opposite directions, while causation means that a change in one variable actually produces a change in the other; treating a correlation as proof of causation is one of the most common analytical mistakes, because two variables can move together for reasons that have nothing to do with either one driving the other.
Cricket analogy: Two players might both score more runs on days their team wins, but that doesn't mean scoring more runs alone is what wins the match; correlation means two things move together, while causation means one actually produces the change in the other, and mixing the two up leads to the wrong conclusion.
Explanation
A frequent source of spurious correlation is a confounding variable, a hidden third factor that independently influences both of the observed variables, making them appear related to each other when in fact neither one is affecting the other directly; a well known illustrative example is ice cream sales and drowning incidents both rising in summer, driven by warm weather rather than one causing the other.
Cricket analogy: A ground's ice-cream sales and its number of sixes hit might rise together during summer, not because sixes cause ice-cream sales or vice versa, but because hot weather is a hidden third factor driving both, the same confounding-variable trap that creates spurious correlations anywhere.
Because observational data alone cannot rule out confounding variables or reverse causation, establishing that one variable truly causes a change in another typically requires a controlled experiment with random assignment, where the only systematic difference between groups is the factor being tested, so any resulting difference in outcome can be attributed to that factor.
Cricket analogy: To prove that a new training drill actually causes better catching, not just correlates with it, a coach would need to randomly assign some players to the drill and others to normal training, then compare results, the same controlled, randomized comparison required to establish causation rather than mere observational correlation.
Example
# Illustrative example: correlation alone does not establish causation
import pandas as pd
df = pd.read_csv('example_monthly_data.csv')
correlation = df['variable_a'].corr(df['variable_b'])
print(correlation)
# A high correlation here still requires further investigation
# (confounders, timing, experiment design) before claiming causationAnalysis
When examining a correlated pair of variables from observational data, the useful discipline is to actively search for a plausible confounder before assuming a causal link, asking whether some third factor, such as season, market conditions, or a shared external event, could independently explain the movement of both variables rather than one variable driving the other.
Cricket analogy: A useful check is asking whether hot weather, not a coach's new drill, could explain why both sixes and drink breaks rose together, the same discipline of hunting for a hidden confounder before crediting one variable with causing the other.
Key Takeaways
- Correlation means two variables tend to move together; causation means one produces a change in the other.
- A high correlation alone is never sufficient proof of a causal relationship.
- Confounding variables, hidden third factors, can make two unrelated variables appear correlated.
- Controlled experiments with random assignment are the standard way to establish causation.
- Always ask whether a plausible confounder could explain an observed correlation before assuming causation.
Practice what you learned
1. What does it mean for two variables to be correlated?
2. What is a confounding variable?
3. What is generally required to establish that one variable causes a change in another?
4. The classic ice cream and drowning example illustrates what concept?
5. Why is observational data alone often insufficient to prove causation?
Was this page helpful?
You May Also Like
Hypothesis Testing
How the null and alternative hypothesis framework, p-values, and significance thresholds are used to judge whether an observed effect is likely real or chance.
What Is EDA
What exploratory data analysis is, why analysts use summary statistics and visuals to understand a dataset before modeling, and what issues it uncovers.
Charts & When to Use Them
How to match a chart type, such as bar, line, scatter, or pie, to the specific comparison, trend, or relationship the underlying data is meant to show.