What is the difference between correlation and causation, and why does the distinction matter?

How to think about it

No: correlation does not establish causation. Correlation says that two variables move together; causation says that deliberately changing one variable would change the other, under a specified intervention.

That difference matters because a prediction can be useful without telling you what will happen when you intervene. A retailer may observe that higher advertising spend coincides with higher sales. Cutting or doubling the advertising budget is a different question.

Why correlation is not causation

Correlation is an observed pattern in data. The common version, Pearson correlation, measures how strongly two variables follow a straight-line relationship. Its value, r, ranges from -1 to 1. An r of 0.9 means a strong positive linear association. It does not mean that advertising caused 90 percent of sales, or that increasing advertising will increase sales by 90 percent.

Causation is about an intervention. The causal question is not merely, “Do weeks with more advertising also have more sales?” It is, “What would sales have been this week if we had spent $10,000 more on advertising, while keeping the relevant circumstances comparable?”

That second question involves a counterfactual: the outcome that would have happened under an alternative action. For a customer, you might write Y(1) for the purchase outcome if the customer sees an ad and Y(0) for the outcome if the same customer does not see it. You cannot observe both outcomes for the same customer at the same time. Experiments and causal methods try to estimate the average difference between those two worlds.

Several different mechanisms can create the same observed correlation:

• Direct causation: advertising increases purchases.
• Reverse causation: expected sales cause the marketing team to increase advertising.
• Confounding: a third variable causes both advertising and sales.
• Selection or coincidence: the way data was collected, or random chance, creates an association.

A confounder is a variable that influences both the apparent cause and the outcome. In the advertising example, holiday demand can increase both ad spending and sales. The ads may help, but the raw relationship cannot tell you how much.

Warning. A large correlation, a small p-value, or a high model fit does not remove confounding. Statistical significance answers whether an observed pattern is surprising under a particular null model. It does not answer whether an intervention will work.

A concrete example: advertising and sales

Suppose an online retailer records the following eight weeks. The demand signal represents the underlying shopping demand in the market. In practice, it may be only partly observed through holidays, search volume, weather, or prior sales.

Week	Demand signal	Ad spend in $ thousands	Sales in $ thousands
1	1.0	10	100
2	1.1	11	102
3	1.8	24	145
4	2.0	27	160
5	1.2	13	105
6	1.0	12	101
7	1.7	23	140
8	1.3	15	110

The scatter plot of ad spend and sales slopes sharply upward. A naive analysis might conclude that the advertising budget caused the sales increase.

But the causal graph could be:

demand → ad spend

and

demand → sales

When demand is high, the retailer buys more ads because customers are already shopping more. The same demand also produces more sales. Even if the ads had no effect at all, the data could still show a strong positive correlation.

Reverse causation can make the problem worse. If the retailer sees strong early sales on Monday, it may increase ad spend on Tuesday. In that case, earlier sales help cause later advertising. A model that treats advertising as the original driver can assign the direction incorrectly.

The business consequence is concrete. Finance may credit every extra dollar of sales to advertising and approve a larger budget. The next month, when demand returns to normal, the extra ads appear to “stop working.” The model did not necessarily fail mathematically. It answered an observational question while the business acted as if it had answered a causal one.

How to establish causation

The cleanest method is a randomized controlled trial, often implemented as a randomized A/B test. Randomly assign comparable customers to an ad group or a holdout group during the same period.

For example:

• 10,000 customers receive the campaign.
• 10,000 comparable customers are randomly withheld from it.
• The ad group produces 600 purchases.
• The holdout group produces 500 purchases.

The purchase rates are 6 percent and 5 percent. The estimated absolute lift is 1 percentage point, or 100 additional purchases in the tested population. The relative lift is 20 percent because 1 percentage point is one-fifth of the original 5 percent rate.

Randomization works because, on average, demand, customer intent, geography, device mix, and other causes are balanced between groups. The groups differ mainly in treatment assignment. That makes the difference in outcomes interpretable as the effect of the campaign, assuming the test was delivered correctly.

The test still needs careful design. Customers may see ads through another channel, people in the treatment group may influence people in the holdout group, or the ad platform may fail to deliver the assigned treatment. If those problems matter, randomizing by household, store, or geographic region may be safer than randomizing individual impressions. Analysts should also report the effect of assignment, known as the intent-to-treat effect, rather than quietly excluding customers who did not comply.

When an RCT is impossible, analysts use quasi-experimental designs. A natural experiment uses an outside event that creates treatment-like variation, such as a policy change, a lottery, or a strict eligibility cutoff. Difference-in-differences compares how a treated group changes relative to a similar untreated group before and after an event. These methods can be powerful, but they rely on assumptions such as comparable pre-treatment trends.

Regression adjustment, matching, and weighting can also reduce confounding when the important confounders are measured. For the retailer, controlling for day of week, holiday status, region, and demand indicators may produce a better estimate. It does not magically create randomization. If an important driver of both spend and sales is missing or measured badly, the estimate can still be biased.

The senior-level nuance

Correlation is not useless. If the goal is forecasting next week’s sales, ad spend may be an excellent predictive feature even when it is not causal. A forecasting model cares whether the feature improves predictions. A budget decision cares whether changing the feature changes the outcome. Those are different objectives.

Do not blindly “control for everything,” either. A causal graph, often represented as a DAG or directed acyclic graph, helps distinguish useful controls from harmful ones. In the retailer example, demand is a confounder because it points to both ad spend and sales. Adjusting for a reliable demand measure can block that non-causal path.

A mediator is a variable through which treatment works. If advertising increases clicks and clicks increase purchases, controlling for clicks estimates only the effect not operating through clicks. It removes part of the total advertising effect.

A collider is a variable caused by two other variables. Conditioning on a collider can create a false association. The general lesson is that a regression coefficient is not automatically a causal effect just because several variables appear in the model.

Time order helps, but it is not enough. Granger causality is a time-series test asking whether past values of X improve predictions of future Y beyond the history of Y. It can show predictive precedence. It cannot rule out a third variable that drives both series, so it is not proof that intervening on X will change Y.

A valid causal estimate also has a scope. An experiment showing that an extra $10,000 of advertising works during the December shopping season may not tell you what the same budget will do in February, in another country, or for a new customer segment. Treatment effects can differ by customer, channel, and demand level.

Failure mode. The first symptom of confusing correlation with causation is often a dashboard that reports strong return on ad spend, followed by a randomized holdout showing almost no incremental lift. Another warning sign is that the estimated advertising effect collapses after adding holiday controls. That usually means the original estimate was partly measuring demand timing rather than advertising impact.

What they’ll ask next

“Does a strong correlation ever prove causation?”

Almost never from observational correlation alone. Causation becomes credible when there is random assignment, a defensible natural experiment, or a clearly stated causal model whose assumptions are supported. The strength of the correlation is not the deciding factor.

“Can regression control for confounding?”

Sometimes. Regression adjustment can estimate a causal effect when the relevant confounders are measured well, treatment groups have enough overlap, the model is appropriate, and no important unmeasured confounder remains. Saying “we controlled for age and region” is not enough if customer intent, seasonality, or targeting rules also affect both treatment and outcome.

“Is Granger causality the same as real causality?”

No. Granger causality is about whether past values improve prediction. It is useful for time-series investigation, but common trends, feedback loops, and omitted variables can produce the result without a true intervention effect.

Say this in the interview

“Correlation is an observed co-movement, while causation is the change produced by an intervention; because confounding and reverse causation can create the same correlation, I would use randomization or a defensible quasi-experimental design before acting on it.”