Correlation Does Not Equal Causation: A Business Owner's Guide

Your email open rate is 35%. Your revenue is $100k this month.

Next month, your email open rate is 38%. Your revenue is $120k.

Correlation: Email open rate and revenue both increased.

Do you conclude: “We should improve email open rates to increase revenue”?

Maybe. Or maybe not.

Here’s the problem: Both might have increased because of a third thing.

You launched a new product in month 2. The new product:

Attracts more engaged customers (they open more emails)
Generates more revenue (the whole product is new)

Email open rate didn’t cause revenue. The new product caused both.

This is correlation vs. causation.

What Is Correlation?

Correlation means two things move together.

When one goes up, the other goes up. When one goes down, the other goes down.

Examples:

As temperature increases, ice cream sales increase (correlation)
As you spend more on ads, revenue increases (correlation)
As customer age increases, purchase frequency increases (correlation)

Correlation is measurable and real. These things genuinely move together.

But correlation does NOT mean one causes the other.

Why Correlation Misleads

Example 1: The Ice Cream Paradox

Ice cream sales are highly correlated with drowning deaths.

In summer, both spike. In winter, both drop.

Does eating ice cream cause drowning? No.

What’s really happening? Warm weather causes both. People eat more ice cream in summer (because it’s hot). People swim more in summer (because it’s warm). More swimming = more drowning.

The correlation is real, but the causation is wrong.

Example 2: The Marketing Spend Problem

Your company spends more on ads in months when revenue is already high.

Q4 (holiday season): High revenue → You spend heavily on ads → Revenue goes up even more

You look at this and think: “Ads are driving revenue!”

But maybe causation is reversed: High revenue is driving high ad spend. You have more cash, so you invest more.

Or maybe a third thing (seasonality) is driving both: People buy more in Q4 regardless of your ads.

Example 3: The Feature Adoption Correlation

Your customers who use Feature X have 50% higher LTV than customers who don’t.

You conclude: “We should push Feature X on all customers to increase LTV.”

But maybe the causation is reversed: High-value customers use more features because they’re more invested in the product.

Feature X didn’t make them valuable; their existing investment in the product made them use more features.

Pushing Feature X on low-value customers might not help because they’re low-value for other reasons (wrong use case, price-sensitive, time-constrained).

The Three Possible Causal Relationships

When you see a correlation, there are only three ways causation could work:

1. A causes B

Example: You run ads (A) → people buy (B)
Intuitive and usually what we assume

2. B causes A

Example: People want to buy (B) → you increase ad spend to meet demand (A)
Less obvious but often real

3. C causes both A and B

Example: Holiday season (C) → more people shop (A) and you spend more on ads (B)
The confounding variable

Most people jump to assumption #1. But #2 and #3 are often the real answer.

How to Test for Causation

Testing for causation requires an experiment. You need to rule out alternative explanations.

Method 1: The Holdout Group (Randomized Test)

You have 1,000 customers. Randomly split:

500 get Feature X (treatment group)
500 don’t get Feature X (control group)

Measure LTV:

Treatment: $5,000 average LTV
Control: $3,300 average LTV
Difference: $1,700 (treatment is higher)

Conclusion: Feature X causes higher LTV (causation confirmed).

Why? Because you randomly assigned customers. There’s no way the “treatment” group was already higher-value. You assigned randomly.

The only thing different between the groups is Feature X. So if LTV is higher, Feature X caused it.

Method 2: The Time-Based Test

You run a campaign on Monday. Measure results Tuesday-Friday.

Before: 100 customers After: 120 customers

Did the campaign cause the increase?

Not necessarily. Maybe Tuesday is always higher traffic. Maybe you had a PR mention that brought traffic.

To control for this: Run the campaign on Monday of Week 1. Don’t run it on Monday of Week 2. Compare:

Week 1 Monday (campaign): 120 customers
Week 2 Monday (no campaign): 100 customers
Conclusion: Campaign caused +20 customers

By controlling for the day of the week and comparing across multiple weeks, you isolate the campaign’s effect.

Method 3: The Cohort Comparison

You have two groups of customers: those who adopted Feature X and those who didn’t.

To test causation, you need to compare apples to apples.

Compare only customers who:

Signed up in the same time period
In the same industry
With the same company size
Same product tier

Then compare LTV:

Matched cohort (Feature X users): $5,000
Matched cohort (non-Feature X users): $3,800

This better controls for confounding variables. The groups are more similar, so Feature X is more likely to be the driver of LTV difference.

The Reality: You Can’t Always Test for Causation

Testing causation is expensive and time-consuming. Most businesses don’t do it.

So how do you make decisions without proof of causation?

1. Start with the obvious direction

If you’re considering “should we spend more on ads,” the obvious direction is: Ad spend → Revenue.

It’s less likely (but possible) that Revenue → Ad spend decides itself.

Start by assuming the obvious direction unless evidence suggests otherwise.

2. Look for temporal ordering

Causation requires timing: The cause must happen before the effect.

If email open rate increased on Monday and revenue increased on Friday, email open rate could have caused it.

If revenue increased on Monday and email open rate increased on Friday, email open rate didn’t cause it (it came after).

3. Look for competing explanations

Before you assume correlation = causation, brainstorm other explanations.

Correlation: “Feature adoption rate went up. Revenue went up.”

Other explanations:

We raised prices (explains revenue, unrelated to feature adoption)
We acquired higher-value customers (explains both, independently)
Market seasonality (explains both, independently)
We improved customer service (explains both, independently)

Which explanation is most likely? That’s your answer.

4. Check the magnitude

If correlation is weak (0.1 correlation coefficient), causation is unlikely.

If correlation is strong (0.9 correlation coefficient), causation is more likely (though still not guaranteed).

5. Ask domain experts

If you’re unsure, ask someone who knows the business deeply:

“Can you think of any way Feature X wouldn’t cause higher LTV even if users of Feature X have higher LTV?”

If they say yes, you found the confounding variable. Investigate further.

The Practical Approach

Most business decisions don’t need proof of causation. They need sufficient evidence to act.

Framework:

See the correlation: Feature X users have 50% higher LTV
Brainstorm alternative explanations: Are they just better customers already? Does X come from customer value or cause it?
Look for evidence against alternatives: How many features do high-LTV customers use? (If they use 5 features and only Feature X is correlated, maybe Feature X caused it. If they use 5 features and all are correlated, they’re just power users—Feature X didn’t cause anything special)
Make a decision with uncertainty: “I think X drives LTV, but I’m not 100% sure. Let me test on new customers and see.”

You don’t need certainty. You need enough confidence to act.

The Takeaway

Correlation is real and useful for identifying patterns. But don’t jump to causation.

When you see two things move together, ask: “Could anything else explain this?”

Usually, the answer is yes.

The harder you look for alternative explanations, the better your decisions will be.

We help you think through correlations in your data and identify which ones likely reflect true causation vs. which ones are just coincidence.