Blog/Correlation ≠ Causation

Statistics

Correlation ≠ Causation

Reading relationships correctly

SK

Skari Team

Skari

July 2026·10 min read

Correlation

r ≈ 0.9

A strong linear pattern — but a trend line alone never proves one variable causes the other.

Two things rise and fall together, so surely one drives the other? That leap is the most common — and most expensive — mistake in data analysis. A correlation says two variables move in step. It says nothing about why.

Note

Correlation is a description of a pattern. Causation is a claim about mechanism. Data alone can show the first; it can almost never prove the second on its own.

What a Correlation Coefficient Says

The correlation coefficient (r) runs from −1 to +1 and captures the strength and direction of a linear relationship — how tightly the points hug a straight line, like the scatter above.

rMeaning
+1Perfect positive — up together, exactly
0No linear relationship
−1Perfect negative — one up, one down
0.8Strong positive, but still just a pattern

Watch out

A high r is not evidence of cause, and a low r doesn't rule cause out — r only measures a straight-line pattern. A perfect curved relationship can even show r near zero.

Three Reasons Two Things Correlate

When A and B move together, causation is only one of several explanations — and rarely the safest bet.

ExplanationExample
A causes BMore exercise → lower resting heart rate
B causes A (reverse)Assumed 'ads → sales' when high sales fund more ads
A third thing causes both (confounder)Heat drives both ice cream and drownings
Pure coincidenceTwo unrelated trends that happen to line up

The confounder is the sneakiest. A hidden common cause makes two effects rise together with no direct link at all — the classic ice-cream-and-drownings trap.

Reverse Causation and Coincidence

  • Reverse causation: you have the arrow backwards — successful companies buy more software, not the other way around
  • Coincidence: with enough variables, some will correlate by pure chance — dredge a big dataset and spurious pairs appear
  • Selection effects: the way data was collected can manufacture a correlation that isn't real in the population

Tip

The more variables you scan for correlations, the more false ones you'll find. Deciding what relationship to test before looking is the cheapest defense against fooling yourself.

What It Takes to Claim Cause

Establishing causation needs more than a coefficient. The gold standard is a randomized controlled experiment; short of that, careful designs can build a strong case.

  1. 1A randomized experiment — randomly assign the treatment so confounders are balanced out.
  2. 2A plausible mechanism — a reason A would cause B, not just a pattern.
  3. 3Correct time order — the cause must precede the effect.
  4. 4Ruling out confounders — control for the obvious common causes, or match on them.

Studying Relationships in the SKARI Statistical Lab

SKARI's Statistical Lab measures relationships honestly — correlation for association, regression to control for other variables, and matching designs to strengthen a causal argument.

  • Correlation analysis with the coefficient and significance
  • Multiple regression to hold confounders constant while estimating an effect
  • Propensity Score Matching (PSM) and Regression Discontinuity (RDD) for stronger causal designs
  • Assumption checks so a linear coefficient isn't read past what it supports

Takeaway

You can move from 'these move together' toward 'this holds up when we control for the obvious alternatives' — the difference between a correlation and a defensible claim.

Frequently Asked Questions

Does a strong correlation ever imply cause?

Not by itself. It can motivate a causal hypothesis, but proof requires experimental design or careful control of confounders.

How do I spot a confounder?

Ask what could plausibly cause both variables. If a common driver exists, control for it in a regression or design around it.

Can regression prove causation?

Not on its own — but controlling for known confounders makes a causal story far more credible than a raw correlation.

Key Takeaways

r shows

Pattern

linear, −1 to 1

Not

Cause

just movement

Beware

Confound

hidden driver

Proof

Design

experiment/control

A correlation is a starting point, not a conclusion. Before acting on 'these move together,' rule out reverse causation and confounders — and remember that only a good design, not a big coefficient, earns the word 'cause.'

Takeaway

When two things move together, ask 'what else could explain this?' before 'which causes which?' — that one question prevents most costly misreads.

Hypothesis Testing Basics

Is the relationship even real?

Regression Analysis

Control for confounders

A/B Testing

The experiment that shows cause