Correlation vs. causation: What's the difference?

A chart can tell you that two numbers moved together. It cannot tell you what would have happened if you changed one of them.

That missing counterfactual is the difference between correlation and causation. Correlation summarizes an observed relationship. Causation answers an intervention question: if you deliberately changed X while holding the relevant alternative explanations in check, would Y change?

Product teams cross this boundary constantly. Activated users retain longer, customers who use integrations spend more, and accounts that adopt a new feature expand faster. Those patterns are useful. They are not yet proof that forcing the behavior will produce the outcome. High-intent customers may cause both adoption and retention, making the feature look more powerful than it is.

This guide explains how to read associations without dismissing them, how causal errors enter ordinary dashboards, and how randomized experiments turn a plausible story into evidence you can use for a decision.

Correlation describes association, not intervention

Correlation measures how variables vary together. A positive correlation means high values of one tend to accompany high values of the other. A negative correlation means high values of one accompany low values of the other. A value near zero indicates little linear association, although a strong nonlinear relationship may still exist.

A coefficient compresses a pattern

Pearson's correlation coefficient, usually written as r, ranges from -1 to 1. It captures the direction and strength of a linear relationship. It does not tell you the slope in business units, prove that the variables are independent when r is zero, or reveal whether extreme observations drive the pattern. The JMP statistics reference also stresses that pairwise correlation cannot account for an underlying third-variable relationship.

That is why a scatter plot should accompany the coefficient. The Khan Academy correlation lesson uses familiar examples to show that a relationship can be real in the data and still support a false causal story. Different datasets can share a correlation while having different shapes, clusters, and outliers.

For a product team, “feature usage and retention have r = 0.62” is an association statement. It says neither that increasing usage will improve retention nor how much retention would change. The coefficient may still be valuable for prediction, segmentation, instrumentation checks, or hypothesis generation.

Prediction can work without causation

A variable does not need to cause an outcome to predict it. Umbrella sales may predict rainy-day traffic even though umbrellas do not cause rain. A model can use device, session pattern, and account attributes to forecast churn without claiming that changing those attributes would prevent churn.

Problems arise when a team treats a predictive feature as an intervention target. A model might learn that customers who contact support frequently are more likely to renew because complex enterprise accounts both need support and have durable contracts. Reducing support contacts would not necessarily improve renewal. It might remove the help that keeps those customers successful.

Use predictive language for predictive evidence: “associated with,” “forecasts,” or “helps identify.” Reserve “drives,” “improves,” and “causes” for designs that justify an intervention claim.

Causation is defined against a counterfactual

Suppose an account receives a new onboarding flow and activates. The causal question is not whether activation occurred. It is whether that same account would have activated under the old flow at the same moment and under the same conditions. You can never observe both outcomes for one unit.

Causal designs construct a credible comparison group for that missing outcome. Randomized experiments do this by assigning similar units to control and treatment in expectation. Quasi-experimental methods do it through assumptions about cutoffs, timing, instruments, parallel trends, or matched comparison groups. In every case, the design carries the causal argument.

Build a testable causal claim

Learn how to define the population, treatment, exposure, metrics, and decision rule before an A/B test starts.

Watch the Webinar

Four mechanisms make correlation look causal

“Correlation is not causation” becomes useful only when you can name a plausible alternative explanation. Most misleading product correlations come from 4 mechanisms.

A confounder influences both variables

A confounding variable is a common cause of the apparent treatment and the outcome. Account maturity can increase both integration adoption and retention. Warm weather can increase both ice cream sales and swimming incidents. The observed variables move together because both sit downstream of something else.

The confounder may be measured, such as plan type, geography, or acquisition channel. It may also be unmeasured, such as motivation or organizational capability. Regression adjustment can address variables you measured correctly. It cannot automatically remove confounders that are missing, badly measured, or chosen after the outcome.

GrowthBook's existing guide to false causality and confounding explores how these hidden variables and Simpson's paradox can reverse an aggregate story.

Cause and effect run in the other direction

Heavy users often adopt more features. A dashboard may suggest that feature breadth creates engagement, while existing engagement actually creates feature discovery. Both directions may operate at once.

Time order helps: a cause must precede its effect. But sequence alone is not enough. A product launch preceding a revenue increase does not rule out seasonality, a promotion, or a competitor outage. Lagged correlations can narrow a theory without proving it.

Selection changes who appears in the data

Conditioning on a treatment-dependent event can manufacture an association. Imagine comparing revenue only among users who completed checkout. If a new flow changes who completes checkout, the selected treatment and control groups are no longer comparable. Revenue per purchaser might rise even while revenue per assigned user falls.

The same error occurs when analysts include only “active” users after a feature changes activity. Analyze the randomized population for the primary intent-to-treat estimate, or define activation before launch and use a method appropriate to the estimand.

Shared trends and chance create convincing patterns

Two metrics can rise because the product, traffic, and market are growing. Raw counts often correlate simply because both contain a time trend. Detrending, using rates with defensible denominators, checking seasonality, and comparing concurrent controls help expose this issue.

Chance becomes more dangerous as the number of metrics, segments, and time windows grows. If analysts search enough combinations, some will look impressive. Predeclared outcomes and multiple-comparison controls limit the freedom to draw a target around noise.

Randomization turns association into causal evidence

A randomized controlled experiment deliberately changes treatment and uses chance to assign eligible units. This breaks the systematic relationship between treatment and pre-existing characteristics in expectation.

Assignment must happen before treatment

Choose the unit that can independently receive the change: user, account, device, request, region, store, or another cluster. NIST describes a completely randomized design as one where treatment levels are randomly assigned to experimental units.

The unit must match the treatment. Randomizing teammates individually for a shared workspace feature creates interference because one person's experience can affect another's outcome. Randomizing accounts protects the boundary, but the analysis must then account for account-level assignment rather than treating every event as independent.

GrowthBook's guide to experimental units explains this connection between the assignment unit, independence, and valid uncertainty.

Exposure must represent the point of divergence

Assignment says what a unit would receive. Exposure records when the treatment could first affect behavior. Logging exposure at application startup may include users who never encounter the feature, diluting the estimate. Logging it after a treatment-dependent click selects on an outcome.

Place the event at the earliest point where variants diverge. Record the experiment key, variant, unit identifier, timestamp, and relevant implementation version. Then verify that exposure precedes measured outcomes and joins to the intended unit.

Outcomes and decision rules must be predeclared

Define one primary outcome, its numerator and denominator, attribution window, exclusions, minimum detectable effect, and stopping rule before opening results. Add secondary metrics for interpretation and guardrails for harms the primary metric might miss.

The CDC causal-inference teaching material describes randomized assignment as particularly strong evidence for causation. Digital experiments benefit from the same logic, but only if implementation and analysis preserve the assignment.

Randomization does not rescue a broken metric, missing exposure data, cross-variant contamination, or selective reporting. It provides a strong foundation whose assumptions still need verification.

Observational evidence can be useful without overclaiming

Some questions cannot be randomized. A team may need to evaluate a historical launch, a pricing change applied to all customers, a policy event, or a rare safety outcome. The correct response is not to discard the data. It is to match the claim strength to the design.

Use associations to generate and prioritize hypotheses

Correlations reveal where to investigate. If integration adoption is strongly associated with retention, ask what mechanism could connect them, which customer characteristics affect both, and whether adoption occurs before retention differences emerge. Interview customers, inspect sequences, and propose an intervention that changes adoption without changing everything else.

Then test the intervention. The experiment may show that guided integration setup changes retention, that it only helps a particular segment, or that adoption was merely a marker of committed customers. Any of those answers is more actionable than the original coefficient.

Stronger nonrandomized designs make assumptions visible

When an experiment is impossible, causal inference may use:

• Difference-in-differences: Compare outcome changes over time between affected and unaffected groups under a parallel-trends assumption.
• Regression discontinuity: Compare units just above and below a treatment cutoff when manipulation around the cutoff is limited.
• Instrumental variables: Use a source of treatment variation that affects the outcome only through treatment.
• Matching or weighting: Balance observed pre-treatment characteristics while acknowledging unmeasured confounding may remain.
• Interrupted time series: Look for a level or trend change at an intervention point while modeling seasonality and other events.

Each method replaces randomization with assumptions that need defense. A model's sophistication does not erase the design problem. The useful Hacker News discussion on causation beyond simple correlation reaches the same practical point: natural experiments and instruments can help when the variation approximates an independent intervention.

Triangulate when no single design can settle the question

Combine evidence with different failure modes. A product team might use observational sequences, customer interviews, a natural experiment from staggered rollout timing, and a smaller randomized nudge. Agreement across methods strengthens confidence because the same hidden bias is less likely to explain every result.

Do not collapse that into “proved.” Report the estimate, population, assumptions, uncertainty, and alternative explanations. Good causal communication tells the reader what design makes the claim credible and where it could still fail.

Audit causal language before it reaches a roadmap

The most practical safeguard is a short review that separates observation from intervention.

Ask what evidence the sentence actually contains

For any claim that X drives Y, ask:

1. Were units randomly assigned to X, or was X observed?
2. Did X occur before Y?
3. Could a third variable affect both?
4. Did selection into the analyzed sample depend on treatment or outcome?
5. Were the outcome and analysis chosen before results?
6. Does the estimate apply to the population and decision at hand?

If the design is observational, rewrite the sentence as an association unless a credible causal method and its assumptions are stated. “Accounts using 3 integrations retained at a higher rate” is accurate. “Adding 3 integrations improves retention” requires intervention evidence.

Check experiment health before reading the effect

For randomized tests, review allocation, pre-experiment balance, exposure coverage, identity joins, metric freshness, and concurrent changes. Sample ratio mismatch can signal assignment or tracking failures. Novelty, interference, and attrition can still make a short-term average misleading.

GrowthBook's experiment analysis workflow combines result interpretation with health checks, while warehouse-native experimentation lets data teams inspect the query logic behind metrics. Transparent SQL does not guarantee causality, but it makes the evidence chain auditable.

Make the decision proportional to uncertainty

A low-cost, reversible change can proceed on weaker evidence with monitoring. A pricing shift, safety-critical algorithm, or irreversible migration requires stronger identification, longer-term guardrails, and often replication. Practical significance matters alongside statistical evidence: a precisely estimated effect smaller than the cost of implementation is not a win.

The goal is not to eliminate uncertainty. It is to prevent an association from silently becoming a causal promise as it moves from dashboard to meeting to roadmap.

Treat correlation as a clue and causation as a design claim

Correlation is not the enemy of causal reasoning. It is often where causal reasoning starts. It identifies patterns, improves forecasts, and helps teams choose where an experiment could create the most learning.

The discipline is to keep 3 statements separate: what the data shows, what mechanism might explain it, and what design could test that mechanism. When those blur together, a compelling chart becomes an expensive product decision. When they remain distinct, observational data and randomized experiments reinforce each other.

Pick one important correlation in your product data. Draw at least 2 alternative causal diagrams, name the confounders and selection points, and design the smallest reversible intervention that distinguishes the stories. That is the move from “these metrics travel together” to “this change produced this outcome.”

Test the cause, not the story

Use GrowthBook to randomize product changes, measure trusted warehouse outcomes, and review the evidence behind the result.

Get Started