Correlation is a statistical measure that shows whether two variables tend to move together and how strong that relationship is. It can be positive, negative, or absent, but it does not by itself show that one variable causes the other.
Correlation
What Is Correlation?
Correlation is a statistical measure of whether two variables tend to move together and how strong that relationship is. It can be positive, negative, or absent, but it does not show that one variable causes the other.
A variable is anything that differs across people, places, or time, such as temperature, income, study time, sales, or website visits. Correlation describes a pattern's direction and strength, not why it exists. For example, ice cream sales and sunburn cases may rise in warm months. Buying ice cream does not cause sunburn. Warm weather affects both.
How Correlation Works
Correlation analysis compares paired observations. Each row needs a value for both variables, such as one person's study hours and exam score.
- Define the two variables and question.
- Collect paired observations from the same people, products, locations, or time periods.
- Plot the data on a scatterplot to check for clusters, curves, outliers, and patterns.
- Select a correlation measure that fits the data.
- Calculate the coefficient, which summarizes direction and strength.
- Assess uncertainty with a confidence interval or statistical test when appropriate.
- Interpret the result using subject knowledge, data quality checks, and alternative explanations.
A scatterplot matters because one number can hide important details. Similar coefficients can reflect curved patterns or results driven by one unusual observation.
The Three Types of Correlation
The usual three types describe the direction of a linear relationship. “No correlation” more precisely means no clear linear correlation.
| Type | Plain-language meaning | Example |
|---|---|---|
| Positive correlation | As one variable increases, the other tends to increase. | Students who spend more time studying may tend to earn higher test scores. |
| Negative correlation | As one variable increases, the other tends to decrease. | As a product’s price rises, sales volume may tend to fall. |
| No linear correlation | There is no consistent straight-line pattern between the variables. | Shoe size and reading ability have no expected linear relationship. |
Understanding the Correlation Coefficient
The correlation coefficient, often written as r for Pearson correlation, ranges from -1 to +1. Its sign shows direction: positive values mean variables tend to rise or fall together, while negative values mean one tends to fall as the other rises. Its absolute size shows the strength of the linear pattern.
A value of +1 is a perfect positive linear relationship, -1 is perfect negative, and 0 means no linear relationship. Zero does not mean no relationship at all. A variable may rise and then fall in a curved pattern, producing a correlation near zero.
There is no universal threshold for a “medium correlation.” In medicine, a small association may matter if it affects many people. In engineering, a stronger relationship may be needed before changing a process. Practical importance depends on the decision, measurement quality, and consequences.
Pearson, Spearman, and Other Correlation Measures
Different correlation measures answer different questions. Choose the measure based on the data, not convenience.
| Measure | Best suited to | Useful note |
|---|---|---|
| Pearson correlation | Two numeric variables with an approximately linear relationship. | It is sensitive to extreme values, called outliers. |
| Spearman rank correlation | Ordered ranks or numeric data with a consistently increasing or decreasing pattern. | It uses ranks, so it can work better when the pattern is not perfectly linear. |
| Kendall rank correlation | Ranked data, especially smaller samples or datasets with many ties. | It compares whether pairs of observations are in the same order. |
| Category-based measures | Variables recorded as categories, such as device type and purchase status. | Measures such as phi or Cramér’s V assess association without treating categories as numbers. |
Correlation vs. Regression vs. Causation
These concepts are related but answer different questions.
| Concept | Main question | What it can support |
|---|---|---|
| Correlation | Do two variables move together? | A summary of direction and strength of association. |
| Regression | How is an outcome related to one or more predictors? | Prediction, adjustment for measured factors, and estimated relationships. |
| Causal inference | Would changing one factor change another? | Cause-and-effect conclusions when supported by a suitable study design. |
Regression can estimate or predict an outcome, such as website visits from advertising spend. But ordinary regression, like correlation, cannot prove causation alone. Strong causal claims usually need experiments, credible comparison groups, or carefully designed observational studies.
Why Correlation Does Not Prove Causation
Correlation may arise from a confounding variable, reverse causation, coincidence, or selection bias. A confounding variable influences both measures. For example, customers who use a support feature more often may also report lower satisfaction. That does not prove the feature causes dissatisfaction. Customers with difficult problems may both use support more and report lower satisfaction.
Reverse causation is also possible: higher sales may lead a business to increase advertising, rather than advertising being the only reason sales increased. A spurious correlation is misleading, accidental, or produced by another factor. Ask what else could influence both variables and whether the direction could be reversed.
How to Read a Correlation Result Responsibly
A careful interpretation combines the coefficient with evidence about the data and decision.
- Inspect a scatterplot before relying on a single coefficient.
- Report the coefficient, sample size, and variables compared.
- Use confidence intervals where possible to show the plausible range of the relationship.
- Understand that a p-value in a Pearson correlation test assesses whether the observed linear association would be surprising if the true correlation were zero. It does not measure importance or prove causation.
- Check for missing data, inconsistent definitions, duplicate records, and outliers.
- Judge practical significance in context, not only statistical significance.
- Use cautious language such as “was associated with” rather than “caused.”
Common Uses of Correlation Analysis
Correlation helps explore patterns, prioritize questions, and identify variables for further study.
- Scientists examine whether exposures, measurements, or outcomes vary together.
- Public health teams explore links between behaviors, environmental conditions, and health indicators.
- Businesses compare customer satisfaction, retention, spending, and service measures.
- Financial analysts examine how assets move in relation to one another.
- Educators study relationships between attendance, study habits, and learning outcomes.
- Product teams compare feature use, engagement, and retention, often alongside app design and development decisions.
- Quality teams monitor whether process settings move with defect rates or production output.
Benefits of Correlation
Used appropriately, correlation offers a fast, understandable way to learn from data.
- It summarizes a relationship with a compact measure.
- It identifies patterns that deserve deeper investigation.
- It can support forecasting models and feature selection.
- It helps teams monitor measures that change together.
- It provides a common language for discussing evidence, especially when paired with charts and definitions from a data and technology glossary.
Limits and Common Pitfalls
Correlation is informative, but weak data or questions can make it easy to misuse.
- Correlation does not establish causation.
- One or two outliers can greatly change Pearson correlation.
- A restricted range, such as studying only high-income customers, can hide a relationship present in the wider population.
- Aggregated data can mislead because group patterns may differ from individual patterns.
- Shared time trends can create false associations, such as two unrelated measures both rising over several years.
- Missing data can bias results if records are absent for systematic reasons.
- Testing many variable pairs increases the chance of finding an apparently meaningful result by chance.
A Practical Example of Correlation
Suppose a retailer compares weekly advertising spend with website visits over a year and finds a positive correlation of 0.62. A careful statement is: “Weeks with higher advertising spend were associated with more website visits in this dataset.” It would not say advertising caused the increase.
Seasonality, promotions, product launches, news coverage, and search demand could influence both spending and visits. Next steps might include comparing similar regions with different campaign levels, controlling for known seasonal factors in a regression model, or running a randomized advertising test. Correlation identifies a pattern worth investigating more rigorously.
Start Building
on Emergent today
on Emergent today
Start Building
