Correlation measures the direction and strength of an association between two variables; regression models how an outcome changes in relation to one or more explanatory variables. Correlation summarizes a relationship without assigning the variables distinct roles, while regression specifies a response and predictors so it can estimate or predict the response.
What Correlation Measures
Correlation is a numerical summary of how two variables vary together. In introductory statistics, the most familiar measure is Pearson’s correlation coefficient, often written as r. It ranges from negative one to positive one: a value near positive one indicates a strong positive linear association, a value near negative one indicates a strong negative linear association, and a value near zero indicates little or no linear association.
“Linear” matters. Pearson correlation describes how closely observations follow a straight-line pattern. A curved relationship can be strong but still have a correlation near zero. Correlation is also sensitive to unusual observations, which can pull its value upward or downward. A scatterplot is therefore a useful companion to the coefficient: it can reveal curvature, clusters, outliers, or a changing spread that one number hides.
Correlation is symmetric: the correlation of X with Y is the same as the correlation of Y with X. It does not designate one variable as the cause, outcome, or predictor of the other. Nor does it establish that either variable causes the association.
What Regression Models
Regression describes an outcome, also called a response or dependent variable, in terms of one or more predictors, also called explanatory or independent variables. In simple linear regression, a fitted line is commonly written as Y = a + bX, where a is the intercept and b is the slope. The slope represents the model’s estimated change in Y associated with a one-unit increase in X.
That interpretation is conditional on the model and its assumptions. It is not automatically a causal effect. A regression can be used to describe a sample, estimate a relationship, or make predictions for relevant values. Prediction does not mean certainty: individual observations usually differ from the fitted line, and predictions outside the range of observed data are extrapolations that can be unreliable.
Regression can include multiple predictors. For example, a course researcher might model exam scores using attendance and prior preparation. A coefficient for attendance then describes its association with scores while holding the other included predictors fixed, according to the fitted model. This adjustment does not guarantee that all confounding has been removed; omitted variables, measurement problems, and poor design can still distort interpretation.
Correlation and Regression Side by Side
| Feature | Correlation | Regression |
|---|---|---|
| Main purpose | Summarize the direction and strength of association | Model an outcome using one or more predictors |
| Roles for variables | No distinction between the two variables | Separates response from predictor variables |
| Typical result | A coefficient, such as Pearson’s r | An equation, fitted values, and estimated coefficients |
| Units | Pearson correlation has no measurement units | Slopes have units tied to the outcome and predictor |
| Direction of calculation | Unchanged when variables are swapped | Usually changes when response and predictor are swapped |
| Common use | Describe association | Explain or predict an outcome under a stated model |
The table describes common introductory methods, not every possible technique. There are different correlation measures and many regression models, including models for binary outcomes or curved relationships. The right method depends on the data and research question.
A Concrete Correlation Example
Imagine a class records the number of practice problems students complete and their quiz scores. A positive correlation would mean that students with more completed practice problems tend, on average, to have higher scores in this sample. The coefficient alone does not tell how many score points correspond to another practice problem. It also cannot establish that practice caused the higher scores: students who practice more may differ in preparation, available time, or other relevant ways.
Correlation is useful here if the question is simply whether the two measurements tend to move together. A scatterplot helps check whether the pattern is roughly linear and whether a small number of students dominate the apparent association. If the plot shows a curve or separate groups, a single Pearson coefficient may give an incomplete summary.
A Concrete Regression Example
Using the same data, a simple regression could treat quiz score as the response and completed practice problems as the predictor. Suppose the fitted slope is positive. Its interpretation is that the model estimates a higher quiz score for students with more completed problems, by the slope’s amount per additional problem. The slope has units of score points per problem; unlike correlation, it depends on the measurement scales.
The fitted line also produces a predicted score for a specified number of practice problems. That prediction is an estimate, not a promise about an individual student. A researcher should examine residuals—the differences between observed and predicted scores—and consider whether the linear form is reasonable. If the data cover only a narrow range of practice amounts, using the line far beyond that range is especially questionable.
Why the Results Are Related but Not Identical
In simple linear regression with an intercept, Pearson’s correlation and the slope are connected. The slope equals the correlation multiplied by the ratio of the variables’ standard deviations. This helps explain why the correlation is unit-free while the slope changes when measurement units change. Converting quiz scores from one scale to another changes the numerical slope, but it does not change the underlying Pearson correlation, provided the conversion is a positive linear rescaling.
In that same simple setting, the square of the correlation equals the proportion of variation in the response accounted for by the fitted linear relationship in the sample. This quantity is commonly called the coefficient of determination, or R-squared. It does not mean that the predictor caused that share of the outcome, and it does not by itself show that a model is appropriate or useful for prediction.
With multiple regression, there is no single ordinary pairwise correlation that summarizes the whole model. The model’s R-squared describes the fit of all included predictors together, while individual regression coefficients describe associations conditional on the other predictors in the model. Those are different summaries and should not be treated as interchangeable.
How to Choose the Right Method
Start by stating what you want to know. If the goal is to describe whether two measured variables are associated, correlation may be an appropriate summary. If one variable is an outcome and you want an equation for estimating it from predictors, regression is the more direct framework.
- Use a scatterplot first when both variables are quantitative; inspect shape, clusters, spread, and unusual points.
- Choose a correlation measure that matches the data and question. Pearson’s measure targets linear association; rank-based measures can be useful for ordered data or some monotonic relationships.
- Use regression when the response and predictors have meaningful, explicitly stated roles, and choose a model suited to the response’s form.
- Check assumptions and residual patterns rather than relying only on a coefficient or a model-fit summary.
- Separate association, prediction, and causal claims. A causal question generally requires a suitable study design and additional assumptions, not merely a regression equation.
For coursework, follow the method and terminology your instructor expects, but explain the result in context. State which variable is the response, identify what a coefficient means in the original units, and avoid implying causation unless the design supports that conclusion.
Common Mistakes and Misconceptions
- “Correlation means causation.” It does not. A third variable, selection effects, or the direction of influence may help explain an observed association.
- “A correlation of zero means no relationship.” It means no linear association as measured by Pearson’s coefficient; a strong curved relationship can still exist.
- “Regression proves a cause.” Regression estimates relationships under a model. Causal interpretation requires defensible assumptions or a design that supports it.
- “A larger correlation means a larger practical effect.” Correlation is unit-free and does not express change in the outcome’s original units. Practical importance also depends on context.
- “A good fit guarantees good predictions.” Fit to existing data does not ensure performance on new data. Overfitting, extrapolation, and shifts in the population can undermine predictions.
Outliers deserve particular attention because they can strongly affect both correlation and regression. A point far from the rest may create an apparent trend or change a fitted slope substantially. Do not remove an observation merely because it is inconvenient; check for data errors, understand its context, and report the reasoning behind any decision to exclude it.
How to Remember the Difference
Think of correlation as a two-variable summary: “Do these variables move together, and in what direction?” Think of regression as a model with assigned roles: “How is this outcome related to these predictors, and what outcome does the model estimate?” The distinction is not that correlation is always descriptive and regression is always predictive—regression can also describe data—but that regression explicitly models a response from predictors.
Frequently Asked Questions
Can correlation be negative while a regression slope is positive?
In simple linear regression with an intercept using the same two quantitative variables, the slope and Pearson correlation have the same sign. Apparent disagreement usually means a different model, different variables or observations, or a different correlation measure is being compared.
Is correlation a type of regression?
No. They are related statistical tools, but correlation summarizes association without assigning variable roles, whereas regression specifies an outcome and predictors. Some simple linear regression results can be calculated from the same quantities used for correlation.
Can regression be used without correlation?
Yes. Regression can model outcomes with multiple predictors or responses that are not well described by ordinary linear methods. A pairwise correlation is not a requirement for fitting every kind of regression model.
Which one should I report in a research paper?
Report the statistic that answers the research question and follow the relevant course or field conventions. A study may report both when each adds useful information, but explain what each number means and do not present either as evidence of causation by itself.
Does a high correlation guarantee accurate predictions?
No. A high correlation describes a strong linear association in the data used to calculate it. Prediction accuracy also depends on the model, the range and quality of the data, and how well the relationship holds for new observations.
Bottom Line
Correlation tells you how strongly and in what direction two variables are associated, while regression gives you a model relating a designated outcome to predictors. Use correlation to summarize a relationship and regression to estimate or predict an outcome, while checking the data, assumptions, and limits of any causal interpretation.