A student emails us: "I found a correlation of r = .54 between study hours and GPA. My supervisor asked me to run a regression instead. What is the difference?" We get this question at least twice a week. Both techniques deal with relationships between variables, but they answer different questions, produce different output, and make different assumptions. Confusing them is one of the fastest ways to lose marks on your methodology chapter.
What Correlation Tells You
Correlation measures the strength and direction of a linear association between two variables. Nothing more. If you compute Pearson's r between weekly study hours and semester GPA for 120 nursing students and get r(118) = .54, p < .001, you know that students who study more tend to have higher GPAs. The coefficient also tells you the relationship is positive (both variables move in the same direction) and moderate-to-strong by Cohen's conventions, where .10 is small, .30 is medium, and .50 is large.
That is all correlation does. It does not tell you which variable drives the other. Swapping the two gives you the same r = .54. Study hours predicting GPA and GPA predicting study hours produce identical coefficients, because Pearson's r is symmetric. There is no predictor and no outcome. There is only a relationship.
You also have a choice of coefficient. Pearson's r requires both variables to be continuous and approximately normally distributed. When one or both variables are ordinal (like a 5-point Likert scale) or when the data are heavily skewed, Spearman's ρ (rho) is the right alternative. Spearman's works on ranks rather than raw values, so it captures monotonic relationships even when the association is not strictly linear. A study with 85 participants comparing self-reported stress (ordinal, 1–10 scale) with sleep quality (ordinal, Pittsburgh Sleep Quality Index categories) would use Spearman's, not Pearson's. (For a deeper comparison of these two coefficients, see our article on Pearson vs Spearman.)
What Regression Adds
Regression does something correlation cannot: it builds a prediction equation. Simple linear regression takes one predictor variable (X) and one outcome variable (Y) and fits the line Y = b0 + b1X. The intercept b0 is the predicted value of Y when X equals zero. The slope b1 tells you how many units Y changes for each one-unit increase in X.
Back to the study-hours example. A simple linear regression on those same 120 nursing students might produce: GPA = 1.82 + 0.09 × StudyHours. That equation means a student who studies zero hours per week is predicted to have a GPA of 1.82, and each additional hour of weekly study is associated with a 0.09-point increase in GPA. A student studying 20 hours per week would have a predicted GPA of 1.82 + (0.09 × 20) = 3.62. Correlation gave you a single number. Regression gives you a formula you can plug numbers into.
The distinction between predictor and outcome matters here. Swapping the variables produces a different equation with a different slope and different predictions. Regressing study hours on GPA is not the same model as regressing GPA on study hours. Regression forces you to commit to a direction, and your research question should dictate which variable is the predictor and which is the outcome.
Regression also scales up. When you have multiple predictors, you move to multiple linear regression: GPA = b0 + b1(StudyHours) + b2(AttendanceRate) + b3(SleepHours). Correlation cannot handle this. You could run three separate Pearson correlations, but that would miss the combined and adjusted effects of all three predictors acting together. In a multiple regression with 200 participants, you might find that study hours (β = .38, p < .001) and attendance (β = .27, p = .002) are significant predictors of GPA, while sleep hours (β = .08, p = .31) are not, after controlling for the other two. Correlation alone would never reveal that pattern.
When to Use Each Method
The choice depends on your research question. If you are asking "is there a relationship between X and Y?" then correlation is the right tool. The research question "Is there a significant correlation between anxiety levels and academic performance among 150 psychology undergraduates?" calls for Pearson's r (or Spearman's ρ if the data are not normally distributed). You report the coefficient, the p-value, and the degrees of freedom, and you are done.
If your question involves prediction, direction, or multiple variables, you need regression. "Do study hours, class attendance, and sleep duration predict GPA?" is a regression question. So is "How much does GPA change for each additional hour of study?" The word "predict" is the clearest signal. If your hypothesis uses it, regression is almost certainly what your supervisor expects.
There is a gray area. Some students phrase their research question as a correlation but their supervisor reads it as a regression. "What is the relationship between social media use and self-esteem?" could go either way. If all you want is to know whether the two are associated, correlation works. But if your literature review frames social media use as a factor that might reduce self-esteem, your supervisor may expect you to treat social media use as a predictor and self-esteem as an outcome. In cases like that, clarify the phrasing before you run anything. If you are unsure which statistical test fits your design, start there.
Correlation vs Regression: Side-by-Side Comparison
| Feature | Correlation | Regression |
|---|---|---|
| Purpose | Measure strength and direction of association | Predict one variable from one or more others |
| Directionality | Symmetric (no predictor/outcome distinction) | Asymmetric (predictor → outcome) |
| Output | Correlation coefficient (r or ρ) | Regression equation (Y = b0 + b1X), coefficients, R² |
| Number of predictors | Two variables only (bivariate) | One (simple) or many (multiple regression) |
| Effect size metric | r itself (or r² as shared variance) | R², adjusted R², standardized β |
| Research question example | "Are X and Y related?" | "Does X predict Y?" or "How much does Y change per unit of X?" |
| Causation | Does not imply causation | Does not imply causation either |
Common Mistakes We See
The most frequent error is running a correlation when the research question clearly asks about prediction. A student writes "this study examines whether parental income predicts children's educational attainment" and then reports a Pearson r = .42. The word "predicts" in the research question demands a regression model. The correlation coefficient answers a different, weaker question. In our experience reviewing over 300 student projects, this mismatch between research question and statistical method is one of the top reasons supervisors send a draft back for revision.
A second mistake is treating r² from a correlation as though it were R² from a regression. Squaring the Pearson r gives you the proportion of shared variance between two variables. In a simple bivariate case, this number equals the R² from a simple linear regression with the same two variables. But students sometimes square a Pearson r from a bivariate correlation and then interpret it using the language of regression ("the model explains 29% of the variance") without ever running a regression model. This is technically not wrong in the simple bivariate case, but it confuses readers and committee members, because you are borrowing the vocabulary of one method while reporting the output of another. Once you add a second predictor, R² from multiple regression and r² from any single correlation become entirely different numbers.
The third mistake is the causation trap, and it runs in both directions. Students know that "correlation does not imply causation." But some then assume that regression does imply causation, because it has a predictor and an outcome. It does not. A regression of ice cream sales on temperature will produce a statistically significant slope, but temperature does not cause ice cream purchases through any mechanism the regression tests. Causation requires experimental design, random assignment, and control of confounding variables. Neither correlation nor regression, on their own, establish cause and effect. (For more on what p-values from these tests actually mean, read our guide to interpreting p-values.)
A fourth issue is ignoring assumptions. Pearson's r assumes both variables are continuous, roughly normally distributed, and linearly related. Linear regression adds assumptions about the residuals: they should be normally distributed, have constant variance (homoscedasticity), and be independent. Running a Pearson correlation on two variables with a clear curvilinear relationship (say, the inverted-U between arousal and performance) will produce a misleadingly low r, because Pearson measures linear association only. With n = 60 and a true curvilinear pattern, you might get r = .12, p = .36, and conclude there is no relationship when there plainly is one.
How to Report Each in Your Thesis
For a Pearson correlation in APA-7 format, report the degrees of freedom, the coefficient, and the p-value: r(118) = .54, p < .001. If using Spearman, write rs(83) = .47, p < .001. Always report the exact p-value rather than just "p < .05," unless the value is below .001.
For simple linear regression, report the overall model fit and the coefficients: F(1, 118) = 47.12, p < .001, R² = .29. Then report the slope: b = 0.09, SE = 0.013, t(118) = 6.86, p < .001, β = .54. For multiple regression, report the overall F-test and R² first, followed by each predictor's unstandardized coefficient (b), standard error, t-value, p-value, and standardized coefficient (β). If you want publication-ready output without formatting each number by hand, Academic Stats Agent generates APA-7 compliant tables and text for both correlations and regressions.
Key takeaway: Correlation tells you whether two variables are related and how strongly. Regression tells you how much one variable changes when the other increases by one unit, and it lets you make predictions. If your research question asks about association, use correlation. If it asks about prediction or involves multiple predictors, use regression. Neither one proves causation.