Correlation and Regression in Psychology
Learning Objectives
- Explain what a correlation coefficient measures and interpret its sign and magnitude.
- Choose the appropriate correlation type (Pearson, Spearman, point-biserial) for a given pair of variables.
- Explain why correlation does not imply causation, with psychological examples.
- Interpret a simple linear regression equation and use it to make a prediction.
- Explain how multiple regression extends simple regression, and identify key limitations of both techniques.
Quick Answer
Correlation and regression are the statistical tools psychologists use to study relationships between variables, rather than just differences between groups. Correlation measures the strength and direction of a relationship — for example, how strongly self-esteem relates to academic performance — using a coefficient (r) that ranges from -1 to +1. Regression goes a step further: it builds an equation that predicts scores on one variable from one or more other variables, such as predicting anxiety levels from stress and sleep quality. Both techniques are everywhere in psychological research, from personality studies to clinical prediction models, but both come with a critical caveat: neither one, by itself, proves that one variable causes changes in the other.
Correlation: Measuring Relationships
A correlation coefficient (r) captures two things at once: direction and strength.
- r = +1: perfect positive relationship (as one variable rises, the other rises in lockstep).
- r = -1: perfect negative relationship (as one rises, the other falls in lockstep).
- r = 0: no linear relationship.
- In practice, psychology rarely sees perfect correlations; r = .10 is considered small, r = .30 medium, r = .50+ large (Cohen's conventions).
Worked Example: Self-Esteem and Academic Performance
| Self-Esteem Score | Exam Score (%) |
|---|---|
| 20 | 82 |
| 25 | 90 |
| 30 | 88 |
| 15 | 70 |
Calculating Pearson's r on this data yields approximately r = 0.75 — a strong positive correlation, meaning students with higher self-esteem scores in this sample tended to have higher exam scores. Importantly, this tells us the two variables move together; it says nothing about which one, if either, is causing the other to change.
Types of Correlation Coefficients
| Type | Used for | Example |
|---|---|---|
| Pearson's r | Two continuous, roughly normal variables | Height and weight |
| Spearman's rho | Two ordinal variables, or non-normal continuous data | Class rank and effort rating |
| Point-biserial | One continuous, one binary variable | Test score and pass/fail status |
| Tetrachoric | Two binary variables (assuming underlying continuous traits) | Presence/absence of two related symptoms |
Correlation Does Not Imply Causation
This is the single most tested concept in this chapter, so it deserves its own worked reasoning. Suppose researchers find a strong positive correlation between ice cream sales and rates of aggressive behavior in a city. Does eating ice cream cause aggression? Almost certainly not — a third variable, hot weather, likely increases both ice cream sales and irritability/aggression independently. This is called a confounding variable, and it's why experimental designs (with random assignment and manipulation) — not correlational designs — are needed to establish causation.
Three possible explanations for any correlation between X and Y:
- X causes Y
- Y causes X
- A third variable (Z) causes both X and Y
Correlational data alone cannot distinguish between these three possibilities.
Regression: From Relationship to Prediction
Simple linear regression predicts a dependent variable (Y) from a single independent variable (X) using the equation:
Ŷ = b0 + b1X
where b0 is the intercept (predicted Y when X = 0) and b1 is the slope (the change in Y for each one-unit increase in X).
Worked Example: Predicting Anxiety from Stress
Suppose a regression analysis on stress level (X) and anxiety score (Y) yields:
Ŷ = 10 + 0.5(Stress Level)
If a student reports a stress level of 20, predicted anxiety = 10 + 0.5(20) = 10 + 10 = 20.
Compare a student with stress level 40: predicted anxiety = 10 + 0.5(40) = 10 + 20 = 30. The slope (0.5) tells us that for every 1-point increase in stress, predicted anxiety rises by 0.5 points — a compact, useful summary of the relationship.
Multiple Linear Regression
Multiple regression extends this to several predictors at once:
Anxiety Score = b0 + b1(Stress Level) + b2(Sleep Quality)
Worked example: Anxiety Score = 10 + 0.5(Stress Level) − 0.3(Sleep Quality)
For a student with stress = 20 and sleep quality = 10: Anxiety = 10 + 0.5(20) − 0.3(10) = 10 + 10 − 3 = 17. Notice the negative coefficient for sleep quality means better sleep is associated with lower predicted anxiety, holding stress constant — multiple regression lets you isolate each predictor's unique contribution while controlling for the others.
Real-World Application
Admissions offices and clinical researchers both use regression heavily. A university might build a regression model predicting first-year GPA from high school GPA and standardized test scores, to inform admissions decisions. A clinical psychologist might use regression to see whether early trauma severity and social support both independently predict adult PTSD symptom severity — letting them identify which risk factors matter most, and by how much, once other factors are statistically controlled.
Limitations to Keep in Mind
- Correlation does not imply causation — always consider confounds and reverse causality.
- Outliers can dramatically inflate or deflate a correlation coefficient or distort a regression line.
- Linearity assumption: Pearson's r and linear regression only capture linear relationships; a strong curved (non-linear) relationship can produce a misleadingly low r.
- Restricted range: if a sample only includes people with similarly high scores (e.g., only honors students), correlations can appear artificially weak even if a strong relationship exists in the full population.
- Overfitting: adding too many predictors to a regression model can make it fit the specific sample very well but generalize poorly to new data.
Key Terms
| Term | Definition |
|---|---|
| Correlation coefficient (r) | A number from -1 to +1 indicating the strength and direction of a linear relationship |
| Confounding variable | A third variable that influences both variables in a correlation, creating a misleading association |
| Simple linear regression | A technique predicting one dependent variable from one independent variable |
| Multiple linear regression | A technique predicting one dependent variable from two or more independent variables |
| Slope (b1) | The predicted change in Y for each one-unit increase in X |
| Intercept (b0) | The predicted value of Y when X = 0 |
| Outlier | A data point far removed from the rest of the data, capable of distorting r or a regression line |
Common Mistakes
-
Misconception: "A strong correlation means one variable causes the other." Why it's wrong: Correlation only shows that two variables move together; it cannot rule out reverse causation or a confounding third variable. Correct: Causation requires controlled experimental designs with manipulation and random assignment, not correlational data alone.
-
Misconception: "A correlation of r = 0 means there's no relationship at all between the variables." Why it's wrong: Pearson's r only detects linear relationships. Two variables can have a strong curved (e.g., U-shaped) relationship and still produce r ≈ 0. Correct: Always inspect a scatterplot before concluding "no relationship" from a low r — a visual check reveals non-linear patterns a correlation coefficient would miss.
-
Misconception: "The regression equation can reliably predict outcomes far outside the range of the original data." Why it's wrong: Regression lines are only validated within the range of X values actually observed; relationships can change shape or direction outside that range. Correct: Restrict predictions to the range of data used to build the model, and treat extrapolation beyond that range with caution.
Comparison and Connections
| Feature | Correlation | Regression |
|---|---|---|
| Purpose | Describe strength/direction of a relationship | Predict one variable from another (or others) |
| Output | A single coefficient (r) | An equation with intercept and slope(s) |
| Directionality | Symmetric (r for X-Y is same as Y-X) | Asymmetric (predicting Y from X differs from predicting X from Y) |
| Number of variables | Exactly two | One dependent, one or more independent |
| Causation claim | Never implies causation | Never implies causation, even though it "predicts" |
Practice Questions
Recall
- What range of values can a Pearson correlation coefficient take, and what does each end of the range represent? Answer guidance: -1 to +1; -1 is a perfect negative relationship, +1 is a perfect positive relationship, 0 indicates no linear relationship.
- Write the general formula for simple linear regression and label each term. Answer guidance: Ŷ = b0 + b1X; Ŷ = predicted value, b0 = intercept, b1 = slope, X = predictor variable.
Understanding 3. Explain, using an example other than ice cream/aggression, how a confounding variable could produce a spurious correlation. Answer guidance: E.g., shoe size and reading ability correlate positively in children only because age drives both — older children have bigger feet and better reading skills; age is the confound. 4. Why is it important to look at a scatterplot before interpreting a Pearson correlation coefficient? Answer guidance: r only captures linear relationships; a scatterplot reveals non-linear patterns, outliers, or restricted range that a single correlation number would hide or misrepresent.
Application 5. A regression equation predicts exam anxiety as: Anxiety = 15 + 0.4(Caffeine intake, in cups). Predict anxiety for a student who drinks 3 cups of coffee. Answer guidance: Anxiety = 15 + 0.4(3) = 15 + 1.2 = 16.2. 6. Two variables, hours of exercise and reported mood, show a correlation of r = 0.6. A student claims "exercising more causes better mood." Evaluate this claim. Answer guidance: The correlation alone can't establish causation; reverse causation (better mood leads to more exercise) or a confound (e.g., general health, social engagement) could also explain the pattern. An experimental design is needed to test causation directly.
Analysis 7. A researcher reports a correlation of r = 0.05 between study strategy diversity and exam scores, and concludes "study strategy diversity doesn't matter." Critically evaluate this conclusion, considering possible limitations of correlational analysis. Answer guidance: A near-zero r could reflect a genuinely weak relationship, but it could also result from a non-linear relationship, restricted range in the sample (e.g., only high-achieving students), or measurement issues; the researcher should check the scatterplot and sample characteristics before concluding "no relationship." 8. Compare simple linear regression and multiple linear regression in terms of what each can tell a researcher about a predictor's contribution to an outcome. Answer guidance: Simple regression tells you the total relationship between one predictor and the outcome; multiple regression tells you each predictor's unique contribution to the outcome after statistically controlling for the other predictors in the model, which can reveal that an apparent relationship shrinks once a related variable is accounted for.
FAQ
If correlation doesn't prove causation, why do researchers bother calculating it? Correlation is still valuable for describing relationships, generating hypotheses, and making predictions — it's a necessary first step before designing an experiment to test causation directly.
What's a "good" correlation coefficient in psychology? By Cohen's widely used convention, around .10 is small, .30 is medium, and .50 or above is considered large — though what counts as "good" also depends on the specific research area and what's being measured.
Can regression be used with categorical predictors? Yes — categorical predictors (like treatment group) can be included in regression using coding schemes (e.g., dummy coding), extending regression's usefulness well beyond purely continuous variables.
What does it mean if my regression's slope is negative? A negative slope means the predicted outcome decreases as the predictor increases — for example, more sleep predicting lower anxiety scores.
Is Spearman's rho just a "backup" for when Pearson's r doesn't work? Not exactly a backup — it's the appropriate primary choice when data are ordinal or when the relationship is monotonic but not strictly linear, since it's based on ranked data rather than raw scores.
Quick Revision
- Correlation coefficient (r) ranges from -1 to +1, capturing direction and strength of a linear relationship.
- Pearson's r: two continuous variables. Spearman's rho: ordinal or non-normal data. Point-biserial: one continuous, one binary variable.
- Correlation never implies causation — always consider reverse causation and confounding variables.
- Simple linear regression: Ŷ = b0 + b1X, predicting Y from one predictor X.
- Multiple linear regression adds more predictors, letting you assess each one's unique contribution.
- Slope (b1) = predicted change in Y per one-unit increase in X; intercept (b0) = predicted Y when X = 0.
- Outliers and restricted range can distort correlation and regression results.
- Regression predictions should stay within the range of the original data; extrapolation is risky.
- A low or zero correlation doesn't rule out a relationship — check for non-linearity via a scatterplot.
Related Topics
Prerequisites: Descriptive Statistics, Inferential Statistics.
Related: Hypothesis Testing, Data Interpretation.
Next: Data Interpretation, which ties descriptive statistics, inferential statistics, and correlation/regression together into a complete research-reading skill set.