Resources / 7: Research Methods & Statistics / Correlation and Regression

Maya draws a bell curve on a whiteboard while the bunny sits above the peak

Correlation and Regression

7: Research Methods & Statistics

Study guide by Anders Chan, PsyD · Updated

Why Correlation and Regression Matter for Your Psychology Career

Let's be honest – statistics can feel like the least interesting part of psychology. You're probably more excited about helping people than calculating coefficients. But here's the reality: understanding correlation and regression is essential for being a competent psychologist. These tools help you evaluate whether a new therapy actually works, determine if a psychological test is worth using, and understand research that guides your clinical decisions. On the EPPP, this material shows up repeatedly, and in your career, it helps you distinguish between solid evidence and empty claims.

Think of correlation and regression as your research detective toolkit. They help you answer questions like: Does therapy attendance predict symptom improvement? Can childhood trauma scores help estimate adult anxiety levels? Do personality test results relate to job performance? These aren't just academic exercises – they're the foundation of evidence-based practice.

The Foundation: What Correlation Actually Tells Us

Correlation measures the relationship between two or more variables. In simple terms, it tells us whether two things tend to move together, move in opposite directions, or have no pattern at all.

Here's the key terminology shift: instead of calling something an "independent variable" like we do in experiments, correlation research uses the term predictor (or X variable). Instead of "dependent variable," we say criterion (or Y variable). Think of it like using your current stress levels (predictor) to estimate how well you'll sleep tonight (criterion).

Most correlation coefficients range from -1.0 to +1.0, and they're usually symbolized with the letter "r."

  • A perfect positive correlation (+1.0) means as one variable increases, the other increases in perfect lockstep
  • A perfect negative correlation (-1.0) means as one variable increases, the other decreases in perfect lockstep
  • Zero Pearson correlation (r = 0.0) means there is no linear association. A nonlinear relationship may still be present (Schober et al., 2018).

Imagine tracking your coffee consumption and your productivity. A positive correlation would mean more coffee = more productivity. A negative correlation would mean more coffee = less productivity (maybe you get too jittery). Zero Pearson correlation would mean no linear association; coffee intake and productivity could still follow a curve (Schober et al., 2018).

Reading the Subscripts: A Critical Detail

Here's something students often miss: the subscript on a correlation coefficient tells you what you're looking at.

When you see two different letters or numbers (like r_xy), you're looking at the relationship between two different variables. This could be:

  • A simple correlation between variables
  • A criterion-related validity coefficient
  • A factor loading

When you see two of the same letters or numbers (like r_xx), you're looking at a reliability coefficient – which is a different topic entirely.

Checks for Correlation and Regression

In plain words: before you trust a correlation, run three quick checks. Do the dots on the scatterplot follow a straight line? Did your sample cover the full range of scores? And if you plan to use regression, do the dots scatter about the same amount along the whole line? A few terms first. Error variance is how far the real scores land from the predicted line. Inference means testing significance or building confidence intervals from your data. Reliability is how consistent a measure is from one use to the next.

Pearson r describes a linear association. A curve can be missed even when r is near zero. Restricting the sample range can change the coefficient and limit generalization. Equal error variance, called homoscedasticity, is an assumption of conventional linear regression inference; it is not required merely to calculate a descriptive correlation. A wider spread of outcomes at one predictor level signals heteroscedasticity. Equal reliability of the two measures is not a prerequisite (Bewick et al., 2003).

Linear Relationships

Pearson r summarizes a linear association: a pattern that follows a straight line. You can calculate it for curved data, but it can miss a strong nonlinear relationship (Bewick et al., 2003).

Consider the relationship between anxiety and performance. At very low anxiety, performance might be poor (you're not motivated). At moderate anxiety, performance peaks (you're engaged). At very high anxiety, performance drops again (you're overwhelmed). This creates a curved, inverted-U relationship. If you calculate a standard correlation coefficient on this curved relationship, you'll underestimate how strongly anxiety and performance are actually related.

Sample Range and Generalization

Check whether the sampled range matches the population you intend to describe. A restricted sample can change a correlation and limit generalization (Bewick et al., 2003).

Imagine you're studying the relationship between therapy experience and treatment effectiveness, but you only survey therapists with 15-20 years of experience. You've restricted your range – you're missing the beginners and the true veterans. Your correlation might show almost no relationship between experience and effectiveness, but that's only because you're looking at a narrow window where everyone has similar experience.

A correlation calculated only among hired applicants need not describe all applicants. Range restriction can attenuate a relationship, but its effect depends on the selection process (Bewick et al., 2003).

Homoscedasticity in Regression

In plain words: picture the band of dots around a regression line. Homoscedasticity means the band is about equally wide all along the line. Heteroscedasticity means the band changes width, like a fan that is narrow at one end and wide at the other. The usual regression standard errors, significance tests, and intervals assume an even band. You can still calculate r as a description when the band is uneven. You just should not trust the usual regression tests and intervals until you deal with the uneven spread.

Suppose you're predicting client satisfaction scores from number of therapy sessions. If clients who attend few sessions have wildly varying satisfaction (some love it, some hate it), but clients who attend many sessions all have consistently high satisfaction, you've violated homoscedasticity. Your prediction accuracy will be different depending on whether you're predicting for low-attendance or high-attendance clients.

Choosing the Right Correlation Coefficient

Not all correlation coefficients are created equal. You choose based on your measurement scales:

Correlation TypeWhen to Use ItExample
Pearson rBoth variables are continuous (interval/ratio) and linearly relatedRelationship between depression scores and sleep hours
EtaBoth variables are continuous but relationship might be nonlinearRelationship between age and cognitive flexibility (might be curved)
Spearman rhoBoth variables are ranksRelationship between class rank and internship preference rank
Point biserialOne variable continuous, one is a true dichotomyRelationship between anxiety scores and biological sex
BiserialOne variable continuous, one is an artificial dichotomyRelationship between therapy hours and pass/fail licensing exam status
Contingency coefficientBoth variables are nominal (categories)Relationship between diagnosis type and treatment modality chosen

Understanding Dichotomies

The distinction between true and artificial dichotomies trips up many students.

A true dichotomy is a variable that naturally exists in only two categories. Pregnancy status is truly binary – you either are or aren't pregnant. There's no underlying continuum being artificially split.

An artificial dichotomy happens when you take a continuous variable and split it into two groups. Exam scores are continuous (0-100), but when you apply a cutoff and create "pass" versus "fail" categories, you've artificially dichotomized a continuous variable.

Why does this matter? The biserial coefficient attempts to estimate what the correlation would have been if you hadn't artificially split that continuous variable. The point biserial coefficient treats the dichotomy as naturally occurring.

Worked example: Pick the coefficient for three studies.

  1. Study A correlates anxiety scores (continuous) with whether each client was assigned to treatment or to a waitlist. Assignment has only two real categories, and no hidden scale sits underneath it. That is a true dichotomy, so use point biserial.
  2. Study B correlates anxiety scores with "high" versus "low" social support. The groups were made by splitting a social support scale at the median. A continuous scale was cut in two, so that is an artificial dichotomy. Use biserial.
  3. Study C has two supervisors each rank the same 10 trainees from best to worst. Both variables are ranks, so use Spearman rho.

So on the exam: words like "cutoff," "median split," "pass/fail," or "high versus low" signal an artificial dichotomy and point you to biserial.

The Coefficient of Determination: Understanding Shared Variance

Here's where correlation gets practical. You can square any correlation coefficient to get the coefficient of determination, which tells you the percentage of variability in one variable that's explained by the other variable.

Let's make this concrete: Suppose you find that the correlation between therapeutic alliance scores and symptom improvement is r = .70. Square that (.70 × .70 = .49), and you discover that 49% of the variability in symptom improvement is explained by differences in therapeutic alliance quality. The remaining 51% is due to other factors – client motivation, life stressors, treatment type, medication, social support, and countless other influences.

This is incredibly useful for setting realistic expectations. When a colleague tells you they found a "significant" correlation between a new assessment tool and treatment outcomes, ask for the coefficient. If r = .30, that means only 9% of the variance is explained (.30 × .30 = .09). That's statistically significant perhaps, but clinically, 91% of the variance is still unexplained.

Worked example: Which is stronger, r = -.60 or r = +.40?

  1. Set the sign aside. The sign only tells you direction. Strength is the distance from zero, so compare .60 with .40.
  2. Square each one. -.60 × -.60 = .36. And .40 × .40 = .16. (Squaring always gives a positive number.)
  3. Turn each into a percent. The first pair of variables shares 36% of their variance. The second pair shares 16%.
  4. So r = -.60 is the stronger relationship. It explains more than twice as much variance.
  5. One more check. Doubling r does not double the variance explained. r = .30 gives 9%, but r = .60 gives 36%, which is four times as much.

So on the exam: a negative r is not a weak r. Ignore the sign, square the number, and compare the percents.

Regression Analysis: From Correlation to Prediction

Once you've established that two variables are correlated, regression analysis lets you make predictions. The correlational study produces a regression equation – essentially a formula that takes someone's predictor score and estimates their criterion score.

If you know that therapy attendance correlates with symptom reduction, regression lets you estimate: "A client who attends 12 sessions will likely show approximately X improvement in their depression scores".

The crucial point: accuracy of prediction increases as the correlation increases. With r = .90, your predictions will be quite accurate. With r = .30, your predictions will be rough estimates at best.

This is why employment tests with low validity coefficients can still be problematic – they're not giving you much predictive power, even if they're "statistically significant."

Reading a Regression Equation

A simple regression equation looks like this: predicted Y = a + bX. The intercept (a) is the predicted score when X is zero. The slope (b) is how much the predicted score changes for each 1-point rise in X.

Worked example: A clinic finds that predicted depression score = 30 - 1.5 × (sessions attended).

  1. A client attends 10 sessions. Multiply: 1.5 × 10 = 15.
  2. Subtract from the intercept: 30 - 15 = 15. The predicted depression score is 15.
  3. Another client attends 12 sessions: 1.5 × 12 = 18, and 30 - 18 = 12. The predicted score is 12.
  4. Each extra session lowers the predicted score by 1.5 points. The slope is negative, so the correlation between sessions and depression is negative too. In simple regression, the slope and r always share the same sign.

So on the exam: a predicted score is the best guess for people with that X score, not a promise for any one client.

How Big Are the Prediction Misses?

The standard error of estimate tells you the typical size of a prediction miss. It uses the standard deviation (SD) of the criterion and the correlation: standard error of estimate = SD of Y × the square root of (1 - r²).

Worked example: The criterion has an SD of 10 points.

  1. If r = .60, then r² = .36. Next, 1 - .36 = .64. The square root of .64 is .8. Multiply: 10 × .8 = 8 points.
  2. If r = .80, then r² = .64. Next, 1 - .64 = .36. The square root of .36 is .6. Multiply: 10 × .6 = 6 points.
  3. If r = 0, then 1 - 0 = 1, and the square root of 1 is 1. The miss is 10 points, the same as the SD. The predictor did not help at all.
  4. If r = 1.0, then 1 - 1 = 0, and the miss is 0. Prediction is perfect.

So on the exam: as r goes up, the standard error of estimate goes down. It ranges from 0 (perfect r) up to the full SD of Y (r of zero). One single error size only fits everyone when homoscedasticity holds.

Going Multivariate: Using Multiple Predictors

Real-world outcomes rarely depend on just one factor. Multivariate techniques let you use multiple predictors simultaneously, which mirrors the complexity of actual psychological phenomena.

Multiple Regression: Multiple Predictors, One Outcome

Multiple regression is used when you have two or more predictors and one continuous criterion variable.

You might want to predict therapy outcome (criterion) using initial symptom severity, therapeutic alliance quality, homework completion rate, and social support level (four predictors). Multiple regression creates an equation that weighs each predictor optimally.

There are two main approaches:

Simultaneous (standard) multiple regression enters all predictors into the equation at once. This approach asks: "Considering everything together, what's the best prediction?"

Stepwise multiple regression adds predictors one at a time, identifying which ones add meaningful predictive power. This approach asks: "What's the minimum set of predictors I need for good prediction?"

The Multicollinearity Problem

Here's an important concept: ideally, each predictor should correlate highly with the criterion but show low correlations with other predictors. This ensures each predictor provides unique information.

Imagine you're trying to predict client engagement, and your predictors include "sessions attended," "appointments kept," and "no-shows." These predictors are measuring almost the same thing – they're highly intercorrelated. This is called multicollinearity, and it creates problems because the predictors aren't giving you unique information.

Good multiple regression uses predictors that each capture something different about the criterion.

Other Multivariate Techniques

Canonical correlation extends multiple regression to situations with multiple continuous predictors AND multiple continuous criteria. You might use several personality measures to predict multiple indicators of job performance simultaneously.

Discriminant function analysis is used when your criterion is categorical (nominal) rather than continuous. You might use therapy engagement variables, symptom measures, and demographic information to predict which treatment group a client should be assigned to: brief therapy, standard therapy, or intensive therapy.

Logistic regression serves the same purpose as discriminant function analysis but works when the statistical assumptions for discriminant function analysis aren't met (like when your predictor variables aren't normally distributed).

Worked example: Pick the technique. Ask two questions: how many criteria are there, and what kind are they?

  1. A program predicts first-year GPA from entrance test score and undergraduate GPA. There is one criterion, and it is continuous. Use multiple regression.
  2. A study uses three personality scales to predict both supervisor ratings and sales totals at once. There are two continuous criteria. Use canonical correlation.
  3. A clinic uses several normally distributed symptom scores to predict which of three treatment tracks a client ends up in. The criterion is a category. Use discriminant function analysis.
  4. Same question, but the predictors are skewed and some are yes/no items. The assumptions for discriminant function analysis are not met. Use logistic regression.

So on the exam: a category as the outcome points to discriminant function analysis or logistic regression. Messy or non-normal predictors tip it to logistic regression.

Structural Equation Modeling: The Advanced Toolkit

Structural Equation Modeling (SEM) represents the sophisticated end of correlational techniques. While you won't need to perform SEM for the EPPP, you should understand what it does because you'll encounter it in research articles.

In plain words: factor analysis finds groups of items that rise and fall together because they seem to reflect the same hidden construct. SEM pairs that with regression. So it can ask how hidden constructs predict each other, not just how single scores do.

SEM combines factor analysis with multiple regression to test complex theories about how variables relate to each other. It's particularly powerful because it handles both observed variables (things you directly measure) and latent variables (underlying constructs you can't directly observe).

Observed variables (also called manifest variables or indicators) are the actual measurements you collect. If you're studying depression, observed variables might include specific items from a depression scale: "I felt sad most days," "I had trouble sleeping," "I lost interest in activities".

Latent variables (also called factors or constructs) are the underlying psychological constructs these measurements represent. Depression itself is a latent variable – you can't directly observe "depression," but you infer it from the observed symptoms.

Exogenous and Endogenous Variables

Variables in SEM are classified as either exogenous or endogenous:

Exogenous variables aren't predicted or explained by other variables in your model. They're like the starting points in your theoretical map – maybe childhood trauma or genetic predisposition.

Endogenous variables are predicted or explained by other variables in the model. These are the outcomes or mediators – maybe current anxiety symptoms or coping strategies.

The Five Steps of SEM

While you won't perform these steps yourself, understanding the process helps you evaluate SEM research:

Step 1: Model Specification – Researchers specify their theoretical model, drawing a path diagram showing how variables relate. Observed variables appear in squares/rectangles, latent variables appear in circles/ovals, and arrows show relationships.

Step 2 & 3: Model Identification and Estimation – These are mathematically complex steps that you don't need to understand in detail for the EPPP.

Step 4: Model Evaluation – Researchers use statistics like the goodness-of-fit index (GFI) to determine whether their theoretical model fits the actual data well.

Step 5: Model Modification – If the fit is poor, researchers revise their model and test again.

In plain words, "fit" means how closely the pattern of relationships the model predicts matches the pattern in the real data. Good fit supports the model. It does not prove the model is the only one that could fit.

Think of SEM like testing whether your theory's roadmap matches the actual terrain. You propose how various psychological factors connect (your map), collect data (observe the terrain), and see whether your map accurately represents reality.

Confidence Intervals vs. Prediction Intervals

In plain words: a confidence interval answers "where is the average for people with this score?" A prediction interval answers "where will this one new person land?"

A regression line estimates an average outcome at a given predictor value. A confidence interval around that estimated mean shows uncertainty about the mean. A prediction interval gives a range for one new person's outcome.

Think of predicting the average finish time of a running club versus the finish time of one new runner. The new runner adds another source of uncertainty.

At the same predictor value and confidence level, a prediction interval is wider than the confidence interval for the mean response. It includes both uncertainty in the fitted mean and the scatter of individual outcomes around it.

For example, predicting one student's score needs a prediction interval. Estimating the average score for students with that study time needs a confidence interval for the mean. Neither interval proves that study time caused the score.

Worked example: A clinic's regression predicts depression score from sessions attended.

  1. Question A: "What is the average score for all clients who attend 12 sessions?" This asks about a group mean. Use a confidence interval for the mean.
  2. Question B: "What score will one new client who attends 12 sessions have?" This asks about one person. Use a prediction interval.
  3. Compare the widths. At the same 12 sessions and the same 95% level, the prediction interval is wider. One person can land well above or below the group average.

So on the exam: one new person means a prediction interval, and it is the wider one. The average of a group means a confidence interval.

Reference: NIST/SEMATECH e-Handbook, section 4.1.3.2, Prediction.

Common Misconceptions to Avoid

Misconception 1: "Correlation Equals Causation"

This is the classic error, but it bears repeating because it's so pervasive. A correlation between X and Y could mean:

  • X causes Y
  • Y causes X
  • Some third variable Z causes both X and Y
  • The relationship is coincidental

If ice cream sales correlate with drowning deaths, that doesn't mean ice cream causes drowning. Both increase in summer due to a third variable: hot weather.

Misconception 2: "Low Correlation Means No Relationship"

A low Pearson r might mean no relationship, OR it might mean the relationship is nonlinear and you're using the wrong statistical tool. Always consider whether a curved relationship might exist.

Misconception 3: "Statistical Significance Means Practical Importance"

With large sample sizes, you can find statistically significant correlations that are practically meaningless. An r = .10 might be statistically significant with 1,000 participants, but it only explains 1% of the variance. Always look at effect size, not just significance.

Misconception 4: "Higher Correlations Are Always Better"

For predictors, yes – higher correlations with the criterion are better. But you want low correlations between your predictors to avoid multicollinearity. The best scenario is predictors that each capture something unique.

Practice Tips for Remembering

Memory Device for Choosing Coefficients

Create a simple decision tree:

  1. Are both variables continuous? → Check if linear → Yes: Pearson r / Maybe not: Eta
  2. Are both variables ranks? → Spearman rho
  3. Is one continuous and one dichotomous? → Is the dichotomy natural (true)? → Yes: Point biserial / No: Biserial
  4. Are both variables nominal? → Contingency coefficient

Remember the Assumptions with "LRH"

Linear relationships Range affects generalization Homoscedasticity for regression inference (not needed just to calculate r)

Squaring to Get Variance Explained

Practice this mental math:

  • r = .70 → .49 (about half)
  • r = .50 → .25 (one quarter)
  • r = .30 → .09 (less than 10%)
  • r = .80 → .64 (about two-thirds)

Multivariate Techniques Quick Reference

TechniquePredictorsCriterion(s)
Multiple regression2+ continuous1 continuous
Canonical correlation2+ continuous2+ continuous
Discriminant function2+ continuous (normally distributed)1 nominal
Logistic regression2+ any type1 nominal (when assumptions aren't met)

Key Takeaways

  • A confidence interval for the regression mean concerns an average; a prediction interval concerns one new outcome and is wider at the same predictor value and confidence level.

  • Correlation measures relationships between variables, ranging from -1.0 to +1.0, with predictors (X) and criteria (Y) instead of independent and dependent variables

  • Pearson r describes linear association. Sample range affects generalization; homoscedasticity concerns conventional regression inference, not the calculation of every descriptive correlation (Bewick et al., 2003).

  • Choose the right coefficient based on your measurement scales – Pearson r for continuous variables, Spearman rho for ranks, point biserial for true dichotomies, biserial for artificial dichotomies, contingency for nominal

  • Coefficient of determination (r²) tells you the percentage of variance explained – always square the correlation to understand practical significance

  • Regression analysis creates equations for prediction, with accuracy depending on the strength of the correlation

  • Multiple regression uses multiple predictors for one continuous criterion; multicollinearity (high intercorrelation among predictors) should be avoided

  • Other multivariate techniques include canonical correlation (multiple predictors and criteria), discriminant function analysis (categorical criterion), and logistic regression (categorical criterion with violated assumptions)

  • Structural Equation Modeling tests complex theories using observed and latent variables, involving model specification, evaluation, and modification

  • Never confuse correlation with causation – relationships don't establish directionality or rule out third variables

  • Statistical significance doesn't equal practical importance – always examine effect sizes and variance explained, not just p-values

Understanding correlation and regression isn't just about passing the EPPP – it's about becoming a psychologist who can critically evaluate research, make informed clinical decisions, and distinguish solid evidence from marketing claims. These tools help you be the evidence-based practitioner your clients deserve.

Correlation and regression scope

This is the same checklist from the start of the lesson, kept here as a recap: straight line, full range, and even spread for regression tests.

Pearson r describes a linear association. A curve can be missed even when r is near zero. Restricting the sample range can change the coefficient and limit generalization. Equal error variance, called homoscedasticity, is an assumption of conventional linear regression inference; it is not required merely to calculate a descriptive correlation. A wider spread of outcomes at one predictor level signals heteroscedasticity. Equal reliability of the two measures is not a prerequisite (Bewick et al., 2003).

Artificial dichotomies and biserial correlation

In plain words: biserial assumes your two groups came from slicing a hidden continuous score, and it estimates the correlation with that hidden score. Point biserial takes the two groups exactly as they are.

A biserial correlation estimates the association between a continuous variable and an underlying continuous variable observed only after a two-category split. Its model assumes a normally distributed underlying response. A point-biserial correlation instead describes the observed binary variable and continuous variable directly (Irwig & Groeneveld, 1988).

Reading a reported coefficient

In plain words: the plus or minus sign tells you direction, and the size of the number (ignoring the sign) tells you strength. For example, r = +.20 and r = +.85 point the same way, but +.85 is far stronger.

The sign of Pearson r gives the direction of a linear association, and its distance from zero describes its strength. A positive coefficient closer to zero indicates a weaker positive linear association than a coefficient near +1. The subject matter alone cannot tell you the numerical value; you need data (Bewick et al., 2003).

Partial correlation

In plain words: a partial correlation asks, "If everyone had the same score on this third variable, would X and Y still go together?" Say study hours and exam scores are correlated, and both also relate to prior GPA. A partial correlation removes the part of each that prior GPA accounts for, then correlates what is left.

A partial correlation describes the association between two variables after statistically controlling for specified measured variables. It does not control unmeasured confounders or establish a causal effect (Hevey, 2018).

Ready to practice?

Get started in the app