Correlation Coefficient

A correlation coefficient is a numerical statistic that quantifies the strength and direction of the relationship between two variables on a standardized scale from negative 1 to positive 1: a linear relationship for Pearson's r, a monotonic one for rank based coefficients such as Spearman's rho. For Pearson's r, a value of positive 1 indicates a perfect positive linear relationship, negative 1 indicates a perfect negative linear relationship, and zero indicates no linear association. Several types of correlation coefficients exist, each suited to different data types and distributional assumptions. The Pearson product moment correlation coefficient (r) measures linear associations between two continuous variables; its usual significance test and confidence interval assume approximately normal data. Spearman's rank order correlation coefficient (rho) measures monotonic associations between ordinal or non normally distributed continuous variables. Other variants include Kendall's tau, the point biserial correlation (for one continuous and one dichotomous variable), and the phi coefficient (for two dichotomous variables).

Correlation coefficients provide a standardized way to communicate the strength of associations, making it possible to compare relationships across studies, variables, and fields. In meta analysis, correlation coefficients are one of the standard effect size metrics, allowing researchers to pool evidence from multiple studies. In psychometrics, correlation coefficients underpin reliability estimates (test retest reliability, inter rater reliability) and validity assessments (convergent, discriminant, criterion validity). Understanding the properties and limitations of different correlation coefficients is essential for selecting the right measure and interpreting it correctly.

The Pearson r is computed by dividing the covariance of the two variables by the product of their standard deviations: r = cov(X, Y) / (SD_X times SD_Y). This standardization ensures that the result is scale free and bounded between negative 1 and positive 1. Statistical significance is assessed using a t test with n minus 2 degrees of freedom. The coefficient of determination, r squared, represents the proportion of variance in one variable that is linearly associated with the other. For Spearman's rho, the data are converted to ranks before the Pearson formula is applied, making it appropriate for ordinal data or when the relationship is monotonic but not strictly linear. Confidence intervals for correlation coefficients are typically calculated using Fisher's z transformation, which stabilizes the variance of r. The interval is built symmetrically on the z scale and then transformed back, so on the r scale it is usually asymmetric, with the longer side toward zero.

Report a correlation coefficient whenever you are describing the strength of the association between two variables. Use Pearson's r for continuous, approximately normally distributed data with a linear relationship. Use Spearman's rho for ordinal data, skewed continuous data, or when the scatterplot suggests a monotonic but nonlinear relationship. Use the point biserial correlation when one variable is dichotomous and the other is continuous. In meta analysis, convert other effect size measures to correlation coefficients (or vice versa) when synthesizing evidence across studies that use different metrics.

Reporting a correlation coefficient without examining the scatterplot is risky because outliers, restricted range, and nonlinear patterns can make the coefficient misleading. Using Pearson's r when the data are ordinal or heavily skewed violates the assumption of interval level measurement and normality. Interpreting a small correlation as unimportant without considering the context is another error; in large scale epidemiological studies, even small correlations can represent meaningful population level associations. Researchers sometimes square the correlation to obtain r squared and then describe the remaining variance as "unexplained," without noting that r squared only captures the linear component of the relationship. Finally, comparing correlation coefficients across samples of very different sizes without accounting for precision differences can be misleading.

A psychometrician evaluates the test retest reliability of a new anxiety measure by administering it to 150 participants on two occasions separated by two weeks. The Pearson correlation between the two administrations is r = 0.87, indicating strong temporal stability. The 95% confidence interval, computed using Fisher's z transformation, ranges from 0.82 to 0.90. The researcher also computes Spearman's rho (0.84) to check whether the result is influenced by a few extreme scores and finds consistent evidence of reliability. Because Pearson's r measures how closely the two sets of scores follow a straight line, and does not change if every participant's score shifts by the same amount on the second occasion, the researcher additionally reports an intraclass correlation coefficient (ICC) for absolute agreement before concluding that the measure is suitable for tracking anxiety levels over time.