Correlation coefficient calculator

The first decision is which coefficient your data can actually support. Pearson's r belongs to two continuous, interval-level variables whose relationship is roughly a straight line and whose distributions are not badly skewed. Spearman's rho works on the ranks instead of the raw values, so it is the right tool for ordinal data such as a Likert response, for skewed measures, and for any relationship that is consistently increasing but curved. If one variable is a survey rating, lean toward Spearman or read our note on how to treat Likert ratings before you commit.

A simple rule helps: plot the two variables first. A tidy oval cloud of points favours Pearson's r; a fan shape, a curve, or a few extreme points argue for the rank-based Spearman's rho, which barely flinches at outliers. When the data are clean and linear the two coefficients land within a few hundredths of each other, so a large gap between them is itself a signal that the Pearson assumptions are strained. The broader question of when to use a rank method sits inside the parametric versus nonparametric choice that governs the whole analysis.

The sign of the coefficient gives the direction and the magnitude gives the strength. A positive r means the two variables rise together, a negative r means one falls as the other climbs, and a value near zero means little linear association. By Cohen's common benchmarks an r of about 0.1 is small, 0.3 is medium, and 0.5 is large, but these are field-dependent: 0.5 is modest in a tightly controlled experiment and substantial in messy survey data. Read your number against what is typical in your own literature rather than treating the cut points as fixed law.

Squaring the coefficient turns it into r-squared, the coefficient of determination, which is the proportion of variance the two variables share. An r of 0.5 gives an r-squared of 0.25, so a quarter of the variation is common to both. This shared-variance framing is often more honest than the raw coefficient because it grows slowly: a relationship has to reach r near 0.71 before the two variables share even half their variance. The same r-squared reappears when you move from association to prediction with the linear regression calculator, where it measures how much of the outcome the fitted line explains.

Pearson's r rests on a short list of assumptions, and ignoring them produces a number that looks authoritative but means little. The relationship must be genuinely linear, because the coefficient only measures the straight-line part of an association and will read near zero for a strong U-shaped pattern. Both variables should be continuous and approximately normal, and the pairs should be independent. A scatterplot is the fastest assumption check there is: it exposes curvature, clustering, and the single stray point that can drag a coefficient up or down on its own.

Outliers deserve special care. One extreme pair can inflate a weak correlation into an apparently strong one, or mask a real relationship entirely, so inspect any point that sits far from the cloud before you trust the result. If an outlier is a genuine observation rather than an error, the rank-based Spearman's rho is a robust fallback that keeps the point without letting it dominate. When distributions are far from normal, confirm the shape first with the descriptive statistics calculator and, if needed, follow the formal checks for a normal distribution.

A correlation shows only that two variables move together; it never shows that one produces the other. Three rival explanations always survive a significant coefficient. A confounding variable may drive both, the direction of causation may run the opposite way, or the association may simply be a fluke of this particular sample. Ice cream sales and drowning deaths correlate strongly, but summer heat causes each, and no scoop of ice cream ever drowned anyone.

Establishing cause needs a design that closes those loopholes, not a larger coefficient. A randomised experiment rules out confounders by assigning conditions at random, while a carefully controlled longitudinal or mediation model can build a credible causal argument from observational data. Until one of those is in place, describe your finding as an association and resist the verb "causes" in the write-up. Our deeper explainer on why correlation never proves causation walks through the classic traps, and mediation versus moderation shows how to model a mechanism properly once you suspect one.

In APA style a correlation is reported in one compact sentence carrying the coefficient, the degrees of freedom, and the exact p-value, for example r(48) = .42, p = .002. Note two conventions that trip people up: the degrees of freedom are n minus two, not the sample size, and you drop the leading zero on any value that cannot exceed one, so it is .42 rather than 0.42. Round the coefficient and the p-value to two and three decimals respectively, and state the direction of the relationship in words alongside the figures.

Strengthen the report by adding context the bare number lacks: the r-squared as a percentage of shared variance, and a confidence interval for r where your reporting style allows it. When you present several correlations together, a labelled matrix is clearer than a string of sentences. For the full set of formatting rules on signs, decimals, and tables, follow our guide to stating each result in APA form.