Pearson Correlation Coefficient Explained

Interpret a coefficient with the plot and design

Pearson’s r measures linear association between paired numerical observations. Its sign gives direction and its magnitude summarizes how closely the observed points follow a straight-line pattern. It does not tell you why the association exists.

For the calculation, see the Pearson lesson. For graphical diagnosis, begin with reading scatter plots.

Work through one complete dataset

Use X = 2, 4, 6, 8, 10 and Y = 65, 75, 80, 88, 92. The means are 6 and 80. The centered cross-product sum is 134, while the squared-deviation sums are 40 and 458.

$r=\frac{134}{\sqrt{40\times458}}\approx0.9900$

These five hypothetical pairs show a tight positive linear pattern. Paste the two lists into the correlation calculator. Preserve the pairing by row; sorting columns independently changes the observations being compared.

Under the usual independent-pair normal model for H₀: ρ = 0, the two-sided p-value is approximately 0.0012. It measures compatibility with that model, not the probability that “correlation is real.”

What r² adds

For simple least-squares regression with an intercept, R² = r². Here it is approximately 0.9801. The fitted line accounts for about 98.01% of the sample’s variation in Y around its mean.

That is a model-fit statement. It does not establish that changing X causes 98% of changes in Y, and it does not guarantee similarly good fit for new observations. With only five teaching points, broad generalization would be unjustified.

Three ways a coefficient can mislead

Curvature: for X = −2, −1, 0, 1, 2 and Y = X², Pearson r is zero despite a deterministic U-shaped relationship.

Influence: an isolated point can dominate the centered sums. Investigate unusual points and show justified sensitivity analyses; do not delete observations simply to obtain a preferred coefficient.

Mixed groups: pooling populations can produce a pattern different from the within-group relationships. Plot relevant group labels and examine the sampling process.

Avoid universal strength labels

The usefulness of r = 0.3 or r = 0.8 depends on the field, outcome, noise, and goal. An association can be statistically detectable but weak for prediction, or uncertain in a small sample despite a large observed coefficient.

Report r, n, the plot, inferential assumptions, and context. If you need predictions in the original response units, use an appropriate regression model and assess prediction error.

Correlation and causation

A confounder can affect both variables, reverse causation can occur, and selection can distort an association. A p-value does not remove these explanations. Causal claims require design and substantive assumptions beyond the coefficient.

Repeated observations from the same person, clustered pairs, and serial data also challenge the ordinary independent-pair significance test. Counting more rows does not automatically provide more independent information.

Reference and related reading

The official SciPy Pearson documentation describes the coefficient and null-distribution calculation. Continue with linear regression, regression diagnostics, or the test-selection guide.