The misconception
A correlation of 0.82 sounds like a compact description of a relationship. It is compact, but it is not a picture. Pearson’s correlation records the strength and direction of linear association after both variables are centered and scaled. It does not report curvature, separated groups, changing spread, or whether one point dominates the calculation.
Different data sets can produce positive, negative, or near-zero correlations, yet the coefficient still cannot preserve their geometry. Reporting r without a scatterplot can therefore remove exactly the structure an analyst needs to see.
A synthetic experiment
The linked experiment generates four selectable patterns. Strong positive places points around an increasing straight line; Strong negative places them around a decreasing line. No linear association uses generated noise constructed to have nearly zero sample correlation with the evenly spaced x values. Nonlinear U-shape makes y rise at both ends of the x range, again with a coefficient near zero because the opposing linear contributions cancel.
The values are synthetic and the interface publishes its seed. They do not describe height, income, education, performance, or any other real characteristic. Neutral axis names x and y keep the lesson focused on geometry.
Choose one pattern at a time and compare the scatterplot with the displayed Pearson r, point count, and pattern identifier. The point-count slider ranges from 20 to 100 in steps of four. Increasing it makes the programmed shape easier to recognize, while New draw replaces the noise using a different deterministic seed. The value table exposes the first 12 generated pairs for direct inspection.
The four choices are deliberately not tuned to share one coefficient. Instead, they establish several basic contrasts: sign distinguishes the two linear directions, a coefficient near zero fits the flat-noise pattern, and approximately the same near-zero summary can accompany a clear U-shape. The last comparison is the central warning: absence of linear association is not absence of structure.
Derivation and reasoning
For paired observations (xi, yi), Pearson’s sample correlation can be written as:
r = sum((xi - x_bar) * (yi - y_bar)) / sqrt(sum((xi - x_bar)^2) * sum((yi - y_bar)^2))
The numerator is a scaled form of covariance. Pairs that are simultaneously above both means or below both means contribute positively. Pairs on opposite sides contribute negatively. The denominator standardizes by the spread of each variable, placing r between -1 and 1.
A value near 1 means the points align closely around an increasing straight line after scaling. A value near -1 means they align around a decreasing straight line. A value near zero means little linear association. It does not imply statistical independence, and it does not rule out a strong curve. For example, a symmetric U-shaped relationship can have correlation near zero because positive and negative cross-products cancel.
Correlation is unchanged by adding a constant to a variable or multiplying it by a positive constant. Multiplying by a negative constant reverses its sign. This scale invariance is useful, but it also means r contains no information about slope units. The same r can accompany a tiny or large practical change in Y per unit of X.
Assumptions
The pairs must correspond correctly. Scrambling one column destroys the pairing and therefore the relationship. Both variables need nonzero variation; if every X value is identical, its standard deviation is zero and correlation is undefined.
Pearson correlation is most interpretable as a summary of linear structure. A scatter plot should be used to assess curvature, outliers, clusters, gaps, and heteroscedasticity. Inference about a population correlation adds assumptions about sampling, independence, and the joint distribution or requires a suitable robust procedure.
Repeated observations from the same unit are not automatically independent pairs. Treating thousands of within-unit points as thousands of independent units can exaggerate precision.
Failure cases
Do not label an influential point as bad data solely because removing it changes r. Verify how it was generated and measured. It may be an error, a valid edge case, or evidence that a proposed linear model is incomplete. Report sensitivity with and without a questionable point when exclusion can be justified.
Do not interpret a group-level correlation as an individual-level relationship. Aggregated and within-group patterns can differ, particularly when groups have different centers. The current lab does not simulate grouped data; this is a separate limitation to check whenever real observations contain meaningful groups.
Do not turn association into causation. A lurking variable may influence both X and Y; selection may create association; or the direction of influence may be reversed. Random assignment and a credible design can support causal claims, but a correlation coefficient cannot supply that design after the fact.
Source notes
The NIST handbook describes scatterplots as tools for detecting linear, nonlinear, grouped, changing-variance, and outlier patterns, and explicitly separates association from causality. OpenStax provides the standard coefficient formula and interpretation. The four selectable structures here are newly generated rather than copies of a famous quartet or a source data set.
Keep these distinctions
- Pearson correlation summarizes the direction and strength of linear association.
- A near-zero correlation can coexist with a strong nonlinear relationship.
- Correlation alone does not establish a causal mechanism.
Sources checked
These sources support the statistical definitions and boundaries. The scenario, prose, calculations, and figure on this page are original to Count & Chance.
- 1.3.3.26. Scatter Plot, National Institute of Standards and Technology. Checked 2026-08-19.
- 12.3 The Regression Equation, OpenStax, Rice University. Checked 2026-08-19.