Independence and correlation: what each statement actually means

Independence is a statement about joint probabilities

For events A and B, independence means P(A and B)=P(A)P(B). When the conditioning event has positive probability, the equivalent conditional statement says that learning B leaves the probability of A unchanged. For random variables, independence must hold across their relevant sets of possible values, not just for one convenient event pair. The Penn State lesson on independent events introduces this concept. A model can assume independence for tractability, but an assumption should remain visible instead of being described as an established feature of the world.

Correlation describes linear co-variation

When the necessary moments exist and both variances are positive, Pearson correlation rescales covariance by the product of the standard deviations. It summarises linear co-variation, with values between minus one and one. The Stanford refresher distinguishes covariance and correlation from independence. A coefficient near zero is not evidence that every possible nonlinear relationship is absent. Nor does a large coefficient establish a causal direction. Before calculating it, inspect what the variables measure, whether observations are paired correctly and whether a shared external factor could affect both.

Work through zero correlation with dependence

Here E(X)=0 and E(XY)=0, so covariance is zero. Both variables have positive variance, making their correlation zero as well. Yet observing Y=0 tells us X=0 with certainty. The variables are dependent because Y is a deterministic function of X. The table provides every input needed to verify that conclusion without relying on a chart’s appearance.

Test the dependence with one event pair

Let A be X=0 and B be Y=0. Each event has probability one third, and their intersection also has probability one third because they identify the same outcome. The product of the marginal probabilities is one ninth. Since one third differs from one ninth, independence fails. One counterexample is enough to disprove independence; one event pair satisfying the equation is not enough to prove full random-variable independence. This asymmetry matters when analysts inspect only a handful of summary statistics and then declare a complicated dataset independent.

Keep mutual exclusivity separate

Mutually exclusive events cannot both occur in the same trial. If each has positive probability, their intersection is zero while the product of their probabilities is positive, so they are not independent. OpenStax’s comparison of the two properties makes this distinction explicit. In the hypothetical model, X=−1 and X=1 are mutually exclusive outcomes of one observation. They are not two independent measurements. By contrast, values recorded from two separately specified trials could be independent if the generating mechanism and model justify that property.

Recognise common dependence in collected data

Measurements taken from the same person, household, device or time period can share information. Recording them in separate spreadsheet rows does not make them independent. Consecutive sensor readings may be related because the underlying condition changes gradually. Pooling several groups can also obscure different within-group relationships. A useful analysis records grouping identifiers and collection order so the dependence structure can be examined. If independence is implausible, an uncertainty calculation based on an independent sample may understate or otherwise misrepresent the variability relevant to the question.

Check pairwise claims separately from a joint claim

Independence can concern more than two events. Consider two independent fair binary labels and a third label recording whether the first two agree. Each label is individually equally likely to take either of its two values, and any pair is independent. However, knowing the first two determines the third. Thus pairwise independence does not establish mutual independence of the whole collection. A model involving many measurements should state which property is assumed; checking pairs alone can miss a constraint involving several variables together.

Relationship assessment checklist

• Specify events or random variables precisely.
• Distinguish independence from zero correlation.
• Check that variances exist before using correlation.
• Look for nonlinear and grouped structure.
• Keep mutually exclusive outcomes separate from independent trials.
• Record collection order and shared sources.
• State assumptions alongside conclusions.

For the three-value example, changing Y from X² to another function creates a different relationship even though the same X values remain. Recalculate the joint table rather than reusing the old coefficient. In an empirical report, include the number of paired observations and explain exclusions. A correlation without its population, pairing rule and missing-data treatment is difficult to interpret or reproduce.

Sources and verification

• Independent events, STAT 414 — Penn State — 2026-10-23
• Probability and statistics refresher — Stanford University teaching materials — 2026-10-23
• Independent and mutually exclusive events — OpenStax, Rice University — 2026-10-23