Is Spearman correlation just Pearson correlation on ranks?

Yes, with appropriate rank and tie handling.

Why can Kendall's tau be smaller than another coefficient?

It uses a different definition and scale interpretation; the values need not match.

Can unordered categories use Pearson correlation by arbitrary numeric codes?

Such coding can create meaningless order and distances; use an association method suited to the categorical question.

AI Glossary / Statistics and probability / Intermediate

Correlational measures

Definition

Correlational measures are statistics that summarize the degree and, where defined, direction of association between variables. In the supplied glossary's context, they concern variables that have at least an ordinal ordering. Common examples include Pearson's r, Spearman's rank correlation, and Kendall's tau. These measures capture different aspects of a relationship and should not be treated as interchangeable numerical labels.

What it means

Pearson's coefficient summarizes linear association using numerical distances from means. Spearman's coefficient summarizes monotonic association by correlating ranks. Kendall's coefficient uses the balance of concordant and discordant pairs, with variants that account for ties.
Choosing among them changes the question. A relationship can be perfectly increasing without being a straight line, so a rank-based measure can indicate perfect ordering when Pearson's coefficient is less than one. Broader association measures for unordered categories also exist, but they do not all have the same directional or scale interpretation as these correlational measures.

How it works

Define the observational units, paired variables, and measurement scales. Check whether ordering is meaningful and whether numerical differences have an interpretable scale. Inspect a plot or table for curvature, outliers, ties, subgroups, and restricted ranges.
Choose a measure that matches the intended relationship. Pearson r uses centered cross-products standardized by the variables' variation. Spearman's rho applies Pearson correlation to ranks, using an appropriate treatment of ties. Kendall's tau compares the ordering of pairs; tau-b adjusts its denominator for ties in either variable.
Do not choose a method solely because it produces the largest number or smallest p-value. Coefficients on different scales or with different definitions need not have equal magnitudes for the same data. A Kendall coefficient and a Pearson coefficient should not be ranked as if they were competing scores for the same property.
Separate the descriptive coefficient from inference about a population. Sample size, sampling design, clustering, time dependence, and the null hypothesis affect tests and intervals. Rank-based methods do not remove every assumption, and small samples or many ties can require special inferential procedures.
Report the coefficient's name, sample size, tie and missing-data treatment, and uncertainty method. Interpret it alongside substantive context. None of these measures, by itself, identifies a causal effect or establishes agreement between measurements.

Three practical examples

A researcher relates two ordered ratings using a rank-based measure and states how tied ratings are handled.
A business analyst summarizes the linear relationship between two numerical operating measures with Pearson r.
An AI team compares two systems' rankings of items using a pair-ordering measure rather than assuming their numerical score scales are directly comparable.

Think of it this way

Think of several tools describing the same path. One asks how closely it follows a straight line, another whether it consistently rises, and another how often pairs appear in the same order. Different readings can all be correct because the tools measure different properties.

Worked example

Consider four paired observations with X = (1, 2, 3, 4) and Y = (1, 2, 4, 8). Both variables increase in exactly the same order, but Y grows faster at the upper end.
For Pearson correlation, the means are 2.5 and 3.75. The centered cross-product sum is 11.5, and the squared-deviation sums are five and 28.75. Thus r = 11.5/sqrt(5 x 28.75), approximately 0.95917.
For Spearman correlation, both rank lists are (1, 2, 3, 4), so their correlation is one. For Kendall's tau, there are six pairs of observations; all six are concordant and none are discordant or tied. Therefore tau = (6 - 0)/6 = 1.
The results do not conflict. Pearson detects that the pattern is not exactly linear, while the rank-based measures reflect perfect ordering. Taking base-two logarithms of Y produces (0, 1, 2, 3), which is exactly linear in X and gives Pearson r = 1. The strictly increasing transformation leaves the original ranks unchanged.
These four observations illustrate definitions, not reliable population estimates. They also do not show that increasing X through an intervention would cause the observed changes in Y.

Illustrative case study

Fictional case: Laurel Review compares human rankings with an AI system's scores. Its first report uses Pearson correlation even though the human assessment is an ordered category scale with many ties. The team clarifies that its main question concerns ordering and examines a tie-aware rank measure.
It also checks whether the system assigns useful numerical scores, which is a separate calibration question. The report explains why ordering agreement and numerical agreement need different evidence. Analysts retain plots and subgroup checks rather than allowing one summary coefficient to stand in for the entire evaluation.

Common misconceptions and mistakes

Do not describe every correlational measure as a test of linearity. Rank methods capture ordering and can miss other forms of dependence. They are not completely immune to outliers, ties, or selection problems. A coefficient of zero does not generally establish independence, and a high coefficient does not establish causation. Comparing magnitudes across different measures without acknowledging their definitions can create misleading judgments about relationship strength.

Business use

Choosing the right correlational measure makes business reports more aligned with the decision. Ranking products, comparing ordered judgments, and evaluating numerical predictions are different tasks. State what association matters and choose the summary accordingly. For AI evaluation, supplement ordering measures with task-specific accuracy, calibration, or error analysis, since a system can rank items well while producing unsuitable scores or decisions.

Check your understanding

Why are Spearman and Kendall equal to one in the example?

Answer: Both variables have identical strict ordering.

Why is Pearson below one?

Answer: The numerical relationship is not exactly a straight line.

Does selecting the measure with the highest coefficient make the analysis better?

Answer: No; the choice should follow the question and measurement properties.

Related terms

Kendall's tau
Concordant pairs
Monotonic relationship
Pearson correlation
Spearman's rho
ordinal data
ties
association measure

Learn next

Compute several measures on data with curvature, ties, and unusual observations. Compare the properties each captures, then examine uncertainty methods that fit the actual sampling design and measurement scale.

Sampling design

Sources

David C. Howell, Statistical Methods for Psychology glossary, correlational measures: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html SciPy documentation, spearmanr: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.spearmanr.html SciPy documentation, kendalltau: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.kendalltau.html

Previous: Correlation matrix Next: Correspondence analysis