Choosing a Measure Starts With the Question
In risk analysis, a single pair of datasets can produce different correlation values depending on the measure used. That is not necessarily a computational problem: Pearson, Spearman, and Kendall summarize different aspects of association. Before calculating one, identify what the variables represent, what kind of relationship matters, and what the resulting number is intended to mean.
For example, a risk committee might ask whether two continuous loss amounts move together in a roughly linear way, whether higher values of one variable generally accompany higher values of another, or how often a randomly selected pair of risks is ordered consistently. These are related but distinct questions. A measure suited to one is not automatically the best measure for another.
What Each Measure Summarizes
For random variables \(X\) and \(Y\) with finite, positive variances, Pearson correlation is
Its sample counterpart, \(r\), is the sample covariance divided by the product of the sample standard deviations. Pearson correlation is most directly interpretable when a linear relationship on the original scales is relevant. A value near zero means little linear association; it does not rule out a strong nonlinear relationship.
Spearman’s rank correlation, denoted \(\rho_s\), is the Pearson correlation of the ranks of the paired observations. With continuous variables and no ties, it can also be calculated from rank differences \(d_i\) for \(n\) pairs:
Here each \(d_i\) is the difference between the two ranks for observation \(i\). With ties, calculate the Pearson correlation of the midranks instead; the rank-difference shortcut above is not generally valid. Spearman is appropriate when the order of values matters more than their exact distances, and it detects monotone patterns, including some nonlinear ones.
Kendall’s tau compares pairs of observations. A pair is concordant if the observation with the higher \(X\) also has the higher \(Y\); it is discordant if the orderings disagree. A tie on just one variable is neither concordant nor discordant. For a sample with ties, Kendall’s tau-b is
where \(C\) and \(D\) are the numbers of concordant and discordant pairs, \(T_X\) counts pairs tied only on \(X\), and \(T_Y\) counts pairs tied only on \(Y\). Pairs tied on both variables are excluded from these counts. Tau-b adjusts for ties and is often useful for ordinal data. When there are no ties, it reduces to \((C-D)/\binom{n}{2}\).
A Practical Selection Framework
If values are meaningful quantities and their distances matter, Pearson may be relevant. If values are ordinal grades or rankings, use a rank measure rather than treating category labels as measured amounts.
Choose Pearson for linear association on the recorded scale; choose Spearman for rank-based monotone association; consider Kendall when pairwise ordering and ties are central.
A scatterplot can reveal curvature, influential points, or clusters that one summary number hides. Pearson can be strongly affected by an extreme point; ranks reduce the effect of its magnitude, though not necessarily its effect on ordering.
State the measure and what it summarizes. Do not treat any of these coefficients as a causal effect or as a complete description of joint tail behavior.
Why Transformations Matter
The distinction between the measures follows from how they respond to transformations. If \(X\) and \(Y\) are replaced by \(a+bX\) and \(c+dY\), with \(b,d>0\), their Pearson correlation is unchanged: covariance is multiplied by \(bd\), while the product of standard deviations is also multiplied by \(bd\). If one multiplier is negative, the sign reverses. Thus positive changes of units, such as dollars to thousands of dollars, do not alter Pearson correlation.
For Spearman and Kendall, any strictly increasing transformation preserves the ordering of observations and therefore preserves ranks and concordance. This is why a strictly increasing nonlinear relationship can have perfect rank correlation without having perfect Pearson correlation. The reverse warning is also important: neither rank correlation is a measure of linearity on the original dollar scale.
Worked Example: An Extreme Claim Changes the Linear Summary
A hypothetical insurer compares five account-level exposure scores \(X\) with corresponding loss indicators \(Y\), in thousands of dollars. The paired data are \(X=(1,2,3,4,5)\) and \(Y=(1,2,3,4,20)\). Calculate Pearson and Spearman correlations, then select a measure for assessing monotone ordering when the last loss is unusually large.
Method. Calculate Pearson from centered observations and Spearman from ranks. All five values within each variable are distinct, so the rank-difference formula applies.
For Pearson, \(\bar X=3\) and \(\bar Y=6\). The centered cross-products sum to \(10+4+0-2+28=40\). The sums of squared deviations are \(10\) for \(X\) and \(250\) for \(Y\). Thus
The ranks of \(X\) and \(Y\) are both \((1,2,3,4,5)\), so every rank difference is zero and \(\rho_s=1-6(0)/(5(25-1))=1.0000\).
The two summaries differ because the unusually large final value changes distances on the original scale, while it does not change the ordering. For the stated question about monotone ordering, Spearman is the more direct summary. Pearson remains relevant if the committee specifically needs linear association in the recorded amounts. Check: both results lie within \([-1,1]\), and the perfect rank result agrees with the strictly increasing order. The mistake to avoid is interpreting \(\rho_s=1\) as evidence that the dollar-scale relationship is linear.
Worked Example: Ordinal Grades With a Tie
A hypothetical underwriting review assigns four portfolios ordinal grades \(X=(1,1,2,3)\). Their separately determined loss-order ranks are \(Y=(2,1,3,4)\). Calculate Kendall’s tau-b and explain why the tied grades matter.
Method. Enumerate the \(\binom{4}{2}=6\) pairs, then apply tau-b. The tied grades are ordinal categories; their numerical labels do not imply equal distances between grades.
The pair of observations 1 and 2 is tied on \(X\), so it is counted in \(T_X\). The other five pairs are ordered on both variables. Pair (1,3) is concordant because \(1<3\) on \(X\) and \(2<3\) on \(Y\). Pair (1,4) is concordant, as is pair (2,3), because \(1<3\) on \(Y\), pair (2,4), and pair (3,4). Therefore \(C=5\), \(D=0\), \(T_X=1\), and \(T_Y=0\).
Tau-b is appropriate here because the grades are ordinal and contain a tie. Check: five pairs are ordered on both variables, one pair is tied only on \(X\), and the denominator is \(\sqrt{6\cdot5}\). A common error is to count the tied pair as concordant merely because the other variable differs; it is neither concordant nor discordant. The reported value is rounded to four decimal places.
Worked Example: A Monotone but Curved Relationship
A hypothetical pricing team compares five increasing risk scores \(X=(1,2,3,4,5)\) with modeled costs \(Y=(1,4,9,16,25)\). The team wants to describe whether higher scores accompany higher costs, not whether costs rise linearly. Calculate both correlations and select the relevant summary.
Method. Use centered values for Pearson and ranks for Spearman. No ties occur.
The means are \(\bar X=3\) and \(\bar Y=11\). The centered cross-products are \(20,7,0,5,28\), summing to \(60\). The squared deviations sum to \(10\) for \(X\) and \(374\) for \(Y\). Therefore
Because both sequences have ranks \((1,2,3,4,5)\), all rank differences are zero, giving \(\rho_s=1.0000\). The modeled costs increase with the score, but the pattern is curved rather than a straight-line relationship. Spearman directly summarizes the perfect increasing ordering; Pearson describes a very strong, but not perfect, linear association. Check: the Pearson denominator is approximately \(61.1555\), so \(60/61.1555\approx0.9811\). The mistake to avoid is selecting Pearson alone and then describing its value as a measure of monotone ordering.
Common Interpretation Traps
- Using Pearson on ordinal labels as if gaps were meaningful. Grades 1, 2, and 3 give an order, but do not necessarily represent equal changes in risk.
- Assuming a small Pearson value means no association. A curved pattern can be strongly monotone even when its linear correlation is modest.
- Calling rank methods immune to outliers. A large magnitude has less direct influence after ranking, but an unusual observation can still change the ordering and the coefficient.
- Comparing coefficients as though they had identical meanings. A Pearson value and a Kendall value summarize association differently; a larger numerical value does not necessarily mean a stronger relationship in a universally comparable sense.
- Inferring independence or tail behavior from one coefficient. Correlation alone does not establish causation, independence, or how often extreme losses occur together.
Check Your Understanding
For each problem, identify the relevant measure or calculate the requested coefficient. Treat each scenario as hypothetical.
- A capital analyst has paired continuous loss amounts and needs a summary of their linear association in dollars. An extreme observation is present, so the analyst will also inspect the data pattern. Which coefficient directly matches the stated target? (Answer: Pearson correlation, with the extreme observation’s influence assessed rather than ignored.)
- For four accounts, the ranks on two review measures are \(X=(1,2,3,4)\) and \(Y=(2,1,4,3)\). Calculate Spearman’s rank correlation. (Answer: \(\rho_s=0.6000\).)
- For four paired observations, \(X=(1,2,3,4)\) and \(Y=(2,1,4,3)\). Calculate Kendall’s tau, with no ties. (Answer: \(C=4\), \(D=2\), and \(\tau=(4-2)/6=0.3333\).)
- A vendor risk rating has ordered categories with repeated ratings, and the question is how often pairs are ranked consistently. Which measure is suitable, and what tie-adjusted form should be considered? (Answer: Kendall correlation, using tau-b to adjust for ties.)
- Two continuous risk variables have a strictly increasing but curved relationship. Which rank measure summarizes their monotone ordering, and what does a perfect value mean? (Answer: Spearman correlation; a value of \(1\) means the observations have identical increasing ranks, not that the original-scale relationship is linear.)