Why does Kendall's tau usually come out smaller in magnitude than Spearman's rho on the same data?
answer
- different scales, not rival estimates
- both equal 1 only at perfect agreement
- one counts pairs, the other squares displacements
- tau roughly two-thirds of rho, moderate case
- bivariate normal at 0.5 gives tau exactly 1/3
basics
~20 sThey are different scales, not competing estimates of one quantity. Tau is a difference of pair-agreement probabilities; Spearman's rho is a correlation of rank positions that weights large rank displacements heavily. Tau typically lands lower.
solid answer
~50 sThe two coefficients agree at the endpoints — both are +1 under perfect monotone agreement and -1 under perfect reversal — but their intermediate values are on different scales, so a gap is expected rather than a sign of error. Tau counts pairs: every disagreeing pair costs the same, whether the two observations were adjacent in rank or at opposite ends. Rho correlates rank positions, so a single large rank displacement is penalised much more than a small one, which pushes moderate association toward larger values. As a benchmark, when two variables follow a bivariate normal distribution with underlying correlation 0.5, Kendall's tau is exactly 1/3 while Spearman's rho is about 0.48. The practical rule: pick one coefficient for a report, name it, and never compare a tau against a rho as if the larger meant stronger.
go deeper
Know that Kendall's tau and Spearman's rho are two different rank coefficients on two different scales, so a smaller tau on the same data is normal rather than a contradiction.
Explain the mechanism: tau counts ordering violations equally while rho is driven by squared rank displacements, so scattered small violations cost tau more relative to its range.
Show you pick a coefficient for reasons — audience interpretability, tie handling, cost — commit to it across a reporting line, and never resolve the gap by quoting the friendlier number.
Own the convention across teams: one named coefficient per metric family, fixed before the data is seen, so that dashboards remain comparable and nobody can move a number by changing measure.
## Not two estimates of the same number A common interview stumble is to compute both rank coefficients on a dataset, see tau = 0.42 and rho = 0.60, and conclude that one of them is wrong or that the association is 'somewhere around 0.5'. Neither reading is right. Tau and rho are **different population quantities**, both of which happen to be scaled to the interval -1 to +1 and both of which hit the endpoints under perfect monotone agreement or reversal. In between, they measure different things and there is no reason for them to coincide. ## The mechanism behind the gap **Tau weights every disagreement equally.** Its definition is a count over pairs: for each pair, do the two variables agree about which observation is larger? A pair that disagrees costs the same whether the two observations sat next to each other in the ordering or at opposite ends of it. Tau is therefore a linear function of the number of ordering violations. **Rho weights large displacements much more.** Spearman's rho is a correlation on rank positions, so it is driven by squared rank differences. An observation dragged 20 rank positions out of place hurts far more than four observations each out of place by five. Real data usually contains many small local scrambles and few large ones, so the quantity that punishes only the large ones lands higher. That asymmetry is why, on the same monotone-but-noisy data, tau lands lower. A useful mental benchmark is `tau` roughly two-thirds of `rho` for moderate association — approximate, not a law, but it stops you reading the gap as a bug. ## A precise reference point For two variables following a bivariate normal distribution with underlying correlation `p`, both rank coefficients have closed forms: - `tau = (2/pi) * arcsin(p)` - `rho_S = (6/pi) * arcsin(p/2)` At `p = 0.5` these give `tau = (2/pi)*arcsin(0.5) = 1/3 = 0.333` exactly, and `rho_S = (6/pi)*arcsin(0.25) = 0.483`. The gap is real, predictable and has nothing to do with sampling noise. There is also a known inequality constraining how far apart the two can drift on any dataset: `3*tau - 2*rho_S` always lies between -1 and +1, so a claimed pair like tau = 0.9 with rho = 0.2 would indicate a computation error. ## Which one to report **Reasons to prefer tau:** - Its interpretation is a plain-language betting statement: the probability that a random pair is ordered the same way by both variables, minus the probability it is ordered oppositely. Non-specialist audiences can act on that sentence. - It has well-defined tie-corrected variants and behaves predictably on small samples. - Its bounded per-observation influence is easy to argue: one observation touches only the `n - 1` pairs containing it. **Reasons to prefer rho:** - It is far more widely recognised, so it needs less explanation to a statistically-trained audience. - Its numbers are on a scale readers already have intuitions for, since it is a correlation of rank positions. - Naively computed it costs a sort plus a correlation, while the naive tau is a quadratic double loop over pairs (there is an `n log n` route, but the simple implementation is slower). **The reporting discipline that matters more than the choice:** 1. Choose one coefficient per analysis line and stay with it across releases, so trends over time are comparable. 2. Name the coefficient and, on tied data, the variant. 3. Never place a tau from one analysis beside a rho from another and rank them — the comparison is meaningless, and the person who computes both and reports the larger has manufactured a result. 4. If both are reported deliberately, say why, and expect the tau to be lower — otherwise a reviewer will read the gap as instability. ## The deeper point All monotone-association coefficients compress the same scatterplot into one number, and each chooses what to sacrifice. Tau sacrifices sensitivity to how badly the ordering is violated in exchange for a clean probabilistic reading. Rho sacrifices that reading in exchange for familiarity and cheap computation. The gap between them is not noise to be averaged away — it is the visible trace of two different summarisation choices applied to the same data.
- Is it acceptable to compute both coefficients and report whichever is larger?No — that is result-shopping. Because tau is systematically lower than rho for the same monotone association, choosing by magnitude guarantees you report rho, and it makes your number depend on a decision made after seeing the data. Pick the coefficient before computing, name it, and keep it fixed across releases so trends stay comparable.
- When do Kendall's tau and Spearman's rho agree exactly?At the endpoints. Perfect monotone agreement between the two orderings gives both exactly +1, and perfect reversal gives both exactly -1, because every pair is concordant or every pair is discordant and every rank difference is at its extreme. Everywhere in between the two scales diverge, with tau typically the smaller magnitude.
saying these in an interview costs you the question
- Treats tau and rho as two estimates of one underlying number
- Averages the two coefficients into a single figure
- Calls the gap between them a sign of a computation bug
- Reports whichever coefficient came out larger
- Compares a tau from one analysis against a rho from another