Mann-Kendall trend testMann-Kendall: no universal best test, only a diagnostics-based legal choice
The Mann-Kendall trend test is a rank-based test for a monotone trend. It is robust to non-normality and resistant to outliers, but it assumes independent observations. Real environmental and business series rarely are. This page explains the six variants in the workspace and the evidence that selects between them.
Reproducible procedure
This is the procedure the workspace executes for this method, in order. Each step has an action, a rule that can stop the step, and a defined output.
- 01
Drop the points removed by the denoising rules and record how many were dropped.
- Rule
- Every method must run on the same rows, otherwise the comparison between them is meaningless.
- Output
- The analysis sample and its size.
- 02
Choose the Mann-Kendall variant that the diagnostics support.
- Rule
- Significant lag-1 autocorrelation selects a variance-corrected variant; a detected cycle selects the seasonal variant.
- Output
- The selected variant and the diagnostic reason for it.
- 03
Compute S, its variance, the z statistic and the p-value.
- Rule
- TFPW is never the sole headline, because pre-whitening can weaken a real trend.
- Output
- S, var(S), z, p and the rank correlation.
- 04
Estimate the rate of change with the Theil-Sen slope and its interval.
- Rule
- The slope is always reported with a confidence interval, never as a point estimate alone.
- Output
- Slope and 95% interval.
- 05
Run the remaining variants and compare their conclusions.
- Rule
- If the significance decision changes between variants, both results are reported side by side.
- Output
- The method sensitivity table.
The standard test and its effect size
The score S adds +1 for every pair where the later value is larger and −1 when it is smaller. The variance includes a tie correction. A continuity-corrected Z gives the p-value; dividing S by the number of pairs gives Kendall’s τ, the effect size. For n ≤ 10 without ties, the workspace uses the exact permutation distribution instead of the normal approximation.
A small p-value with a tiny τ is “significant but weak”. The interface always shows τ, the Theil–Sen slope and its interval next to the p-value so the size of the effect cannot hide behind the significance label.
- What this computes:
- Add +1 for every pair where the later value is larger, −1 when smaller; a positive sum means overall rising, negative means falling.
- Example:
- 10 of 12 pairs larger-later → S is positive.
- Caution:
- Autocorrelation underestimates Var[S].
Mann-Kendall trend test and autocorrelation: the most common silent error
Positive serial correlation reduces the effective number of independent observations. The standard test then underestimates the variance of S and reports false confidence. The Yue–Wang correction inflates Var[S] by n/n*, where n* is the effective sample size computed from the autocorrelation function. The correction can turn a “significant” trend into a non-significant one; that is honesty, not a bug.
- What this computes:
- Autocorrelation reduces the effective independent sample, so the variance is inflated to compensate.
- Example:
- 100 strongly autocorrelated points may behave like 20 independent ones.
- Caution:
- The correction can turn a “significant” trend into a non-significant one — that is honesty, not a bug.
- If lag-1 autocorrelation exceeds the 95% white-noise band, or Ljung–Box rejects independence, a corrected result is mandatory.
- The uncorrected result stays on screen next to the corrected one. The workspace does not let you hide method sensitivity.
TFPW: pre-whitening and its cost
Trend-free pre-whitening removes the Theil–Sen trend, removes lag-1 autocorrelation, adds the trend back and runs the standard test. It can reduce false significance from persistence, but it can also remove part of the real trend and reduce power. The workspace always shows the uncorrected result beside it.
Seasonal Mann–Kendall
SMK compares each season with itself across cycles and combines the evidence. The variance includes covariance terms between seasons. Ignoring a strong seasonal cycle in the standard test can mistake a repeating pattern for a trend; applying SMK to a series with no cycle wastes power and can distort the variance. The cycle length must therefore be inferred or confirmed, never assumed silently.
MKRD and MKC3: available, but unvalidated here
Rank-detrended MK and the third-cumulant correction address autocorrelation created by the trend itself and non-linear dependence. Both are implemented and available in the workspace, but no installed reference implementation was available for independent comparison. They carry a permanent UNVALIDATED badge and never become the headline when a validated method exists.
Multiple columns and false discovery
Testing twenty columns at α = 0.05 gives roughly a 64% chance of at least one false positive. Benjamini–Hochberg FDR adjusts the family of p-values; the workspace reports the number of significant columns before and after correction and refuses to present an uncorrected multi-column result as “significant”.
- What this computes:
- When testing many columns at once, sort the p-values and adjust them upward to reduce false “significant” flags.
- Example:
- 20 columns, smallest p = 0.001 may become q = 0.02 after correction.
- Caution:
- Columns no longer significant must not be called “significant”.
Selection rules used by the workspace
No autocorrelation and no cycle: standard MK. Significant autocorrelation: Yue–Wang primary, standard MK mandatory, TFPW available. Identifiable cycle: SMK primary, standard MK as the comparison. Strong autocorrelation with non-linear evidence: MKRD or MKC3 available, never hidden. Fewer than 20 observations: exact p-value, with an explicit warning that power is limited.
Common questions
How do I correct the Mann-Kendall trend test for autocorrelation?
Use the Yue–Wang variance correction: it inflates Var[S] by n/n*, where n* is computed from the autocorrelation function. TFPW is an alternative that pre-whitens the series. This workspace shows the corrected and uncorrected results side by side so the method sensitivity is visible.
How should I interpret a Mann-Kendall p-value?
The p-value is the chance of seeing a trend at least this strong if the series had no trend at all. It is not the probability that the trend is true. Read it together with Kendall τ and the Theil–Sen confidence interval: a tiny p with a tiny τ is significant but weak.
What is the difference between TFPW and Yue–Wang?
TFPW removes the trend, removes lag-1 autocorrelation, adds the trend back and then runs the test. Yue–Wang keeps the series intact and corrects the variance. TFPW can weaken a real trend; Yue–Wang can change significance in either direction. Both are shown with warnings rather than silently selected.
Why is a Theil–Sen confidence interval shown next to the MK result?
The p-value says whether a monotone direction is distinguishable from chance; the Theil–Sen slope says how large that direction is, in the original units, and the interval says how precisely it is estimated. Reporting only the p-value hides the effect size.
Related method notes
Cite this page: Detrend, “Mann-Kendall: no universal best test, only a diagnostics-based legal choice”, https://detrend.datacorp.top/methods/mann-kendall. Last updated 2026-11-15.
Run this procedure on your own workbook
The workspace executes exactly these steps locally, selects the method from the diagnostics and keeps every alternative visible.
Open the workspace