MerQuR+Statistics Glossary

Statistics Glossary → Sen's slope

Sen's slopeconcept

A robust method that estimates the size of a linear trend in a time series as the median of the slopes between all pairs of observations.

For every pair of observations i < j, the slope (xⱼ − xᵢ) / (tⱼ − tᵢ) is computed; n observations give n(n − 1)/2 slopes, and their median is Sen's slope (also known as the Theil-Sen estimator). Because it uses the median, a few extreme values or faulty measurements barely affect it; its breakdown point is about 29%. Missing observations and irregular time spacing cause no problem, since each pair is evaluated with its own time difference.

Sen's slope is usually reported together with the Mann-Kendall trend test: Mann-Kendall tests whether a trend exists and in which direction, while Sen's slope gives the size of the change per unit of time. Its confidence interval is derived from the variance of Kendall's S statistic. It assumes that the trend is roughly linear and monotonic; with strong seasonality or autocorrelation, adjustments such as the seasonal Kendall test or pre-whitening are needed.

An everyday example

Twenty years of summer water temperatures in a lake are examined. In one year the instrument failed and recorded an unrealistic 35 °C. Because of that value, the least-squares slope comes out at 0.12 °C per year, while Sen's slope gives 0.05 °C per year and better reflects the warming shown by the other 19 years.

In MerQur

In the Mann-Kendall + Sen's Slope analysis, the 'Date (optional) and Slope unit (Automatic / Per year / Per day / Per observation)' option sets the time axis of the slope.

Analyses where this concept appears

Mann-Kendall + Sen's Slope

Related concepts

Autocorrelation correction and prewhitening (Hamed-Rao, TFPW)Seasonal Kendall test