MerQuR+Statistics Glossary

Statistics Glossary → Trimmed mean

Trimmed meanconcept

The arithmetic mean of the values that remain after a fixed proportion of the smallest and largest observations has been removed from each end of the data.

The values are sorted and the same proportion is dropped from each end; for a 5% trimmed mean with n = 100, the 5 smallest and 5 largest values are removed and the remaining 90 are averaged. A trimming proportion of 0% gives the ordinary mean, and trimming close to 50% from each end gives something close to the median. The trimmed mean is a compromise between the two: it is far less affected by outliers than the mean, yet it uses more of the data than the median.

Trimming does not imply that the discarded values are 'wrong'; it simply makes the estimate of location resistant to extreme values. For symmetric distributions it is close to the ordinary mean, while for skewed distributions it can differ noticeably. Its standard error cannot be obtained from the usual formula; the Winsorized variance (in which trimmed values are replaced by the most extreme retained value) or the bootstrap is used instead. Tests comparing trimmed means, such as Yuen's test, belong to the family of robust methods.

An everyday example

Ten judges score a competitor, but one judge gives a very low mark for personal reasons and another gives an inflated one. If, as in many sports, the highest and lowest marks are dropped and the remaining eight are averaged (a 10% trimmed mean), no single judge's extreme view can decide the result.

In MerQur

It is computed in Descriptive Statistics through the 'Robust measures (5% trimmed mean, Huber M)' option, and in Bootstrap CI through the 'Trim / percentile, %' parameter, which sets the trimming proportion.

Analyses where this concept appears

Descriptive StatisticsBootstrap CI

Related concepts

M-estimator (Huber, Tukey biweight)Outlier