Outlier Trimming

mechanism

Extreme estimates can pull an average away from the crowd’s central signal. Trimming clearly implausible values can improve aggregation, provided the rule is set consistently rather than chosen to produce a preferred answer.

In a crowd-estimation experiment, the result improved only after two answers were removed. A large group did not automatically protect its average: a tiny number of extreme values could still bend the apparent consensus.

E1

How a few answers overpower the many

An arithmetic mean gives every value leverage in proportion to its distance from the rest. Most independent overestimates and underestimates may cancel—the engine behind wisdom-of-the-crowds—while one implausibly large or small answer exerts disproportionate pull. Outlier trimming limits that leverage so the aggregate better represents the cluster rather than its most extreme member.

E1

Where it shows up

The two-answer correction

Removing two conspicuous outliers from the experiment shows that aggregation has two separate jobs: collect diverse judgments, then prevent extreme values from dominating their combination. The same safeguard can also improve a crowd-within estimate when one of your independently generated answers is plainly implausible.

E1

Trimming can become disguised cherry-picking

An unusual answer is not necessarily an error; it may contain information the majority missed. Trimming also cannot repair correlated-error or selection-bias: if the whole sample leans the same way, removing its extremes may merely produce a tidier version of the same bias. The method is defensible only when “outlier” is defined independently of the result you want.

Set the exclusion rule before seeing the winner

Tomorrow, when combining estimates, calculate the result twice—once with every value and once after applying a predeclared plausibility rule. If trimming materially changes the answer, report both and explain exactly which values were excluded and why.

Episodes that teach this