Crowds within crowd found to outperform 'wisdom of the crowd'

[Jim__]

From phys.org:

A team of researchers affiliated with institutions in Argentina, the U.S. and Germany has found that there is a way to improve on the "wisdom of the crowd"—separate the people in a given crowd into smaller groups and let them talk about an issue at hand before an answer is given. In their paper published in the journal Nature Human Behavior, the group describes an experiment they carried out with a large crowd of volunteers, and what they learned from it.

Most people have heard of the "wisdom of the crowd," in which individuals in a crowd are privately asked to give an answer to a question, such as how many jelly beans are in a jar. When averaged together, the answer given by the crowd will generally be better than for any given individual. Now, it appears there may be a way to improve the accuracy of a crowd.

...

In their experiment, the researchers asked 5180 people at a 2015 TED talk in Buenos Aires to answer several simple questions, such estimating the height of the Eiffel Tower, or the number of goals scored in the 201 FIFA World Cup. Each of the respondents was asked to give an answer privately, and then to join with a group of four other individuals to discuss the question. After one minute on each topic, each of the groups was asked for mutually agreed-upon answers to the same questions.

The researchers report that the average answers of the 280 groups of five (not all of those who answered the question individually were willing to join a group) was 49.2 percent more accurate than the average crowd response as a whole. This, the team suggests, indicates that allowing some logical discussion into the equation can improve results.

[DetlefK]

1. That reminds me of a method statisticians use to "clean" their data.

Normally one would expect that the more data you have, the more reliable the result is. But the problem is that extreme outliers (e.g. data from biased source) can skew the calculations.

Statisticians use the trick to simply cut the biggest and the smallest value from the data-set. Or the two biggest and two smallest. Or three biggest and three smallest. And so forth.
This in itself skews the result a little bit (because you are leaving out good data), but it protects the data-set from skewing too hard (because you are leaving out bad data).

[Hortensis]

2. Interesting, thanks. Reminds that these days a lot of people

get on the internet and find their own "crowds" to discuss questions with. What are researchers finding from this alteration of a fundamental factor?