« Explain Your Wins | Main | Tolstoy on Patriotism »

March 30, 2007


I think this often goes under the name of "systematic error."

Repeating an experiment with systematic error is a special case of non-independent evidence.

Eliezer: glad you followed up on Robin's suggestion to do this. The examples in my earlier post on "Useful Bias", then involve resorting to systematic error to deal with problems arising out of processes with unsystematic errors. Similarly, the examples other commenters gave, such as alcoholics foreswearing all drinks or persons setting their watches five minutes early, would also be cases of using sytematic error to avoid problems caused by unsystematic errors.

I think this same basic formula is behind the argument for majoritarianism: the crowd's consensus (average) squared-error, plus the variance in the crowd, equals the expected squared-error for a random person in the crowd. Therefore the crowd consensus view has a lower expected squared-error than the average squared error for individuals in the crowd. Hence a random participant will do better to substitute the crowd consensus for his own estimate.

I'm reading a book, "The Difference", by Scott E. Page, which discusses how and when crowds do well, and he calls it the Diversity Prediction Theorem: Given a crowd of predictive models, Collective Error = Average Individual Error - Prediction Diversity.

It is good discription ofsystematic and random error,but how do u differentiat error and bias

The comments to this entry are closed.

Less Wrong (sister site)

May 2009

Sun Mon Tue Wed Thu Fri Sat
          1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30