Andrew Gelman has a great blot post on this:

http://www.stat.columbia.edu/~cook/movabletype/archives/2010...

In a nutshell, the problem arises from trying to estimate small effects and then screening results on statistical significance.