What significance test should you use for a percentage metric with more than two experiments?
For example,
Version | Clicks | Impressions
A | 5 | 1,763
B | 4 | 1,672
C | 2 | 1,689
How sure are we that verison A really is superior to the other two?
In the past I personally have done a pairwise G-tests between the top and the bottom, multiplying the confidence by a fudge factor of n choose 2 to account for the fact that there are n choose 2 possible pairs that could have been the most extreme. Theoretically this is overly conservative, but it worked for me.
See http://elem.com/~btilly/effective-ab-testing/ for more.
