I'm looking for a mathematical algorithm to proof significances in multivariate testing.
E.g. Lets take website tests having 3 headlines, 2 images, 2 buttons test. This results in 3 x 2 x 2 = 12 variations:
h1-i1-b1, h2-i1-b1, h3-i1-b1,
h1-i2-b1, h2-i2-b1, h3-i2-b1,
h1-i1-b2, h2-i1-b2, h3-i1-b2,
h1-i2-b2, h2-i2-b2, h3-i2-b2.
The hypothesis is that one variation is better than others.
I'd like to to know with which significane one of the variations is the winner and how long I have to wait, that I can be sure that I have statistically a winner or at least have an indicator how sure I can be that one variation is the winner.
So basically I'd like to get a probability for each variation telling me wether it the winner or not. As the tests runs longer some variations drop in probability and the winner increases.
Which algorithm would you use? Whats the formula?
Are there any libs for this?
You can use a chi-square test. Your null hypothesis is that all outcomes are equally likely; when you plug in the measured counts for each of the 12 outcomes, you get out a number telling you the probability of getting a set of 12 counts as extreme (i.e. as far away from equally distributed) as this. If the probability is sufficiently small (typically < 5% or < 1%), you conclude that the null hypothesis was wrong.