Saturday, December 05, 2009

Interactions and Data Analysis

Re: Context Matters

Multivariate testing is great. And it a great way to determine interactive factors - which are essentially not possible to determine with one variable at a time testing (though a smart person can see indications within this type to view them).

Your example though seems to largely be about properly stratifying the data for optimization. While it is always difficult to tell with short examples it seems like it may well be that you have 2 different audiences and a solution that, for example, intercepts the search engine traffic and gives them some context might help a lot (you see this on many blogs where they say - "I see you find us search on X - you may also be interested in Y...

But the biggest point I think your story illustrates is the importance of the experimenter. They need to think. Their role is not just to calculate some numbers and whatever number is higher wins. George Box: "it’s not about proving a theorem, it's about being curious about things. There aren't enough people who will apply [DOE] as a way of finding things out."

Related: YouTube Uses Multivariate Experiment To Improve Sign-ups 15% - Using Design of Experiments - Statistics for Experimenters

No comments: