A/A Split Testing | Hacker News Reader

A/A Split Testing

getelastic.com

16 pointsfezzl13 comments

Essentially the article says (paraphrased) "Wow, look, if you A/B test two versions of the exact same design, you get slightly different conversion rates. Can split testing be trusted?"

I don't understand why they find this surprising. Of course there's going to be some variation in the conversion rates. This is the reason why GWO reports statistical significance.

While I don't disagree with your main point, there is another question just under the surface:

Is the method by which you're distributing sessions into your tests somehow biased?

I'm going to call A' the "new" (but identical) test, and A the original test.

If you've been running A for a long time, and now add A', what is the chance that the visitor populations between A and A' will be different enough to drive statisically significant differences that are population-related rather than test-content-related?

Put slightly differently: If your returning session conversion rate is higher than your first session conversion rate, you will need to take some pains to ensure that each of the tests is getting a fair shake at the traffic. In many cases, that means ending test A, and creating a new test a and a' such that neither a nor a' has an advantage. It's easy to ASSume that there is no meaningful test bias, while the reality is that it's quite easy to have test bias creep in.

Another common bias is if your are running multiple split tests in parallel you need to assign each visitor to an option in each split test at RANDOM. If you don't you'll end up with 50% of the users going though all the same 50% of the options and the other 50% of users going through the other 50% of the options.

sokoloff makes a great point.

Something I learned from years as a QA lead: always ask that every test you do has a failure and a success case. Surprisingly often, the tests that show your program works fine aren't actually doing their job properly, so they also need to be checked.

In other words, testing your implementation of A/B testing is a Good Thing. Even if you don't understand that statistics behind it, at least you can see the same variance occurs in A/A testing, which tells you to go read up some more on what you're doing wrong.

Can split testing be trusted? No, not if you misunderstand the most basic concepts. Non-statically significant results are, surprise surprise, not statically significant.

This is exactly the problem with people implementing split testing and they're hurting their business.

They hear about it, run off and implement the a/b versions of their pages and add some tracking. Proceeding then to draw incorrect conclusions and make bad decisions based on them. They don't go read up on it enough to understand the statistics behind it. I used my stats book from Uni (I knew it had a better use than a laptop stand, a java book now has that honour) but there are plenty of tutorials out there. For example: http://visualwebsiteoptimizer.com/split-testing-blog/what-yo...

It really is a toss up between statistics and medicine for which is the most misunderstood of the sciences.

Unfortunately, nobody wants to learn about statistics -- they just want it to work so they can make more money. I just wrote a post about behavioral change and understanding your users here: http://www.iwillteachyoutoberich.com/blog/why-personal-finan...

That's why sites like Visual Website Optimizer are getting increasingly good. They use plain language to explain when to keep running the tests.

Again: It's unrealistic to expect lay users to learn statistics. They don't want to become statistical experts, they usually just want to increase sales. This is where software can help.

I think even laymen should learn the why of the statistics, a few paragraphs about statistical significance would have helped the article's author I should think.

As an aside, wow another one of my favourite bloggers responding to a comment of mine on HN this week!

Even statistics can lie. Having statistically significant results does not guarantee you that the variation will work better than the control. What it actually guarantees is that it is HIGHLY likely that variation will work better than the control. So, it ultimately boils down to what are the stakes involved in an A/B test.

Not just highly likely, likely to specified probability. Typically 95% or 99%. If you want to be more certain (and have the quantity of data) you can look for 99.9% or greater. I'm not sure how you could see that as statistics "lying" :)

EDIT: I just noticed your username and I realise that you probably already understand this :p I still think it's a bit of an odd way of putting it though.

Better not let Zed Shaw read this.

http://www.zedshaw.com/essays/programmer_stats.html

It seems that if an A/A test is showing a variance then your sample size is too small.

There is always a variance. Even with huge tests.

The question is whether it is statistically significant.

Reply on news.ycombinator.com