# Useful Stuff

## Calculating Statistically Significant Split Tests

If, like me, you’ve run lots of split tests on lots of sites, you’ve probably used the split testing tool’s own confidence thresholds to determine whether your variation is the winner.

You may wonder how it’s calculated.  Or if it can be trusted?

In Google Analytics experiments, it looks like this:

Since watching the talented Jason Cohen present on false positives in testing, I prefer to use a simpler method.  The A/B Hamster Method!

Assuming the control and variations have the same sample size:

 Version Visits Conversion Rate Conversions Control 1000 5.00% 50 Variation 1000 6.00% 60

* Define N as the number of conversions: so N = 110 in the example above (50 + 60)

* Define D as  the difference between the winner and loser divided in half: So, using the above data, this = 5 ((60-50) / 2)

* The test result is statistically significant if D2 is bigger than N.  So, above D2 = 25 (5 x 5), and N = 110, so this is not a significantly significant test.

So… what?  Run it for longer?  Sure… lets run it for another week (you should always run tests for a least a week, so cover a full 7 days of weekly seasonality):

 Version Visits Conversion Rate Conversions Control 2000 4.84% 97 Variation 2000 6.43% 129

So, we have slight changes in those conversion rates, but more importantly:

* N = 226 (97 + 129)

* D2 = 256

So... D2 is greater than N, and therefore the above is now statistically significant.  Stop the test, we have a winner!

I just love this and really helps non statistics people get how confidence thresholds are calculated.  Satisfying the above measure to statistical significance actually gives you a 96% confidence threshold, so you have an extra 1% of reassurance!

Summary-Arium:

* If D2 is greater than N, you can be confident you’ve found a winner in your split test.

* Give split tests time to find a conclusion

A Nod to…

Jason Cohen and the AB Hamster method to determine statistical significance