I'm running a AB testing at my company, but the split is somehow different than 50-50.

Is my test still correct? Can the split be 55-45, or 60-40, and so on?

When should I be worried?

Sorry if the question is too vague, but I believe it's very straightforward and I haven't seen any similar question.


There is hardly any problem at performing a A/B test whith unequal sample sizes. But you don't want samples too unbalanced because the precision of the test depends more on the lesser sample size than on the greater (the biggest they are, the better, I bet you already know).

Therefore, an A/B test with sample sizes 50 and 50 is more powerful than the same test with sample sizes 20 and 80.


@carlo has covered the case where the experimenter specifies an uneven split (e.g. 80/20 split) prior to starting the experiment, why that is fine, and scenarios where it might not be ideal.

However, the wording of the original question might also be suggesting that the OP has specified a 50/50 split prior to starting the experiment, and is seeing something that is not 50/50 during analysis.

If the discrepancy is too extreme (beyond randomness), it is said that a Sample Ratio Mismatch (SRM) has occurred. Many things can cause SRM, though it is usually the first indication that something is wrong with your experiment, and the results may not be trustworthy. See e.g. this article for some examples and existing work on the online A/B test space.

Whether a discrepancy is extreme enough to raise alarm depends on how many samples were involved in the first place. If one intended for a 50/50 split, a 49/51 split might be fine if the experiment involved 100 visitors, but definitely not if the number is 10 million. A common way to test for SRM is to use a chi-squared goodness of fit test. This article suggests multiple ways to implement the test.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.