Hello everyone, I have some questions about formula of

proportion metrics

. 1. I read that growthbook use two-tailed t-test in docs. but is it possible to use t-test for 'proportion metrics'? I have known that t-test is used not for categorical(or proportion) but for continuous data. if possible, could you explain me math theory related to it? 2. if proportion metric is based on binomial distribution, as far as I know variance should be

np(1-p)

. but there isn't

n

and only

p(1-p)

is in document. could you explain me where

n

has gone? I used this link as reference.

Hi. I passed on these questions to our data science team. Here's how they responded:

Great questions! 

1. The central limit theorem applies for proportion metrics and for continuous data.  So even if the underlying metric is not normally distributed, its sample mean is.  We are comparing means, which is why the t-test is appropriate. 2. The variance of an individual observation is ​`p(1-p)`​.  Summing ​`n`​ independent random variables, the variance of the sum is ​`np(1-p)`​.  Taking their mean, the variance is ​`p(1-p)/n`​.

Thank you for fast reply which is really helpful to me. anyway, how about supplementing these math basis in document? if you guys do so, it is so helpful to many growthbook users.

We can do that - but we're also open source, you're welcome to edit the documentation as well

oh, there is way of contributing to documents. thanks!