Hi team, I have a question on setting up a duration metric for experiment analysis. One of the options allows me to choose how to aggregate in case the same user has more than one measure of the metric (as per screenshot). However, the only enforced aggregation across users is

AVG

, which makes sense in general but for my specific case could lead to some noisy results given some annoying outliers in the metric distribution. Therefore, I would like to use

MEDIAN

to aggregate the metric across users, and as there’s no option to do so when creating the metric I was wondering if there’s a workaround that people use or if it will be possible in the future to also select the aggregation method across all users.

Hi Fabrizio. It sounds like you might be interested in winsorization which Growthbook does have the ability to do: https://docs.growthbook.io/app/metrics#capped-value

Thanks James! Is the

percentile capping

available from a specific version on?

Are you self hosting?

Yes we are

Ah yes it looks like it was added three weeks ago. So it might not have made it in the last release, but I do believe we are having a new release today.

No it’s fine. I’ll ask the managers of our instance to update the release when it comes out - not urgent. Thank you so much for the help! 🙂

However, the only enforced aggregation across users is

AVG

, which makes sense in general but for my specific case could lead to some noisy results given some annoying outliers in the metric distribution.

You're correct that the enforced aggregation across users is

AVG

however, the field you showing is the user-level aggregation. Normally you actually want to

SUM(value)

as per the default so that your user-level total is the sum of all of their values, and then in analysis we divide by the N users which gives you the average per user (so the

AVG

happens internally in our stats engine). If you choose

AVG(value)

as your custom aggregation, you're doing something very different, where each user will get the average of the rows that are accorded to them. Instead of looking at "duration per user" you're looking at "average duration per user" and comparing those across variations.

The new release with percentile winsorization is live: https://github.com/growthbook/growthbook/releases/tag/v2.3.0