In project settings under Experimentation, specify default covariates for your company:
Covariates can be sourced from the assignment source for the experiment, or from Entity Property Sources.Properties with an Entity Property Source name attached come from a property source and will always be included on the experiment. Properties without an EPS come from an assignment source, and will be used if the column exists on the assignment source of a given experiment.
One tool our framework enables is using metric:metric covariates; for example, using units’ pre-experiment clicks as a covariate for units’ in-experiment revenue. We decided against using this for several reasons:
Consistency. It should not be the case that adding a new metric to your analysis significantly alters the results of other metrics
It’s not necessary. If you have key metrics that function as covariates, we recommend explicitly providing these - for all metrics - as a covariate in an entity property source. This achieves the same result, without “black box” outputs coming from unknown metric covariances
Statsig has a strong opinion that throwing in arbitrary covariates based on an experiment’s metric selection is a bad practice, and it is better to explicitly include key metric covariates as numerical covariates; please reach out in Slack if you would like to discuss!
You can turn off CUPED in your pulse results, and can create a project-level setting to enforce this. This will still run CURE, however, which entails some amount of compute cost. To avoid running CURE, you can turn it off on a given metric by un-checking the CUPED option in the metric’s setup page.
