Get Started With CURE

Defaults

To add additional covariates

Project defaults

In project settings under Experimentation, specify default covariates for your company:

You can source covariates 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 Statsig always includes them on the experiment. Properties without an EPS come from an assignment source and apply if the column exists on the assignment source of a given experiment.

Experiment settings

Per-experiment, specify additional covariates or remove covariates specific to your analysis:

Statsig populates this list from the experiment assignment source columns and relevant Entity Property Sources.

Metric:Metric

One tool this framework enables is metric:metric covariates, such as using units' pre-experiment clicks as a covariate for units' in-experiment revenue. Statsig doesn't use this approach for several reasons:

• Consistency: adding a new metric to your analysis should not significantly alter the results of other metrics.
• Unnecessary complexity: if you have key metrics that function as covariates, explicitly provide these for all metrics as a covariate in an entity property source. This achieves the same result without opaque outputs from unknown metric covariances.

Statsig recommends explicitly including key metric covariates as numerical covariates rather than adding arbitrary covariates based on metric selection. Contact the team in Slack to discuss further.

Preventing adjustments

You can turn off CUPED in your Pulse results, and can create a project-level setting to enforce this. CURE still runs when CUPED is turned off, which incurs some compute cost. To avoid running CURE, turn it off on a given metric by clearing the CUPED option in the metric's setup page.