Frequentist calculator: integrate CUPED variance reduction

[BasLinders]

When looking for statistical significance, the tested data contains a lot of noise (variance). Some of that variance can be explained if we know what the pre-experiment data tells us. To adjust our experiment data with that knowledge, we incorporate CUPED - Controlled experiment Using Pre-Experiment Data. This reduces variance and can be decisive in edge-cases for determining statistical significance.

[BasLinders]

Because the tool works with aggregate data, we can only estimate theta (θ, the scaling factor) online, which can lead to inaccurate results, especially when variance is heterogeneous. To implement CUPED correctly, the calculator needs user-level data, specifically each user's pre-experiment covariate (X) and their outcome KPI (Y) (binomial conversion). This allows for the accurate calculation of the required variance of the covariate (Var(X)) and covariance between the outcome and covariate (Cov(Y, X)), needed to determine the optimal theta and improve accuracy.

This means that an option to upload a CSV for user-level data is needed.