Skip to content

Files

Latest commit

 

History

History

notebooks

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

CausalTune Notebooks

This notebook serves as a good starting when exploring CausalTune. It encompasses an introduction to

  • estimator & metric selection,
  • choosing training time budgets,
  • finding the best estimator and configuration, scoring and visualisation.

Causaltune models require work based on a CausalityDataset instance which includes all relevant data and information to build the causal graph. This notebook briefly outlines how to define a CausalityDataset.

A guide on how to use CausalTune for AB test evaluation.

It explores both variance reduction techniques leveraging additional features and segmentation analysis by feeding conditional average treatment effects (CATEs) into the automated segmentation analytics tool wise-pizza.

This analysis walks the user through estimating effects of hypothetical assignment policies, different from the actual one, even after the experiment has been completed.

The estimators that CausalTune is based on (e.g. doubly robust learners, metalearners etc.) mostly require propensity score weighting. CausalTune therefore requires propensity score weights. This notebook displays the different methods that can be used to bild the propensity score weights.

Those include

  • [Default:] use a dummy estimator that assumes random sampling and estimates treatment probabilities from the dataset,
  • Letting AutoML fit the propensity model,
  • supply an sklearn-compatible classification model,
  • supply an array of custom propensities to treat.

Standard errors are important for statistical inference. CausalTune allows for computation of standard errors of the best estimator as identified by CausalTune. The standard errors are computed with econml methods which includes analytical estimates for some estimators and bootstraps for others.

CausalTune can compare instrumental variable (IV) estimators based on the energy distance, in the special case where the instrumental variable controls access to a feature, and the treatment is the user choice of using this feature (if available to them). This notebook shows how to run the estimator selection and how to compare IV estimators.

It also walks through how to interpret the energy distance and additional result visualisation.