Provided Estimators¶
WhyNot comes equipped with a collection of estimators to enable rapid comparisons of methods on new benchmarks and with new experimental designs. We hope this both allows users of causal inference tools to compare methods and choose methods that perform well in situations of interest. We also hope that WhyNot will allow producers of new causal inference methods to rapidly evaluate their algorithms against a rich set of baselines on common datasets.
Beyond estimators provided by WhyNot, the data generated in Dataset
is easily accessible as NumPy arrays that can be feed into the user’s own
estimators.
>>> import whynot as wn
>>> dset = wn.world3.PollutionRCT.run(num_samples=200)
>>> covariates, treatment, outcomes = dset.covariates, dset.treatments, dset.outcomes
# Replace with your favorite estimator!
>>> estimate = wn.algorithms.ols.estimate_treatment_effect(covariates, treatment, outcomes)
Average Treatment Effect Estimators¶
- Linear Regression
- Matching, using the package Matching.
- IP Weighting, using the package WeightIt.
- Causal Forests, using the package GRF.
- Causal Bart, using the package BartCause.
- TMLE, using the package TMLE.
- Double machine learning, using the package econml.
By default, linear regression and propensity score matching are included with WhyNot. To minimize dependencies for the original package, the remaining estimators ship with the companion WhyNot-Estimators package. For more installation details, see Installation.