facebookresearch/balance 0.10.0 on GitHub

News

This version we transitioned ipw to use sklearn. This enables support for newer python versions as well as the Windows OS!
Updated Python and package compatibility. Balance is now compatible with Python 3.11, but no longer compatible with Python 3.8 due to typing errors. Balance is currently incompatible with Python 3.12 due to the removal of distutils.
Update license from GPL-v2 to the MIT license.

Dependency on glmnet has been removed, and the ipw method now uses sklearn.
ipw method uses logistic regression with L2-penalties instead of L1-penalties for computational reasons. The transition from glmnet to sklearn and use of L2-penalties will lead to slightly different generated weights compared to previous versions of Balance.
Unfortunately, the sklearn-based ipw method is generally slower than the previous version by 2-5x. Consider using the new arguments lambda_min, lambda_max, and num_lambdas for a more efficient search over the ipw penalization space.