PyBalance Utilities
pybalance is a python library for confounding adjustment in non-randomized populations. Given a “pool” and a “target” population that differ systematically on a set of covariates, pybalance finds either a matched subset of the pool or a set of weights on the pool that make it resemble the target – a key step in any causal inference analysis built on observational data.
Everything in pybalance is driven by a researcher-defined balance metric (e.g. standardized mean difference), which is optimized directly rather than treated as a side effect of some other model. This includes propensity score matching: instead of fitting one model and hoping it balances well, pybalance searches over propensity-model hyperparameters and keeps whichever fit optimizes the chosen balance metric. See the Introduction for the underlying problem formulation.
Features
Matching: linear program (integer solver), evolutionary/genetic search, and propensity score matching, plus matching against an aggregate (published Table 1) target.
Weighting: Matching-Adjusted Indirect Comparison (MAIC), general entropy balancing, and propensity-score (IPTW) weighting.
A variety of balance calculators for defining and measuring balance.
Visualization tools for inspecting covariate balance before/after.
Dataset simulation utilities for testing and demos.
Get started by following the Installation instructions instructions, then work through the Demos. Questions or issues are welcome on GitHub.
Getting Started
Demos - Core
Demos - Matching
Demos - Weighting