Overview¶

CForest provides classes and functions to estimate heterogeneous treatment effects in a potential outcome framework.

The algorithms which are implemented in CForest draw heavily on the ideas formulated in Athey and Imbens (2016) and Athey and Wager (2019), who first proposed the Causal Tree and Causal Forest algorithms.

Here it is appropriate to also refer to Athey, Tibshirani and Wager (2019) who combine and generalize the ideas of causal and random forests. Further, they provide an R package (https://github.com/grf-labs/grf) which can be used to compute everything CForest computes and much more.

Example¶

A complete working example can be found in section example.

Warnings¶

Originally there were two reasons for the creation of CForest, given that there already exists a well maintained package. First, through the implementation of algorithms one often learns more about the inner workings of given algorithms. And second, the implementation in grf is written in C++ and wrapped in R, which makes it very hard to explain details of the implementation to students and researchers not trained in C++. This is why we were interested in a Python implementation that uses numpy and numba, which allow for great readibility.

As of right now CForest is still under development and should not be used other than for experimental reasons. An official version will become more likely once we benchmarked our implementation to grf with positive results.