The package is developed for treatment recommendation & pairwise treatment individual effect estimation (ITE/CATE/HTE). It includes some published methods such as S-learner, T-learner, X-learner, and R-learner and some newer/unpublished methods like reference-free simplex R-learner and de-Centralized-learner. Angle-based direct learning (AD-learner), an outcome-weighting-based treatment recommendation method is included as well.
Notably, the S-, T-, and X-learner follows the nomenclature of Kunzel et. al. There are many causal inference methods follows the S- and T-learner structure, like causal boosting, causal forests, etc. X-learner is designed for two-treatment setting and is extended for multiple treatment case in the package.
R-learner is proposed by Nie and Wager mainly for two-treatment setting. Here we extend it for multiple treatment setting as well but we found that different reference group may cause different results.
AD-learning is relatively new method by Qi et. al. which does not involve the estimation of treatment effects but ITR. Individualized Treatment Rules/Regimens/Recommendations (ITR), a mapping from covariate space to treatment space, is a treatment decision rule that determines the optimal treatment directly given the subject's covariates.
Reference-free R-learner is proposed by Zhou et. al. which follows the idea of R-learner but fixes the issue of recommendation inconsistency in multi-armed scenarios.
de-Centralized-learner is now under development. For more details, please wait for publication.
Some technical details can be found on author's website.
To install the package:
devtools::install_github("junyzhou10/MetaLearners")
Now binary outcomes is allowed for S-, T-, and de-Centralized learner. Notice that, binary outcome is not supposed to be supported by X-, and R-learner methods. AD-learner is capable, and will be developed soon.