Debiased Causal Tree: Heterogeneous Treatment Effects Estimation with Unmeasured Confounding

Authors: Caizhi Tang, Huiyuan Wang, Xinyu Li, Qing Cui, Ya-Lin Zhang, Feng Zhu, Longfei Li, Jun Zhou, Linbo Jiang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The computational feasibility and statistical power of our method are evidenced by simulations and a study of a credit card balance dataset.
Researcher Affiliation Collaboration 1Ant Group, 2School of Mathematical Sciences, Peking University
Pseudocode No The paper describes the steps of the algorithm within the text (Section 2.3) but does not provide a formal pseudocode block or an algorithm box.
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of the proposed method (Debiased Causal Tree/GBCT). It mentions 'ECOML (Keith et al., 2019)' for benchmark implementations, which is a third-party library.
Open Datasets No The paper describes a 'credit card balance dataset' used in experiments, stating it 'comes from a randomized controlled trial (RCT) by a commercial finance company'. However, it explicitly mentions that 'The dataset does not contain any Personal Identifiable Information (PII)', 'is desensitized and encrypted', 'was destroyed after the experiment', and 'is only used for academic research', indicating it is not a publicly available or open dataset.
Dataset Splits Yes A total of 20000 samples are generated and randomly split into training and validation sets 10 times.
Hardware Specification No The paper does not specify any hardware details such as GPU models, CPU types, or cloud computing instances used for running the experiments.
Software Dependencies No The paper mentions software like 'Light GBM' and 'ECOML' but does not provide specific version numbers for these or any other software dependencies crucial for replication.
Experiment Setup Yes The number of trees in ensemble models (including boosting and bagging) is 200, the sub-sample ratios of instance and feature are 0.8 and 0.6 respectively, and the learning rate is 0.3. The maximum depth of each tree in forest-based (GRF, DML-RF and DR-RF) and boosting-based (meta learners and GBCT) methods is 10 and 3, respectively, where it should be noted that due to the respective characteristics of the bagging and boosting frameworks, the trees in random forests are generally deeper.