NeurIPS Reproducibility Challenge Report: Adapting Neural Networks for the Estimation of Treatment Effects

Xiaoting Wang, Hao Li, Bozhong Lu

Abstract: Causal inference is a fundamental problem in many fields. In this report, we consider the problem of causal effects estimation based on a semi-synthetic dataset. To be specific, this estimation task was completed by using different neural network architectures including Dragonnet, TARNET and Nednet. The performance of these network architectures is compared. Specially, we focus on the effect of Dragonnet and Dragonnet with target regularization models on treatment estimation, which are two adaptive approaches implemented by original authors. To study the robustness of Dragonnet architecture, we explored different hyperparameters. It can be found that most of the hyperparameters we have tried have a slight impact on the Dragonnet architecture. Also, we tried to modify the number of hidden layers in Dragonnet for the outcome models. We found that a minor improvement has been shown by adding more hidden layers. We further compared the effect of two different plug-in treatment effect estimators. The results have shown that Dragonnet has decent performance even with a conditional-outcome-only estimator.

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=SkeOIVBeUH