A2075
Title: Self-balancing neural network: A novel method to estimate average treatment effect
Authors: Yuqi Qiu - East China Normal University (China) [presenting]
Abstract: In observational studies, confounding variables affect both treatment and outcome. Within this causal framework, instrumental variables further influence the treatment assignment mechanism. Such dependencies distinguish the study context from standard randomized controlled trials, where assignment is random. Consequently, the estimated average treatment effect (ATE) becomes biased. To address this issue, a standard approach is to incorporate estimated propensity scores. However, these methods incur the risk of misspecification in propensity score models. To resolve this limitation, the Self-balancing neural network (Sbnet), which allows the model to endogenously obtain its pseudo propensity score from a balancing net, is proposed. The Sbnet estimates the ATE by using the balancing net as a key part of the feedforward neural network. This formulation resolves the estimation of the ATE in a single step. Moreover, a multi-pseudo propensity score framework is developed, which is estimated from diversified balancing nets and used for enhancing estimation accuracy. Finally, the proposed methods are compared with state-of-the-art methods on four simulation setups and real-world datasets. Empirical results demonstrate that the Self-balancing neural network yields performance superior to state-of-the-art baselines.
