Individualized treatment effect estimation with compromised adversarial nets

Estimating individualized treatment effects (ITE) in causal inference mainly relies on the assumption of strong ignorability, which is often difficult to validate in practice. Moreover, the true value of ITE is unobservable. These factors make it difficult to obtain an appropriate loss function to estimate the ITE. In this paper, a novel framework that leverages generative adversarial networks (GANs) is proposed to estimate ITE using a bounded loss function under the strong ignorability condition. The bound is obtained based on the supervised loss due to the generator, and the unsupervised loss is due to the discriminator. In the proposed method, the discriminator estimates the conditional density of the estimated unobserved outcome and the conditional density of the observed outcome. The discrepancy between these conditional densities accounts for the unsupervised loss. Furthermore, we developed the Compromised Adversarial Network (ITE-CAN), an advanced ensemble model specifically designed to mitigate common limitations of GANs, such as mode collapse. The theoretical foundation of ITE-CAN is established through a series of theorems that validate its efficacy. Through extensive simulations and empirical analysis on two benchmark datasets, we demonstrate that ITE-CAN consistently outperforms existing methods in terms of estimation accuracy at the individual level. This contribution underscores the significance of our approach in enhancing the precision of individualized treatment effect estimation.