Variational Disentangled Generative Model for Individualized Treatment Effect Estimation

Estimating individualized treatment effect (ITE) from observational data is a challenging problem due to the lack of counterfactual outcomes and the presence of selection bias. Existing ITE estimation methods, which primarily focus on disentangled representation learning, have achieved significant success in estimating treatment effects. However, precisely learning disentangled latent factors remains an open problem. In this paper, we propose a Variational Disentangled Generative Model (VDGM) based on a collaborative strategy of Variational Auto-Encoder (VAE), Generative Adversarial Net (GAN), and Mutual Information Network Estimation. VDGM infers latent factors from observed covariates by VAE, disentangling factors into three disjoint sets corresponding to the instrumental, confounding, and adjustment factors. The disentangled latent factors can then be used to balance the distribution discrepancy between treatment and control groups, thereby reducing selection bias. Furthermore, VDGM addresses information loss by preserving useful predictive information under the regularization of a mutual information estimator. It also uses mutual information estimators to maintain the independence of disentangled latent factors. Additionally, our model uses the GAN framework to generate counterfactual outcomes for ITE estimation. Experimental results on synthetic and real-world datasets show that the proposed method outperforms the baselines in achieving low errors in ITE estimation.