Files
Abstract
Causal Inference has been increasingly drawing attention during the past decades, especially in this pandemic when people care more about effective protection approaches against COVID-$19$. The potential outcome framework is the primary theoretical framework for researchers to illustrate and estimate causal effects. However, this framework includes a controversial assumption called \emph{Ignorability Assumption}, which is usually difficult to justify in real-world scenarios. Besides, the data size has grown tremendously during the past decades, bringing the "curse of dimensionality" to many traditional causal inference models. In this thesis, we propose a model named Importance-sampling Causal Effect with Disentangled Variational Auto-Encoder (ICEDVAE), which combines Causal Bayesian Network and Variational Auto-Encoder to estimate causal effects by relaxing the Ignorability Assumption and overcome the curse of dimensionality. Numerical studies show that our model is comparable with other State-Of-The-Art models and has the best performance when the treatment effect is homogeneous over different subjects.