Disentangled Representation Learning for Causal Inference With Instruments.

Area of Science:

• Causal inference
• Machine learning
• Statistics

Background:

• Inferring causal effects from observational data is challenging due to latent confounders.
• The instrumental variable (IV) approach is a key method for addressing this, but often requires known IVs or strong assumptions.
• Existing methods limit the applicability of IV analysis.

Purpose of the Study:

• To propose a novel method for causal effect estimation under latent confounding.
• To relax the requirement of a known instrumental variable (IV).
• To develop a robust IV-based estimator using representation learning.

Main Methods:

• A variational autoencoder (VAE)-based disentangled representation learning approach is proposed.
• The method learns an IV representation from data with latent confounders.
• This IV representation is then used for unbiased causal effect estimation.

Main Results:

• The proposed algorithm successfully learns an IV representation from observational data.
• It enables unbiased estimation of causal effects in the presence of latent confounders.
• Experiments show superior performance compared to existing IV and VAE-based estimators.

Conclusions:

• The VAE-based disentangled representation learning method offers a practical solution for causal inference with unknown IV proxies.
• This approach enhances the applicability of instrumental variable methods.
• The findings demonstrate significant improvements in estimating causal effects from complex observational data.