Long-term causal effects estimation via latent surrogates representation learning.

Area of Science:

• Causal inference
• Machine learning
• Biostatistics

Background:

• Estimating long-term causal effects from short-term surrogates is crucial in fields like medicine and marketing.
• Existing methods often oversimplify by ignoring unobserved surrogates or misclassifying short-term outcomes.

Purpose of the Study:

• To develop a flexible method (LASER) for estimating long-term causal effects in realistic scenarios with mixed observed and unobserved surrogates.
• To address limitations of current approaches that fail to account for partially observed surrogates and proxies.

Main Methods:

• Utilized an identifiable variational autoencoder to learn latent surrogate representations from all available surrogate candidates.
• Developed a theoretically guaranteed and unbiased estimation approach for long-term causal effects using the learned representations.

Main Results:

• Demonstrated the effectiveness of the LASER method on both real-world and semi-synthetic datasets.
• Showcased improved accuracy in estimating long-term causal effects compared to existing approaches.

Conclusions:

• The LASER method provides a more realistic and effective approach to estimating long-term causal effects.
• LASER's ability to handle unobserved surrogates and their proxies enhances its practical utility in various applications.