Propensity score synthetic augmentation matching using generative adversarial networks (PSSAM-GAN)

Area of Science:

• Biomedical Sciences
• Causal Inference
• Machine Learning

Background:

• Causality is vital in biomedical sciences for actionable prediction models.
• Observational data (e.g., EHRs) often contain biases like confounding.
• Randomized controlled trials are ideal but not always feasible for causal effect estimation.

Purpose of the Study:

• To address limitations of existing methods like propensity score matching (PSM) and inverse probability weighting (IPW).
• To propose a novel deep learning approach, PSSAM-GAN, for causal inference.
• To maintain sample size and avoid weighting instability in observational studies.

Main Methods:

• Developed Propensity Score Synthetic Augmentation Matching using Generative Adversarial Networks (PSSAM-GAN).
• PSSAM-GAN generates synthetic matches to balance covariate distributions between treatment and control groups.
• The method was evaluated on semi-synthetic and real-world observational datasets.

Main Results:

• PSSAM-GAN effectively created balanced datasets, reducing the need for weighting or dropout.
• The approach demonstrated competitive performance in estimating treatment effects.
• It showed advantages over simple GANs and comparable results to other deep counterfactual learning architectures.

Conclusions:

• PSSAM-GAN offers a promising deep learning solution for causal inference from observational data.
• It enhances the reliability of treatment effect estimation by improving data balance.
• The method is versatile and can be integrated with various prediction models.