DPN-SA

DPN-SA performs propensity score matching and counterfactual outcome prediction for causal inference in observational studies using a sparse autoencoder-based deep neural architecture to handle high-dimensional covariates and nonlinear or nonparallel treatment assignment.

Key Features:

• Propensity Score Matching: Matches subjects by estimated propensity scores to reduce bias in treatment effect estimation.
• Counterfactual Prediction: Predicts counterfactual outcomes under alternative treatment scenarios to enable causal effect estimation.
• Sparse Autoencoder-based Dimensionality Reduction: Leverages sparse autoencoders to process and extract features from high-dimensional covariates such as electronic health records.
• Robustness to Assignment Bias: Mitigates biases including discrimination in treatment assignment to improve causal estimates.
• Comparative Performance: Demonstrated improvements of 36%–63% over logistic regression and LASSO and 6%–10% over deep counterfactual networks with propensity dropout (DCN-PD) across evaluated datasets.
• Treatment Effect Metrics: Produces accurate average treatment effect (ATE) and average treatment effect on the treated (ATT) estimates with low variance.
• Heterogeneity Precision: Achieves low mean squared error in precision on effect's heterogeneity, capturing variation in treatment effects across subgroups.
• Robustness Testing: Validated robustness through noise injection and the addition of correlated variables.

Scientific Applications:

• Infant Health and Development Program semisynthetic dataset: Validated on a semisynthetic dataset with nonlinear/nonparallel treatment selection bias to assess causal inference accuracy.
• LaLonde employment training program dataset: Applied to a real-world dataset from LaLonde's employment training program for empirical evaluation of treatment effects.
• Comparative evaluation: Used to compare performance against logistic regression, LASSO, and DCN-PD, showing quantitative improvements and estimates close to true values.

Methodology:

DPN-SA employs a deep learning framework that integrates sparse autoencoders to process and analyze high-dimensional covariates, enabling the model to capture complex patterns and relationships in nonlinear or nonparallel treatment assignment.