Can supervised deep learning architecture outperform autoencoders in building propensity score models for matching?

Area of Science:

• Epidemiological research
• Statistical modeling
• Machine learning applications

Background:

• Propensity score matching is crucial for observational studies but sensitive to model specification.
• Accurate propensity score estimation impacts reliable treatment effect inference.

Purpose of the Study:

• To evaluate supervised deep learning and unsupervised autoencoders for propensity score estimation.
• To compare their performance against traditional methods (logistic regression, splines) regarding bias and variance.
• To assess accuracy in treatment effect estimation using real-world and simulated data.

Main Methods:

• Plasmode simulation using the Right Heart Catheterization dataset.
• Evaluation of supervised deep learning and autoencoders against logistic regression and spline-based methods.
• Comparison of bias, standard errors, and coverage probability.
• Validation on real-world data using a double robust approach.

Main Results:

• Supervised deep learning models demonstrated superior variance estimation compared to unsupervised autoencoders, with similar bias.
• Real-world data analysis showed supervised deep learning estimates aligned well with conventional methods.
• Deep learning models performed effectively, even with rare exposure, outperforming traditional methods.

Conclusions:

• Supervised deep learning models offer a promising approach for enhancing propensity score estimation in epidemiology.
• These models provide nuanced confounder adjustment, particularly beneficial for complex datasets.
• Integration of supervised deep learning is recommended for epidemiological research, with reproducible code provided.