Reproduce “Towards Representation Learning for Weighting Problems in Design-Based Causal Inference” on OpenResearch

About this paper

Oscar Clivio, Avi Feller, Chris Holmes

Researchers from the University of Oxford and UC Berkeley developed a framework for learning outcome-agnostic covariate representations in design-based causal inference. Their method uses a neural network to minimize an AutoDML loss, deriving a "balancing score error" (BSE) to quantify information loss, and achieved competitive or superior performance on various benchmark datasets compared to existing weighting approaches.

Key insights

• The "confounding bias," quantifying information loss from covariate representations, can be rigorously bounded by the outcome-agnostic "Balancing Score Error" (BSE).
• The AutoDML loss provides a practical, empirically estimable, and outcome-agnostic proxy for minimizing the BSE, enabling data-driven learning of effective representations.
• Learning robust representations without using outcome information allows machine learning to be integrated into design-based causal inference while preserving theoretical guarantees on bias control.

Summary by alphaXiv.