A0204
Title: Regularizing extrapolation in causal inference
Authors: Harsh Parikh - Yale University (United States) [presenting]
Avi Feller - University of California at Berkeley (United States)
Elizabeth Stuart - Johns Hopkins Bloomberg School of Public Health (United States)
Kara Rudolph - Columbia University (United States)
David Arbour - Adobe Research (United States)
Abstract: Linear smoothers in machine learning and causal inference predict using weighted averages of training outcomes. Traditional approaches either allow negative weights (improving feature balance but increasing variance and model dependence) or restrict weights to be non-negative (reducing variance but worsening imbalance). Replacing hard non-negativity constraints are proposed with soft penalties on extrapolation, introducing a "bias-bias-variance" tradeoff balancing feature imbalance, model misspecification, and estimator variance. A worst-case extrapolation error bound is derived, and an optimization procedure that regularizes extrapolation while minimizing imbalance is developed. The framework unifies existing methods, allowing practitioners to navigate the tradeoff with a single hyperparameter controlling the extrapolation penalty. Synthetic experiments demonstrate that the approach achieves better bias-variance tradeoffs than existing methods across various degrees of positivity violation and model misspecification. In a real-world application, generalizing randomized trial results to target populations, the method produces more reliable estimates while quantifying sensitivity to modeling assumptions. This enables researchers to make informed decisions about appropriate extrapolation levels rather than being restricted to either unconstrained or strictly non-negative weighting.
