Preventing Spurious Interactions: A New Inductive Bias for Accurate Treatment Effect Estimation

In recent years, the use of non-linear machine learning techniques to estimate treatment effects from observational data has gained significant attention. Many of the most effective methods incorporate specific algorithmic inductive biases that encode causal information, thereby enhancing the precision of treatment effect estimates. In this paper, we explore and categorize several of these models based on the biases they address, and we introduce the prevention of spurious variable interactions as a new inductive bias that serves as a central focus of our study. This novel inductive bias, not previously considered in this area, aims to further improve the accuracy of treatment effect estimation. Additionally, we propose a practical method that utilizes knowledge of the data generation processes to restrict variable interactions to those defined by the mechanisms of the underlying causal model. This strategy integrates seamlessly with many of the leading neural network-based models. Our results show that these constraints significantly enhance performance, setting new state-of-the-art benchmarks in treatment effect estimation. Furthermore, we show that our method remains robust even when the causal model is imperfect and that incorporating partial causal information yields better results than disregarding it entirely.