Keywords: Causal Inference, Double Robust, Machine Learning
There has been a surge of interest in double robust treatment effect estimation in observational studies driven by the realization that the combination of double robustness with modern machine learning methods results in asymptotic efficiency and state-of-the-art performance in practice. These methods use a weighted sum of residuals to debias a regression estimate of the treatment effect. Typically the weights used are inverse propensity weights using an estimate of the propensity score; this choice is justified asymptotically. However, there is little evidence that an optimally tuned propensity model yields optimally tuned inverse propensity weights in finite samples. We propose an alternative approach to weighting the residuals that uses weights that directly minimize finite-sample risk bounds. In contrast to typical double robust methods, which estimate weights to minimize error on the propensity scale and then invert them, our weights are effectively estimated on the inverse propensity scale. In extensive experiments, we find that our method compares favorably to recently advocated methods like targeted maximum likelihood and double machine learning.