Navigated Weighting to Improve Inverse Probability Weighting

• In Event: Poster Session: Political Methodology

Abstract

The inverse probability weighting (IPW) is broadly utilized in dealing with missing data problems including causal inference but may suffer from large variances and biases due to propensity score model misspecification. To solve these problems, I propose an estimation method called the NAvigated WeighTing (NAWT), which conveys efficiency and reduces biases due to propensity score model misspecification by tweaking the propensity score estimation depending on a pre-specified quantity of interest. I investigate its large-sample theoretical properties and demonstrate its impressive improvements through simulation studies. Its idea may be utilized to improve other propensity score estimation including the covariate balancing propensity score and cutting-edge machine learning techniques. An easy-to-use R package which implements the NAWT is being developed.

Author

• Hiroto Katsumata, The University of Tokyo