Efficiency of Doubly Robust Estimator of Causal Effects with Trimmed and Winsorized Propensity Score Weights

Angela Fel A. Padecio, Carolina B. Baguio


Estimation of treatment effect with causal interpretation where treatment is not randomized may be biased if confounding is not taken into appropriate account. Adjustment for confounding is often carried out through regression adjustment (ANCOVA) or propensity score (PS) method specifically inverse weighting. When used individually to estimate a causal effect, both \mbox{ANCOVA} and PS method are unbiased only if the model is correctly specified. The doubly protected or doubly robust (DR) estimator combines these two approaches such that only one of the models need be correctly specified to obtain an unbiased effect estimator. This paper focuses on the scenario with correct ANCOVA model but misspecified PS model. PS weighting is sensitive to model misspecification and outlying weights that can excessively influence results. To deal with this problem, robust methods are used to lessen the influence of extreme weights through trimming and winsorization. A Monte Carlo simulation framework is used based on real-world data modeling the use of statin to lower the cholesterol. Empirical results show that trimming the PS weights gives an improvement on the DR estimator efficiency in terms of its root mean square error (RMSE) only for small sample sizes. On the other hand, winsorizing the PS weights considerably improves the efficiency of DR estimator both for small and large samples.


Doubly Robust Estimation; Regression Adjustment (ANCOVA), Propensity Score, Trimming, Winsorization

Full Text:



  • There are currently no refbacks.

Copyright (c) 2015 Angela Fel A. Padecio, Carolina B. Baguio

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.