Zhou XH, Seattle VA Medical Center and University of Washington; Wang LB, University of Washington;
Objectives:
Inferring the causal effect from observational studies is challenging due to difference in pre- treatment variables. Subclassification on quintiles of estimated propensity scores is often used in practice to balance observed covariates. However, residual confounding of the subclassification estimator with a fixed number of subclasses remains even with an infinite sample size. In this talk, we propose a new subclassification estimator, which overcomes the limitation of the existing subclassification methods
Methods:
Unlike the existing subclassification-based propensity score method, our new method allows the number of subclasses to increase with the sample size. The key observation in our paper is that the subclassification estimator can be seen as a coarsened version of the Horvitz-Thompson type inverse probability weighting estimator. We applied this new method to a real-world VA funded project on the effect of hepatitis C antiviral treatment on subsequent healthcare costs.
Results:
We mathematically show that the new estimator is a squared root consistent estimator for the average causal effect when the number of subclasses grows at a certain rate with the sample size. We have also provided a valid bootstrap method for constructing confidence intervals for the average causal effect. We provide an estimate and confidence interval for causal effect hepatitis C antiviral treatment on subsequent healthcare costs.
Implications:
We show that the bias of the existing propensity score subclassification estimator for the average causal treatment effect remains even when the sample size is very large. We then propose a new estimator for the average causal effect and show this new estimator can eliminate the bias asymptotically and can provide more reliable estimators for the average causal effect.
Impacts:
VA HSRandD has funded many projects, which involve the design of observational studies. One of key challenges in the analysis of observational data is how to adjust for confounders. Use of inappropriate statistical methods may lead to wrong conclusions on the effectiveness of a new treatment or VA program. The new method we have proposed will allow VA researchers to use a right method for the analysis of observational data.