Causal Inference in Health Science Research

Abstract

Causal inference methods including propensity score (PS) matching and weighting have been widely used for comparative effectiveness research based on observational clinical databases. There are new challenges of causal inference in medical studies, including how to handle clustered data structure, how to improve efficiency of the traditional methods and etc. To overcome these challenges, we develop novel causal inference method for estimating treatment effects. The proposed methods are motivated by and applied to various studies of cardiovascular diseases and cardiac surgeries.

In Chapter II, we aim to estimate causal treatment effect in clustered observational data and in our application clustered data structure arises from patients being nested within hospitals. We propose a strategy to combine PS matching and outcome regression model for estimating treatment effect while accounting for the hierarchical nature of the data. We show that this method enjoys the double robustness property, i.e. when either the PS or outcome model is correctly specified, the bias is negligible. The proposed method has better performance than the usual PS method and the existing doubly robust PS weighted method, and is more robust than the outcome regression method.

Chapter III is motivated by comparing different types of ventricular assist devices (VAD) for end-stage heart failure patients where patients are likely to receive a heart transplant after receiving a VAD. We propose to treat heart transplants as dependent censoring and propose an augmented inverse probability weighted method to estimate the treatment-specific difference in potential restricted mean lifetimes, had no patients received heart transplant. Specifically, we first derive an estimator that combines inverse probability of treatment weighting and inverse probability of censoring weighting to account for the imbalance in baseline characteristics and censoring that may depend on time-dependent confounders, respectively. Then we propose an augmentation method to improve the efficiency of estimation. Large-sample properties of the proposed methods are studied and simulation studies are conducted to assess the finite-sample performance.

In Chapter IV, we further extend and refine the work in Chapter III and develop methods for estimating more meaningful causal treatment effects as opposed to the average treatment effect. The goal is to overcome two potential problems related to estimating the average treatment effect. Namely, depending on the specific application the average treatment effect may not be the most clinical meaningful and relevant quantity. Also it is known that estimators for average treatment effect often have large variance even with the use of more sophisticated methods for improving efficiency (e.g., augmentation). We propose augmented methods of matching weights to estimate the treatment-specific difference in potential restricted mean lifetimes for the matched population, had no patients received heart transplant. Simulation studies show that the proposed methods considerably improve the efficiency compared to the existing methods.