New Instrumental Variable Methods for Causal Inference.

Abstract

In observational studies, unmeasured differences between treatment groups often confound the relationship of interest. Instrumental variable (IV) methods can give consistent effect estimates in the presence of this unmeasured confounding, and are becoming increasingly popular in health and medical research. In this dissertation, we develop new IV methods and apply them in studies comparing mortality among patients receiving dialysis as treatment for end stage renal disease.

In the first project, we develop a weighted IV estimator that adjusts for instrument-outcome confounders through the IV propensity score. The weights are designed to approximate the probability of being selected into a one-to-one match, though the extension to many-to-one designs is also presented. Advantages of weighting over matching include increased efficiency, straightforward variance estimation, and ease of computation. The estimator is shown to be more efficient than alternatives. Its use is illustrated in a study comparing the relationship between mortality and dialysis session length among hemodialysis patients. While developed for use with binary outcomes, future work on applying the method to survival data is presented as well.

In the second project, we develop a weighting procedure for increasing the strength of the instrument when matching. Compared with existing methods, the proposed weighting procedure strengthens the instrument without compromising match quality. This is a major advantage of the proposed method, as poor match quality can bias estimation. Methods are illustrated with a study comparing early mortality in hemodialysis and peritoneal dialysis patients.

In the third project, we compare estimation with strengthened instruments to estimation with instruments that are naturally stronger. Methods for strengthening the instrument are motivated by the benefits of using stronger instruments, including decreased finite-sample bias, increased efficiency, and results that are more robust to unmeasured instrument-outcome confounders. It has not been shown, however, that strengthened instruments provide these same benefits. Results indicate that while they provide for more efficient estimation, they do not decrease finite-sample bias or improve the robustness to unmeasured instrument-outcome confounders. We highlight an important issue that has thus far been overlooked in the literature, and give guidance for future research related to strengthening the instrument.