A review of instrumental variables estimation of treatment effects in the applied health sciences

Health scientists often use observational data to estimate treatment effects when controlled experiments are not feasible. A limitation of observational research is non-random selection of subjects into different treatments, potentially leading to selection bias. The two commonly used solutions to this problem—covariate adjustment and fully parametric models—are limited by strong and untestable assumptions. Instrumental variables (IV) estimation can be a viable alternative. In this paper, I review examples of the application of IV in the health sciences, I show how the IV estimator works, I discuss the factors that affect its performance, I review how the interpretation of the IV estimator changes when treatment effects vary by individual, and consider the application of IV to nonlinear models.