Instrumental Variable Methods for Effectiveness Research

Many research questions, such as quality of specific providers or guideline-concordant care in typical practices, are commonly studied in an observational setting. These analyses face the risk that covariates related to both an outcome of interest and the probability of treatment are unobserved or uncontrolled. The resulting biases can easily overwhelm true effects or create apparent effects, and small changes in the analytic approach can yield contradictory results, which is demonstrated for antidepressant medication and counseling. An econometric method, instrumental variable estimation (IV), provides a possible solution and permits causal inferences under certain conditions. The central element of IV is the observation that some variables are related to outcomes only through their effect on treatment and have no independent direct effect. The main difficulty of using IV is to identify appropriate instrumental variables and to assure that the sample size is sufficiently large to provide acceptable statistical power, which is substantially lower in IV than in standard regression models. These issues are discussed in the context of determining the effectiveness of depression treatment and illustrated using data from the depression panel of the medical outcomes study.