Diagnosis Measurement Error and Instrumental Variables

As is well understood among clinical researchers, medical diagnoses often suffer from substantial measurement error. This is particularly true for mental health diagnoses.1 Reliance on inaccurate diagnosis indicators can substantially compromise researchers’ and clinicians’ ability to reliably evaluate treatment approaches, measure clinical progress, or estimate relationships between patient attributes and various outcome measures of interest (such as expected treatment costs, hospital and community tenure, or ability to participate in the labor force). Assessing the generalizability of treatment interventions across patient types, for example, is problematic given limitations in the accuracy of diagnosis information. Although much e¤ort has been put into developing better diagnosis instruments and into encouraging the use of standardized codes, the nature of the measurement error problem is such that it will always exist as long as researchers must rely on clinical evaluation or research instruments. While “gold standard” references in the literature (e.g., Roy, et. al, 1997) provide useful benchmarks for comparing measures of “reliability” and “validity” across diagnostic instruments, even instruments meeting these gold standards are likely contaminated with substantial errors. Moreover, even if all diagnosis variables were measured without error, any measurement error in auxiliary control variables correlated with the diagnosis indicators (such as comorbidity or severity) would spill over into inferences about the health e¤ects. The standard approach for handling measurement error in statistical models involves instrumental variables (IV) estimation.2 Suppose we wish to estimate