Diagnosis Measurement Error and Corrected Instrumental Variables

Health diagnosis indicators used as explanatory variables in econometric models often suffer from substantial measurement error. This measurement error can lead to seriously biased inferences about the effects of health conditions on the outcome measure of interest, and the bias generally spills over into inferences about the effects of policy/treatment variables. We generalize an existing instrumental variables (IV) method to make it compatible with the types of instruments typically available in large datasets containing health diagnoses. In particular, we relax the classical IV assumption that the instruments must have uncorrelated measurement errors. We identify and estimate the covariance matrix of the measurement errors and then use this information to derive a correction term to mitigate or eliminate the bias associated with classical IV. Our Monte Carlo simulations suggest that this corrected IV method can produce estimates far superior to those produced by OLS or classical IV.