This project will develop econometric methods which yield consistent estimates of binary models of psychiatric illness when the health indicator is subject to error. This topic is highly relevant to health applications since illness is often measured by a binary variable indicating the presence or absence of a diagnosis. Much is known about the econometric consequences of measurement error in control and outcome variables that are continuous. Only recently have researchers recognized that the theoretical implications of measuring a continuous variable with error do not extend to the situation where the mismeasured variable is dichotomous. In particular, contrary to the classical result, even purely random measurement error in a binary outcome variable will bias the coefficients on control variables in both linear and nonlinear regression models. Moreover, the classic textbook econometric solution for eliminating measurement error bias arising from a mismeasured control variable, the instrumental variables approach, is not a valid solution when the variable subject to error is binary. The methods will be developed in the context of two general applications: one where psychiatric diseases serve as control variables--the labor market consequences of mental illness; and one where mental health is the outcome variable--the risk factors for various psychiatric disorders. The methods will rely on both linear and nonlinear regression techniques as well as the maximum likelihood approach. All approaches will be tractable and have broad applicability to any empirical analysis of health-related behavior. The primary data source will be the National Comorbidity Survey.
