Controlling for mental health in earnings equations: what do we gain and what do we lose?

Abstract

This paper examines the biases in estimating wage equations that may arise from measurement errors in various mental health indicators--two subjective proxies and one clinical assessment. The results suggest that a self-reported measure based on whether the individual reported missing school or work for mental-health-related reasons leads to the smallest measurement error bias in the coefficients of the explanatory variables in an earnings equation. Two specific results lead to this conclusion. For one, the variance of the random component of its measurement error is the smallest among the three indicators leading to the least biased estimates of the impact of mental health on wages. Second, systematic reporting biases which vary with the non-health regressors in a wage equation do not appear to exit in this measure. Consequently, this indicator follows the classical measurement error model. This implies that the coefficient estimates of the impact of non-health variables are always improved, in the sense of having a smaller bias, when this mental health proxy is included in the regression. The measurement error in a self-evaluation according to the scale of excellent, good, fair, or poor has the largest variance thus leading to a substantial understatement of the impact of mental health on earnings. This measure also contains significant reporting biases that vary with gender, race, and education--many of the non-health regressors in a wage equation. The measurement error variance in the simulated diagnostic measure analyzed in this paper is also large. Thus this proxy will yield poor estimates of the impact of mental health on earnings.