The effects of self-assessed health: Dealing with and understanding misclassification bias

Categories of self-assessed health (SAH) are often used as a measure of health status. However,the difficulties with measuring overall health mean that the same individual may select into different SAH categories even though their underlying health has not changed. Thus,their observed SAH may involve misclassification, and the chance of misclassification may differ across individuals. As shown in this paper,if neglected, misclassification can lead to substantial biases in not only the estimation of the effects of SAH on outcomes, but also on the effects of other variables of interest,such as education and income. This paper studies nonlinear regression models where SAH is a key explanatory variable, but where two potentially misclassified measures of SAH are available.In contrast to linear regression models, the standard approach of using one SAH measure as an instrumental variable for the other cannot produce consistent estimates. However, we show that the coefficients can be identified from the joint distribution of the outcome and the two misclassified measures without imposing additional structure on the misclassification, and we propose simple likelihood-based approaches to estimate all parameters consistently via a convenient EM algorithm. The estimator is applied to data from the Household, Income and Labour Dynamics in Australia (HILDA) Survey, where we exploit the natural experiment that in some waves individuals were asked the same question about their health status twice, and almost 30% of respondents change their SAH response. We use the estimator to (i) obtain the first reliable estimates of the relationship between SAH and long-term mortality and morbidity, and to (ii) document how demographic and socio-economic determinants shape patterns of misclassification of SAH.