Essays on Unobserved Heterogeneity and Endogeneity in Health Econometrics

Abstract

Microdata in Health Economics During the past two decades, applied econometric analysis has been widely adopted among health economists. Its adoption is accelerating, producing ever-richer research as electronic recording and collection make available more data about individual patients. In addition, computational power for analyzing large, complex datasets is increasing, facilitating econometric analysis involving latent variables, unobserved heterogeneity, and nonlinear models in the field now established as “health econometrics.” Extensive individual-, household-, and establishment-level microdata are available from cross-sectional and longitudinal sample surveys and the census. Health economics primarily employs cross-sectional data. That is, observations are independent of each other, and pure time series applications are excluded. Microdata used in health econometrics have two notable features. First, they are often measured on a non-continuous scale: data are not only continuous and discrete variables but also on a non-continuous scale, such as quantitative and qualitative (or categorical) variables. This leads inconsistency of linear regression models. For example, analyzing expenditure data is complicated when samples feature a preponderance of observations with zero expenditures. The consistency of standard approaches to the problem relies on the validity of distributional assumptions. To analyze these data, health econometrics requires disparate nonlinear models, including binary responses, multinomial responses, limited dependent variables, integer counts, and measures of duration. Moreover, variables denoting health or quality of life are often unobservable and perhaps measurable only with error (through subjective reports, for example). This situation induces latent variables and selection problems. Second, health data are observational, i.e., they are neither experimental nor collected from surveys and administrative records through randomized experiment. Although availability of