Longitudinal studies are indispensable for social and behavioral research on aging. By following people over time, panel studies of aging help researchers distill changes due to aging or period effects from stable cohort differences. Furthermore, panel studies can bolster causal claims about aging processes, by measuring the association between baseline characteristics and actual changes in people over time. However, sample attrition can undermine the potential benefits of longitudinal studies of aging. This project examines the impact of sample attrition on estimated age-specific changes in physical health in the general U.S. population and in late adulthood. The analyses estimate the change in number of impairments across two waves of the National Survey of Families and Households (1987-88 and 1993-94) and the Asset and Health Dynamics Among the Oldest Old study (1993 and 1995). The extent of potential bias from attrition is first gauged using a comparison of baseline health levels of those who drop out versus those who remain in the sample. Probit models then identify the baseline determinants of overall attrition, supplemented by multinomial logit models that compare the effects of baseline characteristics on different types of attrition (i.e., refusals, can't locate, too ill, deceased). A """"""""tolerance model"""""""" is developed which follows the logic of a reservation wage model and has the unique benefit of directly estimating how baseline characteristics affect survey participants' tolerance to changes in health. The analyses conclude by comparing the tolerance model versus the more commonly used Heckman sample selection model as means of correcting attrition bias in the analysis of age-related changes in impairments. The tolerance model will be made available to other researchers by publishing it as an executable macro for Stata.
