Challenges in Estimating Multi-State Expectancies from Large-Scale Cross-Longitudinal Surveys

Nicolas Brouard, Institut National d'Études Démographiques (INED)
Feinuo Sun, university of texas at arlington

This proposition of communication introduces innovative ways to address the challenges in estimating multi-state expectancies using large-scale longitudinal surveys. We first discuss the challenges in designing longitudinal surveys as well as in examining prevalences of health outcomes considering the rapidly changing health states (i.e., the dynamics of health changing that are captured in the longitudinal surveys). We introduce the method of interpolated Markov chains in order to estimate the incidences of change between states based on the probability of change over a small time interval that is estimated by multinomial regression as a function of age and covariates. More importantly, we discuss how Powell’s algorithm which was useful in optimizing the multinomial regression function to estimate health expectancies for different states based on up to 30 covariates, did not converge when adding more states, covariates or interactions because the likelihood function to be maximized depends on more than 200 variables. We then test the Brent/Praxis algorithm which uses the principal component analysis method to estimate information matrix about probabilities of change, and found that the algorithm is also useful and time-efficient.

Keywords: Longitudinal studies , Mathematical demography , Health and Morbidity, Population Ageing

See extended abstract.