Using the Bootstrap Method to Estimate Variances of Multi-State Life Table Functions

Liming Cai, National Center for Health Statistics (NCHS), CDC
James Lubitz, National Center for Health Statistics (NCHS), CDC
Mark D. Hayward, University of Texas at Austin
Yasuhiko Saito, Nihon University
Aaron Hagedorn, University of Southern California
Eileen Crimmins, University of Southern California

Demographers frequently use multi-state life table (MSLT) method to analyze expected duration in various states. Frequently, life table estimates are derived from sample surveys. Thus, measures of sampling variability are needed for valid statistical inferences. This study will compare variance estimates obtained using the bootstrap method, which takes sampling design into account, with other approaches in common use that don’t. Our data source is the Medicare Current Beneficiary Survey, a longitudinal survey of Medicare beneficiaries with stratification and multi-stage clustering. We will first estimate the variance of total and active (without limitations in activities of daily living) life expectancies at age 65, treating the MCBS as a simple random sample. Then we will estimate variances using the bootstrap approach. The bootstrap variance estimates are larger than the variance estimates ignoring the sampling design because of the clustering effect.