92 – Complex Data Analysis and High-Dimensional Computing: Methods and Applications
Evaluation of Model-based Methods in Analyzing Complex Survey Data: A Simulation Study using Multistage Complex Sampling on a Finite Population
Van Parsons
National Center for Health Statistics
Rong Wei
National Center for Health Statistics
Jennifer D. Parker
National Center for Health Statistics
The usage of traditional design-based methods for complex-survey data often leads to estimation difficulties or unreliable analyses when sample sizes are not sufficiently "large" at some level of multi-stage sampling. In these situations model-based estimation methods are often suggested as alternatives to compensate for data deficiencies. For this study, we focus on both design- and model-based statistical inference based on a sample of ~1000 households taken from a reduced-scale pseudo U.S. population that captures many features of geographical and household clustering within the true U.S. population. This pseudo population was developed from nine years of the National Health Interview Survey (NHIS) data. A simulation study is performed on this pseudo-universe where we imposed a complex design including multilevel sampling from strata, primary clusters, secondary clusters, households, and persons along with post-stratification weighting adjustments. Sampling properties of design-based regression estimators (Binder 1983) and multi-level model-based regression estimators are compared.