Abstract:
|
One of the design decisions for complex multi-stage sampling is how to allocate samples to each stage (e.g., number of sample PSU vs. number of sample households within PSU). In order to maximize efficiency of estimates given fixed costs, the decision can be informed by variance component analysis that decomposes the overall sampling variance into individual components associated with each selection stage. When increasing sample sizes for stages with larger variance than their counterparts, the overall variance can be reduced. However, design-based estimators for this analysis in the literature are limited in that they may result in negative variance components particularly when the between-PSU variance is small compared to within-PSU or between-SSU variance. This paper uses an anticipated variance based on a model and a three-stage sample to reduce or eliminate the problem of negative component estimates. We illustrate its application to three-stage sample allocation for the Health and Retirement Study.
|