Abstract:
|
In survey practice, the collected sample survey data often cannot represent the underlying population data due to unequal probability of selection, nonresponse and sample size constraints. It is crucial to balance the distribution between the sample and population data and generate valid survey inference. Survey weighting borrows the population information to adjust for the discrepancy, and facilitates the inference. When we need to adjust for multiple factors but only the marginal population distributions are available, classical weighting methods rely on raking, as an interactive proportional fitting approach, to construct weights that adjust for the marginal distributions. However, the increasing number of weighting variables and the limited sample size cause computation problems for the raking algorithm that cannot be directly applied. We propose a model-based alternative for raking and induce shrinkage structure via Bayesian hierarchical framework to stabilize the marginal adjustment. We will use simulation studies to compare the performances between model-based and classical raking methods and evaluate the partial pooling properties. The proposal will be applied in a public h
|