Abstract:
|
Survey weighting adjusts for differences between the collected samples and the target population. However, classical weights have lots of problems. Extreme values of weights will cause high variability and blow up the estimates. In practice, weighting construction requires arbitrary choices about selection of weighting factors and interactions, pooling of weighting cells and weight trimming. The general principles of Bayesian analysis imply that models for survey outcomes should be conditional on all variables that affect the probability of inclusion, which are the variables used in survey weighting. We propose to include weighting variables in the model for survey outcomes under the framework of multilevel regression and poststratification at much finer levels with structural prior specification. The procedure will yield the model-based weights after smoothing, which are evaluated via simulations comparing with classical weights. We use Stan for computation and illustrate the performances via the application of the New York Longitudinal Survey of Poverty study.
|