Random Effect Approach to Protection Against Model Error in Logistic Models of Census Coverage

摘要

The application of logistic models and variable selection procedures to modeling Census coverage error has enabled the use of an expanded list of independent variables and has provided a way to eliminate higher order interactions in models; interactions in models which were implicitly included in the post-stratum used in the 2000 Census coverage. In the past, predictive categories of coverage such as a person's Race/ethnicity, age, sex and whether they lived in a rented home needed to be included in a model with all of their resulting cross classications. In the logistic model, only the level of interactions that are important need to be included in the model. In addition, continuous variables, such as single year age splines, Census mail return rates and other local rates can be included. Although the logistic modeling approach has the potential to improve dual-system estimates of coverage, there is still the potential for bias due to erroneously leaving out independent variables that are tested as non-signicant (type II error) such as a subset of the aforementioned higher-order interactions. It may be dicult to add specic interactions terms into the 2010 model after viewing the 2010 data due to the need to guard against perceptions of manipulating the model.