Publication Type: Paper/Poster Proposal
Abstract: Missing data in the form of item or entire survey module non-response is a common problem in survey research (Groves, Dillman, Eltinge, and Little 2001). Approximately 10 percent of the Current Population Survey (CPS) sample refuses to take the demographic supplement. These “full supplement” refusals have the variable values for the entire demographic supplement imputed. Although properly specified imputation can alter basic distributional summary statistics (e.g., means, rates and variances) from the statistics calculated using complete cases only, it should not transform the relationships among variables. In other words, imputation should not create significant correlations between variables that were not there before the imputation, nor should it reduce the magnitude of significant correlations between variables that were there prior to imputation. With this in mind, there are two questions we attempt to answer in this analysis: 1) Is there a difference between the imputed cases and the non-imputed cases with respect to health insurance coverage? 2) Does hot deck imputation create a significant bias in health insurance coverage estimates? In the 2003 Current Population Survey’s Demographic Supplement, 59.7 percent of 18-64 year old adults have commercial health insurance coverage if they have the full supplement imputed. However, 72.4 percent of the non-full supplement imputations have commercial health insurance coverage. Furthermore, full supplement imputations have a 26.7 percent uninsurance rate while all other 18-64 year old adults in the CPS have an uninsurance rate of 14.6 percent. We examine the relationships among key correlates to see whether this difference is due to the characteristics of the full supplement imputations or is due to the Census Bureau’s missing data imputation routines introducing/reducing relationships among variables.