|
Abstract:
|
Failure to treat missing data in survey or census data may introduce bias in the tabulated totals. The Census Bureau is seeking a common adjustment method for more than 6,000 products collected in the Economic Census. Unlike the general statistics items collected from every eligible establishment such as annual payroll and total receipts, product information varies by establishment and administrative data are not available for substitution or validation. Moreover, product data are characterized by poor item response rates, few available predictors, additivity constraints within establishment, and definitional rules such as mutually exclusive products in some cases and required products in others. Little historic data are available for modeling. Hot deck imputation provides a flexible approach to dealing with missing data that retains multivariate relationships without making explicit parametric model assumptions. Instead, hot deck methods impute missing values using reported values from a similar unit. In this paper, we examine the merits of random and nearest-neighbor hot deck imputation using empirical and simulated data for selected Economic Census products.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.