Potential Methodologies for Count Imputation for the Decennial Census

• Activity Number: 28
• Type: Contributed
• Date/Time: Sunday, August 8, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Richard A. Griffin*+
• Companies: U.S. Census Bureau
• Keywords: EM algorithm ; multinomial distribution ; log linear model ; spatial models

Abstract

Count imputation was used for Census 2000 for housing unit records lacking a status designation of occupied, vacant, or nonexistent, as well as for known occupied units with unknown population count. A "hot-deck" imputation methodology was used to determine donors to be used for donees requiring imputation. Count imputation as implemented for Census 2000 was a deterministic method in that given the census data the imputed values are fixed. Alternative stochastic imputation methods which randomly select imputed values from a distribution could also be used. Since the results of count imputation effect many important uses of the Census such as allocation of congressional seats and revenue distribution, the Census Bureau is conducting research on imputation alternatives to the Census 2000 methodology. This paper presents results from simulations of alternative stochastic imputation methodologies using log linear models on Census 2000 data. These methodologies assume a multinomial distribution and take advantage of the monotone missing data pattern to produce explicit maximum likelihood estimates by the factored likelihood method for some log linear models.