Combining Model-Based and Hot-Deck Imputation to Fill Gaps in Longitudinal Surveys

• Activity Number: 116
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Government Statistics
• Author(s): Sharon I O'Donnell*+ and Yves Thibaudeau
• Companies: U.S. Census Bureau and U.S. Census Bureau
• Keywords: Log-linear models ; conditional probabilities ; imputation ; nonresponse ; Survey of Income and Program Participation

Abstract

Early methods to impute missing household panel survey data were designed to produce longitudinally consistent population estimates. Given the growing demand for panel data analysis, new imputation methods are needed to produce longitudinal consistency at the person level. This paper introduces an imputation technique that applies the estimates from models that generate conditional probabilities subject to log-linear constraints. Our approach rests on expanding a specific parameterization for conditional probabilities as proposed by Thibaudeau, Slud and Gottschalck (2011). The parameterization is manipulated so that it applies to the situation where some survey respondents skip waves in the context of a longitudinal survey. The general model we use to represent this situation is a first order Markov transition process between wave, with double transitions to represent the gaps. Our modeling approach is combined with hot-deck imputation to provide imputations for all the requested items. The technique is tested using the Survey of Income and Program Participation.