Does Sequence of Imputed Variables Matter in Hot Deck Imputation for Large-Scale Complex Survey Data?

• Activity Number: 314 - SPEED: Missing Survey Data: Analysis, Imputation, Design, and Prevention
• Type: Contributed
• Date/Time: Tuesday, July 31, 2018 : 9:25 AM to 10:10 AM
• Sponsor: Survey Research Methods Section
• Author(s): Amang Sukasih* and Peter Frechtel and Karol Krotki
• Companies: RTI International and RTI International and RTI International
• Keywords: cyclical tree based hot deck; complex survey data; nonresponse bias

Abstract

Hot deck is an imputation method for complex survey data, especially popular when many survey items are to be imputed. Items are frequently correlated in surveys; one goal of imputation is to preserve these relationships. When imputing many variables and deciding which should be imputed first, one can decide on the sequence in which the variables are imputed---based on order of appearance in the questionnaire (a screener question is imputed first before its follow-up questions) or based on rate of missing data (items with lowest rate would be imputed first, followed by items with higher rates). Iteratively cycling the imputation may address association among variables (once all variables with missing values are imputed, imputation is rerun with previously imputed values in the covariates being treated as reported values). This presentation discusses results from investigating the sensitivity of final estimates to the sequence of imputed variables. We also measure the impact of factors such as missing data rates and number of levels in categorical variables and imputation cycles. We use empirical simulation and focus on bias reduction and preservation of variable relationships.