39 – Modeling and Modeling Error
Comparison of Imputation Techniques for Item Missing Data in the Survey of Income and Program Participation
Sarah McMillan
U.S. Census Bureau
The Survey of Income and Program Participation (SIPP) collects detailed information about income and program participation. These key questions can suffer from higher rates of nonresponse, so a sequential hot deck procedure is used to impute all items with missing data. Single imputation techniques like these do not incorporate the uncertainty of the imputation and can lead to underestimates of the true variance. However, if rates of item missing data are comparatively low, the underestimates may be slight with little impact on significance. Multiple imputation methods were used to impute missing data in the 2008 Panel of the SIPP to incorporate this uncertainty into the imputations. The imputations are made using several different multiple imputation methods and compared based on estimates of variance, as well as, fractions of missing information for key SIPP statistics. Additionally, we compare to similar estimates based on no imputation and based on various single imputation methods, like the current hot deck method.