605 – Imputation of Missing Data
Addressing Nonresponse for Categorical Data Items in Complex Surveys Using Full Information Maximum Likelihood
Susan Edwards
RTI International
Marcus E. Berzofsky
RTI International
Paul Biemer
RTI International/UNC Chapel Hill
This paper presents a comparison of two full information maximum likelihood (FIML) approaches to traditional imputation approaches for addressing item nonresponse in a complex survey. Item nonresponse is an issue researchers using public use files for survey data often encounter. There are several techniques for dealing with item nonresponse in categorical data. A relatively new technique for handling missing data is FIML which incorporates all response patterns during the estimation process rather than ignoring cases with missing values. In 1982, Fuchs' proposed a FIML method to handle nonresponse missing at random (MAR); Fay (1986) expanded the FIML method to handle nonresponse not missing at random (NMAR) by incorporating item nonresponse indicators into the model. The National Survey of Drug Use and Health (NSDUH) is an annual national and state level survey that collects information on the use of tobacco, alcohol, illicit drugs and mental health in the U.S. Several variables on the NSDUH undergo a weighted sequential hot deck imputation (WSHD). Using data from the NSDUH, these two FIML approaches are compared with listwise deletion (i.e., MCAR) and WSHD imputation techniques.