ABSTRACT
It might be difficult to use the joint modeling strategy when data have complex features (e.g., ranges and bounds) and structures (e.g. skip pattern variables). A popular and effective imputation strategy is the fully conditional specification (FCS) approach. FCS is also termed as the multiple imputation by chained equations (MICE) or sequential regression multiple imputation (SRMI). The FCS strategy defines the joint model by specifying univariate models for each variable conditional on others in the dataset. The imputation algorithm goes through these variables sequentially in a cyclic manner. The FCS imputation is implemented by several statistical packages such as R MICE, SAS PROC MI FCS, and IVEware. We present some common options for specifying conditional models used in FCS. We also demonstrate how FCS can effectively handle aforementioned complex data features. We demonstrate that WinBUGS can also be used to carry out FCS imputation. This idea might further improve the utility of FCS due to the wide range of models available in WinBUGS. Since FCS defines the joint model through specification of conditional models, we discuss about its theoretical limitation, that is, the incompatibility issue. However, we show that FCS can work well even if some conditional models are mis-specified. We present a practical example on imputing missing data from U.S. birth data using SAS PROC MI FCS.
