Strategies for Data Analysis with Two Types of Missing Values

Analysis of incomplete data requires assumptions about the mechanism that divides the complete sample into its observed and missing parts. When two different types of missing values occur in the same data set, one should also consider the process that partitions the missing data into the two groups. In this presentation, I extend Rubin's (1976) concept of missing at random to two sets of missing values, describing conditions under which the missing-data processes may be completely or partially ignored. Multiple imputation (Rubin, 1987) may be carried out in two stages, allowing us to measure rates of missing information due to each type of missing value. For illustration, I apply these concepts and techniques to several data examples.