Imputation Using the Other Pair Member
Frechtel, P., Scott, V., Couzens, A. E., & Moore, A. (2012, July). Imputation Using the Other Pair Member. Presented at JSM 2012, San Diego, CA.
The National Survey on Drug Use and Health (NSDUH) provides national, state, and substate data on substance use and mental health in the civilian, non-institutionalized population age 12 and older. In the NSDUH, zero, one, or two people are selected from each selected household. "Pairs" account for about 60% of the annual sample, over 90% of which are members of the same family. The responses to some NSDUH questions have high positive correlations between pair members, especially the questions about family-level characteristics. The current imputation method, predictive mean neighborhood (PMN), does not exploit this correlation, even when the value for one pair member is missing and the value for the other pair member is not missing. This presentation discusses (1) a method for identifying variables for which the other pair member's response may be a better choice for imputation than the PMN-assigned value; and (2) for these variables, a method for identifying exact conditions under which the other pair member's response should be used.