Measuring and Reducing Inconsistency among Questionnaire Items through Imputation: An Application to the NSOPF

Most of the research work on imputation has concentrated on improving methods of imputing missing values a single variable at a time. For large complex surveys a more elaborate imputation process is needed, because the resulting data must satisfy various constraints and dependencies that are often intertwined. These constraints can take the form of one variable being the sum or ratio of others, one variable dictating the possible values of other variables, or some other functional relationship. Missing data are among the inevitable facts of survey research. Oftentimes, sampled units do not respond to certain survey items, or fail to complete different sections of a questionnaire. Also, there are instances where respondents provide responses that are inconsistent with others. Such responses are typically set to missing and then imputed along with the originally missing values.