A split questionnaire design applied to the census of Brazil
Nadia M.C. Rodrigues, Pontific Catholic University of Rio de Janeiro - PUC-Rio
Brazilian Census is conducted with two questionnaires: a short and a long one. The long questionnaire contains 93 items per person and the short one 23. Respondent burden is one of the main causes for missing information. Using real data we simulate a matrix sampling approach: respondents are allocated only items part of one of three components, thus yielding three data sets and lessening respondent burden. Individual data from these three files are linked to create a complete new data set, called synthetic data file, using statistical matching. Every single record is concatenated to similar records from the other two files, using an imputation procedure based on hot deck. We find that all the selected empirical distributions of the complete data are well reproduced in the synthetic data sets, as well as bivariate and conditional distributions. Nearly the same inferences can be achieved using matrix sampling design with a reduced cost and less respondent burden.
Presented in Poster Session 5