Exploring Methodologies to Extend Census Content
The population Census provides a unique opportunity to obtain detailed information from the whole population of Australia in a way that supports tabulation for small geographic areas and fine classificatory items. Unfortunately, space on the Census form is at a premium, and whenever new questions are added, a similar number need to be dropped to limit costs and avoid undue respondent burden.
An intriguing idea to extend the content of the Census is to use multiple "thematic" forms. These would each contain the same "core" questions making up the bulk of the Census, but would include a different set of "theme" questions. This would provide a complete Census of core items, and a very large survey of each theme item.
The South Australian Methodology Unit has conducted an exploration of methods for producing good estimates from the survey component of a thematic forms Census. It examined a situation with three different thematic forms, with each type delivered to approximately every third dwelling. This provides a one-in-three sample for each theme question, which should be enough to provide good estimates for quite small geographic areas and sub-populations.
The study looked at properties of estimates under various approaches. A weighting approach has the disadvantage of providing different survey weights for each theme, and would not reproduce Census totals by the core items. In contrast, an imputation approach would fill in theme values for persons not reporting for that theme, giving consistency of all counts. Unfortunately the standard "hot-deck" approach to imputation would bias the numbers reported for a category, because two-thirds of the theme values were imputed based on assumptions that did not take account of peculiarities of the category concerned. Thus, for example, Greek-born people may have theme values imputed from non-Greek-born people, which will mask any effects peculiar to Greek-born people.
To overcome these limitations, the study developed a "balanced imputation" approach, in which the imputes are chosen to give good agreement with the best available estimates, across a very large set of Census tables at a range of geographic levels right down to CD level. This method requires huge computation, but provides an imputed Census file with very satisfactory properties.
While thematic forms are not under consideration for the 2011 Census, this study opens up the possibility of moving in this direction in the future. The approach promises good small domain estimates for an increased set of questions, while retaining a complete Census of the majority of the current questions. A research paper will be published shortly on the ABS Website. For further information, please contact Philip Bell on (08) 8237 7304.