The Australian Census Longitudinal Dataset (ACLD) has so far linked 5% of 2006 Census records to corresponding records from the 2011 Census. This valuable dataset can help researchers and policy makers understand patterns of change in social and economic conditions over time at the individual level. It can also provide insight into the pathways that lead to particular outcomes and how these pathways vary for different population groups.

The ACLD was created using Felligi-Sunter probabilistic linking methodology, which tolerates missing or disagreeing values on linking fields providing there is agreement on other fields sufficient to offset this. Further, the linking process did not use name and address as these fields are unavailable on the Census. This method can nonetheless correctly identify the majority of matches using fields common to the two Census datasets, however missing and inconsistent field values introduce error into the ACLD - namely, false links and missed links. False links are linked records on the dataset that do not belong to the same individual. They may impact analysis of small sub-populations and individual transitions. Missed links are matches which were not linked and therefore do not appear on the linked dataset. They can be a major source of bias as certain sub-populations can have a higher chance of being missed. The ABS is currently assessing linking errors in the ACLD, and has estimated that 5-10% of ACLD links are false, and approximately 9% of potential links were missed.

This presentation will outline the impact of false and missed links on analysis using the ACLD. It will discuss methods and tools designed to mitigate the impact of linking errors, including weighting to correct for missed links, and consistency flags to help identify potential false links.