Enhancement of reporting of deaths among Aboriginal and Torres Strait Islander peoples using record linkage with a range of population datasets resulted in a substantial increase in the number of reported deaths. Compared to the baseline reporting of 580 deaths in 2007, an algorithm based on assessing the weight of evidence of a person being Aboriginal or Torres Strait Islander increased reporting by an additional 200 (34.5%) deaths. Enhancement using ‘at least one report’ of a person being Aboriginal or Torres Strait Islander increased reporting by an additional 328 (56.6%) deaths. The level of reporting of deaths among Aboriginal and Torres Strait Islander peoples in NSW in the ABS death data is therefore estimated at 74.4% based on enhancement with the algorithm or 63.9% based on ‘at least 1 report’.
In relation to age, the greatest enhancement in reporting of deaths was found in older people. As hospital records comprised 78.9% of the linked records and hospitalisation is more common among older people, there was a greater opportunity to enhance reporting of deaths among older Aboriginal or Torres Strait Islander people compared to younger people. There was also greater enhancement of reported deaths for those with chronic conditions, which are likely to generate many hospital records, compared to acute conditions. Enhanced reporting of deaths resulted in increases in SMRs, with a greater proportional increase in SMRs for cancer and cardiovascular diseases compared to external causes of death.
It is not known whether the observed differential enhancement of number of deaths by age resulted in a biased age distribution in the enhanced dataset, or served to correct a reporting bias in the original dataset. While not examined as part of this project, the observed differential enhancement of death data by age would be expected to change estimates of life expectancy for Aboriginal and Torres Strait Islander peoples. Further research is needed to ascertain whether the age distribution in the enhanced dataset is a true reflection of the age distribution of Aboriginal and Torres Strait Islander peoples who died. For example, linkage of the enhanced dataset with a sample of records from a dataset that is known to have reliable reporting of Aboriginal and Torres Strait Islander peoples, such as records from Aboriginal community controlled health services, could be used to explore this.
For geographic remoteness, enhanced reporting of deaths was associated with decreasing geographic remoteness of residence from remote areas to major cities, resulting in a reduction in the urban-remote differential in median age at death. The percentage increase in number of deaths resulting from enhancement was similar for males and females.
The level of reporting of deaths among Aboriginal and Torres Strait Islander peoples in NSW on ABS death data based on enhancement with the algorithm (74.4%) is similar to that found by ABS for NSW deaths in a eleven-month period in 2006 and 2007 using linked death and Census records (76.3%) [3
]. The pattern of increased enhancement for older persons and non-remote regions observed in this study was also observed by Briffa et al [6
] in Western Australia using the ‘at least one report’ approach.
There is some advantage in using administrative health datasets for linkage as these are available on a continuing basis, whereas Census data are available every 5
years. In Australia, Census data are available for linkage only for a short time after the Census as personal identifiers are removed once the dataset is finalised.
In considering whether to use an approach based on an algorithm that uses the weight of evidence for whether a person is indigenous or an approach based on ‘at least one report’, the likelihood of misclassification of an indigenous person as non-indigenous or vice-versa should be taken into account. A national survey estimated the level of correct reporting of Aboriginal and Torres Strait Islander peoples on NSW public hospital admitted patient data in 2007 to be 88% [16
], while an analysis of linked records estimated the level of correct reporting of Aboriginal and Torres Strait Islander peoples on the PDC to be 68.0% [17
]. There is no information on the quality of reporting of Aboriginal and Torres Strait Islander peoples on the EDDC or RBDM birth registration data. In terms of misclassification of a non-indigenous person as indigenous, incorrect links or incorrect reporting on the source record should be considered. In order to create the observed difference of 128 deaths between the two enhancement methods in this study, a misclassification rate of about 1.9 per 10,000 linked records would be required. Thus, an extremely low misclassification rate in a large linked dataset can make a substantial difference to the number of reported deaths among indigenous peoples when an ‘at least one report’ method of enhancement is used. We suggest that, for enhancement methods using administrative health datasets, the preference should be towards an algorithm that incorporates a weight of evidence. In this study the number of deaths reported among Aboriginal and Torres Strait Islander peoples on the ABS death data is relatively small (n
580), the chance of incorrect reporting of a non-Aboriginal or Torres Strait Islander person as Aboriginal or Torres Strait Islander is also likely to be small, and we suggest that this information be accepted as reported.
It is likely that some deaths of Aboriginal or Torres Strait Islander people are not included in the enhanced counts. There were no linked records for 3.9% of ABS death records. It was not possible to attempt to enhance reporting of deaths for those Aboriginal or Torres Strait Islander people who did not have a relevant health service encounter, or given birth or been born, in the period covered by the study. It is also possible that some Aboriginal or Torres Strait Islander people were not reported as Aboriginal or Torres Strait Islander on the ABS death record or on any of their linked records. It is therefore likely that the count of deaths based on enhancement with the algorithm still represents an under-estimate of the true number of deaths.
Enhancement of reporting of deaths using record linkage does not define whether a person is indigenous. Rather, record linkage results in a statistical construct created for the purposes of planning and research. It provides a mechanism to help reduce the under-reporting of deaths among indigenous peoples in official statistics, and allows adjustment of historical data to obtain improved estimates of the mortality experience of indigenous peoples. Importantly in this study, record linkage resulted in correction of some of the bias in mortality measures resulting from relative under-reporting of Aboriginal and Torres Strait Islander peoples resident in major cities and less remote geographic areas.
We chose to carry out enhancement using all available linked records. A smaller number of years of linked data could have been used, and would have resulted in a different number of reported deaths. Any statistical construct will depend on the purpose for which the data are intended to be used. For example, if the purpose was to examine trends in mortality among over several years, the range of datasets and the years of linked data used should be consistent for each year included in such a study.
Various approaches are possible for algorithms incorporating a weight of evidence, such as a requirement that a certain percentage (e.g. 50%, 75% or 90%) of linked records report that the person is indigenous. Algorithms based on a proportion of records reporting a person as indigenous require a greater weight of evidence than the algorithm used in this study, and would result in a relatively smaller increase in the number of deaths reported as a result of the enhancement. We believe that enhanced reporting of deaths using the algorithm developed in this study provides a balance between achieving a reasonable weight of evidence that a person is indigenous, and maximising the number of additional deaths found through the enhancement.
It would also be possible to develop algorithms where different data collections carry different weights of evidence. For example, linkage could include records from health services that are dedicated to providing services to indigenous people, and these records could provide a greater weight of evidence that a person is indigenous than records collected as part of universal health services. Factors that should be taken into account in determining which datasets should be linked for enhancement purposes include: previous validations studies, representativeness of the community, and the extent to which information in each dataset is collected independently.
Finally, while record linkage provides a mechanism to deal with the issues of under-reporting of deaths among indigenous peoples and to help correct reporting in historical data, it is not a replacement for continued efforts to increase reporting of indigenous peoples on administrative health data collections and death registrations.