Improving the quality of reporting of Aboriginality in police data through record linkage

Summary

Background

A common challenge with administrative records is the accurate recording of Aboriginal identification.  In NSW, data recorded by NSW Police is the primary source of Aboriginality information for people who come into contact with the criminal justice system. Prior to 2022, Aboriginality was frequently missing from police records, with ‘unknown’ recorded for around 15% of persons of interest (POIs) and 30% of victims of violent crime. This limited the ability to monitor over-representation, provide accurate information to communities, and track progress against Closing the Gap targets.

In January 2022 NSW Police introduced a mandatory Aboriginality question applicable in most policing contexts. The key exception to this requirement is traffic-related infringements. This substantially improved data quality with ‘unknown’ Aboriginality records falling to around 3% for in-scope POIs and victims by the end of 2023. However, the change also introduced a break in the time series, meaning figures before and after 2022 could no longer be directly compared. 

To improve the quality of historic data and address the 2022 discontinuity, this study examined whether algorithmic methods could infer likely Aboriginality from linked longitudinal police records to improve the completeness and accuracy of the data. Several approaches were tested, including ever-identified, twice-identified, and a range of percentage-based algorithms. These approaches were evaluated against two more reliable ‘source of truth’ datasets: NSW custody records and post-2022 police data.

Key findings

Percentage-based methods consistently outperformed ever-identified and twice- identified methods. They produced counts closest to the comparison datasets and demonstrated the lowest classification errors and highest precision. In contrast ever- identified and twice-identified approaches substantially overcounted Aboriginal people (by 37% to 96%, and 25% to 50% respectively), due to misclassification of non-Aboriginal individuals, making them unsuitable for ongoing use.

The optimal percentage-based approach differed between adults and young people: the best thresholds for adults overcounted Aboriginal young people, while thresholds for young people undercounted Aboriginal adults. Accurate results for both groups were achieved through a two-step method that applied different thresholds depending on whether a person had any adult police records. 

1. People with an adult record are counted as Aboriginal if at least 25% of their adult records identify them as Aboriginal

2. People with only youth records are counted as Aboriginal if at least 35% of their youth records identify them as Aboriginal.

Figure 1 demonstrates that the two‑step method achieved the strongest overall performance for both groups, delivering accurate Aboriginality counts with high precision. Estimates closely aligned with high‑quality post‑January 2022 police data, matching adult counts within 1%, while substantially reducing the overcounting of Aboriginal young people seen in simpler methods (to around 4–5%). Importantly, the two‑step method resolves the discontinuity introduced in 2022, enabling consistent and comparable measurement over time.

Figure 1. Number of adults proceeded to court and under 18 POIs recorded between 2017 and 2021 who were counted as Aboriginal by various algorithms

Conclusion

Ideally, information about a person’s self-identified Aboriginality would be collected accurately each time an individual comes into contact with the justice system. Despite substantial improvements in recent years, gaps and inconsistencies remain, particularly the high proportion of missing Aboriginality records for traffic infringements.

This study demonstrates that incomplete Aboriginality records can be meaningfully improved through the application of an algorithm that draws on patterns of identification across an individual’s record of police contacts. The resulting data supports more reliable aggregated reporting, statistical analysis, and community data provision. However, derived Aboriginality is not appropriate for determining an individual’s identity for frontline or operational decision-making, or for assessing eligibility for programs or interventions.