Methodology
How suburb crime estimates are made
Most police forces publish recorded crime by council area. South Australia also publishes it by suburb. For the rest, the map shows a modelled estimate of how a council's recorded offences are likely shared between its suburbs, always marked as an estimate with a likely range and a confidence score. It never replaces a published figure.
The method
A Poisson regression, a standard statistical model for counts, learns from 1,198 South Australian suburbs how offences per resident relate to measures available for every Australian suburb:
- the ABS Socio-Economic Indexes for Areas (SEIFA 2021): disadvantage, economic resources and education and occupation, and how the suburb compares with the rest of its council;
- the suburb's population at the 2021 Census, and its distance from the state capital;
- its council's own recorded crime rate.
For every other suburb in a council with published figures, the model predicts a relative level, and the predictions are then scaled so each council's suburbs add up to that council's published total (2025-26, 3-year average). The model decides only how a council's recorded crime is shared between its suburbs, never how much there is. Suburbs with fewer than 200 residents get no estimate, because a rate per resident means little there.
How accurate it is
Tested by leaving out one group of South Australian councils at a time, predicting their suburbs from the rest, and scaling to the councils' real totals exactly as for other states. The comparison is the simplest alternative: sharing each council's total by population.
| Offences | Error, model | Error, by population | Better by | Rank corr. |
|---|---|---|---|---|
| All offences | 93.1 | 125.1 | 26% | 0.35 |
| Violent | 14 | 24.8 | 44% | 0.48 |
| Property | 79.3 | 99.3 | 20% | 0.27 |
| Break-ins | 7.7 | 11.2 | 31% | 0.24 |
Error: mean absolute difference from the published count, offences per suburb over the pooled years. Rank corr.: how well the model orders a council's suburbs from highest to lowest rate (1 is perfect, 0 no better than chance; sharing by population scores about 0). In 80% range: share of held-out suburbs inside their 80% interval, which is set from the same tests.
Limits
- The model learned from one state. Police in other states record offences differently, so estimates outside South Australia carry lower confidence.
- Suburbs are placed in a council by their centre; a suburb spanning two councils is estimated from one.
- Places with many visitors and few residents (city centres, shopping and nightlife precincts, airports) record far more crime per resident than their profile suggests. Their estimates are the least reliable.
- Estimates describe likely recorded crime, not personal risk, and are rebuilt whenever councils' figures are refreshed.
Model version 1, built 2026-10-26. Data: police recorded crime by council and South Australian suburb; ABS SEIFA 2021 and Census 2021 population (CC BY 4.0). The estimates file is public at data/suburb-estimates.json.