How we tested what a suburb says about default

Data and licences

Public, openly licensed data only; nothing scraped, and nothing from any lender. Each file’s address, licence, retrieval date and SHA-256 fingerprint:

AFSA’s figures are the copy on data.gov.au (Creative Commons Attribution 2.5 Australia); AFSA’s own website allows only non-commercial use. Western Australia Police publish crime by locality only in an interactive report with no reuse licence; Tasmania, the Northern Territory and the ACT publish no open suburb-level crime files. ABS material is Creative Commons Attribution 4.0.

Geography

Every mesh block, the ABS’s smallest unit, belongs to exactly one SA1, SA2, SA3, suburb and council, so mesh blocks and their 2021 Census residents give every correspondence. Of 15,353 suburbs and localities, 459 straddle an SA3 boundary; we split them by residents (0.9% of people sit in the smaller part). Areas with fewer than 5,000 adults are left out of the analysis (330 remain, home to 25.4 million people).

The measure

Non-business personal insolvencies per 10,000 adults a year, July 2022 to June 2026. Adults are the ABS estimated resident population of each SA3 (its SA2s added up; each financial year’s middle is the mean of its two Junes; June 2026 extended at 2025’s growth) times the Census share aged 18 and over (2011, 2016 and 2021, straight lines between). AFSA hides counts of one or two; we take them as 1.5 (10% of area-quarters, weighted by adults). Taking them as 1 or 2 instead moves no ratio by more than 0.05 on a log scale. Where only one side of the business split is hidden, it is the total less the other side.

Crime: violent and non-violent

Each state’s categories, mapped to the two kinds. Counts are offences (South Australia, Queensland) or incidents (New South Wales, Victoria) recorded by the location of the offence. New South Wales, Victoria and South Australia publish by suburb, which we match to ABS suburbs by name (Victoria’s council breaks ties; a name used twice in a state with no council to tell them apart is split by residents: 1.5% of New South Wales incidents) and then to SA3s by residents. Queensland publishes by police division: we placed each SA1’s centre in a division and spread the division’s offences over its residents.

Level: offences per 1,000 residents, July 2022 to June 2026, as a log ratio to the state’s adult-weighted average. Growth: the change in that log rate from July 2016 to June 2019 to July 2023 to June 2026, less the state’s average change. Queensland’s recorded assaults roughly doubled from 2020 to 2022, a statewide change that comparing within the state removes. Property crime alone, which every state records alike, gives the same answer (with the rest held equal: 1.2 times the rate for its level, about the same for its growth).

Inequality

The Gini coefficient is the ABS’s, from personal income tax records for 2022–23 (total income of earners, excluding government pensions and allowances). As a check we computed a household Gini from the Census income bands: each band’s households at its middle, nil and negative incomes at zero, the open top band ($4,000 a week or more) at the mean of a Pareto tail fitted to the two highest bands, and straight lines between bands on the Lorenz curve, which slightly understates inequality. The two correlate at 0.74. For timing we used the ratio of mean to median income, published for every year from 2018–19, which tracks the Gini (0.93 in 2022–23). Inequality and median income are nearly independent across areas (correlation 0.20), so the 3 × 3 grid has areas in every box.

At the same median income: Gini 56% lower; Census household Gini 53% lower; top tenth’s income share 52% lower; the 80th to 20th percentile ratio 59% lower. With the full profile held: about the same, 9% lower, about the same and about the same, none clear. The share of earners in the national bottom quarter of income, with the rest held, goes with less insolvency (39% lower), which fits people with little income and few assets rarely filing.

The model behind “with the rest held equal”

Weighted least squares on the log of each area’s rate, weighted by adults, with a separate level for each state; predictors standardised; standard errors allow for correlation within SA4s (groups of neighbouring SA3s). The profile held equal: median household income, income inequality, unemployment, adults with a degree, renting privately, social housing, paying off a mortgage, one-parent families, adults aged 18 to 34, people aged 65 and over, remoteness. A factor outside it is added to it on its own. “Top fifth against bottom fifth” is the model’s ratio between the adult-weighted average area of each fifth; the 95% range comes from the standard error. R² 0.86; on held-out groups of neighbouring areas (ten folds by SA4) 0.82.

How firm each verdict is

Spatial: insolvency clusters (Moran’s I within states 0.44); after the profile, little is left (0.17). Precision: SA3s hide much of the difference between suburbs, so area figures are blunt:

Timing tests

For the Census profile: which 2016 characteristics come before the change in insolvency from 2017–19 to 2022–26, holding each other and the earlier rate (and, in the stricter test, the 2012–17 rate too); and does insolvency in 2012–17 come before the 2016–21 change in each characteristic, holding all the 2016 levels? AFSA changed how it places addresses in September 2017, so its earlier figures are used only as a starting level.

Crime: the level tests hold each later measure’s starting level and the 2021 profile; by state, violent crime came first in New South Wales only. The pre-pandemic growth test (New South Wales and Queensland, three-year blocks 2007–10, 2010–13 and 2013–16; 171 areas) corrects for the noise in small insolvency counts, which would otherwise pull the reverse test toward zero (reliability of the earlier insolvency change 0.62); it uses a method-of-moments correction with a bootstrap over SA4 clusters. The same test across the pandemic (2016–19, 2019–22, 2022–26) is not reported: the 2019–22 counts were so low that the change was 90% noise. The year-by-year test is a count model of each area’s insolvencies with the area’s own level and each state’s year held (2,120 area-years).

Patterns and the decision tree

Flags are set against national figures from the 2021 Census (weighted by adults): unemployment 5.1%, adults with a degree 26%, homes rented 31%; crime against each state’s rate of recorded violent and non-violent offences per resident. Rates for a group are all its insolvencies over all its adult-years; “against its state” in the tables divides by what each area’s state rate would predict. The tree splits on the Poisson deviance of insolvency counts around each state’s rate, three levels deep, at least 15 areas a group; on held-out SA4 groups it explains 63% of the deviance (two levels 53%, four 62%). Given crime as well (four states), it splits on violent crime once, below the education and age questions.

Area types

K-means on twelve standardised Census measures (median household income, income inequality, unemployment, adults with a degree, renting privately, social housing, paying off a mortgage, one-parent families, people aged 65 and over, adults aged 18 to 34, remoteness, flats and apartments, and flats), 100 starts. Seven groups fit about as well as four to eight (silhouette 0.25; the best is 0.25) and keep mining regions and remote communities apart. Refitting on 50 bootstrap samples gives a median adjusted Rand index of 0.68: most areas keep their type. Names follow each group’s defining measure.

Fairness screen

The area score is the profile model refitted without its age measures and predicted without state levels; protected attributes were never inputs. It correlates with the share of Aboriginal and Torres Strait Islander residents (+0.36) and of people needing help with core activities (+0.47).

Limits

• Insolvency stands in for default and undercounts it unevenly, most where people have least.
• Areas, not people: a link between areas need not hold for the people in them.
• Recorded crime depends on reporting and policing; four states only.
• The 2021 Census was taken during lockdowns in Sydney and Melbourne, which affected some work measures.
• Timing tests show what changed first; they cannot rule out a cause that moves both.

Sources

1. Regional personal insolvency time series per quarter, SA3, September 2007 to June 2026. Australian Financial Security Authority, on data.gov.au, CC BY 2.5 AU, retrieved 30 October 2026.
2. Regional population, 2024–25: estimated resident population, 2001 to 2025, SA2 and above. Australian Bureau of Statistics, CC BY 4.0, retrieved 30 October 2026.
3. Census 2021 DataPacks: General Community Profile (SA3, SA2, suburbs and localities) and Time Series Profile (SA3). Australian Bureau of Statistics, CC BY 4.0, retrieved 11 and 30 October 2026.
4. Personal Income in Australia, 2022–23: Tables 1 and 3, income and its distribution by geography. Australian Bureau of Statistics, CC BY 4.0, retrieved 30 October 2026.
5. Australian Statistical Geography Standard, Edition 3: allocation files, SA3 boundaries and mesh block counts. Australian Bureau of Statistics, CC BY 4.0, retrieved 11 and 30 October 2026.
6. Recorded crime: NSW Bureau of Crime Statistics and Research, by suburb (CC BY); Crime Statistics Agency Victoria, by suburb (CC BY 4.0); Queensland Police Service, by police division (CC BY 3.0 AU); South Australia Police, by suburb (CC BY). Retrieved 30 October 2026.
7. Every file, with its address, licence, the day it was retrieved and its SHA-256: Data and licences, above.