Every figure in this atlas is a model estimate, not a measurement. No census, survey or test in Australia records the cognitive ability of suburb populations. What follows is the full chain from raw Census 2021 tables to the number you see on the map, including every assumption that a sceptical reader should attack.
We take seven Census 2021 composition measures of each area — education, occupation, income, unemployment and household structure — standardise them nationally, combine them into a single socio-cognitive factor, and place that factor on an IQ-like scale with a national mean of 100 and a between-area standard deviation of 6.5. The result is an estimate of the average of the adult residents of an area, on a scale where 100 is the Australian mean and 15 is the individual standard deviation.
Individual intelligence testing is expensive, voluntary, and never administered to a geographically representative sample of 25 million people. Australia has no register of IQ by suburb. Any claim of the form "suburb X has an average IQ of Y" is therefore one of three things: a survey of volunteers (hopelessly self-selected), a re-badged socio-economic index, or an ecological inference — estimating an area property from correlated area properties. This atlas is the third kind, and it is honest about it.
Estimating cognitive ability from area characteristics has peer-reviewed precedent, not just internet hobbyists: McDaniel (2006) synthesised state-level US estimates, Carl (2018) examined regional cognitive variation, and Lynn and colleagues published cross-national area estimates. They differ in rigour and in what they claim, and they deserve engagement on their methods rather than dismissal. The design choice that matters most is whether you publish an interval with the point estimate and whether you say plainly that the map describes composition. This one does.
The inferential chain rests on a well-replicated fact: across many countries and datasets, aggregate measures of educational attainment and occupational class in an area predict the average cognitive test performance of that area's residents with moderate-to-strong correlation. Education and occupation are, at the individual level, among the strongest demographic correlates of measured cognitive ability. Aggregating them to an area and using the aggregate to estimate the area mean is legitimate — and it is a different, weaker claim than estimating any individual.
A suburb average tells you nothing reliable about any person living in it. With a between-area standard deviation of 6.5 and an individual standard deviation of 15, the overlap between any two suburbs is enormous: two randomly chosen residents of the highest-scoring suburb and the lowest-scoring suburb are more likely to differ in the "wrong" direction than the right one. Concretely, the 90% range for an individual around any given suburb mean is about ±23 IQ points — wider than the entire span of suburb means on this map. This is the ecological fallacy, and it is not a footnote: it is the main thing to understand before reading the map. Robinson's (1950) classic demonstration found the same variables correlating at −0.53 between areas and +0.20 between individuals — opposite signs. Area-level and individual-level relationships are different objects.
| Dataset | Use | Source / licence |
|---|---|---|
| Census 2021 General Community Profile, SA1 level | All indicator numerators and denominators (61,844 SA1s) | ABS DataPacks, CC BY 4.0 |
| SEIFA 2021 (IRSAD, IRSD, IEO, IER) | Independent cross-check of the composite; displayed in the panel | ABS, CC BY 4.0 |
| ASGS 2021 — SA1, SA2, Suburb & Locality boundaries | Geography and the SA1→suburb allocation | ABS, CC BY 4.0 |
Geography is resolved at the finest published level — the Statistical Area Level 1 — and then pooled into suburbs (Suburbs and Localities, SAL) and districts (SA2). Pooling uses summed numerators and denominators, never averaged percentages, and medians are recomputed from the pooled income-band distribution rather than averaged across SA1s.
The SA1→suburb allocation is not published by the ABS, so we derive it: each SA1 is assigned to the suburb polygon containing its area-weighted centroid. Because both geographies are built from whole SA1s, this is exact in practice — — of SA1s were assigned. As an independent check, a population-weighted aggregation of SA1 SEIFA scores reproduces the ABS's own published suburb-level SEIFA scores at r = — across 8,322 suburbs. A broken allocation would not survive that test.
Seven area-composition measures, each oriented so that a higher value means a higher expected cognitive score. All are census counts of people, so they describe who lives there, not what the place provides.
| Indicator | Block / weight | Why it belongs |
|---|---|---|
| Adults with a bachelor degree or above | Education 0.30 | Educational attainment is the demographic variable most strongly associated with measured cognitive ability at the individual level. |
| Adults with a postgraduate qualification | Education | Sharpens the top of the education distribution, where variation in attainment saturates. |
| Employed residents in managerial or professional occupations | Occupation 0.30 | Occupational class is the other classic individual-level correlate, and adds information not captured by education alone. |
| Employed residents in labouring occupations | Occupation | Negative-loading counterpart; widens the occupational spread beyond a purely professional share. |
| Median household income (log) | Economic 0.13 | Economic advantage is a genuine independent signal, though a weaker one than education; logged because income is heavily right-skewed. |
| Unemployment rate | Economic | Captures exclusion from the labour market that occupation shares alone miss. |
| Share of residents in single-parent households | Household 0.12 | Household structure correlates with area disadvantage in a way income does not fully absorb. |
| Charitable donation deductions claimed per taxpayer | Civic 0.15 | ATO tax data, allocated from postcodes. Included because it is the one behavioural measure that predicted measured school achievement beyond everything above (+0.047 R²), and it is nearly orthogonal to the rest of the index (r = 0.17). See the correlates section for what it does and does not measure. |
These variables are strongly predictive of test scores and were left out on purpose. Their inclusion would have raised the apparent accuracy of the model while changing what it measures from socio-economic composition to ethnic composition.
Superficially this makes the model look less accurate. It also means the number on the map is not a shortcut for the ethnic composition of the area, and we consider that a hard requirement rather than a trade-off.
Each indicator is converted to a z-score across all modelled suburbs, weighted by population, so that a suburb of 30,000 counts more toward the national baseline than a locality of 80. The indicators are grouped into four blocks (education, occupation, economic, household), combined inside each block, then combined across blocks with the weights in the table above.
A theory-weighted scheme invites the objection that the weights were chosen to produce a preferred answer. So we also compute an unweighted principal-component solution over the same indicator matrix — the standard data-driven alternative — and compare the two orderings. They correlate at r = —. The first principal component explains — of the variance in the indicator set, and every indicator loads on it with the expected sign. The weighting choice is therefore not load-bearing: any sensible scheme produces the same ranking.
A split-half check reaches the same conclusion from a different direction. Splitting the seven indicators into two arbitrary halves and scoring each independently yields two composites that correlate at r = —. The construct is stable across the indicators that make it up.
Two constants convert the composite into IQ points, and both are assumptions rather than measurements. We state them plainly.
The mean is free. The population-weighted mean is set to 100 by construction, because the national mean of an IQ scale is 100 by definition. This is the only part of the scale that is imposed rather than estimated.
The spread is an assumption: 6.5. This is the between-area standard deviation — how much the average of a suburb varies. It is emphatically not 15. Fifteen is the spread of individuals; a suburb's mean is an average of hundreds or thousands of people, so its spread is smaller by roughly the square root of the number of people, adjusted for the fact that people who live near each other resemble each other. Published estimates of how much of individual cognitive variance sits between small areas in developed countries cluster between 10% and 26% — PISA 2018 put Australia's between-school share at 18%, and US population data give an intraclass correlation of about 0.17–0.19. That implies a between-area standard deviation of roughly 4 to 7 points. We take 6.5, the top of that range, which is deliberately conservative: it makes the map's contrasts larger, not smaller.
This is the number to attack. It is worth being concrete about what happens if it is wrong:
| If the true between-area SD is | Implied share of individual variance | Then the modelled range becomes | Effect on every number |
|---|---|---|---|
| 5.0 | ~11% | 77 – 115 | shrink all departures from 100 by ×0.77 |
| 6.5 (used) | ~19% | 70 – 119 | as shown |
| 8.0 | ~28% | 63 – 123 | stretch all departures from 100 by ×1.23 |
Crucially, the ranking of suburbs is completely unaffected by this constant — it scales every departure from 100 uniformly. If you disagree with our spread, you can rescale the whole map with one multiplication and every ordering still holds. That is why the panel leads with the percentile rank as well as the point estimate.
Each estimate carries a 90% interval, typically ±6 to ±9 points. It combines two distinct sources of error, and it is worth knowing which dominates:
Because model error is largely common across areas rather than independent, relative comparisons are far more reliable than absolute ones. Two suburbs ten percentile points apart are almost certainly ordered correctly; two suburbs two points apart are indistinguishable. Treat the map as a ranking with a wide absolute scale, not as a set of precise measurements.
Five checks, all reproducible from the pipeline. The first four validate the machinery; the fifth is the only genuine test of whether the index tracks anything real.
| Check | Result | What it rules out |
|---|---|---|
| Geography: aggregated SA1 SEIFA vs ABS published suburb SEIFA | r = —, mean absolute difference — index points | A faulty SA1→suburb allocation, which would scramble the pooled data |
| Coverage: population captured vs Census total | — (25.35M of 25.42M) | Systematic loss of geography |
| Construction: theory weights vs principal component | r = — | Weights tuned to a preferred result |
| Stability: split-half of the indicator set | r = — | An index that depends on one or two variables |
| External criterion: measured NAPLAN achievement (WA public schools, 230 districts from 6,763 school-test records, 2025) | r = —, R² = — | The index having no relationship to measured academic performance |
The NAPLAN test is the important one. NAPLAN scores are a measured outcome that was not used to build the index in any way, and the index explains — of the variance in district-mean achievement across Western Australia.
One honest limit on how far that can be pushed: NAPLAN is a curriculum-anchored test, and there is no published correlation between NAPLAN scores and full-scale IQ. Anyone who quotes one is inventing it. What the literature does establish is that achievement tests generally correlate with latent g at about 0.81 (Deary et al. 2007, >70,000 children), and that NAPLAN tracks PISA reading at roughly r = 0.7 (NCVER Technical Report 86). So this coefficient is best read as a conservative estimate of the index's relationship with area cognitive ability — which means the error bars elsewhere on this page are, if anything, too wide rather than too narrow.
The validation is also geographically partial: school-level NAPLAN is published only for Western Australian government schools (664 schools with 2025 results, about 6.8% of Australian schools). Nothing here validates the independent-school sector, small or remote schools, or any state outside WA. That is a real limitation of the evidence, not a caveat we can wave away.
A second school-data check correlates the index against area-mean ICSEA nationally (r = — across 1,941 districts and 9,623 schools). That is not independent evidence and we do not present it as validation. It is criterion contamination: ICSEA is built from parental education and occupation — the same inputs as our index — and it is designed to predict the same NAPLAN outcome. ICSEA alone explains around 72% of school-level NAPLAN variance, which sounds enormous until you notice that the equivalent student-level figure is nearer 9%; most of that number is aggregation inflation. A high correlation here confirms the geography join and the aggregation, and nothing more.
Licensing note: ACARA's My School terms of use prohibit redistributing school-level data. The school datasets are used here solely to compute the aggregate coefficients printed above; no school-level or district-level derived values are published in this atlas, and the source files are not redistributed.
What none of this can do is test the index against area-level IQ directly, because Australia has no such dataset. The NAPLAN test is the closest available substitute: it measures an outcome, not an input, and it holds up. The step from measured achievement to cognitive ability — correlation around 0.8 in the literature — is where the remaining uncertainty lives, and it is why the map is labelled an estimate.
Behavioural data is more attractive than demographics because it measures what people do rather than who they are. So we tested it rather than assuming either way. Three behavioural measures were built for every district and put through the same question: does this improve prediction of measured NAPLAN achievement once the socio-economic index is already in the model? Raw correlation is meaningless here — with thousands of areas almost everything correlates with everything — so the test is the incremental change in R².
| Behavioural correlate | Source | Incremental R² over the index | Verdict |
|---|---|---|---|
| Knowledge amenities per 10,000 residents (libraries, bookshops, universities) | OpenStreetMap | +0.0009 | Rejected |
| Fast-food outlets per 10,000 residents | OpenStreetMap | +0.0003 | Rejected |
| Net change in business stock over the financial year | ABS CABEE | +0.0082 | Rejected |
| Charitable donation deductions claimed per taxpayer | ATO Taxation Statistics | +0.0473 | Included in the index |
| Work-related self-education claims per taxpayer | ATO Taxation Statistics | +0.0048 | Rejected |
| Patent filings per 10,000 taxpayers (2015+, 2019+, individual and organisation) | IP Australia IPGOD | +0.0003 to +0.0016 | Rejected |
Five of the six other candidates were rejected, and that is the useful result: each is essentially re-measuring the same socio-economic latent variable the index already captures.
Donation claiming was the exception, and it is now in the index. The share of residents claiming charitable donation deductions predicts measured NAPLAN achievement beyond the socio-economic blocks, adding 0.047 R² — and it survives 5-fold cross-validation with a +0.043 out-of-sample gain, so it is not an artefact of overfitting. It is also close to orthogonal to the rest of the index (r = 0.17), which is why it contributes at all. Adding it moved the index's correlation with measured achievement from r = 0.763 to r = 0.790. Its principal-component loading is only +0.116 against roughly 0.42 for education and occupation, which is the signature of a genuinely partly-independent signal rather than another proxy.
Two honest caveats. First, this is not a generosity measure: deductions are claimed by people who itemise, so the variable largely captures financial literacy, tax-agent use and institutional density — having a charity worth giving to and the wherewithal to claim it. Second, it is allocated from postcode data, so it differentiates districts and the suburbs that straddle postcode boundaries, but two neighbouring suburbs inside one postcode receive the same value. Where no allocation was possible the block contributes nothing for that area rather than excluding it.
The patent measure deserves a note on method: our first construction divided every filing in IPGOD — which reaches back a century — by 2021 population, conflating a stock with a snapshot. Re-run as a ten-year and five-year flow, split into individual and organisation applicants, the answer did not change: no meaningful increment in any specification. The rejection stands on merit.
The same discipline applies to the obvious candidates we did not build. Voting behaviour was ruled out on evidence, not squeamishness: once education, occupation, income and urban/rural geography are in the model it is largely redundant, the individual-level IQ–political-preference literature is weak and contested, and presenting party colour beside an IQ estimate would assert something the data cannot support. Crime rates were rejected because they measure policing and reporting intensity as much as behaviour. ATO postcode data (self-education deductions, donations) and IP Australia patent filings are the strongest remaining candidates and would need a postcode→district population weighting before they could be tested on the same footing.
Everything is generated by a pipeline that downloads from the ABS, derives the indicators, builds the index, and loads PostGIS. No hand-edited data, no manual corrections, no scraped mystery numbers. The steps:
All source data is Commonwealth copyright, released under CC BY 4.0. The derived estimates are published in the same spirit: use them, argue with them, but quote the uncertainty.
Built with ABS Census 2021, SEIFA 2021 and ASGS 2021.
This page is generated from the same pipeline output as the map; the figures in it are read from
data/meta.json at load time, so they cannot drift from the model.