Ask a psephologist which demographic groups are most likely to vote Greens, and they’ll have a confident answer.
But turning that intuition into a precise, location-specific estimate – one tied to real election data and adjustable by electorate – requires a more systematic approach.
The model used for The Swing Is On is called ecological regression, a statistical technique for inferring how different types of people voted when you can only directly observe how geographic areas voted. This article explains how it works, where it uses other approaches, and what it can and can’t tell you.
The basic problem
The 2021 Australian Census tells us, for every polling booth’s catchment area, what share of the population falls into each age group, income bracket, ancestry category, and so on. The official 2025 federal election results from the Australian Electoral Commission tell us what share of each Election Day booth voted for each candidate in each seat. What neither dataset tells us directly is how young people voted, or how renters voted, or how people born overseas voted.
Ecological regression bridges that gap. If booths with higher proportions of young residents consistently show higher Greens vote shares – even after accounting for other differences between those booths – that’s evidence of a genuine relationship between youth and Greens voting, even though no individual voter’s demographic characteristics and vote choice are observed together.
The model uses this approach across 574 separate regressions, covering 82 demographic categories across 15 Census variables.
Stage 1: regression
For each demographic characteristic, a regression model is estimated using booth-level data: the proportion of a booth’s estimated eligible voters (Australian citizens aged 18 or over living nearest that booth in the 2021 Census) in a given demographic category on one side, and the booth’s vote shares across all seven party classifications on the other.
Booth size is used to weight the regression, so booths with more nearby eligible voters contribute more to the estimates.
Rather than using extreme predictions at 0% or 100% demographic concentration – values that often don’t actually exist in the data – the core output for each characteristic is a predicted vote rate at the 95th percentile of population proportion across all booths. This is a high-but-observed level of demographic concentration, making it a meaningful and realistic reference point.
Stage 2: national calibration
Raw ecological regression can overstate demographic effects, and for a specific reason. Certain demographic groups are geographically concentrated in ways that aren’t purely about demography. Young people are disproportionately found in inner-city electorates that lean Greens for many reasons beyond age alone – local culture, housing composition, proximity to universities. Without correction, the regression would attribute too much of the Greens vote in those areas to youth, rather than to the broader character of the neighbourhood.
National calibration corrects for this. Each estimated rate is adjusted so that the model’s implied prediction at the national average concentration for a demographic category exactly matches the actual 2025 national primary vote shares. The adjustment is made in logit space – a transformation that keeps probabilities between 0% and 100% – and is applied to all seven classifications simultaneously.
In plain terms: the model is anchored to reality at the national level before being applied anywhere.
Stage 3: location calibration
Nationally calibrated rates are then adjusted for each specific geography. For each electorate, state, geographic type (inner metropolitan, outer metropolitan, provincial, rural), and the national aggregate, an offset is found that makes the location’s demographic rates consistent with the actual 2025 primary vote distribution for that location.
For a seat like Indi, which recorded 45% for Other, Left-leaning candidates in 2025, all demographic rates are anchored to that 45% base. A 25-to-34-year-old in Indi is predicted relative to Indi’s actual vote distribution, not a national average that would produce a very different starting point. A One Nation swing in Queensland is modelled relative to Queensland’s actual One Nation share, not the national figure.
Sex and age: a different approach
For sex and age variables, ecological regression runs into a specific limitation.
The male-to-female ratio is nearly constant across all polling booths in Australia. A regression that can’t see any meaningful variation in a predictor variable across booths can’t estimate that variable’s effect on vote choice. Sex rates are instead drawn from the 2025 Australian Election Study, a post-election survey of a nationally representative sample of voters, and then location-calibrated using the same iterative offset method as the other variables.
Age presents a different problem. The ecological regression for age correctly identifies that younger areas lean left and older areas lean right, but it can’t separate Labor from Greens support within the same geography. Young voters in inner-city seats are concentrated in areas that lean Labor overall, so the regression systematically over-attributes their left lean to Labor and understates the Greens share.
Survey data from the Australian Election Study directly captures this split at the individual level, giving materially more accurate age-based predictions. Age and age-by-sex variables use AES survey rates at the national level, then location-calibrated to each geography.
In the same way that the ecological regression approach isn’t perfect (more on this shortly), neither is the once-off AES poll conducted in the aftermath of the 2025 federal election. Survey respondents are not obliged to be truthful about their voting behaviour, or may not recall their past decision correctly. Ecological regression has the benefit of relying on actual voting data and a sample size of every voter who cast their ballot on Election Day – so, where reasonable, it remains our preference to estimate demographic voting behaviour using ecological regression.
Selecting two characteristics
When two characteristics are selected simultaneously – say, “25–34 year olds” and “renters” – the model combines the two sets of location-calibrated rates using a multiplicative formula that treats the local vote distribution as the reference point.
For each classification, the combined rate is proportional to the product of the two individual rates divided by the local base vote share. The result is normalised to sum to 100%. This formula has a natural property: selecting a single characteristic returns its calibrated rates directly, and adding a second characteristic adjusts the base proportionally in both directions.
Cross-tabulation data from the 2021 Census is used to estimate the joint prevalence of the two characteristics for pool size calculations. For example, voters with Chinese ancestry are much more likely to speak Mandarin at home than Italian – there is not an equal distribution of each demographic variable over every other.
The result is deterministic: the same profile and location always produce the same result.
What ecological regression can and can’t tell you
Ecological regression is a method that infers individual behaviour from group-level statistics. Its strength is that it draws on thousands of polling booths, each providing a data point about the relationship between demographic composition and vote choice. The patterns it identifies are real statistical signals.
Its limitation is that it cannot observe individuals. Renters in a high-Greens electorate are estimated as more likely to vote Greens because that’s what the booth-level data consistently shows – but not every renter in that electorate voted Greens, and some of the correlation reflects the broader character of the neighbourhood rather than rental status itself. The nationally calibrated step reduces this confounding, but does not eliminate it.
Variables set at the dwelling level in the Census – household income, tenure type, and family household composition – carry an additional caveat. A voter in a renting household is recorded as renting regardless of whether they actually pay the bills. A voter in a high-income household is recorded as high-income regardless of their personal salary (though personal income is separately selectable, derived from individual-level data). These variables are better read as indicators of household environment than as confirmed individual characteristics.
The model reflects observed correlations at the booth level, adjusted for location. It is a well-grounded estimate of how a demographic profile relates to voting behaviour – not a precise profile of any actual voter.
