An illustrative case study, TKW Research Group

The brief

With the Victorian state election due in late November 2026, campaigns, media outlets and advocacy groups increasingly want to know what is happening not across the state, but seat by seat. Statewide polling captures the mood of Victoria. It does not tell you who wins a marginal electorate, where a few hundred votes and a shifting preference flow can decide who forms government. That is where seat polling earns its keep, and where method, not the headline number, determines whether the result holds up on election night.

This illustrative case study sets out how TKW Research approaches a single-seat poll in a competitive outer-metropolitan marginal, and the insights that experience surfaces about polling margin seats well.

The scenario

An outer-metropolitan Melbourne seat held on a notional two-party margin of around 2 per cent. Culturally and linguistically diverse population, a large share of mortgage-belt households, and a credible community independent in the contest alongside the two majors. In other words, exactly the kind of seat that swings elections and the kind that statewide polling routinely gets wrong.

The challenge

Three things make a seat like this hard to poll accurately:

1. The margin is smaller than the error. A well-designed seat poll of 400 to 600 respondents carries a margin of error of roughly plus or minus 4 per cent near a 50/50 split. In a seat sitting on a 2 per cent margin, the uncertainty is wider than the contest itself. The design has to work hard to earn a defensible read.
2. The contest is not two-cornered. With a serious independent in the mix, a simple two-party-preferred framing can mislead. You need to capture primary support accurately and model preferences to a two-candidate-preferred result that reflects who is actually left standing after distribution.
3. The population is hard to reach cleanly. Outer-metro marginals are younger, more diverse and more mobile than the state average.

TKW Research’s approach

• Sample design to the seat, not the state. We frame and weight to the electorate’s own profile using ABS small-area data (age, gender within the seat), rather than applying statewide quotas. A marginal is not a miniature Victoria.
• CATI in the respondent’s own language. The real gap in a diverse electorate is not who owns a phone, it is who can and will complete a survey. Online self-completion is effectively English-only, and older voters and non-English-speaking-background voters are far less likely to self-complete online regardless of what device they own. Our CATI AI capability conducts the interview in up to 14 languages with a live-quality experience, which reaches exactly the cohorts an online panel systematically under-counts in a culturally diverse marginal.
• Primary vote first, preferences modelled explicitly. We measure first-preference support for every candidate on the ballot, then estimate the two-candidate-preferred outcome using stated preferences and, where useful, prior distribution behaviour, so the headline reflects the seat’s actual dynamics.
• Nuance that self-completion cannot capture. An online survey can ask “are you undecided?” and record the answer. What it cannot do is hear the hesitation behind it, the voter who says one party but sounds unconvinced, or the reason a soft voter gives when prompted. In a seat where a small soft bloc decides the outcome, that nuance is the intelligence, and it is where live and AI-assisted interviewing earns its place.

What it delivers

A defensible primary and two-candidate-preferred estimate for the seat, reported as a range rather than a single spurious number, alongside the drivers underneath it: which issues move which cohorts, where the independent draws from, and where the soft vote sits. That is intelligence a campaign or newsroom can act on, not just a scoreboard.

Five insights into polling margin seats:

1. You cannot infer a seat from the statewide swing. Swings are never uniform. Local candidate quality, incumbency effects, demographic change and a single dominant local issue routinely produce seat swings that bear little relation to the state figure. Margin seats are precisely where this divergence is largest.
2. The margin of error is often bigger than the margin. In seats decided by one or two points, a single poll is a snapshot with real uncertainty around it. Read it as a range, and treat tracking over time as far more informative than any one result.
3. Compulsory voting changes the question. With turnout above 90 per cent, Australian seat polling is not about predicting who turns up, as it is in voluntary-voting systems. It is about how people rank their choices and where soft and undecided voters land. That shifts the analytical effort to preference modelling and the informal vote.
4. Preferences decide marginals, and independents have scrambled the old assumptions. The rise of community independents means default two-party preference flows can no longer be assumed. Getting the two-candidate-preferred framing right is now central rather than a footnote.
5. Mode and language decide who you actually hear from. The seats that swing are the most culturally and linguistically diverse, and self-completion online is largely English-only. Older and non-English-speaking-background voters are the least likely to complete a survey themselves, whatever device they hold. Interviewing that runs in a respondent’s own language, and can read hesitation as well as record an answer, is not a nicety in these electorates. It is the difference between a poll that reflects the seat and one that flatters a comfortable assumption.

Why TKW Research

Nearly 30 years of Australian field experience, ISO 20252:2019 certification, and a multilingual CATI AI capability built for exactly the hard-to-reach populations that decide marginal seats. When the margin is thinner than the error bar, design is everything.