What moves a vote in Victoria

A campaign ends and everyone involved explains the result as the product of their own tactic. The explanations are rarely tested, because the counterfactual is never run. But there is a real body of measurement underneath the argument: about why voters leave one party for another, about which kinds of contact move turnout, and about how large any of it is. It adds up to a working picture rather than a verdict — and the picture is more useful than either the campaign memoirs or the standard evidence that campaigning is pointless.

People do change their minds

At the 2010 federal election, 70% of Labor's 2007 voters stayed with it — its highest defection rate since the Australian Election Study began in 1987 — with 14% going to the Coalition and 12% to the Greens¹ . In Victoria now, an Accent Research poll of 1,002 voters finds 57% of 2022 Labor voters staying, 13.4% moving to One Nation, 11% to the Coalition and 8.5% to the Greens² . Defection is normal and continuous: stable partisans fell from 36% (Coalition) and 32% (Labor) at the 1967 election to 13% and 11% in 2025³ .

So people move, and they move in numbers. Two things about *when* matter. The first is that defection is mostly over before the campaign begins: fitting a model to 400 national polls between the 2013 and 2016 federal elections, Jackman and Mansillo found the largest single movement was the jump when Malcolm Turnbull replaced Tony Abbott — about five points, with a half-life of roughly four months — while the formal campaign itself produced almost no movement at all⁴ . Victoria's own leadership change fits: five polls after Ben Carroll replaced Jacinta Allan on 28 July showed no honeymoon, and Labor's vote flat to down⁵ . The second is arithmetic. The five-point jump is about 3.5% of voters changing sides; a well-run contact operation moves about 3 points among the people it reaches and under 1 point across a seat. Both numbers are real, and the larger one arrives on its own timetable — which is what leaders, issues and slow dealignment do, and what a six-week campaign mostly does not.

What the defection evidence says drives it

The 2010 study tested what actually separated defectors from loyalists, and the answer was not messages. Leader ratings were the predominant influence on defection, ahead of policy issues and well ahead of the economy: among Labor voters who moved to the Coalition, an unfavourable rating of Julia Gillard and a favourable rating of Tony Abbott were both significant, with Gillard's slightly the larger¹ .

The Victorian poll shows the same shape in current numbers. Among 2022 Labor voters who have moved to One Nation — 70% of whom did not vote for it in 2022 — crime sits in the top three concerns for 79% and cost of living for 62%, and among Labor's defectors cost of living is first for 47%² . One Nation's growth is concentrated in the Coalition's own vote — about one in five 2025 Coalition voters, and one in three among older, non-graduate, regional voters — rather than in Labor's⁶ .

Labour's defection at the 2010 federal election: 70% held — Of Labor's 2007 voters, 70% stayed at the 2010 election — its highest defection rate since the Australian Election Study began in 1987 — 14% going to the Coalition and 12% to the Greens.¹

The scale of the thing

The benchmark for how large a campaign effect can be comes from field experiments rather than from campaign accounts. Kalla and Broockman pooled 40 field experiments and ran nine more, and put the persuasive effect of campaign contact in a US general election at approximately zero — optimistically, one voter in 175 contacted; their best estimate, one in 800⁷ . Forty-nine real 2016 presidential advertisements, randomised across 34,000 people, averaged +0.7 points on vote choice, not statistically significant⁸ .

That is a finding about general elections in the United States, where the partisan cue is already formed and the campaign is the last six weeks of a long argument. It bounds the campaign's marginal effect, not the total movement of public opinion: the same evidence shows voters changing sides between elections in large numbers. What it says is that the change happens somewhere other than the last six weeks of contact. The authors document where contact does show effect: in primaries, where party cues are absent from the ballot, pooled estimates run about **+4.5 points**, and in ballot-measure elections about **+3.2 points**; early contact measured immediately came in at **+3.2 points** and then decayed before election day⁷ .

Where campaign contact showed an effect: +4.5 pts in primaries — Pooled estimates from the same 49 field experiments. General elections: approximately zero. Primaries: +4.51 points (−1.40 to 10.42). Ballot measures: +3.23 points (1.44 to 5.03). Early contact measured immediately: +3.16 points (1.34 to 4.99), decaying before election day.⁷

Turnout is the other half, and Victoria's is compulsory

The other measurable effect of contact is on turnout rather than on choice, and here the evidence is consistently positive — canvassing first. Gerber and Green's founding New Haven experiment, across about 30,000 registered voters, found door-knocking raised turnout by **8.7 points among the people actually contacted**, against 0.5 points per mailing and nothing measurable for phone calls⁹ . Twenty-four years later, Mann and Haenschen re-analysed 240 field experiments, disaggregated by how salient the election was, and canvassing still leads the table¹⁰ .

Turnout effects by tactic, at the point of contact (CACE) and as assigned (ITT)
Tactic | Low-salience election | High-salience election | Attenuation
Canvassing | 4.58 pts (2.97–6.18) | 3.05 pts (1.20–4.90) | 33%
Volunteer phone calls | 3.62 pts (1.51–5.74) | 2.28 pts (1.12–3.43) | 37%
Paid phone calls | 1.66 pts (0.89–2.43) | 0.74 pts (0.31–1.17) | 55%
Social-pressure mail (ITT) | 2.45 pts (1.67–3.23) | 1.38 pts (0.31–2.45) | 44%
Non-social-pressure mail (ITT) | 0.46 pts (0.26–0.66) | 0.28 pts (0.13–0.43) | 39%
Cold SMS (ITT) | 0.58 pts (0.37–0.79) | 0.22 pts (0.05–0.39) | 62%¹⁰

Read the size of that. The larger numbers are complier average effects — effects on the people actually reached — and the smaller ones are effects on everyone the campaign assigned, which is what a vote count sees. Canvassing contacts 25–30% of the doors it works, volunteer phone banks around 10–12%, and professional phone banks 6%, down from 36% in 1997; all effects shrink again in a high-salience contest like a state general election¹⁰ .

Then the Victorian base makes it smaller still, and this is where the level and the margin part company. Compulsory voting means almost everyone already votes: 87.13% of enrolled voters cast a counted vote in 2022, down from 92.96% in 2010¹¹ . A mobilisation tactic has only the non-voters to work on, and the field-experiment effects above were measured in voluntary-turnout elections where that pool is far larger.

What we know about what does move people

Six things the evidence names

• **Being known and liked.** Leader ratings were the strongest single influence on whether Australian voters defected in 2010, ahead of policy and the economy¹ .
• **Who is in the race.** Contact effects appear in primaries (about +4.5 points) and ballot measures (+3.2), the contests where the ballot does not carry a party cue to fall back on⁷ .
• **Correcting the record, with a replacement.** Retracting a false claim leaves it influential — 5.06 references to it against 4.04 for retraction alone, no significant difference — while retraction *plus* a plausible alternative account brings it to 2.22¹² . Correction changes belief and not vote choice, and decays in about a week¹³ .
• **The issue in your own words, not theirs.** Attention to the issue a far-right party owns raises its support: more immigration coverage predicted a higher vote share for anti-immigrant parties at a three-month lag, controlling for unemployment and actual immigration¹⁵ .
• **Neighbours, not strangers.** GetUp's own post-mortems of 2019 found calling can shift votes inside an existing narrative but not create one, and that its election-day scorecards made no difference — 'in some it actually made the results worse'¹⁶ . Contact works best where a resonant story already exists and where it is done by neighbours rather than callers; the ACTU's review of the same election settled on 'the right policy ideas but the wrong slogan'¹⁷ .
• **A reason to vote for you, not only against them.** Nine hundred and thirty-seven of One Nation's 1,000 surveyed supporters said the country is heading in the wrong direction, against 62% of all voters¹⁸ — a mood that is not answered by being told the alternative is worse.

What the Victorian record already shows

The most useful Australian cases are not experiments at all but campaigns that were measured afterwards. The community independents won Warringah and Indi, and six inner-city seats in 2022, through kitchen-table conversations and local hubs rather than a national brand, and more than 470,000 Australians put a community independent first¹⁹ . In the seats where climate campaigners invested the most door-knocking and community engagement, they mostly won; where they did not, they mostly did not²⁰ . The Victorian caution sits beside it: the state-level version of that wave barely registered in 2022, with no independents in the Assembly at all²¹ .

Then there is One Nation's own ground. It ran the 2026 Nepean by-election with no campaign material visible in South Gippsland and no significant field operation, and still outpolled the Liberals at booths including Rosebud and Tootgarook²² . Its vote arrives with no doorstep operation behind it, which is itself a measurement of how little of it the doorstep is doing.

What to do with what is actually known

Three things, and none of them is 'give up'. The first is scale. A tactic that is worth three points among the people reached is worth well under one across a seat, so the question to ask of any operation is what fraction of the seat it actually touches — the door that is knocked and the door that is opened are not the same door.

The second is that this contest is decided by preferences, not by the primary vote: with the right split, the entire majority question rides on where One Nation's preferences break, and a flow shift from 75–25 to 60–40 moves the two-party figure about three points²³ . Preference conversations are a turnout conversation inside a narrative that already exists — the case the turnout evidence is strongest on — and the Victorian count publishes the distribution seat by seat, so the result is checkable where most campaign claims are not.

The third is to write down the counterfactual before the result, not after it. Strong results are about 40 percentage points more likely to be published than nulls²⁴ , which is why the campaign literature is full of tactics that worked. If you are running something between now and 28 November, the cheapest discipline is to decide in advance which seats you are treating and which you are leaving alone, so that the answer is available either way.