Previous Next

Landslide

19

Turnaround: How Labor won voters

Simon Jackman

Abstract

Labor’s turnaround leading up to the 2025 election was one of the most rapid in Australian electoral history, with two-party-preferred support rising by at least 4.3 percentage points (and possibly as much as 7.3 points) over just four months. This chapter analyses a unique four-wave survey panel conducted by The Australian National University, finding that Labor both retained its base and persuaded undecided and Greens voters as well as some Coalition voters. Voters who shifted to Labor were characterised by strong dislike of Peter Dutton and increasingly favourable perceptions of Anthony Albanese. The Coalition lost voters who supported the Voice referendum, as well as those with high educational attainment and who live in areas with high ethnic diversity. Neither gender nor age predicted voters’ propensity to shift to Labor during 2025. The findings emphasise the importance of campaign strategy and party leadership in modern Australian elections.

Keywords: polling; persuasion; vote choice; strategy; campaigning

The 2025 election is historic in the magnitude of Labor’s win and the depth of the Coalition’s loss. But easily forgotten in the magnitude of Labor’s win in May 2025 is just how parlous the party’s position was just five months earlier. My average of public opinion polls had Labor on 48 per cent of the two-party-preferred (2PP) vote on New Year’s Day 2025—its low point for the 2022–25 election cycle and close to Labor’s 2019 election result. By election day on 3 May, that average of public polls had Labor on 52.2 per cent of the 2PP vote, replicating Labor’s 2022 result and corresponding to a 4.2 per cent increase in the 2PP vote share in just four months. Yet, this final poll average was a full 3 percentage points below Labor’s actual 2PP result of 55.3 per cent—easily Labor’s best result in the post-1949 era (see Chapter 17, this volume).

Depending on when and how much the public polls were underestimating Labor’s support, Labor recovered at least 4.3 points of 2PP vote share between January and May and possibly as much as 7.3 points (if early 2025 public polls were unbiased). Either way, this is one of the larger and more rapid movements in Australian voting intentions ever observed.

This chapter reveals the drivers of Labor’s recovery and the Coalition’s decline. I draw on a novel collection of data, a panel of survey data collected by Professor Nicholas Biddle of The Australian National University to explore the dynamics of voting behaviour in 2025. That is, individual respondents were reinterviewed at different stages of the campaign. These data and my analysis help us understand which voters changed their voting intentions between the start of 2025 and the election, and why.

The dataset is unique in Australia. There was no shortage of polling conducted in connection with the election. The flagship Australian Election Survey is a ‘single-shot’ or cross-sectional post-election survey. Commercial pollsters typically utilise cross-sectional designs, with short surveys probing little beyond voting intentions and the most basic demographics, and seldom make their unit-record data available (see Chapter 20, this volume). It is a similar story for the parties’ internal polling which, while more extensive in scope, is not made available for scholarly analysis. Thus, for one of the first times in the study of Australian elections, we can systematically analyse opinion change over a pre-election period and, moreover, one of the more dramatic and rapid changes in voting intentions ahead of an Australian federal election.

The data reveal five key findings. First, Labor pulled votes from all over the political map: from those supporting Greens in January and February, especially from undecideds and from voters who entered 2025 intending to vote for the Coalition (see Table 19.2). The only group with whom Labor did not make significant inroads was those who started 2025 intending to support ‘Other’ parties or candidates, principally Independents or One Nation. Almost one in four Greens supporters later reported voting for Labor and 36 per cent of voters saying they were undecided in early 2025 reported voting Labor in May. Labor also retained more than 80 per cent of its early 2025 support; the Coalition’s retention rate was 76 per cent and the Greens’ only 66 per cent.

Second, the longer the campaign wore on, and the more people saw of the leaders, the more Dutton’s likeability ratings fell and Albanese’s rose (Figures 19.1 and 19.2). And, the more a voter liked Albanese and the less they liked Dutton—or both—the greater was the probability of either a move to or staying with Labor (Figure 19.3). Labor’s research revealed Dutton’s low personal ratings as a key vulnerability for the Coalition. Labor’s campaign was designed accordingly, and—as this analysis indicates—seems to have been effective.

Third, rather than wedging Labor voters, the Voice referendum appears to have wedged voters who as late as early 2025 were intending to vote for the Coalition (Table 19.3). Labor ‘Yes’ and Labor ‘No’ voters were equally likely to stay with Labor between early 2025 and the election. But Coalition ‘Yes’ voters were roughly three times more likely to have moved to Labor by the end of the campaign than Coalition ‘No’ voters.

Fourth, university-educated voters reporting themselves as undecided in early 2025 broke overwhelmingly for Labor: 51 per cent to 21 per cent for the Coalition and 20 per cent for ‘Others’ (Table 19.4). At the other end of the educational attainment continuum, among voters starting 2025 intending to vote Greens and with secondary school education, just 35 per cent remained with the Greens, with 57 per cent reporting a transition to a Labor vote in the May election. Similarly, Labor retained 88 per cent of its early 2025 supporters with secondary school levels of education.

Fifth, undecided voters from the more ethnically and racially diverse areas of the country were much more likely to transition to Labor than undecideds from less diverse areas (Figure 19.4).

Voting intentions

The data span four waves, from late 2024 to the May 2025 post-election wave. Table 19.1 shows estimated voting intentions at each wave. The May wave is when respondents report how they voted in the election; these responses let me weight the entire dataset to closely correspond to the election results (see the two rightmost columns of Table 19.1). Appendix 19.1 contains additional details.

Table 19.1 Voting intentions/recall in the 2025 election campaign, by wave of survey (per cent)

December 2024

January/February 2025

March 2025

May 2025 (post election)

Voting intention

All

Voters

All

Voters

All

Voters

All

Voters

Actual

ALP

26.0

29.6

26.0

29.9

31.3

34.3

29.8

34.6

34.6

Coalition

35.2

40.0

34.3

39.5

33.5

36.8

27.4

31.8

31.8

Greens

13.1

14.9

11.9

13.7

13.6

15.0

10.5

12.2

12.2

Other

13.6

15.5

14.6

16.8

12.7

14.0

18.5

21.4

21.4

DK/NV/INF

12.1

–

13.2

–

8.8

–

13.8

–

–

Effective no.

1,211

1,236

1,080

1,132

1,135

1,115

1,093

1,440

–

Notes: DK = don’t know; NV = didn’t vote; INF = voted informal.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

As shown by public polls fielded at the time, Labor’s estimated first-preference formal vote share in early 2025 was only 29.9 per cent, compared with 39.5 per cent for the Coalition. The Coalition lost almost 8 percentage points of vote share between these estimated January/February levels and the election, ultimately winning 31.8 per cent of formal first preferences. Labor’s vote share improved by more than 5 percentage points from January/February to the election, from 29.9 per cent to 34.6 per cent.

Movers and stayers

The panel data allow us to track individual respondents’ voting intentions over the course of the campaign. Here, I focus on voting intentions recorded in the January/February 2025 wave and the post-election May 2025 wave. Respondents interviewed in both waves supply a basis for estimating rates of ‘moving’ or ‘staying’ with respect to their vote.

We already know that Labor won more votes than it lost between January and May, and that the reverse is true for the Coalition. But Table 19.2 reveals the underlying sources of change and stability, containing estimates of the change in the electorate between early 2025 and the election, computing the distribution of reported voting in May by reported voting intentions in the January/February wave of the survey. Again, the data are weighted to closely correspond to the 2025 election results, bolstering confidence that these estimates validly characterise changes in voting intentions.

Table 19.2 Vote choice reported in May post-election wave by January/February voting intention (per cent)

January/February 2025

May 2025, vote recall

ALP

Coalition

Greens

Other

DK/NV/INF

ALP

81.4

12.1

23.8

8.6

35.6

Coalition

4.6

76.2

1.0

7.4

25.2

Greens

8.8

3.0

65.6

5.2

5.2

Other

5.3

8.7

9.5

78.8

34.0

Notes: DK = don’t know; NV = didn’t vote; INF = voted informal; Rows show post-election vote recollection. Columns show voting intention from January/February. Table excludes May 2025 ‘don’t know’, ‘didn’t vote’ and ‘voted informal’ responses. Cell entries are percentages, summing to 100 within columns, so are interpretable as transition rates or the rate of voting for the party listed in row i of the table given a previously reported intention to vote for the party listed in column j; ‘stay’ rates appear on the diagonal of the table. Effective sample size = 871.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

Voting intentions are more fluid than conventional theories of voting behaviour would suggest: among the five voting categories utilised here, I estimate that 39 per cent of the electorate switched categories between early 2025 and the election. This ‘switching’ rate falls to 23 per cent if we exclude moves to or from the ‘Don’t know/No vote/Informal’ category. Note that I have dropped from Table 19.2 respondents who reported in May not voting or voting informally. The population of non-voters and informal voters is very difficult to survey accurately, especially in this case of a four-wave panel study: conditional on responding to a multi-wave panel study about politics, few respondents report not voting or voting informally. I therefore focus on the four formal voting categories: ALP, Coalition, Greens and Others.

Among voters, Labor retained 81 per cent of its early 2025 supporters, while the Coalition retained only three-quarters. Meanwhile, 12 per cent of Coalition supporters from January/February reported a switch to Labor; just 5 per cent of ALP supporters from early 2025 made the opposite transition.

We also observe considerable churn in the Greens vote, with only two-thirds of January/February Greens supporters reporting in the post-election survey sticking with the Greens. Almost one in four Greens supporters (23.8 per cent) from early 2025 wound up reporting in May that they voted for Labor. Partially offsetting these loses were the 8 per cent of Labor supporters from early 2025 who transitioned to the Greens.

Labor also did well among the disengaged or undecided segment of the electorate, as measured in January and February. Among undecideds, those intending to not vote or vote informally, 36 per cent ended up reporting voting Labor, while just 25 per cent transitioned to Coalition support.

Thus, Labor not only retained more of its vote than any other party grouping but also was the single largest destination for vote switchers. Among those reporting a formal vote in May and switching from their early 2025 voting intention, 38 per cent landed on Labor, 17 per cent with the Coalition and 15 per cent with the Greens. The remaining 30 per cent of these switchers wound up transitioning to ‘Others’, such as Independents (including Teals), One Nation, Trumpet of Patriots, Family First, Legalise Cannabis and so on. In the voter conversion stakes—perhaps the single most important metric of campaign performance—Labor outperformed the Coalition by more than two to one, and the Coalition barely outpaced the Greens.

Diverging views about Albanese and Dutton

At each of the four rounds of interviewing, respondents were asked to rate the two major-party leaders, Peter Dutton and Anthony Albanese, on a zero to 10 likeability scale. Average scores for each leader in each round are shown in Figure 19.1.

Figure 19.1 Anthony Albanese and Peter Dutton average likeability score, by wave (0–10 scale)

Figure 19.1 Anthony Albanese and Peter Dutton average likeability score, by wave (0–10 scale)

Note: Vertical bars cover 95 per cent confidence intervals.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

By late 2024, Dutton had drawn level with Albanese on likeability, with both leaders averaging about 4 on the zero to 10 scale. Albanese had been viewed more favourably than Dutton for most of the 2022–25 election cycle—an advantage that was slowly being muted as the cycle dragged on and cost-of-living pressures dominated the national conversation. Nonetheless, through 2023 and 2024, Labor strategists had repeatedly suggested to me that playing on Dutton’s ‘negatives’—heightening the political salience of Dutton and the contrast with Albanese, reminding voters about Dutton’s past and current policies—was always going to be part of Labor’s re-election campaign.

Labor’s strategy—and its execution—clearly worked. Figure 19.1 shows that, after two years of Albanese’s and Dutton’s ratings slowly converging, these two indicators diverged markedly in the three months before the election. In the post-election wave, the two average scores were separated by a full point: 4.8 for Albanese and 3.6 for Dutton.

Figure 19.2 Change in Anthony Albanese’s and Peter Dutton’s likeability, by paired waves of the panel

Figure 19.2 Change in Anthony Albanese’s and Peter Dutton’s likeability, by paired waves of the panel

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

The panel data let us track the dynamics in evaluations of Albanese and Dutton over the months leading up to the election. Figure 19.2 examines wave-on-wave changes in Albanese’s and Dutton’s likeability, as well as the corresponding change in Dutton’s net likeability (a respondent’s Dutton likeability scores minus their Albanese likeability score). By the time of the May 2025 election, 40 per cent of the electorate found Albanese more likeable than at the start of the year, while 44 per cent found Dutton less likeable. Dutton’s net likeability fell for 56.5 per cent of the electorate between early 2025 and the election, improving among a mere 17 per cent.

These asymmetries in likeability dynamics are especially pronounced in the formal campaign period itself, looking at the March/April to post-election comparisons presented in the third row of Figure 19.2. More than one-third (36 per cent) of respondents reported seeing Albanese as more likeable over this period, while just 20 per cent saw him as less likeable. For Dutton, 42 per cent of respondents reported a fall in their assessments of his likeability and just 18 per cent reported seeing him as more likeable. The result was that more than half the electorate reported Dutton’s net likeability falling over the campaign period and only 19 per cent saw Dutton gaining in likeability vis-a-vis Albanese. Below, the chapter shows that changes in Dutton’s net likeability between early 2025 and the election are strongly associated with changes in voting intentions over the same period.

Voice referendum

In the aftermath of the election, numerous commentators noted that the Coalition’s emphasis on ‘culture war’ issues—Indigenous recognition chief among them—had backfired (Williams et al. 2025). My analysis of these panels reveals the full extent of this ‘backfiring’.

Table 19.3 shows the transition in voting intentions between early 2025 and the election according to how respondents reported voting in the Voice referendum. Recalled Voice vote was asked well before the May post-election wave, so there is no chance that respondents were adjusting their recalled Voice vote to better ‘match’ their self-reported final vote. Labor ‘stay’ rates are statistically indistinguishable across Voice ‘Yes’ and ‘No’ voters. But this is not the case among other segments of the early 2025 electorate.

Table 19.3 Vote choice in May 2025 (rows), by January/February voting intention by Voice vote (per cent)

January/February 2025 by Voice vote

ALP

Coalition

Greens

Other

DK/NV/INF

Post-election vote

No

Yes

No

Yes

No

Yes

No

Yes

No

Yes

ALP

88

81

9

26

26

21

8

11

38

42

Coalition

3

3

83

52

3

0

8

7

42

9

Greens

2

11

0

12

55

71

0

20

0

17

Other

6

5

8

11

16

8

84

61

20

32

Notes: DK = don’t know; NV = didn’t vote; INF = voted informal; Voice vote shown in columns (‘yes’ and ‘no’). Cell entries are percentages, effective sample size = 867. H0 transition probabilities are independent of Voice vote, rejected at p < 0.01. DK/NV/INF on Voice and May 2025 vote choice are excluded.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

One-quarter of voters supporting the Coalition in early 2025 who recalled voting ‘Yes’ for the Voice wound up switching to Labor and only half this segment reported staying with the Coalition. Coalition ‘Yes’ voters were roughly three times more likely to flip to Labor than Coalition ‘No’ voters: 26 per cent to 9 per cent. Reminding voters of the Coalition’s opposition to the Voice—or its claims that Welcome to Country ceremonies were ‘overdone’ and divisive—appears only to have driven votes away from the Coalition. Only one-half of ‘Yes’ voters reporting a Coalition voting intention in early 2025 reported sticking with the Coalition in their votes cast in May.

Among early 2025’s undecided or non-voters, 42 per cent of ‘Yes’ voters wound up voting for Labor and a paltry 9 per cent for the Coalition. The Coalition fared better among undecided ‘No’ voters, picking up 42 per cent of this group, but this transition rate is nonetheless statistically indistinguishable from the 38 per cent of this group who ended up voting for Labor.

Of course, the usual correlation/causation caveats apply—namely, recalled Voice vote might be incorrect, proxying for other policy differences and evaluations of the parties since October 2023. Nonetheless, amid a broad net shift towards Labor between January and May, few variables identify ‘movers’ to Labor as well as one’s recalled position on the Voice. Interaction effects are also apparent, with the Voice and low Dutton net likeability combining to nudge the probability of a Labor vote even higher, when averaged over the electorate.

Education

Educational attainment is a good predictor of the Greens to Labor transition (Table 19.4). More than half those reporting Greens support in early 2025 and secondary school education transitioned to Labor by May; university-educated Greens supporters display much more stability, with just 10 per cent switching to Labor and 79 per cent remaining with the Greens. University education also makes a Coalition to ALP transition more likely, as well as the ‘Others’ to Labor transition.

Table 19.4 Vote choice in May 2025 (rows), by January/February voting intention by educational attainment (per cent)

January/February 2025 voting intention by education

ALP

Coalition

Greens

Other

DK/NV/INF

Post-election vote:

HS

Trade

Uni.

HS

Trade

Uni.

HS

Trade

Uni.

HS

Trade

Uni.

HS

Trade

Uni.

ALP

88

81

77

10

9

15

57

28

10

5

2

15

25

38

51

Coalition

4

0

7

78

78

74

1

0

1

8

15

3

34

14

21

Greens

4

14

10

4

1

2

35

60

79

9

0

5

0

11

8

Other

5

4

6

7

12

9

7

12

10

78

83

77

41

37

20

Notes: DK = don’t know; NV = didn’t vote; INF = voted informal; HS = high school; Cell entries are percentages; effective sample size = 871. H0 transition probabilities are independent of education, rejected at p < 0.01. DK/NV/INF on May 2025 vote excluded.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

A key insight into campaign dynamics comes via the way that educational attainment strongly structures where undecided voters end up. Of those reporting being unsure or planning to not vote or to vote informally in early 2025—and who went on to report a formal vote in the election—half those with tertiary education reported voting for Labor in May and just 21 per cent for the Coalition. The Coalition also trailed Labor in conversions among those with trade or technical qualifications in this segment, 14 per cent to 38 per cent. Only among secondary-school–educated undecideds did the Coalition win more conversions than Labor, with 34 per cent to Labor’s 25 per cent.

This pattern is again consistent with a good portion of early 2025 undecideds comprising many Labor ‘learners’: those who perhaps voted for Labor in 2022 but, for a variety of reasons, reported being undecided at that stage of the electoral cycle. Labor’s resurgence in support over the opening months of 2025 rested not just on flipping voters intending to vote for another party, but also on reactivating soft or dormant preferences for Labor. To the extent that educational attainment is correlated with 1) news consumption and attention to politics and 2) socially progressive policy preferences, Labor’s outperformance among undecideds with tertiary education speaks to the differences in content of the major parties’ campaigns. The statistical modelling that follows helps to disentangle the contributions of two correlated features, ‘Yes’ Voice votes and educational attainment, as sources of Labor gains over the 2025 campaign.

Modelling transitions to or from Labor over 2025

Statistical modelling can rigorously assess the contributions of leader likeability, position on the Voice and other factors to the probability of a Labor vote in May 2025, conditional on each respondent’s voting intention as reported in the January/February 2025 wave.

Specifically, I model the binary outcome yi = 1 if respondent i reported voting Labor in May 2025, and yi = 0 otherwise. The inclusion of indicator variables for each voting intention reported in January/February—ALP, Coalition, Greens, ‘Other’ and ‘Don’t know/Undecided’—as predictors gives the model the interpretation as a transition model. That is, base rates of moving to a Labor vote (or staying with Labor) can be thought of as ‘hardwired’ into the model; these are reported in the top row of Table 19.2. Additional predictors tap into the extent to which ‘move to Labor’ or ‘stay with Labor’ rates vary around these base rates. A wide variety of model specifications was tested via cross-validation, guarding against over-fitting in this modestly sized sample; details appear in Appendix 19.1.

My model for the probability of a Labor vote in May 2025 includes the following predictors:

  • Indicator variables for each voting intention reported in January/February 2025: ALP, Coalition, Greens, ‘Other’ and ‘Don’t know/Undecided’, with the ALP category absorbed into the model intercept.
  • Change in Dutton’s net likeability between January/February 2025 and the post-election wave of the panel: a continuous variable ranging from –19 to 10, with 90 per cent of values between –6 and two.
  • Voice vote: indicator variables for ‘Yes’, ‘No’ and ‘Don’t know’/‘Didn’t vote’, with the last category absorbed into the model intercept.
  • A three-level measure of highest level of educational attainment (as used in Table 19.4): high school (including less than high school), trade and technical education (Certificate III and IV) and tertiary education.
  • A scale measure tapping the ethnic and racial diversity of postcodes, described in Appendix 19.1.

My preferred model is a random forest (Breiman 2001), of a collection of decision trees using these predictors in various combinations. These models do not produce parameter estimates and are inherently nonlinear. Accordingly, I compute predictions from the model and use tables or graphs to show the modelled effects of key predictors. Note, too, that characteristics such as age and gender that often appear in models of voting choice do not appear in this model of transitions from January/February voting intentions to post-election recalled vote; while age and gender do often predict voter intentions in the aggregate (for instance, the Coalition’s repeated poor performance with women and young voters), I find no evidence that age or gender significantly shaped changes in voting intentions between early January and May 2025, which is the focus of this chapter.

Change in leader assessments

The growing gap in likeability assessments of the two leaders has a large effect on the probability of a move to or staying with Labor, as shown Figure 19.3. Each panel corresponds to a January/February 2025 voting intention. Changes in Dutton’s net likeability are plotted against the horizontal axis, restricted to 90 per cent most commonly observed actual changes, ranging from –6 to 2 across the sample (and with very little variation in these 90 per cent bounds across the January/February voting intention segments); the grey vertical bars in each panel in Figure 19.3 show the distribution of changes in Dutton’s net likeability, by January/February voting intention.

Figure 19.3 Modelled effect of change in Peter Dutton’s net likeability between January/February 2025 and the post-election wave on the probability of an ALP vote, by January/February voting intention (each panel) and Voice vote (the orange and black lines, per panel)

Figure 19.3 Modelled effect of change in Peter Dutton’s net likeability between January/February 2025 and the post-election wave on the probability of an ALP vote, by January/February voting intention (each panel) and Voice vote (the orange and black lines, per panel)

Notes: Bars show the distribution of changes in Dutton’s net likeability, by January/February reported vote. Most respondents reported a decline in Dutton’s net likeability over the period, associated with a higher probability of either staying with or transitioning to a Labor vote over the campaign. Note, too, the significant interaction with the Voice vote for respondents who in January/February 2025 were intending to vote for the Coalition. Details on model fitting and construction of marginal effect profiles appear in Appendix 19.1.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

Clearly, where one stood in January/February 2025 is highly predictive of the vote one reported casting in the May election; this explains the large overall differences across the panels in Figure 19.3. Per Table 19.2, Labor retained more of its vote than the Coalition over this interval, and a large portion of Coalition transitions to Labor is concentrated among Voice ‘Yes’ voters.

But for all January/February voting intention segments, staying with or transitioning to Labor is strongly related to changes in the relative assessments of the two leaders. Recall that 44 per cent of voters liked Dutton less by May 2025 than they had in January/February—and 39 per cent liked Albanese more—resulting in 57 per cent of the sample scoring in negative territory on the change in Dutton’s net likeability measure (Figure 19.2). As Dutton’s net likeability falls, the probability of a transition to Labor (or remaining with Labor) rises; conversely, for respondents in January/February intending to vote for the Coalition, the probability of staying with the Coalition rises (the probability of a transition to Labor falls) as evaluations of Dutton vis-a-vis Albanese become more favourable. This is especially true for voters reporting as undecided in January/February, when there is an abrupt increase in the probability of a transition to Labor as changes in Dutton’s net likeability fall from no change to negative.

Ethnic and racial diversity

The ethnic or racial diversity of a respondent’s postcode also predicts the probability of a Labor vote in May 2025, as shown in Figure 19.4. For voters reporting an intention to vote ALP or Greens in early 2025, racial and ethnic diversity has modest predictive power with respect to staying with or moving to Labor by May. Labor’s ‘stay’ rate rises over the least diverse 30 per cent of postcodes, but plateaus and may even taper slightly in the country’s most diverse postcodes. Greens-to-Labor transitions are slightly more likely in the country’s most diverse postcodes, reaching as high as 25 per cent for ‘Yes’ voters in the top decile of postcodes on this scale measure. As most Australians live in postcodes with no concentrations of people not speaking English at home, Muslim religious affiliation and other markers of ethnic and racial diversity (details on scale construction appear in Appendix 19.1), I plot the modelled effects as a function of the deciles of the scale.

Figure 19.4 Modelled effects of racial and ethnic diversity change on the probability of ALP vote, by January/February voting intention (each panel) and Voice vote (the orange and black lines, per panel)

Figure 19.4 Modelled effects of racial and ethnic diversity change on the probability of ALP vote, by January/February voting intention (each panel) and Voice vote (the orange and black lines, per panel)

Notes: There is a significant interaction with Voice vote for respondents who in January/February 2025 were intending to vote for the Coalition. Details on the construction of the racial and ethnic diversity postcode level index and the effects plotted here are in Appendix 19.1.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

The most pronounced and interesting effects of racial diversity are among voters who entered 2025 either intending to vote for the Coalition or undecided. Coalition-to-Labor transitions were much more likely in racially and ethnically diverse postcodes, reaching as high as 35 per cent for ‘Yes’ voters in the top decile of postcodes on this measure and 20 per cent among ‘No’ voters. Labor’s conversion of undecided voters increases almost linearly across the postcode deciles, ranging from a low of 21 per cent in the least diverse decile of postcodes to above 50 per cent for undecided ‘Yes’ voters in the 30 per cent most ethnically diverse postcodes, and in the mid-40 per cent range for ‘No’ voters.

Conclusion: Campaigns or context?

This analysis supports a popular narrative about the 2025 election: 1) that Labor was able to reframe the contest as between Albanese and Dutton and as less of a retrospectively oriented referendum on Labor’s handling of cost-of-living pressures; and 2) given this framing, Albanese’s greater popularity and likeability relative to Dutton would reconstitute the winning majority Labor enjoyed in 2022, shoring up Labor’s core vote and persuading undecided voters to ‘come back’ or transition to Labor.

Of course, two external developments are also widely believed to have lent ‘ecological’ or ‘contextual’ validity to Labor’s reframing strategy. The first was the decision by the Reserve Bank of Australia (RBA) to cut interest rates on 18 February, helping Labor advance the argument that the worst of the cost-of-living crisis was over. The second was the torrent of news from the second Trump administration, including deep and abrupt cuts to the US civil service, the threat of tariffs on Australian exports and complaints about mainstays of Australian public policy such as the Pharmaceutical Benefits Scheme. Dutton and the Coalition seemed remarkably ill prepared and flat-footed in responding to both Labor’s reframing strategy and the extent to which Trump’s dramatic opening months in office galvanised the Australian electorate, radically accelerating the reframing of the election already under way in early 2025. Trump’s persona and his policies—and Dutton’s inability to hose down dalliances with MAGA policies and rhetoric, by himself or Coalition colleagues—put Labor’s reframing strategy on skates. In a remarkably brief window, the election was transformed: no longer retrospection about Labor’s handling of the cost-of-living crisis, but rather a contemporaneous or prospective framing, grounded in assessments about which leader would be least like Trump and would be least likely to import Trump-like policies.

I offer these observations more as commentary than as conclusions supported by the data: the survey data I analyse in this chapter do not contain questions permitting me to rigorously test the causal connections between respondents’ awareness of or reactions to these external events, changes in Dutton’s likeability, changes in issue salience and changes in voting intentions.

The evidence I present here is, however, nonetheless consistent with Labor’s reframing strategy succeeding, whatever role should be attributed to factors such as the RBA rate cut and Trump. We track a broad recovery in Albanese’s likeability over Dutton’s across the five months leading up to the May election. These changes are strongly predictive of either staying with or transitioning to Labor.

It appears that Labor’s campaign helped not only shore up the support it took into the calendar year 2025, but also recapture some of Labor’s 2022 vote that was sitting with the Coalition, the Greens or in the undecided column in early 2025. A long vein of political science stresses that campaigns first reactivate dormant partisan dispositions before they persuade or convert (Campbell et al. 1960; Bartels 2000; Green et al. 2002). The evidence presented here is consistent with that view. The campaign generated outsized transitions to Labor among university-educated undecided voters, secondary-school–educated Greens supporters, those who voted ‘Yes’ for the Voice and those in Australia’s more racially and ethnically diverse postcodes. Labor’s campaign strategy—and the Coalition’s underperformance—saw the party not only reassemble its winning 2022 electoral base, but also consolidate its support in sociodemographic segments already tending to vote Labor.

Longer-term extrapolations from analysis of a campaign are always fraught. My analysis does reveal the capacity for volatility in the electorate over short time frames. Some of this is to be expected in a regime of compulsory voting, where even the least politically engaged are legally obliged to turn out, and most do. Some of Labor’s gains and the Coalition’s losses may therefore be short-lived. This caveat aside, the more consequential finding is the consolidation of Labor support over the past two electoral cycles—and, by corollary, the extent of the erosion of Coalition support.

Labor won not just the campaign—and a massive electoral victory—but also a once-in-a-generation opportunity to govern with broad demographic support and, for now, deep reservoirs of political capital.

References

Bartels, Larry M. 2000. ‘Partisanship and voting behavior, 1952–1996.’ American Journal of Political Science 44, no. 1: 35–50. doi.org/10.2307/2669291.

Biddle, Nicholas. 2025. Hope, Hardship, and Democratic Confidence: Social Wellbeing and Political Sentiment in Election-Year Australia—March/April 2025. 17 April. Canberra: School of Politics and International Relations, The Australian National University. politicsir.cass.anu.edu.au/files/docs/2025/4/Hope-Hardship-and-Democratic-Confidence.pdf.

Breiman, Leo. 2001. ‘Random forests.’ Machine Learning 45, no. 1: 5–32. doi.org/10.1023/A:1010933404324.

Campbell, Angus, Philip E. Converse, Warren E. Miller, and Donald E. Stokes. 1960. The American Voter. Chicago: University of Chicago Press.

Green, Donald P., Bradley Palmquist, and Eric Schickler. 2002. Partisan Hearts and Minds: Political Parties and the Social Identities of Voters. New Haven: Yale University Press.

Hastie, Trevor, Robert Tibshirani, and J.H. Friedman. 2009. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd edn. New York: Springer doi.org/10.1007/978-0-387-84858-7.

Stone, M. 1974. ‘Cross-validatory choice and assessment of statistical predictions.’ Journal of the Royal Statistical Society: Series B 36, no 2: 111–33. doi.org/10.1111/j.2517-6161.1974.tb00994.x.

Williams, Carly, James Vyver, Stephanie Boltje, and Brooke Fryer. 2025. ‘“All the wrong lessons” from Voice referendum “backfired” on Coalition, analysts say.’ ABC News, 6 May. www.abc.net.au/news/2025-05-06/first-nations-people-respond-to-election-result/105260222.

Appendix 19.1

Methodological notes

Data

The data come from a four-wave survey panel study commissioned and authored by Professor Nicholas Biddle of The Australian National University. The data were provided by an online panel provider and the surveys were self-administered online.

The bulk of the analysis presented here focuses on transitions from the second wave of the study, fielded in January/February 2025, to the final, post-election wave administered immediately after the election in early May 2025. The data are weighted to be representative of the Australian electorate, by raking on recalled 2025 House of Representatives vote, age, education and geography.

Since the dataset for analysis spans respondents completing both the January/February wave and the post-election wave, panel attrition produces a slightly different distribution on January/February voting intentions than observed in the original, standalone January/February wave. A final round of weighting ensures that the distribution of January/February voting intentions in the two-wave analysis dataset matches the distribution of voting intentions observed in the January/February wave. The resulting weights were trimmed to lie between 8 and 1/8 and to sum to the number of respondents completing both the January/February and the post-election waves (n = 1,832). This weighting is nonetheless reasonably aggressive, yielding weights with variance of 1.63, resulting in an effective sample size of approximately 1,832/(1 + 1.63) = 696. Paradoxically, some analyses reported in the body of the chapter have larger effective sample sizes, when filtering out respondents with especially large or variable weights.

Postcode-level racial and ethnic diversity

The index of postcode-level racial and ethnic diversity is constructed from 2021 census data, spanning 2,641 postcodes. A broad range of postcode-level census aggregates was assembled, spanning indicators of wealth and income, educational attainment, religious affiliation, housing tenure and type, occupational structure, age structure, language use and ancestry, tabulated with respect to adult citizens by usual place of residence or at the dwelling level.

The effects of these variables were assessed via a wide set of modelling approaches, including tree-based models, a variety of neural architectures and scale construction via factor analysis prior to modelling.

Table A19.1 displays loadings from a factor-analytic model fitting three orthogonal factors via maximum likelihood estimation with a varimax rotation, as implemented in the factanal function in R. The input to the factor analysis is the weighted correlation matrix of the postcode-level indicators, with weights equal to the number of enrolled voters in each postcode (constructed from AEC redistribution files and SA1 to polling place mappings from past elections and normalised to match division-level counts of enrolled voters).

As a model for these indicators, the three-factor solution is probably not sufficient. But cross-validation of the downstream transition models suggests that the second factor produced from this three-factor solution is preferred over other factor-analytic solutions. The point here is less to perform a rigorous factor analysis of postcode-level census data and more to produce a scale measure that has validated predictive power in the models of transitions in voting intentions over the 2025 campaign.

The second factor, labelled ‘Factor 2’ in Table A19.1, taps racial and ethnic diversity, with high loadings from indicators such as percentage of adult citizens speaking non-English languages at home, percentage identifying as Muslim and percentage speaking Chinese languages at home. Note also the large negative loading for the percentage of adult citizens reporting no religious affiliation and positive loadings on dwelling-level measures such as persons per household or persons per bedroom.

Table A19.2 lists the highest-scoring postcodes on this measure, 10 postcodes with scores close to the median score of 0.4 and the 10 lowest-scoring postcodes. Western Sydney dominates the upper end of this diversity index. The pronounced right skew in the constituent indicators passed through to the index itself: although normalised to have a mean of zero by construction, the index has a median score of –0.4, a minimum score of –0.9, but a maximum score of 5.1.

Model fitting

The model relied on in the body of the chapter is a random forest, an ensemble (‘forest’) of decision trees, each tree formed from a random sample of observations and a random selection of predictor variables when forming branches for individual trees. Splits (for categorical predictors) or cut-points (for continuous predictors) are chosen to best discriminate with respect to the outcome being modelled (in this case, a binary outcome, yi = 1 if the respondent, i, reported voting Labor in the May post-election survey and 0 otherwise). The model’s predictions are formed by averaging the predictions from individual trees.

This class of mode—and the configuration of the random forest—was selected from a wide set of candidate models by cross-validation (Stone 1974). A textbook treatment appears in Hastie et al. (2009: Ch. 7). In this case, the data are repeatedly repackaged into five, 80–20 splits of training data and test data, but are stratified to guarantee virtually identical distributions of y between test and training slices. Over the five ‘folds’ of the dataset, each observation takes a turn as an out-of-sample or ‘test’ observation. Model performance is assessed with respect to the average fit across the five batches of out-of-sample fitting—in this case, using a log-loss criterion (equivalent to maximising the log-likelihood of the out-of-sample observations, averaged over the five folds). This procedure to model selection guards against over-fitting and bolsters confidence in the generalisability of the model’s predictions.

Table A19.1 Factor loadings from a three-factor solution to a broad array of postcode-level census indicators, with varimax rotation

Variable

Factor 1

Factor 2

Factor 3

Median total household income ($)

0.934

0.085

–0.097

Professional occupation (%)

0.828

0.045

0.526

Median monthly mortgage repayment ($)

0.827

0.222

0.091

Highest educational attainment: year 9 or below (%)

–0.822

0.165

–0.126

Highest educational attainment: year 10 or above (%)

–0.820

0.051

–0.440

Tertiary qualification (%)

0.818

0.159

0.529

Median weekly rent ($)

0.779

0.190

0.045

Voice, ‘Yes’ vote (%)

0.624

0.228

0.568

Socioeconomic advantage and disadvantage decile 8 (%)

0.449

0.010

0.108

Indigenous (%)

–0.300

–0.037

–0.111

Socioeconomic advantage and disadvantage decile 7 (%)

0.254

0.027

–0.092

Jewish religion (%)

0.215

–0.005

0.176

Language other than English spoken at home (%)

0.096

0.992

0.026

Muslim religion (%)

–0.087

0.735

–0.081

No religion (%)

0.110

–0.669

0.358

Average persons per bedroom

0.206

0.626

0.298

Chinese languages spoken at home (e.g., Mandarin, Cantonese) (%)

0.258

0.587

0.232

Median age (years)

–0.264

–0.438

0.074

Aged under 30 (%)

0.022

0.233

0.183

Average household size (persons)

0.242

0.530

–0.742

Dwellings that are flats, units or apartments (%)

0.302

0.184

0.706

Renting (%)

–0.033

0.287

0.613

Member of a same-sex couple (%)

0.231

–0.080

0.598

Non-traditional Christian religion (%)

–0.153

0.280

–0.397

Lived elsewhere in Australia five years before 2021 Census (%)

0.139

–0.221

0.290

Socioeconomic advantage and disadvantage decile 6 (%)

0.014

–0.057

–0.219

Aged over 70 (%)

0.083

–0.035

0.140

Notes: Loadings greater than 0.5 in absolute value are highlighted in bold. The second factor taps racial and ethnic diversity and is the postcode-level measure used in the transition models in the body of the chapter.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).

Table A19.2 Ten highest-scoring postcodes on racial and ethnic diversity

Postcode

Locality

Division

Diversity score

Non-English language at home (%)

Muslim (%)

No religion (%)

Chinese languages (%)

Persons/bedroom

2144

Auburn

Blaxland (NSW)

5.1

83.4

44.7

18.3

20.8

1.2

2166

Cabramatta

Fowler (NSW)

5.1

84.4

2.0

21.5

17.5

1.1

2195

Lakemba

Watson (NSW)

5.0

80.9

47.9

11.3

6.8

1.3

6799

Home Island

Lingiari (NT)

4.8

77.8

78.4

16.1

0.0

1.1

2165

Fairfield

McMahon (NSW)

4.7

78.8

8.0

14.4

8.1

1.1

2200

Bankstown

Watson (NSW)

4.7

78.9

27.3

14.3

6.6

1.1

0872

Yuendumu

Lingiari (NT)

4.7

77.8

0.1

24.6

0.1

1.4

4876

Bamaga

Leichhardt (Qld)

4.7

78.2

0.3

21.0

0.0

1.3

2177

Bonnyrigg

Werriwa (NSW)

4.6

78.3

5.0

15.0

7.6

1.0

3061

Campbellfield

Scullin (Vic.)

4.6

75.7

35.8

9.9

0.5

1.0

2372

Tenterfield

New England (NSW)

–0.4

3.4

0.0

30.8

0.1

0.7

2480

Goonellabah

Page (NSW)

–0.4

4.4

0.2

44.2

0.2

0.8

4340

Rosewood

Blair (Qld)

–0.4

3.6

0.0

37.1

0.1

0.8

6488

Trayning

Durack (WA)

–0.4

2.4

0.0

47.5

0.0

0.6

2898

Lord Howe Island

Sydney (NSW)

–0.4

4.5

0.0

48.4

1.0

0.9

7264

Gladstone

Bass (Tas.)

–0.4

2.7

0.0

49.6

0.0

0.8

6603

Wongan Hills

Durack (WA)

–0.4

3.5

0.5

42.4

0.0

0.7

2430

Taree

Lyne (NSW)

–0.4

3.6

0.1

35.5

0.2

0.7

6510

Moora

Durack (WA)

–0.4

3.5

0.0

39.3

0.5

0.7

2444

Port Macquarie

Cowper (NSW)

–0.4

4.4

0.2

36.6

0.3

0.7

7175

Bream Creek

Lyons (Tas.)

–0.8

0.0

0.0

57.7

0.0

0.8

3704

Koetong

Indi (Vic.)

–0.8

0.0

0.0

66.7

0.0

0.7

6613

Buntine

Durack (WA)

–0.8

0.0

0.0

61.1

0.0

0.7

3893

Tambo Crossing

Gippsland (Vic.)

–0.8

0.0

0.0

21.7

0.0

0.8

6463

Benjaberring

Durack (WA)

–0.8

0.0

0.0

54.5

0.0

0.8

5138

Basket Range

Mayo (SA)

–0.8

0.0

0.0

66.0

0.0

0.9

3785

Tremont

Casey (Vic.)

–0.8

0.0

0.0

88.6

0.0

0.8

7024

Cremorne

Franklin (Tas.)

–0.8

0.0

0.0

61.7

0.0

0.8

3852

East Sale

Gippsland (Vic.)

–0.9

2.1

0.0

72.0

0.0

0.7

5150

Leawood Gardens

Boothby (SA)

–0.9

0.0

0.0

48.6

0.0

0.7

Note: Includes 10 closest to the median score and 10 lowest-scoring postcodes.

Source: ANU 2025 Election Monitoring Survey Series, Wave 1–4 (Biddle 2025).


Previous Next

Updated:  11 August 2026/ Responsible Officer:  / Page Contact: