The Slow Death of 2pp

For decades, Australian political watchers have leaned on one figure above all others. Not the primary vote share. Not preferred prime minister. The two-party-preferred number, the 2pp, the share each major bloc ends up with once preferences are distributed. It is two things at once: the headline number pollsters publish, and the figure the electoral commission reports on the night. In most elections it has been the number that decides who governs. That dual role is why pundits reach for it (with exceptions in the top left and bottom right quadrants of the next chart).

That number is now less reliable, and not in the way most people assume. The usual complaint is that a fragmenting electorate makes the 2pp fuzzy. The deeper problem is worse. The 2pp is no longer just imprecise. However a pollster builds it, it may be systematically wrong, and the pollster the aggregators treat as the outlier may be the one closest to the truth.

This post works through why, using a Bayesian aggregation of the current cycle.

Why 2pp existed in the first place

Australia uses preferential voting. You rank candidates, and if your first choice is eliminated your vote flows to your next preference. In the past, almost every seat eventually narrows to Labor against the Coalition, so analysts collapse the whole field into that single contest. The collapsed figure is the 2pp.

It worked well for a long time because the primary vote was stable and concentrated. When the two majors took around 80 percent of first preferences between them, the 2pp was a clean summary of a contest that genuinely was two-sided.

That world is gone.

The major-party primary vote has cratered

This trend away from the two major party groups is not new. We can see it dating back some time. The next two charts show the primary vote share for other parties (including the Greens and One Nation) increasing from around 10% in the 1960s to 33.6% at the last election. 

This has seen a rise in crossbench representation in the House of Representatives. That matters for the 2pp in a way separate from how it is measured in opinion polling. The 2pp is a Labor-versus-Coalition number. It says nothing about the crossbench seats, and as those seats multiply, government formation turns on them more and more. So the 2pp is being squeezed from both ends. Its use in polling, the preference assumption, is increasingly shaky (more on this below), and its election output, a two-horse seat race, is increasingly not the race that decides who governs. 

The big news since the last election is the rise of One Nation in the primary voting polls.

Its rise has been at the expense of the Coalition primarily, but also Labor and the other parties. The two major party groups are now polling around 52 per cent of the primary vote (compared with a touch over 66 per cent at the last election). 

The Greens primary vote remains largely unchanged from the last election.

Two ways to guess, and they no longer agree

To turn primary votes into a 2pp, a pollster must assume how preferences will flow. There are two main approaches.

The first is respondent allocation, or RA. You ask voters where they would direct their preference and use their answers. The second is previous-election flows. You assume preferences split the way they did last time and apply those historical ratios.

These are not footnotes. They can produce materially different headline numbers from identical primary votes, and pollsters do not agree on which to use. Essential uses respondent allocation with undecided voters left in. YouGov uses respondent allocation. Resolve publishes RA numbers. Newspoll used previous-election flows before it stopped publishing a 2pp at all in 2026. RedBridge uses last-election flows. Roy Morgan now publishes both side by side because it will not adjudicate the question for its own readers.

When the people producing the number cannot agree on how to produce it, the consumer is in difficulty. But the harder issue is which method is now actually more correct.

The One Nation problem

Here is the crux. The previous-election-flow method assumes 2026 One Nation voters will preference the way 2025 One Nation voters did. But the 2026 One Nation voter is demographically a different animal from the 2025 one. The party has roughly quadrupled its primary vote share, and the new bloc is drawn disproportionately, though not entirely, from former Coalition voters. At the 2025 federal election, One Nation preferences flowed 25.5% to Labor and 74.5% to the Coalition. Whether that ratio still holds is the open question. It was estimated when One Nation polled a quarter of its current size, on a different slice of voters. Stretching it over a bloc four times larger and drawn from a different pool is extrapolation beyond the data it was fitted on, not a mere worry that preferences drift. The ex-Coalition majority would push the real flow further toward the Coalition than the old ratio assumes; the sizeable ex-Labor minority pulls the other way.

Apply the old ratio to this new, much larger, and compositionally different One Nation bloc and you may over-allocate its preferences to Labor, or under-allocate them. The point is not that we know which. The point is that the 2025 ratio is no longer a safe assumption, and the previous-election-flow 2pp now rests on it anyway. RedBridge has warned the traditional 2pp had become "less useful" given the primary spread. Respondent allocation, which asks today's One Nation voters directly rather than assuming they resemble last cycle's, may be the more defensible method in exactly this situation.

That reframes the whole debate. The methodological choice is not a neutral matter of taste. With a quarter of the electorate sitting with One Nation, the previous-election-flow method carries a directional bias whose size, and even whose sign, is now genuinely uncertain.

The Essential divergence, and what it might mean

Now to the pollster the model flags as an outlier. Track Essential's Labor 2pp against the mean of every other pollster and a gap opens and then widens. In late 2025 Essential ran two to three points below the field. By February 2026 it was five points adrift, with Essential at 49.5 against a field average of 54.5. By April the gap was 5.4 points. The aggregation model reads this as a stable house effect of roughly minus three points on Essential's Labor 2pp, with the spread between Essential at the bottom and Newspoll at the top reaching nearly five points. 

The obvious story is that Essential's respondent-allocation method makes it lowball Labor. That story may be wrong, and the way you can tell is Roy Morgan. Roy Morgan publishes both methods from the same survey. Its RA figure has Labor higher than its previous-election-flow figure, not lower. So respondent allocation on its own may not depress Labor.

What is specific to Essential is a combination. Respondent allocation, which is roughly neutral or pro-Labor in most hands, sits alongside an undecided treatment that deflates the primaries, and recalled-vote weighting that pulls those primaries back toward the last election result. That third ingredient is the quiet one. Weighting respondents by how they recall voting last time would tend to drag Essential's primary estimates toward the previous outcome, which in 2026 would mean underweighting the Coalition-to-One-Nation shift. The combination, not the preference method alone, perhaps produces the five-point gap.

Put the two halves of the argument together and something uncomfortable falls out. The previous-election-flow method that most of the field relies on may be mis-stating Labor because of the One Nation composition shift. Essential, through a different route, lands lower. The pollster the aggregator treats as biased may be closer to the underlying truth than the consensus it deviates from. A house effect, in the model's language, is not the same thing as an error.

Pollsters are quietly walking away

The most honest signal is behavioural. Across this cycle, fourteen poll releases published no 2pp at all. Newspoll dropped the 2pp line from every one of its 2026 releases. DemosAU routinely skips it. Some YouGov releases omit it. MRP studies, which model individual seats, do not produce a 2pp by design. When pollsters stop publishing their own former headline metric, and one of them says out loud that it has become less useful, that is worth more than any methodology note.

What this means for reading polls

None of this makes polling worthless. The aggregate still tells a coherent story. Labor's 2pp against the Coalition has fallen from around 56.5 at the start of the cycle to about 53.5 now, and the model expresses honest uncertainty around that path rather than false precision. The trend is real.

The lesson is sharper than "polls are broken." It is that a single 2pp from a single pollster is now a weak and possibly biased signal, because so much of it is assumption rather than measurement, and the dominant assumption is now contestable. Three habits follow.

Read the primary vote first. It is what voters actually said before any modelling, and right now it is the more trustworthy number. Check the preference method, because an RA poll and a previous-election-flow poll are not comparable even when they look identical, and in this cycle the difference could be large and its direction is contestable. And treat an outlier pollster as a hypothesis, not an error. When the consensus shares a questionable assumption, the dissenter deserves a hearing rather than an automatic correction.

There is a parallel lesson for election nights. The 2pp has always been able to diverge from the seat result. Look at the off-diagonal cases in the chart at the top: vote clumping, safe-seat margins, and uneven swing have long let one party win the two-party count and still lose the contest for seats. That is the old problem, and it lived entirely inside a two-party world. The new problem is different in kind. As the crossbench grows, the more likely failure is that neither major party reaches a majority at all, and government is decided by negotiation with members the 2pp never counted. The old divergence was noise around a two-party signal. The new one is the two-party signal ceasing to be the thing that determines power. The 2pp will keep being quoted on the night. It will just keep explaining less of what actually happened.

The 2pp had a good run. It summarised a two-party country for as long as the country stayed two-party. That country is going, and the number is straining to keep up on both counts: harder to measure, and less decisive when measured. The honest reading spends less time on a single decimal and more time on the primary vote, the method behind the headline, and the seats that the headline never counted.