The pollster YouGov has begun releasing survey results generated by the relatively novel technique known as multilevel regression with post stratification, or MRP. Although it demands a level of statistical sophistication beyond the reach of most pollsters, MRP seems destined to become an increasingly important part of the pre-election landscape.
YouGov first used MRP in Britain in 2017. It has since used the technique in Spain (2023), Britain (2024) and Germany (2025). Introduced to Australia by YouGov in 2022, the team that did it then — including Shaun Ratcliff (now at Accent) and Campbell White (now at Pyxis) — is not the same as the team (led by Paul Smith) that’s doing it now.
Polls that use MRP stand out from standard polls in three ways. First, they are based on large numbers — not the usual thousand or two. YouGov’s January–February poll had 40,689 respondents. But in arriving at “the best and most robust sample possible… to project the election” via modelling that explored “over 25,000 different possibilities,” only 8732 responses survived.
Second, MRP incorporates aggregate-level data (for the YouGov poll, electorate-level data covering age, gender, education, past voting behaviour, country of birth, and location) derived from the Census or the Australian Bureau of Statistics. Exactly which depends on the variables pollsters believe (rightly or wrongly) help shape the vote and the variables for which population-wide data are available; these data are not gathered from respondents because the number of respondents in each electorate is too small and what respondents report may be wrong or incomplete.
The third characteristic of an MRP poll is that it shows voting intentions (for YouGov, two-party preferred) seat-by-seat. Using these figures, pollsters can categorise the seats in any number of ways: “safe Coalition,” “likely Coalition,” “lean Coalition,” “tossup Coalition,” “safe independent,” “safe Green,” “safe other,” “tossup Labor,” “lean Labor,” “likely Labor” and “safe Labor” are YouGov’s chosen categories. The number in each category can be added up. Calculating the seat-shares for the Coalition, Labor and others can lead to discussions about possible outcomes — from majority governments to minority governments to hung parliaments in which almost anything might happen. It is the focus on seat-shares rather than vote-shares — the focus of other polls — that is most eye-catching about MRP.
MRP can be used to predict the likely outcome in seats in different parts of the country, in the “key battle-ground constituencies,” where boundaries were changed by the 2024 electoral redistribution, where the sitting member is retiring, and so on. These, however, are not things YouGov chose to pursue, at least not in its initial report of its January–February poll; a webinar is planned for 18 March.
CLASSIFYING SEATS AND PREDICTING OUTCOMES
The election, we now know, will be held in May. What do the seat counts in the YouGov poll show about the probable outcome had the election been held in February? To come to grips with YouGov’s figures we need to understand how it classifies seats — and how this differs from the way psephologists are used to seeing seats classified.
Seats classified as “safe” by the Australian Electoral Commission are those that would change hands on a swing of more than ten percentage points, those classified as “fairly safe” would require a swing of six to ten points, while “marginals” would be lost or gained on a swing of swing of less than six points.
Not so with YouGov. Most of the seats YouGov classifies as “safe” — on YouGov’s figures — would change hands on (further) swings of less than ten percentage points. No fewer than thirty-eight of the sixty seats classified by YouGov as “safe” for the Coalition require smaller shifts — typically, much smaller. For Labor, as many as forty-four of its fifty-one “safe” seats also require shifts of less than ten points. Of the seats it classifies as “safe” for the Coalition, YouGov has the Coalition ahead by just four points, 52-48 in two, Robertson (NSW) and Bass (Tasmania). Both seats, however, are regarded by the AEC as “marginal.” In three of the seats YouGov describes as “safe” for Labor — Hasluck (WA), classified as “safe” by the AEC; Corangamite (Victoria), “fairly safe”; and Hindmarsh (SA), “fairly safe” — Labor also leads by just four points.
Psephologist Malcolm Mackerras first used the two-party swing required for seats to change hands to rank every seat as either “safe,” “fairly safe” or “marginal” (concepts and intervals subsequently adopted by the AEC) in 1972. Every seat that required a swing of ten percentage points or more was regarded as “safe,” seats that required swings of 6.0 to 9.9 points were “fairly safe,” and seats that required smaller swings were “marginal.”
He noted, however, that in some cases categorisation could be misleading. The safety or vulnerability of seats might vary because of the strength or weakness of local members, the presence of a particular industry or of local “discontent,” demographic changes since the last redistribution, and so on. It followed that some “marginal” seats should be considered “safe,” and some “safe” seats were really “marginal.”
What YouGov understands by its own categories — “safe,” “likely,” “lean” or “tossups” — is nowhere made clear. But it appears to have collapsed into one the two things that Mackerras always sought to distinguish — the strictly numerical gap, on one hand, and the factors that might lead to a more nuanced appraisal, on the other — with the numerical gap not YouGov’s main consideration. Hence, its willingness to describe as “safe,” according to its polling, five seats where the gap between the two parties (or candidates) is as low as four percentage points (52–48); to describe not as “lean” to Coalition/Labor but as “likely” Coalition/Labor, seven seats where the gap between the two sides is just two percentage points (51–49); to describe as “lean” rather than “likely,” eight other Coalition/Labor seats where it shows a 51-49 split; to describe as a “tossup” the seat of Ryan where the ALP has a 52-48 lead; and so on.
In arriving at the number of seats the poll has the two sides winning, YouGov assumes that any seat that is “safe Coalition” (sixty), “likely Coalition” (five), “lean Coalition” (five) or even a “tossup Coalition” (three), is a seat the Coalition will win; and that any seat that is “safe Labor” (fifty-one), “likely Labor” (ten), “lean Labor” (four), or a “tossup Labor” (one), Labor will win. By this arithmetic, Coalition has seventy-three; Labor, sixty-six.
Why is “the most likely result” in the three Labor seats rated 50–50 — Macquarie (on a 6.3-point margin, according to the AES); Shortland (6.0); and McEwen (3.8) — “Coalition tossups”? YouGov’s decision to call them for the Coalition matters. Were these seats to have been ignored — regarded as simply too close to call — the count would not have been seventy-three to sixty-six in favour of the Coalition but seventy to sixty-six, making the story’s headline problematic. If the “Coalition tossups” were regarded as “Labor tossups” then the seat count would slip back to seventy to sixty-nine, taking the headline with it.
As well as estimating the most likely distribution of seats, YouGov provides estimates (based on a 90 per cent probability) of the lower and upper limits of “the “true” party shares (and seat totals).” For the Coalition, the range is sixty-five to eighty seats; for Labor, fifty-nine to seventy-two. Lopping off seats from one side categorised as “lean to” or “tossups” and moving them to the other side proves to be a close proxy: sixty-five to seventy-eight for the Coalition; sixty to seventy-three for Labor. Alternatively, had YouGov considered every seat where the polling was at 50–50, 51–49 or 49–51 as the bottom of the range, then the lower limit for the Coalition would have been sixty-four (seventy-three minus nine) and the upper limit, eighty-two (seventy-three plus nine); for Labor fifty-seven (sixty-six minus nine) and seventy-five (sixty-six plus nine) respectively. These three forms of calculation, only one of which is YouGov’s, yield similar results.
Based on what YouGov thinks the “most likely” results, the Coalition picks up seats held by independents who had formerly been Coalition MPs (Calare, NSW; Monash, Victoria; Moore, WA) but not a single seat held by other independents. In the “non-classic contests” — contests where at least one of the last two candidates fighting it out is likely to be neither Coalition nor Labor — Labor is the only party predicted to make inroads — not only against the Greens in Brisbane, Griffith and Ryan but also against the (non-teal) independent Dai Le in Fowler (NSW).
YouGov cautions that in these seats Labor is only “slightly favoured.” Nonetheless, it has Labor ahead by twelve points in Fowler, eight points in Brisbane, six points in Griffith, and four points in Ryan. If having a lead of four to twelve points means a party or candidate is only “slightly favoured,” then the same might be said of no fewer than seventy-seven seats, or almost half the seats.
THE TRACK RECORD
How good is YouGov’s record? According to Patrick English, its UK-based director of political analytics, “YouGov’s approach to MRP modelling is market leading and has a proven track record of success including in recent British and Spanish national elections.”
Two weeks later, he would no doubt have added the 2025 German election, where YouGov’s final MRP model (in its own words) “called 91 per cent of constituencies [630 seats, under list voting] correctly.” Providing a figure for the number of constituencies (meaning “seats”) it had called “correctly” — the number in which it had picked the winner even if not the winning margin — has not been something YouGov has always done.
On most occasions it has only disclosed the difference between its estimate of the number of seats each party would win and the number of seats each party went on to win (a net figure). Always, or almost always, the number of seats that the pollster gets wrong based on a seat-by-seat comparison will exceed the net figure based on a comparison of the party-by-party totals. (This is true, even though YouGov’s method of calculating the net figure involves an element of double counting: underestimating one party’s vote share must involve overestimating the vote share of another),
In Germany, YouGov’s net error was 6.7 per cent but the seat-by-seat error was said to be 9 per cent. Whether the net error matters more than the seat-by-seat error depends on whether one is interested in a pollster’s ability to estimate total seat-shares or a pollster’s ability to pick the seat-by-seat results. (These are the two ways electoral pendulums can be used as well: to predict the net loss of seats, given the overall swing; or to see the seats that are most likely to swing, given the overall swing.)
In the 2024 British general election, YouGov’s prediction, based on 42,758 interviews in 622 seats (across England, Scotland and Wales, but not Northern Ireland), was that “Labour is set to win, and set to win big.” Labour did win and “win big”; every poll had expected that. But YouGov, in common with all the other polls, overestimated Labour’s lead: Labour finished ten points ahead of the Conservatives — less than the seventeen-point lead YouGov had predicted. It wasn’t the only poll to use MRP at that election. Of the others that did so, six had a bigger Labour lead — Survation and Focaldata (nineteen points), We Think (twenty), and Find Out Now and Electoral Calculus (twenty-three) — but two (More in Common and JLPartners) had a smaller Labour lead (sixteen points). On this measure, YouGov enjoyed more success than some but less success than others.
What about its seat success ? YouGov puts that at 92 per cent. This is based on adding up the number by which it overestimated the seats Labour, the SNP and the Greens would win and the number by which it underestimated the seats the Conservatives and Reform would win — fifty-two seats, hence the 92 per cent. In 2017, by the same reasoning, YouGov’s success rate was 93 per cent.
Again, the seat-by-seat errors were not released. No doubt some — perhaps most — will have been among the eighty-nine YouGov characterised as “toss ups” or “marginal” — meaning, here, that the party ahead in the polling had a lead of less than five percentage points.
In the 2023 Spanish general election, where YouGov became “the first pollster” in Spain to use MRP, YouGov appears to have done less well. Six days out from the election (in Spain, the publication of polls any closer to the election is prohibited), YouGov correctly predicted a hung parliament with the People’s Party holding a narrow lead. It also predicted the totals for each party — out, overall, by twenty-eight seats; in a parliament of 350 seats, this meant a net error of 8 per cent.
And in Australia — a precursor to which English did not refer? In 2022, when YouGov published its MRP with 18,923 respondents, the Australian, which had commissioned the poll, rolled out the headline “Labor to win modest majority with eighty seats.” This turned out to be too high: needing seventy-six seats, Labor won seventy-seven, giving it a very modest majority. The Coalition, “most likely” to win sixty-three, won fifty-eight; the Greens, four instead of one; and Other, twelve instead of seven. YouGov failed to pick the winner in sixteen seats (net), an error of 10.6 per cent.
Seat by seat, YouGov says it “called the winner of 92 per cent of electorates.” Unusually, this appears to be an actual seat count, not a net seat count. In a House of 151 seats, it called 11 seats incorrectly. This figure takes into account the six seats that in its print edition the Australian had labelled “too close to call” (seats it found were 50–50, two-party preferred) but which in its online view YouGov had said were “likely” to go to the Liberals in some cases or to Labor in others; two of these were called incorrectly.
A success rate of 91, 92 or 93 per cent — whether based on a party-by-party count or a seat-by-seat count — may sound better than a failure rate of 7, 8 or 9 per cent, even when the underlying reality is the same. But because most seats at an election are safe, their fate predictable, the success rate for YouGov polls may not be quite as good as it sounds, the failure rate worse than it sounds. If there are 150 seats, but only a third, for arguments’ sake, are at any risk of changing hands, then the failure to predict the outcome in eleven seats (as happened in 2022) might be better expressed either by saying that the poll had a 78 (rather than 93) per cent success rate (thirty-nine out of fifty) or a 22 (rather than 7) per cent failure rate (eleven out of fifty). Either way, these scores look very different from those based on the (tacit) assumption that outcomes in every seat are equally difficult to predict. •
