Successes and lessons from the YouGov MRP
Campbell White and Shaun Ratcliff
In the leadup to the 2022 Australian federal election, most commentators were expecting that no party would win a majority in the House of Representatives, and that a minority government was the likely outcome. The consensus appeared to be that with the total major party share projected to be a record low, it was expected the expanded cross bench would be kingmakers, deciding who formed government.
YouGov's MRP, released just under two weeks before the election in partnership with The Australian, predicted that this was very unlikely to be the case. The model estimated that Labor would win enough seats to govern in its own right. This is exactly what happened.
Although Labor’s majority is small, even if it had fallen a few seats short of a majority, the Coalition was so far off the pace that minor parties would have had no option but to back Labor or to remain neutral. The MRP was the method that demonstrated this reality even though the voting intention figures were more modest for Labor than most polls were at the time.
The approach we used to produce these results, MRP (short for multi-level regression with post-stratification), is a statistical technique that combines survey data with electorate-level information from the Census and other government agencies - such as population density and the proportion of the population with a university degree - and previous election results.
The MRP estimated that Labor was on track to win between 76 and 85 seats in the House of Representatives, and that the Coalition would win between 58 and 68 seats. In both cases the actual results were within these ranges, although admittedly at the lower end (Labor won 77 and the Coalition 58).
We said beforehand that our model was not a crystal ball. We also acknowledged there were uncertainties with our estimates, that there would be electorates that behaved in unexpected ways, and there would likely be late shifts in vote intention, particularly in some heavily contested seats.
Despite this, and the data being collected several weeks before Election Day (from 14 April to 7 May), our estimates were more accurate at the national level than any of the final surveys published by other pollsters. The MRP had a national 2PP average figure of 52, which was nearly identical to the final figure of 52.1 and better than any Australian poll by any pollster.
More importantly, we also accurately predicted the overall results: substantial Coalition losses in the House of Representatives and a Labor majority, along with an increased crossbench.
Overall, the MRP called the winner of 92% of electorates correctly.
Our estimates picked up Labor strengths in the inner and middle suburbs of the larger metropolitan areas, including the seats of Reid and Bennelong in Sydney, Higgins and Chisholm in Melbourne, Boothby in Adelaide and Swan in Perth. This included seats that Labor has seldom or never won in the past and clearly showed their pathway to a parliamentary majority.
We also picked up the decline of Labor’s primary vote in certain western suburbs of Sydney and Melbourne, and northern Tasmania – but correctly indicated that this would not result in any gains for the Coalition at Labor’s expense.
Of course, we did not get every seat right. Only one seat that our model estimated would be a Labor win off the Coalition was retained by the Liberal Party: Bass. We missed a couple of Labor gains in Perth, and Labor’s loss of Fowler. While the model picked the Greens primary vote surge in inner-city Brisbane, we thought they would fall short rather than narrowly getting across the line in several seats. In each case, the Greens won despite having just a third of the primary vote.
Our largest miss though was understating the teal independents, especially in affluent parts of Sydney. We estimated they would retain their existing seats everywhere and pick up Kooyong and Goldstein and be competitive in two more seats. In the end the picked up six seats.
MRP relies on voting trends and demographic information in addition to polling data, and it is challenging to estimate something that is completely new. Now we have a lot more information about the sort of voter who selects a teal candidate, and we would expect the model to perform better in such seats in future elections.
Despite not being perfect, we are pleased with these results. This was the first publicly released MRP measuring vote intention during an election campaign in Australia and also the first effort by any organisation to produce and publicly release estimates for all 151 House of Representatives electorates just weeks before Election Day. This result indicates the quality of YouGov’s data and the success of our analysis techniques.
Rather than prove the futility of estimating results for each electorate, even the misses highlight the importance of the endeavour. There was a lot of discussion before the election that the election was close but no coherent picture of what the parliament would look like. The YouGov MRP provided the granular information needed to understand what was happening on the ground during the campaign. The swings in this election were not uniform, with important regional patterns shaping the result. Methods like MRP are vital to understand public opinion and voter behaviour at the local level.
Our goal now is to examine how we can improve on this effort and better estimate idiosyncratic results, and continue to innovate on ways to measure and understand Australian public opinion. YouGov is committed to rolling out our MRP methodology globally, and the Australia 2022 iteration is the next step on a longer journey.