How Weighting by Past Vote Can Improve Estimates of Voting Intentions

Polling error for the 2020 US election was the highest in 40 years and no mode of surveying was unambiguously more accurate. This occurred amid several recent polling failures in other countries. Online panels, as the dominant method now used by pollsters to survey voters, are well-positioned to help reduce the level of bias in pre-election polls. Here, we present a case for those pollsters using online panels for pre-election polling to (re)consider using past vote choice (i.e., whom respondents voted for in the previous election) as a weighting variable capable of reducing bias in their election forecasts under the right circumstances. Our data are from an Australian pre-election poll, conducted on a probability-based online panel one month prior to the 2019 Australian federal election. Three different measures of recalled vote choice for the 2016 election were used in weighting the forecast of the 2019 election outcome. These were (1) a short-term measure of recall for the 2016 vote choice obtained three months after the 2016 election, (2) a long-term measure obtained from the same panelists three years after the 2016 election and (3) a hybrid measure with a random half of panelists allocated their short-term past vote measure for 2016 and the remainder their long-term measure. We then examined the impacts on the bias and variance of the resulting estimates of the 2019 voting intentions. Using the short-term measure of the 2016 recalled vote choice in our weighting significantly reduced the bias of the resulting 2019 voting intentions forecast, with an acceptable impact on variance, and produced less biased estimates than when using either of the other two past vote measures. The short-term recall measure also generally resulted in better estimates than a weighting approach that did not include any past vote adjustment. Implications for panel providers are discussed.