Post developed by Katie Brown and Josh Pasek.
With each election cycle, the news media publicize day-to-day opinion polls, hoping to scoop election results. But surveys like these are blunt instruments. Or so says Center for Political Studies (CPS) Faculty Associate and Communication Studies Assistant Professor Josh Pasek.
Pasek pinpoints three main issues with current measures of vote choice. First, they do not account for day-to-day changes. Second, they capture the present moment as opposed to election day. Finally, they can be misleading due to sampling error or question wording.
Given these problems, Pasek searched for the most accurate way to combine surveys in order to predict elections. The results will be published in a forthcoming paper in Public Opinion Quarterly. Here, we highlight his main findings. Pasek breaks down three main strategies for pooling surveys: aggregation, prediction, and hybrid models.
Aggregation – what news companies call the “poll of polls” – combines the results of many polls. In this approach, there is choice in which surveys to include and how to combine results. While aggregating creates more stable results by spreading across surveys, an aggregation is a much better measure of what is happening at the moment than what will happen on election day.
Prediction takes the results of previous elections, current polls, and other variables to extrapolate to election day. The upside of prediction is its focus on election day as opposed to the present and the ability to incorporate information beyond polls. But, because the models are designed to test political theories, they typically use only a few variables. This means that their predictive power may be limited and depends on the availability of good data from past elections.
Hybrid approaches utilize some combination of polls, historical performance, betting markets, and expert ratings to build complex models of elections. Nate Silver’s FiveThirtyEight – which won accolades for accurately predicting the 2012 election – takes a hybrid approach. Because these approaches pull from so many sources of information, they tend to be more accurate. Yet the models are quite complex, making them difficult for most readers to understand.
So which pooling approaches should you look at? That depends on what you want to know. Pasek concludes, “If you want a picture of what’s happening, look at an aggregation; if you want to know what’s going to happen on election day, your best bet is a hybrid model; and if you want to know how well we understand elections, compare the prediction models with the actual results.”