Likely voter models and projecting turnout

Returning to the Ipsos approach to likely voters, we have set up a method that allows us fine grained control over our model to match the actual turnout rates (here, here and here). Of course, the perceptive polling connoisseur would ask, “great you can match to turnout, how do you know what turnout is going to be?”  That’s a great questions and what I’ll try to address in this post.

Typically opinion researchers will look at past turnout statistics, if they use turnout at all, to model future elections. For example, to model the 2016 presidential election, we’ll look back at what happened in the 2012 and 2008 presidential elections. The idea is that past elections are our best indicators of future elections. The problem is that the past is not always a good predictor of future. The 2014 midterm election had substantially lower turnout rates (~35%) than previous midterms (~45%). This led to a number of pollsters missing the mark in 2014 because they did not have an accurate perception of the electorate for that election.

Beyond past examples, pollsters also have to rely on anecdotal reporting or voter registration statistics to make any adjustments to turnout estimates. My question is “can we have a data-driven indicator of the probable Election Day turnout?” The Reuters/Ipsos poll provides a great research platform to test that idea out. As I’ve previously explained, the Ipsos approach to likely voters scores each and every respondent using a 5 item battery of questions from which we calculate the index score. This approach combines interest in voting with past voting behavior to create a composite view of the electorate. The past voting behavior question is relatively static from election to election but what if the interest in voting question could tell us something.

In the graph below I’ve plotted the average interest in the election from all respondents from each state in the two months preceding the 2014 midterm elections against the actual statewide turnout. The light gray line is the “fit line” showing the correlation relationship between interest in the election and turnout.

lv turnout 1 052615

This indicates that in the 2014 election, there was about a 35% relationship between our interest in the election score and the actual turnout on a state-by-state level. So while there are a lot of other factors at play, a higher score in our interest in the election variable is correlated with a higher turnout.

Now what if we also bring data from the 2012 election into this discussion? The graph below plots the actual state-level turnout in 2012 by our interest in voting data from September and October 2012.

lv turnout 2 052615This chart is plotted using the same scale as the 2014 data. Just looking at it, it looks clear that there is a relationship between turnout and our question AND the levels are lower than in the 2012 presidential election. That is what we’d expect to see since the actual turnout in 2012 was about 20 percentage points higher than seen in 2014. Just for kicks, let’s combine both years into the same plot and see what happens.

lv turnout 3 052615Nice, combining both 2012 and 2014 into the same chart really makes this relationship pop. In fact, over the two elections there is a R2 of .74 – that means that about ¾ of the variation in turnout is correlated with our interest in voting measure.

If this model holds, we can use the regression equation produced by the combined output to predict turnout for upcoming elections based on our current likely voter questions. Moving forward we are going to share these estimates with the public on a regular basis in order to help contribute to the accuracy of polling efforts in the future. Our first 2016 election turnout projection is here (link).