We open this evening’s salon, “Predicting the Presidential Election”, with group introductions as we have a nice collection of new members, the usual core, and members returning from previous years. This meeting represents our fourth conversation on these predictive models in political science. Will these models be able to predict this election in this tumultuous new normal? Let’s talk about that.
David Staley then introduces James (Jim) Bach, a proficient predictor of elections, who successfully predicted the popular vote outcome of the last three elections. Models often use public opinion polls, which are themselves built on a complexity of other factors, and therefor are harder to manage/balance/equate. Here, he describes different assumptions and proposes other models for how voters will vote.
He presents three (original) models built on one of the following variables: 1. Unemployment rate in the third quarter of the election year. 2. Change in GPD over the second and third quarters of the election year. 3. Change in RDPI per capita. X. To each variable the number of the years the party has been in power of the White House. Interestingly, the RDPI appears to depict the incumbent party vote share the best, with a ~0.7 R-squared value.
Historically, these models fit closely with the actual results, where model 1 (unemployment) often is close to predicting the incumbent party vote percentage of the popular vote. Model 2 appears to do even better, where variance is further reduced. Model 3 fits even better. Notable exceptions exist, where political events and social conditions pull votes away from the incumbent in ways the model does not predict for.
2020… is a problem. But for these models, it caused: 1. Pandemic-induced depression. 2. Stimulus injection into the RDPI. The recession hurt GDP but the stimulus increased the RDPI, impacting all of the models, where their inherent dependent variables become hyper skewed.
Model 1 then predicts a 51.3% incumbent vote. Model 2 predicts a 25.6% incumbent vote, the largest landslide in history. Model 3, the most accurate one so far, the incumbent will receive 81.7%. While all of these models were reasonably predictive before, he expects the true vote to be around 47% for the incumbent. So what happened? The changes in the US political landscape have made these models less useful. The popular vote and the Electoral College outcome are different. Voters are possibly less likely to vote due to economic interests. Democrats and Republicans live in increasingly different fact worlds, where they do longer match.
He analyzed the same question in the ANES 2016 and the 2018 Pilot Studies (https://electionstudies.org/data-center/anes-2016-pilot-study/ and https://electionstudies.org/data-center/2018-pilot-study/), thinking about economic changes in the country. An immediate trend is present, where Democrats reflect better on the economy when their president is in the White House and the same is true for Republicans. A question is raised on how we define those voting groups where multiple questions are used to define the strength of those voting blocks from weak, medium, or strong Republicans/Democrats.
If Democrats and Republicans no longer see reality the same, how can we use public attitudes to predict behavior? More importantly, how does one govern with two realities in place? If the suburbs are going Democratic, then what is the core of the Republican party? These complexities are very difficult to model and show that this new normal we live in are highly unpredictable.
An additional model is proposed, http://primarymodel.com/, where the turnout at these events also can be a strong predictors. However, caution is raised that, like public opinion polling, it is another endogenous variable, a dependent variable measuring another dependent variable.
Are there enough undecided swing voters to make a difference? Jim does not believe so. In the crazy last two weeks in particular, no change really is seen from the incumbent. In this case, things appear to be rather diametrically locked, where the incumbent and the contest are not changing and might even be fortifying in their positions. As noted, mass mail-in voting has already shown to be challenging and this factor also is not interior of these models.
Moderately popular incumbent presidents in moderately prosperous times are often likely to win re-election historically. But often, if the scenarios are worse in the country, the incumbent is removed. Will the protests and pandemic change this historical outcome?
Jim describes the theoretical interest on building a model which is able to predict an outcome based on a factor which is not immediately related. Using public opinion polls to predict popular vote election outcome is like measuring someones legs to determine their height, it is measuring the same thing. But to take factors which are exterior of those complex measures and predict the election outcome provides a more theoretically interesting conclusion.
In summary, Jim says that 2020 broke his models. Are we in a period of relative chaos in this arena of political science? It seems so. We may now see a greater weight on qualitative studies over these harder quantitative methods. Now we get to see how this upcoming election supports the opportunity for new models to better predict the political shifting we clearly see.
We thank Jim for his awesome presentation and conversations and we look forward to talking with everyone again next month!