Like most forecasting models, Election Lab uses the past to predict the future. To predict House and Senate elections in 2014, we draw on the elections from 1980-2012. We first look at how well key factors were related to outcomes in those past elections. Then, we gather information about those same factors for 2014. Assuming that these factors will be related to election outcomes in 2014 in the same way they were from 1980 to 2012, we can make a prediction about who will win each race.
The model’s factors fall into three categories:
1) The national landscape. On average, the better things are going in the country, the better the president’s party will do in an election. We capture national conditions using two measures: presidential approval and change in gross domestic product. At the same time, the president’s party usually does worse in midterm years than presidential years even after accounting for the first two measures, so our model takes account of that, too.
2) The partisanship of the state or district. Obviously, House and Senate candidates will do better when their party dominates a district or state. We measure this with Obama’s share of the major-party vote in 2012. In Senate races, we also include the incumbent’s share of the vote from the election six years before, combined with an indicator for whether that incumbent is running or the seat is open. The incumbent’s previous election matters mainly when the incumbent is running again.
3) Key features of the race. The model currently takes account of whether the incumbent is running, which captures the well-known incumbency advantage in congressional elections. For the Senate, we also build in each candidate’s level of experience in elective office. In the Senate, we categorize experience into five levels, from someone who has never held an elective office to an incumbent senator. For states where there hasn’t yet been a Senate primary, we impute candidate experience using historical averages from similar races. (After the primaries, we will also add candidate experience to the House model. There, the measure will be simpler: whether the candidate has held any elective office.)
The difference reflects the much larger range of experience typically found in Senate races. We do not consider appointed senators, such as South Carolina’s Tim Scott, to be true incumbents. Our analysis suggests that appointed senators gain much less of an advantage than elected incumbents. Finally, the model also takes into account fundraising data. For each race, we currently measure fundraising as the sum of money raised by all candidates in each party. After the primaries, we will substitute fundraising by the actual general election candidates.
For House races, our forecast currently relies on this model.
For Senate races, the forecast reflects a combination of the model and, where available, and average of current polls. (We discussed combining the model and polls here, here, and here.) For Senate races with little or no polling, the forecast relies on the model. For races with more polling, the model will gradually give more weight to the polls as the campaign goes on and the number of polls increases, since the predictive power of polls increases with sample size and proximity to Election Day.
Although Election Lab lists candidates in upcoming primary elections as well as the known general-election candidates, our model focuses only on forecasting the November general election. In Senate races where we do not yet know a party’s nominee and thus that person’s prior political experience, we assume that the eventual nominee’s prior experience will be similar to the historical average from similar races.
All forecasting models come with some uncertainty, since we don’t have complete information about every race or even complete certainty about how the information we do have has explained election results in the past. To address this uncertainty we conduct many simulations from our model, each time predicting the winners. Each round of simulations produces a set of outcomes for the individual races along with an overall outcome.
By analyzing these simulations, we can answer questions like, “What is the percent chance that the Democratic candidate wins in such-and-such district?” or “What is the percent chance that Republicans take the Senate?” The goal is always to make the best statement we can about both what we know and what we don’t know.
Who are we?
Ben Highton is professor of political science at the University of California, Davis.
Eric McGhee is a political scientist who studies elections, election reform, and representation.
GWU Ph.D. student Will Cubbison and GWU undergraduates Philip Stein and Lily Fitzgerald assisted with data collection.
This post was updated Aug. 6 to reflect that the model is now incorporating polling data.