Electoral Map 2016: The Structural Model
PoliticalEdu is developing three electoral models for 2016. The first, the structural model, takes data from the 2000, 2004, 2008, and 2012 presidential elections to run a linear regression that determines the relationship between a handful of variables, including state demographics, and number of Democratic public officials, and the Democratic vote share. It is developed by averaging two approaches: one which ignores candidate favorability and a second which includes in the regression the difference between Hillary Clinton and Donald Trump’s net favorability (the results from those models can be found here and here. Clearly, Trump’s historically low favorability ratings could potentially cost him the election).
The structural model assesses the underlying electoral landscape separate from campaign actions. By accounting for factors such as the state partisan voter index (developed by the Cook Political Report), the percent of House seats occupied by a Democrat, and region, we can understand how states are inclined to vote without campaign activities or candidate quirks. Of course, considering Clinton and Trump have high unfavorable ratings, a pure structural analysis will likely miss the mark (hence averaging it with a structural model that includes favorability). We have also developed a state battleground model to analyze poll results.
The structural model serves as a baseline. We can expect these, or similar, results if the campaign ended today. Between now and November 8, one variable will be adjusted: the difference between Clinton and Trump’s net favorabilities. Numbers are from Gallup.
Overall, the model explains around 94 percent of the vote share variation during the four elections.
Predicting third party candidate vote shares is difficult because they fared poorly in previous elections, but polls indicate 2016 will be different. Regression models won’t work. Instead, using a Libertarian and Green Voter Index, vote shares for Gary Johnson and Jill Stein can be modeled. The voter indices approximate each state’s inclination to vote for a Libertarian/Green Party candidate by taking state results from the past four elections and dividing them by the LP/GP national result. This index can then be multiplied by Johnson and Stein’s national polling average to estimate their vote share in any given state.
An example should clarify the method (the following numbers are all made up): Say in Alabama the Libertarian candidate received 0.5% in 2000, 0.25% in 2004, 1% in 2008, and 2% in 2012. Nationally, that candidate earned 1% in 2000, .50% in 2004, 1.5% in 2008, and 3% in 2012. The index for each year is 0.5, 0.5, .67, and .67. Averaging the four, Alabama would have a Libertarian Vote Index value of 0.59. To estimate Gary Johnson’s 2016 vote share in Alabama, I multiple 0.59 by his national polling average (which I have weighted to account for pollster accuracy and date).
With a Libertarian and Green Party candidate included, Clinton and Trump vote shares need to be adjusted. To determine how much to subtract from each, I find the difference in polling averages between the weighted Clinton vs. Trump average and the weighted Clinton vs. Trump vs. Johnson vs. Stein polling averages. From there, I divide the difference between each candidate’s polling average by the total number of percentage points lost between Clinton and Trump. Their initial vote share estimates are then subtracted from the difference quotient multiplied by expected Johnson and Stein vote shares.
These values will obviously change as Johnson and Stein’s poll numbers fluctuate and the difference between the two polling averages changes. As such, this model will be updated weekly (assuming new polls are released during the week).
Including Johnson, this is the electoral map 2016:
The map belies the closeness of many states. Here is a table of states that could very easily change the election.
|Clinton +5-7.5||Clinton +2.5-4.9||Clinton +0-2.4|
|Trump +0-2.4||Trump +2.5-4.9||Trump +5-7.5|
This post will be updated!