How to Predict an Election
Every four years since 2008 Nate Silver has created a model using statistics to attempt to forecast the results of the presidential election. Many political viewers, including myself, find themselves anxiously refreshing FiveThirtyEight constantly to get the most up to date results. While Nate Silver is probably the best known, other websites have built their own probabilistic forecasts. G. Elliot Morris built a model for The Economist, and Jack Kersting built a model on his own website JHKForecasts. As of August 25th, all three models think Vice President Biden is favored, but they disagree on the certainty of that outcome. FiveThirtyEight gives Joe Biden about a seven in ten chance of winning. The Economist is more confident in the Vice President, giving him almost a nine in ten chance. JHKForecasts splits the difference, giving Biden just over an eighty percent chance of victory. I have been interested in how these probabilities are calculated for a long time. Luckily all three of these forecasters have published methodologies.
At their core, all three models work roughly the same way. They start by combining the candidate’s standing in the polls with fundamentals, things like the state of the economy and the partisan lean of each state. When averaging the polls, FiveThirtyEight weights based on their own pollster ratings so that a high-quality poll with a strong track record has more weight in their model than a lower quality one. FiveThirtyEight also makes adjustments based on house effects. If a poll consistently gives one candidate an advantage that is not captured in other polls than FiveThirtyEight takes that into account. JHKForecasts also uses FiveThirtyEight’s pollster ratings to weight polls. FiveThirtyEight’s polling average is also built to be more conservative around major events. If a debate moves polls in the short term, but they quickly revert back then it would not have a huge effect on the FiveThirtyEight forecast.
FiveThirtyEight calculates partisan lean using the presidential vote in the last two cycles, applying more weight to the more recent election. Nate Silver also applies a home-state advantage for the presidential and vice-presidential candidates. An adjustment is made for a states ‘elasticity’ which represents the fact that some states are more likely to move significantly in one direction or the other. FiveThirtyEight also makes an adjustment for how easy it is to vote in a state. JHKForecast’s partisan lean index is far more simple. It is based solely on the two most recent presidential elections and the polling in that state.
All three models take into account the state of the economy. The better the economy, the more likely it is for the incumbent to hold on. FiveThirtyEight creates an economic index using jobs, spending, income, manufacturing, inflation, and the stock market. The economist’s economic index is similar, but it also takes into account quarter 2 GDP. This is part of the reason why The Economist’s model is so much more certain. The Economist is more pessimistic on the economy than FiveThirtyEight which thinks the economy will recover somewhat before election day. JHKForecasts uses the economic index from the picture below, although he says it is largely negligible after July as polls on the general election come in.
After balancing polls and fundamentals all three models simulate the election many times over. FiveThirtyEight runs 40,000 simulations every time the model runs. Each of these simulations accounts for uncertainty and the possibility of a polling error. The Economist and JHKForecast models run 20,000 simulations each, also accounting for the possibilities of polling errors.
At a high level, this is how modelers attempt to predict election outcomes. They combine polling data and fundementals, then simulate the election thousands of times. At each step of the process many decisions have to made on which variables to include and how much weight they should be given, leading to the large discrepancy in results from these three models.
Sources:
