Election Forecasting (The Fundamentals)

POLS 3220: How to Predict the Future

Previously…

Today’s Agenda

  • Is there other information besides polls that we can use to make accurate election forecasts?

  • We’ll build a rudimentary statistical model based on political science fundamentals: quantitative indicators that we expect to be strongly predictive of election results.

  • Along the way, I’ll show you the basics of using linear models for prediction problems.

  • We’ll describe how these models work, and discuss three potential dangers caused by using them carelessly.

Motivating Example

  • Suppose it is June in a presidential election year, and you want to predict which candidate will win.

  • Can you make a reasonably accurate prediction about November’s election based only on information you have in June?

  • Let’s start with an “Outside View”. Are there variables that, historically, have been predictive of presidential election results five months in advance?

  • Discuss: what information would you want to know before making your prediction?

Presidential Election Forecast

Presidential Election Forecast

Presidential Election Forecast

  • The equation describing that line is 3.27 + 0.37 \(\times \text{June Approval}\)

  • For every 1-unit increase in presidential approval, we would predict an additional 0.37 in vote margin for the incumbent’s party.

  • But where did those numbers come from? What is the “line of best fit”?

  • The line of best fit is the line that minimizes average squared error.

    • Minimizing average squared error. Does that remind you of anything?

Presidential Election Forecast

Presidential Election Forecast

Presidential Election Forecast

  • We can make predictions using this model by plugging in values for future elections.

  • In June 2016, the incumbent president’s net approval rating was +5%.

  • So the model would predict 3.27 + 0.37 \(\times 5 = 5.12\%\) vote margin for the incumbent party.

Presidential Election Forecast

Presidential Election Forecast

  • The average forecast error is 4.7 percentage points.

  • And we have some particularly large errors in 1972 (11.9 points) and 1984 (8.5 points).

  • Maybe we’d do better if we added more predictor variables to the model?

Presidential Election Forecast

Presidential Election Forecast

  • Combining the presidential approval and economic growth into the same linear model yields this equation:

    • 0.33 + 0.3 \(\times \text{June Approval}\) + 0.94 \(\times \text{Q2 GDP Growth}\)

  • In 2016, GDP growth was a sluggish 1.2% in the second quarter.

  • So the revised model would predict an incumbent margin of 2.96 percentage points. Much better!

The “Plane of Best Fit”

Presidential Election Forecast

  • The kind of forecast model we’ve been building here is called a fundamentals model.

  • Notice that it doesn’t use polling at all!

  • It makes predictions purely based on political “fundamentals”, historical patterns observed in the data (Abramowitz 2021).

Three Dangers

  • Simple linear models can be a surprisingly useful tool for making predictions.

  • But when using a linear model, keep in mind three dangers that could ruin your forecasts:

    • Nonlinearity

    • Extrapolation

    • Structural instability

Danger 1: Nonlinearity

The linear model faithfully gives you the line of best fit…

…even when a straight line is a terrible model!

Danger 2: Extrapolation

  • Be cautious of making predictions with a linear model if the current situation lies far outside the historical data.

  • Plugging these numbers into our model would predict the incumbent losing by 35.2 percentage points! (Actual margin was -4.5 points).

Danger 3: Structural Instability

  • Predicting with a linear model assumes that the relationships observed in your data will be stable over time.

  • In other words, you’re assuming that the future will follow the same rules as the past.

  • Maybe as voters have become more polarized, the relationship between economic growth and voting behavior has gotten weaker?

    • If so, our fundamentals model will get worse and worse at predicting as time goes on.

Looking Forward

  • Next class, we’ll try fitting this type of model on our Senate election dataset, and see how well it performs.

  • And in our final topic of the module, we’ll combine polls and fundamentals into a single forecasting model.

    • Very foxy move. Don’t miss it!

References

Abramowitz, Alan I. 2021. “It’s the Pandemic, Stupid! A Simplified Model for Forecasting the 2020 Presidential Election.” PS: Political Science & Politics 54 (1): 52–54. https://doi.org/10.1017/S1049096520001389.