Election forecasting is fraught with peril. However, if instead of being like political pundits, we provide our model of election outcomes and the data, it is possible to assess the statistical accuracy of our model and move beyond simple atheoretical correlations.
We should strive to have a forecast with a long lead time. If an announcer predicts the winner of a football game with less than a minute to go, we care very little about the forecast’s accuracy. If a forecast is made before the game begins and is just as accurate, we would be quite interested. Actors who are dependent on the outcome of the election can modify their strategy and behavior if the forecast is provided well before the election.Footnote 1
INFLUENCES ON ELECTION OUTCOMES
Because there are many variables available, it would be impossible to include them all: we have to trim their number down. We can have some sense of election outcomes by looking at the state of the economy. The voting behavior literature is rich with economic models of voting behavior. Because there are so few cases, and because the retrospective and prospective items are so highly correlated (.86), I follow Fiorina (Reference Fiorina1981), Lewis-Beck (Reference Lewis-Beck1988), Lewis-Beck and Whitten (Reference Lewis-Beck and Whitten2013), Lockerbie (Reference Lockerbie2008; Reference Lockerbie2016), and Nadeau, Lewis-Beck, and Bélanger (Reference Nadeau, Lewis-Beck and Bélanger2013), whose forecasts focus on economic expectations. As with the earlier models, I make use of the item from table 8 of the Survey of Consumer Attitudes and Behavior. The score is the average of the responses in the negative for the second quarter of the election year to this question: “Now looking ahead—do you think you (and your family living there) will be better off or worse off financially a year from now, or about the same?”Footnote 2 Given that this item lacks attribution to the parties, it should only serve to understate the relationship between economics and the election outcome, because one could believe that one’s financial situation will change regardless of the election. Moreover, one might think that the opposition party will win, which would improve one’s economic condition. This model would take expectation to indicate support for the incumbent party. Again, this should serve to understate the relationship between prospective economic evaluations and electoral outcomes.Footnote 3
There are also some noneconomic patterns related to election outcomes. Abramowitz (Reference Abramowitz, Campbell and Garand2000) has noted the desire of the public to change parties in the White House after one party has controlled it for two terms.Footnote 4 An incumbent party returning for a third term without the same candidate running for reelection is a rarity in American politics. To account for the diminished support for the incumbent party, I make use of the log of time a party has controlled the presidency.
PRESIDENTIAL ELECTION RESULTS
Table 1 shows the presidential election model.Footnote 5 The results of this iteration are similar to those of earlier years. The more pessimistic people are, the more likely the incumbent party is defeated. The incumbent party loses slightly more than one-half percentage point of the vote for each percentage point increase in pessimism. As expected, there is a negative relationship between the length of time a party has controlled the White House and its share of the vote.
Table 1 Forecasting Equations
Notes: The first number in a column is the regression coefficient. The second number shown in the parentheses is the significance level.
To assess the utility of this model, I use out-of-sample equations.Footnote 6 The average absolute error is 2.70 percentage points. The two independent variables are significant at the .01 level or better in every equation. The equation has successfully predicted the outcome in every election, except in 1960 and 1968; Nixon was forecast to win in 1960 and lose in 1968. In 2000 and 2016, the equation did forecast the actual popular vote winner. For the Democrats, alas, the popular vote and the electoral vote majority were not in agreement.
What does this say for 2020? Even with the atypical political climate resulting from the COVID-19 pandemic and President Trump’s rhetoric, I believe that the model will successfully forecast the outcome of the 2020 election. There is anxiety over the state of the economy, but the public is only modestly more pessimistic than usual. Over the period of the study, the mean level of pessimism is 11.64%. In 2020, the level of pessimism is 12%. Additionally, despite Donald Trump’s tendency to raise hackles, the Republicans have had the White House for only four years. If the model is accurate, with these tenuous circumstances, its strength and precision will be evident. The forecast is a comfortable victory (more than 55% of the two-party vote) for the Republicans. Given that the standard error of the estimate is 2.92 and the forecast is that the Republicans win more than 55% of the vote, using the t-distribution, the estimated probability of the Republicans getting more than 50% of the vote is 95%. If one looks at the out-of-sample errors, there has only been one forecast error greater than 5.17 points. This translates to a .94 probability that the Republicans will win the popular vote.
The forecast is a comfortable victory (more than 55% of the two-party vote) for the Republicans.
US HOUSE ELECTIONS RESULTS
The same model, with one addition, can be used to forecast the seat change in the House of Representatives. The addition is the inclusion of a variable that accounts for incumbency; research has shown that incumbents are overwhelmingly favored to win (Alford and Hibbing Reference Alford and Hibbing1981; Collie Reference Collie1981; Ferejohn Reference Ferejohn1977). We can do well in forecasting House elections simply by stating that incumbents are victorious.
It is among the open seats that there is a greater possibility of seat change. The more open seats there are, the more seats available for pickup by the favored party. Midterms are typically bad for the party of the incumbent president. In on-year races, it is harder to tell which, if either, is the favored party. Here, a rule of thumb is that if one party is predicted to win by more than 60% of those surveyed, it is a good year for the party. So, in midterm elections, the number of open seats is multiplied by –1. In on-year races where the incumbent (opposition) is overwhelmingly predicted to win, the number of open seats (number of open seats multiplied by −1) is the value of this item. If neither party is overwhelmingly predicted to win, the open seat interaction item is scored 0.
As we can see in table 1, this modified presidential model does a reasonably good job of forecasting seat change in the House. The R-squared is not as high, and only two (economic pessimism and the open seat item) of the three variables are statistically significant at conventional levels. In a good (bad) year for the incumbent party, every five open seats yield a two-seat gain (loss). The greater the economic pessimism, the more seats the incumbent president’s party will lose in the House. The logged time in the White House item is both insignificant and in the “wrong” direction.
As with the presidential equation, it is important to examine the results from the out-of-sample equations to assess the model. All the items that are significant in the model are significant at the .05 level, two-tailed or better, in these equations. The average absolute error of the forecasts here is 15 seats out of 435 contests.
What does the model forecast for the 2020 elections? Taking the values of the independent variables from earlier and including the incumbency variable, scored 0, the model forecasts the Republicans will lose six seats in the House. Despite the Republicans retaining the White House, the model forecasts that the Democrats will further solidify their majority in the House.Footnote 7
CONCLUSION
This forecast, along with many others, shows that elections are explicable. With a relatively small number of variables, we can forecast the outcome of the presidential and House elections with relative certainty well before the first vote is cast. We should also note that the models shown here comport with individual-level models that highlight the importance of prospective economic items and incumbency.
DATA AVAILABILITY STATEMENT
Replication materials are available on Dataverse at https://doi.org/10.7910/DVN/A3PNXK.
SUPPLEMENTARY MATERIALS
To view supplementary material for this article, please visit http://dx.doi.org/10.1017/S1049096520001444.
