Forecasting elections is an enterprise that intrigues and amuses political scientists, politicians, and the general public. We have seen office betting pools around political science departments. Politicians and the talking heads will have bets concerning their forecasts and predictions of the upcoming election. Despite the transfers of wealth and honor that come with these wagers, these activities are somewhat less than we might like. Unless, of course, the transfer of wealth flows to us.
What makes these activities less than satisfying is that there is no mechanism by which to assess the forecast. Yes, we know if the prognosticator got it right. We do not, however, know by what means the forecaster made a prediction. Should we trust the forecaster the next time around? The models the academic forecasters come up with are laid out for all to see. Our data are made available with the forecast. Assuming we do not constantly tweak our forecasts, one can assess the accuracy. If we change our forecast so that we can point to one as being right, we are creating a non-falsifiable forecast. If one time I forecast the Democratic candidate will win and another time say the Republican candidate will be victorious, I have guaranteed that that I can point to one forecast as getting it right. If I note that I have had several forecasts that differ, then I have treated the consumer of my forecast properly. Also, one can update one’s forecast with relevant information. When discussing one’s forecast, it should be noted that it has been revised.
Aside from being replicable, what other characteristics should a forecast have? First, it should be tied to a theory of voting behavior. We can find correlations between election outcomes and other events. There are “rules” such as the longest named candidate will win, the National League winning the World Series means the Democrats win, and a Lakers national championship forecasts a Republican victory. These work until they don’t. They then fail spectacularly. Simply look at the abject failure of the World Series rule for the three elections since 1976.
One additional item would should ask of a forecasting model is that the forecast be made well in advance of the event. If one hopes to make use of a forecast, one needs to have some lead time. Imagine that one gives forecast of a football game with less than a minute to go. The forecaster would be correct an overwhelming majority of the time. We, however, would not place much stock in the forecaster. Similarly, a weather forecaster who tells us a storm is brewing as we see the rain coming in the distance has not really provided a service. While we will want updated information about the direction of a storm, we are better able to deal with the storm if we have some advance notice. Similarly, the more lead time with a political forecast, the better able one is to prepare for the outcome, all else being equal of course.
INFLUENCES ON ELECTION OUTCOMES
There are many models of voting behavior that one could include in the analysis here. There are, however, too few cases to include many variables. Moreover, many of the items we might include are highly related to each other. Here, following the work of Fiorina (Reference Fiorina1981), Key (Reference Key1966), Lewis-Beck (Reference Lewis-Beck1988), and Lockerbie (Reference Lockerbie2008), I test an economic model of voting behavior. Because of the limited sample size and the high correlation (.92) between the retrospective and prospective items, I focus on the prospective side. The prospective item is from table 8 of the Survey of Consumer Attitudes and Behavior. The score is the average of the responses in the negative from the second quarter of the election year to the 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 just about the same?” While this item is clearly focused on the future and the individual, it does not have any sense of attribution to either party. One could believe that one’s financial situation is going to get worse because the incumbent will win. One could also believe it will get worse because the incumbent will win. Given the high level of correlation between the retrospective and prospective items, it is reasonable to believe that if one thinks things are going to get worse, the incumbent party is held responsible. Higher values on this item on this item should lead to a diminished vote for the incumbent. Moreover, the respondent could think that changes in one’s financial situation are a reflection of their own efforts and abilities and not either party. If voters do not think either party is responsible, there should be little reason to expect a relationship between these items and electoral outcome. Kramer (Reference Kramer1983) and Lockerbie (Reference Lockerbie2008) make the point that holding officials responsible for economic conditions is required for the economic items to have a bearing on voting behavior. We can take some solace in that if a relationship appears, it probably understates the relationship at the individual level.
Aside from economic expectations, we should note that there is a regular pattern in American politics with regard to presidential approval. After a party has controlled the White House for a while, the number of opportunities to frustrate a portion of the electorate goes up. Abramowitz (Reference Abramowitz, Campbell and Garand2000) has noted that regardless of a president’s popularity or the state of the economy, there is a desire to change the party controlling the White House. Footnote 1 An incumbent party being returned to the White House for a third term without the same candidate running for office is very rare in the last hundred years: George H.W. Bush in 1988, Harry S. Truman in 1948, and Herbert Hoover 1928. To account for the diminished support for the incumbent party, I make use of logged time in the White House for a party as my measure. Footnote 2
As with my earlier models, I do not include presidential popularity. I argue that the same factors that lead to support or opposition to a president also lead to election or defeat of the incumbent party. In a sense, including presidential popularity in the model is akin to putting a lagged version of the dependent variable on the right-hand side of the equation. In some ways, it is like placing a feeling thermometer differential of the two candidates in an individual level model. Yes, the R-squared will probably go up. Does it, however, add to the theoretical power of the model?
The more pessimistic people are about the future of their personal finances, the lower the share of the vote for the incumbent party.
PRESIDENTIAL ELECTION RESULTS
The first equation in table 1 shows the results of the analysis for the presidential election model. Briefly, it performs like it has in earlier iterations. The more pessimistic people are about the future of their personal finances, the lower the share of the vote for the incumbent party. The incumbent’s party loses a little over one-half a percentage point for each one percentage point increase in pessimism. And, as expected, the longer a party has controlled the White House, the more poorly they do. A party that has held the White House for eight years loses approximately 17 points. We should note the constant is over 70, so even with this loss of support, the incumbent should have a reasonable chance of victory. Each of the variables is highly significant. The R-squared of this equation is a relatively strong .72.
Table 1 Forecasting Equations
As this forecasting equation has a small N, I made use of robust regression to minimize the chance that one case could dominate the results. Footnote 3 Given the small sample size, I reran the equations to generate forecasts with out-of-sample forecasts. In short, the equations are rerun excluding one year at time. In each case, the variables that are significant in the original equation are significant here too. This out-of-sample modeling also allows us to generate forecasts without the actual year being forecast in the equation. We are not using the data from a year to generate the equation from which we are checking the accuracy of the equation.
The results of this analysis show a robust equation. Only for 1980 is it close. The economic item’s significance level drops to .058, two-tailed. Given the small number of cases, this is not too terribly surprising. Moreover, we should note that again, except for 1980, the coefficient for the economic item is very stable. Even in 1980, it is reasonably close to that of other years. How effective is the model in forecasting the outcome of elections? Looking at this out-of-sample forecasting, we see that the model does very well (table 2). The average absolute forecasting error is 2.78 percentage points. The equation misses the elections of 1960 and 1968. Richard Nixon is forecast to win in 1960 and lose in 1968. As both of these elections were exceptionally close, this is not too terribly surprising. In 1980, the model has its biggest percentage error, but it does get the ultimate victor. Last, in 2000, the equation does get the popular vote winner. Alas, for Al Gore, this was one of the rare times the popular and electoral votes diverged.
Table 2 Out-of-Sample Forecasting Equations: Presidential Elections
What does this say for 2016? First, now that the Democrats have controlled the White House for eight years, they are at a bit of a disadvantage. Second, the overall pessimism of the electorate is relatively modest, they should do okay. On balance, these two variables portend a very close election. Placing these two pieces of information in the equation, the forecast for 2016 is a narrow, very narrow, victory for the Democratic candidate. The Democratic candidate is forecast to receive 50.4 % of the two-party vote. Given that the election is forecast to be extremely close and the average error of the equation, there is but a 62% likelihood of a Democratic victory. In short, I would not want to wager much on the outcome of the election.
US HOUSE ELECTIONS
We can take the same model, with a modest modification, and forecast House elections (table 3). The one modification is to add a variable for the incumbency advantage. If we know anything at all about House elections, incumbents win (Alford and Hibbing Reference Alford and Hibbing1981; Collie Reference Collie1981; Cover Reference Cover1977; Erikson Reference Erikson1972; Ferejohn Reference Ferejohn1977; Fiorina Reference Fiorina1977). In forecasting House elections, we can do well forecasting that incumbents will win. It is typically in the open seats that there are partisan changes in control. The more open seats there are, the more a favored party can pick up. We know that in midterms the party in opposition to the president is advantaged. In on-year elections, it is bit more difficult ascertain the advantaged party. Here, the rule of thumb is that if a party is forecast by 60% or more to win the election, it is a good year for that party. So, in midterms the number of open seats is multiplied by negative one, as it is in midterm years. In on-year races where the incumbent party is forecast by 60% or more to be victorious, the actual number of open seats is used. If no party is forecast by 60% to be victorious, the open seat item is scored 0.
Taking into account the values of the independent variables (openseat is scored 0, as no one is overwhelmingly favored to win the election), the model forecasts a net change of 0 seats in the House of Representatives.
Table 3 Out-of-Sample Forecasting Equations: House Elections
The pattern of significance is the same for the equation reported in table 1, with the exception of the openseat interaction item for 1964.
The modified presidential model does a reasonably good job of forecasting House elections. Although the explained variation is not as high as that of the presidential equation, it is still a reasonably strong model. Moreover, the pattern of significance is similarly explicable. The incumbency based item is highly significant, as is the economic item. The logged time in the White House item, however, is not statistically significant. It is not even close. Moreover, it is actually in the unexpected direction. How much change is induced by these significant independent variables? In a good (bad) year for the incumbent party, every three open seats yields approximately a three seat gain (loss). The more pessimistic about the economic future, the worse the incumbent party does. On average, every two unit increase in pessimism leads to a three seat loss for the incumbent president’s party.
As with the presidential equation, it is important to examine the out-of-sample equations to assess the model. Here, the strength of the model shows through well. The only items that fail to replicate their level of significance are the open seat interaction item in 1964 and the economic item in 1980. Both of these items are significant at the .05 level, two-tailed. How good a job does the model in forecasting the outcome of House elections? The average absolute error is right at 16 seats. While this is not inconsequential, it shows that the model does a reasonably good job of forecasting the outcome of 435 potential contests.
What does the model forecast for the 2016 House elections? Taking into account the values of the independent variables (openseat is scored 0, as no one is overwhelmingly favored to win the election), the model forecasts a net change of 0 seats in the House of Representatives. In fact, given the number of seats the Republicans currently hold, there is a 98% likelihood that the Republicans will retain control of the House.
CONCLUSION
This forecast, along with others, shows that presidential and House elections are predictable. With a relatively small number of variables that are available well before either party has held their convention, we can offer forecasts of the outcomes. Footnote 4 Moreover, we should note that the models presented here comport with individual-level models that show the power of economic expectations and incumbency.
