THE HISTORIC, YET COMPETITIVE, 2024 US NATIONAL ELECTIONS

For the first time since 1968, the 2024 presidential election features an eligible incumbent president who declined to seek reelection. Given President Biden’s exit from the presidential race in late July 2024 following “a disastrous debate with Donald Trump that raised doubts about the incumbent’s fitness for office” and pressure by Democratic elites, Vice President Kamala Harris assumed the president’s place at the top of the Democratic ticket to oppose former president Donald Trump, despite not winning the nomination during the primary season.Footnote ¹ In the aftermath of President Biden’s decision to forgo a rematch of the 2020 election, press accounts argued that Vice President Harris is “riding a wave of momentum since announcing her candidacy,” and polling suggested that this decision increased the chances of a Democratic victory in November from likely defeat with President Biden at the top of the ticket.Footnote ² However, despite a change in the Democratic nominee and the renomination of a historically unpopular formerly defeated Republican president, the 2024 presidential contest remains hotly contested with election prognosticators, such as The Economist, rating the race as a toss-up and noting the historic unpopularity held by the retiring president.Footnote ³

Extending beyond the presidential backdrop, the battle for both chambers of the US Congress also appears to be very competitive. Despite being saddled with an outgoing president facing a historically low job approval, congressional Democrats are vying to flip control of the US House and maintain control of the US Senate. In contrast to the historical narrative portrayed in the media regarding the 2024 US national elections, this election cycle takes place during a time of incredible partisan continuity and electoral predictability. Current research shows the percentage of major party vote switchers in American elections to be less than 3% (Shino, McKee, and Smith Reference Shino, McKee and Smith2024), whereas the bivariate correlation between the presidential and congressional vote is approaching one (Algara Reference Algara2024). Moreover, scholars note that the polarized era coincides with a decline in the number of battleground states at the presidential level (Cervas and Grofman Reference Cervas and Grofman2017), competitive House and Senate races (Algara, Reference Algara2024), and even competitive US counties (Amlani and Algara Reference Amlani and Algara2021). In short, although the current 2024 election cycle is portrayed as historic and uncertain given the dramatic mid-summer decision by an unpopular president to drop out of the race, it is taking place during a period of remarkable partisan consistency in subnational voting patterns and relatively even partisan competition over a small subset of battleground constituencies.

We make three key contributions to the forecasting literature in this research note. First, we introduce new measures of presidential approval and incumbent party brand since 1937 and show that, even though both concepts are related, they are distinct theoretical and empirical concepts that can be leveraged to predict collective national election outcomes using a unified model of collective accountability. We contribute to the broader forecasting literature by developing a model forecasting the collective accountability of the incumbent party as a function of two core predictors: presidential approval and the incumbent party brand.Footnote ⁴ Second, we use these two main predictors to test how well each predicts the election outcomes of interest: (1) the presidential popular vote, (2) the presidential electoral votes, (3) the number of US Senate seats won by the incumbent party, and (4) the number of US House seats won by the incumbent party. We also leverage out-of-sample predictions to test the accuracy of our forecasting model in predicting presidential and congressional elections from 1938 to 2022. Lastly, we use our models to make predictions regarding collective accountability of the incumbent party (i.e., the Democratic Party) at each level of national partisan competition under a set of potential scenarios.

PRESIDENTIAL APPROVAL AND PARTY BRANDS AS DISTINCT CONCEPTS

Perhaps no variable is used more frequently by scholars to predict American elections than presidential approval. As Victor (Reference Victor2021) points out, the conventional model forecasting presidential elections is Abramowitz’s (Reference Abramowitz1988) “Time for Change” model that leverages three foundational predictors: party incumbency, status of the national economy, and presidential approval. By contrast—and generally within the context of making midterm election predictions—some congressional election models leverage the partisan differential on the generic ballot as their main predictor of seats won in legislative elections (Abramowitz Reference Abramowitz2006; Bafumi, Erikson, and Wlezien Reference Bafumi, Erikson and Wlezien2010). This lack of congruence between presidential and congressional election models can be a bit perplexing, particularly given the literature suggesting that the president plays a large role in shaping the parameters of partisan competition in congressional elections (for foundational work, see Key Reference Key1966 and Tufte Reference Tufte1975).

Theoretically, there are institutional reasons to believe that presidential approval and partisan brands are two distinct concepts. First, although presidential popularity can motivate the popularity of their party (Algara Reference Algara2024), it does not always translate to partisan accountability. Indeed, the literature on presidential coattails notes that presidential popularity plays a limited role in getting weak co-partisan candidates elected (Campbell and Sumners Reference Campbell and Sumners1990). Second, as an institutional matter, although presidents are the leaders of their party, partisan brands in the eyes of voters are generally thought of as being decentralized, weaker, and more ambiguous (Hetherington Reference Hetherington2001). Whereas presidents are held individually (and collectively) accountable because they are the sole elected occupant of the executive branch, parties are a collective of organized interests and individual politicians without the power to directly control their images to voters, given their lack of formal powers to control nominations.

Presidents may be individually popular, but this may fail to translate directly to the popularity of their partisan brand, suggesting that these two mass opinion assessments are distinct concepts. To test this proposition, we construct new measures of presidential approval and the incumbent party’s partisan brand, as constructed by the differential on the congressional generic ballot from survey marginals. The congressional generic ballot is a poll that is “generic” in that it measures partisan preference in the upcoming congressional election, rather than asking about specific candidates or races: the resulting generic congressional ballot measure then provides a preference for one party relative to the other party. We collected 8,412 survey marginals from 148 unique pollsters to estimate the quarterly trend in the congressional generic ballot, and the Roper Center provided 6,597 survey marginals across 99 unique pollsters to construct presidential approval ratings from 1937 through August 2024.Footnote ⁵ We use Stimson’s (Reference Stimson1998) dyad ratios latent variable model to identify shared variance across differently worded surveys designed to measure generic ballot preferences and to derive smoothed quarterly estimates of both concepts. In total, we estimated the presidential approval and incumbent party brand for 349 quarters from 1937 Q3 to 2024 Q3.

In figure 1 we show the bivariate correlation between quarterly presidential approval and the president’s party differential on the congressional generic ballot from 1937 to 2024. Higher values of the generic ballot measure indicate greater preference for the incumbent party (i.e., the president’s party).Footnote ⁶ As seen in figure 1, presidential approval and the incumbent party’s generic brand are weakly correlated at ρ = 0.287. This is also expressed in the relatively weak slope of the bivariate regression line. Moreover, the R ² of the bivariate model is 0.08, indicating that the president’s job approval among the mass public does not explain much variation in the party’s lead on the generic ballot. As the figure shows, popular presidents with greater than 50% approval may still preside over relatively weak parties, just as President George W. Bush’s 65.8% approval rating in 2002 Q1 failed to translate to a meaningful boost for the Republican Party brand on the generic ballot, with Republicans receiving 49.6% on the measure. In table 2 of the appendix, we confirm this substantive finding in more systematic hypothesis testing across four quarterly regression models showing a similar weak relationship between both concepts as conveyed in figure 1. Taken together, we find support that, even though presidential approval and the incumbent party’s standing on the congressional generic ballot are weakly correlated, they are two distinct concepts that can be used to assess the collective accountability of the incumbent party.

Figure 1

Presidential Approval and Incumbent Party Congressional Generic Percentage

Note: N = 349 quarters from 1937 Q3 to 2024 Q3. Bivariate OLS model results for figure 1: βˆ = 0.11 [H2 Robust Std. Error = 0.02; 95% CI: (0.073, 0.153); R² = 0.08]. Appendix figure A1 shows the temporal variation in presidential approval and incumbent party generic ballot percentage over time, whereas appendix figure A2 shows within-president correlation in presidential approval and incumbent party generic-ballot percentage. Appendix table 2 shows a similar relationship between presidential approval and incumbent party electoral brand across four differing model specifications as a bivariate relationship presented in figure 1.

PREDICTING US NATIONAL ELECTIONS, 1938–2022

Now that we have established presidential approval and party brands as distinct theoretical and empirical concepts, we can turn to leveraging them as key individual predictors of collective outcomes in US national elections since 1938. To that end, we specify a comprehensive full model predicting the presidential in-party’s electoral performance in US national elections as measured by the following: (1) the two-party percentage won in the national popular vote, (2) the number of electoral votes won, (3) the number of US Senate seats won by the in-party, and (4) the number of US House seats won by the in-party. We predict variation in each of these four outcomes as a function of presidential job approval, the incumbent party brand, a dummy variable indicating whether the president’s party is Republican or Democratic, a variable indicating the number of quarters the president’s party has controlled the White House heading into Election Day (i.e., the “time in power” counter variable), the unemployment rate at the quarter of the election, and annual growth in the gross domestic product (GDP) at the time of the election. In the congressional election models, we include a dummy variable coded 0 for a presidential election cycle and a 1 for midterm election cycle. Our two key covariates of presidential approval and the incumbent party brand are measured in the third quarter of the election year, which is the quarter preceding the national election.

Figure 2 shows our fully specified model for each outcome variable with respect to our two key covariates, with 95% confidence intervals estimated from the HC2 robust standard errors shown. As one can see, presidential approval is the only key covariate that predicts the popular vote percentage and electoral votes won by the president’s party, with the incumbent party brand being an insignificant predictor of these two presidential outcomes.Footnote ⁷ By contrast, our model finds that presidential approval does not predict congressional election outcomes at the House or Senate level, whereas the incumbent party brand does: this indicates that congressional election outcomes are shaped by the relative popularity of the parties, whereas presidential contests are shaped by the mass public’s assessment of presidential job performance. In appendix tables 4–7, we present the result of additional models predicting each outcome variable—including two bivariate models with just one of our key covariates of interest—and confirm that same substantive result: presidential approval does not predict congressional election outcomes, and party brands do not predict presidential election outcomes.

Figure 2

Marginal Effect of Presidential Approval and Party Brands on Election Outcomes

Note: Full model results available in appendix tables 4–7. The results in figure 2 show the point-estimates for the full comprehensive model, or Model 5 (6) in the presidential (congressional) context, in each of the appendix tables. We also present summary statistics for the annual election models in Appendix table 3. 95% confidence intervals reported in figure 2 are estimated from HC2 robust standard errors.

After evaluating the independent relationship between election outcomes and both of our covariates of interest, we can turn to evaluating the accuracy of our models using a series of jackknife tests to derive out-of-sample predictions for each election in our sample and then calculating the error between these predictions and the observed election results for each of our four election outcomes. These jackknife tests drop a given election year out of the data, re-estimate the model, and then predict the out-of-sample year to derive an out-of-sample estimate. We do this for all election years present in the data. For example, to calculate the out-of-sample popular vote prediction for the 2020 election cycle, we drop 2020 from the dataset, re-estimate the model without this observation and predict the 2020 popular vote percentage for the incumbent party from the re-estimated model results. We then compare this out-of-sample estimate for a given election year with the observed result to calculate the absolute error between the estimate and observed result, which provides a measure of the accuracy of the model. For theoretical cohesiveness, we specify our core collective accountability model with our two predictors of interest taking the form of presidential approval and incumbent party brands.Footnote ⁸

Results of these out-of-sample predictions are presented in figure 3 and appendix tables 8–11 for each presidential election outcome. On the x-axis is the incumbent party model out-of-sample prediction produced by our jackknife test for a given outcome, whereas the y-axis shows the observed election result. The 45-degree line indicates perfect congruence between our out-of-sample model prediction and the observed election result, with observations below the line indicating an incumbent party underperformance relative to our prediction and observations above the line indicating an overperformance relative to our model predictions. Each panel of figure 3 articulates our accuracy test for each election outcome. The median absolute error difference between our out-of-sample predictions and the observed results was 1.68% for the presidential popular vote model, 75.16 electoral votes for the electoral vote model, 4.48 seats in the US Senate seats model, and 17.11 seats for the US House seats.

Figure 3

Forecasting Model Out-of-Sample Predictions and Accuracy

Note: Full out-of-sample predictions, complete with 95% confidence intervals showing uncertainty around our prediction estimates and out-of-sample model fit statistics, for each model presented in appendix tables 8–11.

In terms of discrete predictions, our model correctly predicts the winner of the presidential popular vote in 19 of 21 elections since 1940, with the only misses being the 1960 and 1976 elections in which our model predicted popular vote majorities for Vice President Richard Nixon and President Gerald Ford. Perhaps reflecting the growing polarization and continuity of partisan preferences found in contemporary election cycles, the average out-of-sample absolute error in our popular vote model since 2000 is 1.26%, with the error being 1.18% and 0.02% for the recent 2016 and 2020 election cycles, respectively. Turning to the other election outcomes, our model correctly predicts the (1) electoral college winner in 15 of 21 presidential elections since 1940, (2) the Senate majority party in 29 of 43 election cycles since 1938, and (3) the House majority party in 35 of 43 election cycles since 1938. Of note, our model accurately predicts the correct House majority in more than three-fourths of the elections since 1938. Taken together, our forecasting model shows a good degree of predictive power across each of our electoral outcomes.

2024 ELECTION PREDICTIONS FROM FORECASTING MODELS

Having validated the accuracy of our forecasting models, we can make predictions for the 2024 US national elections. To do this, we take our core collective accountability model for each electoral context and estimate a prediction of the 2024 election over potential values of our key predictor of interest, given observed values of the covariates at the time of the prediction. To best illustrate this prediction method, consider the example of making a prediction of the 2024 two-party popular-vote percentage for incumbent President Joe Biden. First, we take the core model, which predicts this outcome variable as a function of our two key covariates of presidential approval and the incumbent party brand. After estimating the parameters of this model, we then estimate the predicted value of the two-party popular-vote percentage over a series of potential values of our key predictor presidential approval ranging from 38% to 55%, while holding all observed values of the covariates constant at what they are currently observed at the time of the prediction. As such, we set the observed value for the incumbent-party generic ballot covariate at 50.60% because this is what was reported on August 19, 2024, by the polling aggregator FiveThirtyEight when this prediction was derived.

We repeat this process for all election outcomes, with one key difference for congressional elections. Because we find that the generic ballot, rather than presidential approval, is the key predictor for congressional election outcomes, we derive 2024 predictions for the Senate and House outcomes over potential values of the generic congressional ballot (i.e., party brand) while holding presidential approval constant. As of August 19, 2024, President Biden’s approval rating stood at 40.64% according to FiveThirtyEight, which we consider to be the observed value for the calculation of the 2024 prediction. We report our forecasting estimates with 95% confidence intervals estimated from HC2 robust standard errors.

Table 1 shows our popular-vote percentage forecasting estimate for President Joe Biden in the 2024 elections over potential values of his approval rating. Assuming about a roughly 41% approval rating, which is observed at the time of this writing, our model forecasts Democrats winning 47.21% of the popular vote [95% CI: 45.20, 49.22]. Assuming that President Biden does not improve on his relatively low presidential approval rating, our model forecasts a narrow loss in the presidential popular vote for Democratic nominee Vice President Harris. As table 1 also shows, a dramatic increase in President Biden’s approval rating to 49% would predict a robust popular vote majority at 51.52% with the lower bound of the 95% confidence interval being over 50%, indicating a very high degree of confidence in this majority at this presidential approval level.

Table 1

Presidential Popular Vote Prediction over Presidential Approval Levels

Notes: Predictions derived from Model (3) and observed covariate values on August 19, 2024; 95% confidence intervals around the forecast estimates are derived from HC2 robust standard errors.

Table 2

Presidential Electoral Vote Prediction over Presidential Approval Level

Notes: Predictions derived from Model (3) and observed covariate values on August 15, 2024; 95% confidence intervals around the forecast estimates are derived from HC2 robust standard errors.

By contrast, the 2024 forecast is much less optimistic for Democrats with respect to the Electoral College. Our model forecasts Vice President Harris would secure about 168 electoral votes [95% CI: 98.39, 237.49], assuming a presidential approval rating of 41% on Election Day. Given that the upper bound of our 95% confidence interval for this Electoral College vote forecast sits at 237.49, our model is very pessimistic regarding Democratic chances of holding the White House with a co-partisan president holding a roughly 41% approval rating. If this observed approval rating holds, President Biden would have the third-lowest incumbent-party presidential approval rating since 1940 according to our estimates, only beating the 35.9% approval rating for President Bush heading into the 2008 election and the 39.2% approval rate for President Truman on the eve of the 1952 election. Reflecting this unpopularity in retiring incumbent approval, the 1952 and 2008 elections ushered in Electoral College landslides for the out-party along with robust congressional majorities.Footnote ⁹ Given these cases, it is clear why our model is fairly pessimistic regarding Democratic odds in the Electoral College, given the current incumbent’s approval at the time of this writing.

Turning to the US Senate in table 3, our model is fairly optimistic regarding Democratic chances to hold the chamber this November. Assuming the current observed generic ballot percentage for Democrats at the time of this writing at roughly 50%, our model forecasts Democrats to control about 51 Senate seats [95% CI: 47.46, 54.91]. However, we note the fairly large confidence intervals around our forecast estimate, suggesting volatility in this estimate. Reflected across all potential values of generic ballot support ranging from 47% to 53%, the confidence intervals show a great degree of volatility, perhaps owing to the traditional finding that Senate races are much more idiosyncratic candidate-driven contests, compared to presidential elections, that can buck national partisan tides (Algara Reference Algara2024).

Table 3

2024 US Senate Prediction Over Generic Ballot Levels

Notes: Predictions derived from Model (4) and observed covariate values on August 15, 2024; 95% confidence intervals around the forecast estimates are derived from HC2 robust standard errors.

Last, we turn to the 2024 forecasts for the US House found in table 4. As the forecast shows, Democrats are highly competitive in their quest of reclaiming the majority lost in 2022.

Table 4

US House Prediction over Generic Ballot Levels

Notes: Predictions derived from Model (4) and observed covariate values on August 15, 2024; 95% confidence intervals around the forecast estimates are derived from HC2 robust standard errors.

At roughly 50% in the generic congressional ballot, Democrats are predicted to hold 222 seats [95% CI: 210.20, 233.71], which would mirror the number of Democratic seats after the 2020 US House elections that netted the narrowest Democratic majority since 1942. If the incumbent party could increase their generic ballot percentage by roughly 0.4% to 51%, they would be predicted to win about 232 seats [95% CI: 219.54, 244.01], which is 14 more than required to retake the majority in the US House of Representatives and would be similar to what Democrats won during the 2018 midterm elections.

DISCUSSION: LOOKING TOWARD NOVEMBER

In this research note, we make two contributions. First, by leveraging new estimates of presidential approval and party brands, we show that these two considerations are distinct and thus could potentially be used as independent predictors of US national election outcomes within the same collective accountability model. Indeed, although presidential approval and party brands are weakly correlated, we show a large degree of variation in the incumbent party brand that is not explained by the mass public’s job evaluation of the president, who by definition is the leader of the incumbent party. Second, we validate our unified collective accountability model by showing that presidential elections are largely a story of the mass public’s approval of the president, whereas congressional elections are decided by the mass public’s assessment of the incumbent party relative to the out-party. Out-of-sample predictions further validate the accuracy of our model.

In our 2024 forecasts, we find evidence that Republicans are favored to win a robust Electoral College majority and a narrow popular vote majority because President Joe Biden’s historically low approval rating will weigh down Vice President Kamala Harris’s electoral fortunes. This disconnect between our predictions in the popular vote and Electoral College perhaps reflects the pro-Republican bias found in the Electoral College during contemporary elections (Erikson, Sigman, and Yao Reference Erikson, Sigman and Yao2020), with Republicans being more strongly favored in carrying a majority in the Electoral College as opposed to the popular vote. Our forecasts show that while Republicans are well suited to win the presidency, control of both the U.S. Senate and U.S. House are essentially a toss-ups.

We conclude with a potential limitation of our forecasting approach. In addition to standard economic and contextual predictors, our model only considers presidential approval and party brands to generate 2024 election forecasts. This can be potentially limiting given recent work. Indeed, we concur with recent scholarship by Highton and Stone (Reference Highton and Stone2024) showing that presidential election outcomes are more than just mere referendums on the incumbent’s performance in the minds of voters; rather, they are about candidate choice presented to the mass public. Indeed, our model does not incorporate a differential measuring a relative advantage or disadvantage of the incumbent party’s nominee relative to the challenger, independent of other traditional predictors of electoral outcomes, such as presidential approval or economic considerations. However, as Highton and Stone (Reference Highton and Stone2024) suggests, such preelection measures of candidate-based differentials on dimensions such as valence and policy are far less systematically collected than preelection measures such as presidential approval.Footnote ¹⁰ For our purposes, this could be a salient variable to include in forecasting the 2024 presidential elections given the unpopularity of former president Donald Trump and the replacement of an unpopular president at the top of the ticket. Our forecasting model is pessimistic regarding Democratic chances in the presidential election and the ability of congressional Democrats to convincingly garner a majority in both chambers of Congress.