Forecast Models and Outlier in the 2020 Election

To date, researchers have analyzed the influence of economic performance, presidential popularity, and other factors on electoral outcomes, based on the results of presidential elections spanning the years between 1948 and 2020. Several models correctly predict a popular vote winner in 16 out of 19 elections but result in incorrect predictions for the 1960, 1968, and 1976 elections (viz. Lewis-Beck and Tien Reference Lewis-Beck and Tien2021). If Richard Nixon is assumed to have won the popular vote in the 1960 election (Alexander Reference Alexander2019; Campbell Reference Campbell2008; Gaines Reference Gaines2001; Longley and Pierce Reference Longley and Peirce1996), the models make a correct prediction in 17 out of 19 elections.

Donald Trump’s victory in the 2016 election stunned analysts and pundits. Before the vote count, researchers and the media mostly expected a victory for Hillary Clinton. The New York Times announced that Clinton held an 85% chance of winning, and ABC News reported that Clinton had a 95% chance of winning. In parallel, a majority of econometric models employed by scholars forecasted Clinton’s victory. Subsequently, the unexpected triumph of Trump raised questions about the robustness of forecasting models. Nonetheless, most econometric models forecast popular vote results rather than Electoral College vote results, and a great majority of models correctly predicted Clinton as a winner of the popular vote.

Rather than the 2016 election, the 2020 election presented a methodological challenge for forecast models. The year 2020 witnessed abnormal economic performance due to the coronavirus pandemic. The unemployment rate peaked at 14.7% during the year, and the annualized GDP growth rate in the second quarter of 2020 plummeted to -33.3%, marking a record low since the Great Depression. While the nation’s annual GDP growth rate averages approximately 3% from 1947 until 2020, the extraordinary economic conditions during the 2020 election year resulted in outlier values for the economic indicators. These outlier values would weaken the correlation between the economic performance and electoral outcomes, thereby underestimating the influence of economic conditions on electoral results. Also, these outliers inflate the predicted values in electoral results, thereby reducing the predictive power of forecast models.

Presidential Elections and Economic Conditions

For many decades, scholars have noted a significant influence of the economic condition of the nation on the voters’ decisions and the results of presidential elections (Campbell Reference Campbell2008; Erikson Reference Erikson1989; Fiorina Reference Fiorina1981; Lewis-Beck Reference Lewis-Beck1985, Reference Lewis-Beck1988; Lewis-Beck and Stegmaier Reference Lewis-Beck and Stegmaier2000; Tufte Reference Tufte1978). Most of the forecast models for presidential elections incorporate various indicators of economic performance (Abramowitz Reference Abramowitz2021; Campbell Reference Campbell2020; Erikson and Wlezien Reference Erikson and Wlezien2021; Lewis-Beck and Tien Reference Lewis-Beck and Tien2021; Lockerbie Reference Lockerbie2021), and they all have evinced a significant influence of economic conditions on electoral results.

The postulation of economic influence on vote decision is largely based on the retrospective voting thesis. The so-called reward-punishment model exemplifies genuine retrospective voting during elections (Key Reference Key1966; Tufte Reference Tufte1978). Key contends that the vote decisions by the populace are founded on the voters’ evaluation ex post of the president’s performance ex ante (1966). Subsequently, Tufte (Reference Tufte1978) observes that the electorate makes vote decisions largely based on the economic performance under the present administration. Therefore, according to the retrospective voting thesis, elections are quasi-referendums on the heretofore presidential performance, often pertaining to the state of the national economy.

However, due to the partisan vote and voter inertia, a period of economic turmoil or prosperity would not drastically decrease or increase the popular vote won by in-party candidates in presidential elections. Scholars note that a great majority of voters make voting decisions based on their partisanship and that voters’ party identification is very stable. The authors of The American Voter note that “Few factors are of greater importance for our national election than the lasting attachment of tens of millions of Americans to one of the Parties” (Campbell et al. Reference Campbell, Converse, Miller and Stokes1960). The analysts suggest that the results of presidential elections are generally predictable due to the stable and prevalent partisan behavior of the electorate. In comport with the Michigan model, Philip Converse’s (Reference Converse1966) concept of the normal vote exemplifies the assumption of stability in presidential elections. The normal vote is a minimum aggregate proportion of votes that presidential candidates of the two major parties are expected to receive. Converse explains that the results of presidential elections are regularly within the parameter of normal voting due to the solid influence of partisanship on vote decisions and the stability of party identification within the populace. Furthermore, in addition to the direct influence of partisanship on voters’ decisions, partisanship may influence voters’ evaluation of the economy. Several scholars observe that voters have a more positive evaluation of economic conditions when the party of the incumbent president is the party they support (e.g., Bartels Reference Bartels2002; Brady, Ferejohn, and Parker Reference Brady, Ferejohn and Parker2022; Gerber and Huber Reference Gerber and Huber2010).

Also, some voters are risk averse or lack significant information and therefore tend to vote for the incumbent president (or in-party candidates). During an election, many voters know or hear about the incumbent president. In contrast, the information about the challenger is relatively less available, less clear, and less proven. Most voters do not know about the challenger as well as they do about the sitting president. As a result, some voters—specifically undecided voters with less political knowledge—may vote for the incumbent president (Campbell Reference Campbell2008).

Consequently, electoral outcomes are generally stable over time. Even at a time of extreme economic downturns or upswings, electoral results are unlikely to significantly deviate from the outcomes in other elections. For instance, when GDP growth is higher than 10% or falls below -10%, its influence on electoral results would not differ drastically from that of 5% economic growth or -5% growth. In the partisan-bounded economic model, values higher than 5% or lower than -5% in the original measurement of economic growth are re-coded as 5% and -5%, respectively, for the secondary measurement of economic growth.

Method and Variables

In the partisan-bounded economic model, ordinary least squares estimates are analyzed for electoral results spanning from 1948 to 2020. The predicted variables include the percentage of the two-party vote won by in-party candidates, the percentage of total popular votes won by in-party candidates, and the percentage of Electoral College votes won by in-party candidates. The regressors include economic growth, presidential popularity, and shifts in party identification within the electorate.

For economic growth, GDP growth in the second quarter (annualized) of the election year is used. For the secondary measurement of economic growth, values higher than 5% or lower than -5% in the original GDP growth data are re-coded to 5% and -5%, respectively. Thus, the secondary measurement truncates GDP growth values to a fixed range of 5% to -5%. The thresholds of 5% and -5% were determined based on the standard deviation of the sample and the z-scores for outliers. The standard deviation of the original GDP growth measurement is 8.62, indicating significant variation across the 19 elections studied. The z-score for GDP growth during the 2020 election is -3.75, where a z-score higher than 3.0 or lower than -3.0 generally suggests an outlier. Preliminary analysis tested various truncation values, and the chosen threshold values of 5% and -5% were found to minimize the standard deviation and z-scores. After truncation, the standard deviation of GDP growth in the sample is reduced to 2.66, whereas the z-score for the truncated GDP growth (-5%) during the 2020 election is -2.15.

For presidential popularity, the model uses the presidential approval rating in June or a month close to June during an election year, reported by the Gallup Poll. A few analysts posit that in-party candidates under a second-term president would rather face an electoral disadvantage, whereas the voters would prefer out-party candidates (Abramowitz Reference Abramowitz2021; Norpoth Reference Norpoth2021). In-party candidates during the 1952 election (Adlai Stevenson) and 2008 election (John McCain) failed to win popular votes. However, George H. W. Bush, during the 1988 election after the Reagan presidency, Gore, during the 2000 election, and Hillary Clinton, during the 2016 election, won more popular votes than out-party candidates did. In addition, Nixon possibly won more popular votes than Kennedy did in the 1960 election after the Eisenhower presidency (Alexander Reference Alexander2019; Campbell Reference Campbell2008; Gaines Reference Gaines2001; Longley and Pierce Reference Longley and Peirce1996). Therefore, the hypothesis in question is not strongly supported. In a preliminary analysis, a dummy variable for a second-term presidency was included in the empirical test and the variable was insignificant.

Regarding partisanship in the electorate, again, the authors of The American Voter stress the influence of voters’ partisanship on electoral results (Campbell et al. Reference Campbell, Converse, Miller and Stokes1960). However, the relationship between voters’ partisanship and their vote decision could potentially be bidirectional. Morris Fiorina stresses that some voters adjust their partisanship ex post in accordance with their vote decision ex ante for the forthcoming election (Fiorina Reference Fiorina1981, Reference Fiorina2017). Thus, the shift in party identification within the electorate could serve as both a cause and a precursor to the electoral outcomes.

For the partisan shift variable, the percentages of Democratic and Republican voters during an election year (second or third quarter) as reported by Pew Research Center and Gallup Poll are recorded. The change in the party of the incumbent president from four years prior is then calculated. Independent voters leaning to the Democratic or Republican parties are excluded from the calculation. For instance, during the 2020 election, the percentage of partisan voters supporting the Republican Party was 26%, which represented a 1% decrease from 27% in 2016. For the 2024 election, the percentage of Democratic voters was 30%, reflecting a 2% decline from 32% in 2020.

Popular Vote Estimate

Figure 1 reports the scatter plot for the observed values of the percentage of the two-party vote won by in-party candidates (Vote Party) and the GDP growth (GDP). As shown on the graph, the GDP growth in 2020 (Trump, -33.3%) and the one in 1980 (Carter, -11.6%), to a lesser degree, are outliers. Figure 2 presents the scatter plot for the same dependent variable and the re-coded measure of economic growth (GDP truncated), whereas the outlier values of the GDP growth are rescaled within the fixed range between 5% and -5%. The values of GDP (truncated) for the 2020 (-33.3% for GDP) and 1980 (-11.5%) elections are -5%, whereas the values for the 1972 (6.9% for GDP) and 1984 (5.1%) elections are 5%. Figure 2 exhibits a linear pattern among the measured values.

Figure 1

Party Vote and GDP: 1948–2020

Figure 2

Party Vote and GDP (truncated): 1948–2020

Next, multivariate equations are estimated to test the influences of independent variables. Table 2 reports the results of estimates. For the percentage of the party vote won by an in-party candidate, the result of equation 1 shows that presidential popularity (approval) is significant, with a positive coefficient. In-party candidates are likely to win a higher percentage of the party vote when an incumbent president is popular. Partisan shift and GDP and in equation 1 are insignificant. For equation 2, GDP (truncated) replaces GDP. Approval remains significant, and partisan shift remains insignificant. GDP (truncated) is significant, with a positive coefficient. In-party candidates are likely to win a higher percentage of the party vote when the national economy is growing substantially. The value of adjusted R² for equation 2 suggests that the estimate explains 80% of the variation in the percentage of the party vote won by in-party candidates.

Table 2

Popular Vote Won by In-Party Candidates 1948–2020

Note: Entries are ordinary least squares regression coefficients, and standard errors are in parentheses.

⁺ p<0.10;

* p<0.05;

^** p<0.01;

^*** p<0.001.

Figure 3 presents the observed (actual) values, estimated (fitted) values, and errors (residuals) of equation 2. Figure 3 illustrates that equation 2 predicts the party vote results very well. However, figure 3 reveals relatively large residuals for the 1960, 1972, 1976, 2000, and 2020 elections. Nixon in the 1972 election, as well as Trump in the 2020 election, won a large number of votes relative to the values of their popularity and economic condition. Inversely, in the 1960 (Nixon), 1976 (Ford), and 2000 (Gore) elections, in-party candidates won a disproportionately small number of votes relative to the economic performance and the popularity of the incumbent president. Overall, equation 2 accurately predicts the correct popular vote winner for at least 17 out of 19 elections. Assuming Richard Nixon won the popular vote in the 1960 election (Alexander Reference Alexander2019; Campbell Reference Campbell2008; Gaines Reference Gaines2001; Longley and Pierce Reference Longley and Peirce1996), the model’s accuracy increases to 18 out of 19 elections. The model incorrectly predicts a victory by Ford in the 1976 election (predicted 51.5%, actual 48.9%).

Figure 3

Percentage Party Vote Won by In-Party Candidate

Apropos of the percentage of the total vote won by in-party candidates, approval remains significant, with a positive coefficient, in both equation 3 and equation 4. Also, GDP (truncated) remains significant in equation 4. Furthermore, the shift in voters’ party identification (partisan shift) is significant in both equation 3 and equation 4. The positive values of coefficients indicate that in-party candidates are likely to win a larger proportion of total popular votes when the voters of their parties increase. The value of adjusted R² for equation 4 suggests that approval, GDP (truncated), and partisan shift jointly explain 74% of the variation in the percentage of the total vote won by in-party candidates.

Electoral College Vote Estimate

The forecast model for Electoral College vote results is estimated using the same independent variables, which include economic conditions, presidential popularity, and shifts in voters’ party identification. The dependent variable is the percentage of Electoral College votes won by in-party candidates.Footnote ³

Table 3 presents the results of the estimate, and Approval and GDP (truncated) in equation 5 are significant, with positive coefficients. In addition, the shift in party identification in the electorate (partisan shift) is significant, with a positive coefficient. Thus, when partisan voters of in-party candidates increase, the candidates are likely to win more electoral votes. The value of adjusted R² suggests that approval, GDP (truncated), and partisan shift jointly explain 79% of the variation in the percentage of the Electoral College vote won by in-party candidates. Thus, the explanatory power of the estimate is substantial.

Table 3

Percentage of the Electoral College Vote

Note: Entries show coefficients, and standard errors are in parentheses.

* p<0.05;

^** p<0.01;

^*** p<0.001.

Figure 4 presents the observed (actual) values, estimated (fitted) values, and errors (residuals) of equation 5. Figure 4 demonstrates that, overall, equation 5 predicts the percentage of the Electoral College vote won by in-party candidates very well. Still, figure 4 reveals relatively larger residuals for the 1988, 2000, and 2020 elections. Bush in the 1988 election and Trump in the 2020 election won many electoral votes relative to the values of independent variables. During the 2020 election, Trump won 43% of electoral votes, but the equation estimates 24% of electoral votes for him. For the 2000 election, Gore won a disproportionately small number of votes relative to the values of independent variables. Gore won 49.5% of total electoral votes. While when the electoral votes of Florida are discounted, Gore won 51% of electoral votes. However, the equation estimates 71% for his electoral votes.

Figure 4

Percentage of Electoral College Vote: Actual and Predicted Values

Overall, equation 5 predicts accurate Electoral College vote winners for 16 out of 19 elections. The model incorrectly predicts electoral victories by Nixon (1960: actual 44%, predicted 55%), Ford (1976: actual 45%, predicted 52%), and Gore (2000: actual 49.5%, predicted 71%). For the 2000 election, researchers explain that Gore would have won in a Florida statewide recount (e.g., deHaven-Smith Reference DeHaven-Smith2005; Edwards Reference Edwards and George2019; Foley Reference Foley2019).Footnote ⁴ Accordingly, if the 2000 election is discounted, the model predicts accurate Electoral College vote winners for 16 out of 18 elections. For the 2016 election with the split results, the model proffers a largely accurate estimate for Hillary Clinton. In 2016, Clinton won 43% of electoral votes and the predicted value is 45%.