Economics,COVID,Election Forecasting: Did Trump Escape Blame?

Trump is Different

Donald Trump broke so many political expectations and norms that it is hard to try to catalogue them all. There were plenty of media and political science examples that documented Trump’s differences. For example, he lied so often that The Washington Post had to tally them and counted 30,573 lies over four plus years (Kessler et al., 2021). Journalists call him the real “Teflon Don” as they marvel at his ability to escape the consequences of his outrageous behavior. Trump called Americans who died in wars losers (Hunt 2020); he paid off porn stars to keep quiet about their affairs (Crowley 2018); he said John McCain, who was a prisoner-of-war, was not a war hero; and he has called women “fat pigs,” “dogs,” “slobs,” and worse (Byers 2015). He was overtly racist, sexist, and xenophobic on the campaign trail and in office, which actually won him support (see for example Jardina and Traugott 2019; Schaffner et al. 2018; Jardina 2019). Trump’s outrageous behavior has been found to increase support among political independents (James, 2022). His flip-flopping on foreign policy positions did not lead his supporters to abandon him (McDonald et al., 2019). When Trump arrived on the political scene with a splash it was immediately apparent that he was not your typical presidential candidate. Trump’s comments about Mexico were aired repeatedly and well known:

When Mexico sends its people, they’re not sending their best. They're not sending you. They’re not sending you. They’re sending people that have lots of problems, and they’re bringing those problems with us. They’re bringing drugs. They're bringing crime. They're rapists. And some, I assume, are good people. (Washington Post Staff 2015)

Trump also claimed that, “I beat China all the time. All the time.” At his campaign announcement event at Trump Tower in New York City he also boasted that:

I will be the greatest jobs president that God ever created. I tell you that. I'll bring back our jobs from China, from Mexico, from Japan, from so many places. I'll bring back our jobs, and I'll bring back our money.

And of course, he repeatedly called for the building of a wall along the country’s southern border and was famously caught on tape saying that he would grab and kiss women that he was attracted to. Certainly presidents have a long history of being different than their predecessors: FDR running and winning third and fourth terms, JFK giving his inaugural address in the cold without a hat, Nixon visiting China—but Trump seems to take being a different kind of president to a higher level. He had never held elected office before winning in 2016, he was a self-proclaimed successful businessman, and he was twice divorced and twice impeached. Was he so different that he could escape electoral punishment in a bad economy?

Matthew Dickinson (2018), after interviewing many Trump supporters at his rallies, suggests this possibility, arguing that Trump’s supporters were not so much focused on a bad economy or racial resentment, but rather on the rigged political and economic system. Trump supporters believed the system was broken as it was no longer rewarding hard work. Many of the political science forecasting models for 2020 went on to acknowledge the year and Trump himself as an outlier (Lewis-Beck & Tien, 2021). Forecasters grappled with how to forecast an election in a year that included Trump as the incumbent, an electorate that was evermore polarized, and an economy that cratered in the second quarter due to COVID-19 lockdowns. Some dropped economic indicators from their models (e.g. (Abramowitz, 2021)), Winsorized outlier economic data (Lewis-Beck & Tien, 2021), or used multiple economic indicators at the state level (Enns & Lagodney, 2021). Given all these anomalies with his presidency and findings that racial resentment, xenophobia, and sexism all factored in his electoral support, he makes an ideal test case for stress testing the perseverance of economic voting.

The 2016 Presidential Election: A Reward-Punishment Baseline

To explore the impact of the economy on presidential choice, in the forecasting context, a straightforward baseline model provides a convenient test reference. The Political Economy model, in service for 40 years, offers such a frame, with its argument that presidential elections are referenda on incumbent party performance regarding leading economic and non-economic issues, the former measured by economic growth and the latter measured by presidential popularity (Lewis-Beck & Rice, 1982, 1984). In words, it reads as follows:

where the presidential vote = the two-major-party share of the national popular vote for the president’s party; economic growth = the GNP growth in the first two quarters of the election year, and political popularity = the job approval rating for the president in the July Gallup Poll.

Going into the 2016 contest, the model estimates (ordinary least squares, OLS), looked like this (Lewis-Beck & Tien, 2016):

R-squared = .76 Adj. R-squared = .73 Root Mean Squared Error = 2.84.

Durbin-Watson = 2.36, N = 17 elections (1948-2012). Figures in parentheses = t-ratios. Asterisk indicates statistical significance = .05, one-tail.

To forecast the 2016 presidential election, the appropriate values of Popularity and Growth are simply plugged in. The available pre-election data, from 8/26/16, were Growth = .26 and Popularity = 51, yielding the following point estimate:

How accurate was this August 2016 forecast of a Hillary Clinton popular vote win in November? It was spot on, as she actually received this exact percentage, of 51.1. (For comparison to forecasts from the several other structural models, see Campbell 2017). In terms of the leading criterion for evaluating a forecasting model—accuracy—it appears unrivaled (Lewis-Beck, 2005). Furthermore, it does well with respect to other evaluation criteria, such as lead time (August), parsimony (two independent variables), and transparency (readily accessible data).

The model, then, offers a good reference point, at least for starters. What does it say, about the relative importance of the two predictors, popularity and growth? President Obama’s job rating was rather high, at 51 points and so carried a heavy load of votes to Clinton. [Historically, when the president’s popularity registers a majority near election day, the presidential party candidate wins (Lewis-Beck; Lewis-Beck & Charles Tien, 2019a; Lewis-Beck & Rice, 1982; Lewis-Beck & Tien, 2016). How many votes did the economy bring to Clinton? Note that economic growth was not negative but positive, a hopeful sign for her. However, translated into an annual rate, it amounted to less than one percentage point. That says, objectively, she could not expect much of an electoral boost from the economy.

Candidate Clinton herself did reckon that the economy mattered somewhat for her in 2016, dismissing charges that “I lost because I didn’t have an economic message,” but going on to caution that its benefit to her should not be “overstated.” (Clinton, 2017, 395–396, 411). Based on the marginal aggregate positive economic effect reported above in Equation 3, she would seem prescient about its mild role in her vote-gathering. Indeed, individual-level survey data, from the 2016 American National Election Study (ANES), support her assessment. The standard sociotropic retrospective attitude item, asking respondents about the economy, shows a slightly larger number said “gotten better” (30.2%) as opposed to “gotten worse” (27.2%), with “same” close to half (42.6%). This distribution suggests that citizens, collectively, correctly saw the weak national economy, in all its ambiguity. Other ANES analyses demonstrate, over an extended time series that, with one exception, when objective economic growth rises above the average, voters themselves also perceive the economy to be “above average.” (Lewis-Beck et al., 2013).

What does the pattern of perception foretell, in terms of a reward-punishment economic vote? The classic hypothesis would be that, as voter perception of the economy improved, a vote for the incumbent party candidate, Clinton, would be more likely. The 2016 ANES data uphold the hypothesis: Clinton support registers 82% among “better,” 46% among “same,” and 17% among “worse” (Lewis-Beck & Quinlan, 2019, 8). The statistical significance of this economic vote persists, though substantively diminished, in the face of extensive statistical controls (Lewis-Beck & Quinlan, 2019, Table C.10).] Thus, at the individual level of the voter, both incumbent and opposition equivalently sustained their “natural” economic constituencies, e.g., among those who saw the economy as “worse,” 78% expressed support for opponent Trump. Moreover, at the aggregate level of the electorate, there were simply not enough voters who saw the economy as “better” to save Clinton. Trump possessed no unique campaign wizardry in drawing the economically discontented to him. He merely had to stand at the head of the opposition, Republican, ballot.

The 2018 Congressional Elections: Trump Doesn’t Escape Blame

In the 2016 contest, the economic voting mechanism followed its usual rhythm, with the performance of the economy just feeble enough that Clinton could not claim victory. Concerning the Trump economic message in that campaign, it appeared mostly reactive, as he had never sought public office before and so had a meager economic policy portfolio. Nevertheless, in the Republican caucuses and primaries, voters who perceived a worsened economy were more likely to favor Trump over other Republicans; further, this finding persisted in the face of heavy controls on other variables (Tien, 2018, 30). Republican voters who perceived a bad economy were most likely more ready to support an unconventional anti-establishment candidate, like Trump. This hypothesis deserves more attention, but economic voting theory would suggest that voters who saw the economy as poor would be more willing to support non-incumbents and Trump would fit this mold as a first-time office seeker. His surprise arrival at the White House provided him with untried strategic opportunities to develop his economic message. He could employ the bully pulpit and show off his ability to work economic miracles. Facing the 2018 congressional midterms, for example, he could exploit his proclaimed business successes and bury his business failures, thereby giving a “spin” to current growth numbers. It might even be possible, then, to defy the “iron law” of midterm loss and return substantial seat gains for his party. Among his followers, there were certainly those who thought such a turnaround possible.

To test this possibility, we again take up a referendum model, as applied to congressional elections, initially offered by Tufte (1978, 106) and first implemented in ex ante forecasting by Lewis-Beck & Rice (1984a, 1984b, 1992, 62–63). Using this theoretical perspective, forecasts were made for the 2018 House midterms, then assessed in the context of the “special” appeal of Trump as an economic miracle worker. In particular, did Trump as president render outdated 2018 congressional forecasts from this longstanding explanation? Did those variables somehow work differently for him, thereby producing exceptional, positive outcomes? Or, did the economy work its will on the party in power like it always has? Below, some basic facts and obvious empirical connections shaping congressional elections in the United States are offered. Then, a test of the referendum model itself is provided.

Heading into the 2018 midterms, control of the Congress was in the hands of the Republican party, with 240 seats in the House and 51 seats in the Senate. For expository reasons, let us focus on a possible shift of control in the House, the chamber closest to the people’s will. If the Republicans lost 23 House seats or more, its majority would be no more. How likely was such a flip? Consider three historic patterns: institutional (the electoral calendar), political (the popularity of the president), and economic (the health of the economy).

An almost “iron law” of the House electoral calendar reads as follows: the president’s party has a net seat loss after a midterm. Since 1950, in only two midterms (1998 and 2002) has the president’s party had a net seat gain. A forecast based on the electoral cycle alone, then, could herald a Republican loss of control. But that forecast would be dangerous, since it fails to take into account, first, the president’s job performance. A useful rule of thumb has been a popularity number below 50%, since no president with a lower score (in June) has ever seen seat gains (across the time series, 1950–2018). Such a cut off would suggest problems for the Republicans under President Trump, given his June Gallup approval number was only 42 (Lewis-Beck & Tien, 2018).

Besides popularity, other variables have weight, in particular the economy. Hundreds of scholarly papers reveal the impact of the economics on elections, and not only in presidential elections (Stegmaier & Lewis-Beck, 2013). However, most congressional election models in the United States have measured the economy with an income variable, rather than a growth variable (Lewis-Beck & Rice, 1992, 61–62). This income measure was launched by Tufte (1975) who declared that bad midterm economic performance meant less than a two point positive change in personal disposable income. Given such a cut point, it did not look good for the Republicans in 2018, with this income number at 1.73 (December 2017 to July 2018).

These fundamental determinants of congressional election choice — the economy, popularity, and the calendar—operate in tandem. A multiple regression equation provides more precise estimates of their impact, in an ex ante forecast for the 2018 midterm. To quote from the pathbreaking work of Tufte (1978, 106) a congressional election, particularly a midterm, is, “a referendum on the incumbent administration’s handling of the economy and of other issues,” such as expressed in the following Political Economy model (Lewis-Beck & Rice, 1992, 62):

T h e m o d e l e s t i m a t e s ( O L S ) , w i t h m e a s u r e s a n d s t a t i s t i c s , a p p e a r b e l o w ,

where HS = presidential party seat change in the House of Representatives, I = change in real disposable income, for the initial 6 months of the election year (from the Bureau of Economic Analysis’s National Income and Product Account Table 2.6: Personal Income and Its Disposition), P = June Gallup poll presidential popularity rating from Gallup’s Presidential Approval Center, M = midterm dummy (0 = presidential election, 1 = midterm election), figures in parentheses are t-scores, * = statistical significance beyond .05, R² = coefficient of multiple determination, adj. R² = adjusted coefficient of multiple determination, RMSE = the root mean squared error, D-W = Durbin Watson statistic, and N = the elections from 1948 to 2016.

To forecast ex ante (summer of 2018), the appropriate independent variable values can be plugged in: I = 1.73, P = 42, and M = 1.

Accordingly, the model forecast that the House would turn Democratic in 2018. Partly, that was because of the negative weight of the midterm calendar itself. As well, a marginal economic performance and limited presidential popularity worked against seat gains. In fact, the Republicans did lose majority control of the House, with a net Republican loss of 38 seats, meaning their monopoly of Congress was broken. This result, based on a standard model with its standard measures, suggests that Trump brought to these races nothing special, either in a positive or negative sense. In sum, there was no “Trump difference.” However, his economic influence, as measured by income change, may miss other important elements. Therefore, other possible measures of a Trump economic effect are worth exploring.

Unlike 2016, Trump actively engaged the economic issue in the 2018 congressional campaign. Focusing on trade, he tweeted “Trade wars are good, easy to win.” (Borosage 2018). He “thrilled” at the “amazing” GDP growth rate in the second quarter, asserting “these numbers are very, very sustainable.” (Economist 2018, p.24). In Iowa, a pivotal state, he received public praise, like this Des Moines Register headline: “Trump Voters in Iowa Point to ‘Fantastic’ Economy.” (Obradovich 2017). Some Iowa farmers remained skeptical: “For a good number of Iowa farmers, it will be a fourth year of losses,” reported an Iowa State economist (Eller 2017). Men and women of commerce dissented: “The economic sentiment among Iowa’s business leaders has reached the highest level in a decade, according to a new survey from the Iowa Business Council.” (Hardy 2018). Such optimism expressed itself in another news caption: “Iowa incomes grow, poverty rates down.” (Norvell 2018). Not every observer looked at Iowa through rose colored lenses. In the Cedar Rapids Gazette, the story read: “Jobs are growing but pay is not.” (Murphy 2018). Colin Gordon, Senior Consultant to the State of Working Iowa Project, underlined the point, noting that median wages had leveled off since 2000 (stateofworkingiowa dot org, 2018).

Looking at the national scene, New York Times writer Thomas Edsall argued that “Fear of Falling Explains the Love of Trump.” (Edsall 2017). However, left-wing voices carped that “Trump is governing Like a Traditional Republican,” implying that economic “hard times” might harm his party at the midterm (Terkel & Bobic, 2018). In a broad attack on traditional economic measures, Robert Reich (2018, p.13), ex-Secretary of Labor, charged: “Too often, discussions about ‘the economy’ focus on overall statistics about growth, the stock market and unemployment. But most Americans don’t live in that economy.” All the above back and forth indicates the economy was at or near the top of the issue agenda in the mid-term campaign; its special impact on Trump voters might be revealed—if only the proper economic measures were used.

Clearly, the economic question offered an important talking point for Trump’s midterm campaigning. By innuendo, if not more, the congressional economic connection might have been stronger for Trump, when compared to past presidents. Still, according to our estimated House model, the independent effect of disposable income growth provided only a net nine seat gain. [See equation (3), where 4.91 (1.73) = 8.5.] What if the disposable income measure is gotten rid of, and the effect of other, perhaps more relevant, measures assessed? As an experiment, we substitute for the income measure these alternative macro-economic indicators: unemployment rate (change in), jobs growth, trade balance, savings, median earnings, labor productivity, GDP growth, GNP growth, and NBI (National Business Index). Does the Republican seat share show a noteworthy increase, once these different economic measures are used?

In Table 1, the original House referendum model appears in column 1, for comparison to the alternative models. First, in column 2, one sees the unemployment rate does not achieve statistical significance (at .05), implying that, despite Trump hopes, his low unemployment numbers did not yield more Republican seats. Other economic variables ballyhooed by him, like trade balance and median earnings, also do not register statistical significance, see columns 4 and 6. (The savings rate, in column 5, also does not show statistical significance.)

OPEN IN VIEWER

Table 1.

Comparing Economic Indicators on House 2018 Election Forecasts.

	Disp Income (1)	Unemploy Rate Change (2)	Jobs (3)	Trade Balance (4)	Savings (5)	Median earnings (6)	Labor Productivity (7)	GDP (8)	GNP (9)	NBI (10)
Constant	−44.83* (−3.44)	−31.86* (−2.25)	−49.46* (−3.93)	−46.77** (−1.95)	−53.78* (−4.90)	−43.67* (2.46)	−44.12* (−3.41)	−41.52* (−3.00)	−41.47* (−2.98)	−43.69* (−3.44)
Pres approval	0.82* (3.34)	0.84* (3.17)	0.72* (3.08)	1.10* (2.25)	1.15* (5.49)	1.04* (3.10)	.81* (3.33)	0.83* (3.16)	0.84 (3.18)	0.92* (3.80)
Midterm	−28.53* (−4.58)	−30.33* (−4.51)	−27.41* (−4.62)	−27.84* (−2.55)	−27.87* (−5.41)	−27.26* (−3.79)	−29.35* (−4.77)	−29.97* (−4.54)	−30.14* (−4.54)	−27.35* (−5.12)
Disp. Income	4.91* (2.92)
Unemployment rate change		−7.16 (−1.54)
Jobs			8.25* (3.59)
Trade balance				−0.26 (−0.72)
Savings					0.07 (0.45)
Median earnings						4.45 (1.50)
Labor productivity							6.72* (2.97)
GDP								3.52** (1.91)
GNP									3.37** (1.82)
NBI (May)										0.07 (0.83)
2018 forecast	−30	−29	−31	−25	−33	−26	−33	−31	−31	−30
Adj. R2	.55	0.47	0.59	0.41	.65	.54	0.55	0.49	0.48	.61
RMSE	18.06	19.66	17.14	18.85	13.79	15.03	17.99	19.29	19.38	14.72
Durbin-watson	1.87	1.79	2.06	2.26	2.53	2.35	2.03	1.89	1.89	2.38
N	35	35	35	13	29	19	35	35	35	32

There are objective economic variables referenced by Trump supporters that do attain conventional statistical significance, such as jobs growth and labor productivity, in columns 3 and 7. However, when these variables replace disposable income, the point estimates of Republican seat loss are still more negative, i.e., - 31 seats and – 33 seats, respectively. These slightly harsher predictions come from Trump period numbers for these independent variables (i.e., 1.86 and 1.03, respectively) and are slightly below the averages of these variables for the entire series (i.e., 2.16 and 1.33, respectively). Put another way, under Trump, performance on such jobs variables was about average, generating an average estimation of seat loss (rather than less loss, or even gain, owing to a Trump ‘difference’).

What about the trumpeted variable of general economic growth? In Table 1, column 8, real GDP growth for the first 6 months of the election year is included. That model forecasts less well than with real disposable income growth (e.g., the adjusted R-squared is six points less, and its RSME goes up by about 1.25 points). However, the GDP variable itself still manages statistical significance (but at .10). Does utilization of this favored Trump variable (and its 2018 value of 1.58) change the Republican forecast in the upward direction? No. It yields an equivalent losing forecast for the White House party, at −31 seats, instead of – 30 seats. (Moreover, observe that if the traditional GNP growth rate numbers are used, as in column 9, the results hardly change).

Overall, one sees that these alternative macro-economic indicators fail to meaningfully change the 2018 forecast of a Republican loss. Simply put, they show no “Trump difference.” The economy, variously measured, worked more or less as expected, thereby failing to generate a special, one-off Republican boost from this particular president.

The above eight measures are all objective economic indicators. What about the subjective quality of Trump’s allure? A glance at the consumer sentiments data displays promise. This monthly measure (represented in NBI in Table 1, column 10) cumulates public opinion, in a difference score of percentages (business conditions “are better” minus “are worse”). For the entire series, the variable travels from −89 percentage points to +50 percentage points. Before the 2018 midterm, respondents revealed more optimism than pessimism, with a healthy score of +34. Does this number generate a higher prediction of Republican seats under Trump, once plugged into the model? No. In Table 1, column 10, one observes the NBI coefficient lacks significance, substantively or statistically. Furthermore, including it in the model does not alter the initial negative forecast of −30. The perception of economic performance under Trump did not enhance the explanations derived solely from the objective measures of economic performance.

To the degree the economy was shaped by Trump, it showed up in the more usual indicators of economic performance, just as it has with other presidential administrations.

The immediate question here asks whether President Trump, as head of the Republican party, influenced the workings of the model. Did he make an ‘economic difference’ for the Republicans, thus generating more seat wins? The traditional economic measure—real disposable income— did not perform differently under his administration. Non-traditional measures of the economy did not reveal a noteworthy difference in seat gains either. Assessing economic performance in other ways does did not uncover a brighter picture with respect to Trump’s economic impact. His economic forays did not result in a heightened Republican grip on the House. Instead, voters responded in an “average” way to an average performance (i.e., no better or worse). Sound and fury aside, Trump failed to bring about his own extraordinary “economic difference,” so garnering a seats bonus in this 2018 congressional election. For the referendum model, it offered another example of business as usual, with respect to the meting out of economic reward or punishment for incumbent performance.

The 2020 Presidential Election: The Political Economy Model Reapplied

For examination of economic effects on the presidential vote, specifically Trump support, we return to the Political Economy model earlier employed to understand incumbent reward-punishment in 2016. The model argues that the vote share for the party in the White House can be accounted for by presidential popularity and economic growth (Lewis-Beck & Lewis-Beck, 2020; Lewis-Beck & Tien, 2021). Discussion about the uniqueness of the 2020 election campaign has been considerable: we might list COVID-19, the extremism of Trump’s leadership, economic collapse, the street protests sparked by George Floyd’s killing, among other factors. The weight of these events cannot be denied. But these effects, it could be argued, traveled through the causal variables specified in the Political Economy model, which again, reads as follows,

where the Presidential Vote = the two-party share of the national popular vote for the president’s party, Economic Growth = the GNP growth in the first two quarters of the election year, and Presidential Popularity = Gallup’s July job approval rating for the president.

Here are the estimates using OLS, and including the 2016 result.

R-squared = .76. Adj. R-squared = .73, Root Mean Squared Error = 2.75

Durbin-Watson = 2.39, N = 18 elections, 1948–2016, Figures in parentheses = t-ratios, Asterisk indicates statistical significance = .05, one-tail.

The model, as discussed, forecast the 2016 popular vote without error. Such a flawless performance in 2020 would not be expected, but a reasonable performance could be, i.e., at minimum it would pick the winner of the Trump-Biden contest. To make the forecast, then current variable values (from July 27, 2020) can be plugged in, for Popularity (41) and Growth (−4.14). (See discussion below of the outlier nature of the 2020 economy.)

A 95% confidence interval (two-tail) around this point estimate (utilizing the RMSE= 2.75 and degrees of freedom = 15) yields [37.41, 49.6]. This result implies a 95% probability that Trump would lose the popular vote. However, in two for the past five presidential elections, this popular vote outcome did not identify the winner. It becomes useful, then, to forecast the Electoral College winner, as the following OLS regression does:

This equation foresaw an Electoral College forecast of only 68 electoral votes for Trump (Lewis-Beck & Tien, 2021). Given hindsight, that forecast may appear dramatic. However, from a review of the time series data-set, one observes that before the 2020 election, the worst year for the economy consisted of the negative growth rate under Jimmy Carter and he went down in a landslide, garnering only 49 Electoral College votes. Hence, before the election event itself, the landslide loss forecast rested on precedent, plus the extraordinary statistical fit of the Popular Vote-Electoral College equation. In Figure 1, the Electoral College vote share nationwide (in percent), is regressed on the two-party popular vote, for the elections from 1948 to 2020. The actual Electoral College outcomes follow the prediction line closely, in a very tight fit (R-squared = 0.93). Moreover, there exists no slackening of that fit, when the other contests of the 2000s are examined.

OPEN IN VIEWER

Figure 1.

Electoral College Vote, by Percent of Two Party Popular Vote. 1948-2020.

To its credit, then, the Political Economy model did forecast a Trump defeat, as did eight out of the 10 forecasting models offered in the PS:Political Science and Politics symposium (Dassonneville & Tien, 2021. 49). Moreover, Biden’s national two-party popular vote (at 52.3% of the total), showed a comfortable lead over the Trump national popular vote (at 47.7%). Put another way, the difference between these two numbers, of 4.6 percentage points, indicates the race was not “a squeaker.” However, it did not come near the landslide anticipated by the Political Economy model. Of the 10 models mentioned in the PS symposium noted above, five did underpredict the Trump vote share, but not by as much as the Political Economy model. ((Dassonneville & Tien, 2021), 49, Table 2). What lies at the bottom of this condition? Is it the economy or is it Trump?—or is it COVID-19? These questions we explore further below.

OPEN IN VIEWER

Table 2.

The Impact of 2020 on the Model: OLS Results Compared.

	1948–2016	1948–2020
Pres. Approval	.26 a	.29 a
	(4.73)	(5.27)
GNP change	1.18 b	.66 b
	(2.25)	(1.99)
Constant	37.50 a	37.50 a
	(15.37)	(15.09)
Adj. R-squared	.73	.71
RMSE	2.75	2.79
Sample size	18	19

The 2020 Election: A Triptych of Outliers

With regard to the 2020 presidential election, separating out the effects of different forces poses problems because of the several outliers in play. First is the economy, the primary focus of the effort at hand. Across the time series under study, the highly negative economic growth heading into the Biden election stands apart. But, at the same time, we have the public health outlier of COVID-19, along with the obvious candidate outlier—Donald Trump himself. Below we attempt to work though this outlier maze to assess more clearly the degree of punishment President Trump received from presiding over a debauched economy. We examine aggregate-level indicators, before turning to relevant individual-level findings. First, success or failure in handling the pandemic would fall on the president’s shoulders and would show up in the president’s approval ratings. Trump’s approval rating as measured by the Gallup poll was never below 45% in the first five months of 2020 (see Gallup’s online interactive Presidential Job Approval Center (Gallup, 2022)). It was not until the number of COVID-19 deaths in the United States hit 100,000 did Trump’s approval rating from Gallup fall below 40% in 2020. Thus, how Trump’s handling of the government’s COVID-19 response affected his vote share is picked up in the presidential approval variable.

Another way of thinking about the COVID-19 impact comes from performing a counter-factual experiment. COVID-19 stands as a unique event, occurring only once across the time series. Let us suppose, counterfactually, that the 2020 world was like the 2016 world, with the political economy model as estimated in 2016 (see Equation 8), meaning presidential popularity has a regression coefficient of .26 and economic growth has a regression coefficient of 1.18. Then, let us assign pre-COVID-19 values to both independent variables, namely 45% for presidential popularity (as measured by Gallup above) and economic growth at −.36 (the first quarter 2020 rate, see Lewis-Beck & Tien, 2021, footnote 1). Plugging these independent values into the equation forecasts a two-party popular vote share for Trump of 48.78%. In other words, in that pre-COVID-19 world, he still would have not been re-elected.

As we know, after COVID-19 set in, economic conditions worsened considerably. Let us look more closely at economic conditions. Across the time series, 2020 was by far the worst, with a negative 5.40 growth rate (over the first 6 months of the year). The next worst was 1980, with a negative 1.38 growth rate. In other words, the 2020 rate was almost four times as bad as the next worst year. What do forecasters interested in the electoral effects of economic growth do? They could simply trust the linear Political Economy model to continuing working and plug the outlier value into the updated prediction equation. However, that requires quite a leap of faith. Instead, the forecaster might transform the observation in some way, bringing it back to more familiar territory. The authors of the Political Economy model in fact followed that strategy (see equation 9), adjusting the GNP change downward, to be three times the most negative growth number in the series, i.e., −1.38 × 3 = −4.14 (Lewis-Beck & Tien, 2021). Of course, this was not the only “reasonable” adjustment possible. For instance, Gelman and Heidemanns (2020), in The Economist forecasting model, argued that the negative economic impact would be The Great Recession plus 40%. Either of these adjustments, not to mention others, have a degree of arbitrariness. But in the pre-election period, they could serve as “place-holders,” avoiding a throw away of the observation, plus incorporating the plausible assumption that this heavily negative economic growth would yield the Democratic candidate diminishing marginal returns.

Nevertheless, even with this economic growth measurement adjustment, the Political Economy model (see equation 9) returned an outlying forecast, highlighting a deeper non-linearity problem, which merits illustration. Figure 2 displays the scatterplot relating economic growth (on the X-axis) and the two-party popular vote (on the Y-axis). The 2020 coordinate bases itself on the unadjusted economic growth value of −5.40 and is circled, visually revealing its outlier status. Indeed, the 2020 economic growth number falls more than two standard deviations (s.d. = 2.15) from the mean (X-bar = 1.4).

OPEN IN VIEWER

Figure 2.

Percent of Two Party Popular Vote, by Percent Change in GNP. 1948-2020.

Table 2 shows how the statistics change when the original 2020 economic number becomes part of the data-set, and the model re-estimated (OLS). In column 1 the 2016 estimation of Equation 2 is reproduced for purposes of comparison to the estimation of column 2, which adds the 2020 results. While the Presidential Approval slope coefficient barely changes (from column 1 to column 2), the GNP slope is halved and the goodness-of-fit deteriorates. Further, extensive diagnostic testing on the full data-set (1948–2020) demonstrates that 2020 stands as the only case with a large residual and a large leverage, e.g., Cook’s D = .91 (well beyond the cautionary cut-off of .21). Clearly, the 2020 observation, in raw form, falls prey to corrosive non-linearity that begs for transformation.

The question becomes, what transformation will allow restoration of the linearity assumption with its attendant OLS estimation? In Table 3, we explore four transformations: hyperbolic, logarithmic, parabolic and Winsorized. For comparison purposes, the untransformed data estimates are reproduced in column 1, to provide a baseline. The hyperbolic and logarithmic (in columns 4 and 5), clearly perform less well than the baseline, i.e., the coefficient of the economic variable no longer achieves conventional statistical significance and the model fits are less good. However, the second-order parabolic (column 2) shows promise. It offers the highest Adjusted R-squared and the lowest RMSE of all; moreover, a plot of its fitted and actual values over time (see Figure 3) is visually quite impressive. However, it fails at the critical forecasting task of predicting 2020 out-of-sample (in a simple jackknife test, it forecasts Trump to obtain 74.17%, for a whopping error of 26.4 points). This leaves the Winsorized model (in column 2), which we address fully below.

OPEN IN VIEWER

Table 3.

The Political Economy Model Estimated with Different Transformations of GNP.

	Political Economy Model	Political Economy Model Winsorized GNP	2nd Order Parabolic Transfor-mation	Hyperbolic Transfor-mation	Log(10) Transfor-mation
Pres. Approval	.29 a	.27 a	.26 a	.32 a	.31 a
	(5.27)	(5.03)	(5.20)	(5.90)	(5.61)
GNP change	.66 b	.99 b	.84 a
	(1.99)	(2.35)	(2.74)
GNP change squared			.19 b
			(2.24)
1/GNP change				−.24
				(−1.13)
GNP loggedc					3.28
					(1.20)
Constant	37.50 a	37.59 a	37.34 a	37.20 a	34.54 a
	(15.09)	(15.73)	(16.80)	(13.82)	(11.33)
Adj. R-squared	.71	.73	.77	.67	.67
RMSE	2.79	2.70	2.50	3.01	2.99
Sample size	19	19	19	19	19

OPEN IN VIEWER

Figure 3.

Actual and Fitted Values: 2nd Order Power Transformation. 1948-2020.

Winsorizing involves assigning an outlier a lesser value, while still maintaining its extreme status (Tufte, 1974, p.102). Here economic growth is measured in whole units of one, e.g., one percentage point up or down. The most extreme negative value in our series is 2020 (at −5.40) and the next most extreme negative value is 1980 (at −1.38). Therefore, we simply subtract one whole unit from the 1980 value (i.e., −1.38 – 1 = −2.38). Importantly, this elementary transformation preserves, in the data-set, the “most extreme” status of the negative economic growth in 2020. When one observes the supporting statistics of the Winsorized model, in column 2, one sees its desirable properties; the GNP coefficient, is statistically significant and at near unit elasticity (with a coefficient of .99). Moreover, its average expected out-of-sample forecasting error, as measured by the RMSE, is low at 2.70 (actually less the original 2016 Political Economy model with its RMSE of 2.84). Finally, in a jackknife test of its out-of-sample forecast of 2020 it estimates 45.35 for Trump (for an error of only −2.4).

In sum, it seems a promising model to extricate the impact of the economy on the Trump vote in 2020. Substantively, what does it say? The bad economy hurt Trump, but it was not “infinite” hurt. That is, with each percentage decline in growth, the electoral hammer did not fall with equal weight. Of course, every blow mattered, but the later ones mattered less, in a pattern of diminishing returns. Such recognition does not take away from the gravity of the negative economic effect on Trump’s vote share; but it does argue metaphorically that when an elephant steps on your foot, it does not matter much whether it is a big elephant or a little elephant.

Working through the outlier problem of the collapsing 2020 economy, under the discipline of alternative relevant theoretical specifications, allows us to speak with more confidence about the fact that Trump did receive punishment from the electorate. But was the punishment too strict or not strict enough? The above mentioned out-of-sample error for the 2020 forecast, at −2.4, suggests it might not have been strict enough. However, we need to remember that value falls very close, in absolute value, to the error the model would generate, on average, for any contest in the series, i.e., RMSE = 2.70. In that sense, then, Trump basically received expected punishment. Of course, such an inference comes from aggregate level analyses. Does that inference hold up with disaggregated, individual-level, election survey data?

Fortunately, two careful survey research efforts address the impact of the economy on the 2020 presidential vote, especially in comparison to COVID-19. The first, Clarke et al. (2021), deploys large-N national panel surveys to explore the likelihood of voting for Trump, based on a series of issues, plus long-term political and socio-demographic forces. They focus especially on the impact of two leading valence issues—the economy and COVID-19—which topped the issue agenda. Estimating a fully specified binomial probit model (Trump v. Biden), they demonstrate each issue had a statistically significant, equivalent, probability impact, shifting vote intention by 7 points net of extensive controls (Clarke et al., 2021, Table 1). Thus, neither issue crowded out the effects of the other. Regardless of COVID-19, for example, Trump did not escape blame for the failed economy.

The second study of note reinforces the above economic voting results. Neundorf and Pardos-Prado (2021, 12) examine the impact of the economy and COVID-19, in the context of a national survey experiment, where randomly assigned subjects were primed with facts on the economy, e.g., “GDP decreased by 32.9% following the start of the coronavirus outbreak, which is four times higher than after the financial crisis of 2008.” or facts on COVID-19, e.g. “the death toll in the US due to the coronavirus outbreak is the highest in the world, well surpassing 200,000 deaths.” The economic treatment was found to be statistically significant, whereas the COVID-19 treatment itself was not (Neundorf & Pardos-Prado, 2021, 25). After a series of other related tests, the authors conclude: “Trump was assessed as a standard incumbent and not as a perennial political insider with no responsibility over the economic and political turmoil….Our survey experiment revealed exogenous effects of the economic downturn….perfectly consistent with prediction based on standard economic voting and valence models of voting” (Neundorf & Pardos-Prado, 2021, 33–34). These detailed findings of 2020 individual-level voter surveys, support the inference from the foregoing aggregate Political Economy forecasting models. President Trump was blamed for economic failings under his administration, and lost critical votes needed for his re-election. Some might argue that the punishment was not enough, given the seriousness of the economic bust. But it certainly was enough to rob him of his claim that he himself was robbed of a victory.

The political economy forecasting models for presidency and Congress show that presidential approval ratings and economic performance have independent effects on vote totals for the incumbent party. Presidential approval and economic performance may be related to each other, but if there is no high correlation, among other assumptions, between the two then the regression coefficients show the independent effect of each variable on vote percentage. We measured the Pearson’s R correlation coefficient for presidential approval and GNP change for the presidential forecast model and presidential approval and personal disposable income change for the House forecast model. Pearson’s R for presidential approval and GNP change was .47, showing there is a positive relationship between the two, but not so high as to conclude presidential approval is only measuring economic issues. The strength of the relationship between presidential approval and change in disposable income was much weaker with a Pearson’s R of 0.11.

The regression coefficients tell us the independent effects of each independent variable on vote share: thus, a one percent change in presidential approval garners a .26% change in the president’s party’s share of the two-party vote. Or a ten percent increase in presidential approval July would produce a 2.6% increase in the incumbent party’s share of the two-party vote. Which of the two variables has a more substantial impact on vote share? In other words, how can we compare the effects on vote share of presidential approval and economic performance? Using the standardized (beta) coefficients, where the variables are converted to show standard deviations from the mean, will give us the comparative answers. In the presidential forecast model, the estimated beta coefficient for presidential approval is .67; and for economic growth it is .32. This means that a one standard deviation increase in presidential approval is associated with a .67 standard deviation increase in presidential party vote share, and one standard deviation increase in economic growth results in a .32 standard deviation increase in vote share. These estimates tell us that the impact of presidential approval is larger than the impact of economic growth.

How does presidential approval and change in disposable income impact House seat change in comparison? The beta coefficients for the House model are .37 for presidential approval, .24 for change in disposable income, and −.60 for midterm status. Like in the presidential forecast model, presidential approval has a larger impact on the dependent variable than does economic performance. Specifically, presidential approval has about one and one-half times the impact of disposable income (i.e., .37/.24 ˜ 1.5).

The standardized coefficient results for the presidential and House models suggest that presidential approval has a larger impact on vote results than does the state of the economy. It means that even the face of a bad economy, presidents can pull off wins for themselves and their parties in the House with high approval ratings. It is possible then, according to the forecasting models, that if Trump had handled the pandemic brilliantly he might have been reelected even with the poor economy that was devastated by COVID-19 lockdowns around the world. Trump’s low approval ratings, however, confirm that Americans did not see a brilliant performance by the man who said, “Well, accurate is that nobody’s ever done a better job than I’m doing as president” (Bumps, 2018). Apparently, American voters in 2020 did not agree.

Summary and Conclusions

In this data essay, we have focused on a principal question: Did Trump escape blame for economic woes and COVID-19 blows, in the presidential and congressional elections, 2016 to 2020? To answer, we explored Trump’s expected electoral performance, as measured by political economy forecasting models, comparing it to past predictions for other candidates. The short answer, clearly, is “No.” In the 2016 presidential race, it was Clinton, the Democratic incumbent, who made electoral gains for the economy, such as they were, not Trump. In the subsequent congressional battles, Republicans did not manage unusual seat advances, with Trump as the leader of the party. Indeed, the net Republican loses in the House were slightly larger than expected in 2018. With regard to the 2020 Trump loss, the two issues that hurt him the most were the fractured economy and the mismanagement of COVID-19. Thus, we conclude that, while Trump certainly has personal characteristics that distinguish him, the electorate essentially judged him as a politician, on balance finding his policy performance wanting on the critical dimensions of the economy and public health. As a framework, incumbency ties these three contests together, with voters primarily considering the party in the White House as culpable for governance, blaming (or praising) it accordingly. Clinton, as a Democratic incumbent, benefitted somewhat from the state of the economy, while Trump did not. When he became the incumbent, he could not engineer more than mediocre economic performance heading into 2018, and helped bring about socio-economic disaster in 2020. Overall, what we see is the classic reward-punishment mechanism of economic voting playing itself out over the period. Of course, some detractors may say, “Voters did not punish Trump enough.” But few would say “Voters gave him a free pass.” Instead, they held him responsible for damage done under his watch and are not likely to forget that track record as 2024 approaches.