What to expect when you're electing: citizen forecasts in the 2020 election

Abstract

Political divisions in the lead-up to the 2020 US presidential election were large, leading many to worry that heighted partisan conflict was so stark that partisans were living in different worlds, divided even in their understanding of basic facts. Moreover, the nationalization of American politics is thought to weaken attention to state political concerns. 2020 therefore provides an excellent, if difficult, test case for the claim that individuals understand their state political environment in a meaningful way. Were individuals able to look beyond national rhetoric and the national environment to understand state-level electoral dynamics? We present new data showing that, in the aggregate, despite partisan differences in electoral expectations, Americans are aware of their state's likely political outcome, including whether it will be close. At the same time, because forecasting the overall election outcome is more difficult, Electoral College forecasts are much noisier and display persistent partisan difference in expectations that do not differ much with state of residence.

The lead-up to the 2020 election provided a stark setting for understanding contemporary political divisions in the United States. In popular (e.g. Rutenberg, 2020) and academic commentary (e.g. Cornwall, 2020), there was widespread concern that partisan divisions substantially heightened the contest's perceived stakes. Indeed, partisanship is seen as divisive not just because it correlates with beliefs and preferences, but because it is also believed to shape what information individuals encounter (e.g. Garrett and Stroud, 2014, Peterson and Iyengar, 2021) and the biased way in which they process it (e.g. Jerit and Barabas, 2012; Druckman et al., 2021; Groenendyk and Krupnikov, 2021). Some are concerned that these forces are so strong that partisans effectively inhabit different worlds, divided even in their understanding of basic facts (e.g. Gaines et al., 2007).

Against this backdrop, the stakes of losing an election are likely to be perceived as extremely large, and the other side's victory can be cast as “cheating” by the other side (e.g. Sances and Stewart, 2015; Sinclair et al., 2018; Alvarez et al., 2021). Indeed, in the lead-up to the election, both Democratic (e.g. Grove, 2020) and Republican (e.g. Hakim and Saul, 2020) elites preemptively raised arguments that would frame their own defeat as illegitimate. Nationalized political conflict is part of a broader trend in which local, state, and regional-level differences in party positions and policy conflict are supplanted by more uniform nationalized political conflict. Nationalization, reinforced by media coverage focused on national conflict at the expense of local candidates is thought to further weaken attention to local political concerns (e.g. Hopkins, 2018).

In light of this setting, the 2020 campaign provides an important, if difficult, test case for the claim that individuals meaningfully understand their state political environment and the distinction between that environment and the national electoral contest. Following the widespread failure of pre-election polls in 2016 to predict state-level outcomes that led to President Trump's unexpected victory, many individuals expressed widespread concerns about the accuracy of traditional polling reports (e.g. Madson and Hillygus, 2020). Moreover, 2020 presents an interesting case for Americans to accurately appraise their political environment since national surveys indicated large portions of both candidates' supporters doubted published information about the state of the race.¹

As politics become more nationalized, and in turn more polarized, one might expect citizens to be unable to accurately make sense of their own political contexts. Increases in the nationalization of American politics have been associated with greater party loyalty and negative partisanship in the electorate (e.g. Abramowitz and Webster, 2016). In turn, these phenomena have been associated with a greater likelihood of exposure to partisan slanted information (e.g. Ahler and Sood, 2018). Additionally, voters may project national trends onto their home states or partisan bias may influence their perceptions of their local political environment.² But were individuals able to look beyond national rhetoric and the national environment to understand state-level dynamics? What did individuals expect to happen when they were electing? And if individuals do understand their own state environment, how do these beliefs inform, if at all, the much more difficult task of forecasting both the Electoral College outcome and the national popular vote?

In this note we present new data about citizen expectations about both their state and national level outcomes in the 2020 election. Our work is inspired by studies that have used individual-level forecasts (i.e., who will win the election) rather than traditional polls (i.e., who will you vote for) to forecast election outcomes (e.g. Miller et al., 2012; Graefe, 2014; Leiter et al., 2018a, 2018b; Johnston et al., 2019; Murr and Lewis-Beck, 2020; Murr et al., 2021). Many of these studies have demonstrated that at the national level, the wisdom of crowds tends to manifest, more so in less close elections (e.g. Lewis-Beck and Skalaban, 1989; Lewis-Beck and Tien, 1999). Others have focused directly on state-level predictions but have found mixed evidence that voters are grounded in reality. For example, Murr (2015) finds that Americans were quite accurate in predicting their own state's presidential outcomes in 2012, but Lewis-Beck and Murr's found less accurate state-level forecasts in the summer of 2020 for the upcoming presidential contest (2020).³

This prior work lays the foundation for our own study. In an environment with even greater partisan polarization and continued nationalization of American elections, which pattern persists? We focus on the 2020 election, draw from a representative sample, and separately ask about forecasts for state- and national-level outcomes. In this way, we can learn if Americans are able to distinguish between national political outcomes and state political environments and assess how tightly coupled these forecasts are. Moreover, in addition to examining the average accuracy of forecasts, we also investigate acknowledged uncertainty in forecasting. Our key theoretical question is whether partisan divisions and nationalization are so stark as to render citizens' knowledge of their own political environment meaningless in forecasting state-level outcomes. Effectively, can the wisdom of crowds persist in this nationally polarized environment? Do individuals accurately understand the greater uncertainty of state-level forecasts in more competitive states? Additionally, we ask whether individuals are able to answer an even harder question, who will win the national election, and whether answers to those questions are influenced by state-level context.

While we present evidence that average perceptions of political environments are quite grounded in reality, we note that partisan differences still exist. Furthermore, we also acknowledge the limitations of relying on small state samples to reliably predict election outcomes. Our contribution in this paper is not to provide a novel forecasting method; instead, we take advantage of a salient partisan political moment to understand if citizens' views of contemporary politics are entirely nationalized.

How well can citizens predict state and national election outcomes?

Data for this project were gathered on a private team module fielded on the 2020 pre-election wave of the Cooperative Congressional Election Study (CCES).⁴ The survey was conducted by YouGov in the lead up to the presidential election. We asked 1000 survey respondents three questions about whom they expected to win the upcoming presidential election in their home state, the national popular vote, and the Electoral College. They provided their predictions on a five-point scale, with 1 indicating that it was “extremely likely” that Trump would win and 5 indicating that it was “extremely likely” that Biden would win. A value of 3 corresponds to the panelist believing it was “equally likely” that either would win. The CCES also gathered information on respondents' home state, sex, education, age, income, and partisan identification, which we use in our analyses. Finally, YouGov provides poststratification weights for weighting to a nationally representative population.

Survey respondents on average understand state context

Figure 1 presents the average response and associated 95 percent confidence interval for each of the three forecasting questions in separate panels. For these analyses we rescale the variables so 0 corresponds to the highest level of confidence that Trump will win, 1 corresponds to the highest level of confidence that Biden will win, and 0.5 means the respondent is uncertain. Each panel also displays the average subsetted by respondent partisanship, whether the panelist lives in a state won by Donald Trump or Joe Biden, and the combination of these two categories.

Figure 1. Aggregate predictions for the 2020 election by party and state outcome. In panel A, we present the mean value for the question “Who do you think will win your state's popular vote in the upcoming election?” Responses were provided on a five-point scale. We have rescaled the values so that 0 = Certainly Donald Trump, and 1 = Certainly Joe Biden.” Values closer to zero indicate the group was more likely to say Donald Trump would win. Values closer to zero indicate the group was more likely to say Joe Biden would win. In panel B, we present the mean value for the question, “Who do you think will win the Electoral College?” In panel C, we present the mean value for the question, who do you think will win the national popular vote?” The first subset of each panel displays the difference between Republicans' and Democrats' responses to each question. The second subset of each panel displays the difference between those panelists living in states Trump won and those panelists living in states Biden won. The final panel displays the differences between Republicans living in Trump states and Republicans living in Biden states and the differences between Democrats living in Trump states and Democrats living in Biden states.

Source: 2020 Private CCES team module.

Panel A displays the average levels of confidence for the home state question. The top portion of the panel shows that on average, Democrats are significantly more likely to have higher levels of confidence that Biden will win their state. The mean difference between Democrats and Republicans is 0.315 on the 0 to 1 scale, suggesting large differences in expectations about state outcomes between opposing partisans. However, this difference may arise due to true partisan difference in expectations, expressive responding (e.g. Bullock et al., 2015), or differences in where partisans live on average (e.g. Brown and Enos, 2021). When we examine who actually won the state in which a respondent lives, as in the middle portion of the panel, however, we find that those respondents who lived in Trump states were much less likely to indicate that Biden would win than those who lived in Biden states. This difference of 0.383 on the 0 to 1 scale is larger than the average difference between partisans. This difference is also compatible with multiple expectations, including knowledge of state context and different mixes of partisans across states, but it provides strongly suggestive evidence that partisans are not blind to their political context.

When we simultaneously examine the effect of respondent partisanship and home state as shown in the bottom portion of the panel, we find sizable gaps in reported expectations about who will win Trump rather than Biden states among both Republicans and Democrats. For example, among Republicans, those who live in a Biden (rather than Trump) state are 0.395 higher in their belief that Biden will win their state, and for Democrats the effect is similarly 0.382 units. That is, Republicans and Democrats give different home state predictions depending upon political context. They were not entirely driven by partisan differences in beliefs or expressive responding when offering predictions about their home state. Nonetheless, there are still partisan differences holding fixed state context, on the order of 0.284 in states Biden won and 0.297 in states Trump won, although once again we note that these could arise due to sincere beliefs or expressive responding.

To further understand how expected state outcomes vary by state, we estimated a model using ordinary least squares regression (OLS) in which we predict each respondent's confidence that Biden would win their home state using indicators for state of residence.⁵ We omit panelists who indicated they thought their state was a toss-up and created a binary outcome among those who expressed confidence that Trump (0) or Biden (1) would win their state. State averages are therefore the proportion of individuals who thought Biden (rather than Trump) would win their state among those expressing confidence.

Panel A of Figure 2 plots the estimated state coefficients from this regression. The y-axis at the indicates coefficient size, while the x-axis denotes the two-party vote share for Biden. Higher estimates on the y-axis indicate that respondents in that state were more confident that Biden would win the state. These axes are aligned so that above 0.50 on the x-axis indicates Biden won the state and above 0.50 on the y-axis indicates that, on balance, respondents in this state expected Biden to win. Thus, the upper right panel represents states Biden won that respondents expected him to win, while the lower left panel represents states Trump won that respondents expected Trump to win.

Figure 2. Predicting state-level outcomes. Panel A plots Biden's 2-party vote share (x-axis) against the coefficient of a model in which the prediction of Biden winning one's home state was regressed on the respondent's state (y-axis). The outcome variable is coded as 1 if the panelist thinks Biden will win the presidential election in their state and 0 if the panelist thinks Donald Trump will win the election in their state. We omit those who think the state will be a toss-up. We rescale the x-axis value of Washington DC from 0.94 to 0.75 for presentation purposes. Panel B plots Biden's 2-party vote share (x-axis) with the coefficient of a model in which the prediction of the state being a toss-up is the outcome. The outcome variable is coded as 1 if the panelist thinks it is equally likely that Joe Biden or Donald Trump will win the election and 0 if they believe Donald Trump or Joe Biden is likely to win their state. This model includes all panelists. Wyoming serves as the intercept. Models were estimated using ordinary least squares regression. Regression tables are available in Tables A1 and A2 in the Appendix.

The correlation between observed state outcomes and predicted state outcomes is strong. Respondents in states Trump won handily were the least confident that Biden would win their state, with the seven safest Trump states producing an estimated coefficient of zero. Similarly, those panelists in the safest Biden contests were confident Biden would win, with the three safest Biden races producing an estimated coefficient of 1. Those states that were closer produced estimated expectations close to 0.5, indicating that roughly equal numbers of panelists in each state thought Trump or Biden would win the state. Overall, in the aggregate, panelists seem to have a reasonable idea about who would win in their state.⁶ (Maine and Ohio are notable outliers here, although Trump did win one Electoral College vote in Maine.)

Did respondents in closer states recognize the outcome was more uncertain? In Panel B of Figure 2 we present the estimated coefficients for a similar regression in which the outcome variable is 1 if the respondent indicated that they thought their state was a toss-up and 0 if the panelist thought either Trump or Biden would win their state. Once again, we plot Biden's two-party vote share on the x-axis, but now the y-axis represents the coefficient estimate for respondents who thought their state would be a toss-up. Higher values represent greater uncertainty about the state-level outcome.

As the pattern indicates, states where Biden or Trump won by larger margins have smaller estimated coefficients. That is, respondents in these states had less doubt about their state outcome. Among more competitive states, the estimated coefficients increase to near 0.25. This result indicates that in more competitive states, respondents were more likely to express doubt about who would win their state. In battleground states (e.g. Michigan, Nevada, Pennsylvania, Wisconsin, Arizona, Georgia, North Carolina, and Florida), respondents were 29 percentage points more likely to express they didn't know who would win compared to all other states (33 percent vs. 14 percent). Once again, citizens appear to have been grounded in reality when predicting the closeness of their state's electoral outcome.⁷ Of course, it is precisely in these electorally uncertain states where concerns about electoral legitimacy loomed largest, likely because ex ante individuals did not know how the election would turn even if the votes were counted fairly and so the ex post revelation of who was reported to have won would have been most suspect in the face of allegations of fraud.

Cumulatively, these data provide a surprising rejoinder to expectations that an intensively polarized and nationalized political environment would produce survey respondents unable to recognize, in the aggregate, the contours of their state political environments. Even in the presence of substantial partisan differences in stated beliefs about state-level elections outcomes, respondents on average forecast state election outcomes correctly and recognize which state outcomes are more uncertain.

National forecasts are uncertain and largely unaffected by state context

If state residents on average understand their state's politics, how does this realized collective wisdom inform expectations about who will win the national popular vote or the Electoral College? On the one hand, people may extrapolate from their expectations for their state. On the other hand, given that discussion of the Electoral College is ubiquitous, individuals may be aware of the differences between the race in their state and overall. We now turn to those questions. Returning to Figure 1, Panel B provides graphical analysis for predictions about the Electoral College and Panel C displays predictions for the national popular vote that parallel our earlier analysis about state-level election outcomes. As shown in the top portions of the panels, on average, Democrats were much more likely than Republicans to expect Biden to win the Electoral College (a mean difference of 0.388) and the national popular vote (a difference of 0.439). As the middle portion of the panels shows, the effect of living in a Trump versus Biden state is more muted for these outcomes: those in Biden states were more likely to think Biden would win nationally, but for both outcomes the differences are much smaller than for the state-level outcome, roughly one-sixth the size of that difference (0.077 for the Electoral College and 0.068 for the popular vote). Finally, the bottom portion of panels B and C shows that the effect of state of residence is present for Republicans but largely absent among Democrats. More directly, regardless of where people live, on average Republicans (Democrats) are more likely to report they think Trump (Biden) will win both the Electoral College and popular vote. Interestingly, Republicans are more optimistic for the Electoral College, and Democrats for the popular vote, a pattern consistent with what happened in 2016 and evidence that neither party appears, on average, more biased in its forecasts.

In Appendix Figure A8 we replicate these estimates after limiting our sample to respondents in battleground states. We find similar partisan differences as in the overall sample, but which candidate won the state has little predictive power: the average scores in close states won by Biden and close states won by Trump are both near 0.50 (difference = 0.035, 95 percent CI = (−0.104 to 0.034), p = 0.32). (These small differences persist when simultaneously accounting for both state winner and panelist partisanship.) If residents in pivotal toss-up states have great uncertainty in identifying who will win their state, it should not surprise us that the mass public should have difficulty in accurately predicting the winner of the Electoral College, which also turns on the outcomes of those closely contested states, and also that the state-level outcome in one's home state will be of little use in forecasting the national outcome in a close election.

In contrast to the large effect of state context on expected state-level outcomes, the analysis in Appendix Figures A1 and A2, which repeats Figure 2 analysis for forecasts of the Electoral College and the national popular vote, respectively, shows state context had small effects on respondents' expectations or uncertainty about those outcomes. Given the small effect of state of residence on these outcomes in the aggregate analysis shown in Figure 1, these results are not surprising. That is, even in states where people are more confident in their state-level forecast, they remain about equally uncertain in forecasting the national election outcome.

Implications and conclusion

Researchers have noted the rapidly increasing levels of nationalization of American elections (e.g. Hopkins, 2018) and the heightened levels of polarization in nearly all aspects of political life (e.g. Iyengar et al., 2019). This has led to widespread concern that Americans are disconnected both from facts generically and their state-level political environment. We show that despite these factors, and while Americans may engage in expressive partisanship when predicting elections (or have true partisan differences in beliefs), particularly at the national level, survey respondents are reasonably knowledgeable about their state political environments. That is, they do not rely solely on partisanship or national factors to make inferences about their political surroundings. While some recent studies have suggested voters make poor forecasts in relatively uncompetitive states (e.g. Murr and Lewis-Beck, 2020), our results are consistent with those studies that show voters are relatively accurate when forecasting the winner at the state-level when the state outcome is not expected to be too close, and in recognizing which states are likely to be close.

On the one hand, these results suggest reason for optimism. We find that partisanship's influence on political perceptions is limited, and so despite partisan pressures, average citizens understand and report differences across state political environments. Indeed, for most Americans, the outcome of the 2020 election was met with acceptance⁸, perhaps because for most people the outcome matched their expectations for their state or resolved underlying uncertainty. Furthermore, robust electoral competition can be associated with greater perceptions of fairness (Wolak, 2014). However, the facts that even people in uncompetitive states understood the Electoral College would be close and that partisanship is correlated with forecasts for the national outcome mean many people may across the country have been surprised by the outcome. That is, Republicans and Democrats still report expected national outcomes that correlate with partisanship. This pattern may reflect true partisan differences in beliefs or hopeful expressive responding in light of uncertainty about the actual election. Although these two qualities are not worrisome per se, in an era when faith in democratic institutions appears weak, true and unexpected disappointment in electoral outcomes could erode the weak basis of support, and among the small subset of Americans who engaged in post-election violence, these concerns appear real.