Influences on Campaign Perceptions of Closeness

Campaign operatives’ perceptions of electoral closeness are likely to be affected both by contextual and individual-level psychological mechanisms. We assess two key contextual hypotheses. First, we expect to see a difference between campaign workers and candidates associated with incumbent campaigns and those associated with challenger or open-seat campaigns. We link the observation by many political scientists that incumbents strongly fear losingFootnote ²⁹ to findings in psychology and behavioral economics demonstrating that people think differently about potential losses than about potential gainsFootnote ³⁰ and that fear leads to pessimistic assessments.Footnote ³¹ Along with this pessimism, we also expect incumbents, on average, to be more accurate than challengers. Incumbents and their associates have more to lose than challengers, since their jobs are on the line. Furthermore, the uncompetitive nature of many races means that many challengers are low-quality challengersFootnote ³² and may not have the competences or resources to have the same information about their district or the election as incumbents. Altogether, we expect incumbents to be better assessors of the likely outcome than challengers.

A second contextual hypothesis relates to the information environment of the race, which is primarily a function of the type of election at stake and where the election is conducted. Because presidential elections are so thoroughly inundated with polls, we take it as uncontroversial that presidential campaign operatives ought to have a clearer understanding of the closeness of their election than campaign operatives in ‘down-ballot’ races should have about their contests. Similarly, congressional campaigners ought to perceive their competition better than state house campaigners. Notably in down-ballot races, media attention and the availability of polls is often a function of whether an election is in a presidential battleground state. In general, any elections that are perceived as either particularly important or particularly close will be given more attention by the media, and so campaign elites should have more accurate assessments of closeness in these kinds of contests.

As important as contextual considerations are the individual-level dispositions of elite campaigners. First, we expect a floor effect on campaign operatives’ expectations. Invested as they are in their efforts, workers may exercise motivated cognition that makes them unwilling to accept defeat. The worst outcome they may expect is a 50/50 contest. Thus, while a campaigner that expects to win may predict a wide range of outcomes, a losing campaign might insist, until the end, that the race’s worst-case-scenario is a near tie.

Second, we expect that campaign operatives of different backgrounds will be differentially equipped to predict electoral closeness. In the down-ballot races we study, where resources are often limited and staffers are few, variation across individuals is potentially important. Volunteer-activists and professional staffers may see elections in different lights. While both may be motivated by an ideological commitment to their candidate, professional staffers are more likely to have a financial or career stake in the electoral contest. We expect that their experience in politics and financial interest make them more accurate assessors of outcomes. Professional staffers may also have access to private information about the state of the campaign to which volunteers are not privy. As a corollary to this hypothesis, we also expect that campaign workers who claim to be ideologically extreme (e.g. Democratic staffers who report they are very liberal on an ideology measure) ought to be less accurate in their assessments, as they might be less inclined to admit defeat. This is because ideologues may experience more dissonance or distress if their candidate loses and thus may be unable to assess their standing accurately.

Third, we expect the operatives’ level of sophistication to affect their perceptions. Campaign operatives vary substantially in their educational backgrounds. This could matter because predicting an outcome in an election requires considerable sophistication. Moreover, the simple act of offering a number (like the expected vote share in a race) requires a degree of numeracy that might be more common among well-educated workers than poorly-educated ones. Thus, we expect that campaign operatives who are more sophisticated, proxied by their educational attainment, are better at assessing their chances.

At the end of the article, we will explore the implications of the findings of our analysis. But here, it is worthwhile to emphasize that these hypotheses offer insights about the kinds of offices and kinds of political operatives who understand when elections are meaningfully contested. Knowing what types of operatives are good at predicting outcomes helps us understand the circumstances in which political campaigns operating in a world of landslide elections will behave as if the stakes are high.

Predicting Closeness by Electoral Environment

We first approach the contextual hypotheses about the kinds of workers who will be better or worse at predicting accuracy. Here, we focus on basic two-variable and three-variable analyses; later in the article, we will engage in model-based multivariate analysis. We first turn to Figure 2, which shows the down-ballot races in the sample with an eye toward the effect of incumbency. The x-axis measures the average prediction of the Democratic vote share for each campaign. The y-axis measures the actual Democratic vote share in the race. Here, we collapse multiple respondents by race in order to derive an average prediction for the entire campaign. For the modal down-ballot race, we interviewed just one or two respondents during the campaign cycle. There are a few outliers – closely contested Federal Senate and House races – in which we interviewed a relatively large number of respondents in the race. The graph uses different symbols for campaigns associated with incumbent politicians and those associated with non-incumbent politicians. The graph also displays several reference lines. The plot is divided into four quadrants representing predicted victory (NW and NE) versus predicted loss (SW and SE) and actual victory (NE and SE) versus actual loss (NW and SW). The 45° line represents perfect predictions. Dots above that line indicate optimistic predictions and dots below the line indicate pessimistic predictions.

Fig. 2

Incumbent campaigns are more pessimistic predictors Note: Each point represents the mean value of all respondents associated with a particular candidate’s campaign. The name of each campaign cannot be revealed. The plot incorporates eighty-nine non-incumbent campaigns and thirty-eight incumbent campaigns. Among incumbent campaigns, 50 percent over-predict their vote share; among non-incumbent campaigns, 93 percent over-predict their vote share.

Compared to non-incumbents, incumbent campaign operatives are both better at predicting their outcome (i.e. there is a better linear fit between predictions and outcomes) and less likely to overstate their likely vote share. Non-incumbents nearly always overestimate their vote share. Notice, incumbency is correlated with winning, but the propensity to under-predict election outcomes is particular to incumbents who win. If we focus on just the actual winners (points in the NE quadrant), 55 percent of incumbents predict a closer election than the final outcome demonstrates whereas only 25 percent of non-incumbents predict a closer election than the final outcome demonstrates (difference of means two-tailed t-test significant with p-value of 0.04). This result is consistent with the notion, put forth by Fenno, that incumbent legislators operate in a state of fear of being in a close election even when they are not. It is also consistent with the expectation from psychology research that risk aversion is associated with pessimism.Footnote ³⁴

Figure 2 is also informative in two other ways. First, notice the stark absence of data points in the southeastern quadrant. There is not a single campaign in our dataset in which an eventual winner expected to lose. In contrast, the northwestern quadrant is populated by the plurality of campaigns that thought they would win but ended up losing. Second, notice evidence of a ‘floor effect’. No matter their eventual outcome, most of the campaigns that ended up losing expected a 50/50 race. There is a concentration of campaigns along the 50 percent horizontal line. The data are consistent with the idea that campaign workers have a hard time admitting they will lose (i.e. motivated cognition). If signs point to a loss, they might hold out hope, as the Romney campaign apparently did, that they could win in a ‘squeaker’.

We now compare the accuracy of predictions across congressional races (both House and Senate) and state races. The state races consist primarily of races for state legislature. Our attention to the differences in accuracy across different offices stems from our hypotheses about the election environment. Namely, races are generally given more attention as they involve higher ranks of elected office. As a result, we suspect congressional staffers to have much more information they can bring to bear on their estimate of the likely outcome, perhaps leading to a more accurate assessment of the race. Higher offices are also more coveted, possibly leading to operatives devoting more resources to understanding the state of the campaign. This too can lead to more accurate assessments among operatives for higher level campaigns.

Figure 3 shows the absolute deviation between a respondent’s prediction and the eventual outcome for Federal House and Senate and for state races. Campaign workers in federal races are almost 4 percentage points more accurate than workers in state races, a statistically significant difference. We further subdivide this grouping of races into volunteers and professional staffers. In down-ballot races, there is not a significant difference between the perceptions of volunteers and the perceptions of staffers. In fact, even high-level staffers and candidates in down-ballot races are not that much better than down-ballot volunteers at predicting vote share. The highest level staffers (i.e. campaign managers) and candidates are actually slightly worse at predicting the outcome than lower-level staffers, but they are slightly better than the volunteers.Footnote ³⁵

Fig. 3

Workers for higher offices are generally better predictors Note: 95 percent confidence intervals are displayed. The y-axis plots the average absolute deviation between a worker’s prediction of the campaign’s expected vote share and the actual vote share in the election.

A second take on the information environment requires that we look at races that were perceived as toss-up races in comparison to those that were perceived as safe. We expect campaigns operating in races treated as toss-ups to be better informed about the likely outcome than campaigns operating without much informational feedback. Figure 4 shows this hypothesis generally holds. In Figure 4, we show two measures of competitiveness. The first is a pre-election division of races into those that are toss-ups, those that are leaning toward one party, and those that are safe.Footnote ³⁶ The second measure of competitiveness is simply the actual election outcome in the state or race. As Fraga and Hersh note, using the election outcome to study competitiveness is quite common in political science research and has some advantages over ex ante measures of competitiveness.Footnote ³⁷

Fig. 4

Predictions are more accurate in competitive races Note: 95 percent confidence intervals are displayed. Pre-election definitions of toss-up, lean and safe races were taken from http://www.realclearpolitics.com. Only federal races are measured in the left graph (due to data availability of pre-election media estimates of closeness). All races are included in the right graph.

Using the ex post competitiveness measure, as expected, campaign operatives were better at assessing the outcome in races that were closer. In the left graph, the toss-up down-ballot races exhibit the best predictions, but workers in lean races were not more accurate than workers in safe races. Unlike the right plot, the left plot only includes federal races. Among these, the safest are often races associated with incumbent legislators and, as we saw in Figure 2, campaign operatives in these races perceive their election more accurately than operatives in other kinds of races.

So far, we have shown that, consistent with our expectations, incumbent campaigns are both more accurate and more pessimistic at predicting their election outcome than open-race or challenger campaigns. We have also shown that races that are competitive and for higher offices are generally associated with better predictions. With these findings in mind, we now turn to a similar examination of operatives from the presidential campaign. While these workers are obviously less directly responsible for making campaign decisions as those from down-ballot races, an analysis of Obama workers can help us to understand the underlying mechanisms for accuracy: most of these workers operate in a high-information environment, where both the media and the campaign have invested substantial resources in understanding the state of competition. However, as we will demonstrate, these workers still show remarkably similar biases to those observed among down-ballot races, indicating that information effects are only part of story.

Accuracy of Predictions in the Obama Campaign

In Figure 5, we display the responses of Obama workers as we did for down-ballot workers in Figure 1. Recall that for the Obama campaign, respondents were asked about the state in which they were working, not the national race. Obama workers demonstrate a similar pattern to down-ballot workers, with a tendency to over-predict their candidate’s chances. The mean assessment among Obama workers was 8.45 points off the actual outcome in their state (95 percent CI 8.05–8.86), so Obama workers were on average actually slightly worse predictors than down-ballot campaign operatives, despite operating in a high-information environment.

Fig. 5

Summary histogram: Obama workers are over-optimistic, but generally accurate predictors Note: N=2,583. Vertical line at x=0 indicates perfect prediction.

In Figure 6, we show how the predictions of Obama workers are also a function of context. Their responses are related to competitiveness in the expected manner, with competitive states having the most accurate operatives.

Fig. 6

Predictions are more accurate in competitive races Note: 95 per cent Confidence Intervals are displayed. Pre-election definitions of toss-up, lean and safe states were taken from http://www.realclearpolitics.com.

Accuracy of Predictions among the Mass Pubic

Is the optimism we witness a characteristic particular to campaign workers or is optimism a feature of partisans in general? To investigate this, we replicated our closeness question on mass surveys, asking respondents in the 2012 CCES and CCAP how close the presidential race would be in their state.

We display the responses of CCAP and CCESS respondents, by party, in Figure 7. Members of the mass public over-predicted their candidate’s probability of victory, and they were less accurate than the workers. Democrats in the mass public over-predicted Obama’s vote share (mean deviation 18.74, 95 percent CI 17.56–19.92), and the Republicans in the mass public thought their side would do better than it ended up doing (mean deviation 15.19, 95 percent CI 14.01–16.37). Independent voters, who could probably evaluate the expected outcome without their judgement being clouded by partisan loyalties, ended up over-predicting and under-predicting the Democratic vote share about evenly (mean deviation 16.41, 95 percent CI 15.31–17.51). Thus, campaign operatives are much more accurate than members of the mass public, but partisans both among the mass public and among campaign operatives seem to over-predict their chances. This suggests that one reason for the over-prediction among campaign operatives is simply partisan-motivated reasoning. In the supporting material (Figure A.3), we replicate Figure 7 while restricting the sample to highly-educated respondents who are either very liberal Democrats, very conservative Republicans, or moderate independents. Except for the moderate independents, who have balanced predictions, high-information partisans over-predict candidates’ vote share in a manner similar to campaign operatives, demonstrating the powerful effect of partisanship in motivating responses.

Fig. 7

Partisans in the mass public are over-optimistic predictors and less accurate than campaign operatives Note: Data come from a combined sample of two 1,000 person modules of the 2012 CCES and 2012 CCAP. Observation counts for this survey item are 702 Democrats, 476 Republicans, and 631 Independents. Vertical line at x=0 indicates perfect prediction.

Individual-Level Correlates of Accuracy

Turning to individual-level correlates of campaign operatives’ predictions, we focus on their educational background, ideological extremity, and professional status. We have hypothesized that hyper-ideologues ought to be less accurate than moderately ideological workers, as their judgement may be clouded by their political commitments. We have also hypothesized that better educated workers and paid staffers will have more resources and incentives, both cognitive and informational, to gauge the closeness of their contest. As we assess these individual mechanisms, we also engage with larger statistical models that account for the timing of the interview during the campaign season as well as traits of workers’ races, such as the factors of incumbency, closeness, and level of race.

In Table 1, we separate Obama campaign workers from down-ballot campaign workers, and we show three models for each group. The variable college is an indicator variable in which those with a college degree or who appear to be in college have a value of 1.Footnote ³⁸ Paid staff is an indicator representing staff as well as candidates. V. liberal is an indicator distinguishing respondents who consider themselves very liberal from those who chose other options on a seven-point ideology scale (nearly all respondents claimed to be moderately liberal, liberal or very liberal). The Num. of polls in state is a variable representing the total number of presidential polls taken in the state over the course of the fall campaign. The variable relates directly to polling information to which campaign workers have access through the mass media and is thus a representation of the information environment.Footnote ³⁹

Table 1

Which Campaign Workers Accurately Predict Election Outcomes?

Note: Dependent variable is |prediction – outcome|. Standard errors are in parentheses. ***p<0.01, **p<0.05, *p<0.1. Models 1–3 employ robust clustered standard errors, where the state of Obama workers is the clustering variable. Models 4–5 cluster standard errors around the specific down-ballot campaign of the worker. For an alternative specification, Model 6 employs state fixed effects instead of clustered standard errors. The number of presidential polls in the state is collinear with the set of state indicators in Model 6.

The next variable listed in Table 1 represents the weekly change in Obama’s advantage in each state based on state-specific polls. For example, if Obama held 55 percent of the two-party support in a state poll in the first week of October, but 50 percent of the two-party support in the second week of October, then respondents filling out the survey in the second week of October would have a value of −5 on this variable. For down-ballot races, incumbent equals 1 if the respondent is working for a campaign of an incumbent Democrat, and federal race distinguishes races for US Senate and US House from state races. Finally, we include indicator variables for interviews taking place before September, in September, and in October. November interviews are the omitted category.

To aid in the interpretation of Table 1, consider Model 4. According to Model 4, down-ballot campaign workers who were non-college educated and volunteers were, on average, 11.07 points off in their estimate of the expected outcome, as represented by the intercept coefficient. Notice, because smaller deviations between the respondent’s estimate and the outcome represent more accurate responses, negative values in the independent variables indicate higher levels of accuracy. The coefficients on College and Paid Staff indicate that college educated workers were 4.3 points more accurate and paid staff were 1.8 points less accurate.

Across all models in Table 1, education bears a strong relationship to accuracy. The coefficient suggests that well-educated workers are more accurate at assessing electoral closeness.Footnote ⁴⁰ For Obama workers, there is also a consistent effect of being a paid staffer. However, similar to the discussion of volunteers in Figure 3, paid staffers are not better assessors of closeness on down-ballot campaigns relative to volunteers. Table 1 also shows no effect of being very liberal on accuracy. Contrary to our expectation that ideologically extreme respondents would be unable to observe the state of the race clearly, ideology bears no relationship to accurate assessments when other variables are included in the model. Despite ideology seeming to have some effect on responses in the mass public, within campaign operatives, very liberal respondents are no less accurate than others.

In Models 2 and 3, we add a measure of the information environment. The table shows that the more presidential polls conducted in the state, the more accurate are the respondents’ assessments. This variable is also correlated with more accurate predictions in down-ballot races, as shown in Model 5. The variable is absent from Model 6 because Model 6 incorporates state fixed-effects, which are collectively collinear with this variable. Notice that measuring the change in Obama’s state advantage does not significantly alter the coefficients for the individual-level variables, indicating that campaign workers are doing more than just reporting signals in their perceptions of closeness; rather they are also likely drawing on private information and responding to their own psychological motivations. Regarding Models 5 and 6, notice that perceptions are more accurate among operatives on federal races versus state races and elites working for incumbent campaigns versus non-incumbent campaigns. This is consistent with the bivariate analysis above.

It is worth lingering on measures of temporal change. After all, the GCP sample was collected over a period of months ahead of the election. But time, it turns out, is not a strong predictor of accuracy. The way that time is incorporated into Table 1 is through dummy variables for month of interview. Perceptions of closeness were more accurate in the first week of November than at any other time, particularly than in October. But perceptions of closeness were about as accurate in November as in the summer. Other specifications of time indicate similar minimal effects. In Figure A.1 in the supporting material, we show the time trend by itself, for Obama workers and down-ballot workers. There is essentially no movement in the prediction accuracy for Obama workers and a curvilinear movement for down-ballot workers, with better predictions in late August and November, consistent with the coefficients in Table 1. Additionally, we have modeled time in several different ways, including measuring workers’ predictions by weekly changes in Obama’s national poll standing, and by incorporating linear time trends into the models in Table 1. We have also tried restricting the sample just to paid staff. Even among paid staff, there is no relationship between time of interview and accuracy. No matter how we slice it, perceptions of accuracy seem to be driven more by individual-level and race-level measures than by the time of interview. This again indicates that rather than simply relying on public signals of closeness, which should be more accurate as the campaign progresses, campaign operatives are largely influenced by their own information and biases.