Short-term effects: immediate responses to the outcome

The establishment of an election outcome signifies the culmination of a lengthy electoral process, concluding months of campaigning and determining who will assume power. This decisive moment not only identifies the winners and losers but also establishes the trajectory for political decision-making in the years ahead. Such a significant event is likely to attract widespread attention, triggering spontaneous reactions among voters through systematic and heuristic processing.

Systematic processing may contribute to immediate responses in two ways. First, prior to the solidification of the result, some voters may have carefully considered how each outcome might affect the state of democracy. However, since systematic processing tends to be influenced by motivated reasoning (Chaiken and Ledgerwood Reference Chaiken and Ledgerwood2012; Taber and Lodge Reference Taber and Lodge2006), voters are likely to have downplayed indications that their favored candidate could harm the proper functioning of democracy while remaining receptive to signs that the opponent might pose a threat (Kunda Reference Kunda1990). As a result, winners and losers are expected to enter election day with diverging perceptions, and once the outcome solidifies, adjust their satisfaction with democracy accordingly.

Second, even without candidate-specific biases, motivated reasoning may influence how voters respond to the outcome. Winners, perceiving the outcome as positive, are inclined to find justifications for its legitimacy, leading them to believe that the political system, in producing this result, functions as it should. Conversely, losers, dissatisfied with the outcome, become prone to question its output and, in turn, look for flaws in its functioning (Kunda Reference Kunda1990; Taber and Lodge Reference Taber and Lodge2006).

When responding to the result, voters may also rely on simplistic cues, such as their emotions and partisan identity. For instance, as winning and losing trigger distinct emotions, winners, feeling positive about the outcome, are inclined to appreciate the performance of the political system, boosting satisfaction with democracy. In contrast, losers, experiencing negative emotions, are inclined to feel more disturbed over their functioning, leading to more negative evaluations. ^{Footnote 1} Additionally, by adopting heuristics such as ‘my own group can be trusted’, winners may infer that the political system will be managed by reliable in-group actors who respect democratic norms. Conversely, losers applying this same heuristic may see the political system as becoming controlled by less trustworthy out-group actors, heightening concerns about the state of democracy.

The timing at which the outcome solidifies and triggers these voter responses may vary from case to case. In non-competitive elections, the result may be foreseeable during the campaign, enabling voters to anticipate the result early on. Still, due to the many factors that may quickly change public opinion, it is likely unusual for voters to feel entirely certain about the outcome before the official announcement, even when polls indicate a favorite. Therefore, attitudinal responses should typically be expected on election night when preliminary results are reported. However, in situations without a clear winner – where parties must negotiate to form a coalition – the impact may not be visible until these negotiations conclude and an agreement is publicly announced. ^{Footnote 2}

The logic behind these mechanisms is equally applicable to individuals experiencing repeated victories or losses. For instance, current winners are expected to experience happiness and relief upon securing another electoral win, while also being prone to seeking justifications for the accuracy of the result. This discussion leads to the first hypothesis:

Short-Term Hypothesis: As soon as an election outcome is established, winners’ satisfaction with democracy increases, while losers experience a corresponding decline.

Long-term effects: navigating the post-election landscape

In the post-election period, immediate short-term effects are likely to be magnified by various long-term influences. Following the establishment of an election outcome, voters enter a complex political landscape, encountering an abundance of new information related to the election result – from retrospective campaign analyses and expert forecasts to elite statements and peer reactions. As voters learn more about the implications of the outcome, they continue to update their attitudes through systematic and heuristic processing. While some voters may have a clear sense of the outcome’s consequences from the start, leading to only minor attitude adjustments, many will likely gain a deeper understanding gradually, leading to an extended period of attitudinal polarization.

Motivated reasoning fosters biased information processing and selective exposure to arguments. Biased information processing implies that individuals subconsciously perceive arguments supporting their position as stronger and more convincing than counterarguments, while simultaneously investing more time and cognitive resources in opposing arguments that contradict their stance (Kunda Reference Kunda1990; Taber and Lodge Reference Taber and Lodge2006). As voters engage with elite statements, expert analyses, and peer reactions in the aftermath of an election, winners, optimistic about the outcome, are anticipated to be more easily convinced by positive interpretations and analyses. Conversely, when confronting incongruent information, winners are expected to be more critical and allocate more time and cognitive resources to counterargue it. This pattern also holds true for losers, leading to a gradual widening of attitudinal differences as new information is processed after the election (Taber and Lodge Reference Taber and Lodge2006).

Meanwhile, selective exposure implies that individuals are prone to actively seeking out information that aligns with their pre-existing views. Numerous studies illustrate how this behavior prompts voters to mainly seek out news and information that reinforces their existing values (e.g. Arceneaux and Johnson Reference Arceneaux and Johnson2013; Goldman and Mutz Reference Goldman and Mutz2011; Prior Reference Prior2007). Following an election, winners can be expected to turn to sources interpreting the outcome more positively, offering analyses that project a bright future. In contrast, losers likely resort to sources providing more critical interpretations, presenting negative coverage of the new political landscape. This information-seeking behavior leads winners to encounter and accumulate more positive interpretations than losers.

In addition to motivated reasoning, heuristic processing is likely to contribute to further winner–loser polarization following the establishment of the election results. For instance, as voters learn more about the new government’s policy platform, winners may experience additional positive emotions if policies align with their interests, while losers experience heightened frustration, anger, or anxiety. Cue-taking can also be expected to play a particularly prominent role at this stage. This route to attitude formation is especially common when individuals face new and complex situations, forcing them to rely on the judgment of others due to limited knowledge (Murakawa and Gilens Reference Murakawa and Gilens2022).

In the uncharted post-election landscape, voters are anticipated to receive cues from elites and peers regarding the implications of the outcome. When encountering elite cues, voters tend to place greater trust in signals originating from in-party candidates and party-affiliated media outlets (Clayton and Willer Reference Clayton and Willer2023; Daniller Reference Daniller2016; Lelkes Reference Lelkes2016). As signals and statements about future prospects typically are more encouraging and optimistic from winning elites compared to those of the losing camp, it is anticipated that elite cue-taking will contribute to the widening of the ‘winner–loser gap’. While potentially less pronounced, voters may also discern elite skepticism regarding electoral integrity. Senninger et al. (Reference Senninger, Bækgaard and Seeberg2023) find that even in Denmark, one of the world’s most robust democracies, losing candidates express greater concerns about electoral fairness than their winning counterparts. Although Fortin-Rittberger et al. (Reference Fortin-Rittberger, Harfst and Dingler2017) find that actual instances of electoral fraud do not widen the ‘winner–loser gap’ – and may even reduce it – the gap could deepen if defeated elites signal unfounded concerns about electoral integrity to their supporters.

Moreover, as social networks tend to form around shared values, beliefs, and political ideologies (Huber and Malhotra Reference Huber and Malhotra2017; McPherson et al. Reference McPherson, Smith-Lovin and Cook2001), it is likely that a greater proportion of winners’ closest peers are also winners, and vice versa for losers. Consequently, cue-taking from trusted peers in the post-election period is likely to reinforce initial winner–loser attitudes, contributing to further polarization (Levendusky et al. Reference Levendusky, Druckman and McLain2016). ^{Footnote 3}

In summary, the posited long-term effects are driven by voters’ engagement with new information after the election outcome is established. As voters navigate the information-rich post-election environment, they are inclined to biased information processing and partisan cue-taking from elites and peers. This is expected to gradually increase the ‘winner-loser gap’ in the post-election period, leading to the formulation of the second hypothesis:

Long-Term Hypothesis: The ‘winner-loser gap’ grows gradually following the establishment of the election outcome.

Data

The study relies on data from the ESS, rounds 1–10, covering 21 elections. ^{Footnote 4} These are all listed in the online Appendix A, where each election is described and the date of the emergence of the final election result is also uniquely identified. To serve as an example, the first case is the Italian general election held in April 2006. The incumbent Prime Minister Silvio Berlusconi led the center–right coalition ‘House of Freedoms’, which lost against the center–left alliance ‘the Union’, led by Romano Prodi. With over 38 000 000 votes, only 24 000 separated the two blocks, making the winning margin less than 0.1%. On the 11^th of April, Prodi claimed victory, but Berlusconi wanted a recount of some districts and therefore did not admit defeat. On the 19^th of April, the Italian Supreme Court ruled that Prodi had won the election (BBC, 2006). Since the election result was settled by the court, and they are the ones who announced the final outcome, I use the 19^th of April as the cut-off point.

Most of the elections included in the study are ‘least likely’ cases, in the sense that they were held in consolidated democracies with proportional systems for electing their executives. In such settings, the ‘winner–loser gaps’ are typically smaller (Anderson and Guillory Reference Anderson and Guillory1997). However, notable exceptions exist. For instance, Greece in 2009 used a system that awarded the winning party 50 ‘bonus’ seats in parliament, effectively resembling a majoritarian system. ^{Footnote 5}

Figure 1 reveals that the correlation between supporting a party that is currently in opposition and feeling less satisfied with democracy is strongly present within the sample. Only in Estonia and Slovenia, the relationship is missing. The overall ‘winner–loser gap’ is highly statistically significant, and the estimate remains significant even when controlling for conventional confounders, such as age, gender, education, satisfaction with the state of the economy, political interest, and left-right self-placement (see Table B1, Appendix B). This is reassuring, as it suggests that if the short- and long-term mechanisms indeed explain the emergence of the gap, it should be evident in the sample.

Figure 1. Satisfaction with democracy among losers (L) and winners (W), by country. I follow the standard procedure and code those in support of the executive at the time as winners and those in support of the opposition as losers. Observations: 18162 (data retrieved from the European Social Survey).

Variables

The independent variable is whether a person voted for a winning or losing party. I follow the conventional approach and code those who state that they feel closer to a party that ended up in government as winners and those who ended up in the opposition as losers (e.g. Curini et al. Reference Curini, Jou and Memoli2012; Dahlberg and Linde Reference Dahlberg and Linde2017). ^{Footnote 6} Those without a preference for any party are excluded from the sample.

The main dependent variable is satisfaction with democracy. This is measured by the question ‘And on the whole, how satisfied are you with the way democracy works in [country]?’ The respondents could position themselves on an 11-point scale, from 0 (extremely dissatisfied) to 10 (extremely satisfied). While this measure clearly omits several aspects of system support, it is commonly acknowledged to reflect voter approval of the democratic process (Anderson and Guillory Reference Anderson and Guillory1997). Consequently, it is expected to capture something more fundamental about voter system support than a mere evaluation of the incumbent. It is also the survey item most commonly used in the literature on the ‘winner–loser gap’ (see, for instance, Anderson et al. Reference Anderson, Blais, Bowler, Donovan and Listhaug2005, Han and Chang Reference Han and Chang2016).

To control for election-specific effects, the basic model includes dummy variables for country and survey round. At the individual level, I add the standard controls for age (years), gender (male/female), and education (years of education). I also include a range of controls which have been shown to impact people’s evaluation of the political system (Anderson et al. Reference Anderson, Blais, Bowler, Donovan and Listhaug2005). Respondents with missing values in a given analysis were excluded using listwise deletion.

Estimating the short-term effects

The short-term effects are examined using a regression discontinuity in time design (RDiT), which estimates the difference of a value immediately before and after a known temporal cut-off point (Best and Wolf Reference Best and Wolf2015; Hausman and Rapson Reference Hausman and Rapson2018). In this case, it is applied to estimate the difference in system support among winners and losers from the day the election outcome is established to the subsequent day.

As the running variable, I use the dates that the individual interviews were conducted. Aside from the election outcome becoming known, there is no reason to expect the winners’ and losers’ system support to be a discontinued function of the dates of the interviews. Assuming that all other factors relevant in explaining system support are continuous on the day that the election outcome is settled, any positive jump in system support among the winners (or negative among the losers) can be attributed to the causal effect of this event (Angrist and Pischke Reference Angrist and Pischke2009).

I expect the risk of treatment and self-selection bias to be marginal. While the ESS theoretically could make a biased allocation of the treatment (inviting different types of voters to participate in the survey before and after the election outcome becomes known), in practice, their fieldwork is conducted independently of these events, and they use random sampling to get a representative selection of respondents (ESS Sampling Guidelines, 2018). Therefore, there is no reason to suspect there to be a biased difference between the voters they approached before and after the cut-off points.

Perhaps of greater concern is the risk of self-selection bias. For instance, due to a lack of time, some groups of voters might abstain from participating in the survey while the election campaign is still running. Figure 2 illustrates the distribution of observations, indicating a relatively consistent pattern with no substantial evidence of sorting around the threshold. This is also confirmed when running a McCrary test, making a density plot before and after the cut-off (see Figure C1, online Appendix C). ^{Footnote 7} To further ensure robustness, Figure 3 presents balance tests using multiple bandwidths. ^{Footnote 8} These follow the recommendations by Muñoz et al. (Reference Muñoz, Falcó-Gimeno and Hernández2020), covering standard variables such as age, gender, education, and left-right self-placement, as well as ESS paradata on the number of contact attempts made before each interview was completed and the frequency of refusals to participate.

Figure 2. Histogram of 28 282 observations spread over the running variable (data retrieved from the European Social Survey).

Figure 3. Balance tests: covariate differences between treated and control using different bandwidths, with 95% confidence intervals.

With the largest bandwidth, some imbalances are observed among winners and losers. Notably, in both instances, respondents in the treatment group required more contact attempts to complete the survey while being less likely to refuse participation. However, these patterns do not seem to stem from a sharp discontinuity at the cut-off point. Rather, they appear to reflect a trend where later survey participants tend to require more contact attempts and show fewer refusals. When a lagged dependent variable is included to account for gradual temporal effects, as recommended by Hausman and Rapson (Reference Hausman and Rapson2018), these imbalances lose statistical significance (not reported). Moreover, as the bandwidth narrows, the observed imbalances diminish among winners and losers. At the 16-day bandwidth, the only imbalance remaining significant is that losers in the treated group are slightly more likely to report holding citizenship.

The analysis relies on the 16-day bandwidth, which is the most balanced and provides the most rigorous test for causal inference. Using the standard criteria of 80% power, power tests confirm that this bandwidth offers sufficient statistical power to detect shifts in system support of less than 0.4 standard deviations, closely following the criteria set by Muñoz et al. (Reference Muñoz, Falcó-Gimeno and Hernández2020). However, all bandwidths are tested, with the widest bandwidth offering over a 99% chance of detecting shifts as small as a quarter of a standard deviation (see Figures B1–B2, Appendix B). At each bandwidth, I also test including controls for any previously observed imbalances, but these adjustments do not impact the results.

To ensure that the results are not biased by losers being less inclined to participate in the surveys post-defeat, I also run models assessing the number of winners and losers participating before and after the threshold (see Table C2, Appendix C). Here, the groups appear well-balanced across all bandwidths.

To prevent gradual long-term effects from biasing the short-term estimates and to address potential autoregression, I follow the recommendations of Hausman and Rapson (Reference Hausman and Rapson2018). Specifically, I verify that the estimates remain robust both when including a lagged dependent variable from the previous day as control and clustering standard errors at the country/survey round level (Hausman and Rapson Reference Hausman and Rapson2018, pp. 19–22).

I examine the responses of winners and losers separately, letting winners who participated before the results were presented serve as a baseline for winners who participated when the results were known, and correspondingly for the losers. I do this employing the ‘RDrobust-package’ (Calonico, Cattaneo, and Titiunik Reference Calonico, Cattaneo and Titiunik2015), which makes data-driven statistical inferences, constructing two local linear regressions above and below the threshold, where the treatment effect is measured as the difference between the two intercepts. By default, I use robust confidence intervals for the average treatment effect at the cut-off point, unless alternative specifications are explicitly mentioned (Calonico et al. Reference Calonico, Cattaneo and Titiunik2015). Finally, I use the Epanechnikov kernel function, which is commonly applied in RDs, because it minimizes the least mean squared error at the cut-off (Al-Razzaq et al. Reference Al-Razzaq and Mohammad2023), though uniform and triangular are also tried with similar results.

To facilitate the interpretation of the interaction effect between being a winner/loser and knowing the election outcome, the results for the winners and losers are presented in separate figures. The RDiT design is thereby specified as:

(1) $${y_i}=\alpha+\beta 1\left( {{D_i}} \right)+\beta 2\left( {{D_i}\; \times Day{s_i}} \right)+{{X_i}^\prime}\gamma +{\epsilon_i},$$

where ${y_i}$ is the level of satisfaction with democracy for winner (or loser) i, $\alpha $ is the intercept, representing the expected outcome when the running variable (Days) is zero, $\beta 1\left( {{D_i}\;} \right)$ is the treatment effect, where D _i is a dummy variable that takes the value 1 for winners (or losers) if the election outcome has been established, and 0 if it has not, $\beta 2\left( {{D_i}\; \times Day{s_i}} \right)$ is the effect of the running variable, which measures how many days before or after the election results were presented that a respondent participated in the survey. The interaction term ${D_i}\;$ allows for different slopes before and after the election outcome has been decided. ${{X_i}^\prime}\gamma $ is the effect of the covariates represented by ${{X_i}^\prime}$ , and $\gamma $ is the vector of coefficients associated with the covariates. ${\epsilon_i}$ is the error term, representing unobserved factors affecting the outcome that are not included in the model.

Estimating the long-term effects

The long-term effects are anticipated to unfold gradually in the post-election period. Examining this pattern necessitates a model that captures changes in satisfaction with democracy over a continuous timeframe rather than at a single moment. Ideally, I would have been able to assess the development of the ‘winner–loser gap’ day by day. However, the daily number of observations is too few for such estimates. Instead, I organize observations within the same week and analyze the evolving attitudes of winners and losers week by week. Employing each week as a dummy variable enables me to plot the temporal progression of winner–loser attitudes with 95 per cent confidence intervals for each week. This analysis is conducted using the following interaction model:

(2) $${y_i}=\alpha+\beta 1Week+\beta 2Winner+\beta 3\left(Week\times Winner\right)+{{X_i}^\prime}\gamma +{\epsilon_i},$$

where ${y_i}$ is the level of satisfaction with democracy for individual i, $\alpha $ is the intercept, $Week$ and $Winner$ are dummy variables representing the number of weeks before or after the outcome solidified and the winner–loser status, $\beta 1$ and $\beta 2$ are the coefficients for the main effects of $Week$ and $Winner$ , respectively. $\beta 3$ represents the coefficient for the interaction term $Week \times Winner$ . ${{X_i}^\prime}\gamma $ is the effect of the covariates represented by ${{X_i}^\prime}$ , and $\gamma $ is the vector of coefficients associated with the covariates. ${\epsilon_i}$ is the error term, representing unobserved factors affecting the outcome that are not included in the model.