Introduction

Incumbents make unpopular policy decisions all the time. Proposing an unpopular policy is often met with public pressure like protests or opposition resistance. Some incumbents will implement the proposalFootnote 1 nevertheless, while others decide to withdraw the unpopular proposal. For example, President Macron’s government decided to implement the widely unpopular pension reform, despite an enduring and large wave of protests in 2023. On another occasion, Macron withdrew a proposed fuel tax raise after months of ‘gilets jaunes’ protests in France in 2018.

Which of these two strategies—following through and implementing an unpopular policy proposal or withdrawing it—is electorally beneficial? In this paper, we ask whether voters reward a responsive withdrawal of an unpopular policy proposal or punish the disclosed policy intention despite the withdrawal.

We draw on a broad literature on retrospective voting that has extensively worked on how voters react to policies and policy processes. We refer to retrospective voting as circumstances when voters observe policy actions, attribute responsibility for these events to the incumbent, evaluate the incumbent’s performance based on their policy actions, and then vote (Healy & Malhotra, 2013, p. 289). What voters consider informative for their vote has been extensively studied, including seemingly irrational reasons (Achen, 2016; but see Ashworth, 2018), as well as calculating ones. In terms of the latter, for example, good economic performance is assumed to reveal competence (e.g., Alesina & Rosenthal, 1995; Fearon, 1999), which is electorally rewarded (Aytaç, 2020; Duch & Stevenson, 2006; Hopkins & Pettingill, 2018; Larsen et al., 2019). Empirically, most studies on retrospective voting test how voters evaluate popular policy outcomes. Research finds that policies providing public goods increase incumbents’ chances of being re-elected (Adiguzel, 2022; Burnett & Kogan, 2016; Golden & Min, 2013; Kogan, 2020). Hereby, voters are assumed to reward responsive policies for which the incumbent has taken into account popular demand.

Among these rich insights, two research areas remain mostly overlooked. First, few studies investigate voters’ evaluations of unpopular policy outcomes. While voters might reward popular policies, it is not apparent whether they would punish unpopular ones (e.g., public goods removal (see Nyholt, 2024)). Second, due to the focus on policy outcomes, little attention has been paid to the policy process that leads to an outcome. By and large, a critique by Fiorina is still applicable today: “In essence, such a model [of retrospective voting] presumes that the citizen looks at results rather than the policies and events which produce them” (Fiorina, 1978, p. 430).

Yet, policy events (the policy process) and results (the policy outcome) can reveal different information. Understanding whether voters care about the policy process or the outcome is important because there are many cases when process and outcome do not align. If a policy gets withdrawn, voters are left with the information about what the incumbent wanted to implement (the policy proposal) and what the actual outcome was (withdrawing the proposal). We argue that proposing a policy discloses information about the incumbent’s policy intention, whereas withdrawing a (un)popular policy shows that an incumbent is (un)responsive to citizens’ demands. We introduce two types of voters who, in the case of an unpopular policy, either care about the revealed policy intention or the actual outcome and responsiveness.

We test how voters react to unpopular policies that are either withdrawn or implemented in the case of school closures in Sweden. Since the 1990s, over one thousand schools have been proposed to be closed in Sweden (Uba, 2020). In most cases, a municipality’s proposal to close a school is met with community efforts, such as local protests and petitions, to keep the school open (Uba, 2016; Taghizadeh, 2015). At times, citizens succeed in their goal and prevent school closure: Politicians withdraw the initial proposal and agree to keep the school open.

The school closures provide a good case to test our argument as, first, schools matter tremendously to Swedish voters (Solevid, 2009) and we have good reasons to expect that voters follow the whole policy process and care whether a proposal was implemented or not.Footnote 2 Second, school closures are unanimously unpopular in Sweden. This makes sure that we test the effect of an unpopular policy process and that we can expect no party to gain from implementing such a policy. Third, the data allow us to distinguish whether policy proposals were implemented or withdrawn among a universe of highly similar policy proposals. Thus, in contrast to high-profile cases where one reform idea is not implemented, we are able to gather data about similar policies across time and space within the same country. This gives us some leverage for our causal identification strategy.

Empirically, we exploit geographical and temporal variation in school closure proposals from 2002 to 2018 across the 6000 electoral precincts in Sweden. The granular level of analysis and panel structure of the data enables us to account for variation between municipalities, common shocks to all precincts within a municipality, and time-invariant differences between precincts. We match an originally compiled data set of voting outcomes in Sweden at the level of precincts with data on geo-located schools and fine-grained information concerning the policymaking process spanning from proposal to actual school closure gathered by Uba (2016) and Folke et al. (2024).

Our results suggest that in a precinct where a school was proposed to be closed, incumbent parties lose on average one percentage point in vote share if they close the school as well as if they withdraw the proposal to close the school. This finding suggests that voters care about the disclosed policy intention, even if it is not implemented. The effect sizes are substantial for incumbent parties as voting at the local level in Sweden happens in the margins and the effect size is bound by a party’s share in a coalition. Additionally, local elections are closely tied to national elections that occur on the same day, making an electoral shift at the local level even more notable.

The findings contribute to a deeper understanding of retrospective voting and electoral accountability. We show that responsiveness in the sense of withdrawing an unpopular policy is not electorally rewarded, which stands in contrast to previous research emphasizing the importance of responsive policies (e.g., Ashworth, 2012; Boas et al., 2021; Fearon, 1999). Our study highlights the importance of politicians’ policy promises and proposals (see also Elinder et al., 2015; Matthieß, 2020). From an applied perspective, the findings reveal a dilemma for incumbents when making unpopular policies. Once an unpopular idea is suggested, withdrawing it again is unlikely to remedy negative electoral consequences.

Information Cues and Retrospective Voting

We propose two voting types that emerge depending on which information cue voters take into account during the policymaking process of an unpopular policy. The types differ with regard to whether voters consider a policy intention or policy outcome as a relevant cue for their vote. Thus, we expect diverging voting patterns when intention and outcome differ (i.e., when a policy proposal is withdrawn) and expect both types to electorally punish the implementation of an unpopular policy, as policy intention and policy outcome converge. A scope condition is that both voters care about the policy and evaluate the policy as unpopular. In this sense, they consider revealed information cues (e.g., Ashworth et al., 2018) and are retrospective by considering a policymaking process that has happened before their election (see Healy & Malhotra, 2013, p. 289).

We draw on a long line of voting research. A common assumption is that voters base their electoral choice on the little information they have at hand—previous experiences of government activity and performance (see Downs, 1957, p. 45). When choosing an incumbent during an election, voters compare the anticipated performances of different candidates (or parties) to select the candidate from which they expect to gain the most benefits (see further Key, 1966, p. 2).

Anticipated and past performances are discussed as two different sets of information that characterize prospective (Artés et al., 2022; Elinder et al., 2015) or retrospective voting (e.g., Healy & Malhotra, 2013). In the case of prospective voting, voters are assumed to consider future, not past, policymaking. Thus, when voting prospectively, voters take note of a promise for future policy, rather than a past policy outcome. Elinder et al. (2015) find evidence for prospective voting in Sweden by showing that voters responded to policy promises, rather than the previously implemented policy. While we consider similar information cues, our proposed voter types differ from prospective voting as we consider the policymaking process from initial proposal to implementation. Thus, voters look back at the outcome of a policy proposal that has either been implemented or withdrawn.

Our theoretical intuition concerning which information voters consider for an election relates to work by Fearon (1999), who distinguishes between information gained from the policy process and the policy result. Either, voters reward a consistent and principled policy process, or a responsive and popular outcome (Fearon, 1999, p. 56). We adopt the distinction between policy process and policy outcomeFootnote 3 as two information cues but deviate from Fearon’s model by adding that incumbents reveal a policy intention during the policymaking process.

In the following, we outline two potential mechanisms with contrasting implications for incumbents’ electoral support. We argue that when looking back at a policymaking process, voters gain two sets of information. On the one side, the policy proposal indicates an incumbent’s policy intention. On the other side, the policy outcome shows which responsive policy output the incumbent produced during the term.

Policy Intentions and Voting

We assume that proposing a policy reveals information on the incumbent’s policy intentions to voters (see Ashworth, 2012). A proposal is a concrete first step during a policymaking process that goes beyond the mere promise of a policy. It shows which kind of policy the incumbent is willing to implement and thereby signals that similarly unpopular policies are likely to be suggested during the next election term.

Voters have several reasons to mostly take into account the incumbent’s policy intention, despite the eventual policy outcome. To begin with, even if an unpopular proposal is withdrawn, voters have little guarantee that a similar (and unpopular) proposal will not be made again. Since the unpopular policy was proposed in this specific precinct, this choice could reflect the government’s overall lack of effort for this neighborhood (Folke et al., 2024) and increase the chance of similar proposals in the future. Likely, the unpopular policy was avoided by participating in collective action, protesting, and bringing public opinion to the government’s attention (Taghizadeh, 2016; Uba, 2016). However, these activities are costly efforts that citizens aim to avoid for upcoming policymaking processes.

Moreover, an unpopular policy proposal suggests to voters that their opinion was considered only to a limited extent during the proposal stage. It seems that the planning process shared between politicians and the bureaucracy did not involve a thorough consultation of citizens’ opinions. In case it did, public opinion seems to have been miscalculated or ignored. This again, adds to the perception that the government values the neighborhood and its citizens less than other parts of the municipality.

Thus, to avoid similar policy struggles in the future and punish the incumbent for their unpopular idea, voters are likely to consider the first policy proposal over the later outcome of the policy process. Following this line of reasoning, a policy withdrawal does not alleviate the negative effect of the initial unpopular proposal that discloses the incumbent’s policy intention. We formulate the following hypothesis:

H1

If voters mostly consider incumbents’ disclosed policy intention, incumbents are more likely to lose votes when proposing an unpopular policy, even when the policy is not implemented.

Responsive Outcomes and Voting

We assume policy outcomes to reveal something about the responsiveness of an incumbent (Caughey & Warshaw, 2018; Gilens, 2005; Hogan, 2008). Depending on which policy gets implemented incumbents demonstrate whether they care about the public’s opinion. In the case of unpopular policy proposals, an implemented policy signals little responsiveness, while a withdrawn policy implies a responsive policy outcome.

We expect voters to value a responsive policy outcome over the disclosed policy intention for several reasons. From a voter perspective, the outcome of a withdrawn unpopular proposal is a popular result of the policymaking process. Even if the initial proposal was unpopular, the government eventually implemented a popular decision. This means that the incumbent acted responsively to citizen demands. The process of taking back a policy proposal signals to voters that a similar process is possible in future policymaking when a proposal turns out to be unpopular. Thus, voters gain the experience that they have the means to participate at a political level and affect policy outcomes in their favor.

Moreover, that citizen demands were taken into account is shown by the fact that a policy withdrawal involves several costs for the incumbent. The incumbent needs to publicly acknowledge that the proposed policy will not be implemented. In some cases, this acknowledgment amounts to an apology, which can be a risky political maneuver. Further, incumbents responded to citizen demands despite planning and budgetary costs. Anticipated savings cannot be made, which at times will require a re-organization of the budget.

If voters mostly consider policy outcomes rather than politicians’ revealed policy intention, we expect incumbents to be electorally punished if they implement an unpopular policy, but not if they withdraw this policy and prevent an unpopular policy outcome. Here, the policy outcome weighs more for an election than the unpopular proposal. Thus, in our second hypothesis, we formulate the following expectation:

H2

If voters mostly consider incumbents’ responsive policy outcome, incumbents are more likely to lose votes when implementing an unpopular policy, but are less likely to do so when the policy is withdrawn.

Research Design

To test our contrasting expectations for two types of retrospective voting, we analyze electoral outcomes in municipal elections (every four years) in Sweden from 2002 to 2018 at the precinct level. We exploit within municipality differences in electoral outcomes between precincts over time and compare election results for precincts before a proposal to close a school was made to those after the proposal was either implemented or withdrawn. In this section, we first describe the case of school closures in Sweden, then our data, and finally the empirical strategy.

The Case: School Closure Proposals in Swedish Municipalities

The process of school closures in Sweden offers an interesting case to study voters’ responses to different information cues during the policymaking process. School closures are highly salient issues in local elections in Sweden and often generate significant contestation, providing a suitable context to examine electoral responses to proposals that are implemented or withdrawn.

Whether a school is closed in Sweden is decided at the lowest administrative unit, the municipality. The 290 Swedish municipalities have organizational and financial responsibility for the educational system. They have extensive autonomy from the central government, to the extent that local self-government is protected in the constitution (Lidström, 2010). School closures are unpopular among parties across the political spectrum. Instead of ideological reasons, there are often budgetary reasons why a school is supposed to be closed.

Municipal incumbency is reflected by who sits in the municipal council, which is the main political decision-making body in a municipality. Members of the council are elected for four-years on the same day as national elections. After an election, usually a majority coalition is formed and the leading coalition party appoints the chairperson of the executive board. This board has the overarching responsibility for all municipal activities (Lidström, 2010). The chairperson represents the leading politician in a municipality, similar to the position of a mayor in other countries.Footnote 4

The process to close a school is set in motion by members of the council who request a proposal for the closure of a school from the municipal bureaucracy (Uba, 2016). The proposal is discussed by the municipal committee for educational issues. At this point, the proposal becomes public, which often leads to protest or contestation (Taghizadeh, 2016; Uba, 2016). At times, this contestation results in a policy withdrawal, the school is kept open and the budget is reallocated to finance the open school. In the remaining cases, the school is eventually closed.

Among voters, school closures are unpopular for several reasons. For residents of the municipality, it might poorly reflect on an incumbent’s fiscal competence since it is a cost-saving measure for the municipality and implies that the municipality’s budget has not been well managed. For residents in the affected precinct living in proximity to the closing school, in comparison to the rest of the municipality, school closures might be particularly unpopular. First of all, education and service provision of schooling is found to be a highly salient issue for Swedes (Solevid, 2009). Additionally, parents lose an essential public infrastructure for their children. Relocating a child to a new school often implies that daily commuting times increase, as the new school is likely situated farther away and a change in environment can create stress for the family. Property owners face a decrease in the value of the living area and surrounding residents experience changes in who passes by the area and how safe the environment feels.

The high voter turnout in local elections (e.g., 84.1% in 2018) reflects the considerable influence (and possible drawbacks) political decisions at the municipal level have on citizens’ lives. These local elections are polled at the level of precinctsFootnote 5 within a municipality. Each precinct contains 1000 to 2000 eligible voters that go to the same poll and are therefore the most fine-grained level at which election results are published.Footnote 6 Following Folke et al. (2024), we argue that precincts are the most suitable level of analysis to capture local voting within a neighborhood. Each municipality consists of, on average, 20 precincts,Footnote 7 adding to around 6000 precincts across Sweden.

Data

The data include precincts with schools that are proposed to close. We first outline our process for compiling electoral results at the precinct level over time. Then we introduce the data concerning the school closure policymaking process, as well as the composition of municipal assemblies in Sweden.

Electoral Outcomes in Swedish Precincts

For electoral outcomes, we use data from the Swedish Election Authority, which provides election results for all municipal elections along with geodata for precincts. To ensure comparability over time, we match precincts from different election years based on the geographical area they cover. This allows us to follow the same areas and their voting outcomes over time more accurately.Footnote 8

In the main analysis, we consider precincts from two consecutive elections to be comparable if they have at least a geographical overlap of 90%.Footnote 9 Using this method allows for minor changes to the borders of a precinct from one election to the next, since the quality of geodata files across years or minor changes often result in differences. Following this method, about 80% of all precincts match to a similar precinct in the previous election, and 31% of all precincts that existed in 2018 can be observed across all five elections between 2002 and 2018.Footnote 10 In the main analysis, we only include schools proposed to close if they are located in a precinct that has remained unchanged for at least three elections but also run the analysis using other samples as robustness checks (see Table A1).Footnote 11

The outcome variable is the difference in vote shares for the party incumbent when the proposal of school closure was made in the precinct and the rest of the municipality, expressed in percentage points.Footnote 12 Thus, the outcome variable measures how the election results in the affected precinct deviate from the rest of the municipality. Our approach is to estimate how this deviation changes after the proposal and the decision to either close the school or withdraw the proposal, compared to before the proposal was made.

Using our definition of the outcome variable has two main advantages. First, variation in the outcome cannot arise from common shocks that affect the entire municipality equally. Second, by subtracting the aggregated vote share from all other precincts in a municipality, we can follow the same areas over time, even if the borders between unaffected precincts change.Footnote 13

School Closure Policy Process

Data on school closure proposals and implementation are matched to the election data based on the year of the policy proposal and the address of the school.Footnote 14 The data covering the policymaking process of school closures were generously shared by Uba (2016) for the period 1990 to 2009 and by Folke et al. (2024) who extended the original data set by Uba to 2018.Footnote 15 We exclude cases of schools that were proposed to close if the school is located in the same precinct that already had an affected school in a previous election cycle.Footnote 16 However, we merge two school cases if two schools were proposed to be closed during the same election cycle and either both closed or kept open during that election cycle.Footnote 17

We measure whether a school remains open and whether the policy was withdrawn in the absence of school closure. After a proposal, the decision to close a school is made within two years of the initial proposal in a majority of cases. For a school that is indicated to remain open until the end of the data collection period, we assume that it is evident that the school remains open or has a high chance of staying open by the end of the election cycle. We consider this a reasonable assumption as elections occur every four years. The biggest threat to our estimation would be if the proposal is made before an election and a reversal happens after the following election, in response to the election result. This could either happen if the same party remains incumbent or if there is an incumbent turnover. To address these concerns we re-run our main analysis only using cases where there has been no turnover since the proposal was made (see Table A3) and when excluding all proposals made during an election year (see Table A4).

The final estimation sample consists of 599 schools that have been proposed to be closed. Out of these, 382 schools (63%) were eventually closed while the rest remained open until the end of the data collection period. The average time between the proposal and the decision to close a school is 1.73 years and for 106 of the closed schools, an election was held between the proposal and the final decision to close the school. Figure 1 depicts the distribution of school closure proposals across the years in our main estimation sample. The y-axis depicts the number of proposals per year of which some will be eventually implemented (dark gray) or withdrawn (light gray). Interestingly, the graph suggests that local governments tend to propose school closures the year after an election and refrain from making proposals during election years (dashed line). Overall, Fig. 1 shows a steady pattern of proposals to close schools, implemented school closures, and proposal withdrawals.

Fig. 1
figure 1

Distribution of made proposals across the time frame of our estimation sample

Full size image

Besides temporal variation, there is also a substantial variation in the spatial distribution of school closure proposals. Out of the 290 municipalities across Sweden, 213 are included in our main estimation sample. In Fig. 2 we present their spatial distribution by shading the precincts across Sweden that have experienced at least one school closure proposal. The thicker dark line indicates the borders of municipalities in Sweden. Naturally, there are more proposals in more densely populated areas where the sizes of the precincts tend to be smaller. However, the map shows that school closure proposals are common across Sweden.

Fig. 2
figure 2

Shaded areas represent precincts that have experienced at least one school closure proposal in our main estimation sample. The lines show the borders of municipalities in Sweden

Full size image

Incumbent Party in Municipal Assemblies

The party affiliation of the chairperson of the executive board in a municipality is used to determine the incumbent party for all years and in all municipalities. This executive board has the overarching responsibility for all municipal activities and the chairperson usually represents the largest party in the government coalition. Since the governing coalition has control over the political agenda, it is reasonable to ascribe the main responsibility for school closures to the chairperson’s party. To identify incumbent parties in each municipality, we use data compiled by Broms (2022). In Table A9 we show that proposals from different parties have a similar likelihood of ending in closure or withdrawal.

Empirical Strategy

In the analysis, we exploit variation in the outcome variable over time in precincts where a school was proposed to be closed. We make the identifying assumption of parallel trends between the affected precincts and the rest of the municipality. We assume that had the proposal not been made, the vote share for the incumbent party would have followed the same trend in the affected precincts as in the rest of the municipality. This assumption allows us to causally estimate the electoral consequences of implementing or withdrawing school closure proposals even if proposals do not occur randomly.

We run several checks of the parallel trends assumption. First, we follow convention and run a pre-treatment analysis to see if there is a trend in our outcome variable before the proposal is made. Table 1 shows the estimate from an event-study regression (Clarke & Tapia-Schythe, 2021), where the first election following the school closure proposal is defined as period 0 and the election before the proposal (period -1) is used as the baseline. The estimate shows that none of the coefficients for the negative time periods are statistically significant, whereas after the proposal they are negative and statistically significant at the 5 percent level or below. These results provide evidence that the estimated effects are not mainly driven by trends in the outcome variables that already existed before the proposal was made. Second, we re-run our main specification using precinct-specific linear and quadratic time trends. Table B2 shows only minor changes in the estimates and results remain highly comparable.Footnote 18

Table 1 Event study, before and after proposal
Full size table

We look into further threats to identification, which we explain in more detail in the robustness test section and the Appendix. One threat is spill-over effects that could bias our results if voters living outside the affected precinct also react to the proposal. We address this possibility in Tables D3 and D4 and find slightly weaker effect sizes when excluding voters living close to but outside affected precincts. We discuss three potential explanations for these findings in Appendix D.

Relatedly, school closure proposals that were not included in the analysis could affect our results. Some affected schools were excluded as they were located in precincts of which the borders have frequently changed. Although these proposals were excluded, election results from these precincts are still used when we subtract the vote share in the rest of the municipality from the vote share in the treated precincts. While a single precinct in a municipality is unlikely to affect the average vote to a large extent, we make sure to estimate a standard difference-in-differences approach, where precincts unaffected by any proposal are included separately. This yields similar results to our main estimation (see discussion in Appendix B and Table B1).

A last threat to consider is a changing voter composition within precincts affected by school closure proposals. We address this issue by accounting for population change and demographic trends for three election years, see Tables A8, C1, and C2. We find no indication that our results are driven by demographic trends.

Analysis

We present our main results in Table 2. We include precinct fixed effects in all our regressions to account for time-invariant factors that explain differences in voting between affected precincts and the rest of the municipality. Further, as our outcome variable measures the deviation in election results from the rest of the municipality, any time-varying but common shocks within municipalities are accounted for in the analysis. We cluster the standard errors at the precinct level to account for the panel structure in the data.

In column 1, we first regress the outcome variable on a single indicator that takes the value 1 if the school closure proposal has been made and 0 otherwise. This estimation gives us the average effect on voting for the incumbent party in elections following a school closure proposal, regardless of the actual policy outcome. As we are interested in the reaction of voters when the proposal is either implemented or withdrawn, we further run a similar regression that includes three separate binary indicators for proposal, closure, or withdrawal in column 2. In this specification, the first indicator takes the value 1 if the proposal but not yet the final decision has been made. The second indicator takes the value 1 if the decision to close to school has been taken and finally, the third one takes the value 1 when there has been a withdrawal of the initial proposal. The estimated effect of withdrawing a proposal should be interpreted as the net effect of both proposing and withdrawing the proposal. This gives us our quantity of interest, as we are interested in the question if the total effect of first proposing an unpopular proposal and then withdrawing it is negative or not.

Column 1 shows that a party loses votes in precincts where they propose to close a school. On average, the incumbent party loses 0.9 percentage points of their votes in the affected precinct, when accounting for voting in the rest of the municipality and including precinct fixed effects. Note that the coefficient is based on the average across all future elections following the proposal that we observe.Footnote 19

To test our hypotheses, we turn to column 2. Here, we expected retrospective voters to electorally punish school closure and only voters who mostly value the policy intention to also punish a policy withdrawal (H1), whereas voters who value the responsive outcome to refrain from punishing the incumbent in case of policy withdrawal (H2). We include the proposal indicator for cases where there is an election between the initial proposal and the actual decision to close the school, to exclude these from the control group.

Results indicate that when the decision to close a school is made, voters living near the school punish the incumbent.Footnote 20 On average, the incumbent party receives 1.1 percentage points fewer votes in the affected precinct after the decision to close a school. We believe this estimate is substantive, especially in close elections. The difference between the two main political blocs (left and center-right) in Swedish politics in the 2018 local election was less than 1.1 percentage points in about 5% of the municipalities. In 25% of the municipalities, the margin between the blocs was less than 5 percentage points.

Even in case the proposal is withdrawn and a school remains open, incumbents lose on average 0.8 percentage points of votes.Footnote 21 Thus, regardless of the policy outcome (closed or open school), incumbent parties lose electoral support once they publicly propose to close a school.

Table 2 Main results
Full size table

To further analyze the dynamics over time, we examine how the effects depend on the number of elections that have occurred since the decision to either close or withdraw the proposal. Figure 3 displays the estimates for the decision to close a school and the proposal withdrawal, separated by the number of elections since the initial proposal. The plot demonstrates that the negative effect of deciding to close a school remains stable over election years. The decline of 1.35 percentage points in the vote share for the incumbent party in the first election after deciding to close a school persists in subsequent elections, indicating a long-lasting consequence on the voting choices of affected individuals. In cases where the proposal is withdrawn, there is a clear negative response from voters during the first election following the withdrawal. After that, uncertainty increases around the estimates but our results do not suggest any positive shifts in electoral outcome.

Fig. 3
figure 3

Coefficient plot from a regression where the effect of deciding to close a school or withdrawing the initial proposal is separated depending on the number of elections since the initial proposal was made. Dependent variable: Difference in vote share for the party proposing to close a school in the precinct and the rest of the municipality, expressed in percentage points. The inner horizontal lines represent the 90 percent confidence intervals (CIs) and the outer lines are the 95 percent CIs. The full regression output is presented in Table B3

Full size image

Robustness tests

To ensure the robustness of our results, we conducted several tests, which are presented in Appendices A to D. The robustness tests support the reliability and validity of our findings. For some tests, uncertainty around the estimates increases. However, as we discuss in further detail in the Appendix, sample sizes decrease considerably in some models, and point estimates remain highly similar. We therefore have good reasons to conclude that our results remain robust.

First, we examine the robustness of our data construction approach by changing the criteria for a precinct to be included in the estimation sample (Table A1), excluding proposals where the incumbent party has changed during the policymaking process (Table A2) or any time after the initial proposal (Table A3), excluding proposals that have been made during an election year (Table A4), excluding precincts in which two schools were proposed to be closed during the same election cycle (Table A5) excluding proposals that occurred before the first observation in our data for that precinct (Table A6), separately analyzing proposals from the two largest parties (Table A7), and when considering voter mobility (Table A8).

Second, we check our regression specification and use a more standard difference-in-differences approach where we include all precincts that are unchanged for at least three consecutive elections (Table B1). This specification allows us to test for alternative explanations by including precinct-specific linear and quadratic time trends (Table B2). Third, we test for compositional shifts of voters in precincts affected by school closure proposals (Table C1) and how our results hold when controlling for demographic variables measuring the mean annual income, share of highly educated, and share of foreign-born at precinct-level (Table C2). Fourth, we adopt an alternative approach for matching precincts across time (Tables D1 and D2). We further use this sample to test for spill-over effects (Tables D3 and D4).

Discussion and Conclusion

By distinguishing between policy withdrawal and implementation, we find that after proposing to close a school, voters not only punish incumbents for closing a school but also for withdrawing the proposal and keeping a school open. In both cases, the incumbent party loses in the elections following a school closure proposal, on average, one percentage point of the vote share in the affected precinct. We argue that our results can be interpreted causally. We check that results are not driven by voting trends prior to the proposal or a change in the voter composition in affected precincts. Further, the results are robust when testing alterations of the parallel trends assumption between affected precincts and the rest of the municipality as well as several alternative regression specifications.

The results are substantial, as local election wins in Sweden are made in the margins, and differences around one percentage point can make or break an election for the left or right-leaning bloc. It is further important to consider that the effect size is bound to the incumbent party’s share in a coalition where the ruling coalition usually consists of three to four parties. For example, a loss of one percentage point for a party with 20% of the votes would imply that the party lost 5% of its votes. Additionally, voting in Swedish local elections is closely related to national and regional elections on the same day. Voters tend to choose the same party in the national and local elections (Lidström, 2021), an electoral shift based on a local issue is therefore notable.

This paper contributes to a larger debate in democratic theory concerning whether voters hold their government accountable, (e.g., Downs, 1957; Healy & Malhotra, 2013) and if so, which information they consider (e.g., see Fearon, 1999; Fiorina, 1978). First of all, the results show that voters consider policymaking in retrospect. Closing a school and thereby removing a public good is electorally punished for up to three election periods (12 years). By finding negative electoral consequences of public goods removal, this study, in turn, adds novel insights to a growing consensus that the provision of public goods or allocation of resources is electorally beneficial for incumbents (e.g., Adiguzel et al., 2022; Burnett & Kogan, 2016; Kogan, 2020).

Second, we advance the understanding of which information cue voters consider from the policymaking process. While we cannot test mechanisms at the individual level, our findings at the precinct level show a negative electoral effect of withdrawing an unpopular policy proposal, which suggests that voters rather consider the disclosed policy intention than a responsive outcome. This questions what form of responsive policymaking voters reward (see further Ashworth, 2012; Boas et al., 2021).

It is important to mention that interpreting the negative withdrawal effect is limited by a lack of evidence that voters punish a withdrawn policy proposal because it signals a government’s policy intention. It might also be the case that withdrawals go mostly unnoticed and receive much less media attention and voters who notice the unpopular proposal don’t update their beliefs in case of a withdrawal. Additionally, even if they notice the withdrawal, voters might be biased to only retain the negative information of the proposed school closure and not the positive information that a school will remain open (this relates to asymmetric accountability, see Naurin et al., 2019).

While more research is needed to test the mechanism we propose in this paper, all three alternative explanations underscore the importance of proposals as information cues. This relates to work on policy intentions in the form of promises or pledges as cues for voting (e.g., Elinder et al., 2015; Matthieß, 2020) but stands in contrast to research finding voters to discard promises as unreliable information sources (Ellger et al., 2023; Thomson et al., 2017). Future studies could explore whether proposing a popular policy is as sticky in voters’ memories as proposing an unpopular one. Finding that our results do not replicate to popular proposals would provide evidence for a negativity bias among voters and mean that our case is a particularly hard case to find voters rewarding responsive policy withdrawals.

Another explanation could be that voters electorally punish a policy withdrawal because they are dissatisfied with an inconsistent policy process and not, as we suggest, because they care more about the disclosed policy intention than the responsive withdrawal. Yet, our results offer mixed evidence for the importance of consistent policymaking as such a model additionally predicts that a principled and consistent process is rewarded (see Fearon, 1999). In our case, this would be if the unpopular proposal is implemented and a school is closed. We do not find evidence for the latter and further research is needed to test whether the logic of consistency extends to unpopular policies.

Overall, the results point towards a difficult challenge for incumbents. Regardless of the actual outcome of the policy proposal, voters will electorally punish an unpopular proposal. This circumstance suggests that it is crucial for governments to first think carefully about which proposals to make. However, at times, making an unpopular proposal might be inevitable. This study raises the question of whether the negative effect of an unpopular proposal can be mitigated by other activities. Here, research on responsive policymaking and procedural fairness during the policymaking process offers an interesting starting point (e.g., Esaiasson et al., 2016; Grimes & Esaiasson, 2014). While this study is limited in uncovering such variation in participatory policies, future research could test whether communication efforts and citizen participation during the policymaking process increase citizens’ approval of a policy. Additionally, the long-lasting effects of unpopular policies on voting prompt the question of whether we would observe similar, but positive effects in the case of a popular policy. If voters do not reward such popular policies to the extent they punish unpopular ones, the challenge for incumbents to avoid such negative repercussions becomes even more pressing.