1.1 Personal experience and political attitudes

The idea that personal experience matters for political attitudes lies at the core of the economic voting literature (Duch and Stevenson Reference Duch and Stevenson2006). For voters who are less politically engaged, personal experience represents an easy and cheap way of acquiring information that, in turn, might affect their voting behavior or, more generally, their attitudes. Here, we take advantage of disaster data and referendum results, which allows us to study political behavior and how it is affected when voters are exposed to extreme weather events.

Indeed, several papers link personal experience to attitudes on climate change. Stokes (Reference Stokes2016) finds that citizens living in proximity to wind energy projects punish the incumbent government for its climate policy. Brody et al. (Reference Brody2007) note that vulnerability to floods and rising sea levels affects the perception of risk associated with global climate change, whereas temperature has no such impact. Recent papers explore the impact of weather conditions on attitudes toward global warming (Li et al. Reference Li, Johnson and Zaval2011; Egan and Mullin Reference Egan and Mullin2012; Konisky et al. Reference Konisky, Hughes and Kaylor2015; Hazlett and Mildenberger Reference Hazlett and Mildenberger2017). In particular, Egan and Mullin (Reference Egan and Mullin2012) find that an unusual increase in local temperatures strengthens belief in global warming. Their estimates indicate the magnitude of such an increase is substantial, though it survives only in the short-term.

While it provides interesting results and compelling insights, we identify three shortcomings in the existing literature. First, although Erikson and Stoker (Reference Erikson and Stoker2011) and Egan and Mullin (Reference Egan and Mullin2012) are important exceptions, results from previous papers are plagued by the fact that personal experience is not randomly assigned among individuals.

Moreover, previous studies rely on surveys to capture personal experience. This creates several well-documented problems, such as measurement error in self-reporting and political bias in answering questions about personal experience (Achen Reference Achen1975; Bartels Reference Bartels2002). Without linking attitudes to an exogenous shock that is likely to modify opinions on environmental issues, it is difficult to draw conclusions from survey data based on self-reports.

Finally, and importantly for our paper, previous studies capture attitudes toward global warming by relying on surveys among a non-representative sample of the population (Li et al. Reference Li, Johnson and Zaval2011) or a representative sample (Egan and Mullin Reference Egan and Mullin2012). While this approach is surely effective in measuring political attitudes, social scientists are ultimately interested in actual political outcomes. This is particularly important in the case of global warming, where attitudes seem only weakly affected (if at all) by external shocks. By exploring the impact of natural disasters on actual voting in referendums, our study is able to show not only if external shocks change political behavior on climate ballot measures, but also how these effects vary across units and time.

1.2 Natural disasters and political behavior

There is a compelling literature that links natural disasters to political behavior. The theoretical framework is, again, economic voting: Rational voters reward incumbents not only for delivering positive economic performance in good times, but also for organizing prompt rescue and relief programs in bad times, as in the case of a hurricane or major flood.

This literature is heavily grounded in American politics. In a pioneer study, Abney and Hill (Reference Abney and Hill1966) explore voter response to the rescue program after Hurricane Betsy in New Orleans. Chen (Reference Chen2012) shows that disaster relief after the Florida hurricane increased votes for George W. Bush in the 2004 election, but only in Republican precincts. Healy and Malhotra (Reference Healy and Malhotra2009) draw political economy implications from such findings. Since voters reward relief programs, politicians have an incentive to invest in relief aid and under-invest in preparedness measures. Thus, what is an electorally efficient policy turns out to be economically sub-optimal. Finally, Gasper and Reeves (Reference Gasper and Reeves2011) show how the US electorate rewards/punishes the requests/denials for federal assistance.¹

The take-home message from this literature is twofold. First, relief programs, if effective, carry a sizable electoral reward for incumbents and their parties. Second, such a reward is usually short-lived, since voters are quick to forget. This second result squares with the literature on blind retrospective voting with the important exception of the recent contribution by Bechtel and Hainmueller (Reference Bechtel and Hainmueller2011). They provide the most sophisticated analysis on the effect of natural disasters on political behavior outside the US and find that voter gratitude lasts longer than claimed by previous studies.

These studies have convincingly shown the importance of natural disasters on voter behavior, treating relief programs as a form of pork barrel spending in bad times.Footnote ² However, there is another channel through which natural disasters might affect (rational) voter behavior. Given that events such as floods are associated with global warming and climate change, voters who have directly experienced a natural disaster might form new opinions on the salience of the climate issue. In turn, changing such opinions might shape voting behavior on issues related to climate change. To the best of our knowledge, no study has yet explored this channel and this is where we aim to make our contribution to the literature.

3.1 A behavioral measure of climate change attitudes

For the behavioral measure, we focus on voting behavior of villages on ballot measures related to climate change.Footnote ⁵ We identify a number of ballot issues that can be labeled as environmental issues.Footnote ⁶ But not all environmental votes are linked to climate change or global warming.

To select specific votes, we rely on the official government information brochures that are sent to each citizen before a vote on a referendum or initiative.Footnote ⁷ In each brochure, we search for the following keywords to identify votes that can be connected to climate change: emissions, climate, air pollution, exhaust emissions, global warming, greenhouse, and fuel consumption.Footnote ⁸ Based on this, there are nine ballot proposals for which emissions and climate change mattered.

One vote is not included: the 2003 twin-initiative to abandon nuclear power. In this campaign, the center and center-right parties were fighting against a nuclear ban, while the left and the ecological groups supported the ban. This is a highly unusual case because the groups supporting stronger protection of the climate wanted to ban nuclear power while the usual non-environmentalist groups (center, center-right, and the government) embraced climate protection in their argument.Footnote ⁹ While both sides used the climate argument, it was more prominent in the government's position. The problem is that voting for or against the issue is not a clear measure of pro-climate behavior.

Nevertheless, including the 2003 twin-initiative does not change the general inferences—while effect size is somewhat smaller, all significance tests yield the same outcome. This provides us with eight ballot measures from 1998 up to 2009 (see Table A.11 for a full list of all votes used). Finally, the government comprises an oversized coalition of the four or five largest parties, which typically represent 80 percent or more of the citizens. Citizens usually vote four times a year and their vote is based on party recommendations or policy preferences. It is not unusual for a vote not to pass, and therefore the government does not resign in such cases (Kriesi Reference Kriesi2005).

3.2 Measuring exposure

The natural disaster data gathered by the WSL includes all publicly recorded natural disaster events. Data collection is based on more than 1000 newspapers and magazines. The database identifies three different principal events: floods and debris flows, rockfall, and landslides. We focus on medium and large events during the period 1995–2010 (Hilker et al. Reference Hilker, Badoux and Hegg2009). These are events that cause an estimated damage of 400,000 CHF (about $400,000 or euro 340,000) or more.

In Figure 1 all events are plotted. The red dots show the natural disaster events that took place within the 12 months preceding a national environmental vote that was climate-related. We use this cut-off since it allows for long effect decay. We also show later in this paper that 12 months after a natural disaster, there is no longer a difference between affected and non-affected municipalities (see Table 4).

Figure 1. Map of Swiss municipalities natural disaster events (1995–2010).

Notes: Natural disaster events (WSL 2012). Red dots show events that occurred 12 months or fewer prior to a vote related to climate change.

These events are not independent of the topography. From Figure 1, it can be seen that there is an impressive clustering in flat areas along lakes and at the bottom of valleys. This is because floods and mudslides usually occur along rivers and lakes or at the bottom of hills and mountains (e.g. Eng et al. Reference Eng, Milly and Tasker2007).

Our unit of analysis is municipality votes. We are able to rely on more than 2800 municipalities for each vote. Our key variable is exposure and captures whether a municipality has been hit by a flood in the 12months prior to a specific vote. For each vote, we include all treated and control municipalities for which data is available.

3.3 Additional data sources

Apart from the municipality-level ballot outcomes and the disaster database, we also collected electoral results by municipality for all federal elections during the period (BfS 2013). In addition, we gathered numerous geographic and geological variables for all municipalities. Two noteworthy clusters of variables are, first, the surface type in which we have detailed information of how much area in a municipality is underbrush, artificial (houses, lawns, parks), wetland areas, water (lakes or rivers), or forests (Swisstopo 2013). The second relevant part is annual rainfall data per municipality from 1995 to 2011 (MeteoSwiss 2013).

4.1 Identifying the impact of floods on climate behavior: difference-in-differences

A first attempt to identify the impact of a natural disaster event on voting behavior is to compare municipalities hit by a landslide or flood with those not affected. In this very first step, we rely on a difference-in-differences estimator. Model I only includes municipality-level party vote-shares, Model II includes a number of geographic and surface-related variables as well as rainfall data, and Model III includes all covariates.

We include three categories of variables to take the impact of a natural disaster event on voting behavior into account: vote share for parties (as a proxy of the ideological structure of a municipality), rainfall (almost always precedes a natural disaster event), and a number of variables describing the surface. The surface is relevant to how quickly rainfall can be absorbed and thus municipalities with large agricultural spaces, and hence a strong farming element, could differ from those with largely uncultivated spaces that tend to be more oriented toward tourism. Across all models, the difference-in-differences estimator is shown in the top line (labeled as flooded).

The outcome variable in every model is the yes vote in percentage points in a municipality. The explanatory variable of interest is exposure: whether or not a municipality was affected by a natural disaster event in the 12 months preceding a vote. All models indicate that there is a positive and significant effect. Depending on the model specification, this effect is somewhere between 0.9 percent-points and 1.3 percent-points. As mentioned above, the estimated coefficients will only be causal estimates under rare circumstances. The next subsection relies on an alternative identification strategy to produce an estimate of the impact of exposure to a natural disaster event on environmental votes related to climate change.

4.2 Identifying the impact of floods on climate behavior: matching via entropy balancing

Recent contributions to the empiricist's toolbox, namely genetic matching algorithms (Sekhon Reference Sekhon2011) and entropy balancing (Hainmueller Reference Hainmueller2012), enable the retention of the full sample of treated observations while still estimating the ATT. We rely on entropy balancing as it directly achieves balance on our covariates rather than searching for weights for the nearest neighbor and also because it is computationally far less demanding and much faster. Finally, previous Monte Carlo simulations indicate that it performs superior to alternatives.

Entropy balancing enables researchers to find the optimal sets of weights that produce a perfectly balanced sample with respect to exposure. This in turn allows for the estimation of ATT. Two sets of variables are included: first, the political parties, which reflect the political and ideological structure of the municipality and second, an array of geographic and climatic variables—steepness, surface structure, and rainfall.Footnote ¹⁰ One problem is that balance is only achieved on observables. In this application, a known unobservable is the actual risk of a municipality of being affected.

We know that natural disaster events are random within municipalities that share the same risk of being affected. Ideally, we would have a perfect measure of risk and could just compare exposed and non-exposed cases with similar risk measures. While we were working on this project, the Federal Office for the Environment (FONE) concluded a program known as “Aquaprotect.” Together with one of the largest reinsurers (SwissRe), the FONE created a flood model and provided risk estimations. The model is based on grids that are 25 by 25 m. Based on this data, we compute a risk measure for every municipality by estimating the areas that are flood-prone within the next 50 years. In the Appendix, we provide an example of the flood risk and an actual flood in the municipality of Uerkheim; the accuracy of the measure is striking (see Figure A.2). While this measure cannot be a perfect measure, it is the closest we can get to the actual true risk of being affected.

Since the weighted sample is balanced, the ATT is the difference in the means of exposed and non-exposed observations. Table 2 shows the estimate as well as the covariates and balance statistics. Entropy balancing produces a set of weights for each observation and the ATT is estimated by regressing the pro-climate vote share on the variable measuring exposure while using weights.

The estimate is 2.33 percentage points, indicating that if a municipality was affected by a natural disaster in the 12 months preceding a ballot vote, it will on average cast a higher yes vote. This estimate is somewhat larger than the estimated coefficients in Models I, II, and III in Table 1. The ATT is 2.33 percentage points; since the average support for pro-climate environmental ballots lies at 42.2 percent in our sample, a treated municipality will on average cast about 6 percent more yes votes. In the next subsection, we reexamine this effect and explore whether it varies by the structure of the affected municipalities and whether the time elapsed between vote and natural disaster matters. After that, we conclude the empirical section with a number of robustness tests.

Table 1. Voting and weather (OLS)

***p < 0.01, **p < 0.05, $^\ast {\rm p}< 0.1$, full table with all estimated coefficients is presented in the appendix (Table A.1).

4.3 Exploring effect heterogeneity

So far, we have estimated an average effect. But it is most likely not the case that this estimate is really constant over all exposed observations (Gerber and Green Reference Gerber and Green2012, p. 285). In this subsection, we explore effect heterogeneity. First, we look at how the effect varies across different types of municipalities—specifically, whether the educational structure of a municipality is correlated with the effect size. Second, we look at how long-lasting the effect is over time and whether or not it fades out.

4.3.1 Heterogeneity across space

The effect is the average increase in yes votes per municipality when they are hit by a natural disaster event. The assumed underlying mechanism is that people already hold views about the causes of climate change (Beiser-McGrath and Huber Reference Beiser-McGrath and Huber2018). When people are affected immediately, their climate change consideration is activated (Zaller and Feldman Reference Zaller and Feldman1992). This will in turn affect the decision of some individuals, but not all.

One way to validate our assumption is to see if the effect is higher where people are more likely to know that there is a relationship between climate change and extreme weather phenomena. In municipalities with more people who believe that climate change is man-made, the effect is expected to be larger than in those municipalities where fewer believe in the human impact on climate change. Without precise ideological measures for municipalities or a measure of how many people believe in climate change or know how CO₂ is related to climate change, we have to use a proxy. We do not have sufficiently detailed survey data to estimate this knowledge for each municipality in Switzerland, but we do have detailed educational data. We use education—specifically, share of inhabitants with a tertiary degree—to proxy for this awareness. That is, we assume that educated people know that CO₂ is related to climate change and that climate change is related to floods.

We present here three additive models, whereas Model VI presents the interaction effect. The outcome variable is the yes-share in a municipality. To estimate the relationship between a municipality's share of well-educated citizens and the size of the exposure effect, we estimate a weighted linear regression – as in the standard set-up (see Table 2 for more details on the entropy balancing)—with the pro-climate yes-vote share as outcome. But rather than just including the exposure dummy, we now also include the share of well-educated citizens and the interaction.

Table 2. ATT based on entropy balancing (Model IV)

In Table 3, we present the original result based on entropy balancing (Model IV). In Model VI, the interaction effect is integrated. We run a weighted regression with weights that achieve perfect balance between the exposed and non-exposed group (Hainmueller Reference Hainmueller2012). The estimated parameter is significant and positive, indicating that for municipalities with a higher share of well-educated citizens, the effect is larger. For municipalities where there are no well-educated citizens at all, the effect is indistinguishable from 0. The left panel in Figure 2 presents a visualization of this conditionality. These results are consistent with the mechanism based on floods activating the climate change consideration.

Figure 2. Illustration of marginal effects.

Notes: Pseudo-Bayesian approach for uncertainty generation via sampling from the posterior distribution. Dashed line shows 95 percent confidence interval.

Table 3. Heterogeneity across municipalities

***p < 0.01, **p < 0.05, $^\ast {\rm p}< 0.1$.

4.3.2 Heterogeneity in time

The effect of being exposed was estimated by categorizing certain municipalities as exposed and others as non-exposed. We defined an exposed municipality as one affected by a natural disaster event in the 12months leading up to a ballot vote. Here, we explore whether the time lag between event and vote is related to the size of the effect.

To explore any such systematic effect heterogeneity, we use the balanced sample from the entropy balancing and regress the vote outcome on exposure and elapsed time since exposure. The first model (Model IV) only includes a binary indicator whether or not a municipality was exposed. This is the ATT based on entropy balancing since it is the difference in means with optimal weights. The estimated ATT is presented in Table 2.

In Model VIII (Table 4), we also include the amount of time that has elapsed between the natural disaster event and the vote taking place. Since the time variable is set to 0 for non-exposed units, this is the equivalent of the interaction of time and exposure. The interesting result here is that the more time that lapses between, for example, a flood and a climate sensitive vote, the smaller the effect becomes. After 10 months, there is no statistical difference from an untreated unit.Footnote ¹¹ indicating that after 10 months, these two (hypothetical) municipalities are indistinguishable. This is illustrated in the right panel in Figure 2 and shows the decay of the effect over time.

Table 4. Elapsed time and treatment intensity

***p < 0.01, **p < 0.05, $^\ast {\rm p}< 0.1$.

We present here only the simplest functional form, but the results also hold up when using the logarithm or other assumptions of decay. All these functional forms indicate that the greater the passage of time between a natural disaster and a vote, the smaller the effect.

If all municipalities were affected by flooding in the week prior to a referendum, we would on average see a yes share that is 9.4 percentage points higher. This would change the outcome of at least one vote in our sample (that on Subsidy for renewable energies). On average, the votes in this analysis have a vote-share of about 42.2 percent for the pro-climate position. Hence, the effect increases the yes-vote share by roughly 20 percent.

6.1 Future exposure

The first robustness check relies on an alternative exposure variable—municipalities are coded as treated if they experienced a local natural disaster in the 12 months after a vote. In the lead variable, a municipality is not counted as treated if it was hit before a vote but we code it as treated if it was hit after a vote. Since the disaster event is actually after the vote, there should be no effect. This allows a first check of the robustness of our findings. In a second step, we can also add the regular exposure variable and rule out anticipatory effects (Malani and Reif Reference Malani and Reif2015). If we find a significant effect for the lead variable, that would suggest the presence of anticipation which would (partly) refute our argument. We do not find any significant lead effect of natural disasters on pro-climate political behavior across all models (see Tables A.3 and A.4). Importantly, the coefficient of our exposure variable remains positive and significant. This increases the credibility of the presented results.

6.2 Placebo test: environmental votes not related to climate change

An additional robustness test can be performed by reestimating our models on other environmental votes. During the period studied, three other votes were held that were related to the environment but not climate. Two votes in 2003 were on nuclear power plants (one demanded a ban on new nuclear power plants and the other demanded the slow phasing-out of nuclear power) and one vote in 2008 sought to ban military fighter jet training in recreation areas due to noise. The discussion surrounding the two nuclear power plants was based on arguments about the safety of nuclear power (implying support for a ban of nuclear power plants) and the economic costs of such a ban (implying opposition to the ban). The two main arguments against the ban were the loss of jobs and expected price hikes (Blaser et al. Reference Blaser, van der Heiden, Mahnig and Milic2003). All three votes were strongly supported by the Green party but climate change arguments were not a relevant part of the debate on the pro or on the contra side.

We reestimated the effect with both strategies (difference-in-differences and entropy balancing) and show the results in the appendix (see Tables A.5 and A.6). We estimate the models on data from these three votes which are not related to climate change. All estimated effects are not distinguishable from 0, as we would expect.

6.3 Placebo test: non-environmental votes

The final robustness test relies on seven votes about the relationship between the European Union and Switzerland. These votes are unrelated to climate change and represent yet another way of assessing the robustness of the results. These votes also tap into the second dimension but represent a more salient issue (Kriesi et al. Reference Kriesi2005; Linder and Mueller Reference Linder and Mueller2017). We re-estimate the models and find across all specifications no effect (see Tables A.8 and A.9). This further increases the confidence in the main findings.

6.4 Surrounding municipalities

An additional robustness check involves adding information on other municipalities that are close to the flooding events. We add a covariate that captures whether a municipality is close to an affected municipality and use different cutoffs: 2, 4, and 8 km. The substantive results do not change as the estimates remain virtually unchanged. We present this analysis in Table A.10 and the original estimates remain unchanged.

6.5 Summary of empirical results

The last couple of sections show that across a number of different empirical models, there is a clear positive effect—the pro-climate voting share is significantly higher in municipalities recently affected by local extreme weather events. This effect also changes depending on time or educational structure, as expected. The robustness section serves to increase confidence in the identified effect here. It does so by showing that the effect is not found when coding future events as past events or when relying on votes where there should not be an effect.

The effect sizes in Models I-VI are moderate as the average effect is averaged over the events in the data. But in Model VIII, we estimate the decay of the effect and see that an event a week before a vote increases the yes-vote share by 9.4 percentage points, which is a sizable effect that would have changed the outcome of at least one of the votes under consideration here. In addition, the pro-climate vote share is on average 42.2 percent, which implies that being affected right before a vote will increase the yes vote share by about 20 percent.

The argument here is that these small and local events serve as information and make more salient the pro-climate consideration when the individual decides on how to or whether to cast a ballot. Apart from the effect decay, there is also a second verification that must hold. Municipalities with a higher share of people likely to hold the consideration at all should show a larger effect. To test this, we model the effect size as a function of the share of population with a tertiary education. We find again a strong correlation. Based on a model, we can compare the expected effect of a municipality where 7 percent have a tertiary degree (10th percentile) with another where that share is at 22 percent (90th percentile). The effect increases by 4 percentage points. This is a large effect. Overall, the estimation shows a robust causal effect and the correlations of effect size and two theoretically motivated variables support the argument. Voter behavior is affected by living in an area recently hit by a local extreme weather event.