- Email: [email protected]
Come hell or high water: An investigation of the effects of a natural disaster on a local election
Electoral Studies 43 (2016) 85e94
Contents lists available at ScienceDirect
Electoral Studies journal homepage: www.elsevier.com/locate/electstud
Come hell or high water: An investigation of the effects of a natural disaster on a local election Bodet a, *, Melanee Thomas b, Charles Tessier a Marc Andre a b
Department of Political Science, Universit e Laval, 2325 Rue de l'Universit e, Quebec City, G1V 0A6, Canada Department of Political Science, University of Calgary, 2500 University Dr NW, Calgary, T2N 1N4, Canada
a r t i c l e i n f o
a b s t r a c t
Article history: Received 26 January 2015 Received in revised form 31 May 2016 Accepted 4 June 2016 Available online 14 June 2016
How is electoral support for incumbent candidates shaped by natural disasters? Do voters in districts newly recovering from a national disaster punish or reward incumbents for their response to the disaster when compared to their counterparts in unaffected districts? The City of Calgary is used here as a case study. On 20 June 2013, the Bow and Elbow rivers flooded in the Calgary, devastating 26 neighborhoods and displacing approximately 75,000 people, or 7 per cent of the city's population. Four months later, a municipal election was held. When analyzed as a natural experiment, results suggest that support for the incumbent mayor increased city-wide between the 2010 and the 2013 elections, but at a lower rate in areas that experienced residential flooding. However, the flood did not produce equivalent treatment and control groups, as flooded areas differ systematically from areas that were not flooded in ways key to the election outcome. When analyzed more conservatively, results show that the flood had no effect on incumbent support or voter turnout. Thus, this disaster introduces a note of caution into the literature examining the effects of natural disasters on electoral behavior. © 2016 Elsevier Ltd. All rights reserved.
Keywords: Election Incumbency Natural experiment Natural disasters Retrospective voting Turnout
1. Introduction It is a rare day when a politician is labelled as a superhero; it is even rarer when that label is applied without sarcasm. Yet, superhero status was bestowed on Calgary's mayor, Naheed Nenshi, as a result of his reaction catastrophic flooding that hit the city in June 2013. The flood displaced 75 000 residents (approximately 7% of the city's population), and devastated 26 communities within the city limits (Bowman, 2013). Throughout, Mayor Nenshi was seen as the voice of calm, competent leadership. Credited with providing “a voice for all Calgarians”, many in the city viewed the mayor as Superman guiding them through the crisis (Bennett, 2013). This perception underpins the popular view that Nenshis leadership through disaster secured both his re-election to the mayor's office four months after the flood in October 2013 (The Canadian Press, 2013), as well as his 2014 World Mayor Award (Mayor, 2014). Research provides mixed conclusions regarding the impact of exogenous shocks such as natural disasters on electoral outcomes. While some influential research concludes that the impact on political behavior is small and even non-existent (Abney and Hill,
* Corresponding author. E-mail address: [email protected] (M.A. Bodet). http://dx.doi.org/10.1016/j.electstud.2016.06.003 0261-3794/© 2016 Elsevier Ltd. All rights reserved.
1966), others find strong (Sinclair et al., 2011; Arceneaux and Stein, 2006) or even spectacular effects (Achen and Bartels, 2004a). Despite this debate, there are still few empirical studies on the subject; those that have been conducted are almost exclusively focused on the United States. We consequently know surprisingly little in a comparative context. We use the 2013 Calgary flood, as well as the 2010 and 2013 municipal election results, to assess how incumbent vote share and voter turnout vary across districts that have, and have not been affected by a natural disaster. Two approaches are used. The first treats the natural disaster as a natural experiment. This suggests the flood is an exogenous shock that hit certain parts of the city at random, and assesses its effects on the municipal election with these assumptions in mind. This approach suggests that, as found in some of the most cited studies on the topic, the incumbent was penalized by voters for the flood. Though the incumbent mayor's support increased citywide between 2010 and 2013, this increase was smaller in areas that experienced residential flooding in 2013. The second approach assesses rather than assumes that the natural disaster affected voters as-if random. Results show the flood was not, in fact, random and did not, in fact, produce equivalent treatment and control groups. When assessed using a more conservative and appropriate research design matching techniques where
86
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
flooded districts are paired with equivalent, non-flooded districts results show that the flood had no substantial effect on electoral support for the incumbent. Moreover, both approaches show the flood did not meaningfully affect voter turnout in parts of the city affected by the flood. We proceed by outlining the existing literature on the effects of natural disasters on electoral behavior. Then, we present our empirical analysis and results. We conclude by reflecting on the generalizability of Calgary's experience, assessing how our results help build a theory of how exogenous factors such as natural disasters affect democratic elections. 2. Natural disasters and elections One way to probe the political consequences of natural disasters is to ask if voters blame incumbents for exogenous shocks, or if voters reward incumbents for their reaction to disasters that are entirely out of their control. A second common question is to ask how natural disasters affect voter turnout. Each is addressed in turn. 2.1. Support for the incumbent candidate Research provides some clues as to how natural disasters might affect incumbent support. The most compelling theories discuss retrospective voting and blame apportionment. Studies suggest that voters retrospectively evaluate incumbents, such that incumbents who are perceived to have performed well in the past are re-elected while those who are judged as poor performers may lose to a challenger (see Key, 1966; Fiorina, 1981). Research investigating these processes characterizes voters are biased, emotional, and myopic. In other words, it is not plausible to expect or suggest that most retrospective evaluations of elected officials are objective or accurate. Rather, emotions (Bower, 1981), ideology (Bartels, 2002; Anderson et al., 2004), and partisanship (Marsh and Tilley, 2010; Brown, 2010) all prevent voters from objectively or accurately attributing responsibility for events and actions to specific incumbents. These effects are not constant over time, as studies show that voters give more weight to more recent events in their retrospective evaluations than to those that occurred well before the election (Nannestad and Paldam, 2000; Bartels, 2008). Given these issues, what matters about retrospective voting is less about how accurately voters assign blame or reward for past performance, but more simply that they attribute responsibility and judge responsiveness for something to the incumbent, and then act on it. Thus, retrospective voting often leads to voters punishing or rewarding elected officials for things that are well outside their control (Achen and Bartels, 2004b; Arceneaux and Stein, 2006; Healy and Malhotra, 2010; Chang and Berdiev, 2015). For example, Achen and Bartels show how the incumbent president (Woodrow Wilson in their case) was punished for a series of shark attacks in 1916. They conclude that this kind of blind retrospection seriously hampers elections as a form of meaningful democratic accountability (Achen and Bartels, 2004a). By contrast, Abney and Hill (1966) argue instead that the mayor of New Orleans, Victor Schiro, manages to avoid harsh punishment from the voters after his vigorous and active response to Hurricane Betsey in 1965. Other studies suggest that though voters might take the consequences of natural disasters into account (Healy and Malhotra, 2010; Gasper and Reeves, 2011; Cole et al., 2012), this is neither automatic nor always attributed to the correct elected official or incumbent (Arceneaux and Stein, 2006). This seems especially plausible in a federal context, such as Canada or the United States. Moreover, voters sometimes reward politicians if they conclude the incumbent reacted to the disaster in a satisfactory fashion (Gasper
and Reeves, 2011). For the Calgary case, this suggests that the elected officials that might be most likely to experience reward are members of the provincial Legislative Assembly, as the provincial government remains responsible for the bulk of relief payments that came immediately after the flood (see Government of Alberta (2014)), rather than any municipal politician in Calgary. Still, on balance, the literature suggests that incumbents are penalized by voters for natural disasters outside their control. The 2013 municipal election in Calgary is an important test of this generalization for three reasons. First, this case is very similar to New Orleans experience mobilized in a seminal article on the topic of blame attribution (Abney and Hill, 1966). This makes Calgary's 2013 flood important not only for general replication of these results, but also for replication in a non-American context. Second, the positive narrative surrounding Mayor Nenshis leadership during the 2013 flood suggests that voters may not have penalized him for the flood. Third, 2013 marks Nenshis first campaign for re-election, as he was elected mayor in 2010 in an open contest. Since the late 1980s, incumbent mayors seeking re-election in Calgary win with large majorities of the popular vote. For example, the past two mayors prior to Nenshi e Al Duerr and David Bronconnier e were both first elected with 28% of the popular vote. Duerr was subsequently reelected with 90% of the popular vote in his first re-election bid, while Bronconnier was re-elected with 79% of the popular vote. Nenshi was initially elected with a higher proportion of the popular vote in his first election (40%) than his predecessors. This suggests that it is possible, though unlikely that Nenshis support would actually decrease between 2010 and 2013. Instead, it is more plausible that the flood may affect the rate of increase in his support in 2013 from their 2010 levels. 2.2. Turnout A second factor related to incumbent support, but also important in its own right is voter turnout. Rational choice theory suggests that voting itself is an irrational act, as the benefits of voting are disproportionately low when compared to the probability that one will cast the decisive ballot in an election (Downs, 1957; Barry, 1978; Gelman et al., 2004; Blais, 2000). In fact, the benefits of the vote are so low, even small increase of the costs can lead to considerable variations in turnout (Aldrich, 1993). The logical extension of this argument is that natural disasters increase the costs associated with learning about candidates and most importantly voting per se. This should lead a rational citizen in that area to further disengage from the electoral process (for a similar argument regarding cold weather, see Shachar and Nalebuff, 1999). However, other studies examining voter turnout highlight civic duty as an important factor. Citizens who feel it is their duty to vote feel satisfaction in the act itself quite apart from whatever effect their vote might have on the outcome of the election. This is enough to trump the otherwise powerful arguments suggesting that the most rational thing to do is abstain from the process (Riker and Ordeshook, 1968; Blais and Young, 1999; Blais, 2000). Field experiments confirm that when social pressure is applied to voters, the probability they will vote increases considerably (Gerber et al., 2008; Green and Gerber, 2010). Importantly, duty can also be a social or collective property, leading a voter to the ballot box out of a sense of duty for others (Uhlaner, 1999). Framed this way, a natural disaster such as a flood might actually increase voter turnout, particular amongst those who feel a sense of duty to cast a ballot. Studies suggest that evidence supports both approaches to voter turnout in post-disaster elections. For example, in the aftermath of Hurricane Katrina, turnout decreased on average, but increased in the most affected areas (Sinclair et al., 2011).
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
On balance, we find the negative argument more convincing, and expect that the additional costs associated with voting after a natural disaster such as the flood may trump any higher level of civic duty that the disaster may have also produced. Furthermore, it does not necessarily follow that the flood alone could produce the necessary social pressure required to increase voters sense of collective duty. Finally, the scope of the devastation individuals and their families experienced suggests that these particular experiences may have a stronger pull on individuals as voters, rather than any collective pressure that may have been generated by the natural disaster. 3. The 2013 flood in Calgary Calgary's 2013 flood is the textbook definition of an exogenous shock that could, in theory, affect the subsequent civic election. The City of Calgary is located at the confluence of the Bow and Elbow rivers in southern Alberta, roughly 100 km east of the Rocky Mountains, and just over 300 km north of the Canada-US border. As of 2013, approximately 1.15 million people live in Calgary, making it the largest municipality in Alberta and the third largest in Canada (City Clerks Office, 2013; Statistics Canada, 2014). Calgary city council is comprised of 15 members, elected every 4 years: 1 mayor, elected citywide, and 14 councillors elected in local wards (or district). In 2010, there was no incumbent candidate for mayor, leading to higher than average voter turnout, and lower than average support for the winning candidate.1 Social media campaigning is credited, in part, with Nenshis come-from-behind victory in 2010 (CBC News, 2010). Traditionally, Calgary civic politics are relatively centralized, and its local government is tasked with, among other things, planning and zoning the development and expansion of the city (Lightbody, 2006). Flood mitigation infrastructure occurs upstream, well outside the city's jurisdiction or responsibility. Typically, Calgary receives approximately 420 mm of precipitation per year, and approximately 22% of it falls in June (Environment Canada, 2014a). Between 1981 and 2010, Calgary received on average about 94 mm of rainfall in June. June 2013 is unique, as just under 147 mm of rain fell that month, most of it (68 mm) fell on June 20 and 21 (Environment Canada, 2014b). This led to 33 states of emergency to be declared across communities in southern Alberta (Government of Alberta (2013)). The Bow Rivers base flow speed is usually 225 cubic metres per second, while the Elbows base flow speed is normally 30 cubic metres per second.2 By June 20, 2013, the Bows flow rate peaked at 1750 cubic metres per second, and was maintained at this rate for approximately 32 h. At the same time, the Elbow River inflow reached 40 times its normal rate. Because of this, the Glenmore Reservoir the primary source of Calgary's drinking water, created by the Glenmore Dam overflowed. Once this happened, the City could no longer control the downstream flow of the Elbow River. At its peak, the Elbow River outflow from the Glenmore Reservoir reached 700 cubic metres per second. Calgary's state of local emergency was declared at 10:16 a.m. on June 20. By 1:15 p.m., an evacuation plan was developed for six lowlying communities along the Elbow River. By 4 p.m., these communities were placed under a mandatory evacuation order.3 By 7
1 Calgary's 2010 mayoral election was atypical in that the eventual winner (Nenshi) was in a distant third place about a month before election day (Braid, 2010). 2 The City of Calgary compiled their flood-related information at the following website: http://www.calgary.ca/General/flood2013/Pages/timeline.htm. What follows, unless otherwise noted, borrows considerably from these primary sources. 3 The communities evacuated at 4 p.m. are Mission, Elbow Park, Stanley Park, Roxboro, Rideau, and Discovery Ridge.
87
p.m., evacuation orders were expanded to include an additional 12 communities along the Bow and Elbow Rivers.4 Around this time, Calgary Transit CTrain service from the south of the city to the downtown core was disrupted due to rising water and flooding risk. By 10:30 p.m., the flooding had rapidly proceeded, extending the evacuation to a further 13 communities.5 In total, about 75 000 Calgarians were displaced from their homes by the flooding. By 5 a.m. on June 21, flooding had cut off access to the downtown core. By 6:30 a.m. on June 21, the City requested that all Calgarians refrain from non-essential travel. By June 22, approximately 30 000 clients, including the entire downtown core, were without power. Power was restored to approximately 10,000 clients by the evening of June 24. Though a local state of emergency persisted, residents living in portions of 6 evacuated communities were permitted to go home.6 By midafternoon on June 23, the downtown core reopened to building owners only to assess the extent of flood damage. Two days later, just over 50% of the downtown core had power re-established. Most of the power grid was fully restored by June 27. Calgary transit the C-Train service was fully restored on July 3. By the evening of June 23, approximately 65,000 of the 75,000 evacuated residents are given clearance to return home. The City requested that all residents returning to a previously evacuated area ensure three things: 1) that the road and sidewalk in front of their homes are dry; 2) that there is no flood water in their home upon entry, and 3) any water in the basement is below the level of electrical outlets. If all three conditions are met, residents indicated if they needed electricity, power, or water pumping on sheets of paper placed their front windows. Anyone living in an apartment building or condominium routed this process through his or her building management. Calgary's state of local emergency was lifted just after 10 a.m. on July 4, 2013. The estimated cost of the flood related damage to public buildings and infrastructure in Calgary alone is more than $256 million; the total estimated cost of the flood throughout southern Alberta is over $5 billion CAD (Wood, 2013). 4. Data and identification strategy The data used for the analysis presented here were gathered from the office of the City of Calgary Election and Information Services. Most 2013 information, including election results, voter turnout, and census data are available online.7 All mayoral election returns from 1989 to 2010 are available in hard copy from the City Clerks office; the PDFs are available from the authors upon request. In addition to this, the City of Calgary's aerial flood map was overlaid on the 2013 poll division map to determine which poll divisions experienced residential flooding.8 As a result, all variables are operationalized at the voting subdivision level within each
4 The communities evacuated at 7 p.m. are Discovery Ridge, Bowness, Sunnyside, Eau Claire, Downtown East Village, Bonnybrook, Mission, Elbow Park, Elboya, Roxboro, Rideau, Inglewood Erlton, Cligg Bungalow, Victora Park, Westmount, and Mongomery. As a resident of the East Village, one of the authors was evacuated at this time. 5 The communities evacuated at 10:30 p.m. are the Beltline, Bonnybrook, Bowness, the Bridgeland Industrial Area, Chinatown/Eau Claire, Cliff Bungalow, Deer Run, Discovery Ridge, Downtown/East Village, Elbow Park, Erlton, Inglewood, Hillhurst, Mission, Montgomery, Quarry Park, Rideau, Riverbend, Riverdale, Roxboro, Stanley Park/Elboya, Sunnyside, Victoria Park, Westmount, and Windsor Park. 6 Quarry Park, Riverbend, Deer Run, Douglasdale, Hilllhurst, and Bridgeland. 7 http://www.calgary.ca/CA/city-clerks/Pages/Election-and-information-services/ 2013-General-election/default.aspx. 8 The authors would like to thank Peter Peller at the Spatial and Numeric Data Services at the Taylor Family Digital Library at the University of Calgary for his assistance with this task.
88
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
ward. In 2013, Calgary's 14 municipal wards were divided into anywhere between 12 and 23 general voting subdivisions where voters can vote for a mayoral candidate, as well as candidates for city council. In addition to these 165 general voting subdivisions, there is a series of designated stations for specific electorates, such as the elderly housed in nursing homes. Since we are interested in a comparison between the 2010 and the 2013 elections, we are unfortunately forced to eliminate 22 voting subdivisions that have been modified between the two elections. The size of voting subdivisions vary substantially, both in terms of geography and population. Some are small and cover no more than a few city blocks. Others, typically at the fringe of the city limits, are the geographical size of small towns. Similarly, some voting subdivisions are quite small with several hundred electors. Most, though, contain thousands of electors. Voter turnout is considerably lower in municipal elections than in provincial and federal elections. Importantly, turnout varies substantially between voting subdivisions, ranging between 20% in some and 60% in others. The variation in turnout between voting subdivisions is found both between and within wards. Fig. 1 shows how voting subdivisions and wards are superposed. The sizes and organizations vary substantially, as most are organized, at least in part, around natural delimitation such as the Bow and the Elbow Rivers. Fig. 2 shows which voting subdivisions were affected by residential flooding before the municipal election. Most affected voting subdivisions are concentrated around the downtown core, though additional parts of the city were heavily affected, both upstream and downstream of the downtown core. We do not have precise information about the variation in the amount of residential flooding that affected voting subdivision.9 It would have been interesting to have access to such data, as others such as Arceneaux and Stein (2006) have in the past. We also do not have access to individual-level data since electoral data at that level are not publicly available in Canada. Consequently, we do not benefit from the same data setup as other recent studies (Sinclair et al., 2011; Healy and Malhotra, 2010). These limitations bring us closer to what Abney and Hill (1966) and Achen and Bartels (2004b) address in their work. Still, we contend there is value in this analysis, especially with respect to assessing the generalizability of the literature on natural disasters and electoral behavior outside the United States. 5. Empirical findings 5.1. The natural experiment framework
natural experiment.10 When an exogenous factor e independent of the phenomenon under study e leads to a potential change in variables of interest, some argue that the framework is similar to an experiment. The likelihood of being affected by the exogenous factor (or not) is then considered as-if random. In this context, we can measure causal relationships (see Sekhon and Titiunik, 2012 for a cautionary discussion). On the surface, it appears as though the Calgary flooding provides such an opportunity since the overflow of both the Bow and the Elbow Rivers were certainly not correlated with the dynamics of local politics (orthogonality criterion). Similarly, municipal elected officials could not have done anything obvious to avoid the initial damages (exogeneity criterion), especially given the absence of municipal governments' responsibility regarding important civil engineering projects outside their jurisdictions.11 Finally, the likelihood of having a voting subdivisions flooded in wards close to the rivers is potentially as-if random, relative to the pool of voting subdivisions close to the rivers. Table 1 shows the descriptive statistics of our variables of interest.12 In total, half the wards experienced some degree of residential flooding, and 11% of voting subdivisions were flooded. That percentage rises to 18% (not shown) of voting subdivisions if we only include wards that are next to the water and 21% in wards were some flooding occurred. We see that support for the incumbent mayor increases on average by 32 percentage points between the two elections. This average change remains roughly the same as we reduce our pool of subdivisions from all cases included in our analysis, to only those flooded or adjacent to flooded subdivisions. We observe on average a decrease of around 13 percentage points in turnout between 2010 and 2013. This is not surprising considering the absence of real competition between mayoral candidates. This change is reduced to 11 percentage points if we only include subdivisions that were flooded or adjacent to flooded ones. Table 2 shows average differences in change in support for the incumbent and turnout between subdivisions that were flooded and those that were not. There is a clear imbalance in the number of cases as only 15 of the 142 relevant subdivisions were flooded. That said, we find, using difference in means in a Bayesian framework,13 that there is a statistically significant difference between the two groups as the incumbent mayor saw his support increases by “only” 26 percentage points on average in flooded subdivision against 32 percentage points in the others. In other words, there is a 6 percentage point gap in the change in incumbent support that ostensibly could be attributed to the flood. If the comparison is limited to the 73 voting subdivisions situated in wards that suffered some level of flooding, the difference is reduced to 4 percentage points but falls short of statistical significance. Finally, if we further limit the sample
Given that the 2013 flood was unexpected, the analysis first approaches its effects on the subsequent municipal election as a
9 It worth considering why we identify living in a flood-affected voting subdivision as a treatment, rather than, say, working in the downtown core. Our rationale is three-fold. First, the data available to use identify voting subdivisions that experienced residential flooding only. Second, many who work in the downtown core also live within 5 km/3 miles of the core. The overwhelming majority of these communities experienced residential flooding. Given this, our measure of residential flooding captures many, if not a majority of those who work downtown. Third, downtown workers who were not evacuated from their homes were asked to take a “family” day on June 21, 2013 and avoid any unnecessary travel. Though this may have been disrupted some voters, we argue this disruption is minor and considerably more temporary compared to evacuation due to residential flooding. Given that, we suggest that the primary effect the flood could have on the subsequent election rests with those who experienced residential flooding in their neighborhoods. 10 For good illustrations of the potential of natural experiments, see Posner 2004; Dunning 2012.
11 In Canada, full responsibility for these projects rests with either the provincial or federal government; constitutionally, the provincially governments are also technically responsible for all municipalities. 12 Note that the means and medians tend to be similar, showing a tendency toward normal distributions. We have excluded subdivisions that were redrawn between 2010 and 2013. This is not a problem since redistricting is independent of flood risks. 13 We opt for a Bayesian framework since we are dealing with a small population, not a sample. There is discordance in the literature about the appropriate use of frequentist statistics when one benefits from a whole population. Contrary to their frequentist counterparts, Bayesian statistics are solely dependent on the data at hand, without any assumption of random sampling. Some thus argue that this approach is more coherent when working with nonrepeatable data or with nonprobabilistic samples (Jackman, 2009). But others have made a good case in favor of the frequentist approach in these settings (see Abadie et al., 2014). However, in addition to theoretical features, there are also practical benefits to the use of Bayesian statistics. For example, it allows for additional specification tests, notably through the incorporation of subjective priors.
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
89
Fig. 1. A map of Calgary with the flooded areas.
to the 47 voting subdivisions that were either flooded or situated right next to flooded subdivisions, the difference is further reduced
to 2 percentage points and is not statistically different from zero. A similar pattern is found when the analysis is replicated for
90
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
Fig. 2.
voter turnout. There is a statistically significant difference between flooded and non-flooded voting subdivisions when all 142 of them
are included. Turnout has decreased by an average 13 percentage points in subdivisions not affected by the flood, compared to 10
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94 Table 1 Descriptive statistics. Incumbent
Incumbent
Turnout
Turnout
2010
2013
2010
2013
0.41 0.40 0.11 0.21 0.70 142
0.73 0.73 0.07 0.54 0.89 142
0.50 0.50 0.08 0.27 0.70 142
0.37 0.37 0.09 0.20 0.62 142
Flooded wards Mean 0.46 Median 0.47 SD 0.12 Min 0.21 Max 0.70 N 73
0.75 0.75 0.07 0.58 0.89 73
0.52 0.52 0.06 0.34 0.70 73
0.41 0.40 0.09 0.23 0.62 73
Neighboring Mean 0.48 Median 0.50 SD 0.14 Min 0.23 Max 0.70 N 47
0.75 0.76 0.09 0.54 0.89 47
0.51 0.51 0.07 0.34 0.62 47
0.40 0.39 0.09 0.23 0.62 47
All Mean Median SD Min Max N
Flooded
0.11
0 1 142 0.21
0 1 73 0.32
0 1 47
Table 2 Difference in means between 2010 and 2013. Control
Flooded
Difference
Interval 95%
0.32 0.13 127
0.26 0.10 15
0.06 0.03
[0.10, 0.02] [0.01, 0.06]
Flooded wards D Support D Turnout N
0.30 0.11 58
0.26 0.10 15
0.04 0.01
[0.08, 0.01] [0.03, 0.05]
Neighboring D Support D Turnout N
0.28 0.11 32
0.26 0.10 15
0.02 0.01
[0.07, 0.02] [0.03, 0.05]
Matched D Support D Turnout N
0.26 0.09 15
0.26 0.10 15
0.01 0.02
[0.04, 0.04] [0.06, 0.03]
All D Support D Turnout N
percentage points in areas that were flooded. But, if we only include the 73 subdivisions situated in flooded wards, the difference all but disappears completely. The results remain the same in the proximity sample presented last in the table. It is possible that some ward-specific characteristics condition or mediate the impact of floods on incumbent support and turnout. Thus, we conducted a series of linear regressions where a series of fixed effects for each ward are added to take into account the variation within wards' limits. Similar estimates and levels of significance are found.14 We also tested a different treatment. Achen and Bartels (2004a), instead of testing for counties where shark attacks occur, opt for a dichotomous variable that takes the value 1 in what the authors refer to as “beach” counties and 0 otherwise. Our version tests for a “next to water” treatment where all subdivisions situated next to the Bow or the Elbow Rivers were coded 1, and 0 otherwise. Again,
14
Results available on demand.
91
results are similar though weaker, especially for turnout (not shown). What can we conclude at this point from the natural experiment setup? Are these results robust enough to reach a satisfactory conclusion? The answer is not simple. Even though a natural disaster is strictly orthogonal and exogenous to local political dynamics e that is, there is good reason to use the term “Act of God” in this context15 e this particular instance may not be as much of an experiment as might expected. We know that there seems to be a small but not trivial effect of flooding on incumbent support, though a smaller effect is found for voter turnout. Both effects decrease as we limit the size of our sample to the most comparable cases. This tells us that we need to rethink hard about our design. As stated above, the natural experiment approach depends entirely on the assumption of as-if random assignment. However, the difficulty with natural experiments is that even if the natural intervention (treatment) is orthogonal, exogenous, and as-if randomly assigned, it remains an open question whether or not the treatment and control groups are equivalent (enough) to avoid multivariate analysis (Sekhon and Titiunik, 2012). Given this, the first step of a more cautious approach is to assess whether flooded areas are, in fact, equivalent to areas not affected by the flood. To do this, we assessed the equivalence of flooded/treatment and not flooded/control voting subdivisions on a list of key factors e on which data is available e that may be important to municipal vote choice: home ownership rate, age distribution, percentage of government employees, voter turnout, and the level of support the incumbent received in the previous election in 2010. Though we are limited by the availability of data, we remain confident that our list includes crucial variables, notably past level of support for the current mayor and past turnout. Results in Table 3 show that the treatment and control groups are statistically equivalent on most dimensions. But it is not the case for support for the incumbent mayor in the previous election. This is problematic in a natural experiment framework since these data violate the equivalence between treatment and control precondition for an experiment. This suggests that the natural experiment approach shown above must be revisited with a more conservative, observational approach. 5.2. The observational framework There are two challenges at play in these data. First, there are only 15 flooded subdivisions, representing a mere 10% of our potential sample of 142 subdivisions. We thus have to be careful with the use of multivariate analysis, as a few cases could easily skew the estimation in one direction or the other. Second, and more importantly, the variable that breaks the equivalence between treatment and control e support for Nenshi in 2010 e is a lag variable with strong correlation with one of the two outcomes we wish to explain. After all, it is reasonable to expect that voters who supported the incumbent in their first election in 2010 may be willing to do so again in 2013, regardless of whether or not they were flooded. Given this, a multivariate linear regression might not be the most appropriate tool to identify other explanatory factors, since this lag variable artificially captures a large chunk of the variance in our explained variable. Fortunately, Sinclair et al. (2011) provide us with an elegant strategy to solve this particular methodological problem: matching. Here, treated and control cases are matched on a series of covariates. The objective is to reach what Keele and Titiunik (2015) refer
15 Abney and Hill (1966) do the same though more recent work by Dodds (2015) put the use of this expression in perspective.
92
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
Table 3 Balance test. Control
Flooded
Difference
Interval 95%
All subdivisions % Support 2010 % Turnout 2010 % Home owners % Public servant % 34 y/o and % 55 y/o and þ N
0.40 0.49 0.28 0.02 0.47 0.21 127
0.52 0.52 0.26 0.04 0.44 0.24 15
0.12 0.03 0.02 0.02 0.03 0.03
[0.04, 0.20] [0.01, 0.06] [0.05, 0.01] [0.01, 0.05] [0.06, 0.01] [0.01, 0.07]
Flooded wards % Support 2010 % Turnout 2010 % Home owners % Public servant % 34 y/o and % 55 y/o and þ N
0.44 0.52 0.28 0.03 0.46 0.22 58
0.52 0.52 0.26 0.04 0.44 0.24 15
0.08 0.01 0.02 0.01 0.03 0.02
[0.01, [0.03, [0.03, [0.10, [0.07, [0.02,
0.16] 0.03] 0.02] 0.03] 0.01] 0.07]
Neighboring % Support 2010 % Turnout 2010 % Home owners % Public servant % 34 y/o and % 55 y/o and þ N
0.46 0.50 0.25 0.04 0.47 0.21 32
0.52 0.52 0.26 0.04 0.44 0.24 15
0.06 0.02 0.01 0.01 0.03 0.03
[0.03, [0.02, [0.03, [0.02, [0.07, [0.02,
0.15] 0.06] 0.05] 0.03] 0.03] 0.07]
is to see how these two parameters of interest behave if we relax the assumption of the matching process. In our case, we do not have statistically significant effects of flooding on support for the incumbent and turnout, in a Bayesian framework. Still, it is interesting to note that the size of our estimated effects remain the same, and their associated p-values e in a frequentist framework e do not reach significance as the sensitivity parameter (Gamma) diverges from 1.17 These results suggest that, contrary to our expectations and to what we had found using raw differences, the flood did not significantly change support for the incumbent, nor did it have a meaningful effect on turnout. While it may be an overstatement to say that the 2013 flood had no effect on the subsequent 2013 municipal election, there certainly is no evidence to suggest this is a case where voters punished the incumbent for a natural disaster outside their control. 6. Concluding remarks
to as Conditional Geographic Treatment Ignorability, or the fact that “potential outcomes are independent of treatment assignment once we condition on pretreatment covariates (67)”. Following Sinclair et al. (2011) but also other work in the similar vein, we make use of MatchIt, a statistical routine developed by Ho et al. (2013) that find the best dyads of cases in the treatment and control pools to isolate the impact of a specific explanatory variable. In this case, we have opted for a conservative strategy by matching our 15 flooded subdivisions with 15 others based on all sociodemographic variables at hand, plus the lag of incumbent support and turnout previously mobilized for the balance test. The last section of Table 2 presented above shows the differences in means in support for the incumbent and turnout in 2013 between matched flooded and non-flooded subdivisions. We use a Bayesian estimation to ascertain if the differences in incumbent support and voter turnout are statistically significant between flooded and non-flooded areas. We limit the pool of potential subdivisions to those that have either been flooded, or that neighbor flooded subdivisions. This strategy has the advantage of reducing the risk of omitted sources of variability.16 We find that flooded subdivisions have seen, on average, an increase in the support for the incumbent of 26 percentage points, the same result as in our control group. Let us now turn our attention to turnout. When we again limit the potential pool to only subdivisions that were flooded or their neighbors, we find that flooded subdivisions have, on average, seen again a decrease in the support for the incumbent of 10 percentage points against 9 percentage points otherwise. The difference is not statistically significant, however. To make sure our results are able to resist to the presence of an unobserved confounder, we have perform a sensitivity test developed by Rosenbaum (2002) and implemented in R by the package rbound (Keele, 2014). This sensitivity test on both pvalue and estimated effect is usually utilized in the context of a statistically significant effect after a matching process. Its objective
This study introduces a note of caution into the developing literature examining the effects of natural disasters on electoral behavior. We side with Sekhon and Titiunik, and advocate that researchers should first assess whether a natural disaster produces equivalent experimental and control groups. If it does not, then it is not appropriate to analyze its effects as an experiment. Stated differently, assuming the natural disaster acts as an as-if random exogenous shock, and treating it as a natural experiment, can lead to incorrect conclusions. The empirical benefits of this approach rest with its caution. The two analyses above demonstrate that a voters likelihood of being affected by an exogenous factor such as a natural disaster is actually an empirically testable proposition. Given this, no analysts need assume that natural disasters are, in fact, experiments. Rather, studies that wish treat natural disasters as experiments should justify this analytical choice by demonstrating the disaster produces equivalent experimental and control groups. Failing to do so would, in this case, lead us to inappropriately reject the null hypothesis. Theoretically, the 2013 flood in Calgary shows that the assertion that natural disasters typically have negative effects on incumbent support is not as generalizable as some might suggest. Instead, this particular case may be an instance where the popular discourse accurately reflected electoral reality: voters in Calgary did not inappropriately assign blame for the flood to the Mayor. Instead, they fairly assessed the incumbents leadership. In other words, even if voters held the incumbent responsible for their reaction to the flood, it does not necessarily follow that this attribution was automatically negative This introduces a note of caution to a literature that suggests voters e or at least a large portion of the electorate e are unable to identify who is (and is not) responsible of unforeseen or unstoppable events. The shark attacks that affected Woodrow Wilson's administration (Achen and Bartels, 2004b) are not analogous, it seems to a once-in-a-generation flood in a major Canadian city such as Calgary. That being said, it is possible that their data were not balanced either. These null findings have further implications for the literature at large. Certainly, we do not want to inappropriately generalize from one natural disaster and one municipal election. However, this null finding suggests that assertions that voters will blindly punish incumbents for exogenous shocks such as natural disasters should be
16 See Keele and Titiunik (2015) for a similar argument, especially the assumption of Conditional Local Geographic Treatment Ignorability.
17 Results available on demand. The authors would like to thank a reviewer for this suggestion.
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
applied with caution and scepticism. Certainly, a large literature shows that voters unfairly attribute both positive and negative events to incumbents. These results suggest it is problematic to suggest natural disasters play a consistent, negative role in voters retrospective evaluations of incumbent candidates in re-election campaigns. In other words, it is unfair to conclude that most voters hold incumbents to account for natural disasters that are clearly beyond the incumbents control. Given this, future research could examine how retrospective accounts of natural disasters are similar too, or vary from other retrospective considerations, such as the state of the economy. Similarly, future research could assess if voters reactions to natural disasters takes on a prospective character. We cannot assess this idea in this particular case, but it does strike us as plausible that voters may use an incumbents or candidates reaction to a disaster today as a road map for how they might react to future events while in office. Given the importance of prospective economic voting, prospective evaluations of responses to natural disasters may be a compelling avenue for future research. The null finding suggesting that the flood had no effect on voter turnout is arguably surprising. Often, the negative effects of disasters on voter turnout is assumed rather than systematically assessed, albeit with some exceptions (Sinclair et al., 2011). Though other studies suggest that small, even tiny variations in the cost of voting can substantially affect turnout (Aldrich, 1993), our findings suggest that those who suffered directly from a genuine natural disaster voted at about the same rate as their peers who were not directly affected by the flood. Admittedly, voter turnout is rather low in this case, so we cannot rule out the possibility that natural disasters might negatively affect turnout in a context where it is generally higher. That said, we can conclude that the disaster certainly did not significantly increase turnout in this case. Given this, future research could probe when, why, and how these variations in the costs associated with voting actually affect turnout in a meaningful way. Finally, the power of natural experiments for social science warrants consideration, particularly with respect to the study of political and electoral behavior. These opportunities may not allow us to implement research designs that bring us closer to rigorous causality testing as once thought. Here, it is worth underscoring that the broad understanding of the Calgary flood does not misidentify its electoral effects. Without a cautious approach that fully assesses testable assumptions, we may (continue to) erroneously understand how disasters and other such events influence elections and responsibility attribution in politics. Replication in other contexts will help us build a larger theory of when, why, and how natural disasters shape democratic politics. Thus, though this case may read a bit like the disaster at the Little Town on the Prairie, it does make an important contribution to the study of electoral behavior and democratic politics, in part by helping to lay the foundation for a much larger research agenda. Acknowledgements The authors would like to thank the anonymous reviewers for the comments, Peter Peller at the Spatial and Numeric Data Services at the Taylor Family Digital Library at the University of Calgary, the staff at the City of Calgary Elections and Information Services offices, the Fonds de recherche du Qu ebec - Soci et e et culture, and the Centre for the Study of Democratic Citizenship for its financial support. References Abadie, A., Athey, S., Imbens, G.W., Wooldridge, J.M., 2014. Finite Population Causal
93
Standard Errors. Working Paper. NBER, pp. 1e53. Abney, F.G., Hill, L.B., 1966. Natural disasters as a political variable: the effect of a hurricane on an urban election. Am. Political Sci. Rev. 60, 974e981. Achen, C.H., Bartels, L.M., 2004a. Blind Retrospection: Electoral Responses to Drought, Flu, and Shark Attacks (Working Paper). Achen, C.H., Bartels, L.M., 2004b. Musical Chairs: Pocketbook Voting and the Limits of Democratic Accountability (Working Paper). Aldrich, J., 1993. Rational choice and turnout. Am. J. Political Sci. 37, 246e278. Anderson, C.J., Mendes, S.M., Tverdova, Y.V., 2004. Endogenous economic voting: evidence from the 1997 british election. Elect. Stud. 23, 683e708. Arceneaux, K., Stein, R.M., 2006. Who is held responsible when disaster strikes? the attribution of responsibility for a natural disaster in an urban election. J. Urban Aff. 28, 43e53. Barry, B., 1978. Sociologists, Economists and Democracy. University of Chicago Press, Chicago. Bartels, L.M., 2002. Beyond the running tally: partisan bias in political perceptions. Polit. Behav. 24, 117e150. Bartels, L.M., 2008. Unequal Democracy. Princeton University Press, Princeton. Bennett, D., 2013. The Voice of All Calgarians: Flood Leaves Mayor Naheed Nenshi with Superman Status. http://news.nationalpost.com/2013/06/28/the-voice-ofall-calgarians-flood-leaves-mayor-naheed-nenshi-with-superman-status. Blais, A., 2000. To Vote of Not to Vote? the Merits and Limits of Rational Choice Theory. University of Pittsburgh Press, Pittsburgh. Blais, A., Young, R., 1999. Why do people vote? an experiment in rationality. Public Choice 99, 39e55. Bower, G.H., 1981. Mood and memory. Am. Psychol. 36, 129e148. Bowman, J., 2013. Calgary Mayor Naheed Nenshi Gets Online Hero Status. http:// www.cbc.ca/m/touch/blog-post.html?category¼yourcommunity&file¼blogauthor-john-bowman-204.html. Braid, D., 2010. Nenshi Capitalizing on Social-media Buzz. Calgary Herald, p. A4. September 26. Brown, A.R., 2010. Are governors responsible for the state economy? partisanship, blame, and divided federalism. J. Polit. 72, 605e615. CBC News, 2010. Calgary Chooses Nenshi as New Mayor. http://www.cbc.ca/news/ canada/calgary/calgary-chooses-nenshi-as-new-mayor-1.867664. Chang, C.P., Berdiev, A.N., 2015. Do natural disasters increase the likelihood that a government is replaced? Appl. Econ. 47, 1788e1808. Cole, S., Healy, A., Werker, E., 2012. Do voters demand responsive governments? evidence from indian disaster relief. J. Dev. Econ. 97, 167e181. Dodds, G.G., 2015. This was no act of god: disaster, causality, and politics. Risk, Hazard Crisis Public Policy 6, 44e68. Downs, A., 1957. An Economic Theory of Democracy. Harper and Row, New York. Environment Canada, 2014a. Canadian climate Normals. http://climate.weather.gc. ca/climatenormals/. Environment Canada, 2014b. Daily Data Report for June 2013. http://climate. weather.gc.ca/climatedata/. Fiorina, M.P., 1981. Retrospective Voting in American National Elections. Yale University Press, New Haven. Gasper, J.T., Reeves, A., 2011. Make it rain? retrospection and the attentive electorate in the context of natural disasters. Am. J. Political Sci. 55, 340e355. Gelman, A., Katz, J.N., Bafumi, J., 2004. Standard voting power indexes do not work: an empirical analysis. Br. J. Political Sci. 34, 657e674. Gerber, A.S., Green, D.P., Larimer, C.W., 2008. Social pressure and voter turnout: evidence from a large-scale field experiment. Am. Political Sci. Rev. 102, 33e48. Government of Alberta, 2013. Communities Recovering from States of Local Emergency. http://alberta.ca/rainfall-impacted-areas.cfm/. Government of Alberta, 2014. Disaster Recovery Programs. http://alberta.ca/ 2013disasterrecoveryprograms.cfm. Green, D.P., Gerber, A.S., 2010. Introduction to social pressure and voting: new experimental evidence. Polit. Behav. 32, 331e336. Healy, A., Malhotra, N., 2010. Random events, economic losses, and retrospective voting: implications for democratic competence. Q. J. Political Sci. 5, 193e208. Ho, D., Imai, K., King, G., Stuart, E., 2013. MatchIt: Nonparametric Preprocessing for Parametric Casual Inference. R package version 2.4-21. Jackman, S., 2009. Bayesian Analysis for the Social Sciences. Wiley, Chichester. Keele, L., Titiunik, R., 2015. Natural experiments based on geography. Political Sci. Res. Methods 4, 65e95. Keele, L.J., 2014. Rbounds: Perform Rosenbaum Bounds Sensitivity Tests for Matched and Unmatched Data. R package version 2.1. Key, V., 1966. The Responsible Electorate: Rationality in Presidential Voting, 19361960. Harvard University Press, Cambridge. Lightbody, J., 2006. City Politics, Canada. Broadview Press, Peterborough, ON. Marsh, M., Tilley, J., 2010. The attribution of credit and blame to governments and its impact on vote choice. Br. J. Political Sci. 40, 115e134. Mayor, W., 2014. Testimonials for Naheed Nenshi, Mayor of Calgary (Canada). http:// www.worldmayor.com/contest_2014/calgary-mayor-nenshi.html. Nannestad, P., Paldam, P., 2000. What do voters know about the economy?: a study of danish data, 1990e1993. Elect. Stud. 19, 363e392. Riker, W.H., Ordeshook, P.C., 1968. A theory of the calculus of voting. Am. Political Sci. Rev. 62, 25e42. Rosenbaum, P.R., 2002. Observational Studies. Springer, New York NY. Sekhon, J.S., Titiunik, R., 2012. When natural experiments are neither natural nor experiments. Am. Political Sci. Rev. 106, 35e57. Shachar, R., Nalebuff, B., 1999. Follow the leader: theory and evidence on political participation. Am. Econ. Rev. 89, 525e547.
94
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
Sinclair, B., Hall, T.E., Alvarez, R.M., 2011. Flooding the vote: hurricane katrina and voter participation in New Orleans. Am. Polit. Res. 39, 921e957. The Canadian Press, 2013. Nenshi Easily Reelected as Calgary Mayor; Iveson Wins in Edmonton. http://www.thestar.com/news/canada/2013/10/22/nenshi_easily_ reelect-ed_as_calgary_mayor_iveson_wins_in_edmonton.html.
Uhlaner, C.J., 1999. What the Downsian Voter Weighs: a Reassessment of the Costs and Benefits of Action. University of Michigan Press, Ann Arbor. Wood, J., 2013. Province Boosts Cost of Alberta Floods to $6 Billion. Calgary Herald, p. A4. September 25.
Contents lists available at ScienceDirect
Electoral Studies journal homepage: www.elsevier.com/locate/electstud
Come hell or high water: An investigation of the effects of a natural disaster on a local election Bodet a, *, Melanee Thomas b, Charles Tessier a Marc Andre a b
Department of Political Science, Universit e Laval, 2325 Rue de l'Universit e, Quebec City, G1V 0A6, Canada Department of Political Science, University of Calgary, 2500 University Dr NW, Calgary, T2N 1N4, Canada
a r t i c l e i n f o
a b s t r a c t
Article history: Received 26 January 2015 Received in revised form 31 May 2016 Accepted 4 June 2016 Available online 14 June 2016
How is electoral support for incumbent candidates shaped by natural disasters? Do voters in districts newly recovering from a national disaster punish or reward incumbents for their response to the disaster when compared to their counterparts in unaffected districts? The City of Calgary is used here as a case study. On 20 June 2013, the Bow and Elbow rivers flooded in the Calgary, devastating 26 neighborhoods and displacing approximately 75,000 people, or 7 per cent of the city's population. Four months later, a municipal election was held. When analyzed as a natural experiment, results suggest that support for the incumbent mayor increased city-wide between the 2010 and the 2013 elections, but at a lower rate in areas that experienced residential flooding. However, the flood did not produce equivalent treatment and control groups, as flooded areas differ systematically from areas that were not flooded in ways key to the election outcome. When analyzed more conservatively, results show that the flood had no effect on incumbent support or voter turnout. Thus, this disaster introduces a note of caution into the literature examining the effects of natural disasters on electoral behavior. © 2016 Elsevier Ltd. All rights reserved.
Keywords: Election Incumbency Natural experiment Natural disasters Retrospective voting Turnout
1. Introduction It is a rare day when a politician is labelled as a superhero; it is even rarer when that label is applied without sarcasm. Yet, superhero status was bestowed on Calgary's mayor, Naheed Nenshi, as a result of his reaction catastrophic flooding that hit the city in June 2013. The flood displaced 75 000 residents (approximately 7% of the city's population), and devastated 26 communities within the city limits (Bowman, 2013). Throughout, Mayor Nenshi was seen as the voice of calm, competent leadership. Credited with providing “a voice for all Calgarians”, many in the city viewed the mayor as Superman guiding them through the crisis (Bennett, 2013). This perception underpins the popular view that Nenshis leadership through disaster secured both his re-election to the mayor's office four months after the flood in October 2013 (The Canadian Press, 2013), as well as his 2014 World Mayor Award (Mayor, 2014). Research provides mixed conclusions regarding the impact of exogenous shocks such as natural disasters on electoral outcomes. While some influential research concludes that the impact on political behavior is small and even non-existent (Abney and Hill,
* Corresponding author. E-mail address: [email protected] (M.A. Bodet). http://dx.doi.org/10.1016/j.electstud.2016.06.003 0261-3794/© 2016 Elsevier Ltd. All rights reserved.
1966), others find strong (Sinclair et al., 2011; Arceneaux and Stein, 2006) or even spectacular effects (Achen and Bartels, 2004a). Despite this debate, there are still few empirical studies on the subject; those that have been conducted are almost exclusively focused on the United States. We consequently know surprisingly little in a comparative context. We use the 2013 Calgary flood, as well as the 2010 and 2013 municipal election results, to assess how incumbent vote share and voter turnout vary across districts that have, and have not been affected by a natural disaster. Two approaches are used. The first treats the natural disaster as a natural experiment. This suggests the flood is an exogenous shock that hit certain parts of the city at random, and assesses its effects on the municipal election with these assumptions in mind. This approach suggests that, as found in some of the most cited studies on the topic, the incumbent was penalized by voters for the flood. Though the incumbent mayor's support increased citywide between 2010 and 2013, this increase was smaller in areas that experienced residential flooding in 2013. The second approach assesses rather than assumes that the natural disaster affected voters as-if random. Results show the flood was not, in fact, random and did not, in fact, produce equivalent treatment and control groups. When assessed using a more conservative and appropriate research design matching techniques where
86
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
flooded districts are paired with equivalent, non-flooded districts results show that the flood had no substantial effect on electoral support for the incumbent. Moreover, both approaches show the flood did not meaningfully affect voter turnout in parts of the city affected by the flood. We proceed by outlining the existing literature on the effects of natural disasters on electoral behavior. Then, we present our empirical analysis and results. We conclude by reflecting on the generalizability of Calgary's experience, assessing how our results help build a theory of how exogenous factors such as natural disasters affect democratic elections. 2. Natural disasters and elections One way to probe the political consequences of natural disasters is to ask if voters blame incumbents for exogenous shocks, or if voters reward incumbents for their reaction to disasters that are entirely out of their control. A second common question is to ask how natural disasters affect voter turnout. Each is addressed in turn. 2.1. Support for the incumbent candidate Research provides some clues as to how natural disasters might affect incumbent support. The most compelling theories discuss retrospective voting and blame apportionment. Studies suggest that voters retrospectively evaluate incumbents, such that incumbents who are perceived to have performed well in the past are re-elected while those who are judged as poor performers may lose to a challenger (see Key, 1966; Fiorina, 1981). Research investigating these processes characterizes voters are biased, emotional, and myopic. In other words, it is not plausible to expect or suggest that most retrospective evaluations of elected officials are objective or accurate. Rather, emotions (Bower, 1981), ideology (Bartels, 2002; Anderson et al., 2004), and partisanship (Marsh and Tilley, 2010; Brown, 2010) all prevent voters from objectively or accurately attributing responsibility for events and actions to specific incumbents. These effects are not constant over time, as studies show that voters give more weight to more recent events in their retrospective evaluations than to those that occurred well before the election (Nannestad and Paldam, 2000; Bartels, 2008). Given these issues, what matters about retrospective voting is less about how accurately voters assign blame or reward for past performance, but more simply that they attribute responsibility and judge responsiveness for something to the incumbent, and then act on it. Thus, retrospective voting often leads to voters punishing or rewarding elected officials for things that are well outside their control (Achen and Bartels, 2004b; Arceneaux and Stein, 2006; Healy and Malhotra, 2010; Chang and Berdiev, 2015). For example, Achen and Bartels show how the incumbent president (Woodrow Wilson in their case) was punished for a series of shark attacks in 1916. They conclude that this kind of blind retrospection seriously hampers elections as a form of meaningful democratic accountability (Achen and Bartels, 2004a). By contrast, Abney and Hill (1966) argue instead that the mayor of New Orleans, Victor Schiro, manages to avoid harsh punishment from the voters after his vigorous and active response to Hurricane Betsey in 1965. Other studies suggest that though voters might take the consequences of natural disasters into account (Healy and Malhotra, 2010; Gasper and Reeves, 2011; Cole et al., 2012), this is neither automatic nor always attributed to the correct elected official or incumbent (Arceneaux and Stein, 2006). This seems especially plausible in a federal context, such as Canada or the United States. Moreover, voters sometimes reward politicians if they conclude the incumbent reacted to the disaster in a satisfactory fashion (Gasper
and Reeves, 2011). For the Calgary case, this suggests that the elected officials that might be most likely to experience reward are members of the provincial Legislative Assembly, as the provincial government remains responsible for the bulk of relief payments that came immediately after the flood (see Government of Alberta (2014)), rather than any municipal politician in Calgary. Still, on balance, the literature suggests that incumbents are penalized by voters for natural disasters outside their control. The 2013 municipal election in Calgary is an important test of this generalization for three reasons. First, this case is very similar to New Orleans experience mobilized in a seminal article on the topic of blame attribution (Abney and Hill, 1966). This makes Calgary's 2013 flood important not only for general replication of these results, but also for replication in a non-American context. Second, the positive narrative surrounding Mayor Nenshis leadership during the 2013 flood suggests that voters may not have penalized him for the flood. Third, 2013 marks Nenshis first campaign for re-election, as he was elected mayor in 2010 in an open contest. Since the late 1980s, incumbent mayors seeking re-election in Calgary win with large majorities of the popular vote. For example, the past two mayors prior to Nenshi e Al Duerr and David Bronconnier e were both first elected with 28% of the popular vote. Duerr was subsequently reelected with 90% of the popular vote in his first re-election bid, while Bronconnier was re-elected with 79% of the popular vote. Nenshi was initially elected with a higher proportion of the popular vote in his first election (40%) than his predecessors. This suggests that it is possible, though unlikely that Nenshis support would actually decrease between 2010 and 2013. Instead, it is more plausible that the flood may affect the rate of increase in his support in 2013 from their 2010 levels. 2.2. Turnout A second factor related to incumbent support, but also important in its own right is voter turnout. Rational choice theory suggests that voting itself is an irrational act, as the benefits of voting are disproportionately low when compared to the probability that one will cast the decisive ballot in an election (Downs, 1957; Barry, 1978; Gelman et al., 2004; Blais, 2000). In fact, the benefits of the vote are so low, even small increase of the costs can lead to considerable variations in turnout (Aldrich, 1993). The logical extension of this argument is that natural disasters increase the costs associated with learning about candidates and most importantly voting per se. This should lead a rational citizen in that area to further disengage from the electoral process (for a similar argument regarding cold weather, see Shachar and Nalebuff, 1999). However, other studies examining voter turnout highlight civic duty as an important factor. Citizens who feel it is their duty to vote feel satisfaction in the act itself quite apart from whatever effect their vote might have on the outcome of the election. This is enough to trump the otherwise powerful arguments suggesting that the most rational thing to do is abstain from the process (Riker and Ordeshook, 1968; Blais and Young, 1999; Blais, 2000). Field experiments confirm that when social pressure is applied to voters, the probability they will vote increases considerably (Gerber et al., 2008; Green and Gerber, 2010). Importantly, duty can also be a social or collective property, leading a voter to the ballot box out of a sense of duty for others (Uhlaner, 1999). Framed this way, a natural disaster such as a flood might actually increase voter turnout, particular amongst those who feel a sense of duty to cast a ballot. Studies suggest that evidence supports both approaches to voter turnout in post-disaster elections. For example, in the aftermath of Hurricane Katrina, turnout decreased on average, but increased in the most affected areas (Sinclair et al., 2011).
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
On balance, we find the negative argument more convincing, and expect that the additional costs associated with voting after a natural disaster such as the flood may trump any higher level of civic duty that the disaster may have also produced. Furthermore, it does not necessarily follow that the flood alone could produce the necessary social pressure required to increase voters sense of collective duty. Finally, the scope of the devastation individuals and their families experienced suggests that these particular experiences may have a stronger pull on individuals as voters, rather than any collective pressure that may have been generated by the natural disaster. 3. The 2013 flood in Calgary Calgary's 2013 flood is the textbook definition of an exogenous shock that could, in theory, affect the subsequent civic election. The City of Calgary is located at the confluence of the Bow and Elbow rivers in southern Alberta, roughly 100 km east of the Rocky Mountains, and just over 300 km north of the Canada-US border. As of 2013, approximately 1.15 million people live in Calgary, making it the largest municipality in Alberta and the third largest in Canada (City Clerks Office, 2013; Statistics Canada, 2014). Calgary city council is comprised of 15 members, elected every 4 years: 1 mayor, elected citywide, and 14 councillors elected in local wards (or district). In 2010, there was no incumbent candidate for mayor, leading to higher than average voter turnout, and lower than average support for the winning candidate.1 Social media campaigning is credited, in part, with Nenshis come-from-behind victory in 2010 (CBC News, 2010). Traditionally, Calgary civic politics are relatively centralized, and its local government is tasked with, among other things, planning and zoning the development and expansion of the city (Lightbody, 2006). Flood mitigation infrastructure occurs upstream, well outside the city's jurisdiction or responsibility. Typically, Calgary receives approximately 420 mm of precipitation per year, and approximately 22% of it falls in June (Environment Canada, 2014a). Between 1981 and 2010, Calgary received on average about 94 mm of rainfall in June. June 2013 is unique, as just under 147 mm of rain fell that month, most of it (68 mm) fell on June 20 and 21 (Environment Canada, 2014b). This led to 33 states of emergency to be declared across communities in southern Alberta (Government of Alberta (2013)). The Bow Rivers base flow speed is usually 225 cubic metres per second, while the Elbows base flow speed is normally 30 cubic metres per second.2 By June 20, 2013, the Bows flow rate peaked at 1750 cubic metres per second, and was maintained at this rate for approximately 32 h. At the same time, the Elbow River inflow reached 40 times its normal rate. Because of this, the Glenmore Reservoir the primary source of Calgary's drinking water, created by the Glenmore Dam overflowed. Once this happened, the City could no longer control the downstream flow of the Elbow River. At its peak, the Elbow River outflow from the Glenmore Reservoir reached 700 cubic metres per second. Calgary's state of local emergency was declared at 10:16 a.m. on June 20. By 1:15 p.m., an evacuation plan was developed for six lowlying communities along the Elbow River. By 4 p.m., these communities were placed under a mandatory evacuation order.3 By 7
1 Calgary's 2010 mayoral election was atypical in that the eventual winner (Nenshi) was in a distant third place about a month before election day (Braid, 2010). 2 The City of Calgary compiled their flood-related information at the following website: http://www.calgary.ca/General/flood2013/Pages/timeline.htm. What follows, unless otherwise noted, borrows considerably from these primary sources. 3 The communities evacuated at 4 p.m. are Mission, Elbow Park, Stanley Park, Roxboro, Rideau, and Discovery Ridge.
87
p.m., evacuation orders were expanded to include an additional 12 communities along the Bow and Elbow Rivers.4 Around this time, Calgary Transit CTrain service from the south of the city to the downtown core was disrupted due to rising water and flooding risk. By 10:30 p.m., the flooding had rapidly proceeded, extending the evacuation to a further 13 communities.5 In total, about 75 000 Calgarians were displaced from their homes by the flooding. By 5 a.m. on June 21, flooding had cut off access to the downtown core. By 6:30 a.m. on June 21, the City requested that all Calgarians refrain from non-essential travel. By June 22, approximately 30 000 clients, including the entire downtown core, were without power. Power was restored to approximately 10,000 clients by the evening of June 24. Though a local state of emergency persisted, residents living in portions of 6 evacuated communities were permitted to go home.6 By midafternoon on June 23, the downtown core reopened to building owners only to assess the extent of flood damage. Two days later, just over 50% of the downtown core had power re-established. Most of the power grid was fully restored by June 27. Calgary transit the C-Train service was fully restored on July 3. By the evening of June 23, approximately 65,000 of the 75,000 evacuated residents are given clearance to return home. The City requested that all residents returning to a previously evacuated area ensure three things: 1) that the road and sidewalk in front of their homes are dry; 2) that there is no flood water in their home upon entry, and 3) any water in the basement is below the level of electrical outlets. If all three conditions are met, residents indicated if they needed electricity, power, or water pumping on sheets of paper placed their front windows. Anyone living in an apartment building or condominium routed this process through his or her building management. Calgary's state of local emergency was lifted just after 10 a.m. on July 4, 2013. The estimated cost of the flood related damage to public buildings and infrastructure in Calgary alone is more than $256 million; the total estimated cost of the flood throughout southern Alberta is over $5 billion CAD (Wood, 2013). 4. Data and identification strategy The data used for the analysis presented here were gathered from the office of the City of Calgary Election and Information Services. Most 2013 information, including election results, voter turnout, and census data are available online.7 All mayoral election returns from 1989 to 2010 are available in hard copy from the City Clerks office; the PDFs are available from the authors upon request. In addition to this, the City of Calgary's aerial flood map was overlaid on the 2013 poll division map to determine which poll divisions experienced residential flooding.8 As a result, all variables are operationalized at the voting subdivision level within each
4 The communities evacuated at 7 p.m. are Discovery Ridge, Bowness, Sunnyside, Eau Claire, Downtown East Village, Bonnybrook, Mission, Elbow Park, Elboya, Roxboro, Rideau, Inglewood Erlton, Cligg Bungalow, Victora Park, Westmount, and Mongomery. As a resident of the East Village, one of the authors was evacuated at this time. 5 The communities evacuated at 10:30 p.m. are the Beltline, Bonnybrook, Bowness, the Bridgeland Industrial Area, Chinatown/Eau Claire, Cliff Bungalow, Deer Run, Discovery Ridge, Downtown/East Village, Elbow Park, Erlton, Inglewood, Hillhurst, Mission, Montgomery, Quarry Park, Rideau, Riverbend, Riverdale, Roxboro, Stanley Park/Elboya, Sunnyside, Victoria Park, Westmount, and Windsor Park. 6 Quarry Park, Riverbend, Deer Run, Douglasdale, Hilllhurst, and Bridgeland. 7 http://www.calgary.ca/CA/city-clerks/Pages/Election-and-information-services/ 2013-General-election/default.aspx. 8 The authors would like to thank Peter Peller at the Spatial and Numeric Data Services at the Taylor Family Digital Library at the University of Calgary for his assistance with this task.
88
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
ward. In 2013, Calgary's 14 municipal wards were divided into anywhere between 12 and 23 general voting subdivisions where voters can vote for a mayoral candidate, as well as candidates for city council. In addition to these 165 general voting subdivisions, there is a series of designated stations for specific electorates, such as the elderly housed in nursing homes. Since we are interested in a comparison between the 2010 and the 2013 elections, we are unfortunately forced to eliminate 22 voting subdivisions that have been modified between the two elections. The size of voting subdivisions vary substantially, both in terms of geography and population. Some are small and cover no more than a few city blocks. Others, typically at the fringe of the city limits, are the geographical size of small towns. Similarly, some voting subdivisions are quite small with several hundred electors. Most, though, contain thousands of electors. Voter turnout is considerably lower in municipal elections than in provincial and federal elections. Importantly, turnout varies substantially between voting subdivisions, ranging between 20% in some and 60% in others. The variation in turnout between voting subdivisions is found both between and within wards. Fig. 1 shows how voting subdivisions and wards are superposed. The sizes and organizations vary substantially, as most are organized, at least in part, around natural delimitation such as the Bow and the Elbow Rivers. Fig. 2 shows which voting subdivisions were affected by residential flooding before the municipal election. Most affected voting subdivisions are concentrated around the downtown core, though additional parts of the city were heavily affected, both upstream and downstream of the downtown core. We do not have precise information about the variation in the amount of residential flooding that affected voting subdivision.9 It would have been interesting to have access to such data, as others such as Arceneaux and Stein (2006) have in the past. We also do not have access to individual-level data since electoral data at that level are not publicly available in Canada. Consequently, we do not benefit from the same data setup as other recent studies (Sinclair et al., 2011; Healy and Malhotra, 2010). These limitations bring us closer to what Abney and Hill (1966) and Achen and Bartels (2004b) address in their work. Still, we contend there is value in this analysis, especially with respect to assessing the generalizability of the literature on natural disasters and electoral behavior outside the United States. 5. Empirical findings 5.1. The natural experiment framework
natural experiment.10 When an exogenous factor e independent of the phenomenon under study e leads to a potential change in variables of interest, some argue that the framework is similar to an experiment. The likelihood of being affected by the exogenous factor (or not) is then considered as-if random. In this context, we can measure causal relationships (see Sekhon and Titiunik, 2012 for a cautionary discussion). On the surface, it appears as though the Calgary flooding provides such an opportunity since the overflow of both the Bow and the Elbow Rivers were certainly not correlated with the dynamics of local politics (orthogonality criterion). Similarly, municipal elected officials could not have done anything obvious to avoid the initial damages (exogeneity criterion), especially given the absence of municipal governments' responsibility regarding important civil engineering projects outside their jurisdictions.11 Finally, the likelihood of having a voting subdivisions flooded in wards close to the rivers is potentially as-if random, relative to the pool of voting subdivisions close to the rivers. Table 1 shows the descriptive statistics of our variables of interest.12 In total, half the wards experienced some degree of residential flooding, and 11% of voting subdivisions were flooded. That percentage rises to 18% (not shown) of voting subdivisions if we only include wards that are next to the water and 21% in wards were some flooding occurred. We see that support for the incumbent mayor increases on average by 32 percentage points between the two elections. This average change remains roughly the same as we reduce our pool of subdivisions from all cases included in our analysis, to only those flooded or adjacent to flooded subdivisions. We observe on average a decrease of around 13 percentage points in turnout between 2010 and 2013. This is not surprising considering the absence of real competition between mayoral candidates. This change is reduced to 11 percentage points if we only include subdivisions that were flooded or adjacent to flooded ones. Table 2 shows average differences in change in support for the incumbent and turnout between subdivisions that were flooded and those that were not. There is a clear imbalance in the number of cases as only 15 of the 142 relevant subdivisions were flooded. That said, we find, using difference in means in a Bayesian framework,13 that there is a statistically significant difference between the two groups as the incumbent mayor saw his support increases by “only” 26 percentage points on average in flooded subdivision against 32 percentage points in the others. In other words, there is a 6 percentage point gap in the change in incumbent support that ostensibly could be attributed to the flood. If the comparison is limited to the 73 voting subdivisions situated in wards that suffered some level of flooding, the difference is reduced to 4 percentage points but falls short of statistical significance. Finally, if we further limit the sample
Given that the 2013 flood was unexpected, the analysis first approaches its effects on the subsequent municipal election as a
9 It worth considering why we identify living in a flood-affected voting subdivision as a treatment, rather than, say, working in the downtown core. Our rationale is three-fold. First, the data available to use identify voting subdivisions that experienced residential flooding only. Second, many who work in the downtown core also live within 5 km/3 miles of the core. The overwhelming majority of these communities experienced residential flooding. Given this, our measure of residential flooding captures many, if not a majority of those who work downtown. Third, downtown workers who were not evacuated from their homes were asked to take a “family” day on June 21, 2013 and avoid any unnecessary travel. Though this may have been disrupted some voters, we argue this disruption is minor and considerably more temporary compared to evacuation due to residential flooding. Given that, we suggest that the primary effect the flood could have on the subsequent election rests with those who experienced residential flooding in their neighborhoods. 10 For good illustrations of the potential of natural experiments, see Posner 2004; Dunning 2012.
11 In Canada, full responsibility for these projects rests with either the provincial or federal government; constitutionally, the provincially governments are also technically responsible for all municipalities. 12 Note that the means and medians tend to be similar, showing a tendency toward normal distributions. We have excluded subdivisions that were redrawn between 2010 and 2013. This is not a problem since redistricting is independent of flood risks. 13 We opt for a Bayesian framework since we are dealing with a small population, not a sample. There is discordance in the literature about the appropriate use of frequentist statistics when one benefits from a whole population. Contrary to their frequentist counterparts, Bayesian statistics are solely dependent on the data at hand, without any assumption of random sampling. Some thus argue that this approach is more coherent when working with nonrepeatable data or with nonprobabilistic samples (Jackman, 2009). But others have made a good case in favor of the frequentist approach in these settings (see Abadie et al., 2014). However, in addition to theoretical features, there are also practical benefits to the use of Bayesian statistics. For example, it allows for additional specification tests, notably through the incorporation of subjective priors.
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
89
Fig. 1. A map of Calgary with the flooded areas.
to the 47 voting subdivisions that were either flooded or situated right next to flooded subdivisions, the difference is further reduced
to 2 percentage points and is not statistically different from zero. A similar pattern is found when the analysis is replicated for
90
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
Fig. 2.
voter turnout. There is a statistically significant difference between flooded and non-flooded voting subdivisions when all 142 of them
are included. Turnout has decreased by an average 13 percentage points in subdivisions not affected by the flood, compared to 10
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94 Table 1 Descriptive statistics. Incumbent
Incumbent
Turnout
Turnout
2010
2013
2010
2013
0.41 0.40 0.11 0.21 0.70 142
0.73 0.73 0.07 0.54 0.89 142
0.50 0.50 0.08 0.27 0.70 142
0.37 0.37 0.09 0.20 0.62 142
Flooded wards Mean 0.46 Median 0.47 SD 0.12 Min 0.21 Max 0.70 N 73
0.75 0.75 0.07 0.58 0.89 73
0.52 0.52 0.06 0.34 0.70 73
0.41 0.40 0.09 0.23 0.62 73
Neighboring Mean 0.48 Median 0.50 SD 0.14 Min 0.23 Max 0.70 N 47
0.75 0.76 0.09 0.54 0.89 47
0.51 0.51 0.07 0.34 0.62 47
0.40 0.39 0.09 0.23 0.62 47
All Mean Median SD Min Max N
Flooded
0.11
0 1 142 0.21
0 1 73 0.32
0 1 47
Table 2 Difference in means between 2010 and 2013. Control
Flooded
Difference
Interval 95%
0.32 0.13 127
0.26 0.10 15
0.06 0.03
[0.10, 0.02] [0.01, 0.06]
Flooded wards D Support D Turnout N
0.30 0.11 58
0.26 0.10 15
0.04 0.01
[0.08, 0.01] [0.03, 0.05]
Neighboring D Support D Turnout N
0.28 0.11 32
0.26 0.10 15
0.02 0.01
[0.07, 0.02] [0.03, 0.05]
Matched D Support D Turnout N
0.26 0.09 15
0.26 0.10 15
0.01 0.02
[0.04, 0.04] [0.06, 0.03]
All D Support D Turnout N
percentage points in areas that were flooded. But, if we only include the 73 subdivisions situated in flooded wards, the difference all but disappears completely. The results remain the same in the proximity sample presented last in the table. It is possible that some ward-specific characteristics condition or mediate the impact of floods on incumbent support and turnout. Thus, we conducted a series of linear regressions where a series of fixed effects for each ward are added to take into account the variation within wards' limits. Similar estimates and levels of significance are found.14 We also tested a different treatment. Achen and Bartels (2004a), instead of testing for counties where shark attacks occur, opt for a dichotomous variable that takes the value 1 in what the authors refer to as “beach” counties and 0 otherwise. Our version tests for a “next to water” treatment where all subdivisions situated next to the Bow or the Elbow Rivers were coded 1, and 0 otherwise. Again,
14
Results available on demand.
91
results are similar though weaker, especially for turnout (not shown). What can we conclude at this point from the natural experiment setup? Are these results robust enough to reach a satisfactory conclusion? The answer is not simple. Even though a natural disaster is strictly orthogonal and exogenous to local political dynamics e that is, there is good reason to use the term “Act of God” in this context15 e this particular instance may not be as much of an experiment as might expected. We know that there seems to be a small but not trivial effect of flooding on incumbent support, though a smaller effect is found for voter turnout. Both effects decrease as we limit the size of our sample to the most comparable cases. This tells us that we need to rethink hard about our design. As stated above, the natural experiment approach depends entirely on the assumption of as-if random assignment. However, the difficulty with natural experiments is that even if the natural intervention (treatment) is orthogonal, exogenous, and as-if randomly assigned, it remains an open question whether or not the treatment and control groups are equivalent (enough) to avoid multivariate analysis (Sekhon and Titiunik, 2012). Given this, the first step of a more cautious approach is to assess whether flooded areas are, in fact, equivalent to areas not affected by the flood. To do this, we assessed the equivalence of flooded/treatment and not flooded/control voting subdivisions on a list of key factors e on which data is available e that may be important to municipal vote choice: home ownership rate, age distribution, percentage of government employees, voter turnout, and the level of support the incumbent received in the previous election in 2010. Though we are limited by the availability of data, we remain confident that our list includes crucial variables, notably past level of support for the current mayor and past turnout. Results in Table 3 show that the treatment and control groups are statistically equivalent on most dimensions. But it is not the case for support for the incumbent mayor in the previous election. This is problematic in a natural experiment framework since these data violate the equivalence between treatment and control precondition for an experiment. This suggests that the natural experiment approach shown above must be revisited with a more conservative, observational approach. 5.2. The observational framework There are two challenges at play in these data. First, there are only 15 flooded subdivisions, representing a mere 10% of our potential sample of 142 subdivisions. We thus have to be careful with the use of multivariate analysis, as a few cases could easily skew the estimation in one direction or the other. Second, and more importantly, the variable that breaks the equivalence between treatment and control e support for Nenshi in 2010 e is a lag variable with strong correlation with one of the two outcomes we wish to explain. After all, it is reasonable to expect that voters who supported the incumbent in their first election in 2010 may be willing to do so again in 2013, regardless of whether or not they were flooded. Given this, a multivariate linear regression might not be the most appropriate tool to identify other explanatory factors, since this lag variable artificially captures a large chunk of the variance in our explained variable. Fortunately, Sinclair et al. (2011) provide us with an elegant strategy to solve this particular methodological problem: matching. Here, treated and control cases are matched on a series of covariates. The objective is to reach what Keele and Titiunik (2015) refer
15 Abney and Hill (1966) do the same though more recent work by Dodds (2015) put the use of this expression in perspective.
92
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
Table 3 Balance test. Control
Flooded
Difference
Interval 95%
All subdivisions % Support 2010 % Turnout 2010 % Home owners % Public servant % 34 y/o and % 55 y/o and þ N
0.40 0.49 0.28 0.02 0.47 0.21 127
0.52 0.52 0.26 0.04 0.44 0.24 15
0.12 0.03 0.02 0.02 0.03 0.03
[0.04, 0.20] [0.01, 0.06] [0.05, 0.01] [0.01, 0.05] [0.06, 0.01] [0.01, 0.07]
Flooded wards % Support 2010 % Turnout 2010 % Home owners % Public servant % 34 y/o and % 55 y/o and þ N
0.44 0.52 0.28 0.03 0.46 0.22 58
0.52 0.52 0.26 0.04 0.44 0.24 15
0.08 0.01 0.02 0.01 0.03 0.02
[0.01, [0.03, [0.03, [0.10, [0.07, [0.02,
0.16] 0.03] 0.02] 0.03] 0.01] 0.07]
Neighboring % Support 2010 % Turnout 2010 % Home owners % Public servant % 34 y/o and % 55 y/o and þ N
0.46 0.50 0.25 0.04 0.47 0.21 32
0.52 0.52 0.26 0.04 0.44 0.24 15
0.06 0.02 0.01 0.01 0.03 0.03
[0.03, [0.02, [0.03, [0.02, [0.07, [0.02,
0.15] 0.06] 0.05] 0.03] 0.03] 0.07]
is to see how these two parameters of interest behave if we relax the assumption of the matching process. In our case, we do not have statistically significant effects of flooding on support for the incumbent and turnout, in a Bayesian framework. Still, it is interesting to note that the size of our estimated effects remain the same, and their associated p-values e in a frequentist framework e do not reach significance as the sensitivity parameter (Gamma) diverges from 1.17 These results suggest that, contrary to our expectations and to what we had found using raw differences, the flood did not significantly change support for the incumbent, nor did it have a meaningful effect on turnout. While it may be an overstatement to say that the 2013 flood had no effect on the subsequent 2013 municipal election, there certainly is no evidence to suggest this is a case where voters punished the incumbent for a natural disaster outside their control. 6. Concluding remarks
to as Conditional Geographic Treatment Ignorability, or the fact that “potential outcomes are independent of treatment assignment once we condition on pretreatment covariates (67)”. Following Sinclair et al. (2011) but also other work in the similar vein, we make use of MatchIt, a statistical routine developed by Ho et al. (2013) that find the best dyads of cases in the treatment and control pools to isolate the impact of a specific explanatory variable. In this case, we have opted for a conservative strategy by matching our 15 flooded subdivisions with 15 others based on all sociodemographic variables at hand, plus the lag of incumbent support and turnout previously mobilized for the balance test. The last section of Table 2 presented above shows the differences in means in support for the incumbent and turnout in 2013 between matched flooded and non-flooded subdivisions. We use a Bayesian estimation to ascertain if the differences in incumbent support and voter turnout are statistically significant between flooded and non-flooded areas. We limit the pool of potential subdivisions to those that have either been flooded, or that neighbor flooded subdivisions. This strategy has the advantage of reducing the risk of omitted sources of variability.16 We find that flooded subdivisions have seen, on average, an increase in the support for the incumbent of 26 percentage points, the same result as in our control group. Let us now turn our attention to turnout. When we again limit the potential pool to only subdivisions that were flooded or their neighbors, we find that flooded subdivisions have, on average, seen again a decrease in the support for the incumbent of 10 percentage points against 9 percentage points otherwise. The difference is not statistically significant, however. To make sure our results are able to resist to the presence of an unobserved confounder, we have perform a sensitivity test developed by Rosenbaum (2002) and implemented in R by the package rbound (Keele, 2014). This sensitivity test on both pvalue and estimated effect is usually utilized in the context of a statistically significant effect after a matching process. Its objective
This study introduces a note of caution into the developing literature examining the effects of natural disasters on electoral behavior. We side with Sekhon and Titiunik, and advocate that researchers should first assess whether a natural disaster produces equivalent experimental and control groups. If it does not, then it is not appropriate to analyze its effects as an experiment. Stated differently, assuming the natural disaster acts as an as-if random exogenous shock, and treating it as a natural experiment, can lead to incorrect conclusions. The empirical benefits of this approach rest with its caution. The two analyses above demonstrate that a voters likelihood of being affected by an exogenous factor such as a natural disaster is actually an empirically testable proposition. Given this, no analysts need assume that natural disasters are, in fact, experiments. Rather, studies that wish treat natural disasters as experiments should justify this analytical choice by demonstrating the disaster produces equivalent experimental and control groups. Failing to do so would, in this case, lead us to inappropriately reject the null hypothesis. Theoretically, the 2013 flood in Calgary shows that the assertion that natural disasters typically have negative effects on incumbent support is not as generalizable as some might suggest. Instead, this particular case may be an instance where the popular discourse accurately reflected electoral reality: voters in Calgary did not inappropriately assign blame for the flood to the Mayor. Instead, they fairly assessed the incumbents leadership. In other words, even if voters held the incumbent responsible for their reaction to the flood, it does not necessarily follow that this attribution was automatically negative This introduces a note of caution to a literature that suggests voters e or at least a large portion of the electorate e are unable to identify who is (and is not) responsible of unforeseen or unstoppable events. The shark attacks that affected Woodrow Wilson's administration (Achen and Bartels, 2004b) are not analogous, it seems to a once-in-a-generation flood in a major Canadian city such as Calgary. That being said, it is possible that their data were not balanced either. These null findings have further implications for the literature at large. Certainly, we do not want to inappropriately generalize from one natural disaster and one municipal election. However, this null finding suggests that assertions that voters will blindly punish incumbents for exogenous shocks such as natural disasters should be
16 See Keele and Titiunik (2015) for a similar argument, especially the assumption of Conditional Local Geographic Treatment Ignorability.
17 Results available on demand. The authors would like to thank a reviewer for this suggestion.
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
applied with caution and scepticism. Certainly, a large literature shows that voters unfairly attribute both positive and negative events to incumbents. These results suggest it is problematic to suggest natural disasters play a consistent, negative role in voters retrospective evaluations of incumbent candidates in re-election campaigns. In other words, it is unfair to conclude that most voters hold incumbents to account for natural disasters that are clearly beyond the incumbents control. Given this, future research could examine how retrospective accounts of natural disasters are similar too, or vary from other retrospective considerations, such as the state of the economy. Similarly, future research could assess if voters reactions to natural disasters takes on a prospective character. We cannot assess this idea in this particular case, but it does strike us as plausible that voters may use an incumbents or candidates reaction to a disaster today as a road map for how they might react to future events while in office. Given the importance of prospective economic voting, prospective evaluations of responses to natural disasters may be a compelling avenue for future research. The null finding suggesting that the flood had no effect on voter turnout is arguably surprising. Often, the negative effects of disasters on voter turnout is assumed rather than systematically assessed, albeit with some exceptions (Sinclair et al., 2011). Though other studies suggest that small, even tiny variations in the cost of voting can substantially affect turnout (Aldrich, 1993), our findings suggest that those who suffered directly from a genuine natural disaster voted at about the same rate as their peers who were not directly affected by the flood. Admittedly, voter turnout is rather low in this case, so we cannot rule out the possibility that natural disasters might negatively affect turnout in a context where it is generally higher. That said, we can conclude that the disaster certainly did not significantly increase turnout in this case. Given this, future research could probe when, why, and how these variations in the costs associated with voting actually affect turnout in a meaningful way. Finally, the power of natural experiments for social science warrants consideration, particularly with respect to the study of political and electoral behavior. These opportunities may not allow us to implement research designs that bring us closer to rigorous causality testing as once thought. Here, it is worth underscoring that the broad understanding of the Calgary flood does not misidentify its electoral effects. Without a cautious approach that fully assesses testable assumptions, we may (continue to) erroneously understand how disasters and other such events influence elections and responsibility attribution in politics. Replication in other contexts will help us build a larger theory of when, why, and how natural disasters shape democratic politics. Thus, though this case may read a bit like the disaster at the Little Town on the Prairie, it does make an important contribution to the study of electoral behavior and democratic politics, in part by helping to lay the foundation for a much larger research agenda. Acknowledgements The authors would like to thank the anonymous reviewers for the comments, Peter Peller at the Spatial and Numeric Data Services at the Taylor Family Digital Library at the University of Calgary, the staff at the City of Calgary Elections and Information Services offices, the Fonds de recherche du Qu ebec - Soci et e et culture, and the Centre for the Study of Democratic Citizenship for its financial support. References Abadie, A., Athey, S., Imbens, G.W., Wooldridge, J.M., 2014. Finite Population Causal
93
Standard Errors. Working Paper. NBER, pp. 1e53. Abney, F.G., Hill, L.B., 1966. Natural disasters as a political variable: the effect of a hurricane on an urban election. Am. Political Sci. Rev. 60, 974e981. Achen, C.H., Bartels, L.M., 2004a. Blind Retrospection: Electoral Responses to Drought, Flu, and Shark Attacks (Working Paper). Achen, C.H., Bartels, L.M., 2004b. Musical Chairs: Pocketbook Voting and the Limits of Democratic Accountability (Working Paper). Aldrich, J., 1993. Rational choice and turnout. Am. J. Political Sci. 37, 246e278. Anderson, C.J., Mendes, S.M., Tverdova, Y.V., 2004. Endogenous economic voting: evidence from the 1997 british election. Elect. Stud. 23, 683e708. Arceneaux, K., Stein, R.M., 2006. Who is held responsible when disaster strikes? the attribution of responsibility for a natural disaster in an urban election. J. Urban Aff. 28, 43e53. Barry, B., 1978. Sociologists, Economists and Democracy. University of Chicago Press, Chicago. Bartels, L.M., 2002. Beyond the running tally: partisan bias in political perceptions. Polit. Behav. 24, 117e150. Bartels, L.M., 2008. Unequal Democracy. Princeton University Press, Princeton. Bennett, D., 2013. The Voice of All Calgarians: Flood Leaves Mayor Naheed Nenshi with Superman Status. http://news.nationalpost.com/2013/06/28/the-voice-ofall-calgarians-flood-leaves-mayor-naheed-nenshi-with-superman-status. Blais, A., 2000. To Vote of Not to Vote? the Merits and Limits of Rational Choice Theory. University of Pittsburgh Press, Pittsburgh. Blais, A., Young, R., 1999. Why do people vote? an experiment in rationality. Public Choice 99, 39e55. Bower, G.H., 1981. Mood and memory. Am. Psychol. 36, 129e148. Bowman, J., 2013. Calgary Mayor Naheed Nenshi Gets Online Hero Status. http:// www.cbc.ca/m/touch/blog-post.html?category¼yourcommunity&file¼blogauthor-john-bowman-204.html. Braid, D., 2010. Nenshi Capitalizing on Social-media Buzz. Calgary Herald, p. A4. September 26. Brown, A.R., 2010. Are governors responsible for the state economy? partisanship, blame, and divided federalism. J. Polit. 72, 605e615. CBC News, 2010. Calgary Chooses Nenshi as New Mayor. http://www.cbc.ca/news/ canada/calgary/calgary-chooses-nenshi-as-new-mayor-1.867664. Chang, C.P., Berdiev, A.N., 2015. Do natural disasters increase the likelihood that a government is replaced? Appl. Econ. 47, 1788e1808. Cole, S., Healy, A., Werker, E., 2012. Do voters demand responsive governments? evidence from indian disaster relief. J. Dev. Econ. 97, 167e181. Dodds, G.G., 2015. This was no act of god: disaster, causality, and politics. Risk, Hazard Crisis Public Policy 6, 44e68. Downs, A., 1957. An Economic Theory of Democracy. Harper and Row, New York. Environment Canada, 2014a. Canadian climate Normals. http://climate.weather.gc. ca/climatenormals/. Environment Canada, 2014b. Daily Data Report for June 2013. http://climate. weather.gc.ca/climatedata/. Fiorina, M.P., 1981. Retrospective Voting in American National Elections. Yale University Press, New Haven. Gasper, J.T., Reeves, A., 2011. Make it rain? retrospection and the attentive electorate in the context of natural disasters. Am. J. Political Sci. 55, 340e355. Gelman, A., Katz, J.N., Bafumi, J., 2004. Standard voting power indexes do not work: an empirical analysis. Br. J. Political Sci. 34, 657e674. Gerber, A.S., Green, D.P., Larimer, C.W., 2008. Social pressure and voter turnout: evidence from a large-scale field experiment. Am. Political Sci. Rev. 102, 33e48. Government of Alberta, 2013. Communities Recovering from States of Local Emergency. http://alberta.ca/rainfall-impacted-areas.cfm/. Government of Alberta, 2014. Disaster Recovery Programs. http://alberta.ca/ 2013disasterrecoveryprograms.cfm. Green, D.P., Gerber, A.S., 2010. Introduction to social pressure and voting: new experimental evidence. Polit. Behav. 32, 331e336. Healy, A., Malhotra, N., 2010. Random events, economic losses, and retrospective voting: implications for democratic competence. Q. J. Political Sci. 5, 193e208. Ho, D., Imai, K., King, G., Stuart, E., 2013. MatchIt: Nonparametric Preprocessing for Parametric Casual Inference. R package version 2.4-21. Jackman, S., 2009. Bayesian Analysis for the Social Sciences. Wiley, Chichester. Keele, L., Titiunik, R., 2015. Natural experiments based on geography. Political Sci. Res. Methods 4, 65e95. Keele, L.J., 2014. Rbounds: Perform Rosenbaum Bounds Sensitivity Tests for Matched and Unmatched Data. R package version 2.1. Key, V., 1966. The Responsible Electorate: Rationality in Presidential Voting, 19361960. Harvard University Press, Cambridge. Lightbody, J., 2006. City Politics, Canada. Broadview Press, Peterborough, ON. Marsh, M., Tilley, J., 2010. The attribution of credit and blame to governments and its impact on vote choice. Br. J. Political Sci. 40, 115e134. Mayor, W., 2014. Testimonials for Naheed Nenshi, Mayor of Calgary (Canada). http:// www.worldmayor.com/contest_2014/calgary-mayor-nenshi.html. Nannestad, P., Paldam, P., 2000. What do voters know about the economy?: a study of danish data, 1990e1993. Elect. Stud. 19, 363e392. Riker, W.H., Ordeshook, P.C., 1968. A theory of the calculus of voting. Am. Political Sci. Rev. 62, 25e42. Rosenbaum, P.R., 2002. Observational Studies. Springer, New York NY. Sekhon, J.S., Titiunik, R., 2012. When natural experiments are neither natural nor experiments. Am. Political Sci. Rev. 106, 35e57. Shachar, R., Nalebuff, B., 1999. Follow the leader: theory and evidence on political participation. Am. Econ. Rev. 89, 525e547.
94
M.A. Bodet et al. / Electoral Studies 43 (2016) 85e94
Sinclair, B., Hall, T.E., Alvarez, R.M., 2011. Flooding the vote: hurricane katrina and voter participation in New Orleans. Am. Polit. Res. 39, 921e957. The Canadian Press, 2013. Nenshi Easily Reelected as Calgary Mayor; Iveson Wins in Edmonton. http://www.thestar.com/news/canada/2013/10/22/nenshi_easily_ reelect-ed_as_calgary_mayor_iveson_wins_in_edmonton.html.
Uhlaner, C.J., 1999. What the Downsian Voter Weighs: a Reassessment of the Costs and Benefits of Action. University of Michigan Press, Ann Arbor. Wood, J., 2013. Province Boosts Cost of Alberta Floods to $6 Billion. Calgary Herald, p. A4. September 25.