Weather to Vote: How Natural Disasters Shape Turnout Decisions

Abstract

Natural disasters can uproot peoples’ lives in a matter of minutes, leaving behind immeasurable hardships on the people and places that they strike. We examine the impact on voter turnout of one such force majeure in the days leading up to a midterm election. Leveraging the randomness of a rapidly developing, unpredictable Category 5 hurricane, we assemble an original dataset to examine the effects of Hurricane Michael on voting in Florida in the 2018 General Election. Our study assesses whether counties damaged by Hurricane Michael—as determined by relief policies administered by local election officials—affected voter behavior in 2018. Utilizing Difference-in-Difference (DID) models, we test whether voters registered in counties that were affected by Michael voted at rates comparable to their neighbors that were not directly impacted by the Category 5 hurricane. We also test whether voters in affected counties were more likely to alter their usual methods of voting. Our findings—that turnout was lower among those directly impacted by the storm but that early in-person voting helped to mitigate the effects—lend insight into how election administration decisions can offset the deleterious effects of a catastrophic event.

Keywords

weather hurricanes turnout Florida

Natural disasters have the power to uproot peoples’ lives in a matter of minutes. Tornadoes, floods, wildfires, and hurricanes can lead to sudden mass evacuation and untold levels of damage to affected areas. Setting aside for a moment the more significant and longer lasting human and economic costs of such “acts of God,” we are interested here in understanding the potential impact on voter turnout of such force majeure when they occur in the days leading up to an election. To gauge the turnout effects of a natural disaster, we assemble an original dataset to examine how Hurricane Michael altered voter participation in the 2018 General Election in Florida. The study of natural disasters on voter behavior more generally is relatively underdeveloped. Our study leverages a tragic event that disrupted the lives of hundreds of thousands of people in Florida’s Panhandle, including the ability of registered voters to cast their ballots.

As the most destructive storm of the decade to hit Florida, Hurricane Michael developed quickly and with considerable unpredictability. On 6 October 2018, forecasters for the National Oceanic and Atmospheric Administration’s (NOAA) National Hurricane Center (NHC) issued its first advisory on the system, indicating that it would likely develop into a tropical cyclone within 24 hours.¹ By 8 October 2018, the NHC had upgraded the system to a Category 1 hurricane, as it headed north from the Yucatan Peninsula across the Gulf of Mexico towards the Florida Panhandle. With wind speeds hitting 160 miles per hour, Michael was the first Category 5 storm to make landfall on the United States mainland since Hurricane Andrew ravaged South Florida in 1992. In the early afternoon of Wednesday, 10 October 2018, the eye of the storm made landfall near Tyndall Air Force Base and Mexico Beach in Bay County, Florida. After the dust settled, Michael was directly responsible for 16 deaths and roughly $25 billion in damage, displaced thousands of families, and left more than 400,000 people without power.²

When it comes to natural disasters and voter participation, there are two schools of thought about how a storm like Michael might affect voter turnout. Both draw from a rational choice “calculus of voting” framework (Downs 1957). First, and most obvious given the dire warnings of the imminent destruction that a Category 5 storm would bring, it is reasonable to expect that cost of voting increases for voters facing a natural disaster. In the case of Hurricane Michael, it is possible that some registered voters, who otherwise might have cast a ballot in the 6 November 2018, election, decided to prioritize their physical well-being over their civic duty. Voting in a midterm election was likely not their primary concern, especially considering the looming danger of the storm. On 9 October 2018, in the lead up to Michael hitting the Gulf coast, Governor Rick Scott spoke in ominous tones when he ordered those living in the storm’s path to immediately evacuate. “You cannot hide from this storm,” he stated on live TV from the State Emergency Operations Center in Tallahassee. “You can rebuild your home; you cannot rebuild your life.”³ On 7 October 2018, a month before Election Day, the governor declared a state of emergency for 26 counties along the Gulf Coast, from Citrus up to Escambia, which included mandatory evacuations for residents in Bay, Franklin, and Gulf counties. It is certainly plausible that registered voters residing in areas that were likely to be severely impacted by Michael would turn out at a lower rate, all else equal, than those residing in parts of the state that were not expected to be affected by the storm.

In contrast, however, it is possible that cost of voting could become lower for those registered voters in harm’s way should state or local government officials make accommodations in anticipation of imminent destructive forces, or in response to them. Such administrative changes by election officials could mitigate expected dampening effects from a natural disaster. For example, if local election authorities were able to extend early in-person voting days or hours, or ease vote-by-mail regulations, those directly affected by a natural disaster may face lower costs to cast a ballot.

Several states provide considerable discretion to local election officials when it comes to, say, how many early voting locations or days and hours of operation that early voting may be offered. In Florida, for example, early in-person voting may commence three Mondays prior to Election Day and run through the Sunday prior to Election Day, allowing voters registered in a county to cast their ballots outside their local precinct. Other Florida counties in the 2018 General Election provided pre-paid postage on return vote-by-mail envelopes as well as 24/7 drop boxes to return mail ballots. Such flexibility in voting modalities may mitigate a higher cost of voting that a natural disaster may incur. In short, it is plausible that because of election administration efforts, turnout in the affected counties was not as dampened in light of Hurricane Michael.

To be sure, both perspectives are plausible. The devastating impact of Hurricane Michael likely depressed voter turnout in the election, but the actions of local election officials to offer alternative ways to cast a ballot may have alleviated some of the higher costs of voting. We argue that some Florida registered voters in counties that were directly impacted by storm may have strategically altered their behavior, taking advantage of the flexibility in election administration, to ensure that their voting rights would not be compromised by the hurricane.

Election administration flexibility was granted by Governor Scott. On 18 October 2018, just 4 days before the start of early in-person voting and several weeks after vote-by-mail ballots were mailed out to voters, Scott issued a directive (Executive Order 18-283) that authorized eight county Supervisors of Elections (Bay, Calhoun, Franklin, Gadsden, Gulf, Jackson, Liberty, and Washington) to utilize comprehensive election-related relief. The order allowed the county election supervisors to extend early in-person voting days and hours and to relocate early voting locations. It also allowed them to relocate and consolidate Election Day polling locations as well as expedite the delivery and acceptance of mail ballots.⁴ The local officials took full advantage of the executive order. For example, Bay County Supervisor of Elections, Mark Andersen, vigorously defended his decision to allow voters in his county to fax their return VBM ballots. Pointing to Michael’s devastating impact on his county, Andersen decided to accept faxed mail-in ballots despite the fact that doing so directly contravened state law.⁵ Setting any partisan motivations for such actions to the side, it is clear that the creative efforts by the eight local election officials whose counties were in Matthew’s path were designed to make sure that every vote would count.

Considering the official state and local administrative efforts to lower the cost of voting for residents in the counties directly in the eye of the storm, we argue that these measures likely did offset the expected lower turnout of registered voters in the affected counties. It is our aim to determine whether registered voters who faced extreme disruption and devastation due to a natural disaster indeed turned out to vote, and whether the last-minute responses by state and local election officials in anticipation of Hurricane Michael abetted the decision of some registered voters to alter their method of voting. We utilize Difference-in-Difference (DiD) models to test whether voters registered in counties that were affected by Michael turned out in the 2018 General Election at rates comparable to their neighbors who were not directly impacted by the hurricane. We also test to determine whether voters in affected counties altered their methods of voting. To preview our results, we find that while overall turnout indeed was depressed in counties affected by the Category 5 storm, an increase in early in-person voting likely mitigated the hurricane’s negative impact on voting.

Whether (and How) to Vote

The scholarly literature on the cost of voting is considerable (Li, Pomante II and Schraufnagel 2018), with recent studies showing the dampening effect of institutional rules adopted by state actors, such as strict voter ID laws, citizenship qualifications, and changing polling locations (Alvarez, Bailey and Katz 2011; Amos, Smith and Ste. Claire 2017; Brady and McNulty 2011; Biggers and Smith 2018; Hajnal, Lajevardi and Nielson 2017; Haspel and Knotts 2005). Other scholars have examined the expansive effects on turnout when such barriers are lowered, such as the adoption of same-day registration, early in-person voting, and no-excuse vote-by-mail options (Herron and Smith 2014; Leighley and Nagler 2013; Menger, Stein and Vonnahme 2018), as well as added density of early in-person voting sites in a jurisdiction (Fullmer 2015).

Concerns over how weather might affect the calculus of voting, of course, are not new (Campbell 1960). Yet it is important to note that natural disasters should not be considered as ordinary weather events. Hurricanes, specifically, are extreme storms that have sustained winds of at least 74 miles per hour and are typically accompanied by thunderstorms.⁶ Demographers have long tracked the impact of hurricanes, focusing largely on property damages and population displacement. After Hurricane Andrew devastated South Florida in 1992, Smith and McCarty (1996) designed an innovative survey utilizing a network approach to contact those who withstood the hurricane and those who were displaced to determine the extent of damage to the housing stock as well as the diffusion of the affected population. In their subsequent analysis of the evacuation patterns that resulted from four hurricanes that hit the Sunshine State in 2004, Smith and McCarty (2009, 142) found that “many people did not evacuate even when faced with an imminent hurricane threat,” as “[s]ome doubted the severity of the threat or believed their current locations were safe.” Such thinking makes it difficult for officials charged with emergency management to develop evacuation plans, as residents evaluate the proximity and strength of the storm when determining to evacuate (Lindell, Lu, and Prater 2005).

Seeking to expand upon this literature, a growing number of scholars have assessed the impact of natural disasters on voter turnout. The swell in scholarship on this topic should not come as a surprise—even if such catastrophic events are infrequent—given the rise in concern over the impact of climate change and weather-related occurrences. Nationally, as well as globally, myriad studies indicate that rises in global temperatures will likely increase the risk of drought and heighten the intensity and frequency of different storms (Van Aalst 2006; Wuebbles et al. 2014; Brooks 2013; Ibarrarán et al. 2009); in turn, there may well be more displacement of people who anticipate and experience these disasters. As this trend continues, it will be ever more crucial for government officials, including election administrators, to be prepared to face challenges to civic participation.

Scholars who have examined whether weather affects voter behavior have largely investigated the impact of rainfall on turnout. While certainly a regular occurrence across most of the country, and typically not catastrophic, rain can certainly dampen the civic spirit and in doing so, subdue voter turnout. Drawing on meteorological data that provide Election Day estimates of precipitation across US counties, Gomez, Hansford and Krause (2007) improve on previous studies (e.g., Knack 1994; Merrifield 1993), finding that the incidence of rain, and to a lesser degree, snow, decreases turnout in presidential elections (but see Horiuchi and Kang 2018). Also utilizing county-level data on U.S. presidential elections from 1952–2012, Fujiwara, Meng, and Vogl (2016) find that rainfall arriving either on Election Day or on its immediately preceding days decreases turnout, with the effects stronger in poorer and more rural areas, but Fraga and Hersh (2010) find that Election Day rainstorms impact turnout only in non-competitive elections.

Other scholars, who also draw exclusively on aggregate data, have examined voter participation in the aftermath of more calamitous natural disasters. In their study of voter turnout in Calgary, Canada, following a major flood in 2013, Bodet, Thomas, and Tessier (2016) find that although rising water displaced roughly 7% of the city’s population, it had a negligible effect on voter turnout. Utilizing a comprehensive approach incorporating numerous data sources and statistical techniques, Sinclair, Hall, and Alvarez (2011) find mixed results in their study of turnout in the 2006 mayoral election in New Orleans following the destruction wrought by Hurricane Katrina in 2005. Others have probed the effects of Hurricane Sandy on voter participation in states along the Atlantic coast that were affected by the super storm when it made landfall in late October 2012. Drawing on district and city block data, Lasala-Blanco, Shapiro, and Rivera-Burgos (2017) analyzed registered voters in New York City who were affected by Hurricane Sandy in the 2012 presidential election, finding no difference in voter turnout between areas directly affected by the hurricane and those who were not, regardless of the respective district’s competitiveness. Stein (2015), on the other hand, reports that the 220 counties designated by the Federal Emergency Management Agency (FEMA) after Sandy swept through the Northeast saw a 2.8 percentage point decrease in turnout between the 2008 and 2012 presidential elections, relative to turnout in other counties across the country. He also found that the proportion of the total number of votes cast by early in-person voters had a significant and positive effect on overall turnout in counties most negatively affected by Sandy. Meanwhile, other convenience methods of voting, most notably vote-by-mail, did not have similarly positive effects on voter turnout. In another study, Debbage et al. (2014) find that voter turnout decreased between 2008 and 2012 in Connecticut and New Jersey counties and municipalities that were hit by Hurricane Sandy, suggesting that the drop-off in turnout was caused by the close proximity of the election to the storm itself.

Political calculations are also important to take into consideration when analyzing disaster response. As Arceneaux and Stein (2006) show in the case of Hurricane Katrina in Louisiana, voters can hold local officials accountable if they perceive that their elected representatives mishandled their response to natural disasters. They also found that ascriptions of responsibility for disaster response can be molded by whether voters reside in hard-hit areas or possess knowledge about local politics. In addition, Healy and Malhotra (2009) find that voters reward the incumbent presidential party for disaster relief spending, but not for disaster preparedness spending.

Overall, the prevailing evidence from these studies suggests that costs related to voting rise in the wake of precipitation and extreme weather events, leading to a decrease in voter participation. Yet, it remains difficult to generalize from these studies. First, the negative effects on turnout from typical precipitation (either rain or snow) can hardly be equated to the devastation and disruptions caused by natural disasters. Rainfall, or snowfall, in certain parts of the country during a national election is routine; residents in these parts are likely routinized to dress accordingly when choosing to exercise their franchise. Second, the devastation caused by natural disasters is not equable—they involve unique circumstances—making any generalization from a handful of case studies problematic. Third, all of these studies rely on aggregate data, allowing only an assessment of voter turnout in an affected area as share of the overall registered voters in a given jurisdiction (e.g., a county or a city). As such, they are limited by the ecological inference problem, as it is difficult to determine which types of registered voters were more or less affected by the severe weather. Finally, none of these studies considers that Election Day in most American states no longer falls on a single day; rather, with the widespread adoption of convenience voting, casting a ballot takes place over a month’s time and is offered in multiple modalities (Biggers and Hanmer 2015).

Data and Methods

We now turn to our study, which leverages a natural experiment to estimate whether Hurricane Michael decreased individual-level voter turnout or altered voters’ methods of voting in the 2018 General Election in Florida. Natural experiments do not involve an explicit random assignment process, so any causal inferences are subject to greater uncertainty than are studies that deliberately assign subjects to treatments at random (Gerber and Green 2012). With this in mind, we draw on other studies that utilize election administration data and conditions to create as-if random variation in the assignment of control and treatment cases across geographic boundaries in a single state (Hood III and Bullock III 2011; Keele and Titiunik 2016; Walker, Herron, and Smith 2019; Amos, Smith, and Claire 2017). In doing so, we are able to hold constant numerous factors that might condition an individual’s likelihood to turn out by creating a panel dataset in a single state to evaluate turnout and voting method over time in response to a natural disaster.

While residents along the Sunshine State’s coastal counties are especially likely to face the risks that annually come with hurricane season, voter turnout and method of voting should not be affected year-to-year given how relatively infrequently such major storms hit any one area. In other words, there is no evidence that people decide whether to vote, or select their method of voting, by calculating the odds of a hurricane hitting their residence; rather, voters in these areas are likely to change their electoral behavior only if a major natural disaster, such as a hurricane, actually develops and is likely to affect them directly.⁷ Furthermore, as with the potential exogenous cost of precipitation on turnout, we assume that the counties that were impacted by the major hurricane are assigned at random.

Utilizing multiple snapshots of publicly available individual-level administrative data from Florida’s statewide voter files, and leveraging county boundaries, we offer a series of DiD models that estimate voter turnout and changes in method of voting across the 2016 and 2018 General Elections in Florida. We assess whether voters in counties that were damaged by Hurricane Michael—as determined by relief policies administered by state and local election officials—altered their turnout decision and vote method in 2018 compared to 2 years earlier.⁸

Specifically, we utilize a snapshot of the Florida statewide voter registration file (from 1 January 2019), which we merge (using the unique voter ID number) with the statewide voter history file (also from 1 January 2019), to include the vote history of voters in the 2016 and 2018 General Elections.⁹ Conditioning on the 2019 voter file means that there will be marginally fewer voters in a particular county than there were in previous elections due to registered voters attriting from the file over time. Florida’s statewide voter file includes several demographic records which we utilize as control variables in our DiD models. Following others who have analyzed turnout in Florida using the statewide voter file (Herron and Smith 2014; Shino and Smith 2018), we include a voter’s race/ethnicity, party registration, gender, age, and past vote history. All of these well-established factors are more generally known to impact voter turnout (Hill and Leighley 1999; Leighley and Nagler 2013), and we use them as control variables in our DiD models. We recode age into four dummy variables, 18–29, 30–44, 45–64, and 65–104, excluding the oldest cohort from our models. Gender is coded as a binary variable, with male as 1 and female and others with an unreported gender as 0. We code party registration as a series of dummy variables, excluding Republicans from our models. Lastly, we include dummy variables the race/ethnicity of registrants, with white registrants as the reference group.

It is important to acknowledge again that our data drawn from the statewide voter file is an unbalanced panel dataset, as some individuals registered (and thus entered into the dataset) after the November 2016 election, and some registrants exited the dataset following the November 2016 election. We exclude from our analysis individuals who were not yet old enough to vote in 2016 (but who were in the January 2019 voter file). We also filter our dataset to only include registered voters who stayed in the same county between January 2017 and January 2019. As such, our panel dataset includes all voters who were registered in Florida as of 1 January 2019, who were also registered prior to the 2016 General Election in Florida (according to the 1 January 2017 voter file), and who resided in the same county in both the 2016 and 2018 elections.

Our unit of analysis in the DiD models is registrant-year, and our dependent variables are binary: voting in the 2018 election and the method of voting in that same election.¹⁰ In each model, we consider registered voters to be in the treatment group if they reside in a county that received relief policies from the state and in the control group if they reside in a county adjacent to a treated county. That is, registered voters living in the eight treated Florida counties eligible for the relief policies established by Governor Scott’s 18 October 2018, Executive Order 18-283, are coded as one, and voters registered in the four adjacent control counties are coded as zero, as shown in Table 1, along with the number of registered voters in the two groups. Figure 1 displays the treatment and control counties visually, with the eight Florida Panhandle counties colored in yellow specifying the treatment group, and the four counties colored in green specifying the control group.

Table 1.

Treatment Group Characteristics.

	No of counties (Treatment)	No of counties (Control)	No of registrants (Treatment)	No of registrants (Control)
Relief policy	8	4	214,309	258,834

Figure 1.

Florida counties, executive order 18-283.

We specify our generic linear DiD model as follows

\begin{array}{l} y_{i c t} = γ_{c} + λ_{t} + τ I_{c t} + β x_{’}^{ict} + u_{i c t} \\ = α + γ_{1} 1 {c = 1} + λ_{1} {t = 2018} \\ + τ (1 {c = 1} \cdot 1 {t = 2018}) + β x_{’}^{ict} + u_{i c t} \end{array}

where yict is a dummy variable that denotes the turnout/voting method of voter i in county c in time t, and $γ_{c}$ is a dummy variable that captures potential differences between the control and treatment counties prior to the hurricane. $λ_{t}$ is a time-period dummy for the election year 2018 and controls for factors that might have affected voter behavior in the absence of the intervention. The coefficient of interest, $τ$ , on the interaction term I_ct represents the hurricane’s effect on voter behavior in 2018, which is the average treatment effect on the treated (ATT). The indicator function $1 {\cdot}$ takes the value of 1 and zero otherwise. We also control for individual-level characteristics $x_{i c t}$ , and u_ict is the stochastic component of the model. We estimate the linear DiD model using ordinary least squares (OLS). The indicator function, 1{c = 1}, takes a value of one if the registrant resides in the treated county c = 1, and the indicator function for time represented by 1{t = 2018} takes a value of one if year is 2018.

We are particularly interested in whether relief policies that were implemented to mitigate the disastrous effects of Hurricane Michael had any relationship to voter turnout or method of voting in the 2018 General Election. First, we expect that overall turnout in the 2018 midterms was depressed among voters in counties that were affected by the hurricane, regardless of their respective voting methods in 2016. We base this hypothesis on existing literature that has established a negative relationship between higher costs of voting, including adverse weather conditions, and turnout. If this relationship does not hold, it indicates that the window for intense weather events to affect turnout, even powerful and destructive hurricanes such as Michael, is relatively small. If our hypothesis is supported, it is possible that individuals in the treatment counties who typically vote-by-mail were not able to request or return their ballot in time. Similarly, voters who were planning to vote on Election Day might have faced difficulties in exercising their franchise amidst the chaotic circumstances of the storm.

We also theorize about how voters may have altered their method of voting in 2018 when compared to their voting method in 2016, due to circumstances related to the storm. If, for example, those affected by the storm who voted by mail in 2016 were less likely to vote at all in 2018, then this would also likely mean that 2016 mail voters in affected counties were less likely to vote a mail ballot again in 2018. However, committed voters in this group could have plausibly found it more convenient, if not absolutely necessary, to vote early in-person, due to greater flexibility in location and time, if their precinct-specific polling locations were damaged by the storm. This would further suggest that 2018 Election Day turnout was dampened in these counties. Finally, since Governor Scott’s relief directive included expedited delivery and acceptance of mail ballots, it is possible that individuals affected by these policies were more likely to shift to mail voting.

To determine whether early in-person voting is a preferred method of voters in such strenuous circumstances, we test each respective 2016 voting method against 2018 early in-person turnout. If we consistently find positive relationships, this signals that the relief policies offered after Michael were shrewd and should serve as a solid foundation on which other state and local election officials could offer if ever in a similar situation.

We should note that a potential shortcoming of using a DiD approach in this analysis is that there may exist systematic differences between registered voters in the control and treatment counties that might affect voter behavior and are not accounted for in the model formulation and that differences between the two groups are temporally stable (Wing, Simon, and Bello-Gomez 2018). As such, we compare the demographic compositions of our treatment and control counties over the course of four general elections (2012, 2014, 2016, and 2018) and find that the differences between the two groups were quite similar over the time period.¹¹ In other words, exposure to the treatment is not associated with changes in the distribution of the covariates. Simple covariate balancing tables do not ensure the internal validity of our research design, but they do offer some validity in support of the parallel trend assumption, that is, that in the absence of treatment, the unobserved differences in the outcome variable (turnout or vote method) between treatment and control groups are equivalent over time (Angrist and Pischke 2008). This is the most essential assumption to ensure internal validity for the DiD models, and it is also the most difficult assumption to meet. We test this assumption using data from the 2012 and 2014 Florida voter files and find that it largely holds for the DiD models.¹²

We also conduct additional falsification tests in the 2018 General Election using gender (rather than relief policy) as the treatment variable. The intuition behind this set of tests is that gender, or any other covariate, should not be affected by the disaster. As expected, the results from these placebo tests consistently show that the magnitudes of our DiD variable are substantially lower than those from our tests that use the relief treatment variable.¹³ According to Wing, Simon and Bello-Gomez (2018, 460), “it is sensible to consider the magnitude of the change in composition rather than the pure statistical significance of the coefficient estimate.” Thus, we believe that the results from our placebo tests further bolster the validity of our research design. Figure 2 displays a visual representation of the parallel trend assumption, comparing overall turnout data between counties that received relief policies from the state government and those that did not over the course of four general elections. Clearly, the differences in turnout between the treatment and control groups between 2012 and 2016 are relatively constant, but after Hurricane Michael hit the Florida Panhandle we observe a noticeable drop in turnout among voters in the treatment counties relative to those in the control group.

Figure 2.

Parallel trend assumption test: overall turnout by election year, counties that received relief policies and those that did not.

Effects of Executive Order 18-283 on 2018 General Election Turnout and Voting Methods

As previously mentioned, Governor Scott issued a directive on 18 October 2018 that expanded access to early voting methods (both mail voting and early in-person voting) for voters in Bay, Calhoun, Franklin, Gadsden, Gulf, Jackson, Liberty, and Washington counties.¹⁴ This executive order empowered local Supervisors of Elections to extend the number of early voting days, establish additional early voting locations, and expedite the delivery and acceptance of mail ballots. Moreover, “the restriction on vote-by-mail ballots being forwarded to a different address” was waived by the Governor, along with other provisions, in order to “help displaced voters cast a ballot.” The policies were specifically crafted to address what local election supervisors reported to state officials as difficulties they faced administering the midterm elections in the aftermath of Michael. These obstacles included damage to polling centers, lack of electricity, and disruptions in telecommunications services.

The eight counties for which these policies were designed serve as our treatment counties, and the four counties that immediately border these treatment counties serve as our controls. In Figure 3, we plot the relief policies’ ATT on voter turnout and voting methods in 2018. We construct 10 OLS models to test our hypotheses regarding post-hurricane effects. We control for registrants’ demographics, party affiliation, and county-fixed effects, and we cluster standard errors by county. The models estimate the effect that the relief policies possibly had on the counties that received them, in comparison to neighboring counties that were not granted these same privileges.¹⁵

Figure 3.

Effects of Executive Order 18-283 on 2018 General Election Turnout and Voting Methods.

Most strikingly, we observe that registered voters in the treated counties who cast ballots in the 2016 election were 9.2 percentage points less likely to vote on Election Day in 2018 relative to those in proximate counties. By contrast, voters in the treated counties who cast ballots in 2016 were 4.3 percentage points more likely to vote prior to Election Day in 2018 than were voters in the control counties. Examining each respective early voting method more closely, we see that those who cast ballots in 2016 in the treated counties were 6.7 percentage points more likely to vote early in-person in 2018, but were neither more nor less likely to vote-by-mail in 2018, compared to comparable voters in the control counties.

Effects on early in-person voting are even stronger among those who did not vote by that method in 2016. For instance, voters in the treated counties who voted by mail in 2016 were 5.0 percentage points more likely to vote early in-person than were voters from bordering counties who voted by mail in 2016, and those who voted on Election Day in 2016 in counties receiving relief were 7.9 percentage points more likely to vote early in-person in 2018 compared to comparable voters in the four neighboring counties that did not receive relief. This provides strong evidence that Governor Scott’s relief policies facilitated turnout in treatment counties. In short, voters who did not cast early in-person ballots in 2016 and who experienced considerable hurricane damage were much more likely to vote early in-person in 2018. Even those in treatment counties who voted early in-person in 2016 were 2.7 percentage points more likely to vote early in-person again in 2018, compared to similar voters registered in the four control counties.

Voters registered in counties that were devastated by Hurricane Michael who had cast mail ballots in 2016, however, did not fare as well. We find that among voters who cast mail ballots in 2016, there was a 1.9 percentage point drop in voting by mail in 2018 for those registered in the eight treated counties, compared to those in control counties. Indeed, those who voted by mail in 2016 had the most trouble voting, by any method, in 2018. Voters who had cast their ballots by mail in 2016 were 8.9 percentage points less likely to vote in 2018 than were similar voters in neighboring counties. Finally, we find a similarly large 7.2 percentage point drop-off in overall 2018 turnout among those who voted in 2016 in relief-receiving counties, relative to voters who cast ballots in adjacent counties that did not receive relief.¹⁶

Discussion

The results of our study of the impact of a natural disaster on voter turnout and voting methods are bountiful and should stimulate further research into this specific case study and others as well. Undoubtedly, Hurricane Michael had substantial effects on voter behavior in Florida’s 2018 General Election, both in terms of anticipation of the storm and also in terms of actual damage wrought upon devastated areas. Expanding upon previous studies that exclusively use aggregate-level data to make broad conclusions regarding the effects of natural disasters on voter turnout, we introduce to the literature new relationships between the various voting methods with our individual-level DiD models. We do so by focusing on the eight counties that faced the most damage, as suggested by Governor Scott’s Executive Order 18-283.

Notably, we find strong negative effects on overall turnout, regardless of voting method in 2016. We find even stronger negative effects on 2018 Election Day turnout, which supports our hypothesis that affected voters were less likely to wait until November to exercise their franchise. There was, however, a significant increase in early in-person turnout among voters that received relief from the state, regardless of their method of voting in 2016. Meanwhile, effects on mail voting were mixed, with those in treated counties who voted by mail in 2016 notably less likely to do so again in 2018. These findings are consistent with those of Stein (2015), who suggests that voters affected by Hurricane Sandy did not have sufficient time to request a mail ballot. In other words, voters in affected counties would have had to request a mail ballot well before the threat of Sandy was present. Voters affected by Hurricane Michael likely faced similar circumstances when attempting to utilize this particular method of voting. In both states, early voting was available well after each storm made landfall, and the data strongly suggest that voters took advantage of this modality more so than any other. It appears that early in-person voting was the most effective method to vote for those who experienced the most damaging effects of the storm and that the relief policies offered by the state were well-advised. Election administrators would be prudent to take note of these results when preparing for and reacting to future disasters, particularly in the age of COVID-19.

Early in-person voting has surged in recent years.¹⁷ Scholars have consistently found a steady increase in the share of early votes (including mail-in ballots) cast since the 1990s,¹⁸ and scholars have also found that counties with a higher density of early in-person voting locations have higher turnout (Fullmer 2015). Convenience voting methods have expanded the traditional concept of Election Day to include several weeks of voting. Our study suggests that early in-person voting in particular minimizes the costs of voting in the event of natural disasters.

Following the 2018 General Election, county election officials in Florida’s Panhandle successfully lobbied newly elected Governor Ron DeSantis to follow in the footsteps of his predecessor to allow an expansion of early voting locations in the 2020 election cycle.¹⁹ Bay County, for example, offered 13 early voting locations in the 2020 General Election, up from just six in 2018.²⁰ However, in December 2020, following the November presidential election, DeSantis did not renew the executive order, eliminating in-person early voting at Super Voting Centers in Bay and Gulf counties.²¹

While voting by mail also likely reduces the costs of voting for some registered voters, we find that this was not the case for those directly impacted by Hurricane Michael. The increased costs that come with evacuating the address to which a mail ballot is sent, among other difficulties in returning a mail ballots amidst the destruction and chaos brought about by the storm, appear to have taken a toll on those who cast their ballots by mail two years earlier. The additional costs of voting by mail during a natural disaster likely account for why we consistently find a negative relationship between those who voted by mail in 2016 and subsequent turnout in 2018 among individuals in the treated counties.

While drawing on administrative data has certain advantages for studying turnout and method of voting, it has its limits. Most notably, we are unable to directly determine why some of the relationships we discover might exist. Future studies might utilize surveys or field interviews to better pinpoint the motivations of registered voters affected by natural disasters. Another research design might leverage Hurricane Michael’s impact on voter behavior in southwest Georgia, which experienced wind gusts up to 115 mile/hour and saw 400,000 people lose power.²² Others might expand on our study to include voter turnout and voting methods in the Peach State in the 2016 and 2018 General Elections. Lastly, we recognize that our findings are from a single storm affecting an election in a single state; as such, we remain circumspect in generalizing our findings to other settings or elections without further study.

Natural disasters have immeasurable consequences on the people and places that they strike. In no way is our study meant to detract from the loss of life and property in the Florida Panhandle caused by Hurricane Michael. Putting our study in this broader perspective, though, we do think it contributes to the growing literature focusing on outcomes related to voter behavior, as we provide compelling evidence that general turnout is likely dampened when natural disasters strike, but that election administrators may be able to mitigate these effects by expanding access to convenience voting methods. Specifically, we find that early in-person voting emerged as the most convenient method for registered voters directly affected by Hurricane Michael. Our findings should help guide future policy decisions in Florida and other states that share the goal of ensuring that all voters are able to exercise their right to vote in the face of extreme weather conditions or other disruptive events.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Daniel A. Smith

Supplemental Material

Supplemental material for this article is available online.

Notes

References

Alvarez

R. Michael

Bailey

Delia

Katz

Jonathan N.

. 2011. “An empirical Bayes Approach to Estimating Ordinal Treatment Effects.” Political Analysis 19 (1): 20-31.

Amos

Brian

Smith

Daniel A.

Claire

Casey Ste.

. 2017. “Reprecincting and Voting Behavior.” Political Behavior 39 (1): 133-156.

Angrist

Joshua D.

Pischke

Jörn-Steffen

. 2008. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton: Princeton University Press.

Arceneaux

Kevin

Stein

Robert M.

. 2006. “Who Is Held Responsible When Disaster Strikes? The Attribution of Responsibility for a Natural Disaster in an Urban Election.” Journal of Urban Affairs 28 (1): 43-53.

Baringer

Anna

Herron

Michael C.

Smith

Daniel A.

. 2020. “Voting by Mail and Ballot Rejection: Lessons from Florida for Elections in the Age of the Coronavirus.” Election Law Journal: Rules, Politics, and Policy 19 (3): 289-320.

Biggers

Daniel R.

Smith

Daniel A.

. 2018. “Does Threatening their Franchise Make Registered Voters More Likely to Participate? Evidence from an Aborted Voter Purge.” British Journal of Political Science 50 (3): 1-22.

Biggers

Daniel R.

Hanmer

Michael J.

. 2015. “Who Makes Voting Convenient? Explaining the Adoption of Early and No-Excuse Absentee Voting in the American states.” State Politics & Policy Quarterly 15 (2): 192-210.

Bodet

Marc Andre

Thomas

Melanee

Tessier

Charles

. 2016. “Come Hell or High Water: An Investigation of the Effects of a Natural Disaster on a Local Election.” Electoral Studies 43: 85-94.

Brady

Henry E.

McNulty

John E.

. 2011. “Turning Out to Vote: The Costs of Finding and Getting to the Polling Place.” American Political Science Review 105: 115-134.

10.

Brooks

Harold E.

2013. “Severe Thunderstorms and Climate Change.” Atmospheric Research 123: 129-138.

11.

Campbell

Angus

. 1960. “Surge and Decline: A Study of Electoral Change.” Public Opinion Quarterly 24 (3): 397-418.

12.

Debbage

Gonsalves

Shepherd

J. M.

Knox

J. A.

. 2014. “Superstorm Sandy and Voter Vulnerability in the 2012 US Presidential Election: A Case Study of New Jersey and Connecticut.” Environmental Hazards 13 (3): 181-199.

13.

Downs

Anthony

. 1957. “An economic Theory of Political Action in a Democracy.” Journal of Political Economy 65 (2): 135-150.

14.

Fujiwara

Thomas

Meng

Kyle

Vogl

Tom

. 2016. “Habit Formation in Voting: Evidence from Rainy Elections.” American Economic Journal: Applied Economics 8 (4): 160-188.

15.

Fullmer

E. B.

. 2015. “Early Voting: Do More Sites Lead to Higher Turnout?.” Election Law Journal: Rules, Politics, and Policy 14 (2): 81-96.

16.

Gerber

Alan S.

Green

Donald P.

. 2012. Field Experiments: Design, Analysis, and Interpretation. New York: WW Norton.

17.

Gomez

Brad T.

Hansford

Thomas G.

Krause

George A.

. 2007. “The Republicans Should Pray for Rain: Weather, Turnout, and Voting in US Presidential Elections.” The Journal of Politics 69 (3): 649-663.

18.

Hajnal

Zoltan

Lajevardi

Nazita

Nielson

Lindsay

. 2017. “Voter Identification Laws and the Suppression of Minority Votes.” The Journal of Politics 79 (2): 363-379.

19.

Haspel

Knotts

H. G.

. 2005. “Location, Location, Location: Precinct Placement and the Costs of Voting.” The Journal of Politics 67 (2): 560-573.

20.

Healy

Andrew

Malhotra

Neil

. 2009. “Myopic Voters and Natural Disaster Policy.” American Political Science Review 103 (3): 387406.

21.

Herron

Michael C.

Smith

Daniel A.

. 2014. “Race, Party, and the Consequences of Restricting Early Voting in Florida in the 2012 General Election.” Political Research Quarterly 67 (3): 646-665.

22.

Hill

Kim Quaile

Leighley

Jan E.

. 1999. “Racial Diversity, Voter Turnout, and Mobilizing Institutions in the United States.” American Politics Quarterly 27 (3): 275-295.

23.

Hood

M. V.

III Bullock

C. S.

III. . 2011. “An Examination of Efforts to Encourage the Incidence of Early in-Person Voting in Georgia, 2008.” Election Law Journal: Rules, Politics, and Policy 10 (2): 103-113.

24.

Horiuchi

Yusaku

Kang

Woo Chang

. 2018. “Why Should the Republicans Pray for Rain? Electoral Consequences of Rainfall Revisited.” American Politics Research 46 (5): 869-889.

25.

Ibarrarán

María Eugenia

Ruth

Matthias

Ahmad

Sanjana

London

Marisa

. 2009. “Climate Change and Natural Disasters: Macroeconomic Performance and Distributional Impacts.” Environment, Development and Sustainability 11 (3): 549-569.

26.

Keele

Luke J.

Titiunik

2016. “Natural Experiments based on Geography.” Political Science Research and Methods 4 (1): 65-95.

27.

Knack

Steve

. 1994. “Does Rain Help the Republicans? Theory and Evidence on Turnout and the Vote.” Public Choice 79 (1–2): 187-209.

28.

Lasala-Blanco

Shapiro

R. Y.

Rivera-Burgos

. 2017. “Turnout and Weather Disruptions: Survey Evidence from the 2012 Presidential Elections in the Aftermath of Hurricane Sandy.” Electoral Studies 45: 141-152.

29.

Leighley

Jan E.

Nagler

Jonathan

. 2013. Who Votes Now? Demographics, Issues, Inequality, and Turnout in the United States. Princeton: Princeton University Press.

30.

Pomante

M. J.

Schraufnagel

. 2018. “Pomante II and Scot SchraufnagelCost of Voting in the American States.” Election Law Journal: Rules, Politics, and Policy 17 (3): 234-247.

31.

Lindell

M. K.

J.-C.

Prater

C. S.

. 2005. “Household Decision Making and Evacuation in Response to Hurricane Lili.” Natural Hazards Review 6 (4): 171-179.

32.

Menger

Andrew

Stein

Robert M.

Vonnahme

Greg

. 2018. “Reducing the Undervote with Vote by Mail.” American Politics Research 46 (6): 1039-1064.

33.

Merrifield

John

. 1993. “The Institutional and Political Factors that Influence Voter Turnout.” Public Choice 77 (3): 657-667.

34.

Shino

Enrijeta

Smith

Daniel A.

. 2018. “Timing the Habit: Voter Registration and Turnout.” Electoral Studies 51: 72-82.

35.

Sinclair

Hall

T. E.

Alvarez

R. M.

. 2011. “Flooding the Vote: Hurricane Katrina and Voter Participation in New Orleans.” American Politics Research 39 (5): 921-957.

36.

Smith

Stanley K.

McCarty

Christopher

. 1996. “Demographic Effects of Natural Disasters: A Case Study of Hurricane Andrew.” Demography 33 (2): 265-275.

37.

Smith

Stanley K.

McCarty

Chris

. 2009. “Fleeing the Storm (s): An Examination of Evacuation Behavior during Florida’s 2004 Hurricane Season.” Demography 46 (1): 127-145.

38.

Stein

Robert M.

. 2015. “Election Administration during Natural Disasters and Emergencies: Hurricane Sandy and the 2012 Election.” Election Law Journal 14 (1): 66-73.

39.

Van Aalst

M. K.

. 2006. “The Impacts of Climate Change on the Risk of Natural Disasters.” Disasters 30 (1): 5-18.

40.

Walker

Hannah L.

Herron

Michael C

Smith

Daniel A.

. 2019. “Early Voting Changes and Voter Turnout: North Carolina in the 2016 General Election.” Political Behavior 41 (4): 841-869.

41.

Wing

Coady

Simon

Kosali

Bello-Gomez

R. A.

2018. “Designing Difference in Difference Studies: Best Practices for Public Health Policy Research.” Annual Review of Public Health 39 (1): 453-469.

42.

Wuebbles

Donald J.

Kunkel

Kenneth

Wehner

Michael

Zobel

Zachary

. 2014. “Severe Weather in United States under a Changing Climate.” Eos, Transactions American Geophysical Union 95 (18): 149-150.