After a Storm Come Votes: Identifying the Effects of Disaster Relief on Electoral Outcomes

Abstract

The retrospective voting theory suggests that citizens vote for governing parties in response to distributive benefits. Knowing this, governments may reward voters by providing particularistic benefits—i.e., pork—prior to elections. Previous studies, however, do not account for the endogeneity. We address this problem by focusing on disaster relief and exploiting exogeneity of disaster. In particular, by using maximum hourly rainfall as an instrumental variable for disaster relief, we analyze the causal effect of disaster relief on incumbent’s electoral outcomes. Our analyses of Japanese data in the past few decades indicate that disaster relief increased governing parties’ vote share. Specifically, when the disaster relief per capita increases from zero to its mean, the predicted value of the governing parties’ vote share increases by 2.8 and 5.4% points in the lower and upper chambers, respectively. The finding is consistent with retrospective voting behavior. Moreover, our results imply that the incumbent’s electoral gain is brought about by persuading voters from oppositions to governing parties rather than by mobilizing supporters of governing parties.

Introduction

In the distributive politics literature, scholars have claimed that if a government delivers policy benefits to citizens, the recipients will return the benefits by voting for the incumbents (Kitschelt & Wilkinson, 2007; Stokes et al., 2013, ch. 7); this follows the logic of economic voting (Lewis-Beck & Stegmaier, 2007) and retrospective voting in particular (Ferejohn, 1986; Fiorina, 1981). Although the argument is straightforward, it has been difficult to empirically substantiate. Some studies have found that spending leads to incumbent’s electoral gain (Zucco, 2013) while others have not (Samuels, 2002; Stein & Bickers, 1994).

The causality of the relationship has been especially challenging to establish because elections can influence spending. Governing parties may provide pork to voters to solicit their support in the next election—behavior that has been referred to as particularism (Kriner & Reeves, 2015), electoral reciprocity (Bechtel & Mannino, 2022), and vote-purchasing behavior (Dahlberg & Johansson, 2002). If the incumbent rewards its core voters (Cox & McCubbins, 1986), the incumbent vote share positively correlates with government expenditures, and one might misinterpret that the spending increases the incumbent vote share (even though the actual direction of the causality is reverse). By contrast, if the governing parties provide pork to swing voters (Lindbeck & Weibull, 1987), the incumbent vote share negatively correlates with the expenditures, again, in the absence of the expected causality. In both cases, there is an endogenous relationship between elections and distribution, which makes it difficult to identify the causal effect of one on the other (Larcinese et al., 2013; Stein & Bickers, 1994).

While scholars have addressed this identification problem with various research designs (Levitt & Snyder, 1997; De La, 2013; Manacorda et al., 2011; Pop-Eleches & Pop-Eleches, 2012), this article focuses on the literature of natural disasters as exogenous shocks. In fact, studies indicate that horrifying disasters can actually benefit the constituencies financially and, thus, the governing parties politically. Healy and Malhotra (2009) find that the more money the United States (U.S.) spends on disaster relief, the more voters support the incumbent president. Similarly, a president (or governor) issuing a disaster declaration in response to severe weather destruction increases his/her vote share in the next election (Gasper & Reeves, 2011; Reeves, 2011). Conversely, if the president does not issue a disaster declaration despite a governor’s request, the president is punished at the polls while the governor does not lose votes (Gasper & Reeves, 2011; Healy & Malhotra, 2010). Chen (2013) documents that hurricane aid mobilizes the incumbent party’s supporters and demobilizes supporters of the opposition. These patterns are not unique to the U.S. (Bechtel & Hainmueller, 2011; Gallego, 2018).

Despite these extensive findings, the problem of endogeneity persists. Although we can argue that weather events like heavy rain and floods are exogenous and can be exploited as natural “natural experiments” (Rosenzweig & Wolpin, 2000), disaster declarations and relief allocation are human responses to the weather events, and thus they can be endogenous to electoral politics. Indeed, Kriner and Reeves (2015, ch. 4) demonstrate that presidential disaster declarations in the U.S. tend to target supportive states and districts represented by the president’s party (see also Reeves, 2011, p. 1142). Bechtel and Mannino (2022) also show that counties that are partisan strongholds receive larger amounts of disaster relief.

Moreover, previous studies do not clearly distinguish between exogenous weather events and potentially endogenous disaster relief. In the directed acyclic graph below (Fig. 1), we denote the outcome variable (governing parties’ vote share) by V, the treatment variable (disaster relief) by D, the instrumental variable (disaster such as rainfall) by R, and an unobserved confounder by u. Several studies regress electoral outcomes (V) on variables related to disaster relief (D), such as disaster declarations (Gasper & Reeves, 2011; Reeves, 2011), disaster relief spending (Bechtel & Mannino, 2022; Healy & Malhotra, 2009), and approval among hurricane aid applications (Chen, 2013). These studies implicitly assume that these variables are exogenous. Nonetheless, as Reeves (2011) and Bechtel and Mannino (2022) compellingly argue, disaster relief can be endogenous to (would-be) electoral outcomes. This means that the causal relationship is confounded by u and thus that the estimate suffers from endogeneity bias.

Fig. 1

A directed acyclic graph of rainfall, disaster relief, and vote share. We denote the outcome variable (governing parties’ vote share) by V, the treatment variable (disaster relief) by D, the instrumental variable (disaster such as rainfall) by R, and an unobserved confounder by u

Other studies address endogeneity by using weather events like floods (Bechtel & Hainmueller, 2011; Heersink et al., 2017), rain (Gallego, 2018), and tornadoes (Healy & Malhotra, 2010) as exogenous explanatory variables. Thus, they regress V on R. Although this approach may provide internally valid estimates of the weather’s effects, the estimand, the effects of R on V, may not be substantively relevant. Because the natural phenomena themselves (R) are not the distribution of benefits (D), the estimand is different from the quantity of interest, the effects of distributive benefits (D) on electoral outcomes (V).

We address these problems by using exogenous weather events (R)—maximum hourly rainfall—as an instrumental variable (IV) for disaster relief (D). We apply the IV method to the Japanese data over the past few decades. Our IV estimates show that a larger amount of disaster relief (D) leads to a higher vote share for the governing parties (V). The finding is consistent with voters’ retrospective behaviors.

Moreover, once we find distribution increases governing parties’ vote share, a natural next question is how: by mobilizing supporters of governing parties or persuading voters from oppositions to governing parties. Chen (2013) and De La (2013) support the mobilization mechanism, while Pop-Eleches and Pop-Eleches (2012) suggest both. Arguably, our results imply that the incumbent’s electoral gain is brought about by the persuasion mechanism rather than the mobilization mechanism.

Disaster Relief

Municipal Disaster Relief

Although there are several types of government expenditure relating to disasters in Japan, we focus on those for disaster recovery projects (saigai fukkyū jigyō hi in Japanese), which we call “disaster relief.” The disaster relief composes a large share of the expenditure relating to disasters.¹ According to the government’s classification, disaster relief is one of the three components of capital expenditure as well. Of the municipal-level disaster relief we study, 50.5 and 31.6% was for public civil engineering and primary industry facilities, respectively. The former includes roads (59.6%), rivers (14.5%), and fishing ports (14.1%), while the latter consists of agricultural facilities (36.1%), agricultural land (21.3%), and forestry facilities (20.3%) and others.²

Therefore, the disaster relief provides local public goods or geographically targeted club goods, which is considered as pork in the literature (Kitschelt & Wilkinson, 2007, pp. 11–13; Muller, 2007, p. 252). These goods are thought of as typical pork in the Japanese context, too (Fukui & Fukai, 1996; Scheiner, 2007). They are beneficial not only for voters but also local industries which construct and maintain these facilities, hire local workers, and may contribute funds to governing parties’ politicians (Weingast et al., 1981; Muller, 2007, p. 252; Samuels, 2002), though the disaster relief is not paid to individual victims. Pundits sometimes even criticize the political biases in the distribution of the disaster-related expenditure, saying that under the pretext of a good name of “recovery from disaster,” the government takes advantage of it for constructing the infrastructure (e.g., airports, bridges, redevelopment, (rail-)roads, and seawalls) that are not heavily affected by the disasters or involving quality improvement beyond the pre-disaster levels (Harada, 2012; Shiozaki, 2014).

The disaster relief is allocated by national, prefectural, and municipal governments. Even though our outcome variable is the results of national elections, we focus our attention to municipal disaster relief. This is because the geographical variation at the municipal level is essential for our analysis.³ Furthermore, the overwhelming majority of disaster relief is distributed by local authorities. Most national disaster relief takes the form of subsidies for municipal governments—on average, 82.3% of annual disaster relief administered by Ministry of Land, Infrastructure, Transport and Tourism (hereafter MLIT) went to prefectures and municipalities as subsidies.⁴ Prefectures also subsidize municipal disaster relief. To see the other side of the coin, on average, 75.9% of the total municipal disaster relief in a year was spent on projects subsidized by national and/or prefectural governments.⁵ The exact ratio of subsidies relative to each project’s disaster relief depends on the types of facilities being restored, the amounts of the disaster relief, and the severity of damage. For instance, once the national government agrees to subsidize a project that recovers public civil engineering facilities, the minimum of the subsidy ratio is two thirds (the remaining third is paid for by municipalities); furthermore, if additional conditions are met, the ratio can be increased even up to more than 90%.⁶ Given the large share of local disaster relief paid for by the national government, national authorities can conceivably use disaster relief for partisan objectives.

Decision Process

Ostensibly, disaster relief allocation is “programmatic” (Stokes et al., 2013). In fact, the national government has bureaucratic and technical eligibility criteria for disaster relief.⁷ For instance, in the case of rain-related disasters (except for damage to river), a necessary condition for national subsidies is that the 24 h maximum rainfall amount is not less than 80 mms or the one hour maximum rainfall amount is roughly 20 mms.⁸ Nevertheless, once we look closely at the decision process for disaster relief allocation, we find that it is subject to “pork-barrel politics.”

Discretion

Disaster relief is delivered only after the government “assesses” the extent of the damages and recognizes the damages “caused” by a disaster. Therefore, the decision processes can be subject to political maneuvering in a manner similar to the normal budgetary process.

For instance, regulations stipulate that in principle, an eligible project should only restore, not improve, facilities affected by disasters.⁹ However, the law permits an exception; if it is difficult or inappropriate to simply restore facilities, a project to construct a “necessary substitute” (e.g., improved facilities) is allowed.¹⁰ How does the government judge if it is “difficult or inappropriate” to restore the facilities? The government has guidelines that are open to interpretation. One scheme is called “disaster restoration under a single plan” (ittei keikaku niyoru saigai fukkyu or simply ittei sai in Japanese); the government can declare that damaged and undamaged parts of a facility constitute a single damaged component. With this interpretation, the government can subsidize improvements for the undamaged parts as well.¹¹

Another example of government’s discretion relates to the counting of damaged parts. If it is difficult or inappropriate to restore two or more parts of a damaged facility separately, the government can count them as a single part so that the total restoration cost is large enough to qualify for disaster relief.¹² A senior official of the Ministry of Finance wrote in his commentary on disaster relief that “this is a benign measure which regards de facto two parts as de jure one part... one should not abuse it” (Kato, 1982, p. 80, our translation). Governmental discretion determines “appropriateness.”

The third aspect of discretion is timing. Even if the government agrees to pay disaster relief in principle, it actually pays 85% of the cost in the fiscal year the disaster occurs; it usually expends the total cost over three years (Zenkoku Bōsai Kyōkai, 2018, pp. 117–120).¹³ But the pace can be accelerated if politicians push their demand. After the Great Hanshin Awaji Earthquake in January 1995, the governor Kaihara (1995, p. 141) of Hyogo Prefecture pressured the central government so that the cabinet decided to spend the total amount by March 1996.

The fourth relates to the application materials required for government subsidies. The government usually requires cost estimates from three companies. But meeting such a requirement is difficult in the immediate aftermath of a disaster due to the large volume of demands. Following Kumamoto Prefecture’s request after the Kumamoto Earthquakes in 2016, the government only required an estimate by one company (Kumamoto Ken, 2018, p. 261).

In most cases, the issue is not whether the government provides disaster relief. Rather, the issue is how much disaster relief the government delivers and how quickly it does so. Municipalities that are damaged by disasters demand the relief; the national government assesses the applications, considering the limits of its coffers. Politicians can neither distribute disaster relief to municipalities without disasters nor withhold it from municipalities hit by disasters. But they can tweak the degree to which the state meets the demands from the affected municipalities. The devil is in the details.

Lobbying

If disaster relief allocation were purely programmatic, it would be wasteful for politicians to engage in lobbying. But in reality, politicians make enormous efforts to secure disaster relief on behalf of their districts. For instance, soon after the Kumamoto Earthquakes and heavy rain hit Kumamoto Prefecture in 2016, the (vice-)governor and the (vice-)chair of the prefectural assembly repeatedly lobbied governing parties, the prime minister’s office, and national ministries in Tokyo for aid, and ministers visited the prefecture dozens of times. Thereafter, the government decided to spend reserve funds, passed supplementary budgets, relaxed requirements for disaster relief as a special measure, increased the subsidy ratio, and created a new fund to help the prefecture (Kumamoto Ken, 2018, pp. 180, 192, 199, 208, 214–215, 369–370).

Another example is Typhoon Hagibis in 2019. The Prime Minister, Minister of Disasters, and Minister of Land, Infrastructure, Transport and Tourism visited Fukushima Prefecture. The governor of the prefecture asked the national government not only to subsidize larger parts of the recovery costs but also to compensate the remaining costs by intergovernmental transfer. The governors of Ibaraki, Iwate, Miyagi, and Saitama Prefectures made similar requests.¹⁴

Delivery

Last but not least, politicians can exercise direct influence over the allocation of disaster relief by enacting legislation and altering policy implementation. For example, the 2011 Nigata-Fukushima Heavy Rain resulted in flooding and destroyed important infrastructure like the Tadami Line. However, the Act on Improvement of Railroads and Rail Tracks stipulated that the government cannot subsidize private companies that have no financial difficulties. Japan Railway East, the operator of Tadami Line, had no financial problems. Neither did they intend to restore the Tadami Line, as the costs for recovery would have been dramatically larger than any potential profits,¹⁵ A Fukushima-based politician—Ichiro Kanke—founded an all-party parliamentary group, gathering over 140 members. The group drafted and passed a reform bill, allowing the government to use disaster relief for the recovery of Tadami Line.¹⁶

In the 2009 election, Masahiko Kōmura ran for the first district of Yamaguchi Prefecture. Just a month before the election, his district suffered from heavy rain. During the campaign, Kōmura—who had once served as the chairman of Disaster Special Committee of the lower chamber—then promised to make the government recognize the disaster as a “Disaster of Extreme Severity” so that the government’s disaster relief subsidies would increase. He won the election, and a few weeks later, the disaster was designated as a Disaster of Extreme Severity. This was just a day before the long-ruling Liberal Democratic Party (LDP) ceded the office to the former opposition party, the Democratic Party of Japan (DPJ),¹⁷

Taken together, as the above examples illustrate, the decision process of disaster relief allocation is complicated and vulnerable to meddling by politicians from the governing parties.

Rainfall

While the allocation of disaster relief is likely affected by politicians through various channels, it is also subject to exogenous shocks—rainfall. In 2017, 94.2% of the disaster relief administered by MLIT addressed rain-related disasters (summer rain, typhoon, and heavy rain; Zenkoku Bōsai Kyōkai, 2018, p. 55). Although the ratio reduced to 46.0% on average for the period of 2008–2017 due to a few major earthquakes,¹⁸ rain-related disasters happened every year while major earthquakes occurred only intermittently. This suggests that rainfall is the dominant determinant of disaster relief over time.

A similar pattern can be seen with the designation of Disasters of Extreme Severity, which is analogous to the disaster declarations in the U.S. When a disaster meets certain criteria, the national government designates it as a Disaster of Extreme Severity, and the above mentioned ratio of subsidies to the disaster relief can skyrocket.¹⁹ Disasters of Extreme Severity can be declared to be either “national” or “local,” depending on the geographical scope of the disaster. Table 1 summarizes the distribution of disaster types by national and local Disasters of Extreme Severity.²⁰ Heavy rain and/or rainstorm account for over 80% of Disasters of Extreme Severity. Since these disasters are often caused by instantaneous rainfall (usually due to typhoon or seasonal rainy front), we expect that the maximum rainfall is a particularly strong predictor of disasters.

Table 1 Distribution (%) of types of Disaster of Extreme Severity

Research Design

Instrumental Variable Approach

Statistically identifying the causal effect of the disaster relief on an election poses a challenge: the disaster relief can be endogenous to elections. As we explained, the disaster relief depends on the extent to which the government recognizes the damage “caused” by disasters and the government’s “estimates” of the costs for recovering the damage. It would not be surprising if the government would take into consideration electoral concerns in the decision process. Therefore, it is imperative to account for the endogeneity of the disaster relief to elections.

To this end, we exploit the maximum rainfall in the year before an election as an IV for disaster relief. The IV design allows for valid causal inference when a few core assumptions are satisfied. First, the IV should have a strong effect on an explanatory variable (relevance). In our case, the maximum rainfall should be a strong predictor of the disaster relief. As we explained in the previous section, the predominant causes of major disasters are heavy rain and rainstorm. Because rain-related disasters are often caused by near-instantaneous downpours, the maximum hourly precipitation should better predict the disasters than annual averages (Gallego, 2018; Nikolova & Marinov, 2017). The maximum rainfall is therefore expected to increase the severity of disasters, which in turn adds to disaster relief.

Second, the IV must be conditionally independent of the outcome variable (exogeneity). In light of previous studies, which persuasively argue that political and economic factors like elections do not affect rainfall (Fraga & Hersh, 2011; Gallego, 2018; Hansford & Gomez, 2010; Henderson & Brooks, 2016; Nikolova & Marinov, 2017; Tamada, 2009), we can plausibly assume that rainfall (in particular, maximum rainfall) is exogenous. In the following analysis, we also include municipality fixed effects, which account for geographic and climatic conditions relating to rainfall and electoral outcomes. (The specification is mathematically equivalent to subtracting the average of the maximum rainfall over elections from the maximum rainfall for each election.)

Finally, the IV should not affect an outcome variable except for its effect through an explanatory variable (exclusion restriction). In our case, the maximum rainfall should not affect the electoral results in the next year except through its effect on the disaster relief. Although the previous studies find that the rainfall on election day dampens voter turnout or shifts vote shares among parties (Hansford & Gomez, 2010), we employ the rainfall in the previous year of an election, which is unlikely to affect the election directly. Furthermore, even though annual rainfall may indirectly affect the election through, for instance, agricultural production (Gallego, 2018), we use the maximum of hourly precipitation. It is unlikely that extreme precipitation in such a specific moment—an hour—would have a strong impact on an election directly or indirectly through the effects on agricultural production (see also Gallego, 2018, p. 78). As a precaution, we later conduct placebo tests to confirm that the maximum rainfall has no effect on agricultural performance. Moreover, if by any chance the maximum rainfall or its resultant disasters have any direct effect on the following year’s elections, the effect should be negative. As previous studies suggest (Achen and Bartels, 2016; Heersink et al. 2017; Heersink etal., 2022, ch. 5), voters may tend to attribute “unfortunate” disasters such as droughts, floods, and shark attacks to—and hence punish—governing parties by casting ballots for the opposition. If this were the case, our IV design would provide a conservative estimate.

Data

Unit of Observation

The unit of observation is municipality i in Japan as of the last election we analyze (October 22, 2017), and election year t. When multiple municipalities were merged before 2017, we aggregate them for all years. Since we delete 19 irregular municipalities, the number of observations in a given year is 1722 out of 1741 municipalities.²¹

The prime minister dissolved the lower chamber, and the corresponding elections were held, in years $t = 1996, 2000, 2003, 2005, 2009, 2012, 2014$, and 2017. We also examine elections to the upper chamber in years $t = 1992, 1995, \ldots , 2013$, and 2016 (the timing of the elections is fixed to every third year).²² Since our datasets are balanced panels, the numbers of observations are $13,776 (= 1,722 \times 8)$ for the lower chamber and $15,498 (= 1,722 \times 9)$ for the upper chamber.

Elections

Since 1996, the lower chamber has a two-tier electoral system. The upper tier is divided into 11 proportional-representation (PR) districts with 6–33 seats in each, while the lower tier consists of 289–300 single-member districts (SMDs). The upper chamber also has the two-tier electoral system. The upper tier is the national PR district with 48–50 seats, while the lower tier is composed of 47 prefecture districts with one to six seats in each (SMD or multi-member district (MMD)). We focus on the PR districts, neither SMDs nor MMDs, of both chambers because we are interested in the role of the governing parties, the PR votes are especially suitable for analyzing partisan, not candidate, popularity, and the SMD/MMD votes have more idiosyncratic components which make estimation inefficient.²³

The outcome variable is the governing parties’ vote share (in percentage) among total votes to all parties (excluding invalid votes), $V_{it}$.²⁴ (Basic statistics of all variables this section uses are presented in Online Appendix A.5, Table A.3.) A “governing party” is defined as a party whose member joins the cabinet as a minister as of the election day in year t.²⁵

Scholars have argued that Japanese vote choice is affected by the delivery of particularistic goods rather than programmatic policies. This tendency has been accelerated by laws that make it difficult to engage in mass campaigning, such as via television, radio, and print (e.g.,Fukui and Fukai, 1996; Scheiner 2007). In the PR districts, the influence of interest groups is strong. In particular, farmers and construction industries have high stakes in disaster relief, as public civil engineering and primary industry facilities constitute large shares in disaster relief. The voting behavior is also partly based on party support, although, for instance, as of July 2023, 56.8% of citizens do not support any party, and even the supporters of the LDP come next with a large margin (23.6%).²⁶

The LDP won in every election to the lower chamber except in 2009, when it was defeated by the DPJ. Since Japan has a parliamentary system, this means that the LDP and its coalition partners have held power except from 2009–2012. In the elections to the upper chamber, the governing parties lost in 1998, 2007, and 2010, which led to a divided government during the 1998–1999, 2007–2009, and 2010–2012 periods.

Disaster Relief

Our explanatory variable is the logarithm of per capita disaster relief (in thousand yen) in fiscal year $t-1$, $D_{i, t-1}$.²⁷ Due to its nature, the amount of disaster relief, and thus $D_{i, t-1}$, cannot be determined at the time the budget is approved (around the beginning of the fiscal year; in March or April); instead, the amount is confirmed when government accounts are settled toward the end of the fiscal year $t-1$ (namely, March 31, calendar year t).²⁸ Since all of the elections in the study period are held after March of the year, it is the governing parties as of the election day in calendar year t that are responsible for disaster relief in fiscal year $t-1$.

Maximum Rainfall

Our instrumental variable is the logarithm of the maximum amount of one-hour rainfall (in millimeters) in calendar year $t-1$, $R_{i, t-1}$. Specifically, we calculate the maximum amount of one-hour precipitation in each grid cell (one to five kilometers on each side) in municipality i over all hours in calendar year $t-1$ and then take their maximum across all grid cells in municipality i.²⁹ Note that we do not control for normal rainfall or use rainfall deviation (cf. Hansford & Gomez, 2010) because the municipality fixed effects readily account for climatic and geographic conditions in our specification, which we will explain shortly.

Control Variables

The control variables are the logarithms of population, population density, per capita local tax revenue, per capita income, and the ratio of daytime population to the total population; the fractions of old population (65 or older), population in densely inhibited districts (DIDs), and labor force in the total population; the fraction of unemployed labor in labor force; the fractions of the primary and secondary industry employment among employed labor.³⁰ We choose these control variables because most of them are commonly used in electoral studies (e.g., Hirano, 2011), they relate to the degree of urbanization, which can affect both elections and disasters, and they are available across all municipalities over all years. We refer to these control variables in fiscal year $t-2$ because we use the instrumental variable in year $t-1$. We denote a column vector of the control variables by $\varvec{C}_{i, t-2}$.

Model

We make the following model:

$$\begin{aligned} V_{it} = \beta ^{D} D_{i,t-1} + \varvec{\beta }^{C} \varvec{C}_{i,t-2} + \mu _{i} + \tau _{t} + \epsilon _{it}. \end{aligned}$$

(1)

The quantity of main interest is the effect of disaster relief on vote share, $\beta ^{D}$. We include municipality fixed effects $\mu _{i}$ (which, among other things, account for the normal vote shares, normal rainfall, and the size of municipality) and year fixed effects $\tau _{t}$ (which account for factors like nation-wide election tides). Because estimation of $\beta ^{D}$ in Eq. (1) by the ordinary least square (OLS) suffers from endogeneity problems, we use two-stage least square (TSLS). The regression model for the first stage is:

$$\begin{aligned} D_{i, t-1} = \tilde{\beta }^{R} R_{i, t-1} + \tilde{\varvec{\beta }}^{C} C_{i,t-2} + \tilde{\mu }_{i} + \tilde{\tau }_{t-1} + \tilde{\epsilon }_{i,t-1}. \end{aligned}$$

(2)

The regression model for the second stage is Eq. (1) where we replace $D_{i,t-1}$ with its predicted values calculated from the first stage ($\hat{D}_{i, t-1}$).

Results

Second Stage

In Table 2, the first and third columns report the naive OLS results for the lower and upper chambers, respectively. They are contrasted by the corresponding TSLS results in the second and fourth columns. The first row presents the estimates of the disaster relief’s effects on governing parties’ vote share ($\beta ^{D}$ in Eq. (1)). The second row shows the standard errors clustered by municipality.

When we use naive OLS, disaster relief does not have statistically (lower chamber) or substantively (upper chamber) significant effects on the vote share of the governing parties. This, however, does not indicate the absence of causal effect because OLS does not account for endogeneity. The government, for instance, may distribute disaster relief to districts of swing voters (Lindbeck & Weibull, 1987), leading to underestimates of the effects. By contrast, when we use TSLS and thus account for the endogeneity, disaster relief has substantial effects increasing the governing parties’ vote share. In fact, the TSLS estimates are over ten times larger than the corresponding OLS estimates. This fact suggests that there is a reverse causality; that is, governing parties’ vote share may affect the amount of disaster relief.

Table 2 The effects of disaster relief on the governing parties’ vote share

The causal effects are indeed substantively large. For instance, when the disaster relief per capita increases from zero ($D=0$, which is true for 30.0% of the observations) to its mean 7, 704 yen, the predicted value of the governing parties’ vote share increases from 44.5 to 47.2% (by 2.8% points) in the lower chamber, and from 46.7 to 52.1% (by 5.4% points) in the upper chamber.³¹ Figure 2 provides further details about the effect sizes. The vertical axis indicates the governing parties’ vote share, and the horizontal axis represents the amount of disaster relief per capita.³² The dots and triangles correspond to the point estimates for the lower and upper chambers, respectively. The dotted lines present the bounds of the 95% confidence intervals. The figure indicates that a larger shift in the disaster relief per capita results in a greater change in the governing parties’ vote share.

Fig. 2

The effects of disaster relief on governing parties’ vote shares. The vertical axis indicates the governing parties’ vote share, and the horizontal axis represents the disaster relief per capita. The dots and triangles correspond to the point estimates for the lower and upper chambers, respectively. The dotted lines present the bounds of the 95% confidence intervals

First Stage

The results of the first-stage regression (the bottom panel of Table 2) indicate that the maximum rainfall is a strong predictor of disaster relief. The coefficients of the maximum rainfall in the third row ($\tilde{\beta }^{R}$ in Eq. (2)) are positive and significantly different from zero (the clustered standard errors are in the fourth row). Moreover, the F statistics (in the fifth row, Wald test) are far above the conventional criterion of 10 (Stock et al., 2002). Substantively, when the maximum hourly rainfall increases from zero to the mean 57.5 mms per hour, the disaster relief per capita increases by 6, 924 yen in the dataset of the lower chamber and by 3, 480 yen in the dataset of the upper chamber.

When the government designates a rain-related Disaster of Extreme Severity, it also specifies the official period of days. The official periods of the rain-related Disasters of Extreme Severity (heavy rain and/or rainstorm) cover 20.6% of the study period. By contrast, on average across years, in 63.5% of municipalities, the day of the maximum rainfall in a municipality in a year lies in the official periods of the rain-related Disasters of Extreme Severity in the year.³³ The odds ratio is 6.71 $(=(0.635/(1-0.635))/(0.206/(1-0.206)))$. This is consistent with our preposition that the maximum rainfall is a good indicator of a rain-related disaster.

Validity Checks

A critical assumption in the IV analysis is the exclusion restriction. Skeptical readers, however, might suspect that disaster (in particular, the maximum rainfall) directly has a negative impact on in-party’s electoral outcomes, that is, voters could punish the government (Achen & Bartels, 2016; Heersink et al., 2017, 2022). But such violation of the exclusion restriction, if any, would underestimate the disaster relief’s effect on incumbent’s vote share and thus make our estimates conservative.

That said, we check it empirically. One possibility is that the maximum rainfall might affect agricultural outputs, which, in turn, might decrease the governing parties’ vote shares. We address this concern by analyzing the effect of the maximum rainfall on agricultural production and cultivated area. Another potential concern may be the exogeneity of rainfall. Past maximum rainfall, for example, can be a confounder; namely, it might affect both the current maximum rainfall and electoral results. We therefore repeat our analysis, controlling for the past two year lags of the maximum rainfall, $R_{i, t-2}$ and $R_{i, t-3}$.

The two validity checks ensure that we do not have to worry about such concerns (Online Appendix B.1, Tables B.1 and B.2).

Mechanisms: Mobilization or Persuasion

So far, we have shown that disaster relief increases governing parties’ vote share. Is this incumbent’s electoral gain produced by mobilizing supporters of governing parties (Chen, 2013; De La, 2013) and/or persuading voters from oppositions to governing parties (Pop-Eleches & Pop-Eleches, 2012)? For our purpose, we substitute three more outcome variables in Eq. (1): governing parties’ voter turnout (the proportion of citizens who vote for governing parties to eligible citizens, $V^{(G)}_{it}$), opposition voter turnout (the proportion of citizens who vote for the opposition to eligible citizens, $V^{(O)}_{it}$), and the total voter turnout ($V^{(T)}_{it} = V^{(G)}_{it} + V^{(O)}_{it}$). Let us denote the corresponding coefficients of $D_{it}$ by $\beta ^{D(G)}$, $\beta ^{D(O)}$, and $\beta ^{D(T)} (= \beta ^{D(G)} + \beta ^{D(O)})$, respectively. If the mobilization hypothesis is true, we should find $\beta ^{D(G)} = \beta ^{D(T)} > 0, \beta ^{D(O)} = 0$ (De La, 2013). If the persuasion hypothesis works, we should have $\beta ^{D(G)} > 0, \beta ^{D(O)} = -\beta ^{D(G)} < 0, \beta ^{D(T)} = 0$.³⁴ In Table 3, we present the results of the second stages. (We do not report the first stages because they are identical to those in Table 2). For the upper chamber, the effect on governing parties’ turnout is significantly positive ($\beta ^{D(G)} > 0$), and for the lower chamber, the effect is positive and close to statistical significance ($p=0.055$). In contrast, the effects on opposition turnout are significantly negative ($\beta ^{D(O)} < 0$) for both chambers. The effect on the total turnout is not distinguishable from zero. Arguably, the results lend more support for the persuasion mechanism than for the mobilization mechanism.

Table 3 The effects of disaster relief on turnout

Robustness Checks

In addition, we also conduct other robustness checks in the Online Appendix B.2. First, we transform the disaster relief in a few different ways: the dummy variable of whether a municipality obtains any disaster relief in a year (Table B.3), the logarithm of the amount of the disaster relief (not divided by population; Table B.4), and the proportion of the amount of the disaster relief to the total expenditure (Table B.5). Neither changes our conclusion; the disaster relief helps governing parties.

Second, we apply the method of Betz et al. (2020), spatial-two stage least squares, to account for spatial dependency. We use geographically contiguous municipalities in the same year as connected units in the connectivity matrix. The results do not change our findings (Online Appendix B.2, Tables B.6).

Third, we substitute the governing party candidates’ vote share in SMDs of the lower chamber and SMDs and MMDs of the upper chamber as outcome variables (Online Appendix B.2, Table B.7). The analyses also show that our findings are robust.

Conclusion

Our contribution to the literature is twofold. One is methodological. Conventional wisdom in political science suggests that distribution affects elections and vice versa. The endogeneity makes it difficult to identify the causal relationship. Although several studies try to address the problems by considering natural disasters as exogenous shocks in governments’ distribution, political considerations may still filter into allocation of disaster relief, and hence the estimates may still suffer from the endogeneity problems. We address the problem by applying the IV approach to the case of disaster relief in Japan. The results indicate retrospective voting, that is, disaster relief due to heavy rain brings substantial electoral gains to the governing parties. Furthermore, it is suggested that disaster relief persuades opposition supporters to change their minds rather than mobilizes government supporters.

We also intend to make two substantive contributions. First, our study has expanded the scope conditions for the positive effect of disaster relief on governing parties’ electoral outcomes by adding the case of Japan. We conjecture that the effect is explained by general theories such as retrospective voting and economic voting and thus that similar effects can be found in other countries. However, most prior works have focused on the U.S. with only a few exceptions of those about Colombia (Gallego, 2018) and Germany (Bechtel & Hainmueller, 2011). Moreover, Japan is one of the countries that has suffered the most from disaster damage; the country lost four thousand dollars per capita—the largest amount among the OECD countries next to New Zealand (CRED, 2023). Thus, it is imperative to understand their political consequences. Certainly, these points do not preclude further analyses; future studies need to assess the generalizability more rigorously and thoroughly. Second, our study adds a new case to the controversy between mobilization and persuasion mechanisms. It is shown that a certain type of disaster aid in the U.S. (Chen, 2013) and the conditional cash transfer in Mexico (De La, 2013) mobilized government supporters, while the computer coupon in Romania (Pop-Eleches & Pop-Eleches, 2012) persuaded opposition supporters to change their minds as well. In contrast, our study presents internally valid evidence for the persuasion mechanism. Future studies should further explore whether or under what conditions those mechanisms are at work.

While this study focuses on the case of disaster relief to address the endogeneity problems, the case can plausibly be considered a “hard” case. After a natural disaster, the urgent humanitarian and economic needs often dominate the government’s decisions about disaster relief (Bechtel & Mannino, 2022). As a result, not all victims may regard the distribution as a special favor provided by the government. Nonetheless, as our analysis shows, even such disaster relief still boosts the governing parties’ vote share. By contrast, more typical types of pork (e.g., road construction unrelated to disasters) are often greatly influenced by the will of policy makers (Kitschelt & Wilkinson, 2007; Stokes et al., 2013). Knowing this, citizens should be more likely to perceive such benefits as the government’s favor and hence to vote for the governing parties. We leave it to future studies to test those conjectures with empirical rigor.

More broadly, our findings caution against regarding a disaster response as an apolitical process. Indeed, the fact that our TSLS results of the IV analysis substantially differ from the OLS results suggests that reverse causality may exist—namely, the effects of elections on distribution. While natural disasters are often used for the purpose of causal identification not just in electoral studies but also in other contexts (Bueno de Mesquita & Smith, 2010; Quiroz-Flores & Smith, 2013), we may need to rethink the plausibility of these approaches. It is a task of future research to further explore how exogenous government’s responses to “natural” disasters are.