Types of natural disasters and their fiscal impact

Transcription

1 Types of natural disasters and their fiscal impact Ian Koetsier Discussion Paper Series nr: 17-18

2 Tjalling C. Koopmans Research Institute Utrecht University School of Economics Utrecht University Kriekenpitplein EC Utrecht The Netherlands telephone fax website The Tjalling C. Koopmans Institute is the research institute and research school of Utrecht University School of Economics. It was founded in 2003, and named after Professor Tjalling C. Koopmans, Dutch-born Nobel Prize laureate in economics of In the discussion papers series the Koopmans Institute publishes results of ongoing research for early dissemination of research results, and to enhance discussion with colleagues. Please send any comments and suggestions on the Koopmans institute, or this series to How to reach the authors Please direct all correspondence to the first author. Ian Koetsier Utrecht University Utrecht University School of Economics Kriekenpitplein TC Utrecht The Netherlands. This paper can be downloaded at:

3 Utrecht University School of Economics Tjalling C. Koopmans Research Institute Discussion Paper Series Types of natural disasters and their fiscal impact Ian Koetsier Utrecht University School of Economics Utrecht University December 2017 Abstract This study investigates the impact of natural disaster on government debt for different disaster types. It includes 163 countries for the period from 1971 to We apply a panel synthetic control methodology to estimate the impact of natural disasters on government debt. Our findings generally reveal a considerable increase of government debt in the aftermath of a natural disaster, except for droughts. Earthquakes, on average, lead to an increase in government debt of 30.2% of GDP. Floods increase government debt by 7.7% of GDP, while storms increase the level of government debt by 9.5% of GDP compared to the synthetic control group. This study shows that the manner of identifying disaster matters for the estimated disaster impact. Natural disasters are a considerable contingent liability for governments. The effect of natural disasters is even larger than the fiscal costs of financial sector bailouts during the financial crisis. Keywords: government debt, government finances, natural disasters, panel synthetic control method, disaster identification JEL classification: E62, Q51, Q54, Q58 The views expressed here are solely those of the author and do not in any way represent the view of the institution to which he is affiliated.

4 1. Introduction Natural disasters have a tremendous impact on national economies. Large sums of money must be invested in the reconstruction and relief effort. These sums are likely to increase in the coming years and decades because of anthropogenic climate change, which will increase the frequency and intensity of climate events. Policymakers must be aware that weather-related disasters are becoming increasingly frequent (CRED/UNISDR, 2015). A notable example is the hurricane season of 2017, which saw two category 5 hurricanes. 3 Costs of natural disasters are in large part born by the public sector. Governments have the moral obligation to intervene and alleviate the suffering of those affected. The main aim of this study is to establish the impact of different types of disasters on government debt. We estimate the impact on government debt for over 100 large natural disasters. We observe the disaster impact up to 10 years after the disaster occurred. Thereby, we can establish the short, medium and long-term impact of a natural disaster. Different types of natural disaster can have different consequences on government finances. For example, a slow-onset disaster such as a drought might have a minor effect in the short term but potentially an extremely large effect in the long-term, whereas the impact of a sudden-onset disaster, such as an earthquake might be different. The literature on the fiscal impact of natural disasters mostly focusses on a specific type of disaster (e.g. hurricanes) or on a region (e.g. the Caribbean) or on a group of disasters (e.g. seismological, climatological etc.) (see Ouattara and Strobl, 2013; Acevedo, 2014; Melecky and Raddatz, 2011; 2015). Our study investigates multiple types of natural disasters. In this way, we can establish separate estimations of the postdisaster impact of droughts, earthquakes, storms and floods. For comparison reasons, we also present the effects of nonclimate and climate events. This allows us to assess the potential impact of climate change on government finances. The estimations of fiscal impact for different types of natural disasters enable policy makers to improve their predisaster preparations. The occurrence of specific types of natural disaster is not spread equally across the globe. There are large differences in the likelihood of occurrence of a certain type of disaster between countries. A seismologically active area, for example, is more likely to suffer from earthquakes or volcanic eruptions, whereas a country which lies below sea level is more likely to experience flooding than other countries. Another benefit of the distinction between the types of natural disasters is that it allows for a more accurate estimation of the costs of climate change. A climate event might have a different impact on government debt than a nonclimate event. In addition, within various climate events, the impact might differ. An accurate estimation of disaster costs is warranted for the introduction of disaster preventions measures in order to enable policy makers to 3 There were only 33 category 5 hurricanes in the past 150 years, while two occurred in the 2017 season (see National Hurricane Center; Hurricane Research Division, 2017). One was the most powerful Atlantic hurricane ever recorded. 4

5 conduct proper costs-benefit analyses. Furthermore, an estimation of the potential costs is warranted to create awareness of the implicit contingent government obligation that natural disaster represents. This study makes a number of contributions to the literature. Firstly, it investigates the impact of a natural disaster on government debt. Most previous studies focus on the impact of a natural disaster on GDP per capita or output (e.g. Albala-Bertrand, 1993; Skidmore and Toya, 2002; Noy, 2009; Strobl, 2012). Secondly, this study compares the postdisaster impact of different types of natural disasters. To our knowledge, this study is the first to investigate the short, medium and long-term impact on government debt for different types of disasters. Thirdly, our ability to study the entire postdisaster outcomes stems from the use of an innovative panel synthetic control methodology. This method allows us to estimate the impact of a natural disaster on government debt. Although there are some case studies which use the synthetic control method to identify the impact of a specific type of disaster, our study is the first to apply this to multiple types of disasters. This study is structured as follows. Section 2 provides an overview of the empirical literature on the impact of different disaster types. Section 3 presents the data and our synthetic control method, whereas section 4 reveals the estimation results of the synthetic control method. Our final section concludes. 2. Empirical literature The literature on the macroeconomic costs of natural disasters focusses on the effect on output. 4 These studies find mixed evidence; partly attributed to the type of disaster. There are studies which focus on a specific type of natural disaster. Ouattara and Strobl (2013) investigate the fiscal impact of hurricane strikes in the Caribbean. 5 The selection of hurricanes is based on the Saffir-Simpson scale. Hurricanes must have a minimum category 3. For the period , they identify 28 hurricanes. Ouattara and Strobl (2013) reveal that there are considerable fiscal implications in the short run. Government spending increases up to two years after the disaster. In addition, they find a negative effect on the budget balance. Interestingly, they do not find a significant effect on the public investment, debt and tax revenues. In similar vein, Rasmussen (2004) tries to estimate the impact of disasters on the fiscal position of the Caribbean. Although he does not focus on one specific disaster per se, his sample mostly consists of hurricanes. 6 In other words, he investigates the impact of storms on the fiscal position of members of the Eastern Caribbean Currency Union (ECCU). For the period , he finds an increase of median public debt by a cumulative 6.5 percentage points within three years. He determines that this effect is caused by an increase in 4 An extensive literature study on the output effect is beyond the scope of this paper. Cavallo and Noy (2010) provide an excellent overview of the literature. 5 They include 18 Caribbean countries in their analysis. 6 Rasmussen (2004) investigates 44 natural disasters. It consists of 34 storm/hurricanes, and 10 other types of disasters (e.g. floods, volcanic eruptions etc.). 5

6 spending. Our study is closer to another study on the Caribbean. Acevedo (2014) specifically looks at two types of disasters: storms and floods. The sample includes 107 natural disasters, of which, divided 78 are storms and 29 are floods. 7 He conducts a panel VAR analysis to establish an effect on GDP per capita and public debt. He finds mixed effects for the public debt. Floods lead to an increase of public debt, whereas the storms seem to have no effect. Thus, he finds heterogonous outcomes of natural disasters. The Caribbean is a very specific region which lays in the Atlantic hurricane belt. Therefore, we also discuss the studies that use a more diverse set of countries. Bova et al. (2016) try to create a database of governments contingent liability realizations. Contingent liabilities are government obligations that materialize after a specific event, in this case, a natural disaster. Their study shows average fiscal costs of natural disasters equal 1.6% of GDP. They investigate a total of 65 natural disasters, but do not differentiate between the types of disaster. Their findings also reveal maximum fiscal costs of 6% of GDP. A drawback is their definition of fiscal costs, which are limited to the immediate change in the government financial position. Natural disasters potentially have long-term consequences on fiscal outlays, and this definition ignores second-order effects (e.g. macroeconomic effects of the disaster). Noy and Nualsri (2011) try to quantify the fiscal effects of a natural disaster using quarterly data. Their sample includes a variety of countries: 22 high income countries, 11 upper-middle income countries and 9 lower-middle income countries. They apply the Eichenbaum and Fisher (2005) panel VAR methodology to assess the fiscal response following a large natural disaster. They find a heterogeneous response. Developed countries conduct counter-cyclical policies, increasing spending and cutting taxes. In contrast, developing countries engage in pro-cyclical policies, decreasing spending and increasing revenues. The result is that developed countries accumulate additional debts equal to 8% of GDP within 1.5 years. There are a few drawbacks of this study. They only use direct damages to identify the occurrence of a natural disaster, whereas different types of natural disaster have different outcomes in terms of damage, deaths or total affected. This might result in an overrepresentation of a specific type of natural disaster. Furthermore, there is no overview of the types of disasters which are included. Consequently, we are unable to assess the heterogeneous impact of natural disasters. Melecky and Raddatz (2011) estimate the effect of a large natural disaster on government expenditure and revenue. Their sample consists of 81 high and middle-income countries from 1975 to They include the types of natural disaster in large groups: geological, climate and other 7 The sample consists of 12 Caribbean countries: Antigua and Barbuda, The Bahamas, Barbados, Dominica, Dominican Republic, Grenada, Haiti, Jamaica, St. Kitts and Nevis, St. Lucia, St. Vincent and the Grenadines, Trinidad and Tobago. 6

7 disasters. 8 They include a total of 32 geological disasters, 430 climate disasters and 15 other disasters. Melecky and Raddatz (2011) find heterogeneous outcomes for the different disaster groups. Climate disasters, on average, lead to a worsening of the budget deficit. However, the lower-middle income countries experience a deterioration of the deficit for all disaster groups. A drawback of the approach of Melecky and Raddatz (2011) is that it potentially hides heterogeneous outcomes within the disaster group. For example, a storm might have a different impact on fiscal indicators than a flood. There have been some attempts to estimate the impact of a certain type of disaster; however, these attempts mostly focus on one specific type of disaster. This study uses a consistent methodology to assess the impact of different types of disasters on government debt. 3. Data and methodology Data We have an unbalanced panel with 163 countries for the period from 1971 to Table 1 provides an overview of the countries selected for this study. The disaster data is obtained from the EM-DAT database. 9,10 This database includes different types of disasters: extreme temperature, storm, wildfire, mass movement (dry), volcanic activity, earthquake, landslide, and flood. Figure 1 shows the geographical spread of the different natural disasters. The EM-DAT database includes indicators of disaster severity, such as total number of affected, total number of deaths and direct damages. The macroeconomic indicators are mainly obtained from UN statistics and the World Bank (for more details, see Table 23). Following Peduzzi (2005), we define extreme temperatures, drought, storm and floods as climate events. Earthquakes and volcanic activity, we define as nonclimate events. We only discuss results for a type of natural disasters which have a sufficient number of natural disasters of a specified type (in our case, a minimum of ten natural disasters disasters). Moreover, we do not use wildfires, mass movements (dry) and landslides. Mass movements, wildfires and landslides are often local phenomena. As a consequence, they are unable to affect the government s financial position. [Insert Table 1 and Figure 1 here] 8 The group geological disasters consist of earthquakes, volcanic eruptions and tsunamis. Climate disasters include floods, drought, extreme temperatures and storms, and the group other disaster is a mix of manmade and medical disasters (e.g. famines, epidemics, plagues and accidents). 9 For a more detailed discussion on the EM-DAT database, see Koetsier (2017a). 10 EM-DAT (2016a). The CRED/OFDA International Disaster Database, D. Guha-Sapir, R. Below, P. Hoyois, Université Catholique de Louvain, Brussels. 7

8 Our study focuses on large natural disasters which potentially influence the macro economy and government finances. Therefore, we impose a stricter disaster threshold than the EM-DAT database. 11 We use the indicators of disaster severity to construct our disaster identification measures. Firstly, we construct the standard disaster measures: deaths over population, total affected over population and damage over GDP. We name this standard disaster identification strategy (henceforth, the standard strategy). The reader should note that the denominators are lagged by one period because the natural disaster can potentially influence them as well. 12 Secondly, we apply a disaster identification strategy using land area (henceforth, land area strategy). This measure incorporates the potential benefits of country size. Furthermore, land area is believed to be less influenced by the level of development, and as such, is a more exogenous indicator for a disaster. Table 2 shows the standard strategy and the land area strategy and the type of disasters, they identify. There are some differences in the types of disaster under investigation depending on the mode of identification (e.g. total affected, damage or deaths) and the identification strategy (e.g. standard or land area). Later in this study, we use disaster magnitude to assure the exogeneity of disaster identification. [Insert Table 2 here] This study uses a panel synthetic control methodology. This method needs a treatment group (disaster countries) and a control group (nondisaster countries). Countries are divided into three categories: disaster, nondisaster and nonclassified countries. All natural disaster that classify in the largest 2.5% of disaster severity distribution result in a disaster country qualification. Thus, for the standard strategy, one disaster in the top 2.5% of natural disasters for total affected or deaths or damages is sufficient to classify a country as a disaster country. The same procedure holds for the land area strategy. The other countries which did not classify as a disaster country are divided in two groups. We start with the nonclassified countries. These countries experience a (sizable) natural disaster in between the 2.5% and 5% largest natural disasters. For this classification, again, one natural disaster is sufficient to classify a country as a nonclassified country. Countries which only experience natural disasters below the 5% largest natural disasters in the disaster severity distribution are defined as nondisaster countries. 13 Methodology This study uses the panel synthetic control methodology developed by Cavallo et al. (2013). This methodology builds on the case study method by Abadie and Gardeazabal (2003) and Abadie et al. (2010). The synthetic control method allows us to construct a counterfactual which represents the 11 EM-DAT (2016b). EM-DAT guidelines. Accessed 30 March 2016 at 12 Population and GDP are influenced by large natural disasters. 13 The occurrence of one type of natural disaster is sufficient to classify a country as a disaster country. 8

9 disaster county if it had not experienced a natural disaster. The nondisaster countries are used to construct this counterfactual. This counterfactual can be made up by multiple nondisaster countries. In addition, the weight of the different nondisaster countries can differ, which results in the best possible approximation of the disaster country in the predisaster period. The first step in the estimation of the panel synthetic control method is an estimation of the individual case studies. In the predisaster period, we still observe the trajectory of government debt of the disaster country without the occurrence of a natural disaster. DDDDDDTT NN iiii gives government debt of a country ii at time tt. NN indicates that no natural disaster has occurred yet. Our study includes JJ + 1 countries, ii = 1,, JJ + 1, and 21 time periods, tt = TT 0 10,, TT We include ten predisaster periods to assure an accurate match of the synthetic control group with the disaster country. In similar vein, we include the trajectory up to 10 periods after the occurrence of the disaster. This allows us to capture the short, medium and long-term disaster consequences. We define the predisaster period as TT 0 10 < TT 0, and the postdisaster period as TT 0 to TT DDDDDDTT II iiii denotes the outcome for country ii at time tt if country ii experienced a natural disaster. Cavallo et al. (2013) notes that the occurrence of a natural disaster is unpredictable. This is true for some aspects, such as timing, exact location and severity. This study will later relax this assumption. For now, government debt is not affected before the occurrence of the natural disaster, resulting for tt {TT 0 10,, TT 0 1} and ii {1,, NN} in DDDDDDTT II iiii = DDDDDDTT NN iiii. The problem is that we do not observe the postdisaster period without a natural disaster for country ii as the natural disaster occurred in this country. This study wants to establish the difference between a business-as-usual scenario and the disaster scenario, αα iiii = DDDDDDTT II iiii DDDDDDTT NN iiii, for country ii in the postdisaster periods TT 0, TT 0 + 1, TT 0 + 2,, TT The size of the disaster effect can differ over the postdisaster period (αα 1,TT0,, αα 1,TT0 +10). An additional benefit is that we can present the effect over different time frames, short, medium and long run. For tt TT 0, αα 1tt = DDDDDDTT II 1tt DDDDDDTT NN NN 1tt = DDDDDDTT 1tt DDDDDDTT 1tt (1) NN The synthetic control method estimates DDDDDDTT 1tt in equation (1). 15 NN DDDDDDTT 1tt will be a composition of nondisaster countries, which get weights between 0 and 1. The sum of these weights is equal to one. The higher the weight, the better the explanatory power of the nondisaster country for the disaster country in the predisaster period. The match is made based on a selection of macroeconomic, institutional, and geographical indicators. These indicators capture the underlying economic structure 14 If the average disaster occurs halfway the year, we have present a postdisaster trajectory for 10.5 years. 15 It is important that there is no contamination of the control group. Therefore, we exclude disaster country and nonclassified countries from our control pool. In addition, we exclude nondisaster countries that experience idiosyncratic shocks in the postdisaster period. 9

10 of the disaster country ii. 16 We assume that only country jj = 1 is exposed to a natural disaster. The other JJ countries do not experience a natural disaster; therefore, we can use them as our donor pool. 17 There is a (JJ 1) vector of weights WW = (ωω 2,, ωω JJ+1 ). Conditions assure that these weights are positive or zero ωω jj 0 for countries jj = 2,, JJ + 1, and sum to one ωω 2 + ωω ωω JJ+1 = 1. This procedure allows us to estimate the disaster effect in equation (2): JJ+1 αα 1tt = DDDDDDTT 1tt ωω jj DDDDDDTT jjtt jj=2 (2) The second step is to extent our case study analysis to a panel analysis. Following Cavallo et al. (2013), we repeat previous estimations for multiple disasters (GG). In other words, we estimate the respective synthetic control case studies and reclassify the time periods accordingly. Equation (3) represents this panel estimation. αα = αα TT0,, αα TT0 +10 = 1 GG (αα gg,tt 0,, αα gg,tt0 +10) GG gg=1 (3) This study does not only estimate the average disaster effect as represented in equation (3). It also estimates the median disaster effect, which is less influenced by potential outliers. As mentioned earlier, the match between the disaster and nondisaster countries is based on a selection of macroeconomic, geographical and institutional indicators. This selection is taken from a combination of the economic growth literature (see Islam, 1995; Barro and Sala-i-Martin, 2003) and the literature on macroeconomic costs of natural disasters (see Noy, 2009; Cavallo et al., 2013). The indicators are GDP growth, current account (% of GDP), openness (% of GDP), population density, population growth, GDP per capita, GDP share of agriculture, hunting and minerals 18, general government consumption (% of GDP) and gross capital formation (% of GDP), the average latitude 19, years of schooling and the total societal and interstate major episodes of political violence. These indicators capture the underlying economic structure of a country and partly incorporate the possible vulnerability to natural disasters. Table 3 provides summary statistics of the included indicators. It reveals that the average government debt, our variable of interest, equals 61.6% of GDP. 16 If an exact match between the nondisaster countries and disaster country in the predisaster period is not attainable, then the method selects the best available option. 17 The donor pool is much more restrictive. We exclude, for example, nonclassified countries. 18 Sectors that are closely related to climate, such as agriculture, tourism, and water, face a great burden from extreme events (IPCC, 2012). 19 Latitude controls for the climatological circumstances. 10

11 [Insert Table 3 here] This study follows Abadie et al. (2010), and includes government debt in the part of the predisaster period in some specifications. We do not include the entire predisaster period. Kaul et al. (2016) stress that the inclusion of the entire predisaster period renders the other indicators irrelevant in the predisaster match. Table 3 also shows different specifications of our model. These differ in terms of the number of indicators used to match the disaster and nondisaster countries in predisaster period. There are differences in data availability between indicators, and this will lead to estimations with fewer natural disasters for some models. There is a trade-off between the number of indicators included and the number of natural disasters used in our panel synthetic control estimation. An additional problem is that data availability depends on the level of development. Thus, we must be aware that inclusion of additional indicators might lead to the exclusion of developing countries from our sample. As a consequence, we choose to estimate different specifications which enable us to compare the results across models, and investigate whether our findings are robust for different specifications. 4. Results The numbers which are presented in this section are the differences between the disaster country and the synthetic control group. Those numbers are expressed as a percentage of GDP. For simplicity reasons, we normalize the outcome for the year preceding the disaster year. This year is set equal to zero. A positive effect indicates the government debt in the disaster country is higher than in the synthetic control group. Thus, we find a positive effect of the natural disaster on government debt. For the opposite sign, it is the other way around. This study starts by presenting the effect of climate events on government debt in Table 4. In recent decades, climate change has impacted natural and human systems on all continents and across the oceans (IPCC, 2014). For the standard strategy, we find mixed evidence of the effect of climate events on government debt. For total affected over population, our study finds a negative effect on government debt, whereas, for damages and deaths, we find a clear positive effect on the level of government debt compared to the synthetic control group. For damages, we find an average disaster effect between 7.5% and 10.1% of GDP. These estimates are significant at the 5%-level. Our results indicate the timing of the largest impact differs between the models. The median disaster effect is considerably lower, and sometimes even negative (see model (5)). The median increase of government debt equals 3.8% of GDP when all models are combined. For deaths, the average and median disaster effect is relatively similar with an increase between 4.9% and 7.7% of GDP for the average effect for models (1)-(4), and an increase between 4.0% and 9.9% of GDP for models (1)-(5). Melecky and 11

12 Raddatz (2011) find an increase in the budget deficit for climatological disasters. Our study proves that budget deficits, because of climate disasters, translate into a higher level of government debt. Nonclimate events result in a different postdisaster trajectory (see Table 5). When we regard the total affected population, a relative low number of nonclimate events qualify using this disaster identification criterion. Therefore, we should be careful in interpreting these results as being similar to the impact of a nonclimate disaster, even though, the models reveal a positive impact on government debt. Government debt increases by 20.6% of GDP compared to the synthetic control group. For the deadliest disasters, the average disaster effect reveals a sharp increase in the level of government debt just after the natural disaster. The greatest effect on government debt is in the first two years after the disaster. In the first years, debt increases by 4.7% of GDP compared to the control group. For the median effect, the pattern is different. Government debt grows until the first four years after the natural disaster. It peaks between the second and the fourth year after the disaster, equaling 8.9% of GDP. The size of the effect is roughly equal to that reported in Noy and Nualsri (2011). They find that government outstanding debt accumulated to more than 8% of GDP over a year and a half. The pattern of the impact on government finances is, however, somewhat different. In our study, the full impact is only reached after more than two years. [Insert Table 4 and Table 5 here] The previous results aggregate the different types of disaster in disaster groups. We are interested in the effects of specific disaster types. Therefore, we distinguish between storms, droughts and floods for climate events. Notice we are unable to the present the results for extreme temperature events, there are insufficient of these events in our sample. In contrast to Acevedo (2014), we do find that storms increase the level of government debt (see Table 6). We study between 10 and 55 storms depending on the model, whereas Acevedo (2014) investigates 78 storms. Our sample is for the entire world, where Acevedo (2014) only uses Caribbean countries. Therefore, the average storm in our sample is much larger than the average storm in Acevedo s sample. Our findings for storms are mixed. There is no evidence of a positive effect on government debt if we identify natural disasters using the total affected. For damages, we find a very large effect on government debt compared to the synthetic control group. Depending on the model, the average disaster effect might be as large as 26.3% of GDP. This result is significant at the 5%-level. The other estimates also point to very large increases of the government debt, although the estimates are not always statistically significant at the end of the estimation period (between eight and ten years after the natural disaster). This sizeable effect is partly the result of some extremely costly disasters as the median disaster effect only reveals an increase of government debt of 5.7% of GDP. If we investigate the deadliest disasters, we observe more mixed results. The combined models reveal a modest increase 12

13 of government debt compared to the synthetic control group. Government debt increases by 3.5% of GDP, this result is significant at the 5%-level. Droughts are only identified when we use the total affected to identify the occurrence of a natural disaster. The other disaster identification strategies identify far less than ten droughts. We find a clear decline of government debt compared to the synthetic control group (see Table 7). Although we do not assess the cause of this decline in detail, a possible explanation might be sovereign default. Almost 70% of the countries which experiences a drought default on their debt obligations in our research period. Koetsier (2017a) finds similar results when he uses the total affected over population. Floods do increase government debt in most specifications, except for the total affected over population. For the deadliest and most damaging disasters, the postdisaster trajectories of floods reveal a bell-shaped curve (see Table 8). Initially, government debt increases rapidly but in later periods, government debt declines. For damages and deaths, we find convincing evidence of an increase in government debt compared to the synthetic control group. For damages, we find, on average, a substantial increase, between 9.0% and 12.4% of GDP in models (1)-(4). The notable exception is model (5) with insignificant results. For all models, we find a median increase of government debt. The combined models reveal an increase of debt of 6.9% of GDP. The deadliest floods also result in a substantial increase of government debt between 8.7% and 13.3% of GDP. A substantial increase in the level of government debt is also observed for the median disaster effect. Floods increase government debt between 4.9% and 17.2% of GDP. [Insert Table 6, Table 7 and Table 8 here] We also distinguish the nonclimate events in volcanic eruptions and earthquakes. However, the sample of nonclimate events is dominated by earthquakes. Therefore, our findings stay the same compared to the nonclimate events estimates (see Table 9). Furthermore, the number of earthquakes is insufficient when we use total affected over population to identify earthquakes. Consequently, these results are not presented here. The reader should note that the nonclimate results already present an indication of the potential impact. For the most damaging and deadliest disasters, the initial impact shows a substantial increase of government debt. For example, the median disaster effect for damages equals 8.9% of GDP. However, this effect gradually disappears towards the end of our postdisaster period. For deaths, we also find an initial increase of government debt for the average and median disaster effect. The increase of government debt equals 7.6% and 6.7% of GDP, respectively, which are both significant at the 1%-significance level [Insert Table 9 here] 13

14 In general, we find an increase of government debt after a climate event. In other words, government debt is considerably higher after the occurrence of a natural disaster than in the what-if scenario of no disaster. Our study, however, reveals considerable heterogeneity in the postdisaster outcomes. Floods and storms lead to a substantial increase of government debt compared to the synthetic control group. The increases are most pronounced when natural disasters are identified using deaths or damages. In extreme cases, storms can even have a great impact on government finances. The impact of floods is also large, and the increase of government debt proves to hold for most estimations. In contrast, droughts do not lead to an increase of government debt compared to the control group. We attribute this partly to sovereign defaults. For earthquakes, we find a relative large short-term impact on government debt. Large countries benefit from geographical diversification. If a natural disaster strikes a large country, parts of the country are likely to be unaffected by the natural disaster. Thus, the negative effects of natural disaster on the country s economy are less pronounced in a large country. Other potential benefits are that federal fiscal transfers can help finance immediate disaster relief and reconstruction. Deryugina (2016) and Miao et al. (2016) show for United States that parts of the disaster costs are covered by federal transfers. The unaffected states are net-payers for government interventions in the aftermath of a natural disaster. To assess the effect of country size, we estimate the EM-DAT disaster outcomes (total affected, damage and deaths) over land area in squared kilometers. This gives another indication of the severity of the disaster, which corrects for the potential benefits of geographical diversification. We start again by investigating the impact on government debt for the disaster groups, like climate and nonclimate events. Here we estimate the disaster impact from climate events in Table 10. For total affected over population, we find some evidence of an increase of government debt, especially for the median disaster effect. This shows a debt increase of 4.4% of GDP. For damage and deaths, our findings show a consistent increase of government debt compared to the synthetic control group. The most damaging disasters, on average, result in an increase of government debt ranging from 17.1% to 24.4% of GDP. Over the entire postdisaster period, we observe an increase of the disaster effect. The median disaster effect also reveals a substantial increase of government debt. This ranges from 11.4% to 14.5% of GDP at a 1%-significance level. The deadliest disasters also lead to an increase of government debt, although the effect is less pronounced than that found for the most damaging disasters. We only identify a small number of nonclimate events when we use total affected to identify them. Most of the findings are insignificant which can partly be attributed to low number of natural disaster for this disaster identifier. The findings for the most damaging and the deadliest disasters are also presented in Table 11. In particularly, the most damaging disasters lead to a substantial increase in government debt. For the combined models, the increase of government debt equals 26.8% of GDP, 14

15 which is also statistically significant at the 1%-level. The median disaster effect is comparable to the average disaster effect. The median effect reveals an increase of government debt of 20.8% of GDP. In similar vein, the median effects for the deadliest nonclimate events are comparable to the average effects. The median effect ranges from 6.3% to 16.3% of GDP. [Insert Table 10 and Table 11 here] Our approach remains the same for the land area strategy compared to the standard strategy. We present the estimation results for the different disaster types in Table 12-Table 15. The effect of storms on government debt is somewhat ambiguous, as shown in Table 12. For total affected and deaths over land area, we find mixed evidence. For total affected over land area, the average disaster effect reveals mostly insignificant effects, whereas the median disaster effect shows a modest increase of government debt. When this study identifies disasters using deaths over land area, there is an average negative impact on government debt. For the median disaster effect, we find an insignificant impact on government debt. Contrary to these results, the most damaging disasters show a clear increase in government debt. These disasters can lead to substantial increase in government debt. Some disasters result in an average increase of between 20.8% and 26% of GDP. The median disaster effect somewhat lower but an increase of government debt equaling 12.3% of GDP is still very substantial. In Table 13, droughts are only identified if we use total affected over land area. We find similar results compared to the standard strategy. In general, droughts result in a reduction of government debt compared to the synthetic control group. As mentioned earlier, this might result from external or internal sovereign default. When we investigate floods in Table 14, this study finds increases of government debt in the entire postdisaster period. Thus, floods increase government debt compared to the synthetic control group. The growth of government debt is in line with the findings of Acevedo (2014) for floods. Another interesting finding is that the estimated effects are very consistent across different modes of disaster identification, like total affected, damage or deaths over land area. The average debt increase is 12.1% of GDP for total affected, 15.2% of GDP for damage and 13.1% of GDP for deaths. The findings for median disaster effect also prove relatively similar. Earthquakes are the only nonclimate event that we can investigate due to the limited number of volcanic eruptions in our sample. Just as in the standard disaster identification strategy, the total affected over land area results in an insufficient number of observations. Therefore, we only present the damage and deaths over land area in Table 15. We find a substantial average and median disaster impact for the disasters with highest damages per square kilometer. The disaster impact equals 26.8% of GDP for the average disaster impact and 20.8% for the median disaster impact. These effects are statistically and economically significant. The number of deaths per square kilometer also shows an increase of government debt compared to the synthetic control group. For the average disaster effect, 15

16 we find a debt increase of 8.7% of GDP, whereas, for the median disaster effect, our findings show a debt increase of 10.3% of GDP. [Insert Table 12, Table 13, Table 14 and Table 15 here] In summary, the findings for land area identification strategy reveal relative similar results when they are compared with the standard identification strategy. Our findings show a substantial increase of government debt for storms, floods and earthquakes. Droughts lead to a decline of government debt compared to synthetic control group. In general, storms have the largest adverse effect on the government s fiscal position, but the increase of government debt is observed across almost all specifications for floods. Thus, the positive effect on government debt is most consistent for floods. Disaster magnitude and disaster identification Previous natural disaster identification strategies are based on disaster outcomes. Although the outcomes are likely to affect the fiscal costs of the natural disaster, these measures do not account for the possibility of a severe disaster with a minimal loss of life or a minimal number of total affected. 20 The minimal impact could have to do with the level of preparation. A country that suffers from earthquakes regularly is likely to build its infrastructure accordingly (Jaramillo, 2007). This level of preparation also is related to the level of development. Koetsier (2017b) presents a clear positive relationship between the level of development and the preparedness of a country. Another way to identify a natural disaster is to investigate the magnitude of the natural disaster event. The level of development and preparation do not influence the magnitude of the natural disaster, but the magnitude of natural disaster influences the extent of the economic impact of the disaster. We apply a method that is developed by Klomp (2016a). To identify natural disasters, we relate the magnitude of the disaster to the country s inhabitants, as we are interested in natural disasters and not natural hazard. Natural hazards are natural events which occur in an unpopulated region (e.g. a desert). We study the fiscal impact of a natural disaster which requires deaths, people affected or damages resulting from the disaster. This study uses the equation (4) to identify natural disasters. ddddss iiii = αα ii + ββ 1 pppppp iiii + ββ 2 mmmmgg iiii + θθ 3 (pppppp iiii mmmmgg iiii ) + υυ iiii (4) We use the predicted values of equation (4), which includes the disaster dummies identified by our standard strategy. ddddddtt iiii is a dummy which is equal to one if a disaster occurs, identified as the deadliest, the most affected and the most damaging disasters. There is no immediate overlap between 20 The EM-DAT database includes indicators on deaths, total affected and direct damage. From these indicators, the direct damages are most incomplete. The completeness of the direct damages is partly related with the availability of disaster insurance. 16

17 the natural disasters identified by the standard strategy and the estimates for equation (4). This study investigates multiple types of natural disasters which are identified by a different intensity scale (e.g. wind speed or Richter scale). As a consequence, we cannot compare the magnitudes between disaster types. Therefore, the disaster composition identified using the standard strategy is kept constant. Equation (4) uses the population density of the country (pppppp iiii ) because this approximates the extent of the possible impact. 21 A higher population density can lead to a higher number of people affected, killed or a higher amount of damage. mmmmgg iiii is a vector of the magnitude of a natural hazard. We have data on the Richter scale, precipitation, temperature, wind speed and volcanic activity from Felbermayr and Gröschl (2014). To account for the time-invariant geographical, climatological and seismological characteristics, a country intercept (αα ii ) is included. An interaction term (pppppp iiii mmmmgg iiii ) is constructed to account for the possibility of a natural hazard occurring in nonpopulated area. Thus, this interaction term makes sure that the natural hazard indeed becomes a large natural disaster. The magnitude of disaster data is obtained from Felbermayr and Gröschl (2014). The reader should note that the magnitude data is only available for the period from 1979 to It contains monthly data on Richter scale, Volcanic Explosivity Index (VEI), wind speed, difference in mean temperature compared to the long-run mean and difference in rainfall compared to the long-run mean. We must adjust the temperature and rainfall measures to identify floods and droughts. This study uses the droughts identified by Felbermayr and Gröschl (2014). 23 However, it is necessary to differentiate between the different severities of droughts. The severity is indicated by the negative difference in mean rainfall. A similar approach is used for floods. The severity of floods is indicated by the positive difference in mean rainfall. The exogenous disaster identification strategy is used to assure that our previous estimates are not influenced by our choice to identify disaster using their outcomes. When we use the disaster magnitude to identify disasters, we find somewhat more mixed results for climate events (see Table 16). For total affected and damage, this study finds mixed evidence for the disaster effect on government debt. The deadliest climatological disasters reveal a positive effect on government debt. The average disaster effect shows an increase of government debt equal to 15.1% of GDP. The median disaster effect reveals a more moderate increase of government debt of 5.8% of GDP. These findings are all significant at the 1%-level. 21 City states are excluded due to their high population density. 22 We apply a predisaster period of ten years. Thus, the first disaster estimates are estimated from 1981 onwards. Even though, we do account for the disaster in the period In the previous estimations, we also use the years 2011 and This study follows the definition of drought by Felbermayr and Gröschl (2014). To measure the intensity of a drought, we must adjust this measure to allow for differentiation in drought severity. This study uses the difference in precipitation during the drought. We conduct additional sensitivity analysis on the existence of drought which continues for multiple years. This does not alter our results. 17

18 The nonclimate events substantially increase the level of government debt when we apply the exogenous disaster identification strategy. For total affected, this study identifies only a limited number of natural disasters in Table 17. The estimations, however, do reveal significant increases of government debt. We find a large average disaster effect on government debt. The most damaging disasters lead to an increase of government debt totaling 59.2% of GDP. However, the separate models show insignificant results. The median disaster effect shows a debt increase of 11.1% of GDP. Similar results are found for the deadliest disasters. These disasters show an increase of government debt equal to 68.1% of GDP. However, the median disaster effects reveal a considerably lower effect on government debt. The increase of government debt ranges from 8.8% to 16.0% of GDP. These results have a high statistical significance. In other words, there are some indications that several extreme disaster impacts might drive the average disaster impact. [Insert Table 16 and Table 17 here] Storms, on average, lead to a substantial increase in government debt (see Table 18). The size of the impact is relatively similar across disaster identification modes. There are insufficient disasters for total affected over population in model (4). For total affected over population, government debt increases by 6.4% of GDP in the aftermath of a natural disaster compared to the synthetic control group. The most damaging disasters cause an increase of government debt of 8.0% of GDP. The deadliest disasters cause an increase of government debt of 18.7% of GDP. The median effects are more moderate. The largest impact is observed for the deadliest disasters. Debt in this case increases by 2.9% of GDP at a 1%-significance level. Other climatological events (droughts and floods) are presented in Table 19 and Table 20, respectively. For droughts, we find a decline of government debt compared to the synthetic control group. This is similar to our previous results. However, we find mixed evidence for the debt impact of floods. For the deadliest disasters, we find a positive effect on debt. Government debt, on average, increases between 6.7% and 12.2%. The median disaster effect shows a debt increase of 7.6% of GDP. The findings for the deadliest disasters are in line with our previous results for floods. However, the other results clearly differ. For the other identification modes, we find a decline of government debt. [Insert Table 18, Table 19 and Table 20 here] Earthquakes clearly result in an increase of government debt. These results are highly similar to the findings for nonclimate events. 24 For the most damaging disasters, the models are statistically insignificant. When we combine the models, our estimations, however, show a large positive impact 24 This results from the dominance of earthquakes in the sample of nonclimate disasters. 18

19 on government debt. Government debt increases by 59.2% of GDP compared to the synthetic control group. Notice that the median effect is substantially lower than the average disaster effect. This effect shows an increase in government debt of 11.1% of GDP in Table 21. We obtain similar results for the deadliest disasters. The combined models reveal a debt increase of 74.0% of GDP compared to the synthetic control group. However, our findings for the median disaster impact show substantially lower fiscal impact. The median effect of the deadliest disasters equals 14.8% of GDP. These estimates are significant at the 1%-level. [Insert Table 21 here] The exogenous disaster identification strategy reveals somewhat different results. The most damaging disasters show more mixed results compared to the previous identification strategies. The explanation is that data on direct damages are predominately available for countries with higher economic development. Data on damages, for example, is obtained from damage assessments by large insurance firms. For the exogenous disaster identification strategy, the availability of direct damages is not necessarily a condition for disaster identification. This strategy uses the disaster magnitude, which does not require an estimate of damages. In general, we find additional government financing needs in the aftermath of a natural disaster compared to the synthetic control group in Table 22. Government debt increases are most pronounced for damages or deaths. For all disaster identification strategies, we find a positive effect on government debt for nonclimate events. For climate events, our findings generally show debt increases for storms and floods, whereas droughts lead to lower government debts. Our study also finds that grouping disaster types, such as done in other studies (such as, Ouattara and Strobl, 2013; Acevedo, 2014; Melecky and Raddatz 2011; 2015; Koetsier 2017a) is not a suitable strategy to estimate the impact of a natural disasters. Our findings show that different types of climatological disasters have very different postdisaster trajectories. In other words, its impact on the government s fiscal position differs between the different types of climate events. [Insert Table 22 here] We combine the different findings across disaster identification strategies. This enables us to provide one average and one median disaster effect for the type of disaster. There are no additional resources needed when government debt declines compared to the synthetic control group. We also assume that this is the case when we find mixed results. If we take the simple average of all disaster identification strategies 25, it gives an indication of the average or median impact per disaster type. In general, our 25 This approach gives equal importance to the different disaster identification strategies. 19