The impact of climate change on the occurrence of floods and the resulting economic costs in Europe

Master Thesis

International Economics and Business (M.Sc. & M.A.)

University of Groningen – Faculty of Economics and Business University of Göttingen – Faculty of Economic Sciences

The impact of climate change on the occurrence

of floods and the resulting economic costs

in Europe

Esther Platzbecker

Student number: S3191265 Date: 19th _{of June 2017}

Supervisor: Prof. Dr. C.J. Jepma (University of Groningen) Co-assessor: Dr. L. Birg (University of Göttingen)

Abstract

This thesis assesses the relationship between the anthropogenic climate change, precipitat io n patterns, and direct monetary flood losses. Panel regression models are applied to test whether the projected precipitation trends for Europe are already observable and to what extent these trends affect the consequential costs. Using a panel dataset of merged meteorologica l, economic and disaster data, the results show that the flood characteristics indeed changed and that the marginal climate change effect on direct flood losses is stronger than the effect resulting from increased wealth in exposed areas. Hence, this thesis demonstrates the severe consequences of intensified global warming and calls for actively taking preventive measures.

I

T

ABLE OF

C

ONTENT

List of Figures ... III List of Tables... IV

1 Introduction ... 1

2 Literature Review ... 3

2.1 Past occurrence of flood events ... 3

2.2 The influence of climate change on precipitation patterns ... 5

2.2.1 Observed trends in the past ... 5

2.2.2 Future projections... 6

2.3 Addressing flood damage ... 7

2.3.1 Types of flood losses ... 8

2.3.2 Flood risk assessment... 9

2.3.3 Economic determinants and estimation strategies ... 10

2.4 Research aim and hypotheses ... 12

3 Empirical Analysis ... 12

3.1 Data sources... 13

3.2 Part I: Detecting a climate change trend over the past ... 14

3.2.1 Description of variables ... 14

3.2.2 Methodology ... 16

3.3 Part II: Estimating the effect of climate change on flood losses ... 18

3.3.1 Description of variables ... 18

3.3.2 The empirical model... 20

4 Results ... 22

4.1 Part I: Detecting a climate change trend over the past ... 22

4.1.1 Frequency of flood occurrence... 22

4.1.2 Duration of flood events... 23

4.1.3 Precipitation extremes ... 24

4.2 Part II: The effect of climate change on direct flood losses ... 26

4.2.1 Baseline regression... 26

4.2.2 The climate change effect ... 29

II

5 Conclusion... 33

References ... 36

Appendix ... 41

A.1 Flood statistics ... 41

A.2 Climate change projections ... 42

A.3 Description of variables... 43

A.4 Summary statistics... 46

A.5 Results Part I ... 48

A.5.1 Frequency ... 48

A.5.2 Duration ... 51

A.5.3 Precipitation... 52

A.6 Results Part II ... 56

A.7 Robustness checks ... 57

A.7.1 Part I ... 57

III

L

IST OF

F

IGURES

Figure 1: N umber of relevant flood events worldwide 1980 – 2016 ... 41

Figure 2: N umber of catastrophic flood events worldwide 1980 – 2016... 41

Figure 3: O verall and insured losses in US$ of relevant flood events worldwide 1980 – 2016 .... 42

Figure 4: Projected changes in annual and summer precipitation 1961-1990 and 2071-2100 ... 42

Figure 5: Annual occurrence of all floods... 48

Figure 6: Annual occurrence of severe floods ... 48

Figure 7: Annual occurrence of general floods ... 48

Figure 8: Annual occurrence of flash floods ... 48

Figure 9: Annual number of floods North... 48

Figure 10: Annual number of floods Northwest ... 48

Figure 11: Annual number of floods Central ... 49

Figure 12: Annual number of floods East ... 49

Figure 13: Annual number of floods South... 49

Figure 14: Duration of all flood events ... 51

Figure 15: Duration of flood events North... 51

Figure 16: Duration of flood events Northwest ... 51

Figure 17: Duration of flood events Central ... 51

Figure 18: Duration of flood events East ... 51

Figure 19: Duration of flood events South... 51

Figure 20: Highest five day precipitation... 52

Figure 21: Highest five day precipitation North ... 52

Figure 22: Highest five day precipitation Northwest ... 53

Figure 23: Highest five day precipitation Central ... 53

Figure 24: Highest five day precipitation East... 53

Figure 25: Highest five day precipitation South ... 53

Figure 26: Proportional change in RX5day... 54

Figure 27: Proportional change in RX5day North ... 54

Figure 28: Proportional change in RX5day Northwest ... 54

Figure 29: Proportional change in RX5day Central... 54

Figure 30: Proportional change in RX5day East... 54

IV

L

IST OF

T

ABLES

Table 1: Final specification including a climate change effect ... 27

Table 2: Description of variables ... 43

Table 3: Summary statistics ... 46

Table 4: Correlation matrix ... 47

Table 5: Negative binomial regression model with country dummies: Frequency... 49

Table 6: Negative binomial regression with country dummies: Frequency of general floods ... 50

Table 7: Zero- inflated negative binomial regression: Frequency of flash floods ... 50

Table 8: Fixed-effects regression model: Average duration ... 52

Table 9: Random-effects regression model: RX5day ... 53

Table 10: Random-effects regression model: Proportional change in RX5day... 55

Table 11: Baseline regression ... 56

Table 12: Fixed-effects regression model: Frequency ... 57

Table 13: Fixed-effects model of frequency per region and flood type ... 58

Table 14: Negative binomial regression: Duration ... 59

Table 15: Panel regression models with large regional samples: Frequency a nd duration... 59

Table 16: Random-effects regression with large regional samples: Precipitation variables ... 60

Table 17: Random-effects regression model: Normalized losses ... 61

Table 18: Random-effects regression model with hazard interaction terms ... 62

1

1 I

NTRODUCTION

Over the last fifteen years, the number of major weather events in Europe has noticeably increased. Especially in the years 2002, 2005, 2010, and 2014, Europe was hit by nine to eleven catastrophic flood events that each caused between 50 and 11,600 million US$ of visible damage to vehicles, buildings, and infrastructure and in some cases resulted in more than 50 deaths (NatCatSERVICE, 2017). But also on a global level the high number of extreme weather events is alarming. The Global Risk Report 2017 ranks the risk from extreme weather events first in terms of likelihood and second in terms of impact (World Economic Forum, 2017). The increasing risk is very likely to be driven by the widely discussed anthropogenic climate change. Even though the climate has always been changing over time, for long periods it was questioned whether the currently observed changes exceeded the natural fluctuations due to human activities. However, the discussion ended when the Intergovernmental Panel on Climate Change (IPCC) published its Fourth Assessment Report providing clear evidence for a global warming that is caused by greenhouse gas emissio ns (Pachauri & Reisinger, 2007).

2 and temperature. While the annual temperature has been increasing in all European regions, precipitation patterns across Europe show a high spatial and temporal variability (EEA, 2012). Therefore, it is necessary to investigate the effect of climate change on precipitation patterns on a sub-continental level before making suggestions about the impact of climate change on monetary losses from flood events. In my thesis, I follow this path by first investigating the changes in flood frequency, flood duration and the precipitation amounts in order to detect a climate change trend on a sub-continental level. Thereafter, I relate these results to monetary flood losses. In particular, I am aiming to answers the following two research questions:

1. How did anthropogenic climate change affect the precipitation patterns and thereby the frequency, duration, and severity of flood events in Europe since 1980?

2. To what extent does the change in climate-related flood characteristics affect the annual amount of monetary flood losses per country?

While the first question aims at providing evidence for the impact of the anthropogenic climate change in Europe already being observable in terms of altered precipitation patterns, the second research question is of central importance not only to national governments but also to the insurance sector, which both are concerned about the potential implications of increasing flood losses. Given the positive association between climate change and monetary flood losses, future scenarios of a further increase in the average surface temperature are likely to involve immense monetary losses from extreme weather events. The fact that the insurance penetration against natural risks in Europe is still considerably low (Maccaferri et al., 2011) implies, that national governments have to engage in individual financial support, posing a burden to the national budget of many countries possibly resulting in long-lasting economic consequences (Melecky & Raddatz, 2011). Therefore, a climate change driven increase in monetary flood losses would entail substantial consequences not only for insurance systems of companies and private households but also for insurance companies themselves since they need to adjust their portfolio to the increasing risk of flooding.

In order to find answers to the research questions above, I divide the empirical analysis into two parts. In the first part, I use graphical analysis tools and panel regression models to investigate the trend in the frequency, the duration of floods, and precipitation extremes since 1980. I do this on a subcontinental level in order to detect variation both across time and across regions. My results provide strong evidence for a climate change since 1980 in Europe. The change is indicated by an increase in the number of floods in all regions, longer flood periods in Eastern Europe, and a strong increase in extreme precipitation amounts in Central Europe. Based on these findings, I use a random-effects model to estimate the impact of these changes on the annual amount of direct flood losses. I find that a marginal increase in either the frequency, the average duration of flood events, or in precipitation amounts leads to a strong increase in monetary flood losses. Overall, my findings suggest a strong impact of the anthropogenic climate change on flood losses.

3

2 L

ITERATURE

R

EVIEW

2.1 P

Over the last decades there has been an increasing number of flood events in Europe. Especially during the 1990s, where Europe was hit by several destructive floods each causing more than 1 billion US$ in losses. Some of the most severe floods occurred in 1994 in Italy, one year later in the Netherlands, in 1996 in Spain, and also in Eastern Europe, where Poland and the Czech Republic were hit by severe flood events in 1997 (Rojas et al., 2013). Following this flood- rich decade, several countries were affected by devastating floods in 2000. The most severe occurred in Italy and the UK, but also Switzerland, France, Spain, and Ireland were heavily affected. The absolute record in terms of total damages due to flood events in Europe was reached in 2002, when Germany, Italy, Austria, the Czech Republic, and France experienced the largest damages (NatCatSERVICE, 2017). Later years with high occurrence of large floods include the year 2007, when the UK was hit by a severe summertime flood, and 2013, when Central Europe suffered from several devastating floods (Rojas et al., 2013).

A potential reason for the high frequency of severe floods is the increase in extreme precipitat io n as well as in mean precipitation levels (Christensen & Christensen, 2003; Kundzewicz & Schellnhuber, 2004). A comprehensive analysis of major flood disasters in Europe between 1950 and 2005 is provided by Barredo (2007). He finds that regarding the geographical distribution of major floods in Europe, the most affected countries were Italy, Spain, France, and Germany. Barredo (2007) thereby differentiates between different types of floods such as between flash floods and river floods. Flash floods are commonly defined as “swift flood responses to intense rainfall or release of water over a small area” (Saharia et al., 2016, p.397) that bear the potential of a substantial impact on lives and infrastructure (Hong et al., 2013). Usually, they are characterized by localized, short, but very intensive rainfalls over basins with response times of only minutes or few hours (Saharia et al., 2016). Compared to this type of flood, river floods are caused by intense and/or persistent rain fall for several days or longer periods over large areas. The severity of these floods depends on several factors including weather and soil conditions, but also flood protection measures and land use. In Europe, river floods are strongly related to seasonal conditions. While in summer and autumn they are initiated by high regional precipitation amounts, in winter they are partly caused by snow and ice melt. Barredo (2007) finds that Spain was the most affected country by flash floods, whereas Italy, Germany, France, and Poland experienced the highest numbers of river floods in the considered time period (Barredo, 2007).

4 increasing between 1980 and 2016 (see A.1 Figure 1). However, in the case of catastrophic floods1_,

the time trend is not as strong since the global number of these events usually varies between three and fourteen events per year (see A.1 Figure 2). In regards to the absolute value of total damage caused by relevant floods2_{, the overall losses, expressed in 2016 values, abruptly jumped after 1992}

without showing a clear time trend afterwards (see A.1 Figure 3). The same holds true for catastrophic flood events. These sudden increases are partly based on a reporting bias (Kron et al., 2012) caused by the development of new communication technologies over the two past decades, which has made more information and data available and has facilitated its spread via the internet. As opposed to earlier times, nowadays, it is possible to identify even local and small-scale events such that there used to be severe underreporting of flood events. Nevertheless, in highly developed countries such as the Western European countries, the bias related to the past 35 to 45 years is assumed to be negligible (Kron et al., 2012).

Another aspect that contributes to the historical reporting bias is the development in politica l restrictions and boundary conditions. Many political regimes, in Latin America and Eastern Europe, for example, restricted the global sharing of information by the specific type of events that occurred. Hence, significantly less historica l data on both the frequency and loss amounts is available for those countries for earlier time periods (Kron et al., 2012). Furthermore, the observation of a positive trend in the overall frequency of floods but no clear time trend in the case of severe catastrophes relates to the fact that the effort in producing accurate data is smaller if the consequences of these events for politics and business are relatively small. The situation, however, changes in the case of large catastrophes where the consequences can be traced in society’s records, even decades after the event. As a result, the reporting bias seems to be much smaller for large disasters (Kron et al., 2012).

Apart from the reporting bias, there exist other explanations for the rise in flood losses over the past decades. According to the IPCC (2012), the major cause for the long-term increases in economic loss from weather-related disasters stems from both changes in climate and the socio-economic development. In line with this, Barredo (2009) reports that increased losses from river floods are driven by changes in population size, in economic wealth, and by extended activities in hazard-prone areas. Correcting for these changes in wealth and population, Visser et al. (2012) show that the normalization of disaster trends reveals stable trend patterns for economic losses for OECD countries. However, none of the scholars have explicitly considered the influence of changed precipitation patterns which are likely to be driven by the anthropogenic climate change. The following section provides some possible explanations for the existence of this research gap.

1 _{Accounted events have caused ≥ 1,000 fatalities and/or produced normalized losses ≥US$ 100m, 300m, 1bn, or 3bn}

depending on the assigned World Bank income group of the affected country (Munich Re, 2017).

2 _{Accounted events cause more than one fatalities or exceeded a normalized o verall loss of 100K, 300K, 1m or 3}

5

2.2 T

Despite the above mentioned issues, the authors of recent studies still suggest that the number of extreme flood events and also the associated damages in Europe have increased over the last decades (Kundzewicz et al., 2013). Nevertheless, until now, there exists no clear evidence that proves the causal relationship between the higher frequency or severity of floods, flood losses, and climate change, indicated by changed precipitation patterns (Rojas et al., 2013). Possibly, the detection of a trend is hindered by the interaction between socio-economic factors and climate driven physical causes (Barredo, 2009; Feyen et al., 2009; Elmer et al., 2012). As Barredo (2009) describes in his paper, floods are “the result of both societal and climatological factors” (p.97). Since after 1970 Europe has seen improvements in the standard of living, real per capita wealth and in population size, people started to move to flood-prone areas and therewith the exposure of people and assets to flooding increased (Barredo, 2009).

Another complicating factor affecting the detection of a possible trend is related to the statistica l analysis of extreme river discharges that form the basis in the assessment of trends in floods. The reason for this lies in the presence of the large uncertainties due to the natural variability of extreme events and the infrequency of the large events (Kundzewicz et al., 2005; Wilby et al., 2008). Since major floods occur relatively rarely there is only limited availability of observational records which can be used for an empirical analysis. Hence, the increase in the frequency and amount of damage in the last few years could either be driven by a real climatic trend, or be the result of randomne ss and/or socioeconomic changes (Barredo, 2007). Despite these complicating factors, several authors provide partial evidence for a positive relationship between the anthropogenic climate change and changed precipitation patterns, the most relevant of which will be discussed in the following two sections.

2.2.1 Observed trends in the past

According to Trenberth et al. (2007), the number of heavy precipitation events outside the tropical regions has been growing over the last 50 years, both at the continental and the global level. This is mainly driven by higher precipitation amounts per wet day due to a significant increase in the water vapor amount in the warmer atmosphere. Nevertheless, there exists a high inter-annual and inter-decadal variability that affect the rainfall statistics. Also with respect to spatial patterns, the changes in heavy precipitation are region- and season-specific and inconsistent across studies (Kundzewicz et al., 2013).

In contrast to the detection of climate change trends related to temperature, changes in precipitat io n across Europe show more spatial and temporal variability. While the average temperature in Europe in the last decade has significantly increased by 1.3°C as compared to the pre-industrial level indicating the warmest decade on record, there is no clear overall trend in precipitation levels on a continental level (EEA, 2012). However, the EEA suggests that since the mid-20th _{century, annual}

6 the same amount in some parts of Southern Europe. Nonetheless, the seasonal precipitation trends are influenced by large inter-annual variations. Between 1960 and 2012, in most Scandinavian and Baltic countries, as well as the northern parts of the UK, annual precipitation levels increased by more than 14 mm per decade with the strongest increase in western Norway. Compared to that, the largest decreases were seen in the Iberian Peninsula, especially in the northwestern regions, and in the northwestern regions of Italy.

The estimation of trends in precipitation extremes is more intricate than in the case of trends in mean precipitation since extreme precipitation events occur with low frequency only. Thus, it results in greater uncertainties when assessing the statistical significance of variations. Letting this this aside, it could be found that in the western European countries, the number of intense precipitation events has significantly rose (EEA, 2012). However, there is no significant trend in the number of consecutive wet days across Europe. On a subcontinental level, it could be found that in the northern European countries, in particular in the western parts of Finland, and in the western parts of the UK and Ireland the number of wet days per decade significantly increased. However, in the exact same regions the number of consecutive dry days also increased (EEA, 2012). One major threat from this development relates to the water retention capacity of the soil. Given that during longer periods of dry days the soil becomes more impermeable to water, a flood following a period of extreme rainfall becomes more probable. Hence, even in regions with observed decreases in mean precipitation levels, the flood risk can still be high (EEA, 2012). 2.2.2 Future projections

Given the past trends in precipitation patterns, future projections suggest even more severe changes. According to the EEA (2012), results from two EU-funded research projects predict a general increase in annual precipitation in Northern Europe and a decrease in Southern Europe until the end of this century. According to the ENSEMBLES project (Van der Linden & Mitchell, 2009), it is assumed that in Northern Europe precipitation levels will increase between 10 and 20%, whereas it will probably decline by 5 to 20% in the Mediterranean region (see A.2 Figure 4). In the case of northeastern European countries precipitation patterns are projected to remain constant or to slightly increase (Van der Linden & Mitchell, 2009).

7 Besides that, all future projections of climate change effects are highly dependent on the assumed emission scenarios published by the Intergovernmental Panel on Climate Change (IPCC) (Rojas et al., 2013). In general, the IPCC emission scenarios are divided into four qualitative storylines (families) that differ with respect to e.g. the expected development of economic growth, the changes in economic structures, the assumed introduction of new and efficient technologies, and the environmental protection levels. These four families provide the basis for six scenario groups. For instance, the often assumed A1B-scenario builds upon a future world of very rapid economic growth, a global population that peaks in the mid-century and declines thereafter, and the rapid introduction of new and more efficient technologies. Furthermore, it is assumed that the technological emphasis is balanced between fossil intensive and non-fossil energy sources (Nakićenović & Swart, 2000). Based on this A1B emission scenario, Rojas et al. (2012) estimate a pan-European hydrological model and observe a strong increase of over 40% in the future river flood hazard for the UK, southeastern and northwestern regions in France and northern Italy, while in Central Europe only decent increases of 10-30% are projected. Building on this paper, Rojas et al. (2013) analyze the socio-economic impacts of river floods in Europe by considering not only climate and socio-economic changes but also adaptation scenarios. The authors come to the result that on a country level France, the UK, and Italy, as well as Romania, Hungary, and the Czech Republic will probably suffer from the highest absolute damage costs.

Regarding possible explanations for this development, most scholars relate to the potential intensification of the hydrological cycle as a consequence of climate change (Christensen & Christensen, 2007; Van der Linden & Mitchell, 2009), which is likely to lead to an increase in both frequency and magnitude of intense precipitation levels in many parts of Europe (Christensen & Christensen, 2007; Nikulin et al., 2011; Van der Linden & Mitchell, 2009). This in turn would increase the risk of flooding in the future (Dankers & Feyen, 2009; Whitfield, 2012). In Northern Europe, the flood hazard is altered by the non-linear relationships between temperature and snowfall or rainfall. According to Kundzewicz and Radziejewski (2006), the rise in temperatures is likely to reduce early spring floods. However, this may also lead to compensation effects between rainfall and snow-driven river floods in currently snow-dominated areas. Hence, the projections of future flood risks are highly uncertain in those regions (Dankers & Feyen, 2009; Rojas et al., 2012).

2.3 A

8 2.3.1 Types of flood losses

Most scholars divide flood losses into tangible and intangible damages as well as into direct and indirect losses. While tangible damages can be specified in monetary terms and refer to damage s to man-made capital or the resource flow, intangible damages are defined as the damage to assets which can hardly be monetarized and which is moreover not traded in the market (Merz et al., 2010). Apart from that, direct damages refer to all damages resulting from the physical contact with flood water, whereas indirect losses are induced by direct impacts but occur outside to the flood event. An example for direct and tangible losses is the damage to buildings, whereas human losses are in most cases considered to be direct and intangible (Merz et al., 2010). It is highly controversial whether human life can be addressed in monetary values, i.e. whether the loss of life is tangible or intangible (Viscusi & Aldy, 2003). This reveals that even though these four categories are commonly used, their interpretation and allocations differ strongly.

The global reinsurance company Munich Re distinguishes between direct losses, indirect losses and secondary or consequential losses (Munich Re, 2001). According to their classification, direct losses are tangible and comprise all immediately visible and countable losses such as the loss or damage to household property, vehicles, building, and infrastructure. The amount of direct losses is calculated on the basis of replacement and repair costs. Nevertheless, this calculation method does not come without limitations. First, it is problematic when historical quarters or cultura l heritage is involved as their value is difficult to estimate, and second, the resulting numbers reflect costs rather than the value of damaged objects since damaged structures such as dykes are often upgraded while being repaired or replaced (Kron et al., 2012). Indirect losses, on the other hand, refer to all costs and losses that are related to the natural catastrophe resulting from business interruption and power failure. This involves the loss of jobs or of rental income, or higher transport costs due to infrastructure damage, for example (Kron et al., 2012). Lastly, consequential losses or secondary losses reflect the economic consequences of a natural catastrophe over a longer term. For example, floods may reduce the gross domestic product (GDP), or may lower tax revenues and weaken the currency (Kron et al., 2012). However, some scholars report that floods also bear the potential to stimulate the economy in the long-term. Fomby et al. (2013) for example find a positive economic growth effect of floods that is driven by a positive response of agricultural growth one year after the event and a later positive response of non-agricultural growth. In the latter case, the authors suggest that the growth effect in the non-agricultural sector is driven by transmiss io n mechanisms based on supply chain relationships across sectors.

9 Munich Re. I refrain from using human losses since I am predominantly interested in the monetary damages as they are highly relevant for both governments and the insurance sector.

2.3.2 Flood risk assessment

In order to assess the determinants of flood losses, various factors need to be taken into account. As mentioned above, these factors do not only include external effects such as climate conditions but also wealth effects and parameters that determine the vulnerability of regions and countries to flooding. Kunreuther and Michel-Kerjan (2007) suggest developing a model that combines data on the loss history of a country or region with experts’ knowledge on a particular risk in order to estimate the direct economic losses of flood events. According to them, such a model consists of four components: hazard, inventory, vulnerability, and the loss. The IPCC (2012) defines hazard as “the potential occurrence of a natural or human-induced physical event that may cause loss of life, injury, or other health impacts, as well as damage and loss to property, infrastruct ure, livelihoods, service provision, and environmental resources”(IPCC, 2012, p. 32). In the case of floods, hazard can be determined by e.g. the maximum discharge levels, the water depth, or the water flow velocity, whereas inventory refers to the properties at risk. The IPCC (2012) alternatively uses the expression exposure, which is referred to as “the presence (location) of people, livelihoods, environmental services and resources, infrastructure, or economic, social, or cultural assets in places that could be adversely affected by physical events and which, thereby, are subject to potential future harm, loss, or damage” (IPCC, 2012, p. 32). These properties can be further specified with respect to their geographic coordinates, the construction type, the number of stories of a building, the age, etc. If both, the hazard and the inventory are known, the vulnerabilit y of a certain structure or a region can be calculated, which is needed to evaluate the resulting direct economic loss (Kunreuther & Michel-Kerjan, 2007).

Until the beginning of the last decade, most scholars and insurance companies used the output of such a catastrophe model to construct an exceedance probability curve. The curve identifies the probabilities that a certain loss level, in form of the monetary value of damages or fatalities, will be surpassed in a given location over a specific time period (Kunreuther & Michel-Kerjan, 2007). However, during the last decade, more research on the adequate risk assessment of floods has been done that additionally includes possible climate change scenarios and the socio-economic development of a region (Alfieri et al., 2015; Rojas, et al., 2013). In general, flood risk is defined as the combination of the probability of the occurrence of a flood event and of the potential adverse consequences for human health, the environment, cultural heritage, and economic activit y associated with a flood event (EU Floods Directive, 2007).

10 adequate assessment of flood risks requires extensive information, on the asset distribution, the topography of the region, the population density, and the potential damage function as explained before (Alfieri et al., 2015), until now, most scholars have evaluated flood risk at river basin scales (Falter et al., 2015; te Linde et al., 2011) or national scales (Hall et al., 2005). Only a few scholars have applied the risk of floods on a global or continental scale (Alfieri et al., 2015; Dankers et al., 2014) , which this paper does, however.

2.3.3 Economic determinants and estimation strategies

Apart from the literature on the determining factors of flood losses through risk assessment strategies, there exists a number of studies that aim at estimating the direct costs of floods and other natural disasters from an economic perspective. According to Cavallo and Noy (2010), most papers that analyze the determinants of direct disaster costs use panel regression analysis of the form

𝐷𝐼𝑆𝑖𝑡=∝ +𝛽𝑿𝑖𝑡+ 𝜀𝑖𝑡, (1)

where 𝐷𝐼𝑆_𝑖𝑡 is a measure of direct damages of disaster in country i and time t, and 𝑿_𝑖𝑡 a vector of control variables of interest. Apart from applying panel regression methods, some papers use cross-section estimation techniques by aggregating data across time (Cavallo & Noy, 2010). In that case, the basic estimation model is very similar to the panel estimation model with the only difference being that the variables are expressed as averages across the estimated time period. In both models, the vector of control variables usually includes a measure of disaster magnitude. In the case of flood events one could think of precipitation amount, water depth, or duration of a flood as adequate measure of magnitude. Water depth is relevant since it strongly affects the damage to buildings, for instance, because absorptive materials swell up and burst (Hausmann, 1998; Merz et al., 2007). The flood volume will e.g. determine the design of flood retention measures (de Moel et al., 2015) and longer periods of inundation result in greater losses because metals corrode, organic materials rot, or watertight buildings are torn from their moorings as ground water rises (Hausmann, 1998). As explained in the previous section, the vector of control variables should additionally capture the vulnerability of a country or a region to the disaster (Cavallo & Noy, 2010). One major factor that determines the vulnerability of a country to an external shock is the level of economic development (Kahn, 2005). It is argued that on the one hand, advanced countries are hit more strongly by natural disasters in terms of absolute capital losses as households are richer and thus more value can be possibly damaged. On the other hand, richer countries have the means to spend money on preventive measures such as the implementation of building codes, land-use planning and engineering interventions (Freeman et al., 2003; Jaramillo, 2009). Due to this ambiguity in the effect of higher economic development on vulnerability, Kellenberg and Mobarak (2008) suggest a non-linear relationship, stating that due to changing behaviors the disaster risk initially increases with higher incomes as people move to popular coastal more flood-prone locations within a city. Sadowski and Sutter (2005) support this finding for the case of hurricanes in the United States, where residents seem to move to coastal areas if the level of preparedness is increased.

11 hazard exposure measure they use takes both the likelihood of a disaster event and the local exposure in terms of population into account. They come to the result that for countries that face a low level of hazard, losses first increase with rising economic development and later decrease with it. Whereas in countries that face high hazard of disasters, losses first decrease and then increase with rising incomes. The authors explain that given the preferences of low hazard countries, less preventive measures are taken for low levels of wealth but undertaking such measures becomes more profitable as wealth levels rise. On the contrary, in high hazard countries, they will invest in preventive measures even at low levels of economic development since this can mitigate losses, given that the marginal benefits from these expenditures outweigh their costs. However, the marginal benefits from further preventive investments are declining and in addition, the potential for preventive expenditures is limited. That is why with higher levels of wealth, those countries are likely to experience increasing losses (Schumacher & Strobl, 2011).

Apart from the level of development, most papers control for country size, measured e.g. by the size of population, the land area, or GDP. The underlying idea is that in bigger countries more value is exposed to direct damages. In line with this, Cavallo et al. (2010) observe higher direct economic losses from natural disasters for larger countries. Another relevant factor that was found to determine the vulnerability of a region is the geographic location. Even though it is straight forward that being located close to a river or in floodplains increases the risk of being exposed to a flood, the critical locations may change depending on the type of flood. While river floods can only occur in areas close to river basins, the locations exposed to flash floods are more complex to define as they occur as independent and random events (Munich Re, 2015). So with respect to the relationship between the geographic location and the vulnerability of a region or country, if possible, one should distinguish between flood types.

Moreover, several scholars come to the conclusion that better institutions reduce the impact of a disaster. However, most of these studies measure the direct impact of disasters in terms of fatalit ies. For example, it could be shown that countries with greater security of property and a stable democratic regime suffer from less deaths due to natural disasters (Kahn, 2005). In terms of both human and economic losses Toya and Skidmore (2007) find that greater openness and more

complete financial systems are associated with fewer losses. Raschky (2008) provides support for

12 effect of urbanization is found to be smaller than the wealth effect. According to Choi (2016), a possible reason is the large difference in control and management mechanisms within the urbanization process across the OECD countries. In the case of well-managed urbanization, in form of land use regulations, for instance (Choi, 2010), people may gain access to superior socio-economic institutions, better designed infrastructure, and urban planning that is likely to reduce the disaster hazard in cities. Hence, urbanization does not necessarily lead to an increase in economic losses from natural disasters (Choi, 2016).

All in all, many factors have already been found to determine the monetary economic losses from natural disasters. Nevertheless, none of the existing papers has explicitly considered the effect of the anthropogenic climate change on these losses. This paper closes this gap for a pan-European country sample by adding a climate change variable as well as flood characteristics variables to the standard estimation models that are directly influenced by the anthropogenic climate change.

2.4

R

I aim at investigating two major aspects that are related to climate change and the occurrence of flood events in Europe. Based on the previous findings of other scholars, I first try to detect a climate change trend that is related to extreme precipitation events. I therefore analyze not only the changes in the frequency and duration of flood events, but also in precipitation patterns in five European regions that were hit by a flood event between 1980 and 2016. My first hypothesis hereby consists of the following parts: (1) I expect to find a significant increase in both the annual number of floods, as well as in their duration in all European regions except for the southern region, where I expect a decrease in both parameters since precipitation is projected to decrease over time (Seneviratne et al., 2012). (2) In line with the finding of the European Environment Agency (EEA, 2012), I expect that extreme precipitation has been over-proportionally increasing in the northern, northwestern, central, and eastern European region, while it has been decreasing for the Southern Europe.

In the second part, I aim at assessing both the channels and the extent to which these changed precipitation patterns affect the direct economic flood losses. I hereby control for other loss determining factors, in particular for the wealth and urbanization effect that have been found to be major drivers of these losses. In order to evaluate which effect prevails, I furthermore compare not only the significance of the climate change effect with the two other effects but also their magnitudes. In particular, I test the hypothesis that the anthropogenic climate change leads to higher economic flood losses due to its impact on the duration, frequency and severity of floods.

3 E

MPIRICAL

A

NALYSIS

13 and investigates the effect of the previously detected climate change on direct flood losses. I therefore control for the wealth effect, the urbanization effect, and other characteristics that influence the vulnerability to floods of countries.

I choose to evaluate the climate change effect on flood damages for a pan-European sample since the effect of global warming on the frequency and intensity of heavy precipitation events is found to be among the strongest for this region (Allen et al., 2014). Another reason relates to the factors that determine the monetary value of flood losses but that are not included in my empirical model given the existing data constraints. Some of these factors relate to the degree and quality of established preventive measures or the quality of institutions that are responsible for damage mitigation. In order to mitigate a potential omitted variable bias, I choose a country sample that is likely to differ less with respect to these unobserved factors given the spatial, historical, and economic similarities. Furthermore, I choose the study period between 1980 and 2016 since the data availability for earlier periods is very limited.

3.1 D

In my empirical analysis I use three different data sources. The first source is the NatCatSERVICE dataset that contains information about flood events in Europe covering the period 1980-2016. It is provided by the global reinsurance company Munich Re and includes the begin and end date of a flood, the catastrophe class, the event type (general flood or flash flood), the country of occurrence, the affected areas within a country, the geocoding of the location that was most-affected, the number of fatalities, the monetary losses in terms of direct overall losses and insured losses, as well as estimated loss values in the case where no official institution has published loss estimates. The data is based on systematic data sourcing from surveys that are conducted through internet queries, news portals, direct contacts to institutions, international offices, and companies that collect or provide regional or local information on natural disasters. In the case of contradictory information from different sources internal experts of the Munich Re re-assess the information so that only consolidated data enters to the NatCatSERVICE database (Munich Re, 2017). Therefore, I consider the data very reliable.

As second data source, I make use of the World Development Indicators (WDI) provided by the World Bank, which are collected from officially recognized international sources and involve the most current and accurate global development data that is available (World Bank, 2017). All variables of my regression model that describe country characteristics are based on the WDI, i.e. GDP per capita, the share of urban population, the population size, the land area, the population density, and the share of permanent cropland.

14 quality controlled data from their national meteorological station networks. Since 2010, it has gained the status of Regional Climate Centre for high-resolution observation data in Region VI (Europe and the Middle East) of the World Meteorological Organization (Klein Tank et al., 2002). I use this third data source for variables that are related to precipitation amounts.

By merging the data from these three data sources, I obtain an unbalanced panel dataset with 659 observations that includes 44 pan-European countries and covers the time period between 1980 and 2016, i.e. a time span of 37 years. While half of the countries are observed over 14 years or less still 25 percent of all countries are observed over 23 and more years.

3.2 P

I:

D

3.2.1 Description of variables

In order to detect a climate change trend over the time period of 1980-2016 on a subcontine nta l level in Europe, I consider three parameter groups that are related to flood events and that are likely to be influenced by the anthropogenic climate change, according to the IPCC (2012). In particular, I consider the frequency and the duration of flood events, as well as a precipitation index, i.e. the annual highest 5-day precipitation amount (RX5day) that serves as indicator for extreme precipitation levels in the year of flood occurrence.

Regarding the first parameter of interest, the variable frequency describes the annual number of floods that occurs in a given country. It is constructed by aggregating the number of floods that is given in the event-based dataset from the NatCatSERVICE per year and country. As can be seen in the descriptive statistics in Table 3 (see Appendix A.4), the count variable ranges from 1 to 13 with a mean value of 2.3. Its distribution is positively skewed and has a standard deviation of 1.93, so the variance of around 3.71 is greater than the mean, which will be considered when choosing the specification of the econometric model. Furthermore, I distinguish between small and severe floods by making use of the catastrophe classification given in the NatCatSERVICE database. The classification is based on both the number of fatalities caused by the events and loss values relative to the income group the country is assigned to by the World Bank (Munich Re, 2017). Since it ranges from 0 (marginal event) to 4 (catastrophic event), I define a severe flood as a flood that is classified 3 or higher. Moreover, I distinguish between the frequency of occurrence for two different flood types, i.e. for general floods and flash floods. The first type includes mostly river floods since they exclude coastal floods and floods due to storms; the second type refers to flash flood events that are caused by short but intense rainfall. While the variable that describes the annual number of general floods per country ranges from 0 to 8 with a mean value of 1.3 and a standard deviation of 1.25, the annual number of flash floods ranges from 0 to 11 and has a mean value of 1 and a standard deviation of 1.56.

15 Appendix A.4), the variable ranges from 0.4 to 31 days and has a mean value of 4.3. Regarding the distribution, it can be noticed that it is positively skewed and that the standard deviation of 5.47 is considerably large. Hence, most floods last for only few days.

The last parameter that is used in order to detect a climate change trend in Europe over the last decades is concerned with the amount of extreme precipitation. I use the precipitation index

RX5day from the ECA&D because it is the most related to extreme precipitation events. Compared

to the other precipitation indices that are provided by the ECA&D, it particularly relates to a continuous time interval of heavy precipitation rather than to mean precipitation values. I refrain from using mean values since an increase in mean precipitation is not automatically associated with a higher number of floods, whereas heavy precipitation over a certain time period is more likely to cause flooding.

Since the RX5day index is based on raw rainfall data from meteorological stations of the ECA&D project, I need to construct a new variable for my panel dataset in the following way. I use the

RX5day index which gives the highest 5-day precipitation amount measured by the differe nt

meteorological stations on an annual, half-annual, seasonal, and monthly basis for a maximum time period of 1860-2016. Based on the exact location of each flood event, which is given in the NatCatSERVICE dataset, I choose the weather station within the country of occurrence that is closest to the event in terms of longitude and latitude3_{. However, if there are less than five}

meteorological stations available for one country, I choose the station with the highest value in the year of occurrence because the distance between the event and the weather stations is large in any case. Given the selection of stations, I assign the annual value of the highest 5-day precipitat io n amount for the year in which the respective flood occurred to the considered flood event. Since I aggregate the flood events per year and country, I then compare the selected precipitation amount of each flood event in country i and year t and include the highest value into the panel dataset. This gives the variable RX5day which is used in both parts of the empirical analysis. Additionally, in order to control for large differences in the absolute amount of precipitation across weather stations within one country, I calculate the average value of the RX5day index over the 20-year period previous to my period of study, so 1960-1980. For every year and country, I take the difference of the RX5day index to this pre-average and divide the difference by the same in order to obtain the variable propchange that describes the proportional change in the highest 5-day precipitat io n relative to the period 1960-1980.

Concerning the statistical distribution of these two precipitation variables, the variable RX5day ranges from 5.9 mm to 328.3 mm with a mean value of 106.1 mm and a standard deviation of 54mm. It is positively skewed and has a kurtosis that exceeds the one of a normal distribution. The second parameter, propchange, covers values between – 0.89 and 2.85 and has a mean value of 0.25. This indicates that aggregating across all countries and years, on average, the highest 5-day precipitation amount has increased compared to the amounts in the time between 1960 and 1980.

3 _{The only case where a weather station was chosen that lies in a different country is Andorra because no data was}

16 Moreover, I define indicator variables for five geographical regions that, based on the future climate change projections of the EEA (2012), are likely to be differently affected by the anthropogenic climate change. The first region, north, includes seven countries, the second region,

northwest, six, the central region also covers six countries, the eastern region contains another

fourteen countries, and finally the region south contains eleven countries. 3.2.2 Methodology

In order to detect a time trend in the above described variables that are potentially affected by the anthropogenic climate change, I use both graphical analysis tools that plot the variables of interest against time as well as panel regression analysis, where the basic regression model is given by the following equation:

𝐹𝑙𝑜𝑜𝑑 𝑚𝑒𝑎𝑠𝑢𝑟𝑒_𝑖𝑡= 𝛼 + 𝛽₁𝐷90_𝑖𝑡+ 𝛽₂𝐷00_𝑖𝑡+ 𝛽₃𝐷10_𝑖𝑡 + 𝜀_𝑖𝑡, (2)

In equation (2), the dependent variable 𝐹𝑙𝑜𝑜𝑑 𝑚𝑒𝑎𝑠𝑢𝑟𝑒_𝑖𝑡 indicates the four alternative variables of interest, i.e. the frequency and the average duration of flood events that occur in country i in year t, as well as the associated precipitation index RX5day and its propchange. The three explanatory variables D90, D00, and D10 are indicator variables for the decades 1990s, 2000s, and 2010s, so I choose the 1980s as a baseline decade. The estimated coefficients of the regression model thus indicate the marginal change in the respective flood measure for a given country and the considered decade compared to the 1980s. I prefer using indicator variables for decades over year dummies since the trend over decades should exceed the trend over single years. Furthermore, I make use of the regional division of my sample into the regions North, Northwest, Central, East, and South and estimate the regression model for each of those five regions in order to detect the spatial distribution of precipitation patterns.

Concerning the exact specification of the regression model, I first consider the type of the dependent variable in order to choose a different regression model for count and continuo us variables. In the case of count outcome variables, I base my decision between a Poisson and a negative binomial regression model on the statistical distribution of the considered dependent variable. Given a continuous dependent variable, I use the Hausman test to test whether the error terms are significantly correlated with the regressors. If they are, I use the fixed-effects model that controls for country-specific time-invariant characteristics. Otherwise, I use the random-effects model, which in that case produces a more efficient estimator. Moreover, I use country-clustered standard errors in order to control for potential heterogeneity and correlation.

17 estimates (Kellenberg & Mobarak, 2008). I choose the negative binomial regression model over a Poisson regression model since the count outcome variable frequency is over-dispersed as indicated by a variance that exceeds the mean (see Table 3 in A.4), and thus the central assumption of a mean-variance-ratio to equal 1 in the Poisson model is violated. Since I aim at detecting a time trend in the frequency of floods and therefore try to control for all other time-invariant factors that are not included in the model, a fixed-effects model would be an appropriate choice. However, one major problem with regard to negative binomial regression models in combination with panel data is that the fixed-effects negative binomial regression model introduced by Hausman, Hall, and Griliches (1984) which is used by the statistical software Stata is not a true fixed-effects model since it allows for an arbitrary intercept for each individual or country which is not possible in the usual conditional fixed-effects models (Allison & Waterman, 2002). I hence follow the suggestio n by Allison and Waterman, (2002) and specify a conventional negative binomial regression (NB2) model with country dummy variables to estimate the fixed effects as the authors show that in such a model the resulting coefficients are still consistent. Furthermore, I use country-clustered error terms.

In line with my first hypothesis, I expect to find a positive trend in the annual number of floods over time especially in the northern, northwestern, the central, and eastern region, where it is projected that climate change leads to an increasing frequency of extreme weather events due to changes in precipitation patterns (IPCC, 2012). Nevertheless, as explained in the first section of this paper, there might exist a reporting bias that artificially drives the frequency of floods for later periods. Even though it is suggested that for high-income countries the bias is rather small (Kron et al., 2012), in my graphical analysis I test for this bias by comparing the results that include all types of flood events with the results from subsamples that are expected to be less prone to a reporting bias. In particular, since it has been found that the reporting bias is much smaller for

severe flood events (Kron et al., 2012), in my graphical analysis, I consider a subsample that

includes only events that are classified as “severe” according to the NatCatSERVICE, so events that are classified 3 or higher (Munich Re, 2017). I do not test for a potential reporting bias in my panel regression model since the flood events are aggregated across countries and years and thus, the severity of single floods is not observable.

18 Furthermore, I investigate the trend in the annual highest 5-day precipitation amounts (RX5day) for countries that were hit by flood events between 1980 and 2016 by using scatterplots first and a random-effects model afterwards. I use a random-effects model because the Hausman test reveals no significant differences between the coefficients of the fixed-effects model and the random-effects model and thus the coefficients of the latter are more efficient. Since the absolute amount of precipitation is however likely to vary strongly across regions, e.g. between mountainous regions and urban regions, I will use the proportional change of the 5-day precipitation amount as alternative outcome variable. In line with my first hypothesis that is based on the climate change projections of the ENSEMBLES project (Van der Linden & Mitchell, 2009), I expect to find a positive and significant time trend in extreme precipitation amounts, especially in the proportional change of the RX5day index, for the northern, northwestern, central and eastern European regions. For the southern European region, however, I expect to find a decreasing trend in precipitat io n patterns.

3.3 P

II:

E

3.3.1 Description of variables

19 availability. Concerning the distribution of the outcome variable (see Table 3 in A.4), it can be noticed that the inflation adjusted flood losses vary between 7,287 US$ and 11.6 billion US$. However, the distribution is highly positively skewed and the mean value of 230 million US$ lies to the right of the median which is slightly less than 5 million US$.

As the review on the central natural disaster literature shows, the determining factors for direct economic flood losses can be broadly divided into three groups: the variables that describe the flood hazard, those that indicate the exposure, and finally those variables that characterize the

vulnerability of countries to flood losses. While the first group consists of all flood characterist ics

that are analyzed in the first part of my analysis, the second group primarily relates to the exposed wealth in a region indicated by the per capita income of a country. Finally, the vulnerability is determined by other country characteristics such as size of total population, population density or the share, of urban population.

Since I consider the same variables that describe the characteristics and precipitation patterns related to flood events, the definition and data sources for the first group of explanatory variables remain the same as in the first part of my empirical analysis. However, in order to directly estimate the effect of changed precipitation patterns due to climate change, I additionally construct a climate change indicator variable, climate trend. It is based on the proportional change of the highest 5-day precipitation amount from the ECA&D project. In a first step, I therefore calculate the mean proportional change for each country in the last decade of my study period (meanprop) since it indicates the most current proportional change in extreme precipitation relative to the time period between 1960 and 1980. Hence, it is an indicator for a climate change trend over the whole period of consideration and equals 1 if the calculated mean proportional change in the current decade is greater than 0 and 0 otherwise. In order to check whether my results are robust to other specifications of the climate change variable, I will later extend the time period that is used to calculate the mean proportional change and additionally use a higher threshold value than 0 that needs to be exceeded so that the indicator variable climate trend is equal to 1.

Apart from the variables related to the climate change trend, the main variables that I expect to have a significant impact on relative flood losses are GDP per capita and the degree of

urbanization. The first is taken from the World development indicators (WDI) and is expressed in

purchasing power parity (PPP) and constant 2011 values by converting GDP to international dollars using purchasing power parity. It ranges from 1,810 US$ to 65,072 US$ in the given sample with a mean of 27,767.5 US$. The variable urban is given by the share of people living in urban areas as percentage to the population and covers a range between 15.41% and 100%. Aggregating over all countries and years in the sample, the average degree of urbanization is equal to 68.3%. All other country characteristic variables that characterize both the exposure as well as the vulnerability of countries to flood events are taken from the WDI. In particular, I consider the total

population of country i which is given by the midyear estimates of the total population based on

20 area. The land area refers to the respective country's total area expressed in squared kilometers. Finally, I consider two variables that, at least partly, describe the land use and land cover. The first is the share of permanent cropland, which refers to the land that is cultivated with crops occupying the land for long periods, the second is an indicator variable for mountainous countries that equals 1 if a country is mostly covered by mountain regions (EEA, 2012).

Testing for multicollinearity between all explanatory variables, the correlation matrix (see A.4 Table 4) indicates that none of the explanatory variables are highly correlated since all correlation coefficients are below 0.8, which is a common threshold value for multicollinearity. The only coefficient that slightly exceeds this threshold is found between land area and total population. In order to avoid problems related to multicollinearity, I therefore use only the population size as a control for country size. This leads me to the specification of the empirical model that is used in order to detect the impact of climate change on the direct economic flood losses.

3.3.2 The empirical model

Similar to most scholars that have analyzed the economic losses of natural disasters (Cavallo & Noy, 2010), I apply a panel regression model in order to observe the changes over time for differe nt countries. In my estimation model, I decompose the amount of flood losses into a climate change effect and two other effects that have been found to play a strong determining role for economic losses from natural disasters. In particular, apart from other country characteristics, I control for a wealth effect and an urbanization effect. The baseline estimation regression is characterized by the following equation:

ln(𝑙𝑜𝑠𝑠)𝑖𝑡= 𝛼 + 𝛽1ln(𝑑𝑢𝑟𝑎𝑡𝑖𝑜𝑛)𝑖𝑡+ 𝛽2ln(𝑅𝑋5𝑑𝑎𝑦)𝑖𝑡+ 𝛽3ln(𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦)𝑖𝑡+ 𝛽4ln(𝐺𝐷𝑃𝑝𝑐)𝑖𝑡+

𝛽5ln(𝐺𝐷𝑃𝑝𝑐)2𝑖𝑡+ 𝛽5𝑢𝑟𝑏𝑎𝑛𝑖𝑡+ 𝜆𝑗∑𝑚𝑗=1𝑋𝑖𝑡+ 𝑢𝑖+ 𝜀𝑖𝑡, (2)

where the dependent variable, ln(𝑙𝑜𝑠𝑠)𝑖𝑡, is given by the natural logarithm of the annual and

inflation adjusted flood losses in country i and year t. I refrain from normalizing losses to wealth and population as suggested by Neumayer and Barthel (2011) since my regression model controls for these effects in form of explanatory variables. The first three independent variables in the baseline regression equation capture the flood characteristics and flood magnitude. They are expressed in natural logarithms so that the coefficients can be interpreted as elasticities. The parameter ln(𝑑𝑢𝑟𝑎𝑡𝑖𝑜𝑛)𝑖𝑡 is given by the average number of days of inundation due to a flood event

in country i and year t. Ln(𝑅𝑋5𝑑𝑎𝑦)𝑖𝑡 refers to the absolute highest 5-day precipitation amount in

the affected region in country i and year t and serves as flood severity indicator. The third flood variable, ln(𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦)𝑖𝑡, is given by the total number of flood events that occurred in country i

21 and Mobarak (2008), I expect an inverted U-shaped relationship between income and economic flood losses, i.e. a positive coefficient for the normal and a negative coefficient for the squared term. However, my results may differ because Schumacher and Strobl (2011) show that the relationship between income and disaster losses depends on the disaster hazard that the country of consideration faces. In order to control for this, as a robustness check, I will subdivide my sample into two groups, one that faces a relatively low hazard and one that is affected by a high number of floods over the research period. Thereby, I investigate whether the results by Schumacher and Strobl (2011) also apply to my pan-European country sample.

Furthermore, the explanatory variable 𝑢𝑟𝑏𝑎𝑛𝑖𝑡 captures the urbanization effect that is determined

by the degree of urbanization in a country and for which mixed results were found before. While Kellenberg and Mobarak (2008) find a positive and significant impact of urbanization on disaster risk for the case of earthquakes and a negative but insignificant coefficient for flood events, Choi (2016) comes to the conclusion that the impact of urbanization on disaster losses depends on the management and institutions related to the urbanization process. He argues that on the one hand, urbanization can lead to higher losses from natural disasters since it comes with a large number of people that is exposed to disaster events. On the other hand, urbanization can reduce disaster losses because it enables people the access to well-designed infrastructure and urban planning, given that the process is well-managed. Regarding my country sample, I expect that the latter is the case since I assume the overall quality of institutions in most European countries to be relatively high compared to countries in other regions. Therefore, I suggest a negative coefficient of the variable

𝑢𝑟𝑏𝑎𝑛𝑖𝑡.

The last explanatory variable of the baseline regression, ∑𝑚𝑗=1𝑋𝑖𝑡, is a set of time and/or

cross-country varying variables that capture the vulnerability of a cross-country to flood events. It includes the natural logarithm of the population size, the natural logarithm of population density, an indicator variable for mountainous countries, and natural logarithm of the share of permanent cropland. I do not only expect a positive coefficient for the two population variables, but also for the mountain dummy because the degree of inclination is associated with a higher flow velocity of water which then affects flood damages (Merz et al., 2010). In contrast to that, I expect the variable of permanent cropland to have a negative impact on flood losses since the overall water retention capacity of the soil should increase with a higher share.

Again, I base the decision between a fixed-effects and the random-effects model on the results of the Hausman test, which this time speaks in favor of a random-effects model. Therefore, the term

𝑢𝑖 represents a random individual effect for each country and 𝜀𝑖𝑡is the usual regression random