Impact of Natural Disasters on Financial Development

Subhani Keerthiratne1 Richard S. J. Tol1

1

Department of Economics, University of Sussex, Falmer, Brighton, BN1 9SL, United Kingdom. Correspondence: Subhani Keerthiratne, Department of Economics, Jubilee Building, University of Sussex, Falmer, Brighton, BN1 9SL, United Kingdom. Tel: +447404282802. E-mail: [email protected]

Abstract We attempt to ascertain the impact of natural disasters on financial development proxied by private credit. We employ a panel fixed effects estimator as our main estimation tool on a country level panel data set of natural disasters and other economic indicators covering 147 countries for the period from 1979 to 2011. We find that companies and households get deeper into debt after a natural disaster. This effect is stronger in poorer countries whilst the effect is weaker in countries where agriculture is more important. Accordingly, it appears that natural disasters have a significant positive effect on contemporaneous private per capita credit. This positive impact is mitigated by higher per capita income and further dampened by higher agriculture dependency in the economy. In quantifying results, it is apparent that the impact of natural disasters on credit is country specific as well as time specific. Therefore, in addition to the numerical quantifications, findings are presented by way of graphs and also as a world map. Our findings are robust to various checks.

Keywords:

Natural disasters, economic impact, financial development, private credit

JEL Classification:

Q54, O11, G00, G20, G21

1.

Introduction

Natural disasters are inherently destructive, disruptive and costly, especially when welfare is concerned. Nevertheless, they can act as a catalyst to the economy through potential impact prevention and mitigation activity in pre-disaster era and reconstruction activity in the postdisaster era. In line with the definitions given by the United Nations International Strategy for the Disaster Risk Reduction (UNISDR) for the terms ‘disaster’ and ‘natural hazard’, a natural disaster can be defined as a natural process or phenomenon which causes a serious disruption to the functioning of a community or society through loss of life, injury or other health impacts, property damage, loss of livelihoods and services, social and economic disruption, or environmental degradation (UNISDR, 2009). In the midst of human and physical capital destructions natural disasters affect social, mental and spiritual well-being of human beings. Climatic, hydrological, geophysical, meteorological, oceanic or biological sources acting individually or in combination can give rise to a natural disaster. The impact of natural disasters may vary according to the disaster magnitude, intensity, frequency, extent of the exposure and duration. For instance, while earthquakes are generally of a shorter life span restricted to a small locality, droughts can be prolonged and they can affect much larger regions compared to earthquakes. Further, the magnitude of disaster impact is also dependent on the conditions of vulnerability, coping and mitigating capacities, and levels of adaptation prevailing in the affected territories. In 1991, the deadly tropical cyclone occurred in Bangladesh with a wind speed of around 250km per hour cost more than 138,000 lives. However, in Los Angeles, United States in 1992, the aftermath of a hurricane approximately of the same calibre, amazingly limited to a single digit death toll. The debate in the academic and empirical literature with respect to the impact of natural disasters on economic growth is yet, inconclusive. The UNISDR identifies natural disasters as a major impediment for global development efforts (UNISDR, 2002) and the resolution dated 18 February 2009 adopted by the United Nations General Assembly stresses the fact that the impacts of natural disasters heavily hinder the achievement of internationally agreed development targets of the world including the millennium development goals. Although, natural disasters are considered as negative barriers for growth in general (Raddatz, 2007), some literature suggest positive correlation between natural disaster frequencies and economic growth (Albala-Bertrand, 1993; Dacy & Kunreuther, 1969; Skidmore & Toya, 2002).

1

Tol and Leek (1999) discuss the economic analysis of natural disasters in depth. They admit that natural disasters are costly events which can have negative impact on the economy, initially. However, they are of the view that in the medium to long run a disaster can act as a Keynesian impulse which stimulates economic activity. Given the economic situation, the long term effects of a disaster can be positive, negative or negligible. Moreover, where there is effective and efficient disaster management in place with the anticipation of disasters and as a reaction to the same, the long term impact of natural disasters may not be negative indeed. In addition, reconstruction activity may derive long living positive effects on the economy through multiplier effects, rather than inducing the economy momentarily. On the other hand, it should be noted that the ulterior motive of disaster management is defensive as opposed to the norm of economic optimisation objective. Hence, foregone other economic investments and consumption as a result of disaster management investments may contribute to reduce contemporary and future welfare of agents as illustrated by Tol and Leek in their mathematical demonstration. There are number of studies which explore the effect of natural disasters on economic growth. However, as Tol (2014) points out, although the disruption of economic activity and regaining activities in the recovery phase resultant from a natural disaster are accounted for in calculating gross domestic product (GDP), the destruction of physical and human capital are not readily reflected in economic growth measures such as GDP. In this context, growth may not be a superior indicator to measure the impact of natural disasters on the economy. The level of financial development plays a key role in determining the level of economic growth. In an economy where there is high level of credit availability, there is an enabling environment for investments which is an essential instrument for higher growth. Specifically, in the recuperation subsequent to a catastrophic disaster which causes a huge devastation through destruction of physical capital, mainly infrastructure, it is necessary that the agents should have quick and unconstrained access to finances for immediate and smooth recovery. Delays in the availability of funds for reconstruction activities naturally cause delays in regaining the growth momentum the economy was enjoying just prior to the disaster. If the recovery investments bring in better and advanced technology, it not only ensures the speedy recovery but also paves the way for a higher economic growth. As mentioned afore, finances are needed in the recovery phase of a severe natural calamity for reconstruction, restoration and rehabilitation. Proceeds of realised insurance claims, own 2

savings, aid and grants from the government and third parties, third party investments and indebtedness are the means to meet this financial need. It is observed that there is a higher propensity to save in disaster vulnerable countries like Japan (Skidmore, 2001). As Tol and Leek (1999) point out required finances can be acquired through assistance (credit or aid), savings or insurance. They show that in the aftermath of a natural disaster savings are redistributed to reconstruction activities from their original purposes. Such redistributions may impact financial markets as the demand of affected parties for finances might fall due to selffinancing. Further, insurance companies experience downturn in profits if the magnitude of the honoured claims is overwhelming. Furthermore, subject to the severity and coverage of the disaster damage, a possible crash in the entire financial system also cannot be ruled out. Nevertheless, given that insurance penetration and savings rates are very low in lower income countries, their main source for finance is credit. In reducing economic damages caused by disasters, not only income rather a strong financial sector is also important (Toya & Skidmore, 2007). If finances are readily available, it facilitates the speedy recovery which in turn enhances the development and regaining of the pre-disaster economic growth. Countries with higher levels of domestic credit better able to withstand and endure natural disasters without affecting their economic output much (Noy, 2009). McDermott, Barry, and Tol (2014), in their recent paper find that natural disasters have a significant negative contemporaneous impact on economic growth which is mitigated by higher credit. While claiming that the available literature is mostly empirical and theoretically unfounded on mechanisms and channels through which effects of natural disasters are transmitted to the economic growth, McDermott et al present a convincing theoretical model supported by empirical evidence which demonstrates how financial development proxied by the availability of credit determines the impact of natural disasters on medium-term growth dynamics. Using a panel data set covering the period from 1979 to 2007, they explore the medium to long term impact of natural disasters on economic growth at country-year level. Examining an economy with unconstrained access to capital markets as opposed to a credit-constrained one, they find persistent negative growth effects of natural disasters over the medium term in countries with lower level of financial development. On the contrary, in the rich world where there is high levels of financial development, although there could be a temporary economic slowdown owing to the occurrence of a major natural disaster, in the long run it eventually catches up the pre-disaster growth path, in terms of their theoretical analysis. 3

Even though, many researchers agree on the positive correlation observed between natural disasters and economic growth as stated before, it is apparent from the study of McDermott et al that in the aftermath of a natural disaster the economic growth beyond short run clearly dependent upon the level of financial development prevailing in the affected territory. This raises the question whether natural disasters also affect financial development in an economy. However, the said study does not discuss the direct impact of natural disasters, if any, in turn on financial development. Belke (2013) explores the consequences of 2011 earthquake in Japan which led to a major tsunami and then to a nuclear accident in Fukushima on the world financial markets and the development of the national debt, however, this is only a micro study which discusses the ongoing impact of an individual natural disaster at that time. In his recent paper Klomp (2014) highlights that natural disasters increase the likelihood of banks’ default and in his opinion disaster impact depends on financial regulation and development. Apart from this piece of work which focuses on bank Z-scores and not on financial development per se, we do not find any other study in the existing literature which explores the impact of natural disasters on financial development. As such, the aforesaid findings of McDermott et al (2014) motivate the instant research to ascertain the direct impact of natural disasters on financial development of an economy and our work is closely based on the analysis of McDermott et al (2014). Accordingly, we explore whether there is any impact of natural disasters on financial development proxied by credit, if so in which direction and in what magnitude and how it depends on other economic factors. At a broader level, financial development can be defined as the improvement in the quality of five key financial functions: (1) producing and processing information about possible investments and allocating capital based on these assessments; (2) monitoring individuals and firms and exerting corporate governance after allocating capital; (3) facilitating the trading, diversification, and management of risk; (4) mobilizing and pooling savings; and (5) easing the exchange of goods, services, and financial instruments (Čihák, Demirgüč-Kunt, Feyen, & Levine, 2013, p. 9). Financial system is constituted of financial institutions, financial markets and payment systems. As Levine et al highlight financial development is indicated through high level of access to, depth, efficiency and stability of financial institutions and markets. In this context, they present a 4x2 matrix of financial system characteristics where they compile panel data of financial measures categorized under each group of the matrix which can be used as proxies 4

for financial development. Of all the measures, credit availability to the real sector by domestic banks as a percentage of GDP appears to be the highly favoured candidate in the literature. Levine et al (2013) also agree that private credit is the variable that has received much attention in the empirical literature with respect to financial development. The vital role plays by credit to the private sector by domestic banks in economic and financial development and its wide data coverage across time and space may be the mere reasons for the dominance of the said variable to represent financial development in cross country studies over the years. In the current study too, we use the availability of credit to the private sector by domestic banks as a percentage of GDP to construct the dependent variable i.e. private credit per capita (hereinafter referred to as private credit or credit) to proxy financial development. However, it should be noted at the outset that although this variable can be considered as the best available option, it reflects only one aspect of financial development, i.e., the depth of financial institutions.

5

2. 2.1

Empirical Analysis Data

The source of natural disaster data for this study is the EM-DAT, the International Disaster Database maintained by the Centre for Research on the Epidemiology of Disasters (CRED) at the Université Catholique de Louvain in Brussels, Belgium (Guha-Sapir et al., 2014). The EMDAT database contains inter-alia data on world-wide natural disasters occurred since 1900. Over 13,000 natural disaster events occurred in about 220 countries from 1900 to 2014 are reported in the database. As per the database, from 1979 to 2011, the period on which the instant study is focused for the reasons of data quality and availability, over 10,000 natural disaster events have occurred in 219 countries affecting more than six billion people. The EM-DAT classifies natural disasters into sub-groups, namely, biological, climatic, hydrological, geophysical, meteorological and extraterrestrial disasters. Each natural disaster sub-group contains data on relevant types and sub-types of natural disaster events. In terms of the definitions given by EM-DAT, a biological disaster is a hazard caused by the exposure to living organisms and their toxic substances or vector-borne diseases that they may carry. Epidemics of bacterial, parasitic and viral diseases, insect infestations and animal accidents are categorized under this sub-group. A hazard caused by long-lived meso to macro scale atmospheric processes ranging from intra-seasonal to multi-decadal climate variables is defined as a climatic disaster. Accordingly, droughts, extreme temperature events, glacial lake outbursts and wild fires are regarded as climatic disasters. A geophysical disaster is defined to be a hazard originating from solid earth. Earthquakes (seismic activity), dry mass movement (avalanche, landslide, rock-fall and subsidence) and volcanic activity (ash fall and lava flow) are the types of natural disasters under this sub-group. A hydrological disaster is a hazard caused by occurrence, movement, and distribution of surface and subsurface freshwater and saltwater, namely, floods, landslides (avalanche, landslide, rock-fall and subsidence) and wave action. A hazard caused by short-lived, micro to meso scale extreme weather and atmospheric conditions that last from minutes to days is defined to be a meteorological disaster. This subgroup consists of extreme temperature events, fogs and storms. A hazard caused due to space incidents which originate beyond the earth such as asteroids, meteoroids and comets is said to be an extraterrestrial disaster.

6

For a natural disaster to enter into the EM-DAT database, at least one of the setout criteria needs to be fulfilled, i.e., reported death toll of 10 or more, 100 people reported affected, a call for international assistance or the declaration of state of emergency. Although many disaster studies across countries and time have used publicly accessible EM-DAT data, it is to be noted in limine that the accuracy of these data has been questioned due to the said arbitrary thresholds for reporting disasters and the humanitarian focus of EM-DAT and its data sources, as highlighted by Miao and Popp (2014). However, it is clear that the measurement error is a common issue with natural disaster data given that there is a tendency for national governments to exaggerate the disaster damage in reporting as a strategy for attracting external aid, especially in developing countries (Noy, 2009). Ferreira and Ghimire (2012) refrain from using EM-DAT data in their study on flood frequency in developing countries claiming that a flood is reported only if it is observed meaning that it should occur in a highly populated area. It is also to be noted that in the EM-DAT data set, there are instances where there are reported deaths due to a particular natural disaster event, nevertheless the number of affected is zero, which seems to be rather unrealistic from a pragmatic point of view. Still, EM-DAT is the source of data that has been used widely in disaster literature. Having considered strengths and weaknesses of the database, we are of the view that EM-DAT data is trustworthy and comprehensive enough for our purposes. The EM-DAT database contains disaster outcomes measured as the number of total deaths, number of people affected (injured, became homeless, displaced or affected otherwise) and the total monetary damage caused by a disaster. As McDermott et al (2014) show, the death toll and the physical damage of a disaster primarily depend on the level of preparedness and coping capacity of the affected economy. As such, in rich countries where it can be assumed to be high levels of development and better institutions, it can be expected the resultant number of deaths and physical damages of a natural disaster to be less compared to more vulnerable poor countries. On the other hand, the economic data of a disaster may be gathered by the individuals and institutions who attend the affected area in the aftermath of a disaster primarily with the intention of providing medical care and physical aid to the affected. Therefore, they may be lack of necessary expertise to carry out an accurate estimate of the economic loss caused by the disaster ensuing more noise and inconsistency in the reported damage. In this context, total damage of natural disasters reported in the database may not be a proper and reliable indicator to represent disaster outcome. Of the numbers of people killed and affected, the preferred variable appears to be the number of people affected. Unless the disaster is a sudden and 7

extremely catastrophic severe event which does not allow anybody to escape alive, the number of people affected far exceeds the number killed and in some instances even for a severe disaster the death toll may be nil. In contrast, when it comes to the number of people affected, all the countries suffer to some extent, at least on immediate or temporary basis irrespective of the level of development and institutions they enjoy. Further, as Gassebner, Keck, and Teh (2010), Cavallo and Noy (2011) and Klomp (2014) show disaster casualties and damages reported by EM-DAT seem to be surprisingly low. Hence, in this study, the number of people affected by natural disasters in a country year is chosen as the variable of interest: the natural disaster explanatory variable. McDermott et al (2014) also focus on this disaster measure owing to similar reasoning. Given that the number of people affected depends on the population and population density of the affected country, following Noy (2009) and McDermott et al (2014), the disaster variable is normalized as the “percentage of population affected” by dividing the number of people affected in the current period by the recorded population in the previous period, allowing a fair comparison across countries. By taking the lagged population into account instead of the contemporaneous one, it is expected to avoid the denominator being influenced by the numerator.

Table 1: Severity of disasters by % of population affected Disaster Type

Observations

Mean

Std. Dev.

Min

Max

All Disaster Events Biological Climatic Hydrological Geophysical Meteorological

2,712 710 522 1,718 528 861

3.64 0.23 8.98 1.21 0.67 3.01

11.27 1.36 18.11 3.62 3.37 11.59

1.65e-06 6.85e-07 2.13e-06 1.65e-06 5.03e-07 1.39e-06

156.78 25.16 118.47 45.24 48.51 156.78

During the period under concern, there are 2,712 country-year observations with at least one disaster event that qualifies to enter into the data set by affecting 100 people or more. On average, disasters affect 3.6% of the population in a country-year and the maximum percentage of population affected by natural disasters in a single country-year surpasses 150%. As apparent from Table 1, hydrological disasters appear to be the mostly common natural disasters with highest frequency of country-years of 1,718. However, when the disaster severity is considered, climatic disasters affect the highest percentage of population, i.e., on average 9.0% 8

of population and meteorological disasters are the next in line with the percentage of population affected being 3.0%. Of all the sub-groups of natural disasters, biological disasters affect the least percentage of population of 0.2%, on average in a country-year. As enunciated already, the number of people affected by a disaster is dependent on the nature of the disaster as well as on the underlying socio-economic status and disaster management strategies of the affected economy leading to potential endogeneity issues in the models attempt to quantify the economic impact of natural disasters (Kellenberg & Mobarak, 2008; McDermott et al., 2014; Sen, 1983; Tol & Leek, 1999). In order to reduce the endogeneity problem, while Noy (2009) and Klomp (2014) develop a count disaster measure, McDermott et al (2014) construct a binary disaster variable imposing a threshold of 0.5 percent on the fraction of population affected to capture only the relatively severe disasters in the model. As a robustness check, McDermott et al (2014) carry out their analysis using a binary disaster variable constructed without imposing any such threshold. They admit the fact that the binary variable reduces the variation of data and the explanatory power of the model. In spite of this they opt for a binary disaster variable as it reduces not only the influence of measurement error in disaster data on the analysis but also the possibility of results are being driven by outliers at the upper bound of the disaster data distribution. However, by doing this they equalize minor disaster events which affect a very few individuals with severe disaster events which affect hundreds of thousands of people. Further, it can be argued that the imposition of an arbitrary threshold to segregate large disasters would cause biases in the estimates. Yet, it is not less common in disaster studies to adopt such decision rules to isolate severe disasters to include in the model. For instance, Becerra, Cavallo, and Noy (2014) and Klomp (2014) deploy such decision rules to limit their investigation to major disasters. Exploring disaster effects on bank solvency, Klomp (2014) limits his sample to 170 severe disasters which caused highest economic damage and the time period from 1997 to 2010 in quantifying the impact of natural disasters on bank Z score which reflects banks’ distance to default. Since there is a clear trade-off in using a binary disaster variable with or without a decision rule, the current analysis employs a continuous disaster variable, namely the percentage of population affected by natural disasters in a country year. Nevertheless, as a supportive identification strategy and a robustness check, the baseline model is run using a binary disaster variable with various thresholds to segregate severe disasters in constructing the disaster dummy, as morefully described later on, to see whether it derives consistent results.

9

The instant study explores the impact of natural disasters on financial development. As already explained, widely used private credit measure is chosen to proxy financial development given its broad data coverage in space and time, although its representation is limited only to the depth of financial institutions. The main credit data, namely, private credit by deposit money banks as a percentage of gross domestic product (GDP) and other measures of financial development are obtained from the Global Financial Development Database, an open data source of the World Bank constructed by Levine et al (2013) covering 205 economies from 1960 to 2011. They have constructed the said credit variable using International Financial Statistics (IFS) published by the International Monetary Fund (IMF) and the same is defined as the domestic private credit to the real sector by deposit money banks as percentage of local currency GDP. Accordingly, private credit does not include credit issued to governments, government agencies and public enterprises, and credit issued by central banks. This credit measure is used to construct the dependent variable of our model i.e., the private credit per capita to avoid any potential endogeneity of using credit as a percentage of GDP for the reason that logged per capita GDP is included in the regression model as a key explanatory variable. As such, the use of per capita credit circumvents the issue of GDP being present in both sides of the equation. Further, when private credit is considered as a percentage of GDP, the variation in the data may not be due to the variation in credit per se, but may be due to the dominant denominator effect of GDP. Therefore, the use of per capita private credit measure likewise resolves the latter issue. The measure of private credit as a percentage of GDP is converted to constant 2005 US dollar per capita credit using constant 2005 US dollar GDP data as this measure accounts for dollar inflation over time. However, given that a dollar of credit in a poor economy does not have the same value in a rich economy, the analysis is repeated using purchasing power parity (PPP) constant 2005 US dollar per capita credit as the latter unit of measurement accounts for price differences across countries. The level of credit clearly depends on the level of income as income determines one’s credit necessity and also credit worthiness through repayment capacity and availability of collateral. Accordingly, the natural log of the output based real GDP per capita of the current year enters the regression along with its interaction with the current disaster variable as regressors. Constant 2005 US dollar per capita GDP and PPP constant 2005 US dollar per capita GDP are calculated using relevant data contained in the World Bank’s World Development Indicators and the Penn World Tables (PWT) Version 8.0 database, respectively.

10

Political institutions play a vital role in developing disaster mitigation and adaptation capacities in an economy. As Tol and Leek (1999) point out, defensive investment needs sacrifice in opportunities because protection does not necessarily mean direct economic development. Nevertheless, if the political regime has the will and ability to duly identify potential impediments to the economic development and welfare of the society, and take corrective measures promptly, it is not difficult for any country to achieve prosperity. Over the past 150 years the Netherlands could easily surpass Bangladesh to achieve a remarkable economic development as a result of good governance, desire and focus of political institutions to successfully manage continuous severe floods which was a devastating common problem to both the countries due to their lower altitude in location. So, it is apparent that political institutions substantially contribute in deciding the effects of natural disasters on any economic indicator. In this context, political institutional variable enters the regression. In the opinion of Plumper and Neumayer (2010), polity2 variable from the Polity IV Project is the most appropriate and popular measure of a country’s political regime. In this light, political institutional data in the current analysis are taken from the Polity IV Project and polity2 which indicates openness of a country’s political institutions is used as the political institutional variable. In Polity IV database, the democracy indicator (democ) which varies in an additive eleven-point scale (0-10) represents the institutionalized democracy of a state. It is dependent on 3 elements which cover the democratic rights of citizens and the necessary constraints on the executive in exercising its powers. Similarly, the institutionalized autocracy indicator (autoc) is also an additive eleven-point scale (0-10) which measures authoritarian regime of a country. These two scales democ and autoc do not share any contributor categories in common. Value of polity2 is obtained by subtracting the institutionalized autocracy (autoc) from the democracy variable (democ) and it ranges between +10 (strongly democratic) and -10 (strongly autocratic). If a country’s economy is heavily relied upon agriculture, especially weather dependent traditional agriculture, in addition to the agriculture’s direct contribution to determine per capita income and credit, it can be expected severe natural disasters to cause indirect impact on such agricultural economies through credit. In the normal course, cultivations and livestock are more prone to catastrophic destruction. Natural calamities heavily interrupt livelihood of farmers and fishermen creating a high demand for finances, however, simultaneously deteriorating their credit worthiness and repayment capacities as most of the time their wealth accumulation is also in the form of cultivation and livestock. As such, agriculture share of the 11

economy as a percentage of GDP together with its interaction with disaster variable is included in the benchmark specification. Data on share of agriculture as a percentage of GDP and other control economic variables such as inflation, government consumption as a percentage of GDP, share of trade as a percentage of GDP, net official development assistance (ODA) received as a percentage of gross national income (GNI), financial sector rating, lending interest rate, private savings rates and insurance penetration are taken from the World Bank’s World Development Indicators. Data on resources of countries are obtained from the Wealth of Nations data series maintained by the World Bank. Sample for the baseline model consists of 147 countries during 1979 to 2011. With the inclusion of more control variables sample size decreases due to non-availability of data. Post estimation summary statistics for the variables used in the baseline analysis are provided in Table 2.

Table 2: Summary statistics Variable

Mean

Std. Dev.

Min

Max

Disaster (% of Population Affected) Credit per capita (constant 2005 US$) GDP per Capita (constant 2005 US$) Polity2 Share of Agriculture (as % of GDP)

1.69 6,547 8,100 3.01 17.46

6.89 15,583 12,980 6.78 14.76

0 0.85 112 -10 0

118.47 163,982 87,717 10 73.48

12

2.2

Empirical Model

We employ a panel regression estimator with country and year fixed effects as the main estimation strategy in our analysis. Fixed effects estimator is chosen since country and year fixed effects control for time-invariant country heterogeneity and time-variant shocks that simultaneously affect all the countries, respectively. As such, this approach reduces any potential endogeneity issue. Further, results of the Hausman tests carried out show that fixed effects estimator is preferred to a random effects estimator. Furthermore, as country-fixed effects capture variation of country specific effects within a country which do not change over time, fixed effects also arrest any selection biases which may arise due to over representation of poor countries in the disaster data distribution as a result of their higher vulnerability to disasters (McDermott et al, 2014). Year fixed effects capture the effects of time- varying factors common to all countries such as world business cycle, global technological advancement and world-wide economic and financial crises. In terms of the results of ‘Testparm’ test, time-fixed effects are needed to be included in the regression. Errors are clustered at country-level as natural disasters are not evenly distributed across countries and also to obtain robust standard errors as a remedial measure for heteroscedasticity. Given the constraints on availability and reliability of data, the analysis is restricted to the time period 1979 – 2011. The baseline model covers 147countries. The panel regression equation of the baseline model is as follows; 𝐶𝑟𝑒𝑑𝑖𝑡𝑖𝑡 = 𝛽0 + 𝛽1 𝐷𝑖𝑠𝑖𝑡 + 𝛽2 𝐶𝑟𝑒𝑑𝑖𝑡𝑖,𝑡−1 + 𝛽3 𝑙𝑛𝐺𝐷𝑃𝑖,𝑡 + 𝛽4 𝐷𝑖𝑠𝑖𝑡 ∗ 𝑙𝑛𝐺𝐷𝑃𝑖,𝑡 + 𝛽5 𝐴𝑔𝑟𝑠ℎ𝑟𝑖𝑡 + 𝛽6 𝐷𝑖𝑠𝑖𝑡 ∗ 𝐴𝑔𝑟𝑖,𝑡 + 𝛽7 𝑃𝑜𝑙𝑖𝑡𝑦𝑖𝑡 + 𝜃𝑖 + 𝜃𝑡 + 𝜖𝑖𝑡

(1)

Where, credit per capita valued at constant 2005 US$ in country i for year t is the dependent variable. A lagged credit term is included as an explanatory variable because it can be reasonable to assume that the current credit level is heavily determined by its past level and to defend the existence of autocorrelation in the regression. However, as by construction lagged dependent variable and error term are correlated, one may argue that the use of a lagged dependent variable in the fixed effects estimator poses a serious econometric problem. Such use can cause negative biases on estimates for positive coefficients in short panels with small time periods. To overcome this issue the best remedy would be the use of a valid instrument variable, however, it is very hard to find such an instrument. As McDermott et al (2014) show this is a serious concern only in the event the panel is short. They claim that the issue is being

13

addressed by using a long panel of 29 years and they support their findings with consistent results obtained in dynamic panel estimators. Ours is an even lengthier panel of 33 years. We also get consistent results using System GMM. It may be worthwhile mentioning that we observe consistent results even when the specification is modelled without including the lagged dependent variable but including only disasters, logged GDP per capita and disaster-income interaction with and without further control variables as specified under the robustness checks. Dis is our variable of interest, disaster measured as the percentage of population affected due to all the natural disasters occurred in a single country year. As the percentage of population affected increases, it can be expected the private credit to rise as a result of higher demand for financing aimed at recovery, reconstruction and rehabilitation in the aftermath of a natural disaster. As private credit availability is an indicator of financial development, if we find a positive coefficient on the disaster variable we would be able to establish positive effects of disasters on financial development. GDP is the logged GDP per capita in constant 2005 US$ included in the model as the level of credit is clearly dependent on income level and as there appears to be a very high correlation between income and credit. Besides, demand for and the availability of credit are importantly different in poor and rich economies and the main difference between the two is the higher per capita income enjoy by the latter. In poor countries, dependency on private credit for financing appears to be much higher in the recovery phase subsequent to a natural disaster given that their private savings rate and insurance penetration are substantially lower than rich countries on average. Disaster variable is interacted with per capita GDP and it can be expected this interaction term to derive a negative coefficient as it is sensible to presume higher income to crowd out the necessity for external financing for recovery may be through self-savings and insurance. If that is the case, then the interaction term will dampen the aforesaid any positive effect of disasters on financial development. Share of agriculture within GDP and its interaction with disaster are included in the bench mark specification to capture the effect of disaster on credit in the presence of economy’s dependency on agriculture. As a country’s preparedness and management strategies for natural disasters depend on the political will and institutions of that country, polity2 which reflects the political institutional regime in place enters into the baseline model as a control variable. Terms

14

𝜃𝑖 and 𝜃𝑡 are the country and year fixed effects included in the model, respectively. The final term 𝜖𝑖𝑡 in the equation is the independently and identically distributed error term. When using a longer panel in regression, one has to be extremely careful because nonstationary series of data might give rise to spurious results as suggested by Nelson and Plosser (1982). As one can suspect of the non-stationarity or the presence of unit root in credit data, we model our regression taking the change in private credit (first difference) as the dependent variable. Further, we also model a first difference linear estimator with and without additional control variables to overcome any possible non-stationarity issue in the data, to show that it yields consistent results. Furthermore, we run fixed effects estimator using five-year and tenyear averages of data in addition to the use of annual data in our primary regressions to avoid any potential stationarity issues and we repeat this exercise using PPP data as an additional measure. To show our original results are not driven by outliers at the lower or upper bounds of the credit and disaster data distribution, we repeat our regressions removing alternatively and jointly substantial amount of observations at the lower and upper bounds in the credit distribution and upper bound of the disaster distribution. A major concern one can raise regarding the variable of interest, disaster is that whether we identify the true impact of disasters on credit originated through exogenous variation in disasters or whether there is a possibility for disasters and credit to jointly determine each other. We assume exogeneity of disaster measure as the probabilities are very remote for contemporaneous credit to influence disaster affected percentage of population in the same year for the reason that it takes long for credit to be converted into effective and defensive disaster impact preventive or mitigating projects. So in our opinion, at least there is no instantaneous impact of credit on the percentage of population affected by disasters. Especially, given that what is under consideration is the formal credit disbursed by domestic banks to the private sector, we can reasonably rule out any possibility of mutual determination between contemporaneous credit and disasters. Disaster exogeneity assumption is adopted by other disaster literature such as Noy (2009), Raddatz (2007), Ramcharan (2007) and Skidmore and Toya (2002) in different contexts. If the exogeneity assumption does not hold, then the best solution again to avoid potential reverse causality would be to employ a valid instrument. However, it is extremely difficult to find such an instrument for disasters as good as randomly assigned and which satisfies both 15

instrument relevance and exclusion restriction. Noy (2009) is apologetic of being unable to find such an instrument which may not readily exist indeed, however, indicates the ability of constructing a unique index of disaster intensity, if one wishes to do so, that depends purely on physical features of a disaster, which may entail enormous ground work of obtaining first-hand information on characteristics of each disaster. To infer causality as there need to be a proper identification strategy, disaster related studies use binary disaster variable with and without decision rules to segregate severe disasters although it reduces the variation in disaster variable as all disasters are given an equivalent weight without considering the magnitude or severity of the disaster outcome thereby resulting a decline in the explanatory power of the model. Following McDermott et al (2014) we also construct a binary disaster variable, however in our case at various different thresholds and use our baseline specification with binary disaster variable as an identification strategy to show that it yields consistent results. However, we prefer to stick to our original baseline model with continuous disaster variable as we can be more precise in quantifying disaster effect on private credit. Robustness of results in the baseline model could be checked by adding more control variables that represent macro stability, magnitude of the government spending, foreign links, etc. which can be expected to have any influence on per capita credit, the dependent variable and independent variables in the model. By doing this we can overcome any omitted variable biases the baseline model suffers from, which are not taken care of by already included controls and, country and year fixed effects. The inclusion of additional control variables is done at different stages. Firstly, we add main control variables one by one to the baseline model and subsequent to each addition, an interaction term of that control with the disaster variable is included so that their impact on the baseline model can be observed clearly. These main control variables are inflation which control for macroeconomic stability of the country, government expenditure as a percentage of GDP and the trade share which reflects the degree of trade openness. Secondly, we control for other factors which seem to either stimulate or hinder private credit in connection with disasters, by using simple variant models of the baseline specification. Accordingly, we control for financial sector regulation using CPIA (Country Policy and Institutional Assessment) financial sector rating, non-life insurance premia volume as a share of GDP, lending interest rate, share of resource rent (including rent received on coal, oil gas, iron ore and minerals such as gold, silver, copper, etc. but not including rent on forestry) within the

16

GDP, and share of forestry rent as a percentage of GDP and net official assistance received as a percentage of gross national income. Apart from the panel fixed effect regression, estimators of different types are employed to support the validity of results, namely, ordinary least squares (OLS) and system generalized method of moments (GMM), a dynamic panel estimator which uses GMM procedure. The difference GMM estimator which uses the lagged level variables as GMM-style instruments in a more desirable way was proposed by Manuel Arellano and Bond (1991). A main advantage of this method is that it does not reduce the panel length as this instrumentation is done on period by period basis. By enhancing and refining this model further, M. Arellano and Bover (1995) and Blundell and Bond (1998) introduced system GMM, which we also employ as an alternative estimation in our study. As discussed in Roodman (2006 as revised in 2007) in depth, these estimators are designed to use with panel data when the explanatory variables are not strictly exogenous (if it is suspicious that they are correlated with past and possibly current realizations of the error) and suitable instrument variables are not available (McDermott et al, 2014, p11). Of the available popular dynamic panel estimators, system GMM is preferred to difference GMM considering the superiority of the former over the latter as it successfully addresses the issues pertaining to the latter by using a two-equation set of simultaneous equations. In addition to the OLS and system GMM, we run a quantile regression as there is a substantial variation in our credit data with a high range. As our per capita credit and GDP are measured in constant 2005 US$ we account for dollar inflation over time. However, one can argue that the findings are misleading as we do not control for price differences across space. Hence, we repeat the baseline analysis with alternative estimators using per capita credit and GDP measured in purchasing power parity (PPP) constant 2005 US dollars to check whether this yields consistent results. We explore the impact of different categories of natural disasters on private credit by modelling each subgroup of natural disasters, viz., biological, climatic, hydrological, geophysical and meteorological disasters and we also run our regression for different geographical regions, individually. Finally, we attempt to ascertain the impact of natural disasters on other financial measures which are proxies for financial development and represent financial access, efficiency and stability as credit only reflects financial depth.

17

3.

Results

Results of our baseline model are given in Table 3 and we restrict our concern to the marginal effect of natural disasters on private credit. As per the results, disasters show a significant 2 positive effect on contemporaneous credit, however, this positive effect is dampened down by higher income. It appears that disaster-agriculture interaction also yields a negative coefficient suggesting that the positive impact of disasters on credit is further mitigated by higher share of agriculture in the economy. However, as this interaction is less significant3, we ignore it for the time being.

Table 3: Baseline model Dependent variable: Credit per capita Fixed Effects Disaster (% Population Affected) Lagged Credit per capita GDP per capita (in logs) Disaster * GDP per capita Share of Agriculture Disaster * Agriculture Polity2 Observations Number of Countries R-squared

35.35** (14.08) 1.00*** (0.0172) 654.0*** (176.3) -4.669** (1.807) 15.64** (7.006) -0.135* (0.0763) -1.064 (5.273) 3,189 147 0.958

Notes: Annual data 1979-2011, except where lost due to lags. All models include a constant term, country and year fixed effects. Errors clustered at the country level. Robust standard errors in parentheses. *** p