WPS4968 Policy Research Working Paper 4968 Assessing the Macroeconomic Impacts of Natural Disasters Are there Any? Stefan Hochrainer The World Bank Sustainable Development Network Vice Presidency Global Facility for Disaster Reduction and Recovery Unit June 2009 Policy Research Working Paper 4968 Abstract There is an ongoing debate on whether disasters cause event), comparing counterfactual with observed gross significant macroeconomic impacts and are truly a domestic product, the authors find that natural disasters potential impediment to economic development. This on average can lead to negative consequences. Although paper aims to assess whether and by what mechanisms the negative effects may be small, they can become disasters have the potential to cause significant GDP more pronounced depending mainly on the size of the impacts. The analysis first studies the counterfactual shock. Furthermore, the authors test a large number of versus the observed gross domestic product. Second, the vulnerability predictors and find that greater aid and analysis assesses disaster impacts as a function of hazard, inflows of remittances reduce adverse macroeconomic exposure of assets, and, importantly, vulnerability. In a consequences, and that direct losses appear most critical. medium-term analysis (up to 5 years after the disaster This paper--a product of the Global Facility for Disaster Reduction and Recovery Unit, Sustainable Development Network Vice Presidency--is part of a larger effort in the Network to disseminate the emerging findings of the forthcoming joint World Bank-United Nations' Assessment of the Economics of Disaster Risk Reduction.. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at hochrain@iiasa.ac.at. We are grateful to Apurva Sanghi, Reinhard Mechler and participants of the seminar at the World Bank held on this topic for their suggestions and constructive comments. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team ASSESSING THE MACROECONOMIC IMPACTS OF NATURAL DISASTERS: ARE THERE ANY? Stefan Hochrainer 1 International Institute for Applied Systems Analysis (IIASA) JEL: C22, C53, E01 Keywords: Natural disasters, macroeconomic consequences, time-series analysis, ARIMA process, vulnerability. 1 We gratefully acknowledge support by the UN/World Bank project "Economics of Disaster Risk Reduction." We would like to thank the team leader of this project, Apurva Sanghi, as well as Sebnem Sahin of the World Bank for ongoing support and stimulating discussions, and Jesus Crespo Cuaresma, as well as a number of anonymous referees for very helpful and stimulating comments. 1 INTRODUCTION A small, but growing literature has emerged over the last few years on the macroeconomic and development impacts of natural disasters. Interestingly, there is as yet no agreement on whether disasters are important from a macroeconomic perspective, and two positions can be identified. The first considers natural disasters a setback for economic growth and is well represented by the following citation: It has been argued that although individuals are risk-averse [to natural disasters risk], governments should take a risk-neutral stance. The reality of developing countries suggests otherwise. Government decisions should be based on the opportunity costs to society of the resources invested in the project and on the loss of economic assets, functions and products. In view of the responsibility vested in the public sector for the administration of scarce resources, and considering issues such as fiscal debt, trade balances, income distribution, and a wide range of other economic and social, and political concerns, governments should not act risk- neutral (OAS, 1991). The other position sees disasters as entailing little growth implications and consider disasters and their reduction a problem of, but not for development (e.g. Albala-Bertrand, 1993, 2006; Caselli and Malhotra, 2004). These authors find natural disasters do not negatively affect GDP and "if anything, GDP growth is improved" (Albala-Bertrand, 1993: 207). This paper can be understood as an attempt at reconciling this body of literature. There are two entry points for the analysis. The first is to look at counterfactual vs. observed GDP, the second entry point is to assess disaster impacts as a function of hazard, exposure of assets (human, produced, intangible), and, importantly vulnerability. Overall, the evidence reveals adverse macroeconomic consequences of disasters on GDP. In a medium-term analysis, natural disasters on average seem to lead to negative effects on GDP. The negative effects may be small, yet they can become more pronounced depending on the size of the shock. We tested a large number of vulnerability predictors and found that higher aid rates as well as higher remittances lessen the adverse macroeconomic consequences, while capital stock loss is the most important predictor for the negative consequences. 2 The paper is organized as follows. Section 2 reviews the literature on the macroeconomic impacts of disasters and locates the proposed analysis within the disaster risk management paradigm. In section 3, we present the data and methodology used for projecting the economic impacts for a medium term horizon (up to 5 years after an event), as well as the regression analysis used for identifying predictor variables explaining potential impacts. Section 4 ends with a discussion of possible implications of our analysis. 2 LITERATURE REVIEW The literature on the macroeconomic effects of disasters can be divided into studies looking into the short-to-medium term (1-5 years in economic analysis) and the longer term (beyond 5 years), with almost all studies taking a shorter-term perspective. A key response variable analyzed in this line of work is GDP. In principle, after a disaster event the following trajectories may be distinguished (see figure 1) leading to no, positive or negative follow-on effects. Projected line without GDP Positive long term disaster event effect Negative long term effect No long term effect Time Disaster Event Fig. 1: Possible trajectories of GDP after a disaster. Source: Hochrainer, 2006 3 Two positions can be distinguished as shown in table 1. Position 1 broadly suggests the post-disaster trajectory will fall short of the planned trajectory, while position 2 contends that there is no negative effect beyond the first year and the planned GPD path can be achieved or even surpassed. Table 1: Synopsis of macroeconomic perspectives on natural disasters Position 1 Position 2 "Natural disasters are setbacks for "Disasters have no effects on economic economic growth" growth" Methodologies involving Methodologies involving · Supply side focus · Supply side and demand side · Model projections · Empirical evidence · Neoclassical intuition · Empirical evidence Studies by Benson (various); ECLAC Studies by Albala-Bertrand, 1993, 2006; (various); Otero and Marti, 1995; Crowards, Skidmore and Toya, 2002; Caselli and 2000; Charveriat, 2000; Murlidharan and Malhotra, 2004. Shah, 2001; Freeman et al., 2002; Mechler, 2004; Cuaresma, Hlouskova, and Obersteiner, 2004; Hochrainer, 2006; Noy, 2009; Okuyama, 2009 Source: Adapted from Zenklusen, 2007 The body of research subscribing to position 1 generally finds significant short-to- medium-term macroeconomic effects (Otero and Marti, 1995; Benson, 1997a,b,c; Benson, 1998; Benson and Clay, 1998, 2000, 2001; ECLAC 1982, 1985, 1988, 1999, 2002; Murlidharan and Shah, 2001; Crowards, 2000; Charveriat, 2000; Mechler, 2004; Hochrainer, 2006; Noy, 2009) and considers natural disasters a barrier for development in disaster-vulnerable developing countries. ECLAC (various studies) has been conducting numerous case studies on disaster impacts in Latin American countries since 1972. Otero and Marti (1995) summarized the 4 results and generally found serious shorter-term impacts as national income decreases, an increase in the fiscal deficit as tax revenue falls, and an increase in the trade deficit as exports fall and imports increase. Substantial longer term impacts on development prospects, perpetual external and fiscal imbalances due to increased debt service payments post-disaster and spending requirements, and negative effects on income distribution were also found (ECLAC and IDB, 2000; Otero and Marti, 1995). They generally hold that the significance of the impact depends on the size of the disasters, the size of the economy and the prevailing economic conditions (Otero and Marti, 1995). Benson (1997a,b,c) and Benson and Clay (1998, 2000, 2001) produced a number of case studies on Fiji, Vietnam, the Philippines, and Dominica. The timeframe of this analysis was mainly short-term, i.e. the period up to one year after a disaster. They detected severe negative economic impacts, with agriculture being hit most strongly, an exacerbation of inequalities, and reinforcement of poverty, however also finding it difficult to isolate disaster impacts on economic variables from other impacts. Murlidharan and Shah (2001) by means of a regression analysis analyzed a large data set of 52 catastrophes in 32 developed and developing countries with a the short-term focus (year before event compared to year of event). They found catastrophes for all country income groups to affect short-term growth very significantly. In the medium-term (average of two preceding years compared to average of event and two following years), the effect on growth was still significant. Over time, they detected impact on economic growth to subside. They also discovered associations between disasters and the growth of external debt, the budget deficit and inflation. Crowards (1999 discussed in Charveriat, 2000) examined the impacts of 22 hurricane events in borrowing member countries of the Caribbean Development Bank and found that GDP growth slowed by 3% points on average post-event, but rebounded due to the increase in investment the following year. He also detected large variations around averages. 2 Charveriat (2000) for most cases in her disaster sample identified a typical pattern of GDP with a decrease in the year of an event and a recuperation of the growth rate in the following two years due to high investment into fixed capital. She detected the scale of short-term impacts to depend on the loss-to-GDP-ratio and whether the event was localized or country-wide. For high- 2 This study could not be obtained and we rely on Charveriat (2000) as a secondary source. 5 loss-to-GDP ratios and country-wide events she found larger impacts. She found the following crucial variables affecting the scale of aggregate effects: structure of the economy and general conditions prevailing, the size of economy, the degree of diversification and the speed of assistance of the international community. Another study, Rasmussen (2004), is in accordance with above studies and for a cross-country sample identified a median reduction of the growth rate by 2.2% points in the year of the event. Raddatz (2007) generally assessed the role of external shocks (such as commodity price fluctuations, natural catastrophe, and adverse influences from an international economic environment) on output volatility of low-income countries. While he found external shocks to explain a fraction of output variance, their contribution to output fluctuations was dwarfed by more important contributors from internal sources such as level of inflation, a possible overvaluation of the real exchange rate and large public deficits. Noy (2009) took a look at the reduction of GDP growth rates for a large sample of disaster events, for which while using a linear regression modeling approach he concluded that the ability to mobilize resources for reconstruction as well as the financial condition of the country are important predictors of GDP growth effects. As one of the few longer term studies, Cuaresma et al. (2004) concluded that the degree of catastrophic risk has a negative effect on knowledge spillovers between industrialized and developing countries. Further, they suggested that only countries with relatively high levels of development may benefit from capital upgrading through trade after a natural catastrophe. There are only a few studies adopting position 2 and the key papers here are Albala- Bertrand (1993) and to a lesser extent Caselli and Malhotra (2004). In (partial) contrast to the above studies, Albala-Bertrand (1993) came to different conclusions and finds himself partially in opposition to accepted views when analyzing impacts mainly on developing countries. He first statistically analyzed part of the ECLAC data set discussed above and found that natural disasters do not negatively affect GDP, public deficit and inflation in the short to medium term. His findings on the trade deficit are in accordance with ECLAC and other research. These findings he explains with a sharp increase in capital inflows and transfers (private and public donations). He holds that natural disasters do not lower GDP growth rates and "if anything, they might improve them" (1993: 207). Albala-Bertrand also examined longer-term effects for a number of 6 developed and developing countries and found no significant long-term effects in developed countries; he came to the conclusion that in developing countries aggregate effects fade away after two years, but that some negative effects on income distribution and equality persist. Overall, Albala-Bertrand considered disasters "a problem of development, but essentially not a problem for development." (Albala-Bertrand 1993). According to his analysis, while the number of deaths and people affected and the extent of monetary losses are determined by the current state of a country's development, disasters do not normally hinder long-term development, with the sole exception being widespread droughts. 3 Further, Caselli and Malhotra (2004) based their analysis on neoclassical growth theory and analyzed the losses in relation to country growth rates after disaster events using a dataset of 172 countries for events between 1975 and 1996. They concluded that their hypothesis that losses of labor and capital stock have no effect on short-term economic growth could not be rejected. Finally, Skidmore and Toya (2002) discovered a robust positive correlation between the frequency of natural disasters and long-run economic growth after conditioning for other determinants, which they explain by some type of Schumpeterian creative destruction. Overall, while the balance of evidence and studies seems to imply that there are adverse economic disaster effects in terms of the "negative" trajectory stylized above, there are important "outliers" that merit more investigation. Another observation is that the studies generally have a short-term focus, and in their analyses often do not go beyond the year following an event. Finally, analyses generally compare key indicators of interest after the fact to their pre-disaster states, rather than comparing the counterfactual, i.e. the system without a shock, to the observed. The latter point seems important, as important opportunity costs, e.g. in terms of economic growth foregone, are consequently often not accounted for in analyses on the macro effects of disasters. 3 Albala-Bertrand (1993) started fruitful discussions about some assumptions and estimating issues in the literature, and his findings were discussed and replicated by various other authors including Mechler (2004) and Hochrainer (2006). For example, Hochrainer (2006) extended Albala-Bertrand's sample to 85 disaster events in 45 countries and found GDP growth (on average) negatively affected in the disaster year and no significant increases in growth for the subsequent post-disaster years, which implies that, due to a lack of recovery, a net loss of GDP. 7 2.1 Economic effects and vulnerability In order to set the stage for the analyses, we hold it important to locate the discussion within the disaster risk management framework. The standard approach here is to understand natural disaster risk as a function of hazard, exposure and (physical) vulnerability (see figure 2). Hazard analysis entails determining the type of hazards affecting a certain area with specific intensity and recurrence. Assessing exposure involves analyzing the relevant elements (population, assets) exposed to relevant hazards in a given area. Vulnerability is a multidimensional concept encompassing a large number of factors that can be grouped into physical, economic, social and environmental factors as outlined on the figure. We refer mostly to physical vulnerability as the susceptibility to incurring harm of people and engineered structures leading to direct risk in terms of people affected and, important from the perspective taken in this paper, capital stock destroyed. As a consequence of such direct impacts, follow-on effects may materialize leading to indirect potential and actual impacts. Economic vulnerability may refer to the economic or financial capacity to absorb disaster events, e.g. the ability to refinance asset losses and to recover quickly to a previously planned economic growth path. It may relate to private households and businesses as well as governments, the latter often bearing a large share of a country's risk and losses. Based on assessments of disaster risks and its determinants, risk management measures may be systematically planned for risk reduction and risk transfer. 8 Hazard Exposure Physical Vulnerability Direct losses (risk) Produced capital Human capital Risk Management Environmental capital Socio-economic vulnerability Economic Consequences GDP Fig. 2: Conceptual framework used in this study for explaining economic risk due to natural disasters The literature on the economic impacts discussed above can be related to this framework, and table 2 lists the key studies and general factors contributing to a discussion of (macro) economic risk. Determinants of impacts and risk can be distinguished according to (i) the type of natural hazard (hazard variable), (ii) geographical area and spatial scale of impact (exposure), (iii) the overall structure of the economy, (iv) the stage of development of the country, (v) prevailing socio-economic conditions, and (vi) the availability of formal and informal mechanisms to share risks (the latter four variables related to economic vulnerability). 4 4 It should be mentioned that in the studies discussed and our analysis, observed losses are used for examining future economic consequences. However, when it comes to risk management, losses should be based on probabilities and the discussion framed in terms of risk in order the incorporate the full possible range of potential losses (and its probabilities) in the analysis. 9 Table 2: Studies assessing macroeconomic consequences and economic vulnerability to natural hazards. Study Vulnerability variables for predicting Response variables economic impacts and risk Charveriat, 2000 · Size of the economy, degree of · GDP diversification and size of the informal and agricultural sectors. ECLAC and IDB, · Ability to refinance losses and · GDP, fiscal variables 2000; provide relief to the affected Freeman et al. population (financial vulnerability) 2002; · Availability of implicit (aid) and Mechler,2004; explicit (insurance) risk sharing Hochrainer, 2006 arrangements Burton et al.,1993; · Income · Deaths due to natural Kahn, 2005. disasters Benson and Clay, · Structure of the economy · Total GDP annual change 2004 · Size · Agricultural GDP annual · Income level and stage of change development · Non-Agric. GDP annual · Prevailing socioeconomic conditions change Toya and · Educational attainment in population · Disaster-related deaths Skidmore, 2007 aged 15 and over · Damages/GDP · Economic openness (exports+imports)/GDP · Financial sector level of development (M3/GDP) · Government consumption · Additional variables that determine the deaths caused by disasters (population, land area, disaster type). Noy, 2009 · Literacy rate · GDP · Quality of institutions · Per capita income · Openness to trade · Levels of government spending · Foreign exchange reserves · Levels of domestic credit · Openness of capital accounts Raschky, 2008 · Availability of financial risk sharing · GDP institutions Source: extended from Barrito, 2008. 10 All of the indicators used for explaining the response variables mentioned above are valid candidates as proxies for hazard, exposure and vulnerability and most of them will be used in the analysis in the next section. 3 ASSESSING ECONOMIC DISASTER CONSEQUENCES AND RISK In order to identify the macroeconomic effects of disasters, we suggest comparing a counterfactual situation ex-post to the observed state of the system ex-post. This involves assessing the potential trajectory (projected unaffected economy without disaster) versus the observed state of the economy. This contrasts with observing economic performance post-event and actual performance pre-event, as usually done in similar analysis. Our analysis requires projecting economic development into a future without an event. The approach is illustrated via the case of Honduras, which was heavily hit by Hurricane Mitch at the end of 1998. In figure 3 absolute GDP with the event and projected GDP without an event were estimated. The chart exhibits GDP growth to become negative in the year after, then rebound later; yet, overall the net effect would seem to be a loss. 8,000 GDP in Honduras 7,500 Million constant 2000 USD 7,000 6,500 6,000 Projected w/o event-ECLAC Projected w/o event-IIASA 5,500 Observed 5,000 1996 1997 1998 1999 2000 2001 2002 2003 2004 Fig. 3: Observed GDP in Honduras with events vs. projected growth without events. Source: Zapata, 2008; World Bank, 2007; own calculations Note: Zapata (2008) uses a model based projection, IIASA projects growth statistically based on pre-disaster observed GDP. 11 Using this approach for Honduras, a "GDP gap" as a follow-on consequence after the hurricane can be identified. For example, in 2004, about 6 years after the event, this gap can be considered to have, ceteris paribus, amounted to about 6% of potential GDP given extrapolation of pre-disaster GDP with a 4-year average growth rate, and to 8.6% percent based on the ECLAC projection. In the following, similarly we compare GDP effects in terms of counterfactual vs. observed trajectories by projecting absolute GDP into the future under the assumption of a no disaster event scenario and comparing it with observed GDP values. A 5 year time horizon is chosen as it is the minimum data requirement for estimating time series projections into the future and reflects the trade-off between data requirements and number of samples (the larger the sample the lower the time horizon). There are two avenues for deriving the counterfactual: (i) running a (statistical or behavioral) economic model without a disaster event, for which a large number of models calibrated to the respective countries would be necessary; (2) using time series models. We adopt the second option to eliminate as much possible business cycles in the dataset. We use econometric models which seem to be able to handle empirically observed patterns, which is important as a large number of the countries examined are of developing nature and exhibit strong growth volatility. 3.1 Estimation methodology We use autoregressive integrated moving average models, also called ARIMA(p,d,q) (Box and Jenkins, 1976) for forecasting GDP into the future after the disaster event. ARIMA modeling approaches are chosen because they are sufficiently general to handle virtually all empirically observed patterns and often used for GDP forecasting (see for example Abeysinghe and Rajaguru, 2004). While such a type of modeling may be criticized for its black box approach (Makridakis and Wheelwright, 1989), it here serves well due to the large number of projections to be made and the difficulty identifying suitable economic model approaches, such as input-output models for all the different countries within the sample and over a time period starting from 1965. 12 The ARIMA process Recall, an autoregressive process of order AR(p) can be defined as x t = 1 x t -1 + 2 x t - 2 + + p x t - p + t A moving-average process of order MA(q) may be written as xt = t + 1 t -1 + 2 t - 2 + + q t - q and an ARMA(p,q) process, with p autoregressive and q moving average terms can be defined to be xt = 1 xt -1 + + p xt - p + t + 1 t -1 + + q t - q where and are parameters to be estimated and are white noise stochastic error terms. Now, let yt be a non-stationary series and define the first order regular difference of yt as yt = yt - yt -1 or more generally using a back-shift operator denoted as B k zt = zt - k d yt = (1 - B) d yt An ARIMA(p,d,q) model can then be expressed as p ( B)(1 - B) d yt = q ( B) t with p ( B) = 1 - 1B - - p B p and q ( B) = 1 - 1B - - q B q 13 The Box-Jenkins methodology (Box and Jenkins, 1976) is applied for determining the components of the ARIMA process; i.e. we test different ARIMA(p,d,q) models with p and q to be smaller or equal 4 (due to the limited amount of data) and estimate and using Maximum likelihood techniques and the Akaike Information Criterion (AIC) as well as diagnostic checks to detect a suitable model. The data requirements were set thus that at least 5 observed data points are needed for projections into the future. This is the smallest number of observations which are needed to estimate ARIMA(4,1,4) models (however, the majority of the sample (greater 90 percent) has at least 10 data points). Furthermore, all models are tested to be stationary (usually d=1 suffices to assure a stationary process) and all series are demeaned. To include uncertainty in the projections, also 95 percent confidence forecasts were calculated and analyzed. Forecasts into the future are performed with the selected models and then compared to the observed variables. Increases or decreases of GDP in future years are measured as a percentage increase or decrease to baseline GDP (i.e., baseline =100) which is defined 5 to be GDP a year before the disaster event. Furthermore, the differences between observed values and projected ones are calculated and called Diff(t), which indicates the percentage difference between the observed and projected value of GDP in year t. We focus on projections with a medium term perspective (up to 5 years into the future). This limitation is due to important data constraints for the ARIMA models within the sample and increasingly large uncertainties beyond the medium-term time horizon. 3.2 Data used Our sample consists of 225 large natural disaster events during 1960-2005. The sample is based on information from two databases and was compiled by Okuyama (2009) with the threshold for a large event defined arbitrarily to a loss exceeding 1 percent of GDP. 6 One database is the open-source EMDAT disaster database (CRED, 2008) maintained by the Centre for Research on the Epidemiology of Disasters at the Université Catholique de Louvain. EMDAT currently lists information on people killed, made homeless, affected 5 To decrease variance a logarithmic transformation of GDP was performed at the beginning. 6 In order to define the "event set" the threshold of stock losses is set as a share (1%) of flow effects (GDP). While it would have been more systematic to define an asset threshold, yet we responded to the larger intuitive appeal of using GDP as a denominator, and the fact that this threshold was also used by another paper in the EDRR working paper series which we wanted to be in line with. 14 and financial losses for more than 16,000 sudden-onset (such as floods, storms, earthquakes) and slow-onset (drought) events from 1900 to present. Primary data are compiled for various purposes, such as informing relief and reconstruction requirements internationally or nationally, and data are generally collected from various sources and, including UN agencies, non-governmental organizations, insurance companies, research institutes and press agencies. The other database is the proprietary Munich Re NatCat Service database, which mainly serves to inform insurance and reinsurance pricing. This database contains fewer entries focusing on the about 300 largest events since 1950, yet data exhibit a higher reliability as often crosschecked with other information. We focus on the monetary losses (direct impacts or risk) listed in constant 2000 USD terms. In both datasets, loss data follow no uniform definition and are collected for different purposes such as assessing donor needs for relief and reconstruction, assessing potential impacts on economic aggregates and defining insurance losses. We distinguish between sudden and slow onset events. Key sudden-onset events are extreme geotectonic events (earthquakes, volcanic eruptions, slow mass movements) and extreme weather events such as tropical cyclones, floods and winter storms. Slow-onset natural disasters are either of a periodically recurrent or permanent nature; these are droughts and desertification. We broadly associate the loss data with asset losses, i.e. damages to produced capital. This is a simplification, as indirect impacts, such as business interruption, may also be factored into the data. Yet, generally, at least for the sudden onset events, analysts generally equate the data with asset losses, and an indication that this assumption can be maintained is the fact that loss data are usually relatively quickly available after a catastrophe, which indicates that flow impacts emanating over months to years are usually not considered. Losses are compared to estimates of capital stock from Sanderson and Striessnig (2009), which estimated stocks using the perpetual inventory method based on Penn World table information on investments starting in 1900 and assuming annual growth and depreciation of 4 percent. 15 3.3 Projecting disaster impacts on GDP We project differences (in percent) between observed and projected GDP up to five years after a disaster event. A negative value indicates a situation where the projection surpasses the observation leading to a negative effect. Figure 4 charts out these differences for the years 1 to 5. Due to the heterogeneity of the data, it is not very surprising that the results are heavily skewed and as an average value the median should be looked at. 40 67 60 45 36 108 39 40 36 108 30 67 60 39 76 15 145 104 0 -15 156 71 37 92 106 89 89 -30 102 122 106 120 -45 114 114 120 114 102 120 Difference Difference Difference Difference Difference (Year 1) (Year 2) (Year 3) (Year 4) (Year 5) Fig. 4: Box-plots for differences between observed and projected GDP (in percent of observed, baseline GDP in the event year) The mean, median, standard deviation as well as the skewness coefficients for the whole sample are shown in table 3. Table 3: Summary results for differences of observed and projected GDP levels t+1 t+2 t+3 t+4 t+5 Mean -1.27 -1.43 -1.68 -1.75 -2.02 Median -0.53 -1.03 -1.86 -2.27 -3.98 Std. Dev 7.19 11.01 14.99 18.37 22.53 16 Skewness -1.54 -0.76 -0.13 0.42 0.98 According to the skewness and standard deviation the results are asymmetric with a large spread. The results, however, clearly indicate a trend. All post-disaster years show negative values with an increasing "gap," indicating that "on average" one can expect negative economic follow-on consequences in the short-medium term, leading to a median reduction of GDP of about 4% points (of baseline GDP in to) in year 5 after the event. We further test whether the differences are statistically different from zero and, due to non-normality of the data, used the non-parametric one-sample Wilcoxon test (table 4). The null hypothesis H0 is that the median is equal to zero, while the alternative hypothesis H1 is that the median is smaller than zero. Table 4 shows the p-values for this test using the (mean) projections. Table 4: p-values of the Wilcoxon test for differences to be smaller than zero (H1) and H0: equal to zero. t+1 t+2 t+3 t+4 t+5 p-value 0.0138 0.0379 0.0258 0.0171 0.0129 Hypothesis H1 H1 H1 H1 H1 Clearly, the null hypothesis is rejected for all post-disaster years, and therefore one can conclude that there are significant negative follow-on effects. Furthermore, also 95 percent forecast confidence intervals to include uncertainty of the projections within the analysis are used. Additionally, also sub-sample analysis to include uncertainty regarding the influence of multiple occurrences of disasters is performed. The sub-sample is chosen so that only events are considered with no other event (with losses higher than 1 percent of GDP) occurring 5 years before and 5 years after the event considered in the sample. Results related to this sub-sample corroborate our findings on the negative economic consequences (details can be found in Appendix D). 3.4 Explaining the variation: vulnerability predictors 17 As a next step, we test key variables, particularly those relating to economic vulnerability, as to their suitability as predictors for explaining the differences of projected and observed GDP in year 5 post event. Based on the literature review and discussion above, the following variables listed in table 5 are assessed. Table 5: Predictor variables used in the analysis 7 Predictors Variables Source Direct impact and risk Direct monetary losses EMDAT, 2009, Munich Re, 2008 as compiled by Okuyama, 2009 Losses in percent of GDP Okuyama, 2009 Losses in percent of capital stock Own calculations Exposure GDP WDI, 2008 Capital stock Sanderson and Striessnig, 2009 Total number of population WDI, 2008 Hazard Hazard type: EMDAT, 2008 Storm, Flood, Earthquake, Munich Re, 2008 Drought, others Economic vulnerability Indebtedness WDI, 2008 Income level WDI, 2006 Land area WDI, 2008 Literacy rate WDI, 2008 Aid WDI, 2008 Remittances WDI, 2008 Small island development state WDI, 2008 (SIDS) In the following, we first use multivariate models, then employ general linear regression modeling approaches (GLM) using fixed factors, covariates and mixed models as independent variables and Diff(5) as the dependent variable. 7 We did not look at physical vulnerability factors (for example, the quality of building stock in an economy) as predictors, as those do not seem to be of importance in isolation and are accounted for in the direct impact variable. 18 First, exploratory analyses are performed (see tables A-1). Pearson correlation analysis (which assumes a linear relationship) between the continuous variables and Diff(5) leads to (highly) significant results with (log) capital stock losses (correlation of - 0.317, p-value 0.000). Interestingly, such a correlation cannot be found for GDP losses, indicating that capital stock losses may serve as a better predictor. Furthermore, total population (correlation of 0.200, p-value 0.013) as well as aid (in percent of capital formation) are found to be significant (correlation of 0.187, p-value 0.032). Descriptive statistics for Diff(5) within sub-groups according to the income, indebtedness, SIDS and hazard type indicators are considered next (see tables A-2 to A- 6). Using the income indicator, the mean of Diff(5) for all sub-groups exhibits negative values. Also, with regards to the indebtedness indicator, there are negative mean (median) values. As to the type of hazard, storms and earthquakes as well as droughts (if the median is looked at) show negative values. In addition, additional "layers" (or sub- sub groups) are examined; however, the number of observations quickly becomes very small, and therefore average values should be treated with caution. Results of Diff(5) for the interaction of two indicators (which means 6 possible sub-groups) can be found in tables A-6 to A-11. For example, low income in combination with high indebtedness leads to more pronounced negative consequences. Overall, however, a general interpretation of these results is difficult as no clear trend can be discerned. Therefore, we use regression models in the following. Multivariate regression model A forward stepwise regression procedure to detect the most important independent variables from table 5 for the dependent variable Diff(5) is employed. In the first round of the iteration, the independent variables are each added to the starting model (i.e. intercept only model), and the improvement in the residual sum of squares for each of these resulting models is calculated. Next, for each model the p-value for the change in the sum of squares is determined (based on the F-distribution). The variable associated with the lowest p-value is the first model candidate. If the p-value is below 0.1 (significance at the 10% level), then this model is taken. In the next round, this model will be the starting 19 model and the subsequent rounds follow the same procedure as the first. The forward procedure stops if the lowest candidate p-value in subsequent rounds is not lower than 0.1. Table 6 lists the initial model 1 and the final model 2 (all output tables for the full regression model can be found in Appendix B). Table 6: Multivariate Regression results using a forward algorithm( Model=1: Starting model, Model=2: Final model) Model Coefficients Standardized (Unstandardized) Coefficients t p-value B Std. Error Beta 1 Constant 3.254 3.247 1.002 0.322 Percent of Capital -4.600 2.076 -0.317 -2.216 0.032 stock loss (log) 2 Constant -3.095 4.276 -0.724 0.473 Percent of Capital -5.934 2.086 -0.409 -2.844 0.007 stock loss (log) Remittances 1.946 0.897 0.312 2.170 0.036 The final regression model is already reached at step 2, which indicates that the selected variables already have good predictive power. Regarding the fit of the model, while not very satisfactory from a predictive point of view (R square is around 19 percent), two variables are significant at the 5 percent level: capital stock losses (p 0.007) and remittances in the disaster year (p 0.036). While the capital stock loss variable has a negative coefficient suggesting a larger direct shock will lead also to larger negative GDP effects, the remittances parameter has a positive value suggesting that stronger remittances inflow will decrease negative consequences. In line with the exploratory analysis, the direct impacts variable (capital stock losses) seems to be a strong predictor. To summarize, the size of the direct impact (losses) strongly predicts the magnitude of follow-on effects. The fact that it significantly explains the variation in Diff(5), which is based on the time series approach, seems to suggest some validity of the regression results so far. However, interdependencies between variables are not used in this model and are looked at next. 20 General linear regression model A general linear regression modeling approach 8, which also allows for inclusion of interdependencies of several indicator variables, is used next. The model is restricted to selected key variables first identified in the literature review, the further limited by the exploratory analysis (partly presented already in the tables). The model has 4 fixed factors (indicators), including country income group, indebtedness, countries relating to SIDS and hazard type (see table 7). 9 Table 7: Indicators used for the GLM regression Name [abbreviation] Value Label Observations high income 19 Income [I_Income] middle income 96 low income 46 Nan 20 Indebtedness [debt] less indebted 59 medium indebted 18 highly indebted 62 Yes 41 SIDS [I_SIDS] No 118 Storm 55 Flood 41 Hazard [I_Hazard] Earthquake 26 Drought 24 Other 13 8 GLM underlies most of the statistical analyses used in applied and social research due to its widespread applicability. With general linear models many statistical tests can be handled as a regression analysis, including t-tests and ANOVA (Analysis of Variance). 9 The covariates (continuous variables) are chosen based on table 2 and full order effects up to level 2 are included, i.e. relationships between up to two fix factors (indicators) and one covariate are explored within the model. 21 We thus define different sub-samples according to these indicator variables. For example, the whole sample can be split by the income group indicator into 3 sub-samples, the high income sub-sample (19 observations), the middle (94 observations) and low income sub- samples (46 observations). As mentioned, the limitation of higher order effects is mainly due to the decreasing number of observations within sub-groups. Table 8 shows the tests for the different main factors as well as their interactions with the indicators. 10 Full output details can be found in Appendix C. Table 8: GLM Findings: tests of between-subjects effects Dependent Variable: Difference (year 5) Type I Sum of Mean Source Squares df Square F Sig. Corrected Model 21220a 40 531 6.446 .023 Intercept 1337 1 1337 16.243 .010 Literacy rate 244 1 244 2.969 .145 Aid (capital formation) 13 1 13 .162 .704 Aid (percent of import and exports) 764 1 764 9.284 .029 Capital Stock loss (log) [logCapLoss] 1802 1 1802 21.888 .005 Aid (percent of GNI) 2230 1 2230 27.093 .003 Remittances [Remit] 1849 1 1849 22.467 .005 Capital Stock (log) 20 1 20 .238 .646 GDP (log) 80 1 80 .971 .370 Land Area (log) 0 1 0 .003 .956 I_debt * Remit 4108 2 2054 24.959 .003 I_Income * Remit 1 1 1 .008 .931 I_SIDS * Remit 97 1 97 1.174 .328 I_debt * I_Income * Remit 965 1 965 11.723 .019 I_debt * I_SIDS * Remit 653 1 653 7.932 .037 I_debt * I_Hazard * Remit 4155 8 519 6.310 .029 I_Income * I_SIDS * Remit 369 1 369 4.483 .088 I_Income * I_Hazard * Remit 106 1 106 1.291 .307 I_SIDS * I_Hazard * Remit 245 3 82 .991 .468 I_debt * logCapLoss 727 2 364 4.418 .079 I_Income * logCapLoss 698 1 698 8.475 .033 I_SIDS * logCapLoss 5 1 5 .063 .812 I_Hazard * logCapLoss 1805 4 451 5.482 .045 I_debt * I_Income * logCapLoss 82 1 82 .998 .364 I_debt * I_SIDS * logCapLoss 140 1 140 1.706 .248 I_debt * I_Hazard * logCapLoss 63 2 31 .381 .702 I_Income * I_SIDS * logCapLoss 0 0 . . . I_Income * I_Hazard * logCapLoss 0 0 . . . I_SIDS * I_Hazard * logCapLoss 0 0 . . . Error 412 5 82 Total 22969 46 Corrected Total 21632 45 a. R Squared = .981 (Adjusted R Squared = .829) 10 A least squares criterion is used to obtain estimates of the parameters models. 22 As to the model specification (table 8 bottom), the model itself is significant (p-value 0.021) with about 83 percent of the variation explained (R-square 0.829), which is quite satisfactory. Significant variables (p-value smaller than 0.05) include aid (in percent of import and exports), capital stock loss (logged), aid (in percent of GNI), remittances, and interactions of capital stock losses and remittances with some of the other indicators, such as indebtedness, income and hazard. The parameter estimates in Appendix C for the dependent variables cannot be used for interpretation purposes, because GLM models usually have systematic colinearity between the dependent variables and therefore the impact of one single dependent variable is not captured within the parameter estimate. Hence, the variables found to be significant in table 8 are analyzed according to scatter-plots, profile plots as well as comparisons of averages. In line with the observations made above the results lead to the conclusion that especially the direct impact, measured in percent of capital stock loss leads to negative long-term consequences. Remittances as well as various forms of aid decrease the negative effects, however, not as strongly as direct losses. Unfortunately, it has not been possible to refine the analysis with further sub-sub groups, such as looking at country debt levels which seems promising, as the number of observations became too small. Overall, we also find that in general natural disasters can be expected to entail negative consequences in the medium term (five years after an event). As in the multivariate regression, adverse macroeconomic effects can be related to the direct impact in terms of asset losses. Higher aid rates as well as higher remittances (pre- disaster) seem important in lessening the adverse macroeconomic consequences. 4 DISCUSSION There is an ongoing debate on whether disasters cause significant macroeconomic impacts and are truly a potential impediment to economic development. Given the divergent positions, this analysis aimed at better defining a sort of "middle ground" identifying circumstances under which disasters have the potential to cause significant medium-term economic impacts. In a medium-term analysis, comparing counterfactual 23 GDP derived by time series analysis with observed GDP, natural disasters on average lead to significant negative effects on GDP. The negative effects may be small, yet can become more pronounced depending on the direct impact measured as a loss of capital stock. Using regression analysis, we further test a large number of predictors and find that higher aid rates as well as higher remittances importantly lessen the adverse negative macroeconomic consequences, while direct capital stock losses had the largest effects in causing adverse GDP effects. A number of other variables, such as country debt, seemed promising in terms of explaining the variability of GDP, yet it was not possible to further refine the analysis due to small number of observations. Beyond these findings, final conclusions are difficult to draw and the uncertainty in loss data and socioeconomic information has to be acknowledged. One reason is the challenge associated with determining the size and type of impacts as well as identifying additional key predictors. For example, particularly for middle and high income countries, capital stock losses probably play a minor role and other variables such as human and natural capital increasingly become important. Obvious steps for improving the analysis should thus focus on increasing the sample size and quality of data generated, particularly as relates to those lower income and hazard-prone countries supposed to be most vulnerable and of highest interest for the analysis. Finally, another key extension of the analysis would be to also look at disaster impacts on human and environmental capital and its economic repercussions, in isolation as well as in conjunction with produced capital. 24 5 REFERENCES Abeysinghe, T. and Rajaguru, G. (2004). Quarterly Real GDP Estimates for China and ASEAN4 with a Forecast Evaluation. Journal of Forecasting 23: 431-447 Albala-Bertrand, J. M. (1993). Political Economy of Large Natural Disasters With Special Reference to Developing Countries. Oxford, Clarendon Press. Albala-Bertrand, J. M. (2006). The Unlikeliness of an Economic Catastrophe: Localization & Globalization. Working Papers 576, Queen Mary, University of London, Department of Economics. Baritto, F. (2008). Disasters, Vulnerability and Resilience. Working paper. UN-ISDR. Benson, C. (1997a). The economic impact of natural disasters in Fiji. Working Paper No.97. London, UK, Overseas Development Institute (ODI). Benson, C. (1997b). The economic impact of natural disasters in the Philippines. Working Paper No.99. London, UK, Overseas Development Institute (ODI). Benson, C. (1997c). The economic impact of natural disasters in Viet Nam. Working Paper No.98. London, UK, Overseas Development Institute (ODI). Benson, C. (1998). The cost of disasters. Development at risk? Natural disasters and the third World. J. Twigg. Oxford, Oxford Centre for Disaster Studies, UK National Coordinated Committee for the International Decade for Natural Disaster Reduction (IDNDR): 8-13. Benson, C. and Clay, E. (1998). The impact of drought on Sub-Saharan African economies. World Bank Technical Paper Series No. 401. Washington, D.C., The World Bank. Benson, C. and Clay, E. (2000). Developing countries and the economic impacts of catastrophes. Managing disaster risk in emerging economies. A. Kreimer, Arnold, M. (eds.). Washington, D.C., World Bank Publication. Benson, C. and Clay, E. (2001). Dominica: Natural disasters and economic development in a small island state. Disaster Risk Management Working Paper Series No.2. Washington, D.C., The World Bank. Benson, C. and Clay, E. (2004). Understanding the Economic and Financial Impacts of Natural Disasters. Disaster Risk Management Series. The World Bank. Box, G.E.P., and Jenkins, G.M. (1976). Time Series Analysis: Forecasting and Control. San Francisco, Holden-Day. Burton, K., Kates, R. and White, G. (1993). The Environment as Hazard, 2nd edition. New York, Guilford Press. 25 Caselli, F. and Malhotra, P. (2004). Natural Disasters and Growth: from Thought Experiment to Natural Experiment. Washington DC, IMF. Charveriat, C. (2000). Natural disasters in Latin America and the Caribbean: An overview of risk. Working Paper No. 434. Washington, D.C., Inter-American Development Bank (IDB). CRED (2008). EM-DAT: International Disaster Database. Brussels, Belgium, Centre for Research on the Epidemiology of Disasters, Université Catholique de Louvain. Crowards, T. (2000). Comparative vulnerability to natural disasters in the Caribbean. Charleston, South Carolina, Caribbean Development Bank. Cuaresma, J., Hlouskova, J. ,Obersteiner, M (2008). Natural disasters as creative destruction. Evidence from developing countries. Economic Inquiry 46(2): 214-226 ECLAC (1982). Nicaragua: Las inundaciones de mayo de 1982 y sus repercusiones sobre el desarrollo economico y social del pais. Report No.E/CEPAL/MEX/1982/R.2/Rev.1. New York, Economic Commission for Latin American Countries, United Nations. ECLAC (1985). Damage caused by the Mexican earthquake and its repercussions upon the country's economy. Santiago de Chile, Economic Commission for Latin America and the Caribbean. ECLAC (1988). Damage caused by Hurricane Joan in Nicaragua: Its effects on economic development and living conditions, and requirements for rehabilitation and reconstruction. New York, Economic Commission for Latin American Countries, United Nations. ECLAC (1999). Manual for estimating the socio-economic effects of natural disasters. New York City, United Nations Economic Commission for Latin America and the Caribbean, International Decade for Natural Disaster Reduction. ECLAC (2002). Handbook for estimating socio-economic and environment effects of disasters. Mexico City, ECLAC. ECLAC and IDB (2000). A matter of development: how to reduce vulnerability in the face of natural disasters. Seminar "Confronting Natural Disasters: A Matter of Development", New Orleans, 25-26.3.2000. Freeman, P. K., L. A. Martin, R. Mechler, K. Warner with P. Hausmann (2002). Catastrophes and Development. Integrating Natural Catastrophes into Development Planning. Washington DC, World Bank. Hochrainer, S. (2006). Macroeconomic risk management against natural disasters. Wiesbaden, German University Press (DUV). Kahn, M. (2005). The Death Toll From Natural Disasters: The Role of Income, Geography, and Institutions. Review of Economics and Statistics 87: 271-284. 26 Makridakis, S. and Wheelwright, S.C. (1989). Forecasting Methods for Management (5th edition), Chichester: Wiley. Mechler, R. (2004). Natural Disaster Risk Management and Financing Disaster Losses in Developing Countries. Karlsruhe, Verlag für Versicherungswissenschaft. Munich Re (2008). Disaster loss data. Munich Re NatCat Service. Munich, Munich Re. Murlidharan, T. L. and Shah, H.C. (2001). Catastrophes and macro-economic risk factors: An empirical study. Conference on 'Integrated Disaster Risk Management: Reducing Socio- Economic Vulnerability', Laxenburg, Austria, International Institute for Applied Systems Analysis (IIASA). Noy, I. (2009). The macroeconomic consequences of disasters. Journal of Development Economics 88: 221-231 OAS (1991). Primer on Natural Hazard Management in Integrated Regional Development Planning. Washington DC, Organization of American States. Okuyama, Y. (2009). Impact estimation of higher order effects. Background paper for EDRR report. Otero, R. C. and R.Z. Marti (1995). The impacts of natural disasters on developing economies: implications for the international development and disaster community. In M. Munasinghe and C. Clarke (eds.). Disaster Prevention for Sustainable Development: Economic and Policy Issues. Washington DC, World Bank: 11-40 Raddatz, C. (2007). Are external shocks responsible for the instability of output in low-income countries? Journal of Development Economics 84, 155-187 Rasmussen, T.N. (2004). Macroeconomic implications of natural disasters in the Caribbean. IMF Working Paper WP/04/224. Raschky, P. (2008). Institutions and the losses from natural disasters. Natural Hazards and Earth System Sciences, 8, 627-634. Sanderson, W.C. and Striessnig, E. (2009). Demography, Education, and the Future of Total Factor Productivity Growth. IIASA Interim Report IR-09-002, Laxenburg, Austria. Skidmore, M. and Toya, H. (2002). Do Natural Disasters Promote Long-Run Growth? Economic Inquiry 40(4): pages 664-687 Toya, H. and Skidmore, M (2007). Economic development and the impacts of natural disasters. Economics Letters: 94(1): 20-25 Zapata, R. (2008). Quantification of disaster impacts. Seminar Power point Presentation. 27 Zenklusen, O. (2007). Natural disasters and economic development: A neoclassical review of theoretical perspectives and empirical evidence. Dissertation, University of St. Gallen, Switzerland. World Bank (2006). Where is the wealth of nations?, Washington DC, World Bank. World Bank (2007). World Development Indicators (WDI), Washington DC, World Bank. World Bank (2008). World Development Indicators (WDI), Washington DC, World Bank. 28 Appendix A: Tables Table A-1: Correlation matrix Correlations Loss in Loss in Loss in Literacy rate Difference percent of monetary percent of Total (percent of Government (year 5) GDP Capital Stock GDP terms Capital Stock Population adult) Aid Difference (year 5) Pearson Correlation 1 -.105 .051 -.128 -.142 -.184* .200* .098 -.092 Sig. (2-tailed) .195 .528 .117 .083 .025 .013 .388 .653 N 155 155 155 150 150 149 155 80 26 Loss in percent of GDP Pearson Correlation -.105 1 -.102 -.052 .261** .334** -.099 .093 -.065 Sig. (2-tailed) .195 .131 .466 .000 .000 .152 .338 .689 N 155 220 220 199 199 193 210 108 40 Capital Stock Pearson Correlation .051 -.102 1 .242** .174* -.025 .693** .107 -.035 Sig. (2-tailed) .528 .131 .001 .014 .728 .000 .269 .832 N 155 220 220 199 199 193 210 108 40 GDP Pearson Correlation -.128 -.052 .242** 1 .422** .014 .101 .084 -.066 Sig. (2-tailed) .117 .466 .001 .000 .846 .156 .399 .692 N 150 199 199 199 199 193 199 102 39 Loss in monetary terms Pearson Correlation -.142 .261** .174* .422** 1 .948** .035 .073 -.071 Sig. (2-tailed) .083 .000 .014 .000 .000 .628 .463 .666 N 150 199 199 199 199 193 199 102 39 Loss in percent of Pearson Correlation -.184* .334** -.025 .014 .948** 1 -.023 .017 -.057 Capital Stock Sig. (2-tailed) .025 .000 .728 .846 .000 .749 .864 .734 N 149 193 193 193 193 193 193 99 38 Total Population Pearson Correlation .200* -.099 .693** .101 .035 -.023 1 .028 -.044 Sig. (2-tailed) .013 .152 .000 .156 .628 .749 .776 .789 N 155 210 210 199 199 193 210 105 40 Literacy rate (percent of Pearson Correlation .098 .093 .107 .084 .073 .017 .028 1 .112 adult) Sig. (2-tailed) .388 .338 .269 .399 .463 .864 .776 .629 N 80 108 108 102 102 99 105 108 21 Government Aid Pearson Correlation -.092 -.065 -.035 -.066 -.071 -.057 -.044 .112 1 Sig. (2-tailed) .653 .689 .832 .692 .666 .734 .789 .629 N 26 40 40 39 39 38 40 21 40 *. Correlation is significant at the 0.05 level (2-tailed). **. Correlation is significant at the 0.01 level (2-tailed). Table A-1: Correlation matrix (continued) Correlations Loss in Aid (percent Loss in percent of Difference Aid (capital of imports percent of Capital (year 5) formation) and exports) Land area GDP (log) Stock (log) Aid (% of GNI) Remittances Difference (year 5) Pearson Correlation 1 .187* .132 .118 -.149 -.317** .061 .107 Sig. (2-tailed) .032 .162 .143 .064 .000 .494 .277 N 155 132 113 155 155 149 130 106 Aid (capitall formation) Pearson Correlation .187* 1 .763** -.171* .034 -.034 .813** .009 Sig. (2-tailed) .032 .000 .025 .661 .668 .000 .921 N 132 171 133 171 171 160 161 122 Aid (percent of imports Pearson Correlation .132 .763** 1 -.147 .052 .049 .636** .041 and exports) Sig. (2-tailed) .162 .000 .081 .540 .572 .000 .656 N 113 133 142 142 142 133 136 121 Land area Pearson Correlation .118 -.171* -.147 1 -.203** -.338** -.137 -.195* Sig. (2-tailed) .143 .025 .081 .002 .000 .065 .016 N 155 171 142 220 220 193 183 152 Loss in percent of Pearson Correlation -.149 .034 .052 -.203** 1 .714** .208** .355** GDP (log) Sig. (2-tailed) .064 .661 .540 .002 .000 .005 .000 N 155 171 142 220 220 193 183 152 Loss in percent of Pearson Correlation -.317** -.034 .049 -.338** .714** 1 .100 .210* Capital Stock (log) Sig. (2-tailed) .000 .668 .572 .000 .000 .210 .015 N 149 160 133 193 193 193 160 133 Aid (% of GNI) Pearson Correlation .061 .813** .636** -.137 .208** .100 1 .172* Sig. (2-tailed) .494 .000 .000 .065 .005 .210 .049 N 130 161 136 183 183 160 183 132 Remittances Pearson Correlation .107 .009 .041 -.195* .355** .210* .172* 1 Sig. (2-tailed) .277 .921 .656 .016 .000 .015 .049 N 106 122 121 152 152 133 132 152 *. Correlation is significant at the 0.05 level (2-tailed). **. Correlation is significant at the 0.01 level (2-tailed). 29 Table A-1: Correlation matrix (continued) Correlations Difference Capital Money Land (year 5) Stock (log) GDP (log) loss (log) Area (log) Difference (year 5) Pearson Correlation 1 .117 -.065 -.177* .043 Sig. (2-tailed) .147 .428 .030 .598 N 155 154 150 150 155 Capital Stock (log) Pearson Correlation .117 1 .833** .618** .624** Sig. (2-tailed) .147 .000 .000 .000 N 154 204 193 193 204 GDP (log) Pearson Correlation -.065 .833** 1 .837** .593** Sig. (2-tailed) .428 .000 .000 .000 N 150 193 199 199 199 Money loss (log) Pearson Correlation -.177* .618** .837** 1 .368** Sig. (2-tailed) .030 .000 .000 .000 N 150 193 199 199 199 Land Area (log) Pearson Correlation .043 .624** .593** .368** 1 Sig. (2-tailed) .598 .000 .000 .000 N 155 204 199 199 220 *. Correlation is significant at the 0.05 level (2-tailed). **. Correlation is significant at the 0.01 level (2-tailed). Table A-2: Diff(5) vs. Income Difference (year 5) * Income level Difference (year 5) Income level N Mean Std. Deviation Median Skewness high income 19 -10.0428 10.28454 -8.2346 -.610 low income 46 -1.5493 28.08414 1.8748 .661 middle income 90 -.1570 21.37437 -4.1126 1.075 Total 155 -1.7820 22.73418 -3.4932 .951 Table A-3: Diff(5) vs. Debt Difference (year 5) * Indebtedness Difference (year 5) Indebtedness level N Mean Std. Deviation Median Skewness NanN 20 -8.5480 12.31746 -7.4272 -.033 highly indebted 62 -.6998 26.53054 1.7900 .629 medium indebted 17 1.4293 33.09615 -8.4707 1.283 less indebted 56 -1.5386 16.55988 -4.8505 .396 Total 155 -1.7820 22.73418 -3.4932 .951 30 Table A-4: Diff(5) vs. SIDS Difference (year 5) * SIDS Difference (year 5) SIDS N Mean Std. Deviation Median Skewness no 114 -1.0722 21.51452 -2.5134 1.009 yes 41 -3.7554 26.01534 -3.9810 .944 Total 155 -1.7820 22.73418 -3.4932 .951 Table A-5: Diff(5) vs. Hazard type Difference (year 5) * Hazard type Difference (year 5) Hazard type N Mean Std. Deviation Median Skewness Storm 53 -3.2304 15.29672 -5.1644 1.287 Flood 41 2.5940 22.90447 3.0448 -.032 Earthquake 25 -3.6452 23.32322 -4.4723 .998 Drought 23 4.6507 31.28664 -5.4178 1.711 other 13 -17.4760 23.65540 -9.8835 -.427 Total 155 -1.7820 22.73418 -3.4932 .951 Table A-6: Diff(5) vs. Income vs. Debt. Difference (year 5) Income level Indebtedness level N Mean Std. Deviation Median Skewness high income NanN 16 -9.8812 10.82718 -7.4272 -.679 less indebted 3 -10.9044 8.45075 -11.5879 .362 Total 19 -10.0428 10.28454 -8.2346 -.610 low income highly indebted 41 -1.1036 29.47572 3.3870 .603 medium indebted 5 -5.2039 12.89095 -8.1523 .716 Total 46 -1.5493 28.08414 1.8748 .661 middle income NanN 4 -3.2148 18.09280 -3.6352 .114 highly indebted 21 .0887 20.20344 .4068 .857 medium indebted 12 4.1931 38.78767 -10.0309 .991 less indebted 53 -1.0084 16.79157 -4.5365 .331 Total 90 -.1570 21.37437 -4.1126 1.075 Total NanN 20 -8.5480 12.31746 -7.4272 -.033 highly indebted 62 -.6998 26.53054 1.7900 .629 medium indebted 17 1.4293 33.09615 -8.4707 1.283 less indebted 56 -1.5386 16.55988 -4.8505 .396 Total 155 -1.7820 22.73418 -3.4932 .951 31 Table A-7: Diff(5) vs. Income vs. Hazard type Difference (year 5) * Income level * Hazard type Difference (year 5) Income level Hazard type N Mean Std. Deviation Median Skewness high income Storm 6 -9.9249 7.93491 -8.9508 -.233 Flood 3 -8.7336 7.81854 -12.3034 1.626 Earthquake 6 -15.3454 14.40122 -16.6655 .329 Drought 3 -4.5909 4.97845 -6.6197 1.529 other 1 .7820 . .7820 . Total 19 -10.0428 10.28454 -8.2346 -.610 low income Storm 14 1.4656 9.33077 4.1589 -.602 Flood 16 4.5497 28.45303 7.6761 -.186 Earthquake 3 -11.8533 37.53206 3.5834 -1.538 Drought 9 .8741 39.00068 -7.8058 2.163 other 4 -34.2222 24.94859 -41.9039 1.528 Total 46 -1.5493 28.08414 1.8748 .661 middle income Storm 33 -4.0054 17.78638 -6.9198 1.482 Flood 22 2.7163 19.84732 1.7985 -.168 Earthquake 16 2.2814 22.53234 .6409 2.057 Drought 11 10.2610 29.30183 4.9994 1.213 other 8 -11.3851 21.02976 -7.6592 -.918 Total 90 -.1570 21.37437 -4.1126 1.075 Total Storm 53 -3.2304 15.29672 -5.1644 1.287 Flood 41 2.5940 22.90447 3.0448 -.032 Earthquake 25 -3.6452 23.32322 -4.4723 .998 Drought 23 4.6507 31.28664 -5.4178 1.711 other 13 -17.4760 23.65540 -9.8835 -.427 Total 155 -1.7820 22.73418 -3.4932 .951 Table A-8: Diff(5) vs. Income vs. SIDS Difference (year 5) * Income level * SIDS Difference (year 5) Income level SIDS N Mean Std. Deviation Median Skewness high income no 16 -11.0912 10.62976 -9.9112 -.457 yes 3 -4.4515 6.98729 -2.1327 -1.329 Total 19 -10.0428 10.28454 -8.2346 -.610 low income no 33 1.5991 21.89418 4.9307 .033 yes 13 -9.5415 39.78641 -5.4178 1.464 Total 46 -1.5493 28.08414 1.8748 .661 middle income no 65 .0377 22.82711 -4.2906 1.281 yes 25 -.6632 17.44403 -3.9810 -.339 Total 90 -.1570 21.37437 -4.1126 1.075 Total no 114 -1.0722 21.51452 -2.5134 1.009 yes 41 -3.7554 26.01534 -3.9810 .944 Total 155 -1.7820 22.73418 -3.4932 .951 32 Table A-9: Diff(5) vs. Hazard vs. SIDS Report Difference (year 5) Indebtedness level SIDS N Mean Std. Deviation Median Skewness NanN no 17 -10.4743 11.15400 -8.2346 -.494 yes 3 2.3679 15.35486 1.0816 .374 Total 20 -8.5480 12.31746 -7.4272 -.033 highly indebted no 42 .7537 23.01306 3.3392 .329 yes 20 -3.7520 33.20379 -1.0841 1.013 Total 62 -.6998 26.53054 1.7900 .629 middle indebted no 12 6.2663 34.49705 -6.2215 1.500 yes 5 -10.1795 29.49825 -11.5911 .481 Total 17 1.4293 33.09615 -8.4707 1.283 low indebted no 43 -1.1866 17.74184 -5.4348 .300 yes 13 -2.7030 12.38017 -3.9810 1.078 Total 56 -1.5386 16.55988 -4.8505 .396 Total no 114 -1.0722 21.51452 -2.5134 1.009 yes 41 -3.7554 26.01534 -3.9810 .944 Total 155 -1.7820 22.73418 -3.4932 .951 Table A-10: Diff(5) vs. Debt. vs. SIDS Difference (year 5) Indebtedness level SIDS N Mean Std. Deviation Median Skewness NanN no 17 -10.4743 11.15400 -8.2346 -.494 yes 3 2.3679 15.35486 1.0816 .374 Total 20 -8.5480 12.31746 -7.4272 -.033 highly indebted no 42 .7537 23.01306 3.3392 .329 yes 20 -3.7520 33.20379 -1.0841 1.013 Total 62 -.6998 26.53054 1.7900 .629 medium indebted no 12 6.2663 34.49705 -6.2215 1.500 yes 5 -10.1795 29.49825 -11.5911 .481 Total 17 1.4293 33.09615 -8.4707 1.283 less indebted no 43 -1.1866 17.74184 -5.4348 .300 yes 13 -2.7030 12.38017 -3.9810 1.078 Total 56 -1.5386 16.55988 -4.8505 .396 Total no 114 -1.0722 21.51452 -2.5134 1.009 yes 41 -3.7554 26.01534 -3.9810 .944 Total 155 -1.7820 22.73418 -3.4932 .951 33 Table A-11: Diff(5) vs. Debt. vs. Hazard Difference (year 5) Indebtedness level Hazard type N Mean Std. Deviation Median Skewness NanN Storm 3 -8.9454 9.11666 -6.3138 -1.191 Flood 5 -6.6942 7.78935 -10.3152 .548 Earthquake 6 -15.3454 14.40122 -16.6655 .329 Drought 5 -3.8723 15.37481 -6.6197 .333 other 1 .7820 . .7820 . Total 20 -8.5480 12.31746 -7.4272 -.033 highly indebted Storm 28 .9441 17.26948 2.6662 .951 Flood 16 4.6616 29.24047 7.6761 -.252 Earthquake 4 -7.9115 31.64260 3.7486 -1.811 Drought 9 2.7294 38.47571 -5.4178 2.081 other 5 -27.4645 26.36579 -37.2340 .369 Total 62 -.6998 26.53054 1.7900 .629 medium indebted Storm 4 -9.1192 3.59991 -10.0309 1.009 Flood 5 -7.6672 16.27240 -8.1523 .013 Earthquake 1 71.3230 . 71.3230 . Drought 5 12.6623 43.04089 -12.3435 1.152 other 2 -17.7618 43.65925 -17.7618 . Total 17 1.4293 33.09615 -8.4707 1.283 less indebted Storm 18 -7.4628 12.97722 -8.3785 1.456 Flood 15 6.9051 19.91528 6.3466 -.382 Earthquake 14 -2.7668 13.83809 -2.2192 .650 Drought 4 9.6128 13.16919 10.9818 -.443 other 5 -11.0247 15.71403 -9.8835 -1.098 Total 56 -1.5386 16.55988 -4.8505 .396 Total Storm 53 -3.2304 15.29672 -5.1644 1.287 Flood 41 2.5940 22.90447 3.0448 -.032 Earthquake 25 -3.6452 23.32322 -4.4723 .998 Drought 23 4.6507 31.28664 -5.4178 1.711 other 13 -17.4760 23.65540 -9.8835 -.427 Total 155 -1.7820 22.73418 -3.4932 .951 34 Appendix B: Linear (forward) regression: Details Table B-1: Model Summary Model Summary Adjusted Std. Error of Model R R Square R Square the Estimate 1 .317a .100 .080 21.03051 2 .435b .189 .151 20.19663 a. Predictors: (Constant), Loss in percent of Capital Stock (log) b. Predictors: (Constant), Loss in percent of Capital Stock (log), Remittances Table B-2: ANOVA c ANOVA Sum of Model Squares df Mean Square F Sig. 1 Regression 2171.570 1 2171.570 4.910 .032a Residual 19460.421 44 442.282 Total 21631.991 45 2 Regression 4092.124 2 2046.062 5.016 .011b Residual 17539.867 43 407.904 Total 21631.991 45 a. Predictors: (Constant), Loss in percent of Capital Stock (log) b. Predictors: (Constant), Loss in percent of Capital Stock (log), Remittances c. Dependent Variable: Difference (year 5) Table B-3: Coefficients Coefficientsa Unstandardized Standardized Coefficients Coefficients Model B Std. Error Beta t Sig. 1 (Constant) 3.254 3.247 1.002 .322 Loss in percent of -4.600 2.076 -.317 -2.216 .032 Capital Stock (log) 2 (Constant) -3.095 4.276 -.724 .473 Loss in percent of -5.934 2.086 -.409 -2.844 .007 Capital Stock (log) Remittances 1.946 .897 .312 2.170 .036 a. Dependent Variable: Difference (year 5) 35 Table B-4: Excluded Variables Excluded Variablesc Collinearity Partial Statistics Model Beta In t Sig. Correlation Tolerance 1 Capital Stock (log) -.163 a -.807 .424 -.122 .506 GDP (log) -.131 a -.878 .385 -.133 .916 Money loss (log) -.116 a -.807 .424 -.122 1.000 Land Area (log) -.221 a -1.332 .190 -.199 .728 Remittances .312a 2.170 .036 .314 .913 Aid (% of GNI) -.083 a -.570 .572 -.087 .983 Loss in percent of a .043 .211 .834 .032 .512 GDP (log) Aid (capital formation) a .047 .319 .751 .049 .968 Aid (percent of a .102 .696 .490 .105 .954 imports and exports) Literacy rate (percent a .011 .075 .940 .011 .908 of adult) 2 Capital Stock (log) -.123 b -.629 .533 -.097 .501 GDP (log) -.100 b -.688 .495 -.106 .906 Money loss (log) -.087 b -.629 .533 -.097 .990 Land Area (log) -.070 b -.376 .709 -.058 .561 Aid (% of GNI) -.027 b -.187 .853 -.029 .948 Loss in percent of b .034 .177 .861 .027 .512 GDP (log) Aid (capital formation) b .108 .756 .454 .116 .933 Aid (percent of b .169 1.183 .243 .180 .918 imports and exports) Literacy rate (percent b -.049 -.330 .743 -.051 .876 of adult) a. Predictors in the Model: (Constant), Loss in percent of Capital Stock (log) b. Predictors in the Model: (Constant), Loss in percent of Capital Stock (log), Remittances c. Dependent Variable: Difference (year 5) 36 Appendix C: General Linear Regression Table C-1: Between-Subject factors Name [abbreviation] Value Label N high income 19 Income [I_Income] middle income 96 low income 46 Nan 20 Indebtedness [debt] less indebted 59 medium indebted 18 highly indebted 62 Yes 41 SIDS [I_SIDS] No 118 Storm 55 Flood 41 Hazard [I_Hazard] Earthquake 26 Drought 24 Other 13 37 Table C-2: Tests of between-Subject factors Dependent Variable: Difference (year 5) Type I Sum of Mean Source Squares df Square F Sig. Corrected Model 21220a 40 531 6.446 .023 Intercept 1337 1 1337 16.243 .010 Literacy rate 244 1 244 2.969 .145 Aid (capital formation) 13 1 13 .162 .704 Aid (percent of import and exports) 764 1 764 9.284 .029 Capital Stock loss (log) [logCapLoss] 1802 1 1802 21.888 .005 Aid (percent of GNI) 2230 1 2230 27.093 .003 Remittances [Remit] 1849 1 1849 22.467 .005 Capital Stock (log) 20 1 20 .238 .646 GDP (log) 80 1 80 .971 .370 Land Area (log) 0 1 0 .003 .956 I_debt * Remit 4108 2 2054 24.959 .003 I_Income * Remit 1 1 1 .008 .931 I_SIDS * Remit 97 1 97 1.174 .328 I_debt * I_Income * Remit 965 1 965 11.723 .019 I_debt * I_SIDS * Remit 653 1 653 7.932 .037 I_debt * I_Hazard * Remit 4155 8 519 6.310 .029 I_Income * I_SIDS * Remit 369 1 369 4.483 .088 I_Income * I_Hazard * Remit 106 1 106 1.291 .307 I_SIDS * I_Hazard * Remit 245 3 82 .991 .468 I_debt * logCapLoss 727 2 364 4.418 .079 I_Income * logCapLoss 698 1 698 8.475 .033 I_SIDS * logCapLoss 5 1 5 .063 .812 I_Hazard * logCapLoss 1805 4 451 5.482 .045 I_debt * I_Income * logCapLoss 82 1 82 .998 .364 I_debt * I_SIDS * logCapLoss 140 1 140 1.706 .248 I_debt * I_Hazard * logCapLoss 63 2 31 .381 .702 I_Income * I_SIDS * logCapLoss 0 0 . . . I_Income * I_Hazard * logCapLoss 0 0 . . . I_SIDS * I_Hazard * logCapLoss 0 0 . . . Error 412 5 82 Total 22969 46 Corrected Total 21632 45 a. R Squared = .981 (Adjusted R Squared = .829) 38 Table C-3: Parameter estimates a mte s a s Pra e rEtimte ee dn a b : if re c e r ) Dpn etVrialeDfe ne(y a 5 5 Cn ece te a 9% o fid n In rv l a mt r Pra ee B td rro S .E r t S. ig o e on LwrBu d pe on UprBu d te et In rc p 54 6.0 8 22 8.0 0 .73 9 6 .44 4.7 1 -1 5 9 7.8 6 25 8 i ray Lte c 9 -.3 4 6 .32 -1 8 .08 2 .36 .34 -1 2 3 .57 Agidcf 9 -.1 2 4 .22 -.7 6 9 6 .42 1 -.8 3 2 .49 id e A imx 9 .31 2 .44 .93 2 9 .38 9 -.6 8 .49 17 g aL s lo Cpos 15 -2 .60 2.5 5 29 4 -.0 4 9 2 .99 1.7 5 -6 1 1 6.4 4 58 1 id N A GI 9 -.2 7 5 .99 -.3 9 0 7 .70 .72 -2 6 .19 26 Rm e it 68 -1 .47 9.7 4 14 5 -.0 5 8 3 .96 1.1 0 -5 7 2 8.1 5 44 4 g aS ck lo Cp to .90 25 .85 68 .48 2 8 .66 44 -1 .79 04 2.6 9 gD lo GP 14 -1 .16 .86 78 -1 1 .44 1 .27 11 -3 .47 .14 92 g ad re lo Ln A a 12 1.3 0 .27 48 24 .60 4 .06 9 .29 24 2.3 0 db 1 0 e it [I_e t=.0 ]*Rm 7.3 7 23 3 7.0 6 32 6 .75 3 9 .46 8.0 8 -6 3 8 29 6 12 .72 db 2 0 e it [I_e t=.0 ]*Rm 2.8 2 12 7 3.2 3 22 4 .59 2 1 .69 7.1 7 -4 4 2 1.8 2 79 7 db 3 0 e it [I_e t=.0 ]*Rm 0 . . . . . I co e 7.0] e it [I_n m=6 0*Rm 4.6 9 -2 5 6 0.3 7 56 2 -.4 5 8 4 .68 57 2 -1 4.2 3 05 8 15 .86 I co e 7.0] e it [I_n m=7 0*Rm 0 . . . . . S S.0 ] e it [I_ ID= 0*Rm 42 2.3 2 0.0 7 22 9 .10 2 0 .99 9.1 3 -4 5 8 4.8 8 53 2 S S1 0 e it [I_ ID= .0]*Rm 0 . . . . . db 1 0 In m=6 0 e it [I_e t=.0 ]*[I_ co e 7.0]*Rm 53 -6 .84 2.9 6 15 0 -.5 3 2 2 .63 8.4 7 -3 9 8 5.8 8 27 1 db 1 0 In m=7 0 e it [I_e t=.0 ]*[I_ co e 7.0]*Rm 0 . . . . . db 2 0 In m=6 0 e it [I_e t=.0 ]*[I_ co e 7.0]*Rm 0 . . . . . db 2 0 In m=7 0 e it [I_e t=.0 ]*[I_ co e 7.0]*Rm 0 . . . . . db 3 0 In m=7 0 e it [I_e t=.0 ]*[I_ co e 7.0]*Rm 0 . . . . . db 1 0 S S.0] e it [I_e t=.0 ]*[I_ ID= 0*Rm 6.4 7 -1 3 1 7.0 1 43 5 -.3 5 4 4 .74 39 3 -1 7.4 3 02 9 15 .59 db 1 0 S S1 0 e it [I_e t=.0 ]*[I_ ID= .0]*Rm 0 . . . . . db 2 0 S S.0] e it [I_e t=.0 ]*[I_ ID= 0*Rm 0 . . . . . db 3 0 S S.0] e it [I_e t=.0 ]*[I_ ID= 0*Rm 0 . . . . . db 3 0 S S1 0 e it [I_e t=.0 ]*[I_ ID= .0]*Rm 0 . . . . . db 1 0 Hz rd 1 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 7.4 2 -2 1 4 8.0 4 30 5 -.7 4 1 0 .57 28 0 -1 4.4 1 0.5 7 75 1 db 1 0 Hz rd 2 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 54 5.5 2 56 6.9 1 .82 4 3 .48 1.0 5 -1 4 1 2.0 9 25 9 db 1 0 Hz rd 4 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 0 . . . . . db 1 0 Hz rd 5 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 0 . . . . . db 2 0 Hz rd 2 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 22 -3 .40 4.6 7 13 8 -.2 6 2 3 .80 0.7 0 -4 1 8 3.9 0 36 4 db 2 0 Hz rd 3 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 3.7 2 -1 0 2 3.2 6 28 7 -.5 9 4 0 .67 4.2 9 -7 3 2 8.7 6 41 8 db 2 0 Hz rd 4 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 0 . . . . . db 3 0 Hz rd 1 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 28 1.8 6 4.1 8 54 6 .04 2 8 .92 35 4 -1 8.9 3 41 1 11 .75 db 3 0 Hz rd 2 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 57 5.1 2 9.8 6 10 5 .29 8 8 .74 3.4 9 -4 5 3 4.7 4 55 8 db 3 0 Hz rd 3 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm .35 -5 8 .59 79 -.7 9 0 1 .50 41 -2 .99 44 1.1 9 db 3 0 Hz rd 4 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 0.6 8 13 1 2.8 8 28 1 .43 5 7 .60 8.5 6 -4 4 7 9.8 2 61 1 db 3 0 Hz rd 5 0 e it [I_e t=.0 ]*[I_ aa = .0]*Rm 0 . . . . . I co e 7.0] _ ID= 0 e it [I_n m=6 0*[I S S.0]*Rm 0.8 1 36 7 8.7 0 41 2 .67 3 5 .52 3.4 9 -9 1 2 55 7 14 .11 I co e 7.0] _ ID= . 0 e it [I_n m=6 0*[I S S10 ]*Rm 0 . . . . . I co e 7.0] _ ID= 0 e it [I_n m=7 0*[I S S.0]*Rm 0 . . . . . I co e 7.0] _ ID= . 0 e it [I_n m=7 0*[I S S10 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=6 0*[I Hz rd 10 ]*Rm 7.9 3 18 0 5.0 6 19 2 12 .15 1 .32 2.8 7 -2 9 8 8.6 2 57 9 I co e 7.0] _ aa = . 0 e it [I_n m=6 0*[I Hz rd 20 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=6 0*[I Hz rd 40 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=7 0*[I Hz rd 10 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=7 0*[I Hz rd 20 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=7 0*[I Hz rd 30 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=7 0*[I Hz rd 40 ]*Rm 0 . . . . . I co e 7.0] _ aa = . 0 e it [I_n m=7 0*[I Hz rd 50 ]*Rm 0 . . . . . S S.0 ] Hz rd 1 0 e it [I_ ID= 0*[I_ aa = .0]*Rm 64 -1 .43 5.1 6 50 8 -.0 0 3 7 .97 40 4 -1 3.7 1 37 5 19 .85 S S.0 ] Hz rd 2 0 e it [I_ ID= 0*[I_ aa = .0]*Rm 6.5 7 -1 8 8 9.8 5 17 1 -.8 2 5 3 .43 7.0 7 -6 7 8 3.9 3 39 1 S S.0 ] Hz rd 3 0 e it [I_ ID= 0*[I_ aa = .0]*Rm 0 . . . . . S S.0 ] Hz rd 4 0 e it [I_ ID= 0*[I_ aa = .0]*Rm 1.0 6 -1 3 8 3.7 2 27 3 -.4 6 7 5 .64 2.1 5 -7 4 9 9.0 4 48 2 S S.0 ] Hz rd 5 0 e it [I_ ID= 0*[I_ aa = .0]*Rm 0 . . . . . S S1 0 I Hz rd 1 0 e it [I_ ID= .0]*[_ aa =.0 ]*Rm 0 . . . . . S S1 0 I Hz rd 2 0 e it [I_ ID= .0]*[_ aa =.0 ]*Rm 0 . . . . . S S1 0 I Hz rd 4 0 e it [I_ ID= .0]*[_ aa =.0 ]*Rm 0 . . . . . S S1 0 I Hz rd 5 0 e it [I_ ID= .0]*[_ aa =.0 ]*Rm 0 . . . . . db 1 0 g aL s [I_e t=.0 ]*lo Cpos 93 -8 .04 1.8 5 14 3 -.7 5 7 7 .43 8.2 6 -3 4 2 0.1 9 26 5 db 2 0 g aL s [I_e t=.0 ]*lo Cpos 47 -7 .51 24 3.9 5 -2 6 .24 7 .03 5.2 8 -1 9 5 01 1.1 6 db 3 0 g aL s [I_e t=.0 ]*lo Cpos 0 . . . . . I co e 7.0] g a Lss [I_n m=6 0*lo Cp o 95 5.2 4 45 3.8 3 10 .70 5 .10 03 -3 .39 4.8 8 18 4 I co e 7.0] g a Lss [I_n m=7 0*lo Cp o 0 . . . . . S S.0 ] g aL s [I_ ID= 0*lo Cp os 3.8 6 -1 7 9 97 7.0 5 -1 4 .74 4 .12 4.1 6 -3 1 6 57 6.3 3 S S1 0 g a Lss [I_ ID= .0]*lo Cp o 0 . . . . . Hz rd 1 0 g a Lss [I_ aa = .0]*lo Cp o 3.9 3 17 4 9.2 5 13 7 .74 1 0 .57 5.8 6 -3 8 8 3.7 2 64 7 Hz rd 2 0 g a Lss [I_ aa = .0]*lo Cp o 5.8 4 13 9 0.8 2 22 4 .79 5 8 .42 6.5 9 -3 7 2 7.3 7 65 1 Hz rd 3 0 g a Lss [I_ aa = .0]*lo Cp o 6.2 6 11 0 9.4 2 19 0 .88 0 5 .46 5.3 5 -3 1 7 7.7 6 63 8 Hz rd 4 0 g a Lss [I_ aa = .0]*lo Cp o 6.1 6 10 0 0.7 6 22 6 .70 9 6 .46 6.1 1 -3 1 2 8.3 4 61 3 Hz rd 5 0 g a Lss [I_ aa = .0]*lo Cp o 0 . . . . . db 1 0 In m=6 0 g a Lss [I_e t=.0 ]*[I_ co e 7.0]*lo Cp o 82 -4 .94 23 4.2 7 -1 5 .18 9 .29 5.4 8 -1 7 9 94 5.6 9 db 1 0 In m=7 0 g a Lss [I_e t=.0 ]*[I_ co e 7.0]*lo Cp o 0 . . . . . db 2 0 In m=6 0 g a Lss [I_e t=.0 ]*[I_ co e 7.0]*lo Cp o 0 . . . . . db 2 0 In m=7 0 g a Lss [I_e t=.0 ]*[I_ co e 7.0]*lo Cp o 0 . . . . . db 3 0 In m=7 0 g a Lss [I_e t=.0 ]*[I_ co e 7.0]*lo Cp o 0 . . . . . db 1 0 S S.0] g aL s [I_e t=.0 ]*[I_ ID= 0*lo Cp os 1.5 2 14 2 0.9 8 12 7 11 .12 1 .37 5.1 2 -1 0 9 7.2 5 39 3 db 1 0 S S1 0 g a Lss [I_e t=.0 ]*[I_ ID= .0]*lo Cp o 0 . . . . . db 2 0 S S.0] g aL s [I_e t=.0 ]*[I_ ID= 0*lo Cp os 0 . . . . . db 3 0 S S.0] g aL s [I_e t=.0 ]*[I_ ID= 0*lo Cp os 0 . . . . . db 3 0 S S1 0 g a Lss [I_e t=.0 ]*[I_ ID= .0]*lo Cp o 0 . . . . . db 1 0 Hz rd 1 0 g a Lss [I_e t=.0 ]*[I_ aa = .0]*loCp o 04 -2 .24 09 4.6 7 -.4 7 9 4 .60 2.8 0 -1 4 6 47 8.3 1 db 1 0 Hz rd 2 0 g a Lss [I_e t=.0 ]*[I_ aa = .0]*loCp o 69 -2 .51 34 3.4 5 -.7 5 9 6 .43 1.5 4 -1 2 6 98 5.3 2 39 Appendix D: Uncertainty analysis To assess the uncertainty in the projections based on the ARIMA models, 95% forecast confidence intervals were calculated. For each observation in the sample, we calculated the 95% forecast confidence intervals and used the upper and lower bounds for comparison with the observed GDP data; i.e., we calculate the differences to observed data based on these two values. Hence, there are two additional samples: one on the upper and one on the lower confidence region. The mean and median for these two samples are shown in table D1. Table D1: Mean and median of the sample differences using either the lower bound projections or the upper bound projections of the 95 percent forecast confidence intervals. t+1 t+2 t+3 t+4 t+5 low up low up low up low up low up Mean -11.09 6.97 -22.95 14.18 -37.94 20.95 -56.15 27.06 -80.47 33.02 Median -9.14 5.86 -19.10 13.10 -31.10 20.31 -44.95 27.79 -59.29 34.15 A large range can be found for the differences in the post-disaster years according to these 95 percent upper and lower confidence intervals of the projections; yet there is a clear trend to negative differences. The test for the lower and upper confidence bounds of the projections are however not useful for interpretational purposes due to the high standard errors associated with mean projections, leading either to a full rejection of the Null hypothesis or not. One remaining question regarding the ARIMA model projections and the validity of the results above is the influence of multiple disaster events. We tackle this issue by looking at a sub-sample within the full sample where 5 years before and 5 years after the disaster event no other major disaster (with losses higher than 1 percent of GDP) occurred. Table D2 again shows the mean and median as well as the sample size. 40 Table D2: Summary results for differences of real and projected GDP levels for sub-sample t+1 t+2 t+3 t+4 t+5 Mean -2.0558 -3.0284 -4.1281 -5.2683 -7.0973 Median -.8355 -1.4487 -2.0793 -3.5084 -5.9910 Std. Dev. 7.75618 12.15134 17.14314 23.01776 30.86930 Skewness -1.721 -1.764 -2.201 -3.200 -4.172 Observations 136 129 128 123 120 As in the full sample case, the average values are all negative, even with higher negative values. Statistical non-parametric Wilcoxon tests reveal that all of the average results are significantly lower than zero on the 95 percent confidence interval. 41