WPS5002
Policy Research Working Paper 5002
The Growth Aftermath of Natural Disasters
Thomas Fomby
Yuki Ikeda
Norman Loayza
The World Bank
Development Research Group
&
Global Facility for Disaster Reduction and Recovery
July 2009
Policy Research Working Paper 5002
Abstract
This paper provides a description of the macroeconomic can be beneficial when they are of moderate intensity,
aftermath of natural disasters. It traces the yearly response severe disasters never have positive effects. Third, not all
of gross domestic product growthboth aggregated and natural disasters are alike in terms of the growth response
disaggregated into its agricultural and nonagricultural they induce, and, perhaps surprisingly, some can entail
componentsto four types of natural disasters benefits regarding economic growth. Thus, droughts have
droughts, floods, earthquakes, and storms. The paper a negative effect on both agricultural and nonagricultural
uses a methodological approach based on pooling the growth. In contrast, floods tend to have a positive effect
experiences of various countries over time. It consists of on economic growth in both major sectors. Earthquakes
vector autoregressions in the presence of endogenous have a negative effect on agricultural growth but a
variables and exogenous shocks (VARX), applied to a positive one on nonagricultural growth. Storms tend to
panel of crosscountry and timeseries data. The analysis have a negative effect on gross domestic product growth
finds heterogeneous effects on a variety of dimensions. but the effect is shortlived and small. Future research
First, the effects of natural disasters are stronger, for should concentrate on exploring the mechanisms behind
better or worse, on developing than on rich countries. these heterogeneous impacts.
Second, while the impact of some natural disasters
This papera product of the Development Research Group and the Global Facility for Disaster Reduction and Recoveryis
part of a larger effort in the department to study the main sources of vulnerability and to disseminate the emerging findings
of the forthcoming joint World BankUN 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 yikeda1@
worldbank.org and nloayza@worldbank.org.
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
The Growth Aftermath of Natural Disasters*
Thomas Fomby Yuki Ikeda Norman Loayza
Southern Methodist University Georgetown University The World Bank
JEL Classification: O11, O40, Q54
Key Words: Natural disasters, economic growth, sectoral value added
*
We are grateful to Apurva Sanghi, S. Ramachandran, Jamele Rigolini, Eduardo Olaberría,
Claudio Raddatz and seminar participants at the World Bank for valuable comments, suggestions,
and advice. Tomoko Wada provided excellent research assistance. This paper was
commissioned by the Joint World Bank  UN Project on the Economics of Disaster Risk
Reduction. Partial funding of this work by the Global Facility for Disaster Reduction and
Recovery is gratefully acknowledged. 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.
1
I. Introduction
This paper provides a description of the macroeconomic aftermath of natural
disasters, specifically tracing the economic growth response in the wake of these events.
Its purpose is to contribute to the analysis of the path of adjustment and recovery by
tracing the yearly response of GDP growth both aggregated and disaggregated into its
agricultural and nonagricultural components to four types of natural disasters 
droughts, floods, earthquakes, and storms. As has been shown in recent papers (see, for
instance, Loayza, Olaberría, Rigolini, and Christiaensen 2009), the analysis by sector of
economic activity and by type of natural disaster is crucial to measure and interpret its
complex effects on the economy.
Apart from this disaggregated analysis, this paper has four other features that set it
apart. First, it traces the growth response in every year of and after the event. This focus
on the annual frequency is necessary to characterize the details of the adjustment path,
rather than only explaining its net permanent effect. For instance, it is conceivable that,
say, an earthquake has no longrun consequences on economic growth while having a
growth path of decline followed by recovery, whose characterization would be of interest
for the present analysis.
Second, the paper uses a methodological approach based on pooling the
experiences of various countries over time to arrive at mean responses of growth to
natural disasters. While losing country specificity, the methodology allows describing
basic patterns in a sensible and robust manner. The econometric methodology consists of
vector autoregressions in the presence of endogenous variables and exogenous shocks,
applied to panel, crosscountry and timeseries, data (for short, the methodology is
described as panel VARX). The full sample consists of 87 countries representing all
major regions of the world and 48 years covering the period 19602007.
Third, the paper considers the difference between advanced and developing
countries. Some key papers in this literature have noted that although the impact of
2
natural disasters is not the same across countries, it is not erratically heterogeneous either
(see Skidmore and Toya 2007, and Noy 2009, among others). Rather the impact follows
a more or less clear pattern, where poorer nations (in terms of economic, social, or
institutional wellbeing) tend to experience stronger effects from natural disasters. In
order to take this important insight into consideration, while preserving the panel nature
of the analysis, the paper conducts the econometric study not only on the full sample of
countries but also on two separate groups: poor or developing countries (62) and rich or
advanced countries (25).
Fourth, the paper expands the analysis by considering the potentially different
effect of severe vs. moderate natural disasters. Disasters of moderate magnitude are less
difficult to handle than severe ones. Thus, in the presence of moderate natural disasters,
governments and private organizations can deploy, redistribute, and relocate their
physical and human resources to compensate for the losses and reactivate the economy.
Under some conditions, moderate disasters may even bring about an increase in
economic growth by raising land productivity (in the case of floods) or inducing capital
transformation (in the case of earthquakes). However, if the disaster is of such magnitude
that it overwhelms public and private responses, its effect is likely to be more
detrimental.
At the end of this introduction, the paper offers a comprehensive review of the
new and interesting literature dealing with the macroeconomic impact of natural
disasters. Nevertheless, at this point, we highlight three papers that are most closely
connected with this study. The first is the paper by Loayza, Olaberría, Rigolini, and
Christiaensen (2009). In a sense, that paper and the present study can be regarded as
companion papers. Produced almost concurrently, the two studies take advantage of
disaggregation by type of disaster, sector of economic activity, and level of economic
development in order to enrich the analysis and elucidate the interpretation of results.
The focus of Loayza et al. (2009), however, is not on the path of adjustment and recovery
but on the net effects in the medium to long terms, for whose analysis it uses period
3
averages rather than annual data. Therefore, instead of employing a panelVARX
approach to trace yearly responses, Loayza et al. uses GMMSystem estimator (designed
for panels with large crosssection and short timeseries dimensions) to obtain average
net effects.
The second is the paper by Hochrainer and Mechler (2009). It assesses the
macroeconomic consequences of natural disasters by comparing the gap between a
counterfactual GDP and observed GDP. The counter factual is constructed using the
projection of past GDP under the assumption of a nodisaster scenario. The paper finds
that natural disasters on average lead to negative effects on GDP. Although Hochrainer
and Mechler's paper differs from ours regarding the methodological approach, it is
similar on the importance of separating natural disasters according to type and estimating
their effects independently. Thus, it finds that typical (or median) storms, earthquakes,
and droughts have a negative impact on GDP, while floods show a positive impact one.
As shown below, these results are consistent with most of our findings.
The third paper is by Raddatz (2009). In this case, the methodological approach
seems to be similar to ours regarding the use of an autoregressive model applied to panel
data to assess the macroeconomic consequences of natural disasters. There are, however,
some important differences. Raddatz concentrates on the effects of disasters on
aggregate GDP growth, while we also analyze the effects on agricultural and non
agricultural sectors, finding differing effects on each of these sectors of the economy.
Although Raddatz also recognizes the importance of disaggregating by type of disaster,
he emphasizes a way of grouping them that, while popular in the literature, may mask
contrasting effects. Such is the case of "climatic" natural disasters, which group together
floods and droughts. We separate them and find that they have radically different
impacts on economic growth. Another difference between Raddatz' analysis and ours is
that we differentiate between relatively moderate disasters and extremely severe disasters
to capture possible nonmonotonic effects. On the other hand, Raddatz' contribution
extends in dimensions that we do not explore. He finds that neither the inflow of foreign
4
aid nor the initial level of indebtedness of the country significantly affects the growth
impact of natural disasters. On the other hand, he finds that the level of economic
development does influence the impact of natural disasters. It is this dimension of the
heterogeneity across countries that we emphasize in this paper.
Before proceeding with the literature review, we now provide the outline of the
paper. Section II presents the description of the data, including details on the sample
regarding countries, periods, and frequency of observations; and on the variables used in
the analysis concerning definitions, sources, and summary statistics, with special
attention to the measures of moderate and severe natural disasters. Section III introduces
the econometric methodology, including an exposition of the VARX method, and two
important specification tests dealing with exogeneity assumptions and lag structures.
Section IV presents the basic results, discussing and contrasting the effects of droughts,
floods, earthquakes, and storms, focusing mostly on the sample of developing countries.
Section V offers some concluding remarks.
Literature review
Economic research in this field is still in an early phase of development. In
general, the results on the macroeconomic impact of natural disasters seem to be
ambiguous. A close examination in recent studies further demonstrates that these effects
may depend on economic, social, and institutional conditions, as well as on the type of
natural disaster and sector of the economy.
Rasmussen (2004) assessed the impacts of natural disaster incidences using a
crosscountry sample for the period 1970 through 2002. The data were obtained from the
EMDAT database of the Centre for Research on the Epidemiology of Disasters (CRED),
which is the major source of data on natural disasters used in most studies. According to
CRED, a natural disaster is defined as a situation or event which overwhelms local
capacity, necessitating a request for external assistance. The database consists of disaster
events which fulfill at least one of the following criteria: ten or more people reported
5
killed; 100 or more people reported affected; declaration of a state of emergency; or call
for international assistance. These disasters include hydrometeorological disasters such
as floods, wave surges, storms, droughts, landslides and avalanches; geophysical disasters
such as earthquakes, tsunamis and volcanic eruptions; and biological disasters such as
epidemics and insect infestations. To provide a comprehensive picture, he compared the
frequencies and impacts of disasters across countries by employing four measures,
including the number of events divided by land area, the number of events divided by
population, the number of affected persons divided by total population, and damage
divided by GDP. He found that developing countries, particularly small island states in
the Eastern Caribbean Currency Union (ECCU), face higher relative costs than advanced
countries when measured in terms of the number of person affected and the value of the
damage. The author also assessed the shortterm impacts of 12 major disasters occurred
in the ECCU and observed its negative effects on economic output as well as external and
fiscal balances. The analysis showed that natural disasters led to a median reduction of
2.2% in the sameyear real GDP growth. Moreover, a median increase in the current
account deficit amounted to 10.8% of GDP in the disaster year. The median public debt
was also observed to increase by a cumulative 6.5% over three years following disaster
events.
Closely related to this approach, Heger, Julca, and Paddison (2008) investigated
the macroeconomic impact of natural disasters with the specific focus on the Caribbean
region. Their analysis was based on the annual dataset that included sixteen Caribbean
states over the 19702006 period, drawn from the EMDAT database. The authors first
selected proxies for natural disasters through a simple OLS estimation. They identified
the frequency of disasters, the estimated costs of disasters, and the number of total
affected as the major explanatory variables for different macroeconomic outcomes. With
those variables in the corresponding OLS regression analysis, the results illustrated that
natural disasters negatively impact growth, fiscal balance, and external balance. These
results coincide with those of Rasmussen's presented above. Another significant finding
6
was that when a country relies on export or import specialization, larger damages occur
in response to disasters. The authors conclude that diversification of the economy can
help mitigate the effects of natural disasters.
Using a panel vector autoregression model, Raddatz (2007) examined the
dynamic impacts of external shocks, including natural disasters, on the volatility of
output. Focusing on lowincome countries, he uses a sample of 40 countries over the
period from 1965 to 1997. For the disaster measurement, the author employs the annual
data on the number of disastrous events, compiled from the EMDAT database. The
analysis indicated that the effects of external shocks in general on per capita GDP are
modest and contribute to only a small portion of its volatility, leading the author to
conclude that output volatility is largely determined by internal causes rather than
external shocks. However, shocks derived from some natural disasters did appear to have
and important effect. In particular, it was observed that climatic disasters lead to a
decrease of 2% in real per capita GDP one year after the disaster, while humanitarian
disasters reduce it by 4%. On the other hand, geological disasters were found to be
insignificant in terms of contribution to the variance of output.
A recent study by Noy (2009) investigated the shortrun macroeconomic response
to natural disasters using a panel dataset over the period 19702003. Taken from the EM
DAT database, three measures of disaster damages were employed: the number of people
killed; the number of people affected; and the amount of direct damage. In light of
potential factors that can influence the disaster impacts, the author took into account
differences in population size, size of economy, and timing of incidences. The regression
of annual GDP growth rate on the disaster measure and other control variables revealed
that the impact of natural disaster is statistically significant when it is measured as the
amount of property damage incurred. As other studies suggest, it was also found that the
macroeconomic costs were much higher in developing countries than in developed
countries. Noy further analyzed the determinants of these negative macroeconomic
effects following disasters. He concluded that higher level of literacy, better institutional
7
quality, higher per capita income, higher government spending, and more open
economies along with better financial conditions are likely to contribute to countries'
macroeconomic performance after natural disasters.
On a similar line, several studies have documented that economic development
plays an important role in mitigating a countries' vulnerability to catastrophic incidences.
Skidmore and Toya (2007) investigated the effects of the level of development on
disaster impacts, using a dataset of natural disasters incurred in 151 countries over the
period from 1960 to 2003. The analysis included two patterns of dataset obtained from
the EMDAT. One used the number of killed to assess the disaster impacts, while the
other considered economic damages. The OLS regression analysis demonstrated that
human and economic damages from natural disasters are generally reduced along with
economic development. In particular, the results showed that deaths and damages were
lower in countries with higher level of educational attainment, greater degree of
openness, more developed financial sector, and smaller governments. The authors
suggest that policymakers could consider further efforts in developing economic and
social infrastructures, which can contribute to decreasing natural hazards.
Taking a different approach in exploring the impacts of capital and labor losses on
shortterm growth, Caselli and Malhotra (2004) tested the empirical validity of the
predictions of the Solow growth theory. The theory suggests that a decline in the capital
labor ratio resulting from a natural disaster would lead to an increase in the country's
growth rate, while an increase in the capitallabor ratio would curtail it. In their empirical
analysis, the total number of people killed, injured, and affected by disasters were used to
calculate the percentage loss in the labor force, while the immediate damage as a
percentage of GDP was used as a proxy for the loss in capital stock. The data were
compiled from the EMDAT database for a sample of 172 countries for the period
between 1975 and 1996. Using the real per capita GDP growth rate in the disaster year to
estimate the Solow model, their empirical analysis found that sudden losses of capital and
labor did not bring about a change in the economic growth as expected by the Solow
8
growth model. The results, however, remain questionable given the proxies used to
measure capital and labor destructions and the timing of the growth response.
Jaramillo (2007) presented a comprehensive analysis of the link between natural
disasters and economic growth both in the shortrun and longrun using a panel dataset of
113 countries over the period 19601996. However, disasters that develop through
extended periods, such as droughts and famines, as well as insect infestations and
epidemics are excluded from the analysis. The type of disasters examined by Jaramillo
include earthquakes, floods, wild fires, wind storms, waves and surges, extreme
temperatures, volcano episodes, and slides. Taking country and year fixed effects into
account and controlling for trade openness and foreign aid, the author examined the
shortrun effects of disasters on economic growth, followed by an analysis of the long
run effects. For the shortrun, Jaramillo assessed the impacts on GDP growth in the
disaster year and the following year, whereas for the longrun, he tested for the
cumulative disaster effects over the period 19601996 on the GDP per capital level in
1996. The regression results indicated that short and longterm effects of disasters are
determined by countries' income level, population, and the type of disaster. On the
whole, it was found that the effects of disasters on GDP growth rate varied from 0.9%
decrease to 0.6% increase depending on the disaster type.
Focusing specifically on the longterm macroeconomic impacts of natural
disasters, the first comprehensive empirical research was done by Skidmore and Toya
(2002). In their crosscountry analysis, the authors use average per capita GDP growth
over the period 19601990 and the total number of significant disaster events observed in
respective countries during the same period. The disasters studied cover climatic and
geologic disasters. The results revealed that climatic disasters have positive effects on
the longrun economic growth as they induce higher capital accumulation and total factor
productivity than before. It is argued that total factor productivity is the predominant
factor in promoting growth after disasters. By contrast, geologic disasters were observed
9
to affect growth negatively as it deteriorates physical capital and decreases human capital
due to the initial loss of life.
Following Skidmore and Toya's findings, Cuaresma, Hlouskova, and Obersteiner
(2008) examined the longrun effects of natural disasters by analyzing the direct
relationship between foreign technology absorption and disaster incidences. Earlier
studies argued that disasters can provide countries with opportunities to renew
technologies, thereby promoting longrun growth. The authors assess this argument by
using gravity model to analyze foreign knowledge spillovers between the G5 countries
and a sample of 49 developing countries. According to the regression results, natural
catastrophic risk negatively affected knowledge transfers from the industrialized to
developing countries. The authors further found that countries with higher levels of
development are more likely to be better off than countries with lower levels of
development through capital upgrading following natural disasters.
Hallegatte and Ghil (2007) added business cycle framework to the study of
disaster impacts. They analyzed the effects of exogenous shocks, including natural
disasters and stochastic productivity stocks, on economic behavior. Employing a Non
Equilibrium Dynamic model with endogenous business cycles, they found that total GDP
losses resulting from natural disasters are higher when occurring during expansions than
during recessions. The reason is that because preexisting disequilibria are widened by
exogenous shocks in the former phase, whereas the shocks are mitigated by the existence
of unused resources in the latter case. The paper drew the conclusion that the phase of the
business cycle during which a disaster occurs affects the degree of the macroeconomic
response.
As discussed above, while some studies found common patters in the
determinants of a country's vulnerability to catastrophic events, researchers have not
come to a consensus on the impacts of natural disasters on economic growth. This paper
attempts to help disentangle this ambiguity by using a bettergrounded econometric
10
methodology and a conceptually driven disaggregation by type of natural disaster and
sector of economic activity.
II. Data
A. Periods, frequency, samples (groups of countries)
To perform our estimations, we use pooled crosscountry and annual timeseries
data covering 87 countries over the period 19602007. The panel is unbalanced, with
some countries having more observations than others. We refer to the data as "all
countries". Then we split the data into two groups: "rich countries" and "developing
countries". We classify 25 Arab and OECD countries into the first group and the other
62 counties into the second group. Table II.1 gives the list of countries of these groups.
B. Variables, definitions, sources
The main variables used in the paper are divided in three groups. First, to study
the impact of natural disasters on the economy, we define three types of growth variables.
The first is the growth rate of real per capita Gross Domestic Product (GDP). The others
are the growth rates of real per capita value added in the two major sectors of the
economy, the agricultural sector and the nonagricultural sector. All of them are
measured as the log difference of per capita output (in 2000 US dollars), where per capita
output is obtained by dividing the value added of each sector by the total population.
Second, as a variable which represents the role of external conditions that may
affect the growth performance across countries, we use shocks to the Terms of Trade
(TOT). Terms of trade shocks are measured by the growth rate of the terms of trade
(export prices relative to import prices). The idea is to capture shifts in the demand for a
country's exports. Data for all the above variables were obtained from the World Bank
(WDI, 2008).
11
The last set of variables represents the role of natural disasters on the growth
performance across countries. Data for natural disasters were obtained from the
Emergency Disasters Database (EMDAT) maintained by the Center for Research on the
Epidemiology of Disasters (CRED). EMDAT provides the number of casualties (people
confirmed dead, reported missing, and presumed dead), the number of people injured,
and the number of people affected. People affected are those requiring immediate
assistance during a period of emergency. Also, people reported injured or homeless are
aggregated with those affected to produce the total number of people affected (we refer to
this number as "total affected"). Throughout the paper, we assume that natural disaster
variables are (block) exogenous with respect to the growth variables and shocks to the
terms of trade. 1
C. Moderate and severe natural disasters
As mentioned above, we divide natural disasters into four categories: droughts,
floods, earthquakes, and storms. The measure of intensity of natural disasters, NDtk,i , is
given by:
droughtt ,i if k = 1,
flood t ,i if k = 2,
NDtk,i = (1)
earthquaket ,i if k = 3,
stormt ,i if k = 4,
where
killed tk,i , j + 0.3 * total affected tk,i , j
intensitytk,i , j = , (2)
populationt ,i
= 1, if intensitytk,i , j > 0.0001,
NDtk,i , j (3)
= 0, otherwise,
1
For the exogeneity of natural disaster variables, see section III.b.i.
12
J
NDtk,i = NDtk,i , j , (4)
j =1
and J describes the total number of typek events (k = 1, 2, 3, and 4 correspond to
drought, flood, earthquake, and storm, respectively) that took place in country i during
year t. The following steps describe how to create the intensity measure. First, for each
event of typek disaster, we create a variable intensitytk,i , j measuring the magnitude of the
event relative to the size of the economy, that is, the sum of the number of casualties
(killed ) and 30% of the total number of people affected (total affected ) divided by
k
t ,i , j
k
t ,i , j
the population (equation (2)) 2. Then we construct a dummy variable NDtk,i , j which takes
the value of 1 if intensitytk,i , j is greater than 0.01% (equation (3)). Finally, for each type
of disaster, the respective dummy variables NDtk,i , j , j = 1, ..., J are summed up to obtain
the indicator value NDtk,i to assess the total magnitude of typek disasters in country i
during year t (equation (4)).
Many practitioners point out that the impact of moderate disasters and extremely
severe disasters on the economic performance differ, not only in their magnitude, but also
in their dynamic characteristics. To capture the particular effects of severe disasters, we
construct a second measure of intensity, sevNDtk,i , as follows:
sev. drought t ,i if k = 1,
sev. flood t ,i if k = 2,
sevNDtk,i = (5)
sev. earthquaket ,i if k = 3,
sev. stormt ,i if k = 4,
where
2
This intensity measure is similar to the one established by the International Monetary Fund
(IMF, 2003), and used by Becker and Mauro (2006).
13
killed tk,i , j + 0.3 * total affected tk,i , j
intensity k
t ,i , j = , (6)
populationt ,i
= 1, if intensity tk,i , j > 0.01,
sevNDtk,i , j (7)
= 0, otherwise,
J
sevNDtk,i = sevNDtk,i , j . (8)
j =1
Here, for the dummy variable for the intensity of individual severe disaster, sevNDtk,i , j ,
we set the threshold at 1% of the population, while we applied the threshold of 0.01% for
general or moderate disasters. In section IV, we show the results of two types of
estimation, in which (i) only moderate disaster variables are included (the basic model),
and (ii) both moderate and severe disaster variables are included.
D. Summary statistics
Regarding the growth variables introduced in the early part of this section, a few
observations deserve some comments. First, we should point out that the growth
performance of the different sectors varies widely in each country. As shown in Table
II.2, during the period 19602007, the nonagricultural sector has had much higher
average growth rate (1.7% in developing countries, 2.1% in rich countries) than the
agricultural sector (0.31% in developing countries, 0.93% in rich countries). Also, Table
II.3 shows that the correlation between the growth rates of nonagricultural sector with
the agricultural sector is quite low (0.1095 in developing countries and 0.0173 in rich
countries). The considerable disparities among the growth performances provide some
grounds to suspect that natural disasters could have had diverse effects on the different
sectors of the economy.
14
III. Methodology
A. Econometric method
The econometric model we adopt here is a fixedeffects Panel VARX model,
namely,
y t ,i = i + 1y t 1,i + 2 y t 2,i + 3 y t 3,i + 0 x ti + 1x t 1,i + 2 x t 2,i + t ,i , (9)
where the country index is i = 1, 2, ..., M and the time index for each country is t = 1, 2,
..., Ti. The fixed effect for each country is represented by i . Hereafter, the total
M
number of observations for all countries in the panel is denoted by T = Ti . The
i =1
endogenous variables vector is denoted by the (2 × 1) vector y ti while the (4 × 1)
exogenous variables vector x ti represents the occurrences at time t of the disasters,
respectively, drought, flood, earthquake, and storm. In equation (9) we assume the
homogenous error structure E ( t ,i ,i ) = for all t and i where t ,i is the (2 × 1) vector of
t
errors of the system. Furthermore, we assume independence of the errors within
equations, E ( t ,i , j ) = 0 , i j , and across equations, E ( t ,i s , j ) = 0 , for any t and s
t
where i j .
Model (9) is applied to three different groups of countries: All of the countries,
Developing countries, and Developed Countries. We choose to estimate Model (9) by
OLS to the demeaned series resulting in the socalled withinfixedeffects estimator. As
pointed out by Nickell (1981), given that Model (9) is dynamic, if T is small and fixed,
such an estimator is inconsistent as the number of countries, M, goes to infinity.
However, in our case we consider the number of countries fixed and since in each
grouping of the countries considered here the number of available observations, T, is at
15
least 778. 3 In this case, the bias of the withinfixedeffects estimator should be
negligible. Hereafter, we refer to the withinfixedeffects estimator simply as the OLS
estimator, with the coefficient estimates being denoted by i , i = 1, 2, 3, and i , i = 0,
^ ^
1, 2.
Model (9) can be written more compactly as
(I  1 L  2 L2  3 L3 )y t ,i = i + ( 0 + 1 L + 2 L2 )x t ,i + t ,i ,
or
( L)y t ,i = i + ( L) t ,i , (9')
where L denotes the usual lag operator. To insure that (9') produces a steady state, we
require that all of the roots of the determinant equation (I  1 L  2 L2  3 L3 ) = 0 lie
outside of the unit circle. Inverting (9') produces the multiplier form of Model (9):
y t ,i = ( L) 1 ( L)x t ,i + ( L) 1 t ,i . (10)
The mean responses from the occurrences of natural disasters are therefore captured by
the lag polynomial
( L) = ( L) 1 ( L). (11)
It follows that the coefficients of the lag polynomial (L) can be obtained by matching
the coefficients in the expression
( L)( L) = ( L). (12)
This gives rise to the solutions
3
The number of observations available in the sample of rich countries, with nonagricultural
growth rate as an endogenous variable.
16
0 = 0 (13)
1 = 1 + 0 1 (14)
2 = 2 + 1 2 + 0 2 (15)
3 = 2 1 + 1 2 + 0 3 (16)
s = s 11 + s  2 2 + s 3 3 for s 4 . (17)
Now let
= [1 2 3 0 1 2 ] (18)
denote the coefficient matrix of (9). The coefficient matrix i is 2 × 2 for i = 1, 2, 3 and
i is 2 × 4 for i = 0, 1, 2. Therefore, the coefficient matrix is (2 × (6 + 12)) = (2 ×
18). Let = vec( ) . Then is a (2(18) × 1) vector with the first 18 elements being the
autoregressive and current and lagged natural disaster coefficients from the first equation
and the second 18 elements being the corresponding coefficients from the second
equation.
Let s = vec( s ) denote the (2(4) × 1) vector of the speriod delay mean
responses due to natural disasters. The first 4 elements represent the speriod delay mean
responses of the first endogenous variable to the natural disasters while the second 4
elements represent the speriod delay mean responses of the second endogenous variable
to the natural disasters. Moreover, let denote the vector of the OLS estimates of
^
equation (9). Then it can be shown under fairly general conditions that
T (  ) N (0, ( Q 1 ))
^ (19)
17
where = E ( t ,i ,i ) is the variancecovariance matrix of the error terms of (9) and
t
Q = plim( X' X / T ) where X is a (T × 18) design matrix of the form
X1
'
'
X
X= 2 (20)
X'
T
where X t' = (y t' 1y t' 2 y t' 3 x t' x t' 1x t' 2 ).
In implementing the result of equation (19), we need consistent estimates of
and Q . These estimates are obtained as follows:
T
^ 1
= t ^ ^t (21)
T t =1
and
Q = X' X / T .
^ (22)
Let s ( ) denote the estimated speriod delay mean responses to the exogenous
^ ^
vector x t where the dependence of these estimates on the coefficient estimates is
^
made explicit. One way to obtain standard errors for these estimates is to use Monte
Carlo methods. First, randomly draw a (36 × 1) vector from the distribution
1 ^
N ( ,
^ ( Q 1 )). Denote this vector by (1) . Calculate s ( (1) ). Repeat this process
^ ^
T
for, say, a total of 10,000 times. Then to get, for example, the 90% confidence interval
for the first element of s , say s1 , we need the 5th percentile, s1 , and the 95th
percentile, s1 , from the simulated values of s1 resulting in the 90% confidence
18
interval for s1 , namely, ( s1 , s1 ). The confidence intervals for the remaining
elements of s are similarly constructed.
B. Diagnostic tests
i. Individual and panel unit root tests
Before we can proceed to build a VARMAX panel model for analyzing the
effects of natural disasters on various endogenous variables, we need to determine the
stationary forms of the endogenous variables we are going to be using in our analysis. In
this study we chose as the endogenous variables of interest (1) the log of real GDP per
capita, (2) the log of real agricultural value added per capita, (3) the log of real non
agricultural value added per capita, and (4) the log of terms of trade. We chose to use the
log transformation of the variables because of the variance stabilizing characteristics of
the transformation and the fact that, if a unit root is contained in the logged variables,
then differencing them yields a very straightforward interpretation of the differenced
data, namely percentage change.
We proceeded to pursue unit root testing in these variables in two ways: series
byseries unit root tests and panel unit root testing with individual country effects as in
the Levin, Lin, and Chu (2002) and ImPesaranShin (2003) panel unit root testing
frameworks. These unit root tests are, of course, dependent on the specification of the
deterministic parts of the unit root test equations. That is, does the data contain a trend or
not? Is the data without trend but has a nonzero mean as compared to a zero mean? To
obtain consistent statistical hypothesis test results one must properly specify the
deterministic parts of the data under the alternative hypothesis of stationarity. In this vein
we tested the significance of the trend in the above four series by testing the significance
of the intercept in the following AR(2) equation of the variable in question, countryby
country:
zt = + 1zt 1 + 2 zt 1 + t . (23)
19
In equation (23) z t represents a particular country's variable in question and
represents the first differencing operator. We specified a secondorder autoregression to
ensure that the residuals of the equation would be white noise thus implying that OLS t
statistics involving the intercept would be appropriate for testing for the presence or
absence of trend. In the case that the null hypothesis H 0 : = 0 was supported, we
concluded that the data does not have a trend in it. On the other hand, if the alternative
hypothesis of H 1 : 0 was supported, we concluded that the data has trend in it. With
respect to the log of real GDP per capita and log of real nonagricultural value added per
capita, the preponderance of tests indicate trend is present (52 of 87 null hypotheses
rejected for the former and 47 of 87 null hypotheses rejected for the latter). In contrast,
for the log of real agricultural value added per capita and the log of terms of trade, the
preponderance of tests indicated that trend is absent (15 of 87 null hypotheses rejected for
the former and 1 of 87 null hypotheses rejected for the latter). Thus, for the production
run of unit root tests, we choose to treat all of the log of real GDP per capita and log of
real nonagricultural value added per capita series as having trends in them while the log
of real agricultural value added per capita and log of terms of trade series had no trend in
them but nonzero means. 4
As a result of these trend tests we chose to use an intercept and deterministic trend
in testing for unit roots countrybycountry in the log of real GDP per capita and log of
real nonagricultural value added per capita series in the augmented DickeyFuller and
PhillipsPerron unit root test equations while for the log of real agricultural value added
per capita and the log of terms of trade, we chose to use only an intercept in the
augmented DickeyFuller and PhillipsPerron unit root test equations. Of course, when
testing for the sufficiency of the first difference in producing stationarity in a series, we
checked the first difference of the series for unit roots using the appropriate deterministic
terms implied by differencing. In particular, when testing for the stationarity of the first
4
Detailed test results are available from the authors upon request.
20
difference of the log of real GDP per capita and the first difference of the log of real non
agricultural value added per capita we included only an intercept in the test equation. In
contrast, when testing for the stationarity of the first difference of the log of real
agricultural value added per capita and the log of terms of trade we set the intercept to
zero in the test equation.
In contrast to the countrybycountry unit root tests, the panel unit root tests of
specific time series assume as the null hypothesis that a unit root exists for all of the
countries, with country distinction coming only from having separate deterministic terms
for each country (i.e. different intercept effects or different intercept effects as well as
different trend effects for each country). The difference between the Levin, Lin, and Chu
(2002) and ImPesaranShin (2003) panel unit root tests resides in the form of the
alternative hypotheses assumed by the tests. In the Levin, Lin, and Chu test the
alternative hypothesis takes the form of a common stationary firstorder autoregressive
coefficient across all of the countries whereas the ImPesaranShin test assumes all of the
firstorder autoregressive coefficients are stationary but that they can possibly take on
different stationary values. Both tests are, of course, allornone tests in the sense that
test results imply that either (1) all of the countries' given series have unit roots in them
or (2) all of the countries' series are stationary of the same degree (as in the Levin, Lin,
and Chu test) or different degrees (as in the ImPesaranShin) test. The benefit of the
panel unit root tests are that, in the case of short time series in the panel, the power of the
unit root tests are increased when one or more of the panel series are nonstationary as
compared with countrybycountry unit root tests.
The results of the above unit root tests applied to the four series are summarized
in Table III.1. 5 The left half of the table pertains to unit root tests of the nontrending
series (log of real agricultural value added per capita and log of terms of trade) while the
right half of the table pertains to the unit root tests of the trending series (log of real GDP
5
All of the results reported in Table III.1 were produced by EViews 6.0.
21
per capita and log of real nonagricultural value added per capita). In addition, the top
half of the table (Section A) reports the unit root tests of the levels while the bottom half
of the table (Section B) reports the unit root tests of the first differenced data.
Furthermore, in each section the results of four unit root tests are reported, the first two
tests being countrybycountry unit root tests while the latter two tests are the panel unit
root tests. 6
The results reported in Table III.1 are summarized as follows:
· Log of real agricultural value added per capita. The preponderance of the
individual unit root tests indicates the presence of a unit root. The panel unit
root tests likewise indicate the presence of unit roots. After first differencing
the series seems to be stationary.
· Log of Terms of Trade. The results for this series are similar to those of the
previous nontrending series except for the significance of the LevinLinChu
panel test where the pvalue is less than 5% in the levels of the data. In
contrast the ImPesaranShin panel test (with a flexible alternative
hypothesis) indicates a unit root at the 10% level. Evidently, the log of terms
of trade is "near" stationary. Despite this "split decision" on the existence of
a unit root we decided to treat this series as having a unit root and to model its
differences as being stationary.
· Log of Real GDP per capita. The preponderance of the individual unit root
tests indicates the presence of a unit root. The panel unit root tests likewise
indicate the presence of unit roots. After first differencing the series seems to
be stationary.
6
Note in the case of the first difference of the nontrending data, the ImPeseranShin test is not
reported as EViews does not accommodate the zero mean case.
22
· Log of real nonagricultural value added per capita. The same conclusions
hold that hold for the log of real GDP per capita. Unit roots are present and
the first differenced series appears to be stationary.
In summary, the test results of Table III.1 indicate that, when building meaningful
VARMAX panel models to examine the impacts of various natural disasters on
developing countries' GDP and agricultural, nonagricultural value added, and terms of
trade, the growth rate forms of these endogenous variables should be used.
ii. Block exogeneity tests
The VARX model presented in the previous subsection is dependent on the
assumption of exogeneity of the natural disaster variables. While all variables in the
model are assumed to be endogenous in a simple VAR model, a VARX model allows
some of the variables to be exogenous. In this section, we present the hypothesis testing
method about the exogeneity of the disaster variables and its results.
Here, we are interested in the exogeneity of the disaster variables as a group, with
respect to shocks to the terms of trade and one of the growth variables (GDP growth,
agricultural growth, or nonagricultural growth). Without assuming the exogeneity of the
disaster variables, we can rewrite Model (9) as a simple VAR of order p as follows:
p p
x t ,i = 1 + A h x t h ,i + B h y t h ,i + u t ,i ,
i
h =1 h =1
p p
(24)
y t ,i = + C h x t h ,i + D h y t h ,i + v t ,i ,
2
i
h =1 h =1
where
23
droughtt ,i
flood
x t ,i = t ,i ,
earthquaket ,i
(25)
stormt ,i
TOTt ,i
y t ,i = ,
GDP / Agr. / Non  agr. growtht ,i
and 1 and i2 are the fixed effects for country i. In equation (24) we assume the
i
homogenous error structures: E (u t ,i u,i ) = 11 , E (u t ,i v,i ) = 12 , E ( v t ,i u,i ) = 21 , and
t t t
E ( v t ,i v,i ) = 22 for all t and i, where u t ,i and v t ,i are the errors of the system. The group
t
of variables represented by x is said to be blockexogenous with respect to the variables
in y if B h = 0 for h = 1, ..., p.
To check the exogeneity of the disaster variables, we can perform a likelihood
ratio test with the null hypothesis, H 0 : B h = 0, h = 1, ..., p. This test can be done with
running OLS regressions of each of the disaster variables on p lags of all of them and p
^
lags of all of the elements of y. Let denote u t ,i the (4 × 1) vector of sample residuals
from these regressions and 11 their variancecovariance matrix.
^ Next, run OLS
regressions of each of the disaster variables only on p lags of them, without lagged
^
variables of y. Let denote u t (0) the (4 × 1) vector of sample residuals from the second
set of regressions and 11 (0) their variancecovariance matrix. If
^
T * {log  11 (0)   log  11 },
^ ^ (26)
where T is the number of observations, is greater than the critical value for a 2 (4 × 2 p )
variable, then the null hypothesis is rejected and the conclusion is that some of the
disaster variables are helpful in forecasting y, i.e., the disaster variables are not block
exogenous with respect to the variables in y.
24
Table III.1 displays the results of the block exogeneity test. As it shows, the null
hypothesis is not rejected in any of three samples, with any of growth variables, and with
p = 1, 2, 3, at 5% of statistical significance. At 10% of significance, the null hypothesis
is rejected only in 2 cases out of 27 cases, when we use the sample of rich countries and
include the agricultural growth in y, with p = 1 and 3. These results strongly suggest the
use of VARX model, over the use of VAR model in which all variables are treated as
endogenous.
iii. Lag structure
Before estimating the panel VARX model, we need one crucial piece of
information. That is the number of lags to include for each variable in the model. To
identify the lag structure, some statistical criteria can be used.
A wellknown criterion is Akaike's information criterion (AIC) (Akaike (1973)),
given by
l K
AIC = 2 , (27)
T
and an alternative is Schwarz's Bayesian information criterion (SBC) (Schwarz (1978)),
which is given by
 2l + log(T ) × K
SBC = , (28)
T
where T is the number of observations, K is the number of parameters in the model 7,
1 ^^
l = T × 1 + log(2 ) + log det , (29)
2 T
7
In our basic model, K = 2(2p + 4(q + 1)).
25
and is the (T × 2) matrix of the error terms of Model (9). Models with a lower AIC or
^
SBC are preferred. Both criteria add a penalty that increases with the number of
regressors or lags.
Table III.2 shows the AIC and SBC statistics for the models with three different
endogenous variables (GDP growth, agricultural growth, and nonagricultural growth)
and three different groups of countries (all countries, developing countries, and rich
countries). p and q represent the number of lags for the endogenous variables and the
exogenous variables, respectively. In most cases, the results suggest either the models
with p = q = 1, or the models with p = q = 2. Clearly, SBC tends to favor more
parsimonious models than AIC, because the penalty for increasing the number of lags is
larger for SBC.
Based on the information criteria values, we selected the lag length 2 as our basic
lag structure. From a statistical point of view, there is little to choose between the lag
length 1 and 2, since we have the mixed results from the information criteria. The latter
one, however, provides much richer dynamics of the mean responses of the endogenous
variables to exogenous shocks. As the goal of this paper is to study the dynamic effects
of natural disasters, this is reason enough to select the lag length 2. We apply this lag
structure to all of our models homogenously to simplify the interpretation.
IV. Results
We now report and discuss the main results on the growth consequences of
natural disasters. We organize the presentation by type of disaster droughts, floods,
earthquakes, and storms. For each of them, we consider the effects on GDP per capita
growth and its major components, agricultural and nonagricultural per capita value
added growth. We first estimate these effects using the sample of all countries (Table
IV.1). Then, to gain further insight on the development angle of the issue, we divide the
26
sample into developing countries (Table IV.2) and advanced countries (Table IV.3).
Focusing on the sample of developing countries (for which the effects are stronger), we
then consider the differing impact of moderate and severe natural disasters (Table IV.4).
The estimation of the VARX model renders a wealth of results, from which we
choose those that are most pertinent to the main objective of the paper. Since we are
interested in tracing out the dynamic path of adjustment in the aftermath of the disaster,
the most relevant results are the mean response of growth to a given natural disaster for
each year after the event. Since the effects are small and nonsignificant a few years after
the event, we only report the mean responses for years 0, 1, 2, and 3 of the event (where
year 0 is when the disaster occurred). We indicate whether these responses are
statistically greater or smaller than zero, according to the Monte Carlo simulations
explained in the methodological section of the paper. Furthermore, we report the
cumulative effect of the event, which corresponds to the sum of mean responses for the 4
years after the event. We organize and present these results in several tables, as indicated
above. In addition, we present a graphical representation of the mean responses for each
natural disaster for the sample of developing countries, together with their corresponding
confidence bands indicating 10% tails of the distribution of effects (Figures IV.14). The
confidence bands are obtained through the Monte Carlo simulations mentioned above.
The majority of the discussion refers to the results obtained with the sample of
developing countries. For comparison purposes, we also discuss the results from the
sample of all countries (of which developing countries represent nearly 80%) and the
sample of advanced countries.
Finally, we offer some robustness analysis regarding the lag structure of the
VARX model (Appendix Tables A.1 and A.2, and Figures A.1A.4). In particular, we
use a more restrictive lag structure, p = q = 1, which, as mentioned in the previous section
also received support from the information criteria tests. The results are broadly similar
to those using the preferred longer lag structure. The main difference is that when only
27
one lag is allowed, the mean responses corresponding to later years are smaller and less
significant.
A. Droughts
Droughts have an overall negative effect on GDP growth. As expected, the effect
is stronger for agricultural growth, but it is also negative for nonagricultural activities.
For agricultural growth, the negative effect of droughts is larger on the year of the event.
There is a significant recovery on the following year, but the cumulative effect remains
significantly negative. For nonagricultural growth, the negative impact is felt on the
year of the drought and also a couple of years afterwards, indicating the presence of
delayed effects. In the sample of developing countries, the cumulative negative response
to droughts is 1.7 percentage points (pp) for GDP growth and 1.6 pp for agricultural
growth.
The pattern of results just described applies to the samples of all countries and of
developing countries. For advanced countries, there is also a negative response on the
year of the drought but it only applies to agricultural growth. Furthermore, in the
subsequent years agricultural growth recovers so substantially that the cumulative effect
of droughts for advanced countries is essentially zero.
Turning to the analysis of severe vs. moderate cases, the strongest negative effects
(in size and statistical significance) come from severe droughts. The year of the event,
severe droughts have twice the negative impact on GDP growth than moderate droughts.
Furthermore, severe droughts induce larger volatility of growth, which means that they
produce a larger drop the year of the event and a stronger recovery in the following year.
In the case of GDP growth, this recovery is sufficiently strong so that the cumulative
effect of severe droughts is comparable to that of moderate droughts (1.52.0 pp).
However, in the case of agricultural growth, the recovery is insufficient and, then, the
negative cumulative impact of severe droughts (2.0 pp) is twice as large as that of
28
moderate ones. For nonagricultural growth, severe droughts also have the strongest
impacts and the most volatile ones.
B. Floods
In contrast to droughts, floods tend to have a positive effect on economic growth.
The mean response of GDP growth is positive and significant in years 2 and 3 after the
event. This coincides with the mean response of nonagricultural growth, which indicates
that the positive impact of floods for industry and services occurs with some delay. The
timing of the effect highlights the importance of transmission mechanisms based on
supply chain relationships (for instance, larger cotton production inducing a later
expansion in textile production) and electricity generating capacity (as plentiful water
supply facilitates electricity generation, leading to a future expansion of industry and
services).
The response of agricultural growth is significantly positive one year earlier than
nonagricultural growth, in year 1, but not the same year of the event. This may indicate
that the potentially beneficial effects of floods on land productivity emerge in the
subsequent harvesting cycle. For the sample of developing countries, the cumulative
mean effect of floods on GDP growth is 0.5 pp and on agricultural growth, 0.6 pp.
This description of results applies to the samples of all countries and developing
countries only. For advanced countries, only agricultural growth is significantly affected
by floods. Although in year 3 after the event the mean response of agricultural growth is
significantly negative, the previous mean response had been consistently positive so that
the cumulative effect of floods is also positive and significant for advanced countries.
Regarding the comparison between moderate and severe floods, the annual mean
responses indicate that the significantly positive effects observed above come only from
moderate floods. Severe floods do not produce positive and significant mean responses
of GDP growth or its two components. Regarding the cumulative effects, moderate
29
floods induce an increase of 0.6 pp for GDP growth and 0.5 pp for agricultural growth.
As something of an anomaly, the cumulative impact of severe floods is positive and
significant, despite the fact that none of the annual mean responses is statistically
significant.
C. Earthquakes
The results on the mean response of growth to earthquake shocks are weaker in
terms of statistical significance than in the case of droughts and floods. Earthquakes do
not seem to have a significant effect on GDP growth in any of the three samples of
countries. However, there are some noteworthy results regarding sectoral growth,
particularly for the sample of developing countries.
Focusing on the sample of developing countries, earthquakes appear to have a
negative impact on agricultural growth, rendering a negative cumulative effect of about
1.4 pp. The fact that this effect is not due to a sharp response in any given year but,
rather, to the accumulation of effects over some years may elucidate its likely channels.
They may consist of, first, the disruption of transport and other infrastructure services
that supports the distribution of agricultural inputs and outputs, and, second, a diversion
of resources to reconstruction efforts in other sectors, particularly in urban areas.
In contrast, earthquakes elicit a positive mean response of nonagricultural growth
in years 0 and 1 of the event. The latter one is statistically significant and amounts to an
increase of 0.7 pp of valueadded growth. This positive effect is consistent with the
reconstruction activity that follows an earthquake in residential housing, public
infrastructure, and production plants.
These results are further clarified when considering the effect of moderate vs.
severe natural disasters. The negative cumulative impact of earthquakes on non
agricultural growth appears to occur with larger strength and significance for severe
earthquakes. They produce a cumulated decrease in agricultural growth of almost 5 pp
30
over the first years after the event. Similarly, the positive impact of earthquakes on non
agricultural activity seems to derive from moderate earthquakes only severe earthquakes
do not produce a significantly positive mean response of nonagricultural growth. In the
case of severe earthquakes, the destruction of capital stock and labor force is large
enough so as to cancel out the positive effect of reconstruction activity.
D. Storms
As in the case of earthquakes, the mean responses of growth to storms are weaker
in statistical significance than those of droughts and floods. Nonetheless, some results do
emerge from the data. Storms tend to have a negative effect on GDP growth and non
agricultural growth the same year of the event. This observation holds for the samples of
all countries and developing countries. The effect is shortlived and small. In fact, for
the sample of developing countries only, the negative impact of storms amounts to 0.3 pp
of GDP growth and 0.4 pp of nonagricultural growth. In the following years,
particularly for nonagricultural growth, there is a growth rebound representing most
likely reconstruction efforts.
For the sample of rich countries, the effect of storms is minimal. There seems to
be a negative response of agricultural growth in year 2 after the event, but in the
surrounding years the mean response is positive, albeit non significant.
Turning to the comparison between moderate and severe storms, the main point to
observe is that the negative growth effect noted above comes from the severe cases. For
both GDP growth and nonagricultural growth, the cumulative effect of severe storms is
negative and statistically significant, amounting to about 3 pp. For severe storms, the
largest or most significant negative effects appear with some delay, in years 2 or 3 after
the event. Conversely, for moderate storms, the mean response of growth and non
agricultural growth in those years is positive, reflecting in all likelihood the importance of
reconstruction activities. How are these results consistent with those presented above?
The negative and positive effects in later years of, respectively, severe and moderate
31
storms would tend to cancel each other out when estimation does not differentiate by
severity of the disaster.
Finally, regarding agricultural growth, moderate storms have a negative and
significant effect in year 1 of the event. However, in the following year, the effect is
positive, significant, and of about the same size, cancelling the previous one.
V. Concluding Remarks
This study has analyzed the path of macroeconomic adjustment and recovery in
the aftermath of four types of natural disasters, namely, droughts, floods, earthquakes,
and storms. Specifically, we have measured and examined the mean response of GDP
per capita growth and its major components, agricultural and nonagricultural per capita
valueadded growth. Applying a VARX methodology on a panel of 87 countries and 48
years (19602007), we find heterogeneous effects on a variety of dimensions. First, the
effects of natural disasters are stronger, for better or worse, on developing than on rich
countries. Second, while the impact of some natural disasters can be beneficial when
they are of moderate intensity, severe disasters do never have positive effects. Third, not
all natural disasters are alike in terms of the growth response they induce, and, perhaps
surprisingly, some can entail benefits regarding economic growth. Even within
commonly used categories of natural disasters (e.g., climatic), different types of disasters
can and do have different effects (e.g., droughts vs. floods).
Let's focus the conclusion on the results for developing countries. Droughts have
an overall negative effect on GDP growth. As expected, the effect is stronger for
agricultural growth, but it is also negative for nonagricultural activities. For agricultural
growth, the negative effect of droughts is immediate, while for nonagricultural growth,
the negative impact is felt also with some delay. The cumulative negative response to
droughts is 1.7 percentage points (pp) for GDP growth and 1.6 pp for agricultural growth.
32
In contrast to droughts, floods tend to have a positive effect on economic growth.
The response of agricultural growth is significantly positive one year after but not on the
same year of the event. The positive response of nonagricultural growth appears even
later, which suggests the importance of transmission mechanisms based on supply chain
relationships across sectors. The cumulative positive effect of floods on GDP growth is
0.5 pp and on agricultural growth, 0.6 pp.
Earthquakes do not seem to have a significant effect on GDP growth. However,
there are some noteworthy results regarding sectoral growth. Earthquakes appear to have
a negative impact on agricultural growth, rendering a negative cumulative effect of about
1.4 pp. In contrast, earthquakes elicit a positive mean response of nonagricultural
growth one year after the event of 0.7 pp. This positive effect is consistent with the
reconstruction activity that follows an earthquake in residential housing, public
infrastructure, and production plants.
Storms tend to have a negative effect on GDP growth and nonagricultural growth
the same year of the event. The effect is shortlived and small, however. In fact, the
negative impact of storms amounts to 0.3 pp of GDP growth and 0.4 pp of non
agricultural growth. In the following years, particularly for nonagricultural growth,
there is a growth rebound representing most likely reconstruction efforts.
In our opinion, future research should concentrate in exploring and clarifying the
mechanisms through which the heterogeneous impacts of natural disasters on economic
growth are produced. This paper has contributed to describing this heterogeneity, but
much remains to be done in explaining it. For this purpose, both panel and individual
country analysis should prove to be useful.
33
References
[1] Akaike, H. 1973. Information theory and an extension of the maximum likelihood
principle. In Petrov, B. N., Cszaki, F. eds., Second International Symposium on
Information Theory, 267281. Akademiai Kiado, Budapest.
[2] Becker, T., and P. Mauro. 2006. Output drops and the shocks that matter. IMF
Working Paper WP/06/172.
[3] Benson, C., and E. Clay. 2004. Understanding the economic and financial impacts
of natural disasters. Disaster Risk Management Series No.4. World Bank,
Washington DC.
[4] Caselli, F., and P. Malhotra. 2004. Natural disasters and growth: from thought
experiment to natural experiment (unpublished).
[5] Cuaresma, J. C., J. Hlouskova, and M. Obersteiner. 2008. Natural disasters as
creative destruction? Evidence from developing countries. Economic Inquiry 46 (2),
214226.
[6] Jaramillo H., C. R. 2007. Natural disasters and growth: evidence using a wide panel
of countries. Documento CEDE 200714.
[7] Hallegatte, S., and M. Ghil. 2007. Endogenous business cycles and the economic
response to exogenous shocks. FEEM Working Paper No. 20.
[8] Hamilton, J. D. 1994. Time Series Analysis, Princeton University Press, Princeton,
NJ.
34
[9] Heger, M., A. Julca, and O. Paddison. 2008. Analysing the impact of natural
hazards in small economies. UNUWIDER Research Paper No.2008/25.
[10] Im, K. S., M. H. Pesaran, and Y. Shin. 2003. Testing for unit roots in heterogeneous
panels. Journal of Econometrics 115, 5374.
[11] International Monetary Fund, 2003. Fund assistance for countries facing exogenous
shocks. Policy Development and Review Department, International Monetary Fund.
[12] Levin, A., C. F. Lin, and C. S. J. Chu. 2002. Unit root tests in panel data:
Asymptotic and finitesample properties. Journal of Econometrics 108, 124.
[13] Loayza, N., E. Olaberría, J. Rigolini, and L. Christiaensen. 2009. Natural Disasters
and MediumTerm Economic Growth: The Contrasting Effects of Different Events
on Disaggregated Output. Mimeo, The World Bank.
[14] Nickell, S. 1981. Biases in dynamic models with fixed effects. Econometrica 49 (6),
14171426.
[15] Noy, I. 2009. The macroeconomic consequences of disasters. Journal of
Development Economics 88 (2), 221231.
[16] Raddatz, C. 2007. Are external shocks responsible for the instability of output in
lowincome countries? Journal of Development Economics 84 (1), 155187.
[17] Rasmussen, T.N. 2004. Macroeconomic implications of natural disasters in the
Caribbean. IMF Working Paper WP/04/224.
35
[18] Schwartz, G. 1978. Estimating the dimension of a model. Annals of Statistics 6 (2),
461464.
[19] Skidmore, M., and H. Toya. 2002. Do natural disasters promote longrun growth?
Economic Inquiry 40 (4), 664687.
[20] Skidmore, M., and H. Toya. 2007. Economic development and the impacts of
natural disasters. Economics Letters 94, 2025.
[21] World Bank, 2008. World Development Indicators. Washington, DC
36
Table II.1
List of countries
Country name
All countries 87
Developing countries 62
Rich countries 25
* Algeria France * Pakistan
* Argentina Gabon * Panama
Australia Germany * Papua New Guinea
Austria Ghana * Paraguay
* Bangladesh Greece * Peru
* Barbados Guatemala * Philippines
Belgium GuineaBissau Portugal
* Belize Guyana * Rwanda
* Benin Honduras Saudi Arabia
* Bolivia Hungary * Senegal
* Botswana Iceland * Seychelles
* Brazil * India * South Africa
* Brunei Darussalam * Indonesia Spain
* Burkina Faso Italy * Sri Lanka
* Cameroon Japan * St. Vincent and the Grenadines
Canada * Jordan * Swaziland
* Central African Republic * Kenya Sweden
* Chad * Korea, Rep. Switzerland
* Channel Islands * Lesotho * Syrian Arab Republic
* Colombia Luxembourg * Thailand
* Congo, Dem. Rep. * Madagascar * Togo
* Costa Rica * Malawi * Trinidad and Tobago
* Cote d'Ivoire * Malaysia * Tunisia
Denmark * Mexico United Arab Emirates
* Dominican Republic * Morocco United Kingdom
* Ecuador Netherlands United States
* Egypt, Arab Rep. New Zealand * Uruguay
* El Salvador Norway * Venezuela, RB
Finland * Oman * Zambia
* indicates developing countries
37
Table II.2
Descriptive Statistics
Sample: Developing countries
Obs Mean Std. Dev. Min Max
Growth 2843 0.0167517 0.0546948 0.4422607 0.4798489
Agr. Growth 2348 0.0031466 0.0826112 0.4797475 0.4935743
Nonagr. Growth 2305 0.017093 0.0558703 0.4585984 0.3568618
Sample: Rich countries
Obs Mean Std. Dev. Min Max
Growth 1136 0.0236359 0.0330033 0.2331791 0.1951103
Agr. Growth 858 0.0093051 0.0718126 0.2801243 0.4300981
Nonagr. Growth 843 0.0211919 0.0358027 0.2625033 0.1856347
Table II.3
Piecewise correlation among variables
Sample: Developing countries
Nonagr.
Growth Agr. growth growth Droughts Floods Earthquakes Storms
Growth 1
Agr. growth 0.3878 1
Nonagr. growth 0.7969 0.1095 1
Droughts 0.0735 0.1048 0.0294 1
Floods 0.0377 0.0172 0.0386 0.0994 1
Earthquakes 0.0098 0.0202 0.0071 0.0169 0.1175 1
Storms 0.0124 0.0116 0.0057 0.0353 0.1747 0.0682 1
Sample: Rich countries
Nonagr.
Growth Agr. growth growth Droughts Floods Earthquakes Storms
Growth 1
Agr. growth 0.0737 1
Nonagr. growth 0.9684 0.0173 1
Droughts 0.0011 0.0392 0.005 1
Floods 0.015 0.0271 0.0037 0.0306 1
Earthquakes 0.044 0.0098 0.0582 0.01 0.026 1
Storms 0.0247 0.0517 0.005 0.037 0.0451 0.0249 1
38
Table III.1
Unit Root Tests
With coutnryspecific Agr. value With coutnryspecific
intercept added Terms of trade intercept GDP per capita Nonagr. value added
per capita and countryspecific trend per capita
A. Tests for Series in levels A. Tests for Series in levels
I. Fraction of countries that reject UR in ADFtest I. Fraction of countries that reject UR in ADF test
2/75 11/76 5/87 3/73
II. Fraction of countries that reject UR in PP test II. Fraction of countries that reject UR in PP test
17/75 16/76 5/87 4/73
III. Pvalues of LevinLinChu test III. Pvalues of LevinLinChu test
0.123 0.0419 0.123 0.621
IV. Pvalues of ImPesaranShin test IV. Pvalues of ImPesaranShin test
0.969 0.101 1 1
B. Tests for Series in Differences B. Tests for Series in Differences
I. Fraction of countries that reject UR in ADF
test I. Fraction of countries that reject UR in ADF test
46/75 63/76 59/87 49/73
II. Fraction of countries that reject UR in PP test II. Fraction of countries that reject UR in PP test
75/75 76/76 87/87 72/73
III. Pvalues of LevinLinChu test III. Pvalues of LevinLinChu test
0 0 0 0
IV. Pvalues of ImPesaranShin test
0 0
(i) The significance level is at 10 percent.
(ii) For all unit root tests, both individual and panel, the default settings of EViews 6.0 were used.
39
Table III.2
Block Exogeneity Tests for the Disaster Variables
Significance level
Sample Lag 1 Lag 2 Lag 3
All GDP growth 0.5112839 0.67821097 0.5277162
countries Agr. growth 0.3291263 0.77459891 0.8643032
Nonagr. growth 0.5947947 0.41047053 0.415458
Developing GDP growth 0.6395072 0.82498702 0.7428414
countries Agr. growth 0.3762702 0.84914854 0.9066644
Nonagr. growth 0.6364801 0.53872097 0.5864839
Rich GDP growth 0.1871027 0.29234087 0.1342733
countries Agr. growth 0.056833 * 0.11869545 0.0635466 *
Nonagr. growth 0.2488736 0.63553462 0.5321749
* denotes statistical significance at 10 percent level.
40
Table III.3
Information Criteria Values
Number of lags
Sample p=q=1 p=q=2 p=q=3
All GDP growth AIC 15.5853 15.5933 15.5777
countries SBC 15.516 15.4855 15.4314
Agr. growth AIC 13.3562 13.3932 13.3752
SBC 13.2769 13.2698 13.2078
Nonagr. growth AIC 15.4666 15.4645 15.4527
SBC 15.3864 15.3396 15.2832
Developing GDP growth AIC 14.5324 14.5318 14.5089
countries SBC 14.442 14.3911 14.3179
Agr. growth AIC 12.6936 12.72 12.6939
SBC 12.591 12.5605 12.4774
Nonagr. growth AIC 14.4328 14.4216 14.4004
SBC 14.3292 14.2603 14.1815
Rich GDP growth AIC 22.0632 22.1027 22.061
countries SBC 21.865 21.7944 21.6426
Agr. growth AIC 17.6826 17.777 17.699
SBC 17.4524 17.4189 17.213
Nonagr. growth AIC 22.1215 22.1395 22.0814
SBC 21.8882 21.7766 21.5889
Bold figures indicate the minimum AIC / SBC.
41
Table IV.1
Mean responses of the growth rates of each sector to natural disaster shocks
Sample: All countries
Mean responses of
GDP growth Agr. growth Nonagr. growth
AIC 15.6239 13.4292 15.501
SBC 15.5623 13.3587 15.4297
Year 0 0.013223 ** 0.031715 ** 0.004852 *
Droughts Year 1 0.0023064 0.020766 ** 0.0001746
Year 2 0.0044188 * 0.0021294 0.0053657 *
Year 3 0.00028832 0.0017919 0.00056442
Cumulative effect 0.015047 ** 0.014870 ** 0.010957 *
Year 0 0.0017396 0.000020262 0.0019237
Earthquakes Year 1 0.002263 0.0096375 0.0048984
Year 2 0.0022208 0.0047104 0.0038522
Year 3 0.0005828 0.0024555 0.0012232
Cumulative effect 0.001199 0.011872 * 0.0017467
Year 0 0.0014053 0.001213 0.0006851
Floods Year 1 0.000036711 0.0050796 * 0.0008682
Year 2 0.0026809 * 0.0013598 0.0025976 *
Year 3 0.0006529 ** 0.0008888 0.0009695 **
Cumulative effect 0.0047024 * 0.0067636 ** 0.003384
Year 0 0.0031159 * 0.0009415 0.0037988 *
Storms Year 1 0.0010362 0.0030481 0.0011004
Year 2 0.00062423 0.0026836 0.0002100
Year 3 0.000022213 0.0005090 0.0000592
Cumulative effect 0.0035057 0.001815 0.0029677
* (**) denotes statistical significance at onetail 10 (5) percent level
42
Table IV.2
Mean responses of the growth rates of each sector to natural disaster shocks
Sample: Developing countries
Mean responses of
GDP growth Agr. growth Nonagr. growth
AIC 14.5741 12.7693 14.4715
SBC 14.4937 12.6781 14.3794
Year 0 0.014091 ** 0.031021 ** 0.0052742 *
Droughts Year 1 0.0022092 0.021654 ** 0.00025094
Year 2 0.0050741 * 0.0056886 0.005939 *
Year 3 0.00033704 0.00059208 0.00062653
Cumulative effect 0.016619 ** 0.015648 ** 0.012091 *

Year 0 0.0020709 0.001304800 0.0022349
Earthquakes Year 1 0.0032152 0.0093886 0.0070503
Year 2 0.0034535 0.005758 0.0049593
Year 3 0.00084333 0.0026175 0.0016999
Cumulative effect 0.00098927 0.013834 * 0.002626
Year 0 0.0014411 0.0011730 0.0005088
Floods Year 1 0.000070818 0.0049745 * 0.0012603
Year 2 0.0029372 * 0.0004760 0.0030079 *
Year 3 0.00067184 ** 0.0005522 0.0010942 **
Cumulative effect 0.0049793 * 0.0060713 ** 0.0033506
Year 0 0.0032138 0.0002941 0.0042957 *
Storms Year 1 0.0008594 0.0055736 0.0016841
Year 2 0.00083222 0.0053877 0.0000693
Year 3 0.000072543 0.0011998 0.0001012
Cumulative effect 0.0033135 0.0016798 0.0026434
* (**) denotes statistical significance at onetail 10 (5) percent level
43
Table IV.3
Mean responses of the growth rates of each sector to natural disaster shocks
Sample: Rich countries
Mean responses of
GDP growth Agr. growth Nonagr. growth
AIC 22.2151 17.9134 22.2782
SBC 22.0389 17.7088 22.0709
Year 0 0.007907 0.060068 ** 0.0018634
Droughts Year 1 0.0063094 0.020022 0.0055474
Year 2 0.0093003 0.10835 ** 0.0080284
Year 3 0.0047549 0.0238 ** 0.0041009
Cumulative effect 0.015653 0.00446 0.0084453
Year 0 0.00075515 0.0061523 0.0010407
Earthquakes Year 1 0.0022573 0.0075171 0.0034712
Year 2 0.0021551 0.0006452 0.00021396
Year 3 0.0012077 0.0017494 0.00046867
Cumulative effect 0.0018607 0.0010298 0.0017479
Year 0 0.0021692 0.0044266 0.0030229
Floods Year 1 0.00021644 0.0049304 0.0019032
Year 2 0.00011382 0.011557 0.0006401
Year 3 0.000032614 0.0037681 * 0.0002934
Cumulative effect 0.0020340 0.017146 * 0.0039926
Year 0 0.0011873 0.0067637 0.0002764
Storms Year 1 0.0017083 0.010002 0.0032889
Year 2 0.00014199 0.015712 * 0.0001117
Year 3 0.00052518 0.0026164 0.0005428
Cumulative effect 0.0022284 0.0098573 0.0031342
* (**) denotes statistical significance at onetail 10 (5) percent level
44
Table IV.4
Mean responses of the growth rates of each sector to moderate/severe natural disaster shocks
Sample: Developing countries * (**) denotes statistical significance at onetail 10 (5) percent level.
Mean responses of
GDP growth Agr. Growth Nonagr. Growth
Moderate Severe Moderate Severe Moderate Severe
AIC 14.5476 12.7444 14.4379
SBC 14.4069 12.5849 14.2766
Year 0 0.0085714 * 0.018847 ** 0.011008 0.046423 ** 0.0027887 0.0078248 *
Droughts Year 1 0.0043316 0.006814 * 0.0014815 0.035768 ** 0.0044605 0.0028619
Year 2 0.006477 0.0040597 0.00044072 0.0078779 0.0040232 0.0070282 *
Year 3 0.00040188 0.00092473 0.000025381 0.0015406 0.00066348 0.0004734
Cumulative effect 0.019782 ** 0.015168 ** 0.0099418 0.020074 ** 0.011936 0.0124645
Year 0 0.0010823 0.0022661 0.0003944 0.007899 0.0005302 0.00183
Floods Year 1 0.000602630 0.0078984 0.0056914 * 0.0013757 0.0010570 0.0066199
Year 2 0.0031417 * 0.000058796 0.00057858 0.0073623 0.0034817 * 0.001819
Year 3 0.000787 ** 0.00052586 0.00028351 0.0024108 0.0012580 ** 0.0004639
Cumulative effect 0.0056136 * 0.0062170 0.0052237 * 0.014226 * 0.0042129 0.0107328
Year 0 0.0013108 0.005904 0.0012498 0.0080319 0.0010324 0.0051802
Earthquakes Year 1 0.003685 0.0011218 0.0083234 0.025484 0.008449 * 0.0010412
Year 2 0.0048286 0.0043999 0.0027261 0.028254 0.006611 0.0080918
Year 3 0.00093429 0.00066329 0.0014124 0.012082 * 0.0020765 0.00082313
Cumulative effect 0.00076709 0.010762 0.0083873 0.049688 ** 0.0007939 0.015136
Year 0 0.002562 0.0061287 0.00039477 0.0000557 0.0037605 0.0018873
Storms Year 1 0.00030924 0.0072597 0.0062362 * 0.0025188 0.0024532 0.0054874
Year 2 0.0021152 0.014758 * 0.0065535 * 0.010492 0.0018227 0.018303 **
Year 3 0.00010718 0.0019053 0.00152 0.0040787 0.0003248 0.0043348 *
Cumulative effect 0.00064886 0.030052 * 0.0015975 0.0088764 0.00084024 0.0300125 *
45
Figure IV.1: Response to Drought Shock
Mean response of GDP growth
Mean response of agricultural growth
Mean response of nonagricultural growth
46
Figure IV.2: Response to Flood Shock
Mean response of GDP growth
Mean response of agricultural growth
Mean response of nonagricultural growth
47
Figure IV.3: Response to Earthquake Shock
Mean response of GDP growth
Mean response of agricultural growth
Mean response of nonagricultural growth
48
Figure IV.4: Response to Storm Shock
Mean response of GDP growth
Mean response of agricultural growth
Mean response of nonagricultural growth
49
Table A.1
Mean responses of the growth rates of each sector to natural disaster shocks (with p = q = 1)
Sample: Developing countries
Mean responses of
GDP growth Agr. growth Nonagr. Growth
AIC 14.5606 12.7264 14.4661
SBC 14.5103 12.6694 14.4085
Year 0 0.013799 ** 0.031014 ** 0.0053130 *
Droughts Year 1 0.0013177 0.022584 ** 0.0010444
Year 2 0.0000049136 0.006832 ** 0.00047764
Year 3 0.0000089841 0.0019891 ** 0.000090109
Cumulative effect 0.012477 ** 0.013273 ** 0.0069251
Year 0 0.00156690 0.00133190 0.00074075
Floods Year 1 0.00018370 0.00501190 * 0.00094847
Year 2 0.0001215 0.0013814 * 0.000070547
Year 3 0.0000047314 0.00039035 0.000028721
Cumulative effect 0.0018768 0.0053528 ** 0.0003070
Year 0 0.001798 0.00020651 0.0024988
Earthquakes Year 1 0.0027774 0.0086748 0.0073422 *
Year 2 0.00027501 0.0021892 0.0011007
Year 3 0.0000114 0.00059905 0.00030342
Cumulative effect 0.0043118 0.0072912 0.011245
Year 0 0.0028280 0.00041656 0.0040363 *
Storms Year 1 0.00071057 0.0058776 0.0015851
Year 2 0.00035935 0.0015246 0.000018088
Year 3 0.000015152 0.00042154 0.000033405
Cumulative effect 0.0039131 0.0043580 0.0023997
* (**) denotes statistical significance at onetail 10 (5) percent level.
50
Table A.2
Mean responses of the growth rates of each sector to moderate/severe natural disaster shocks (with p = q = 1)
Sample: Developing countries * (**) denotes statistical significance at onetail 10 (5) percent level.
Mean responses of
GDP growth Agr. growth Nonagr. Growth
Moderate Severe Moderate Severe Moderate Severe
AIC 14.5435 12.7126 14.4412
SBC 14.453 12.61 14.3375
Year 0 0.0079178 * 0.018701 ** 0.011129 0.046391 ** 0.0026107 0.0077559 *
Droughts Year 1 0.0053518 0.0062294 * 0.0027189 0.037006 ** 0.0052728 0.0021805
Year 2 0.00087276 0.00055808 0.0010556 0.01101 ** 0.0015441 0.00021606
Year 3 0.00006324 0.000060463 0.00032598 0.0031707 ** 0.00032876 0.000074007
Cumulative effect 0.014206 * 0.011853 * 0.0091397 0.017224 ** 0.0097564 0.0052853
Year 0 0.0011946 0.0027187 0.00051129 0.0082707 0.00074202 0.0013555
Floods Year 1 0.0010416 0.006966 0.0061001 ** 0.00087786 0.00043716 0.0052264
Year 2 0.00014743 0.000099217 0.0017522 * 0.0010216 0.000046472 0.000079873
Year 3 0.000011664 0.000046840 0.00049898 * 0.00036199 0.0000153 0.00011316
Cumulative effect 0.0023953 0.0041949 0.0053582 * 0.0080525 0.00024309 0.0067749
Year 0 0.0016463 0.0037947 0.0022036 0.010158 0.0022439 0.0035697
Earthquakes Year 1 0.0029503 0.00079829 0.0077088 0.021632 0.0084635 * 0.00076896
Year 2 0.00020976 0.00066901 0.0018478 0.0061573 0.001313 0.000080931
Year 3 0.000015021 0.000013519 0.00049239 0.0017483 0.00035681 0.000003059
Cumulative effect 0.0044019 0.0039105 0.0041498 0.027381 0.012377 0.0042608
Year 0 0.0028685 0.0053638 0.00022475 0.00014246 0.0043831 * 0.0013487
Storms Year 1 0.00026787 0.0053952 0.0062769 * 0.00062212 0.0019576 0.0040608
Year 2 0.00044568 0.00029306 0.0015361 0.00078555 0.00002783 0.00010281
Year 3 0.00001471 0.000030068 0.00041283 0.00027976 0.000033918 0.000093899
Cumulative effect 0.0035968 0.010496 0.0053784 0.00025879 0.0024194 0.0056062
51
Figure A.1: Response to Drought Shock (with p = q = 1)
Mean response of GDP growth
Mean response of Agricultural growth
Mean response of Nonagricultural growth
52
Figure A.2: Response to Flood Shock (with p = q = 1)
Mean response of GDP growth
Mean response of Agricultural growth
Mean response of Nonagricultural growth
53
Figure A.3: Response to Earthquake Shock (with p = q = 1)
Mean response of GDP growth
Mean response of Agricultural growth
Mean response of Nonagricultural growth
54
Figure A.4: Response to Storm Shock (with p = q = 1)
Mean response of GDP growth
Mean response of Agricultural growth
Mean response of Nonagricultural growth
55