The role of trade in adaptation to
natural disasters
Thomas van den Berg*
August 12, 2019
(Research) Master Thesis
Abstract
In the context of increasing natural hazards, finding policy mechanisms that can mitigate their costs is an important task. In this study, I explored whether trade openness has a mitigative effect on the damages incurred by natural disasters. To this end, I studied the effects of trade openness using physical measures of natural disaster intensity in a macro panel, using both conventional growth regressions as well as survival analysis. I argue that survival analysis is especially suited to answering questions on ex ante mitigative mechanisms in the context of natural disasters. In the end, I found evidence that trade can indeed mitigate the macroeconomic damage associated with natural disasters, although the effects are heterogeneous in disaster type and trade sector type, and statistical significance is moderate. Consequently, both policy makers and economists should consider the potential that trade openness has in this dimension, yet also be aware of the conditions necessary for trade to succeed as adaptive mechanism.
1. Introduction
Natural disasters cause severe human and economic losses. The most recent 10-year average global cost of natural disasters was found to exceed $250 billion per year (Daniell et al., 2016), while damage estimates for the 2011 Tōhoku earthquake and accompanying flooding that struck Japan alone exceed $300 billion. Indeed, natural disasters and their associated damage costs are spread heterogeneously over time and space, which exacerbates their local macroeconomic impacts when they happen.
Recently, economic literature showed increased interest in the growth effects of natural disasters. This relates to the proliferation of global warming, which is predicted to increase both the rate and severity of many types of natural hazards. These predictions are synthesized in the 5th and most recent IPCC report (Seneviratne et al., 2012). For example, there will be an increase in extreme temperature and precipitation events (Banholzer, Kossin & Donner, 2014) and in heat strokes (IPCC, 2013), while rising sea levels are predicted to increase occurrences of flooding with strong confidence.
In the context of increasing natural hazards, studying factors that can mitigate the negative effects associated with disasters is of large policy relevance (Kahn, 2005). Indeed, in this study, the focus is specifically on investigating the interaction between natural disaster impacts and the degree to which countries are open to trade. Trade is a salient mechanism to explore empirically because there is strong recent policy interest within the context of climate change. This relevance is exemplified by several prominent institutions who emphasize the importance of studying trade in relation to climate change and its effects. Last year, the World Trade Organisation (WTO, 2018) organised its first “Natural Disasters and Trade Symposium”, stressing the importance of further research. Another major policy institution emphasizing the importance of more research into the intersection between trade and economic effects of climate change is the United Nations Intergovernmental Panel on Climate Change (see Jones et al., 2014).
future natural disasters, such policies may be misguided. More generally, a mitigative effect contrasts with potential policy conclusions actors might take based on strands of literature and the public debate highlighting the negative effect of globalisation on the environment (e.g. Wenz & Levermann, 2016), either through increased transport through global supply chains (although Arto & Dietzenbacher, 2014, find no effect of trade pattern changes on global greenhouse gas emissions) or through a supposedly induced race-to-the-bottom in environmental standards (although country-panel empirical studies, e.g. Frankel & Rose, 2005, find no effect on global greenhouse gases).
To study whether trade openness has a mitigative effect or not, this study investigates the impact of natural disasters within a multiple-country panel setup, because this allows us to compare disaster impacts on GDP over time with trade openness at the policy-relevant country level. The study especially adds to the literature through its investigation of the effect of trade using survival (or duration) analysis. Survival analysis relates the time that passes before an event occurs with covariates that are associated with increasing or decreasing that quantity of time. Taking the event to be some level of intensity of the natural disaster with certain threshold of minimal impact on GDP growth, I then relate this to the covariate of trade openness. Indeed, this method of analysis does not depend as strongly on assumptions from closed-economy neoclassical growth models as conventional GDP growth regressions would. Moreover, duration analysis has the benefit that it allows for specification of a hazard, which determines the risk of an event occurring at baseline levels of covariates. Thus, taken together, the use of this method allows this study to specifically pinpoint the effect that trade openness will have on the natural-disaster-GDP relation, where natural disasters are also related directly to the natural hazard.
The rest of this paper will be structured as follows. In the next section, I discuss the existing economic literature, signalling deficiencies this study fills in the literature on determinants of disaster risk, and elaborating on theoretical and empirical reasons for an interaction between trade and natural disaster damages. Section 3 describes the data and explains the empirical strategies that are used. Section 4 presents results. Finally, section 5 offers concluding remarks.
2. Theoretical considerations
2.1 Natural disaster impacts
Research on economic impacts of natural disasters is plentiful over the last decade. Many studies examine the growth effects of natural disasters in a country panel framework and find a negative effect of disasters on economic growth (Raddatz, 2007; Noy, 2009; Felbermayr & Gröschl, 2014). Klomp and Valckx (2014) perform a meta-analysis of the field, in which they signal 30 comparable studies to Noy (2009) and 600 relevant estimates. They find that only 40% of these estimates is significant at the 10% level. Also, about a third of the significant estimates has a positive sign. Based on qualitatively analyzing the literature, they conclude in favor of a negative genuine effect of natural disasters but find some evidence of publication bias as well. Furthermore, they argue that it is likely that studies that find ‘false positives’ suffered from data endogeneity issues.
It must be noted that these studies focused on the question whether there actually is an ex post negative growth effect of natural disasters (e.g. Skidmore & Toya, 2002; Felbermayr & Gröschl, 2014; Klomp, 2016), as opposed to focusing on ex ante policy mechanisms, such as (partially) trade openness, that might prevent or mitigate such negative effects. Although some of these studies do find suggestive (Benson & Clay, 2003) and concrete (Noy, 2009; Felbermayr & Gröschl, 2014; Wen & Chang, 2015) evidence for a mitigative effect of trade, they do not comment much on it, as it is not their main focus. However, there is reason to believe that there is a mitigative effect of trade for both theoretical and empirical reasons, which warrants specific research. Consequently, the next section identifies these strands of research that give foundation to the claim that there is a relation between trade openness and disaster damages, through discussing various channels by which trade openness may affect it.
2.2 Trade as determinant of natural disaster impacts
The most salient channel through which trade openness may affect the natural disaster risk is through its effects on the sectoral composition. However, it is ambiguous what the sign of this interaction effect is. Openness to trade can have both an effect of specialization and an effect of diversification.
According to the Heckscher-Ohlin theorem, trade openness leads to an increased specialization in goods that a country is relatively factor-abundant in. This may leave a country more vulnerable if they specialize in sectors that are more at risk from suffering negatively from natural disasters (Benson & Clay, 2004), a mechanism especially noted in the context of developing countries shifting to specific crop and manufacturing types that are especially at risk. Indeed, Benson and Clay (ibid.) suggest that the effect of trade openness therefore depends on a country’s initial sectoral composition and is therefore ambiguous.
and Smith (2016). Moreover, in response to aforementioned calls for more research (Dell, Jones & Olken, 2014), some concrete empirical developments have been made recently, albeit in more micro-level contexts.
Most notably, a case study in India (Allen & Atkin, 2016) building on Costinot, Donaldson and Smith (2016) as well as earlier work (Costinot, Donaldson & Komunjer, 2011) proved that better internal market access to trade corresponded with less volatile agricultural crop choices and, subsequently, higher average yields in income terms. This empirical extension concerns long-term average temperature increases rather than natural disasters, but the rationale should be the same. In the context of the 2011 Tōhoku earthquake and accompanying flooding that struck Japan, Hamano and Vermeulen (2019) relate natural disaster impact at the firm level with the amount of ports available for them to trade in, and also find a mitigative effect of trade. Essentially, these recent developments imply that trade openness functions as an ex ante risk-sharing mechanism, as noted by Allen and Atkin (2016), through diversification across space.
Next to the direct impacts of trade openness on natural disaster damages through sectoral composition shifts, trade openness can also indirectly interact with the natural-disaster-growth relation. I categorize this interaction into three channels. First, disaster impacts are largest among poorer sectors and poorer countries (Kahn, 2005; Skidmore & Toya, 2007), and trade has direct effects on GDP and the wealth distribution. Nonetheless, the direction of this effect is again ambiguous, as some studies find that a reduction in relative natural disaster costs only appears at some level of development, and that it follows a Kuznets-type pattern (Schumacher & Strobl, 2011).
Second, trade openness fosters knowledge spillovers in other drivers of disaster risk reduction. Namely, trade openness may improve the financial system through FDI expertise spillovers (Benson & Clay, 2003; Skidmore & Toya, 2007) such that it can better accommodate potential underinvestment in natural disaster risk reduction (Neumayer, Plümper & Barthel, 2014). Also, trade openness fosters sharing of technological improvements (Yanikkaya, 2003) that mitigate disaster impacts.
distribution of resources to cope with disaster consequences (Vaillancourt & Haavisto, 2015). However, this final effect may also turn out negatively, as governments may underinvest in disaster risk prevention through increased expectations of humanitarian disaster aid (Strömberg, 2007; Cohen & Werker, 2008).
A related final concern is that there may be a strong negative causal effect of natural disasters on trade, through these governmental influences but also through the destruction or blocking of vital transport infrastructure. Gassebner, Keck and Teh (2010) address this question. However, they conclude very small impacts of disasters on international trade, finding a decrease in imports of on average 0.2% and a decrease in exports by 0.1%, only in the very short run, and conclude that international trade is “shaken, not stirred” by natural disasters.
All in all, the literature provides many theoretical mechanisms by which trade openness interacts with the effects of natural disaster on GDP growth, but offers ambiguous predictions regarding the sign of the effect. Although most empirical studies imply that trade openness indeed mitigates disaster risks, functioning as risk-sharing mechanism across geographical space, the sign of the aggregate effect at the macro level is yet unclear. This presents a puzzle: Is there a positive or a negative effect of country-level trade openness in the context of dealing with natural disasters?
Indeed, while some studies at the macro level do include interactions with institutional variables including trade in their analyses (Noy, 2009; Felbermayr & Gröschl, 2014), they do not focus on them, and I test this hypothesis in a more general case than theirs, by the use of survival analysis. The following section explains the empirical strategies in more detail, while also explaining the data used.
3. Empirical strategies and Data
focused on improving the underlying data samples. Indeed, studies all use a specification similar to the original studies, employing a Barro-type (Barro, 1991) growth regression, where growth is expressed as a function of GDP in some earlier period and country-level economic or institutional indicators. The covariate of interest is that of natural disasters. The predominant finding is that natural disasters negatively affect GDP growth.
In this section, I explain how the Barro-type model can be used to predict the effect of trade openness on the disaster-income relation. I then contrast this empirical strategy with survival analysis, which this paper pioneers in the field. Indeed, I juxtapose results for both the Barro-type regressions and survival analysis in the context of my sample in the results section. In the final part of this section, I use the data considerations that stem from the previous literature to justify the data I use in this study and describe this data.
3.1 Barro-type growth regressions
The model specification used looks as follows:
∆𝑦𝑖,𝑡= (𝜌 − 1)𝑦𝑖,𝑡−1+ 𝛼𝐷𝑖,𝑡+ 𝛽𝑋𝑖,𝑡−1+ 𝛾𝐷𝑖,𝑡× 𝑋𝑖,𝑡−1+ 𝜈𝑖+ 𝜈𝑡+ 𝜀𝑖,𝑡 (1)
where ∆𝑦𝑖,𝑡 is the natural logarithm of the real GDP per capita growth rate, 𝑦𝑖,𝑡−1 is the lagged natural logarithm of GDP per capita, 𝐷𝑖,𝑡 is a disaster measure, 𝑋𝑖,𝑡−1 is a set of Barro (1991) type macro-economic or institutional variables, potentially including trade openness, and 𝜈𝑖 and 𝜈𝑡 are country- and time-fixed effects. Estimation is typically done using the OLS FE estimator. To find the effect of trade openness, we look at the sign of 𝛾, the coefficient on the interaction term. The use of only one lag of GDP growth is supported by the literature in the field (e.g. Raddatz, 2007; Noy, 2009).
an official state of emergency, or (4) when there is a call for international assistance (Munich Re, 2006; Gassebner, Keck & Teh, 2010).
The reporting probability thus depends on insurance claims and is therefore increasing with GDP. Also, there is potential measurement error because it depends on historical sources that are not standardized across countries. Finally, disaster data itself is operationalized in the form of monetary damages, which are likely higher as the economic activity increases. Indeed, these issues lead to concerns of selection bias. Klomp and Valckx (2014) argue that this bias is likely severe, as they signal ‘false positives’ for the coefficients on the disaster measures in many studies.
Recent literature accounts for this selection bias in two ways. First, they turn towards physical disaster data instead, that is, data that describes the physically measured severity of disasters, rather than the personal or economic impacts of them. Second, they opt for using some systems GMM estimation method to estimate equation (1) instead. Indeed, in this study, I take both approaches.
The former approach is taken by Felbermayr and Gröschl (2014) at the country-panel level. They compile an alternative database with geophysical and meteorological measures for disasters, from 1979 to 2010, for most of the world (Felbermayr & Gröschl, 2014) and for the same sorts of natural disasters as in EM-DAT. They repeat the country-panel analysis using model (1), but no longer require the use of systems GMM methods. They find a negative effect of natural disasters on growth. Additionally, they also find a mitigative effect of trade openness in one of their tests, where they add it as interaction term. They comprehensively analyse each individual type of natural disaster, consistently finding negative effects. Their conclusion is that the effect of natural disasters on short-term economic growth is “naturally negative”.
The second approach, using a systems GMM procedure to account for endogeneity in the disaster measure, is taken by Noy (2009) and many subsequent studies (e.g. Loayza et al., 2012; Klomp, 2016). It must be noted that they still state a preference for using actual physical intensity disaster measures, as such measures “can potentially fully overcome the exogeneity issue,” signalling some degree of uncertainty about the systems GMM approach still. Indeed, although they employ the three-step Hausman-Taylor estimator (Hausman & Taylor, 1981) which does overcome correlation between country-specific effects and the covariates, this estimator still requires the assumption that there is no reverse causality in the GDP growth variable itself. They do not use the more efficient Arellano-Bond estimator because it is not designed for typical macro panel samples (small N, high T).
3.2 Survival analysis
Natural disasters are hazardous events, much like diseases. From this perspective, it makes intuitive sense to statistically treat them as hazards. Indeed, employing a survival analysis, a method often used in biomedical sciences and micro-level empirical economics, relates the time it takes for an event to happen with covariates that increase or decrease that value of time. For example, a medical trial might use it to determine whether a new drug (the covariate) can increase the time until patient death (the ‘failure event’). Alternatively, a labour economist may be interested in the effect of a job stimulation plan on the time it takes for the unemployed to enter the labour market again. In the latter disciplinary context, it is more widely known as duration analysis.
The link with natural disasters is quite intuitively made. Namely, we can relate the time it takes for an event (namely, a natural disaster with certain threshold impact on the economy) to happen with covariates, such as trade openness, that increase or decrease that time-to-happen.
One important benefit of survival analysis that makes it better than Barro-type regressions is that it is not as grounded in neoclassical growth theory. Specifically, Barro-type models are grounded in the Solow-Swan growth model (Solow, 1956; Swan, 1956), built from the closed-economy perspective. In the survival model of this study, I do still include Barro-type indicators as covariates, but the dependent variable becomes the time until some economically damaging natural disaster occurs, making the relation between trade openness and the dependent variable more suitable with respect to the underlying economic-theoretical reasons to expect an interacted effect of trade openness (as outlined in the previous section) than in the Barro-type case. For example, the Barro-type growth model contains lagged GDP as explanatory variable. However, if trade indeed functions as ex anterisk sharing mechanism across space, then the positive effects of trade will be at least partly captured by the term on lagged GDP. Survival analysis prevents such unfavorable conditions.
As is common, I conduct the survival analysis using the Cox proportional hazards model (Wooldridge, 2010). This is a semiparametric model that works as follows. The hazard, which means the instantaneous rate of an event occurring after time 𝑡 has passed since the onset of the hazardous period, is denoted ℎ. The Cox proportional hazards regression assumes the relationship that
ℎ(𝑡, 𝑥𝑖) = ℎ0(𝑡) exp(𝑥𝑖′𝛽) (2)
Thus, ℎ(𝑡, 𝑥𝑖) is a function of some baseline hazard function ℎ0(𝑡), which is unknown and based partially on climate factors, and a vector of covariates 𝑥𝑖, which are associated with increasing or decreasing the time until the hazardous event – let us call it a ‘failure’ – happens.
In the context of this study, the most important qualitative decision is what constitutes a ‘failure’. Namely, a disaster only has an economic effect when economic activity is impacted by the natural disaster. Therefore, in this study, I define a ‘failure’ as a natural disaster that depresses GDP to some extent. Because it is a normative task to find such events and attribute GDP losses to them, I vary the extents to which GDP is affected, and insert determinants of GDP growth into the vector of covariates 𝑥𝑖.
stronger negative effect on GDP growth (e.g. Felbermayr & Gröschl, 2014; Strobl, 2012), while smaller natural disasters have almost no effect. Thus, to filter out as much as possible those natural disasters that substantially depressed GDP growth, I denote failures in the country-panel based also on criteria of disaster intensity. In sum, the vector of covariates 𝑥𝑖 contains trade openness as most important factor, while it also contains other determinants of the ‘failure’ state, including country-fixed effects, year-fixed effects, and some theory-based predictors of GDP growth.
Any panel database on disasters is potentially multiple-failure data, which means the same subject (in this case country) could experience multiple failures. The Cox proportional hazard model is equipped to deal with this. A more serious issue, however, is that of potential left-truncation in the sample. Left truncation occurs when countries are unable to enter the sample or enter the sample at later dates because they suffered too severe consequences from a pre-sample period failure to still enter. Indeed, low-developed countries that were continuously in the path of growth-harming natural disasters might not have had the data institutions required to supply enough macroeconomic indicators to enter the sample here, or they might have had changing borders for too long to still be considered while the border-changing conflicts were partly caused by natural disasters and their effect on economic growth. The bias left truncation causes may be severe when frailties are used to control for unobserved homogeneous factors driving multiple failures in unbalanced panels (Van den Berg & Drepper, 2016). The Cox proportional hazards model including a country-specific frailty term is:
ℎ𝑘𝑖(𝑡, 𝑥𝑖) = 𝜁𝑘ℎ0(𝑡) exp(𝑥𝑘𝑖′ 𝛽) (3)
where 𝜁𝑘 is the frailty for country 𝑘, and ℎ𝑘𝑖(𝑡, 𝑥𝑖) is the hazard rate for country 𝑘 at a time without having experienced failure of 𝑡.
developing countries, is almost balanced, and the balanced subset contains an even higher proportion of developing countries.
3.3.1 Data: Disaster Data
In light of the above discussion on disaster data endogeneity, this study uses the dataset of physical disaster intensity measures as described in Felbermayr and Gröschl (2014) and juxtaposes it with the EM-DAT dataset that is used in almost all other related literature. The Felbermayr-Gröschl data is more commonly known as GeoMet data from the ifo GAME database (ibid.). It covers the period 1979 to 2010, containing data on 108 countries. Out of these 108 countries, the database contains twenty-seven high income OECD member countries, six high income non-OECD members, twenty-two low income countries, thirty lower middle-income countries, and twenty-three upper middle-middle-income countries. Thus, the share of developing versus developed countries is roughly equal. The non-high-income countries are distributed as follows: nine in East Asia and the Pacific, twelve in Eastern Europe and Central Asia, seventeen in Latin America and the Caribbean, seven in the Middle East and North Africa, five in South Asia, and twenty-five in Sub-Saharan Africa. Indeed, the dataset contains many countries in areas that are known to have large disaster risk or be of lower development level. The dataset is combined with the EM-DAT database, which contains their potentially endogenous measures of disasters for the same countries and years.
half of the long-run average. Finally, extreme temperature events are also sourced from meteorological records of worldwide weather stations, and the measure is the difference of the monthly temperature over the long-run (1979-2010) monthly mean temperature.
Records at the country-year level are of these six disasters’ highest intensity scores for that year. Figure 1 below shows a map of the world which displays the country means of those highest intensity scores for each year, separately for each disaster type. There are some surprising results, particularly that floods seem to appear mostly in the northern half of Africa, and that there is not much difference across regions in their scores for earthquakes and storms. Of course, this is partly due to the way the measures are set up. For example, earthquake Richter scales are treated linearly, even though this is a logarithmic scale in reality. Floods are recorded as strongly positive rainfall shocks, which explains their high incidence in relatively arid regions. The conclusion is that this dataset also suffers from potential bias, and therefore, it makes sense to juxtapose it with EM-DAT data, to still get a more robust picture of reality, even when the EM-DAT data suffers from selection bias that is likely more concerning in the context of the empirical specifications here.
A relevant determinant of the observed patterns is that some countries simply have higher chances to experience certain natural disasters by their absolute land area size. Of course, in the context of relating them to their economic effects, we must control for this. Indeed, Felbermayr and Gröschl (2014) control for this by scaling the disaster measures by countries’ land areas, which I also do.
In the empirical analyses, it is usual to only include a single indicator for natural disasters. In this study, I follow the strategy of Felbermayr and Gröschl to construct this measure as the sum of the six individual measures weighted by land area and by the inverse of the standard deviations of the different disaster indices, the latter to control for scale differences. Figure 2 below shows the distribution of values for this aggregated disaster index across the countries in the sample.
only the single most disastrous event of the year counts. This might be aggravated by the fact that weather stations in larger countries may not offer as much coverage in large, sparsely populated areas. In that case, the economic effects rationale behind dividing disaster measures by total land area of countries also becomes weaker.
For these reasons, figure 2 also includes maps for the aggregated disaster index only weighted by inverse standard deviations, the aggregated disaster index only weighted by area, and the unweighted aggregate disaster index. The finding of high intensity measures for natural disasters in Europe persists, though. A look at figure 1 suggests this may be due to relatively many extreme weather and high wind events.
Figure 1 - Disaster measures, physical intensity, GeoMet database, mean of all years (at most 1979-2010) per country. Circles
Figure 2 – Disaster measure, aggregate of physical intensities for the six subcategory disasters, weighted by 0, 1 or 2 factors.
Data from GeoMet Database mean of all years (at most 1979-2010) per country. Circle sizes are based on subgroups that are calculated using the Jenks natural breaks optimization algorithm.
We can juxtapose these observed relative patterns of physical intensity of natural disasters over space with the respective EM-DAT measures of natural disasters. In figure 3 below, I present maps for some main operationalizations of disaster intensity in EM-DAT: the number of recorded “large” disasters1, the number of recorded large disasters weighted by area, the sum of yearly monetary damage incurred from these disasters, and the sum of yearly monetary damage incurred from these disasters weighted by GDP.
From figure 3, it becomes apparent that there indeed seems to be higher reporting probability of large disasters based on their monetary and personal costs, although the painted picture over geographical space is quite similar to figure 2 for the area-weighted “large” disaster measures. The reason for this is that large disasters are probably still reported in lower income countries, and that the reporting probability bias is not as serious as implied earlier. However, taking the sum of monetary damages as explanator for GDP leads to obvious endogeneity bias, as the bottom panels in figure 3 exemplify.
1 In line with Munich Re (2006) and Gassebner and coauthors (2010), a large disaster occurs (1) when 10 or
Figure 3 – Disaster measures, EM-DAT, mean of all years (1979-2010) per country. Circle sizes are based on subgroups that
are calculated using the Jenks natural breaks optimization algorithm.
Finally, in the context of climate change, it is interesting to see whether disasters actually increased over time already. Figure 4 below plots the values for the aggregated physical intensity disaster measure weighted by area and inverse standard deviations for the whole balanced subsample (containing 88 countries) for each year. From this figure, we can conclude that there has not yet been a visible increase in the disaster hazard occurring over the sample period for any of the six disaster types.
Figure 4 – Disasters by disaster type over time. Physical disaster measures, global total. Disasters are weighted by their
3.3.2 – Data: Macroeconomic variables
Literature in the field (Skidmore & Toya, 2002; Noy, 2009; Loayza et al., 2012; Felbermayr & Gröschl, 2014) uses Barro-type growth regressions (see eq. 1), relating GDP per capita growth to some external and domestic policy factors and some structural factors. Indeed, this study attempts to use a set of covariates as close as possible to the study by Felbermayr and Gröschl to explain GDP growth.
Specifically, I source GDP per capita and aggregate trade volume data from the Penn World Tables (Feenstra, Inklaar & Timmer, 2015), and use the World Bank’s World Development Indicators (World Bank Group, 2014) for all remaining covariates (population size, real interest rate as percentage, domestic credit provided by the banking sector as percentage of GDP, gross capital formation as annual percentage growth of capital, foreign direct investment as net inflows relative to GDP, inflation at consumer prices as annual percentage, and the current account balance as share of GDP) except those for the financial openness and polity indices. Namely, following Felbermayr and Gröschl, I use the Chinn-Ito index as measure of financial openness (Chinn & Ito, 2008), while I use the Polity IV index as measure of democracy (Marshall, Gurr & Jaggers, 2014).
A novelty of this study is to consider trade in subsectors of the macro economy. Data for this is taken from COMTRADE (United Nations, 2003). After aggregating this 4-digit level-of-disaggregation sectoral trade data to the 2-digit industrial classification nomenclature, we remain with each country’s exports and imports at the slightly disaggregated level of approximately 30 sectors. Consequently, importance of trade in subsectors is defined as exports in the specific subsector plus imports in the specific subsector divided by total trade. This allows to pinpoint which sectors a country trades relatively much in, which may uncover heterogeneous effects on the disaster interaction depending on the traded sector type.
Summary statistics of all the covariates named in this section as well as the disaster measures named earlier are provided in the appendix, table A1.1. There, a division is also made between the developed and developing halves of the sample (table A1.2). In the following sections, all these variables will be used in at least one empirical specification.
4. Results
This section is divided into three main parts. The first part discusses results of the baseline Barro-type GDP growth regressions, the second part discusses results of the baseline survival analysis. After comparing these results, the final part of this section presents an extensive number of robustness checks.
4.1 – Barro-type GDP growth regressions
In line with Felbermayr and Gröschl, I also include interaction terms with the Chinn-Ito financial openness index and with the polity index. The baseline regression uses GeoMet data, the aggregate disaster indicator that is weighted by area and the inverse of the standard deviations of the individual disaster indicators, and uses the Sachs-Warner measure for trade openness in the interaction between trade and the disaster measure to foster interpretability.
Table 1 – GDP Growth regressions, estimating the effect of natural disasters and the natural-disaster-trade interaction on
GDP growth in the short run. The econometric method OLS FE is more efficient, while Hausman-Taylor better accounts for Nickell (1981) bias. Other than that, different physical disaster intensity measure indicators are used. The disaster measures are taken from Felbermayr & Gröschl (2014)’s GeoMet data of physical intensities and are thus exogenous.
VARIABLES OLS FE Hausman-Taylor
GMM (1)
(preferr ed)
(2) (3) (4) (5) (6)
𝐷𝑖,𝑡, area- and inv.sd. -weighted -2.27***
(.734) -2.27*** (.756) 𝐷𝑖,𝑡, area-weighted -2.11** (.906) -2.11** (.933) 𝐷𝑖,𝑡, inv.sd. -weighted -.002* (.001) 𝐷𝑖,𝑡, unweighted -.0002* (.0002) Trade openness, t-1: (Imp + Exp) / GDP 0.040*** (.012) .039*** (.012) .040*** (.012) .037*** (.011) .040*** (.012) .039*** (.012) (Trade openness, t-1) × 𝐷𝑖,𝑡 : (Sachs-Warner) 1.12*** (.010) 1.15* (.579) .001 (.001) -.0001 (.0002) 1.20** (.469) 1.15* (.596) ln GDP per capita, t-1 -.090*** (.015) -.092*** (.016) -.089*** (.016) -.086*** (.016) -.090*** (-.016) -.092*** (.017) Ln population, t-1 -.053** (.028) -.054* (.029) -.065** (.033) -.067** (.033) -.053* (.029) -.053* (.030) Polity, t-1 -.004 (-.028) -.003 (.015) .012 (.020) -.003 (.018) -.004 (.014) -.003 (.015) (Polity, t-1) × 𝐷𝑖,𝑡 2.00*** (.730) 1.81** (.815) .0001 (.844) .0002 (.0002) 2.00*** (.751) 1.81** (.840) Interest, t-1 -.080 (.157) -.096 (.151) -.057 (.170) .086 (.181) -.080 (.162) -.096 (.156) Domestic credit, t-1 -.036*** (.008) -.036*** (.008) -.034*** (.007) -.038*** (.008) -.036*** (.008) -.036*** (.008) Gross capital formation, t-1 .020***
(.008) .019** (.008) .022*** (.008) .025*** (.008) .020** (.008) .019** (.008) FDI, t-1 .035 (.043) .035 (.045) .059 (.050) .059 (.051) .035 (.044) .035 (.046) Ln Inflation, t-1 -.004* (.002) -.003* (.002) -.003* (.002) -.003** (.002) -.003** (.002) -.003* (.002) Current account balance, t-1 .063*
(.036) .057 (.036) .077** (.038) .075** (.038) .063* (.037) .057 (.037) Financial openness, t-1 .008 (.009) 010 (.010) .023 (.018) .016 (.010) .008 (.009) .010 (.010) (Financial openness, t-1) × 𝐷𝑖,𝑡 .244*** (.071) .249*** (.084) .0002 (.0003) .0001 (.0003) .244*** (.073) .249*** (.086) Constant 1.28*** (.335) 1.30*** (.358) 1.39*** (.396) 1.49*** (.386) 1.50*** (.457) 1.52*** (.481) Adjusted 𝑅2 0.289 0.290 0.275 0.279 - -
Note: The dependent variable is ∆ ln GDP per capita. There are 1,902 observations from 108 different countries. Country- and time-fixed effects included in every model, but their coefficients are omitted from this table. Standard errors are reported in parentheses and are clustered at the country level. * p<.1, ** p<.05, *** p<.01
estimates, implying that there is indeed no correlation between country-specific effects and the covariates. Given this fact, the use of the Hausman-Taylor estimator is slightly less efficient, as exemplified by the slightly higher standard errors in these estimates.
Some final econometric issues regarding this estimation are whether the data is stationary, the potential bias relating to the use of a lagged endogenous variable on the right-hand side of the panel regression equation (Nickell, 1981), and the unbalanced nature of the sample. The sample also contains one gap: the country of Kuwait was missing observations for the years 1990 and 1991, which is likely a result of military conflict. Because the Kuwait sample only started midway through the 1980s anyways, I dropped all observations on Kuwait before 1992. The unbalanced nature of the sample should not be a large issue, considering most the sample (about 80%) has observations on all time periods.
Nickell bias should also not be severe, as previous studies have argued it is small in panels with long time dimensions (Judson & Owen, 1999). Finally, I run the Levin-Lin-Chu (2002) unit root test on the balanced subsample of 88 panels containing observations on all periods. I reject the null hypothesis that panels contain unit roots (p < .001). A weaker test for unbalanced subsamples is the Im-Pesaran-Shin (2003) test. Again, on the subsample of 107 countries with more than 9 observations, I reject the null hypothesis, this time being that all panels contain unit roots (p < .001). Indeed, these results imply that the data are covariance stationary.
Finally, it is worth noting that these results are quite volatile. This already becomes apparent from using non-area-weighted disaster measures, which decreases estimated coefficients for the effect of disasters so strongly that the interaction effects involving them are no longer significant. Note, however, that these outcomes are likely invalid from the use of non-area-weighted disaster scores overstating the impacts of disasters in larger countries where nothing was felt. The stylized fact that larger countries tend to have lower openness to trade corresponds to this.
4.2 – Survival Analysis
takes seven years (1979 to 1985). Results of the survival analysis are presented in table 2 below. As explained in the previous section, failure states are defined as a combination of a drop in GDP growth and a strong natural disaster. To investigate drops in GDP growth, in the baseline estimation, I take a moving average of GDP and compare the observed GDP with that of the moving average. A strong natural disaster is exemplified by a top 10 percentile value for the area- and standard-deviation-weighted disaster measure, which turns out to be at a value of about 0.04. The six columns of table 2 represent different operationalizations of this moving average.
Although findings regarding the coefficient of trade openness are quite robust across different operationalizations of the failure state, especially compared to other included explanators of the time it takes to reach the failure state, and although these coefficients on trade openness are of the expected sign – that is, a positive sign which indicates that higher trade openness relates to a longer time until a failure, in this case a damaging natural disaster, occurs – there is no statistical significance. This is most likely due to small sample size. Indeed, p-values on this coefficient range between the 0.1 and the 0.4 mark, and standard errors may decrease further as the sample size increases.
Nonetheless, the finding contrasts with the findings of significant interaction effects in the Barro-type growth regressions of previous literature (e.g. Noy, 2009). This result has important implications. Namely, the coefficient on trade in the survival analysis says something about the total effect of trade openness on reaching a natural disaster that hurts GDP to a certain extent – that is, it encapsulates both direct effects through sectoral composition shifts, as well as the indirect channels of trade in the disaster-GDP relation through spillovers and better aid distribution. In the case of Barro-type growth models, arguing that such effects are captured by the trade or trade interaction term is more contrived. In other words, this result of insignificance suggests that, indeed, there is strong ambiguity in the way trade interacts with disaster risk mitigation at the macroeconomic level.
and 6 repeat the failure states of columns 1, 2 and 3 respectively, but presume that the disaster has to happen in the previous year, not the current year, to contribute towards being a failure state. I find that changing this does not strongly alter the results.
Table 2 – Cox proportional hazards model with shared frailty – estimated coefficients (note: not Hazard ratios)
VARIABLES (1) (2)
(preferred)
(3) (4) (5) (6)
Is failure when both 𝐷𝑖,𝑡 > 0.04 and
𝑦𝑡 is smaller than…
Is failure when both 𝐷𝑖,𝑡 > 0.04
and 𝑦𝑡+1 is smaller than… 𝑦𝑡−1+ 𝑦𝑡+1 2 ∑ 𝑦𝑡+𝑗 6 3 𝑗= −3 −𝑦𝑡 6 𝑦𝑡−1 𝑦𝑡+ 𝑦𝑡+2 2 ∑ 𝑦𝑡+𝑗 6 4 𝑗= −2 −𝑦𝑡+1 6 𝑦𝑡 Trade openness: (Imp + Exp) / GDP .680 (.511) 1.074 (.717) .822 (1.06) .621 (.536) .501 (.595) .879 (.875) Polity 1.89 (1.18) 2.33* (1.23) 2.79 (2.00) .951 (1.15) .464 (1.14) .928 (1.39) Interest 5.00 (20.7) 11.02 (20.69) 52.7 (35.5) -8.19 (20.3) 34.8 (21.3) 15.6 (26.3) Domestic credit .321 (.651) .402 (.699) .076 (1.40) .404 (.702) .524 (.733) -.164 (1.06) Gross capital formation -1.55* (.811) -1.12 (.847) -3.82*** (1.47) -.107 (.772) -.200 (.767) .090 (1.31) Foreign direct investment -2.68 (4.32) 2.50 (4.41) 4.29 (9.12) -1.42 (3.49) 3.12 (3.45) 6.46 (4.99) Ln Inflation -.159 (.141) -.113 (.135) -.397 (.323) .015 (.140) -.026 (.139) .176 (.299) Theta 9.639 (3.244) 9.629 (3.341) 13.03 (6.778) 13.58 (4.669) 12.92 (4.760) 9.098 (4.291) LR test of theta (𝝌̅𝟐) (p-value in parentheses) 161.86 (0.000) 162.8 (0.000) 48.28 (0.000) 189.84 (0.000) 162.04 (0.000) 45.18 (0.000) No. of subjects 107 108 108 108 108 108 No. of failures 93 90 27 95 92 29 Time at risk 1917 1963 1963 1963 1963 1963 Log-likelihood -250.53 -240.93 -72.39 -262.96 -245.31 -77.78 Note: Standard errors are reported in parentheses and are conditional on theta. * p<.1, ** p<.05, *** p<.01. If the LR test of Theta is significant, there is frailty that needs to be accounted for, as is done in each of the columns here. Estimates on the year-interval fixed-effects are omitted from this table but were entered in the estimation procedure.
values for what constitutes a large enough disaster, and through different cut-offs and cut-off operationalizations for impacts on GDP. Finally, I add to the literature by considering sector-level interaction effects between trade on the one hand and the disaster-GDP relation on the other hand.
4.3.1 – Robustness checks: Developed and developing countries
The first robustness check splits the sample into developed and developing countries. Specifically, I do this because of a possible nonlinear effect of trade openness on the natural-disaster-GDP relation. Results are presented in table 3 below. Developing countries are identified through the World Bank income groups as not belonging to the high-income category.
Table 3.1 – GDP Growth regressions, estimating the effect of natural disasters and the natural-disaster-trade interaction on
GDP growth in the short run, dividing the sample in developed and developing countries.
Panel 1: Barro-type OLS FE
VARIABLES (1)
(Developing)
(2) (Developed) 𝐷𝑖,𝑡, area- and inv.sd. -weighted -3.64** (1.51) -.877 (.626)
Trade openness, t-1: (Imp + Exp) / GDP .051*** (.017) .050*** (.011) (Trade openness, t-1) × 𝐷𝑖,𝑡 : (Sachs-Warner) .055 (1.50) .350 (.356)
ln GDP per capita, t-1 -.107*** (.020) -.114*** (.013) Ln population, t-1 -.135** (.040) .016 (.048) Polity, t-1 -.018 (.012) .026 (.021) (Polity, t-1) × 𝐷𝑖,𝑡 3.49*** (1.26) .591 (.644) Interest, t-1 -.061 (.150) -.452 (.462) Domestic credit, t-1 -.036*** (.013) -.019** (.009)
Gross capital formation, t-1 .014 (.009) .038** (.018)
FDI, t-1 .038 (.056) -.001 (.032)
Ln Inflation, t-1 -.003 (.002) .001 (.003)
Current account balance, t-1 .068 (.046) .094** (.038)
Financial openness, t-1 .023* (.012) 009 (.013) (Financial openness, t-1) × 𝐷𝑖,𝑡 -.567 (.425) .290*** (.099) Constant 2.114*** (.459) 1.30*** (.358) Number of observations 1189 625 Number of countries 70 31 Adjusted 𝑅2 0.269 0.263
Table 3.3 – Cox proportional hazards model with shared frailty – estimated coefficients (note: not Hazard ratios), sample
split into developing and developed countries.
Panel 2: Survival analysis
VARIABLES (1) (Developing) (2) (Developed) Trade openness: (Imp + Exp) / GDP 2.02 (3.06) .865 (.818) Polity 3.74 (2.48) .112 (1.80) Interest -13.0 (32.3) 17.8 (33.1) Domestic credit -2.27 (3.57) .318 (.774)
Gross capital formation -2.97
(1.91)
-.351 (1.09)
Foreign direct investment .595
(10.99) 3.30 (6.11) Ln Inflation -.221 (.327) .035 (.177) Theta 24.19 (17.68) 4.29 (1.92) LR test of theta (𝝌̅𝟐) (p-value in parentheses) 46.98 (0.000) 60.41 (0.000) No. of subjects 75 33 No. of failures 30 60 Time at risk 1293 670 Log-likelihood -53.23 -125.27
Note: A failure is taken to be the case where a strong (𝐷𝑖,𝑡> 0,04) disaster (physical intensity, aggregate index, weighted by
standard deviations and by area) hits and the same year sees a decrease in GDP compared to the 3-period forward-and-back moving average. Standard errors are reported in parentheses and are conditional on theta. * p<.1, ** p<.05, *** p<.01. If the LR test of Theta is significant, there is frailty that needs to be accounted for, as is done in each of the columns here. Estimates on the year-interval fixed-effects are omitted from this table but were entered in the estimation procedure.
An interesting finding regarding the Barro-type growth regressions is also that I find no significant effect anymore of natural disasters on GDP growth in the developed-country subsample. This implies that natural disasters have very little impact on growth in more developed countries. In line with theory, the interaction effect of democracy and natural disaster risk mitigation also depends on whether the country is still developing. Namely, the coefficient on this interaction term is no longer significantly positive in the developed country sample. Results from the survival analysis also imply this. Although again finding insignificant coefficient estimates, I find higher estimates for the effects of trade openness and polity on the time-till-failure in the developing-country subsample than in the developed-country sample.
4.3.2 – Robustness check: EM-DAT data
The second robustness check I perform is to repeat the analyses using EM-DAT data instead of GeoMet data. The disaster measures I use are twofold. First, I use area-weighted total number of “large” natural disasters, large again defined by the decision rule of enough people affected or enough damages claimed. Second, I use the GDP-weighted total yearly disaster costs. The exercise is done to prove that using EM-DAT data suffers from severe endogeneity bias, and results can be taken with a grain of salt. Therefore, I only present these results in table A2 of the appendix. Indeed, I find results using EM-DAT that contradict most of the previous empirical and theoretical background, finding no significant growth effects of natural disasters and even positive interaction effects of trade openness.
4.3.3 – Robustness check: heterogeneous effect of disasters
Table 4 – GDP Growth regressions, estimating the effect of natural disasters and the natural-disaster-trade interaction on
GDP growth in the short run, focusing on specific disaster types. The disaster measures are taken from Felbermayr & Gröschl (2014)’s GeoMet data of physical intensities and are thus exogenous.
VARIABLES (1) (2) (3) (4) (5) (6) 𝐷𝑖,𝑡 = Earthquakes (area-weighted) -37.4*** (9.86) 𝐷𝑖,𝑡 = Volcanic eruptions (area-weighted) .083 (.117) 𝐷𝑖,𝑡 = Storms (area-weighted) -8.64** (3.90) 𝐷𝑖,𝑡 = Extreme temperature events (area-weighted) -.027 (.152) 𝐷𝑖,𝑡 = Droughts (area-weighted) -.032 (.068) 𝐷𝑖,𝑡 = Floods (area-weighted) -.1.08 (.807) Trade openness, t-1: (Imp + Exp) / GDP .040*** (.012) .040*** (.012) .040*** (.012) .041*** (.012) .040*** (.012) .040*** (.012) (Trade openness, t-1) × 𝐷𝑖,𝑡 : (Sachs-Warner) 16.6*** (5.59) .009 (.031) 4.55* (2.46) .097 (.098) .086* (.046) .151 (.537) ln GDP per capita, t-1 -.090*** (.016) -.088*** (.017) -.092*** (.016) -.088*** (.017) -.088*** (.016) -.084*** (.015) Ln population, t-1 -.056* (.029) -.060* (.032) -.050* (.028) -.061* (.032) -.060* (.032) -.055* (.028) Polity, t-1 .006 (.013) .017 (.014) -.001 (.014) .016 (.013) .016 (.013) .016 (.014) (Polity, t-1) × 𝐷𝑖,𝑡 39.0 (11.1) -.064 (.129) 7.53** (3.54) .069 (.157) .018 (.073) .329 (.946) Interest, t-1 -.104 (.154) -.115 (.156) -.113 (.147) -.122 (.155) -.117 (.157) -.113 (.156) Domestic credit, t-1 -.034*** (.008) -.033*** (.008) -.034*** (.008) -.033*** (.008) -.033*** (.008) -.035*** (.008)
Gross capital formation, t-1 .020** (.008) .019** (.008) .019** (.008) .019** (.008) .019** (.008) .019** (.008) FDI, t-1 .038 (.045) .053 (.049) .031 (.044) .054 (.049) .052 (.049) .044 (.046) Ln Inflation, t-1 -.003* (.002) -.003* (.002) -.003* (.002) -.003* (.002) -.003* (.002) -.003* (.002)
Current account balance, t-1 .069* (.038) .077* (.039) .059 (.037) .078** (.039) .075* (.039) .068* (.039) Financial openness, t-1 .009 (.009) .016 (.010) .009 (.009) .017 (.009) .016 (.010) .010 (.010) (Financial openness, t-1) × 𝐷𝑖,𝑡 6.03*** (2.15) .004 (.061) 1.07*** (.365) -.184* (.110) .024*** (.009) .904*** (.329) Constant 1.29*** (.341) 1.31*** (.387) 1.26*** (.353) 1.32*** (.390) 1.31*** (.384) 1.24*** (.334) Adjusted 𝑅2 .287 .277 .289 .277 .277 .281
In the context of the survival analysis, dividing disasters into subcategories is difficult, because there is too little variation for finding the global maximum likelihood values that solve the model. Therefore, I do not extensively present results for the individual-disaster-types within the survival analysis case. The relatively high significance of the coefficients on the direct and interactive effects of earthquakes in table 4, though, warrant a survival analysis using just this case. However, again, using the six-year moving average of GDP and the top 10 percentile of earthquake intensities to find cases where disasters actually hurt GDP, I find no significant coefficient of trade openness (𝛽𝑇𝑂 = 1.69, 𝜎𝑇𝑂= 1.39).
4.3.4 – Robustness check: heterogeneous effect of traded goods
Perhaps the effect of trade openness changes in the context of which goods are traded. This relates back to the rationale explained earlier, that trade openness may leave a country more vulnerable if they specialize in sectors that are more at risk from suffering negatively from natural disasters (Benson & Clay, 2004). They specifically name the example of developing countries shifting towards riskier industries in agriculture and manufacturing. Using the sectoral trade data I added to the sample, I test for this by (1) adding the sectoral trade importance measures, defined as sector-level imports plus sectoral exports divided by total trade, as interaction-with-disaster covariates to the Barro-type regression model, and (2) by adding this sectoral trade importance measure to the survival analyses as covariate within 𝑥𝑖′, the vector of covariates at the right-hand side of the Cox proportional hazards model.
Note that only one interaction of trade importance and the disaster measure enters each estimation because trade importance interactions are highly collinear, and it should not matter to the conclusion here that each of the terms is estimated separately. Table 5 below presents results by sector in the Barro-type model as well as in the survival analysis model. The most striking result from this is that, indeed, the estimated effects are heterogeneous across sectors.
until a disaster that does visible damage to GDP strikes. In fact, the survival analysis finds a strong negative effect of the importance of logging products in the trade mix and the length of time until a damaging disaster hits. Apparently, trading in logging products is associated with higher damages as result of natural disasters, which makes sense considering the sensitivity of the logging industry to natural disasters. Moreover, it could also signal the direct impact that logging has on natural disasters occurrence (Kahn, 2005) through the effect of deforestation on flood risk.
Another interesting finding is that there is a negative effect of prominence in trading of manufactured transport equipment on GDP in the case of disasters (sector number 384). Finally, I find a positive effect of prominence in trading of manufactured furniture on GDP in the case of disasters (sector number 332). Both these cases align well with predictions by Benson and Clay (2004), and especially the estimate for furniture aligns well with predictions by Strulik and Trimborn (2019).
4.3.5 – Robustness check: Further alternatives for the ‘failure state’ in survival analysis
In table 2, I juxtaposed different operationalisations for what constituted a ‘failure’ in the estimated survival models. Specifically, I compared the following moving averages: one-period back and one-one-period forward; three-one-periods back and three-one-periods forward; one-one-period back; and the remaining three moving averages were the respective equals of these, but with the assumption that disasters affected GDP growth in the subsequent period, rather than the contemporary period. Estimates did not differ much across these cases.
However, we can alter the operationalizations of the failure state further. Namely, first, we can choose the cut-off value for the physical severity of a natural disaster to be higher or lower. Second, we can decide to require more (or less) than just any positive difference between predicted GDP and observed GDP. Third, we can estimate whether there had been a decrease in GDP in a different way than by taking a moving average. Indeed, in this final section of the robustness checks, I explore all three of these channels.
Table 6 – Explanation of operationalizations of failure state to be used in the subsequent table. Note: Di,t is the aggregate GeoMet disaster index weighted by inverse standard deviation of the individual disaster measures and by land area.
Failure state number Disaster impact threshold Predicting GDP in current period method Comparison of GDP threshold Baseline (see column 2, table 2) 𝐷𝑖,𝑡 > 0.04 (90th percentile) ∑ 𝑦𝑡+𝑗 6 3 𝑗= −3 −𝑦𝑡 6 (six-period moving average) 𝑦𝑡 < [∑ 𝑦𝑡+𝑗 6 3 𝑗= −3 − 𝑦𝑡 6]
(observed 𝑦𝑡 below the prediction of
GDP from the previous column) (I) 𝐷𝑖,𝑡 > 0.065
(95th percentile) Same as baseline Same as baseline (II) 𝐷𝑖,𝑡> 0.019
(80th percentile) Same as baseline Same as baseline (III) Same as baseline
Forecast using AR(4) regression
of GDP growth
Same as baseline, although different prediction of GDP for comparison
(IV) Same as baseline Same as baseline
𝑦𝑡 < [∑ 𝑦𝑡+𝑗 6 3 𝑗= −3 − 𝑦𝑡 6] − 𝑦𝑡 200
(observed 𝑦𝑡 below the six-period
moving average minus an additional 0.5% of total GDP)
(V) Same as baseline Same as baseline
𝑦𝑡 < [∑ 𝑦𝑡+𝑗 6 3 𝑗= −3 − 𝑦𝑡 6] + 𝑦𝑡 200
(observed 𝑦𝑡 below the six-period
moving average plus an additional 0.5% of total GDP)
Indeed, this robustness check proves that the finding of no significance is not due to the operationalization of the failure state. Namely, both increasing and decreasing the number of failures does not lead to more significant estimates. This makes sense: at increased rates of failure, we are no longer truly identifying shocks to GDP as a result of natural disasters, and this rationale holds also for slight increases in GDP or more irrelevant natural disasters up to the 80th percentile, as is clear from the estimates on trade openness. On the other hand, decreasing the number of failures runs into a too low number observations with failures, which strongly increases standard errors.
Table 7 – Cox proportional hazards model with shared frailty – estimated coefficients (note: not Hazard ratios) – Different
specifications of what constitutes a ‘failure’ in the survival model.
VARIABLES (Baseline) (I) (II) (III) (IV) (V)
Trade openness: (Imp + Exp) / GDP 1.074 (.717) 1.153 (.900) .690 (.475) .432 (.682) .806 (1.60) .502 (.498) Polity 2.33* (1.23) 3.22 (2.06) .951 (.648) 3.04 (1.93) 1.42 (1.89) .417 (1.04) Interest 11.0 (20.7) 57.99 (37.63) -3.85 (13.20) 29.4 (27.3) 33.4 (33.2) -1.65 (14.6) Domestic credit .402 (.699) .107 (1.42) .097 (.465) -2.97** (1.25) .990 (1.60) -.049 (.576) Gross capital formation -1.12 (.847) -.627 (1.21) -.651 (.494) -1.38 (1.37) -2.98** (1.46) -.716 (.579) Foreign direct investment 2.50 (4.41) 4.61 (6.51) -.645 (2.78) -3.09 (8.29) 14.0 (9.57) .792 (2.63) Ln Inflation -.113 (.135) -.260 (.232) -.172* (.091) -.127 (.301) -.422 (.295) -.073 (.108) Theta 9.629 (3.341) 19.9 (10.3) 6.34 (2.25) 18.5 (8.97) 17.3 (9.04) 15.6 (4.82) LR test of theta (𝝌̅𝟐) (p-value in parentheses) 162.8 (0.000) 91.8 (0.000) 223.1 (0.000) 72.0 (0.000) 58.5 (0.000) 395.6 (0.000) No. of subjects 108 108 108 108 108 108 No. of failures 90 45 195 38 29 174 Time at risk 1963 1963 1963 1963 1963 1963 Log-likelihood -240.93 -89.57 -622.02 -77.43 -75.85 -428.86 Note: Standard errors are reported in parentheses and are conditional on theta. * p<.1, ** p<.05, *** p<.01. If the LR test of Theta is significant, there is frailty that needs to be accounted for, as is done in each of the columns here. Estimates on the year-interval fixed-effects are omitted from this table but were entered in the estimation procedure.
5. Conclusion
The study has also set out an empirical method that is novel to the field of study to quantify the effects of trade openness on coping with natural disasters. Although data limitations prevent a fully successful implementation of this empirical method, the benefits of setting up an empirical model like this are clear: one can endogenize physical environmental factors related to how climate change induces increases in the natural disaster hazard in the model setup, while the model is less assumptive than traditional growth regressions in terms of measuring ex ante factors interacting with disaster risk, such as trade openness.
The findings will also be of interest to policymakers at the country level with regards to their trade policy, especially when their country suffers from high disaster risk. For example, Klomp and Hoogezand (2018) find that domestic agricultural protectionism increases after natural disasters. On the other hand, I find out in this study that, even if agriculture is a major part of the trade mix of a country, it does not affect growth negatively in the context of a natural disaster. This implies that such agricultural protectionism is, indeed, misguided. More generally, the study finds the first suggestive evidence that, also at the macroeconomic country level, openness to trade can function as risk-sharing mechanism across geographical space.
Nonetheless, considerably more work will need to be done to determine what exactly drives whether trade ‘works’ as a mitigative mechanism or not at the country level. A fruitful area for future research is then, perhaps, the development of more microeconomic models of geographical interaction to the country level. This work is important to understand not just that trade interacts with natural disaster impacts, but also how exactly it interacts with natural disaster impacts.
The finding of little significance is, as indicated at the end of the results section, likely attributable also to the fact that there is too little variation in reported GDP as result of natural disasters. Indeed, then, one could study an analogous mechanism at smaller geographical level, for example by defining trade endogenously as the potential market access from different locations and how this market access interacts with impacts from natural disasters.
for such events. Then, a natural progression of this work is to add expectations somewhere in the empirical models, for example using disaster risk data from insurance companies.
Second, there is one yet unnamed but potentially severe source of upward bias in coefficients on mitigative effects. Namely, using officially reported GDP figures generates selection bias, as the reliability of the data is positively related to the level of GDP. The problem is aggravated in the context of natural disasters, because the shadow economy is an important contributor to economic activity in developing countries. Moreover, shadow economic activities are often precisely among the most vulnerable to certain natural disasters (Loayza et al., 2012), for example when the activity is subsistence farming.
Following the recent idea (e.g. Chen & Nordhaus, 2011; Henderson, Storeygard & Weil, 2012) to use nighttime luminosity satellite data as proxy for GDP to account for these biases, Klomp (2016) estimates that the effect of natural disasters on nighttime light growth is much more strongly negative than the estimated effect of natural disasters on GDP per capita growth. This suggests that not accounting for bad GDP data quality in developing countries positively biases the estimated effects of natural disasters. Although this implies that the persistence of insignificant estimates in the literature is less concerning, which strengthens the results of this study somewhat, it is desirable to integrate this idea in future work.
References
Allen, T., & Atkin, D. (2016). Volatility and the Gains from Trade. NBER Working Paper 22276. Arto, I., & Dietzenbacher, E. (2014). Drivers of the growth in global greenhouse gas emissions.
Environmental science & technology, 48(10), 5388-5394.
Banholzer, S., Kossin, J., & Donner, S. (2014). The impact of climate change on natural disasters. In Z. Zommers & A. Singh (Eds.), Reducing disaster: early warning systems for climate change (pp.
21–49). Berlin: Springer Publishers.
Barro, R. J. (1991). Economic growth in a cross section of countries. The quarterly journal of
economics, 106(2), 407-443.
Benson, C., & Clay, E. (2003). Disasters, vulnerability and the global economy. Building safer cities:
The future of disaster risk, 3, 3-32.
Benson, C., & Clay, E. (2004). Understanding the economic and financial impacts of natural disasters. The World Bank.
Van den Berg, G. J., & Drepper, B. (2016). Inference for shared-frailty survival models with left- truncated data. Econometric Reviews, 35(6), 1075-1098.
Berlemann, M., & Wenzel, D. (2018). Hurricanes, economic growth and transmission channels: Empirical evidence for countries on differing levels of development. World Development, 105, 231-247.
Bertinelli, L., & Strobl, E. (2013). Quantifying the local economic growth impact of hurricane strikes: an analysis from outer space for the Caribbean. Journal of Applied Meteorology and
Climatology, 52(8), 1688-1697.
Chen, X., & Nordhaus, W. D. (2011). Using luminosity data as a proxy for economic statistics.
Proceedings of the National Academy of Sciences, 108(21), 8589-8594.
Chinn, M. D., & Ito, H. (2008). A new measure of financial openness. Journal of comparative policy
analysis, 10(3), 309-322.
Costinot, A., Donaldson, D., & Komunjer, I. (2011). What Goods Do Countries Trade? A Quantitative Exploration of Ricardo’s Ideas. The Review of Economic Studies, 79(2), 581–608.
Costinot, A., Donaldson, D., & Smith, C. (2016). Evolving comparative advantage and the impact of climate change in agricultural markets: Evidence from 1.7 million fields around the world.
Journal of Political Economy, 124(1), 205-248.
Daniell, J., Wenzel, F., & Schaefer, A. (2016). The economic costs of natural disasters globally from 1900-2015: historical and normalised floods, storms, earthquakes, volcanoes, bushfires, drought and other disasters. [In EGU General Assembly Conference Abstracts, vol. 18.] Based on CATDAT Database: see Landgraf, M. (2016). Natural Disasters since 1900: Over 8 Million Deaths and 7 Trillion US Dollars damage. Karlsruhe Institute of Technology Press Release, 058/2016.
Feenstra, R.C., Inklaar, R., & Timmer, M. (2015). The Next Generation of the Penn World Table.
American Economic Review, 105(10), 3150-3182. Downloadable at www.ggdc.net/pwt.
Felbermayr, G., & Groeschl, J. (2014). Naturally negative: the growth effects of natural disasters.
Journal of Development Economics 111, 92-106.
Frankel, J. A., & Rose, A. K. (2005). Is trade good or bad for the environment? Sorting out the causality. Review of economics and statistics, 87(1), 85-91.
Gassebner, M., Keck, A., & Teh, R. (2010). Shaken, Not Stirred: The Impact of Disasters on International Trade. Review of International Economics, 18(2), 351–368.
Hamano, M., & Vermeulen, W.N. (2019). Natural Disasters and Trade: The Mitigating Impact of Port Substitution. Journal of Economic Geography. https://doi.org/10.1093/jeg/lbz020
Hausman, J. A., & Taylor, W. E. (1981). Panel data and unobservable individual effects. Econometrica:
Journal of the Econometric Society, 1377-1398.
Henderson, J. V., Storeygard, A., & Weil, D. N. (2012). Measuring economic growth from outer space.
American economic review, 102(2), 994-1028.
Hsiang, S.M., & Jina, A.S. (2014). The Causal Effect of Environmental Catastrophe on Long-Run Economic Growth: Evidence From 6,700 Cyclones. NBER Working Papers 20352.
Im, K. S., Pesaran, M. H., & Shin, Y. (2003). Testing for unit roots in heterogeneous panels. Journal of
