Quantitative Research Paper: The impact of natural disasters on GPD: a global review
Introduction
Natural disasters have been a feature of life for millennia and have caused countless deaths, injuries and damage to life and property. Each disaster event is unique in terms of its cause, location, magnitude that creates unique disaster exposures for each locations. At a human level the consequences of a disaster event are in part determined by the social and economic characteristics of the people affected (CARE 2009). Their access to information about a coming event such as a cyclone, their ability to understand it, their capacity to take action and the resources available to them to cope are part of a complex picture of human disaster vulnerability. (IFRC 2008) A national economy is also affected in different ways – directly in terms of damage and costs to recover and indirectly in terms of lost productivity and lost human capital. (Okuyama 2008)
In 2011 332 natural disasters killed 30,773 people and caused the highest ever estimated damages of US$ 366.1 billion. The same a small number of countries carry the main burden of disaster impacts year after year. Whilst eight out of the top ten countries for total deaths from disasters were in Asia, for total deaths as a proportion of total population African countries dominate. (Guha-Sapir et al 2012)
People around the world are already experiencing massive impacts from natural disasters and impacts are felt at a global, national and human scale. Increasing population, increasing wealth and assets and increasing urbanisation are also contributing to increasing impacts in recent decades (Baker 2012). And, with climate change looking likely to cause more unpredictable weather and climate hazard events, impacts from these will also increase. (IPCC 2013)
From an economic point of view a natural disaster can be described as “a shock that leads to a decline in human, social and physical capital stocks as well as a fall in economic activities” . (Sadeghi 2009 p342) In this paper we explore the impact of natural disasters on GDP and consider a range of related variables at a global scale, using continuous cross-sectional data for GDP in 2012.
What follows is a brief literature review of approaches to the economic analysis of the impact of disasters, our data sources and data limitation, analysis results and conclusions. The hypothesis being explored is whether, as the literature suggests, disasters have a positive impact on GDP.
Literature review
The economics of natural disasters is a new area of research so, as Noy says, it is in its infancy (2009, p 222). Noy identifies two areas where there has been more research done: research into case studies of particular events, and the ways people and households prepare and cope with shocks. The latter area of research relates to work in the disaster risk reduction field of humanitarian practice that looks at ways to strengthen resilience and the human scale. (Mercer 2010)
There are, though, not many examples of multi country, multi disaster event research and there do not appear to be established methodologies. The earliest effort found was an analysis of the macroeconomic impact of disasters (Albala-Bertrand in 1993) based on various economic indicators before and after 28 disasters in 26 countries during 1960–1979. Albala-Bertrand found that, if looking at disaster losses as proportion of GDP, there was a slight positive impact on long term growth. He asserts two main reasons for this: disaster loss calculations are normally focused on ‘crude’ measures of the immediate loss of assets and loss of life and imply an exaggeration of the long- term impact; and the full economic impact is not captured because the poorest who are least active in the formal economy bear the brunt of disasters. He concludes that research is needed these gaps and on the contribution of underlying economic and social factors that mean the economy can recover beyond just disaster loss data. (Lewis 1994)
The potential positive economic impact was given further attention by other researchers such as Tol and Leek (1999) who assert that this can be in part explained because GDP measures new production and reconstruction involves expenditure. In other words GDP measures the investment in the replacement of capital not damages to existing stocks of capital. Thus, looking at the short-term economic impacts of disasters and immediate recovery is not sufficient to understand the real impact of disasters on economic growth.
There are some examples of research into longer-term economic consequences of disasters. Skidmore and Toya (2002) look at the frequency of natural disasters from 1960–1990 for each country, and also factor in land size. They find that higher frequencies of events correlate to GDP growth. They separate their analysis into climate and geological events and find a greater positive impact on climate events. They assert that this difference is because there is more chance to prepare for climate hazards than for geohazards such as earthquakes. Hallegatte and Dumas 2009 consider the contribution of the ‘productivity effect ’ to the long term positive impact, in other words that that replacement infrastructure is newer and more efficient than what was destroyed but find that it is not a fully effective descriptor of long term growth.
Another aspect in the literature is research into the disaster–income–safety relationship. This theory states that as income increases the demand for safety also increases, and a higher income means individuals and countries can take more expensive risk actions. (Toya and Skidmore 2007) Disaster exposure was a factor in their investment in safety measures as well as wealth. Schumacher and Strobl (2011) looked at the relationship between disaster damages and wealth, and found an inverse U shaped relationship between losses and wealth in low to medium exposure countries, and a U shaped relationship for high exposure countries. This follows the narrative from Toya and Skidmore (2007) that as countries become wealthier they can invest more in safety measures, but that at some point further investments cannot further reduce the impact of disasters.
As noted by Albala-Bertrand (in 1993) the underlying “social/economic fabric that increases safety for all of society” (Toya and Skidmore 2007). This “underlying fabric” and two of the factors affecting economic impacts explored in Noy 2009 and Toya and Skidmore 2007 are considered further in our discussion on selection of independent variables for this research.
Hallegatte and Dumas 2009 suggest that further research into the long term impact of disasters on economic growth is an important area in part to better understand interventions to address this trap.
In this paper we explore the link between natural disasters and long run GDP, and include variables that reflect the “underlying fabric” of the countries in the analysis.
Data sources and limitations
To analyse the impact on long run economic growth, GDP was chosen as the dependent variable, and we use a cross section of annual country data for 2013. Whilst other researchers have used GDP growth (such as Toya and Skidmore 2002) we are using a cross section of GDP rather than time series growth data. It was sourced from the World Bank Data Bank. The Development Data Group of the bank coordinates a database that draws much of its data from the statistical systems of member countries, so the quality of the data depends on these national systems.
To get a picture of disaster exposure EM-DAT, the Emergency Events Database maintained by the Centre for Research on the Epidemiology of Disasters (CRED) since 1988 (Guha-Sapir 2014) was used. Whilst this is the best available data on disasters worldwide, some authors have suggested there are limitations in the data. Damages are only direct costs and not indirect costs of the disaster; there may be an “incentive to exaggerate damages in order to secure international assistance” and accurate damage estimates from developing countries is more difficult in developed countries because insurance cover is less extensive and informal markets are more significant. (Toya Skidmore 2007 p22, Noy 2009)
In relation to the use of the disaster data, initial plans were to just consider total disaster losses as a measure of the magnitude of an event but from reading the literature additional indicators of disasters severity were added (people affected and number of events). Noy (2009) and Skidmore and Toya (2002) identify the potential existing relationship between disaster losses and GDP, meaning that greater GDP could cause greater losses so it is not a good measure of disaster magnitude. They refer to this as exogeneity but this is not explored here.
Drawing on the literature, there were a lot of other independent variables to consider. As noted, several authors have highlighted the underlying factors that contribute to the resilience of an economy (or people) to a hazard event. (IFRC 2008, Noy 2009). To take these factors into account in the model two additional variables were selected: adult literacy as a percentage of the population and exports as a percentage of GDP. The former came from the UN data (an internet-based data service which brings UN statistical databases within easy reach of users through a single entry point). This data is based on nationally collected data so with the same potential limitations due to national data system quality. The latter came from the World Bank Data Bank.
Whilst there are many variables that contribute to GDP what we are considering here are instead variables that affect the ability of an economy to recover from a shock, rather than variable to affect GDP. Hence variables like industry mix, education levels, urbanisation etc were not used as variables but do have an affect on GDP, hence it is quite possible that this will come up in the model analysis. Further, the relationship between these variables and GDP may not be linear – theory would suggest for example that literacy would increase as GDP increases up to a point and then beyond that it would not follow the same pattern.
GDP data for 2013 is the dependent variable; data for the period 2003-2013 was used to calculate average values for the independent variables. It could be argued that there is a lag time between a disaster event and long term impact on GDP – so a country that has recently experience a major event may not have felt the economic impacts. To try to take this into account an average of disaster events over the previous ten years was used.
Model design and basic model testing
Initial data analysis looked at the relationship between variables using transformed and untransformed variables. Theory suggests that as disasters increase (in this analysis damages, affected people as a proportion of population and events per square km increase) so does GPD; and that as literacy and openness go up so does GDP. However theory also suggests that these may not be linear relationships.
First a kernel density distribution of the dependent variable GDPmln was done.
|
kernel density estimate |
|
>
|
|
|
kernel = epanechnikov, bandwidth = 5 . 5e+04 |
||
|
Kernel density of GDPmln showed it was not a normal distribution, and using a logarithmic transformation did appear normal. This may suggest that log transformed variables maybe a better model – this is explored in further model testing.
Next the relationship between GDPmln and disaster exposure was explored. The average number of natural disasters per square kilometre, the average population affected as a proportion of the average population and average disaster damages in the preceding ten years were included in the analysis. Basic correlations were generated for level level and log log combinations of GDPmln and disaster exposure.
+
+
|
Graph : GDpmln 2 2 and avera e total damage |
|||
|
20000000 30000000 total dam (000 Us$) |
40000000 |
5000000 |
Level level combination of GDPmln and damages gave a strong correlation of 0.9120, although this may be an indication of exogeneity as suggested in the literature. To explore the relationship between disaster intensity and GDP further additional variables are considered: average people affected as a proportion of average total population and average events per square kilometre.
|
Correlation of 0.1760
|
Correlation of - 0.0519
These graphs indicate that there is heteroskedasticity in the data as it appears that the errors would not be randomly distributed. This is explored in further model testing in the next section.
The additional variables of openness (exports as a percentage of GDP) had a negative correlation of -0.0611 and literacy was 0.1475. Openness had been expected to be positive, again perhaps an indication of a non linear relationship.
Graphs 4, 5 and 6 show these relationships between the logarithmically transformed variables – both dependent and independent.
Graph 4 lngdpmln and lndamages
|
Correlation of 0.6303 and a visual indication of a relationship that may not be linear but this is not clear.
Graph 5: lngdpmln and lnaffected population
|
Correlation of -0.2834 and visually the relationship is not clear. Graph 6: lngdpmln and lnevents per square km
|
Correlation of -0.5361 and a visual indication of a linear relationship.
Correlation signs: Theory suggested that as disasters increase, so would GDP. However there was a contradiction in the results of the analysis – the theory held true for the correlations between GDP and disaster damages and people affected but not for events. In log transformed variables events and affected people had unexpected negative signs. These mixed results could reflect the potentially non linear relationship between each variable and GDP, the relationship between damages and GDP (as noted in the literature review – exogeneity), multicolinearity and / or a simply that the theory did not hold true for this data.
Of the three models using un transformed GDPmln keeping all variables level gave the higher adjusted Rsquared (85% of the variation inGDPmln could be explained by the variables in the model). Of the three models the models using a log transformed GDPmln, using all log transformed or all log transformed except for literacy and openness gave very similar adjusted Rsquared values (in the former 63.5% could be explained and in the latter 64.4% could be). The model with quadratic independent variables gave the highest Rsquared however as noted this model may be incorrect (80% of the variation could be explained).
Regression results:
The models generated a mix of results in the regression analysis as summarised in the table in Appendix 1
In the level level model for each increase in events per square km by 1 event, GDPmln would decrease by 3.80e+07 units (million USD); for each increase in affected people by 1 unit, GDPmln would decrease by 4843143 units (USD) and for each increase in damages (000 USD) GDP would increase by 0.3544345 million USD. The results for the impact of an additional 1 person affected as a proportion of average population is significant and may indicate a problem with this data.
We can reject the null hypothesis that the slope for affectedprop, damages and literacy is zero but not for eventspersq or open, indicating that the latter two may have a slope of zero, as reflected in the inconclusive graph of eventspersq and GDPmln above.
In the log log model for each increase in events per square km by 1 event, GDPmln would decrease by .4545612 % for each increase in affected people by 1 %, GDPmln would decrease by .2229837 % and for each percentage increase in damages (000 USD) GDP would increase by .3510787 %.
We cannot reject the null hypothesis that the slope coefficient is zero for all variables except openness.
In the mixed model with log dependent variable and log disaster variables but level variables of the social fabric for each increase in events per square km by 1 %, GDPmln would decrease by -.4475 %; for each increase in affected people by 1 %, GDPmln would decrease by .34956 %
We cannot reject the null hypothesis that the slope coefficient is zero for all variables except openness.
Further models and further model testing
Transformed and untransformed variables were used in a total of eight combinations of transformed logarithmic and untransformed variables were considered and full results are in the attached log file and in Appendix 1. One model was run with a quadratic transformed variable but this was inconclusive perhaps due to errors in the model setup.
Further testing of these models involved reviewing the results of tests for heteroskedasticity (Breusch-Pagan / Cook-Weisberg ) and the Ramsey Reset test for omitted variables / model design as well as analysis of residuals and analysis of results indicating multicolinearity. As noted above the initial analysis indicated a problem with heteroskedasticity and multicolinearity in the level level model. These results are explored further now through applying some model tests.
As a result of the testing it appears that the level level model has a problem with heteroskedasticity so this is not a good model fit further confirmed by the test for omitted variables that is also an indication of a poor model fit.
The model of all log transformed variables indicates there is a constant error variance, multicolinearity and potentially omitted variables. 63.4% of the variation in the logarithm of GDPmln is explained by the logarithm of the model variables.
After testing eight combinations of logarithmic transformed and untransformed variables and one with quadratic forms (see Appendix 1) it appears that a log log model is the best fit for the data of the models tested, however this was not conclusive because there was evidence of multicolinearity and potentially omitted variables. The testing involved reviewing the results of tests for heteroskedasticity (Breusch-Pagan / Cook-Weisberg) and the Ramsey Reset test for omitted variables / model design.
Conclusions / Gaps / Limitations
At the conclusion of model testing there are several persistent problems. Each will be briefly explored followed by some suggestions for future research
Signs: As noted, theory suggested that as disasters increase, so would GDP. However there was a contradiction in the results of the analysis and the theory did not hold true for each the variables chosen as a measure of disaster exposure. These mixed results could reflect the potentially non linear relationship between each variable and GDP, the relationship between damages and GDP (as noted in the literature review – exogeneity) and hence a poor model fit.
Model fit / missing variable testing: As noted in the discussion above on the selection of variables we chose two that were considered in both of those analyses: literacy and a measure of economic openness. However we did not seek to include all indicators of the economic resilience of a country, which means there was likely to be an indication of missing variables. This was true for each models although for the log log and the log log level model the result was very close to the critical value. Numerous additional factors affecting economic resilience should have been included to get a picture of all the variables relevant to this analysis. Furthermore there are other variables that were not included such as distance from the equator (this may affect exposure to climate hazards) as well as relative length of coastline.
Non linear relationships: as suggested in the literature there may not be a linear relationship between wealth and damages because wealth and exposure drive different levels of investment in safety measures. None of our models were able to demonstrate this.
Further research: Hallegatte and Dumas 2009 discuss the “poverty trap” that refers to the countries where they are not able to fully recover between disaster events and hence become trapped in poverty, unable to achieve GDP growth. To explore this futher data could be separately analysed for OECD and non OECD countries. Further, separating the data into climate and geohazard related events (as suggested by the results of Skidmore and Toya in 2002) would tease out the different impacts and preparation for each kind of event. This is particularly interesting given climate projections.
版权所有:编程辅导网 2021 All Rights Reserved 联系方式:QQ:99515681 微信:codinghelp 电子信箱:99515681@qq.com
免责声明:本站部分内容从网络整理而来,只供参考!如有版权问题可联系本站删除。