EM-DAT emergency database
All the analyses in this article are based on the EM-DAT emergency database. This database is open source and maintained by the World Health Organization (WHO) and the Centre for Research on the Epidemiology of Disasters (CRED) at the University of Louvain, Belgium (Guha-Sapir et al. 2012). Comparable databases are NatCat (Munich Re) and Sigma (Swiss Re), which are maintained on a commercial basis. The EM-DAT database contains disaster events from 1900 onwards, presented on a country basis. Details and reliability aspects are given in Appendix A of ESM. CRED applies the following classification for weather-related disaster events (Below et al. 2009):
-
Meteorological events, including hurricanes (typhoons), extratropical storms and local storms
-
Hydrological events, including coastal and fluvial floods, flash floods and mass movements
-
Climatological events, including cold waves, heatwaves, other extreme temperature events, drought and wildfire
We have adopted this terminology in this article (although it is rather arbitrary why a storm is denoted as a meteorological event and a heatwave as a climatological event). It should be noted that the term ‘natural disasters’ is used in the literature if geophysical events are also included (earthquakes, tsunamis and volcano eruptions). These latter events are excluded here.
The EM-DAT database provides three disaster impact indicators for each disaster event: economic losses, the number of people affected and the number of people killed. Here, economic losses are direct damage costs and a direct consequence of weather or climate events. The number of people affected is the sum of people injured, people needing immediate assistance for shelter and people requiring immediate assistance during a period of emergency (this may include displaced or evacuated people; see ESM for more detailed definitions).
We aggregated country information on disasters to three economic regions: OECD countries, BRIICS countries (Brazil, Russia, India, Indonesia, China and South Africa) and the remaining countries, denoted hereafter as Rest of World (RoW) countries. OECD countries can be seen as the developed countries, BRIICS countries as upcoming economies and RoW as the developing countries. An example of these annually aggregated data is given in Fig. 1, in which extreme disasters are highlighted using catchwords. The panels show large differences across disaster indicators and regions: economic losses are largest in the OECD countries, the number of people affected is largest in the BRIICS countries and the number of people killed is largest in the RoW countries.
Normalization
As described in the introduction, disaster burden can be formulated as a function of exposure, climate drivers and vulnerabilityFootnote 1. Exposure can be described as a function of population size and wealth, or
$$ {D}_{\mathrm{t}} \approx {E}_{\mathrm{t}}*\mathrm{g}\left({C}_{\mathrm{t}},{V}_{\mathrm{t}}\right) = \mathrm{h}\left({W}_{\mathrm{t}},{P}_{\mathrm{t}}\right)*\mathrm{g}\left({C}_{\mathrm{t}},{V}_{\mathrm{t}}\right) $$
(2)
where W
t is an indicator for wealth; here GDP and expressed as PPP (purchasing power parity), and P
t is the population size.
For normalization purposes we formulated the following multiplicative models:
$$ \mathrm{economic}\kern0.5em \mathrm{Losses}:{L}_{\mathrm{j},\mathrm{t}}={W}_{\mathrm{j},\mathrm{t}}\ast {g}_{\mathrm{j},1}\left({C}_{\mathrm{j},\mathrm{t}},{V}_{\mathrm{j},\mathrm{t}}\right) $$
(3a)
$$ \mathrm{people}\ \mathrm{Affected}:{A}_{\mathrm{j},\mathrm{t}} = {P}_{\mathrm{j},\mathrm{t}}*{g}_{\mathrm{j},2}\left({C}_{\mathrm{j},\mathrm{t}},{V}_{\mathrm{j},\mathrm{t}}\right) $$
(3b)
$$ \mathrm{people}\ \mathrm{Killed}:{K}_{\mathrm{j},\mathrm{t}} = {P}_{\mathrm{j},\mathrm{t}}*{g}_{\mathrm{j},3}\left({C}_{\mathrm{j},\mathrm{t}},{V}_{\mathrm{j},\mathrm{t}}\right) $$
(3c)
with the subscript j denoting one of the regions ‘OECD’, ‘BRIICS’, ‘RoW’ or ‘Global’. For all three equations, it holds that the indicator sum of regions equals ‘Global’. The functions g
j,1, g
j,2 and g
j,3 are assumed to be monotonic with respect to their arguments; in other words, disaster burden will not decrease if weather events become more extreme in the region of interest and/or the region becomes more vulnerable. We note that Eq. (3a) equals the formulation of Neumayer and Barthel (2011; Eqs. (1) and (4)). Now, normalized indicators are simply defined as:
$$ {L}_{\mathrm{j},\mathrm{t}}\hbox{'} = {L}_{\mathrm{j},\mathrm{t}}/{W}_{\mathrm{j},\mathrm{t}},\ {A}_{\mathrm{j},\mathrm{t}}\hbox{'} = {A}_{\mathrm{j},\mathrm{t}}/{P}_{\mathrm{j},\mathrm{t}}\ \mathrm{and}\ {K}_{\mathrm{j},\mathrm{t}}\hbox{'} = {K}_{\mathrm{j},\mathrm{t}}/{P}_{\mathrm{j},\mathrm{t}} $$
(4)
In Section 3 we analyze the trends in L
j,t′, A
j,t′ and K
j,t′; the relative role of meteorological, hydrological and climatological events are given in Section 4. Normalization can be performed in various ways. Using Eqs. (3a)–(3c), we followed the method proposed by Neumayer and Barthel (2011), see their equation 4. Other functional methods are given by Pielke et al. (2008) and Neumayer and Barthel (2011, see their equation 1). Once a functional method has been chosen, the relations given in Eq. (4) can be calculated on much smaller spatial scales and summed afterwards to the region of interest. Neumayer and Barthel applied this approach for economic losses by defining gridded affected areas with corresponding GDP data. This approach is not followed in this study.
Data for W
j,t and P
j,t are taken from OECD (2012, see Chap. 2) and further described in Visser et al. (2012).
Trend estimation
There are many methods for estimating trends in historic data (Chandler and Scott 2011; Visser and Petersen 2012); however, generally speaking, there is no best method as this depends on the characteristics selected. We have chosen here to apply a model from the class of structural time series models, in combination with the Kalman filter: the integrated random walk (IRW). The advantage of this model is its ability to generate uncertainties for trend statistics. More specifically, this trend approach generates uncertainties for trend estimates μ
t and any trend difference (μ
t − μ
s). Furthermore, the OLS linear trend, often applied in disaster research, is a special case of the IRW trend approach. The IRW trend can therefore be seen as a natural extension of the straight line, in which the full uncertainty information is retained (Visser 2004; see Appendix B of ESM for details). The importance of choosing a particular trend method in disaster research has been illustrated by Visser and Petersen (2012, their figure 7a and b).
An example for the number of people affected in BRIICS countries is given in Fig. 2. Since the data are skewed to higher values, we log-transformed the original data y
t: z
t = log (y
t). After trend estimation, the trend estimates were back-transformed by taking exponentials (Appendix B of ESM). One consequence of the transformation is that any trend difference (μ
t – μ
s) estimated for z
t, transforms to a trend ratio (μ
t/μ
s) for y
t. The upper panel shows the normalized data along with the IRW trend estimate and 95 % confidence limits. The lower panel shows trend ratios (μ
2010/μ
t), along with 95 % confidence limits. This panel shows three important findings: (1) the trend ratio over the full sample period is high: (μ
2010/μ
1980) = 5.2, (2) the 95 % confidence limits are wide: (1.5–18.7), and (3) the rise in the trend occurs mainly over the first 5 years: the trend ratios (μ
2010/μ
t) are significant for the 1981–1985 period only (α = 0.05)Footnote 2.
IRW trend estimates do not give satisfactory results for all the disaster time series, due to outliers in the data. For these cases, we applied extreme value models (Coles 2001) and a non-statistical trend estimation technique known as Lowess estimators (Chandler and Scott 2011, see their section 4.3.1). We used the S-PLUS 8.1 software for these analyses.