Vol. 161, No. 12
Printed in U.S.A.
DOI: 10.1093/aje/kwi149
American Journal of Epidemiology
Copyright ª 2005 by the Johns Hopkins Bloomberg School of Public Health
All rights reserved
Analysis of Worldwide Earthquake Mortality using Multivariate Demographic and
Seismic Data
E. Gutiérrez1, F. Taucer1, T. De Groeve2, D. H. A. Al-Khudhairy2, and J. M. Zaldivar3
1
European Laboratory for Structural Assessment Unit, Joint Research Centre–Institute for the Protection and Security of the
Citizen, Ispra, Italy.
2
External Security Unit, Joint Research Centre–Institute for the Protection and Security of the Citizen, Ispra, Italy.
3
Inland Marine Waters Unit, Joint Research Centre–Institute for Environment and Sustainability, Commission of the
European Communities, Ispra, Italy.
Received for publication July 8, 2004; accepted for publication February 16, 2005.
In this paper, mortality in the immediate aftermath of an earthquake is studied on a worldwide scale using
multivariate analysis. A statistical method is presented that analyzes reported earthquake fatalities as a function
of a heterogeneous set of parameters selected on the basis of their presumed influence on earthquake mortality.
The ensemble was compiled from demographic, seismic, and reported fatality data culled from available records of
past earthquakes organized in a geographic information system. The authors consider the statistical relation
between earthquake mortality and the available data ensemble, analyze the validity of the results in view of the
parametric uncertainties, and propose a multivariate mortality analysis prediction method. The analysis reveals
that, although the highest mortality rates are expected in poorly developed rural areas, high fatality counts can
result from a wide range of mortality ratios that depend on the effective population size.
demography; mortality; multivariate analysis; natural disasters
Abbreviations: GDP, gross domestic product; USGS, US Geological Survey.
The scientific issue we address in this paper is that of
analyzing which physical and demographic factors most
affect earthquake mortality. This analysis responds to the
need of organizations involved in humanitarian aid in developing countries to intervene promptly when confronted
with the problem of preparing emergency relief and disaster
mitigation programs for natural disasters. This need motivates development of models capable of estimating potential
human losses from natural disasters that may occur anywhere on the globe, which significantly constrains the quality of methods used to estimate mortality. Thus, in contrast
to the developed world, where the most pertinent demographic and physical parameters are readily accessible and
are of high quality, in developing countries, certain data are
scarce, unreliable, or nonexistent.
New data retrieval and satellite remote sensing technologies are opening new avenues that may overcome the deficiencies mentioned above. Recently, with the development
of geographic information systems, earthquake engineers,
seismologists, and emergency relief agencies have collaborated to develop information technology systems to help
decision makers rapidly assess potential damage and mortality in the immediate aftermath of an earthquake (1–7).
Such models have been calibrated by using detailed classification of buildings, soil types, and other key components
from past, well-documented earthquakes. Although these
models are continuously improving, the sophistication required, in terms of both the quality of input data and modeling resources, limits implementation of these techniques
to developing countries. This limitation is primarily the result of the incompleteness of pertinent parametric data in
some of the affected areas.
In this paper, we start by first identifying fatal earthquakes
(those that resulted in deaths) and their location (events); we
then report on other relevant physical and environmental
information concerning each event, such as the magnitude
Reprint requests to Eugenio G. Gutiérrez, Joint Research Centre–Institute for the Protection and Security of the Citizen, European Laboratory
for Structural Assessment, Via E. Fermi, 1, TP 480 Ispra 21020 (VA), Italy (e-mail: [email protected]).
1151
Am J Epidemiol 2005;161:1151–1158
1152 Gutiérrez et al.
of the earthquake according to the Richter scale and where
on the earth’s crust the earthquake took place. The second
and more challenging phase is to compute the affected population and how it is distributed with respect to where the
earthquake took place. Hence, from an epidemiologic point
of view, earthquakes can be seen as a disease whose mortality can be linked to a number of contributing factors.
MATERIALS AND METHODS
Data retrieval and compilation
We compiled a data ensemble so that each of the selected
parameters meets the following two requirements: 1) it can
be geographically referenced to a sufficiently precise position on the globe, and 2) it has been deemed by the engineering and scientific community to have some potential
bearing on earthquake mortality.
Each event is tagged to the relevant, quantifiable factors
that may have a bearing on mortality. The worldwide earthquake casualty set was compiled from fatality and injury
reports in the US Geological Survey (USGS) (8) database
of significant earthquakes (those that either exceeded a magnitude of 6.5 on the Richter scale or caused fatalities, injuries, or substantial damage) dating back to 1980. The
database also contains other attendant seismologic factors
pertaining to the location and depth of the earthquake, discussed in more detail below.
We divided the data set into two classes. The first includes
statistically derived demographic data. The second includes
parameters that are physical or pertain to physically related
environmental factors; here, we incorporate reported fatalities and injuries resulting from the earthquake.
Demarcation of the affected area
It is evident that the earthquake mortality analysis presented herein is inherently dependent on identification and
demarcation of the affected area, for it establishes not only
the boundaries of the affected population but also the values
of the other—physical and environmental—parameters to
which the population was exposed during the earthquake.
The toponomy types provided in the USGS data set are not
homogeneous. Sometimes, the name of an affected town or
city is given; in other instances, only the district or county is
provided. To homogenize the definition of the affected area
(and hence its population data sets), we demarcated the
affected areas with circles, as described in the following
three paragraphs.
Locate the affected area. We used a geographic information system to locate a town or administrative unit provided by USGS. If the toponomy given by the USGS
database could not be located, no circle was digitized and
the event was discarded.
Demarcate the affected area with circles. For a city, we
ensured that most of the high-density areas were contained
in the circle. We decided that suburban regions should be
included unless they had a radial orientation, which would
imply inclusion of a large portion of a rural area. The following information pertains to an administrative unit: 1) in
FIGURE 1. Reported fatalities and injuries worldwide, 1980–2001.
Diamonds, injuries; dashed curve, fatalities. The data are plotted on
a natural logarithmic scale and are ordered along the x-axis according
to a monotonically increasing fatality count. Unfiltered (top) and
filtered (bottom) data indicate that extreme events would appear to
cause more fatalities than injuries.
a sparsely populated unit, we drew the circle in such a way
that it covered most of the administrative unit and coincided
as much as possible with its borders; 2) in a unit with population centers (towns or cities with higher population densities), we drew the circle so that it contained the major
population centers.
Assign attributes. For each circle, we assigned an identification tag along with specific demographic, physical, and
environmental attributes. These attributes were compiled by
mapping the circle to the other information layers contained
in the geographic information system database (discussed in
the sections below).
We provide a representative example from the El Quindı́o
earthquake of January 25, 1999, in Colombia. For this earthquake, the USGS database reports 1,185 persons killed and
more than 700 missing and presumed killed. In the city of
Armenia, 907 fatalities are reported, with 60 percent of the
buildings destroyed. However, whereas similar percentages
of structural damage were reported in the neighboring towns
of Calarcá and Pereira, no fatality figures are given for either. It is highly unlikely that such structural damage did not
result in a high number of fatalities even though no fatalities
are explicitly ascribed. In the case of the El Quindı́o earthquake, the affected area is taken as the large circle encompassing the urban centers mentioned in the USGS report,
as shown in Web figure 1. (This figure is posted on the
Journal’s website (http://aje.oupjournals.org/).)
Reported casualties: fatalities and injuries
After examining the USGS database, we compiled a total
of 366 earthquake events (those in which at least one fatality
Am J Epidemiol 2005;161:1151–1158
Analysis of Worldwide Earthquake Mortality
TABLE 1. Demographic parameters
Parameter
1153
TABLE 2. Physical and environmental parameters
Description
Maximum density
The value of the maximum population
density in the demarcated area in
which fatalities have been reported,
that is, the highest population pixel
covering an area of 1 square
kilometer
Mean density
The value of the mean population
density in the demarcated area
Mode of density
Parameter
Description
Depth
The distance from the hypocenter or focus
(the point in the earth’s crust where a
seismic rupture begins) to the epicenter
(the point on the earth’s surface vertically
above the focus)
Magnitude
A measurement scale of the maximum
motion recorded by a seismograph, the
most common being the Richter scale
The value of the most common
population density value in the
demarcated area
Distance
The radial distance from the epicenter to
the center of the circle that defines the
area affected by the earthquake
Median of density
The halfway density value: in the
demarcated area, half of the pixels
have values below and half above
Activity
index
A measure devised in this paper to project
the local time that an earthquake occurs
to human diurnal or nocturnal activities
Mean (standard
deviation) density
The standard deviation of the
population density in the
demarcated area
Radius of
demarcated
area
Total population
The summation of the population
densities over the demarcated area
The radius of the circle that encompasses
the geographic, political, or urban area
in which the US Geological Survey
records reported fatalities
Xm
The ratio of the mean of the squares
of the population densities over the
mean population density of the
demarcated area
Maximum
hazard
The maximum hazard pixel value in the
demarcated area; earthquake hazard
includes indicators for surface faulting,
ground shaking, landslides, liquefaction,
tectonic deformation, and tsunamis
Minimum
hazard
The minimum hazard pixel value in the
demarcated area
Maximum
slope
The maximum topographic slope in the
demarcated area
Minimum
slope
The minimum topographic slope in the
demarcated area
Mean slope
The mean topographic slope in the
demarcated area
Maximum
soil index
The maximum soil-texture pixel value in
the demarcated area; soil index reflects
the percentages of sand, clay, and silt
and assigns the values of 1, 2, and 3 to
coarse, medium, and fine textures,
respectively
Minimum
soil index
The minimum soil-texture pixel value in
the demarcated area
Mean soil
index
The mean soil-texture pixel value in the
demarcated area
GDP* per
capita
The gross domestic product of an affected
country in US dollars (2004)
Fatalities/
injuries
The number of fatalities and/or injuries
ascribed to the demarcated area
is reported and the geographic region is well defined) for our
analysis. Although the sources from which the USGS draws
its casualty reports are not provided, such figures are usually
obtained from local governmental authorities as well as international nongovernmental aid agencies. Whereas casualty
reporting in the developed world is means-supported by governmental and financial institutions for legal and financial
claims reasons, it is expected that casualty reporting in the
developing world is inaccurate. In these countries, the lack of
a strong governmental social infrastructure where, quite often, population censuses are lacking and emergency relief
mechanisms cannot always cope with large-scale emergencies, casualty reporting can at best be described as estimated.
Demographic data
Having established the locations where fatalities are reported for each significant earthquake, we computed the
demographic data from a raster-based global population
set (9) assembled on a geographic information system.
The worldwide data set comprises geographically referenced digital elements, referred to as pixels; each spans an
area of 1 square kilometer of the earth’s surface. A synopsis
of the demographic parameters is given in table 1. The
population density (inhabitants per square kilometer) is automatically defined by the value of the population for each
pixel. The ensembles of the pixels in the demarcated area
constitute the affected population.
From these data, it is also possible to calculate the other
statistical population attributes. Statistical moments on demographic distribution can be used to differentiate urban
from rural areas that may have similar total populations
but widely differing statistical properties. Such differences
in demographic distribution can be used as surrogate paramAm J Epidemiol 2005;161:1151–1158
* GDP, gross domestic product.
eters for different construction types (e.g., high-rise buildings or other). For example, demographic statistics may
discern urban areas surrounded by topographic features associated with scarcely populated zones such as lakes or
mountains; for these cases, we devised a statistical moment
ratio, obtained from the mean of the squares of the individual population pixel values (x) divided by the mean popux:
lation density (
x) of the area, Xm ¼ x2 =
Physical and environmental data types
Physical parameters (table 2) such as the earthquake’s
magnitude and the depth of the hypocenter from the
1154 Gutiérrez et al.
epicenter (the point on the earth’s surface vertically above
the hypocenter) are intrinsic properties of the earthquake
event that can be obtained from the USGS, which also provides their location with geographic coordinates. The position of the earthquake itself is therefore included as an
information layer on the geographic information system.
The type of fault that caused the earthquake was not considered in the analysis, nor was the distance of the population to major tectonic fault lines.
Other data types in this group relate to environmental
factors; they are cataloged as subsidiary layers of the geographic information system and are matched to the demographic data layers. Thus, the distance from the center of the
demarcated area to the epicenter; the radius of the circumscribed area; the type of soil; the topographic slope, which
highlights the placement of the population in special topographic areas such as alluvial planes or hilly ground; and the
hazard—a composite parameter linked to factors such as
historical seismic activity, maximum ground acceleration,
landslides, and ground liquefaction (10)—can all be referenced geographically. Thus, to each population pixel in the
demarcated area there corresponds a quantified attribute of
each of the parameters included in our study. From this
ensemble, it is possible to derive statistical properties of
these attributes that uniquely characterize the population
under investigation for each specific earthquake event.
There has been much speculation as to the time that an
earthquake hits a population. There would seem to be compelling reasons to include this information in our analysis,
so the local time of day when the event occurs is taken into
account with a simple parameter, activity index, ranging
from 0 (sleep) to 2 (active or working).
In the field of earthquake engineering (a branch of engineering that studies the behavior of structures and buildings
subject to seismic action), certain parameters play a key role
in the capacity of an earthquake to cause destruction. Of
these parameters, the deep geology and the stratigraphic
profiles that determine the local ground conditions influence
the attenuation or amplification of seismic waves (11) and
are considered of fundamental importance in structural seismic design (12). We are not aware of a worldwide raster data
set for the aforesaid parameters; instead, the topsoil texture
(13) is used as a surrogate, which may be indicative of the
substrate parent rock that generated it.
Another key factor is the quality of construction, that is,
the architectural, structural, and material qualities that enable a structure to withstand or dissipate the input energy
from the earthquake without collapse. In developed countries, the quality of housing in earthquake-prone areas has
improved considerably, and the use of structural steel and
reinforced concrete has, in general, improved the capacity of
most structures to withstand earthquakes. In the developing
world, the quality of the materials and workmanship and the
more predominant use of brick masonry and adobe as the
main load-carrying elements make the housing stock more
vulnerable to earthquakes. We are not aware of a worldwide
database of housing stock that may be used to assess structural build quality; hence, we used the gross domestic product (GDP) per capita (14) of the affected country as an
indicator of the overall quality of construction and as a sur-
rogate for vulnerability (15) of the affected population to
natural disasters.
Mortality
Mortality is obtained by dividing the declared number
of fatalities by the total population in the demarcated area.
The error sources in estimating the total population inevitably result in over- and underestimation of mortality. Conversely, over- and underestimation of fatalities will have
a similar effect. For example, the May 1995 Sakhalin Island
7.1-magnitude earthquake in Russia resulted in 1,998 reported fatalities, with a circumscribed population (according to our population data set and the demarcated area
definition) of only 850. In spite of this incongruence, we
have no systematic reason to exclude such cases because it
is to be presumed that examples of overreporting of fatalities
may statistically balance other examples in which the population is overestimated. Other errors concern the impact of
the evolution of demographic trends over time that is not
taken into account by the global population data set. How
good can any mortality prediction be when confronted with
such calibration data? We can claim to study worldwide
earthquake mortality in statistical terms only.
Analysis methods
The analysis assumes that some of the parametric data
(particularly those corresponding to demographic and casualty reporting) are contaminated by errors. From the unfiltered data, it is not possible to discern whether the
correlations of the presumed casualty-inducing parameters
with mortality are low because they have no influence or
simply because they represent the complexity of some nonlinear dependence of mortality on the parameters selected.
Moreover, a low correlation value is not a sign of noncausality if the distribution is not Gaussian. We conjecture that
if a nonlinear correlation exists, filtering will reduce noise
and highlight the relation between parameters. To do so, we
screened the data with a moving window averaging filter.
We examined both the value and the rate of convergence of
the correlation factors as a function of the filter bandwidth.
Having first performed a null-hypothesis test on randomly
generated data, we consider that for the highest filter value
(32-point average), only those correlations higher than 0.75
can be deemed significant. The aim of the analysis, then, is
to monitor the parametric correlations of mortality as a function of the filter bandwidth. We claim that, by so doing, we
can construct a ranking of the sensitivity of mortality to the
parameters selected for this study.
RESULTS
First, we examined the relation between fatalities and
injuries. Figure 1 shows, on a natural logarithmic scale,
the evolution of reported casualties ordered in terms of increasing fatality count for 366 events. Plotting the data sequence in order of increasing fatality count (as opposed to
chronological order) allows us to observe that the ratio of
reported fatalities to reported injuries would seem to depend
Am J Epidemiol 2005;161:1151–1158
Analysis of Worldwide Earthquake Mortality
1155
TABLE 3. Correlation coefficients for mortality associated with the other parameters studied
Filter window size
0
2
4
8
16
32
Demographic parameters
Maximum density
0.582
0.696
0.812
0.876
0.934
0.971
Mean density
0.497
0.611
0.709
0.789
0.849
0.879
Mode of density
0.102
0.146
0.225
0.294
0.429
0.519
Median of density
0.317
0.426
0.534
0.621
0.718
0.767
Mean (standard
deviation) density
0.563
0.684
0.788
0.857
0.909
0.939
Total population
0.649
0.746
0.840
0.896
0.929
0.945
Xm
0.552
0.676
0.795
0.864
0.921
0.962
0.185
0.254
0.364
0.498
0.651
0.788
0.209
0.278
0.404
0.548
0.696
0.805
0.252
0.344
0.485
0.625
0.759
0.868
0.047
0.089
0.151
0.229
0.335
0.549
Radius of the target area
0.189
0.234
0.303
0.403
0.478
0.560
Maximum hazard
0.098
0.112
0.156
0.252
0.333
0.464
Physical and environmental
parameters
Depth
Magnitude
Distance
Activity index
Minimum hazard
0.009
0.013
0.031
0.099
0.149
0.236
Maximum slope
0.077
0.089
0.113
0.181
0.204
0.229
Minimum slope
0.282
0.365
0.460
0.541
0.641
0.731
Mean slope
0.156
0.215
0.285
0.348
0.439
0.516
0.134
0.175
0.236
0.379
0.507
0.600
Maximum soil index
Minimum soil index
0.111
0.143
0.181
0.222
0.273
0.351
Mean soil index
0.019
0.034
0.062
0.115
0.167
0.188
GDP* per capita
0.253
0.332
0.462
0.606
0.722
0.827
0.717
0.810
0.892
0.935
0.962
0.973
Fatalities
* GDP, gross domestic product.
on the severity of the event. Thus, for the most severe earthquakes (thousands of fatalities), the ratio of reported injuries
to reported fatalities is on the order of 1:1 or less; for moderate severity earthquakes (hundreds of fatalities), it is approximately 10:1; and for the most moderate earthquakes,
that is, those where tens of fatalities are reported, it is 20:1.
Overall, the number reported injured (~553,000) exceeds
that of fatalities (~190,000); however, in the database constructed herein, numerous examples may be found where
more fatalities than injuries are reported. We cannot claim
to determine whether the cause of this trend is due to the
manner of reporting or is a direct result of the severity of the
event; however, a parallel can be drawn with classes of
severe man-made disasters, extreme events, and diseases
that may result in near-complete mortality within the affected cohort.
We now discuss the correlations of mortality with the
remaining parametric data by monitoring the evolution of
the correlation factors as a function of the size of the filter
window applied. Given that strong, nonlinear dependencies
may be introduced by the wide range and standard deviations within a species (e.g., populations range from tens of
Am J Epidemiol 2005;161:1151–1158
millions to tens), we take the natural logarithm of the demographic and human loss data. The correlation coefficients
for mortality associated with the other parameters for both
filtered and unfiltered data are given in table 3.
For the unfiltered data (0-point window filter size), we
note that the only significant correlations of mortality are
with the demographic parameters and fatality. The maximum population density, the total population, the mean,
the mean (standard deviation), and Xm all approach or exceed a negative correlation factor of 0.5. The absolute values
of the correlation coefficients for depth, magnitude, and
distance from the epicenter are rather low. This finding reflects the fact that, on an individual event basis, earthquakes
of high magnitude may not necessarily induce high mortality if they do not combine other key factors, such as depth
and demographic criteria. This statement may seem to play
down the effect of magnitude, but, as we show later in this
paper, the average value for the highest-fatality-count earthquakes is in excess of magnitude 7.
The absolute values of the correlations of magnitude and
depth increase (respectively, positive and negative, i.e., high
magnitude, shallow earthquakes increase mortality) with
1156 Gutiérrez et al.
increasing filter size. Likewise, the trend in the correlation
coefficient of the distance from the demarcated population
to the epicenter is negative (short distances). Most of the
other physical parameters seem to play a negligible role,
the only exception being that flat terrain is moderately positively correlated with mortality. The correlations of soil
texture with mortality are not particularly strong, indicating
that such a parameter cannot be used as a surrogate of soil
type as intended in earthquake engineering. Likewise, it may
seem surprising that the correlation of seismic hazard with
mortality is rather low, but, because most of these events
take place in hazardous zones, the comparatively small variations in hazard may be hidden by the conditional dependence on other factors. The activity coefficient does not
seem to play a significant role either, indicating perhaps that
mortality can be equally high when people are congregated
in buildings either when asleep or during working hours.
The correlation trends with the demographic data become
more strongly negative (i.e., nonurban) with filter size, particularly regarding maximum population density; the mean,
mean (standard deviation) density; and Xm. The increased
negative coefficients with total population cannot be trivially explained by the fact that mortality is directly derived
from the ratio of fatalities and total population: from the
available demographic data at our disposal, there is a clear
trend that indicates higher mortality in less populated zones.
For example, the 2001 Bhuj, northwest India, earthquake
was a relatively low mortality event even though it resulted
in a high fatality count (above 20,000) where the circumscribed analysis area encompassed a high total population
(more than 40 million).
Conversely, the December 2003 earthquake in Bam in
southeast Iran (not included in this data set) is an example
of a high-mortality event, for it is believed to have caused
more than 30,000 fatalities in a population of less than
200,000 within a 50-km radius of the epicenter. The latter
is representative of an earthquake occurring in sparsely
populated semirural areas with no large, highly developed,
urban centers. These findings are corroborated by the fact
that Xm (which tends to highlight built-up urban areas
surrounded by sparsely populated zones) is also highly
negative.
Also of great relevance, from a humanitarian relief point
of view, is the increasing negative correlation of GDP per
capita with mortality. The initial correlation for the unfiltered set is 0.25, whereas, for the 32-point average, it is
0.83. However, even though GDP values vary considerably at both ends of the mortality spectrum for earthquakes
of similar mortality rank, on average it was observed that
the high end of the mortality range is populated with lower
GDP values. Thus, conditional parametric sensitivity implies that, for all other parameters being equal, we expect
a higher mortality if the GDP of the affected population is
low, which is as expected in poor rural communities where
the structural quality of buildings is lowest.
We now turn our attention to high-fatality events (i.e.,
those for which more than 500 fatalities are reported);
of these, 33 events (as shown in figure 2) were compiled
from the complete 366-event ensemble. It is important to
note that the average magnitude for the high-fatality earth-
FIGURE 2. Prediction of mortality and fatalities for severe earthquakes worldwide, 1980–2001. In both plots, the data are ordered
along the x-axis by monotonically increasing mortality rank. The top
plot shows the predicted and reported mortality values; the bottom plot
shows the predicted and reported fatality counts. For both plots,
predicted values are circles and reported values are diamonds.
quake set was just higher than 7. For ease of identification,
the sequence number now refers to mortality rank.
For this reduced subset, we tried to predict mortality and
fatality by using an automatic regression technique based on
multivariate principle component analysis (16–18). The results are shown in figure 2. The most severe earthquakes
(more than 10,000 fatalities) are identified by their corresponding mortality sequence rank number on the graphs.
Severe, high-fatality earthquakes were found at either end
of the mortality scale; that is, earthquakes of similar mortality rank may result in widely differing fatality counts.
The number of fatalities predicted for the 1988 Armenia
earthquake (figure 2, #31) was 1,500, whereas 25,000 fatalities were reported. Although the prediction for the 1988
west Iran earthquake (#28) was of the order of 10,000 and
the reported value was 45,000, we propose that, in terms of
early warning, this would still be a valid reason to set up the
appropriate emergency procedures. However, the 2001 Bhuj
(#7) and 1999 Turkey (#29) fatality predictions were more
accurate. The remaining fatality predictions were consistent
with the reported values. Finally, the cumulative predicted
fatality count for the 33-event ensemble was ~130,000,
whereas the reported one was ~182,000. Considering that
the cumulative fatality toll for the complete 366-event ensemble was ~190,000 implies that the 33-event high-fatality
ensemble accounted for nearly 95 percent of the reported
death toll from 1980 until 2001.
DISCUSSION
As stated, we have not been able to quantify the errors in
reported casualties (fatalities and injuries). Likewise, the
Am J Epidemiol 2005;161:1151–1158
Analysis of Worldwide Earthquake Mortality
demographic data are subject to a number of error sources.
In the first instance, even if we assumed an error-free population database, the fact that the area of the affected population is not well defined precludes precise quantification
of the affected population. Secondly, the population database is itself subject to errors. Finally, we presume that the
population has remained invariant over the period covered
by the earthquake data set (1980–2001). Thus, we have not
considered the effects of seasonal trends such as tourism or
enforced migration, nor have we attempted to account for
the long-term demographic trends.
From our analysis, a pattern emerges of highly vulnerable
rural and semirural areas with a poorly built environment. In
fact, of the 50 highest mortality ratios, two thirds are for
populations with fewer than 500,000 inhabitants. However,
in terms of total fatalities, the worst earthquakes vary over
a wide mortality range; the 2001 Bhuj and the 2003 Bam
earthquakes are contrasting examples. Even though the
worldwide postearthquake mortality prediction correlation
for a selected set of high-fatality earthquakes is about 0.93,
commission (false high fatality alerts) and omission (low
numbers of fatalities predicted for high-fatality events) errors occur. We claim that the accuracy of the method improves when the causal correlations between parametric
input data and mortality are sufficiently strong, and that,
when used with more accurate input parameter sets (available for the developed world), the accuracy will improve
further. However, as yet, this is not a truly blind prediction
method because the demarcated area considered in the prediction was known beforehand.
The damaging effects of earthquakes are not limited to the
immediate morbidity and mortality resulting from the structural collapse of buildings. The sequelae of earthquakes may
result in other pathologies linked to the breakdown of essential social services, such as sanitation and the health care
infrastructure, as well as concomitant factors from chronic
diseases such as cardiovascular disorders (19–21). This study
should complement other epidemiologic earthquake mortality studies by providing an approximate basis for assessing
worldwide earthquake mortality and how regional differences may attenuate or aggravate the attendant mortality.
ACKNOWLEDGMENTS
The authors thank Dr. P. Billing of the European
Commission Humanitarian Office for support and motivation regarding the research work presented herein. They
are also grateful to colleagues Dr. Y. Drossinos, Dr. N.
Stilianakis, and S. Primi for suggestions concerning
preparation of the manuscript.
REFERENCES
1. Peek-Asa C, Ramı́rez MR, Shoaf K, et al. GIS mapping of
earthquake-related deaths and hospital admissions from the
1994 Northridge, California, earthquake. Ann Epidemiol
2000;10:5–13.
Am J Epidemiol 2005;161:1151–1158
1157
2. Blais NC, Seligson HA, Petrow AJ. Use of rapid damage assessment and geographic information systems for emergency
response in the Northridge earthquake. Presented at the 11th
World Conference on Earthquake Engineering, Acapulco,
Mexico, June 23–28, 1996.
3. Davidson R. An urban earthquake disaster risk index. Stanford,
CA: The John A. Blume Earthquake Engineering Center,
1997. (Report no. 121).
4. Zolfaghari MR. Earthquake risk assessment model for loss
estimation in Italy. Presented at the 11th European Conference
on Earthquake Engineering, Paris, France, September 6–11,
1998.
5. HAZUS: loss estimation earthquake model. Washington, DC:
US Federal Emergency Management Agency, 2004. (http://
www.fema.gov/hazus/hz_eq.shtm).
6. Nichols JM, Beavers JE. Development and calibration of an
earthquake fatality function. Earthquake Spectra 2003;19:
605–33.
7. Shakharamanian MA, Larionov VI, Nigmetov GM, et al.
Assessment of the seismic risk and forecasting consequences
of earthquakes while solving problems on population rescue.
Moscow, Russia: Russian Civil Defense and Disaster Management Research Institute, 2000.
8. Significant earthquakes of the world. Reston, GA: US
Geological Survey, 2004. (http://neic.usgs.gov/neis/eqlists/
significant.html).
9. LandScan: Global Population Project. Oak Ridge, TN: Oak
Ridge National Laboratory, 2004. (http://www.ornl.gov/sci/
gist/landscan/index.html).
10. Giardini D, Grünthal G, Shedlock K, et al. The global seismic
hazard assessment program (GSHAP). Ann Geofis 1999;
42:1225–8.
11. Evernden JF, Thomson JM. Predictive model for important
ground motion parameters associated with large and great
earthquakes. Reston, VA: US Geological Survey, 1988.
(Bulletin 1838).
12. Eurocode 8 EN1998-1: design of structures for earthquake
resistance, part 1: general rules, seismic actions and rules
for buildings. Brussels, Belgium: European Committee for
Standardization, European Commission, 2004.
13. Digital soil map of the world and derived properties. Land
and water digital media series no. 1. Rome, Italy: Food and
Agriculture Organization of the United Nations/United
Nations Educational, Scientific and Cultural Organization
(FAO/UNESCO), 1998. (http://www.fao.org/catalog/
book_review/giii/w7189-e.htm).
14. The world factbook 2004. Washington, DC: Central Intelligence
Agency. (http://www.cia.gov/cia/publications/factbook/).
15. Peduzzi P, Dao H, Herold C. Global risk and vulnerability
index trends per year (GRAVITY), phase II: development,
analysis and results. Geneva, Switzerland: United Nations
Development Programme, 2002. (Scientific report UNDP/
BCPR).
16. Gutiérrez E, Taucer F, De Groeve A, et al. Prediction of mortality
after an earthquake based on a multivariate analysis of demographic and seismic data. Ispra, Italy: Directorate General, Joint
Research Centre, Institute for the Protection and Security of the
Citizen, Commission of the European Communities, 2004.
(European Commission report no. EUR 21059 EN).
17. Krzanowski W. Principles of multivariate analysis. Oxford,
United Kingdom: Oxford University Press, 2000.
18. Iyengar RN, Prodhan KC. Classification and rating of strongmotion earthquake records. Earthquake Eng Struct Dyn
1983;11:415–26.
1158 Gutiérrez et al.
19. Trevisan M, Jossa F, Farinaro E, et al. Earthquake and coronary heart disease risk factors: a longitudinal study. Am J
Epidemiol 1992;135:632–7.
20. Armenian HK, Melkonian AK, Hovanesian AP. Long term
mortality and morbidity related to degree of damage following
the 1988 earthquake in Armenia. Am J Epidemiol 1998;
148:1077–84.
21. Bland SH, Farinaro E, Krogh V, et al. Long term relations
between earthquake experiences and coronary heart disease
risk factors. Am J Epidemiol 2000;151:1086–90.
Am J Epidemiol 2005;161:1151–1158