VULN MOD Fragility Compendium 201501 v03 1

Existing Empirical
Fragility and Vulnerability

Vulnerability
Relationships: Compendium
and loss and Guide for Selection
modelling
T. Rossetto, I. Ioannou, D. N. Grant
GEM Technical Report

2015-01 v1.0.0
GEM
global earthquake model
Existing Empirical Vulnerability and
Fragility Functions: Compendium
and Guide for Selection

Vulnerability Global Component Project
Version: 1.0
Author(s): T. Rossetto, I. Ioannou, D.N. Grant
Date: July, 2015

Copyright © 2015 T. Rossetto, I. Ioannou, D.N.Grant Except where otherwise noted this work is made
available under the terms of Creative Commons Attribution-‐ShareAlike 4.0 (CC BY-‐SA 4.0).
The views and interpretations in this document are those of the individual author(s) and should not be
attributed to the GEM Foundation. With them also lies the responsibility for the scientific and technical data
presented. The authors do not guarantee that the information in this report is completely accurate.
Citation: Rossetto, T., Ioannou, I., Grant, D.N. (2015) Existing Empirical Fragility and Vulnerability Functions:
Compendium and Guide for Selection, GEM Technical Report 2015-‐1, 10.13117/GEM.VULN-‐MOD.TR2015.01.
www.globalquakemodel.org

ii

ABSTRACT
This document reviews existing empirical vulnerability and fragility functions worldwide collected until April
2014 in terms of their characteristics, data sources, and statistical modelling techniques. A qualitative rating
system is described and applied to all reviewed functions to aid users to choose between existing functions
for use in seismic risk assessments. The MS Access database developed by GEM VEM of all reviewed
empirical functions and associated ratings is also described in this document. The database may be freely
downloaded and includes all existing empirical vulnerability and fragility functions.

Keywords: empirical; fragility functions; vulnerability functions; rating system

iii
TABLE OF CONTENTS

Page
ABSTRACT ............................................................................................................................................................. ii
TABLE OF CONTENTS ........................................................................................................................................... iii
LIST OF FIGURES ................................................................................................................................................... v
LIST OF TABLES ..................................................................................................................................................... v
1 Introduction ..................................................................................................................................................... 1
2 Existing Empirical Vulnerability and Fragility Functions .................................................................................. 2
2.1 Form of Functions ................................................................................................................................... 4
2.2 Quality of Loss or Damage Database ....................................................................................................... 6
2.2.1 Damage ......................................................................................................................................... 7
2.2.2 Economic Loss from Direct Damage ........................................................................................... 11
2.2.3 Downtime ................................................................................................................................... 12
2.2.4 Casualties .................................................................................................................................... 12
2.2.5 Building Classification ................................................................................................................. 13
2.2.6 Non-‐Sampling Errors ................................................................................................................... 15
2.3 Quality of Ground Motion Intensity ...................................................................................................... 16
2.3.1 Isoseismic Unit ............................................................................................................................ 16
2.3.2 Measures of Ground Motion Intensity ....................................................................................... 17
2.3.3 Evaluation of Ground Motion Intensity ...................................................................................... 21
2.4 Vulnerability or Fragility Assessment Procedures ................................................................................. 22
2.4.1 Sample Size ................................................................................................................................. 22
2.4.2 Data Manipulation and Combination ......................................................................................... 22
2.4.3 Shape of Vulnerability or Fragility Functions .............................................................................. 24
2.4.4 Statistical Modelling Techniques ................................................................................................ 26
2.4.5 Method of Conversion of Damage to Loss in “Indirect” Vulnerability Curves ............................ 29
2.5 Validation Procedures ........................................................................................................................... 29
3 Rating System for Empirical Fragility and Vulnerability Functions ................................................................ 30
3.1 Application ............................................................................................................................................ 34
4 Compendium of Existing Vulnerability or Fragility Functions ........................................................................ 39
4.1 Basic Information .................................................................................................................................. 39
4.2 Global Coverage of Existing Empirical Vulnerability and Fragility Functions ........................................ 45
4.3 Harmonisation of Damage and Intensity Measures for Comparison of Existing Functions .................. 47

iv

4.4 Overview of Quality Ratings of Existing Functions ................................................................................ 47
5 Final Comments ............................................................................................................................................. 48
REFERENCES ........................................................................................................................................................ 49
APPENDIX A Tabulated Data for Compendium ................................................................................................. I
APPENDIX B Rating of Existing Empirical Vulnerability and Fragility Functions .............................................. XI

v
LIST OF FIGURES

Page
Figure 2.1. Illustration of a) a column of a DPM for intensity im, b) fragility curves corresponding to n=3
damage states for the same building class. ................................................................................................ 4
Figure 2.2. ATC-‐20 (1989) flowchart for normal building safety evaluation and posting. .................................. 9
Figure 3.1. Components, attributes and criteria of the proposed rating system. ............................................. 31
Figure 4.1. Fragility curves constructed by Karababa and Pomonis (2010). ..................................................... 44
Figure 4.2. Distribution of reviewed a) vulnerability and b) fragility functions according to the
country/countries of data origin. ............................................................................................................. 45
Figure 4.3. Frequency of construction material used in the reviewed a) vulnerability and b) fragility functions.
.................................................................................................................................................................. 46
Figure 4.4. Distribution of reviewed a) vulnerability b) fragility functions according to whether they use one
or more database. .................................................................................................................................... 46
Figure 4.5. Frequency of ground motion intensity measures used in the reviewed a) vulnerability and b)
fragility functions. ..................................................................................................................................... 47

LIST OF TABLES

Table 2.1. Important aspects of reviewed vulnerability and fragility assessment methodologies. .................... 3
Table 2.2. Database typologies and their main characteristics. .......................................................................... 6
Table 3.1. Rating the overall quality of existing empirical fragility functions. .................................................. 33
Table 3.2. Rating the relevance of existing empirical fragility functions to the needs of future application. ... 34
Table 3.3. Rating for the damage probability matrices proposed by Braga et al. (1982). ................................. 35
Table 3.4. Rating for the fragility curves proposed by Rota et al. (2008b). ....................................................... 36
Table 3.5. Rating for the indirect economic vulnerability relationship proposed by Yang et al. (1989). .......... 37
Table 3.6. Rating for the indirect economic vulnerability relationship proposed by Murakami et al. (1992). . 38
Table 4.1. Basic information provided in the compendium of existing vulnerability or fragility functions. ..... 40
Table 4.2. Parameters of the fragility curves (mean μ and standard deviation σ) constructed by Karababa and
Pomonis (2010). ........................................................................................................................................ 44

vi

Table A.1. Glossary ............................................................................................................................................... I

Table A.2. Notation for the vulnerability or fragility functions .......................................................................... III
Table A.3. Main characteristics of existing methodologies developed for the construction of DPMs. ............. IV
Table A.4. Main characteristics of existing indirect methodologies developed for the construction of
vulnerability functions. ............................................................................................................................... V
Table A.5. Main characteristics of existing methodologies developed for the construction of alternative
fragility functions (function of probability being in a damage state against intensity). ............................ VI
Table A.6. Main characteristics of existing methodologies developed for the construction of fragility curves.
.................................................................................................................................................................. VII
Table A.7. Main characteristics of existing direct methodologies developed for the construction of
vulnerability functions. .............................................................................................................................. IX
vulnerability functions (downtime). ........................................................................................................... X
vulnerability functions (fatalities). .............................................................................................................. X

Table B.1. Rating of existing direct and indirect vulnerability functions. ........................................................... XI
Table B.2. Rating of existing fragility functions. ................................................................................................ XII

1
1 Introduction
In seismic risk assessment, vulnerability functions express the likelihood that assets at risk (e.g. buildings,
people) will sustain varying degrees of loss over a range of earthquake ground motion intensities. In the GEM
Vulnerability Estimation Methods (GEM VEM) working group, the seismic loss is expressed in terms of the
cost of the direct damage, casualties, and downtime. In addition, ground shaking is considered the only
source of seismic damage to the building inventory. Vulnerability functions are based on the statistical
analysis of loss values recorded, simulated, or assumed over a range of ground motion severities. The loss
statistics can be obtained from past earthquake observations, analytical or numerical studies, expert
judgement, or a combination of these. This document focuses on the review and rating of existing empirical
vulnerability functions developed from post-‐earthquake observations. Other GEM VEM reports review and
rate existing vulnerability functions developed by analytical methods and engineering judgement (D’Ayala et
al, 2015; Jaiswal et al, 2012). A recent study instead (Rossetto et al, 2014) attempts to harmonise the ranking
system for the selection of the best fragility functions from available analytical and empirical functions.
Post-‐earthquake surveys of the performance of asset classes are commonly regarded to be the most realistic
source of loss statistics, since all aspects of an earthquake source as well as the wide variety of path, site,
foundation, exposure, structural and non-‐structural components are, at least in theory, represented in the
sample. These observations collected after one or more seismic events are effectively considered capable to
predict the vulnerability of specified assets for ground motion intensities occurring in future events.
However, empirical vulnerability curves may not yield reliable predictions given the typically poor quality and
(often) quantity of available observations and the often questionable procedures adopted to capture the
uncertainty in the observations. Furthermore, empirical vulnerability curves constructed from data obtained
from a single event may not appropriately account for the variability in structural response due to aleatory
uncertainty in the characteristics of the ground shaking (e.g. number of cycles, duration or frequency content
of the ground motion) at any given intensity. This perhaps makes them unable to reliably predict the
vulnerability of this building class affected by a future event. By contrast, databases corresponding to
multiple seismic events from diverse tectonic environments might be associated with observations from
buildings having considerably diverse structural characteristics, resulting in functions with very large
uncertainties and questionable applicability (Rossetto and Elnashai, 2003).
Empirical vulnerability functions can be constructed directly from past earthquake observations of losses
collected over sites affected by different intensities of strong ground motion. If an intensity measure level
(IML) has not been recorded at each site, one can be assigned using an appropriate ground-‐motion
prediction equation. A model is typically fit to the data so as to minimise hindcasting errors. Commonly this is
done through regression analysis, which aims to estimate the parameters of a chosen functional form. Such
“direct” vulnerability curves are often developed for the estimation of casualties or economic losses at a
national or regional level and may or may not distinguish between casualties associated with different
building classes (Jaiswal and Wald 2010; 2013).
An alternative “indirect” procedure involves constructing empirical vulnerability functions through the
coupling of the empirical fragility of given building class located in the affected area, with appropriate
damage-‐to-‐loss functions, which convert the damage estimates to loss estimates. In this case, the empirical

2

fragility functions express the likelihood of differing levels of damage sustained by a structure class estimated
from post-‐earthquake building damage survey data. Similar to the direct functions, an IML is typically
assigned in each site using an appropriate ground-‐motion prediction equation due to the scarcity of ground
motion recordings. The damage-‐to-‐loss functions may be empirical in nature if based on field observations
and specific surveys, or they can be based on expert judgement. A detailed review of these functions can be
found in the GEM report by Rossetto et al (2015).
Both direct and indirect empirical vulnerability assessment procedures require the quantification and
modelling of substantial uncertainty from a number of sources. Depending on the model, aleatory (inherent
in the model) and epistemic (due to lack of knowledge or modelling capabilities) components of this
uncertainty can be identified. This classification of uncertainty may assist the user to make informed
decisions on how to improve the model. In empirical vulnerability functions, aleatory uncertainty is
introduced by the natural variation in earthquakes and their resulting ground shaking, or the variation of the
seismic response of the buildings of a given class. Epistemic uncertainty is introduced mainly by the low
quantity and/or quality of the damage/loss databases and the inability to account for the complete
characteristics of the ground shaking in the selection of measures of the ground motion intensity. Not every
source of uncertainty is modelled in existing studies, which typically do not appropriately communicate the
overall uncertainty in the vulnerability or fragility functions and are often incapable to distinguish the effect
of the two components, i.e. aleatory and epistemic.
This document presents a review of existing empirical vulnerability functions for buildings constructed using
“direct” and “indirect” approaches and published until April 2014. As part of the latter, it also reviews
existing empirical fragility functions which have (or not) been used for vulnerability assessment. This is
because more research has been done to date on fragility than vulnerability, and lessons learned from the
former can be applied to the latter. The review highlights issues related to: the functional form of the
relationships, quality issues regarding the loss/damage databases and the ground motion intensity as well as
empirical vulnerability or fragility assessment and validation procedures. A qualitative rating system is also
proposed and applied to the reviewed functions to aid users to assess the quality of existing empirical
functions for use in future seismic risk studies. Finally, the content of the MS Access 2010 database (here
called “Compendium”) of empirical fragility and vulnerability functions is described. This compendium has
been compiled specifically for GEM and is freely available to download from the GEM Nexus site and the
EPICentre website (http://www.epicentreonline.com/).
2 Existing Empirical Vulnerability and Fragility Functions

A number of empirical vulnerability and fragility functions have been developed from post-‐earthquake data
mostly by individual researchers or small research groups rather than a united research community.
Consequently, there is a large variation in the empirical vulnerability or fragility assessment procedures
presented in the literature. These result from differences in the selected functional form of the relationship,
in the quality of the adopted loss or damage observation databases and the selected ground motion
intensity, as well as the statistical modelling techniques and validation methods.

3
This chapter presents a review of existing methodologies for constructing empirical vulnerability and fragility
functions. The review is organised in terms of the aspects of the fragility or vulnerability functions deemed
important, which are listed in the first column of Table 2.1. As many of these aspects relate both to
vulnerability and fragility functions these are treated together in the following sections. The main
characteristics of the individual functions mentioned can be found in Table A.3 to Table A.6 in Appendix A.

Table 2.1. Important aspects of reviewed vulnerability and fragility assessment methodologies.
Form of Relationship: Discrete.
Continuous function.
Damage or Loss Damage Damage scale.
Database Quality: Characterisation: Consideration of non-‐structural damage.
Loss Measures of economic loss, casualty and downtime.
Characterisation:
Building Single or multiple building classes.
Classification:
Data Quality: Post-‐earthquake survey method.
Coverage error in surveys.
Response error in the surveys.
Measurement error in surveys.
Degree of refinement in building class definition.
Number of damage states (DS).
Data Quantity: Country/Countries of data origin.
Number of seismic events.
Structural unit.
Quantity of data in each isoseismic unit (e.g. number of buildings
or loss observations).
Range of IM and DS covered by data.
Total number of data.
Quality of Ground Intensity Measure (IM).
Motion Intensity: Isoseismic unit (e.g. zip-‐code area, town etc.).
IM estimation method (e.g. GMPE or recorded).
Statistical Modelling Data manipulation or combination.
Techniques: Relationship model.
Optimisation method.
Method of conversion of damage to loss in “indirect” vulnerability
curves.
Treatment and communication of uncertainty in the vulnerability
or fragility curves.
Validation With independent data.
Procedures: With other existing functions/methods.

4

2.1 Form of Functions
Vulnerability and fragility functions correlate loss and damage, respectively, to ground motion intensity and
their form is found to be either discrete or continuous. This section discusses the general forms of
vulnerability and fragility functions found in the literature, with a detailed discussion of the functional shapes
and parameters of the functions provided in Section 2.4.3.
In the case of empirical fragility assessment, the vast majority of existing functions are expressed either in
terms of damage probability matrices (e.g. Whitman et al, 1973; Yang et al, 1989; Gulkan et al, 1992;
Decanini et al, 2004; Eleftheriadou and Karampinis, 2008) or fragility curves (Rossetto and Elnashai, 2003;
Colombi et al, 2008; Rota et al, 2008; Liel and Lynch, 2009). Damage probability matrices (DPM) are
composed of sets of values defining the probability of a level of damage being reached in a given building
class at specified intensity measure levels. By contrast, fragility curves express the probability of a level of
damage being reached or exceeded given a range of intensity measure levels. The two outcomes are linked
as follows:

⎧ 1 − P (DS ≥ dsi IM ) i =0
0 < i ≤ n − 1
⎪
P (DS = dsi IM ) = ⎨P (DS ≥ dsi IM )− P (DS ≥ dsi +1 IM ) (2.1)
⎪
⎩ P (DS ≥ dsi IM ) i =n

where P (DS ≥ dsi IM ) is the probability of a level of damage, dsi , being reached or exceeded given seismic
intensity measure IM (fragility curves); P (DS = dsi IM ) is the probability of the buildings being within dsi for
IM (DPM); and dsn is the highest state of damage. The relationship between these two expressions of fragility
is also illustrated in Figure 2.1b, where the fragility curves corresponding to three damage states (ds1-‐3) are
depicted. For a given intensity measure level im, the probability of being in one damage state is represented
by the distance between one fragility curve and the one below it.

0.4 1.0
P(DS≥ds1 |IM)
0.3
P(DS ≥ dsi | IM)
P(DS = dsi | IM)
0.2 0.5 P(DS=ds2 |IM)

P(DS≥ds2 |IM)
0.1
P(DS≥ds3 |IM)
0.0 0.0
ds0 ds1 ds2 ds3 DS| im im IM
a) b)
(a) (b)

Figure 2.1. Illustration of a) a column of a DPM for intensity im, b) fragility curves corresponding to n=3 damage states
for the same building class.

A variant of the traditional definition of fragility relationship above can also be found in the literature (e.g
Scawthorn et al, 1981; Petrovski and Milutovic, 1990) where instead of the exceedence probability, the
probability of a building class sustaining a given level of damage as a continuous function of ground motion

5
intensity is used. In the case of empirical fragility functions, this variation has a subtle disadvantage over the
traditional definition, i.e. as each damage state curve is regressed for separately, the sum of the probabilities
obtained from each damage state curve for a given level of ground intensity does not always equate to one.
DPMs and fragility curves have been used in some studies for the construction of vulnerability functions
through an indirect approach. In its generic form, the indirect approach constructs the intensity-‐to-‐loss
relationship by coupling damage probabilities for building classes at specified intensities to damage-‐to-‐loss
functions (for more details on the latter functions see Rossetto et al, 2014) using the total probability
theorem:

n
P ( L > l IM ) = ∑ P ( L > l dsi ) P (DS = dsi IM ) (2.2)
i =0

where P(L>l|IM) is the complementary cumulative distribution of the loss given intensity IM; P(L>l|dsi) is the
complementary cumulative distribution of loss given a damage state dsi; P(DS=dsi|IM) is the damage
probability matrix. In practice, most indirect empirical vulnerability assessment methods (Benedetti et al,
1988; Yang et al, 1989; Spence et al, 2003; Eleftheriadou and Karabinis, 2011) ignore the uncertainty in the
damage-‐to-‐loss functions and focus on the estimation of the average loss at discrete IMLs, i.e.:

n
E ( L IM ) = ∑ E ( L dsi ) P (DS = dsi IM ) (2.3)
i =0
( ) ( )
where E L dsi is the mean loss L suffered by a class of structures for a given damage state; E L IM is the
mean loss for intensity IM.
To date, few direct vulnerability functions have been constructed due to the scarce availability of good
quality empirical loss data. The majority of these existing functions correlate a measure of the cost of the
direct damage to an intensity measure type. A small minority correlate a measure of casualty with intensity
and only Comerio (2006) are found to correlate downtime with ground motion intensity. With regard to the
form of these functions, discrete functions of loss are proposed by Scholl (1974) and Cochrane and Schaad
(1992) in terms of average economic loss and by Comerio (2006) in terms of downtime. These are
represented in similar matrix formats to DPMs with mean losses presented for each intensity level.
Continuous functions have also been proposed to correlate the economic loss of the direct damage
(Scawthorn et al. 1981; Petrovski et al. 1984) or fatality ratio (i.e. the deaths divided by the total exposed
population of an area) (Jaiswal and Wald 2010) with a range of IM parameters. These take a range of forms,
further discussed in Section 2.4.3.
An alternative representation of vulnerability or fragility is applied by Yang et al (1989), Sabetta et al (1998)
and Wesson et al (2004). These studies fit probability distributions to the loss or damage data for each
isoseismic unit, i.e. an area associated with a specific level of ground motion intensity whose spatial
distribution across the municipality is considered negligible. Then, they use a continuous relationship to
correlate the parameters of the latter distribution to a range of intensities. This approach, however, may lead
to wider confidence intervals around the parameters since it does not accounts for the overall number of
data, but rather concentrates on the number of isoseismic units available.

6

2.2 Quality of Loss or Damage Database
The level of quality of the loss or damage databases strongly affects the accuracy of the vulnerability or
fragility functions. Ideally, empirical functions should be based on a high quality database, i.e. one that
includes detailed damage and loss data from all the structural units (e.g. buildings), or at least a statistically
significant representative sample, located in the area affected by a strong event. However, this is rarely the
case. Instead, epistemic uncertainty is introduced in the vulnerability or fragility assessment by the lower
quality of databases, typically adopted in the reviewed studies, which are associated with errors or low
degree of refinement in the definitions of damage scales, loss measures, building classes or the location of
the structural units as outlined in Table 2.2.
The errors affecting existing databases used for the construction of vulnerability or fragility curves can be
classified as sampling or non-‐sampling errors (UN, 1964):
• Sampling errors occur because a subset of the population of buildings, located in the affected area,
has been selected (surveyed) instead of the total affected population. The subset is selected to be
representative of the target population of buildings, with the required number of buildings for a
given level of error typically being calculated from standardised procedures based on the adopted
sampling technique (see Levy and Lemeshow, 2008).
• Non-‐sampling errors represent the remaining sources of error occurring in the survey design, as well
as errors in the collection and processing of the data. Three main sources of non-‐sampling error have
been identified in the empirical vulnerability or fragility literature (United Nations, 1982):
o Coverage errors are unobservable errors that occur in cases where the database of damage
or loss observations does not accurately reflect the target population of the considered
asset, e.g. total number of structural units in the affected area.
o Response errors occur when incomplete or no data for some structural units are collected
during the survey due to rapid assessment, an inexperienced team, poor supervision or
problems with questionnaire (e.g. lengthy and/or unclear).
o Measurement errors describe the deviance of the observed value of a variable from its
‘true’, but unobserved value.
In what follows, a critical review of the definitions of damage scales, loss measures and building classes is
presented, together with a discussion of the common sources of error found in seismic loss/damage
databases due to the survey design and data collection (non-‐sampling errors). Section 2.2.6 specifically
reviews how common non-‐sampling errors are dealt with in the literature. As sampling errors are typically
reduced by aggregating data from different classes or combining loss or damage databases, they are instead
further discussed in Section 2.4.2 where data combination is reviewed.

Table 2.2. Database typologies and their main characteristics.
Type Survey Method Typical Typical Building Typical No. Reliability Typical issues

7
Sample Classes of of
Sizes Damage observations
States
Rapid Surveys Large All buildings 2-‐3 Low Safety not damage
evaluations.
Detailed Large to Detailed Classes 5-‐6 High Possibility of
“Engineering” Small unrepresentative
Surveys samples.
Surveys by Very Detailed classes 5-‐6 High Possibility of
Damage
Reconnaissance Small unrepresentative

Teams samples.
Remotely sensed Very All buildings 3-‐5 Low Only collapse or very
Large heavy damage states
may be reliable.
Misclassification
errors.
Tax assessor data Very large All buildings/ -‐ High Often focus on
Detailed classes damaged buildings
Economic
only.
Loss
Claims data Very large All buildings -‐ High Often focus on
damaged and/or
insured buildings only.
Government Surveys Very large All buildings 1-‐2 Low Unlikely association
with building damage
Surveys by NGOs/ Varies All buildings 1-‐5 Low and causes of injuries.
Casualties
hospitals
Possibility of
Detailed Casualty Very Detailed classes 3-‐5 High unrepresentative
Surveys Small samples.

2.2.1 Damage
The quality of a damage database compiled from post-‐earthquake observation depends on the adopted
measure of damage and the errors in the design and execution of the survey. In existing fragility functions,
damage is characterised via descriptive damage states associated with a discrete damage scale. The choice of
a damage scale for the assessment of buildings is therefore fundamental to the definition of fragility
functions and resulting indirect vulnerability functions. As discussed in Hill and Rossetto (2008), for post-‐
earthquake damage observations to be useful in the development of empirical fragility functions they must
clearly and unambiguously define the damage states in terms of visually observable structural and non-‐
structural damage characteristics for different structural classes, ideally using both text and figures. The
inclusion of non-‐structural damage in the damage scale is particularly important for loss evaluation. The
survey should also be carried out by engineers trained in the identification of building damage using the
damage scale. All these measures reduce potential misclassification errors.

8

In reviewing the empirical fragility and vulnerability literature, a wide variety of damage scales is seen to
have been used. These scales vary in the number of damage states and the level of detail in the description
of damage for each state according to the purpose of the survey. With very few exceptions (e.g. Liel and
Lynch, 2009) building damage statistics are not collected from the field for the purpose of constructing
fragility or vulnerability functions. Rather, they have become available from surveys carried out to assess
structural safety or evaluate cost of repair for insurance, tax reductions, or government aid distribution
purposes or for reconnaissance purposes. Hence, care must be taken when adopting these for fragility
analysis.
Post-‐earthquake surveys commissioned by state authorities as part of their disaster response are the most
commonly used source of data for vulnerability or fragility relationship construction, e.g. Greek authorities
(Eleftheriadou and Karabinis, 2008; Karababa and Pomonis, 2010), Italian authorities (e.g. Braga et al, 1982;
Benedetti et al, 1988; Yang et al, 1989; Goretti and Di Pasquale, 2004; Decanini et al, 2004), Japanese
authorities (e.g. Yamazaki and Murao, 2000). These databases tend to be the most extensive in terms of the
size of the affected area surveyed and number of buildings. The survey area is seen in the literature to vary in
geographical scale according to the earthquake size and attenuation/amplification characteristics of the
affected area. According to the availability of the survey data, the reviewed studies select a target area for
the empirical vulnerability or fragility curve generation, which can range in geographical extent from the
entire earthquake affected area (e.g. Braga et al, 1982; Petrovski et al, 1984; Goretti and Di Pasquale, 2004;
Karababa and Pomonis, 2010), a city within the affected area (e.g. Scholl, 1974; Benedetti, 1988; Sengezer
and Ansal, 2007; Eleftheriadou and Karabinis, 2008), or smaller political units such as a ward (e.g. Yamazaki
and Muraio, 2000). The smaller the chosen target area the smaller the sample sizes (number of buildings) for
the building classes that can be used for vulnerability or fragility assessment. This target area is typically
subdivided into smaller, units, which are considered isoseismic. As stated in Section 2.3.1, the chosen
isoseismic units are seen to vary substantially in geographical size and population of buildings, even within
the same study. In some surveys, all units within the target area have been surveyed (e.g. Benedetti et al,
1988; Goretti and Di Pasquale, 2004; Eleftheriadou and Karabinis, 2008; Karababa and Pomonis, 2010). In
others, only a fraction of these units has been surveyed (Braga et al, 1982; Petrovski et al, 1984; Sengezer
and Ansal, 2007). Nonetheless, with the exception of Scholl (1974), who selected two sites with similar
number of buildings but different soil conditions, the reviewed studies did not provide the unit selection
criteria. This raises questions on whether the effects of the examined earthquake are appropriately
represented in the survey. For example, what is the proportion of less affected versus severely affected
units? Are all the sites close to the epicentre and few located far away? Are buildings randomly selected?
This level of detail regarding the survey design is missing from most empirical studies.
It is not uncommon that successive surveys with increasingly detailed structural evaluations, and potentially
different damage scales, are carried out in earthquake-‐affected areas, as shown in Figure 2.2. Initially, a rapid
survey is conducted in order to assess the safety of the building inventory affected by the earthquake. These
surveys commonly adopt damage scales consisting of three limit states (such as in the Greek level 1
approach, OASP 1997 (Dandoulaki et al, 1998) or ATC-‐20 1989 in the USA) but some may only consist of two
broad damage states (i.e. damaged and undamaged, or safe and unsafe), grouping together a wide range of
seismic damage levels, e.g. buildings that suffered collapse or severe damage are grouped together. For
those buildings identified as being damaged, this first stage of evaluation is commonly followed a second
more detailed assessment: the “detailed damage evaluation” or “engineering evaluation” (see Figure 2.2).
The damage scales used for the latter evaluations commonly comprise between three and eight discrete and
mutually exclusive states. Damage evaluation forms are used to record damage to the structural and non-‐

9
structural components (see for example the Italian AeDES forms, Baggio et al, 2000). The procedures and
forms adopted for damage data collection vary for different countries and have changed over time, (see
Goretti and Di Pasquale, 2002 for a summary of the evolution of damage data collection procedures and
tools in Italy).
The aforementioned surveys commonly suffer from a number of significant non-‐sampling errors. For
example, rapid damage evaluations commonly adopt damage scales with two or three damage states. No
distinction is often made between different structural classes. In addition, these surveys are carried out
through an external inspection of the structure only, potentially introducing a significant misclassification
error in the assignment of the damage. For all these reasons, the resulting vulnerability or fragility functions
are of limited usefulness and reliability. Whilst rapid surveys can cover the entire affected region, detailed
damage evaluations are typically carried out over a subset of the affected regions (usually the worst affected
regions) and cover a sample of total number of buildings (e.g. buildings that are heavily damaged, or
buildings for which a specific request for detailed damage evaluation has been made by the owner). For
example in the 2002 Molise earthquake in Italy, Goretti and Di Pasquale (2004) report that for areas away
from the epicentre, where damage was less, engineering surveys were only carried out on the building
owner’s request. This may result in significant coverage error in the database due to the practical exclusion
of the majority of the undamaged buildings from the sample. This error, if untreated, could lead to a bias in
the damage statistics (see Section 2.2.6).

Figure 2.2. ATC-‐20 (1989) flowchart for normal building safety evaluation and posting.

10

Partial non-‐response errors can also be introduced when information regarding the building typology, level
of damage, building’s location or repair cost is omitted from the survey forms. This is seen to be a significant
problem in Eleftheriadou and Karabinis (2008), who, due to incomplete reporting, only used data from
73,468, out of 180,945 surveyed reinforced concrete buildings in the 117 municipalities of Athens affected by
the 1999 earthquake to build their DPMs. It is unclear what influence this has on the final result bias,
especially as it is unclear whether the spatial distribution of these buildings is taken into account. The
proportion of unreliable data due to incomplete survey forms or claims were also found to be substantial
(ranging from 27% to 66%) in other damage databases (i.e. Colombi et al, 2008) but insignificant in others
(i.e. Rota et al, 2008 and Wesson et al, 2004).
Measurement errors are also commonly found in survey forms/claims. These errors occur when
inexperienced engineers carry out the surveys, when the survey form is not sufficiently detailed or
appropriate training has not been provided to reduce the surveyors’ bias. Measurement errors in a database
also occur from inclusion of the combined losses or damage caused by successive large aftershocks or other
main events (e.g. Goretti and Di Pasquale, 2004; Karababa and Pomonis, 2010) and/or secondary hazards (e.g
liquefaction in Scawthorn et al, 1981) with those caused by ground shaking in the main event. Such errors
limit the application of functions based on these databases to other locations. The importance of these
errors is not clear given that existing studies did not (and in some cases could not) independently check their
damage data.
An alternative source of building damage statistics are databases compiled by institutions or specialist
reconnaissance teams, such as those of the UK Earthquake Engineering Field Investigation Team (EEFIT) and
the US Earthquake Engineering Research Institute (EERI). Such databases are smaller in geographical scale
than those carried out by government authorities. They mostly consist of damage surveys carried out
through external building inspection and over small areas (at road or block level). Such surveys commonly
use damage scales linked to macroseismic intensity scales (such as those in EMS-‐98, (Grunthal,1998)), as part
of the remit of the teams may be to evaluate the macroseismic intensity distribution caused by the event.
Most such databases are associated with low measurement errors as they are carried out by experts but high
sampling errors due to the small sample sizes (of the order of 10-‐50) of surveyed buildings. Data collected by
reconnaissance teams have been used in combination with other larger databases for the construction of
fragility functions (e.g. Coburn and Spence, 2002; Rossetto and Elnashai, 2003; Liel and Lynch, 2009). For
example, Liel and Lynch (2009) examined the vulnerability of Italian RC buildings by inspecting a sample of
483 out of the total population of 14,088 RC buildings located in the municipality of L’Aquila in the aftermath
of the 2009 event. Despite the high reliability of the observations (i.e. low damage misclassification error),
the samples can often biased with under-‐coverage error due to the traditional focus of the reconnaissance
teams on the severely damaged buildings/areas. However, recently reconnaissance teams have been
targeting locations close to ground motion recording stations (e.g. Sarabandi et al, 2004 and Rossetto et al,
2010). These data could, in theory at least, improve the reliability of the fragility functions by minimising the
uncertainty in the value of ground motion intensity. Such a sampling technique was used by King et al (2005)
who adopted a database of 500 inspected buildings affected by the 1994 Northridge earthquake, which were
located within 1000 feet from 30 ground motion stations.
Following recent events, large databases of building damage data have been obtained using remote sensing.
These damage assessments have been made by visual interpretation or change analysis of satellite images
taken after, or before and after, an earthquake event, respectively. Such sources are becoming more viable
as the number of satellites and availability and resolution of satellite images increases, and the cost of these
images decreases. The satellite images can be used alone or adopted together with aerial reconnaissance

11
images. Of note is the recent establishment of the Virtual Disaster Viewer (Bevington et al. 2010), supported
by several international institutions which provides images and a forum for experts to interpret damage from
satellite images. The subsequent GEO-‐CAN initiative (Ghosh et al. 2006) also brought together experts from
all over the world to interpret damage in the 2010 Haiti and 2011 Christchurch, New Zealand earthquakes.
Three major issues have been identified with the use of remotely sensed damage data (Booth et al. 2011,
Saito and Spence 2004 and Clasen et al (2010) :
• Difficulty in recognising soft-‐storey type collapses from vertical satellite images of rooftops,
• Inability to reliably identify damage states lighter than collapse,
• Difficulties with automating the change analysis procedure for damage detection.
Some of these difficulties may be overcome with time as more “angled” satellite or aerial images become
available and as technologies develop. To date, remotely sensed damage data has only recently been used
for the construction of empirical fragility functions for earthquakes (Hancilar et al. 2011) and tsunami (e.g.
Koshimura et al 2009).
In order to reduce epistemic uncertainty, empirical fragility assessment studies should adopt the same
damage scale as used by the damage surveys. However, in the literature it is seen that damage statistics are
sometimes interpreted in terms of a different damage scale (e.g. the 8 damage states of the Irpinia 1980
damage statistics converted to the 6 damage states of the MSK scale in Braga et al, 1982). This may be in
order to use a damage scale more closely related to the IM used, if the latter is expressed in terms of a
macroseismic intensity. The mapping of damage states between different damage scales is instead necessary
when damage data from surveys for the same earthquake (e.g. Karababa and Pomonis, 2010, Yamaguchi and
Yamazaki, 2001), or different earthquake events from the same country (e.g. Colombi et al, 2008; Rota et al,
2008, Sabetta et al, 1998) or different countries (e.g. Spence et al 1992, Rossetto and Elnashai, 2003) are to
be combined. This is discussed further in Section 2.4.2.
2.2.2 Economic Loss from Direct Damage

Vulnerability functions that correlate the economic loss with levels of ground motion intensity have been
obtained with direct as well as indirect vulnerability assessment procedures. Existing studies have been
limited to the estimation of economic loss from direct damage to structural and non-‐structural elements of
buildings, and have ignored the indirect loss of revenue or the rent paid by the owners during the repairs of
their dwelling or business.
Jara et al (1992) measured vulnerability in terms of monetary cost of repairing the damaged RC buildings in
dollars. However, the adoption of such a measure of economic loss is strongly location-‐ and time-‐ specific. A
more useful measure of economic loss is the damage factor (DF) estimated in terms of repair to replacement
cost ratio for the examined building class. Damage factors are assumed to vary much less over time for any
location and structure type, and not to vary greatly for similar structures in different locations, thus allowing
comparisons between existing functions to be made. However, the validity of these assumptions is
questionable and is not explored in the literature. Moreover, there is not a wide consensus on the definition
of ‘repair’ or ‘replacement’ cost. For example, repair cost is usually assumed to be the cost of repairing the
structural elements of the building to their original state without any strengthening, and may or may not
include the cost of replacing non-‐structural elements. Similarly, replacement cost is typically estimated as the
average cost of constructing a building of a given class without specifying whether the cost of site clearance
or the provision of a new structure with design compliant with current seismic code is taken into account. By

12

contrast, Dowrick and Rhoades (1990) use the property value instead of the replacement cost, and Wesson
et al (2004) consider the replacement cost to be equal to the fire-‐insured value of the examined buildings.
Detailed post-‐earthquake economic loss databases have been compiled from claims submitted by building
owners to the authorities (Dowrick and Rhoades, 1990) and/or the private insurance industry (Dowrick and
Rhoades, 1990; Cochrane and Shah, 1992; Wesson et al, 2004). An acknowledged bias in the latter databases
is the under-‐coverage error occurring due to the lack of data regarding losses below the deductible, i.e. a
threshold of economic loss below which the policyholder, rather than the insurance company, is expected to
pay for the repair cost. Another source of under-‐coverage error can be introduced by the potentially large
proportion of uninsured buildings present in the building inventory, which may lead to the sample (insured
buildings) misrepresenting the distribution of damage in the exposed inventory. Some studies have ignored
this issue (Cochrane and Shah; 1992; Wesson et al, 2004). Others (Dowrick, 1991; Dowrick and Rhoades,
1990) have instead shown that for their data, the number of uninsured data although notable (15%) did not
lead to substantial errors. Questionnaires have also been used in the past to collect loss data after
earthquake events. Whitman et al (1973) identified an area affected by the 1973 San Fernando earthquake
where approximately 1,600 mid-‐ and high-‐rise buildings were located. They collected loss data from a sample
of 370 buildings by sending questionnaires to the building owners. In general, the latter databases can be
considered unbiased and of high quality, with their level of epistemic uncertainty depending on the sample
size.
Post-‐earthquake cost data, (like building damage state data), can also be obtained from detailed or
engineering surveys (see Section 2.2.1), where engineers are asked to provide estimates of repair or
replacement costs from the observed degree of building damage. However, such data is rare, may be
inconsistent, and is dependent on the experience and training given to the engineers carrying out the
assessment. Such cost data can be used directly for generating vulnerability functions but, in recognition of
its high variability, should only really be used to define average Damage Factors (or DF distributions) for
specified building classes, which can then be adopted in indirect vulnerability curve generation. More
information on DFs can be found in Rossetto et al (2015).
2.2.3 Downtime
Downtime is defined as the time (i.e. in years or months) necessary to plan, finance and complete the repairs
of damaged buildings (Comerio, 2006). Interestingly, the time for the replacement of severely damaged or
collapsed buildings is not included in this definition. Although downtime is a measure of loss necessary for
the estimation of business disruption and the overall impact of an earthquake on society, it is largely ignored
in the empirical vulnerability literature. This may be due to the lack of systematically collected observations
on downtime caused by the significant span of time required to complete repairs (which can extend to
several years). It may also be due to the extremely large expected variation in downtime estimates for
buildings with different usage in different locations. Such variations are often influenced by socio-‐economic
and governance factors (e.g. availability of funds for repair/rebuilding) as well as the size of the earthquake
affected area. Furthermore, building-‐owners and/or state-‐authorities often play an important role in
prioritising and determining works, which influence the time of repair.
2.2.4 Casualties
In direct vulnerability functions, casualties are usually characterised by fatality counts (the number of people
who have died in an event) or by fatality rates (the number of deaths as a proportion of the exposed
population considered). Fatality rate is a more useful parameter than fatality counts for use in comparing

13
vulnerabilities in different locations. Indirect methods for casualty estimation usually involve combining
building damage or collapse estimates from fragility analyses with lethality rates defined for the considered
building class. Lethality ratio is defined by Coburn and Spence (2002) as the ratio of number of people killed
to the number of occupants present in the collapsed buildings of a particular class. The lethality ratio can be
affected by a number of factors including: building type and function, failure mechanism, occupant
behaviour, time of the event, ground motion characteristics, and search and rescue effectiveness (Coburn
and Spence, 2002).
The empirical evaluation of fatality counts, fatality and lethality rates depend on the availability of reliable
casualty data from past earthquake events. However, earthquake casualty data are reported to be scarce and
associated with significant quality issues (Lagorio, 1990; Petal, 2011). A number of sources of casualty data
have been used for the development of vulnerability functions. These include official statistics published by
government agencies, hospital admissions data, coroner and medical records, and specific surveys carried
out by casualty researchers (e.g. De Bruycker et al 1985; Petal 2009). Methods used for the data collection
vary and are commonly only specified in the case of researcher-‐led surveys. Hence, fatality counts, fatality
ratios, and lethality are plagued by inconsistencies in definitions and data collection methods.
Also, Li et al (2011) show that the reported fatality count changes as a function of time following the
earthquake event, however, there is no apparent guidance or consensus in the literature as to how long after
the event deaths should stop being attributed to the earthquake. Furthermore, in highly fatal earthquakes,
the number of deaths is seen to be highly unreliable and to vary widely, with official figures sometimes being
incorrect, e.g. aggregating different causes of death, (e.g. cardiac arrest, deaths from secondary hazards or
from epidemics following earthquakes), to deaths directly from structural and non-‐structural earthquake
damage. Petal (2011) also states that in some cases official death counts can be deliberately exaggerated or
understated for political reasons. In some casualty studies, official statistics are replaced or supplemented
with data from hospital admissions (e.g. Pomonis et al, 1992) or specific field studies conducted after
earthquakes in order to determine the number of deaths and their causes, (associating them, for example,
demographic characteristics or with mechanisms of structural damage). Such studies provide data of higher
quality, but the geographical scale of areas covered is a small subset of the entire affected area, resulting in
an insufficient range of ground motion intensities for the determination of direct vulnerability functions.
Hence, the casualty data are used mainly to calibrate damage-‐to-‐casualty functions used in indirect
vulnerability estimation. Different approaches are followed to collect casualty data in field surveys. For
example, De Bruyker et al. (1985), collect casualty data from a random sample of one-‐third of villagers in a
selected area affected by the 1980 Southern Italy earthquake, and Petal (2009) carry out geo-‐stratified
random sampling of families affected by the 1999 Kocaeli (Turkey) earthquake. Although small in sample size,
such data can be considered unbiased and complete for the target area if significant migration of people out
of the affected area has not taken place. However, in many other instances, (e.g. So, 2011), convenience
sampling is carried out, whereby only easily accessible groups or locations are sampled. This may result in
biases if the sample is not representative of the demographics of the affected population, or if the full range
building types and failure mechanisms within a selected building class are not represented. Such biases result
in estimates of fatality rates and lethality rates, respectively, that cannot be projected to the affected
population.
2.2.5 Building Classification

Buildings have unique structural and non-‐structural characteristics that are (partially) responsible to the wide
differences in their performance under similar seismic excitation. Nonetheless, in empirical vulnerability

14

studies building classes with similar earthquake response characteristics need to be defined in order to
obtain the basic damage statistics required to construct the vulnerability functions. In theory, the more
detailed the building class the more homogenous the group of buildings and the smaller the variation in
seismic response. However, in practice a narrowly defined building class often results in small sample sizes. A
careful balance is therefore struck between level of detail in the building class definition and sample size.
The structural (and non-‐structural) characteristics mostly affecting seismic response depend on which loss
parameter is being assessed. For example, the structure’s deformation affects the state of damage and
consequently cost, downtime, and human casualties. The structural characteristics influencing this
deformation are well documented (e.g. FEMA 547, 2007 and include (but are not limited to):

• Lateral load resisting system (e.g. MRF, Shear wall etc)
• Construction materials
• Layout in plan and elevation
• Presence of irregularities
• Seismic design code used (i.e. capacity design, level of structural detailing)
• Height of structure
• Weight of structure
• Redundancy of structure
• Type and amount of strengthening intervention

A building classification system that accounts for all these influencing factors allows a high-‐degree of
differentiation in fragility studies. However, in the case of cost, downtime and casualty evaluation, the
presence of non-‐structural elements and occupancy/use of the structure also have a strong influence on the
accuracy of the vulnerability evaluation, and hence should also be considered as part of the building
classification system. Furthermore, in the case of human casualty the lethality of the building materials, floor
and roof system should be considered in addition to their contribution to the structure dynamic response.
However, the latter is not commonly done, a primary relationship often being assumed between material or
roof type and the vertical elements of the structure (Shiono et al, 1991).
In the case of empirical vulnerability functions, the building class definition used can be dictated by the level
of detail present in the loss or damage data. In particular, rapid post-‐earthquake survey data often do not
distinguish between damage/losses observed in structures of different types, and vulnerability or fragility
functions developed from these data are highly specific to a particular built environment (e.g. Tavakoli and
Tavakoli, 1993). Different degrees of refinement are seen in the building classification systems adopted by
existing studies according to their data. For example, Gulkan et al (1992) classify buildings by construction
material only, Colombi et al (2008) by construction material and height, Karababa and Pomonis (2010) by
construction material, lateral load resisting system and age (representative of seismic code level), and
Spence et al (2003) classify reinforced concrete buildings in terms of their design code, height and quality of
construction. Dolce et al (2006) is the only study seen to account explicitly for the presence of seismic
strengthening measures, and none of the reviewed empirical functions takes explicit account of structures
with irregularities that may precipitate their damage/failure under earthquake loading.
Despite the existence of various building classification systems (e.g. adopted by HAZUS99 (FEMA, 1999) or
PAGER (Porter et al, 2008), these are not commonly adhered to by existing studies; rather, bespoke building
classes are defined. For example, there is no agreement as to how many storeys are included in different
height categories. Low-‐rise structures are fairly uniformly defined as being between 1-‐3 storeys, but

15
definitions of mid-‐rise and high-‐rise categories differ significantly. For example, Scawthorn et al (1981) define
mid-‐rise as containing buildings with between 3 and 12 storeys, whereas Rossetto and Elnashai (2003) adopt
5-‐8 storeys. Seismic code classes depend on the location being assessed, and on the year of the study. Care
must be taken when adopting studies carried out long ago, as the rapid evolution of seismic codes in recent
years may mean that the classes of structure they regard as “modern” may be considered “low-‐code” by
current standards.
Instead of using strictly typological building classes, some existing studies have adopted the approach of
assigning structures to vulnerability classes. These vulnerability classes are either described through bespoke
vulnerability indices that are evaluated from the characteristics of the structure (e.g. in Benedetti et al. 1988)
or are taken from earthquake macroseismic intensity scales. An example of the latter is Goretti and Di
Pasquale (2004), who assign the MSK vulnerability classes to the masonry buildings in their datasets through
consideration of their layout, the flexibility of their floors as well as the presence of tie beams.
2.2.6 Non-‐Sampling Errors

Several sources of non-‐sampling errors have been mentioned in Sections 2.2.1-‐2.2.5. Here, specific examples
of non-‐sampling errors found in the literature are discussed together with the procedures that have been
followed to reduce or deal with them.
Large under-‐coverage errors in existing damage databases have been acknowledged in the literature as
major sources of bias. In the case of government led surveys, this bias commonly occurs from procedures
that require the building owner to request their building to be inspected (Goretti and Di Pasquale, 2004;
Eleftheriadou and Karabinis, 2008; Karababa and Pomonis, 2010) resulting in a database mainly comprising
damaged buildings. Such a bias is likely to be larger for areas least affected by the ground shaking (typically
more distant from the earthquake fault). This bias, which occurs due to the incomplete surveying of
buildings, is addressed in two different ways by existing studies. In the first approach, a minimum proportion
of surveyed buildings within a geographical unit (of assumed uniform ground motion intensity) is defined as
the “threshold of completeness”. If a smaller proportion of the building stock in the geographical unit is
surveyed, then the unit is discarded. The value of the ”threshold of completeness” is, however, seen to vary
arbitrarily. For example, Gorreti and Di Pasquale (2004) selected a level of completeness ≥ 80%. This
threshold was reduced to 75% for Sabetta et al (1998) and 60% for Rota et al (2008) who used larger
proportion of their database but with lower reliability. Whatever the threshold, this approach results in a
(potentially significant) reduction in the number of suitable data, especially for less populated building
typologies. By contrast, in the second approach non-‐surveyed buildings are assumed undamaged and the
surveys are “completed” using census data (e.g. Karababa and Pomonis, 2010; Colombi et al, 2008). In some
studies (Karababa and Pomonis, 2010; Eleftheriadou and Karabinis, 2011), the census dates to a time close to
that of the earthquake event. However, other studies (Yang et al, 1989; Colombi et al, 2008) are seen to use
census data compiled up to 11 years before or after the examined event. This introduces a bias due to the
likely substantial differences in the built environment, e.g. old buildings could be demolished, and new
buildings, especially the ones built after the census year, are misrepresented. The lack of a contemporary
census led Tavakoli and Tavakoli (1993) to estimate the number of buildings located in the affected area by
projecting the population recorded in the 1976 census to 1986. Undoubtedly, this is associated with large
uncertainties and raises questions about the validity of their curves. The completion process can also be
complicated if the census is reported in structural units that are not consistent with that of the survey, with
further uncertainties introduced by the harmonization of the structural units in the two databases (e.g. the
census used by Colombi et al. (2008) uses dwellings rather than buildings).

16

Response errors may also occur in a single earthquake damage/loss database due to a large number of
incomplete survey forms. To date, existing empirical vulnerability and fragility studies consider that this error
is random, and as such can be addressed by simply ignoring the incomplete forms and constructing
vulnerability or fragility functions using only the complete forms. This assumption, however, is not
appropriately validated, and considering the survey process, it is likely not to be random. It is possible that
the same set of surveyors consistently does not complete forms and that there may be a geographical bias, if
these surveyors are assigned to particular sites. If the error is not random, ignoring incomplete forms would
introduce bias in the results especially in cases where the proportion of the eliminated data is significant; e.g.
Colombi et al (2008) eliminated 50% of the available data.
Measurement errors are also found to contaminate existing databases mainly due to the inclusion of loss or
damage data associated with secondary hazards. In existing studies, such errors are seen to be removed from
the databases where the survey data are available in sufficient detail (e.g. Yamazaki and Murao, 2000,
removed buildings damaged due to liquefaction from their database), or where the areas affected by the
secondary hazard were geographically constrained (with damage data from these areas excluded). However,
this is not always possible (e.g. Scawthorn et al, 1981). By contrast, measurement errors occurring due to the
inclusion of loss or damage data from successive strong aftershocks or other large earthquakes (e.g. Goretti
and Di Pasquale, 2004; Karababa and Pominis, 2011) cannot be removed from the database. In the latter
cases, this bias is acknowledged by the studies, and should be considered in interpreting the results from
these vulnerability or fragility studies, which may over-‐predict the damage/loss, (especially if applied to
estimate risk from small magnitude events).

2.3 Quality of Ground Motion Intensity
The construction of empirical vulnerability or fragility functions requires reliable loss and/or damage statistics
from the typically large area/s affected by strong ground shaking. In practice, the affected area is subdivided
into smaller units over which the seismic demand is considered uniform (isoseismic units). Ideally, these
isoseismic units should be areas of uniform soil conditions, in close proximity to ground motion recording
stations. These are necessary conditions for an accurate and uniform seismic demand to be assumed across
the building population, which minimises the uncertainty in the ground motion and results in damage/loss
statistics that reflect only the uncertainty in buildings resistance (Rossetto and Elnashai 2005).
In practice, the size of the isoseismic unit can be dictated by the minimum building sample size, survey
method used or way in which the damage data is reported. Scarcity of ground motion recording instruments
means that ground-‐motion prediction equations are used to assign IMLs to the isoseismic units. Moreover, a
range of different intensity measures has been used in the derivation of fragility and vulnerability functions.
In this section, existing fragility and vulnerability studies are critically reviewed with respect to these three
aspects of ground motion intensity estimation and implications on epistemic uncertainty are discussed.
2.3.1 Isoseismic Unit

In the literature, a multitude of sizes of isoseismic units is assumed (see Appendix A). Most existing post-‐
earthquake surveys report observed building damage over very large areas (e.g. a town in the case of
Murakami, 1992), which are unlikely to have uniform soil types or sustain a uniform level of ground shaking.
This is especially true when government led survey data is used as the empirical source, or when a measure
of macroseismic intensity is used as the measure of the seismic hazard. However, no existing empirical study

17
is seen to evaluate the uncertainty associated with the representation of the ground motion intensity across
an area by a single (average) parameter value.
2.3.2 Measures of Ground Motion Intensity

The epistemic uncertainty in the vulnerability or fragility assessment due to ground motion intensity partly
depends on the ability of the parameter chosen for its representation, IM, to describe the damage potential
(and consequent loss potential) of strong ground shaking, This reasoning has led to the widespread use of
macroseismic intensity as the measure of ground motion intensity in existing fragility and vulnerability
functions, (see Appendix A). Many different macroseismic intensity scales have been used for this in the
empirical literature, and the majority of existing DPMs use measures of macroseismic intensity as their IM.
Those most frequently adopted are the Modified Mercalli Intensity (MMI), Japan Meteorological Agency
(JMA), Mercalli-‐Cancani-‐Sieberg (MCS), Medvedev-‐Sponheuer-‐Karnik (MSK) and its successor the European
Magnitude Scale (EMS98). Different criteria are adopted in the definition of the macroseismic intensity levels
within each scale. Consequently, vulnerability or fragility functions obtained for different macroseismic
intensity scales are not directly comparable.
Despite the good correlation reportedly observed between macroseismic intensity and observed damage and
loss, their use in vulnerability or fragility assessment is not without disadvantages. Macroseismic intensities
are subjective measures, and large discrepancies have been noted in their evaluation at the same sites by
different reconnaissance teams (e.g. Rossetto et al, 2007). Macroseismic intensity scales are discrete
measures associated with intervals between unit values which are not necessarily equal, and intermediate
decimal or fractional values have no meaning. Furthermore, few ground motion prediction functions for
macroseismic intensity exist, and conversion equations between macroseismic intensity and other ground
motion intensity measures introduce additional uncertainty. This is a severe limitation in terms of the
application of macroseismic intensity-‐based fragility and vulnerability functions in risk assessments. Despite
these drawbacks macroseismic intensities have been extensively used in existing empirical vulnerability
studies (e.g. Samaradjieva and Badal (2002), Thráinsson (1992)). One appeal of macroseismic intensity is that
empirical loss and damage data pre-‐dating the widespread use of earthquake measurement devices can be
used. A further appeal is that macroseismic intensity values form natural “bins” of IM within which a large
number of observations of loss and damage can be obtained, due to the large spatial coverage of any given
macroseismic intensity value (e.g. Eleftheriadou and Karambinis, 2011). This helps provide statistically
significant samples of buildings in each IM value. However, in the case of data from more than one event, a
very large scatter in the vulnerability data associated with each macroseismic intensity value will be
observed, which results in considerably larger error distributions than present in the original single event
data (Spence et al, 1991).
In recognition of the disadvantages posed by the discrete nature of macroseismic intensity Spence et al
(1992), devised a continuous parameterless intensity measure, termed PSI or Ψ. PSI can be evaluated for a
site based on the distribution of damage observed in unreinforced masonry constructions. This parameter
has been adopted by Orsini (1999) and Karababa and Pomonis (2010) for their fragility curves. Functions
between PSI and peak ground acceleration (PGA) are proposed by Pomonis et al (2002) for use together with
existing ground motion prediction equations, which allows the PSI vulnerability curves to be used in
predictive earthquake risk assessments. However, these PSI-‐PGA functions are based on a limited number of
field observations and corresponding ground motion recordings. Therefore, they are associated with large
scatter and require additional verification before they can be deemed reliable.

18

Regardless of the scale continuity, the use of macroseismic intensity, including PSI, to characterise ground
motion in vulnerability or fragility functions can introduce inter-‐dependence between the predicted
vulnerability or fragility and the IM. This is because macroseismic intensity is evaluated directly from
observed earthquake effects and, hence, its value is partly dictated by the building stock fragility. In the case
of single earthquake events use of macroseismic intensity as the IM results in a reduced data scatter, as
damageability of an earthquake is better correlated with its observed damage than with direct measures of
ground shaking. Intuitively, one could assume that this would be offset by higher uncertainty on the
prediction of the IM, since it shifts the main predictive responsibility to GMPEs that estimate macroseismic
intensity (i.e. Intensity Prediction Equations, which include both damage observations and felt effects) from
seismological features such as magnitude, hypocentral location, and style of faulting. On the other hand,
limited evidence on this topic (e.g. Allen et al, 2011) suggests that there is actually a lower uncertainty on
predictions of macroseismic intensity than peak motions. This results from the fact that macroseismic
intensity estimates involve a spatial average over a study area, while peak motions are point measurements.
The authors are not aware of any studies that have compared uncertainty on loss estimates, propagated
through from the estimation of the IM and vulnerability, but such a study would be valuable for the
assessment of the relative merits of each approach.
A few studies propose continuous functions correlating the economic loss (Thráinsson and Sigbjörnsson
1994) or death rate (Samardjieva and Badal, 2002; Nichols and Beavers, 2003) to earthquake magnitude and
epicentral distance, due to a limited amount of available data. However, earthquake magnitude is seen to be
a very poor predictor of fatalities by Ferreira et al (2011). This is because earthquake magnitude is not
representative of the strong ground shaking caused by the earthquake, and hence the applicability of
functions using this parameter as their IM is severely limited.
A different approach for the selection of ground motion intensity measures has been promoted by the
current seismic risk assessment framework, where it is important to decouple the uncertainties associated
with the seismic demand from their counterparts introduced by the structural fragility. This can be achieved
by selecting a measure of seismic demand capable of representing the influence of source, path, and site on
the strong ground motion. This measure should characterize the ground shaking at each isoseismic unit and
ideally its level should be determined from earthquake ground motion records. However, due to the scarcity
of strong ground motion recording stations, in practice ground motion parameter values are obtained from
ground motion prediction equations, whose availability determines the applicability of the measure of
intensity. For the latter reason, and due to its traditional use in the definition of design loads for structures
and seismic hazard maps, peak ground acceleration (PGA) is the main parameter used to represent ground
motion intensity in empirical fragility studies (e.g. Sabetta et al, 1998; Rota et al, 2008). However, it is well
known (e.g. Rossetto and Elnashai, 2003) that PGA does not correlate well with observed damage (especially
for ductile structures or large damage states). In engineering terms, PGA is related to the dynamic loads
imposed on very stiff structures and, therefore, may not be adequate for the characterisation of seismic
demand in medium or high rise building populations. Peak ground velocity (PGV) is more representative of
the seismic demand on structures of intermediate natural period of vibration than PGA. PGV is generally
calculated through direct integration of accelerograms and may be sensitive to both the record noise content
and filtering process. Fewer ground motion prediction equations for PGV exist than for PGA, but PGV seems
to be the preferred IM in empirical fragility studies from Japan (e.g. Yamazaki and Murao 2000). Few
empirical vulnerability or fragility studies (e.g. King et al, 2005) currently exist that adopt peak ground
displacement (PGD) as their IM. Few attenuation functions exist for the prediction of PGD, due to its high
sensitivity to noise in accelerograms and to filtering techniques adopted in the elimination of unwanted

19
frequencies from raw records. The PGD attenuation functions that do exist, (e.g. Sadigh and Egan 1998), are
based on few digital records and cover limited areas and fault mechanisms. Rossetto (2004) observes a
better correlation of PGD than PGA with her database of reinforced concrete building damage statistics,
however notes that in both cases the overall correlation is poor. Sarabandi et al (2004) also notes poor
general correlation of all the ground motion indices they studied with a limited set of damage data for
reinforced concrete structures (60 buildings maximum sized database), but also notes slightly better
correlation of PGD than PGV and PGA with their data.
Parameters such as PGA, PGV, and PGD are unable to capture features of the earthquake record, such as
frequency content, duration, or number of cycles, which affect the response of a structure and its
consequent damage and loss. Hence, in an attempt to better capture the influence of accelerogram
frequency content, more recent empirical fragility and vulnerability studies have favoured the use of
response spectrum based parameters as measures of ground motion intensity.
Acceleration spectra are representative of the imposed seismic forces on the structure over a wide range of
frequencies, and are used by most modern design codes in the determination of structure loads. In view of
this, ground motion prediction equations for elastic spectral acceleration at 5% damping (Sa5%(T)) have been
proposed by various authors (e.g. Ambraseys et al 1996) and Sa5%(T) is commonly used in seismic hazard
maps. Sa5%(T) has been adopted in several recent empirical fragility and vulnerability studies (e.g. Wesson et al
2004; Colombi et al, 2008). Spence et al (1992), Singhal and Kiremidjian (1997) and Rossetto (2004) noted
that Sa5%(T) provided a better correlation with their empirical damage data than PGA. However, damage is
more closely related to the seismic energy imparted to the structure and imposed relative displacements
than to imposed forces.
Spectral velocity, Sv(T), is indicative of the peak (not total) earthquake energy, and although ground motion
prediction equations exist for the parameter evaluation, Sv(T) is rarely used in the development of empirical
vulnerability curves (e.g. King et al, 2005; Scholl 1974). A few empirical fragility studies have adopted spectral
displacement (Sd(T)) as their IM (e.g. Rossetto and Elnashai, 2003; Sarabandi et al, 2004; Colombi et al, 2008).
Sd(T) has gained importance in seismic risk assessment due to the development of displacement-‐based
methods of seismic design and assessment (e.g. Priestley et al, 2007). Furthermore, the installation of digital
strong-‐motion measuring instruments in countries worldwide has eliminated the uncertainty in the spectral
displacement determination from accelerograms, associated with noise contamination and record frequency
filtering procedures. Consequently, ground motion prediction functions for Sd(T) have been derived (e.g.
Akkar and Bommer 2007). Rossetto (2004) judges Sd(T) to correlate better with the observational damage
statistics in her database, compared to Sa(T) and PGA, and this is also observed for limited data by Sarabandi
et al (2004).
One of the main difficulties associated with the use of elastic spectral values are the determination of
equivalent vibration periods and damping coefficients for the characterisation of buildings as single degree of
freedom systems (SDOF). Consideration must be given to the likely structural failure mode and its strength
and stiffness degradation during ground excitation. It may be possible to identify the likely failure mode from
the structural configuration and seismic code used in construction. However, estimation of the probable
degradation is extremely difficult as this depends not only on the structure, but also on the ground motion
characteristics. However, Rossetto (2004) sustains that in the case of empirical damage/loss data, the effects
of inelasticity in both the seismic demand and structural response are implicitly included in the observations.
Consequently, inelasticity and period elongation during ground shaking is taken into account in the
determination of the exceedence probabilities. The inclusion of inelasticity in the IM is, probably unnecessary

20

and the structure elastic period of vibration can be used to characterise the ground motion demand for all
damage state curves.
This however does not solve the question as to which elastic period of vibration should be used to
characterise a given building class, which will naturally be composed of structures with a range of
geometrical and material characteristics. Empirical functions relating building height to fundamental period
of vibration have been used by several empirical studies, (e.g. Scawthorn et al, 1981 for mid-‐rise building
fragility curves). Seismic building codes typically include such empirical relationships (e.g. Rossetto and
Elnashai, 2003 used those in Eurocode 8), however, they tend to provide conservatively low values of the
structural elastic period as this results in higher spectral acceleration values, and consequently higher design
forces. This conservatism is not appropriate for use vulnerability studies. Crowley et al. (2004) instead
present empirical relationships for the cracked elastic period of different structures, which provide a more
appropriate estimate of fundamental period for use in empirical fragility studies but have not yet been
applied in this context. All, these relationships for estimating the structure fundamental period assume that
the building height is known or that an average height for a building class can be adopted. Another solution is
to derive the structural period that results in the best fit of the damage data to a regression curve. This is
done by Scawthorn et al (1981) who adopt spectral acceleration evaluated at a fundamental period of 0.75s
in deriving fragility curves for their low-‐rise wooden building class. However, when considering the height of
structures in the building class (1-‐2 storeys), it is questionable whether this value of fundamental period
value is representative of the elastic response of the buildings. In order to overcome the issue of
fundamental period evaluation for a population of buildings, and in recognition of the variation of
fundamental period in a building class, some reviewed studies adopt spectral response parameters averaged
over a range of structural periods (relevant to their building class) for their IMs (e.g. Scholl 1974).
Colombi et al (2008) are the only authors to derive empirical fragility curves using inelastic spectral
displacement for their IM. They assume each building class to have achieved a given ductility level at each
damage state in their fragility relationship and use this to calculate the effective period and inelastic
(overdamped) spectral displacement at this structural period. Use of inelastic spectral values presents a
problem when interpreting and/or applying the resulting fragility functions. In the first case, the IM values
associated with one structure at different damage states is different which means that the proportion of
buildings in each damage state (which should sum to 1) cannot be determined directly from a single IM
value. In addition, determination of the ductility values corresponding to the damage states implies carrying
out a pushover analysis of the building being assessed if the fragility curves are to be applied to a structure
type that differs from those in the study (for which ductility values have been proposed). The plastic
deformation associated with a damage state for buildings within a structural class will differ significantly and
this source of uncertainty should be quantified and included in the fragility calculation, but is not.
Finally, it is worth noting that there exist different definitions for peak ground motion and spectral
parameters, depending on the treatment of the two measured horizontal components of ground
acceleration. For engineering applications, Baker and Cornell (2006) noted that it is important that
seismologists and engineers use a consistent measure when estimating seismic hazard and structural
response, and this applies similarly for loss assessment. GMPEs are available for the geometric mean of two
measured components, the maximum of two measured components, and other mean or maximum
definitions. Existing studies on fragility and vulnerability relationships are not explicit on which definition has
been used; when instrumental intensity measures have been estimated from GMPEs, the original GMPE
reference would need to be consulted to determine which definition was used. Demand estimates based on
a maximum measure are likely to be better correlated with damage than geometric mean measures (Grant,

21
2011), but intensity estimates of the latter are more stable, as extreme values are averaged out (Baker and
Cornell, 2006).
2.3.3 Evaluation of Ground Motion Intensity

When evaluating an existing empirical fragility or vulnerability relationship care must be taken to assess the
method used to determine the IM values at the damage/loss sites. Ideally, ground motion recordings would
be available across the affected area. In practice, such instruments are scarce and may not cover the entire
affected area. Two exceptions are found from the literature. Yamaguchi and Yamazaki (2001) propose
fragility curves that are developed using data from 346,078 low-‐rise wood frame buildings surveyed after the
1995 Kobe earthquake in areas close to 17 accelerometers. For areas of low ground shaking (roughly half the
points), the isoseismic units used in order to get a non-‐zero value of damage are however so large they are
unlikely to represent areas of uniform ground shaking. Sarabandi et al (2004) use damage data collected near
recording stations during the Northridge 1994 (USA) and Chi Chi 1999 (Taiwan) earthquakes to develop
fragility functions for several structural classes. However, the constraints imposed by similar soil conditions
and proximity to a recording station, result in small numbers of datasets, containing small numbers of
buildings for each building class considered (e.g. 69 steel buildings in total). Hence, there exists a practical
trade-‐off between the uncertainties introduced into the fragility/vulnerability relationship from the use of
small samples versus that introduced from error in the ground motion determination.
In the majority of cases, IM values are determined from ground motion prediction equations (GMPE). In
these cases, the adequacy of the chosen GMPE for the earthquake event and site should be assessed.
Modern GMPEs are based on numerous ground motion recordings and account for differences in focal
mechanisms and soil types. However, modern GMPEs may not be available for the particular country
assessed and adopted from regions with similar tectonic environments, or the study may pre-‐date the
derivation of reliable GMPEs for the particular intensity measure and site. For example, in the fragility curves
of Rossetto and Elnashai (2003), a very limited set of GMPEs for spectral acceleration and displacement IM
evaluation are used to represent damage data from different tectonic settings. A better approach would
have been to take country or region specific GMPEs. Soil type should always be accounted for in the
determination of IMs. However, this is not always done; for example, in Rota et al (2008) all municipalities in
Italy for which they have damage data are assumed to be founded on rock. Rota et al (2008) do carry out a
sensitivity analysis to see the effect of randomly varying the PGA at their sites by up to 50%. They report little
differences in the resulting mean fragility curves. Despite this assertion, it is possible to observe from
produced figures differences of up to 10% in the exceedence probabilities of the larger damage states.
Furthermore, since Rota et al (2008) “bin” their PGA values for regression, this “binning” process may result
in less observed variability than if a direct regression were carried out.
In empirical fragility and vulnerability studies, macroseismic intensities are usually evaluated from field
observations of damage, which, results in a dependence of the x-‐axis of the derived fragility curve on the y-‐
axis. Although strong ground motion based IM values are usually derived from accelerogram readings or
GMPE, in a few cases in the literature observed damage distributions have been used to update and modify
such IM values. For example, Yamazaki and Murao (2000) adopt an iterative derivation methodology for their
fragility curves wherein the peak ground velocity (PGV) distribution derived by Yamaguchi and Yamazaki
(1999) for the Nada Ward after the Kobe (Japan, 1995) earthquake, is modified using preliminary fragility
functions. The resulting PGV distribution is then re-‐applied in the derivation of new fragility curves. The
process of PGV re-‐estimation and fragility curve updating is repeated several times and the final PGV
distribution obtained differs significantly from the original. Such processes are not recommended as the axes

22

of these fragility curves are effectively inter-‐dependent and the horizontal axis can no longer be regarded as
representative of the observed strong ground motion.
Care must be taken in adopting fragility or vulnerability curves from older studies that use spectral ordinates
as IMs. In past studies based on relatively limited computing power, standard spectral shapes (such as
Trifunac 1977 by Scawthorn et al 1981) or pseudo spectral values may have been used (e,g, pseudo spectral
displacement in Scawthorn et al 1981). The use of pseudo spectral values was investigated by Colombi et al
(2008) who considered two methods for deriving spectral displacement with 5% damping in their fragility
study of Italian buildings: (1) the pseudo-‐spectral velocity GMPE of Sabetta and Pugliese (1996), transformed
to spectral displacement via the pseudo-‐spectral functions; and (2) the use of a GMPE for spectral
displacement Sd(T) by Faccioli et al (2007). They report that the epistemic uncertainty in the GMPE has a
large influence on the fragility curves generated, as they predict widely different Sd(T) values for the same
building classes.

2.4 Vulnerability or Fragility Assessment Procedures
This section aims to critically review existing procedures used for the construction of empirical fragility and
vulnerability functions. In particular, the focus is on issues regarding the treatment of uncertainty, especially
the manipulation or combination of damage and loss data in databases, the selection of a model to express
the functions, the adopted optimisation processes, and procedures and tests used to establish confidence of
the chosen relationship to the data and to communicate the overall uncertainty in the relationship.
2.4.1 Sample Size

The sample size of a high quality database determines the reliability of the mean vulnerability or fragility
functions. The sample size depends on the type of post-‐earthquake survey, i.e. surveys undertaken by
reconnaissance teams contain small samples, whilst surveys conducted by state authorities produce large
databases. Existing empirical functions are based on a wide range of sample sizes varying from 20 (Sarabandi
et al. 2004) to 346,078 (Yamaguchi and Yamazaki, 2001). It should be mentioned that most existing functions
are based on samples of 200 buildings or above, which can be, theoretically at least, considered an
acceptable sample size. However, loss or damage observations often do not encompass a range of ground
motions sufficient for the reliable determination of vulnerability or fragility functions for all damage states
over a wide IM range (e.g. Karababa and Pomonis, 2010 acknowledged the lack of data for most building
classes to construct partial collapse and collapse damage state fragility curves for some of their building
classes). This is particularly true for the damage statistics based on data collected in the aftermath of small
magnitude events, e.g. in Dolce et al (2006), the ML 5.2 Potenza earthquake of 1990 had a maximum-‐
recorded MCS intensity of VII.
2.4.2 Data Manipulation and Combination

Section 2.2 highlighted the serious issues with data quality of available post-‐earthquake damage and loss
surveys. This means that data from post-‐earthquake surveys, even from single events, is almost never
adopted for constructing fragility and vulnerability relationship without some level of data manipulation or
database combination aiming at reducing the sampling errors (note: non-‐sampling errors are discussed in
Section 2.2.6). Data manipulation is also required for transforming each database into a form suitable for

23
combination with other databases. In what follows, the main procedures for data manipulation and
combination found in empirical vulnerability or fragility assessment literature are discussed.
Some studies address high sampling errors by aggregating data from various detailed building classes into
more general ones. For example, Braga et al (1982) combine the detailed building typologies used by the
1980 Irpinia database into three vulnerability classes in line with the requirements of the MSK-‐76
macroseismic scale. However, as buildings with different geometries and structural systems are grouped
together, uncertainty in the resulting vulnerability or fragility increases. Moreover, the ability of the obtained
general functions based on heterogeneous data to represent subclasses of buildings with very small
contribution to the overall sample size of the class is questionable (e.g. Rossetto and Elnashai, 2003).
Some of the reviewed studies combine loss or damage databases from multiple events. Several empirical
vulnerability or fragility studies combine loss or damage data from several earthquakes in the same country
in an attempt to overcome data scarcity, e.g. Colombi et al (2008) and Rota et al (2008) for Italy, Gulkan et al
(1992) for Turkey and Amiri et al (2007) for Iran, Jaiswal and Wald (2010) for casualty data in Italy, India and
Iraq. These studies largely retain the advantage of consistently assembled damage survey data for similar
asset classes (if post-‐earthquake survey procedures have not substantially changed over the considered time
frame), but significant differences may be found in adopted methods for casualty and economic loss data
collection. Although functions derived from multiple earthquake databases tend to include data over a wider
range of IMs than single event functions, due to the infrequency of large earthquake events near urban
areas, the datasets are still seen to be highly clustered in the low-‐damage/loss, low-‐IM range. In areas of
medium and low seismicity, this may be acceptable if the IM values covered by the data sufficiently
represent the ground shaking that can be generated by locally feasible events. A few studies are seen to
combine empirical loss or damage data for similar asset types from multiple events and countries worldwide
in order to obtain data over a wider range of IM values, or simply a larger quantity of data from which to
construct the functions. However, even in these cases the number of datasets available for high-‐damage
states and high-‐IM values are few (e.g. Rossetto and Elnashai, 2003). Implicit in the studies is the assumption
that the uncertainty in the seismic performance of individual buildings is larger than the uncertainty in the
performance of buildings in different earthquake. This assumption has not been addressed anywhere in the
literature. In addition, none of the aforementioned studies provided a thorough discussion on the criteria for
selecting the seismic events. Existing loss estimation methods (e.g. Coburn and Spence, 2002) and some
fragility functions (e.g. Rossetto and Elnashai, 2003) are seen to combine databases from different tectonic
environments and/or significantly different fault systems. These methods assume that the variability in IM
and resulting damage are sufficiently accounted for by the ground motion prediction equations (GMPEs)
used to estimate the IMs. However, few use GMPEs that appropriately distinguish between different faulting
mechanisms. In addition, the combined databases are associated with different degrees of reliability. This is
typically not taken into account in the curve fitting procedures used (but should be). Finally, it is well known
that fatality rates are affected by a number of factors that are specific to the event and location (such as time
of day of the event, working practices and local customs, effectiveness of response teams, etc.). Hence, the
combination of multiple event data from diverse environments into simple fatality ratio estimation models
that do not explicitly take such factors into account can add significant uncertainty to the loss estimate. For
example, this is true of the empirical fatality vulnerability functions proposed by Jaiswal and Wald (2010),
who combine casualty data for multiple earthquake events at a national level, without accounting for the
influence of population demographic characteristics, season or time of day.
Most empirical fragility and vulnerability studies aggregate the damage or loss data into isoseismic units,
which assumes that the buildings in the isoseismic unit are subjected to the same ground motion intensity.

24

The number of isoseismic units (and hence data points) that have been used to construct individual
vulnerability or fragility curves of this type range from 3 (e.g. Karababa and Pomonis, 2010) to approximately
79 (Yang et al 1989). In larger databases, individual datapoints are also seen to be aggregated into “bins” of
similar ground motion intensity values, resulting in a smaller number of datapoints for regression (e.g.
Rossetto and Elnashai, 2004; Rota et al. 2008). This tendency for aggregation might be traced to two main
reasons. The first reason is the general lack of very detailed building-‐by-‐building databases. The data are
often only available in an aggregated form. The second reason stems from the common use of least squares
methods of regression for model fitting. In the unweighted form of the least squares method the reliability of
the fitted model is based on the number of data points, which in the case of fragility, is the number of points
expressing the probability of a building reaching or exceeding a damage state given various levels of ground
motion intensity, and not the total number of buildings in the database. This means that a data point based
on 5 buildings is as important as a data point based on 1000 buildings. To avoid datapoints deriving from
small numbers of observations strongly influencing the model fit, some studies specify a minimum threshold
number of buildings for data points, below which the data points are considered unreliable and ignored (e.g.
20 buildings for Spence et al, 1992 and Karababa and Pomonis, 2010). In the reviewed empirical fragility and
vulnerability literature, the effect of different aggregation assumptions (e.g. chosen minimum thresholds of
buildings in a datapoint, size of bins, bin ranges) is not fully understood and has not been systematically
studied.
Finally, manipulation of damage and loss data in a database to prepare it for combination with other
databases usually involves the interpretation or mapping of the damage states of the data onto a different
damage scale. Direct mapping of data discretised according to a large number of damage states to scales
with fewer (or equal number of) damage states (e.g. in Braga et al, 1982, and assumed in Spence et al 1992)
introduces uncertainties. However, it is assumed in the literature that these are small and are never seen to
be explicitly quantified. Mapping from a small number to large number of damage states must assume some
distribution of the damage data within the larger damage state classes. This should not be done unless the
original damage survey forms are available and are detailed enough to allow for a larger number of damage
states to be defined (as is the case in Rota et al, 2008). Despite this, Sarabandi et al (2004) and Rossetto and
Elnashai (2003) have carried out such damage state mapping, both following a similar procedure. In the
former, a mean damage factor value is assigned to the damage scales within each damage scale used,
according to the damage scale definition. In the latter a maximum interstorey drift (ISD) value is assigned to
the damage states of all damage scales used in their database, through comparison of the damage state
descriptions with those of their proposed and experimentally calibrated damage scale. These DF/ISD values
are used to fit lognormal/beta distributions to each reported damage frequency plot for a given IM and site.
These fitted distributions are then used to derive the proportion of buildings in damage states defined by
their damage scale DF/ISD thresholds. Neither study explicitly quantifies the uncertainty introduced by
damage scale mapping.
2.4.3 Shape of Vulnerability or Fragility Functions

Vulnerability and fragility of the examined assets have been characterised as discrete or continuous functions
(see Section 2.1). In this section, the shape and parameters of the functions selected by the reviewed studies
to represent loss or damage data at one or more isoseismic units are discussed.
Post-‐earthquake survey damage data for buildings is most commonly represented as histograms, expressing
the probability of being in the predefined damage states for a specified level of intensity. Some empirical
studies are seen to transform these discrete damage functions into parametric probability distributions.

25
Braga et al (1982), Sabetta et al (1998), Di Pasquale et al (2005) and Roca et al (2006) all fit discrete binomial
distributions to their histograms of damage. This distribution is fully described by a single parameter, which
Sabetta et al (1998) correlated through a third degree polynomial with the ground motion intensity in order
to develop their representation of fragility. Instead, Rossetto and Elnashai (2003) adopt continuous beta
distributions, which are fully described by two parameters. In theory, using a model with more parameters
may improve the fit. However, the use of a continuous distribution in this case requires the correlation of
each discrete qualitative damage state with a threshold of a loss or response parameter. Such functions are
mainly judgement based, and rarely validated with experimental or observational data, thus this can
introduce additional, and perhaps substantial, uncertainty in the model.
Where continuous functions expressing the probability of buildings being in a damage state versus ground
motion intensity are proposed, these functions are seen to be either linear (e.g. Scholl 1974, and Scawthorn
et al, 1981) or exponential (e.g. Petrovski and Milutovic, 1990).
Another continuous representation of the fragility is the fragility curve. In this representation, the majority of
empirical studies adopt lognormal cumulative distribution functions (LN_CDF). The popularity of LN_CDF can
be attributed to three properties. Firstly, this function is constrained in the y-‐axis between [0, 1], which is
ideal for fitting data points expressing aggregated probabilities. Secondly, with regard to the x-‐axis, the
values are constrained in (0, +∞). This agrees with the range of almost all the ground motion intensity
measures. Thirdly, this distribution appears to be skewed to the left, and can thus, theoretically at least,
better reflect the frequency of the observations which are mostly clustered at low ground motion intensities.
The normal cumulative distribution function (N_CDF) is the second most popular regression curve in the
empirical fragility assessment literature. It is mostly preferred by studies (e.g. Spence et al, 1992; Orsini,
1999; Karababa and Pomonis, 2010) that use intensity measures that range from (-‐∞, +∞), e.g. PSI.
Nonetheless, Yamaguchi and Yamazaki (2000) express their fragility curves in terms of this distribution
despite their intensity measure being discrete and positive.
Instead of using a cumulative probability distribution, an exponential function, which is unconstrained in
both x-‐ and y-‐axis has been adopted by Rossetto and Elnashai (2003) and Amiri et al (2007). The use of a non-‐
probability distribution function to express the fragility curves may have implications in the risk assessment,
which requires its coupling with a hazard curve to produce the annual probability of reaching or exceeding a
damage state.
In the reviewed vulnerability functions, Jaiswal and Wald (2010) adopt the lognormal distribution to express
death rate in terms of the ground motion intensity, in line with most fragility assessment methodologies. By
contrast, many other casualty vulnerability functions correlate the total death rate with earthquake
magnitude due to the lack of more refined data. The use of magnitude ignores a number of significant
contributors to loss such as the source-‐to-‐site distance, the soil conditions, and the tectonic environment,
thus increasing the uncertainty in the model. In such studies, linear functions are often used to correlate
death rates and magnitudes. Similar functions are proposed for economic loss by Thráinsson and
Sigbjörnsson (1994), who correlate loss to earthquake magnitude and source-‐to-‐site distance. In the
reviewed literature, economic loss is directly correlated to other IMs through a number of different function
forms, such as linear (e.g. Petrovski et al, 1984) and power functions (e.g. Scawthorn, 1981). A different
approach is proposed by Wesson et al (2004) who fit a gamma distribution to insurance economic loss data
in each isoseismic unit, and then correlated the two parameter of this distribution with the intensity through
the use of continuous exponential and polynomial functions. Yang et al (1989) follow a similar approach to
Wesson et al (2004) in their indirect vulnerability assessment method. They fit normal distributions truncated
in [0,1] to the data in each IM value (in this case MMI Intensity), and then correlate the two parameters of

26

this distribution (i.e. the mean and standard deviation) to the IM through the use of linear functions. With
regard to the remaining indirect vulnerability functions, the mean economic loss given discrete levels of
intensity is obtained in most studies with the exception of Elefteriadou and Karabinis (2011), who obtain
multi-‐linear piecewise vulnerability functions from the aforementioned discrete values. The general forms of
relationship here described are all presented in Appendix A.
2.4.4 Statistical Modelling Techniques

Most empirical fragility and vulnerability studies reviewed focus on fitting a parametric function to post-‐
earthquake building damage and loss observations, respectively. This is effectively treated as an optimisation
problem by the reviewed studies, who thus assume that the damage/loss data is of high quality and that the
intensity measure levels can be determined with negligible uncertainty. Through these assumptions an
objective function can be formulated, which correlates the loss/damage (considered the dependent variable)
to the ground motion intensity (considered the explanatory variable). Values of the ‘true’ but unknown
parameters of the selected function (e.g. LN_CDF or N_CDF) have mainly been estimated by minimising the
objective function, and rarely by its maximisation. A discussion of the two approaches follows.
According to the minimisation approach, used almost exclusively in the reviewed empirical literature, the
objective function has been expressed in terms of the sum of least squares errors (i.e. observed -‐ predicted
values):

⎡ m ⎤ ⎡ m 2 ⎤
⎢ j =1 ⎥∑ ⎡
⎢ j =1 ⎣ ∑ ⎤
⎦ ⎥
(
θopt = argmin ⎢ ε 2j ⎥ = argmin ⎢ ⎢w j y j − f θ, im j ⎥ ⎥ ( )) (2.4)
⎣ ⎦ ⎣ ⎦

where yj is data point j based on loss or damage observations (a data point can express the damage ratio
sustained by an isoseismic unit with intensity IM=imj or the probability of buildings sustaining or exceeding a
specified damage state in this isoseismic unit); m is the number of data points; wj is the weight for data point
j; m is the number of data points; f(θ,imj) is the predicted value based on a predetermined function f(.) with
parameters θ=[θ1, θ2, …,θN]; εj is the error at point j which is considered independent and normally
distributed for each j with mean zero and constant standard deviation.
In most direct vulnerability assessment studies, the parameters, θ, are linearly combined in the proposed
functions, therefore their values are estimated by a closed form solution through the linear least squares
method. This method is also overwhelmingly adopted in the fragility assessment studies to fit cumulative
lognormal or normal distributions to the damage data. The application of this method for the estimation of
parameters nonlinearly combined in the fragility functions requires the linearisation of the fragility functions
(e.g. Yamazaki and Murao, 2000; Yamaguchi and Yamazaki, 2000; Beneddetti et al. 1998) in the form:

f −1 y j = θ1im j + θ2 + ε j
( ) (2.5)
Also required is the transformation of the data points into the form: (ln(imj), Φ-‐1(yj)) or (imj, Φ-‐1(yj)) for the
cumulative lognormal and normal distributions, respectively. However, this transformation may introduce
bias to the estimation of the mean fragility curves. This bias is more evident when yj is very close to the
extreme values of 0 and 1, as the transformation is not feasible for these values, limiting the applicability of

27
this approach. Procedures (e.g. Porter et al, 2007) that have attempted to deal with the transformation of
yj=0 seem questionable (Baker, 2011). However, the transformation of the selected model does not lead to
bias in cases where an exponential function (e.g. Eq.(A.6)) is fitted.
A handful of fragility assessment studies avoid this bias (Rossetto and Elnashai, 2003; Amiri et al. 2007; Rota
et al. 2008) through the numerical estimation of the parameters by the use of the nonlinear least square
method. However, the nature of damage data typically is seen to violate the assumptions of this approach
with regard to error normality and error homoscedasticity (i.e. constant error for each level of IM). For
example, the normality of the error is clearly violated by the fact that exceedence probability is a variable
bounded between 0 and 1. In addition, the homoscedasticity assumption is violated given that the
uncertainty in seismic performance for extreme levels of intensity is considered lower than for the
intermediate intensity measure levels. Some studies attempt to address the heteroskedasticity requirement
by using weighing techniques. Sabetta et al. (1998) weight the least squares by the number of buildings in
each isoseismic unit. The use of buildings by the latter study, however, implies sample sizes that vary from
few tens to few thousands, and this difference is unlikely to reduce the heteroskedasticity. A different
weighting technique is adopted by Rota et al (2008), where datasets falling within an IM bin are grouped into
one data point for the regression. The bootstrap technique is adopted to find the variance values to
represent the scatter in the mean data point in each bin. The inverse of this variance is used to weight the
least squares in Eq. (2.4) in order to reduce the impact of bins with small numbers of data points.
Nonetheless, the effect of these schemes on heteroskedasticity is questionable. Instead of weighting
techniques Jaiswal and Wald (2010) address the heteroskedastic error by use of a different objective
function. The level of bias introduced in the generated mean fragility curves by fitting models using a
nonlinear least squares method with violated assumptions is an issue of ongoing research (e.g. Ioannou et al.
2012). There are indications, however, that the differences compared to more realistic models (e.g. the
generalized least squares model) are evident in the tails of the distributions (e.g. Lallemant and Kiremidjian
2012).
Few existing studies adopt the maximum likelihood method in order to estimate the parameters of statistical
models used in the vulnerability or fragility assessment. Such studies maximise the likelihood function, which
has the form:
⎡ m ⎤
θopt = argmax ⎢ ∏( )
f y j ; θ, x j ⎥ (2.6)
⎢ j =1
⎥
⎣ ⎦

where f(.) is the probability density function of a variable with parameters θ;. This approach was used by
Yang et al (1989) and Wesson et al (2004) to fit truncated normal and gamma distributions, respectively, to
their isoseismic unit loss data. Ioannou et al. (2012) also adopt this method in order to fit a generalised linear
model to damage data. According to the latter study, individual fragility curves were constructed from data
points having the form (xj,(yj, mj-‐yj)), which express the intensity level imj that affect an isoseismic unit j
causing yj buildings to suffer damage DS≥dsi and causing mj-‐yj buildings to suffer damage DS<dsi. A
generalized linear model is then constructed, which considers that the counts of data with DS<dsi and DS≥dsi
for each unit j follow a binomial distribution, having a likelihood function in the form:

28

M
⎛ m j ⎞ yj m j −y j
L ( θ ) = ∏ ⎜ ⎟ µ ( im j ; θ ) ⎡⎣1 − µ ( im j ; θ )⎤⎦ =
y
j =1 ⎝ j ⎠
(2.7)
M
⎛ m j ⎞ yj m j −y j
( )
= ∏ ⎜ ⎟ Φ θ1 ln ( im j ) + θ 2 ⎡1 − Φ θ1 ln ( im j ) + θ 2 ⎤
j =1 ⎝ y j ⎠
⎣ ⎦ ( )

The mean μ(imj;θ), which fully defines the binomial distribution, is assumed to follow the probit function
with log(IM), which is essentially equivalent to the lognormal cumulative distribution function as presented
in the right side of Eq.(2.7). The parameters of this statistical model are estimated by maximizing the
likelihood function. Contrary to the model expressed in Eq.(2.4), this model provides a better, in theory at
least, representation of the damage data given that (i) it recognizes that the fragility curves are bounded in
[0,1], (ii). it successfully relaxes the assumption of constant variance of residuals by accommodating for
smaller uncertainty in the tails of the mean µ, and higher in the middle, and (iii). It takes into account that
some points have a larger overall number of buildings than others.
Determination of the goodness of fit of the selected models to the adopted data involves procedures for
assessing the validity of the assumptions on which the fitted statistical model fitted is based, and on ways to
identify the best model amongst acceptable models. Despite the importance of these procedures in
determining the reliability of proposed models, only Ioannou et al. (2012) provide any detail on goodness of
fit checks carried out, and recognize, in their case, the need for more detailed damage data and perhaps
more sophisticated statistical models. A small number of existing fragility and vulnerability studies provide an
incomplete goodness-‐of-‐fit assessment by accompanying their curves with the coefficient of determination
(R2) in the case that linear least squares regression is used or the sum of the least squares if nonlinear
regression analysis is adopted. However, these measures cannot highlight potential violation of the
assumptions on which the statistical models are based. In two studies (King et al. 2005; Frolova et al. 2011),
the Smirnov-‐Kolmogorov goodness-‐of-‐fit test was used to assess the suitability of the selected probability
distribution.
Existing empirical vulnerability and fragility studies overwhelmingly fail to provide a deeper insight on the
goodness of fit of the selected model by quantifying the uncertainty in the estimated parameters and
constructing confidence intervals around the mean fragility curves, which can illustrate the uncertainty
introduced by the often limited number of data. Notable exceptions in the fragility literature are Braga et al
(1982), Orsini (1999) and Amiri et al (2007). Braga et al (1982) propose upper and lower bounds for the
parameter of their discrete damage probability matrices but do not associate these bounds with a confidence
level. Similarly, Orsini (1999) estimates upper and lower bounds for their damage probability matrices for
given levels of MSK, corresponding to the upper and lower bound of PSI for this intensity measure. Amiri et al
(2007) also estimate the 90% confidence intervals of the parameters of their fragility curves. However, the
envelope of the fragility curves obtained from the four combinations of the upper and lower bound of the
parameters often takes negative values, perhaps due to the small number of data points in combination with
the inappropriate statistical model selected. With regard to vulnerability assessment studies, Wesson et al
(2004) estimate confidence intervals for the parameters of the gamma distribution for each isoseismic unit
for which loss data is available by using the bootstrap technique. Only two studies are instead seen to
construct prediction intervals for their fragility curves, which account for the above uncertainty as well as the
typically large scatter in the data points from different isoseismic units. In particular, Rossetto and Elnashai
(2003) attempt to estimate the 90% prediction intervals by taking into account the non-‐constant error noted
in their analysis. It is noted that these prediction intervals could be biased given that the non-‐constant error
violates one of the assumptions on which their regression is based. Ioannou et al (2012) use a bootstrap in

29
order to construct point wise prediction intervals corresponding to their data points and highlight the
difficulties in projecting these intervals for future groups of buildings and the need for further research.
2.4.5 Method of Conversion of Damage to Loss in “Indirect” Vulnerability Curves
The estimation of indirect vulnerability from fragility functions requires the use of damage-‐to-‐loss conversion
functions. The treatment of uncertainty in these functions by the reviewed studies is discussed here.
In the case of economic losses, damage factors are commonly correlated to the discrete qualitative damage
states by the use of field data, insurance claims, engineering judgement, or combination of the
aforementioned sources. With the exception of Yang et al (1989) and Spence et al (2003) who adopt different
damage factors for each building class, overwhelmingly the damage factors adopted are assumed
independent of the building typology in existing vulnerability studies. This is counterintuitive, as different
building types damaged to the same level are unlikely to have the same cost of repair.
Damage-‐to-‐loss functions are commonly expressed deterministically in terms of a best estimate level of the
damage factor given a damage state. The only exception is noted in the study of Dolce et al (2006) who
introduce the aleatory uncertainty in these functions by adopting beta distributions of the damage factor for
each damage state. It should be mentioned that the epistemic uncertainty of these functions, expressed in
terms of a family of probability distributions of the loss given a damage state, is not explicitly treated
anywhere in the reviewed studies.

Indirect vulnerability approaches to casualty estimation typically only include fragility functions for the
damage state of collapse in their calculation of fatalities (e.g. Murakami, 1992). The estimated number of
collapsed (and sometimes heavily damaged) buildings is converted to number of fatalities by considering that
a defined proportion of the people likely to be in the building type are killed (lethality ratio). Quite complex
models are seen to exist for determining both the stated proportion and level of building occupancy. Such
models (e.g. Coburn and Spence, 2002; Nichols and Beavers, 2003) combine a number of parameters that are
supposed to represent the proportion of occupants directly killed, those trapped, and the proportion of
those trapped that are then rescued. The numerical values of the parameters and the parameter
distributions change according to building type, and are based on data but undoubtedly contain some level
of judgement. In existing casualty studies, the level of occupancy of buildings is often estimated from census
data. These values of occupancy can be used directly, or modified through empirical or judgement-‐based
factors to take into account the influence on occupancy of tourism, occupant behaviour, season and time of
day of earthquake, (amongst others). Very complex models for converting building damage into fatalities are
not currently justified by the data used for their calibration, as this is rarely very detailed, plentiful or reliable
(see also Section 2.2.4).

2.5 Validation Procedures
Ideally, the validity of vulnerability or fragility functions should be determined through comparison with
independent post-‐earthquake observations. However, in the reviewed literature only two studies are seen to
do this. Orsini (1999) compare the overall loss obtained indirectly from their fragility curves (which are
constructed from the 1980 Irpinia damage database) with the overall loss caused by the 1997 Umbria-‐
Marche earthquake. Spence et al (2003) compare their DPMs (developed from observations from several
Turkish earthquakes) with the damage frequencies in two sites affected by the Kocaeli 1999 (Turkey)

30

earthquake. However, it is unclear from the study whether or not the damage data from Kocaeli used for the
validation was also included in the construction of their DPMs. The lack of interest in this type of validation
can perhaps be attributed to the common practice of using all available datasets for the construction of
vulnerability or fragility functions in order to get adequate sample sizes.
The main procedure adopted for validation of the vulnerability or fragility functions in existing empirical
studies is the comparison with other functions constructed for similar asset types by the same or other
authors. Orsini (1999) compares the upper and lower bound of his DPMs for each level of intensity with their
counterparts obtained for the Umbria-‐Marche 1997 earthquake. Yamaguchi and Yamazaki (2001) compare
their fragility curves with their counterparts obtained for similar building types by Miyakoshi et al (1998)
using the same database. Colombi et al (2008) and Rossetto (2004) validate their empirical fragility functions
with the analytical vulnerability curves they developed. Rota et al (2008) use the judgement-‐based approach
proposed by Giovinazzi and Lagomarsimo (2006) to validate their fragility curves. Eleftheriadou and Karabinis
(2011) compare the multi-‐linear piecewise loss curves obtained indirectly from their DPMs with their
counterparts estimated by hybrid and empirical methods for similar Greek building types. In all these cases,
the comparisons made are visual only, and no systematic comparison of quantitative values is carried out.
3 Rating System for Empirical Fragility and Vulnerability Functions

The here proposed rating system, is a modification and extension of the rating system for all vulnerability
functions proposed by Porter (2012). The rating system here is specific to empirical fragility and vulnerability
relationships and consists of two components, namely the overall quality of the fragility/vulnerability curves
and their relevance to the needs of a future application as depicted in Figure 3.1. The description of the four
main attributes to these two components, also presented in Figure 3.1, is briefly outlined in what follows.

Data quality: An overall rating of data quantity, constrained categories, excitation observations, and loss
observations.

• Damage/Loss observations: Each estimate of loss or damage state is reasonably accurate, verified by
some quantitative measure.

• Excitation observations: The estimates of excitation (shaking, deformation, ground failure, etc.) to
which the specimens were subjected are reasonably accurate. In addition, the observed levels of
excitation span up to the highest level that most assets are likely to experience in their design life.

• Data quantity: The number of observations is sufficient to draw statistically useful conclusions. The
number of observations for each considered asset and damage/loss state represented are sufficient
to reasonably model the uncertainty in the damage/loss at given IM values.

31
• Constrained categories: The observations are clearly made from specimens of the asset category in
question, and the category is clearly defined (passing the so-‐called clarity test).

Components Attributes Criteria
Damage Observations
Excitation Observations
Database Quality
Constrained Building Class
Data Quantity
Treatment of uncertainty
Hindcasting
Overall Quality Rationality
Cross-Validation
Rating First Principles
Scheme
Documentation Quality Documentation Quality
Relevance Representativeness

Figure 3.1. Components, attributes and criteria of the proposed rating system.

Representativeness: The specimens observed are broadly representative of the diversity of assets in the
category, considering material properties, building configuration, detailing, geographic extent, and variety of
failure modes.
Note: This criterion refers here to the representativeness of the vulnerability or fragility curves for the needs
of a particular future project. For example, the indirect empirical vulnerability assessment of casualties is
based on the presence of reliable fragility functions corresponding to the collapse and heavy damage.
However, the rating of this criterion for the functions presented in the compendium (Appendix B) is not
presented. Instead, future users are encouraged to rate the representativeness of existing studies according
to the specific needs of their project.

Rationality: An overall rating of hindcasting, cross validation, first principles, and treat uncertainty:

• Treat uncertainty: The author identifies and treats the major sources of uncertainty in asset category
and value, excitation, and loss or damage. By “uncertainty in asset category,” the question is how
well constrained the observations are as to the membership of the observed assets in the asset
category. If replacement value or construction cost is a parameter of the model, how well
established are those values? Unless specimens were instrumented or very near instrumentation,
the excitation to which specimens were subjected can be highly uncertain, especially if estimated
using ground-‐motion prediction equations or in terms of macroseismic intensity. Is the excitation so
crisply defined that there can be no ambiguity? For example, spectral acceleration response at some
specified damping ratio and period can be ambiguous unless direction is specified. Damage is more
certain if it is defined in terms of measurable quantities, less if qualitative and highly subject to

32

interpretation. Loss is more certain if the repair costs (or other loss measures) were actually
recorded and distinct from modifications or other costs unrelated to damage, less if loss were
estimated, or otherwise subject to different interpretation by different observers.
• First principles: The vulnerability function or fragility function seems reasonable in light of
engineering principles of hazard analysis, structural analysis, damage analysis, and loss analysis. The
rationality of analytically derived structural vulnerability and fragility should reflect the degree to
which the structural analysis follows current, peer-‐reviewed procedures.

• Hindcasting: The vulnerability function or fragility function reasonably hindcasts loss or damages in
some significant past event, especially if the data source and events being hindcast are independent.

• Cross validation: The vulnerability function at least roughly agrees with some prior accepted model.
If it disagrees, the disagreement seems reasonable in light of shortcomings in the past model, or
differences between the asset classes of the past model and the one in question.

Documentation quality: All the necessary inputs, outputs, and analytical steps are clearly documented to a
level that will allow the study to be reproduced by others. The documentation is readily available to future
users. The documentation has been independently peer reviewed.

The proposed system allows for four possible values to be assigned to each criterion:
• ‘H’: Superior, meaning little, if anything, could have been done better.
• ‘M’: Average, meaning the work is of acceptable quality, though there are areas for improvement or
further research.
• ‘L’: Marginal, means that the work is acceptable for use but only if there are no practical
alternatives; and much improvement or further research is needed.
• ‘N/A’: Not applicable means that the rating measure cannot be applied.

The above rates are assigned to each criterion when the details conditions presented in Table 3.1-‐Table 3.2
are met for a given empirical function. It is acknowledged that the rating exercise is prone to subjectivity, and
that interpretations of the general criteria may change according to intended use and understanding of the
reviewer.

33

Table 3.1. Rating the overall quality of existing empirical fragility functions.
Attribute Criterion Rate Assigned when:

Data Loss or Damage H Damage scales or loss measures are clearly defined. Negligible non-‐sampling errors. Significant non-‐sampling errors have been
quality Observations acknowledged and reduced using adequate methods.
M Damage scales or loss measures are clearly defined but some significant non-‐sampling errors have been treated by relying on
assumptions which are not checked.
L Damage scales or loss measures are defined with ambiguity. Significant non-‐sampling errors have not been reduced or reduced
with questionable procedures.
Excitation H The IMLs have been determined from ground motion recording stations or GMPEs, and more than one intensity measure has
Observations been used in order to identify the one that fits the data best. The influence of the uncertainty in the ground motion in the
fragility or vulnerability functions has been investigated.
M The uncertainty in IM has been partially investigated or if more than one IMs has been used for the vulnerability or fragility
L assessment.
IMLs are interdependent with the observed damage data. If they used a single intensity measure and did not explore any other
Constrained H sources ocf lasses
Building uncertainty.
are defined in terms of building material, lateral-‐load resisting system, height and seismic code (age).
Building Class M Building classes are defined in terms of building material, lateral-‐load resisting system or in terms of vulnerability class, e.g.
L EMS98.
Crude
building classes are defined, e.g. RC buildings, RC frames, adobe buildings from worldwide databases.
Data Quantity H For continuous functions: Sample sizes ≥200 damage or loss observations. For aggregated damage data, a minimum of 20
observations per bin of IM is used for a minimum of 10 bins.
For discrete functions: Sample sizes ≥200 observations with a minimum of 20 structural units per bin of IM.
M For continuous functions: sample sizes between 20 and 200. For aggregated damage data, number of bins of IM between 5
and 10 are used with a minimum of 20 observations per bin.
For discrete functions: A minimum of 20 observations per bin, for most bins of IM.
L For continuous functions: Sample sizes <20 units or units aggregated in <5 bins of IM.
For discrete functions: Less than 20 observations per bin, for most bins of IM.
Rationality First Principles H Obtained curves follow expected trends and they do not cross.
M Not applicable.
L Obtained curves violate the first principles, e.g. fragility curves cross.
Treatment of H Acceptable assumptions regarding data manipulation. Appropriate statistical models are selected and diagnostic tools
uncertainty M Acceptable
demonstrate assumptions regarding
their goodness of fit. data manipulation. Appropriate statistical models are selected, but diagnostic tools fail to
demonstrate their goodness of fit.
L Unacceptable assumptions regarding data manipulations. Inappropriate statistical models are selected.
Hindcasting H Independent damage or loss data are predicted by the fragility or vulnerability function, respectively.
M The overall damage or loss of an independent event is well predicted by the available functions. However, a more detailed of
damage or loss hindcasting has not been done.

34

Table 3.1. Rating the overall quality of existing empirical fragility functions (continued…).
Attribute Criterion R Assigned when:

Hindcasting L Independent data are not reasonably predicted by the fragility or vulnerability function.
Independent data has not been used to assess the predictive capacity of the functions.
H Obtained functions are compared with existing functions and the agreement or disagreement of the results is thoroughly and reasonably
Rationality Cross-‐ justified.
Validation M Obtained functions are compared with existing functions and the agreement or disagreement is not accompanied by reasonable
L justification.
Obtained functions are not compared with existing functions.
Documentation quality H All the necessary inputs, outputs, and analytical steps are clearly documented and the work is reproducible.
M Only partial information regarding the aforementioned issues has been addressed in the work.
L Insufficient information is provided to the fragility or vulnerability function or the methodology.

Table 3.2. Rating the relevance of existing empirical fragility functions to the needs of future application.
Attribute Criterion Rate Assigned when:

Representativeness H The range of IMs (for which the functions have been obtained) includes the required levels or range of IM.
The building class and region for which the function has been obtained is exactly the same as with the required class.
The description of the damage scale (or state) is appropriate for the needs of the user’s study.
M The required intensity measure level is 20% greater or smaller than the min or max value of the IM range for which loss or damage data
are available. The building class of the function is a subset or includes the required class. The same applied to region.
L The range of IMs (for which the functions have been obtained) is very different than the levels or range of required IMs.
When the building class is not a subset or does not include the required class or region. The description of the damage scale is too crude
for the needs of the user’s study.
3.1 Application
The rating exercise has been carried out by the authors for empirical fragility and vulnerability functions found in the literature and the results are presented in
Tables 3.3-‐3.6. The resulting qualitative scores are included as fields in the compendium of empirical functions described in Section 4. The ratings are based on
the documentation easily available to the authors and might change if further information is made available. Where the same reference proposes multiple
functions, each is rated separately in the compendium with Table B.1 and Table B.2 only presenting a general overall rating. Examples of how the rating system

35
is applied are presented in detail in Tables Table 3.3 to Table 3.6 for four functions: the single event DPMs of Braga et al. (1982), the multiple event fragility
curves of Rota et al (2008b), the indirect economic vulnerability relationship of Yang et al. (1989), which are all for Italian earthquake events, and the indirect
fatality vulnerability functions of Murakami et al. (1992) for Armenia.

Table 3.3. Rating for the damage probability matrices proposed by Braga et al. (1982).
Attribute Criterion Rate Comments
Data Quality Data Quantity H 41 datasets are used for each of the three vulnerability classes of structures considered. These span the three MSK Intensity
levels considered.
Constrained M The 13 building classes of the survey data are combined into three vulnerability classes, judged to show similar earthquake
Building Class response characteristics and defined to be consistent with MSK-‐76 vulnerability classes.
Excitation L Macroseismic intensity (MSK-‐76) is used as the IM, which has an interdependence with the damage data, especially as, in this
study, its values are determined from the damage data directly rather than from field observation.
Damage H The surveys are very detailed and evaluate damage to both the vertical and horizontal elements of the structure. The surveys
Observations are carried out by military engineers and architects. No large aftershocks are reported.
Rationality Hindcasting L The proposed DPMs represent the data used but are not verified with independent data.
Cross validation L No cross-‐validation is carried out with other studies.
First principles H The proposed method and data is reproducible and the results do not invalidate engineering principles.
Treatment of M Uncertainty is taken into account to an extent in the proposal of binomial distributions for the DPM entries and the provision of upper
uncertainty and lower limits. However, these limits are not associated confidence levels. Also bias in sampling the municipalities is not addressed.
Documentation Documentation H Well-‐documented procedure and data collection survey method.
quality

36

Table 3.4. Rating for the fragility curves proposed by Rota et al. (2008b).
Data Quality Data Quantity H For most of the fragility curves for the 23 building classes specified, more than 10 datasets are available for each damage state
curve. However, in the compendium, an ‘M’ rating is given to the fragility curves associated with some of the building classes
where data was insufficient for the definition of collapse state curves.
Constrained H Masonry building categories account for structural system, floor system, presence of tie rods and irregularities. RC building
Building Class classes include structural systems, seismic code, and height categories.
Excitation M PGA is used as the main IM and is evaluated using a GMPE. Housner Intensity is also used for the IM. However, rock conditions
are assumed for all affected locations.
Damage M The surveys from the five Italian earthquakes considered use a detailed and approximately consistent survey methodology and
Observations damage scale. Where the damage scale differs, the original survey data has been consulted in order to correctly map damage
states.
Rationality Hindcasting NA The study results are not verified with independent data but represent well their used data.
Cross validation H The PGA-‐based fragility curves are cross-‐validated with the judgement-‐based curves proposed by Giovinazzi and Lagomarsino
(2007). They do not show a good agreement due to differences in relationship form, the use of macroseismic intensity to PGA
conversion by the latter, etc.
First principles H The proposed method and data is reproducible and the results do not invalidate engineering principles.
Treatment of L An inappropriate statistical model is fitted.
uncertainty
Documentation Documentation H Well-‐documented and reproducible procedures.
quality

37
Table 3.5. Rating for the indirect economic vulnerability relationship proposed by Yang et al. (1989).
Data Quality Data Quantity H 76 datasets are used which report damage to rooms in 76 municipalities affected by the 1976 Friuli, Italy, earthquake.
Constrained Building L Rooms are used rather than buildings as the structural unit. The rooms are grouped by age of construction; however all are built
Class prior to 1975 when the first seismic code was introduced in Italy. However, the age classes are used to better estimate the DFs.
Excitation L Macroseismic intensity (MSK) is used as the IM, which has been evaluated from field observation. There are insufficient
observations for intensities VI and X, for which data is extrapolated.
Loss Observations L Details of the survey are not given in detail but a government-‐led reconnaissance based report is referenced. Average
replacement costs for the rooms are taken from the survey report. The authors estimate the number of undamaged rooms from
census data collected in 1971.
Rationality Hindcasting L The proposed matrix of average damage factors when used by the authors to recalculate losses in municipalities does not give a
good estimate of the observed losses.
Cross validation L No cross-‐validation is carried out with other studies.
First principles H The proposed procedure does not violate any first principles.
Treatment of L Sources of uncertainty in the damage and loss data are not taken into account explicitly or quantified. These are insufficient to
uncertainty determine the vulnerability (mean damage factor) values at MSK intensities VI and X (the limits of their vulnerability matrix).
Documentation Documentation M Moderately well documented procedure.
quality

38

Table 3.6. Rating for the indirect economic vulnerability relationship proposed by Murakami et al. (1992).
Data Quality Data Quantity L Adopts damage data from 10 towns surveyed by EERI after the 1988 Spitak, Armenia, earthquake. The town surveys are
associated with three MSK intensity levels. However, in the case of pre-‐cast concrete panel structures and precast concrete
frame structures, only two observations are available for deriving the DPMs. Also, in the case of stone masonry and
composite stone masonry buildings, respectively 3 and 6 of the 10 observations have sample sizes less than 20 buildings. The
damage survey concerns only damaged buildings.
Constrained Building L Multistorey residential buildings are used that belong to four categories of construction material.
Class
Excitation L Macroseismic intensity (MSK) is used as the IM, estimated from observed data.
Loss Observations L The damage surveys are carried out by experts following a sound and unbiased damage survey method. Occupancy level is
assumed from another empirical study. Lethality ratios are determined for the different building types from casualty data
collected for five areas by another author, who does not state the building class where the deaths occurred. Predominant
building classes are assigned to each area from background knowledge, in order to estimate the lethality ratios. There is bias
in the collected data towards damaged buildings.
Rationality Hindcasting L No hindcasting using independent data is available.
Cross validation L No cross-‐validation is carried out with other studies
First principles H The proposed method is reproducible.
Treatment of L Sources of uncertainty in the damage and loss data are not taken into account explicitly or quantified.
uncertainty
Documentation Documentation quality H Well documented procedure.

39
4 Compendium of Existing Vulnerability or Fragility Functions
4.1 Basic Information
A compendium of existing vulnerability or fragility functions has been compiled in MS Access (2010), wherein
each row of information is termed “record” and each column “field”. For the constructed compendium, the
fields are classified into 10 general categories as presented in Table 4.1. Each record provides information
regarding the vulnerability or fragility functions obtained for a specified building class by an existing study.
Each vulnerability or fragility relationship is also rated according to the rating system described in Section 3,
and each entry is accompanied by a brief commentary. The continuous functions are all included in the
compendium, illustrated in the appropriate field in terms of their parameters (see Table 4.2) and/or shape
(see Figure 4.1). With respect to the discrete functions, only those providing values of the levels of loss or
damage corresponding to two levels of intensity or more are included in the compendium. In the case of
indirect vulnerability functions, only those that adopt empirically constructed fragility functions are included
in the compendium, irrespective of the nature of the damage-‐to-‐loss functions used. The latter damage-‐to-‐
loss functions are also reported in the compendium.

40
Table 4.1. Basic information provided in the compendium of existing vulnerability or fragility functions.
General category Field Description Properties Example
Existing Study PText Karababa and
Reference Reference based on the author-‐date reference system used by GEM.
Pomonis (2010)
Type of assessment Type of assessment followed by study, e.g. fragility, direct or indirect vulnerability. PText Fragility
Damage and Loss measures Damage scale The main damage scale adopted by the study. PText EMS-‐98
No of DS Number of damage states used by the main damage scale. Number 6
Other DS? Did the study adopt more damage scales? (Yes-‐No) Y-‐N No
Other DS The alternative damage scales used by the examined study. PText
No of other DS Number of damage states used by the alternative damage scale. Number
Definition of the loss adopted by a vulnerability assessment study, e.g. fatality PText
Loss Parameter
rate.
Building Classification GEM Building Class The building class according to GEM taxonomy system. PText
Does the building class account for the construction material of the buildings? Y-‐N
Construction material? Y
(Yes-‐No)
Structural System? Does the building class account for the structural system? (Yes-‐No) Y-‐N Y
Age? Does the building class account for the age? (Yes-‐No) Y-‐N N
Height? Does the building class account for the design code? (Yes-‐No) Y-‐N Y
Does the building class account for the construction material of the buildings? Y-‐N
Design Code? Y
(Yes-‐No)
Irregularity? Does the building class account for the irregularity? (Yes-‐No) Y-‐N N
Vertical Material The material of the vertical structure of a building class, e.g. RC, M. Ptext RC
Infill? Do the buildings of each class have infill walls? (Yes-‐No-‐Unknown) Y-‐N-‐NA Yes
PText Hollow clay
Infill Material Material of the infill walls, RC.
masonry
Structural System Description of the structural system. PText Non-‐ductile frames
Horizontal Material The material of the horizontal structure of a building class. PText RC
Flooring system The structural system of the floors, e.g. rigid, flexible. PText
Irregularity Horizontal or vertical irregularity of the building class, e.g. PYLOTIS. PText*
GEM Height General description of the height of a building class, e.g. low-‐rise. PText Low-‐Rise
Max No of storeys Maximum number of storeys of the examined building class, e.g. 4 storeys. Number 4
41
Table 4.1. Basic information provided in the compendium of existing vulnerability or fragility functions (continued).
Building Classification Occupancy Use of the buildings PText
Ground Motion PText
Type of IM Type of IMs, e.g. macroseismic intensity (MI), ground motion parameters (GMP). MI
Intensity
IM The main ground motion intensity measure used by each study, e.g. PGA. PText PSI
Range of IM Range of IM values of the data, e.g. 0g-‐1g. FText
Main IM Estimation Method The main way of determining the levels of IM, e.g. observed data.
Other Methods IM? Did the study adopt alternative ways to determine levels of IM? Y-‐N
Other Methods IM Alternative methods used to estimate levels of IM, e.g. recorded ground motions. PText
GMPE_TE The ground motion prediction equation and or transformation equation used. PText*
Other IM? Did the study adopt alternative intensity measures? (Yes-‐No) Y-‐N No
Other IM_1(-‐N) Alternative IM adopted by a study, e.g. Arias Intensity. PText
Other Methods_IM_1(-‐N) Methods used to estimate the level of the alternative IM. PText
Other GMPE_TE_1(-‐N) The equation that has been used to estimate the IM (if available). PText
Damage-‐to-‐Loss Source of D-‐L Relationship The methodology used to obtain these functions, e.g. empirical, expert PText

functions judgement.
D-‐L Relationship The D-‐L Relationship, e.g. parameters, shape of the probability distribution of loss File.xcl

for each damage state.
D-‐L Uncertainty Whether deterministic D_L function, or whether the aleatory and/or epistemic PText

uncertainty are taken into account
Data quality/quantity Country/ies Name of the country of each database used, e.g. Greece. PText Greece
(used for the Source Source/s of data, e.g. 1973 San Fernando database. PText* 2003 Leukada
construction of Mechanism The focal mechanism of each event. PText Strike-‐slip
vulnerability or fragility
Depth The focal depth of each event, e.g. 15km. Number 12km
functions)
No Event Number of events, e.g. single event. FText Single
No Assets Number of suitable assets (e.g. buildings, casualties) used for the construction of Number
3079
the examined relationship, e.g. 1000 buildings.
Non-‐sampling Errors? Is the damage/loss databases contaminated with non-‐sampling errors? (Yes-‐No-‐ Y-‐N-‐NA
Yes
NA(Unknown))
Non-‐sampling errors The non-‐sampling errors PTest
Addressed non-‐sampling The errors which were addressed by the use of a rigorous procedure. PTest

errors
Isoseismic Unit The isoseismic units adopted for the regression, e.g. municipality. PTest District
No of Data Points Number of data points used for the construction of the regression analysis, e.g. 10 Number
33
data points.

42

General category Field Description Propeties Example

Data quality/quantity Min No of Assets Data Minimum number of assets, e.g. a minimum of 20 buildings is used in Number
20
(used for the construction of Point each data point.
vulnerability or fragility Min No of non-‐zero Data Minimum number of non-‐zero data points used for the construction of Number
6 for DS4
functions) Points per DS each fragility curve.
Method for constructing the Type of analysis The analysis used by the examined study, e.g. regression, univariate PText
vulnerability or fragility distribution fitting. Regression
functions.
Method Methods used to manipulate the data, e.g. maximum likelihood or least PText Non-‐linear least
squares. squares
Algorithm The algorithm used. PText Newton-‐Raphson
Other Analysis? Did they use additional analysis to construct vulnerability or fragility Y-‐N
No
functions? (Yes-‐No)
Other Analysis Additional analysis used to manipulate the data in order to correlate PText

them with IM, e.g. regression.
Other Method Methods used to manipulate the data, e.g. maximum likelihood or least PText

squares.
Other Algorithm Algorithm used to perform the additional methods. PText*
Type of relationship Type of relationship adopted by a study, e.g. DPMs, Fragility Curves, PText
Fragility Curves
Loss Curves.
Form of relationship Functional form of the relationship, e.g. linear, discrete. PText Normal CDF
Relationship File.xcl or Error! Reference
Contains the plot of vulnerability or fragility functions and where File.png source not found.,
available the corresponding parameters. Error! Reference
source not found.
Goodness of fit? Does the study comment on the goodness of fit of the relationship to Y-‐N
No
the data? (Yes-‐No)
Goodness of fit The method of goodness of fit test adopted for testing the fit of a PText

relationship.
Measurement Error in X Was the epistemic uncertainty in the X axis of the examined relationship Y-‐N
No
axis? taken into account? (Yes-‐No)
Type of Confidence The measure of confidence provided by the study, e.g. 90% confidence PText

intervals
Val. Raw Data? Are the functions validated with independent data? (Yes-‐No) Y-‐N No
Source of Raw data Source of independent data, e.g. 2003 Leukada. PText
Val. Existing Study? Are the functions validated by comparing them with existing functions?
Y-‐N No
(Yes-‐No)

43
Val. Existing study The existing study adopted for the validation. PText M

Ranking system Data Quantity PText H
Constrained Categories PText L
Excitation PText H
Loss Observations PText H
Representativeness PText M
Check section 3.1 for definition.
Hind Casting PText H
Cross Validation PText L
First Principles PText L
Documentation Quality PText H
Overall Rating Score PText
General Comments General Comments General comments regarding the ranking of the examined FText Well reported survey. Survey focussed
study. on damaged structures, undamaged
determined from 2000 census. Damage
data contains buildings damaged from a
major aftershock as well as the
mainshock. Constant standard deviation
assumed for all fragility curves and mean
regressed for. Curves for high damage
states (D4 and D5), “borrowed” from
Coburn and Spence (2002), where
no/insufficient statistics to construct
curve. Curves adopt PSI as IM and are
compared to curves by Coburn and
Spence (2002). No measurement error
accounted for in IM or loss values. No
confidence or prediction intervals
defined.
Notation PText: Text with predefined options.
PText* Text with predefined options, which allows for multiple entries.
N-‐Y(-‐NA): Yes-‐No(-‐Unknown) options.
FText: Free text.
_1(-‐N): This implies that N fields are included which account for the alternative IM measures.

44

Table 4.2. Parameters of the fragility curves (mean μ and standard deviation σ)
Fragility Curve constructed by Karababa and Pomonis (2010).
1.0
DS Μ σ
ds1 7.2 2.5
ds2 9.9 2.5
P(DS≥dsi|PSI)
ds1 ds3 11.5 2.5

0.5 ds2 ds4 13.5 2.5
ds3 ds5 14.1

2.5
ds4
ds5
0.0
0 5 10 15 20
PSI
Figure 4.1. Fragility curves constructed by Karababa and Pomonis (2010).

45
4.2 Global Coverage of Existing Empirical Vulnerability and Fragility Functions
The compendium includes existing empirical vulnerability and fragility functions, which provide loss or
damage for at least two intensity measure levels. The compendium contains 245 empirical functions (see
Appendix A). Fragility functions constitute approximately 80% of these functions. Of the included 55
empirical vulnerability functions, 65% are obtained directly from loss data. 80% of the reviewed existing
vulnerability functions express loss as the economic loss due to direct damage sustained by affected
buildings, whilst the remaining 20% as casualties in both cases using varying definitions of loss. There is a
single function that correlated the downtime with a single level of ground motion intensity and for this
reason, it is not included in the compendium.
Vulnerability and fragility functions have been constructed for only a few seismic-‐prone countries as depicted
in Figure 4.2. In particular, Figure 4.2a shows that vulnerability functions have been constructed mainly from
data from earthquakes in the USA or by combining data worldwide. A very different picture is presented in
Figure 4.2b for fragility functions, which have mainly been developed from data obtained from events in
Southern Europe and Japan.
The fragility curves have overwhelmingly been constructed for building classes defined predominantly via
construction material, followed by structural system (which in some cases includes structural irregularities),
and height, in line with the first layer taxonomy used in the GEM project. With regard to construction
material, more than half of the fragility functions included in the compendium correspond to reinforced
concrete (RC) or masonry buildings. By contrast, the distribution of the construction material is
approximately uniform for the vulnerability functions.

Iceland
Greece Greece
Mexico USA Worldwide
Italy
Worldwide
Montenegro Armenia Taiwan
Japan Iran Italy
Armenia
Japan
USA
New Zealand
Turkey
Mexico
Spain
Montenegro
a) Vulnerability functions Europe b) Fragility functions

Figure 4.2. Distribution of reviewed a) vulnerability and b) fragility functions according to the country/countries of data
origin.

46
Undefined Abode
MIxed
Masonry Multiple
Undefined
Multiple Masonry
Wood
Mixed
RC
Steel
Wood
Steel RC
a) Vulnerability functions b) Fragility functions

Figure 4.3. Frequency of construction material used in the reviewed a) vulnerability and b) fragility functions.

Most (over 60%) of the reviewed vulnerability functions have been based on multiple databases as depicted
in Figure 4.4a. By contrast, most empirical fragility functions are based on data from a single earthquake, as
depicted in Figure 4.3b. As mentioned in Section 2.6, single event damage/loss data often covers a small
range of IM levels and typically contains few observations for high levels of loss or damage.

2 2 3

Figure 4.4. Distribution of reviewed a) vulnerability b) fragility functions according to whether they use one or more
database.

The adopted intensity measure type is also a significant factor for the future application of a vulnerability or
fragility relationship (see Section 2.3). Figure 4.4a depicts the distribution of measures of intensity in the
vulnerability studies. It is observed that macroseismic intensity types, (mostly MMI followed by MSK), are
predominant in the reviewed vulnerability studies, with peak ground acceleration (PGA) also being a popular
choice. Earthquake magnitude has been used for the construction of vulnerability functions, which functions
show high levels of uncertainty. A greater variety of intensity measure types have been used in the fragility
literature as illustrated in Figure 4.4b. Macroseismic intensities are still the most frequently observed when
all the literature is considered together, but are seen to appear less frequently in recent fragility functions.
Peak and spectral ordinates of strong ground motion are also popular with PGA and PGV dominating the
recent literature. The use of other measures (e.g. Housner Intensity in Rota et al. 2008b), also appears to be
significant. However, most such functions derive from a single study, King et al, (2000) who constructed
fragility curves for the same database for over 20 intensity measure types.

47
MSK MMSK
M,R MMI
OTHER OTHER
Sa M
PSI
Housner
PGA AI
Sd JMA
MSK MMI Sa EMS

PGD MCS
MMSK
PGV PGA

Figure 4.5. Frequency of ground motion intensity measures used in the reviewed a) vulnerability and b) fragility
functions.
4.3 Harmonisation of Damage and Intensity Measures for Comparison of Existing Functions
Comparison of existing vulnerability or fragility functions may be desirable for validation purposes, sensitivity
checking or to help choose a relationship for use in seismic risk assessment. This is straightforward if two
functions for identical building classes exist with the same IM and loss parameter/damage scale. However,
this is not usual, as multiple taxonomies are seen to exist in the empirical functions even for the same
geographic area. Structures of different heights, seismic codes, materials, and lateral load resisting systems
can be grouped in different ways, and their damage expressed through different damage scales. Casualties,
downtime, and cost can also be defined in different ways. This makes it difficult to compare existing fragility
and vulnerability functions. The same differences can also be seen in the data itself, when collected by
different authorities or for different earthquakes. This hinders data comparison and combination (e.g. for the
development of fragility and vulnerability functions from multiple events). Vulnerability or fragility functions
are also seen to vary in terms of intensity measure (IM) used, which again hinders comparison of existing
functions.
Converting or mapping of damage states and intensity measures will introduce large uncertainties (e.g. see
Hill, 2011) and is not recommended for quantitative comparisons. However, for qualitative comparisons, the
damage state conversion tables proposed in Rossetto et al (2013) and IM conversions suggested by Cua et al
(2010) could be used.

4.4 Overview of Quality Ratings of Existing Functions
In Section 3, a new rating system is proposed for empirical fragility and vulnerability relationships. The
summary results of the application of this rating system to the relationships present in the compendium are
presented in Appendix B. From Table B.1 and Table B.2, it is evident that no fragility or vulnerability
relationship obtains a “High” rating in all criteria. The rating exercise highlights the profusion of data quality
and quantity issues (discussed in Sections 2.2 and 2.4) in existing empirical studies. It is also clear that data
quantity alone cannot be taken as a measure of relationship reliability, as some vulnerability or fragility

48
functions that rate highly for data quantity may not have good ratings in constrained categories and
excitation parameters/observations (e.g. Yang et al. 1992), which counterbalance the advantages posed by
large quantities of data.
It is a real concern that there is a failure in most existing functions to appropriately treat uncertainty. Very
few studies are also seen to cross-‐validate their proposed functions with the results of other studies, and
fewer still compare their functions with independent data. Difficulties were experienced by the authors in
carrying out the rating exercise caused by poor documentation regarding the survey data, almost universal
lack of explanation of the rationale for selection of survey sites, unclear definition of parameters used, and
lack of detail on the exact methodology followed for manipulation, combining data and generating the
relationship. It is recommended that all these observations be taken into account in the development of the
GEM VEM guidelines for empirical vulnerability functions.
5 Final Comments
This report provides an overview of the state of art in empirical fragility and vulnerability relationships, and
presents a database of all existing such functions found by the authors in the literature. A rating system is
also proposed that can be used to qualitatively assess which empirical fragility or vulnerability functions are
most reliable for use in future seismic risk assessments.
Overall, it is noted that a number of authors have developed empirical seismic fragility and vulnerability
functions but that the quality and geographic scope of these functions vary. Existing empirical fragility and
vulnerability functions are seen to typically be based on databases associated with important quality issues,
which include low levels of refinements on the building class, damage states, and often substantial non-‐
sampling errors. Paucity of observations at high shaking intensities and damage states is common and very
few detailed loss databases are available for direct vulnerability evaluation. There is no consensus in the
literature concerning the functional form of empirical vulnerability and fragility functions or on best-‐practice
methodologies for the treatment of the non-‐sampling and sampling errors. Instead, a variety of methods for
manipulating, combining data, and curve-‐fitting procedures has been noted. Finally, existing studies are
severely limited in their modelling and communication of uncertainty in the vulnerability or fragility
assessment.
These observations highlight the need for improved protocols for the collection of loss and damage data in
post-‐earthquake scenarios, in order to provide a sound base for the derivation of future empirical fragility
and vulnerability functions. There is also an urgent need for a rational, statistically correct, widely accepted
approach to be developed for the construction of empirical fragility and vulnerability, which explicitly
quantifies and models the uncertainty in the data and clearly communicates the uncertainty in the
vulnerability or fragility functions. To provide such guidance is the aim of the GEM vulnerability estimation
methods working group.

49
REFERENCES
Access, M. [2010] MS Access.
Akkar S., Bommer J. [2007] "Prediction of elastic displacement responce spectra in Europe and the Middle East",
Earthquake Engineering & Structural Dynamics, Vol. 36, No. 10, pp. 1275-‐1301.
Allen T.I., Wald D.J., Worden C.B. [2011] "Prediction of macroseismic intensities for global active crustal
earthquakes", Journal of Seismology, Vol. 16, No. 3, pp. 409-‐433.
Amiri G.G., Jalalian M., Amrei S.A.R. [2007] ''Derivation of vulnerability functions based on observational data for
Iran'', Proceedings of International Symposium on Innovation & Sustainability of Structures in Civil Engineering,
Tongji University, China.
ATC-‐20 [1989] ''Procedures for post-‐earthquake safety evaluation of buildings'', Report, Applied Technology
Council, 555 Twin Dolphin Drive, Suite 550, Redwood City, California 94065.
Baggio C., B. A., Colozza R., Corazza L., Della Bella M., Di Pasquale G., Dolce M., Goretti A., Martinelli A., Orsini G.,
Papa F., Zuccaro G. [2000] ''Field manual for the 1 level post-‐earthquake damage, usability and emergency
measures form'', Report.
Baker J. [2011] Fitting fragility functions to structural analysis data using maximum likelihood estimation Working
Paper -‐PEER.
Baker J.W., Cornell C.A. [2006] "Which spectral acceleration are you using?", Earthquake Spectra, Vol. 22, No. 2,
pp. 293-‐312.
Benedetti D., Benzoni G., Parisi M.A. [1988] "Seismic vulnerability and risk evaluation for old urban nuclei",
Earthquake Enginnering & Structural Dynamics, Vol. 16, No. 2, pp. 183-‐201.
Bevington J., Adams B., Verrucci E., Rossetto T., Evans R., Amyx P., H. M. [2010] ''The Virtual Disaster Viewer:
Understanding disasters through shared knowledge. The story so far...'', Proceedings of 7th International
Workshop on Remote Sensing for Disaster Management University of Texas at Austin, MCEER.
Booth E., S. K., Spence R., Madhabushi G., Eguchi R.T. [2011] "Validating assessments of seismic damage made
from remote sensing", Earthquake Spectra, Vol. 27, No. S1, pp. S157-‐S177.
Braga F., Dolce M., Liberatore D. [1982] ''Southern Italy November 23, 1980 earthquake: a statistical study of
damaged buildings and an ensuing review of the M.S.K.-‐76 scale'', Report, Rome, Italy.
Clasen C. [2010] '' A comparative study of the 2010 Haiti earthquake damage using a proposed semi-‐automated
methodology and community remote sensing method'', Report, MSc Thesis for MSc Remote Sensing, University
College London.
Coburn A.W., Spence R.J.S. [2002] ''Earthquake protection'', John Wiley and Sons, Chichester, UK.
Cochrane S.W., Schaad W.H. [1992] ''Assessment of earthquake vulnerability of buildings'', Proceedings of 10th
World Conference on Earthquake Engineering, Madrid, Spain.

50
Colombi M., Borzi B., Crowley H., Onida M., Meroni F., Pinho R. [2008] "Deriving vulnerability curves using Italian
earthquake damage data", Bulletin of Earthquake Engineering, Vol. 6, No. 3, pp. 485-‐504.
Comerio M.C. [2006] "Estimating Downtime in Loss Modeling", Earthquake Spectra, Vol. 22, No. 2, pp. 349-‐365.
Crowley H., Pinho R., Bommer J. [2004] "A probabilistic displacement-‐based vulnerability assessment procedure
for earthquake loss estimation ", Bulletin of Earthquake Engineering, Vol. 2, No. 2, pp. 173-‐219.
Cua G., Wald D. J., Allen T. I., Garcia D., Worden C.B, Gerstenberger M., Lin K., Marano K. [2010] '' „Best Practices”
for using macroseismic intensity and ground motion intensity conversion equations for hazard and loss models
in GEM'', Report, GEM Foundation, Pavia, Italy.
D'Ayala D., Meslem A. [2015] ''Guide for selection of existing analytical fragility curves and compilation of the
database'', Report, GEM Foundation, Pavia, Italy.
Dandoulaki M., Panoutsopoulou M., Ioannides K. [1998] ''An overview of post-‐earthquake building inspection
practices in Greece and the introduction of a rapid building usability evaluation procedure after the 1996
Konitsa earthquake'', Proceedings of 11th European Conference on Earthquake Engineering, Paris, France.
De Bruycker M, Greco D, L. M.F. [1985] "The 1980 earthquake in Southern Italy: mortality and morbidity",
International Journal of Epidemiology, Vol. 14, pp. 113-‐117.
Decanini L.D., De Sortis A., Goretti A., Liberatore L., Mollaioli F., Bazzurro P. [2004] "Performance of Reinforced
Concrete Buildings During the 2002 Molise, Italy, Earthquake", Earthquake Spectra, Vol. 20, No. S1, pp. S221-‐
S255.
Di Pasquale G., Orsini G., Romeo R.W. [2005] "New developments in seismic risk assessment in Italy", Bulletin of
Earthquake Engineering, Vol. 3, No. 1, pp. 101-‐128.
Dias W. P. S., Yapa H. D., Peiris L. M. N. [2009] "Tsunami vulnerability functions from field surveys and Monte Carlo
simulation", Civil Engineering and Environmental Systems, Vol. 26, No. 2, pp. 181-‐194.
Dolce M., Kappos A., Masi A., Penelis G., Vona M. [2006] "Vulnerability assessment and earthquake damage
scenarios of the building stock of Potenza (Southern Italy) using Italian and Greek methodologies", Engineering
Structures, Vol. 28, No. 3, pp. 357-‐371.
Dolce M., Masi A., Goretti A. [1998] ''Damage to buildings due to 1997 Umbria-‐Marche earthquake'', Proceedings
of International Workshop on Measures of Seismic Damage to Masonry Buildings, Monselice, Padova, Italy.
Dowrick D.J. [1991] "Damage costs for houses and farms as a function of intensity in the 1987 Edgecumbe
earthquake", Earthquake Engineering & Structural Dynamics, Vol. 20, No. 5, pp. 455-‐469.
Dowrick D.J., Rhoades D.A. [1990] "Damage costs for domestic buildings in the 1987 Edgecumbe earthquake",
Bulletin of the New Zealand National Society for Earthquake Engineering, Vol. 23, No. 2, pp. 137-‐149.
Earthquake Planning and Protection Organisation [1997] ''Guidelines and forms for immediate post-‐earthquake
screening of reinforced concrete buildings (in Greek)'', Report, Athens, Greece.
Eleftheriadou A.K., Karabinis A.I. [2008] ''Damage probability matrices derived from earthquake statistical data'',
Proceedings of 14th World Conference on Earthquake Engineering, Beijing, China.
Eleftheriadou A.K., Karabinis A.I. [2011] "Development of damage probability matrices based on Greek earthquake
damage data", Earthquake Engineering and Engineering Vibration, Vol. 10, No. 1, pp. 129-‐141.

51
Faccioli E., Cauzzi C., Paolucci R., Vanini M., Villani M., D. Finazzi [2007] ''Long period strong ground motion and its
use as input to displacement based design'', Earthquake Geotechnical Engineering: 4th International
Conference on Earthquake Geotechnical Engineering -‐ Invited lectures'', Pililakis K.D., Springer, Netherlands.
FEMA 547 [2007] ''Techniques for the seismic rehabilitation of existing buildings'', Report, Federal Emergency
Management Agency.
FEMA [1999] ''HAZUS: Earthquake loss estimation methodology, Technical manual SR2'', Report, Federal
Emergency Management Agency through aggreements with National Institude of Buildings Science,
Washington D.C., USA.
Ferreira M.A., Oliveira C.S., M. d. S. F. [2011] ''Estimating human losses in earthquake models: A discussion'',
Human Casualties in Earthquakes: Progress in Modelling and Mitigation'', Spence R., S. E.K.M and S. C.,
Springer, London.
Florova N., Larionov V., Bonnin J. [2011] ''Earthquake casualties estimation in emergency mode'', Human
Casualties in Earthquakes: Progress in Modelling and Mitigation'', Spence R., So E.K.M and Scawthorn C.,
Springer, London, UK.
Giovinazzi S., Lagomarsimo S. [2006] "Macroseismic and mechanical models for the vulnerability and damage
assessment of current buildings", Bulletin of Earthquake Engineering, Vol. 4, No. 4, pp. 415-‐443.
Goretti A., Di Pasquale G. [2002] ''An overview of post-‐earthquake damage assessment in Italy'', Proceedings of
EERI invitational workshop: An action plan to develop earthquake dmage and loss data protocols, Pasadena,
California, Earthquake Engineering Research institute.
Goretti A., Di Pasquale G. [2004] "Building inspection and damage data for the 2002 Molise, Italy earthquake",
Earthquake Spectra, Vol. 20, No. S1, pp. 167-‐190.
Grant D.N. [2011] "Response spectral matching of two horizontal ground-‐motion components", Journal of
Structural Engineering, Vol. 137, No. 3, pp. 289-‐297.
Grünthal G. [1998] ''European macroseismic scale 1998'', Report, Vol. 15, Cahiers du Centre Européen de
Géodynamique et de Séismologie, Luxembourg.
Gűlkan P., Sucuoğlu H., Ergűnay O. [1992] ''Earthquke vulnerability, loss and risk assessment in Turkey'',
Proceedings of 1th World Conference on Earthquake Engineering Madrid, Spain.
Hancilar U., Taucer F., Corbane C. [2011] ''Empirical fragility assessment after the January 12, 2010 Haiti
earthquake'', Risk Analysis VIII'', B. C.A., WIT Press, Southampton, UK.
Hill M., Rossetto T. [2008] "Comparison of building damage scales and damage descriptions for use in earthquake
loss modelling in Europe", Bulletin of Earthquake Engineering, Vol. 6, No. 2, pp. 335-‐365.
Hill M.P. [2011]. ''A framework for the seismic risk assessment of existing buildings in Europe'', PhD Thesis,
University College London, London.
Ioannou I., Rossetto T., Grant D.N. [2012] ''Use of regression analysis for the construction of empirical fragility
curves'', Proceedings of 15th World Conference on Earthquake Engineering, Lisbon, Portugal.

52
Jaiswal K., Aspinal W.P., Perkins D, Wald D., Porter K.A. [2012] ''Use of expert judgement to estimate seismic
vulnerability of selected building types'', Proceedings of 15th World Conference on Earthqueak Engineering,
Lisbon, Portugal.
Jaiswal K., Wald D. [2010] "An empirical model for global earthquake fatality estimation", Earthquake Spectra, Vol.
26, No. 4, pp. 1017-‐1037.
Jaiswal K., Wald D. [2013] "Estimating economic losses from earthquakes using an empirical approach",
Earthquake Spectra, Vol. 29, No. 1, pp. 309-‐324.
Jaiswal K.S., Wald D.J., Hearne M. [2009] ''Estimating casualties for large worldwide earthquakes using an empirical
approach'', Report, U.S.G.S. .
Jara M., Guerrero J.J., Aguilar J. [1992] ''Seismic vulnerability of Mexico City buildings'', Proceedings of 10th World
COnference on Earthquake Engineering, Madrid, Spain.
Karababa F.S., Pomonis A. [2010] "Damage data analysis and vulnerability estimation following the August 14,
2003 Lefkada Island, Greece, Earthquake", Bulletin of Earthquake Engineering, Vol. online.
King S., Kiremidjian A., Sarabandi P., Pachakis D. [2005] ''Correlation of observed building performance with
measured ground motion'', Report, Department of Civil and Environmental Engineering, Stanford University
USA.
Koshimura S., O. T., Yanagisawa H., Imamura F. [2009] "Developing fragility functions for tsunami damage
estimation using numerical model and post-‐tusnami data from Banda Aceh, Indonesia", Coastal Engineering
Journal, Vol. 51, No. 3.
Lagorio H.J. [1990] ''Status of medical inputs to vulnerability studies conducted by the U.S. Federal and State
agencies'', Proceedings of International Workshop on Earthquake Injury Epidemiology: Implications for
Mitigations and Response, Reston VA/Baltimore MD, U.S. Geological Survey.
Lallemant D., Kiremidjian A.S. [2012] ''Fitting fragility curves to empirical data'', Report, GEM foundation, Pavia,
Italy.
Levy P.S., Lemeshow S. [2008] ''Sampling of Populations: Methods and Applications, 4th edition'', John Wiley &
Sons, New Jersey, USA.
Li Y, Wu Z., Zhao Y. [2011] "Estimating the number of casualties in earthquakes from early field reports and
improving the estimate with time", Natural Hazards, Vol. 56, No. 3, pp. 699-‐708.
Liel A.B., Lynch K.P. [2009] "Vulnerability of reinforced concrete frame buildings and their occupants in the 2009
L'Aquila", Natural Hazards Review, Vol. In press.
Miyakoshi J., Hayashi Y., Tamura K., Fukuwa N. [1998] ''Damage ratio functions of buildings using damage data of
the 1995 Hyogo-‐Ken Nambu earthquake'', Proceedings of 7th International Conference on Structural Safety
and Reliability, Kyoto, Japan.
Murakami H.O. [1992] ''A simulation model to estimate loss for occupants of collapsed buildings in an
earthquake'', Proceedings of 10th World Conference on Earthquake Engineering, Madrid, Spain.
Nations, U. [1964] ''Recommendations for the Preparation of Sample Survey Reports (Provisional Issue)'', Report,
Statistical Papers, Series C, ST/STAT/SER.C/1/Rev.2, New York, USA.

53
Nichols J.M., Beavers J.E. [2003] "Development and calibration of an earthquake fatality function", Earthquake
Spectra, Vol. 19, No. 3, pp. 605-‐633.
Orsini G. [1999] "A model for buildings' vulnerability assessment using the parameterless scale of seismic intensity
(PSI)", Earthquake Spectra, Vol. 15, No. 3, pp. 463-‐483.
Petal M. [2011] ''Earthquake casualties research and public education'', Human casualties in earthquakes: Progress
in modelling and mitigation'', Spence R., So E. and S. C., Springer, London.
Petrovski J., Milutinovic Z. [1990] ''Development of Vulnerability Functions Based on Empirical Data from
September 19, 1985 Michoacan, Mexico Earthquake'', Proceedings of 9th European Conference on Earthquake
Engineering, Moscow, USSR.
Petrovski J., Stankovic V., Nocevski N., Ristic D. [1984] ''Development of Empirical and Theoretical Vulnerability and
Seismic Risk Models'', Proceedings of 8th World Conference on Earthquake Engineering San Francisco,
California, USA.
Pomonis A., Coburn A.W., et al. [1992] ''Part three: casualty estimation in the collapse of reinforced concrete
buildings, human casualties in building collapse -‐ second year report'', Report, Cambridge, U.K.
Porter K., Jaiswal K., Wald D.J., Greene M., Comartin C. [2008] ''WHE-‐PAGER Project: A new initiative in estimation
global building inventory and its seismic vulnerability'', Proceedings of 14th World Conference on Earthquake
Engineering, Beiijing, China.
Porter K., Kennedy R., B. R.. [2007] "Creating fragility functions for performance-‐based earthquake engineering",
Earthquake Spectra, Vol. 32, No. 2, pp. 471-‐489.
Porter K.A., Farokhnia K., Cho I.H., Rossetto T., Ioannou I., Grant D., Jaiswal K., Wald D., D' Ayala D., Meslem A., So
E., Kiremidjian A.S., Noh H-‐Y. [2012] ''Global vulnerability estimation methods for the global earthquake
model'', Proceedings of 15th World Conference on Earthquake Engineering.
Priestley M.J.N., Calvi G.M., Kowalsky M.J. [2007] ''Displacement-‐based seismic design of structures'', IUSS, Pavia,
Italy.
Roca A., Goula X., Susagna T., Chávez J., González M., Reinoso E. [2006] "A simplified method for vulnerability
assessment of dwelling buildings and estimation of damage scenarios in Catalonia, Spain", Bulletin of
Rossetto T. [2004]. ''Vulnerability curves for the seismic assessment of reinforced concrete structure populations '',
PhD Thesis, Imperial College, London, UK.
Rossetto T., D' Ayala D., Ioannou I., Meslem A [2014] ''Evaluation of Existing Fragility Curves'', SYNER-‐G: Typology
Definition and Fragility Functions for Physical Elements at Seismic Risk'', C. H. Pitilakis K., Kaynia A.M.,, Springer.
Rossetto T., Elnashai A. [2003] "Derivation of vulnerbaility functions for European-‐type RC structures based on
observational data", Engineering Structures, Vol. 25, No. 10, pp. 1241-‐1263.
Rossetto T., Elnashai A. [2005] "A new analytical procedure for the derivation of displacement-‐based vulnerability
curves for populations of RC structures", Engineering Structures, Vol. 27, No. 3, pp. 397-‐409.

54
Rossetto T., Free M., Hill M. [2007] "Report on building damage from the 28th April 2007 Kent (Folkestone):
Earthquake and intensity evaluation", The Society for Earthquake and Civil Engineering Dynamics (SECED)
Newsletter, Vol. 20, No. 3, pp. 6-‐10.
Rossetto T., Ioannou I., Grant D.N., Maqsood T. [2014] ''Guidelines for empirical vulnerability assessment '', Report,
GEM Foundation, Pavia, Italy.
Rossetto T., Novelli V., Porter K. [2015] ''The conversion of damage and loss in GEM-‐VEM indirect vulnerability
curve derivation'', Report, GEM Foundation, Pavia, Italy (under review).
Rossetto T., Peiris N., Alarcon J.E., So E., Sargeant S., Free M., Sword-‐Daniels V., Del |Re D., Libberton C., Verrucci
E., Sammonds P., Faure Walker J. [2010] "Field observations from the Aquila, Italy earthquake of April 6, 2009 ",
Bulletin of Earthquake Engineering, Vol. 9, No. 1, pp. 11-‐37.
Rota M., Penna A., Strobbia C.L. [2008] "Processing Italian damage data to derive typological fragility curves ", Soil
Dynamics and Earthquake Engineering, Vol. 28, No. 10-‐11, pp. 933-‐947.
Rota M., Penna A., Strobbia C.L., Magenes G. [2008b] ''Derivation of empirical fragility curves from Italian damage
data'', IUSS Press, Pavia, Italy.
Sabetta F., Goretti A., Lucantoni A. [1998] ''Empirical fragility curves from damage surveys and estimated strong
ground motion'', Proceedings of 11th European Conference on Earthquake Engineering Paris, France.
Sabetta F., Pugliese A. [1996] "Estimation of response spectra and simulation of nonstationary earthquake ground
motions", Bulletin of the Seismological Society of America, Vol. 86, No. 2, pp. 337-‐352.
Sadigh R.K., Egan J.A. [1998] ''Updated relationships for horizontal peak ground velocity and peak ground
displacement for shallow crustal earthquakes. '', Proceedings of 6th US National Conference on Earthquake
Engineering, Seattle, USA.
Saito K., Spence R. [2004] ''Rapid damage mapping using post-‐earthquake satellite images'', Proceedings of
International Geoscience and Remote Sensing Symposium, IEEE International.
Samardjieva E., Badal J. [2002] "Estimation of the expected number of casualties caused by strong earthquakes",
Bulletin of Seismological Society of America, Vol. 92, No. 6, pp. 2310-‐2322.
Sarabandi P., Pachakis D., King S., Kiremidjian A. [2004] ''Empirical fragility functions from recent earthquakes'',
Proceedings of 13th World Conference on Earthquake Engineering, Vancouver, Canada.
Scawthorn C., Iemura H., Yamada Y. [1981] "Seismic damage estimation for low-‐ and mid-‐rise buildings in Japan",
Earthquake Engineering & Structural Dynamics, Vol. 9, No. 2, pp. 93-‐115.
Scholl R.E. [1974] "Statistical analysis of low-‐rise building damage caused by the San Fernando earthquake",
Bulletin of Seismological Society of America, Vol. 64, No. 1, pp. 1-‐23.
Scholl R.E., Kustu O., Perry C.L., Zanetti J.M. [1982] ''Seismic damage assessment for high-‐rise buildings'', Report,
URS/JAB 8020, URS/John A. Blume & Associates, Engineers, San Francisco, USA.
Şengezer B., Ansal A. [2007] "Probabilistic evaluation of observed earthquake damage data in Turkey", Natural
Hazards, Vol. 40, No. 2, pp. 305-‐326.
Shiono K., Krimgold F., Ohta M. [1991] "Post-‐event rapid estimation of earthquake fatalities for the management
of rescue activity", Comprehensive Urban Studies, Vol. 44, pp. 61-‐104.

55
So E.K.M. [2011] ''Challenges in collating earthquake casualty field data'', Human Casualties in Earthquakes:
Progress in Modelling and Mitigation'', Spence R., So E.K.M. and Scawthorn C., Springer, London, UK.
Spence et al, So E. [2008] ''Estimating shaking-‐induced casualties and building damage for global earthquake
events '', Report, Cambrige Architectural Research Ltd, Cambridge, UK.
Spence R., Bommer J., del |Re D., Bird J., Aydinoğlu N., Tabuschi S. [2003] "Comparing loss estimation with
observed damage: a study of the 1999 Kocaeli earthquake in Turkey", Bulletin of Earthquake Engineering, Vol.
1, No. 1, pp. 183-‐113.
Spence R., So E., Jenkins S., Coburn A., Ruffle S. [2011] ''A global earthquake building damage and casualty
database'', Human CAsualties in Earthquakes: Progress in Modelling and Mitigation'', S. E. K. M. Spence R.,
Scathorn C.,, Springer London, UK.
Spence R.J.S., Coburn A.W., Pomonis A. [1992] ''Correlation of ground motion with building damage: The definition
of a new damage-‐based seismic intensity scale'', Proceedings of 10th World Conference on Earthquake
Engineering, Balkema, Rotterdam.
Spence R.J.S., Coburn A.W., Sakai S., Pomonis A. [1991] ''A parameterless scale of seismic intensity for use in
seismic risk analysis and vulnerability assessment'', Proceedings of Earthquake, Blast and Impact: Measurement
and Effects of Viration, Manchester, UK, Elsevier.
Steinbrugge K.V. [1982] ''Earthquake, Volcanoes and Tsunamis: An anatomy of hazards'', Skandia America Group,
New York, USA.
Tavakoli B., Tavakoli S. [1993] "Estimating the vulnerability and loss functions of residential buildings", Natural
Hazards, Vol. 7, No. 2, pp. 155-‐171.
Thiel C.C., Zsutty T.C. [1987] "Earthquake characteristics and damage statistics", Earthquake Spectra, Vol. 3, No. 4,
pp. 747-‐792.
Thráinsson H. [1992] ''Damage assessment ofor two Icelandic earthquakes'', Proceedings of 10th World
Conference on Earthquake Engineering, Madrid, Spain.
Thráinsson H., Sigbjörnsson R. [1994] ''Seismic vulnerability of Icelandic URC shear wall buildings'', Proceedings of
10th European Conference on Earthquake Engineering, Vienna, Austria.
Trifunac M. D. [1977] ''Forecasting the spectral amplitudes of strong earthquake ground motion'', Proceedings of
Sixth World Conference on Earthquake Engineering, New Delhi, India.
United Nations [1982] ''National household survey capability programme. Non-‐sampling errors in household
survyes: sources, assessment and control'', Report, Department of Technical Co-‐Operation for Development
and Statistical Office, New York, USA.
Wesson R.L., Perkins D.M., Leyendecker E.V., Roth R.J., Petersen M.D. [2004] "Losses to single-‐family housing from
ground motions in the 1994 Northridge, California, earthquake", Earthquake Spectra, Vol. 20, No. 3, pp. 1021-‐
1045.
Whitman R.V., Reed U.W., Hong S.T. [1973] ''Earthquake damage probability matrices'', Proceedings of 5th World
Conference on Earthquake Engineering, Rome, Italy.

56
Yamaguchi N., Yamazaki F. [2000] ''Fragility curves for buildings in Japan based on damage surveys after the 1995
Kobe earthquake '', Proceedings of 12th World Conference on Earthquake Engineering, Auckland, New Zealand.
Yamaguchi N., Yamazaki F. [2001] "Estimation of strong motion distribution in the 1995 Kobe earthquake based on
building damage data", Earthquake Engineering and Structural Dynamics, Vol. 30, No. 6, pp. 787-‐801.
Yamazaki F., Murao O. [2000] ''Vulnerability functions for Japanese buildings based on damage data due to the
1995 Kobe earthquake'', Implication of Recent Earthquakes on Seismic Risk, Series on Innovation and
Construction, Imperial College Press 2, London, UK.
Yang W.S., Slejko D., Viezzoli D. [1989] "Seismic risk in Friuli-‐ Venezia Giulia: an approach", Soil Dynamics and

I
APPENDIX A Tabulated Data for Compendium

Table A.1. Glossary

General category Category Options
Source of Statistics Country/Countries of data origin Standard two letter abbreviations for countries (e.g. AM for Armenia,
GR for Greece, IT for Italy).
Sample size -‐: The total number of buildings of all types in the dataset is not
mentioned.
No of events S: single seismic event.
M: ≥4 events.
M*: Italian events only.
M**: Events affected mainly the USA.
M***: Mainly European events.
Quality of Surveys Source The name of the event on which the dataset is based.
C: Census.
FS: Field Study data.
Completeness: A database is Y: The survey is complete.
complete if all the buildings in an N: Only some buildings from each isoseismic unit have been considered.
isoseismic unit has been surveyed.
N*: A census is used to estimate the number of undamaged buildings.
N**: High levels of completeness have been considered.
-‐: The level of completeness is not available.
Other biases B1: Damaged buildings from aftershocks or other main events are
included.
B2: Burnt buildings have been included in the database.
B3: Buildings damaged by liquefaction have been included in the
database.
B4: Size of building stock estimated from projecting the size of
population.
B5: The Census and field surveys used different structural units.
B6: Errors in compilation of the survey forms.
B7: Data from tsunami included.
B8: Damage was not concentrated in the vertical elements. Undamaged
buildings >10 storeys. The weakest storey according to the method
was not the most
damaged.
B9: The exposed population has been estimated.
B10: Misclassification error due to remote sensing data.
-‐: Biases are not mentioned.
Building classification Structural Unit B: Building.
A: Apartment.
R: Room.
S: Structural elements.

II
Table A.1. Glossary (continued).

General category Category Options
Building classification Material M: Masonry.
M: Masonry
RC: Reinforced Concrete.
URM: unreinforced masonry.
S: steel.
Loss parameter Loss FR: fatality rate.
FC: fatality count.
DF: Damage factor.
Intensity Measure Isoseismic Unit A: Isoseismic Contour (Areas with a given intensity level).
Co: Communities.
D: District.
GBB: groups of building blocks.
IC:
M: Municipality.
S: sites.
Se: Settlement.
T: Town.
-‐: the type of isoseismic unit is not available.
Other (if other ground motion <BLANK>: No additional measures were used.
intensity measures have been AI: Arias Intensity.
adopted)
EPA: effective peak acceleration.
HI: Housner Intenstiy.
k: base shear coefficient.
M: Magnitude of the seismic event (unspecified type).
Ms: Surface magnitude of a seismic event.
R: Source to site distance.
Source of IMs GMPE: ground motion prediction equations.
RGM: recorded ground motion.
TE: Empirical transformation equation from one measure to another.
Analysis CFS: Closed Form of Solution.
LS: Least Squares
LLS: Linear Least Square.
ML: Maximum Likelihood.
NLS: Nonlinear Least Square.
_GN: Gauss-‐Newton Algorithm.
_LM: Levenberg-‐ Marquardt Algorithm.
_NR: Newton-‐Raphson Algorithm.
_W: Weighted Least Squares.
PP: Probability Paper.
<Blank>: No optimisation technique was adopted.
-‐: The type of optimisation technique is not mentioned.
Confidence refer to CI: Confidence intervals.
methods or ways the I: Upper and lower intervals without levels of confidence associated
confidence or the with them.
quality of fit is taken I*: intervals produced with fuzzy logic.
into account in the K-‐S: Kolmogorov-‐Smirvov test.
literature 2
R : coefficient of determination.

III
Table A.2. Notation for the vulnerability or fragility functions

General category Options
Vulnerability or Be: Beta distribution.
fragility Functions Bi: Binomial distribution with parameter p.
D: Discrete values of loss given im.
H: Histogram.
G_PDF: Gamma Probability Density Function: P (L = l ) = 1 ⎛ 1 ⎞

l a −1 exp⎜ − ⎟ , with parameters a, b estimated
b a Γ(a ) ⎝ b ⎠
from Eq.(A.7), Eq.(A.8).
LN_CDF: Lognormal Cumulative Distribution Function, Φ⎛⎜ ln(im ) − λ ⎞⎟ , λ and ζ the log mean and log standard
⎜ ζ ⎟
⎝ ⎠
deviation.
MLP: Multi-‐Linear Piecewise Curve.
ModBe: Modified Beta distribution.
ModBi: Modified Binomial distribution.
N_CDF: Normal Cumulative Distribution Function, Φ⎜

⎛ im − µ ⎞ , μ and σ the mean and standard deviation.
⎟
⎝ σ ⎠
TR_N_PDF: Truncated in [0,1] Normal Probability Density Function with parameters μ, σ estimated from Eq.(A.9),
Eq.(A.10).
Eq.(A.1): P (DS ≥ dsi IM ) = αIM b or P (DS = dsi IM ) = αIM b where α, b parameters.
bIM 2
(
Eq.(A.2): P DS = dsi IM = e αIM e ) where α,b parameters.

2 3
p = α0 + α1IM + α 2IM + α3IM
Eq.(A.3): 0 1 2 3 where α , α , α , α .
Eq.(A.4): P (DS ≥ dsi IM ) = 1 − exp (− αIM b ) where α, b parameters.

Eq.(A.5): ln(L ) = α + βM − ln(R ) + bR where α, β, b parameters.

Eq.(A.6): Log(N)=a+bM
Eq.(A.7): a = α exp⎡− ⎛⎜1 + 1 ⎞⎤ where α, β, δ parameters.

⎢ ⎜ ⎟⎟⎥
⎢⎣ ⎝ βIM δ ⎠⎥⎦
Eq.(A.8): b = α 0 + α1IM + α 2IM 2 where α0, α1, α2 parameters.

Eq.(A.9): ln(µ ) = α + bIM where α, b parameters.
Eq.(A.10): σ = α + b ln(IM ) where α, b parameters.
2
Eq.(A.11): log(fatality counts)=-‐αΜ +βΜ+c

Structural system
Age/Design code
Structural unit
Isoseismic Unit
Country(ies) of
Completeness
No of events
Other biases
Sample size
data origin
Reference
Material
Source
Height
Other
Main
Whitman et al USA -‐ S 1973 San-‐Fernando -‐ -‐ B High-‐rise RC (-‐’33; 47’-‐) A MMI
(1973) Medium-‐rise RC (-‐’47)
High-‐rise S (-‐’33; 47’-‐)
Medium-‐rise S (-‐’47)
Scholl et al Worldwide 5,676 M FS -‐ -‐ B High rise RC A EI MMI
(1982) High-‐rise S
Yang et al IT -‐ S 1979 Friuli,1981C N* -‐ R Rooms built -‐’45 79 M MSK
(1989) ’46-‐‘60
’61-‐‘75
Gűlkan et al TR 30,000 M FS -‐ -‐ B Brick M A MSK
(1992) Wood
RC
Stone M
Adobe
Dolce et al IT 4,745 S 1998 Umbria-‐Marche -‐ -‐ B M 5 M MCS
(1998)
Spence et al TR -‐ M FS -‐ -‐ B URM -‐ EMS-‐

(2003) Low-‐rise RC frame (good, poor design) 98
Mid-‐rise RC frame (good, poor design)
C (MSK)
Goretti and Di IT 23,300 S 2002 Molise N** B1 B A,B,C (MSK) for masonry 13 M MCS
Pasquale RC
(2004) S
Eleftheriadou GR 73,468 S 1999 Athens,2001C N* B6 B Non-‐ductile MRFs, dual frames or mixed 117 M MMI
and masonry and RC load bearing
Karampinis systEMS-‐98 (’59-‐’85;‘85-‐’95;’95)
(2008) M with RC, S , W floors
Braga et al IT 29,157 S 1980 Irpinia Y -‐ B A,B,C (MSK) 41 M MSK
(1982)
Sabetta et al IT 47,677 2 1980 Irpinia -‐ -‐ B A, B, C (MSK) 68 M PGA EPA,
(1998) 1984 Abruzzo AI
Di Pasquale et IT 50,000 M* FS, 1991C -‐ -‐ A A, B, C1, C2 (MSK) (>41)M MCS
al (2005)

Roca et al ES 29,157 S 1980 Irpinia Y -‐ B A, B, C1, C2 (MSK) 41 M EMS-‐
(2006) 98
Thiel and Mainly 10,971 M** FS -‐ -‐ B URM (mainly) A MMI
Zsutty (1987) USA Wood
Adobe
RC
Şengezer and TR 9,366 3 1992 Erpizan -‐ -‐ B RC -‐ MSK
Structural system
Age/Design code
Structural unit
Isoseismic Unit
Country(ies) of
Completeness
No of events
Other biases
Sample size
data origin
Reference
Material
Source
Height
Other
Main

Murakami (1992) AM 2,175 S 1988 Spitak N -‐ A 9 storey precast RC 10T MSK
B 4-‐5 storey composite stone
1-‐5 storey stone URM
1-‐2 storey stone URM
Spence and So Worldwide 53,446 M Wordwide -‐ B A-‐E according to EMS98 5 IC MMI
(2008)
Pomonis et al GR 2,950 S 1986 Kalamata Incomplete B Masonry 7 Iso-‐seis-‐ MMI
(2011) Mixed Masonry mal Areas
Frolova et al (2011) Russian, Uzbekistan, -‐ M -‐ B A,B,C,D,E7-‐9 -‐ MMSK-‐
Turkmenistan 86
Romania, Moldova,
Armenia, Georgia

Table A.4. Main characteristics of existing indirect methodologies developed for the construction of vulnerability functions.
Definition of loss

Damage-‐to-‐Loss functions
Vulnerability
Confidence
Reference
Analysis
Source
curves
Form
Yang et al (1989) CFS BE FD for a room DF (dependent on class) -‐ TR_N_PDF, Eq.(A.9), Eq.(
Spence et al (2003) CFS BE FD+ Insurance Data DF (dependent on class) -‐ D ( expected scenario loss E(L|
sites)
Spence et al (2008) CFS Discrete Martin Centre Death rate, injury rate -‐ Discrete
Frolova et al (2011) CFS Discrete Observed data Fatality rate -‐ Discrete
Eleftheriadou and Karabinis CFS BE EJ DF (independent of class) -‐ MLP (E(L/IM) vs IM)
(2011)

Structural system
Age/Design code
Structural unit
Isoseismic Unit
Country(ies) of
Completeness
No of events
Other biases
Source of IM

Sample size
data origin
Reference
Material
Source
Height
Other
Main
Scholl (1974) USA 1,043 S 1972 San-‐Fernando -‐ -‐ S >80% low rise wood frames >40 years old 2 S Sa RGM 2
Tavakoli and IR 314,000 S 1990 Manjil N B4 B Residential -‐ PGA TE 2

Tavakoli (1993) Hospitals
Education buildings
Scawthorn et al JNP 60,755 S 1975 Miyagikenoki -‐ -‐ B Low-‐rise wooden frames 13 S Sa(T=0.75s) 3
(1981) Mid to High rise RC+SRC (3-‐12 storeys) Sd(T)

Petrovksi et al ME 40,000 S -‐ -‐ -‐ B Strengthened masonry 7Co, Equi-‐valent Recorded 2
(1984) PGA PGA
RC frames 700
Se
Petrovski and MX 2,100 S 1985 Mexico City -‐ -‐ B 5 ranges according to fundamental periods 44 S Sa RGM and 3
MIlutovic Soil
(1990) conditions
Benedetti et al IT 500 S 1979 Friuli -‐ -‐ B Masonry details accounted M MSK Observed 5
(1988) for by the vulnerability index (VI) data

Structural system
Country/Countrie
Age/Design code
Isoseismic Unit
Structural unit
s of data origin
Sample size of
Completeness
Source of IM

No of events
Other biases
Reference
surveyed
buildings
Material
Source
Height
Other
Main
Jara et al (1992) MX 429/200 S 1985 Mexico City FS -‐ B8 B RC up to 13 storeys. 1 Area k Surveyed
subdivided into buildings
2 zones drawings
Spence et al (1992) Worldwide 70,000 M Martin Centre -‐ -‐ B Brick with ringbeam or diaphragm -‐ Ψ MSK Observed
(13) FS Unreinforced brick M data
Non-‐ductile RC frame
Rubble stone M
Adobe
Concrete block M
Dressed stone M
Reinforced unit M
Orsini (1999) IT 53,774 S 1980 Irpinia FS Y B6 A A, B, C1, C2 (MSK) 41 M Ψ -‐ Observed
data
Karababa and Pomonis GR 43,353 S 2003 Leukada N* B1 B Low-‐rise with: 3-‐33 D Ψ -‐ Interpretation
(2010) FS,2001C RC non-‐ductile frame infill (no code; code of URM
1959;code 1984; >’95) damage
Stone M with lime mortar, W floors (nocode,’19-‐ statistics
’45, 46-‐’60, >’60)
Post beam, W floors
European style wood frames, W floors
Non-‐ductile stone of brick M and RC bearing
system, RC or W floors (no code; >’60 )
Miyakoshi et al (1998) JNP -‐ S 1995 Kobe -‐ B2 B Wood frame (-‐’50;’51-‐‘60;’61-‐’70;’71-‐’80;’81-‐‘94) GBB PGV -‐ RGM,
FS1 RC (-‐’71;’72-‐’81;’82-‐‘94) overturned
S (-‐’81;’82-‐‘94) tombstones,
Light Gauge S observed data
Yamaguchi and JPN 96,261 S 1995 Kobe -‐ -‐ B Low-‐rise wood frames GBB PGV Sa,JMA RGM,
Yamazaki FS2 observed data
(2000)
Yamazaki and Murao JPN 30,544 S 1995 Kobe -‐ B3 B Wood frame (-‐’51;’52-‐‘61;’62-‐’71;’72-‐’81;’82-‐‘94) GBB PGV RGM,
(2000) FS3 RC (>’71;’72-‐’81;’82-‐‘94) observed data
S (’71;’72-‐’81;’82-‐‘94)
Light Gauge S (’71;’72-‐’81;’82-‐‘94)
Yamaguchi and JPN 346,078 S 1995 Kobe -‐ -‐ B Low-‐rise residential wood frames GBB PGV Sa,JMA RGM,
Yamazaki FS4 observed data
(2000)
Sabetta et al (1998) IT 47,677 2 1980 Irpinia N** -‐ B A, B, C (MSK) 68 M PGA -‐ GMPE
1984 Abruzzo
Sarabandi et al (2004) USA, 83 2 1994 Northridge -‐ B5 B C1,C2 30 S Sd MMI, Sa, RGM
TWN 1999 Chi-‐Chi (<1000 ft from a PGA, PGV
station) etc
Rossetto and Elnashai Europe* 340,000 M European events -‐ -‐ B RC 99 S Sd Sa,Sdinelastic, RGM, GMPE
(2003) (19) mainly PGA,
Amiri et al (2007) IR 686,548 3 2003 Bam -‐ -‐ B Adobe 13S Sa PGA RGM, GMPE
1997 Ghaen M
1990 Manjil W
RC

Source of IM
No of events
Other biases
Structural sy
Isoseismic U
Structural u
Completene
Sample size
Age/Design
Reference
s of data ori

Country/Co
surveyed
buildings
No of DS
Material
Source
Height
Other
Main

Colombi et al (2008) IT 113,262/ M 1980 Irpinia N* B6 B Low-‐rise M M Sd GMPE 4
96,282 1990 Sicily B5 Mid-‐rise M
1997 Umbria-‐Marche Low-‐rise RC
1998 Umbria Mid-‐rise RC
1998 Pollino
2002 Molise, 1991C
Rota et al (2008bb) IT 163,000/ M 1980 Irpinia N** B6 B Mixed 1–2 M grouped in 10 PGA HI GMPE 5
91,394 1984 Abbruzzo Mixed X3 PGA ranges
1997 Umbria-‐Marche Reinforced concrete—seismic design 1–3
1998 Pollino Reinforced concrete—no seismic design 1–3
2002 Molise Reinforced concrete—seismic design ≥4
Reinforced concrete—no seismic design ≥4
Masonry—irregular layout—flexible floors—with
tie rods or tie beams 1–2
Masonry—irregular layout—flexible floors—w/o
tie rods and tie beams 1–2
Masonry—irregular layout—rigid floors—with tie
rods or tie beams 1–2
Masonry—irregular layout—rigid floors-‐w/o tie
rods and tie beams 1–2
Masonry—irregular layout—flexible floors—with
tie rods or tie beams ≥3
Masonry—irregular layout—flexible floors –w/o
tie rods and tie beams ≥3
Masonry—irregular layout—rigid floors—with tie
rods or tie beams ≥3
Masonry—irregular layout—rigid floors—w/o tie
rods and tie beams ≥3
Masonry—regular layout—flexible floors—with
tie rods or tie beams 1–2
Masonry—regular layout—flexible floors—w/o
tie rods and tie beams 1–2
Masonry—regular layout—rigid floors—with tie
rods or tie beams 1–2
Masonry—regular layout—rigid floors—w/o tie
rods and tie beams 1–2
Masonry—regular layout—flexible floors—with
tie rods or tie beams ≥3
Masonry—regular layout—flexible floors—w/o
tie rods and tie beams ≥3
Masonry—regular layout—rigid floors—with tie
rods or tie beams ≥3
Masonry—regular layout—rigid floors—w/o tie
rods and tie beams ≥3
Steel All
Liel and Lynch (2009) IT 483 S 2009 L’ Aquila N -‐ B RC buildings (mainly frames of various heights) 4 areas of 1 M, PGA 5
each building
had different IM
Spence et al (2011) Wordwide 40,000 M Martin Centre -‐ -‐ B Brick and block masonry without an RC slab 4 isoseismal MMI 6
areas
Hancilar et al. (2011) HT 240,672 2 Remote Sensing, Field B10 B No classification, 15 PGA MMI ShakeMap 6
data 1-‐2 storey simple RC frames, municipalities PGV
3-‐5 storey simple RC frames,
Structural system
Age/Design code
Country/Countri
Structural unit
s of data origin
Isoseismic Unit
Completeness
No of events
Other biases
Sample size
Reference
Material
Source
Height
Other
Main
Scholl (1974) USA 205 S FS -‐ -‐ B >80% low rise wood frames >40 years 2 S Average Sa
old between
T=0.05 and
0.2s
Swathorn (1981) JNP S FS -‐ -‐ B Low-‐rise residential wood frames Sa
Steinbrugge (1982) USA -‐ M Insurance claims -‐ -‐ B Wood frame-‐small, family dwellings 4C MMI Repair cost ove
mostly Wood-‐frame large
All Metal –small
All Metal-‐large
Steel frame-‐earthquake resistive
Steel-‐frame ordinary
Steel frame-‐other
RC-‐earthquake resistive
RC-‐ordinary
RC-‐precast
RC-‐other
Mixed construction-‐earthquake
resistive
Mixed construction-‐ordinary resistive
Mixed construction-‐ordinary non-‐
resistive
Mixed construction-‐hollow masonry,
adobe
Petrovski et al (1984) ME 105 S -‐ -‐ -‐ B Strengthened masonry 7 Co,700 Se equivalent repair and stren
RC frames PGA cost of new co
Dowrick and Rhoades (1990) NZ 3131 S 1987 Edgecumbe Y B Houses 3 isoseismal MMI Indemnity co
units property/In
Thráinsson (1992) IS -‐ 2 1934 and 1976 -‐ -‐ B Low-‐rise wooden frames V MMI
earthquakes Unreinforced concrete shear walls
Thráinsson and Sigbjörnsson IS -‐ 2 1934 and 1976 -‐ -‐ B Unreinforced concrete shear walls V M, R
(1994) earthquakes
Cochrane and Schaad (1992) W -‐ M Swiss-‐Re data -‐ -‐ B Wood frame with lights exterior wall -‐ MMI
finish
Wood frame with brick veneer finish
Steel frame with steel bracing or RC
shear walls or with light cladding
Steel frame w/o steel bracing or shear
walls and with non-‐load bearing walls
of RC, brick, glass etc
RC frame with RC or M shear walls
RC frame w/o shear walls but with load
or non-‐load bearing walls of precast
concrete, brick, glass etc
Precast RC frame with lift slab buildings
with or w/o shear walls
Precast RC tilt up, reinforced masonry
or reinforced hollow clock bearing (or
non bearing with plaster)
Unreinforced hollow block bearing
walls
Unreinforced masonry of block bearing
Source of statistics Quality of Surveys Building classification Intensity Measure
Structural system
Country/Countrie
Age/Design code
Structural unit
s of data origin
Isoseismic Unit
Completeness
No of events
Other biases
Sample size
Reference
Material
Source
Height
Other
Main
Comerio (2006) USA ~ 440,000 S 1994 Northridge -‐ -‐ B Single family buildings, mobile homes Downtime in
400 1989 Loma Pieta , multi-‐family buildings from the time o
completio

Table A.9. Main characteristics of existing direct methodologies developed for the construction of vulnerability functions (fatalities).
Source of statistics Quality of Surveys Building Intensity Measure Loss
Classification
Structural system
Country/Countrie
Age/Design code
Structural unit
s of data origin
Isoseismic Unit
Completeness
No of events
Other biases
Reference
Sample size
Definition
Material
Source
Height
Other
Main
Samardjieva and Badal (2002) Worldwide -‐ M 1900-‐1950 B9 -‐ IC M, MSK Casualty counts
1950-‐1999
Nichols and Beavers (2003) Worldwide M -‐ M Casualty counts
So and Spence (2008) Worldwide -‐ M IC MMI % of injured, died in buildin
Jaiswal et al (2009) Worldwide -‐ M Past surveys from 1973 -‐ IC MMI FR

XI
APPENDIX B Rating of Existing Empirical Vulnerability and Fragility

Functions

Table B.1. Rating of existing direct and indirect vulnerability functions.
Reference Data quality Documentatio
Rationality
n
Documentation
Cross validation
First Principles
Constrained
Observations
Data quantity
Damage/Loss
observations
Excitation
Hindcasting
uncertainty
categories
quality
Treat
Direct Vulnerability Functions for Economic Loss
Scholl (1974) H M H H L L H L M
Scawthorn et al (1981) H M L M L L H L M
Steinbrugge (1982) M M L M L L H L L
Petrovski et al (1984) H L L M L L H L L
Dowrick and Rhoades (1990) H L L H L L H M M
Thráinsson (1992) M H L L L H H L L
Thráinsson and Sigbjörnsson (1994) M H L M L L H L M
Cochrane and Schaad (1992) M M M L L L H L L
Wesson et al (2004) H M M H L L H H H
Vulnerability Functions for Casualty
Murakami et al. (1992) H M L L L L H L M
Samardjieva and Badal (2002) H L L L L L H M M
Nichols and Beavers (2003) H H L H L L H M M
So and Spence (2008) H H L M L L H L M
Jaiswal and Wald (2010) H H L M L L H M M
Frolova et al. (2011) H L L M L L H L M
Vulnerability Functions for Downtime
Comerio (2006)* M L L L L L L L L
Indirect Vulnerability Functions for Economic Loss
Yang et al (1989) H L L L L L H M M
Jara et al. (1992) M L L H L L H L L
Eleftheriadou and Karabinis (2011) H L-‐M L L L H H L H
*It is extremely difficult to rate this function due to the lack of information provided

XII
Table B.2. Rating of existing fragility functions.

Reference Data quality Documentatio
Rationality
n
Documentation
Cross validation
First Principles
Constrained
Data quantity
Observations
observations
Excitation
Hindcasting
uncertainty
categories
quality
Treat
Loss
DPMs
Whitman et al (1973) M L L H L L H L H
Scholl et al (1982) L-‐H L M M L L H L H
Gűlkan et al (1992) M L-‐M L M L L H L M
Spence et al (2003) M H L M M L H L H
Goretti and Di Pasquale (2004) H-‐M M L H L L H H H
Eleftheriadou and Karabinis (2011) H L-‐M L H L M H L H
Braga et al (1982) H M L H L L H H H
Sabetta et al (1998) H M M H L L H L H
Di Pasquale et al (2005) H L-‐M L L L L H L H
Roca et al (2006) H M L H L L H L M
Şengezer and Ansal (2007) H M L M L L M M M
Spence and So (2008) M L L H L L H L H
Murakami (1992) M L L L L L L L M
Fragility curves
Jara et al (1992) M H L H L L H L M
Spence et al (1991) H L L H L L H L H
Tavakoli and Tavakoli (1993) H L L L L L H L L
Orsini (1999) H M L M M L H L H
Karababa and Pomonis (2010) L-‐H H M H L L H L H
Miyakoshi et al (1998) H M-‐H H M L H H L M
Yamaguchi and Yamazaki (2000) H L-‐H L M L L-‐H H L M
Yamazaki and Murao (2000) H M L H L L H L M
Yamaguchi and Yamazaki (2001) H H M H L H H L M
King et al (2005) L-‐M M M H L H H L H
Rossetto and Elnashai (2003) H L M H L L H L H
Amiri et al (2007) H L L H L L L L L
Colombi et al (2008) H M L L L H L L H
Rota et al (2008b) M-‐H H M-‐H H L H H L H
Liel and Lynch (2009) M L L H L L H L H
Frolova et al. (2011) M M L M L L H L M
Pomonis et al. (2011) M M L L L L H L H
Spence et al. (2011) M M L L L L H L M
Ioannou et al (2012) H L L M L L H M H
Alternative fragility functions
Scholl (1974) M M H H L L H L H
Tavakoli and Tavakoli (1993) H L L L L L H L M
Scawthorn et al (1981) L-‐M L-‐M L L L L L-‐H L H
Petrovksi et al (1984) H L L M L L H L L
Petrovski and MIlutovic (1990) H L L M L L H L L
Benedetti et al (1988) H L H H L H H M H

The Global
Earthquake Model
The mission of the Global Earthquake Model
(GEM) collaborative effort is to increase
earthquake resilience worldwide.
To deliver on its mission and increase public

understanding and awareness of seismic
risk, the GEM Foundation, a non-profit public-
private partnership, drives the GEM effort by
involving and engaging with a very diverse
community to:
- Share data, models, and knowledge

through the OpenQuake platform
- Apply GEM tools and software to inform

decision-making for risk mitigation and
management
- Expand the science and understanding of

earthquakes.
GEM Foundation
Via Ferrata 1
27100 Pavia, Italy
Phone: +39 0382 5169865
Fax: +39 0382 529131
[email protected]
www.globalquakemodel.org
Copyright © 2015 GEM Foundation, T. Rossetto, I.

Ioannou, D. N. Grant.
This work is made available under the terms of the

Creative Commons license CC BY 3.0 Unported
april 2015

VULN MOD Fragility Compendium 201501 v03 1

Uploaded by

Copyright:

Available Formats

VULN MOD Fragility Compendium 201501 v03 1

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

VULN MOD Fragility Compendium 201501 v03 1

Uploaded by

Copyright:

Available Formats

Existing Empirical

Fragility and Vulnerability

T. Rossetto, I. Ioannou, D. N. Grant

GEM Technical Report

TABLE OF CONTENTS

LIST OF FIGURES

LIST OF TABLES

Table A.1. Glossary ............................................................................................................................................... I

2 Existing Empirical Vulnerability and Fragility Functions

2.1 Form of Functions

0.2 0.5 P(DS=ds2 |IM)

2.2 Quality of Loss or Damage Database

Reconnaissance Small unrepresentative

2.2.2 Economic Loss from Direct Damage

2.2.5 Building Classification

2.2.6 Non-­‐Sampling Errors

2.3 Quality of Ground Motion Intensity

2.3.1 Isoseismic Unit

2.3.2 Measures of Ground Motion Intensity

2.3.3 Evaluation of Ground Motion Intensity

2.4 Vulnerability or Fragility Assessment Procedures

2.4.1 Sample Size

2.4.2 Data Manipulation and Combination

2.4.3 Shape of Vulnerability or Fragility Functions

2.4.4 Statistical Modelling Techniques

2.5 Validation Procedures

3 Rating System for Empirical Fragility and Vulnerability Functions

Attribute Criterion Rate Assigned when:

Attribute Criterion R Assigned when:

Attribute Criterion Rate Assigned when:

4 Compendium of Existing Vulnerability or Fragility Functions

4.1 Basic Information

General category Field Description Propeties Example

ds1 ds3 11.5 2.5

a) Vulnerability functions b) Fragility functions

a) Vulnerability functions b) Fragility functions

MSK MMI Sa EMS

a) Vulnerability functions b) Fragility functions

4.4 Overview of Quality Ratings of Existing Functions

APPENDIX A Tabulated Data for Compendium

Table A.1. Glossary

Table A.1. Glossary (continued).

Table A.2. Notation for the vulnerability or fragility functions

G_PDF: Gamma Probability Density Function: P (L = l ) = 1 ⎛ 1 ⎞

N_CDF: Normal Cumulative Distribution Function, Φ⎜

Eq.(A.7): a = α exp⎡− ⎛⎜1 + 1 ⎞⎤ where α, β, δ parameters.

Spence et al TR -­‐ M FS -­‐ -­‐ B URM -­‐ EMS-­‐

Definition of loss

Source of IM

Tavakoli and IR 314,000 S 1990 Manjil N B4 B Residential -­‐ PGA TE 2

Sample size of

Source of IM

s of data ori

APPENDIX B Rating of Existing Empirical Vulnerability and Fragility

Table B.2. Rating of existing fragility functions.

To deliver on its mission and increase public

- Share data, models, and knowledge

- Apply GEM tools and software to inform

- Expand the science and understanding of

Copyright © 2015 GEM Foundation, T. Rossetto, I.

2.2.6 Non-‐Sampling Errors

Spence et al TR -‐ M FS -‐ -‐ B URM -‐ EMS-‐

Tavakoli and IR 314,000 S 1990 Manjil N B4 B Residential -‐ PGA TE 2