EARTHQUAKE ENGINEERING & STRUCTURAL DYNAMICS

Earthquake Engng Struct. Dyn. 2017; 46:2317–2335

Published online 20 April 2017 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/eqe.2906

Empirical seismic fragility for the precast RC industrial buildings

damaged by the 2012 Emilia (Italy) earthquakes

Nicola Buratti1 , Fabio Minghini2,*,† , Elena Ongaretto1, Marco Savoia1 and

Nerio Tullini2
1
Department of Civil, Chemical, Environmental, and Material Engineering (DICAM), University of Bologna, Viale
Risorgimento 2, Bologna, Italy
2
Department of Engineering, University of Ferrara, Via Saragat 1, Ferrara, Italy

SUMMARY
The paper analyses the seismic fragility of precast reinforced concrete buildings using observational damage
data gathered after the 2012 Emilia earthquakes that struck Northern Italy. The damage level in 1890
buildings was collected, classified and examined. Damage matrices were then evaluated, and finally,
empirical fragility curves were fitted using Bayesian regression. Building damage was classified using a
six-level scale derived from EMS-98. The completeness of the database and the spatial distribution of the
buildings investigated were analysed using cadastral data as a reference. The intensity of the ground motion
was quantified by the maximum horizontal peak ground acceleration, which was obtained from ShakeMaps.
Copyright © 2017 John Wiley & Sons, Ltd.

Received 13 July 2015; Revised 26 October 2016; Accepted 13 March 2017

KEY WORDS: precast RC buildings; seismic damage; seismic fragility; seismic risk; Emilia earthquakes

1. INTRODUCTION

In the present economy, seismic loss estimation is extremely important for planning civil protection
strategies and for predicting costs for restoring or retrofitting damaged buildings after earthquakes.
Fragility curves are a fundamental tool for seismic risk assessment. These curves relate the probability
of exceeding a particular damage level to ground motion intensity [1]. They can be obtained using
different approaches, mainly statistical analysis of observational damage data or numerical modelling.
Observational damage data from past earthquakes are commonly used worldwide for the
development of new empirical fragility curves or for validating existing ones based on mechanical
models. D’Ayala et al. [2] used damage data from the 1755 Lisbon (Portugal) earthquake for
estimating fragility functions suitable for Europe’s historic city centres. Yamaguchi and Yamazaki
[3] developed fragility functions for five different building typologies in Japan using damage data from
the 1995 Kobe earthquake. Rossetto and Elnashai [1] derived empirical vulnerability curves for
reinforced concrete (RC) buildings from 99 datasets collecting field observations from 19 earthquakes
and including about 340 000 buildings. Karababa and Pomonis [4] obtained a set of vulnerability
curves for five building typologies in Lefkada Island, Greece. In their proposal, the authors used
damage data collected after the earthquake that occurred in the island on 14 August 2003 and related
the vulnerability of the buildings to the parameterless scale of seismic intensity [5, 6]. Molina et al. [7]
recently used a damage database concerning about 67 500 buildings struck by the 2010 Haiti
earthquake for calibrating vulnerability curves for the city of Port-au-Prince. For the same destructive

*Correspondence to: Fabio Minghini, Department of Engineering, University of Ferrara, Via Saragat 1, Ferrara, Italy.
†
E-mail: [email protected]

Copyright © 2017 John Wiley & Sons, Ltd.

 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2318 N. BURATTI ET AL.

event, new fragility functions based on two separate methods of damage assessment, including field
surveys and remote sensing, were obtained in [8]. With reference to Italian buildings, mainly
comprising low-rise to mid-rise masonry and RC structures, Rota et al. [9] proposed typological
fragility curves based on earthquake damage data collected in the past 30 years. In [10], the possible
sources of uncertainty that can affect empirical vulnerability curves were identified, such as the errors
in ground shaking prediction and building exposure, the use of census data to establish the number of
buildings in each municipality for each building typology, the incompleteness and deficiencies in
survey forms and the errors in data post-processing. For a recent comprehensive review of the existing
empirical fragility functions, see [11].
When numerical procedures are used to evaluate the seismic fragility of structures, damage is
generally estimated using the results obtained from numerical models. In nonlinear analysis
methods, the attainment of a particular damage level, corresponding to a given limit state, can be
defined in terms of material strains [12, 13], interstorey drifts or chord rotations [14, 15] and other
engineering demand parameters [16]. Although the analysis method influences the vulnerability
assessment of structures, Silva et al. [17] showed that, for a given structural typology, fragility
curves are not particularly sensitive to the type of numerical analysis adopted. Of course, in order to
be able to make reliable predictions, numerical models must be able to represent all the possible
damage mechanisms that can affect the category of buildings under consideration, as well as
possible inhomogeneities and irregularities. This aspect is crucial, for instance, for masonry
structures, where both local and global collapse mechanisms must be captured by models, but also
for precast RC industrial buildings. In fact, the dynamics of these structures can be strongly affected
by infill walls or intermediate floors covering only a limited portion of the plan. These elements,
because of limited structural redundancy, can facilitate the onset of failure mechanisms. For these
reasons, observational models based on field results are very important to assess and calibrate
numerical prediction models.
The present paper focuses on the definition of observational fragility curves for RC precast buildings
using damage data collected after the Emilia seismic sequence that struck the north of Italy in 2012. In
this sequence, two main shocks can be identified: the 20 May earthquake, with Mw = 6.1, and the 29
May earthquake, with Mw = 6.0.
The region struck by these earthquakes is one of the most productive areas in Italy and is
characterized by medium-to-small industrial zones, located in various municipalities. The number of
industrial buildings located in the Emilia-Romagna region is almost 80 000, corresponding to
approximately 12% of the industrial buildings in Italy [18, 19]. Most of the precast RC buildings in
Emilia have a single-storey structure, typically made of a series of simple portal frames. Each frame
comprises columns clamped on either cast-in-place or precast pocket foundations, main beams simply
supported on the column tops and precast slab elements, also simply supported on the main girders.
Some buildings might have two floors or intermediate floors in specific areas, typically along one of
the two short edges, where offices are usually located.
Modern seismic design codes require adequate structural ductility and compliance with capacity
design, with particular regard to connections between precast members [20–23]. However, up until
2005, the region struck by the Emilia earthquakes was not considered as a seismic area by building
codes. Therefore, beam–column and slab–beam supports were typically friction based, without
mechanical connectors to prevent large relative displacements or loss of support failures during
seismic ground motions.
After the 2012 seismic events, the vulnerability of industrial precast buildings drew much attention
from the research community. In [24], a review of the precast structural typologies and construction
practice in Northern Italy was presented, and for industrial buildings located in Tuscany and in Emilia
(with some of them struck by the 2012 earthquake sequence), the probabilistic distributions of some
geometrical characteristics, such as column height, span length of main girders and roof slab elements,
were obtained. The total number of precast RC buildings included in that study is 670, with 40 located
in Emilia. The authors reported that more than 85% of the 40 buildings struck by the Emilia earthquake
were built without suitable seismic design rules. Finally, by means of a series of finite element
nonlinear static and dynamic analyses, the authors evaluated the seismic fragility of the building stock
in terms of spectral acceleration. The methodology has recently been extended by Casotto et al. [25],

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2319

who defined an automated procedure for generating series of precast RC structures compatible with
the properties of the building stock through a probabilistic characterization of material and
geometrical building properties. The damage levels, related to loss of support, column bar first
yielding and flexural collapse limit states, were estimated through both 2D and 3D nonlinear finite
element models.
The present paper presents the damage data collected by the authors, using both field surveys and
technical reports prepared for obtaining public funds for reconstruction. The distribution of the
buildings for which damage data were collected is analysed using cadastral data as a reference.
Damage data are then used to define damage matrices from which empirical fragilities are estimated.
Finally, parametric fragility curves for the different damage states under consideration are fitted
using a Bayesian approach. Two main categories of models are adopted: fragility models based on
the exceedance of individual damage states and ordinal models, which maintain ordinality among
the fragility of damage states. The uncertainty on the ground motion estimate is discussed and
included in the ordinal model. When compared with fragility curves for RC buildings reported in
the literature, the results presented here show that Italian precast structures for industrial buildings
are characterized by much higher seismic vulnerability than cast-in-place RC frame structures and
therefore require specific fragility models.

2. DAMAGE DATABASE FOR INDUSTRIAL BUILDINGS AFTER THE EMILIA

EARTHQUAKES

Damage data were collected for a vast region around the epicentres of the earthquakes that struck Emilia
in 2012. The region of interest, considered in the present paper, was identified according to the definition
introduced by the Legislative Decree no. 74/2012 [26]. The boundaries of this area were defined as the
envelope of circles with a radius of 10 km centred on the epicentres of all the earthquakes with ML ≥ 3.5
[27]. The resulting area is approximately 100 km long (east–west) and 40 km wide (north–south) and
includes 52 municipalities, 35 of which are in the Emilia-Romagna region.

2.1. Typological features of the precast industrial buildings in the area of interest
A review of the main typologies of precast structures used in Italy since the 70s is presented in [28, 29].
According to [19], approximately 70% of the industrial buildings in the Emilia-Romagna region are
single-storey precast RC structures. For the industrial buildings struck by the 2012 Emilia earthquakes,
two main categories of precast RC structures were identified in [30]: (i) buildings constructed from
1970 to 1990 (type 1), featuring double slope beams with spans from 12 to 20 m, roof slabs spanning
from 6 to 10 m and masonry infills, and (ii) buildings constructed after 1990 (type 2), with significantly
longer spans of beams and roofing elements, and either horizontal or vertical prefabricated RC cladding
panels. These two building types approximately correspond to those subsequently adopted by Casotto
et al. [25]. Furthermore, in [25], the two building categories, also named type 1 and type 2, were
characterized by means of probability distributions for the main geometrical parameters (Table I).
The construction date may represent an important factor for the analysis of the seismic behaviour of
the precast buildings struck by the Emilia earthquakes, because of the changes in construction practice
and technology occurred over time. However, most of the territory struck by the earthquakes was not
considered a seismic area by design codes until October 2005. As a consequence, most of the partial
and full collapses were caused by the usage, both in type 1 and type 2 buildings, of friction-based
slab–beam and beam–column connections.

2.2. Definition of the building stock

In the present paper, the total number of industrial buildings located within the area of interest was
estimated using cadastral data. In Italy, the cadastre has the role of public registry of real estates and
land properties and is established mainly for fiscal purposes.
The elementary urban real estate unit is defined as the smallest real estate asset with functional
autonomy and ability to produce income. The Italian cadastre is divided into categories related to

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2320 N. BURATTI ET AL.

Table I. Geometrical dimensions used by Casotto et al. [25] for the generation of their industrial building
stock.
Building configuration Geometrical parameter Median value [m] Log. SD [
] Min [m] Max [m]
a
Type 1 Beam span 14.9 0.3 8 30
Frame spacing 6.8 0.28 8 10
Column height 6.5 0.25 4 12
SD [m]
Type 2b Beam span 8.7 2.1 8 10
Frame spacing 16.5 3.7 10 25
Column height 6.5 1.3 4 11
Log. SD, logarithmic standard deviation.
a
Lognormal distribution.
b
Normal distribution.

the activities undergoing in real estate units. In the present study, the building stock was defined
with reference to two specific cadastral categories, labelled D/1 and D/7, corresponding to
‘factories’ and ‘buildings hosting a specific industrial activity’, respectively. Because in some
cases one building can be constituted by more than one real estate unit, the actual number of
industrial buildings forming the reference population does not correspond to the number of
cadastral units included into categories D/1 and D/7. A detailed analysis was then performed, and
the number of actual (independent) buildings, evaluated using aerial photography and some field
surveys, was compared with the number of cadastral units for 18 representative municipalities
selected among the total of 35 in the area of interest. This analysis showed that the ratio of actual
buildings over cadastral units is, on average, about 0.52. The total size of the building stock
(number of independent industrial buildings) for the 17 municipalities not analysed in detail was
then estimated by multiplying the total number of real estate units obtained from the cadastral
register [18] by 0.52. It is worth noticing that the procedure adopted possibly overestimated the
actual number of prefabricated industrial RC buildings because the cadastre might classify in
categories D/1 and D/7 also cast-in-place concrete and masonry structures that might not have
been identified from aerial photography and field surveys.
The distribution of the building stock within the region of interest, estimated as described above, is
reported in the map of Figure 1, where green circles are located on the administrative centres of the
municipal territories. The maximum concentration of industrial buildings is observed in the Carpi
district.

Figure 1. Map of the territory hit by 2012 Emilia earthquake reporting the distribution of the industrial
building stock (green circle data points) and the epicentres of the two mainshocks (red stars). [Colour
figure can be viewed at wileyonlinelibrary.com]

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2321

2.3. The damage database

The peculiar seismic behaviour of precast RC buildings, and in particular the vulnerability of the
connections between structural elements, made the standard survey forms used in Italy for post-earthquake
damage and safety assessment [31] not suitable to collect data for this structural typology. In the present
paper, given the absence of information from a specific fast survey procedure, damage data were
collected from reports prepared by structural engineers, obtaining more detailed and accurate damage
estimates than from fast surveys. These reports were prepared by professional engineers, representing
building owners, as partial requirement for obtaining regional funds for either reconstruction or retrofit,
in accordance with Regional Decree 57/2012 [32]. These reports were validated by a public in-house
company in charge of assessing the coherence of public funding with the interventions planned in
design. It is worth noticing that funding was available also for retrofitting non-damaged buildings and
that retrofitting of industrial buildings with structural deficiencies (e.g. lack of mechanical connections
between elements) was mandatory. Nonetheless, some building owners decided to not apply for funding
as per Regional Decree 57/2012 for three main reasons: (i) buildings were not occupied; (ii) they
preferred to apply for national funds for building refurbishment, which were convenient in some cases;
and (iii) they had private insurances covering seismic damage.
Accurate descriptions of the typical damage occurred to industrial buildings, and its possible causes
are discussed in [30, 33–35]. In the present study, the damage data were classified according to the six-
level damage scale reported in Table II [36], adapted from the provisions of Regional Decree 57/2012
[32], and substantially coincident with the European Macroseismic Scale [37]. Examples of recurrent
modes of damage, observed in precast buildings struck by the Emilia seismic sequence, are reported in
Figure 2. In particular, the formation of plastic hinges at the base of columns (Figure 2(a)) may be
related to damage levels D4 or D5 depending on the number of damaged columns. The loss of
support of roof slabs (Figure 2(b)) and collapse of RC cladding panels (Figure 2(c)) may correspond
to damage levels D3, D4 or D5 depending on the extent of the collapsed surfaces (Table II), that is,
the number of cladding panels involved.
The total number of precast buildings included into the damage database gathered for the present
study is 1890. The number of buildings belonging to each damage level is reported in Table III. The
normalized cumulative number of buildings investigated, defined as the cumulative number of
buildings divided by the total number of buildings (i.e. 1890), is plotted, in Figure 3 (curve labelled
D ≥ D0), against the epicentral distance, defined as the distance of each building from the nearest

Table II. Damage levels considered in the study.

Type of damage
Damage grade Definition
Non-structural Structural

D0 (no damage) – – No damage

D1 (slight damage) Slight – Local or distributed damage to up to 20% of
horizontal and/or vertical partitions without
collapses
D2 (moderate damage) Moderate Slight Local or distributed damage to more than 20%
of horizontal and/or vertical partitions
without collapses
D3 (severe damage) Heavy Moderate Severe damage to up to 15% of horizontal and/or
vertical covering surfaces with collapses or at
least
one column with permanent drift larger than 2%
D4 (heavy damage) Very heavy Heavy Severe damage to up to 30% of horizontal and/or
vertical covering surfaces with collapses or to up
to 20% of columns, with permanent drift larger
than 2%, or plastic hinge at the base section
D5 (collapse) Total or partial collapse Collapses of more than 30% of horizontal
and/or vertical covering surfaces or damage to
more than 20% of columns

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2322 N. BURATTI ET AL.

Figure 2. Typical damage observed in precast buildings struck by the Emilia earthquakes: (a) plastic hinge at
the base of a column, (b) loss of support of a simply supported roof slab from main girders and (c) out-of-
plane overturning of RC cladding panels. [Colour figure can be viewed at wileyonlinelibrary.com]

Table III. Number of buildings analysed for each damage level.

Damage level D0 D1 D2 D3 D4 D5 D1 + … + D5
No. of buildings in the database 967 371 174 105 76 197 1890

Figure 3. Normalized cumulative number of industrial buildings versus distance from the nearest epicentre.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2323

epicentre, between those of the two mainshocks. For 1341 buildings, this distance is associated with
the second mainshock (29 May).
In order to check the level of completeness of the data collected, we also report the normalized
cumulated number of buildings, estimated from cadastral data as described in Section 2.2, in the
same figure (curve labelled ‘cadastral data’). This curve is normalized to the total number of
buildings estimated from cadastral data. The positions of these buildings were defined based on the
main industrial areas identified by aerial photography: in particular, all the buildings in each
industrial area were assumed at its centre. The shapes of the curves corresponding to the database
and to the cadastral estimate are in good agreement. The sudden increase in the building density
between distances of 16 and 20 km corresponds to a series of large industrial zones in Carpi, in the
Modena district. That area is peculiar in the region. In fact, it contains mainly large textile
manufactories, a production sector that, in the Emilia region, was severely affected by an economic
crisis started in 2009. For this reason, many buildings in the Carpi area were not in use at the time
of the earthquakes and their owners did not submit reports to the authority to obtain funds;
therefore, their damage was not classified. The number of buildings in the database represents about
30% of the building stock estimated from cadastral units. It is worth noticing that the building
distribution in the area is not uniform; otherwise, the curve representing the cumulated distribution
of buildings would be quadratic in terms of epicentral distance [36]. Finally, Figure 3 also shows the
normalized cumulative number of buildings against the distance to the nearest epicentre for D ≥ D1
(i.e. the total number of damaged buildings in the database) and D ≥ D3 (the number of buildings
with severe damages up to partial or total collapse). Note that most of the buildings with D ≥ D3 are
located within 15–20 km from the nearest epicentre. For epicentral distances shorter than 10 km, a
clear predominance of damaged buildings is observed.

3. DAMAGE DISTRIBUTION VERSUS GROUND MOTION INTENSITY

Over the largest part of the territory, the maximum ground motion intensity was recorded during the
two mainshocks, occurred on 20 May (Mw = 6.1) and 29 May (Mw = 6.0) 2012 [38]. Therefore, the
measures of ground motion intensity adopted in the present study refer to these two seismic events only.

3.1. Definition of ground motion intensity

The ground motion intensity at the different building locations was obtained from the official
ShakeMap data published online by Istituto Nazionale di Geofisica e Vulcanologia [39]. These data
provide information on the intensity of ground shaking in terms of Peak Ground Acceleration
(PGA), Peak Ground Velocity (PGV) or pseudo-spectral accelerations at 0.3, 1.0 and 3.0 s,
combining actual ground motion recordings and predictions from attenuation relationships. These
ShakeMaps are computed assuming that the ground motion intensity at each location is lognormally
distributed. For the fragility assessment presented in the following, after analysing the ground
motion accelerograms from the recording stations and the site-to-site variability of different possible
ground motion intensity measures, the maximum horizontal PGA was chosen as measure of ground
motion intensity. In fact, spectral accelerations at different natural periods were characterized by a
very large variability.
Figure 4(a) and (b) show the ShakeMaps for the median value of the horizontal PGA referred to the
20 May and 29 May earthquakes, respectively. Figure 4(c) and (d) shows the maps of the standard
deviation of the natural logarithm of PGA for the two events. The uncertainty on the ShakeMaps for
the two earthquakes is very different. In fact, many temporary ground motion recording stations
were installed after the first mainshock, and therefore, the ShakeMaps for the subsequent shocks
(and in particular for the earthquake occurred on 29 May 2012) are more accurate. The logarithmic
standard deviation of PGA will be used in the derivation of the fragility curves presented in
Section 5.4. For a discussion on the level of approximation of the official ShakeMaps for the Emilia
earthquakes, see [40, 41].

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2324 N. BURATTI ET AL.

Figure 4. ShakeMaps (a, b) for the median value of the horizontal PGA and (c, d) for its logarithmic standard
deviation (SD), for (a, c) the 20 May and (b, d) 29 May earthquakes. Grey dots and black squares indicate
the locations of the buildings in database associated with the PGA of 20 May and 29 May, respectively.
[Colour figure can be viewed at wileyonlinelibrary.com]
Because the strong-motion parameters provided by Istituto Nazionale di Geofisica e Vulcanologia
are referred to a dense spatial grid with nodes every 1 km, according to the rule suggested by the
Italian Building Code [42], the PGA at each building location was computed as the weighted mean
value of the PGAs at the four closest grid nodes, with the ith weight being the reciprocal of the
distance between the location and the ith node.
For each building, the value of the ground motion intensity considered was the maximum between
those related to the two mainshocks of 20 and 29 May. Grey dots and black squares in Figure 4 indicate
the locations of the buildings in the database associated with the 20 May and 29 May PGA,
respectively. Note that for most of the buildings associated with the 20 May ground motion, the
logarithmic standard deviation (Figure 4(c) and (d)) is as large as 0.6, which corresponds to the total
standard deviation of the attenuation relationships used to compute the ShakeMaps.
Considering, as ground motion intensity, the maximum PGA generated by the two mainshocks
corresponds to assuming that the damage produced by the two seismic events was not correlated.
This assumption is supported by the outcomes of field surveys [30, 36]. In fact, the prefabricated
RC structures in the area were typically characterized by extremely fragile failure modes because
they did not have any structural redundancy and featured friction-based connections between
elements (Section 2.1). Even when mechanical connectors were present, their strength was
insufficient because they were not designed for seismic actions, but only to facilitate the assembly of
prefabricated structural members during construction [30]. Therefore, these structures had no
redistribution capacity. This behaviour was highlighted by some field surveys carried out after both
the mainshocks, especially in the Mirandola area. This industrial area is particularly interesting
because, during the two events, it experienced similar PGA values (recorded by a permanent
accelerometric station), that is, 258 and 288 cm/s2, respectively. In spite of the similar PGA values,
there were many cases of buildings totally undamaged after 20 May that collapsed on 29 May
because of failures in friction-based connections; the most common of which were either the
unseating of beams from columns or the roofing of elements from beams [30]. Of course, near-
source effects might have contributed to the collapses occurred on 29 May [30].

3.2. Damage distribution versus PGA

The cumulative number of buildings with damage level D greater than or equal to D0, D1 or D3 is
reported in Figure 5, together with the cumulative number of buildings estimated from cadastral data
(Sections 2.2 and 2.3), versus the maximum horizontal PGA.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2325

Figure 5. Cumulative number of industrial buildings, in the survey area, that experienced a peak ground
acceleration less or equal to the PGA values reported in abscissa: building stock from cadastral data (right
axis), all buildings in the database (D ≥ D0) and buildings with damage level D ≥ D1 and D ≥ D3.
[Colour figure can be viewed at wileyonlinelibrary.com]

For PGA ≤0.28 g, there are 1267, 385 and 72 buildings with damage D ≥ D0, D ≥ D1 and
D ≥ D3, respectively. For high accelerations, the curves feature a sudden step. Because this
peculiar shape can be observed also for D ≥ D0 (i.e. the whole database) and for the building
stock estimated from cadastral units, the step must be a consequence of the non-uniform
distribution of buildings (Section 2.3) and of the spatial distribution of PGA. In fact, the curve
derived from cadastral data is completely independent from damage and unaffected by the
procedure used to collect data. On the other hand, it is worth noticing that the relative increment
in the number of damaged buildings before and after the step is much higher for D ≥ D3
(+323%) than for D ≥ D1 (+122%) and D ≥ D0 (+45%). Therefore, these different percentages
might suggest that a portion of the sudden increment in the number of buildings with at least
severe damage could be related to the activation of specific damage modes. Finally, the similitude
of the shapes of the curves for D ≥ D0 and for the building stock estimated from cadastral data is
an indicator of the soundness of the data collection procedure.

4. DAMAGE ANALYSIS AND FRAGILITY

4.1. Damage data

In order to analyse the fragility of the buildings in the database, the 1890 damage data were categorized
into the damage matrix reported in Table IV [43], considering seven intervals for the PGA (column
IPGA,i). The criteria used to define these intervals will be discussed at the end of Section 4.2 being
related to fragility estimation. Table IV provides, for each PGA interval, the number of buildings
associated with each damage level. For example, among the 257 buildings that experienced a PGA
between 0.297 and 0.313 g, 35 were undamaged (D = D0) and 44, 52, 35, 25 and 35 buildings were
classified in damage levels D1, D2, D3, D4 and D5, respectively.
Moreover, from Table IV, the cumulative damage matrix reported in Table V was obtained. This
table shows the number of buildings that were exposed to a PGA belonging to the interval indicated
in the first column and that were associated with a damage level greater than or equal to Dj. For
instance, 126 of the 257 buildings that experienced a maximum horizontal PGA between 0.297 and
0.313 g had damage levels greater than D3. Of course, the column D ≥ D0 indicates the total
number of buildings for each PGA interval.

4.2. Point estimates of fragility

Using the cumulative damage matrix, we can obtain a first estimate of the fragility of the buildings. In
fact, the probability of observing ni,j buildings with damage D ≥ Dj in the ith ground motion intensity
interval IPGA,i can be represented by the following binomial distribution [44]:

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2326 N. BURATTI ET AL.

Table IV. Damage matrix for all buildings collected in the database: number of buildings for each damage
level versus intervals of PGA.
IPGA,i [g] D = D0 D = D1 D = D2 D = D3 D = D4 D = D5

[0.000–0.076[ 249 21 1 0 1 0
[0.076–0.112[ 205 57 6 0 0 0
[0.112–0.159[ 158 80 10 4 8 10
[0.159–0.216[ 153 74 18 7 3 15
[0.216–0.297[ 133 51 27 19 14 26
[0.297–0.313[ 35 44 52 35 25 66
[0.313–0.349] 34 44 60 40 25 80
Total 967 371 174 105 76 197

Table V. Cumulative damage matrix for all buildings in the database.

IPGA,i [g] D ≥ D0 D ≥ D1 D ≥ D2 D ≥ D3 D ≥ D4 D = D5

[0.000–0.076[ 272 23 2 1 1 0
[0.076–0.112[ 268 63 6 0 0 0
[0.112–0.159[ 270 112 32 22 18 10
[0.159–0.216[ 270 117 43 25 18 15
[0.216–0.297[ 270 137 86 59 40 26
[0.297–0.313[ 257 222 178 126 91 66
[0.313–0.349] 283 249 205 145 105 80
Total 1890 923 552 378 273 197

!

 Ni  N i
ni; j
n
P ni; j in N i with D ≥ Dj ¼ pi;i;j j 1
pi; j (1)
ni; j

!
Ni
where indicates the binomial coefficient; Ni is the total number of buildings in the the ith PGA
ni; j
interval, IPGA,i; and pi, j represents the probability of observing damage D ≥ Dj in that interval. This
probability can be estimated as

^pi; j ¼ ni; j =N i (2)

and its variance as

   
v^ar pi; j ¼ ^pi; j 1
^pi; j =N i (3)

where ^ indicates estimates.

Using the data of the damage matrix reported in Table V together with Eq. (2), we estimated the
failure probabilities for D ≥ Dj (j = 1 to 4) and D = D5. These probabilities are given in Table VI.
The central value μPGA,i of each interval, defined as the arithmetic mean of the PGA values, is also
reported. The damage probabilities in Table VI are also plotted in Figure 6 (black circles), together
with ±1 standard deviation intervals obtained from Eq. (3).
The boundaries of the seven PGA intervals were selected in order to (i) have a similar number of
buildings in each interval (approximately 270) [43] (see the column D ≥ D0 of Table V) and (ii)
obtain, for each damage level, non-decreasing exceedance probability values for increasing PGAs. It
should be noted that the intervals adopted were used only to provide a graphical representation of
fragility and do not affect the parametric fragility models discussed in the following.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2327

Table VI. Estimated failure probability for each damage level.

IPGA,i [g] μPGA,i [g] D ≥ D1 D ≥ D2 D ≥ D3 D ≥ D4 D = D5
[0.000–0.076[ 0.04 0.08 0.01 0.00 0.00 0.00
[0.076–0.112[ 0.09 0.24 0.02 0.00 0.00 0.00
[0.112–0.159[ 0.14 0.41 0.12 0.08 0.07 0.04
[0.159–0.216[ 0.19 0.43 0.16 0.09 0.07 0.06
[0.216–0.297[ 0.26 0.51 0.32 0.22 0.15 0.10
[0.297–0.313[ 0.31 0.86 0.69 0.49 0.35 0.26
[0.313–0.349] 0.33 0.88 0.72 0.51 0.37 0.28

Figure 6. (a–e) Comparison among failure probabilities for damage levels D1 to D5, point estimates from
damage matrices (black circles) and lognormal (LN) and log-logit (LL) parametric fragility curves obtained
by maximum likelihood estimation; (f) LL parametric fragility curves for the various damage levels. [Colour
figure can be viewed at wileyonlinelibrary.com]

5. PARAMETRIC FRAGILITY CURVES

5.1. General approach

Parametric fragility curves were fitted starting from the damage data described earlier. Various models
and regression procedures have been proposed in the literature to obtain fragility curves from
observational data, as described in the comprehensive review recently published by Lallemant et al.
[44]. In the present work, different models were considered, adopting a Bayesian approach in order
to estimate their parameters [45].
In the Bayesian framework adopted, a general parametric fragility model can be defined as a
function dependant on ground motion intensity IM and on a set of unknown regression
parameters Θ:

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2328 N. BURATTI ET AL.

Pf ðIM Þ ¼ f ðIM ; ΘÞ (4)

In Bayesian statistics, the current knowledge of Θ is defined by a joint density f(Θ) function, referred
to as priori distribution. Once a vector y of observed data is available, the Bayes’ theorem can be used
to update the knowledge of the parameters, so obtaining a posterior distribution:

PðyjΘÞf ðΘÞ PðyjΘÞf ðΘÞ

f ðΘjyÞ ¼ ¼ (5)
PðyÞ ∫PðyjΘÞf ðΘÞdΘ
where the function P(y|Θ) is referred to as likelihood function, L. The integrals involved in Eq. (5) can
be complicated for some combinations of priori distributions and likelihood functions, but they can be
easily solved using computational algorithms based on Markov chain Monte Carlo (MCMC) methods
and Gibbs sampling [45].

5.2. Fragility curves based on individual damage levels

The present section describes the procedure adopted for fitting parametric fragility models for the
general damage exceedance condition D ≥ Dj. To this aim, the observed damage data are first
transformed, for each damage level Dj, into a binary variable yi,j that is equal to 1 if, in the ith
building, damage is not less than Dj and 0 otherwise. Assuming that the damage data yi,j are
independent and identically distributed, the likelihood function Lj for the general damage level Dj
can be defined as [44, 46, 47]

   N 
ð1
yi; j Þ
y
Lj ¼ P yj Θj ¼ ∏ 1
pi; j IM i ; Θj pi; j IM i ; Θj i; j (6)
i¼1

where N indicates the total number of buildings observed, pi, j represents the probability to exceed the
damage level under consideration for the ground motion intensity IMi, Θi indicates the model
parameters and yj is a vector collecting the binary observations yi, j for the damage level Dj, that is, yj ¼
j k
yi; j ; …; yN ; j . Equation (6) corresponds to assuming that each binary damage observation, yi,j,
follows a Bernoulli distribution, B, with probability pi,j:


yi; j ∼B 1; pi; j IM i ; Θj (7)

It is worth noticing that the whole dataset is used for evaluating the likelihood function (6) for each
damage level Dj.
In the present work, two different models were considered for expressing pi,j as a function of ground
motion intensity: a lognormal (LN) model and a log-logit (LL) model. In the first case, the failure
probability pi,j is written as
 
lnðIM i Þ
μj
pi; j ¼ Φ (8)
σj

where Φ(·) indicates the standard normal cumulative distribution function. For each damage level Dj,
the model parameters are Θj = [μj, σ j]. In the second case, the failure probability is written as
 
logit pi; j ¼ β0; j þ β1; j lnðIM i Þ (9)

and the model parameters are Θj = [ β0, j, β1, j]. In both models, the logarithm of the ground motion
intensity was considered as covariate, in order to avoid non-zero damage probability for IM = 0.
Bayesian regression (Section 5.1) was carried out using the software R and JAGS [48, 49] in order to
estimate the parameters of the models. Convergence of the MCMC chains was checked by computing
the potential scale reduction factor [45]. Three MCMC chains were used. Uninformative distributions

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2329

were adopted as priori distributions of the model parameters. The two different models, that is, LN and
LL, were compared using the deviance information criterion (DIC), which is preferable than other
criteria as AIC and BIC when using MCMC Bayesian regression [45]. The DIC is computed based
on the deviance of a model and its number of parameters and, given the same goodness of fit to a
dataset, will favour models with less parameters [45].
Figure 6 shows the fragility curves obtained using the LN and LL models. Black dots at the top and
bottom of each panel represent the binary damage data, yj, used for fitting the models for each damage
state Dj. Table VII lists the mean values (used as estimates of the parameters) and the standard
deviation of the posterior distribution for the two parameters of the LL model. For all the damage
states considered, the DIC indicated the LL model as preferable, even if by a low margin.
Note that the curve for D ≥ D1 slightly overlaps the curve for D ≥ D2 starting from PGA values larger
than 0.45 g, which is clearly unjustifiable from a theoretical point of view. It is worth noticing that the
dataset used has a maximum PGA value of 0.35 g and, obviously, extrapolations of the fragility curves
are more uncertain. A possible solution to overcome this issue is using statistical models that force
ordinality of the damage states [44]. This approach will be discussed in the following.

5.3. Fragility curves using ordinal models

In the present section, an ordinal log-logistic (OLL) model is presented with the aim of avoiding
overlapping fragility curves. In particular, a link generalized linear model was used [50]. Alternative
approaches are presented in Agresti [51]. In the model adopted in this section, all the damage states
are considered together and a single likelihood function is defined. In this section, the damage for
the ith building is defined in terms of an ordinal damage variable yi, which can assume integer
values from 0 to 5, corresponding to damage levels ranging from D0 to D5, respectively.
The OLL model first requires to define a continuum latent variable y*, which is here assumed to have
a logistic distribution, on which linear regression is carried out, considering the logarithm of the ground
motion intensity as covariate:

yi ¼ lnðIM i Þβ þ ϵ; ϵ∼Logisticð0; sÞ; i ¼ 1; …; N (10)

where β is an unknown regression parameter and ε is a logistically distributed random variable with 0
mean and scale parameter s. Using a normal distribution for ε would generate a cumulative probit
model. In the present work, a proportional odds model was assumed; that is, the β parameter value
does not depend on the damage level [51]. This assumption is required in order to avoid
overlapping fragility curves. The continuum latent variable yi is mapped to the ordinal damage
variable yi corresponding to damage levels D0 to D5 (Table II), using the following scheme:

yi ≤ τ 0 ⇒ yi ¼ 0
τ j
1 < yi ≤ τ j ⇒ yi ¼ j; for 1 ≤ j ≤ 4 (11)
yi > τ 4 ⇒ yi ¼ 5

where j indicates the general damage level and τ 0 to τ 4 are unknown thresholds, to be defined by
regression, fulfilling the ordering constraint τ 0 < τ 1 < τ 2 < τ 3 < τ 4. The probability of observing the
different damage levels can be computed as

Table VII. E and SD of the posterior distribution of the parameters of the log-logistic models for the different
damage states.

D ≥ D1 D ≥ D2 D ≥ D3 D ≥ D4 D = D5
E[β0][
] 3.902 5.268 4.329 3.360 3.001
SD[β0][
] 0.192 0.301 0.344 0.371 0.445
E[β1][1/ ln(g)] 2.247 3.987 3.839 3.467 3.527
SD[β1][1/ ln(g)] 0.108 0.214 0.258 0.281 0.344
E, mean value; SD, standard deviation.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2330 N. BURATTI ET AL.

P½ yi ¼ 0 ¼ P yi ≤ τ 0
P½ yi ¼ j ¼ P τ j
1 < yi ≤ τ j ; for 1 ≤ j ≤ 4 (12)
P½ yi ¼ 5 ¼ P yi > τ4

which can be easily evaluated using the cumulative logistic distribution function as illustrated by
Figure 7. Finally, assuming that data are independent and identically distributed and defining an
indicator variable δi, j, which is 1 if yi = j (i.e. if damage in the ith building is equal to Dj) and 0
otherwise, the likelihood function L for the model can be written as

N 5
L ¼ ∏ ∏ P½ yi ¼ jδi; j (13)
i¼1 j¼0

The parameters of the so defined model are unidentifiable. In fact, any change in the scale parameter
s in Eq. (10) can be balanced by changes in τ and β. Therefore, this model requires a set of
normalization constraints. In the present work, the scale parameter s was set to 1 [50]. For the same
reason, no intercept parameter was defined in Eq. (10). In fact, this latter would be balanced by
shifting all the threshold values τ 0 , … , τ 4. Under these assumptions, the parameters of the model
are Θ = [ β, τ 0, τ 1, τ 2, τ 3, τ 4].
As suggested in [50], uninformative normal priori distributions were used for the model parameters.
Furthermore, β and τ 0 , … , τ 4 were assumed as independent a priori [50]. Convergence of the MCMC
chains was checked, as discussed in Section 5.2. The mean value and standard deviation of the
posterior distribution of the parameters are reported in Table VIII.
Considering the linear structure of the regression model in Eq. (10), associated with a non-negative
value of the estimate for the β parameter, and that cumulative distribution functions are non-decreasing,
it is evident that this model will lead to non-overlapping fragility curves.

5.4. Ground motion uncertainty

An important factor to consider in fragility estimation is the uncertainty in ground motion data. The
adopted ShakeMaps (Section 3.1) assume that the ground motion intensity, that is, PGA in this
paper, at each site has an LN distribution with a median value and a logarithmic standard deviation
as provided in Figure 4. In other words, the true value of the ground motion intensity, at the ith
building location, is not known and can be written as [52]

 
Figure 7. (a) PDF and (b) CDF for yi jIM i ; βj and yj IM j ; β, with IMj > IMi and β > C. The boundaries of the
intervals used to map the continuum latent variables yi and yj to the ordinal damage variables yi and yj are
indicated by dashed lines. The areas corresponding to the probability of observing damage Di = D1 (i.e. yi = 1)
and Dj = D1 (i.e. yj = 1) are hatched in (a).

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2331

Table VIII. E and SD of the posterior distribution of the parameters of the ordinal log-logistic models with
and without OLL-R and OLL ground motion uncertainty, respectively.
Model E[β][1/ ln(g)] E[τ 0][
] E[τ 1][
] E[τ 2][
] E[τ 3][
] E[τ 4][
]

OLL 2.554
4.508
3.299
2.616
2.136 -1.708

OLL-R 3.306
5.727
4.243
3.404
2.823 2.315
SD[β][1/ ln(g)] SD[τ 0] SD[τ 1] SD[τ 2] SD[τ 3] SD[τ 4]
OLL 0.103 0.183 0.167 0.161 0.161 0.163
OLL-R 0.187 0.318 0.276 0.255 0.245 0.239
E, mean value; SD, standard deviation; OLL, ordinal log-logistic; OLL-R, ordinal log-logistic model taking
ground motion uncertainty.



ln IM true
i ∼N lnðIM i Þ; σ lnðIM Þ;i (14)

where N indicates the normal distribution and IMi is the median ground motion intensity at the ith
building location (provided by the ShakeMaps in Figure 4(a) and (b)) and σ ln(IM),i its corresponding
logarithmic standard deviation (Figure 4(c) and (d)). Figure 8 shows the uncertainty on the PGA

Figure 8. Dataset used for the ordinal regression. (a–f) Damage levels D0 to D5, respectively. Each point
represents the median PGA value assigned to a building; colours indicate the earthquake that produced the
ground motion (20 May and 29 May). Vertical bars represent ±standard deviation intervals on the logarithm
of PGA, centred on the median PGA value. [Colour figure can be viewed at wileyonlinelibrary.com]

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2332 N. BURATTI ET AL.

associated with each building in the database (see also Figure 4). Vertical bars represent ±1 standard
deviation intervals on the logarithm of PGA, centred on the median PGA value. In general, the
PGAs associated with larger standard deviations are mostly those obtained from the ShakeMaps for
the 20 May earthquake (see red points in Figures 8 and 4(c)). It is worth noticing that for the
buildings with damage states D3 to D5 (Figure 8(d)–(f)), small PGA values systematically feature
larger standard deviations than large PGA values.
The error model defined in Eq. (14) is normally referred to as Berkson error model [53]. It differs
from traditional covariate error models (e.g. error in variables models) in which one assumes that
the measured value of the covariate can be defined as the summation of a true value and a random
error term with zero mean [53]. Equation (14) assumes that, on average, the ground motion
prediction is unbiased as suggested by Straub and Der Kiureghian [52]. This assumption is also
justified based on the procedure adopted for computing ShakeMaps [54]. The Berkson error model
can be included in MCMC-based Bayesian regression by considering that the IM value at each
building location is random. In particular, the median ground motion intensity IMi in Eq. (10) is
replaced by the random variable IM true i . Then, during the MCMC simulations, random samples of
this latter variable are generated using the PDF defined in Eq. (14).
Markov chain Monte Carlo Bayesian regression was carried out according to the procedure and
criteria described in the previous sections. The mean values and standard deviations of the posterior
distributions of the model parameters are reported in Table VIII, and the corresponding fragility
curves are plotted in Figure 9, together with curves obtained from the models discussed in the
previous sections. It is worth noticing that the curves obtained considering the uncertainty on PGA
(OLL-R) are, in general, steeper than those provided by the OLL model. This result can be
explained considering the non-uniform uncertainty of the PGA associated with the damage data
(Figure 8). In fact, as discussed earlier, low PGA values have larger uncertainties. Therefore, these
data are penalized in the regression that will favour points with smaller uncertainties. On the other
hand, the standard deviation of the posterior distribution of the OLL-R parameters (Table VIII) is
larger than that for the OLL model, and therefore, the confidence on the fragility curve (not plotted
here) is reduced.
Clearly, the proportional odds ordinal models provide a worst fit to the data if compared with those
discussed in Section 5.1, being these latter fitted independently for each damage level. A better fit to the
data could be possible, introducing higher-order terms (e.g. β1 ln(IMi)2) to the linear regression in Eq.
(10), but in that case, it might be possible to obtain decreasing curves that are not justified theoretically
[44]. Furthermore, a direct comparison of the goodness of fit of the ordinal model with those in
Section 5.2 is not possible because each of these latter uses the full dataset, which is converted into
a binary observation variable that will assume different values for each damage level. The models
related to the different damage states are therefore fitted independently and have different
likelihoods. The ordinal model, on the other hand, uses the full dataset for defining a single
likelihood function to obtain fragilities for all damage states. Given these considerations and the

Figure 9. Comparison of fragility curves obtained from the log-logit models (LL), the ordinal log-logistic
model (OLL) and the ordinal log-logistic model taking ground motion uncertainty into account (OLL-R).
[Colour figure can be viewed at wileyonlinelibrary.com]

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2333

different number of regression parameters – in the LL and LN models – two parameters per damage
level are adopted, while in the OLL models, six parameters in total – in the OLL and OLL-R
models – obviously provide a worst fit to the data if compared with those in Section 5.2 but have
the important advantage of providing non-overlapping fragility functions.
Finally, comparing the fragilities presented in the present work with those available in the literature
for cast-in-place RC frame structures, for example [1], it is easy to notice the much higher vulnerability
of prefabricated structures, especially as far as the most severe damage levels are considered (D3, D4
and D5). For example, in Figure 5(c) of [1], providing fragility curves for European-type RC
buildings derived from a large observational dataset, the PGA values corresponding to 50% failure
probability for the ‘extensive’, ‘partial collapse’ and ‘collapse’ damage states are 1.65, 2.11 and
2.27 g, respectively. On the contrary, the PGA values corresponding to 50% failure probability for
the fragility curves proposed in the present paper for damage states D ≥ D3, D ≥ D4 and D = D5 are
as low as 0.36, 0.43 and 0.50 g, respectively.

6. CONCLUSIONS

A database of seismic damage on 1890 precast RC buildings was assembled using data collected after
the 2012 Emilia earthquake. Both field surveys and information provided by structural engineers
appointed, by owners, to design retrofit/strengthening interventions for damaged buildings were
used. The consistency of the building database was analysed using cadastral data as reference.
The damage was classified on a five-level scale derived from EMS-98. Damage matrices were
evaluated from the database, and observational parametric fragilities were computed using a
simulation-based Bayesian approach. Two different classes of models were fitted: (i) models
considering the different damage levels independently and (ii) an ordinal logistic model that leads to
non-overlapping fragility curves. In fact, being the dataset limited to 0.35 g, fragility curves
obtained from the individual damage states were slightly overlapped for larger PGA values.
Furthermore, uncertainty on PGA was discussed and included in the ordinal model adopting a
Berkson error model.
The fragility curves obtained in the present work, when compared with literature fragilities for cast-
in-place RC frame buildings, indicate that precast industrial buildings are significantly more
vulnerable. Therefore, specific fragility models should be used for assessing the seismic risk related
to prefabricated buildings.
Finally, it should be noted that Emilia earthquakes caused PGA values not larger than 0.35 g and no
information was available on the behaviour of the buildings under consideration for stronger ground
motions. Therefore, the fragility curves obtained, in particular those related to the most severe
damage states considered, could be biased and should be used with care for stronger ground
motions. Nevertheless, the fragility models presented may provide important information for
validating fragility curves obtained from numerical models.

ACKNOWLEDGMENTS
The present investigation was developed within the activities of the (Italian) University Network of Seismic
Engineering Laboratories – ReLUIS in the research programme funded by the (Italian) National Civil
Protection– Progetto Esecutivo 2017 – Research Line ‘Reinforced Concrete Structures’, WP2. The cadastral
data reported in Section 2.2 were provided by the Italian Revenue Agency – Observatory of the Real Estate
Market. The damage data were made available by the Emilia-Romagna region.

REFERENCES
1. Rossetto T, Elnashai A. Derivation of vulnerability functions for European-type RC structures based on observational
data. Engineering Structures 2003, 25(10):1241–1263.
2. D’Ayala D, Spence R, Oliveira C, Pomonis A. Earthquake loss estimation for Europe’s historic town centres.
Earthquake Spectra 1997, 13(4):773–793.
3. Yamaguchi N, Yamazaki F. Fragility curves for buildings in Japan based on damage surveys after the 1995 Kobe
earthquake. Proceedings of the 12th World Conference on Earthquake Engineering, Auckland, New Zealand,
January 30–February 4, 2000.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2334 N. BURATTI ET AL.

4. Karababa FS, Pomonis A. Damage data analysis and vulnerability estimation following the August 14, 2003 Lefkada
Island, Greece, Earthquake. Bulletin of Earthquake Engineering 2011, 9(4):1015–1046.
5. Spence RJS, Coburn AW, Sakai S, Pomonis A. A parameterless scale of seismic intensity for use in the seismic risk
analysis and vulnerability assessment. Proceedings of the International Conference on Earthquake, Blast and Impact,
Manchester, United Kingdom, September 18–20, 1991.
6. Orsini G. A model for building’s vulnerability assessment using the parameterless scale of seismic intensity (PSI).
Earthquake Spectra 1999, 15(3):463–483.
7. Molina S, Torres Y, Benito B, Navarro M, Belizaire D. Using the damage from 2010 Haiti earthquake for calibrating
vulnerability models of typical structures in Port-au-Prince (Haiti). Bulletin of Earthquake Engineering 2013, 12
(4):1459–1478.
8. Hancilar U, Taucer F, Corbane C. Empirical fragility functions based on remote sensing and field data after the 12
January 2010 Haiti earthquake. Earthquake Spectra 2013, 29(4):1275–1310.
9. Rota M, Penna A, Strobbia CL. Processing Italian damage data to derive typological fragility curves. Soil Dynamics
and Earthquake Engineering 2008, 28(10–11):933–947.
10. Colombi M, Borzi B, Crowley H, Onida M, Meroni F, Pinho R. Deriving vulnerability curves using Italian
earthquake damage data. Bulletin of Earthquake Engineering 2008, 6(3):485–504.
11. Rossetto T, Iannou I, Grant DN. Existing empirical fragility and vulnerability functions: compendium and guide for
selection. GEM Technical Report 2013-X, GEM Foundation, Pavia, Italy, 2013. Available from http//www.
globalquakemodel.org Accessed 28 March 2017
12. Crowley H, Pinho R, Bommer JJ. A probabilistic displacement-based vulnerability assessment procedure for
earthquake loss estimation. Bulletin of Earthquake Engineering 2004, 2(2):173–219.
13. Buratti N, Ferracuti B, Savoia M. Response surface with random factors for seismic fragility of reinforced concrete
frames. Structural Safety 2010, 32(1):42–51.
14. Erberik MA. Fragility-based assessment of typical mid-rise and low-rise RC buildings in Turkey. Engineering
Structures 2008, 30:1360–1374.
15. Babič A, Dolšek M. Seismic fragility functions of industrial precast building classes. Engineering Structures 2016,
118:357–370.
16. Porter KA. An overview of PEER’s performance-based earthquake engineering methodology. 9th International
Conference on Applications of Statistics and Probability in Civil Engineering, San Francisco, CA, U.S., July 6–9,
2003.
17. Silva V, Crowley H, Varum H, Pinho R, Sousa R. Evaluation of analytical methodologies used to derive
vulnerability functions. Earthquake Engineering and Structural Dynamics 2014, 43(2):181–204.
18. Italian Revenue Agency. Observatory of the real estate market, www.agenziaentrate.gov.it, 2014 accessed 28 March
2017.
19. Nuti C, Vanzi I. Retrofitting of industrial structure. International seminar and exhibition on recent developments
in design and construction for precast concrete technology–REDECON 2014, Bangalore, India, November 9–13,
2014.
20. Restrepo JI, Park R, Buchanan AH. Design of connections of earthquake resisting precast reinforced-concrete
perimeter frames. PCI Journal 1995, 40(5):68–80.
21. Belleri A, Riva P. Seismic performance and retrofit of precast concrete grouted sleeve connections. PCI Journal
2012, 57(1):97–109.
22. Zoubek B, Fischinger M, Isakovic T. Estimation of the cyclic capacity of beam-to-column dowel connections in
precast industrial buildings. Bulletin of Earthquake Engineering 2015, 13(7):2145–2168.
23. Tullini N, Minghini F. Grouted sleeve connections used in precast reinforced concrete construction – experimental
investigation of a column-to-column joint. Engineering Structures 2016, 127:784–803.
24. Bellotti D, Casotto C, Crowley H, Deyanova MG, Germagnoli F, Fianchisti G, Lucarelli E, Riva S, Nascimbene R.
Single-storey precast buildings: probabilistic distribution of structural systems and subsystems from the sixties.
Progettazione Sismica 2014, 5(3):41–70 (in Italian).
25. Casotto C, Silva V, Crowley H, Nascimbene R, Pinho R. Seismic fragility of Italian RC precast industrial structures.
Engineering Structures 2015, 94:122–136.
26. Legislative Decree 74/2012. Interventi urgenti in favore delle popolazioni colpite dagli eventi sismici che hanno
interessato il territorio delle province di Bologna, Modena, Ferrara, Reggio Emilia, Mantova e Rovigo, il 20 e 29
maggio 2012. Rome, Italy, 2012. Available from http://www.gazzettaufficiale.biz/atti/2012/20120131/012G0096.
htm (in Italian) accessed 28 March 2017.
27. Dolce M, Di Bucci D. National Civil Protection organization and technical activities in the 2012 Emilia earthquakes
(Italy). Bulletin of Earthquake Engineering 2014, 12(5):2231–2253.
28. Bonfanti C, Carabellese A, Toniolo G. Strutture prefabbricate: catalogo delle tipologie esistenti. Report of the
(Italian) University Network of Seismic Engineering Laboratories–ReLUIS, February 2008. Available from http://
www.reluis.it/images/stories/Catalogo%20tipologie%20strutture%20prefabbricate.pdf (in Italian) accessed 28 March
2017.
29. Mandelli Contegni M, Palermo A, Toniolo G. Strutture prefabbricate: schedario di edifici prefabbricati in c.a. Report
of the (Italian) University Network of Seismic Engineering Laboratories–ReLUIS and the National Association of
Industrial Concrete Products–ASSOBETON, May 2008. Available from http://www.reluis.it/images/stories/
Schedario%20edifici%20prefabbricati%20in%20ca.pdf (in Italian) accessed 28 March 2017.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe
 10969845, 2017, 14, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/eqe.2906 by University Of California - Irvine, Wiley Online Library on [31/10/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SEISMIC FRAGILITY OF PRECAST BUILDINGS HIT BY THE EMILIA EARTHQUAKES 2335

30. Savoia M, Buratti N, Vincenzi L. Damage and collapses in industrial precast buildings after the 2012 Emilia
earthquake. Engineering Structures 2017, 137: 162–180.
31. Baggio C, Bernardini A, Colozza R, Corazza L, Della Bella M, Di Pasquale G, et al. Guide to filling in the 1st level
form for damage survey, prompt interventions and usability of ordinary buildings in the post-earthquake emergency
(AeDES). Italian Department of Civil Protection (ed), Rome, Italy, 2000 (in Italian).
32. Emilia-Romagna Regional Decree 57/2012. Available from http://www.regione.emilia-romagna.it/terremoto/gli-atti-
per-la-ricostruzione/2012/ordinanza-n.-57-del-12-ottobre-2012 (in Italian) accessed 28 March 2017.
33. Liberatore L, Sorrentino L, Liberatore D, Decanini LD. Failure of industrial structures induced by the Emilia (Italy)
2012 earthquakes. Engineering Failure Analysis 2013, 34:629–647.
34. Bournas DA, Negro P, Taucer F. Performance of industrial buildings during the Emilia earthquakes in Northern Italy
and recommendations for their strengthening. Bulletin of Earthquake Engineering 2014, 12(5):2383–2404.
35. Biondini F, Dal Lago B, Toniolo G. Role of wall panel connections on the seismic performance of precast structures.
Bulletin of Earthquake Engineering 2013, 11(4):1061–1081.
36. Minghini F, Ongaretto E, Ligabue V, Savoia M, Tullini N. Observational failure analysis of precast buildings after
the 2012 Emilia earthquakes. Earthquakes and Structures 2016, 11(2):327–346.
37. EMS-98. European Macroseismic Scale 1998. In: Grünthal G (ed) Cahiers du Centre Européen de Géodynamique et
de Séismologie, Vol. 15, Luxemburg, 1998.
38. Braga F, Gigliotti R, Monti G, Morelli F, Nuti C, Salvatore W, Vanzi I. Speedup of post earthquake community
recovery: the case of precast industrial buildings after the Emilia 2012 earthquake. Bulletin of Earthquake
Engineering 2014, 12(5):2405–2418.
39. Italian National Institute of Geophysics and Volcanology (INGV). ShakeMap Home Page, 2012. Available from
http://shakemap.rm.ingv.it/shake/index.html accessed 28 March 2017
40. Cultrera G, Faenza L, Meletti C, D’Amico V, Michelini A, Amato A. Shakemaps uncertainties and their effects in the
post-seismic actions for the 2012 Emilia (Italy) earthquakes. Bulletin of Earthquake Engineering 2014, 12
(5):2147–2164.
41. Braga F, Gigliotti R, Monti G, Morelli F, Nuti C, Salvatore W, Vanzi I. Post-seismic assessment of existing
constructions: evaluation of the shakemaps for identifying exclusion zones in Emilia. Earthquakes and Structures
2015, 8(1):37–56.
42. NTC08. Italian Building Code-D.M. 14/01/2008. Rome, Italy, 2008 (in Italian).
43. O’Rourke MJ, So P. Seismic fragility curves for on-level steel tanks. Earthquake Spectra 2000, 16(4):801–815.
44. Lallemant D, Kiremidjian A, Burton H. Statistical procedures for developing earthquake damage fragility curves.
Earthquake Engineering and Structural Dynamics 2015; https://doi.org/10.1002/eqe.2522.
45. Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB. Bayesian Data Analysis (3rd edn). Chapman &
Hall/CRC Press: Boca Raton, Florida, US, 2013.
46. Koutsourelakis PS, Pradlwarter HJ, Schuëller GI. Reliability of structures in high dimensions. Proceedings in
Applied Mathematics and Mechanics 2003, 3(1):495–496.
47. Shinozuka M, Feng MQ, Lee J, Naganuma T. Statistical analysis of fragility curves. Journal of Engineering
Mechanics 2000, 126(12):1224–1231.
48. R Development Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria: R
Foundation for Statistical Computing; 2008.
49. Plummer M. JAGS: A Program for analysis of bayesian graphical models using gibbs sampling. Proceedings of the
3rd International Workshop on Distributed Statistical Computing, Wien, Austria, 20–22 March: 2003.
50. Jackman S. Bayesian Analysis for the Social Siences. John Wiley & Sons: Chichester, United Kingdom, 2009.
51. Agresti A. Analysis of Ordinal Categorical Data. John Wiley & Sons: Hoboken, New Jersey, 2010.
52. Straub D, Der Kiureghian A. Improved seismic fragility modeling from empirical data. Structural Safety 2008, 30
(4):320–336.
53. Carroll RJ, Carroll RJ. Measurement Error in Nonlinear Models: A Modern perSpective. Chapman & Hall/CRC:
Boca Raton, Florida, US 2006.
54. Worden C., Wald DJ. ShakeMap Manual Online: technical manual, user’s guide, and software guide. 2016. DOI:
https://doi.org/10.1234/012345678.

Copyright © 2017 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2017; 46:2317–2335
DOI: 10.1002/eqe