EARTHQUAKE ENGINEERING & STRUCTURAL DYNAMICS

Earthquake Engng Struct. Dyn. 2015; 44:2057–2073

Published online 20 March 2015 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/eqe.2575

Average spectral acceleration as an intensity measure for collapse

risk assessment

Laura Eads1,*,†, Eduardo Miranda2 and Dimitrios G. Lignos3

1
Risk Management Solutions, Newark, CA 94560, U.S.A.
2
Department of Civil and Environmental Engineering, Stanford University, Stanford, CA 94305, U.S.A.
3
Department of Civil Engineering and Applied Mechanics, McGill University, Montreal, QC H3A 2K6, Canada

SUMMARY
This paper investigates the performance of spectral acceleration averaged over a period range (Saavg) as an
intensity measure (IM) for estimating the collapse risk of structures subjected to earthquake loading. The
performance of Saavg is evaluated using the following criteria: efficiency, sufficiency, the availability or ease
of developing probabilistic seismic hazard information in terms of the IM and the variability of collapse risk
estimates produced by the IM. Comparisons are also made between Saavg and the more traditional IM:
spectral acceleration at the first-mode period of the structure (Sa(T1)). Though most previous studies have
evaluated IMs using a relatively limited set of structures, this paper considers nearly 700 moment-
resisting frame and shear wall structures of various heights to compare the efficiency and sufficiency of
the IMs. The collapse risk estimates produced by Saavg and Sa(T1) are also compared, and the variability
of the risk estimates is evaluated when different ground motion sets are used to assess the structural
response. The results of this paper suggest that Saavg, when computed using an appropriate period range,
is generally more efficient, more likely to be sufficient and provides more stable collapse risk estimates than
Sa(T1). Copyright © 2015 John Wiley & Sons, Ltd.

Received 31 July 2014; Revised 17 February 2015; Accepted 23 February 2015

KEY WORDS: ground motion intensity measure; nonlinear structural response; performance-based
earthquake engineering

1. INTRODUCTION

An intensity measure (IM) quantifies the intensity of the ground motion and can depend solely on the ground
motion properties (e.g. peak ground acceleration) or on both the ground motion and structural properties
(e.g. spectral acceleration at the first-mode period of the structure, Sa(T1)). An IM serves as a useful link
between the ground motion hazard at a particular site and the response of a given structure and facilitates
risk-based calculations. The IM-based approach to collapse risk assessment involves combining the
probability of collapse conditioned on certain ground motion properties with the likelihood of
observing those ground motion properties. Assuming a homogenous Poisson process for earthquake
occurrence (as is typically done), the mean annual frequency of collapse (λc) is computed as follows
1 mmax r max
X
n
λc ¼
i¼1
λðM i > mmin Þ ∫0 m∫ ∫0 PðCjim; m; r Þ  f IM jM i ;Ri ðimjm; rÞ  f M i ;Ri ðm; rÞ dr dm d ðimÞ (1)
min

where n is the number of seismic sources considered, λ(Mi > mmin) is the mean annual frequency of
exceeding an event with magnitude mmin at source i, mmax and rmax are the maximum magnitude and

*Correspondence to: Laura Eads, RMS, 7575 Gateway Blvd, Newark, CA 94560, USA.
†
E-mail: laura.eads@rms.com

Copyright © 2015 John Wiley & Sons, Ltd.

2058 L. EADS, E. MIRANDA AND D. LIGNOS

distance, respectively, considered, P(Cjim,m,r) is the probability that the structure will collapse when
subjected to a ground motion with intensity level im produced by an event with magnitude m at a
distance r, fIM jMi,Ri(imjm,r) is the probability density function for the value of IM given values of
magnitude and distance (which is provided by a ground motion prediction equation, GMPE) and
fMi,Ri (m,r) is the joint probability distribution function for magnitude and distance. Equation (1) is
often simplified by separating the structural response and ground motion hazard components as follows
1  
dλIM ðimÞ
∫
λc ¼ PðCjimÞ  
0
dðimÞ
d ðimÞ 
(2)

where dλIM(im)/d(im) is the slope of the seismic hazard curve. In this study we adopt the definition of IM
sufficiency proposed by Luco and Cornell [1] in which an IM is said to be sufficient if it
renders a probabilistic measure of structural response (in this study the probability of collapse
given IM) that is conditionally independent of earthquake magnitude M and source-to-site
distance R (i.e. P(C j IM=im,M=m,R=r) = P(C j IM=im)). This means that the probability of
collapse at a given intensity level is not influenced by the earthquake magnitude and source-to-site
distance associated with the ground motion. Sufficiency is a desirable property of the IM because it
means that the probability of collapse estimates will not depend on the particular set of ground motions
used in structural response analyses to obtain an estimate of P(Cjim). Luco and Cornell [1] proposed
using M and R for quantifying sufficiency because in a probabilistic seismic hazard analysis (PSHA)
the IM is typically conditioned on these two parameters and therefore allows one to use Equation (2)
with the readily available results from PSHA instead of having to use Equation (1). In other words the
simplified IM approach implicit in Equation (2), although not perfect, has the enormous advantage of
decoupling the probabilistic structural response analysis (i.e. in this study the estimation of P(Cjim))
from the probabilistic seismic hazard analysis λIM(im). Given the complex relationship between ground
motion history and nonlinear structural response and the difficulty of capturing relevant ground motion
features with a single parameter, it is unlikely that a scalar IM will render a nonlinear structural
response that is conditionally independent of other ground motion parameters that may affect structural
response. The same is true for a vector IM that uses only a small number of parameters. So in the
context of performance-based earthquake engineering a sufficient IM (scalar or vector) would be one in
which P(CjIM=im,M=m,R=r) ≈ P(CjIM=im) to minimize the influence of the ground motion set that is
used to estimate P(CjIM=im). Therefore, in this paper an IM is said to be more sufficient than another
if P(C|IM=im,M=m,R=r) is closer to P(CjIM=im). To facilitate computation of λc via Equation (2) one
can use ground motions that are ‘hazard consistent,’ i.e. the distributions of the ground motion
parameters used in structural response analyses are consistent with the site hazard at the intensity
level of interest as identified through a probabilistic seismic hazard deaggregation. Bradley [2] and Lin
et al. [3] have demonstrated that similar risk-based estimates can be obtained using different IMs when
performing rigorous ground motion selection to ensure hazard consistency.
Another desirable property of an IM is efficiency, which is a measure of the level of variability
in the peak structural responses obtained using different ground motions with the same intensity level
(as measured by the IM). Thus, efficiency measures the ability of an IM to predict the structural
response. The more efficient an IM is, the smaller the number of response history analyses required
to estimate the structural response with a given level of confidence.
The ability to compute hazard information related to the IM is essential when performing any type of
risk-based structural response assessment using the IM approach. Specifically, one must be able to
develop a GMPE that estimates the probability of exceeding a given intensity level (as measured by
the IM) conditioned on the occurrence of a given seismic event (e.g. a given magnitude and
distance) in order to compute the seismic hazard associated with the IM.
A number of previous studies have focused on evaluating IMs based on their efficiency, sufficiency
and/or the ease of obtaining probabilistic seismic hazard information (e.g. [1, 4–6]), and some have
also considered the risk estimates generated by the IM (e.g. [7–11]). Though Sa(T1) is commonly
used as an IM, a number of researchers have found that spectral acceleration averaged over a period
range (Saavg) can be a significantly more efficient predictor of displacement-based nonlinear
structure response, including collapse, (e.g. [12–17]) and more likely to be sufficient (e.g. [15, 17])

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2059

than Sa(T1). Some researchers have improved the performance of Sa(T1) by including it in a vector IM
with other ground motion parameters (e.g. [8, 13, 18]), including those that have used a parameter
based on Saavg (e.g. [15, 19]); however, the present study is largely focused on evaluating Sa(T1)
and Saavg as scalar IMs. While most previous studies have considered a relatively limited set of
structures, this paper uses nearly 700 moment-resisting frame and shear wall structures of various
heights to evaluate the efficiency and sufficiency of Saavg and Sa(T1) for predicting structural
collapse. This paper also compares the collapse risk estimates produced by Saavg and Sa(T1) and
evaluates the variability of the risk estimates when different ground motion sets are used to
assess the structural response.

2. EVALUATING THE EFFICIENCY AND SUFFICIENCY OF SAAVG AND SA(T1)

This section evaluates the efficiency and sufficiency of Saavg and Sa(T1) as IMs for collapse prediction
using nearly 700 moment-resisting frame and shear wall structures of various heights. Saavg is
computed as the geometric mean of spectral acceleration values between periods c1T1 and cNT1
!1=N
Y
N
Saavg ðc1 T 1 ; … ; cN T 1 Þ ¼ Saðci T 1 Þ (3)
i¼1

where N is the number of periods used to compute Saavg and the ci terms are non-negative values. This
paper uses the 5%-damped response spectral values and computes Saavg using a period range between
0.2·T1 and 3·T1 with a uniform period spacing of 0.01 s unless noted otherwise. The period range
between 0.2·T1 and 3·T1 was chosen because it generally yielded good results in terms of the
efficiency and sufficiency of the IM with respect to collapse prediction. Examining the impact of the
period range and spacing using to compute Saavg is outside the scope of this paper but is discussed
in detail in Chapter 7 of [20].
While the geometric mean and arithmetic mean values are generally very similar, the former has
some advantages for computing the average spectral acceleration as it is less sensitive to extreme
spectral ordinates (i.e. very high or very low), which may significantly influence the arithmetic mean
value. More importantly, using the geometric mean permits GMPEs for Saavg, which are used in the
computation of seismic hazard curves, to be computed based on existing GMPEs for spectral
acceleration at a single period. Further discussion on obtaining GMPEs and hazard curves for Saavg
is provided later in Section 3.2.
The structures used to evaluate the IMs for collapse risk assessment include the 396 generic
moment-resisting frame (MRF) and the 252 generic shear wall structures used by Zareian and
Krawinkler (see Chapter 4 in [21]) as well as the 30 modern reinforced concrete (RC) MRF
structures used by Haselton and Deierlein (see Table 3.3 in [22]), which are representative of typical
structures designed in California. The generic MRF and generic shear wall structures range from
four to sixteen stories with fundamental periods between T1 = 0.4 s and 3.2 s and T1 = 0.2 s and 1.6 s,
respectively, while the RC MRF structures range from one to twenty stories with fundamental
periods between T1 = 0.42 s and 2.63 s. All structural models are two dimensional, include P-Delta
effects and capture material nonlinearity using rotational springs with the Ibarra–Medina–Krawinkler
hysteretic material that incorporates strength and stiffness deterioration [23]. Selected ground
motions from the PEER NGA database [24] representative of those expected at far-field sites in
California are used. Full details about the structures including geometry, material properties and
modeling assumptions and the ground motions used are provided in [21] and [22].
The collapse intensity is defined as the minimum IM value to which a particular ground motion must
be scaled to cause collapse (lateral dynamic instability) of a particular structure. The collapse intensities
in terms of Sa(T1) (denoted Sa(T1)col) were generously provided by the authors of [21] and [22].
The values were obtained via incremental dynamic analysis [25], in which nonlinear response
history analyses of a given structure subjected to a given ground motion are repeated at different
levels of ground motion intensity (as measured by the IM value) until the collapse intensity is
identified. The collapse intensity in terms of Saavg (denoted Saavg,col) is obtained by

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2060 L. EADS, E. MIRANDA AND D. LIGNOS

multiplying the Saavg value of the unscaled (i.e. as-recorded) ground motion with the collapse
scale factor (i.e. the ratio between Sa(T1)col and the Sa(T1) value of the unscaled ground
motion). The collapse intensities measured by Saavg and Sa(T1) are used to assess the
efficiency and sufficiency of the IMs in the following sections.

2.1. Efficiency
The efficiency of an IM is related to the number of response history analyses required to estimate the
structural response with a given level of confidence. In the case of collapse prediction, efficiency is
related to the variability of the collapse intensities. In this paper the efficiency of the IM is
quantified by the lognormal standard deviation, σ ln, (also called the dispersion) of the collapse
intensities for a particular structure. The lognormal distribution is used as the Sa(T1)col and Saavg,col
values are typically fit with this distribution (e.g. [16, 20–22]).
A summary of results provided in Table I shows that measuring the intensity of the ground motion
using Saavg reduces the dispersion in collapse intensities by an average of approximately 40% for
generic MRF and generic shear wall structures and over 50% for RC MRF structures compared to
dispersions computed using Sa(T1). In addition to being more efficient on average, Saavg was also
more efficient than Sa(T1) for all but four of the structures examined. For these four structures
σ ln[Saavg,col] was no more than 6% greater than σ ln[Sa(T1)col]. These results provide compelling
reasons for using Saavg as an IM instead of Sa(T1) if efficiency of the IM when estimating structural
collapse is a primary concern.
These results are consistent with those of other researchers who compared the dispersion in collapse
intensities using Saavg versus Sa(T1). For example, Vamvatsikos and Cornell [13] found that using an
average spectral acceleration with different weights assigned to different periods could reduce the
dispersion in the collapse intensities of a 20-story steel MRF structure by approximately 50%
compared to Sa(T1), and Tsantaki et al. [16] found a 25% – 45% reduction in dispersion for single-
degree-of-freedom (SDOF) systems with periods between 0.2 s and 5.0 s when using Saavg versus
Sa(T1). Note that these studies did not necessarily use the same period range to compute Saavg as the
range used in the present paper.
So why is Saavg significantly more efficient at predicting collapse than Sa(T1) (i.e. why is there
less dispersion in the collapse intensities when using Saavg as an IM)? Though both IMs are based
on the peak response of linear elastic SDOF systems, Saavg incorporates more information about the
ground motion than Sa(T1) that is relevant to the nonlinear response of a structure. In particular,
Sa(T1) does not consider the (elastic) response at other periods which might be important in
estimating nonlinear and MDOF structural response. Furthermore, the peak structural response
can be very sensitive to the particular displacement and velocity conditions of the system when
it is hit by a significant pulse in the ground motion history, and while Saavg is also based on
elastic ordinates, the influence of the ‘initial’ conditions (which are generally different for
inelastic and/or MDOF systems compared to an elastic SDOF system) at the arrival of significant
pulses is reduced by averaging the response over a range of periods [20: Chapter 5]. In this
paper the term ‘pulse’ refers to a segment of the ground acceleration history between zero
crossings (i.e. a half-cycle acceleration pulse). Finally, by taking an average spectral value over a
range of periods, Saavg indirectly accounts for the underlying pulses in the ground motion that
produce the response spectrum and cause structural damage.

Table I. Summary of dispersion of collapse intensities using Saavg and Sa(T1).

Mean value of dispersion Reduction in σ ln using Saavg (%)*

No. of
Structure type structures σ ln[Saavg,col] σ ln[Sa(T1)col] Mean Range

Generic MRFs 396 0.22 0.38 40 (6) – 66

Generic Shear Walls 252 0.28 0.48 41 17 – 67
RC MRFs 30 0.20 0.44 54 40 – 63
*Difference between σ ln[Sa(T1)col] and σ ln[Saavg,col], expressed as a percentage of σ ln[Sa(T1)col].

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2061

2.2. Sufficiency
This section evaluates the sufficiency of collapse intensities measured by Saavg and Sa(T1) with respect
to magnitude, distance and spectral shape. Essentially, this section explores whether some ground
motion parameter influences the collapse intensity (e.g. whether records from large magnitude
events are likely to produce either larger of smaller collapse intensities compared to records from
smaller magnitude events). It is desirable that the collapse intensities be conditionally independent
of (i.e. not affected by) the values of the ground motion parameters that are already incorporated in
estimating the IM in the seismic hazard analysis so that no bias exists in the structural response
results because of the particular ground motions used for response analysis when scaled to common
values of the IM. Note that there are many ground motion properties that may affect structural
response (e.g. duration and peak ground velocity), but only three are considered here because of
space limitations.
The sufficiency is evaluated by performing a standard linear regression of the natural logarithm of
the collapse intensities with the ground motion parameter of interest to identify possible trends (e.g.
systematic increments or reductions) in the response parameters computed using records with
different ground motion parameters. The regression is of the form

E ½Y jX ¼ x ¼ β0 þ β1  x (4)

where Y is the natural logarithm of the collapse intensity, X is the ground motion parameter of interest,
x is some value of X, β0 and β1 are coefficients determined by the regression and E[Y | X = x] is the
expected value of Y given some value x. A hypothesis test is then performed where the null
hypothesis is that β1 = 0 (i.e. that the expected value of Y does not depend on the value of X). The
probability of observing a β1 value at least as large (in absolute value terms) as the β1 value found
by the regression given that the true value of β1 equals 0 is calculated. This probability is called the
“p-value” and is used to determine whether the null hypothesis can be rejected at some predefined
significance level. A 5% significance level is typically used when judging the sufficiency of an IM
(e.g. [1, 10]) and is also used here. This means that a p-value less than 0.05 leads to a rejection of
the null hypothesis, and the IM is deemed not sufficient with respect to the ground motion
parameter of interest.
2.2.1. Sufficiency with respect to magnitude. The sufficiency of the collapse intensities with respect to
the earthquake magnitude (Mw) is evaluated for each structure. Though at least 98% of the generic
MRF and generic shear wall structures exhibited p-values ≥ 0.05 for both IMs, the narrow range of
magnitudes (Mw = 6.53 – 6.93) combined with the relatively small number of ground motions (40)
used to assess these structures is inadequate to judge the dependence of collapse intensity on
magnitude. The RC MRF structures, which are evaluated using 78 ground motions with a wider
magnitude range (Mw = 6.53 – 7.62), show that the collapse intensities are sufficient with respect to
magnitude for only 53% and 20% of these structures when using Saavg and Sa(T1), respectively.
However, when excluding structures with T1 < 1.2 s, 94% of the RC MRF structures show that the
collapse intensities are sufficient with respect to magnitude when using Saavg.
Earthquake magnitude influences the frequency content and thus the spectral shape of a ground
motion. This is reflected in GMPEs, which use a magnitude-dependent term to predict the response
spectrum. Figure 1 shows median response spectra predicted by the 2008 Boore and Atkinson
GMPE [26], denoted BA08. Spectra are computed for Mw = 6.5, 7.0 and 7.5 conditioned on a
Joyner–Boore distance (Rjb) of 12 km, a shear wave velocity of Vs30 = 285 m/s and an unspecified
fault type. The unscaled spectra in Figure 1(a) show that the spectral ordinates at a given period
increase with magnitude, as expected, and that the width of the spectral peak in the short period
range also increases with magnitude. Figure 1(b) shows the same spectra scaled to a common value
of Saavg (equal to 0.12 g), where the dashed vertical lines bracket the period range used to compute
Saavg (from 0.2·T1 and 3·T1, where T1 = 1.33 s). The scaled spectra shown in Figure 1(b) are re-
plotted in Figure 1(c) using a logarithmic axis for the spectral values, which makes it easier to
observe differences at longer periods. For the conditions used to create the spectra in Figure 1, it is
expected that ground motions from Mw = 7.5 events will have greater spectral values at moderate to

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2062 L. EADS, E. MIRANDA AND D. LIGNOS

Figure 1. Response spectra predicted by BA08 for different magnitudes with Rjb = 12 km, Vs30 = 285 m/s and
an unspecified fault type: (a) unscaled spectra; spectra scaled to a common value of Saavg shown with a (b)
linear Sa(T) axis and (c) logarithmic Sa(T) axis. The dashed vertical lines bracket the period range used to
compute Saavg.

long periods and smaller spectral values at short periods relative to ground motions from Mw = 6.5 or
7.0 events for equal values of Saavg.
As previously stated, magnitude typically affects the frequency content of the ground motion
(i.e. the intensity and duration of the pulses in the ground motion), which can affect the response of the
structure. While neither of the IMs directly accounts for differences in the relative frequency content
between ground motions, Saavg is thought to have a higher rate of sufficiency with respect to
magnitude because it ensures that the average spectral value over a given period (frequency) range is
equal among the ground motions, whereas Sa(T1) ensures that the spectral value at only a single
period (frequency) is equal among ground motions.
Interestingly, the RC MRF structures for which Sa(T1)col is sufficient or almost sufficient with
respect to magnitude are typically the same structures for which Saavg,col is not sufficient with
respect to magnitude and vice-versa. This is illustrated in Figure 2, which shows the p-value of the
relationship between collapse intensity and magnitude for each of the 30 RC MRF structures plotted
versus T1. The sufficiency of Sa(T1)col or Saavg,col with respect to magnitude appears to be related to
the first-mode period of the structure. The Saavg,col values are sufficient with respect to magnitude
for all but one of the structures with T1 > 1.5 s and for none of the structures with T1 < 1.2 s. (Note
that none of the structures analyzed have periods between T1 = 1.2 s and 1.5 s.) The median shape of
the Sa spectrum predicted by many GMPEs [e.g. 26] has a spectral peak in the short-period region
and decreasing Sa values at longer periods, with the slope decreasing as T increases as seen in
Figure 1. The slope of the spectrum (i.e. the relative frequency content) is expected to change
significantly as a function of T in the short-period range because of the spectral peak. As discussed
earlier, magnitude affects the absolute and, for a given value of Saavg, the relative frequency content
of the ground motion; however, Saavg does not directly account for the differences in relative

Figure 2. p-Values for the relationship between Mw and collapse intensity for the RC MRFs.

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2063

frequency content. When using a period range of 0.2·T1 to 3·T1 as done in this paper, the portion of the
period range associated with the spectral peak will be greater for short-period structures, which may
explain why the Saavg,col values of structures with shorter periods were more likely to be affected by
earthquake magnitude than those of structures with longer periods. For the RC MRF structures with
T1 < 1.2 s the average slope of the linear regression between ln(Saavg,col) and Mw is 0.24 meaning
that, on average, the Saavg,col value found using a ground motion from a Mw = 6.50 event is expected
to be approximately 30% larger than the Saavg,col value found using a ground motion from a
Mw = 7.62 event. It should be noted, however, that the correlation between ln(Saavg,col) and Mw is
only moderate, with an average correlation coefficient of ρ = 0.38 and an average R2 value of 0.14
for these structures, which indicates significant scatter in the relationship.

2.2.2. Sufficiency with respect to source-to-site distance. The sufficiency of the collapse intensities
with respect to the source-to-site distance of the record measured by Rjb is evaluated for each of
the nearly 700 structures. The results are summarized in Table II, which presents statistics on the
p-values of the linear relationship between the natural logarithm of collapse intensity and Rjb. As
shown in Table II, the results from nearly all the structures suggest that the collapse intensities
are sufficient with respect to source-to-site distance for both IMs. Given the previous results
concerning earthquake magnitude, this is somewhat expected because distance typically has much
less effect on spectral shape than magnitude in GMPEs [27].

2.2.3. Sufficiency with respect to ε. The spectral shape proxy epsilon (ε) measures the number of
logarithmic standard deviations a pseudo-acceleration spectral ordinate of a ground motion is from
the median ordinate predicted by a GMPE at a given period. Previous research (e.g. [8, 21, 22]) has
shown that structural response estimates such as peak inter-story drift ratios and collapse intensities
can be influenced by the ε(T1) values of the ground motions chosen for response analysis when the
IM is Sa(T1). The sufficiency of the collapse intensities is evaluated with respect to ε, which is
computed using the BA08 GMPE. Table III presents statistics on the p-values of the linear
relationship between the natural logarithm of collapse intensity and ε(T1). Saavg,col is sufficient with
respect to ε for only slightly more than half of the nearly 700 structures considered. For perspective,
nearly none of the structures show that Sa(T1)col is sufficient with respect to ε.
No trends were observed between structural properties such as number of stories or T1 and the
sufficiency of Saavg,col with respect to ε; however, a trend was observed with respect to the period
range used to compute Saavg. When using a period range of 0 to 3·T1 (instead of 0.2·T1 to 3·T1) to
compute Saavg, the percent of structures for which Saavg,col is sufficient with respect to ε increases to

Table II. Summary of p-values for the relationship between distance and collapse
intensity as measured by Saavg and Sa(T1).
% of structures with p-values ≥0.05
No. of Rjb range
Structure type structures [km] Saavg Sa(T1)

Generic MRFs 396 0.0 – 37.7 99 100

Generic shear walls 252 0.0 – 37.7 100 100
RC MRFs 30 0.9 – 74.2 93 100

Table III. Summary of p-values for the relationship between ε and collapse
intensity as measured by Saavg and Sa(T1).

% of structures with p-values ≥0.05

No. of
Structure type structures Saavg Sa(T1)

Generic MRFs 396 61 3

Generic shear walls 252 41 1
RC MRFs 30 43 0

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2064 L. EADS, E. MIRANDA AND D. LIGNOS

77% for the generic MRF structures and 81% for the RC MRFs. On average, this change in period
range does not affect the efficiency of Saavg as an IM for these structures. When Saavg is computed
over a period range between 0 and 6·T1 for the generic shear wall structures, the percent of
structures for which Saavg,col is sufficient with respect to ε increases to 83%, and the efficiency of
the IM increases as well.

3. EVALUATING COLLAPSE RISK ESTIMATES PRODUCED BY SAAVG AND SA(T1)

The results of the previous section suggest that with respect to collapse intensities Saavg is generally a
significantly more efficient IM than Sa(T1) and more likely to be sufficient with respect to magnitude
and spectral shape as measured by ε than Sa(T1). Both IMs were found to be generally sufficient with
respect to distance. This section evaluates the performance of Saavg and Sa(T1) with respect to their
collapse risk predictions for a case study. The ε adjustment proposed by Haselton et al. [28] to the
Sa(T1)-based collapse fragility curve to account for spectral shape is also evaluated. The discussion
focuses on how IM properties such as sufficiency and efficiency can affect the collapse risk estimate
and examines how the collapse risk estimate of a given IM varies when different ground motions
are used to estimate the structural response. As previously noted, it has been demonstrated that
different IMs can produce similar risk-based estimates when using hazard-consistent ground motions
to assess the structural response [2, 3]. Attempts at ensuring hazard consistency at all intensity
levels, such as the approaches proposed in [2] and [3], are not used here to permit the effects of IM
properties to be reflected in the collapse estimates. Therefore, the results here are examples of the
variation that can occur in collapse risk estimates using different IMs in the absence of hazard
consistency.

3.1. Structure and site

The structure is a four-story office building designed for the metropolitan area of Los Angeles,
California according to the 2003 International Building Code [29] and the 2005 AISC seismic
provisions [30, 31]. The lateral force-resisting system is a steel special MRF with reduced beam
sections with a design base shear coefficient of V/W = 0.082. Addition details about the design are
provided in [32: Section 5.2.] A two-dimensional model of one of the MRFs is created in OpenSees
[33]. A concentrated plasticity concept is used with the frame members modeled as elastic elements
with nonlinear rotational springs at their ends. The hysteretic behavior of the springs is governed by
a modified version of the bilinear Ibarra–Medina–Krawinkler deterioration model [32] that captures
stiffness and strength deterioration, including in-cycle degradation. P  Δ effects are simulated via a
leaning column carrying gravity loads that is connected to the frame by axially rigid beams pinned
at both ends. The first three modal periods of the structure are 1.33 s, 0.43 s and 0.22 s. Rayleigh
damping is implemented following the approach of Zareian and Medina [34] with 2% of critical
damping assigned to the first and third modes. Additional details on the modeling of the structure
are provided in [20].
The site is adapted from a hypothetical site used by Baker [35]. The seismicity at the site is
dominated by a single fault located 10 km from the site. The magnitude distribution is described by
the bounded Gutenberg–Richter recurrence law, which imposes a limit on the maximum earthquake
(mmax) a source can produce. In this example magnitudes between 6 and 8 are considered. The
relative ratio of small to large earthquakes in the region is taken as 1, and the mean annual rupture
rate of the fault is taken as λ(M > mmin) = 0.02. The fault rupture is assumed to include the segment
of the fault located nearest to the site so that Rjb, the minimum horizontal distance from the site to
the surface projection of the rupture area, is constant and equal to 10 km. The average shear wave
velocity at the site is Vs30 = 285 m/s (NEHRP Site Class D).

3.2. Seismic hazard curves

The IMs considered are Saavg and Sa(T1), where T1 = 1.33 s and Saavg is computed using a period range
from 0.2·T1 to 3·T1 with a uniform period spacing of 0.01 s. The hazard curves for each IM are

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2065

computed using standard PSHA calculations (see, e.g. [35]). Though adaptations to standard PSHA
calculations exist for incorporating the effects of near-fault directivity, which can be significant for
sites located close to faults (see, e.g. [36, 37]), they were not used in the simplified example
presented here. An essential step of PSHA is computing the probability that the IM exceeds a given
intensity level conditioned on the values of magnitude and distance (which is provided by a GMPE).
The lognormal distribution has been shown to be an appropriate model for both Sa(T1) and Saavg
[38], so their probability distributions are fully defined by an expected value and standard deviation.
Though many GMPEs provide the expected value and standard deviation of a spectral ordinate at a
given period, the authors are unaware of an explicit GMPE for Saavg. However, the GMPEs for
spectral ordinates at given periods can be used to estimate the expected value and variance of
ln(Saavg) as shown by Equations (12) and (13), respectively, of [39]. These equations show that
the expected value of ln(Saavg) is a function of the expected values of the ln(Sa(T)) values used
to compute ln(Saavg), while the variance of ln(Saavg) is a function of the standard deviation of
the ln(Sa(T)) values used to compute ln(Saavg) and the correlation of ln(Sa(T)) values at different
periods. The BA08 GMPE is used to compute the hazard curves for both Sa(T1) and Saavg, and the
correlation coefficients are taken from [39]. It should be noted that the expected value and variance of
ln(Saavg) computed via Equations (12) and (13), respectively, of [39] may differ from those computed
via a regression analysis directly on the ln(Saavg) values of individual ground motions (i.e. values
from an explicit GMPE for ln(Saavg)). Reasons for these differences include: (i) different records may
be used in a regression of ln(Sa(T)) versus a regression of ln(Saavg) because of limitations associated
with the maximum usable period of some records; (ii) equations for predicting the expected value
of ln(Sa(T)) with existing GMPEs do not yield exact matches to the empirical data (i.e. residuals may
differ from zero); and (iii) equations for predicting the correlation coefficients used to compute the
variance of ln(Saavg) also do not yield exact matches to the empirical data. Nevertheless, these
equations are used here for convenience. The resulting seismic hazard curves are shown in Figure 3(a).

Figure 3. Collapse risk components for different IMs using the MRCD 137 ground motion set: (a) seismic
hazard curves, (b) collapse fragility curves and (c) λc deaggregation curves.

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2066 L. EADS, E. MIRANDA AND D. LIGNOS

3.3. Ground motion sets

Seven ground motion sets are used to assess the collapse risk of the structure and evaluate the risk
estimates produced by each IM. An overview of the ground motion sets is provided in Table IV,
including the name of the set, number of ground motions, magnitude and distance range and a
source for further information about the ground motions. Records from Set #1A are used for two
different ground motion sets, which differ based on the orientation of the ground motions. In one set
(denoted ‘FN/FP’) the records are rotated to fault-normal and fault-parallel orientations, and in the
other set (denoted ‘unrotated’) the records are not rotated and are used as provided in the PEER
NGA database [24]. With the exception of the LMSR-N set, both horizontal components from a
given recording are used and are treated as independent records (e.g. MRCD 137 is comprised of
137 recordings, each with two horizontal components, which gives 274 records). Further
information about the ground motion sets can be found in the sources listed in Table IV.

3.4. Collapse risk analysis results

Incremental dynamic analysis is used to find the collapse intensity of each ground motion, where
collapse is defined as lateral dynamic instability. The collapse risk is quantified by λc as shown in
Equation (2). A lognormal cumulative distribution function is fit to the collapse intensities to
describe P(C|im), which is commonly known as a collapse fragility curve. The collapse fragility
curves for all IM and ground motion set combinations pass the Kolmogorov–Smirnov goodness-of-
fit test [40] at the 5% significance level, meaning that the lognormal distribution is a reasonable
assumption for the collapse fragility curve. The probability of at least one collapse in 50 years (Pc,50)
is also used here to quantify the collapse risk. It is essentially a transform of λc that is easier to
communicate to project stakeholders. Making the common assumption that the occurrence of
collapses in time follows a homogenous Poisson process, the probability of at least one collapse
over t years can be computed as

Pc ðin t yearsÞ ¼ 1  expðλc  t Þ (5)

Collapse risk analysis results using the MRCD 137 ground motion set are examined in detail to
study the effects of efficiency and sufficiency on the risk estimates produced by the different IMs.
The focus is then expanded to study the variability of the risk estimates when different ground
motion sets are used to assess the structural response. Components of the collapse risk analysis
using the MRCD 137 ground motion set are presented in Figure 3, which shows the seismic hazard
curve in Figure 3(a), the collapse fragility curve in Figure 3(b) and the λc deaggregation curve in
Figure 3(c). The λc deaggregation curve is the inner product in the integral of Equation (2) and gives
the contribution to λc by IM level.
A summary of collapse risk metrics and related parameters is provided for each IM in Table V.
Collapse risk metrics λc and Pc,50 are computed for each IM using Equations (2) and (5), respectively.
As shown in Table V the IMs produce different collapse risk estimates, with Sa(T1) producing a Pc,50
estimate of approximately 1.0% versus approximately 0.3% using Saavg. One would like to know
which IM gives the better estimate of the collapse risk. While the ground motions a structure will
experience in its lifetime are unknown, one can estimate λc using information about the types and
frequencies of seismic events that are likely to occur at the site and information about how the

Table IV. Summary of ground motion sets.

Name Number of records Mw range Rjb range [km] Source
MRCD 137 274 6.9 – 7.6 0 – 27 [20: Appendix A]
LMSR-N 40 6.5 – 6.9 0 – 38 [43: Section 3.3]
Set One 78 6.5 – 7.6 1 – 74 [22: Appendix 3B]
Set #1A 80 6.1 – 7.9 0 – 40 [41: Section 3.1]
CS, ε(T1) = 1.7 80 6.1 – 7.6 0 – 147 [20: Appendix C]
CS, ε(T1) = 1.8 80 6.1 – 7.6 0 – 106 [20: Appendix C]

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2067

Table V. Collapse risk metrics and related parameters using the MRCD 137 ground motion set.
IM Median collapse intensity σ lnIMcol σ lnIM,GMPE* λc [104] Pc,50
Sa(T1) 1.03 g 0.39 0.68 2.03 1.01%
Sa(T1) + ε 1.33 g 0.39 0.68 0.88 0.44%
Saavg 0.70 g 0.22 0.59 0.59 0.29%
*Note: σ lnIM,GMPE is the standard deviation from the GMPE, while σ lnIMcol is the standard deviation of the collapse
intensities.

structure is likely to respond to those events (i.e. whether or not they produce collapse). This
is accomplished as shown in Equation (2). The following discussion uses properties of the IMs
(e.g. efficiency, sufficiency, etc.) to argue why one IM gives a better estimate of λc (or, equivalently, Pc,50)
than the other, given that careful ground motion selection to ensure hazard consistency at all intensity
levels is not performed.

3.4.1. Effects of efficiency and sufficiency. Both the Sa(T1)col and Saavg,col values are found to be
sufficient with respect to magnitude and distance for this case study (i.e. the case study discussed in
Section 3). The Saavg,col values are also found to be sufficient with respect to ε for this case study;
however, the Sa(T1)col values are not. Figure 4 shows the Sa(T1)col values plotted against ε(Τ1). The
positive slope of the linear regression in Figure 4 indicates that the Sa(T1)col values tend to increase
with the value of ε. To reduce the bias in the collapse risk estimate arising from using ground
motions with ε values that are not consistent with the site hazard, the collapse fragility curve is
adjusted using the approach of Haselton et al. [28]. Note that no adjustment is necessary for the
Saavg-based risk estimate because it is sufficient with respect to ε for this case study (i.e. from a
statistical perspective the ε value of the ground motion does not affect Saavg,col). If the Saavg,col
values were found to be insufficient with respect to ε and to introduce a significant bias in the
collapse estimate, a bias-correction procedure similar to that developed by Haselton et al. [28] could
be applied (note that such a procedure is not developed here), or ground motions could be reselected
such that the ε values are appropriate for the particular site and hazard level or such that the
reselected records produce Saavg,col values that are sufficient with respect to ε. Based on the results
presented in Section 2.2.3 when computing Saavg using period range of 0.2·T1 to 3·T1, Saavg,col
values are expected to be sufficient with respect to ε in slightly more than 50% of cases where
careful ground motion selection is not used.
In the approach of Haselton et al. [28], the median collapse intensity is calculated as the expected
collapse intensity at the target ε level, which is based on a regression between collapse intensities
and ε values as shown in Figure 4. The target ε level is taken as the ε value with the greatest

Figure 4. Natural logarithm of Sa(T1)col vs. ε(T1) for the MRCD 137 ground motion set.

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2068 L. EADS, E. MIRANDA AND D. LIGNOS

contribution to the seismic hazard at a given intensity level as identified through a PSHA
deaggregation. A PSHA deaggregation by magnitude and ε at Sa(T1) = 0.72 g is shown in Figure 5
and is used to determine the target ε value. This intensity level is chosen because it is at the peak of
the λc deaggregation (Figure 3(c)), meaning it has the most significant contribution to the collapse
risk. The PSHA deaggregation in Figure 5 shows that a Mw = 6.7 event, which is associated with
ε(T1) = 1.8, has the greatest contribution to the seismic hazard at this intensity level. Based on the
regression presented in Figure 4, at the target value of ε(T1) = 1.8 the expected collapse intensity is
Sa(T1)col = 1.33 g (up from 1.03 g, see Table V). Haselton et al. note that adjusting the median
collapse intensity in this manner also reduces the dispersion of the collapse intensities;
however, they found that the reduction in dispersion was on the order of 10% and had only a
minor effect on the collapse risk. They proposed simply using the dispersion calculated from
the original (i.e. unadjusted) collapse intensities, which is adopted here.
The collapse risk results based on the ε-adjusted collapse fragility curve are denoted ‘Sa(T1) + ε’ and
are shown in Figure 3 and Table V. The collapse risk measured by Sa(T1) + ε is Pc,50 = 0.44% as shown
in Table V. This is approximately a 60% reduction in Pc,50 compared to the value calculated for Sa(T1)
without considering ε; however, this adjusted value of Pc,50 is still approximately 1.5 times greater than
the Pc,50 computed using Saavg. It is believed that a significant contributor to Sa(T1) producing a higher
collapse risk estimate is the larger variability in the collapse intensities (a larger σ lnIMcol) combined with
the larger variability from the GMPE perspective (σ lnIM,GMPE) (i.e. the variability in the expected value
of the IM for a given magnitude and distance) for Sa(T1) compared to Saavg, as shown in Table V.
The standard deviations indicate how well the IM values observed from recorded ground motions
match the GMPE’s prediction of the expected value and as such indicate, loosely speaking, how
easy or difficult it is to predict the value of the IM. Perhaps it is not surprising that Saavg has a
smaller value of σ lnIM,GMPE than Sa(T1) (Table V) as one might expect that the process of averaging
spectral ordinates over a range of periods reduces extreme values. The mathematical basis for why
σ lnSaAvg,GMPE is smaller than σ lnSa(T1),GMPE can be understood by considering that σ lnSaAvg,GMPE is a
function of the σ lnSa(T),GMPE values for periods used to compute Saavg and the correlation
coefficients for spectral values at different periods. Figure 6(a) shows the former while Figure 6(b)

Figure 5. PSHA deaggregation by magnitude with associated ε(T1) values for Sa(T1) = 0.72 g.

Figure 6. Components influencing the value of σ lnSaAvg,GMPE: (a) values of σ lnSa(T),GMPE; and (b) contour of
correlation coefficients between spectral values at different periods.

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2069

shows the latter obtained from [39]. For the case study Saavg is computed between periods of 0.27 s and
3.99 s (which is the period range between 0.2·T1 and 3·T1, where T1 = 1.33 s), which means this period
range has over twice as many periods greater than T1 than less than T1. Although periods greater than
T1 have greater values of σ lnSa(T),GMPE compared to the value at T1 as seen in Figure 6(a), the spectral
values are not perfectly correlated as seen in Figure 6(b). The reduction in the correlation of spectral
ordinates as one departs from T1 leads to a reduction in σ lnSaAvg,GMPE compared to the value of
σ lnSa(T1),GMPE in nearly all cases.

3.4.2. Effect of ground motion selection. To test the sensitivity of the previous observations to the
particular ground motions used, the collapse fragility curve is computed using different sets of
ground motions, and the collapse risk estimates are calculated for each set. Ideally, a good IM will
give relatively stable estimates of the collapse fragility and collapse risk no matter what ground
motions are used (i.e. a small variability in the various estimates computed with different ground
motion sets) because it is hoped that the IM is sufficient with respect to magnitude, distance, ε and
other properties of the ground motions that may affect the structural response.
The ground motion sets are listed in Table IV. The ground motions in the MRCD 137, Set One and
LMSR-N sets were chosen primarily based on magnitude, distance, soil type and fault type, whereas
ground motions in Set #1A and the CS sets were chosen to match target spectra. As explained in
[41], Set #1A ground motions were selected to match the means and standard deviations of spectral
ordinates between periods of 0 and 5 s predicted by the BA08 GMPE for a Mw = 7 strike–slip
earthquake 10 km from a site with an average shear wave velocity of Vs30 = 250 m/s. The target
spectra for the CS sets are based on a PSHA deaggregation for the case study site at an intensity of
Sa(T1) = 0.68 g, which corresponds to the 2/50 hazard level and is expected to have a significant
contribution to the collapse risk. Based on the deaggregation, the magnitude with the largest
contribution to the hazard is Mw = 6.7, which corresponds to ε(T1) = 1.7. The target spectrum is
a conditional spectrum (CS), which describes the full distribution of the spectrum at all periods
(i.e. it captures both the expected spectral values and the variability of those values) [3]. It is
constructed using these Mw and ε(T1) values with the other site parameters (i.e. Rjb = 10 km,
Vs30 = 285 m/s, and an unspecified fault type) using the BA08 GMPE. A similar CS is also created
with ε(T1) = 1.8 to be consistent with the target ε(T1) value of the MRCD 137 set as shown in
Figure 5 and to be closer to the target ε(T1) values of other ground motion sets as discussed later.
Note that the CS can be used to select hazard-consistent ground motions; however, one generally
needs to compute the CS (and reselect ground motions) for each intensity level of interest as the
hazard distributions generally change with the intensity level, and this was not done here.
Although the record selection methodology and the values of many target parameters (e.g.
magnitude, distance, etc.) are similar for the Set #1A and the CS sets, there is an important
difference between these two sets: the CS sets were conditioned on non-zero values of ε(T1), which
affects the target spectrum because of correlations between ε values at different periods. The positive
ε(T1) values used to construct the target spectra for the CS sets imply that these spectra typically
have a peak at or near T1. In contrast, the target spectrum used to select the Set #1A ground motions
is based on the median prediction from the GMPE, which implies that the ε values are zero for all
periods, which results in different spectral shape characteristics between the ground motion sets, as
illustrated in Figure 7. Figure 7 compares the (geometric) mean spectra of the ground motion sets,
where all records are scaled to a common value of Sa(T1) in Figure 7(a) and a common value of
Saavg in Figure 7(b). Figure 7 shows that the overall (i.e. mean) relative frequency content differs
among ground motion sets. For example, the MRCD 137 and both Set #1A ground motion sets
generally have higher spectral ordinates compared to the other ground motion sets between periods
of 3 s and 4 s, while the LMSR-N and Set One sets generally have higher spectral ordinates
compared to the other ground motion sets at short periods, regardless of which IM is used to scale
the records. The CS sets exhibit peaked spectra near T1 and tend to have higher spectral ordinates
near T1 compared to the other ground motion sets when scaled to a common value of Saavg.
The collapse risks computed by each IM for the different ground motion sets are summarized in
Figure 8. The ε-adjusted collapse risk results for Sa(T1) are also presented in Figure 8 and are
denoted as ‘Sa(T1) + ε’. A PSHA deaggregation at the intensity associated with the peak of the λc

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2070 L. EADS, E. MIRANDA AND D. LIGNOS

Figure 7. Geometric mean spectrum of different ground motion sets scaled to a common value of (a) Sa(T1)
and (b) Saavg. Note the vertical line in (a) is drawn at T1 = 1.33 s, and the horizontal line in (b) denotes the
common value of Saavg over the period range between 0.2·T1 and 3·T1.

Figure 8. Summary of collapse risk results for different IMs and different ground motion sets: (a) median
collapse intensity; (b) standard deviation of collapse intensities, σ lnIMcol; (c) mean annual frequency of
collapse, λc, and probability of collapse in 50 years, Pc,50.

deaggregation is used to determine the target ε value for each ground motion set. This intensity varies
between ground motion sets (Sa(T1) = 0.70 g to 0.87 g), which results in the target ε value varying
between the ground motion sets (ε(T1) = 1.8 to 2.1). All PSHA deaggregations show that the
magnitude with the greatest contribution to the seismic hazard is Mw = 6.7.
Figure 8(a) presents the median collapse intensities for each IM. As previously discussed, it is
desirable that the median collapse intensity of a given IM remains relatively stable for different
ground motion sets because this indicates that the results are likely not affected by properties of the
ground motions such as magnitude, distance, ε or other properties that differ between the ground
motion sets, and therefore the results are more reliable as they are less affected by decisions made
when selecting ground motions. Figure 8(a) shows that the median collapse intensities measured by
Saavg are relatively stable as they only vary between Saavg = 0.70 g and 0.77 g (a 5% variation with
respect to the value in the middle of this range of intensities). The median collapse intensities
measured by Sa(T1) + ε, on the other hand, vary considerably more as the values range from Sa(T1)
= 1.33 g to 1.72 g (a 13% variation with respect to the value in the middle of this range of

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2071

intensities). The dispersions on the median collapse intensity are presented in Figure 8(b) and show that
Saavg is the more efficient IM for all ground motion sets, with an average dispersion value of
σ lnSaAvgCol = 0.25. Note that, for reasons previously stated, the dispersion of collapse intensities using
the Sa(T1) + ε method is taken as equal to the dispersion of collapse intensities based on Sa(T1) alone.
The values of λc and Pc,50 for each IM are presented in Figure 8(c). Discussion herein will focus on
the Pc,50 values as they are simply a transform of the λc values as shown by Equation (5). Figure 8(c)
shows that collapse risk estimates are most stable for Saavg, which has Pc,50 values ranging from 0.22%
to 0.33% (a 20% variation with respect to the value in the middle of this range). This indicates that
compared to Sa(T1) the collapse risk estimated using Saavg is relatively insensitive to the particular
set of ground motions used, which is a very desirable property. This relatively narrow range of
collapse risk estimates is obtained using records sets with very different characteristics, and yet they
lead to very similar collapse risk estimates. For example Figure 7(b) shows that when scaled to the
same Saavg value records in the MRCD 137 set have, on average, significantly larger spectral
ordinates than the LMSR-N set between periods of 3 s and 4 s while between periods of 0 and 1 s
the trend is reversed. Despite differences in relative frequency content that Saavg does not capture,
the collapse risk estimates using these two sets are very similar as the Pc,50 values are 0.29% and
0.23% for the MRCD 137 and LMSR-N sets, respectively.
Sa(T1) + ε produces collapse risk estimates ranging from Pc,50 = 0.14% to 0.44% (a variation of
approximately 50% with respect to the value in the middle of this range), which indicates that even
after correcting for the different ε values between the ground motion sets the collapse risk computed
by this IM is still sensitive to the records used in the evaluation.

4. SUMMARY AND DISCUSSION

Studies of nearly 700 MRF and shear wall structures of various heights showed that while both Sa(T1)
and Saavg were generally sufficient with respect to the distance of the ground motions used to assess
collapse, Saavg was approximately 40% more efficient, on average, at predicting the collapse
intensities and was more likely to be sufficient with respect to magnitude and spectral shape proxy
ε. Analysis of 30 RC MRF structures showed that while Sa(T1) was sufficient with respect to
earthquake magnitude for only 20% of the structures, Saavg was sufficient for approximately 50% of
the structures and for over 90% of the structures with T1 > 1.5 s. The results presented in the
collapse risk assessment case study suggest that Saavg is a good IM to use for collapse risk
assessment, as compared to Sa(T1) it is a better (more efficient) predictor of structural response, is
predicted with less dispersion from a GMPE perspective and produces relatively stable collapse risk
estimates using different ground motion sets (which is a reflection of its sufficiency with respect to
ground motion properties that affect structural response). Though seismic hazard curves in terms of
Sa(T1) are relatively widely available, this is not the case for Saavg and is a significant factor
preventing the widespread use of the IM in practice. However, hazard curves for Saavg can be
generated using existing GMPEs for spectral values at given periods and existing equations for
correlations between spectral values at different periods, and implementing these for Saavg is a goal
of the OpenSHA project [www.opensha.org].
It has been demonstrated that one can compute similar collapse risk estimates with different IMs
using rigorous ground motion selection to ensure hazard consistency [2, 3]. In practice collapse risk
assessment involves using a limited number of ground motions to estimate the structural response,
and, in some cases, there may be a lack of available ground motions that are consistent with the
ground motions expected at a particular site and intensity level, especially if one tries to match joint
distributions of several ground motion parameters. Though shortcomings in the selected ground
motions can be accounted for through adjustment procedures, they may be cumbersome and do not
guarantee convergence to the correct answer. Because of these practical issues associated with
assessing the collapse risk there is a benefit to using an IM that is predicted with less uncertainty
from the GMPE perspective, is an efficient predictor of structural response and is sufficient with
respect to ground motion properties affecting structural response, or at least one producing collapse
fragility estimates that are less dependent on ground motion properties, and therefore less biased,

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
2072 L. EADS, E. MIRANDA AND D. LIGNOS

than estimates obtained using alternative IMs. Furthermore, the relatively stable collapse risk estimates
obtained across the ground motion sets when using Saavg are indicators of the sufficiency of this IM
with respect to various ground motion properties and suggest that collapse risk estimates obtained
using this IM are not particularly sensitive to the ground motions used in structural response
(collapse fragility) analysis. This implies that the careful record selection and/or modification of
structural response results required to obtain a good estimate of the collapse risk when using Sa(T1)
as the IM may not be required or may be significantly reduced if Saavg is the IM. Future work is
necessary to confirm this, however, as the results presented here are based on only a single case
study. In particular, it would be valuable to study the sufficiency of Saavg with respect to ground
motion duration and pulse-like ground motions as these parameters were not explicitly examined
here and have been shown to bias structural response estimates when Sa(T1) is the IM (e.g. see [42]
and [1], respectively).

ACKNOWLEDGEMENTS
The authors would like to thank Professor Farzin Zareian and Professor Curt Haselton for sharing their
collapse analysis results for generic moment-resisting frame and shear wall structures and for modern
reinforced concrete moment-resisting frame structures, respectively. They would also like to thank the
two reviewers, whose comments helped improve the quality of this paper. This research is funded by the
US National Science Foundation (NSF) under Grant No. CMS-0936633 within the George E. Brown, Jr.
Network for Earthquake Engineering Simulation Research (NEESR) Consortium Operations. The financial
support of NSF is gratefully acknowledged. Any opinions, findings and conclusions or recommendations
expressed in this paper are those of the authors and do not necessarily reflect the views of NSF.

REFERENCES
1. Luco N, Cornell CA. Structure-specific scalar intensity measures for near-source and ordinary earthquake ground
motions. Earthquake Spectra 2007; 23(2):357–392. DOI:10.1193/1.2723158.
2. Bradley BA. The seismic demand hazard and importance of the conditioning intensity measure. Earthquake Engi-
neering and Structural Dynamics 2012; 41(11):1417–1437. DOI:10.1002/eqe.2221.
3. Lin T, Haselton CB, Baker JW. Conditional spectrum-based ground motion selection. Part I: Hazard consistency for
risk-based assessments. Earthquake Engineering and Structural Dynamics 2013; 42(12):1847–1865. DOI:10.1002/
eqe.2301.
4. Shome N, Cornell CA, Bazzurro P, Carballo JE. Earthquakes, records, and nonlinear responses. Earthquake Spectra
1998; 14(3):469–500. DOI:10.1193/1.1586011.
5. Aslani H, Miranda E. Probabilistic earthquake loss estimation and loss disaggregation in buildings. Report No. 157;
The John A. Blume Earthquake Engineering Center, Stanford University, Stanford, CA, 2005.
6. Riddell R. On ground motion intensity indices. Earthquake Spectra 2007; 23(1):147–173. DOI:10.1193/1.2424748.
7. Cordova PP, Deierlein GG, Mehanny SSF, Cornell CA. Development of a two-parameter seismic intensity measure
and probabilistic assessment procedure. Proc., The Second U.S.–Japan Workshop on Performance-Based Earth-
quake Engineering Methodology for Reinforced Concrete Building Structures, Sapporo, Hokkaido, Japan, 2000;
187–206.
8. Baker JW, Cornell CA. Spectral shape, epsilon and record selection. Earthquake Engineering and Structural Dynam-
ics 2006; 35(9):1077–1095. DOI:10.1002/eqe.571.
9. Tothong P, Luco N. Probabilistic seismic demand analysis using advanced ground motion intensity measures. Earth-
quake Engineering and Structural Dynamics 2007; 36(13):1837–1860. DOI:10.1002/eqe.696.
10. Bradley BA, Dhakal RP, MacRae GA, Cubrinovski M. Prediction of spatially distributed seismic demands in
specific structures: ground motion and structural response. Earthquake Engineering and Structural Dynamics
2010; 39(5):501–520. DOI:10.1002/eqe.954.
11. Bradley BA, Dhakal RP, MacRae GA, Cubrinovski M. Prediction of spatially distributed seismic demands in
specific structures: structural response to loss estimation. Earthquake Engineering and Structural Dynamics
2010; 39(6):591–613. DOI:10.1002/eqe.955.
12. Kennedy RP, Short SA, Merz KL, Tokarz FJ, Idriss IM, Power MS, Sadigh K. Engineering characterization of
ground motion. Task I: Effects of characteristics of free-field motion on structural response. Report No.
NUREG/CR-3805, U.S. Nuclear Regulatory Commission, Washington, DC, 1984.
13. Vamvatsikos D, Cornell CA. Developing efficient scalar and vector intensity measures for IDA capacity estima-
tion by incorporating elastic spectral shape information. Earthquake Engineering and Structural Dynamics 2005;
34(13):1573–1600. DOI:10.1002/eqe.496.
14. Bianchini M, Diotallevi PP, Baker JW. Prediction of inelastic structural response using an average of spectral accel-
erations. 10th International Conference on Structural Safety and Reliability (ICOSSAR09), Osaka, Japan, 2009.
15. Bojórquez E, Iervolino I. Spectral shape proxies and nonlinear structural response. Soil Dynamics and Earthquake
Engineering 2011; 31(7):996–1008. DOI:10.1016/j.soildyn.2011.03.006.

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe
 AVERAGE SPECTRAL ACCELERATION AS AN IM FOR COLLAPSE RISK ASSESSMENT 2073

16. Tsantaki S, Jäger C, Adam C. Improved seismic collapse prediction of inelastic simple systems vulnerable to the
P-delta effect based on average spectral acceleration. 15th World Conference on Earthquake Engineering, Lisbon,
Portugal, 2012.
17. De Biasio M, Grange S, Dufour F, Allain F, Petre-Lazar I. A simple and efficient intensity measure to account for
nonlinear structural behavior. Earthquake Spectra 2014; in press. DOI: 10.1193/010614EQS006M.
18. Shome N, Cornell CA. Probabilistic seismic demand analysis of nonlinear structures. Report No. RMS-35, RMS Pro-
gram, Stanford University, Stanford, CA, 1999.
19. Modica A, Stafford PJ. Vector fragility surfaces for reinforced concrete frames in Europe. Bulletin of Earthquake En-
gineering 2014; 12(4):1725–1753. DOI:10.1007/s10518-013-9571-z.
20. Eads L, Miranda E, Lignos DG. Seismic collapse risk assessment of buildings: effects of intensity measure se-
lection and computational approach. Report No. 184, The John A. Blume Earthquake Engineering Center,
Stanford University, Stanford, CA, 2014.
21. Zareian F, Krawinkler H. Simplified performance-based earthquake engineering. Report No. 169, The John A. Blume
Earthquake Engineering Center, Stanford University, Stanford, CA, 2009.
22. Haselton CB, Deierlein GG. Assessing seismic collapse safety of modern reinforced concrete moment frame build-
ings. Report No. 156, The John A. Blume Earthquake Engineering Center, Stanford University, Stanford, CA, 2007.
23. Ibarra LF, Medina RA, Krawinkler H. Hysteretic models that incorporate strength and stiffness deterioration. Earth-
quake Engineering and Structural Dynamics 2005; 34(12):1489–1511. DOI:10.1002/eqe.495.
24. Chiou B, Darragh R, Gregor N, Silva W. NGA project strong-motion database. Earthquake Spectra 2008;
24(1):23–44. DOI:10.1193/1.2894831.
25. Vamvatsikos D, Cornell CA. Incremental dynamic analysis. Earthquake Engineering and Structural Dynamics 2002;
31(3):491–514. DOI:10.1002/eqe.141.
26. Boore DM, Atkinson GM. Ground-motion prediction equations for the average horizontal component of PGA, PGV,
and 5%-damped PSA at spectral periods between 0.01s and 10.0s. Earthquake Spectra 2008; 24(1):99–138.
DOI:10.1193/1.2830434.
27. Abrahamson N, Atkinson G, Boore D, Bozorgnia Y, Campbell K, Chiou B, Idriss IM, Silva W, Youngs R. Compar-
isons of the NGA ground-motion relations. Earthquake Spectra 2008; 24(1):45–66. DOI:10.1193/1.2924363.
28. Haselton CB, Baker JW, Liel AB, Deierlein GG. Accounting for ground-motion spectral shape characteristics
in structural collapse assessment through an adjustment for epsilon. Journal of Structural Engineering 2011;
137(3):332–344. DOI:10.1061/(ASCE)ST.1943-541X.0000103.
29. ICC. International building code. International Code Council (ICC): Falls Church, VA, 2003.
30. AISC. Prequalified connections for special and intermediate steel moment frames for seismic applications, AISC
358–05. American Institute of Steel Construction (AISC): Chicago, IL, 2005.
31. AISC. Seismic provisions for structural steel buildings, including supplement no. 1, AISC 341–05. American Insti-
tute of Steel Construction (AISC): Chicago, IL, 2005.
32. Lignos DG, Krawinkler H. Sidesway collapse of deteriorating structural systems under seismic excitations. Report
No. 177, The John A. Blume Earthquake Engineering Center, Stanford University, Stanford, CA, 2012.
33. Open system for earthquake engineering simulation (OpenSees). Version 2.1.1. Pacific Earthquake Engineering Re-
search Center, 2009. Available from: http://opensees.berkeley.edu [2 May 2014].
34. Zareian F, Medina RA. A practical method for proper modeling of structural damping in inelastic plane structural
systems. Computers and Structures 2010; 88(1–2):45–53. DOI:10.1016/j.compstruc.2009.08.001.
35. Baker JW. An introduction to probabilistic seismic hazard analysis (PSHA). White Paper, Version 1.3, 2008. Avail-
able from: http://www.stanford.edu/~bakerjw/publications.html [2 May 2014].
36. Shahi SK, Baker JW. An empirically calibrated framework for including the effects of near-fault directivity in prob-
abilistic seismic hazard analysis. Bulletin of the Seismological Society of America 2011; 101(2):742–755.
DOI:10.1785/0120100090.
37. Chioccarelli E, Iervolino I. Near-source seismic hazard and design scenarios. Earthquake Engineering and Structural
Dynamics 2013; 42(4):603–622. DOI:10.1002/eqe.2232.
38. Jayaram N, Baker JW. Statistical tests of the joint distribution of spectral acceleration values. Bulletin of the Seismo-
logical Society of America 2008; 98(5):2231–2243. DOI:10.1785/0120070208.
39. Baker JW, Jayaram N. Correlation of spectral acceleration values from NGA ground motion models. Earthquake
Spectra 2008; 24(1):299–317. DOI:10.1193/1.2857544.
40. Benjamin JR, Cornell CA. Probability, Statistics, and Decision for Civil Engineers. McGraw-Hill: New York, NY,
1970.
41. Baker JW, Lin T, Shahi SK, Jayaram N. New ground motion selection procedures and selected motions for the PEER
transportation research program. PEER Report 2011/03, Pacific Earthquake Engineering Research (PEER) Center,
Berkeley, CA, 2011.
42. Chandramohan R, Lin T, Baker JW, Deierlein GG. Influence of ground motion spectral shape and duration on seis-
mic collapse risk. 10th International Conference on Urban Earthquake Engineering, Tokyo, Japan, 2013.
43. Medina RA, Krawinkler H. Seismic demands for nondeteriorating frame structures and their dependence on ground
motions. Report No. 144, The John A. Blume Earthquake Engineering Center, Stanford University, Stanford, CA, 2003.

Copyright © 2015 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2015; 44:2057–2073
DOI: 10.1002/eqe