Prof. C.

Allin Cornell
Professor Emeritus
Stanford University
110 Coquito Way
Portola Valley, Ca, US, 94028
[email protected]

Probabilistic methods in structural engineering

Luis Esteva has been one of my two most knowledgeable and influential tutors in earthquake
engineering and probability. In the summer of 1966 Luis and I were both working at UNAM
beside my second such tutor, Emilio Rosenblueth. But Luis was the one who was then also
deeply interested in the questions of earthquake occurrence and ground motion prediction, and
their representation for engineering hazard assessment purposes. My own work in that field owes
a great deal to the openness and patience with which Luis discussed his ongoing studies. His
example has also set a standard for me with respect to the importance of keeping our sometimes
seemingly esoteric work in probabilistic methods grounded in the needs of engineering practice.
The two of us have since shared many concentrated technical and pleasant social hours not only
at UNAM and at MIT and Stanford, where I sought his multi-month visits, but also in the many
cities around the world that this profession takes you. For all that, Luis Esteva, I thank you and
salute you on this special day in your career.

Allin Cornell,
August 10, 2005
ON EARTHQUAKE RECORD SELECTION FOR NONLINEAR DYNAMIC ANALYSIS

ABSTRACT

The apparently simple act of selecting accelerograms for use in conducting

nonlinear dynamic analysis has been the subject recent study (e.g., Baker 2005c)
and conclusions are emerging. This paper attempts to elucidate those conclusions
by starting from very ideal and simplified cases to deduce observations that are
supported by more through numerical studies of nonlinear MDOF structural
models. A direct and an indirect approach are discussed. To insure both lack of
bias and adequate confidence limits in the estimation of response and response
likelihoods, the selection of the spectral shape and number of the records is found
to be dependent on the site, the structure and ground motion level. With proper
choice of the spectral shape scaling of the records should not cause significant
bias.

Introduction

It might be said that both Luis Esteva’s career and my own have been highly focused on a
single problem: estimating the annual frequency, λ, that an earthquake induces in a particular
structure some specified behavior state. We have worked separately, together and with many
colleagues and students on pieces of and the whole of this issue. There is a half century of
literature on this subject with many approaches, simplifications, and results. Yet many of us are
still working hard on the subject. Modern computational resources have accelerated the
progress, opening opportunities, for example, for large statistical samples of nonlinear dynamic
analyses of multi-degree-of-freedom models of structures both for research and for applications.
But the increased computational power is also being used by the structural modelers to improve
the detail and accuracy of their FEM models. This verisimilitude is especially important as we
attempt to push our studies to the extreme damage and collapse domains of behavior important to
life safety assessment. Therefore there will always be a demand to limit the number of dynamic
analyses used to estimate λ. Even when, as in research contexts, computational limits may be
less restrictive than in practice, there remain open many questions about which accelerograms to
run in any particular situation. These questions arise both in practice and research.

Current best record-selection practice even in research (e.g., ASCE 2005, Stewart 2002) has the
seismologist providing of the order of 10 or fewer records that represent magnitudes and
distances identified by probabilistic seismic hazard analysis (PSHA) disaggregation (e.g.,
Bazzurro 1999) to be the most likely to have caused the event that a particular response spectral
ordinate equals the level associated with a particular mean annual rate of exceedance. These
records are then scaled as necessary to match the level of the uniform hazard spectrum in one
manner or another. An important implication of this practice is that the result is dependent on
the character of the seismicity that surrounds the site and the mean return period of interest. It is
implicit too that it believed that the magnitude and distance of a record do or may affect the
structural response. It is also clear that nothing beyond single-degree-of-freedom (SDOF) linear
structural behavior is used in the selection. While many questions surround various elements of
this practice (e.g., the number of records, the impact of the scaling, etc.), especially for
application to rare, severe response of nonlinear multi-degree-of-freedom (MDOF) structures, it
has only recently begun to be investigated for accuracy and effectiveness.

The objective of this paper is to back up briefly to look at the problem from a simple but perhaps
fundamental perspective that may give us insights into what to look for and what to avoid in
selecting accelerograms for nonlinear time history analysis. This may also suggest some of the
compromises we are making in current practice and research.

With these objectives I shall be making, for the purpose of clarity of exposition, various
specializations and restrictions and simplifications, most of which could be expanded or
generalized without significant effort. For example I take it as a given that the fundamental
objective is the estimation of λ. Note that special cases include, for example, the annual
frequency (or, approximately, probability) of collapse, of roof drift angle greater than 3%, of
maximum interstory drift in the first floor greater than its (random) capacity, etc. Such limit state
frequencies are at the basis modern approaches to earthquake codes, guidelines, advanced
practice, and performance-based earthquake engineering. They are the first step toward
estimation of more direct decision variables involving consequences such as lives lost, economic
damage to structure or non-structural elements, and lost occupancy time. Other limitations in the
paper will be simple site hazard and structural representations.

Direct Approaches

In the best of circumstances λ would be estimated by monitoring the response of the

building itself for a sufficient number of years to estimate λ with the desired accuracy. Suppose
for simplicity that single response variable, θ, is of concern (say the maximum interstory drift
ratio - MIDR - in a frame). Then, just as with empirical flood frequency analysis, we would
need only order and plot these values versus i/n (where n is the number of years of observation)
as an empirical complementary cumulative mean annual rate of exceedance function (CCDF), λ θ
(x). The value of x at λ θ (x) = 0.001 would be the estimated 1000 year mean return period
MIDR. Even, however, if the structure had been built and monitored, for safety level λθ’s of the
order of 10-3 or less, this would require 10,000 or more years of data1.

Somewhat more realistically the response will be estimated from linear/nonlinear dynamic
analysis of a numerical model of the structure, be it an existing building or a proposed design.
(We assume here that such numerical models are precise.) Then we would need “only” to have
had an accelerometer in place at the site for these 10,000 years. All the records (above some

1
The basis for this statement is that simple binomial trials statistics suggests it requires a sample size of about 10/p
to estimate small probability p with a standard error of about +/- 30% of p. To reduce this number to 15% would
require 4 times as many years.
practical threshold, say 1000 in total) would need to be run and their resulting values of θ plotted
as above. While this case remains completely unrealistic it is an excellent hypothetical model to
repeatedly return to as we consider how to address practical record selection. This record set of
some 1000 records would implicitly contain records generated by a multitude of magnitudes
from different sources at various distances and azimuths over various travel paths, and this set
would fully and properly represent all the characteristics of ground motion that might possibly
affect the structure’s response. What is more, all these characteristics, including site effects, all
appear in this large hypothetical set of in situ recordings in exactly the correct relative frequency
of occurrence, both marginally and jointly. Our objective in practical record selection is to
reproduce this condition or, rather, to approximate it as best we can.

Lacking this ideal set of records at our site we turn either to the catalogue of recordings or to
simulation. Not wishing in this short note to take on the questioning of just how realistic various
current modes of simulating accelerograms are, I limit myself to the catalogue2. I further
presume that all these recordings are free from instrumentation limitations. Many good
catalogues (or virtual catalogues) are readily available today with thousands of records (e.g.,
PEER 2005). Going to the catalogue of accelerograms recorded elsewhere to represent what has
happened at my single site over years is an example of “trading space for time”. It carries with it
the need to try to understand what is important to the problem at hand to gain confidence that the
trade has been fair and the conclusions accurate. So we must ask which records in the catalogue
with which characteristics are “right” for my site and in what proportion should select them.

Suppose first, again for simplicity of exposition, that the threat at our site is a single fault
segment, located R kilometers from the site3, which produces only “characteristic” events, i.e.,
full segment ruptures with very similar magnitudes4, M, and which does so in a Poisson way
with known5 mean rate λo. Then, hoping/assuming (again for simplicity, but close to current
practice) that M and R are sufficient representations of the source and path and that, say, “firm
soil” (or some like category among the current catalogue representations) is a sufficient
characterization of our site, we might logically sort the catalog for all such records and run them
through the numerical analysis to obtain values of θ. Ordering and plotting them versus i/m
(where m is the number of records found) would produce what we would hope to be a reasonable
and accurate estimate of G θ (x), the complementary cumulative distribution function (CCDF) of
θ given an event on the fault. With this result it follows that λ θ(x) = λ’ P[θ>x| event] = λ‘ G θ
(x). How much data does this require? Suppose, as in coastal California, λ’ is 1/ (several
hundred years), then for λ θ(x) of the order of 10-3 to 10-4 we need G θ (x) to be of the order of
0.1. Reasonably confident estimation requires m to be about6 100. Apart from the
computational cost, there, of course, are not nearly enough such specific (M, R, firm soil)
records in current or foreseeable catalogues (especially as one would like them to be from 100

2
For approaches using simulated records see, for example, Han 1997, Luco 2002 or Jalayer 2005.
3
Measured in some relevant way, e.g., as closest distance to the rupture surface.
4
A more fundamental approach might be to use rupture length rather than the more heavily processed M to
characterize events and select records.
5
Again we assume that this is known accurately. Indeed we shall assume throughout this and all such seismicity
information to follow is perfectly known.
6
Making “a distribution assumption”, i.e., fitting this data to a named probability law such as one with an
exponential or power form, and estimating its parameter values can reduce this number by a factor of 2 to 3 at the
risk of inducing errors in the upper tail by having made the wrong parametric modeling assumption.
independent events). Therefore we must relax the constraints and accept events within an
interval7 M+/-∆M and R+/- ∆R. Immediate questions are: how wide need these bins be to gather
an adequate sample size? How much accuracy in θ is given up by using the “wrong” M and R,
and how does that increase as say ∆M gets larger? And are there ways to modify the records to
make them “more nearly right”? The simplest, common illustration of this is, say, scaling up by
some amount the accelerograms that are from “too small” M’s or “too large” R’s. And does such
scaling induce biases of it own? Before addressing such questions, let us release at least one of
the previous unrealistic limitations.

We need to recognize that the assumption of a simple single {M, R} scenario is not only
unrealistic but impacts our record selection discussion. Even with a single dominant neighboring
fault there may be lesser magnitudes at various locations (and distances) that contribute to the
likelihood of exceeding any θ level, x. Further there are inevitably other seismic sources that
also contribute to the hazard. These additional scenarios complicate record selection to the
degree that they needed to be represented in the analysis. For such cases we need write:

λθ ( x) = ∑ Gθ |M , R ( x | mi , ri ) λ (mi , ri ) (1)
i
in which the individual mean rates of occurrence, λ, and conditional CCDF’s, G, are identified
for each of the interesting (here discretized) set of {M,R} scenarios. Records would have to be
selected as above for each such scenario. Even if there are only 5 to 10 such scenarios the
number of records and analyses needed may mount once again into the 1000 range.

Intensity Measure-based Approaches: In Principle

To address these challenges the seismic community has used for decades the notion of what
some of us now call an “intensity measure” (IM); examples include PGA and first-mode-period
spectral acceleration, Sa(T1), to serve as an intermediate variable in such assessments. The
estimation of the mean rate of exceedance of the IM is the subject of PSHA and the
responsibility of earth scientists not structural analysts. The equation for λ IM(y) looks just like
Eq. 1 with Gθ|M,R replaced by GIM|M,R. The last function is obtained from standard strong ground
motion attenuation “laws”. This PSHA problem is again outside the scope of this discussion; we
assume that λ IM(y) is available for almost any IM we might chose. It is important to recognize
that this a site-specific product reflecting all the {M,R} scenarios. Therefore it has captured a
major portion of the specific nature of our site. In particular it has released us from the need to
have 10,000 years of recordings at the site in order to measure how frequently important
magnitudes occur at critical distances from the site, and how certain measures (IM’S) of the
amplitudes of their motions depend on magnitude and attenuate with distance. How completely
and adequately this job has been completed for the objective of structure-specific λ θ(x)
estimation remains to be discussed.

7
Let us drop further explicit discussion of soil conditions. This is not to say that this is not a very important issue.
Some portion of the variability in records from similar but different sites would presumably not be found from event
to event at the same site. This question is related to the current discussion ergodicity (Anderson 1999)
Returning to the estimation of λ θ(x), it is clear that IM’s can likely play a key role. Let us see
how. The total probability theorem (e.g., Benjamin 1970) always permits an expansion of λ θ(x)
into (e.g., Bazzurro 1998).

λθ ( x) = ∫ Gθ |IM ( x | y ) | dλ IM ( y ) | (2)

in which the last factor is the absolute value of the derivative of λ IM(y), or, loosely, the mean
rate of occurrence of a value of the IM equal to y. Having, as we do, λ IM(y), our problem is
transformed into estimating Gθ|IM, the conditional probability that θ > x given IM = y. Consider
first our original ideal case where we have some 10,000 years of recordings directly from our
site. For a given level of y, we would select a random sample from those records that have this
(or approximately) this IM level, analyze the structure under them, and process as usual to
estimate Gθ|IM versus x for this y level.

How many records would this require? Suppose we are interested again in about 10-3 to 10-4.
Then we should look at levels of y such that λ IM(y) is in this range as well, since it is well
known that λ IM(y) not Gθ|IM dominates the integral in Eq. 2. Indeed this is so true that generally8
the estimation of Gθ|IM can be reduced to estimating well the mean of θ and only roughly its
standard deviation and distribution shape. Accepting this statement as at least a rough
approximation we can estimate the required sample size to be of order9 only 10. This enormous
reduction in the necessary sample size is possible because the variability in θ given IM is small
relative that in θ marginally. This reduction is the consequence of the PSHA analysis of the IM
“absorbing” most of the total variability in θ, leaving only the conditional variability of θ given
IM to be estimated from the structural dynamic analyses. Under some practical circumstances
only one such level (or “stripe”) may be adequate, in particular if there is interest in only a single
level of x ( e.g., 3% drift) or of mean frequency (e.g., 2% in 50 years); in this case this an
extremely attractive approach to the record selector and structural analyst. Unlike the previous
approach this number is independent of the number of sources contributing to the hazard.
However more than one level of y may be necessary; 3 to 5 or more levels may be required if λ
10
θ(x) is needed for a broad range of x values (Jalayer 2003), raising the required sample size to
order 100. In any case this IM-based approach is in principle very efficient11 and accurate, at
least in the ideal case when the large site-specific sample is available. Even with this nominal
perfect sample there remain interesting questions as to what variable is best to use as the IM. For
MIDR prediction, PGA for example leads to larger required sample sizes than Sa(T1) simply
because it is less well correlated with MIDR for moderate to long period structures. We shall not
pursue this question of which IM to use here except (later) to the degree it impacts our focus:
record selection.

8
Exceptions include case where the IM hazard curve is very steep and the dispersion of θ given IM is large in the
region of interest.
9
The basis for this is that the log of λ θ (x) is roughly proportional 2 to 3 times the mean of the log of x. Therefore
the +/-30% standard error in λ θ mentioned above requires about a 10% standard error in the mean of x. This in turn
requires a sample size equal to the square of the coefficient of variation (COV) of x divided by this 0.1. The COV
of say MIDR of an MDOF frame structure given a reasonable choice of IM (such as Sa(T1)) is less than 0.3 to 0.5
plus for very severe degrees of nonlinearity.
10
Especially near global collapse when the dependence on y may be very nonlinear and the dispersion broad.
11
It can be interpreted as an application of conditional Monte Carlo analysis.
 Intensity Measure-based Approaches: In Practice
Now we must return to the real world and ask how records are to be selected in the IM-base
approach when we must depend not on a ideal site-specific catalog but on the existing catalog of
recordings. This question is commonly asked and answered in practice but rarely in a very
formal way. Given the discussion above it should not be a surprise that this is not a trivial formal
question. There is no error in the application in the formulation of Eq. 2 and we have assumed
that we have an accurate site-specific assessment of λ IM(y). Therefore the problem reduces to
how one should select records to estimate Gθ|IM. The simple answer is we chose them so that we
get the right answer. This is more complex than simply what is an adequate sample size because
there is now the question of suitability of records recorded elsewhere to this site. The contention
is that this question can be addressed only if one considers carefully the structure of concern (and
the IM at hand). Let us consider two simple structures to get a sense of what is involved.

Single DOF Linear Structure

Suppose to begin that our structure is simply a linear oscillator, and that we have selected the IM
to be the common one: Sa(T1) where T1 is the natural period of the structure. Our objective is to
select records from an available catalog to estimate Gθ|IM (x|y) accurately. In this case record
selection for dynamic analysis purposes is clearly trivial. One can select any record (no matter,
for example, what its M, R, or relative strength), scale it such that its IM = y, run the dynamic
analysis and get, of course12, θ = y because in this case IM and θ are the same. Further as all
records will produce the same value, only a single record is necessary for perfect accuracy. In
addition the same record can be used perfectly accurately for all levels, y, of the IM, only scaling
is required. This absolute robustness with respect to record selection and scaling and this low
dispersion (with its implied computational efficiency) are all a direct result of the selection of the
IM, and the simple structure. Had the IM been another popular choice, PGA, none of these
conclusions would hold. Because many structures are known to be “first-mode-dominant” it
follows further that these desirable properties of Sa(T1) as an IM are likely to hold to some
degree for many real structures, at least in the linear range. Therefore, it can be anticipated that
many concerns about record selection for such cases are unwarranted. Dynamic analysis of linear
first-mode-dominated structures is seldom needed, however. Nonetheless it suggests that this IM
is a natural starting point from which to look for improvements for more complex structures.
Certainly those candidates for IM, such as PGA, that create record selection sensitivity for this
simplest structural problem are not strong initial choices for the more realistic structural
problems.

Two DOF Linear Structure

Suppose next that our structure is, say, a two-story frame that can be represented accurately by a

12
Provided, as I assume here for argument’s sake, our structure has 5% damping as is standard for the attenuation
laws used in PSHA. If our structure has a different damping level there will be an offset dependent on average on
the degree to which the two damping levels differ and there will be some comparatively mild dispersion from record
to record. There is no evidence that this effect is dependent on M or R or other such record properties.
2-DOF linear model. Further continue to assume that the IM is still Sa(T1), where T1 is the first-
mode period. How now should the catalog be searched for records for dynamic analyses to be
used? For the purposes of illustrating the principles, let us further assume that in fact the simple
square-root-of-sums-of-squares (SRSS) approximation is exact13. In this case given that IM =
Sa(T1) = y, the response of interest can be written:

θ = c12 y 2 + c 22 S a2 (T2 ) (3)

in which c1 and c2 are coefficients depending on the dynamic properties of the structure. This
form makes it clear that (given Sa(T1)) θ is simply a function of the random variable Sa(T2),
where the probability distribution of Sa(T2) is the conditional distribution of Sa(T2) given Sa(T1).
We shall take advantage of this below. One can also re-write Eq. 3 as

c 22 S a2 (T2 )
θ = c1 y 1 + (4)
c12 y 2

In this form it is clear, first, that our concern is only with cases in which the second mode makes
a comparatively strong contribution to θ, as indicated by the ratio under the radical, and, second,
that θ is simply a function of the (random) spectral ratio R2/1 = Sa(T2)/ Sa(T1) - still
conditioned, of course, on Sa(T1) = y. The former observation implies that the record selection is
trivial and robust in the first-mode dominated case, as discussed above. The latter observation
emphasizes that once the IM level is given it is the only relative value, Sa(T2)/y, (or spectral
shape) that matters. This notion carries over to other structures as well. Knowledge of this fact
supports the practice of selecting records from {M, R} bins that dominate the site hazard,
because M is known to have some effect (in the mean at least) on spectral shape. The fact is
even more evident in the common practice of first selecting records whose spectral shapes
closely match that of the UHS or of the median spectrum given the dominant M and R, and then
scaling them to match the level of the target spectrum. Neither of these practices, however,
reflects the conditional nature of this dependence. We address this issue this next.

To estimate well Gθ|IM(x|y) when the second mode is important we clearly need to select records
from the catalog that capture accurately the conditional distribution of Sa(T2) given Sa(T1) = y (or
equivalently of R2/1 given Sa(T1) = y). This distribution is not readily available today14. But we
can get guidance as follows.

In order to understand better how in principle record selection should proceed in this case, we
shall once again simplify by assuming, for the moment, that a single M = m and R = r scenario
dominates our site’s hazard. Conventional attenuation laws are based on the fact any Sa(T) value
has a lognormal distribution with the mean of the natural log, call it µlnS, equal to specified
function of m and r and with standard deviation of the log, σlnS, equal to another such function. It
is reasonable to assume further that the spectral accelerations at two different periods are jointly

13
Again one might ask: then why is dynamic analysis necessary? It is not, of course, but I again appeal to the sake
of the argument.
14
It happens that they can be found from disaggregation of vector-valued PSHA (Bazzurro 2002), public tools for
which are under development (Somerville 2005).
lognormal with correlation coefficient15 of the logs ρ. In this case the conditional mean of log
Sa(T2) given Sa(T1) = y is:

σ ln S (T )
E[ln(S a (T2 ) | S1 = y ] = µ ln Sa (T2 ) + ρ a 2
(ln y − µ ln S ) (5)
σ ln S (T )
a 1
a (T1 )

The median, η, of Sa(T2) given Sa(T1) = y is e raised to this power or:

ηS a ( T2 ) | S1 (T1 ) = y
= η S a (T2 ) exp[ ρ σ ln S a (T2 ) ε S a (T1 ) ] (6)

in which ε S a (T1 ) = (ln y − µ ln S a (T1 ) ) / σ ln S a (T1 ) is the deviation (in log terms) of Sa(T1) away from its
expected value (given M =m and R=r). The conditional median spectral ratio or shape is Eq. 6
divided by y or

ηS a (T2 )
σ ln S a (T2 )
ηR = exp[−ε S a (T1 )σ ln S a (T1 ) (1 − ρ
)] (7)
η S (T )
2 /1
a 1
σ ln S (T ) a 1

Note that this conditional shape is the marginal median shape times an exponential that is
negative and proportional to ε and (1-ρ) under the reasonable assumption that the two σ’s are
about equal. For the larger values of y of most engineering interest ε is positive and the
conditional median shape at other periods is below the one we expected “on average”. The
degree to which it is below depends on how far the selected level y of Sa(T1) is above its median
and how weakly correlated the two spectral ordinates are. This correlation decays with
separation between the two periods16. The implication is that the conditional median spectral
shapes of interest are somewhat peaked around T1 relative to the median shape. This in turn
means that if the response of the first mode is higher than expected the response of the second
mode will be relatively less than one would expect otherwise. These conclusions have been
verified empirically (Baker 2005c). Fig. 1 shows plots of predicted spectra, including one
median spectrum and two that are conditional on Sa(0.8s) being at the 1.5 ε level. It also shows
these three spectra scaled to a common value of Sa(0.8s). Note that the two positive ε spectra are
more peaked than the median spectrum and have very similar shapes despite the half-unit
magnitude difference in their causative earthquake.

If the structure has an important second-mode contribution to θ, a first-order requirement

towards estimating well Gθ|IM(x|y) would be to select records that have approximately this
median shape value. Note that for higher values of y that this shape is not the median shape.
Therefore the practices of using the UHS shape or the median shape of the dominant M and R
scenario17 are not as accurate as they might be for the rare events of common engineering
15
A recent study of this correlation coefficient is to be found in (Baker 2005b). It depends primarily on the ratio
T1/T2 and is effectively independent of magnitude and distance.
16
For separation of modal periods by a factor of 3 the value of ρ might be about 0.6 and σln Sa about 0.7 implying for
ε equal 1 the shape is about 25% lower and for ε equal 2 the shape is 40% lower. The effect on θ will depend on the
importance of the second mode to the response quantity in question.
17
For our special case here that the seismic threat is limited to single {M,R} scenario, the UHS shape and the
median shape (given these M and R values) are virtually identical (differing only to the degree that the log standard
deviations are different at different periods).
interest.

A direct way to select the records in this simple linear 2-DOF case is to calculate the spectral
ratio at T1 and T2 for all the records and select records such that their median is about that given
in Eq. 7. Then they may be scaled such that Sa(T1) equals level y and the dynamic analyses run
to find Gθ|IM(x|y). Note in particular that for this linear 2-DOF case the records selected will
need to change as y changes, that the M and R of the records are immaterial per se (only their
spectral ratio matters), and that the selected records may be scaled to any degree without loss of
accuracy.
Even in the linear 2-DOF case we have made simplifications. In the record selection we have
tried only to match the conditional central value not the variability of the spectral ratio. This is
also quite feasible but will not be pursued further here. Further if we do not have a single {M,R}
pair the determination of this conditional median shape is not trivial. In principle there must be a
weighting over all {mi,ri} pairs18. It may be sufficiently accurate in some cases to use
disaggregation to determine a single dominant {M,R}and then apply the reasoning above as if it
were the only threat. This would already be a step beyond current practice.

A second more indirect way to select the records for this case is to chose records (from the
general magnitude range; R is less critical as it has little mean effect on strong motion spectral
shape) that have the required level of ε (consistent with the level y), and then scale them to the
correct level of Sa(T1). This too should on average at least capture the more peaked spectral
shape associated with the higher levels of y and ε of primary interest.

More General Structures

While the conclusions from the simple 1 and 2-DOF linear structure are in principle limited to
these simple cases, they suggest several generalizations that can be made, which have been
verified in recent studies with nonlinear SDOF and MDOF frame models (e.g., Iervolino 2005,
Baker 2005c, Tothong 2005). The objectives in record selection and scaling for accurate
estimation of Gθ|IM(x|y) are to capture primarily the proper general amplitude (via the IM level)
and secondarily the spectral shape given that IM level. For the common IM Sa(T1) once the first
objective is met the estimation of Gθ|IM(x|y) is fairly robust with respect to the records selected as
long as the structure is first-mode dominated and only moderately nonlinear.

For other structures it can be important to select the records to capture the appropriate spectral
shape. This was demonstrated for higher modes (T2/T1 < 1), but it is clear from our
understanding of nonlinear dynamic behavior that it is equally important to reflect properly the
longer periods when the structure experiences substantial “softening”. While there are no direct
ways (similar to that used above for the 2-DOF case) to identify one or a few unique longer
periods to focus upon and analyze, it should be clear that an objective of matching the
conditional median spectral ratio would apply to each period of interest. Hence it follows that
the entire conditional median spectral shape is logical first-order target for record selection for
all structures. It should be re-emphasized that this shape is not the same as the median or UHS
shape unless the level of y is near the median value of the IM Sa(T1), and that this shape will
change as y changes, being most different from the median shape for large, rare values of the IM,
18
As mentioned in footnote 14 such information will become available in time.
which is when strong degrees of nonlinearity may occur. It is the author’s belief, while not
proven here, that for more general structures (as was shown here for the 1 and 2-DOF linear
structures) it is not critical to capture the “causative” magnitude19 if the spectral shape itself has
been selected well. Magnitude is primarily just a proxy for the median shape.

We saw above that arbitrary levels of scaling of the records was not a cause of response bias in
the 1 and 2-DOF linear structures, provided (in the 2-DOF case) that the spectral shape was
correct. It is the author’s experience that, with this same proviso, this conclusion is more
generally true (e.g., Shome 1998, Baker 2005c).

It should be noted that the schemes discussed above are based on the common use of linear first-
mode period spectral acceleration as the IM. It has been found that other choices of the IM may
provide even more robustness with respect to record selection (Luco 2002, Luco 2005, and
Tothong 2005), just as Sa(T1) provides more such robustness than PGA. These new IM’s are
based on inelastic spectral acceleration. Another major objective of seeking improved IM’s is the
reduction of the number of records and analyses needed to achieve a specified level of
confidence in the estimate (recall it was set here as a standard error of estimation of λθ(x) of
about +/- 30%, or of the conditional mean of θ of about +/- 10% ). This is achieved by finding
“better response predictors”, i.e., IM’s that reduce the variance of θ given IM = y for various
levels of y (i.e., degrees of structural nonlinearity). This so-called IM efficiency (Luco 2005)
issue has not been addressed here.

Further this study of the record selection problem presumes that the model of the structure is
available at least to the level of knowing the general range of its first-mode period. (Because of
the high correlation between two comparatively nearby periods there is little loss of accuracy or
efficiency if the period T of the Sa(T) used as the IM is some distance from that of the final first-
natural period of the structural model. (Of course whether that model estimates well the first-
natural period of the real structure is another problem, which does not influence how we should
best analyze the model we have.) While the general principles above hold for all cases the
judgments stated as to accuracy and effectiveness depend on the author’s experience with
building-like structures, which at least in the linear range tend to have no more than two or three
most-important elastic modes. Other cases, including those in which the record selection is done
(for good or bad reasons) without knowledge of the structure or those where there are many very
important response measures sensitive individually to different portions of the input spectrum,
have not as yet been studied in this way. Short of using different IM’s, records, and analyses for
the different subsets of these cases (a strategy which is permitted in the nuclear arena, e.g, NRC
1997), it is clear that some compromise well have to be made with respect both to efficiency
(implying larger sample sizes or larger standard errors) and perhaps to the accuracy (or record
selection robustness).

19
Exceptions may be when the response is duration sensitive and magnitude serves as a proxy for duration. The
peak displacements of nonlinear framed structures do not seem to be duration sensitive even when strength
degradation is involved.
 Conclusions
We conclude that the selection of records for use in nonlinear seismic time history analysis of
MDOF models of structures can benefit from starting from a defined structural objective, here
estimation of the mean annual frequency of some structural response measure, θ, exceeding level
x, i.e., λθ(x), and then asking how that might be directly and/or indirectly estimated in various
ideal and simplified cases. Discussion of the ideal case in which 10,000+ years of recordings
have been made at the site reveals that numerous (order 1000) records and time history analyses
will be required (save, perhaps, special techniques to reduce this number). Recognizing that the
recorded accelerograms must come from catalogs of data recorded at many sites and caused by
many sources, it becomes clear that record selection must allow in some way for the failure of
such data to reflect the relative frequency with which various events at various distances will
affect the site. In short in this form the record selection problem is very site-seismicity
dependent. This complicates the selection problem and increases the number of analyses
necessary for accurate estimation of λθ(x).

Turning to an intensity-measure-based formulation of the analysis and estimation problem, one

finds that the PSHA analysis leading to the IM hazard curve, λIM(y), captures much of the site-
specific seismicity issue and of the variability in θ. Now one needs to select records to estimate
Gθ|IM(x|y), the CCDF of the θ conditional on IM = y. Even though this CCDF may have to be
evaluated at several levels of y, this approach leads to significantly reduced record selection and
sample size needs. But again the ideal problem is compromised by the need to select records
from the available record set rather than a site-specific catalogue. Two, simplest structural
dynamic models, namely linear 1- and 2-DOF systems, are discussed with the objective of
understanding the benefits and challenges of the record selection and scaling problem for “real”
(nonlinear MDOF) structural models. The conclusions are that, while the sample sizes may be
smaller with the IM-based procedure, to the degree that the structure is not first-mode dominated
or not nearly linear, care may be needed in selection of the records (and particularly with respect
to their spectral shape), to insure that a bias is not introduced in the estimation of Gθ|IM(x|y) and
hence in λθ(x). In particular it appears that the spectral shape needs to reflect properly the level
of ε, which is a measure of the Sa(T1) IM level relative to its expected value (given the M and R
scenarios of primary interest at the site). This shape is, for the positive ε’s of engineering
interest, more peaked than the UHS or median spectral shape (given a dominant scenario {M,R}
pair). Provided this shape is captured it appears that the M and R selection criteria may be
significantly relaxed and that scaling records to match the level y of the IM will not induce
significant bias. These conclusions are supported by current research on nonlinear MDOF
frames (e.g., Baker 2005a, Baker 2005c, Iervolino 2005).

It is noted that, while the focus here has been on estimating λθ(x), the conclusions here apply to
“more deterministic” current-code-practice objectives, such as estimating the mean of θ “given
the 2% in 50 ground motion”. Since this concept in quotes exists only for a scalar ground motion
measure (and not, for example, for an entire spectrum), it can be taken here to mean the 2% in 50
Sa(T1) IM level. Setting y equal to that level y* for which λIM(y*) = 0.0004, this problem can be
stated as estimating the mean of the distribution Gθ|IM(x|y), i. e., the conditional mean of θ given
Sa(T1) = y*. It will be recalled that concerns about the required sample sizes and about biasing
that distribution were cast above in terms of that mean. More modern codes are based on a target
value of λθ(x) (rather than λIM(y)). Therefore the discussion above applies directly to them.
 Acknowledgements

This simple paper depends strongly on the far more elegant and complete PhD thesis work of Dr.
Jack Baker (Baker 2005c). I want to acknowledge his primary contributions, and the many good
discussions with him, Polsak Tothong, and Iunio Iervolino on this subject. This work was
supported by the Earthquake Engineering Research Centers Program of the National Science
Foundation, under Award Number EEC-9701568 through the Pacific Earthquake Engineering
Research Center (PEER). Any opinions, findings and conclusions or recommendations expressed
in this material are those of the authors and do not necessarily reflect those of the National
Science Foundation.

References

American Society of Civil Engineers (2005). Seismic design criteria for structures, systems, and
components in nuclear facilities, Structural Engineering Institute, Working Group for Seismic
Design Criteria for Nuclear Facilities, ASCE/SEI 43-05, Reston, Va., 81 pp.

Anderson, J. and J. N. Brune (1999). “Probabilistic seismic hazard analysis without the ergodic
assumption”, Seismological Research Letters, 10 (1) 19-28.

Baker, J. W. and C. A. Cornell (2005a). “A vector-valued ground motion intensity measure

consisting of spectral acceleration and epsilon”, Earthquake Engineering & Structural Dynamics,
34, 1193-1217.

Baker, J. W. and C. A. Cornell (2005b) ”Correlation of response spectral values for multi-
component ground motions”, Accepted for publication, Bulletin of the Seismological Society of
America.

Baker J.W. and C. A. Cornell (2005c). Vector-valued ground motion intensity measures for probabilistic
seismic demand analysis, John A. Blume Earthquake Engineering Center Report No. 150,
Stanford University, Stanford, CA.. http://blume.stanford.edu/Blume/Publications.htm.

Bazzurro, P., C. A. Cornell, N. Shome, and J. E. Carballo (1998). “Three Proposals for Characterizing
MDOF Nonlinear Seismic Response,” Journal of Structural Engineering, ASCE 124 (11), 1281-
1289.

Bazzurro, P., and C. A. Cornell (1999). “On Disaggregation of Seismic Hazard,” Bulletin of
Seismological Society of America, 89 (2), 501-520.

Bazzurro, P. and C. A. Cornell (2002). “Vector-Valued Probabilistic Seismic Hazard Analysis

(VPSHA),” Proceedings 7th U.S. National Conference on Earthquake Engineering, Boston, MA.

Benjamin J.R. and C. A. Cornell (1970). Probability, Statistics, and Decision for Civil Engineers,
McGraw-Hill, New York, 684 pp.

Han S.W. and Y. K. Wen (1997). Method of reliability-based seismic design: Equivalent nonlinear
systems, Journal of Structural Engineering 123 (3), 256-263.
Iervolino, I., and C. A. Cornell (2005). “Record selection for nonlinear seismic analysis of
structures”, Earthquake Spectra 21 (3), 685-713.

Jalayer F., (2003). Direct probabilistic seismic analysis: Implementing non-linear dynamic assessments,
Ph.D. Thesis, Stanford University, Stanford, CA, 244 pp. http://www.stanford.edu/group/rms/

Jalayer F. and C. A. Cornell (2003). A technical framework for probability-based demand and capacity
factor design (DCFD) seismic formats, Pacific Earthquake Engineering Research Center Report
PEER 2003-08, University of California at Berkeley, Berkeley, CA, 106 pp.

Jalayer F., J. L. Beck, and K. A. Porter (2004). Effects of ground motion uncertainty on predicting the
response of an existing RC frame structure, Proceedings 13th World Conference on Earthquake
Engineering, Vancouver, Canada.

Luco, N. and C. A.Cornell (2005). “Structure-specific scalar intensity measures for near-source and
ordinary earthquake ground motions,” Earthquake Spectra, Submitted.

Luco, N., M. Mai, C.A. Cornell, and G. Beroza (2002). “Probabilistic seismic demand analysis at a near-
fault site using ground motion simulations based on a kinematic source model,” Proceedings 7th
U.S. National Conference on Earthquake Engineering, Boston, MA.

Nuclear Regulatory Commision (1997). “Identification and characterization of seismic sources and
determination of safe shutdown earthquake ground motion”, Regulatory Guide 1.165.
www.nrc.gov/reading-rm/doc-collections/reg-guides/power-reactors/active/01-165/ (accessed
8/17/2005)

PEER strong motion database (2005). http://peer.berkeley.edu/smcat/ (accessed 6/29/2005).

Shome, N., C. A. Cornell, P. Bazzurro, and J. E. Carballo (1998). “Earthquakes, records and nonlinear
responses,” Earthquake Spectra 14 (3), 469-500.

Somerville P.G. and H. K. Thio (2005). Probabilistic vector-valued seismic hazard analysis for near-fault
ground motions, Southern California Earthquake Center Project Report (in preparation).

Stewart J.P., S. J. Chiou, J. D. Bray, R. W. Graves, P. G. Somerville, and N. A. Abrahamson (2002).

Ground motion evaluation procedures for performance-based design, Soil Dynamics and
Earthquake Engineering, 22 (9-12), 765-772.

Tothong P. and C. A. Cornell (2005). Near-fault ground motions for seismic demand analysis (manuscript
in preparation).
 1
10
Magnitude = 7.5, Epsilon = 2
Magnitude = 7.5, Epsilon = 0
Magnitude = 6.5, Epsilon = 2
Spectral Acceleration (g)

0
10

-1
10
-1 0
10 10
Period (s)

1
10
Magnitude = 7.5, Epsilon = 2
Magnitude = 7.5, Epsilon = 0
Magnitude = 6.5, Epsilon = 2
Spectral Acceleration (g)

0
10

-1
10
-1 0
10 10
Period (s)

Figure 1. (Upper) expected response spectra for three scenario events; (lower) expected
response spectra for three scenario events, scaled to have the same Sa(0.8s) value.
Source: Baker 2005a.