Bulletin of the Seismological Society of America, Vol. 104, No. 1, pp. 503–512, February 2014, doi: 10.

1785/0120130129

Issues and Approaches for Implementing Conditional

Mean Spectra in Practice
by Brian Carlton and Norman Abrahamson

Abstract Ground-motion selection for dynamic analysis of structures is often

based on the conditional mean spectrum (CMS). The CMS is a spectrum that matches
the uniform hazard spectrum (UHS) at a conditioning period but that also represents
the response from a single earthquake scenario more realistically than the UHS. This
paper discusses and offers solutions for several issues related to implementing CMS
in practice. The issues covered are the selection of the conditioning period, the number
of CMS necessary to cover the range of the UHS, how to broaden the CMS to match
the UHS with fewer conditioning periods, how to compute a CMS for several ground-
motion prediction equations that is consistent with probabilistic seismic-hazard analy-
sis, the application of correlation coefficients to hard-rock sites, and the robustness of
correlation coefficients.

Introduction Review of CMS

Hazard analysts often use the uniform hazard spectrum The CMS is an estimate of the response spectrum that
(UHS) to select suites of design ground motions. The UHS is has the same spectral acceleration value as the UHS at a se-
determined by conducting probabilistic seismic-hazard lected conditioning period, T
, but which also represents the
analysis (PSHA) to estimate the hazard at each period sepa- response spectrum during a single earthquake. The engineer
rately. Then, for a given hazard level (e.g., 1=2500 years), the can then use a suite of CMS as the design spectra for a dy-
spectral acceleration is determined for each period. The re- namic analysis. Baker (2011) provides a comprehensive
sulting spectrum tells the engineer that the chance of any one overview of how to determine CMS. We offer a brief review
of the spectral acceleration values being exceeded is the of the necessary steps below:
same. The UHS does not give any information about the 1. Estimate the hazard curve for the conditioning period T

chance of more than one spectral acceleration value being using PSHA.
exceeded during the same earthquake. The UHS is analogous 2. Deaggregate the hazard into magnitude M and distance R
to using an envelope of multiple earthquake spectra in a bins for T
at the hazard level of interest, and select a
deterministic analysis. In the same way, the UHS is not rep- representative controlling M and R scenarios.
resentative of the spectrum from any one ground motion just 3. Compute the controlling scenario spectrum mean natural
as the envelope of two spectra in a deterministic analysis is log value μT and natural log standard deviation σT at
not representative of the ground motion during a single earth- all periods T using a GMPE.
quake. In many cases, moderate earthquakes at short distan- 4. Using equation (1), calculate εT
, the number of stan-
ces control the response at short periods, and large far away dard deviations of difference between the mean natural
earthquakes control the response at long periods (Bommer log spectral value of the controlling scenario spectrum
et al., 2000). μT
 and the natural log of the UHS value SAUHS T
.
Baker and Cornell (2006) and Baker (2011) proposed a
εT
  SAUHS T
 − μT
=σT
: 1
method to reduce the UHS into realistic scenario spectra,
called the conditional mean spectrum (CMS). This paper
5. Calculate the mean epsilon, ε
T, at other periods based
begins with a review of CMS and introduces a technique on the correlation between epsilon at T
and epsilons at
to broaden the CMS to match the UHS with fewer condition- different periods ρT i ; T
.
ing periods. We then address the issue of how to include
more than one ground-motion prediction equation (GMPE) ε
T i   εT
 × ρT i ; T
: 2
when developing CMS, followed by a discussion on the
correlation coefficients necessary to develop CMS and a 6. Compute the natural log of the CMS by combining the
correction of the correlation coefficients for high-frequency ε
T i  from step 5 with the μT and σT calculated
content. in step 3, using the equation

503
504 B. Carlton and N. Abrahamson

(a) 1×10-1 (b)

1×10-2

1×10-3

Hazard 1×10-4

1×10-5

1×10-6

1×10-7
0.01 0.1 1 10
SA ( g)

(c) (d) 1 T*=0.2

T*=1.0
0.8
1

Correlation p(T,T*)
0.6
SA (g)

0.1 0.4

UHS 0.2
Zhao Median
0.01 0
0.01 0.1 1 0.01 0.1 1 10
Period (s) Period (s)

Figure 1. (a) Hazard curve for T  1 s at a site near Portland, Oregon (step 1); (b) M and R deaggregation for a hazard of 2:5 × 10−5 at
T  1 s, with the controlling scenario of M w 8.4 and R  88:3 km (step 2); (c) median spectrum computed with one GMPE (Zhao et al.,
2006) compared with UHS for hazard 2:5 × 10−5 , εT
 1  1:97 (steps 3–4); and (d) correlation coefficients pT; T
 computed using
Baker and Jayaram (2008) (step 5). The color version of this figure is available only in the electronic edition.

CMST
T i   μT i   ε
T i  × σT i ; 3 tance event (Mw 8.4, R  88:3 km). Figure 1c shows the
median spectral acceleration computed for this scenario by
in which CMST
T i  is the CMS at period T i conditioned one GMPE compared with the UHS. Using equation (1) we
at period T
in natural log units and all other parameters compute εT
  1:97. Figure 1d shows the correlation co-
are as previously defined. efficients according to Baker and Jayaram (2008). Using
The name CMS comes from the fact that, given epsilon these correlation coefficients and the value of epsilon at T

at the conditioning period T
, we compute the mean values we can calculate the mean epsilons at other periods ε
T i .
of epsilon at other periods T i and then combine these mean Finally, using equation (3), we combine the mean epsilons
epsilons with the median spectral accelerations to produce with the median spectral values to create the CMS. We repeat
the CMS. this procedure for multiple T
until we have a suite of CMS
that cover the range of the UHS, as shown in Figure 2.

Example Calculation of CMS

Broadening of the CMS
Figure 1 shows an example calculation of CMS in
graphical form. This example is used later in the paper to Two important questions when creating CMS are how to
demonstrate the technique of broadening CMS and how to choose T
and how many CMS to develop for a given UHS.
incorporate multiple GMPEs in a CMS calculation. For the The selection of T
is not critical if the suite of CMS cover the
example we chose a site near Portland, Oregon. Figure 1a range of the UHS. However, this could require a large number
shows step one, the hazard curve computed using PSHA of conditioning periods, which is not practical for many
for a conditioning period of one second (T
 1). Figure 1b projects. On the other hand, selecting too few conditioning
shows step two, the deaggregation for a hazard level of periods will result in CMS that fall significantly below the
2:5 × 10−5 , as well as the mean magnitude and distance. UHS at periods away from their conditioning periods. An al-
The deaggregation shows that, for this site, the hazard at this ternative is to develop two or three CMS and then artificially
level and period is dominated by a large magnitude and dis- broaden their spectra. We recommend broadening the CMS
Issues and Approaches for Implementing Conditional Mean Spectra in Practice 505

UHS UHS
CMS T* = 0.2 CMS T *=0.2
CMS T* = 0.5 CMS T*=1
CMS T* = 1
CMS T* = 2
1 1
SA (g)

SA (g)
0.1 0.1
0.01 0.1 1 0.01 0.1 1
Period (s) Period (s)

Figure 2. Suite of four CMS that cover the range of the UHS Figure 3. Two CMS broadened to cover the range of the UHS.
(step 6). The color version of this figure is available only in the The solid lines are the original spectra, and the dashed lines are the
electronic edition. broadened spectra. The color version of this figure is available only
in the electronic edition.

symmetrically about each conditioning period so that the PSHA and k is the index of GMPE. The standard deviation
mismatch of spectral accelerations between the CMS and σ
CMS;T
of CMST
is computed in the same way (equation 5)
the UHS is not greater than 10% of the UHS. The broadening but has a small adjustment to account for the additional un-
should be done to a period halfway in log space to the next certainty due to the variation in median spectral values pre-
conditioning period for those between two conditioning peri- dicted by each GMPE. Both the CMST
and the σ
CMS;T
are in
ods. Figure 3 gives an example of two CMS from Figure 2 natural log units:
that we broadened to capture the UHS more closely, reducing
X
the number of CMS from four to two. Broadening CMS is CMST
T i   wlk × CMST
;k T i  4
subjective and requires engineering judgment. Therefore, k
it is ultimately the responsibility of the engineer conducting
the analysis to determine to what degree it is worth broad- σ CMS;T
T i 
ening the CMS to reduce the number of scenarios considered. rX












































































































Abrahamson (2010) suggests using T
s that capture the  wlk fσ 2CMS;T
;k T i CMST
;k T i −CMST
T i 2 g:
short- and long-period range of the UHS, with the T
s based k

on the project. There is no need, however, to require that T
5

be the fundamental mode of the structure. It is also not prac- In method 3, step 1 is the same as described in the Re-
tical because the modal periods often change during the view of CMS section, but for step 2 the hazard is deaggre-
structural design phase of a project. gated by M, R, and GMPE. Lin and Baker (2011) describe
GMPE deaggregation in detail. The analyst then determines
Weighting and More Than One GMPE the controlling M and R scenarios for each GMPE. Steps 3–6
are computed for each GMPE using its controlling M and R
When using only one GMPE, the procedure to compute scenario. Finally, the CMS from all GMPEs are combined in
the CMS is straightforward. However, if multiple GMPEs are the same manner as method 2 (equations 4 and 5), but instead
used in the PSHA through a logic tree, then the calculation of of using the logic-tree weights wl , the GMPE deaggregation
the CMS becomes more complex. weights wd are used. The GMPE deaggregation weights are
Lin et al. (2013) suggest four different methods to cal- the fractional contribution of GMPE k to the total hazard for a
culate CMS. Method 1 uses only one GMPE and is the same given period and hazard level. The deaggregation weights are
method outlined above in the Review of CMS section. Meth- determined using equation (6):
ods 2–4 address multiple GMPEs and increase in complexity,
with method 4 being the most intricate. In method 2, steps 1 wdk  wlk × Hk =Ht ; 6
and 2 are completed as described in the Review of CMS
section, but steps 3–6 are done for each GMPE separately. in which H t is the total hazard, H k is the hazard based on
Then, after step 6, the mean CMS conditioned on period GMPE k, and wlk is the logic-tree weight of GMPE k.
T
CMST
 is determined using equation (4), in which wl Method 4 from Lin et al. (2013), which they term the
is the logic-tree weight determined at the beginning of the exact method, is the most complex method. In method 4,
506 B. Carlton and N. Abrahamson

-2
1×10
steps 1 and 2 are completed the same as in method 3, but, Youngs et al 1997
instead of finding the deaggregation weights of the control- Atkinson and Boore 2003
ling M and R scenarios for each GMPE, the deaggregation -3
Zhao et al 2006
1×10 Weighted Average Hazard
weights are computed for each M and R bin for each GMPE.
For example, if we deaggregate the hazard for a given period
and GMPE into six distance bins and five magnitude bins,

Hazard
-4
and we are using four GMPEs, then we would create 6 × 5 × 1×10

4  120 M, R, and GMPE scenarios. The deaggregation

weight is then the fractional contribution of each of the
-5
120 M, R, and GMPE combinations to the total hazard. Steps 1×10
3–6 are then completed for each of the 120 scenarios. Finally,
the CMST
is computed as the weighted sum of all 120 com-
-6
binations in the same manner as equations (4) and (5). 1×10
0.1 1 10
We propose an alternative method, termed method 2.5, SA (g)
that is simpler and faster than methods 3 and 4, but without a
significant reduction in accuracy. Instead of computing all of Figure 4. The hazard curve for individual subduction zone
the CMS first and then weighting and summing them, method GMPEs and the weighted average hazard from all three combined
for T  1 s. The arrow shows the hazard level for 2:5 × 10−5. The
2.5 sums the weighted median and standard deviation values
color version of this figure is available only in the electronic edition.
from each GMPE before calculating εT
. Steps 1 and 2 are
the same as described in the Review of CMS section, but step
3 is conducted for each GMPE. Then, μT and σT for each
GMPEs; therefore, this site is ideal for testing whether the
GMPE are weighted by their deaggregation weights and
two methods give similar predictions of the CMS.
summed at each period, giving mean natural log response
Figure 4 shows the hazard curves by GMPE for a period
spectral values μ
T and mean natural log standard deviation
of one second for the site near Portland, Oregon. The GMPEs
values σT,

as shown in equations (7) and (8). Steps 4–6 are
used in this study are three subduction zone GMPEs; Youngs
then completed as described in the Review of CMS section,
et al. (1997), Atkinson and Boore (2003), and Zhao et al.
but using μ
T and σ
T instead of μT and σT.
(2006). The weighted average hazard curve is the weighted
X sum of these three, of which the logic-tree weights are 0.25
μ
T i   wdk × μk T i  7
k
for both the Atkinson and Boore (2003) GMPE and the
Youngs et al. (1997) GMPE and 0.5 for the Zhao et al. (2006)
X GMPE. To determine the deaggregation weights, it is
σ
T i   wdk × σ k T i : 8 necessary to find the hazard H k for each curve at the same
k
spectral acceleration, in which the weighted average hazard
The advantage of method 2.5 is that, by weighting and equals the design hazard, in this case 2:5 × 10−5 . For the
summing the μT and σT of each GMPE instead of their present example, when the weighted average hazard equals
CMS, only one εT
 and one CMS need to be calculated. In 2:5 × 10−5 , the spectral acceleration for one second is 0:8g.
addition, method 2.5 uses only the deaggregation from the Using the hazard Hk and logic-tree weight wlk for each
mean hazard to determine the controlling M and R, instead GMPE, and a weighted average hazard H t of 2:5 × 10−5 , we
of the GMPE-specific deaggregation. This is consistent with compute the deaggregation weights using equation (6).
commonly used hazard programs. Table 1 lists the logic-tree weights, hazard, and deaggrega-
tion weights for each GMPE.
Next, we compute μT and σT for each subduction
Comparison of Methods
GMPE using M w 8.4 and R  88:3 km. We then take the
The following example illustrates the differences be- weighted average of each GMPE using the deaggregation
tween the CMS computed using methods 2.5 and 3 and the weights to determine μ
and σ,
as shown in Figure 5a and b.
steps necessary to complete method 2.5. It also demonstrates Steps 4–6 are completed as described in the
the difference in CMS when using logic-tree weights instead Review of CMS section. Figure 6 shows the CMS computed
of deaggregation weights for combining results from differ- using method 2.5 with deaggregation weights as well as with
ent GMPEs. We continue with the earlier example of a site logic-tree weights and methods 2 and 3 from Lin et al.
near Portland, Oregon, at which the hazard for one second (2013). Figure 6 shows that there is a significant difference
is dominated by a subduction zone interface event. We chose between the CMS when deaggregation weights are used in-
Portland for the comparison because GMPEs for subduction stead of logic-tree weights, regardless of the combination
zone interface events have large deviations in their median procedure. In addition, Figure 6 shows that methods 2 and
values, and the greatest uncertainty when developing CMS is 2.5 using logic-tree weights are similar, and methods 3
when there are differences in the predictions given by the and 2.5 using deaggregation weights are similar. Method
Issues and Approaches for Implementing Conditional Mean Spectra in Practice 507

1
(a) UHS
Weighted Avg.
Zhao Median
Youngs Median
1 B&A Median

SA ( g)
SA ( g)

Method 2
0.1 Method 2.5 wl
Method 2.5 wd
Method 3

0.1
0.01 0.1 1 10
0.01 Period (s)
0.01 0.1 1
Period (s)
Figure 6. CMS computed using logic-tree weights, deaggrega-
tion weights, and both averaging methods. The color version of this
(b) 0.9
figure is available only in the electronic edition.

strates why small differences between the methods are ex-

Standard deviation ln units

0.8
pected in general. Equation (9) is method 3, and equation (10)
is the expanded form. Equation (11) is method 2.5, and
equation (12) is the expanded form. When we equate
0.7 equations (10) and (12), we can see that the methods are sim-
ilar: the main differences are the terms identified as equa-
tions (13) and (14). When we further expand equations (13)
0.6 and (14) into equations (15) and (16), respectively, it is easier
to see their differences. Each method contains the same parts.
The main difference is the order in which the parts are multi-
0.5 plied and summed. The equations are all in terms of natu-
0.01 0.1 1 ral log.
Period (s) Method 3:
X
Figure 5. Comparison of the weighted average (a) median spec- CMST
T i   wdk × CMST
;k T i : 9
tral acceleration and (b) standard deviation values to those com- k
puted for the individual GMPEs for M w 8.4 and R  88:3 km.
The color version of this figure is available only in the electronic This expands to
edition.
X 
CMST
T i   wdk × μk T i   ρT i ; T

2.5 with deaggregation weights is never more than 5.5% dif-
k
ferent than method 3 at periods between 0.03 and 3 s for this X 
example. × wdk × εk T
 × σ k T i  : 10
k

Mathematical Basis Method 2.5:

The following mathematical comparison between taking

CMST
T i   μ
 ρT i ; T
 × εT
 × σ:

11
the weighted average of the CMS (method 3) and taking the
weighted average of μT and σT (method 2.5) demon- This expands to
X 
Table 1. CMST
T i   wdk × μk T i   ρT i ; T
 × εUHS T

k
GMPE Specific Logic-Tree Weights (wlk ) and Hazards X 
(Hk ) Used to Determine Deaggregation Weights (wdk )
for Each GMPE × wk × σ k T i  :
d
12
k
wlk Hk wdk
Then extract the portions that are different:
Zhao et al. (2006) 0.50 2:18 × 10−5 0.434
Atkinson and Boore (2003) 0.25 5:04 × 10−5 0.501 X
Youngs et al. (1997) 0.25 6:63 × 10−6 0.066 Method 3 : wdk × εk T
 × σ k T i  13
k
508 B. Carlton and N. Abrahamson

X
Method 2:5 : εT
 × wdk × σ k T i : 14
k

Method 3:
X SAUHS T
 X wd × μk T
 × σ k T i 

× wdk × σ k T i  − k
:
k
σ k T  k
σ k T

15

Method 2.5:
SA T
 X
P dUHS × wdk × σ k T i 
wk × σ k T
 k
k
P d P
wk × μk T
 × wdk × σ k T i 
− k P d k
: 16
wk × σ k T

k

We can deduce several interesting facts from equa-

tions (15) and (16). If the standard deviations are indepen-
dent of period for each GMPE, regardless of differences
between GMPEs [σ k T in   σ k T i≠n ], then methods 3 Figure 7. Correlation coefficients for total residuals
and 2.5 give the same CMS. Additionally, if each GMPE has ρTotal T i ; T
 from the Abrahamson and Silva (2008) GMPE. The
color version of this figure is available only in the electronic edition.
the same standard deviation for a given period [σ kn T i  
σ k≠n T i ], then the methods give the same CMS. Because q
















σ k T i  are similar for different GMPEs, usually within the σT i   τ2  ϕ2 ; 19
range of 0.5–0.8 natural log units, methods 3 and 2.5 will
give similar results. in which μT is the mean natural log spectral acceleration,
fM; R; T i  is the ground-motion prediction equation, δtotal
is the total residual, δB is the between-event residual, δW
Correlation-Coefficient Models
is the within-event residual, τ is the standard deviation of
The CMS method requires a model for the correlation of δB, ϕ is the standard deviation of δW, and σ is the total stan-
εT between periods. Not all GMPEs have published corre- dard deviation. The between-event residual (δB) is the aver-
lation models, however, generic models are available (e.g., age difference between the observed spectral acceleration
Baker and Jayaram, 2008). Figure 1d shows correlation co- from a specific earthquake and the median spectral acceler-
efficients plotted for two different T
values. Figure 7 shows ation predicted by the GMPE. The within-event residual (δW)
correlation values of epsilon plotted as contour lines versus is the difference between the spectral acceleration at a spe-
period. When the periods are the same, the epsilons are the cific site for a given earthquake and the median spectral
same and the correlation is one. The greater the difference acceleration predicted by the GMPE plus δB (Al Atik et al.,
between periods the smaller the correlation, as seen by a drop 2010).
in the contours. This is because the peak responses of oscil- Because of the two error terms in GMPEs, correlations
lators with short and long periods tend to be affected by dif- for epsilon computed directly from the total residual, δtotal ,
ferent aspects of the ground motion, and thus the oscillator divided by the total standard deviation, σ, do not work. To
responses are less correlated as the difference between oscil- account for the two error terms, the correlation for the total
lator periods increases. The plot is symmetric about the one residuals ρtotal T i ; T
 must be calculated as
to one line because it does not matter which period is the
τT i  × τT

conditioning period, the correlation will be the same for ρtotal T i ; T
  ρ T ; T

the epsilons of the same two periods. σT i  × σT
 B i
Epsilon represents the total normalized residual, but in ϕT i  × ϕT

 ρ T ; T
; 20
GMPEs the total residuals are composed of two parts, as σT i  × σT
 W i
shown in equations (17), (18), and (19).
in which ρB T i ; T
 is the correlation of the normalized be-
μT i   fM; R; T i   δB  δW 17 tween-event residuals, and ρW T i ; T
 is the correlation of
the normalized within-event residuals. Although not explic-
δtotal  observed − predicted  δB  δW 18 itly stated in the paper, Baker and Jayaram (2008) deter-
mined correlations from within-event residuals using the
Issues and Approaches for Implementing Conditional Mean Spectra in Practice 509

Figure 9. T amp1:5 for records used to develop the Abrahamson

et al. (2013) GMPE. The color version of this figure is available only
in the electronic edition.

Figure 8. Comparison of correlation coefficients for within- periods are farther from T
, and in Figure 7 as a widening
event residuals ρW T i ; T
 and total residuals ρTotal T i ; T
 from of the contour lines at short periods.
the Abrahamson and Silva (2008) GMPE. The color version of this The peak period of every response spectrum is different
figure is available only in the electronic edition.
and depends on many factors. For example, as magnitude
Pacific Earthquake Engineering Research Center Next and distance increase, the peak period increases, but as
Generation Attenuation (PEER NGA) West 1 Database V S30 increases, the peak period decreases. As a result, cor-
(Chiou et al., 2008), not from the total residuals. However, relation coefficients of epsilon values cannot be calculated
τ is typically smaller than ϕ, and ρB T i ; T
 and ρW T i ; T
 based on period alone because the widening effect seen in
are similar, so the difference between ρW T i ; T
 and Figure 7 begins at different periods for every response spec-
ρtotal T i ; T
 is small. Figure 8 compares the ρW T i ; T
 trum. Determining T p is difficult, however, so we quantify
and ρtotal T i ; T
 calculated from the Abrahamson and Silva the period when the high-frequency content affects the re-
(2008) GMPE residuals and shows the difference between the sponse spectrum as T amp1:5 . T amp1:5 is the lowest period when
two is negligible. the spectral acceleration is equal to 1.5 times the PGA.
Figure 9 shows the T amp1:5 values for the records used in
the Abrahamson et al. (2013) GMPE. Figure 9 shows a wide
Correlation Coefficients and High Frequency range of T amp1:5 values for shallow crustal earthquakes. The
low T amp1:5 bin in Figure 9 bounds the T amp1:5 values
An important issue regarding correlation coefficients is between minus one and minus two standard deviations of the
their application to hard-rock sites with increased high- median T amp1:5 value, and the high T amp1:5 bin bounds the
frequency content (i.e., low kappa). This is seen in response T amp1:5 values between plus one and plus two standard de-
spectra as the decrease in spectral values from the peak viations of the median.
period T p to the peak ground acceleration (PGA). At short The dashed lines in Figure 10 are the correlation coef-
periods there is little energy left to resonate, so the oscillator ficients for records in the low T amp1:5 bin, and the solid lines
follows the response of the ground and not the response of a are the correlation coefficients for records in the high T amp1:5
single degree-of-freedom system with its given natural bin. The period at which the contour lines begin to flare out is
period. The response at short periods is then controlled by different for each subset of data due to their different T amp1:5
the dominant period of the acceleration time series, which is values. To correct for the effect of high-frequency content,
the peak period of the response spectrum, more than the we normalize the periods of each record by its T amp1:5 value.
response at its own period. Therefore, at periods smaller than We then calculate the correlation coefficients of epsilon
T p , the epsilon values will be more correlated with the between different T=T amp1:5 . Figure 10b shows the resulting
epsilon values of T p , and hence more correlated with epsilon normalized correlation model. We can see that, for T=T amp1:5
values of periods greater than T p , than for other periods with ratios smaller than one, the same flaring effect of the contour
similar spacing. This is seen in Figure 1d as a rise in the lines occurs due to the effect of high-frequency content. At
correlation coefficients at small periods, even though these T=T amp1:5 ratios greater than one, the contour lines converge
510 B. Carlton and N. Abrahamson

Figure 11. T amp1:5 values for records in the PEER NGA West 1
database. The color version of this figure is available only in the
electronic edition.

Baker and Jayaram (2008) derived their model using

data from the PEER NGA West 1 Database (Chiou et al.,
2008). Figure 11 is a histogram of the T amp1:5 values for this
database. The mean value of T amp1:5 for this database is 0.1 s,
which is the same period at which the correlation coefficients
begin to increase in Figure 1d. This indicates that the Baker
and Jayaram (2008) model is calibrated for response spectra
with T amp1:5  0:1 s. When calculating CMS from a control-
ling scenario spectrum with a T amp1:5 value different than
0.1 s, the following procedure should be used. First, multiply
T
by 0:1=T amp1:5 to get T
′ . Then, use T
′ in the Baker
and Jayaram (2008) model to estimate ρT i ; T
′ . Finally,
multiply the periods, not the correlation coefficients, by
T amp1:5 =0:1. Figure 12 shows examples of correlation coef-
ficients computed with the Baker and Jayaram (2008) model
for T
 1 s and T amp1:5 values of 0.05, 0.1, and 0.2 s. As
expected, the trough of each curve corresponds to its T amp1:5
value, otherwise the shapes of the curves are the same.

Robustness of Correlation Coefficients

Using different datasets, other studies have found
Figure 10. (a) Correlation coefficients for records with low
ρT i ; T
 values similar to Baker and Jayaram (2008). For
T amp1:5 and high T amp1:5 versus period and (b) period normalized
by T amp1:5. The color version of this figure is available only in example, Figure 13 shows ρT i ; T
 found from the residuals
the electronic edition. of the BC Hydro Model (N. Abrahamson, N. Gregor, and K.
Addo, unpublished report, 2014; see Data and Resources),
which is a GMPE for subduction zone events, overlaid on
and follow the more expected pattern. Comparing Figure 10a the ρT i ; T
 from the Abrahamson and Silva (2008) GMPE
and Figure 10b, we see that the contour lines are very similar, developed for shallow crustal earthquakes. The correlation
except that in Figure 10b the contours are shifted so that the coefficients from these two different datasets are similar be-
effect of the high-frequency content occurs at the same point cause they have similar average T amp1:5 values. Figure 14 is a
for both subsets of data. If the effect of the high-frequency histogram of the T amp1:5 values for the subduction zone
content (i.e., kappa) is not taken into account then the cor- events used to develop the BC Hydro GMPE (N. Abraham-
relation coefficients at short periods will be overpredicted for son, N. Gregor, and K. Addo, unpublished report, 2014; see
hard-rock sites and underpredicted for soft soil sites. Data and Resources). The mean T amp1:5 value is 0.08 s,
Issues and Approaches for Implementing Conditional Mean Spectra in Practice 511

Figure 12. Correlation coefficients computed using Baker and

Jayaram (2008) with different values of T amp1:5 . The color version of Figure 14. T amp1:5 values for subduction zone events used to
this figure is available only in the electronic edition. develop the BC Hydro Model (N. Abrahamson, N. Gregor, and
K. Addo, unpublished report, 2014; see Data and Resources).
The color version of this figure is available only in the electronic
edition.

Conclusions
The CMS is a spectrum that matches the UHS at a con-
ditioning period but that also represents the response from a
single earthquake scenario more realistically than the UHS.
The technique of broadening the CMS to cover the range of
the UHS is a useful way to reduce the number of CMS nec-
essary to perform a seismic analysis. Determining how much
to broaden the CMS is the responsibility of the project engi-
neer, and it involves a trade-off that is specific to each
project. We recommend broadening the CMS symmetrically
about each conditioning period so that the mismatch of spec-
tral accelerations between the CMS and the UHS is not greater
than 10% of the UHS. The broadening should be done to a
period halfway in log space to the next conditioning period
for periods between two conditioning periods.
The proposed method 2.5 presented in this paper for us-
ing multiple GMPEs in calculating CMS gives similar results
to method 3 proposed by Lin et al. (2013). Method 2.5, how-
ever, is easier to use in practice. Both methods found that the
Figure 13. Comparison of correlation coefficients for total logic-tree weights give different answers from the results us-
residuals ρtotal T i ; T
 of the BC Hydro subduction zone GMPE
(N. Abrahamson, N. Gregor, and K. Addo, unpublished report, ing the deaggregation weights, and we recommend using the
2014; see Data and Resources) and the Abrahamson and Silva deaggregation weights.
(2008) active crustal GMPE (figure from Al Atik, 2011). The color Correlation-coefficient models based only on the within-
version of this figure is available only in the electronic edition. event residuals are close enough to correlation coefficients
based on the total residuals for all practical purposes. How-
ever, correlation coefficients depend on the high-frequency
whereas for the shallow crustal dataset it is 0.1 s. This sup- content of the response spectra. This effect is removed by
ports the argument that any variation in correlation coeffi- normalizing the periods by T amp1:5, the shortest period when
cients comes from spectral shape rather than tectonic the response spectral value is 1.5 times the PGA. Figure 10b
region and that generic correlation models are robust and shows correlation coefficients normalized by T amp1:5. If this
can be used in determination of CMS regardless of the correction is not used then the correlation coefficients for
GMPEs considered. hard-rock sites will be overpredicted at short periods and
512 B. Carlton and N. Abrahamson

underpredicted for soft-soil sites at short periods. No correc- Al Atik, L., N. Abrahamson, J. Bommer, F. Scherbaum, F. Cotton, and
tion is necessary for long periods. N. Kuehn (2010). The variability of ground-motion prediction models
and its components, Seismol. Res. Lett. 81, no. 5, 794–801.
Finally, correlation coefficients found using residuals Atkinson, G., and D. Boore (2003). Empirical ground-motion relations for
from the subduction zone BC Hydro GMPE (N. Abrahamson, subduction-zone earthquakes and their application to Cascadia and
N. Gregor, and K. Addo, unpublished report, 2014; see Data other regions, Bull. Seismol. Soc. Am. 93, no. 4, 1703–1729.
and Resources) are similar to those from the Abrahamson Baker, J. (2011). Conditional mean spectrum: Tool for ground motion
selection, J. Struct. Eng. 137, no. 3, 322–331.
and Silva (2008) GMPE for shallow crustal events because
Baker, J., and C. Cornell (2006). Spectral shape, epsilon and record
both datasets have similar average T amp1:5 values. This sug- selection, Earthq. Eng. Struct. Dynam. 35, no. 9, 1077–1095.
gests that as long as the target response spectrum’s T amp1:5 Baker, J., and N. Jayaram (2008). Correlation of spectral acceleration
value is taken into account, correlation-coefficient models values from NGA ground motion models, Earthq. Spectra 24,
can be used when calculating CMS regardless of the GMPEs no. 1, 299–317.
considered. Bommer, J., S. Scott, and S. Sarma (2000). Hazard-consistent earthquake
scenarios, Soil Dynam. Earthq. Eng. 19, 219–231.
Chiou, B., R. Darragh, N. Gregor, and W. Silva (2008). NGA project strong-
Data and Resources motion database, Earthq. Spectra 24, 23–44.
Lin, T., and J. Baker (2011). Probabilistic seismic hazard deaggregation of
The example PSHA for the site near Portland, Oregon, is ground motion prediction models, in Proc. of the 5th International
proprietary information of Pacific Gas and Electric Company Conference on Earthquake Geotechnical Engineering, Santiago,
(PG&E). Norm Abrahamson provided the residuals for the Chile, 10–13 January 2011.
Abrahamson et al. (2013), and the Abrahamson and Silva Lin, T., S. Harmsen, J. Baker, and N. Luco (2013). Conditional spectrum
computation incorporating multiple causal earthquakes and ground
(2008) GMPEs. The NGA West 1 (http://peer.berkeley.edu/ motion prediction models, Bull. Seismol. Soc. Am. 103, no. 2A,
ngawest/databases.html, March 2013) and NGA West 2 1103–1116.
(http://peer.berkeley.edu/ngawest2/databases/, April 2013) Youngs, R., S. Chiou, W. Silva, and J. Humphrey (1997). Strong ground
ground-motion databases were downloaded from the PEER motion attenuation relationships for subduction zone earthquakes,
website. The BC Hydro model by N. Abrahamson, N. Gregor, Seismol. Res. Lett. 68, no. 1, 58–73.
Zhao, J., J. Zhang, A. Asano, Y. Ohno, T. Oouchi, T. Takahashi, H. Ogawa,
and K. Addo has been submitted to Earthquake Spectra and is K. Irikura, H. K. Thio, P. G. Somerville, Ya. Fukushima, and Yo.
in review. Fukushima (2006). Attenuation relations of strong ground motion
in Japan using site classification based on predominant period, Bull.
Acknowledgments Seismol. Soc. Am. 96, no. 3, 898–913.

The first author completed this research while supported by a National

Science Foundation Graduate Research Fellowship.
University of California, Berkeley
418 Davis Hall
References Berkeley, California 94720-1710
[email protected]
Abrahamson, N. (2010). Selection of T0, multiple CMS, NRC approach,
(B.C.)
COSMOS Technical Session: Issues Associated with Applying
Conditional Mean Spectra to the Selection and Scaling of Design
Ground Motions, Millbrae, California, 19 November 2010.
Abrahamson, N., and W. Silva (2008). Summary of the Abrahamson and
Pacific Gas & Electric Company
Silva NGA ground-motion relations, Earthq. Spectra 24, no. 1, 67–97. Geosciences Department
Abrahamson, N., W. Silva, and R. Kamai (2013). Update of the AS08 P.O. Box 770000, Mail Code N4C
Ground-Motion Prediction equations based on the NGA-west2 data San Francisco, California 94177
set, Pacific Engineering Research Center Report 2013/04. (N.A.)
Al Atik, L. (2011). Correlation of spectral acceleration values for subduction
and crustal models, COSMOS Technical Session: Ground Motions
from Subduction Earthquakes and Issues for Seismic Design, Manuscript received 17 May 2013;
Emeryville, California, 4 November 2011. Published Online 14 January 2014