See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/305276733

Ground‐Motion Prediction Equations for Subduction Slab Earthquakes in Japan

Using Site Class and Simple Geometric Attenuation Functions

Article in Bulletin of the Seismological Society of America · July 2016

DOI: 10.1785/0120150056

CITATIONS
READS
21
799

13 authors, including:

Hao Xing
Ruibin Hou
Southwest Jiaotong University
Southwest Jiaotong University
5 PUBLICATIONS 49 CITATIONS
2 PUBLICATIONS 49 CITATIONS

Yingbin Zhang
Pengcheng Yu
Southwest Jiaotong University
Southwest Jiaotong University
110 PUBLICATIONS 1,275 CITATIONS
22 PUBLICATIONS 276 CITATIONS

Some of the authors of this publication are also working on these related projects:

Attenuation of seismic waves (Q-value) View project

earthquake induced landslides View project

All content following this page was uploaded by Hao Xing on 13 July 2020.

The user has requested enhancement of the downloaded file.

 Bulletin of the Seismological Society of America, Vol. 106, No. 4, pp. –, August 2016, doi: 10.1785/0120150056

Ground-Motion Prediction Equations for Subduction Slab Earthquakes in

Japan Using Site Class and Simple Geometric Attenuation Functions
by John X. Zhao, Fei Jiang, Pan Shi, Hao Xing, Haifeng Huang, Ruibin Hou,
Yingbin Zhang, Pengcheng Yu, Xiaowen Lan, David A. Rhoades, Paul G.
Somerville, Kojiro Irikura, and Yoshimitsu Fukushima

Abstract The frequency content of strong ground motions from subduction slab
earthquakes differs significantly from that of ground motions produced by other cat-
egories (tectonic locations: shallow crustal, upper mantle, and subduction interface) of
earthquakes in subduction zones. In the last two decades, a large number of records
from subduction slab events have been obtained in Japan. We present a ground-motion
prediction equation (GMPE) for this category of earthquakes. We used a large
dataset from reliably identified slab events up to the end of 2012. The GMPEs were
based on a set of simple geometric attenuation functions, site classes were used as
site terms, and nonlinear site amplification ratios were adopted. A bilinear
magnitude-scaling function was adopted for large earthquakes with≥ moment
magnitude Mw 7:1, with the scaling rates for large events being much smaller than
for the smaller events. A magnitude-squared term was used for events with Mw < 7:1
as well as the bilinear magnitude-scaling function. We also modeled the effect of
volcanic zones using an anelastic attenuation coefficient applied to a horizontal
portion of the seismic-wave travel distance within possible volcanic zones. We found
that excluding the records from sites with inferred site classes improved the model
goodness of fit. The within-event residuals were approximately separated into within-
site and between-site residuals, and the corresponding standard deviations were
calculated using a random effects model. The separation of within-event residuals
into within-site and between-site components allows for the possibility of adopting
different standard deviations for different site classes in a probabilistic seismic-
hazard analysis if desired.

Online Material: Figures showing the distribution of between-event residuals

with respect to magnitude and fault-top depth and the distribution of within-event
residuals with respect to magnitude and source distance.

Introduction
In the last decade, many modern ground-motion
predic- tion equations (GMPEs) have been published. These
include the Next Generation Attenuation models, which are
mainly based on strong-motion records from California but
supple- mented by shallow crustal records from Taiwan,
Japan, and Turkey, including Abrahamson and Silva
(2008), Boore and Atkinson (2008), Campbell and
Bozorgnia (2008), and Chiou and Youngs (2008). For
subduction zones, the tectonic set- tings are complex and a
relatively small number of modern GMPEs have been
developed, for example, Atkinson and Boore (2003),
Kanno et al. (2006), McVerry et al. (2006), Zhao, Zhang,
et al. (2006), and Ghofrani and Atkinson (2014). The
models by Zhao, Zhang, et al. (2006) and Kanno
et al. (2006) were based on strong-motion records from Japan,
and nonlinear site terms were not used. The Zhao, Zhang, et
al. (2006) model used site class based on site period as the
site term, because many strong-motion recording stations
had no site information. For most stations in Japan, the site
class was inferred by Zhao, Irikura, et al. (2006) using
response spectral ratios of the horizontal and vertical
components (H/V). The site classes are defined in Table 1
together with the approxi- mate National Earthquake
Hazards Reduction Program (NEHRP) site classes
(Building Seismic Safety Council [BSSC], 2000). The use
of site classes by Zhao, Irikura, et al. (2006) produced
consistent site amplification ratios for three soil site classes
(SC II, III, and IV) with respect to SC I sites.
BSSA Early Edition / 1
 2 J. X. Zhao, et al.

Table 1
Site Class Definitions Used in the Present Study and the Approximately Corresponding NEHRP Site Classes (BSSC,
2000)
Site Class Description Natural Period VS30 Calculated from Site Period NEHRP Site Classes
SC I Rock T < 0:2 s VS30 > 600 A+B+C
SC II Hard soil 0:2 ≤ T < 0:4 s 300 < VS30 ≤ 600 C
SC III Medium soil 0:4 ≤ T < 0:6 s 200 < VS30 ≤ 300 D
SC IV Soft soil T ≥ 0:6 s VS30 ≤ 200 E+F

NEHRP, National Earthquake Hazards Reduction Program.

The Zhao, Zhang, et al. (2006) model used strong-motion re-
cords up to the end of 2003, and since then many more strong-
motion records have been obtained by the K-NETand KiK-net
strong-motion networks in Japan. In the past, subduction inter-
face and slab earthquakes were grouped together (Youngs
et al., 1997), and usually a constant in the GMPE was used
to describe the different attenuation characteristics. Zhao,
Zhang, et al. (2006) used common site terms for all three
categories of earthquakes: shallow crustal, subduction
interface, and subduc- tion slab categories. The subduction
slab events have an addi- tional geometric attenuation term.
When the three categories of events are combined, between-
and within-event standard de- viations are the same for all
three earthquake categories.
The reasonably large number of records in each earth-
quake category allowed Zhao and Rhoades (2014) to develop
separate GMPEs for the three categories of events. The first
model was for the shallow crustal and upper-mantle events,
the second model is for subduction interface events, and the
third model is for subduction slab events. Deriving separate
GMPEs has an important advantage: separate within-event
and between-event standard deviations can be obtained. Be-
cause of the different frequency content in the strong-
motion records from different event categories, the standard
devia- tions may differ statistically. This may be significant
for a probabilistic seismic-hazard analysis in subduction
zones. For
example, the standard deviations for subduction slab events
may be larger than those for shallow crustal earthquakes (Zhao
and Rhoades, 2014). If the same model standard deviation
is used, the hazard from subduction slab events would be
under- estimated whereas the hazards from the other two
types of earthquakes are overestimated.
Another advantage of deriving three separate models is
to differentiate the site terms among the three event groups.
Zhao et al. (2009) and Zhao and Zhang (2010) suggested that
site amplification ratios may depend on the frequency con-
tent and thus may depend on earthquake categories. There-
fore, it may be more appropriate to use different site terms
derived from separate GMPEs than using the same site
terms for all event categories. We derive site terms in this
study that differ from those in the Zhao, Liang, et al., 2016
study for shallow crustal and upper-mantle events and the
Zhao, Zhou, et al. (2016) study for subduction interface
moment magnitude Mw > 7:0, using strong-motion data
from worldwide crustal earthquakes and from large subduc-
tion interface earthquakes in Japan. Their studies showed that
the magnitude-scaling rates were very small, varying in a
range of 0.0–0.3, for large crustal earthquakes. For large sub-
duction interface earthquakes (Mw > 7:0), Zhao (2014) and
Zhao and Rhoades (2014) showed that the magnitude-scaling
rate was also much smaller than that for smaller events. In the
present study, a magnitude-squared term and a bilinear
mag- nitude-scaling function hinged at Mw 7.1 from the
Zhao and Rhoades (2014) study was used.
Zhao and Rhoades (2014) and Zhao, Zhou, et al. (2015)
examined the earthquake locations reported in the catalogs
of the Japan Meteorological Agency, an Engdahl–van der
Hilst–Buland group in the catalogs from the International
Seismological Centre, and the U.S. Geological Survey Na-
tional Earthquake Information Center catalog. They assigned
earthquake categories for 312 events, among which are the
records from the subduction slab events used in this study.
In the present study, we use site classes as defined in
Table 1. Even though site classes do not provide a continuous
predicted response spectrum across all sites, they are still
useful for some design codes, such as the New Zealand de-
sign code NZS1170.5:2004 (Standards New Zealand,
2004),
and for some engineering sites that may have a site class but
no accurate site period or V S30 (the travel-time averaged
events.
Zhao and Lu (2011) and Zhao and Xu (2012) investi- gated
the magnitude-scaling rates for large earthquakes with

BSSA Early Edition

 GMPEs for Subduction Slab Earthquakes in Japan 3
shear-wave velocity in the top 30 m).
Zhao, Zhou, et al. (2016) presented the site
information quality effect, that is, the change in the
goodness-of-fit parameters for GMPEs with or
without the strong-motion records from sites with an
inferred site class, using H/V spec- tral ratios (Zhao,
Irikura, et al. 2006) or geological descrip- tions of
the surface soils. In the present study, we also
investigated this aspect for subduction slab records.
We used the maximum log likelihood ( MLL),
rather than the model standard deviation, as the
indicator of goodness of fit for a GMPE, as described
by Zhao and Rhoades (2014). We found that the
MLL was a good way to identify an ap- parently
biased distribution of residuals with a parameter that
is strongly influenced by an outlier in the dataset. In
this case, when an additional term is used to correct
this type of biased residual distribution, the MLL
does not change, meaning the correction is not
statistically necessary.

BSSA Early Edition

 Table 2
Number of Events in Each Focal Mechanism Group
Focal Mechanism
Dataset Strike Total in Each
Number Reverse Slip Normal Earthquake Category
1 98 13 25 136
2 95 10 20 125

Table 3
Number of Records in Each Site Class
Dataset Number SC I SC II SC III SC IV Total
1 2031 1354 443 882 4710
2 2002 1292 414 847 4555

events. The first dataset includes records from all earth-

Figure 1. (a) The distribution of earthquakes used in the present
study with respect to fault depth and moment magnitude and (b)
the distribution of strong-motion records with respect to source
distance and magnitude.
We adopted the site terms accounting for soil nonlinear
response derived by Zhao, Hu, et al. (2015) and extended by
Zhao, Zhou, et al. (2016). The number of strong-motion re-
cords from subduction slab events in Japan with significant
nonlinear soil response is still small, because subduction
events are usually deep and distant from the land in Japan.
The Zhao, Hu, et al. (2015) and Zhao, Zhou, et al. (2016)
models were derived from 1D models based on the shear-
wave velocity profiles from a number of selected KiK-net
stations. The model parameters that were controlled by
weak motions were separated from the crossover rock
spectrum that distinguishes amplification from
deamplification ranges of a rock spectrum, as described by
Zhao, Hu, et al. (2015). The parameters controlled by weak
motions were determined from GMPE amplification ratios,
whereas the crossover rock spectrum and the slope of the
attenuation in the amplification ratios for a very high rock-
site spectrum were determined from 1D model analyses.
quakes since 1968, with 98 earthquakes having reverse-
faulting mechanisms, 13 strike-slip, and 25 normal-faulting
mechanisms, as listed in Table 2. In the first dataset, 2031
records are from SC I sites, 1354 from SC II sites, 443 from
SC III sites, and 882 from SC IV sites, with a total of 4710
records, as listed in Table 3. Seven earthquakes have a
mag- nitude greater than 7.0, with the largest magnitude
being 8.25, and 539 records are from large events with≥ Mw
7:1 in the first dataset.
For the records from earthquakes before 1996 and
some records from a number of K-NET stations, the site
class was inferred from response spectral ratios of H/V
components. In the second dataset, the exclusion of sites
with an inferred site class leads to the removal of 11 events
and 155 records, in- cluding 29 SC I records, 62 SC II
records, 29 SC III records, and 35 SC IV records. In the
second dataset, 125 events were used, and the number of
records in each site class is pre- sented in Table 3. We used
the first dataset to derive the magnitude-scaling rate for
large events. Then we derived the subduction slab model
≥
from the second dataset, using the fixed magnitude-scaling
rate for large events with Mw 7:1 derived from the first
dataset. Figure 1a shows the distribu- tion of earthquakes
with respect to fault-top depth and mag- nitude, and Figure
1b shows the distribution of records with respect to source
distance and magnitude for the second data- set. The
maximum magnitude in the second dataset is 7.92, as
shown in Figure 1.
We have fault-rupture models for a number of earth-
quakes. The references for these models can be found in the
electronic supplement of Zhao, Liang, et al., 2016.

Strong-Motion Dataset Model Function Forms

We constructed two datasets, as in the study by Zhao, In this study, we employed the following functions to
Liang, et al., 2016 for shallow crustal and upper-mantle model the source effect, that is, the magnitude and fault depth
for subduction slab events:
4 J. X. Zhao, et al.

to gravity, and μ denotes the recorded PGA or spectrum.

Variable e denotes the anelastic attenuation rate; x denotes
the shortest distance from a recording station to the fault
plane if a fault model is available and otherwise is the hy-
pocentral distance; γSL is a constant; and gSL denotes the
geometric attenuation rate. The term gSLL is a large-distance
geometric attenuation rate. The superscript v indicates as-
sociation with the volcanic path. The anelastic attenuation
rate evSL is applied to the horizontal distance passing
through volcanic zones, denoted by xv and illustrated in
Figure 2. ASL is the site amplification ratio for slab events
and contains both linear and nonlinear site terms. The sub-
script i indicates the ith event in the dataset, and j indicates
the jth record in the ith event. The random variable ξ is the
within-event residual with an average value of 0.0 and a
standard deviation of σ (i.e., the within-event standard
devia- tion). Random variable ηi is the between-event
residual with an average value of 0.0 and a standard
deviation τ (i.e., the between-event standard deviation). The
Figure 2. (a) The definition of volcanic path for four cases distance used for geo- metric spreading is defined by
and
(b) the horizontal and slant volcanic distance. The color version of ri;j = xi;j + exp(c1 + c2Cm); (3)
this figure is available only in the electronic edition.
and
fmSL(mi;hi)= bSLhi
cSL1mi + cSL2(mi − msc)2 if mi ≤mc mi if mi ≤ Cmax
+ Cm = : (4)
c m + (m − )2 + (m − m ) if m >m ; Cmax if mi > Cmax
c m d
SL1 c SL2 c SL i i c
sc c
(1) We used a maximum magnitude C m in equation (4)
in which SL indicates that the term is associated with al. (2016)
subduc- tion slab earthquakes, bSL is the coefficient for the as described by Zhao, Zhou, et max = c . Coefficient
fault-top depth term, and cSL1 and cSL2 are the coefficients for c2 =1:151 was selected and justified by the relationship
the linear magnitude and the magnitude-squared terms, of magnitude and fault length relations (see Zhao, Zhou,
respectively, for events with a moment magnitude mi less et al., 2016).
than or equal to mc. The magnitude-squared term with a The anelastic attenuation term qSLH is defined by
positive value was found
statistically significant with msc = 6:3 selected and msc is a 0 if h < 50
magnitude constant. In the Zhao, Zhang, et al. (2006) model, a qSLH = eSLH : (5)
positive magnitude-squared term was used, and this term leads 0:02h − 1:0 if h ≥ 50
to an unrealistically large short-period spectrum for large
events. Coefficient dSL is the magnitude term for large The fault-top depth is denoted by h in equation (5) for the
= mc 7:1 based on the
events. In the present study, we selected depth-dependent anelastic attenuation rate. For slab events,
results of Zhao and Lu (2011), Zhao and Xu (2012), Zhao equation (5) does not appear to be reasonable because we
(2014), and Zhao and Rhoades (2014). would expect that the seismic waves from a deep slab event
The GMPE in the present study is would have a long travel path within the subducting slab
that has high-Q values. From this simple reasoning the
loge(yi;j)= fmSL + gSL loge(ri;j)+ gSLL loge(xi;j + 200:0) anelastic attenuation rate for deep slab events should be
inversely pro-
portional to the fault depth, as shown by Eberhart-Phillips
+ eSLxi:j + qSLHxi:j + evSLxvi;j + γSL
and McVerry (2003). We attempted to use an anelastic at-
+ loge(A)+ ξi;j + ηi; (2a) tenuation rate inversely proportional to the fault depth.
How- ever, this term is not statistically significant, while
and the term
qSLH in equation (5) leads to a sizable increase in the MLL.
BSSA Early Edition
GMPEs for Subduction Slab Earthquakes in Japan 5
μi;j
log = + η: The physical cause behind equation (5) is probably related
e y i;j i (2b) to the geometry of the subduction interface and the location
ξ i;j
of the strong-motion recording stations relative to the
subduc-
Variable y is for peak ground acceleration (PGA) or the 5% tion trench, as shown in Figure 3. For the North Island of
damped response spectrum in units of the acceleration due New Zealand, the onshore depth of the interface is
relatively

BSSA Early Edition


Figure 3. Different seismic-wave propagation paths from a
deep subduction slab event to a recording station in Japan (left)
and a recording station in New Zealand (right). The Japanese islands
are much further from the subduction trench than is the North Island
of New Zealand. The color version of this figure is available only
in the electronic edition.
small (starting from about 20 km), and the trench formed
by the subducting slab is also relatively close to the
shoreline. For many slab earthquakes, a large portion of the
travel path
for the seismic waves recorded by New Zealand onshore
stations lies in the subduction slab, as illustrated by the
seis- mic-wave travel path to the recording station on the
right side of Figure 3. Therefore, the length of the travel
path within the high-Q slab increases with increasing
earthquake depth. This travel path leads to an anelastic
attenuation co- efficient inversely proportional to the fault
depth. For Japan, the subduction trench is usually far
offshore, and the seis- mic waves reaching the recording
stations travel in the slab and also through the upper mantle
and crust that have smaller Q values than the subducting
slab. We expect that the Q value in the upper mantle is
likely to be smaller than the Q value within the subducting
slab. The travel path within the mantle would increase with
increasing depth, and therefore the apparent anelastic
attenuation rate may in- crease with increasing depth, as
shown in equation (5). Zhao (2010) showed that, if the
constructive interference between the seismic waves
traveling along the direct path and the waves traveling
along a refracted path through the slab were modeled by
depth-scaled geometric attenuation functions, the apparent
anelastic attenuation rate for the slab events would
decrease.
We adopted the method of modeling volcanic path
attenuation from Zhao, Zhou, et al. (2016), using the sum
of the horizontal portions of the path (along a straight line
be- tween a station and the fault plane) that pass through
volcanic zones xv as the measure of volcanic path. An
anelastic attenu-
ation rate evSL was applied to the volcanic path. We adopted
the minimum and the maximum values for xv as 12 and
80 km, respectively. For example, when 0:0 < x≤v 12:0 km,
= ≥ =
Following Zhao, Zhou, et al. (2016), the random vari-
able ξ can be divided into a within-site component ξS and a
between-site component ηS, using the algorithm of Abra-
hamson and Youngs (1992). We fit the following random
effects model to the within-event residuals:
ξk;n = ξsk; + ηsk; (6)
n
in which k stands for the site number and n denotes the nth
record from the kth site. The within-site component ξk;s has
a zero mean and a within-site standard deviation of σnS. The
between-site component ηks has a zero mean and a between-
site standard deviation of τS. The between-site standard
deviation is an indicator of how well the site effects are
mod- eled. The within-event residuals contain the random
errors associated with path effects, as well as any other
effects that are not modeled.
The total site standard deviation can be calculated by
= qσ2
σSTk + τ2 ; (7)
S S
k k
in which k denotes the site class number.
SC I sites are neither rock nor engineering bedrock
sites. Many SC I sites have a layer of stiff soil with a
thickness up to 24 m and a shear-wave velocity as small as
200 m=s for some sites with a thin surface soil layer. The
average imped- ance ratio (defined by Zhao, Hu, et al.,
2015) is 3.7, and many sites have an impedance ratio
between 4 and 8. The sites with a thin soil layer usually have
a small average shear- wave velocity and a large impedance
ratio. The sites with a thick soil layer usually have a large
average shear-wave velocity and a relatively small
impedance ratio for a given site period. These
characteristics of the SC I sites lead to small nonlinear soil
response, even when subjected to strong rock motions,
because the shear stress from the inertial force in the thin
soil layer can be smaller than the soil yield stress. Also the
definition for rock sites, with a shear-wave velocity of 760
m=s or larger at the ground surface, means that the VS30 for
these sites could be over 1000 m=s and the site could be
classified as A or B in the NEHRP (BSSC, 2000). Because
the amplification ratio of the nonlinear site model in Zhao,
Hu, et al. (2015) and Zhao, Zhou, et al. (2016) is the
spectral ratio for all sites over the rock sites, we need to
estimate the amplification ratio for SC I sites relative to the
rock sites.
We used the following method to determine the ampli-
fication ratios for SC I sites or deamplification ratios for
rock sites. We examined the within-event residuals of the
SC I
sites and fitted a linear function of site periods to the SC I
xv 12:0 km, and when xv 80:0 km, xv 80:0 km, as suggested by
Zhao and Rhoades (2014) and Zhao, Zhou, et al. (2016).
within-event residuals from all three GMPEs (one for
shallow crustal and upper-mantle events presented
by Zhao, Zhou, et al., 2016, one for subduction slab
events presented in this study, and one for
subduction interface events by Zhao, Liang, et al.,
2016). We combined the residuals from three
 Figure 5. (a) Smoothed deamplification ratio for a rock-site
spectrum with respect to SC I sites and (b) an example SC I and
rock spectrum. The rock-site spectrum equals the SC I spectrum
divided by the deamplification ratio presented in (a). Rock sites have
a surface shear-wave velocity of 760 m=s or larger. The color
version of this figure is available only in the electronic edition.

as the hard-rock-site term. Figure 5b shows an example SC

I spectrum and the rock-site spectrum derived by dividing
the SC I site spectrum by the deamplification ratio
presented in Figure 5a. The spectral period at the peak
spectrum for rock sites is close to 0.05 s, shifted slightly
from that of the SC I spectrum, as it should be. This method
Figure 4. Distribution of within-event residuals for site class
(SC I) sites with respect to site period for (a) peak ground accel-
does not require iterations because very few records from
eration (PGA), (b) 0.5 s, and (c) 2.0 s spectral periods. The color SC I sites contain the effect of significant soil nonlinear
version of this figure is available only in the electronic edition. response. Table 4 presents the rock-site ( )AmSC I exp
= factor
S1N . The maxi- mum site amplification ratio is defined by
GMPEs in order to have a large number of records. Figure 4
shows the distribution of within-event residuals with AN max k = AmSC I for SC I sites (k = 1); (8a)
respect to site period, and the distributions are clearly
biased. The solid trend lines represent the average within- and
event resid- uals, and the negative value of the average
residuals at zero
site period is then defined as S1N , the negative intercept of AN max k = AmSC I exp(Sk)
the linear trend line. We used linear functions for all spectral (8b)
for SC II; III; IV sites (k = 2; 3;
periods.
Figure 5a shows the smoothed deamplification ratios 4):
for rock sites with respect to SC I sites (or the average am- Dropping k for simplicity, the nonlinear amplification ratio
plification ratios for SC I sites with respect to rock sites). is given by
At short spectral periods, the deamplification ratios vary
between 1.2 at 0.1 s spectral period and 2.05 at 0.3 s, and loge(A)
then decrease with increasing spectral periods down to 1.27 loge(Sα + β) − loge(β)
MR
at 5.0 s. The PGA deamplification ratio is 1.38. These deam- by large impedance ratios for SC I sites. We referred
plification ratios are surprisingly large and may be caused
 ;
to S1N = loge(AN max) − loge(Amax) e α + β) − loge(β)
log (S
ReffC
(9)
 Table 4
Rock-Site Deamplification Factor AmSC I
Period Period Period Period
Number Period (s) AmSC I Number Period (s) AmSC I Number Period (s) AmSC I Number Period (s) AmSC I

1 PGA 1.381 11 0.1 1.231 21 0.4 2.025 31 2.0 1.574

2 0.01 1.228 12 0.12 1.334 22 0.45 1.999 32 2.5 1.500
3 0.02 1.087 13 0.14 1.448 23 0.5 1.975 33 3.0 1.439
4 0.03 1.042 14 0.15 1.510 24 0.6 1.931 34 3.5 1.387
5 0.04 1.035 15 0.16 1.573 25 0.7 1.891 35 4.0 1.341
6 0.05 1.047 16 0.18 1.707 26 0.8 1.855 36 4.5 1.301
7 0.06 1.071 17 0.2 1.833 27 0.9 1.822 37 5.0 1.265
8 0.07 1.103 18 0.25 1.954 28 1.0 1.791
9 0.08 1.141 19 0.3 2.034 29 1.25 1.724
10 0.09 1.184 20 0.35 2.052 30 1.5 1.667

PGA, peak ground acceleration.

Table 5
Adjustment Factors for Nonlinear Site Model
Site Classes Site Classes
Period Number Period (s) I II III IV Period Number Period (s) I II III IV
1 PGA 1.0 1.0 1.0 1.0 17 0.2 0.0 0.565 0.650 1.006
2 0.01 1.0 1.0 1.0 1.0 18 0.25 0.0 0.601 0.479 1.027
3 0.02 1.0 1.0 1.0 1.05 19 0.3 0.0 0.579 0.449 1.021
4 0.03 1.0 1.0 1.0 0.58 20 0.35 0.0 0.679 0.482 1.003
5 0.04 1.0 1.006 1.0 0.482 21 0.4 0.0 0.655 0.499 1.010
6 0.05 1.0 0.851 1.0 0.472 22 0.45 0.0 0.615 0.515 0.985
7 0.06 1.0 0.803 1.044 0.506 23 0.5 0.0 0.550 0.530 0.990
8 0.07 1.0 0.918 0.975 0.587 24 0.6 0.0 0.0 0.530 1.006
9 0.08 1.0 1.062 0.964 0.683 25 0.7 0.0 0.0 0.499 1.000
10 0.09 1.0 1.106 0.980 0.782 26 0.8 0.0 0.0 0.369 1.000
11 0.1 1.0 1.071 0.970 0.823 27 0.9 0.0 0.0 0.3 0.960
12 0.12 0.0 0.952 1.022 1.029 28 1.0 0.0 0.0 0.2 0.904
13 0.14 0.0 0.672 0.889 0.991 29 1.25 0.0 0.0 0.0 0.738
14 0.15 0.0 0.631 0.861 0.983 30 1.5 0.0 0.0 0.0 0.535
15 0.16 0.0 0.600 0.831 0.973 31 2.0 0.0 0.0 0.0 0.358
16 0.18 0.0 0.571 0.748 0.979 32 2.5 0.0 0.0 0.0 0.0

SNC
} Parameters α = 2:0 and β = 0:6 were used for all
e F
α periods, and Amax, SRC, and Imf are the maximum
e Nmax e
1 amplification
log (A )log (SReffC +β)−log
(S )ln(β)
= exp ; −β ratio, crossover rock-site spectrum, and impedance ratio
loge(Amax) factor, respectively, of the 1D models defined by Zhao and
(10) Rhoades (2014) and Zhao, Hu, et al. (2015). All parameters
are presented in tables 4 and 5 in Zhao, Zhou, et al. (2016),
with 12 spectral periods being added and smoothed to the
SNC
SMR = Reff
f SR ; (11) model by Zhao, Hu, et al. (2015). An adjustment factor fSR
SReffC is introduced so that a broadly smoothed spectrum for non-
S
linear soil site can be obtained. This parameter has a value
SReff = SRImf; (12) in the 0–1.106 range, as shown in Table 5 of this article.
The adjustment factor fSR is zero for spectral periods over
2.5 s, which means that only linear amplification ratios are
SReffC = SRCImf; (13) necessary.
The adjustment factor fSR was determined in the
and follow- ing manner:
 AN max
SF = : 1. Select an expected largest magnitude, such as Mw 8.5 or
max (14) the magnitude of the largest event in the dataset; a
A

Figure 6. The effect of site information quality: the increase in
the weighted maximum log likelihood after excluding the strong-
motion records from stations with inferred site class.
shortest possible source distance for most subduction
zones, such as 25 km; and a possible fault depth that is
consistent with the source distance (e.g., the distance
must be less than the fault depth).
2. Set fSR 1:0
= for all spectral periods for which a nonlin-
ear site term is required.
3. Fit a smoothed curve to the calculated spectrum.
4. Manually adjust fSR so that the calculated spectrum
equals the smoothed spectrum derived in the last step.
When Amax is less than 1.25, SReffC can be calculated by
equation (13) in Zhao, Hu, et al. (2015). When AN max is less
than 1.25, SNC can be calculated by equation (15) in Zhao,
Hu, et al. (2015).
Model Coefficients and Standard Deviations
We established two GMPEs with identical functional
forms for each dataset. The first GMPE used the strong-
motion records in the first dataset, and the second GMPE used
the second dataset, excluding those records from the sites
with an inferred site class. We used MLLs from the two mod-
els to identify the effect of excluding the sites with an in-
ferred site class. Figure 6 shows the differences between
the weighted MLL (defined by Zhao, Zhou, et al., 2016)
from the second dataset and those from the first dataset.
Zhao, Zhou, et al. (2016) referred to the differences as the site
information quality effect, with a positive value suggesting
that a better model can be derived by excluding the sites
with an inferred site class. The increase varies between 18
and 75, suggesting a better model by excluding the sites
with an inferred site class.
We used the first dataset to determine the magnitude-
scaling rate dSL for large slab events with Mw≥7:1. We then
used the second dataset to determine the other terms for the
GMPE presented in this study. Figure 7 shows the values of
dSL and the ratio of dSL=cSL1. The magnitude-scaling rates
Figure 7. Magnitude-scaling coefficient dSL for large subduc-
tion slab events and the ratio dSL=cSL1. The color version of this
figure is available only in the electronic edition.
for events with Mw ≥7:0 are much smaller than those for the
smaller events.
The strategy of determining model parameters and
smoothing the model coefficients is presented in the Zhao,
Liang, et al. (2016) study.
Tables 6 and 7 present the smoothed coefficients in the
model for the GMPEs described in equations (1)–(5). Figure 8
shows the magnitude-squared term cSL2. This coefficient is
positive, varying between 0.0454 at spectral periods over
2.5 s and 0.507 at 0.07 s. The subduction slab model by
Zhao, Zhang, et al. (2006) has a magnitude-squared term for
the slab events. This term is also positive, leading to a rapid
increase in the predicted spectrum with increasing magnitude
when magnitude is large.
Figure 9 compares the anelastic attenuation rate,
expressed as a percentage, with that of the depth-dependent
anelastic attenuation rate for events with a depth over
50.0 km (calculated at 150 km depth) and with the attenu-
ation rate for the volcanic path. The volcanic attenuation
rate is much larger than the other two terms for spectral
periods up to about 1.0 s. The depth-dependent anelastic
attenuation rate is not large compared with the linear
anelastic attenua- tion rate at many spectral periods.
Figure 10 shows the site class terms, which are similar to
those for the shallow crustal and upper-mantle events pre-
sented by Zhao, Zhou, et al. (2016) and to those for the sub-
duction interface model by Zhao and Rhoades (2014) and
Zhao, Liang, et al. (2016) at some spectral periods.
Ⓔ Figures S1–S14, available in the electronic supple-
ment to this article, show the distribution of the between-
event and within-event residuals for PGA, 0.5, 1.0, 2.0, 3.0,
4.0, and 5.0 s spectral periods. Linear trend lines are also
presented. The slope of each trend line seems to be satisfac-
torily small. For each spectral period, we fitted a linear
func- tion of magnitude and source depth to the between-
event residuals and a linear function of magnitude and
source distance to the within-event residuals. The
coefficients of the linear functions should not differ from
0.0 at a significance
 Table 6
Model Coefficients, Part 1
T (s) c1 cSL1 cSL2 dSL bSL gSL gSLL eVSL
PGA −5.30119 1.44758 0.37625 0.42646 0.01826 −1.98471 1.12071 −0.01499
0.01 −5.28844 1.45400 0.38099 0.42075 0.01826 −1.96360 1.03278 −0.01503
0.02 −5.27568 1.46625 0.39101 0.40055 0.01826 −1.91839 0.94715 −0.01517
0.03 −5.26822 1.49246 0.41976 0.36433 0.01826 −1.89271 0.93420 −0.01567
0.04 −5.26293 1.50129 0.45746 0.32072 0.01826 −1.87260 0.97168 −0.01616
0.05 −5.25882 1.51051 0.48601 0.30000 0.01826 −1.85351 1.01492 −0.01676
0.06 −5.25547 1.51380 0.50311 0.31147 0.01826 −1.83395 1.06854 −0.01722
0.07 −5.25263 1.51111 0.50704 0.32673 0.01826 −1.81345 1.13401 −0.01752
0.08 −5.25017 1.50406 0.50004 0.34289 0.01826 −1.79189 1.20364 −0.01768
0.09 −5.24801 1.49423 0.48071 0.35921 0.01826 −1.76931 1.25808 −0.01772
0.1 −5.24607 1.48300 0.45759 0.37000 0.01826 −1.74581 1.30112 −0.01768
0.12 −5.24271 1.45559 0.41355 0.40606 0.01826 −1.73746 1.39137 −0.01742
0.14 −5.23988 1.44277 0.37828 0.43450 0.01826 −1.74463 1.47084 −0.01700
0.15 −5.23861 1.43314 0.36308 0.45000 0.01826 −1.74972 1.50784 −0.01676
0.16 −5.23742 1.43253 0.34919 0.46055 0.01826 −1.76259 1.54326 −0.01649
0.18 −5.23525 1.43710 0.32464 0.48439 0.01826 −1.78989 1.60985 −0.01594
0.2 −5.23331 1.44781 0.30358 0.50900 0.01826 −1.82110 1.67146 −0.01537
0.25 −5.22921 1.48260 0.26174 0.55500 0.01826 −1.90412 1.80738 −0.01395
0.3 −5.22585 1.51881 0.23036 0.59300 0.01826 −1.98439 1.92242 −0.01261
0.35 −5.22302 1.55291 0.20580 0.62500 0.01826 −2.05756 2.02102 −0.01139
0.4 −5.22056 1.58443 0.18597 0.65200 0.01826 −2.12282 2.10642 −0.01029
0.45 −5.21839 1.61360 0.16960 0.67500 0.01826 −2.18047 2.18097 −0.00931
0.5 −5.21645 1.64075 0.15585 0.69500 0.01826 −2.23118 2.24651 −0.00843
0.6 −5.21310 1.69020 0.13405 0.72900 0.01826 −2.31475 2.35602 −0.00694
0.7 −5.21026 1.73450 0.11757 0.75600 0.01826 −2.37885 2.44331 −0.00574
0.8 −5.20781 1.77474 0.10476 0.77800 0.01826 −2.42769 2.51391 −0.00477
0.9 −5.20564 1.81162 0.09458 0.79600 0.01826 −2.46450 2.57166 −0.00398
1 −5.20370 1.84561 0.08636 0.81200 0.01826 −2.49170 2.61931 −0.00333
1.25 −5.19959 1.92015 0.07173 0.84100 0.01808 −2.52758 2.70638 −0.00215
1.5 −5.19624 1.98274 0.06258 0.86100 0.01786 −2.53359 2.76244 −0.00142
2 −5.19095 2.08214 0.05327 0.88400 0.01718 −2.49565 2.82205 −0.00067
2.5 −5.18684 2.15841 0.05036 0.90000 0.01628 −2.42623 2.84475 −0.00039
3 −5.18349 2.22046 0.04536 0.90000 0.01549 −2.34726 2.84988 −0.00030
3.5 −5.18065 2.27406 0.04536 0.90000 0.01489 −2.27002 2.84667 −0.00026
4 −5.17819 2.32307 0.04536 0.90000 0.01458 −2.19947 2.83992 −0.00021
4.5 −5.17602 2.37009 0.04536 0.90000 0.01459 −2.12528 2.82802 −0.00021
5 −5.17409 2.37009 0.04536 0.90000 0.01459 −2.02646 2.82521 −0.00021

msc = 6:3.

deviations of the
level of 5%. If this criterion was not satisfied, a new regres-
sion analysis was carried out.
Figure 11 shows the variation of standard deviations
with spectral period. The largest between-event standard
deviation is 0.598 at 0.07–0.08 s. The largest within-event
standard deviation is 0.713 at 0.2 s, and the largest total
stan- dard deviation is 0.884 at spectral periods of 0.08–0.09 s.
The standard deviations are larger than those from the other
cat- egories of events (Zhao and Rhoades, 2014; Zhao,
Liang, et al., 2016; Zhao, Zhou, et al., 2016) at many
spectral peri- ods. Figure 12a compares the interevent and
total standard deviations from the present study with those
from the Abra- hamson et al. (2015) model (digitized from
fig. 7 in that study), in which the subduction interface and
slab events were combined together as a single group. The
between- event standard deviations of the two models are
very similar at many spectral periods. The total standard
present study are larger than those of the
Abrahamson et al. (2015) study in the 0.04–2.5 s
period range, with the largest difference being about
18%. Figure 12b shows the within- event standard
deviation. The values from the present study are
considerably larger than those of the Abrahamson et
al. (2015) model; the largest difference is about 19%.
However, the standard deviations from subduction
interface events in the Zhao, Liang, et al. (2016) study
are considerably smaller than those for slab events in
the present study for spectral periods up to 2 s. This
may mean that the smaller standard deviations in the
Abrahamson et al. (2015) study are caused by the
interface events, whereas the data in the present
study are all from slab events. Zhao, Liang, et al.
(2016) suggested the possibility that the use of VS30 in
the Abrahamson et al. (2015) study may lead to a
reduction in within-event standard deviations. This
cannot be confirmed without comparing between-
site standard deviations from the two studies.
Figure 8. The magnitude-squared term for slab events with a
magnitude less than 7.1. The color version of this figure is
available only in the electronic edition.
Figure 9. Comparison of anelastic attenuation rates for slab
events, the depth-dependent anelastic attenuation rate calculated
at a depth of 150 km, and the anelastic attenuation rates for
volcanic path. The color version of this figure is available only in
the elec- tronic edition.

Table 7
Model Coefficients, Part 2
Period (s) eSL eSLH γ S2 S3 S4 σ τ σT
PGA −0.00340 −0.00050 −9.880 0.2320 0.1437 0.1470 0.587 0.457 0.744
0.01 −0.00331 −0.00050 −9.513 0.2289 0.1398 0.1328 0.587 0.458 0.745
0.02 −0.00345 −0.00050 −9.266 0.2183 0.1260 0.1443 0.587 0.465 0.749
0.03 −0.00391 −0.00050 −9.332 0.1874 0.0616 0.0660 0.588 0.480 0.759
0.04 −0.00454 −0.00050 −9.508 0.1233 −0.0171 −0.0171 0.599 0.521 0.794
0.05 −0.00510 −0.00050 −9.729 0.0721 −0.0633 −0.0731 0.607 0.555 0.823
0.06 −0.00552 −0.00050 −9.966 0.0270 −0.1010 −0.1196 0.623 0.584 0.854
0.07 −0.00588 −0.00049 −10.226 −0.0062 −0.1468 −0.1601 0.638 0.600 0.876
0.08 −0.00615 −0.00048 −10.551 0.0157 −0.1448 −0.1243 0.651 0.598 0.884
0.09 −0.00635 −0.00048 −10.807 0.0509 −0.1267 −0.0729 0.662 0.585 0.883
0.1 −0.00652 −0.00048 −11.022 0.0956 −0.0932 −0.0146 0.674 0.567 0.881
0.12 −0.00660 −0.00049 −11.365 0.2004 −0.0088 0.0825 0.689 0.534 0.872
0.14 −0.00652 −0.00051 −11.730 0.3037 0.0893 0.1715 0.692 0.504 0.856
0.15 −0.00647 −0.00052 −11.880 0.3428 0.1360 0.2093 0.696 0.486 0.849
0.16 −0.00636 −0.00053 −12.056 0.3740 0.1775 0.2412 0.697 0.465 0.838
0.18 −0.00614 −0.00056 −12.420 0.4270 0.2531 0.2990 0.704 0.430 0.825
0.2 −0.00590 −0.00059 −12.785 0.4630 0.3201 0.3459 0.713 0.406 0.821
0.25 −0.00526 −0.00067 −13.635 0.5086 0.4530 0.4423 0.711 0.385 0.808
0.3 −0.00468 −0.00075 −14.381 0.5078 0.5488 0.5178 0.684 0.365 0.775
0.35 −0.00415 −0.00083 −15.035 0.4971 0.6171 0.5760 0.665 0.371 0.762
0.4 −0.00369 −0.00091 −15.616 0.4807 0.6663 0.6224 0.657 0.383 0.761
0.45 −0.00327 −0.00099 −16.138 0.4616 0.7011 0.6598 0.647 0.391 0.756
0.5 −0.00290 −0.00107 −16.613 0.4422 0.7256 0.6907 0.640 0.403 0.756
0.6 −0.00227 −0.00124 −17.453 0.4054 0.7529 0.7380 0.633 0.412 0.755
0.7 −0.00178 −0.00139 −18.181 0.3734 0.7625 0.7723 0.632 0.432 0.766
0.8 −0.00139 −0.00154 −18.825 0.3462 0.7612 0.7974 0.635 0.438 0.772
0.9 −0.00109 −0.00166 −19.403 0.3236 0.7538 0.8162 0.636 0.438 0.772
1 −0.00086 −0.00178 −19.928 0.3048 0.7428 0.8301 0.636 0.439 0.773
1.25 −0.00052 −0.00199 −21.058 0.2703 0.7083 0.8504 0.635 0.444 0.775
1.5 −0.00043 −0.00213 −21.996 0.2483 0.6726 0.8573 0.645 0.448 0.786
2 −0.00070 −0.00225 −23.488 0.2253 0.6107 0.8499 0.633 0.425 0.762
2.5 −0.00127 −0.00219 −24.647 0.2154 0.5640 0.8276 0.607 0.413 0.735
3 −0.00198 −0.00207 −25.597 0.2115 0.5261 0.7991 0.582 0.407 0.710
3.5 −0.00271 −0.00193 −26.410 0.2098 0.4977 0.7678 0.562 0.395 0.687
4 −0.00341 −0.00180 −27.132 0.2088 0.4769 0.7359 0.540 0.381 0.661
4.5 −0.00421 −0.00170 −27.793 0.2077 0.4622 0.7041 0.526 0.367 0.641
5 −0.00500 −0.00158 −28.313 0.2067 0.4527 0.6722 0.522 0.378 0.645
Figure 10. Comparison of elastic site terms for three site similar to the within-site standard de-
classes. The color version of this figure is available only in the
elec- tronic edition.

Figure 11. Variation of between-event standard deviation (τ),

within-event standard deviation (σ), and total standard deviation
(σT ) with spectral period. The color version of this figure is avail-
able only in the electronic edition.

Table 8 presents the within-site and between-site stan-

dard deviations derived from the within-event residuals using
a random effects model described in equation (6). Figure
13a shows the between-site standard deviations, and Figure
13b shows the within-site standard deviations. SC I sites
have the largest between-site standard deviation for spectral
periods (up to 0.16 s) compared with those for the other site
classes. The largest value is 0.643 at a spectral period of 0.08
s, which is close to the average site period of SC I sites. SC
II sites have the second largest between-site standard
deviations at spectral periods up to 0.15 s. The largest value
is 0.601 at about 0.18–0.25 s spectral period. The between-
site standard deviations for SC III sites are similar to those
of SC IV sites up to about 0.06 s and smaller than those for
SC IV sites in the 0.07–0.3 s spectral period range. At long
spectral periods over 0.5 s, the between-site standard
deviations are very sim- ilar among all site classes. Figure
13b shows that the within- site standard deviations among
the four site classes are very similar for all site classes.
These do not vary with spectral periods as much as the
between-site standard deviations do. In this aspect, they are
Figure 12. Comparison of (a) between-event and total
standard deviations and (b) within-event standard deviations
from the present study and the Abrahamson et al. (2015)
study. The color version of this figure is available only in
the electronic edition.

viations from shallow crustal and upper-mantle

events in Zhao, Zhou, et al. (2016). Figure 14
compares the total site standard deviations σST for all
site classes with the within- event standard deviation
σ. The total site standard deviation σST is larger than
σ at very short periods but is very similar to σ at
spectral periods over 1.0 s. At short periods, SC I
sites have the largest total site standard deviation;
the largest dif- ference among the four site classes is
close to 20%. For SC I, II, and III sites, the total site
standard deviation has the larg- est peak value at a
spectral period close to the average site period of
each site class, suggesting that site resonance
response may increase model variability.
We also evaluated whether the model standard devia-
tions depend on earthquake magnitude. We divided
the residuals into magnitude bins, each of which covers
0.5 mag- nitude units if there is a reasonable number
of events for between-event residuals and a
reasonable number of records for within-event
residuals. Then we calculated the standard deviations
of the residuals in each magnitude bin. Figure 15a
shows the standard deviations for the between-event
resid- uals in four magnitude bins, with the last bin
containing all events with Mw > 6:5 so as to have
enough events. Clearly, the between-event standard
deviations do not decrease with increasing
magnitude in a consistent manner, as shown in
Figure 15a. Figure 15b shows that the standard
deviations
 Table 8
Within-Site and Between-Site Standard Deviations
SC I SC II SC III SC IV
Period (s) σS1 τS1 σST1 σS2 τS2 σST2 σS3 τS3 σST3 σS4 τS4 σST4
PGA 0.398 0.511 0.648 0.417 0.449 0.613 0.409 0.431 0.594 0.415 0.422 0.592
0.01 0.397 0.517 0.651 0.417 0.450 0.614 0.409 0.431 0.594 0.415 0.418 0.589
0.02 0.395 0.518 0.652 0.417 0.449 0.613 0.408 0.431 0.594 0.416 0.425 0.594
0.03 0.389 0.537 0.663 0.418 0.449 0.614 0.409 0.430 0.593 0.417 0.422 0.593
0.04 0.387 0.572 0.691 0.420 0.456 0.620 0.413 0.428 0.595 0.420 0.431 0.602
0.05 0.387 0.586 0.702 0.422 0.479 0.639 0.409 0.429 0.592 0.422 0.439 0.609
0.06 0.397 0.613 0.730 0.416 0.507 0.656 0.401 0.449 0.602 0.423 0.445 0.614
0.07 0.403 0.633 0.750 0.412 0.532 0.673 0.394 0.456 0.603 0.420 0.473 0.633
0.08 0.413 0.643 0.765 0.412 0.555 0.692 0.389 0.455 0.599 0.423 0.491 0.648
0.09 0.422 0.631 0.759 0.412 0.568 0.702 0.391 0.475 0.615 0.427 0.518 0.671
0.1 0.429 0.626 0.759 0.418 0.566 0.704 0.394 0.507 0.642 0.426 0.559 0.703
0.12 0.440 0.614 0.755 0.429 0.570 0.713 0.428 0.548 0.695 0.445 0.576 0.728
0.14 0.441 0.613 0.755 0.435 0.597 0.739 0.433 0.507 0.667 0.442 0.561 0.714
0.15 0.450 0.602 0.752 0.440 0.599 0.743 0.420 0.490 0.645 0.440 0.555 0.708
0.16 0.452 0.599 0.750 0.444 0.597 0.744 0.424 0.494 0.651 0.437 0.552 0.704
0.18 0.454 0.599 0.752 0.447 0.601 0.749 0.446 0.500 0.670 0.436 0.553 0.705
0.2 0.462 0.590 0.749 0.450 0.596 0.747 0.441 0.491 0.660 0.432 0.562 0.709
0.25 0.474 0.555 0.730 0.470 0.601 0.763 0.459 0.450 0.643 0.432 0.507 0.666
0.3 0.472 0.532 0.711 0.475 0.557 0.732 0.437 0.499 0.663 0.432 0.489 0.653
0.35 0.468 0.505 0.688 0.478 0.518 0.705 0.444 0.532 0.692 0.432 0.463 0.633
0.4 0.457 0.479 0.663 0.484 0.488 0.687 0.453 0.546 0.709 0.416 0.466 0.624
0.45 0.450 0.461 0.644 0.477 0.479 0.676 0.477 0.529 0.713 0.408 0.461 0.616
0.5 0.445 0.451 0.633 0.470 0.471 0.665 0.472 0.495 0.684 0.409 0.467 0.620
0.6 0.448 0.436 0.625 0.460 0.467 0.655 0.459 0.470 0.657 0.407 0.434 0.594
0.7 0.440 0.430 0.615 0.460 0.464 0.653 0.461 0.473 0.660 0.403 0.430 0.589
0.8 0.441 0.427 0.614 0.460 0.456 0.647 0.457 0.457 0.646 0.407 0.454 0.610
0.9 0.435 0.431 0.612 0.456 0.466 0.652 0.449 0.439 0.628 0.408 0.456 0.612
1 0.427 0.440 0.613 0.448 0.466 0.647 0.442 0.442 0.625 0.408 0.462 0.616
1.25 0.413 0.440 0.603 0.442 0.476 0.649 0.428 0.432 0.608 0.412 0.444 0.606
1.5 0.416 0.449 0.612 0.448 0.473 0.651 0.418 0.461 0.622 0.418 0.443 0.609
2 0.409 0.441 0.601 0.437 0.462 0.636 0.406 0.472 0.623 0.413 0.439 0.603
2.5 0.398 0.425 0.582 0.429 0.427 0.606 0.387 0.484 0.620 0.414 0.427 0.595
3 0.390 0.396 0.556 0.423 0.407 0.587 0.369 0.447 0.580 0.412 0.431 0.596
3.5 0.386 0.386 0.545 0.410 0.401 0.574 0.376 0.434 0.574 0.402 0.413 0.576
4 0.377 0.373 0.530 0.410 0.386 0.564 0.362 0.406 0.544 0.394 0.391 0.555
4.5 0.360 0.364 0.512 0.412 0.373 0.556 0.368 0.383 0.531 0.385 0.366 0.532
5 0.361 0.359 0.509 0.447 0.322 0.551 0.381 0.316 0.495 0.380 0.324 0.499

Subscripts 1, 2, 3, and 4 with σ and τ denote site classes.

for the within-event residuals do not decrease or increase Mw 6 events

with magnitude either. Abrahamson et al. (2015) also
used a magnitude-independent standard deviation.

Predicted Response Spectra

Next, we present the predicted spectra for various mag-
nitude, depth, and distance ranges. The smoothing of each
model parameter with respect to the logarithm of spectral
periods does not lead to smoothed spectra at all magnitude
and distance ranges.
Figure 16 shows the rock-site spectra from events with
magnitudes of 5, 6, 7, and 8 at a source distance of 30 km
and a fault-top depth of 30 km, and the predicted PGAs are
listed in Table 9. Among the periods that have been
modeled, the peak rock-site spectrum is at a spectral period
of 0.1 s. The differences between the spectra for Mw 7 and
are considerably larger than those between the
spectra for Mw 6 and Mw 5 events at spectral periods
up to about 0.4 s because of the magnitude-squared
term.
Figure 17a shows the predicted spectra for SC I
sites for four magnitude units, and the peak of the
SC I spectrum is at about 0.1 s. Figure 17b presents
the predicted spectrum for SC II sites, and the peak
of the spectrum for all events is at
0.15 s. Figure 18a shows that the peak of the
predicted spec- trum is at about 0.15–0.16 s for SC
III sites, similar to that in Figure 17b. Figure 18b
shows the predicted spectrum for SC IV sites. The
corresponding PGAs for the predicted spectra in
Figures 17 and 18 are presented in Table 9.
Figure 19a compares the spectra from an Mw 7.0
event with a fault depth of 30 km at a source
distance of 30 km for rock sites and four soil site
classes. The PGA is 0.394g for rock site, 0.542g for
SC I sites, 0.651g for SC II sites, 0.577g for SC III
sites, and 0.553g for SC IV sites. The reduced
Figure 13. Comparison of (a) between-site standard deviations Figure 20a shows the attenuation of PGA, and Figure 20b
and (b) within-site standard deviations for four site classes. The
color version of this figure is available only in the electronic edition.

Figure 14. Comparison of site total standard deviations with

the within-event standard deviations. The color version of this figure
is available only in the electronic edition.

short-period spectra for SC III and SC IV sites are mainly

caused by the reduced elastic site term for these two site
classes compared with the elastic site term for SC II sites.
The nonlinear site terms for SC III and SC IV sites also
lead to the reduction in predicted spectrum at short
periods.
Figure 19b compares the nonlinear soil spectrum with
the elastic spectrum for an SC IV site from an Mw 8 event
with fault depth of 30 km at a source distance of 30 km.
The nonlinear soil spectrum has a PGA of 0:76g, reduced
from 1:04g in the elastic spectrum. The largest reduction
is at
0.16 s, and the nonlinear spectrum is 1:42g, reduced from
2:44g in the elastic spectrum.
Figure 15. Variation of standard deviations for (a)
between- event and (b) within-event residuals in a number
of magnitude bins. The color version of this figure is
available only in the electronic edition.

Figure 16. Predicted rock-site spectra for slab events with

Mw 5–8 and a depth of 30 km at a source distance of 30 km.
The color version of this figure is available only in the
electronic edition.

Table 9
Predicted PGA (g) for Four Magnitude Units
and Five Site Classes
Magnitude (Mw)
 Site Class 5.0 6.0 7.0 8.0
Rock site 0.071 0.136 0.394 0.651
SC I 0.099 0.187 0.542 0.893
SC II 0.124 0.235 0.651 0.997
SC III 0.113 0.214 0.577 0.845
SC IV 0.114 0.213 0.553 0.760
shows the spectrum at 0.5 s from events with magnitudes of
 Figure 17. Predicted spectra for slab events with Mw 5–8 and the subduction slab events as one group. We found that
a depth of 30 km at a source distance of 30 km for (a) SC I sites
and
(b) SC II sites. The color version of this figure is available only in
the electronic edition.

5, 6, 7, and 8, and with a fault-top depth of 25 km for SC II

sites. The distance range is 30–320 km. These figures clearly
show the reduced scaling for events with Mw > 7:0 and the
increased magnitude scaling for the events with a magnitude
between 6.0 and 7.0 as a result of the magnitude-squared
term. The magnitude scaling between magnitudes 5 and 6 is
also markedly smaller than that between magnitudes 6 and 7.
Figure 21 shows the attenuation of spectra at 1.0 s and 3.0 s
spectral periods; again, the magnitude-scaling rates for large
events are significantly less than those for events with an Mw
up to 7.1. The magnitude-scaling rate for magnitude 5 and 6
events is similar to that for magnitude 6 and 7 events because
the magnitude-squared term is small for these spectral periods.
Figure 22 shows the effect of volcanic path on the pre-
dicted response spectra for an event with Mw 8.0 and a depth
of 30 km at a source distance of 65 km. The PGA is reduced
from 0:372g to 0:278g at a volcanic distance of 20 km,
0:207g at 40 km, and 0:153g at 60 km. The reduction is sig-
nificant for spectral periods up to about 0.7 s.

Conclusions
We assembled 4710 strong-motion records from sub-
duction slab events in Japan up to the end of 2012. The
large number of records allows us to develop a GMPE for
Figure 18. Predicted spectra for slab events with Mw 5–8
and a depth of 30 km at a source distance of 30 km for (a)
SC III sites and
(b) SC IV sites. The color version of this figure is
available only in the electronic edition.

the standard deviations from the present study are

larger than those from the shallow crustal and upper-
mantle events (Zhao, Zhou, et al., 2016) and the
subduction interface events (Zhao, Liang, et al.,
2016). The different model stan- dard deviations
may improve the seismic-hazard estimation for each
type of earthquake.
The model in the present study adopted a
bilinear magnitude-scaling function hinged at Mw
7.1, leading to a considerably reduced magnitude
scaling for events with an Mw 7:1. The coefficient ≥
for this magnitude term was much smaller than that
for events with Mw < 7:1. We also adopted a
magnitude-squared term for events with a magnitude
up to 7.1. The effect of volcanoes on the at-
tenuation of seismic waves was modeled by applying
an anelastic attenuation term to the volcanic path, the
horizon- tal portion of a straight-line distance (the
closest distance from a station to the fault plane if
available) that passes through the assumed low-Q
zones around the active volca- noes. The absolute
values of the anelastic attenuation rate for volcanic
path are much larger than the anelastic attenu- ation
rate at spectral periods up to 1.0 s. A volcanic path
of 60 km can lead to nearly 60% reduction in the
predicted PGA.
Site classes based on site period were used for
the site term. We tested the effect of site information
quality. The over-
 (a) 1
PGA

Acceleration spectrum (g)

0.1

0.01

0.001
Mw=5
Mw=6 Slab SC II
Mw=7
0.0001 Mw=8
20 40 80
Source distance (km) 160 320

(b) (b)
1

Acceleration spectrum (g)

0.5s
0.1

0.01

Mw=5
0.001 Mw=6 Slab SC II
Mw=7

0.0001 Mw=8
20 40 80 160 320
Source distance (km)

Figure 20. Attenuation of predicted response spectra for four

Figure 19. (a) Predicted spectra from slab events with Mw 7.0 magnitude units of subduction slab events with a fault-top depth
and a depth of 30 km at a source distance of 30 km for four site of 25 km at SC II sites for (a) PGA and (b) 0.5 s. The color
classes and (b) the elastic and nonlinear spectra for SC IV sites from version of this figure is available only in the electronic edition.
an Mw 8.0 event. The color version of this figure is available only
in the electronic edition.
effects model. We found that the between-site standard
deviations are generally larger than the within-site standard
all goodness of fit for the attenuation models improved signifi- deviations for spectral periods up to 0.6 s. The between-site
cantly after records from sites with inferred site classes standard deviations tend to vary significantly with site class
were excluded. The model coefficients presented in this and spectral period. The values at spectral periods close to
study were derived based on the records from sites that the average site period of each site class tend to be large.
have a measured shear-wave velocity profile down to The within-site standard deviations from the four site
engineering bedrock. classes are generally similar and do not vary significantly
The model standard deviations derived from the with spectral periods.
present study are generally similar to those by
Abrahamson et al.
(2015) at some periods, but the within-event standard without comparing between-site standard deviations.
devia- We approximately separated the within-event residuals
tions are significantly larger than those from Abrahamson into within-site and between-site residuals using a random
et al. (2015) in the 0.2–2.5 s period range. One possible
explanation is that subduction slab events may have larger
standard deviations, because the Abrahamson et al. (2015)
model used both interface and slab events whereas the
present study is for slab events only. Compared with the re-
sults from the Zhao, Liang, et al. (2016) model for interface
events, the slab events do have substantially larger standard
deviations. Another possible explanation is that
Abrahamson et al. (2015) used VS30 as the site term,
whereas site classes are used in the present study. VS30 may
lead to a reduced within-event standard deviation compared
with site classes. However, this cannot be confirmed
 We investigated whether model standard
deviations depend on magnitude by comparing the
standard deviations computed from the residuals in a
number of magnitude bins. We found that these
standard deviations do not depend on magnitude,
which is consistent with the Abrahamson et al.
(2015) study.
We adopted a nonlinear site model based on 1D
model- ing. The nonlinear site term reduced the
elastic spectrum sig- nificantly, up to the 0.7 s period
for records from large events at short distances.

Data and Resources

The strong-motion records are from K-NET and
KiK- net, administered by the National Research
Institute for Earth Science and Disaster Prevention of
Japan. A small number of records are from the Port
and Airport (Port and Harbour) Re-
 (a) 1 1.28
0km
1.0s

Acceleration spectrum (g)

0.64 20km
Acceleration spectrum (g)

0.1 40km
0.32 60km

0.01
0.16

0.08 Slab
0.001 Mw=5
MW =8.0
Mw=6
Slab
Mw=7 0.04 Depth=30km
0.0001 Mw=8 SC II Dist.=67km
20 40 80 160 320 0.02
Source distance (km) 0.02 0.04 0.08 0.16 0.32 0.64 1.28 2.56 5.12
Spectral period (s)

(b) 1 3.0s
Slab SC II Figure 22. Predicted response spectra from a slab event with an
Mw 8.0 and a depth of 30 km at a source distance of 65 km, with
0.1
Acceleration spectrum (g)

volcanic travel paths of 0, 20, 30, 40, and 60 km. The color
version of this figure is available only in the electronic edition.
0.01

0.001 Abrahamson, N. A., and W. J. Silva (2008). Summary of the Abrahamson & Silva
Mw=5
Mw=6 NGA ground-motion relations, Earthq. Spectra 24, no. 1, 67–97.
0.0001
Mw=7
Mw=8
0.00001
20 40 80 160 320
Source distance (km)

Figure 21. Attenuation of predicted response spectra for four

magnitude units of subduction slab events with a fault-top depth
of 25 km at SC II sites for (a) 1.0 s and (b) 3.0 s. The color
version of this figure is available only in the electronic edition.

search Institute. A number of rock-site strong-motion records

are from the Pacific Earthquake Engineering Research
Center strong-motion database
(http://peer.berkeley.edu/smcat/; last accessed August
2015).

Acknowledgments
The work reported here is partially supported by research grants
from the National Science Foundation of China (51278432) and the
Southwest Jiaotong University (SWJTU12ZT04), 973 Project from the
Ministry of Science of China (2013CB036204), and by the New Zealand
Earthquake Commission 2010 Biennial Research Grant. At an early stage
(2011), sup- port was received from the New Zealand Foundation for
Research Science and Technology, New Zealand Hazards Platform
(Contract C05X0907). The authors would like to thank Jim Cousins and
Chris Van Houtte of GNS Science for their review of this manuscript. We
would like to thank Kimiyuki Asano from Disaster Prevention Research
Institute (DPRI) of Kyoto University for supplying the fault model
parameters for a number of earthquakes. Finally, we would like to thank
Eric Thompson and two anonymous reviewers for their constructive
review comments.

References
Abrahamson, N. A., and R. R. Youngs (1992). A stable algorithm
for regression analysis using the random effect model, Bull.
Seismol. Soc. Am. 82, no. 1, 505–510.
Abrahamson, N. A., N. Gregor, and K. Addo (2015). BC Hydro
ground motion prediction equations for subduction
earthquakes, Earthq. Spectra 32, no. 1, 23–44, doi:
10.1193/051712EQS188MR.
Atkinson, G. M., and D. M. Boore (2003). Empirical ground-
motion relations for subduction-zone earthquakes and their
application to Cascadia and other regions, Bull. Seismol.
Soc. Am. 93, no. 4, 1703–1729.
Boore, D. M., and G. M. Atkinson (2008). Ground-motion
prediction equa- tions for the average horizontal component
of PGA, PGV, and 5%- damped PSA at spectral periods
between 0.01 s and 10.0 s, Earthq. Spectra 24, no. 1, 99–
138.
Building Seismic Safety Council (BSSC) (2000). The 2000
NEHRP recommended provisions for new buildings and
other structures, Part I (Provisions) and Part II
(Commentary), Federal Emergency Manage- ment Agency,
368/369, Washington D.C.
Campbell, K. W., and Y. Bozorgnia (2008). NGA ground motion
model for the geometric mean horizontal component of PGA,
PGV, PGD and 5% damped linear elastic response spectra for
periods ranging from 0.01 to 10 s, Earthq. Spectra 24, no. 1,
139–171.
Chiou, B. S.-J., and R. R. Youngs (2008). An NGA model for the
average of horizontal component of peak ground motion and
response spectra, Earthq. Spectra 24, no. 1, 173–216.
Eberhart-Phillips, D., and G. H. McVerry (2003). Estimating slab
earthquake response spectra from a 3D Q model, Bull. Seismol.
Soc. Am. 93, no. 6, 2649–2663.
Ghofrani, H., and G. M. Atkinson (2014). Ground-motion
prediction equations for interface earthquakes of M7 to M9
based on empirical data from Japan, Bull. Earthq. Eng. 12,
no. 2, 549–571.
Kanno, T., A. Narita, N. Morikawa, H. Fujiwara, and Y.
Fukushima (2006). A new attenuation relation for strong
ground motion in Japan based on recorded data, Bull.
Seismol. Soc. Am. 96, no. 3, 879–897.
McVerry, G. H., J. X. Zhao, N. A. Abrahamson, and P. G. Somerville
(2006). Crustal and subduction zone attenuation relations for
New Zealand earthquakes, Bull. New Zeal. Soc. Earthq.
Eng. 39, no. 1, 1–58.
Standards New Zealand (2004). Structural Design Actions–Part 5
Earthquake Actions–New Zealand, New Zealand Standard
NZS 1170.5:2004.
Youngs, R. R., S.-J. Chiou, W. J. Silva, and J. R. Humphrey
(1997). Strong ground motion attenuation relationships for
subduction zone earthquakes, Seismol. Res. Lett. 68, no. 1,
58–73.
Zhao, J. X. (2010). Geometric spreading functions and modelling
of volcanic zones for strong-motion attenuation models
derived from records in Japan, Bull. Seismol. Soc. Am. 100,
no. 2, 712–732.
GMPEs for Subduction Slab Earthquakes in Japan 17

Zhao, J. X. (2014). Magnitude scaling rates for large earthquakes in

School of Civil Engineering
ground motion prediction equations for Japan, Annual Meeting of
Southwest Jiaotong University
Seismologi- cal Society of America, Anchorage, Alaska, 30 April–2 111 1st Northern Section of Erhuan Road
May 2014. Chengdu 610031
Zhao, J. X., and M. Lu (2011). Magnitude-scaling rate in ground-motion Sichuan, China
prediction equations for response spectra from large shallow crustal J.Zhao@gns.cri.nz
earthquakes, Bull. Seismol. Soc. Am. 101, no. 6, 2643–2661. (J.X.Z., F.J., P.S., H.X., H.H., R.H., Y.Z., P.Y.)
Zhao, J. X., and D. A. Rhoades (2014). Ground-motion prediction
equations for subduction zones based on strong-motion records from
Japan, GNS Science Consultancy Report 2014/236, 161 pp.
Zhao, J. X., and H. Xu (2012). Magnitude-scaling rate in ground-motion Institute of Crustal Dynamics
prediction equations for response spectra from large subduction China Earthquake Administration
1 Anningzhuang Road
interface earthquakes in Japan, Bull. Seismol. Soc. Am. 102, no. 1,
Haidian District
222–235. Beijing 100085, China
Zhao, J. X., and J. Zhang (2010). Side-effect of using response spectral (X.L.)
amplification ratios for soft soil sites—Earthquake source-type
dependent amplification ratios, Soil Dynam. Earthq. Eng. 30, 258–
269.
Zhao, J. X., J. S. Hu, F. Jiang, J. Zhou, Y. Zhang, X. An, M. Lu, and D. A. GNS Sciences
Rhoades (2015). Nonlinear site models derived from 1-D analyses 1 Fairway Drive
for ground-motion prediction equations using site class as the site Avalon
param- eter, Bull. Seismol. Soc. Am. 105, no. 4, 2010–2022. Lower Hutt 5010
New Zealand
Zhao, J. X., K. Irikura, J. Zhang, Y. Fukushima, P. G. Somerville, A.
(D.A.R.)
Asano,
Y. Ohno, T. Oouchi, T. Takahashi, and H. Ogawa (2006). An empirical
site-classification method for strong-motion stations in Japan
using H/V response spectral ratio, Bull. Seismol. Soc. Am. 96, no. 3, AECOM
914–925. 915 Wilshire Boulevard, 7th Floor
Zhao, J. X., X. Liang, F. Jiang, H. Xing, Y. Zhang, C. Zhao, X. Lan, D. A. Los Angeles, California 90017
Rhoades, P. G. Somerville, K. Irikura, et al. (2016). Ground-motion (P.G.S.)
prediction equations for subduction interface earthquakes in Japan
using site class and simple geometric attenuation functions, Bull.
Seismol. Soc. Am. 106, no. 4, doi: 10.1785/0120150034.
Aichi Institute of Technology
Zhao, J. X., J. Zhang, A. Asano, Y. Ohno, T. Oouchi, T. Takahashi, H.
Aichi Prefecture
Ogawa,
Toyota 470-0392, Japan
K. Irikura, H. K. Thio, P. G. Somerville, et al. (2006). Attenuation (K.I.)
relations of strong ground motion in Japan using site classification
based on predominant period, Bull. Seismol. Soc. Am. 96, no. 3, 898–
913.
Zhao, J. X., J. Zhang, and K. Irikura (2009). Side-effect of using response International Seismic Safety Centre
spectral amplification ratios for soil sites—Variability and Division of Nuclear Installation Safety
earthquake- magnitude and source-distance dependent amplification Department of Nuclear Safety and Security
ratios for soil sites, Soil Dynam. Earthq. Eng. 29, 1262–1273. International Atomic Energy Agency
Zhao, J. X., S. L. Zhou, P. J. Gao, T. Long, Y. B. Zhang, H. K. Thio, M. Vienna International Centre
P.O. Box 100
Lu, and D. A. Rhoades (2015). An earthquake classification scheme
1400 Vienna
adapted for Japan determined by the goodness-of-fit for ground- Austria
motion prediction equation, Bull. Seismol. Soc. Am. 105, no. 5, (Y.F.)
doi: 10.1785/0120150013.
Zhao, J. X., S. Zhou, J. Zhou, C. Zhao, H. Zhang, Y. Zhang, P. Gao, X. Lan,
D. Rhoades, Y. Fukushima, et al. (2016). Ground-motion prediction Manuscript received 19 February 2015;
equations for shallow crustal and upper-mantle earthquakes in Japan Published Online 12 July 2016
using site class and simple geometric attenuation functions, Bull.
Seismol. Soc. Am. 106, no. 4, doi: 10.1785/0120150063.

BSSA Early Edition

View publication stats