The Professional Journal of the Earthquake Engineering Research Institute

PREPRINT
This preprint is a PDF of a manuscript that has been accepted for
publication in Earthquake Spectra. It is the final version that was
uploaded and approved by the author(s). While the paper has been
through the usual rigorous peer review process for the Journal, it has
not been copyedited, nor have the figures and tables been modified for
final publication. Please also note that the paper may refer to online
Appendices that are not yet available.
We have posted this preliminary version of the manuscript online in
the interest of making the scientific findings available for distribution
and citation as quickly as possible following acceptance. However,
readers should be aware that the final, published version will look
different from this version and may also have some differences in
content.

The DOI for this manuscript and the correct format for citing the paper are
given at the top of the online (html) abstract.
Once the final, published version of this paper is posted online, it will
replace the preliminary version at the specified DOI.

Turkey-Adjusted NGA-W1 Horizontal Ground

Motion Prediction Models

Zeynep Glerce

M.EERI,

Bahadr Kargolu and Norman A. Abrahamson

M.EERI

Corresponding (first) author: ZEYNEP GLERCE

Mailing address: MIDDLE EAST TECHNICAL UNIVERSITY CIVIL ENGINEERING
DEPARTMENT K1-308 ANKARA TURKEY 06531
Phone: (+90) 312 2107481
Fax: (+90) 312 2105401
E-mail address: zyilmaz@metu.edu.tr
Submission date for review copies: March 14, 2014
Submission date for revised copies: August 14, 2014 October 30, 2014
Submission date for camera-ready copy:

GLERCE-1

Turkey-Adjusted NGA-W1 Horizontal Ground

Motion Prediction Models
Zeynep Glerce a) M.EERI, Bahadr Kargolu a) and Norman A. Abrahamson
b)

M.EERI

The objective of this paper is to evaluate the differences between the Next
Generation Attenuation: West-1 (NGA-W1) ground motion prediction models
(GMPEs) and Turkish strong ground motion dataset and to modify the required
pieces of the NGA-W1 models for applicability in Turkey. A comparison dataset
is compiled by including strong motions from earthquakes occurred in Turkey and
earthquake metadata of ground motions consistent with the NGA-W1 database.
Random-effects regression is employed and plots of the residuals are used to
evaluate the differences in the magnitude, distance, and site amplification scaling.
Incompatibilities between the NGA-W1 GMPEs and Turkish dataset in small-tomoderate magnitude, large distance and site effects scaling are encountered. The
NGA-W1 GMPEs are modified for the misfit between the actual ground motions
and the model predictions using adjustments functions. Turkey-Adjusted NGAW1 models are compatible with the regional strong ground motion characteristics
and preserve the well-constrained features of the global models.
INTRODUCTION
The Next Generation Attenuation (NGA) project (Power et al., 2008) developed ground
motion prediction equations (GMPEs) for shallow crustal earthquakes in active tectonic
regions. Even though the target area was Western US (WUS), the GMPEs were intended to
be applicable in other shallow crustal and active tectonic regions around the world. Slowly,
NGA GMPEs have been used in probabilistic seismic hazard assessment (PSHA) studies in
various regions, raising the issue of their applicability outside WUS. Recent studies by
Stafford et al. (2008), Scasserra et al. (2009), Shoja-Taheri et al. (2010), and Bradley (2013)
tested the applicability of the NGA-W1 GMPEs for Euro-Mediterranean Region, Italy, Iran,
and New Zealand. Scasserra et al. (2009) have taken a further attempt and modified the
distance scaling of the NGA-W1 GMPEs by re-regressing the model coefficients related to
a)
b)

Middle East Technical University, Civil Engineering Department, Ankara, TURKEY, 06531
Pacific Gas & Electric Company, Geosciences Department, 245 Market St., San Francisco, CA 94105

the distance attenuation and the constant term for PGA, 0.2, 0.5, 1 and 2 second spectral
periods. Bradley (2013) found that the Chiou and Youngs (2008) model performed better
than the other NGA-W1 models in capturing the difference between small magnitude scaling
of the NGA-W1 and New Zealand datasets, especially after the small magnitude model
proposed by Chiou et al. (2010) was included. However, several other features of the Chiou
and Youngs (2008) GMPE had to be adjusted for New Zealand and coefficients of the
modified model were provided by Bradley (2013).
The primary objective of this study is to evaluate the compatibility of the NGA-W1 GMPEs
in terms of magnitude, distance and site effects scaling with respect to the Turkish strong
ground motions. The NGA-W1 database (Chiou et al., 2008) includes only a small sample of
the available strong ground motions from the earthquakes that occurred in Turkey. Table 1
presents the subset of the NGA-W1 database that was used in each model. Only seven
earthquakes and, at most, 52 recordings were included in the datasets of the models due to
two fundamental reasons: i) the limitations of the NGA-W1 project (especially the magnitude
constraint), and ii) missing earthquake metadata and recording site information for strong
ground motions recorded in Turkey during the database compilation period (2000-2003).
Table 1. Number of earthquakes that occurred in Turkey and the number of recordings from these
events that was included in the NGA-W1 developers datasets.
Earthquake
Name
Izmir
Dursunbey
Erzincan
Dinar
Kocaeli
Dzce
aldran
Total

Number of recordings included by

Event ID in
NGA-W1
Database

Year

Mw

44
47
121
134
136
138
141

1977
1979
1992
1995
1999
1999
1976

5.30
5.34
6.69
6.40
7.51
7.14
7.21

AS
2008
0
1
1
2
17
13
1
35

BA
2008
0
0
0
4
26
22
0
52

CB
2008
0
0
1
2
22
14
1
40

CY
2008
1
1
1
2
17
12
1
36

ID
2008
0
1
0
0
6
7
0
14

Turkey is one of the most seismically active regions in the world and many more strong
ground motion recordings are available. However, the earthquakes are recorded by different
agencies (mainly by AFAD: General Directorate of Disaster Affairs and KOERI: Kandilli
Observatory and Earthquake Research Institute) and a unified database, including the
complete earthquake source information, could not be established until recently. A

GLERCE-3

compilation of the Turkish earthquakes and strong motion data from the last 50 years was
completed as part of the Turkish Strong Motion Database (TSMD) Project (Akkar et al.,
2010). This expanded dataset of Turkish ground motions is used in this study to evaluate, and
modify as needed, the NGA-W1 models for applicability to Turkey. The TSMD is used as the
preliminary dataset, however, many parameters required for the NGA-W1 models were
missing and had to be estimated. Additionally, we screened the ground motion waveforms for
data quality and calculated the orientation-independent intensity measures. Details of the
database compilation efforts are elaborated in the following section.
Incompatibilities between the NGA-W1 GMPEs and Turkish strong motion dataset in smallto-moderate magnitude scaling, large distance scaling and site amplification scaling were
encountered during the evaluation of the residuals. These features are systematically
modified for all NGA-W1 GMPEs by introducing carefully designed adjustment functions to
keep the better-constrained features of the NGA-W1 GMPEs while reflecting the regional
ground motion characteristics. We propose the new set of Turkey-specific versions of the
NGA-W1 GMPEs, the TR-Adjusted NGA-W1 models that are applicable to M4-M8 events
and distances up to 200 km. Median predictions and standard deviations of the TR-Adjusted
models are compared with the regional and global GMPEs in the subsequent sections.
COMPILATION OF THE TURKISH COMPARISON DATABASE
The TSMD includes 4067 sets of recordings from 2996 events and almost 80% of the
recordings are from small magnitude events (Mw<5). We have not applied a criterion for
selecting well-recorded earthquakes for Mw 5 events since the since the amount of data
from moderate-to-large magnitude events is limited. All of the 173 earthquakes in the
database with magnitude 5 or larger were selected. Even if the NGA-W1 GMPEs are
applicable for moderate-to-large magnitude earthquakes (M5-M8), they are being
extrapolated for smaller magnitudes in PSHA applications, which commonly ignore any
stated magnitude limits in the GMPEs. The magnitude scaling of the models should be
compatible with the small magnitude scaling of the target regions dataset; therefore, smaller
magnitude events with 3 or more recordings were also included. No magnitude information
was available for 37 of the selected earthquakes; therefore, these events (with 151 recordings)
were discarded from the dataset. The moment magnitudes for 119 earthquakes were not
available; for these small magnitude earthquakes (ML<5), moment magnitudes were
estimated from local magnitude (ML) values using the regional magnitude conversion
GLERCE-4

relationships proposed by Akkar et al. (2010). One of the earthquakes in the dataset was
missing focal depth information; therefore, this event was removed. Also, style of faulting
classification for 47 events, which have 106 recordings, was missing. For 30 of these
earthquakes, the style-of-faulting was estimated by associating the event with other events in
the sequence or dominant fault mechanism in the region. These 30 events, with 68
recordings, were preserved in the dataset but the remaining 17 events with 38 records were
removed.
Unfortunately, no site information (VS30 or any site classification) could be found for 431 of
the recordings obtained from 50 different strong motion recording stations. Coordinates of
these stations were compared to the station coordinates in NGA-W1 (Chiou et al., 2008) and
NGA-W2 (Ancheta et al., 2014) datasets to search for the closest match. VS30 values of 9
stations (49 recordings) were estimated from NGA database, assuming that small errors
(0.001) in latitude and longitude of recording stations might occur due to measurement
errors and some stations might be duplicated or moved. Remaining 382 records with no VS30
value or no site class information were removed from the database.
Considering the applicable distance range of the NGA-W1 predictive models, 118 records
with rupture distance (Rrup) or Joyner-Boore distance (Rjb) larger than 200 km were discarded
from the dataset. Source-to-site distance metrics for 96 remaining records were missing.
Fortunately, these ground motions were recorded during small magnitude earthquakes;
therefore, the Rrup and Rjb were estimated from the hypocentral distance and epicentral
distance, respectively. At this stage, our Turkish comparison dataset included 288
earthquakes and 1179 strong motion records. The mainshock and aftershock classification of
the earthquakes in the dataset were performed by using rupture plane geometries in NGA-W1
database for Erzincan (1992), Dinar (1995), Kocaeli (1999), and Dzce (1999) earthquakes
based on the method proposed by Wooddell and Abrahamson (2014), and by using Gardner
and Knopoff (1974) methodology for other small-to-moderate magnitude earthquakes. The
entire TSMD (before the elimination of recordings) was analyzed for event classification.
Two types of events, Class 1 (mainshocks and aftershocks located away from the mainshock
rupture plane) and Class 2 (aftershocks located close to the mainshock rupture plane) were
declustered using a distance limit to the mainshock rupture (CRJB) of 20 km. Of the 288
events in the comparison data set, 156 events (%54) were classified as Class 1 events and 132
events (%46) were classified as Class 2.

GLERCE-5

Some of the parameters required for the NGA-W1 predictive models were missing in the
TSMD such as; depth to the engineering rock (denoted by Z1.0 and Z2.5) and depth to the top
of the rupture (Ztop). These values were estimated using Equations 1 to 3:
ln(Z1.0 ) 28.5 0.4775* ln(VS830 378.78 )

(1)

Z2.5 0.519 3.595* Z1.0

W
Z top Z hyp * sin( )
2

(2)
(3)

where Zhyp is the hypocentral depth, W is the fault width, is fault dip angle, Z1.0 and Z2.5 are
basin effect terms which represents depths to the 1.0 and 2.5 km/s shear wave velocities in
the soil profile, respectively. The Ztop was restricted to be positive. The depth of the
earthquakes was not used as a selection criterion during the establishment of the comparison
dataset. Approximately 10% of the recordings in the comparison dataset (156 out of 1142)
were from events with hypocentral depths greater than 20 km, however the Ztop values for 53
of these recordings are estimated (using Equation 3) as less than 20 km.
The data censoring is an important issue for the older data in the NGA-W2 dataset with small
dynamic range, but is less of an issue for the digital data with larger dynamic range.
Truncation for small magnitude data (less than 0.005g to 0.02g) was observed in PGAdistance plots of NGA-W2 dataset, especially for large magnitudes. To evaluate the same
issue, the PGA values for M3.5-M4.5 and M4.5-M5.5 events are plotted with respect to
distance in Figure S1 of the electronic supplement. For other magnitude ranges, the data is
too sparse. As Figure S1 indicates, the data censoring problem for small ground motions is
not observed in the Turkish comparison dataset for the M3.5 to M4.5 range with PGA values
as small as 0.0004g; however, there is an indication of truncation at 0.001g for distances
greater than 100 km for the M4.5 to M5.5 range. The distribution of the data below 0.00020.0001g is significantly affected from missing observations since a distance limitation of 200
km is applied in the dataset. Because the truncation is not seen at smaller magnitudes (M3.5 M4.5), we have not applied a distance truncation to the selected data.
Majority of the recordings in the selected dataset were processed by Akkar et al. (2010). We
aimed to preserve as much data as possible to obtain a representative dataset. Therefore, 284
unfiltered recordings were not excluded from the dataset. The waveform data of all ground
motion recordings were checked for data quality and 37 unfiltered recordings were
eliminated from the dataset due to spike, insufficient digitizer resolution, multi-event or Swave trigger problems. During the data quality check for the waveforms, we observed that
GLERCE-6

the initial excitation time of the three orthogonal components is not consistent for a large
number of processed records. This time lag results from the separate a-causal low-cut
filtering applied to the components of the record by adding zero pads in different lengths
(Glerce et al., 2012). We performed a systematic screening procedure on the waveforms and
shifted the short horizontal component by adding zero pads to align with the longer
horizontal component in each recording with a time lag.
Final dataset used in the comparison includes 1142 recordings from 288 events with the
earthquake metadata, source-to-site distance metrics for the recordings (RRup and RJB), VS30
values for the recording stations, and horizontal component spectral values in terms of
GMRotI50 (Boore et al., 2006) for 23 spectral periods between 0.01 and 10 seconds.
Calculated GMRotI50 values are consistent with the values in NGA-W1 database as shown
in Figure 1(a) and Figure 1(b) for two well-recorded, large magnitude earthquakes (1999
Kocaeli and Dzce Earthquakes). These figures indicate that the distance measures of the
mutual recordings are generally consistent for the Kocaeli Earthquake in both datasets;
however, some discrepancies may occur in the spectral values due to different filtering
procedures employed. The estimates of Rrup from the NGA-W1 and the TSMD datasets are
different due to different assumptions of the extent of the rupture for the Dzce earthquake
(Figure 1b). Figure 1(c) shows the mean and standard deviation of the average difference (in
natural log units) of common recordings in the TSMD and NGA-W1 dataset as a function of
period. At short periods (T<1 sec), the average difference in the spectral accelerations for the
two datasets is small (less than 5%). At periods between 1 and 5 seconds, the NGA-W1
spectral accelerations are, on average, up to 10% larger than the TSMD spectral
accelerations. The difference increases to 30% at a period of 10 seconds, but within a large
uncertainty range.
To be consistent with the data selection methods used by the NGA-W1 developers, for
Abrahamson and Silva (2008, AS08) and Chiou and Youngs (2008, CY08) models, the
aftershocks (Class 2 events) in the comparison dataset were included, whereas, these events
were excluded for Boore and Atkinson (2008), Campbell and Bozorgnia (2008), and Idriss
(2008) (BA08, CB08 and ID08) models. Because the ID08 model only includes the rock sites
(450 m/s < Vs30 < 900 m/s), only the data within this range was included in the analysis. The
response spectral values for the recordings in the dataset are used in the regression analysis if
the spectral frequency is greater than 1.25 times the high-pass corner frequency, consistent

GLERCE-7

with the constraints used for the NGA models. As a result, the size of the dataset used in the
analysis decreases as the spectral period increases. The resulting period dependence of the
number of recordings used in the regression analysis is shown in Figure 2. The significant
drop in the number of recordings at 2.6 seconds indicates that the results in the long-period
range are not well constrained by the empirical data.

Average difference in the common

recordings of TSMD and NGA-W1

(a)

(b)

1
0.8
0.6

0.4
0.2

0
-0.2
-0.4
0.01

0.1

10

Period (s)
(c)
Figure 1. GMRotI50 values at PGA in the NGA-W1 database and in the Turkish comparison dataset
(a) of the ground motions recorded during 1999 Kocaeli Earthquake, and (b) of the ground motions
recorded during 1999 Dzce Earthquake (only common recordings are plotted). (c) Mean and
standard deviation of the average difference of common recordings.

GLERCE-8

1400

Number of Records

1200
1000
800
600

400
200
0
0.1

10

Period (s)

Figure 2. Period dependence of number of ground motions in the Turkish comparison dataset based
on the lowest useable frequency of the average horizontal component.

The magnitude-distance distributions for peak ground acceleration (PGA) and spectral
acceleration at T=3 seconds are shown in Figure 3(a) and 3(b), respectively. Figure 3 shows
that the recordings obtained from events with magnitudes bigger than 6.0 and recordings
from the moderate-to-large magnitude events within 30 kilometer from the rupture are rather
sparse. This feature of the dataset is not the outcome of the excluded data points as discussed
above, same phenomenon was also observed by Akkar et al. (2010) for the TSMD.

Moment Magnitude

Moment Magnitude

8
7
6
5
4
3

7
6
5
4
3

0.1

10

100

Distance (km)

0.1

10

100

Distance (km)

(a)

(b)

Figure 3. Distribution of magnitude-distance pairs for (a) PGA and (b) T=3 sec

MODEL RESIDUALS AND ADJUSTMENTS FOR THE TURKISH DATASET

The preferred methodology for evaluating the misfit between the model predictions and
actual ground motion data is the analysis of model residuals. The actual spectral accelerations
GLERCE-9

and NGA-W1 model predictions for each recording (i) from each event (j) at each period (k)
are denoted by aijk and pijk, respectively. Total model residuals are calculated and separated
into three components: the mean offset representing the average bias of the actual data
relative to the model predictions (ck), the event terms (or inter-event residuals, Bjk), and the
intra-event residuals (Wijk) using random effects regression as given in Equation 4:

Rijk ln(aijk ) ln(pijk ) ck Bjk Wijk

(4)

After calculating the residuals, the procedure described in Figure 4 is applied. As the first
step, the total inter-event residuals (ck+ Bjk) are plotted against moment magnitude to test the
ability of the NGA-W1 GMPEs to capture the magnitude scaling of the Turkish data (Step 1
in Figure 4). If a trend is observed, an adjustment function is fitted to the total event terms to
modify the original GMPE at each period. To define the adjustment function, the functional
forms and cut-off or hinge coefficients of the NGA-W1 models are taken into consideration
(Step 2 in Figure 4).
The mean offset (ck), event terms (Bjk) and the intraevent residuals (Wijk) are calculated
repeated
using the
modified
form

The event terms (Bjk) are

plotted
against
moment
magnitude.

1
4

2
If a trend is observed, an
adjustment function is fitted
to the event terms to modify
the original GMPE at each
period.

The intra-event residuals

(Wijk) are plotted against the
distance and Vs30
5
If a trend is observed, an
adjustment function is fitted
to the intra-event residuals to
modify the original GMPE at
each period.

3
repeated
using the
modified
form

The remaining mean-offset

after the modifications is
neglected if small, or added
to the constant term of the
modified GMPE.

Figure 4. Flowchart describing the methodology of applied residual analysis

Residual analysis is repeated using the modified forms of the original model (Step 3 in Figure
4). Next, the intra-event residuals (Wijk) are plotted against Rrup and VS30 to test the ability of
the NGA-W1 models to capture the distance and site effects scaling of the actual data (Step 4

GLERCE-10

in Figure 4). If a trend is observed, adjustment functions are fitted to the intra-event residuals
to modify the original GMPE (Step 5 in Figure 4). Residual analysis is repeated using the
modified forms of the original model (Step 6 in Figure 4). The remaining mean offsets after
the modifications may be neglected if small, or may be added to the constant term of the
modified GMPE (Step 7 in Figure 4). Finally, standard deviations of the intra-event residuals
are compared to the intra-event standard deviations of the original NGA-W1 models. We note
that when using random effects to compute the residuals, the estimation of the event terms
may be sensitive to large changes in the distance attenuation. If there are large changes in the
distance scaling terms, then the event terms should be re-estimated using the revised distance
scaling; however, for the Turkey data, the distance scaling is not very different from the
NGA-W1 models so the event terms estimated in the first step are stable.
Steps 1-3: Magnitude Scaling of NGA-W1 GMPEs:
As the first step towards checking the compatibility of NGA-W1 models with the Turkish
comparison dataset, total inter-event residuals are plotted with respect to moment magnitude
for PGA as shown in Figure 5 (total inter-event residuals are presented by grey dots in each
figure). Figure 5 shows that the ground motions in the dataset are overestimated by all NGAW1 GMPEs significantly, especially for small-to-moderate magnitude earthquakes. The trend
is persistent for all spectral periods as seen in the 0.2, 0.5 and, 1 second residual plots
(provided in the electronic supplement in Figures S2-S3) for all models. Residual plots for all
periods suggested that the magnitude scaling of the Turkish comparison dataset is drastically
different than the NGA-W1 models and this feature needs to be fixed to consider the models
applicable in Turkey. On the other hand, preserving the well-constrained pieces of the NGAW1 models is critical, since the large magnitude earthquakes are poorly represented in the
Turkish comparison dataset (see Figure 3) and any changes on large magnitude scaling
parameters will have a major impact on the hazard calculations.
Fortunately, the magnitude scaling functions of the all NGA-W1 models except the Chiou
and Youngs (2008) (CY08) model includes linear magnitude scaling terms with either one
(AS08, BA08 and ID08) or two breaks (CB08) between small, moderate and large magnitude
earthquakes. The basic form of AS08 is given as an example in Equation 5:
2

a1 a 4 ( M c1 ) a8 (8.5 M ) a 2 a 3 ( M c1 ) ln(R) for( M c1 )

f 1 ( M , Rrup )
2

a1 a 5 ( M c1 ) a8 (8.5 M ) a 2 a 3 ( M c1 ) ln(R) for( M c1 )

GLERCE-11

(5)

where M is the moment magnitude, R is distance measure that includes the fictitious depth
term, a4 and a5 are the linear magnitude scaling coefficients for moderate (5<Mw<6.75) and
large (Mw6.75) magnitude events, and c1 is the break in the magnitude scaling (equal to 6.75
at all periods in AS08 model). The same magnitude break is used in the ID08 model for all
periods and in the BA08 model up to 5 seconds spectral periods. The hinge magnitude is
increased to 8.5 in BA08 model for periods longer than 5 seconds. The CB08 model used two
hinge magnitude values and a tri-linear magnitude scaling term with breaks at Mw=5.5 and
Mw=6.5. Considering the hinge magnitude values for all models, the misfit between the
actual data and predictions is modeled by Equation 6:

f 1 _ TA ( M )

c1 _ TA ( M M h ) for( M M h )
0 for( M M h )

(6)

where 1_ is the adjustment function coefficient determined separately for each model by
least-square regression and Mh is the hinge magnitude. To be consistent with the functional
forms of the models, Mh value is selected as 6.75 for AS08, BA08, CY08 and ID08 models
and 6.50 for CB08 model. The model fit to the residuals by Equation 6 is presented in Figure
5 by the black lines. Note that the selected adjustment function affects the magnitude scaling
only for small and moderate events; therefore, the well-constrained large magnitude scaling
of the NGA-W1 models are not modified. To develop a smooth model as a function of the
spectral period, c1_TA values are smoothed across period. Depending on the selected
functional form of the magnitude scaling by the model developers, the c1_TA coefficient can
be used to modify the original model coefficients or f1_TA term can be added to the GMPE
individually. In AS08 model, the linear magnitude coefficients (a4) are modified by adding
the c1_TA values. Similarly, the 3 coefficients of ID08 model are modified based on the
smoothed the c1_TA values. These modified coefficients (a4* and 3* and) are listed in Table
S1 of the electronic supplement. Since the hinge magnitude of the BA08 model depends on
the period, a direct modification of the linear magnitude scaling term for small-moderate
magnitude events (e5) is not possible. Tri-linear magnitude scaling of the CB08 model and
the functional form of the CY08 are also not suitable for direct modification of the magnitude
scaling; therefore, for BA08, CB08 and CY08 models, f1_TA term needs to be added to the
magnitude scaling. The smoothed c1_TA values for these models are also provided in Table S1
of the electronic supplement.

GLERCE-12

Total Inter-Event Residuals

Total Inter-Event Residuals

4
3

2
1
0
-1

-2
-3

4
3
2
1

0
-1
-2
-3
-4

-4
4

6
7
Magnitude (Mw)

6
7
Magnitude (Mw )

(b)
4

Total Inter-Event Residuals

Total Inter-Event Residuals

(a)

3
2
1
0
-1
-2
-3

-4
4

6
7
Magnitude (Mw )

3
2
1
0
-1
-2
-3
-4

6
7
Magnitude (Mw )

(c)

(d)

Total Inter-Event Residuals

4
3
2

1
0
-1
-2

-3
-4

6
7
Magnitude (Mw )

(e)
Figure 5. Distribution of total inter-event residuals of PGA for the unmodified models by (a) AS08,
(b) BA08, (c) CB08, (d) CY08, and (e) ID08 (grey dots) with magnitude. The adjustment function
fitted to the total inter-event residuals is shown by black lines in each plot. The plots are truncated at
M=4 to improve the visibility.

GLERCE-13

Slope:-0.19450.1047
T-ratio: 1.86

Slope:-0.11990.083
T-ratio: 1.44

(a)

(b)
Slope:-0.31040.0916
T-ratio: 3.39

Slope:-0.33170.1073
T-ratio: 3.09

(c)

(d)
Slope: 0.16670.1260
T-ratio: 1.32

(e)
Figure 6. Distribution of total inter-event residuals of PGA after the modification of magnitude
scaling of the models by (a) AS08, (b) BA08, (c) CB08, (d) CY08, and (e) ID08 (grey dots) with
magnitude. Standard error of the slope and the T-ratio is calculated for 4<M<6 range where the data is
not sparse.

GLERCE-14

(a)
(b)
Figure 7. The median spectral displacement curves for vertical strike-slip magnitude 5 earthquakes at
30 km distance for rock sites (Vs30=760 m/s) for each model (a) before and (b) after the magnitude
scaling adjustment.

After the magnitude adjustment, the model residuals are re-calculated using the modified
forms of the NGA-W1 models. The total inter-event residuals after the magnitude adjustment
are plotted with respect to moment magnitude in Figure 6 for each model. Figure 6 shows
that the total inter-event residuals of the adjusted models are evenly distributed along the zero
line up to the hinge magnitude value; therefore, the magnitude scaling of the adjusted models
are compatible with the magnitude scaling of the Turkish strong motion dataset. For larger
magnitudes, especially for magnitudes 7 and above, the residuals are negative but within the
expected scatter. The event terms for these two data points (1999 Kocaeli and 1999 Dzce
Earthquakes) were also negative for the original NGA-W1 GMPEs indicating that the event
terms of these two earthquakes are below the global average for large magnitude events.
The slope of the residuals, the standard error of the slope, and the T-ratio are calculated after
the adjustment and displayed in Figure 6 for each model. Statistical analysis of the residuals
indicated that the slope of the remaining residuals after the magnitude adjustment is slightly
negative as expected, except for the ID08 model. For AS08, BA08 and ID08 models, the Tscore is insignificant (smaller than 2), but for CB08 and CY08 models, a higher t-score is
estimated (between 3-3.4). The selected adjustment function (Eq. 6) is constrained to zero at
large magnitudes; therefore, the adjustments did not eliminate the trends in the residuals
completely. An adjustment function without a hinge magnitude would have worked better
statistically. Given that there are only two large magnitude events in the Turkish dataset and
the ground motions for large magnitude earthquakes are not well constrained, we preferred an

GLERCE-15

adjustment function with the hinge magnitude and used the global models to define the
contents for large magnitude range.
The median spectral displacement plots for vertical strike slip and magnitude 5 earthquakes at
30 km distance for rock sites (VS30=760 m/s) for each model are shown in Figure 7 before
and after the magnitude scaling adjustment. Figure 7(a) shows that the AS08 and CB08
models have constant spectral displacement values at long periods before the adjustment. The
additional constraint used in AS08 model to preserve this shape (given in Equation 21 of
Abrahamson and Silva, 2008) had to be removed during the magnitude adjustment; but the
same behavior was captured by the f1_TA (M) function. The additional magnitude scaling
introduced by the f1_TA (M) function in each model is constrained so that the spectral
displacement values are approximately constant at long periods (Figure 7(b)).
Steps 4-6: Distance Scaling of NGA-W1 GMPEs and Large Distance Effects:
After the magnitude adjustment, the intra-event residuals for the modified form of the models
are plotted with respect to rupture distance for PGA in Figure 8 and for 0.2, 0.5, and 1 second
spectral periods in Figures S4-S5 of the electronic supplement, suggesting no trend within the
applicability range of the NGA-W1 models for tectonic regions other than Western US (100
kilometers, Power et al., 2008). Figure 8 indicates that the NGA-W1 models slightly
underestimate the ground motions in the Turkish comparison dataset for rupture distances
within the range of 100-200 km. We note that the truncation of below average ground
motions for M4.5 to M5.5 events at distances greater than 100 km could be part of the reason
for the slight positive trend in the residuals versus distance from 100-200 km.
The distance scaling of BA08, CB08 and ID08 models do not include a separate large
distance scaling term even though the regional differences in the attenuation were
acknowledged by the authors. Small underestimation of the data for large distances (100-200
km) was also observed in some of the residual plots for the original models provided by the
developers (e.g. Campbell and Bozorgnia, 2008). Because these models do not treat the large
distance scaling effects separately, we did not impose adjustment functions for the Turkish
dataset on BA08, CB08 and ID08 models for large distance scaling. On the other hand, AS08
and CY08 models used additional functions for the large distance scaling of moderate
magnitude events (the gamma term) based on the broadband data for small magnitude
earthquakes from California (Boatwright et al., 2003). Gamma term of AS08 model is in
form of Equation 7:
GLERCE-16

a18 * ( RRUP 100) * T6 ( M ) for( RRUP 100)

0 for( RRUP 100)

(7)

where a18 is the regressed coefficient and T6(M) is the magnitude taper defined in AS08 as:
1 for ( M 5.5)

T6 ( M ) 0.5 * (6.5 M ) 0.5 for (5.5 M 6.5)

0.5 for ( M 6.5)

(8)

Based on the functional form used by AS08 model, the misfit between the actual data and
model predictions is modeled by Equation 9:
2_ = {

0 ( < 100)
2_ ( 100) ( 100)

(9)

where Rrup is the rupture distance and 2_ is the adjustment model coefficient determined
separately for each GMPE by least-square regression. The magnitude taper of AS08 model
(Equation 8) was included in the adjustment function for the AS08 model, whereas only
Equation 9 is used for CY08 model. The 2_ coefficients were used to modify the original
large distance scaling coefficients (a18) in AS08 model and the modified coefficients (denoted
by a18*) are listed in Table S1. In CY08 model, the anelastic attenuation was modeled as
magnitude dependent for PGA using the California broadband data. For other spectral
periods, the gamma term for the PGA was scaled based on the relative difference in the
broadband data for other periods with respect to the PGA value. Functional form of the CY08
model does not allow us to modify the model coefficients directly; therefore Equation 9 was
added to the original CY08 model form individually. To arrive at a smooth model across the
period, the 2_ values estimated for CY08 model are smoothed and listed in Table S1.
After the distance adjustment, the intra-event residuals are re-calculated using the modified
forms of AS08 and CY08 models. Intra-event residuals of the modified AS08 and CY08
models for PGA are plotted with respect to rupture distance in Figure 9(a) and (b) showing
that the modified large distance scaling of both models are more compatible with the Turkish
comparison dataset after the adjustment.

GLERCE-17

(a)

(b)

(c)

(d)

(e)
Figure 8. Distribution of intra-event residuals of PGA for the unmodified models by (a) AS08, (b)
BA08, (c) CB08, (d) CY08, and (e) ID08 (grey open circles) with rupture distance. The adjustment
function fitted to the intra-event residuals is shown by black lines in (a) and (d).

GLERCE-18

(a)
(b)
Figure 9. Distribution of intra-event residuals of PGA after the modification of distance scaling of the
models by (a) AS08 and (b) CY08 with rupture distance.

Steps 4-6: Site Effects Scaling of NGA-W1 GMPEs:

Next step is to evaluate the site effects scaling of the NGA-W1 models by plotting the intraevent residuals of the models modified for magnitude and distance. Plots of intra-event
residuals for PGA with respect to the shear-wave velocity for the top 30 meters (VS30) for
each model are shown in Figure 10. Figure 10 shows that the AS08, BA08, CB08, and CY08
models under-predict the ground motions in the Turkey comparison dataset at stiff
soil/engineering rock sites. The slope of the residuals, the standard error of the slope, and the
T-ratio are calculated and displayed in Figure 10 for each model. Statistical analysis of the
residuals indicated that the slopes of the intra-event residuals are negative for all models. For
BA08, CB08 and ID08 models, the T-score is insignificant (smaller than 2), but for AS08 and
CY08 models, significantly higher T-scores are estimated. The trends in the residuals for
BA08 and CB08 models are negligible, but for AS08 and CY08 models which included the
aftershocks in their datasets, the underestimation is quite noticeable. This observation is
expected because approximately 40% of the data for AS08 and CY08 models datasets consist
of data from Taiwan (1999 Chi-Chi Earthquake and its aftershocks) and correlation between
the VS30 and the deeper shear-wave velocity profiles, which controls the amplification of the
recording stations in Taiwan and in Western US, could be different. The trends in AS08 and
CY08 models are also visible for other spectral periods as seen in the residual plots for 0.2,
0.5 and, 1 seconds (provided in the electronic supplement Figure S6-S7). A linear adjustment
function for AS08 and CY08 models for the site effects is chosen as given below:
3_

30
= 3_ (
)
450

(10)
GLERCE-19

where VS30* is the shear wave velocity definition used by AS08 model and 3_ is the
regression coefficient.
Linear site amplification coefficients of AS08 and CY08 models (a10 and 1) cannot be
modified directly, since the hinge VS30 value used in Equation 10 (450 m/s) is not equal to the
period dependent VS30,LIN values of AS08 model. Also, the linear site term of CY08 model is
normalized by a higher hinge VS30 (1130 m/s). Instead, Equation 10 should be added to the
models separately. The smoothed c3_TA coefficients for AS08 and CY08 models are provided
in Table S1 (electronic supplement). Figure 11 presents the intra event residuals of the
modified models for PGA with respect to VS30 (Figure 11(a) for AS08 model and Figure
11(b) for CY08 model). Distribution of the residuals for the adjusted models shows no trends
up to VS30 = 600 m/s, however, the residuals are still negative for VS30 between 600 and 800
m/s. We chose to maintain a smooth scaling with VS30, which does not accommodate the bias
in this higher velocity range.
Step 7: Remaining Mean Offsets and the Constant Terms of NGA-W1 GMPEs:
The mean offset values (ck) for NGA-W1 GMPEs before and after the adjustments are shown
in Figure 12(a) and 12(b), respectively. In both figures, the mean offsets follow the same
trend up to 2 seconds: they are approximately equal to one negative ln unit before the
adjustment and very close to zero after the modifications. For periods longer than 2 seconds,
the mean offsets are more scattered and lay on the positive side due to the reduced number of
recordings used for comparison at those periods. We believe that the remaining mean offset
values after the adjustments are insignificantly small; therefore, the constant terms of the
NGA-W1 models are not modified.
Final check for standard deviations:
After the modifications, the inter-event and intra-event residuals are re-calculated for the TRadjusted models and the standard deviations for both set of residuals are evaluated. Figures
13 (a) and (b) compare the magnitude dependence of the within-event standard deviations of
the original and TR-adjusted models for PGA and T=1 sec, respectively. Solid lines in both
figures represent the averaged values for the AS08, CB08 and CY08 models that provided
separate and terms. Both figures indicate that the intra-event standard deviations of the
adjusted models are comparable with the original NGA-W1 models for magnitudes up to 6.5.

GLERCE-20

Slope: -0.0001350.0001T-ratio: 0.85

Slope: -0.0007510.0001 T-ratio: 5.95

(a)

(b)
Slope: -0.0007320.0001 T-ratio: 5.70

Slope: -0.0004640.0001 T-ratio: 2.66

(c)

(d)
Slope: -0.00050.0004 T-ratio: 1.40

(e)
Figure 10. Distribution of intra-event residuals of PGA for the unmodified models by (a) AS08, (b)
BA08, (c) CB08, (d) CY08, and (e) ID08 (grey dots) with VS30. The adjustment function fitted to the
intra-event residuals is shown by black lines in (a) and (d).

GLERCE-21

(a)
(b)
Figure 11. Distribution of intra-event residuals of PGA after the modification of linear site
amplification scaling of the models by (a) AS08 and (b) CY08 with VS30.

For larger magnitudes, for which Turkish dataset is very sparse of data, the values of the
adjusted models deviate from the original model values. Since the data is very limited at large
magnitudes, the values are not statistically reliable; therefore, the intra-event standard
deviation coefficients of the original models should be used without any modification.
Figures 13(c) and (d) present the standard deviation of the inter-event residuals for PGA and
T=1 sec. Again, the solid lines in both figures represent the averaged values for the AS08,
CB08 and CY08 models that provided separate and terms. A much larger difference is
observed in the inter-event standard deviations of the adjusted models with respect to the
values. Difference in the inter-event standard deviations have been observed in the previous
studies for other regions such as New Zealand (Brendon, 2013) but the cause for this
difference has not been identified. Because the NGA-W1 database is statistically more stable
than the Turkish dataset, we recommend the use of between-event standard deviation
coefficients of the original models until further evaluation of the underlining reason for the
difference is completed.

GLERCE-22

BA08
ID08

CB08

1.5

0.5

0.5

Mean Offset

Mean Offset

1.5

AS08
CY08

0
-0.5
-1

BA08
ID08

CB08

0
-0.5
-1
-1.5

-1.5
-2
0.01

AS08
CY08

0.1
1
Period (s)

10

-2
0.01

(a)

0.1
1
Period (s)

10

(b)

Figure 12. Mean offset values for (a) original models, (b) Turkey-Adjusted models.

(a)

(b)

(c)

(d)

Figure 13. Magnitude dependence of intra-event standard deviations of the original and TRAdjusted models for a) PGA and b) T=1 sec. Magnitude dependence of inter-event standard
deviations of the original and TR-Adjusted models for c) PGA and d) T=1 sec.

GLERCE-23

TURKEY-ADJUSTED NGA-W1 MODELS

The average median response spectra of the NGA-W1 models (broken black lines) and
Turkey-Adjusted NGA-W1 models (solid black lines) are compared for different scenarios in
Figure 14. Median predictions of NGA-W2 models are also provided in Figure 14 (gray
lines). Figures 14(a) and (b) show the median predictions for magnitude 5 strike-slip
earthquakes at the rupture distance of 10 kilometers for average rock (VS30=760 m/s) and
average soil (VS30=270 m/s) conditions, respectively. For both of these scenarios, the TRAdjusted NGA-W1 models are significantly lower than the NGA-W1 and NGA-W2 models,
since the small-to-moderate magnitude scaling of TR-adjusted models are modified. Figures
14(c) and (d) present the median predictions for magnitude 7 strike-slip earthquakes at the
same distance for the same site conditions given in Figures 14(a) and (b). The median
predictions of the TR-Adjusted models are only slightly different from the NGA-W1 and
NGA-W2 models for large magnitude earthquakes due to the effect of linear site
amplification scaling adjustment. For large magnitude events at long distances (M7, D150
km rock and soil scenarios provided in Figures 14(e) and (f)), the TR-adjusted NGA-W1
model predictions are quite similar to NGA-W1 models in the short period range; however,
they are closer to the NGA-W2 model predictions, especially for the longer periods.
The median response spectra of the TR-adjusted NGA-W1 models and the Akkar and anan
(2010) (AC10) model are presented in Figures 15(a) through 15(f) for the same scenarios.
AC10 model is the regional model that was developed using a subset of the same database
(TSMD, Akkar et al, 2010); therefore, agreement of the median predictions for the range
well-constrained by the Turkish dataset (M4 to M6) is expected. For the small magnitude
scenarios provided in Figures 15(a) and (b), predictions of the TR-adjusted models and AC10
model are very similar indicating that the small-to-moderate magnitude scaling of the Turkish
ground motions are captured by the magnitude adjustment functions. A significant difference
is observed in Figures 15(c) and (d); predictions of the TR-adjusted NGA-W1 models are
very close to the predictions of global models for large magnitudes, but significantly higher
than the predictions AC10 model. The spectral shapes for large magnitudes at large distances
provided in Figure 15(e) and (f) are quite different. This difference is related to the
constraints in the magnitude scaling from small to large magnitudes; in the TR-Adjusted
models the large magnitude scaling is constrained by the global data, whereas in the AC10
model the large magnitude scaling is estimated from the sparse Turkish dataset.

GLERCE-24

0.1

0.1

Sa (g)

Sa (g)

0.01

TR_NGA

TR_NGA

0.001

0.01

0.001

NGA_W1

NGA_W2

NGA_W2

0.0001
0.01

NGA_W1

0.1

10

0.0001
0.01

0.1

10

Period (s)

Period (s)
(a)

(b)
1

0.1

0.1

Sa (g)

Sa (g)

0.01

0.01

TR_NGA

NGA_W1

NGA_W1

NGA_W2

NGA_W2

0.001
0.01

0.1

TR_NGA

10

0.001
0.01

0.1

10

Period (s)

Period (s)
(c)

(d)

1
TR_NGA

TR_NGA

NGA_W1

NGA_W1

NGA_W2

NGA_W2

0.1

Sa (g)

Sa (g)

0.1

0.01

0.001
0.01

0.01

0.1

10

0.001
0.01

0.1

10

Period (s)

Period (s)

(e)
(f)
Figure 14. Median predictions of NGA-W1, NGA-W2 and TR-Adjusted NGA_W1 models for
following scenarios; (a) M5, D10km, Vs30=760 m/s, (b) M5, D10km, Vs30=270 m/s, (c) M7,
D10km, Vs30=760 m/s, (d) M7, D10km, Vs30=270 m/s, (e) M7, D150km, Vs30=760 m/s, and (f)
M7, D150km, Vs30=270 m/s.

GLERCE-25

TR_NGA

TR_NGA

AC_2010

AC_2010

Sa (g)

Sa (g)

0.1

0.1

0.01

0.001
0.01

0.01

0.001

0.1

0.0001
0.01

0.1

Period (s)

Period (s)
(a)

(b)

Sa (g)

Sa (g)

0.1

0.1

0.01
TR_NGA

TR_NGA

AC_2010

AC_2010

0.001
0.01

0.1

0.01
0.01

0.1

Period (s)

Period (s)
(c)

(d)

1
TR_NGA

TR_NGA

AC_2010

AC_2010

0.1

Sa (g)

Sa (g)

0.1

0.01

0.001
0.01

0.01

0.1

0.001
0.01

0.1

Period (s)

Period (s)

(e)
(f)
Figure 15. Median predictions of TR-Adjusted NGA_W1 and Akkar and agnan (2010) (same
dataset) models for following scenarios; (a) M5, D10km, Vs30=760 m/s, (b) M5, D10km, Vs30=270
m/s, (c) M7, D10km, Vs30=760 m/s, (d) M7, D10km, Vs30=270 m/s, (e) M7, D150km, Vs30=760
m/s, and (f) M7, D150km, Vs30=270 m/s.

GLERCE-26

COMPARISON WITH THE UPDATED NGA-W2 MODELS

The proposed modifications on the NGA-W1 models based on the Turkish ground motion
dataset (small magnitude scaling, regionalization of the VS30 and large distance terms) were
also amended by the revision within the NGA-W2 project (Gregor et al., 2014). In order to
compare the small magnitude adjustments proposed for the Turkish dataset with the small
magnitude modifications published after the NGA-W1 project (Chiou et al. (2010) and NGAW2 models), the magnitude scaling of NGA-W1, NGA-W2, and TR-NGA-W1 models for
vertical strike-slip events at Rrup=30 km and for VS30=760 m/s are presented in Figure 16,
individually for each model. Figure 16(a) and (b) compares the AS08 and BA08 models with
the Turkey adjusted versions and the updated NGA-W2 models by the same developer
groups (Abrahamson, Silva and Kamai, 2014 (ASK14) model and Boore, Stewart, Seyhan
and Atkinson, 2014 (BSSA14) model). Both figures indicate that the magnitude scaling of all
models is indifferent above the hinge magnitude value (M>6.75). The NGA-W1 and W2
models are only slightly different than each other for magnitudes larger than the small
magnitude break point (M>5 for ASK14 and M>5.5 for BSSA14); however, TR-Adjusted
models deviate from the global models for small-to-moderate magnitude range (4<M<6.75)
significantly. For smaller magnitudes (M<4), predictions of NGA-W2 models and TRAdjusted NGA-W1 models are similar. CB08 and CY08 models are compared with the
updated CB14 (Campbell and Bozorgnia, 2014) and CY14 (Chiou and Youngs, 2014) models
in addition to the TR-adjusted versions in Figure 16(c) and (d), respectively. Magnitude
scaling of the NGA-W2 and TR-adjusted NGA-W1 models is very similar for these two
models with more flexible functional forms, except for the small difference in 5<M<6.75
range. The updated model by Idriss (2014, ID14) has approximately the same magnitude
scaling with the previous version (ID08); therefore, both models are significantly different
than the TR-adjusted version (Figure 16(e)). The average magnitude scaling of NGA-W2 and
TR-adjusted NGA-W1 models are noticeably different as shown in Figure 16(f); with TRadjusted models being smaller and smoother for magnitudes smaller than 6.75. The
interaction between the constraints of the magnitude scaling functions of the original models
and the applied adjustment functions has a significant effect on the difference between the
predictions of the NGA-W2 and TR-NGA-W1 models, especially within the M5-M6 range.
Magnitude corrections applied to the NGA-W1 models are in good agreement with the
magnitude scaling of NGA-W2 models, but they are not consistent in the M3-M4 range.

GLERCE-27

0.1

0.1
PGA (g)

PGA (g)

0.01
TR_AS08
AS08 (NGA_W1)
ASK14 (NGA_W2)

0.001
0.0001
3

5
6
Magnitude

0.01
TR_BA08
BA08 (NGA_W1)
BSSA14 (NGA_W2)

0.001
0.0001
3

0.1

0.1

0.01
TR_CB08
CB08 (NGA_W1)
CB14 (NGA_W2)

0.001
0.0001
3

5
6
Magnitude

TR_CY08
CY10_SM
CY08 (NGA_W1)
CY14 (NGA_W2)

0.001
0.0001
8

0.1

0.1
PGA (g)

PGA (g)

0.01
TR_ID08
ID08 (NGA_W1)
ID14 (NGA_W2)

5
6
Magnitude

(d)

0.0001

0.01

(c)

0.001

(b)

PGA (g)

PGA (g)

(a)

5
6
Magnitude

5
6
Magnitude

0.01
Average (TR_NGA_W1)
Average (NGA_W1)
Average (NGA_W2)

0.001
0.0001
8

5
6
Magnitude

(e)
(f)
Figure 16. Magnitude scaling of vertical strike-slip events for Rrup=30 km and VS30=760 m/s of; (a)
AS08, TR-AS08 and ASK14 models, (b) BA08, TR-BA08 and BSSA14 models, (c) CB08, TR-CB08
and CB14 models, (d) CY08, TR-CY08 and CY14 models, (e) ID08, TR-ID08 and ID14 models, and
(f) average of NGA-W1, TR-NGA-W1 and NGA-W2 models.

GLERCE-28

The dataset is sparse below M4 (see Figure 3) thus we recommended that the adjusted models
are used in M>4 range.
All NGA-W2 models except for the ID14 model acknowledged the regional differences in
large distance scaling; however, regional terms were developed for regions like Taiwan,
Japan, China and Italy with significant amount of data in the NGA-W2 database. The number
of Turkish recordings in the NGA-W2 database is sparse (only 55 recordings); therefore,
individual anelastic attenuation terms are not provided for the Turkey. BSSA14 model
combined the recordings from Turkey with the recordings from China (defined as high Q
regions) and provided a China-Turkey modification on their c3 term. In ASK14 model,
Turkish recordings were grouped with ground motions from Iran under the Middle East
subgroup; however, separate anelastic attenuation or site amplification coefficients are not
provided for Middle East region after the evaluations. Similarly, CB14 and CY14 models did
not find a difference between the large distance scaling of California and Turkish ground

0.1

0.1

TR_AS08
ASK14 (NGA_W2)
AS08 (NGA_W1)

0.01

0.001
1

PGA (g)

PGA (g)

motions.

TR_CY08
CY14 (NGA_W2)
CY08 (NGA_W1)

0.01

0.001

10
100
Rupture Distance (km)

(a)

10
100
Rupture Distance (km)
(b)

PGA (g)

1
0.1
TR_BA08
BSSA14 (NGA_W2)
BA08 (NGA_W1)

0.01

0.001
1

10
100
Rupture Distance (km)
(c)

Figure 17. Distance scaling of vertical strike-slip events (M=7 and VS30=760 m/s) of; (a) AS08, TRAS08, ASK14 models, (b) CY08, TR-CY08, CY14 models, c) BA08, TR-BA08, BSSA14 models

GLERCE-29

Results of this study support the findings of NGA-W2 project: the distance scaling of the
Turkish dataset is not significantly different than the distance scaling of NGA-W1 and W2
models. The anelastic attenuation terms of AS08 and CY08 were only slightly modified for
Turkey and the effect of this modification is insignificant as shown in Figure 17.
SUMMARY AND CONCLUSION
A large number of global and regional GMPEs were developed in the last 20 years that are
applicable to shallow crustal and active tectonic regions. Constructing the GMPE logic tree
for PSHA applications is a controversial issue since: (i) local GMPEs are developed from the
regional datasets so they are expected to reflect the regional tectonic characteristics better
than the others, (ii) the uncertainties introduced by local GMPEs are higher than those of the
global GMPEs because they are based on statistically less stable and limited datasets. The
objective of this study is to introduce another alternative to local and global GMPEs by
regionalizing the global GMPEs that are developed for the same tectonic region as the target
area. Our study is inspired by the work of Scasserra et al. (2009) on the Italian strong motion
database (ITACA); however, a much larger dataset including the ground motions recorded
from the earthquakes that occurred in Turkey in the last 50 years is used. NGA-W1 GMPEs
are selected as the starting point and the earthquake metadata of the ground motions
consistent with the NGA-W1 database is compiled. Using the random-effects regression with
a constant term, model residuals between the actual strong motion data and NGA-W1 model
predictions are calculated for a period range of 0.01-10 seconds.
Plots of the residuals are used to evaluate the differences in the magnitude, distance, and site
amplification scaling between the Turkish dataset and the NGA-W1 models. Inter-event
residuals indicated that the ground motions in the dataset are overestimated by all 5 NGA-W1
models. We have only modified the small-to-moderate magnitude scaling of the NGA-W1
GMPEs in order to preserve the well-constrained large magnitude scaling of the global
dataset. No trends in the residuals are observed in the intra-event residuals vs. rupture
distance plots up to 100 kilometers; therefore, the distance scaling of the NGA-W1 models is
not adjusted. The large distance scaling (between 100 and 200 km) of the AS08 and CY08
models with separate gamma terms are modified. Regional differences in those parameters
were also recognized by the NGA-W2 model developers. Median predictions for scenarios
important for engineering applications (large magnitude and short distance) after the
adjustments were almost identical to the median predictions of the original GMPEs. The
GLERCE-30

slight difference in the median predictions arises from the small adjustments applied to the
site effects scaling of the AS08 and CY08 models. Even if the median predictions for large
magnitudes at short distances remained unchanged, the constant bias in the model predictions
was removed after the adjustments. Therefore, we believe that the new set of TurkeyAdjusted NGA-W1 models proposed in this study is suitable for ground motion
characterization studies in Turkey.
Some important pieces of the NGA-W1 GMPEs such as the style of faulting and hanging
wall effects could not be evaluated since the majority (65-70%) of the events in the Turkish
database consists of strike slip earthquakes. Also, the non-linear site response effects
incorporated in the models are adopted without a change. These features may be tested in the
future once the dataset is enriched by dip-slip mechanism events and site-response simulation
results representative of the deeper shear-wave profiles of the recording stations in Turkey
are available. Further analyses are required to address the possible data censoring issues in
the Turkish dataset. A standard deviation model for the ground motions in the Turkish dataset
that acknowledges the repeatable path effects is under development. We recommend the use
of standard deviations of the original NGA-W1 models until the single-station sigma model
of Turkey is developed.
It is not the intention of this paper to suggest that the TR-Adjusted NGA-W1 models should
be used in place of the NGA-W2 models. We proposed the TR-Adjusted NGA-W1 models as
an alternative to NGA-W1 models and regional GMPEs developed for Turkey.
Implementation process of the NGA-W2 models in US is currently ongoing; however, we
foresee that the NGA-W1 models will be slowly replaced with the new versions in other
regions such as Turkey. During this time of transition, we believe that the TR-adjusted NGAW1 models will be a convenient and reasonable alternative for PSHA studies in Turkey.
ACKNOWLEDGEMENTS
We would like to thank Emre Akyz, Katie E. Wooddell and Robert B. Darragh and for their
help and guidance during the database compilation. Efforts of Nick Gregor on incorporating
the Turkey-Adjusted NGA-W1 models in HAZ43 software are gratefully acknowledged. We
appreciate the anonymous reviewers and members of the PEER-NGA working group for
valuable suggestions during the review process.

GLERCE-31

REFERENCES
Abrahamson, N. A. and Silva, W. J., 2008. Summary of the Abrahamson and Silva NGA groundmotion relations, Earthquake Spectra 24, 6797.
Abrahamson, N., Atkinson, G., Boore, D., Bozorgnia, Y., Campbell, K., Chiou, B., Idriss, I. M., Silva,
W., Youngs, R., 2008. Comparisons of the NGA ground-motion relations, Earthquake Spectra
24, 4566.
Abrahamson, N. A., Silva W. J., and Kamai, R., 2014. Summary of the ASK14 Ground-Motion
Relation for Active Crustal Regions, Earthquake Spectra 30, 1025 - 1057.
Akkar, S. and Cagnan, Z., 2010. A local ground-motion predictive model for Turkey and its
comparison with other regional and global ground-motion models. Bulletin of the Seismological
Society of America 100, 2978-2995.
Akkar, S., anan, Z., Yenier, E., Erdoan, ., Sandkkaya, A., Glkan, P., 2010. The recently
compiled Turkish strong motion database: preliminary investigation for seismological parameters.
Journal of Seismology 14, 457 479.
Ancheta, T.D., Darragh, R. B., Stewart, J. P., Seyhan, E., Silva, W., Chiou, B. S. J., Wooddell, K. E.,
Graves, R. W., Kottke, A. R., Boore, D. M., Kishida, T., Donahue, J. L., 2014. NGA-West 2
Database, Earthquake Spectra 30, 1007 - 1025.
Boatwright, J., Bundock, H., Luetgert, J., Seekins, L., Gee, L., and Lombard, P., 2003. The
dependence of PGA and PGV on distance and magnitude inferred from northern California
ShakeMap data, Bull. Seismol. Soc. Am. 93, 20432055.
Boore, D. M., Watson-Lamprey, J. and Abrahamson, N. A., 2006. GMRotD and GMRotI:
Orientation-independent measures of ground motion, Bull. Seismol. Soc. Am. 96, 12021511.
Boore, D. M. and Atkinson, G. M., 2008. Ground-motion prediction equations for the average
horizontal component of PGA, PGV, and 5%-damped PSA at spectral periods between 0.0 1s and
10.0 s, Earthquake Spectra 24, 99139.
Boore, D. M., Stewart, J. P., Seyhan, E., and Atkinson, G. A., 2014. NGA-West 2 Equations for
Predicting PGA, PGV, and 5%-Damped PSA for Shallow Crustal Earthquakes, Earthquake
Spectra 30, 1057 - 1087.
Bradley, B. A., 2013. A New Zealand-Specific pseudospectral acceleration ground-motion prediction
equation for active shallow crustal earthquakes based on foreign models, Bull. Seismol. Soc. Am.
103, 18011822.
Campbell, K. W. and Bozorgnia, Y., 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, Earthquake Spectra 24, 139173.
Campbell, K. W., and Bozorgnia, Y., 2014. NGA-West2 Ground Motion Model for the Average
Horizontal Components of PGA, PGV, and 5%-Damped Linear Acceleration Response Spectra,
Earthquake Spectra 30, 1087 - 1117.
Chiou, B., Darragh, R., Gregor, N. and Silva, W., 2008. NGA project strong motion database,
Earthquake Spectra 24, 2344
Chiou, B. S. J. and Youngs, R. R., 2008. Chiou-Youngs NGA ground motion relations for the
geometric mean horizontal component of peak and spectral ground motion parameters,
Earthquake Spectra 24, 173217.
Chiou, B. S. J., and Youngs, R. R., 2014. Update of the Chiou and Youngs NGA Model for the
Average Horizontal Component of Peak Ground Motion and Response Spectra, Earthquake
Spectra 30, 1117 - 1155.

GLERCE-32

Gardner, J. K. and Knopoff, L., 1974. Is the sequence of earthquakes in southern California, with
aftershocks removed, Poissonian? Bull. of the Seis. Soc. of America 64, 1363-1367.
Glerce, Z., Kargolu, B. and Akyz E., 2012. Comparison of the NGA-W1 vertical-to-horizontal
spectral acceleration ratio prediction equations with Turkish strong ground motion database,
Proceedings of 15th World Conference on Earthquake Engineering (15th WCEE), 24-28
September 2012, Lisbon.
Glerce, Z. and Akyz, E., 2013. NGA-W1 Vertical-to-Horizontal Spectral Acceleration Ratio
Prediction Equations Adjusted for Turkey, Seismological Research Letters 84, 678 - 688.
Gregor, N., Abrahamson, N. A., Atkinson, G. M., Boore, D. M., Bozorgnia, Y., Campbell, K.W.,
Chiou, B., Idriss, I. M., Kamai, R., Seyhan, E., Silva, W., Stewart, J. P., Youngs, R., 2014.
Comparison of the NGA-W2 GMPEs, Earthquake Spectra 30, 1179 - 1199.
Idriss, I. M., 2008. An NGA empirical model for estimating the horizontal spectral values generated
by shallow crustal earthquakes, Earthquake Spectra 24, 217243.
Idriss, I. M., 2014. An NGA-West2 Empirical Model for Estimating the Horizontal Spectral Values
Generated by Shallow Crustal Earthquakes, Earthquake Spectra 30, 1155 - 1179.
Power, M., Chiou, B., Abrahamson, N., Bozorgnia, Y., Shantz, T., and Roblee, C., 2008. An overview
of the NGA project, Earthquake Spectra 24, 321.
Scasserra, G., Stewart, J. P., Bazzurro, P. Lanzo, G. and Mollaioli, F., 2009. A comparison of NGA
ground-motion prediction equations to Italian data, Bull. of the Seis. Soc. of Am. 99, 2961-2978.
ShojaTaheri, J., Naserieh, S., and Hadi, G., 2010. A test of the applicability of NGA models to the
strong ground-motion data in the Iranian plateau, J. Earthquake. Engineering 14, 278292.
Stafford, P. J., Strasser, O. and Bommer, J. J., 2008. An evaluation of the applicability of the NGA
models to ground-motion prediction in the Euro-Mediterranean region. Bulletin of Earthquake
Engineering 6, 149-177.
Wooddell, K.E. and Abrahamson, N. A., 2014. Classification of Mainshocks and Aftershocks in the
NGA-West2 Database, Earthquake Spectra 30, 1257 - 1269.

GLERCE-33