Bradley EpistemicUncertaintiesNzPsha BSSA 2012 PDF

Bulletin of the Seismological Society of America, Vol. 102, No. 4, pp. 15541568, August 2012, doi: 10.
1785/0120110257
Consideration and Propagation of Epistemic Uncertainties

in New Zealand Probabilistic Seismic-Hazard Analysis
by Brendon A. Bradley, Mark W. Stirling, Graeme H. McVerry, and Matt Gerstenberger
Abstract
This article presents results from the consideration of epistemic uncertainties in New Zealand (NZ) probabilistic seismic-hazard analysis. Uncertainties in
ground-motion prediction are accounted for via multiple ground-motion prediction
equations within the logic-tree framework. Uncertainties in the fault-based seismicity
of the earthquake rupture forecast due to uncertainties in fault geometry, slip parameters, and magnitude-scaling relationships are considered in a Monte Carlo simulation
framework. Because of the present lack of fault-specific data quantifying uncertainties
for many faults in NZ, representative values based on judgement and available data for
NZ and foreign faults were utilized. Uncertainties in the modelling of background
seismicity were not considered. The implications of the considered epistemic uncertainties in terms of earthquake magnitudefrequency distributions and probabilistic
seismic-hazard analyses for two spectral acceleration ordinates, two soil classes,
and two locations (Wellington and Christchurch) are examined. The results illustrate
that, for the uncertainties considered, the variation in seismic hazard due to the
adopted ground-motion prediction model is larger than that due to the uncertainties
in the earthquake rupture forecast. Of the earthquake rupture forecast uncertainties
considered, the magnitude-geometry scaling relationships was the most significant,
followed by fault rupture length. Hence, the obtained results provide useful guidance
on which modelling issues are the most critical in the reliability of seismic-hazard
analyses for locations in NZ.
Introduction
The location of New Zealand (NZ) astride the boundary
of the Australian and Pacific plates makes it a country of high
seismicity. For the purposes of mitigation, seismic hazard in
New Zealand is routinely computed on the basis of a probabilistic seismic-hazard analysis (PSHA). The two basic ingredients of a PSHA are (i) an earthquake rupture forecast
(ERF), quantifying the location and likelihood of all possible
earthquake ruptures that may occur; and (ii) a ground-motion
prediction equation (GMPE), quantifying the ground motion
shaking at a specific location due to the occurrence of an
earthquake rupture.
Because of the complexity of earthquake rupture, wave
propagation, and local site effects, ERFs and GMPEs are
typically probabilistic in that they incorporate uncertainties
in the parameters they attempt to quantify. In PSHA, it is
common to distinguish between two main types of uncertainties. The first is uncertainties that, for the given models
adopted, are deemed to be purely random and unpredictable
and are referred to as aleatory variability. The second is that
which arises due to limited knowledge of the phenomena
being predicted and is referred to as epistemic uncertainty.
An example of the aleatory variability is the variability in
ground-motion amplitudes at a given distance from an

earthquake predicted using a GMPE, while an example of
epistemic uncertainty is the assessment of which GMPE is
most appropriate to be used for a particular problem under
consideration.
The benefit of making the distinction between aleatory
variability and epistemic uncertainties is that, in principle,
epistemic uncertainties can be reduced with improved
knowledge (both empirical and theoretical), while aleatory
variability is assumed to be purely random. Clearly, such
a distinction is somewhat idealistic in that some of the
observed aleatory variability could be due to systematic
effects (hence, strictly being a source of epistemic uncertainty), resulting from, for example, the simplified nature
of GMPEs. Despite this idealization, the separation of aleatory variability and epistemic uncertainties is still important
in assessing performance over time (e.g., Der Kiureghian and
Ditlevsen, 2008).
The consideration of epistemic uncertainties in PSHA via
logic trees was first proposed by Kulkarni et al. (1984).
Logic trees allow the consideration of alternative models
and the values of their parameters to be used in PSHA, with
1554
Consideration and Propagation of Epistemic Uncertainties in New Zealand PSHA

each alternative possibility given a degree of belief such that
the sum of the possibilities adds to 1.0. Although, there is
relatively little literature on the logic-tree concept (as a result
of the field of PSHA being largely practice driven), Bommer
et al. (2005) and Bommer and Scherbaum (2008) provide
useful overviews and potentially adverse consequences.
Since their inception, logic trees have become commonly
employed in PSHA to attempt to account for epistemic uncertainties. However, at present, nationwide seismic-hazard
analyses for New Zealand (Stirling et al., 2011, 2002) do not
explicitly account for epistemic uncertainties, instead using
only preferred values of parameters in the ERF and a single GMPE.
This article presents a summary of the results obtained
from consideration of epistemic uncertainties in nationwide
PSHA in NZ. First, the methodology by which epistemic uncertainties are considered is discussed. Second, the influence
of epistemic uncertainties on the nationwide magnitude
frequency distribution of fault-based seismicity and seismichazard analyses for various intensity measures and locations
are examined. The observations are discussed with reference
to the presently employed PSHA methodology, as well as
the significance of epistemic uncertainties obtained in relation to other studies. Finally, the limitations of the study are
addressed. This article is intended to provide an overview of
this study, and more elaborate discussion of the methodology
and presentation of results can be found in Bradley
et al. (2011).
Adopted Epistemic Uncertainty Methodology

Consideration of epistemic uncertainties in a hazard
assessment serves two primary objectives. First, it provides
an assessment of the significance of various model uncertainties of the estimates of the particular hazard considered and
subsequently can be used to provide confidence estimates.
Second, the results can be used to understand the sensitivity
of the results to specific inputs and hence which sources of
modelling uncertainty are critical in precisely quantifying the
hazard. In this study, emphasis was placed primarily on the
latter of these two objectives. The primary reason for this is
that no attempt was made to consider epistemic uncertainties
in the inputs for a PSHA in an exhaustive manner; and, therefore, the results obtained cannot be considered to represent
the full distribution of uncertainty in the hazard estimates.
Second, the study was conducted in the context of continuing
improvement of seismic-hazard estimation in NZ and therefore ascertaining how the precision of such hazard estimates
can be most efficiently improved with limited resources. As
such, and in light of the fundamental nature of several of the
assumptions used in PSHA (as elaborated by Field, 2007), the
consideration of epistemic uncertainties is kept relatively
simple. The adopted methodology is separately explained
for the GMPE and ERF components of PSHA.
1555
Epistemic Uncertainty in NZ GMPE

Epistemic uncertainty in ground-motion prediction results from (i) uncertainties in the values of input parameters
of a GMPE and (ii) uncertainty regarding the appropriateness
of one or more prospective GMPEs. Uncertainties in GMPE
parameter inputs relating to fault and wave-propagation details are incorporated within the ERF. Uncertainties in GMPE
parameter inputs relating to local site effects are not considered herein; the primary reason for this omission is that the
conventional development of GMPEs does not remove the
epistemic uncertainty of such parameters (Gehl et al., 2011;
Moss, 2011), which would result in double-counting of such
uncertainty. Uncertainty regarding the appropriateness of
prospective GMPEs is considered via the conventional use
of logic trees as discussed subsequently in this article.
The McVerry et al. (2006) model, developed based on
pre-1995 NZ strong-motion data, is the only NZ-specific
GMPE developed to predict both peak ground accleration
(PGA) and spectral acceleration (SA) and consequently is the
only GMPE considered in conventional NZ PSHA (Stirling
et al., 2011), referred to herein as the 2010 New Zealand
national seismic hazard model (NSHM). Hence, the present
consideration of GMPE uncertainty requires the consideration of foreign GMPEs. Bradley (2010) and Bradley et al.
(2011) examined the applicability of various foreign GMPEs,
as well as the McVerry et al. (2006) model, based on observed strong-motion data in NZ up to 2009. These studies
considered four GMPEs for active shallow crustal events and
three for subduction slab and subduction interface events. A
comparison of the magnitude and source-to-site distance
scaling of the considered GMPEs is given in Figure 1. For
brevity, specific details of these studies are omitted here,
other than the recommended logic tree and associated
weights to account for GMPE uncertainty provided in Table 1.
In particular, it is noted that the Atkinson and Boore (2003)
model was assigned zero weight due to its poor comparison
with recorded data (Bradley, 2010). Te considered GMPEs of
McVerry et al. (2006), Zhao et al. (2006), Boore and Atkinson (2008), Chiou et al. (2010), and Atkinson and Boore
(2003) henceforth are abbreviated as McV06, Z06, BA08,
C10, and AB03, respectively.
Epistemic Uncertainty in NZ ERF
The current methodology adopted for PSHA in NZ, as
implemented by Stirling et al. (2002, 2011), is based on a
combination of fault-based sources and distributed seismicity
as depicted in Figure 2. The fault-source model uses the
dimensions and slip rates of mapped fault sources to develop
magnitudefrequency estimates for characteristic earthquakes, and the spatial distribution of historical seismicity is
used to develop magnitudefrequency estimates for the background seismicity model. Because the fault-based seismicity
is assumed to be characteristic in nature, it is assumed that
events less than this characteristic magnitude are modeled by
the distributed background sources.
1556
Inter-event Std Dev,
(e)
10km
-1
rup
10
=50km
McV06
Z06
BA08
C10
-2
6
7
8
Moment magnitude, M w
(d)
Subduction interface
PGA
10
-1
rup
=50km
Rrup=120km
10
-2
McV06
Z06
AB03
0.4
6
7
8
Moment magnitude, M w
(f)
Mw= 5
0.3
Mw= 7
0.2
0.1
0 -2
10
McV06
Z06
BA08
C10
-1
10
Period, T (s)
10
10
10
Active shallow crustal

PGA
Mw =7.5
-1
Mw =5.5
10
McV06
Z06
BA08
C10
-2
10
Spectral acceleration, SA(T ) (g)
10
(b)
Active shallow crustal

PGA
10
10
Source-to-site distance, R rup (km)
Subduction interface
PGA
10
-1
M w=7.5
10
10
-2
McV06
Z06
AB03 M w=5.5
-3
10
0.7
Intra-event Std Dev,
(c)
(a)
B. A. Bradley, M. W. Stirling, G. H. McVerry, and M. Gerstenberger
10
Source-to-site distance, R rup (km)
Mw = 5
0.6
0.5
0.4
0.3
McV06
Mw = 7
Z06
BA08
C10
0.2 -2
10
-1
10
Period, T (s)
10
Figure 1. Variation in PGA amplitudes predicted by the various GMPEs considered in this study: (a,b) PGA scaling with magnitude and
source-to-site distance for active shallow crustal earthquakes; (c,d) PGA scaling with magnitude and source-to-site distance for subduction
interface earthquakes; and (e,f) interevent and intraevent standard deviation dependence on vibration period. Figures for subduction slab
events are omitted and can be found in Bradley (2010). (McV06, McVerry et al., 2006; Z06, Zhao et al., 2006; BA08, Boore and Atkinson,
2008; C10, Chiou et al., 2010; AB03, Atkinson and Boore, 2003.) The color version of this figure is available only in the electronic edition.
Only epistemic uncertainties in the fault-based component of the ERF are considered in this study, which is consistent with similar studies conducted elsewhere (e.g.,
Working Group on California Earthquake Probabilities
[WGCEP], 2003). That is, no uncertainties in background
seismicity are considered. The implications of this are discussed in light of the results obtained subsequently. Furthermore, given underlying fundamental assumptions (e.g., fault
segmentation and time-independence), the consideration of
uncertainties in the fault-based component of the ERF is limited to the consideration of uncertainties in the parameters
used to specify the characteristic distribution of each fault
source rather than to consider different magnitudefrequency
distributions for fault sources. Hence, the consideration and
propagation of such uncertainties first requires an overview
of the methodology by which the characteristic magnitudefrequency distribution of modeled faults is determined.
Deterministic Calculation of Source Magnitudes and Rates

of Occurrence. The seismic hazard for characteristic
ruptures, as modeled in the fault-model of Stirling et al.
Table 1
Logic-Tree Weights Used for Various GMPEs for
Different Tectonic Environments (Bradley et al., 2011)
Model
Abbreviated Name
Active Shallow Crustal Sources

McVerry et al. (2006)
McV06
Zhao et al. (2006)
Z06
Boore and Atkinson (2008)
BA08
Chiou et al.(2010)
C10
Subduction Slab and Interface Sources
McVerry et al. (2006)
McV06
Zhao et al. (2006)
Z06
Atkinson and Boore (2003)
AB03
Weight
0.2
0.2
0.28
0.32
0.4
0.6
0.0
1557
Figure 2. (a) Modeled characteristic fault sources and (b) seismotectonic background seismicity regions in the 2010 New Zealand National Seismic Hazard Model (NSHM; figure modified from Stirling et al., 2011). The locations of Christchurch and Wellington, which are
considered as case study locations in this article, are also annotated. (KMF, Kapiti-Manawatu faults; North Is, North Island; MFS, Marlborough fault system). The color version of this figure is available only in the electronic edition.
(2002, 2011), is governed by the magnitude M w and recurrence interval of the characteristic rupture. The fault is
described by several geometrical parameters (length, L; top
of rupture extent, Dtop ; bottom of rupture extent, Dbottom ;
dip, ) and deformation parameters (slip rate, s_ ; coupling
coefficient, c). Based on the fault geometry, a magnitudescaling relation is used to compute the magnitude of the rupture. For New Zealand faults, four different magnitudescaling relations are used, depending on the classification of
the fault. For crustal plate-boundary faults, the relationship
of Hanks and Bakun (2002) is utilized:
4
Median: Mw 3:39 log10 A;
3
Sigma: Mw 0:22:
and
(1)
For normal faults in volcanic and rift environments, the

relation of Villamor et al. (2001) is utilized:
4
Median: Mw 3:39 log10 A;
3
Sigma: Mw 0:195:
and
where W is the fault width.
Median: Mw 4:441 0:846 log10 A;

Sigma: Mw 0:286:
and
(4)
No scaling relation is given for subduction slab events

because these are modeled as point sources in the background seismicity model. The median of equations (1)(3)
are used in computing the characteristic magnitudes of crustal sources in the NSHM (i.e., Stirling et al., 2002, 2011), but
the standard deviations are not considered. For subduction
interface events, Stirling et al. (2011) used assumed values
of average displacement and computed magnitudes from the
definition of moment magnitude (equation 6), however, the
scaling relation of equation (4) will be used herein.
For the given moment magnitude estimated from a magnitude-scaling relation, the seismic moment M 0 (in units of
Nm) can be computed from (Hanks and Kanamori, 1979)
log10 M 0 9:05 1:5M w :
(5)
(2)
For all other crustal faults, the relationship of Berryman

et al. (2001) is utilized:
2
4
Median: M w 4:19 log10 W log10 L;
3
3
Sigma: Mw 0:18:
For subduction interface events, the relationship of

Strasser et al. (2010) is utilized:
and
(3)
The total seismic moment rate for the fault can be

computed as follows (Brune, 1968):
M_ 0 A_sc;
(6)
where is the shear rigidity of the fault interface; s_ is the

average slip rate over the area of the fault surface, and c is
the coupling coefficient. Based on the assumption that all the
moment rate accumulating on the modeled fault surface is
released in characteristic events (recalling the assumption
1558
for i=1:nsimulation
for j=1:nfaults
)
1. Generate a random set of geometrical fault parameters (
and deformation parameters ( ).
2. Using the appropriate magnitude scaling relation (Equations (1)(4)), determine the mean and standard deviation of the characteristic magnitude.
3. From the mean and standard deviation of the characteristic magnitude, generate a randomly realized magnitude,
(for fault j and realization i).
get the associated seismic moment (Equa4. For the generated magnitude
tion (5) and (6)), and determine the mean annual rate of occurrence,
(Equation (7)).
end
end
Figure 3. Monte Carlo procedure for fault-based epistemic

uncertainty consideration.
that noncharacteristic events are modeled as background
sources), the mean occurrence rate can be computed from
M_ 0
:
M0
(7)
Consideration of Epistemic Uncertainties in the Fault

Model. The values of the characteristic magnitude and frequency of modeled faults in the NSHM are uncertain as a
result of uncertainties in fault geometry and fault deformation and of uncertainty in empirical magnitudegeometryscaling relationships (i.e., equations 14). To account for
such uncertainties, the characteristic magnitude and frequency of modeled faults used in PSHA were obtained based
on the Monte Carlo simulation procedure given in Figure 3.
Specific Values of Fault-Based Uncertainties Used in the
Present Study. Given that epistemic uncertainties are a result of a lack of knowledge (both theoretical and empirical),
those faults that have had less attention devoted to them will
be modeled with parameters/relationships that have a greater
epistemic uncertainty. Therefore, ideally one would have
estimates for the magnitude of the parameter/relationship
uncertainties that are fault specific. Unfortunately, this is
not the case for the faults in the NZ seismic-hazard model,
with the majority of such data not presently cataloged. As a
result, the approach taken here is to make use of available
fault-specific data and judgement in the absence of such data
to assign uncertainties to each of the parameters. A key consideration in the subsequent analyses conducted is a sensi-
tivity study to assess the importance of each parameter

uncertainty in the overall picture of seismic hazard, therefore
indicating which parameters deserve more rigorous estimation in future.
In determining the magnitude of uncertainties to assign
to faults without specific parameter uncertainty estimates,
use was made of both NZ-specific and foreign data, primarily
from the Working Group on California Earthquake Probabilities (WGCEP, 2003). The following information from
WGCEP was obtained (Bradley et al., 2011): (i) coefficients
of variation (COV) for fault length range from 0.06 to 0.28,
with a mean of 0.15; (ii) fault width COVs range from 0.08 to
0.12, with a mean of 0.10; (iii) fault slip rate COVs range
from 0.06 to 0.34, with a mean of 0.18. In addition to the
parameter uncertainty magnitudes in the WGCEP study,
various estimates are also available for well-studied and/or
recently studied NZ faults. Examples of such estimates
include the following slip rate COVs: 0.14, Wellington fault
(Van Dissen, 2010); 0.10, Ostler and Irishman Creek faults
(Amos et al., 2007); 0.16, Ohariu fault (Heron et al., 1998);
0.1, Hope fault (Langridge and Berryman, 2005); 0.15,
Porters Pass fault (Howard et al., 2005); 0.16, Blue Mountain
fault (Pace et al., 2005); 0.15, Ohariu fault (Litchfield et al.,
2006); 0.25, Taupo rift faults (Villamor and Berryman,
2006); 0.13, Wairau fault (Zachariasen et al., 2006); 0.08,
Alpine fault (Zachariasen et al., 2006); 0.08, Wairarapa fault
(Van Dissen and Berryman, 1996); and 0.1, Hikurangi subduction zone sources (Wallace et al., 2009); among others.
On the basis of these and the WGCEP values, a COV of 0.20
was assigned to slip-rate uncertainties for general faults.
Estimates for fault length, depth, and dip uncertainties for
NZ faults are few and far between; and, therefore, use was
largely made of the WGCEP (2003) study in assigning
values for these uncertainties.
Table 2 provides the values of parameter uncertainties to
assign to faults without fault-specific data, as well as the assumed distribution. Mean values of parameters for individual
faults are documented in Stirling et al. (2011). For use in the
later Monte Carlo simulations, these parameter distributions
were truncated at two standard deviations from the mean.
Truncated normal distributions were used for representing
Table 2
Uncertainties Assigned to Fault Parameters in the Absence of Fault-Specific Data
Parameter
Uncertainty
Assumed Distribution
Length of fault plane, L

Top of rupture extent, Dtop
Bottom of rupture extent, Dbottom
Fault dip,
Fault slip rate, s_
Coupling coefficient, c
Magnitude-scaling relationship
COV 0:15*
p1
km
12
1 km
5
COV 0:20
COV 0:15 [interface sources only]
Fault-type specific [i.e., equations (1)(4)]
Truncated normal
Bounded uniform
Truncated normal
Truncated normal
Truncated normal
Truncated normal
Truncated normal
*COV, coefficient of variation; equal to the standardpdeviation

divided by the mean.
For a uniform distribution, a standard deviation of 1= 12 refers to maximum and minimum values that are
0.5 units above and below the mean value.
If the minimum depth was negative, then it was set to zero and the distribution function renormalized.
Table 3
that use different definitions of magnitude, source-to-site

distance, and fault mechanism to be consistently utilized.
All results are presented for the geometric mean horizontal
component of ground motion (with the rotated geometric
mean definition used by BA08 and C10 considered to be practically equivalent (Beyer and Bommer, 2007). Hence, the
only parameter-consistency consideration that remains
involves those parameters that relate to the classification
of local site effects. In line with the results of Bradley (2010),
Table 3 provides the parameters used for site class B and D
sites according to the NZ loadings standard (NZS 1170.5;
Standards New Zealand, 2004).
A total of 50 Monte Carlo simulations for each GMPE
were performed, which was found to provide a sufficient level of accuracy.
Site Class Compatibility of the Various GMPEs

McVerry et al.
(2006)
Zhao et al.
(2006)
V S30 (m=s)
Z1:0 (m)
B
D
B
D
Soft rock
Medium soil
760
250
50
500
*Standards New Zealand, 2004.

V
S30 as was utilized in the BA08 and C10 GMPEs.
Z1:0 as was utilized in the C10 GMPE.
uncertainties in parameter values with the exception of fault

depth, for which a uniform distribution was used. Because
there is a paucity of data to estimate even the size of faultspecific uncertainties (i.e., COV), then it is not possible to
scrutinize appropriate distributions for each parameter considered. Given that the size of uncertainty is likely more
important than the distribution considered, the normal distribution was adopted for simplicity.
Uncertainty in Nationwide MagnitudeFrequency

Distribution of Modeled Faults
Figure 4 illustrates the nationwide magnitudefrequency
distributions from fault-source seismicity due to the consideration of various parameter uncertainties. Each of the plots in
Figure 4 illustrate the NSHM distribution obtained from
Stirling et al. (2011), the individual distributions from 50
Monte Carlo simulations (gray lines), as well as the mean,
and the 16th and 84th percentiles of the simulated distributions. Generally, the mean distribution and the NSHM distribution are relatively similar. The main exception to this is for
Application of Methodology to NZ
Seismic-Hazard Analysis
Implementation in OpenSHA
(a)
Annual exceedance frequency, M w
The aforementioned methodology was implemented in

the open-source seismic-hazard-analysis software OpenSHA
(Field et al., 2003). Within OpenSHA, earthquake ruptures
are treated as objects, therefore easily allowing GMPEs
10
(b)
NSHM
Mean
10
10
10
16th, 84th
-1
-2
-3
Considering fault-length uncertainty
10
-4
NZS 1170.5*
10
NSHM
Mean
10
10
10
16th, 84th
-1
-2
-3
Considering fault depth-of-rupture uncertainty
10
-4
10
(d)
NSHM
Mean
10
10
10
10
16th, 84th
-1
-2
-3
Considering slip-rate uncertainty
-4
Magnitude, M w
Magnitude, M w
Magnitude, M w
(c)
1559
10
NSHM
Mean
10
10
10
10
16th, 84th
-1
-2
-3
-4
Considering M w -scaling uncertainty
Magnitude, M w
Figure 4. Epistemic uncertainty in the nationwide magnitudefrequency relationship due to fault-source seismicity considering (a) faultlength uncertainty only; (b) depth-of-rupture uncertainty only; (c) slip-rate uncertainty only; and (d) magnitude-scaling relationship uncertainty only. Values are included for the 16th and 84th percentiles. The color version of this figure is available only in the electronic edition.
1560

0.8
0.7
Sigma, ln
0.6
0.5
0.4
1.2
(b)
Mw Scaling
Length
Rup Bottom
Rup Top
Slip Rate
Dip
Coup Coef
Mw Scaling
Geometry
Deformation
All
Sigma, ln
(a)
0.3
0.8
0.6
0.4
0.2
0.2
0.1
0
Magnitude, M w
Magnitude, M w
Figure 5. Lognormal standard deviation in the exceedance rate of various magnitudes considering various fault parameter uncertainties:
(a) individual parameter uncertainties and (b) parameter uncertainties by group. (See text for details.) The color version of this figure is
available only in the electronic edition.
16th and 84th percentiles) of 73%88%. For a magnitude 8
or greater event, the NSHM probability is 35%, while the
mean probability is 25% with a 68% confidence interval
of 1231% for this study. As a result of the fault parameter
uncertainties not correlating between different faults, a
magnitudefrequency distribution for all faults has a smaller
uncertainty than that for a smaller subset of faults. Hence, if
the seismic hazard at a single site is dominated by a small
subset of seismic sources, such uncertainties will be more
pronounced.
Figures 46 illustrate that the magnitude-scaling relation
uncertainty tends to be the dominant uncertainty for modeled
faults. Magnitude-scaling relation uncertainty was also treated on a fault-specific basis through the use of equations (1)
(4). Hence, this suggests that the approximate uncertainty
values that are prescribed for the other fault parameters
(in cases where fault-specific parameter uncertainties were
10 0
NSHM
large magnitudes, where the parameter uncertainties result in

a smoother variation in frequency compared with the step
changes that result in the case of no uncertainties. It is also
worth recalling that, in this study, magnitudes for subduction
interface events (which dominate the magnitudefrequency
distribution around M w 8) were estimated using the scaling
relation of Strasser et al. (2010), compared with the assumed
average displacements utilized by Stirling et al. (2011). At
small magnitudes, the consideration of epistemic uncertainties leads to a higher exceedance rate. The reason for this
difference is that if the simulated characteristic magnitude is
smaller than the mean value, then the corresponding occurrence rate will be higher. Comparison of the size of the uncertainty in the magnitudefrequency distribution due to each
of the parameter uncertainties illustrates that magnitudescaling-relation uncertainty is the most significant, followed
by fault-length uncertainty. This observation is seen explicitly
in Figure 5a, which illustrates the lognormal standard deviation of the exceedance frequency as a function of magnitude
for various parameter uncertainties. Figure 5b presents the
uncertainty in the magnitudefrequency distribution when
multiple uncertainties are considered. Specifically, all the
parameter uncertainties related to fault geometry (i.e., fault
length, rupture top, rupture depth, fault dip), fault deformation
(i.e., slip rate and coupling coefficient for interface sources)
were considered simultaneously. Figure 5b illustrates that,
despite the grouping of uncertainties, magnitude-scaling
relation uncertainty remains the dominant uncertainty in the
nationwide fault-based magnitudefrequency distribution.
Figure 6 presents the uncertainty in the magnitude
frequency distribution due to the consideration of all fault
parameter uncertainties. As with the previous figures, the
NSHM and mean curves are similar and show the aforementioned variation at larger magnitudes. Based on the timeindependent nature of the ERF (i.e., a Poisson temporal
distribution), the 50-year exceedance probability of an event
with magnitude greater than 7 is essentially 100%. The
NSHM 50-year probability of a magnitude 7.5 or greater
event is 79%, while this study results in a mean probability
or 84% and a 68% confidence interval (i.e., that between the
Mean
16 th, 84th
10 -1
10 -2
10 -3
Considering all fault uncertainties
10 -4
Magnitude, M w
Figure 6. Uncertainty in the nationwide fault-based seismicity

magnitudefrequency distribution, considering all fault parameter
uncertainties. Values are included for the 16th and 84th percentiles.
The color version of this figure is available only in the electronic
edition.
Seismic-Hazard Analysis Case-Study Sites

Considered
In order to examine the importance of various epistemic
uncertainties for a range of situations, seismic-hazard
analyses for two different sites, soil classes, and intensity
measures are presented. Generic locations in Wellington and
Christchurch are selected, both because they are NZs largest
two cities and because at the 475-year return period, Wellingtons hazard is dominated by fault sources, while in contrast, background sources provide a significant contribution
for Christchurchs hazard. Hazard analyses are performed at
these two generic locations for both site class B (weak rock)
and site class D (deep soil) site conditions (NZS 1170.5,
Standards New Zealand, 2004). Finally, hazard analyses are
performed for both PGA and SA(2.0), in order to generally
depict the magnitude of epistemic uncertainties at short and
long vibration periods. A minimum magnitude of M w 5.0
was considered in calculations.
Uncertainty in Seismic Hazard for Wellington
(a)
50-year exceedance probability
Figure 7 illustrates seismic-hazard curves for PGA and

site class D at a generic location in central Wellington
(latitude 41:2889, longitude 174.7772) for an exposure
period of 50 years. In Figure 7, and similar subsequent
figures, 50 simulations for each different GMPE logic-tree
combination (i.e., Table 1) were used to sample the considered ERF epistemic uncertainties (i.e., Fig. 3). These individual simulations are shown in thin lines, with the mean of
this ERF uncertainty (but for a single GMPE) shown in thicker
lines. From the observation that the individual simulations
tend to be in distinct groups for each GMPE combination,
it is apparent that the GMPE uncertainty tends to be larger
than that due to ERF uncertainty. The importance of GMPE
uncertainty for different tectonic source types can be seen by
examining the differences in the seismic-hazard curves due
to different GMPE combinations in Figure 7. For example,
Wellington
Site Class D
10 -1
Mean of realizations
with a single GMPE
Individual
realizations
10 -3
10 -2
C10
Z06
McV06
BA08
Uncertainty in Seismic Hazard for Christchurch

Figure 9 illustrates the seismic-hazard curves for PGA
and site class D at a generic location in central Christchurch
(latitude 43:5300, longitude 172.6203) for an exposure
period of 50 years. Because Christchurch is subject to
negligible seismic hazard from subduction zone sources, only
results for the four different crustal GMPEs being considered
in this study are explicitly annotated. An obvious observation
(b)
10 0
10 -2
Figure 7a illustrates the difference between the hazard using

the Z06 subduction GMPE but different crustal GMPEs, while
Figure 7b illustrates the difference in the hazard using the
McV06 subduction GMPE but different crustal GMPEs. Both
Figure 7a and 7b clearly illustrate that the Z06 crustal GMPE
leads to a significantly larger hazard than that using other
crustal GMPEs, which give similar hazards and are consistent
with the median and sigma GMPE scaling in Figure 1.
While the seismic hazards obtained using the C10, BA08,
and McV06 GMPEs in Figure 7a and 7b are similar relative to
each other, they differ due to the differences in the subduction zone GMPE used. Figure 8 explicitly illustrates the significance of subduction zone GMPE uncertainty for PGA and
site class D conditions in Wellington. Both Figure 8a and 8b
illustrate that the Z06 GMPE tends to result in a larger hazard
compared with the McV06 GMPE.
Comparison of Figures 7 and 8 illustrates that both
active shallow crustal and subduction zone sources contribute significantly to the hazard, and its epistemic uncertainty,
for the given conditions. This observation is consistent with
the knowledge that Wellington is located close to several active shallow crustal faults (most notably the Wellington fault)
and to the Hikurangi subduction zone, for which the interface
passes approximately 20 km underneath Wellington (Stirling
et al., 2011).
In interpretation of the observed epistemic uncertainties
for Wellington, it is worth bearing in mind that the seismic
hazard at this location is generally dominated by fault-based
seismicity rather than background sources. Hence, the
neglect of background seismicity epistemic uncertainties is
likely inconsequential for Wellington.
Effect of crustal GMPE

All results use the Z06
subduction GMPE
10 -1
10 0
Peak ground acceleration, PGA (g)
unavailable) are likely to have a second-order effect on the

size of the total uncertainty.
1561
10 0
Wellington
Site Class D
with a single GMPE
10 -1
Individual
realizations
10 -2
10 -3
10 -2
C10
McV06
Z06
BA08
Effect of crustal GMPE

All results use the McV06
subduction GMPE
10 -1
10 0
Figure 7. Seismic-hazard curves for site class D PGA in Wellington: (a) using the Z06 subduction GMPE and various crustal
GMPEs; and (b) using the McV06 subduction GMPE and various crustal GMPEs (which tend to be clustered together and appear as a
thicker gray line here). The color version of this figure is available only in the electronic edition.
1562
Wellington
Site Class D
with a single GMPE
10 -1
Individual
realizations
10 -2
Z06
McV06
10 -3
10 -2
Effect of subduction GMPE

All results use the C10
crustal GMPE
10 -1
10 0
(b)
10 0
(a)
Wellington
Site Class D
10 -1
10 0
Individual
realizations
10 -2
10 -3
McV06
Z06
Effect of subduction GMPE
All results use the McV06
crustal GMPE
10 -2
with a single GMPE
10 -1
10 0
Figure 8. Seismic-hazard curves of peak ground acceleration for site class D in Wellington: (a) using the Z06 subduction GMPE and
various crustal GMPEs; and (b) using the McV06 subduction GMPE and various crustal GMPEs. Individual realizations are shown in thin
lines, while thicker lines represent the mean hazard of the 50 realizations of fault-based seismicity for each GMPE combination (which tend
to be clustered together and appear as a thicker gray line here). The color version of this figure is available only in the electronic edition.
in the nature of epistemic uncertainty in the seismic-hazard

curves in Figure 9, relative to that for Wellington (i.e., Figs. 7
and 8), is the reduction in the seismic-hazard uncertainty due
to ERF uncertainty. This observation is the result of two
factors. First, unlike Wellington, which is a high seismic
region dominated by a handful of fault-based seismic sources,
Christchurch is a region of relatively lower seismicity, and significant contributions to the seismic hazard are made by both
numerous fault-based sources and background seismicity as
depicted in Figure 10. As no epistemic uncertainties were considered in the background seismicity model, there is consequently less uncertainty in the total seismic hazard due to
both fault and background seismicity. The second factor in the
reduction in ERF uncertainty for Christchurch is a result of the

there being numerous fault-based sources that significantly
contribute to the seismic hazard. As previously noted with
respect to the uncertainty in the nationwide fault-based
magnitudefrequency distribution (i.e., Fig. 6), because faultsource uncertainty is not correlated between different faults, a
larger number of fault sources providing a substantial contribution to the seismic hazard will consequently lead to a reduction in the magnitude of the epistemic uncertainty in the
seismic hazard. The effect of removing the background
sources for the epistemic uncertainty in the Christchurch
seismic hazard is explored in the following section.
Christchurch Hazard with the Removal
of Background Seismicity
10 0
0.15
Christchurch
Site Class D
0.1
0.05
0.1
0.2
0.3
10 -1
10 -2
C10
BA08
Z06
McV06
10 -3 -2
10
10 -1
10 0
Figure 9. Seismic-hazard curves of peak ground acceleration

for site class D in Christchurch. Individual realizations are shown
in thin lines, while thicker lines represent the mean hazard of the 50
realizations of fault-based seismicity for each GMPE combination
(see inset) (which tend to be clustered together and appear as a
thicker gray line here). The color version of this figure is available
only in the electronic edition.
Figure 11 illustrates the seismic hazard computed with

and without the consideration of background seismicity
sources. Comparison with Figure 9, in which fault and background source are considered, illustrates that the background
sources are significant, with the 10% in 50-year exceedance
probability being approximately 0:20:24g when background
sources are considered (i.e., Fig. 9) and 0:120:19g when
background sources are neglected (i.e., Fig. 11). In additiona,
Figure 11 shows that, even in the fault-based seismicity
only case, the ERF component of the total epistemic uncertainty is insignificant compared with the GMPE uncertainty
(i.e., the variation in the seismic hazard considering ERF
uncertainty, but only a single GMPE is relatively small).
Interestingly, the McV06 GMPE results in the largest seismic
hazard for exceedance probabilities greater than 10% in
50 years when background sources are considered but results
in the smallest seismic hazard when background sources are
neglected. This occurs because the McV06 model is known to
significantly overpredict ground motions produced by small
magnitude (M w < 6) events (e.g., Fig. 1a), which are the predominant background sources for the PGA hazard at the 10%
in 50-year exceedance probability (e.g., Fig. 10a).
1563
Figure 10.
Seismic-hazard deaggregation for Christchurch (site class D) for 2% in 50-year exceedance probability: (a) PGA and (b) SA
(2.0), considering both fault and background seismicity sources. The color version of this figure is available only in the electronic edition.
Magnitude of Epistemic Uncertainties in Terms

of Design Seismic Demands
In order to understand the salient differences in epistemic uncertainties for NZ PSHA, Figures 12 and 13 illustrate
the seismic-hazard curves for the eight different cases considered with explicit indication of the mean, median, and 5th
and 95th percentiles, while Table 4 provides these values for
the 10% and 2% exceedance probabilities in 50 years. The
magnitude of epistemic uncertainties is clearly significant in
determining the ground-motion hazard level for a given exceedance probability or vice versa. For example, while the
mean hazard for the 10% and 2% exceedance probabilities
in 50 years for PGA and site class B in Wellington are 0:48g
10
0.15
10
10
Christchurch
Site Class D
No Background
Sources
-1
0.1
0.05
0.1
0.2
0.3
-2
C10
BA08
Z06
10
McV06
-3
10
-2
Figure 11.
-1
10
10
Seismic-hazard curves of peak ground acceleration

for site class D in Christchurch as a result of only fault-based seismicity (i.e., neglecting background sources). Individual realizations
are shown in thin lines, while thicker lines represent the mean hazard of the 50 realizations of fault-based seismicity for each GMPE
combination (see inset) (which tend to be clustered together and
appear as a thicker gray line here). The color version of this figure
is available only in the electronic edition.
and 0:93g, respectively, the 90% confidence intervals are

0:41g; 0:57g and 0:76g; 1:19g. That is, the 90% confidence interval for the 10% in 50-year exceedance probability
is almost 0:2g and for the 2% in 50-year exceedance probability is slightly over 0:4g. On the other hand, for SA(2.0)
and site class D in Christchurch, the mean SA(2.0) hazard for
the 10% and 2% exceedance probabilities are 0:16g and
0:28g, respectively, while the 90% confidence interval
ranges on these mean values are 0:11g; 0:21g and 0:21g;
0:35g, respectively. That is, the range of the 90% confidence
interval is 0:1g and 0:14g for these two exceedance probabilities, respectively. Table 4 also illustrates the values from the
New Zealand loadings standard NZS 1170.5 (Standards New
Zealand, 2004), for the considered cases. For the most part,
the values are in agreement, although the mean values obtained in this study are generally larger than the NZS
1170.5 values for Wellington and smaller than the corresponding values for Christchurch.
Table 5 provides another interpretation of the data in
Table 4 based on the ratio of the difference between the 95%
and 5% fractiles and mean ground-motion intensity for a
given exceedance probability. From Table 5, the average
ratio is approximately 0.49, but such ratios vary significantly
from 0.200.85. There is relatively little difference between
the size of the uncertainty for Wellington and Christchurch,
with average ratios of 0.46 and 0.52, respectively. Again, it is
noted that the magnitude of the uncertainty for Christchurch
in particular is artificially low due to the neglection of background seismicity uncertainty, which will be most noticeable
for PGA, in which frequent small magnitude events are most
important (i.e., Fig. 10). The size of the ratios in Table 5
generally increases with reducing exceedance probability
for Wellington but are essentially constant for Christchurch.
Finally, it is noted that the ratios are not systematically larger
for site class D cases than for site class B cases, which may
be assumed a priori because of the different manner in which
GMPEs account for local site response, something which
should be relatively insignificant for soft rock (class B) sites.
1564
10
(b)
Wellington
Site Class B
10
-1
5%, 95% confidence

bounds
Mean
Hazard
NSHM Hazard
10
-2
10
(a)
Median Hazard
10
-1
(d)
Wellington
Site Class B
10
10
NSHM Hazard
-1
5%, 95% confidence

bounds
-2
Median Hazard
-1
NSHM Hazard
5%, 95% confidence
bounds
10
-2
Median Hazard
Mean Hazard
10
Mean
Hazard
-3
10 -2
10
10
10 -2
10
10
10
(c)
Wellington
Site Class D
-3
-3
10 -2
10
-1
10
10
Wellington
Site Class D
10
NSHM Hazard
-1
5%, 95% confidence

bounds
10
-2
Mean
Hazard
Median Hazard
-3
-1
10 -2
10
10
10
Spectral acceleration, SA(2.0) (g)
-1
10
10
Figure 12. Comparison of seismic-hazard analyses with epistemic uncertainty and those of the conventional NZ NSHM for the four
cases considered in Wellington. The color version of this figure is available only in the electronic edition.
10
Christchurch
Site Class B
10
10
(b)
NSHM Hazard
-1
5%, 95%
confidence
bounds
-2
10
(a)
Mean Hazard
Median Hazard
-3
(c)
10
10
NSHM Hazard
-1
-2
5%, 95%
confidence
bounds
Mean Hazard
Median Hazard
-3
10 -2
10
Figure 13.
10
NSHM Hazard
-1
5%, 95% confidence

bounds
10
-2
Mean Hazard
Median Hazard
10 -2
10
Christchurch
Site Class B
10
Christchurch
Site Class D
-3
-1
10
10
(d)
10
10 -2
10
-1
10
10
Christchurch
Site Class D
10
10
NSHM Hazard
-1
-2
5%, 95%
confidence
bounds
Mean Hazard
Median Hazard
-3
-1
10
10
10 -2
10
-1
10
10
Comparison of seismic-hazard analyses with epistemic uncertainty and those of the conventional NZ NSHM for the 4 cases
considered in Christchurch. The color version of this figure is available only in the electronic edition.
1565
Table 4
Summary of Mean, 5th, and 9th Percentile Hazard Values for 10% and 2%
Exceedance Probabilities (PE) in 50 Years
PE 10% in 50 years
IM
Site Class
Wellington
PGA
Rock, B
Soft/deep
SA(2.0) Rock, B
Soft/deep
Christchurch
PGA
Rock, B
Soft/deep
SA(2.0) Rock, B
Soft/deep
soil, D
soil, D
soil, D
soil, D
PE 2% in 50 years
Mean
[5%,95%]
Code*
Mean
[5%,95%]
Code*
0.48
0.53
0.21
0.41
[0.41,0.57]
[0.44,0.72]
[0.16,0.26]
[0.35,0.47]
0.40
0.45
0.23
0.48
0.93
1.01
0.47
0.92
[0.76,1.19]
[0.72,1.50]
[0.39,0.63]
[0.78,1.11]
0.72
0.80
0.42
0.86
0.16
0.22
0.07
0.16
[0.12,0.21]
[0.20,0.24]
[0.05,0.11]
[0.11,0.21]
0.22
0.25
0.13
0.26
0.29
0.36
0.13
0.28
[0.22,0.35]
[0.32,0.41]
[0.09,0.19]
[0.21,0.35]
0.40
0.44
0.23
0.47
*NZS 1170.5 (Standards New Zealand, 2004).
Comparison of Preferred Hazard and the Hazard

with Explicit Epistemic Uncertainties
10
10
Figure 14.
Lower and upper

bounds on uncertainty
from PSHA for SFBA
Bradley (2009)
-1
-2
Wellington
10
One of the fundamental difficulties with assessing epistemic uncertainties in seismic-hazard analyses is that while
the ultimate aim is to represent the uncertainty in the seismichazard estimate, most often the consideration of epistemic
Christchurch
10
Comparison of Epistemic Uncertainty Magnitude for

NZ with that of the San Francisco Bay Area, USA
ERF
uncertainty
only
-3
0.2
0.4
0.6
0.8
Dispersion in hazard probability, lnP(IM=im)
(b)
Mean 50-year exceedance probability
(a)
Mean 50-year exceedance probability
A key benefit of the consideration of epistemic uncertainties is developing an understanding of the impact of
certain assumptions on outcomes of a seismic-hazard analysis. Previous sections in this article have demonstrated that,
of the considered uncertainties, GMPE epistemic uncertainty
is the largest source of uncertainty on PSHA results. Therefore, it is insightful to scrutinize how the present NZ seismichazard analyses using the McV06 model compare with the
results presented here (considering multiple GMPEs). In
addition to the previously discussed results in Figures 12
and 13, which are based on seismic-hazard analyses with
epistemic uncertainties, the single hazard curve obtained
using the NSHM methodology (that is, neglecting epistemic
uncertainties in ERF parameters and using the McV06 GMPE)
is also depicted. This is referred to as the NSHM Hazard
in Figures 12 and 13. For the PGA hazard in Wellington
(for both site classes B and D), the NSHM hazard is approximately equal to the 5% fractile of the hazard analysis
considering epistemic uncertainties for the 10% and 2% in
50-year exceedance probabilities. For the SA(2.0) hazard in

Wellington (for both site classes B and D), the NSHM hazard
is approximately the 95% fractile for the 10% in 50-year
exceedance probability but is approximately the 5% fractile
for the 2% in 50-year exceedance probability. For Christchurch, with the exception of PGA on site class D, the NSHM
hazard is approximately equal to the 95% fractile of the
hazard considering epistemic uncertainties. As GMPE uncertainty produces the largest variation in the considered seismic-hazard analyses, the observation of the NSHM hazard (as
compared with the mean and the 5th and 95th percentiles
from the seismic hazard considering epistemic uncertainties)
primarily results from the use of only the McV06 GMPE in the
NSHM hazard calculations.
10
10
10
Lower and upper

bounds on uncertainty
from PSHA for SFBA
using NGA GMPEs
Bradley (2009)
-1
Using 1997
version GMPEs
-2
Christchurch
10
-3
ERF and
GMPE
uncertainty
Wellington
0.5
1
1.5
Dispersion in hazard probability, lnP(IM=im)
Comparison of the magnitude of epistemic uncertainty in seismic hazard (dispersion) as a function of exceedance probability
for Wellington and Christchurch compared with that of the SFBA from the PSHA conducted by Bradley (2009): (a) considering ERF
uncertainties only; and (b) considering both ERF and GMPE uncertainty. The color version of this figure is available only in the electronic
edition.
1566
uncertainties simply reflects the range of scientific models
available (Abrahamson, 2006). A consequence of this is that
using available models for a site with little or no data will
indicate smaller epistemic uncertainty compared with a wellstudied site with many available models when clearly the
poorly studied site will have a larger epistemic uncertainty.
Along this line of thought, Bradley (2009) examined the
magnitude of epistemic uncertainty in PSHA conducted for
the San Francisco Bay area (SFBA), using the ERF developed
by the WGCEP (WGCEP, 2003), which extensively considered epistemic uncertainties and for which a range of four
different western USA-specific GMPEs are available. Hence,
it is insightful to compare the magnitude of epistemic uncertainties observed for the SFBA in Bradley (2009) with those
obtained in this study.
Figure 14a and 14b illustrate the magnitude of the epistemic uncertainties obtained in this study in comparison with
the bounds obtained by Bradley (2009) for locations in the
SFBA when considering ERF uncertainty only and considering both ERF and GMPE uncertainty, respectively. The
seismic-hazard uncertainty due to the consideration of ERF
uncertainty in this study is less than that for the SFBA. With
respect to the total seismic-hazard uncertainty (i.e., due to
both ERF and GMPE uncertainty), the results obtained in
this study are similar to those for the SFBA using the Next
Generation Attenuation (NGA) relations but less than that
obtained for the SFBA using the 1997-version GMPEs.
In order to reconcile the differences between the magnitude of the ERF uncertainties observed in this study and those
obtained from SFBA, it is necessary to consider first the ERF
uncertainties that WGCEP (2003) accounted for, which were
(Bradley, 2009, WGCEP, 2003): (i) time dependence of characteristic ruptures; (ii) uncertainty in M w -geometry scaling
relationships; (iii) fault segmentation endpoints; (iv) seismogenic thickness; (v) slip rate; (vi) frequency of multisegment
ruptures; (vii) anelastic slip; and (ix) magnitudefrequency
distribution (i.e., either characteristic or GutenbergRichter).
Thus, with respect to the present study, it can be stated that
uncertainties (ii), (iii), (iv), (v), and (vii) were considered;
however, uncertainties (i), (vi), and (ix) were not. The NZ
ERF methodology (Stirling et al., 2011) is a time-independent
ERF (with the exception of the Wellington, Wairarapa, and
Ohariu faults, which are given time-independent rates
equivalent to their conditional probabilities over 50 years),
with implementation of the characteristic rupture hypothesis
(i.e., without considering multisegment ruptures, or characteristic versus GutenbergRichter magnitudefrequency distributions). One exception to this statement is the treatment
of subduction zone sources, in which two scenarios are
considered based on the unknown seismogenic potential of
some areas of this interface source (Stirling et al., 2011). Field
(2007) noted that, of all of the uncertainties considered in the
WGCEP (2003) forecast, the assumed time-dependence model resulted in the largest uncertainty. Hence, given the omission of time-dependent models in the NZ ERF, as well as the
other omitted uncertainties noted previously in this article, it is
logical to see the reason for the reduced seismic-hazard uncertainty due to consideration of ERF uncertainty observed
in the present study relative to the WGCEP (2003) ERF.
Figure 14b illustrates that despite the lower epistemic
uncertainty in the NZ ERF compared with that of the WGCEP
(2003) ERF, the total epistemic uncertainty (due to both ERF
and GMPE uncertainty) is similar to that obtained for the
SFBA using the NGA GMPEs. Hence, it can be roughly stated
that the GMPE uncertainty for NZ PSHA presented in this
study is larger than that which exists for the SFBA using the
NGA GMPEs, but likely still less than that using the 1997version GMPEs. Again, this result agrees with intuition given
that the NGA GMPEs were developed specifically for the
western USA, while for NZ only the McV06 GMPE (based on
pre-1995 data) is NZ-specific and the remaining models are
foreign.
Discussion and Limitations

The consideration of epistemic uncertainties serves both
to quantify the range of seismic hazard for a given exceedance probability and also importantly to illustrate which
model assumptions lead to the most significant variation in
the estimated hazard. Of the uncertainties being considered,
the previous results illustrated that the results due to the selection of an appropriate GMPE lead to the largest variability
in seismic hazard. Therefore, improvement in NZ-specific
ground-motion prediction is considered a high priority.
Unfortunately, because of the timing at which the GMPE
assessment was completed as part of this study (July 2010),
it was not possible to incorporate the recent earthquakes and
aftershocks that have since occurred in the Christchurch
region. The wealth of empirical information obtained from
these events will be utilized for the development of (multiple)
NZ-specific ground-motion prediction equations, in an attempt to better represent, and ultimately reduce, epistemic
uncertainty resulting from GMPEs.
In this study epistemic uncertainties were considered in
the fault-based component of the earthquake rupture forecast. However, as mentioned in comparison to the WGCEP
(2003) ERF, not all epistemic uncertainties in the fault-based
ERF were considered. Furthermore, for many faults, the magnitudes of the epistemic uncertainties considered were assigned based on judgment because fault-specific estimates
Table 5
Ratio of the Difference between the 95% and 5% Percentiles
Divided by the Mean Ground-Motion Intensity for a Given
Exceedance Probability (PE)
Wellington
IM
PGA
SA(2.0)
Christchurch
Site Class
10% PE
2% PE
10% PE
2% PE
Rock, B
Soft/deep soil, D
Rock, B
Soft/deep soil, D
0.33
0.53
0.44
0.28
0.46
0.77
0.53
0.36
0.49
0.20
0.85
0.63
0.46
0.25
0.75
0.53

were not available; therefore, maintenance of the fault database that is used to develop the fault-based component of the
NZ ERF should devote further attention to the cataloguing of
epistemic uncertainties in raw data.
Epistemic uncertainties were not considered in the
parameters (including their spatial correlation) that define
the GutenbergRichter distribution for background seismicity sources. For regions in which the background seismicity
represents a significant component of the seismic hazard
(e.g., Christchurch), clearly this leads to a systematic underestimation of the epistemic uncertainty in PSHA results. A
methodology for consideration of uncertainties in background source is therefore clearly warranted in order to
assess the magnitude of such uncertainties.
Finally, this study focused on the uncertainties of the
parameters of the fault-based component of the ERF but not
on the ERF methodology itself. Field (2007) noted that the
level of complexity in treating time dependence in the model
in WGCEP (2003) was inconsistent with other fundamental
assumptions such as fault segmentation. A data-driven methodology, with fewer ideologically driven assumptions is
therefore a high priority for future seismicity models that integrate both fault-based and background sources. The same
comment is applicable to the manner in which alternative
GMPEs are considered in PSHA.
Conclusions
This study presented the results of considering epistemic
uncertainties in probabilistic seismic-hazard analyses for
locations in New Zealand. The methodology accounted for
uncertainties in ground-motion prediction via the use of multiple ground-motion prediction equations in a logic tree. Uncertainties in the characteristic rupture magnitude and recurrence
rate of the fault-based component of the seismicity due to
uncertainties in fault deformation, geometry, and empirical
magnitude-scaling relationships were considered via Monte
Carlo simulation. Uncertainties in background seismicity
were not considered. Because of the present lack of faultspecific data quantifying uncertainties for many faults in NZ,
representative values based on judgement and the data available for NZ and foreign faults were utilized where required.
Probabilistic seismic-hazard analyses were conducted for
two vibration periods of spectral acceleration [PGA and
SA(2.0)] for site class B (soft rock) and D (soft/deep soil) conditions in Wellington and Christchurch. The obtained results
illustrated that, of the uncertainties considered, those due to
ground-motion prediction produced the largest variation in
seismic hazard. Of the earthquake rupture forecast uncertainties considered, that due to the magnitude-scaling relationships was the most significant, followed by rupture length.
Data and Resources

All data used in this paper are from the published
sources listed in the references.
1567
Acknowledgments
Financial support for this study was provided by the New Zealand
Earthquake Commission under award EQC 10/593. Constructive comments
from two anonymous reviewers improved the article and are greatly
appreciated.
References
Abrahamson, N. A. (2006). Seismic hazard assessment: Problems with
current practice and future developments, in Proc. of 1st European
Conf. on Earthquake Engineering and Seismology, Geneva, Switzerland, 38 September 2006, 17.
Amos, C. B., D. W. Burbank, D. C. Nobes, and S. A. L. Read (2007).
Geomorphic constraints on listric thrust faulting: Implications for active deformation in the Mackenzie basin, South Island, New Zealand,
J. Geophys. Res. 112, no. B03S11, 24 pp, doi 10.1029/2006JB004291.
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, 17031729.
Berryman, K. R., T. H. Webb, N. Hill, D. Rhoades, and M. W. Stirling
(2001). Seismic loads on damsWaitaki system. Development of fault
scaling relationships and their application to fault source characterisation, GNS Client Report 2001/129, 66 pp.
Beyer, K., and J. J. Bommer (2007). Selection and scaling of real accelerograms for bi-directional loading: A review of current practice and
code provisions, J. Earthq. Eng. 11, 1345.
Bommer, J. J., and F. Scherbaum (2008). The use and misuse of logic trees in
probabilistic seismic hazard analysis, Earthq. Spectra 24, 9971009.
Bommer, J. J., F. Scherbaum, H. Bungum, F. Cotton, F. Sabetta, and
N. Abrahamson (2005). On the use of logic trees for ground-motion
prediction equations in seismic hazard assessment, Bull. Seismol. Soc.
Am. 95, 377389.
Boore, D. M., and G. M. Atkinson (2008). Ground-motion prediction
equations 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, 99138.
Bradley, B. A. (2009). Seismic hazard epistemic uncertainty in the
San Francisco Bay area and its role in performance-based assessment,
Earthq. Spectra 25, 733753.
Bradley, B. A. (2010). NZ-Specific Pseudo-Spectral Acceleration Ground
Motion Prediction Equations Based on Foreign Models, Department
of Civil and Natural Resources Engineering, University of Canterbury,
Christchurch, New Zealand, 324 pp.
Bradley, B. A., M. W. Stirling, G. H. McVerry, and M. Gerstenberger (2011).
Consideration and propagation of epistemic uncertainties in
New Zealand probabilistic seismic hazard analysis, New Zealand
Earthquake Commission EQC Report 10/593, 61 pp.
Brune, J. N. (1968). Seismic moment, seismicity and rate of slip along major
fault zones, J. Geophys. Res. 73, 777784.
Chiou, B., R. Youngs, N. Abrahamson, and K. Addo (2010). Ground-motion
attenuation model for small-to-moderate shallow crustal earthquakes
in California and its implications on regionalization of ground-motion
prediction models, Earthq. Spectra 26, 907926.
Der Kiureghian, A., and O. Ditlevsen (2008). Aleatory or epistemic? Does it
matter?, Struct. Saf. 31, 105112.
Field, E. H. (2007). A summary of previous working groups on California
earthquake probabilities, Bull. Seismol. Soc. Am. 97, 10331053.
Field, E. H., T. H. Jordan, and C. A. Cornell (2003). OpenSHA: A developing community modelling environment for seismic hazard analysis,
Seismol. Res. Lett. 74, 406419.
Gehl, P., L. F. Bonilla, and J. Douglas (2011). Accounting for site characterization uncertainties when developing ground-motion prediction
equations, Bull. Seismol. Soc. Am. 101,11011108.
Hanks, T. C., and W. H. Bakun (2002). A bilinear source-scaling model for
M-log A observations of continental earthquakes, Bull. Seismol. Soc.
Am. 92, 18411846.
1568
Hanks, T. C., and H. Kanamori (1979). A moment magnitude scale,
J. Geophys. Res. 84, 23482350.
Heron, D., R. van Dissen, and M. Sawa (1998). Late Quaternary movement
on the Ohariu fault, Tongue Point to MacKays Crossing, North Island,
New Zealand, New Zeal. J. Geol. Geophys. 41, 419439.
Howard, M., A. Nicol, J. Campbell, and J. R. Pettinga (2005). Holocene
paleoearthquakes on the strike-slip Porters Pass fault, Canterbury,
New Zealand, New Zeal. J. Geol. Geophys. 48, 5974.
Kulkarni, R. B., R. R. Youngs, and K. J. Coppersmith (1984). Assessment of
confidence intervals for results of seismic hazard analysis, in 8th
Proc. of World Conf. on Earthquake Engineering, San Francisco, California, 2128 July 1984, 263270.
Langridge, R. M., and K. R. Berryman (2005). Morphology and slip rate of
the Hurunui section of the Hope fault, South Island, New Zealand, New
Zeal. J. Geol. Geophys. 48, 4357.
Litchfield, N., R. Van Dissen, D. Heron, and D. Rhoades (2006). Constraints
on the timing of the three most recent surface rupture events and
recurrence interval for the Ohariu fault: Trenching results from
MacKays Crossing, Wellington, New Zealand, New Zeal. J. Geol.
Geophys. 49, 5761.
McVerry, G. H., J. X. Zhao, N. A. Abrahamson, and P. G. Somerville (2006).
New Zealand acceleration response spectrum attenuation relations for
crustal and subduction zone earthquakes, Bull. New Zeal. Natl. Soc.
Earthq. Eng. 39, 158.
Moss, R. E. S. (2011). Reduced sigma of ground-motion prediction
equations through uncertainty propagation, Bull. Seismol. Soc. Am.
101, 250257.
Pace, B., M. W. Stirling, N. J. Litchfield, and U. Rieser (2005). New active
fault data and seismic hazard estimates for west Otago, New Zealand,
New Zeal. J. Geol. Geophys. 48, 75 83.
Standards New Zealand (2004). Structural Design Actions, Part 5: Earthquake ActionsNew Zealand, Standards New Zealand, Wellington,
New Zealand, 82 pp.
Stirling, M. W., G. McVerry, M. Gerstenberger, N. Litchfield, R. Van Dissen,
K. Berryman, P. Barnes, J. Beavan, B. A. Bradley, K. Clark, K. Jacobs,
G. Lamarche, R. Langridge, A. Nicol, S. Nodder, J. Pettinga,
M. Reyners, D. Rhoades, W. Smith, P. Villamor, and L. Wallace
(2012). National seismic hazard model for New Zealand: 2010 update,
Bull. Seismol. Soc. Am., 102, 15141542.
Stirling, M. W., G. H. McVerry, and K. R. Berryman (2002). A new seismic
hazard model for New Zealand, Bull. Seismol. Soc. Am. 92, 1878
1903.
Strasser, F. O., M. C. Arango, and J. J. Bommer (2010). Scaling of the source
dimensions of interface and intraslab subduction-zone earthquakes
with moment magnitude, Seismol. Res. Lett. 81, 941950.
Van Dissen, R., P. Barnes, J. Beavan, J. Cousins, G. Dellow, C. FrancoisHolden, B. Fry, R. Langridge, N. Litchfield, T. Little, G. McVerry, D.
Ninis, D. Rhoades, R. Robinson, W. Saunders, P. Villamor, K. Wilson,
P. Barker, K. Berryman, R. Benites, H. Brackley, B.A. Bradley, R.
Carne, U. Cochran, M. Hemphill-Haley, A. King, G. Lamarche, N.
Palmer, N. Perrin, N. Pondard, M. Rattenbury, S. Read, S. Semmens,

E. Smith, W. Stephenson, L. Wallace, T. Webb, and J. Zhao (2010). Its
Our Fault: Better defining earthquake risk in Wellington, in Proc. of
11th International Association for Engineering Geology (IAEG) Congress, Auckland, New Zealand, 510 September 2010, 8 pp.
Van Dissen, R. J., and K. R. Berryman (1996). Surface rupture earthquakes
over the last 1000 years in the Wellington region, New Zealand,
and implications for ground shaking hazard, J. Geophys. Res. 101,
59996019.
Villamor, P., and K. R. Berryman (2006). Late Quaternary geometry and
kinematics of faults at the southern termination of the Taupo volcanic
zone, New Zealand, New Zeal. J. Geol. Geophys. 49, 121.
Villamor, P., K. R. Berryman, T. Webb, M. W. Stirling, P. McGinty,
G. Downes, J. Harris, and N. Litchfield (2001). Waikato seismic
loadsTask 2.1. Revision of seismic source characterisation, GNS
Client Report 2001/59, 115 pp.
Wallace, L. M., M. Reyners, U. Cochran, S. Bannister, P. M. Barnes,
K. Berryman, G. Downes, D. Eberhart-Phillips, A. Fagereng, S. Ellis,
A. Nicol, R. McCaffrey, R. J. Beavan, S. Henrys, R. Sutherland, D. H.
N. Barker, N. Litchfield, J. Townend, R. Robinson, R. Bell, K. Wilson,
and W. Power (2009). Characterizing the seismogenic zone of a major
plate boundary subduction thrust: Hikurangi margin, New Zealand,
Geochem. Geophys. Geosyst. 10, Q10006, 32 pp., doi 10.1029/
2009gc002610.
Working Group on California Earthquake Probabilities [WGCEP] (2003).
Earthquake probabilities in the San Francisco Bay region: 2002
2031, USGS Open-File Report 03-214, 235 pp.
Zachariasen, J., K. Berryman, R. Langridge, C. Prentice, M. Rymer,
M. Stirling, and P. Villamor (2006). Timing of late Holocene surface
rupture of the Wairau fault, Marlborough, New Zealand, New Zeal. J.
Geol. Geophys. 49, 159174.
Zhao, J. X., J. Zhang, A. Asano, Y. Ohno, T. Oouchi, T. Takahashi,
H. Ogawa, K. Irikura, H. K. Thio, P. G. Somerville, and Y. Fukushima
(2006). Attenuation relations of strong ground motion in Japan using
site classification based on predominant period, Bull. Seismol. Soc.
Am. 96, 898913.
Department of Civil and Natural Resources Engineering
University of Canterbury
Private Bag 4800
Christchurch, New Zealand
brendon.bradley@canterbury.ac.nz
(B.A.B.)
GNS Science
P.O. Box 30368
Lower Hutt 5044, New Zealand
(M.W.S., G.H.M., M.G.)
Manuscript received 11 September 2011

Bradley EpistemicUncertaintiesNzPsha BSSA 2012 PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bradley EpistemicUncertaintiesNzPsha BSSA 2012 PDF

Uploaded by

Copyright:

Available Formats

Bulletin of the Seismological Society of America, Vol. 102, No. 4, pp. 15541568, August 2012, doi: 10.

Consideration and Propagation of Epistemic Uncertainties

ground-motion amplitudes at a given distance from an

Consideration and Propagation of Epistemic Uncertainties in New Zealand PSHA

Adopted Epistemic Uncertainty Methodology

Epistemic Uncertainty in NZ GMPE

Inter-event Std Dev,

Active shallow crustal

Spectral acceleration, SA(T ) (g)

Spectral acceleration, SA(T ) (g)

Active shallow crustal

Intra-event Std Dev,

Spectral acceleration, SA(T ) (g)

Spectral acceleration, SA(T ) (g)

B. A. Bradley, M. W. Stirling, G. H. McVerry, and M. Gerstenberger

Deterministic Calculation of Source Magnitudes and Rates

Active Shallow Crustal Sources

Consideration and Propagation of Epistemic Uncertainties in New Zealand PSHA

For normal faults in volcanic and rift environments, the

where W is the fault width.

Median: Mw 4:441 0:846 log10 A;

No scaling relation is given for subduction slab events

For all other crustal faults, the relationship of Berryman

For subduction interface events, the relationship of

The total seismic moment rate for the fault can be

where is the shear rigidity of the fault interface; s_ is the

B. A. Bradley, M. W. Stirling, G. H. McVerry, and M. Gerstenberger

Figure 3. Monte Carlo procedure for fault-based epistemic

Consideration of Epistemic Uncertainties in the Fault

tivity study to assess the importance of each parameter

Length of fault plane, L

*COV, coefficient of variation; equal to the standardpdeviation

Consideration and Propagation of Epistemic Uncertainties in New Zealand PSHA

that use different definitions of magnitude, source-to-site

Site Class Compatibility of the Various GMPEs

*Standards New Zealand, 2004.

Z1:0 as was utilized in the C10 GMPE.

uncertainties in parameter values with the exception of fault

Uncertainty in Nationwide MagnitudeFrequency

Annual exceedance frequency, M w

The aforementioned methodology was implemented in

Considering fault-length uncertainty

Annual exceedance frequency, M w

Considering fault depth-of-rupture uncertainty

Considering slip-rate uncertainty

Annual exceedance frequency, M w

Annual exceedance frequency, M w

Considering M w -scaling uncertainty

B. A. Bradley, M. W. Stirling, G. H. McVerry, and M. Gerstenberger

Annual exceedance frequency, M w

large magnitudes, where the parameter uncertainties result in

Considering all fault uncertainties

Figure 6. Uncertainty in the nationwide fault-based seismicity

Consideration and Propagation of Epistemic Uncertainties in New Zealand PSHA

Seismic-Hazard Analysis Case-Study Sites

50-year exceedance probability

Figure 7 illustrates seismic-hazard curves for PGA and

Uncertainty in Seismic Hazard for Christchurch

Figure 7a illustrates the difference between the hazard using

Effect of crustal GMPE

Peak ground acceleration, PGA (g)

50-year exceedance probability