You are on page 1of 5

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

net/publication/278014945

Alternative Approaches to Modeling Epistemic Uncertainty in Ground Motions


in Probabilistic Seismic-Hazard Analysis

Article  in  Seismological Research Letters · November 2014


DOI: 10.1785/0220140120

CITATIONS READS

32 307

3 authors, including:

Gail M. Atkinson Julian Bommer


The University of Western Ontario Imperial College London
259 PUBLICATIONS   10,420 CITATIONS    206 PUBLICATIONS   11,050 CITATIONS   

SEE PROFILE SEE PROFILE

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

Mega Quakes: Cascading Earthquake Hazards and Compounding Risks View project

NGA-West2 View project

All content following this page was uploaded by Gail M. Atkinson on 12 June 2015.

The user has requested enhancement of the downloaded file.


SRL Early Edition

Alternative Approaches to Modeling reasonably accomplished. In this article, we comment on the


two main alternative approaches that have been used to model
Epistemic Uncertainty in Ground Motions in epistemic uncertainty in GMPEs. These two alternatives are
Probabilistic Seismic-Hazard Analysis (1) direct use of multiple GMPEs drawn from the technical
literature, with weights assigned to each GMPE based on the
Ground-motion prediction equations (GMPEs) describe the judgment of the analyst concerning their relative merits or
amplitude of shaking as a function of earthquake magnitude, applicability, and (2) use of GMPEs, data analysis, and
distance from the source, site conditions, and sometimes other judgment to define a representative suite of models to capture
variables. Amplitudes are predicted for peak ground accelera- the uncertainty, including one or more central (or backbone)
tion and velocity and for such measures of engineering interest models along with high and low alternatives. We argue that the
as the response of a damped oscillator of a specified natural development of a representative suite is a superior approach to
period. Selection of appropriate GMPEs is a critical step in building ground-motion characterization (GMC) logic trees, in
a probabilistic seismic-hazard analysis comparison with the more widely prac-
(PSHA), and therefore their selection is ticed use of multiple GMPEs drawn from
an issue of major practical significance. the literature.
For PSHA applications, we are interested The multiple-GMPE The multiple-GMPE method has
in more than just a best estimate of the approach has the advantage been used in standard PSHA practice
GMPE; equally important is the range of of expediency, in that it for many site-specific analyses and is also
alternative models used to characterize makes use of previous work used in national seismic hazard maps (e.g.,
the uncertainty in the GMPEs and and thus minimizes the need Montaldo et al., 2005; Petersen et al.,
thereby to express the limitations of our for model development. 2008) as well as for the Global Earthquake
current knowledge. Explicitly, we wish to However, it also has many Model (Stewart et al., 2014). The multi-
capture the center, the body, and the shortcomings. ple-GMPE approach has the advantage of
range of defensible technical judgments expediency, in that it makes use of pre-
regarding the GMPEs (Budnitz et al., vious work and thus minimizes the need
1997; U.S. Nuclear Regulatory Commis- for model development. However, it also has many shortcom-
sion, 2012). ings. The first of these is that none of the available models may
Two types of uncertainty are recognized in GMPEs and actually be very applicable to the target region or the actual site
modeled in a PSHA : (1) aleatory uncertainty, expressing random conditions, which means that not even the center of the dis-
variability of amplitudes about a median prediction equation, tribution of earthquake ground motions may be captured prop-
and (2) epistemic uncertainty, expressing uncertainty concern- erly. This is why it is often necessary to extensively adjust
ing the correct value of the median. Aleatory variability is GMPEs for specific conditions (e.g., Cotton et al., 2006;
handled relatively easily in a PSHA by integrating over the dis- Van Houtte et al., 2011; Ghofrani and Atkinson, 2014). More-
tribution of ground-motion amplitudes about the median over, the selected set of GMPEs may use different definitions
(Bommer and Abrahamson, 2006). Epistemic uncertainty is for explanatory and predicted variables, thereby necessitating
usually handled by considering alternative GMPEs in a logic- adjustments for compatibility (Bommer et al., 2005). It may
tree format (e.g., Bommer and Scherbaum, 2008; Bommer, even be necessary to develop new GMPEs for site-specific
2012). Construction of a logic tree involves populating the PSHA studies (Bommer, 2012) if the existing GMPEs are
branches with alternative models and then assigning weights not sufficient.
to the branches; together, the branches and their weights are in- For many smaller projects, it can be tempting to assign and
tended to represent the underlying continuous distribution of weight the GMPEs that make up the logic tree without suffi-
possible ground motions. cient analysis of ground-motion data (or lack thereof ) that
The logic tree must capture both the best estimates of might inform the exercise. Simply populating the logic-tree
what is known and the potential range of alternatives in light branches for median motions with a wide selection of existing
of what is currently not known. There has been considerable GMPEs is unlikely to correctly capture the range of epistemic
confusion, misunderstanding, and controversy in the technical uncertainty in the magnitude–distance ranges that drive the
community over what this really means and how it may be hazard analysis. A better judgment regarding appropriate

doi: 10.1785/0220140120 Seismological Research Letters Volume 85, Number 6 November/December 2014 1
SRL Early Edition
GMPEs may be made through analysis of ground-motion data (e.g., Toro et al., 1997; Douglas et al., 2013). With the hybrid
to explore the relative performance of different models (e.g., empirical approach (Campbell, 2003), multiple versions of the
Scherbaum et al., 2009; Atkinson and Adams, 2013; Kale backbone GMPE can be generated using multiple sets of model
and Akkar, 2013). Methodologies are also available to facilitate adjustments to represent various effects. The representative
less arbitrary choices in the assignment of branch weights suite approach has been implemented in major projects, includ-
(Runge et al., 2013). However, the relationship between ing the SSHAC Level 3 PSHA studies for the Hanford nuclear
weights on GMPEs and weights on ground-motion amplitudes site in Washington (Coppersmith et al., 2013), the BC Hydro
is unavoidably obscured in the process of assigning weights to PSHA for hydro-electric dams in British Columbia, and for a
models, rather than to the outcomes of models. nuclear site characterization project in South Africa (Bommer
Even if appropriate GMPEs are available, there is no reason et al., 2014). A partial use of the approach is encountered in the
to expect that the selected multiple equations will correctly U.S. national hazard maps (Petersen et al., 2008), in which the
represent the underlying epistemic uncertainty in median am- GMC logic tree for the western United States included addi-
plitudes over all important magnitudes and distances. For ex- tional branches on the median predictions of three selected
ample, ground motions from crustal earthquakes in California GMPEs to provide additional epistemic uncertainty. The na-
are now commonly estimated using the Next Generation tional seismic hazard maps in Canada have used the represen-
Attenuation (NGA)-West 2 GMPEs (see Douglas and Kramer, tative suite approach since ∼2000, including the GMPE model
2014, special issue), which were derived from a common suite for the 2015 maps (Atkinson and Adams, 2013). The
strong-motion database and which yield similar predictions common feature in all of these applications, which vary signifi-
for scenarios well represented in the database. All of the data cantly in detail, is that some (or all) of the branches of the
for earthquakes of magnitude greater than M 7.2 come from ground-motion logic tree are populated by scaled versions of
regions outside California (Alaska, Iran, Turkey, China, and a backbone GMPE.
Taiwan), with the exception of a single re- The choice between the repre-
cording from the 1952 Kern County event. sentative suite or the more traditional
Thus, the application of the alternative NGA multiple-GMPE approaches to episte-
GMPEs to large events in California may not In our view, there are no mic uncertainty in ground motions
be properly centered and may underestimate advantages to the traditional has been controversial and often
epistemic uncertainty. When exported to approach, other than that of hotly debated. In our view, there
other applications, there are often other allowing one to quickly and are no advantages to the traditional
problems, such as poor constraint of the effortlessly produce a logic approach, other than that of allowing
equations at large distances (important in
tree, albeit one of unknown one to quickly and effortlessly pro-
low-hazard regions) and poor constraint of
worth. duce a logic tree, albeit one of un-
moderate events at very short hypocentral known worth. Proponents of the
distances (important in induced-seismicity multiple-GMPE approach argue that
applications). Thus even if the overall approach is to be based use of multiple models with alternative functional forms is re-
on applicable and well-founded existing GMPEs, there is often quired to properly capture uncertainties in form as well as am-
a need to modify these GMPEs, or create additional branches, plitude. However, there is no guarantee that the use of multiple
to properly represent the center and range of defensible tech- equations will fully capture uncertainty in functional form. In
nical assessments. fact, if uncertainty in some aspect of the functional form (such
We conclude that in most cases the multiple-GMPE ap- as how amplitudes scale with magnitude at close distances) is a
proach will not meet the objective of describing the center, driving uncertainty, then one should examine this uncertainty
body, and range of technically defensible interpretations of explicitly—which is unlikely to happen by simply assigning
the data. We believe that this objective is better met using the multiple GMPEs. Ultimately, what matters to the hazard result
representative suite approach. The essence of this approach is is not the functional form so much as the dependence of the
to make adjustments to one or more central or backbone ground-motion amplitudes on magnitude, distance, and so on.
GMPEs as required, so that each model is transformed into a Advantages of the representative suite approach are that it
suite of alternative models that describe the epistemic uncer- facilitates explicit judgments regarding magnitude and distance
tainty explicitly and transparently. The explicit modeling of scaling and enables greater control over how the selected mod-
epistemic uncertainty is consistent with the original suggestions els will satisfy data constraints that are important to the
made in the Senior Seismic Hazard Analysis Committee project. Moreover, for projects of large geographic scale, it al-
(SSHAC) guidelines (Budnitz et al., 1997) and has already been lows control over how both the median GMPEs and their un-
used widely in practice (Bommer, 2012). certainty will behave across regions and event types. Thus,
The representative suite approach can be implemented in one can ensure that the epistemic uncertainty is larger in regions
a number of ways. For GMPEs developed with the stochastic with sparse data, for example, regardless of whether alternative
method, the alternative models required to capture the body published GMPEs fortuitously happen to be similar (e.g., Atkin-
and range of the distribution can be generated using suitably son and Adams, 2013). By contrast, the traditional approach can
sampled ranges of the stochastic source and path parameters never represent a wider range of uncertainty than that covered by

2 Seismological Research Letters Volume 85, Number 6 November/December 2014


SRL Early Edition
the GMPEs that happen to be available. Finally, the representative Bommer, J. J. (2012). Challenges of building logic trees for probabilistic
suite approach has significant practical utility, enabling a complex seismic hazard analysis, Earthq. Spectra 28, no. 4, 1723–1735.
Bommer, J. J., and N. A. Abrahamson (2006). Why do modern probabi-
problem to be represented by a minimum number of branches listic seismic hazard analyses lead to increased hazard estimates?
for hazard calculations, which is efficient and transparent. Bull. Seismol. Soc. Am. 96, no. 6, 1967–1977.
An additional advantage of the representative GMPE Bommer, J. J., and F. Scherbaum (2008). The use and misuse of logic
approach, when used with a single backbone GMPE, is that trees in probabilistic seismic hazard analysis, Earthq. Spectra 24,
997–1009.
the models on the branches of the GMC logic tree become
Bommer, J. J., K. J. Coppersmith, R. T. Coppersmith, K. L. Hanson,
mutually exclusive. Moreover, if they have been scaled suffi- A. Mangongolo, J. Neveling, E. M. Rathje, A. Rodriguez-Marek,
ciently to capture the full range of epistemic uncertainty, they F. Scherbaum, R. Shelembe, P. J. Stafford, and F. O. Strasser
are collectively exhaustive. In this case, the treatment of the (2014). A SSHAC Level 3 probabilistic seismic hazard analysis
weights on the branches as probabilities becomes uncontrover- for a new-build nuclear site in South Africa, Earthq. Spectra doi:
10.1193/060913EQS145M.
sial (Abrahamson and Bommer, 2005; McGuire et al., 2005; Bommer, J. J., F. Scherbaum, H. Bungum, F. Cotton, F. Sabetta, and
Musson, 2005). Finally, use of the representative suite approach N. A. Abrahamson (2005). On the use of logic trees for ground-
facilitates consideration of new ground-motion information as motion prediction equations in seismic hazard assessment, Bull.
it becomes available. New GMPEs can be compared with the Seismol. Soc. Am. 95, no. 2, 377–389.
existing distribution rather than requiring that the hazard be Budnitz, R. J., G. Apostolakis, D. M. Boore, L. S. Cluff, K. J. Copper-
smith, C. A. Cornell, and P. A. Morris (1997). Recommendations for
recalculated to evaluate their potential impact. This is very probabilistic seismic hazard analysis: Guidance on uncertainty and use
helpful in meeting the SSHAC objective of stability and lon- of experts, NUREG/CR-6372, two volumes, U.S. Nuclear Regulatory
gevity of seismic-hazard results. Commission, Washington, D.C.
Admittedly, there is a significant degree of judgment that Campbell, K. W. (2003). Prediction of strong ground motion using the
must be exercised in the development of the representative hybrid empirical method and its use in the development of ground-
motion (attenuation) relations in eastern North America, Bull. Seis-
suite regarding the central model(s) and mol. Soc. Am. 93, no. 3, 1012–1033.
its upper and lower branches, and this ex- Coppersmith, K. J., J. J. Bommer, R. W. Bryce, S.
erts significant influence on the hazard McDuffie, and G. A. Lisle (2013). An ap-
results. However, such subjective judg- We conclude that in most plication of the SSHAC Level 3 process to
ments are equally important when select- cases the multiple-GMPE the probabilistic seismic hazard analysis for
approach will not meet the nuclear facilities at the Hanford site, eastern
ing and weighting models for use with Washington, USA, Proc. 22nd International
the multiple-GMPE approach. Using objective of describing the Conference on Structural Mechanics in Reac-
published GMPEs may make the analysts center, body, and range of tor Technology (SMiRT22), San Francisco,
feel that they are avoiding subjective judg- technically defensible California, 18–23 August.
ments, but this is an illusion. And they interpretations of the data. Cotton, F., F. Scherbaum, J. J. Bommer, and H.
Bungum (2006). Criteria for selecting and
are implicitly making the judgment that adjusting ground-motion models for specific
median motions cannot be outside the target regions: Applications to central
range of motions predicted by their selected equations; this Europe and rock sites, J. Seismol. 10, 137–156.
is a critical judgment that is generally offered without justifi- Douglas, J., B. Edwards, V. Convertito, N. Sharma, A. Tramelli, D.
Kraaijpoel, B. Mena Cabrera, N. Maercklin, and C. Troise
cation. (2013). Predicting ground motion from induced earthquakes
In conclusion, both the multiple-GMPE and the represen- in geothermal areas, Bull. Seismol. Soc. Am. 103, no. 3,
tative suite approaches have been widely used in practice to 1875–1897.
characterize epistemic uncertainty in median ground-motion Douglas, J., and S. Kramer (2014). Special Issue, NGA-West2 Ground
Motion Prediction Equations, Earthq. Spectra, 30, no. 3, available at
amplitudes for PSHA applications. Whichever approach is http://earthquakespectra.org/page/NGA‑West2 (last accessed August
adopted, the analyst should (1) document and justify the 2014).
epistemic uncertainty in amplitudes that has been incorporated Ghofrani, H., and G. Atkinson (2014). Ground motion prediction equa-
as a function of magnitude and distance and (2) discuss how tions for interface earthquakes of M 7 to M 9 based on empirical
the resulting range of models addresses the available informa- data from Japan, Bull. Earthq. Eng., doi: 10.1007/s10518-013-
9533-5.
tion and its limitations. Our opinion is that the representative Kale, O., and S. Akkar (2013). A new procedure for selecting and ranking
suite approach is strongly preferred as the means to this end, ground-motion prediction equations (GMPEs): The Euclidian dis-
due to its flexibility, explicit handling of the key issues, and tance-based ranking (EDR) method, Bull. Seismol. Soc. Am. 103,
transparency. no. 2A, 1069–1084.
McGuire, R. K., C. A. Cornell, and G. R. Toro (2005). The case for the
mean hazard curve, Earthq. Spectra 21, no. 3, 879–886.
Montaldo, V., E. Faccioli, G. Zonno, A. Akinci, and L. Malagnini (2005).
REFERENCES Treatment of ground-motion prediction relationship for the
reference seismic hazard map of Italy, J. Seismol. 9, 295–316.
Abrahamson, N. A., and J. J. Bommer (2005). Probability and uncer- Musson, R. M. W. (2005). Against fractiles, Earthq. Spectra 21, no. 3,
tainty in seismic hazard analysis, Earthq. Spectra 21, no. 2, 603–607. 887–891.
Atkinson, G., and J. Adams (2013). Ground motion prediction equations Petersen, M., A. Frankel, S. Harmsen, C. Mueller, K. Haller, R. Wheeler,
for application to the 2015 national seismic hazard maps of Canada, R. Wesson, Y. Zeng, O. Boyd, D. Perkins, N. Luco, E. Field, C.
Can. J. Civil Eng. 40, 988–998.

Seismological Research Letters Volume 85, Number 6 November/December 2014 3


SRL Early Edition
Wills, and K. Rukstales (2008). Documentation for the 2008 update ments factors for GMPEs, Bull. Seismol. Soc. Am. 101, no. 6,
of the United States National Seismic Hazard Maps, U.S. Geol. 2926–2941.
Surv. Open-File Rept. 2008-1128, 61 pp.
Runge, A. K., F. Scherbaum, A. Curtis, and C. Riggelsen (2013). An Gail M. Atkinson
interactive tool for the elicitation of objective probabilities in
probabilistic seismic-hazard analysis, Bull. Seismol. Soc. Am. 103, University of Western Ontario
no. 5, 2862–2874. Richmond Road
Scherbaum, F., E. Delavaud, and C. Riggelsen (2009). Model selection in London, Ontario, Canada N6A 5B7
seismic hazard analysis: An information-theoretic perspective, Bull. gmatkinson@aol.com
Seismol. Soc. Am. 99, no. 6, 3234–3247.
Stewart, J., J. Douglas, M. Javanbarg, Y. Bozorgnia, N. A. Abrahamson, D.
M. Boore, K. W. Campbell, E. Delavaud, M. Erdik, and P. J. Stafford Julian J. Bommer
(2014). Selection of ground motion prediction equations for the Civil & Engineering Dept.
Global Earthquake Model, Earthq. Spectra doi: 10.1193/ Imperial College London
013013EQS017M. South Kensington Campus
Toro, G. R., N. A. Abrahamson, and J. F. Schneider (1997). Models of London UK SW7 2AZ
strong ground motions from earthquakes in central and eastern
North America: Best estimates and uncertainties, Seismol. Res. Lett.
68, no. 1, 41–57. Norman A. Abrahamson
U.S. Nuclear Regulatory Commission (2012). Practical Implementation Dept. Civil & Environmental Engineering
Guidelines for SSHAC Level 3 and 4 Hazard Studies, NUREG- University of California Berkeley
2117, Rev. 1, April, U.S. Nuclear Regulatory Commission, Washing- Berkeley, California 94720-1710 U.S.A.
ton, D.C.
Van Houtte, C., S. Drouet, and F. Cotton (2011). Analysis of
the origins of κ (kappa) to compute hard rock to rock adjust- Published Online 1 October 2014

4 Seismological Research Letters Volume 85, Number 6 November/December 2014

View publication stats

You might also like