Model Selection in Seismic Hazard Analysis: An Information-Theoretic Perspective

Scherbaum, Frank; Delavaud, Élise; Riggelsen, Carsten

doi:10.1785/0120080347

Cited by 299 publications

(169 citation statements)

References 37 publications

Supporting

Mentioning

162

Contrasting

Order By: Relevance

“…Such equations often have an empirical nature and are developed based on vast databases of observed values of ground motion parameters (Boore and Joyner 1982). The selection of the most appropriate GMPE is not a trivial task and some guidance and criteria for choosing the most appropriate GMPE for the application in a PSHA for a particular site can be found in Scherbaum et al (2009) and Arroyo et al (2014). A comprehensive list of GMPEs developed during the period 1964-2010 is presented in Douglas (2011).…”

Section: Methodsmentioning

confidence: 99%

Effect of alternative distributions of ground motion variability on results of probabilistic seismic hazard analysis

Pavlenko

2015

Nat Hazards

View full text Add to dashboard Cite

Probabilistic seismic hazard analysis (PSHA) is a regularly applied practice that precedes the construction of important engineering structures. The Cornell-McGuire procedure is the most frequently applied method of PSHA. This paper examines the fundamental assumption of the Cornell-McGuire procedure for PSHA, namely, the log-normal distribution of the residuals of the ground motion parameters. Although the assumption of log-normality is standard, it has not been rigorously tested. Moreover, the application of the unbounded log-normal distribution for the calculation of the hazard curves results in non-zero probabilities of the exceedance of physically unrealistic amplitudes of ground motion parameters. In this study, the distribution of the residuals of the logarithm of peak ground acceleration is investigated, using the database of the Strong-motion Seismograph Networks of Japan and the ground motion prediction equation of Zhao and co-authors. The distribution of residuals is modelled by a number of probability distributions, and the one parametric law that approximates the distribution most precisely is chosen by the statistical criteria. The results of the analysis show that the most accurate approximation is achieved with the generalized extreme value distribution for a central part of a distribution and the generalized Pareto distribution for its upper tail. The effect of replacing a log-normal distribution in the main equation of the Cornell-McGuire method is demonstrated by the calculation of hazard curves for a simple hypothetical example. These hazard curves differ significantly from one another, especially at low annual exceedance probabilities.

show abstract

Section: Methodsmentioning

confidence: 99%

Effect of alternative distributions of ground motion variability on results of probabilistic seismic hazard analysis

Pavlenko

2015

Nat Hazards

View full text Add to dashboard Cite

show abstract

“…Starting from a set of viable candidates (e.g., Bommer et al, 2010), the selection of models to populate the branches of the logic tree (e.g., Bommer and Scherbaum, 2008;Bommer, 2012) is often supported by measuring the goodness-of-fit with respect to selected data sets (e.g., Delavaud et al, 2012). In particular, Scherbaum et al (2009) introduced an information-theoretic perspective for ranking a set of GMPEs with respect to a set of observations. To perform a relative ranking, Scherbaum et al (2009) proposed using the difference between the base-2 average log likelihood (logLH) computed for the selected models considering a validation data set that is independent of the calibration one.…”

Section: Introductionmentioning

confidence: 99%

“…In particular, Scherbaum et al (2009) introduced an information-theoretic perspective for ranking a set of GMPEs with respect to a set of observations. To perform a relative ranking, Scherbaum et al (2009) proposed using the difference between the base-2 average log likelihood (logLH) computed for the selected models considering a validation data set that is independent of the calibration one. Recently, Roselli et al (2016) proposed a weighting scheme based on the Bayesian information criterion (BIC, Schwarz, 1978) to build an ensemble model from a set of GMPEs.…”

Section: Introductionmentioning

confidence: 99%

“…Explanatory modeling (Shmueli, 2010) is often preferred to data-driven predictive approaches, and the analysis of the residual distribution is generally used to support the model qualification. Following previous studies (Kuehn et al, 2009;Scherbaum et al, 2009), this work aims to again stress the importance of validation for assessing the predictive power of GMPEs and for avoiding, or at least limiting, data overfitting. Considering the strong-motion data and models recently analyzed by Roselli et al (2016), I exemplify the application of standard validation approaches based on predictive metrics (Akaike and Bayesian information criteria), resample techniques (cross validation and bootstrap), and validation against new data.…”

mentioning

confidence: 99%

See 1 more Smart Citation

The Predictive Power of Ground‐Motion Prediction Equations

Bindi

2017

Bulletin of the Seismological Society of America

View full text Add to dashboard Cite

Although model calibration should be always performed alongside model validation, this is rarely the case when deriving new ground-motion prediction equations (GMPEs). Explanatory modeling (Shmueli, 2010) is often preferred to data-driven predictive approaches, and the analysis of the residual distribution is generally used to support the model qualification. Following previous studies (Kuehn et al., 2009;Scherbaum et al., 2009), this work aims to again stress the importance of validation for assessing the predictive power of GMPEs and for avoiding, or at least limiting, data overfitting. Considering the strong-motion data and models recently analyzed by Roselli et al. (2016), I exemplify the application of standard validation approaches based on predictive metrics (Akaike and Bayesian information criteria), resample techniques (cross validation and bootstrap), and validation against new data. The results confirm that GMPEs overfitting the calibration data have a limited predictive power, whereas, regarding validation against new data, the selection of the validation data set should take into account possible regional effects in the ground motion.

show abstract

“…"CU= OQm}wQ u}=Q@=v@ "CU= QwmPt VwQ |=yha[ R= VQ}PB |xv=DU; u}}aD , qFt x}rw= [7] "CU= xOt; O}OB xOWQmP |=yha[ Q@ x@re |= Q@ LLH |D=aq]=-|Q_v R= uDiQo s=yr= =@ w LLH w LH |}=tvCUQO Q@ |vD@t VwQ wO 'u}= Q@ xwqa |UO}rk= |xrY=i Q@ |vD@t |WwQ u= QoWywSB '=yxOv=t}k=@ p}rLD |=yVwQ Q}=U h= QLv= |Q}oxR=Ov= w u}tR |=yCmQL C}a]k sOa x@ xHwD =@ |Ov@x@DQ |= Q@ = Q |@=} RQ= CyH Q=@ u}rw= |= Q@ xm Ov=xOQm |iQat |v}tND w |D=Oy=Wt |=yxO=O u}@ R= |}=yCW=ov?=DW T=U= Q@ xv=}tQw=N w =B wQ= x@ \w@Qt Q}N= |oOvy=m \@=wQ Q}N= |=yp=U QO u}vJty [8] "CU= xOW xO=iDU= u=DUm=B w uOQ= 'R=kik 'u= Q}= 'x}mQD |wQ Q@ |oOvy=m \@=wQ |Ov@x@DQ w ?=NDv= |xv}tR QO |at=H Cq=kt u= QoWywSB ?NDvt |v=yH w |rLt |oOvy=m \@=wQ T=U= Q@ XNWt |xO=O xawtHt l} xawtHt =yu; |xty QO xm Ov=xOQm x = Q= uQOt w l}Uqm |=yVwQ |Q}oQ=mx@ =@ 'p=Lu}= =@ …”

mentioning

confidence: 99%