Direct Probabilistic Seismic Analysis Dissertation

DIRECT PROBABILISTIC SEISMIC ANAYSIS:
IMPLEMENTING NON-LINEAR DYNAMIC ASSESSMENTS
A DISSERTATION
SUBMITTED TO THE DEPARMENT OF CIVIL AND ENVIRONMENTAL
ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Fatemeh Jalayer
March 2003
I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.
____________________________________
C. Allin Cornell
(Principal Advisor)
Philosophy.
____________________________________
Helmut Krawinkler
Philosophy.
____________________________________
Charles Menun
Philosophy.
____________________________________
Eduardo Miranda
Approved for the University Committee on Graduate Studies:
____________________________________
Abstract
A formal probabilistic framework for seismic assessment of a structural system can be built
around the expression for the probability of exceeding a limit state capacity, a measure of the
reliability of system under seismic excitations. Common probabilistic tools are implemented in
order to derive a simplified closed-form expression for the probability of exceeding a limit state
capacity. This closed-from expression is particularly useful for seismic assessment and design of
structures, taking into account the uncertainty in the generic variables, structural demand and
capacity as well as the uncertainty in seismic excitations.
This framework implements non linear dynamic analysis procedures in order to estimate
variability in the response of the structure (demand) to seismic excitations. Alternative methods
for designing a program of nonlinear analyses and for applying the results of dynamic analysis,
particularly as it relates to displacement-based demand and capacity estimation, are
discussed. These alternative methods are presented through a comprehensive case study of the
Holiday Inn Hotel in Van Nuys, CA. This structure represents an older reinforced concrete
structure with degrading behavior in nonlinear range.
The onset of global dynamic instability in the structure is used to define the system capacity in
this study. The probabilistic model describing the structural demand in the vicinity of system
capacity is modified in order to explicitly account for the large displacement demands particular
to a system close to the onset of global instability. This leads to an alternative presentation of the
probabilistic framework in the range of global instability in the structure.
Ground motion record selection is potentially significant in implementing a program of nonlinear

dynamic analyses. However, it is demonstrated that even under the most extreme cases, namely
structures with very short and very long first-mode periods, the structural response is
conditionally independent of the ground motion characteristics such as magnitude and source to
site distance for a given seismic intensity level. This conclusion may justify random record
selection for the sites soil condition and considering the relative position of the site with respect
to the major faults around it.
Acknowledgements
The same as most of the things I do, this text is written in the very last minute!! And the fact that
my advisor, Professor Allin Cornell is not going to revise it, makes my agitated hands even more
shaky. I hope that I manage to express my thanks to some of the people who helped me to get to
this very page of my thesis and also of my life.
I would like to express my deepest gratitude to Professor Cornell for his enthusiasm, generosity,
constant support, and his patience with me. I feel truly honored to have the opportunity of being
Professor Cornells student, a unique experience of being guided by a master. And surely, I
cannot miss acknowledging his legendary notes. Yes, I did finally learn how to read them!
I would like to thank Professor Helmut Krawinkler, Professor Charles Menun, and Professor
Eduardo Miranda for taking the time to read this thesis and for providing me with very helpful
comments. I would like to thank Professor Steven Winterstein for being the very enthusiastic
teacher of CE 203, more formally known as, Probability and Statistics for Civil Engineers.
My many thanks to Dr. Mohsen Rahnama, my advisor for the masters degree in Sharif Univeristy
of Technology, Iran. I am very grateful for his amazing support through all these years. My very
special thank you to Mr. Ali A. Taheri, The Big Boss!!. I have had my only real work
experience at his design office, in Abbas-Abad Avenue, in Tehran, when I was a junior. It is not
an exaggeration if I say that he has been one of the first people to ignite my enthusiasm for
structural engineering and also my desire for experiencing something different, something like
pursuing my studies in a country very far away Thank you. Another very special thank you to
Ms. Eshghi my 9th grade algebra teacher, who believed in me in a time when I had many doubts
about my ability to learn mathematics.
Among my probabilistic colleagues, I would like to thank Dimitris Vamvatsikos the Greek
man for his many helpful comments through the years, and also for providing me with the acid
free paper on which these words are written! I would like to thank Tina Kashef, Nico Luco, Paolo
Bazzurro, Nilesh Shome, Jorje Carballo, and Leroy Fitzwater. I would also like to thank my
office-mates, Jack Baker and Gee Liek Yeo. Be prepared Gee Liek, now its your turn, you are
the senior
Many thanks to my dear friends and colleagues in Blume Center, Babak Alavi, Pablo Sanz, Luis
Ibarra (Luis, believe me, I am graduating after all!), Ricardo Medina, Farzin Zareian, Keith
Porter, Shahram Taghavi, Hesaam Aslani, Arash Altoontash, Pooya Sarabandi, Christoph Adam,
Toko Hitaka, Jerry Lynch, Paul Cordova, Kerri Tokoro, Maziar Mottahari, Amit Kanvinde, Rohit
Kaul, Konstantin Savov, Medji Sama, Jose Andrade and Jorje Ruiz-Garcia. A special thank you
to Paolo Giovenale for reading my chapters so fast, even before they were properly written!
Paolo, you still have to read Maggie (also known as Chapter 6), and it is quite long
And how can I not thank our wonderful admin, Racquel Hagen for her charismatic presence and
enthusiam. Thank you dear Blumesters for making the second floor like a second home for me.
I wish I could still linger some more in this wonderful place, but alas, everything has an end!
My many thanks to my friends, Azita Emami, Yasi Mostofi, Dara Ghahremani, Kaveh Hosseini,
and Maryam Fazel. I guess we all made it, it looked so far-fetched one day. Azita, thank you for
your wonderful support and also your wonderful laptop on which I gave my most important
presentations. And, Dara, I will certainly remember our many discussions about our careers and
our dilemmas about the future. My special thanks to Nogol Rashidi and Arash Hassibi my
adopted parents!! for their amazing support in the times that I needed it the most, thank you.
My many thanks to my wonderful room-mates, Ericka Beckman and Sanja Perovic. Ericka, your
enthusiam for your work has been truly inspiring for me, and Sanja, last autumn in your
company, I had one of the most wonderful and productive autumn quarters at Stanford! Also my
thanks to Maurizio Franzini for being such a good friend and making the best brunches in the
world ...
And it is time for me to acknowledge my parents and my brother Kaveh for being unbelievably
supportive. My uttermost thanks to you, it is not possible for me to fully express my gratitude to
you.
Fatemeh Jalayer
Blume Center, Stanford, March 16th, 2003
Contents
Abstract vi
Acknowledgements v
1 Introduction 1
1.1 General overview 1
1.2 This thesis in the context of probabilistic PBD 2
1.3 The organization of this thesis 5
2 A Technical Framework for Probability-based (DCFD) Seismic Formats 8

2.1 Abstract 8
2.2 Introduction 8
2.3 Sources of uncertainty in engineering problems 9
2.4 Document map 10
2.5 Foundation development 10
2.5.1 Structural limit states 11
2.5.2 Structural demand variable (state variable) 11
2.5.3 Structural capacity variable (Limit state variable) 11
2.5.4 Limit state probability 12
2.5.5 General solution strategy 12
2.5.6 Ground motion intensity measure 14
2.5.7 Randomness: The only source of variability 14
2.5.7.1 Spectral acceleration hazard 15
2.5.7.2 Median relationship between S a and drift demand 16
2.5.7.3 Annual frequency of exceeding demand Drift hazard 18
2.5.7.3.1 Numerical example 25
2.5.7.4 Annual frequency of exceeding a limit state 28
2.5.7.4.1 Numerical example: 32
2.5.7.5 Limit state frequency, the IM-based approach 34
2.5.8 Randomness and uncertainty as the sources of variability 36
2.5.8.1 Spectral acceleration hazard: 37
2.5.8.2 Probability of exceeding a drift demand value Drift hazard 40
2.5.8.3 Annual probability of exceeding a limit state 48
2.5.8.4 Limit state frequency: the IM-based approach 54
2.5 Summary 56
2.6 Acknowledgements 56
3 Probability-based Design (DCFD) Seismic Formats 57

3.1 Introduction 57
3.2 Randomness: The only source of uncertainty 58
3.2.1 Spectral acceleration corresponding to a demand equal to d 59
P0
3.2.2. Spectral acceleration s a for a hazard level equal to P0 60
3.2.3 DCFD Format 61

3.2.3.1 Displacement-based demand 64
3.2.3.2 Displacement-based capacity 65
3.2.3.3 Demand factor 65
3.2.3.4 Capacity factor 66
3.2.3.5 Factored demand and demand hazard 66
3.2.3.6 General Form for the DCFD design format 67
3.2.3.7 Numerical Example 68
3.2.4 Fragility/Hazard Format An IM-based probabilistic format 73
3.2.4.1 Hazard curves 75
3.2.4.2 Fragility curves 76
3.2.4.3 The IM-based limit state frequency and fragility/hazard 77
3.3 Randomness and uncertainty, the sources of uncertainty 79
3.3.1 A confidence-based DCFD format 81
3.4 Summary and conclusions 87
4 Alternative Nonlinear Demand Estimation Methods for Probability-
Based Seismic Assessment 89
4.1 Abstract 89
4.2 Keywords 90
4.3 Introduction 90
4.4 A Comprehensive assessment example 92
4.5 Record selection 94
4.6 Probabilistic framework for design and assessment of structures 95
4.7 Drift hazard - the annual frequency of exceeding drift demand 96
4.8 Demand and capacity factored design format 97
4.9 Site-specific hazard curve: estimation of k value 100
4.10 Evaluation of factored demand by numerical integration 100
4.11 Narrow-range methods 103
4.11.1 Single-stripe analysis 103
4.11.2 Cloud analysis 107
4.11.3 Cloud analysis scaled 111
4.11.4 Narrow-range methods for a lower spectral acceleration 113
4.12 Wide-range methods 115
4.12.1 Multiple-stripe analysis (MSA) 116
4.12.2 Incremental dynamic analysis 118
4.12.2.1 Capacity estimation using the results of IDA 121
4.12.3 Evaluation of factored capacity by numerical integration 123
4.13 Number of Records 124
4.14 Summary and Conclusions 126
5 Probability-Based Seismic Assessment: Non-linear Dynamic Displacement

Estimation in the Region of Global Instability 130
5.1 Abstract 130
5.2 Introduction 131
5.3 Organization of this chapter 132
5.4 Global dynamic instability capacity (collapse) estimation 134
5.5 Spectral acceleration hazard 135
5.6 Probability distribution of demand for a given spectral acceleration 135
5.6.1 A Three-parameter distribution and its percentiles 137
5.6.2 Methods incorporating the three-parameter distribution 139
5.6.2.1 Fully parametric using power-law functions 139
5.6.2.2 Fully parametric using the IDA procedure 141
5.6.2.3 Semi-parametric using multiple-stripe analysis 144
5.6.2.4 Pseudo-parametric method 146
5.6.2.5 Comparison of the fully-parametric IDA method to
semi-parametric with MSA 146
5.6.3 Non-parametric methods 147
5.6.3.1 Non-parametric considering the collapse cases 147
5.6.3.2 Non-parametric not considering the collapse cases 148
5.6.3.3 Comparison of the non-parametric methods
considering and not considering the collapse cases 148
5.6.3.4 Comparison of the non-parametric method not
considering collapse and semi-parametric method with MSA 149
5.6.3.5 Discussion 149
5.7 Drift hazard 151
5.7.1 Estimating drift hazard by disaggregating into collapse and
non-collapse parts 151
5.7.1.1 Drift hazard estimation in limiting cases 151
5.7.2 Non-parametric calculation of drift hazard 152
5.7.3 Semi-parametric calculation of drift hazard 153
5.7.4 Fully parametric (closed-form) evaluation of drift hazard 154
5.7.5 Discussion 157
5.8 Factored Demand 157
5.8.1 Non-parametric using the non-parametric drift hazard 158
5.8.2 Fully-parametric using IDA procedure with simple DCFD 158
5.8.3 Factored demand: Semi-parametric using parametric IDA curves 160
5.8.4 Factored demand: Semi-parametric using single and
double-stripe analysis 161
5.8.5 Discussion 166
5.9 Factored capacity 167
5.10 Practical Applications 167
5.11 Summary and Conclusions 168
6 Non-linear Response Dependence on Ground Motion Characteristics with
Implications for Ground Motion Record Selection 171
6.1 Abstract 171
6.2 Keywords 173
6.3 Introduction 173
6.4 Records used in this chapter 176
6.5 Ground motion characteristics 177
6.6 The spectral acceleration and the spectral shape factor 177
6.7 Maximum inter-story drift angle: The displacement-based response 178
6.8 Residual-residual plot 179
6.9 Establishing sufficiency based on the un-scaled set of records 181
6.9.1 System with long first-mode period: LA 20 story Results 182
6.9.2 System with short period: Generic SDOF structure 188
6.10 Studying the dependence of the displacement-based response given
S a (T FM ) on the spectral shape factor at period T, R (T , T FM ) 193
6.10.1 System with large first-mode period: LA 20 story Results 194
6.10.2 SDOF systems with short period 196
6.11 Discussion: Correlation between spectral-shape factor and magnitude 198
6.12 Discussion: Does magnitude matter? 202
6.13 Observations for the SDOF system with a different yield strength 203
6.14 Observations for LA 20 story building when the ground motion
records are scaled 206
6.15 Sufficiency and ground motion record selection 210
6.16 Investigating the sufficiency of PGA as the intensity measure with
respect to magnitude 214
6.17 Adjusting for possible dependencies using weighted regression 216
6.18.1 Limitations 223
6.19 Acknowledgments 224
7 Conclusions 225
7.1.1 Chapter 2 225
7.1.2 Chapter 3 226
7.1.3 Chapter 4 227
7.1.4 Chapter 5 228
7.1.5 Chapter 6 229
7.2 Limitations and future work 230
Appendix A 233
Appendix B 235
Appendix C 236
Bibliography 238
List of Tables
Table 5-1. Parameter estimates for s a 0 and c 140
Table 6-1- Modal periods for LA 20 M1+ model structure 183
Table 6-2 - The results of the regression of response on S a for different yield strengths 205
Table 6-3 - The importance of ground motion characteristics, M , R rup , and , and two
different spectral shape factors as the second regression variable, SDOF system
M P = 10000 (moderate ductilities) 206
Table 6-4 - The importance of ground motion characteristics, M , R rup , and , and two
different spectral shape factors as the second regression variable, SDOF system
M P = 15000 (low ductilities) 206
Table 6-5 The results of the regression of response on S a (TFM ) for the case where the
ground motion accelerations are scaled by a factor 3, LA 20 story building 208
Table 6-6 - The importance of ground motion characteristics and spectral shape factor
as the second regression variable, LA 20 building S .F . = 3.0 208
Table 6-7 The results of the linear multi-variable regression of ln max on ln PGA
and M for the case where the ground motion accelerations are scaled by a factor 3,
LA 20 story building 220
List of Figures
Figure 2-1 - A typical hazard curve for spectral acceleration corresponding to a damping
ratio equal to 5% and a structural fundamental period of 1.8 seconds. 15
Figure 2-2 - A set of spectral acceleration and demand data pairs, showing the model fit to
these data points. 17
Figure 2-3 - Basic elements for drift hazard evaluation, H S s (x) , and the distribution of D
given Sa characterized by D (x) and D|S a . 20
Figure 2-4 - Spectral acceleration corresponding to the demand value, d 25
Figure 2-5 - Hazard curve for spectral acceleration values corresponding to an oscillator
with a period equal 1.0 and damping ratio of 2%. 26
Figure 2-6 - Hazard curve derived for maximum inter-story drift values. 28
Figure 2-7 - Basic elements for limit state probability evaluation, H S s (x) , distribution of
drift variable D given Sa characterized by D (x) and D|S a , distribution
of capacity variable C characterized by C and C . 29
Figure 2-8 - The spectral acceleration hazard curve. The hazard value for a spectral
acceleration equal to, 2.15, is shown on the figure. 33
Figure 2-9 - 16th, median and 84th percentile spectral acceleration hazard corresponding
to a damping ratio equal to 5% and a structural fundamental period of 1.8
seconds. 39
Figure 2-10 - Basic components for the evaluation of drift hazard with uncertainty
~
in the estimation of spectral acceleration hazard, H S a ( x) . 40
Figure 2-11 - Spectral acceleration corresponding to the inter-story drift ratio value d for a
given value of deviation of median drift. 43
Figure 2-12 - Basic components for the evaluation of drift hazard with uncertainty in
the estimation of spectral acceleration hazard and median drift. 46
Figure 2-13 - Basic components for the derivation of the limit state probability when
there is uncertainty in the estimation of the spectral acceleration hazard, median
drift demand, and median drift capacity. 52
Figure 2-14 - The median estimate for the spectral acceleration hazard curve. The
hazard value for a spectral acceleration equal to 2.15 is shown on the figure. 54
Figure 3-1 - Spectral acceleration corresponding to a displacement-based demand equal

to d. 60
Figure 3-2 - Spectral acceleration for a hazard level equal to P0 . 61
Figure 3-3 - Graphical presentation of median demand for a given spectral acceleration. 64
Figure 3-4 - Graphical presentation of median drift capacity. 65
Figure 3-5 - The hazard curve for S a (T = 1, = 2%) . 70
Figure 3-6 - Spectral acceleration plotted versus maximum inter-story drift angle,
and fitted power-law relation (a line on the two-way logarithmic paper) 71
Figure 3-7 - A schematic plot of the hazard and fragility curves. The main
parameters of hazard/fragility format are also shown on the figure. 77
Figure 4-1 - A 7-story frame structure in LA: The modeling characteristics of a transverse
Frame. 93
Figure 4-2 - Hysteretic response of a beam-column to cyclic loading. 93
Figure 4-3 - The static pushover curve for the structural model. 94
Figure 4-4 - Site-specific hazard curve, Van Nuys CA. Approximating the hazard curve
with a line in the region of interest. 101
Figure 4-5 The drift hazard curve derived by numerical integration. 102
Figure 4-6 - The Spectral Acceleration Corresponding to a given acceptable probability

Level. 104
Figure 4-7 - Stripe response obtained by applying the selection of ground motion records
scaled to the spectral acceleration of the stripe. 104
Figure 4-8 - Estimated statistical parameters for the stripe response 105
Figure 4-9 - Estimation of factored demand using two stripes 107
Figure 4-10 - The Cloud response of the structure subjected to a set of ground motion
records(plotted in the Logarithmic Scale). 108
Figure 4-11 - The statistical properties of the cloud response. The parameter estimates
for the median, and dispersion measure are shown on the figure. 110
Figure 4-12 - The scaled cloud response of the structure subjected to the selection of
records, scaled by a factor of two. 111
Figure 4-13 The statistical properties of the scaled cloud response. 112
Figure 4-14 - Stripe response obtained by applying the selection of ground motion
records scaled to the spectral acceleration of the stripe,, to the model
Structure. 114
Figure 4-15 Multiple-stripe analysis. 116
Figure 4-16 - Evaluation of factored demand by multiple-stripe analysis. 118
Figure 4-17 - Factored demand predicted by multiple-stripe analysis (MSA) for multiple
limit state frequency values, Po , and the drift hazard curve obtained by
numerical integration. 119
Figure 4-18- Individual IDA Curves 119
Figure 4-19 - Results of the Incremental Dynamic Analysis when the model structure is
subjected to the selection of ground motion records. 120
Figure 4-20 - Evaluation of the factored demand using the results of incremental
dynamic analysis. 121
Figure 4-21 - Collapse Points marked on the IDA curves. 122
Figure 4-22 - Evaluation of the factored capacity using the results of incremental
dynamic analysis (IDA) 123
Figure 4-23-a) Median and standard error of the estimates of the median response as a
function of the sample size (number of records). 125
Figure 4-23-b) The (fractional) standard error of the estimates of median response as a
function of the sample size (number of records). 126
Figure 5-1 - Estimation of collapse points from IDA curves. a) IDA curves with the
collapse points marked on them. b) Collapse Points plotted Separately. 134
Figure 5-2 - Conditional probability of non-collapse for a given spectral acceleration

represented both empirically and analytically. 140
Figure 5-3 - Some IDA curves in the non-collapse regime approximated by straight lines
in the log-log representation. 142
Figure 5-4 - Fully parametric 16th, 50th and 84th percentiles of the drift response given
spectral acceleration obtained using parametric IDA curves. 143
Figure 5-5 - Response results of multiple-stripe analysis with cases identified as collapse
marked. 144
Figure 5-6 - The 16th, 50th and 84th percentiles of the semi-parametric three-parameter
distribution calculated from the results of the multiple-stripe analysis 145
Figure 5-7 - The 16th, 50th and 84th percentiles of drift response given spectral
acceleration estimated semi-parametrically using multiple-stripe analysis
(MSA) and fully parametrically using the IDA procedure. 147
Figure 5-8 - Multiple-stripe analysis counted 16th, 50th and 84th percentiles of the response
with and without consideration of the collapse information 149
Figure 5-9 - The counted 16th, 50th and 84th percentiles not considering the collapse
cases (as per Chapter 4) and the semi-parametric 16th, 50th and 84th percentiles
using multiple-stripe analysis. 150
Figure 5-10 - Non-parametric estimation of drift hazard: a) collapse cases considered

b) Plot 10-a together with drift hazard when the collapse cases are not
considered. 153
Figure 5 -11 - Drift hazard calculated by numerical integration using the

non-parametric and semi-parametric (using the lognormal distribution for
the non-collapse cases) methods. 154
Figure 5-12 - Full-parametric drift hazard estimation using single-stripe analysis 156
Figure 5-13 - Fully parametric factored demand estimation: Simple DCFD formulation
using the parametric IDA curves and non-parametric spectral acceleration
hazard 159
Figure 5-14 - Semi-parametric factored demand estimation: Using the closed-form for
drift hazard in Equation 5-23 considering the collapse cases, with the
parametric non-collapse (linear) IDA curves and non-parametric
spectral acceleration hazard. 161
Figure 5-15 - Double-stripe, semi-parametric factored demand estimation: Using the

closed-form drift hazard in Equation 5-23 considering the collapse cases,
with results for S a = 0.6 g and S a = 0.70 g and non-parametric spectral
acceleration hazard. 165
Figure 6-1 - The dynamic displacement response of LA20 model structure to the
combined record selection plotted versus the spectral acceleration for T=3.9 sec
and %5 damping. 184
Figure 6-2 - Residual-residual plot for rupture distance, as a second independent

variable for predicting structural response for LA20 structure, combined
record selection. 186
Figure 6-3 - Residual-residual plot for magnitude as a second independent variable for
predicting the structural response for LA20 structure, combined record
selection 187
Figure 6-4 - Residual-residual plot for the epsilon of the attenuation law, , as
a second independent variable for predicting the structural response for LA 20
structure, combined record selection. 188
Figure 6-5 - The static pushover curve for a bi-linear SDOF system with T=0.1sec,
= 0.05 , (strain hardening ratio) = 0.03 , and M P = 5000 k in . 189
Figure 6-6 - The dynamic displacement response of the bi-linear SDOF model
structure with T=0.1, damping = 5% , and Mp = 5000 k in to the combined
record selection. 190
Figure 6-7 - Residual-residual plot for magnitude as a second independent variable for
predicting the response of an SDOF system, combined record selection 191
Figure 6-8 - Residual-residual plot for distance as a second independent variable for
predicting the maximum inter-story drift angle, SDOF system, combined
selection 192
Figure 6-9 - Residual-residual plot for the normalized residual of the attenuation law,
, as a second independent variable for predicting the structural response
for SDOF system, combined record selection. 192
Figure 6-10 - residual-residual plot for spectral shape factor at SMF, as a second
independent variable for predicting the structural response for LA 20
structure, combined record selection. 195
Figure 6-11 - The reduction in the sigma of the regression when the spectral shape
factor at period T is used as a second independent variable in the regression,
LA 20 structure, combined record selection. 196
Figure 6-12 - The residual-residual plot for the shape factor at T=0.3 as a second
independent variable for predicting the structural response for SDOF system
with T=0.1, combined record set. 197
Figure 6-13 - The normalized sigma of the residual-residual plot for the spectral shape
factor at period T with respect to the sigma of the original regression on
spectral acceleration at T=0.1. 198
Figure 6-14 - Combined selection, shape factor for T=1.3 sec and T=3.9 sec plotted
against moment magnitude 199
Figure 6-15 - Attenuation prediction for shape factor at T=1 sec ( ) and T=4 sec ( ) plotted
against magnitude (thick lines). The thin is the linear regression prediction
based on the combined ground motion record set. 200
Figure 6-16 - Combined selection, shape factor for T=0.3 sec and T=0.1 sec plotted
against moment magnitude 201
Figure 6-17 - Combined selection, shape factor for TFM=0.1 sec and T=0.3 sec plotted
against magnitude 202
Figure 6-18 - The SDOF system with M P = 15000 subjected to the combined record
selection 204
Figure 6-19 - Maximum inter-story drift angles for LA20 structure, combined selection
scaled by a factor of 3 207
Figure 6-20- The residual-residual plot for magnitude as a second regression variable,
the combined record set scaled by a factor of 3. 209
Figure 6-21 - Comparing the (square) RMS of the residual-residual regression for the shape
factor at different periods with the (square) RMS of the original regression
on FMF spectral acceleration, LA 20 structure, the combined record set scaled
by a factor of 3. 210
Figure 6-22 - The original regression of the displacement-based response on PGA,

LA 20 building, the records scaled by a factor of 3. 215
Figure 6-23 - The residual-residual plot for magnitude as a second regression variable,
LA 20 structure, combined record set scaled by a factor of 3. 216
Figure 6-24 - The weighted regression of the displacement-based response on PGA,

LA 20 structure, the combined set of record scaled by a factor of 3. 217
Figure 6-25 - The weighted residual-residual plot for magnitude, LA 20 structure, the
combined record set scaled by a factor of 3. 218
Figure 6-26 -The hazard curves for the displacement-based response with PGA as the
intensity measure. The hazard curve in dashed lines is based on weighted
regression. 219
Figure 6-27 - The hazard curve for the displacement-based response with the pair
[PGA,M] as the intensity measure (thick line) compared to the curves in
Figure 6-26. 221
Chapter 1
Introduction
1.1 General overview
Every now and then a big earthquake shatters a part of our planet and sends seismic waves to the
built-environments across the world. The level of seismic awareness in the society surges and a
myriad of questions are raised as to how well the structures dealt with the earthquake and what
can be done to improve their future performance.
The history of formation of seismic building codes is marked by these big earthquakes; often
serving as landmarks for the evolution of seismic codes. As time passes, the design criteria
become more and more articulate and ambitious, just as the design objectives evolve from
collapse prevention only to also ensuring the possibility of immediate occupancy after an
earthquake. Loma Prieta Earthquake 1989, Northridge Earthquake 1994, and Kobe Earthquake
1995 are distinct landmarks in this sense, marking the point where the term Performance-Based
Design (PBD) found its way into seismic guidelines such as, VISION 2000, ATC-40, FEMA 273,
FEMA 350-352. The definition of the desirable performance for a structure was more-or-less
qualitative and based on the engineering judgment and intuition before then, whereas, the PBD
objectives are quantified based on life-cycle cost considerations. Life cycle considerations
encompass various parameters that affect the structural performance, such as, structural damage,
non-structural content damage, and human casualties. Hence, quantified performance objectives
are among the fundamental qualities distinguishing PBD from traditional earthquake engineering.
Apart from being rare events and having large consequences, earthquakes events also have large
uncertainty associated with them1. This makes the quantification of performance objectives even
more challenging and complicated. Application of the theory of probability in modeling the
uncertainty in earthquake hazard or Probabilistic Performance-Based Design has been the subject
of ongoing research in the past years (see Bazzurro and Cornell, 1994, Wen 1995, Beck 1999).
Probabilistic PBD is hardly a new concept; it has been first implemented in specialized design
guidelines such as the nuclear power plant procedures (see, e.g., Kennedy and Short, 1994, and,
1
Earthquakes are low-probability, large-consequence, and large-uncertainty hazards. Quoting Y.K. Wen,
from the paper: Reliability and Performance-Based Design, Structural Safety 23 (2001).
DOE, 1994). The performance objectives in the context of probabilistic PBD can be defined as
the probability of exceeding a performance level (see Krawinkler, 1996, 1999).
Various factors such as, structural damage, non-structural content damage, human casualties and
life cycle costs, are instrumental in measuring the structural performance under seismic
excitations. Nonlinear time-history analyses of structural response can be implemented in order to
make probabilistic estimations of the structural damage and to facilitate estimation of non-
structural damage and life cycle costs. On the ground motion side, there has always been the need
for variables that are going define and quantify the intensity level of the ground motion records as
they are applied to the structure. This is going to affect the selection of the ground motion records
to be used in the time history analyses and also the assessment of structural performance.
This dissertation resembles a journey through probabilistic performance-based design equipped

with common nonlinear dynamic analysis tools. The itinerary for this journey includes the
development of a theoretical foundation for assessment of the performance objectives using the
theory of probability, shaping the results into design formats that can be (indeed are)
implemented by seismic guidelines, using the nonlinear dynamic procedures within the developed
design format(s), and observing the validity of such formats for the case where the dynamic
analysis tools indicate that global instability is imminent. The final stop in the journey elaborates
on the selection of the ground motion records that are going to be applied in the non-linear
dynamic analyses. This is the place for testing the validity of some of the major assumptions that
were implicitly or explicitly made in order to make the performance assessments stated in the
thesis.
1.2 This thesis in the context of probabilistic PBD
Most of the current seismic design procedures (FEMA 356, ATC-40) are based on defining
design earthquakes with prescribed probabilities of exceedance in a given time period. One of
the main attributes of these procedures is that the probabilistic nature of the seismic excitations
are only implicitly accounted for in the definition of the design response spectra (see Bazzurro
et. al, 1998), and not in the statement of the performance objectives. In the recent years,
specifically after the Northridge Earthquake, considerable research effort has been focused on
defining probabilistic performance objectives (e.g., Collins et. al., 1995) that balance desirable
structural performance and life cycle costs (e.g., Wen, 2001). A performance objective can be
expressed in terms of the (mean) annual frequency of exceeding a specified limit state (limit state
frequency in short), LS 2. The limit state frequency can calculated by summing up the
contributions from all the events that can be generated from the faults that surround the site:
LS = v P
i
i LS |eventi (1 1)
where PLS |eventi is the probability of exceeding the limit state given an event on a specific fault, i,
and v i is the mean annual rate3 of activity of events generated by fault, i. Estimation of the term,
PLS |eventi directly would entail performing several (e.g., thousands of) non-linear dynamic analyses
in order to represent all of the events and ground motions that can be generated by a given fault.
Y.K. Wen and his co-workers (Collins et. al., 1996, Han and Wen 1997) have followed this
approach by employing simple equivalent non-linear systems in order to reduce the
computational effort.
The limit state frequency could also be evaluated, in a potentially more efficient manner, by
applying the total probability theorem (see Appendix B and Benjamin and Cornell, 1970) and
regression analysis in order to predict the variation of the conditional probability of exceeding the
limit state given the event and motion parameters such as moment magnitude, M, and source-to-
site distance, R:
LS = v P
i
i LS | M , R ( m, r ) f M , R (m, r ) dr dm (1 2)
where PLS |M , R ( x) is the conditional probability of exceeding the limit state for a given magnitude
and distance, and f M ,R (m, r ) is the joint probability density function (PDF) for magnitude and
distance.
The evaluation of the limit state probability can be greatly simplified by decoupling the structural
analysis and the ground motion predictions by means of introducing a site and structure-specific
interface variable referred to as the ground motion intensity measure (IM):
2
In this thesis, we have used the notation H (.) and (.) interchangeably to refer to the mean annual
frequency of exceedance.
3
Note the way we have used the terms probability and frequency in Equation 1-1, where the product of
the probability of exceeding the limit state times a mean annual rate results in mean annual
frequency of exceedance. We may sometimes alternate between the two terms, probability and
frequency, in this thesis, since, for the type of rare events we are studying, the corresponding numerical
values are very close (and small).

LS = PLS | IM ( x) f IM ( x) dx = PLS | x d IM ( x) (1 - 3)
where PLS | IM ( x) is the conditional probability of exceeding the limit state for a given ground
motion intensity level, x, f IM (x) is the PDF for the intensity measure, and IM (x) is the mean
annual frequency of exceeding (also known as the hazard function) a given intensity level x. The
practical implementation of Equation 1-3 is usually (implicitly or explicitly) based on the
assumption that the probability of exceeding the limit state for a given ground motion intensity
level is conditionally independent of ground motion characteristics such as magnitude and
distance (Shome et al., 1998). The term PLS | IM ( x) can be estimated by employing a very limited
number of non-linear structural analyses (see Shome et. al. 1998), whereas the IM ( x) term can
be calculated by performing conventional probabilistic seismic hazard analysis (PSHA),
IM ( x) = v G
i
i IM | M , R ( x | m, r ) f M , R (m, r ) dr dm (1 4)
(where G IM |M , R (.) is the conditional complementary cumulative distribution function for the
intensity measure given magnitude and distance). Such hazard information is routinely generated
by earth science community (e.g., the hazard maps available on USGS website). This approach
has been developed by Cornell and co-workers in the recent years (e.g., Bazzurro and Cornell,
1994, Shome et. al. 1998, Luco and Cornell 2003, Vamvatsikos and Cornell 2002, and, Jalayer
and Cornell 2003).
In the present work, an analytic framework will be derived for the evaluation of limit state
probability based on this last approach. An analytic closed-form solution for the limit state
probability in Equation 1-3 will be derived based on a set of simplifying assumptions. This
analytic expression forms a technical basis for developing design and assessment formats suitable
for implementation in seismic guidelines. It will be shown how non-linear dynamic procedures
can be implemented in order to provide parameter estimates for this analytic closed-form
solution. The performance objective(s) underlying the limit state probability calculations in this
work is purely based on structural response parameters and does not take into account life cycle
cost considerations. The methodology presented in this work could also be applied to alternative
interpretations of the limit state such as, for example, the life cycle cost of exceeding a specified
level, provided that appropriate procedures for defining the probability of the limit state given
IM = x (as a function of x) can be established.
1.3 The organization of this thesis
This thesis is organized in the same order it has been formed. Chapter 2 with the deserving title of
Foundation Development is the central piece that connects the other chapters. It is intended to
serve as a step-by-step derivation of a closed-form expression for the annual probability of
exceeding structural performance level(s) based on some simplifying assumptions. The
formulation of every piece is explained in detail in order to provide an insight into probabilistic
assessments for the interested reader, including those with limited experience with such
probabilistic derivations. The derivations start with hazard estimations for the intensity measure
of choice, which is the first-mode spectral acceleration. The next level of complexity is to derive
the hazard expressions for structural displacement response based on the available expression for
spectral acceleration hazard and expressions for the displacement probability distribution given
the intensity level. In the final step, structural limit state capacity information is added in order to
derive the expression for annual probability of exceeding a structural limit state, which also
serves as the primary product of this chapter. Another layer of complexity is added by
considering the uncertainty due to limited knowledge (epistemic uncertainty) in the formulation
of the limit state frequency.
Chapter 3 or Format Development starts from where Chapter 2 ends. This chapter discusses
several of the many alternative design formats that can stem from the expression for probability
of exceeding a limit state. Demand and Capacity Factored Design (DCFD) is a closed-form
design and assessment format that directly results form the original formulation for probability of
exceeding a limit state derived in the previous chapter. This format has been implemented in
FEMA 350, 351, and, 352 and in an ISO offshore structure guideline (Banon et. al., 2001). The
Fragility-Hazard design format is another way of transforming the closed-from expression for
probability of failure from Chapter 2 into a potentially graphical design format. A variation of this
format has been implemented in Department of Energy Guidelines (DOE 1020) for nuclear power
plants (PRA 19383). Similar to the previous chapter another level of complexity is added by
including the epistemic uncertainty in the formulations. The consideration of this type of
uncertainty may manifest itself in the form of a confidence statement about the performance
objective being met (which may in effect modify the demand and capacity factors in DCFD
format), or in the mean estimate for the limit state frequency (as in DOE 1020, 1994).
Chapter 4, Alternative Non-Linear Demand Estimation Methods for Probability-Based Seismic
Assessment was formed around the question of how the design formats developed are going to
be practically implemented, as the previous two chapters introduced many equations and
parameters that need to be estimated in practice. It began as a brief two or three page outline of
alternative ways to estimate the parameters in the DCFD formulation developed in Chapter 3. It
became important however, to explore the various ways that non-linear dynamic analyses can be
implemented in the design format for varying levels of analysis effort. The division of Chapter 4
into narrow-range and wide-range methods reflects a variety of possibilities in terms of the
number of analysis runs performed. The basic idea was to be able to estimate the parameters in
the DCFD format locally while performing the fewest number of nonlinear dynamic analyses at a
ground motion intensity level that lies in the region of interest, which is defined by the allowable
probability of exceeding a limit state. The wide-range methods are discussed in the context of
multiple performance levels. Although they require a comparatively large amount of analysis
effort, the wide range methods can provide a great deal of information about the structural
response to wide ranges of seismic excitation. These methods are also useful for assessing the
onset of global dynamic instability, also referred to as global collapse in the structure.
One of the underlying assumptions for the derivation of DCFD design format is that the
distribution of response for a given level of intensity can be modeled by a lognormal distribution.
However, for the displacement-based response in the highly non-linear region of global dynamic
instability in the structure this assumption may well be unrealistic. Hence, the question rises as to
how effective the DCFD format would be in the region of displacements close to the dynamic
instability in the structure. Chapter 5 is written as a sequel to Chapter 4 aiming to explore
structural response prediction in this region. It is devoted to the prediction of response when the
displacements are so large as cause non-convergence of the structural analysis computer
algorithm used and/or not explained properly by a common two-parameter distribution such as
the lognormal. In this manner, a modified version of DCFD format is introduced based on a three-
parameter distribution proposed by Shome (1999).
The probabilistic assessments coupled with non-linear dynamic time history analyses in the
context of DCFD format are typically based (at least implicitly) on the assumption that the
structural response given the intensity measure is conditionally independent of the ground motion
record characteristics such as magnitude and source-to-site distance. In simple words it is
assumed that the intensity measure is adequate for conveying the ground motion record
characteristics for the purpose of response predictions. Based on the assumption that this was true
for first-mode spectral acceleration and the structure under investigation, we took the liberty, in
Chapters 4 and 5, of choosing more-or-less arbitrary selections of ground motion records, and
also of scaling them whenever we needed multi-level assessments. Chapter 6 investigates whether
such assumptions are realistic for the two extreme non-linear cases of a 20-story long-period
structure and an SDOF system with a very short period. It is observed in both these illustrations
that first-mode spectral acceleration does indeed render the displacement response effectively
independent of record characteristics such as magnitude, source-to-site distance and the
normalized residual of attenuation prediction, . Also, in the case where the response is
dependent on one or more record characteristics, it is demonstrated how the output of non-linear
dynamic analyses can be post-processed in order to enhance the resulting demand estimations.
This is done by coupling the actual distribution of the ground motion characteristics for a
particular level of intensity, also know as the PSHA disaggregation results (e.g., Bazzurro and
Cornell, 1998), with the structural response results for that level.
Chapters 4,5 and 6 are (in part) based on the anticipation that assessment (and even design) will
be more and more based on systematic non-linear dynamic analysis procedures in the future.
Most common use of these procedures today is in the assessment of existing and/or damaged
buildings (e.g., FEMA 351, 352 and 356). These chapters show how to organize the ground
motion records, intensity levels, number of analyses, and processing of results. One can envision
that this could all be one day automated for the engineer once she has modeled the structure.
Chapter 2
A Technical Framework for Probability-based Demand
and Capacity Factor Design (DCFD) Seismic Formats
F. Jalayer and C. A. Cornell (2003). RMS Report No. 43- Part I Foundation Development
2.1 Abstract
This chapter presents an analytical foundation for probability-based formats for seismic design
and assessment of structures. These formats are designed to be suitable for code and guideline
implementation. The framework rests on non-linear, dynamic seismic analysis. The formats can
be used to ensure that the structural seismic design can be expected to satisfy specified
probabilistic performance objectives, and perhaps (more novel) that it does so with a desired,
guaranteed degree of confidence. Performance objectives are presumed to be expressed as the
annual probability of exceeding a structural performance level. Structural performance levels are
in turn defined as specified structural parameters (e.g., ductility, strength, maximum drift ratio,
etc.) reaching a structural limit state (e.g. onset of yield, collapse, etc.). The degree of confidence
in meeting the specified performance objective may be quantified through the upper confidence
bound on the (uncertain) probability. In order to make such statements, aleatory (random)
uncertainty and epistemic (knowledge limited) uncertainty must be distinguished. The single
seismic design foundation can be formatted into the alternative conventional design methods such
as LRFD design and fragility-hazard design. Versions of the new developments reported here are
already in place in recently completed seismic guidelines (see FEMA 350-352, and Banon et al.,
2001).
2.2 Introduction
This chapter is the first part to a report (Jalayer and Cornell 2003) that develops a framework for
probability-based seismic demand and capacity design and assessment procedures. In this
framework, the natural randomness inherent in seismic phenomena and the uncertainty involved
in the evaluation of parameters related to such phenomena are both considered.
This chapter (part 1 of the report) develops the necessary formal foundation for this probability-
based design framework; producing the annual probability of exceeding a prescribed performance
level as the main output. The foundation derivation is performed in two sections. The first
considers only the variability due to randomness. In the second section the uncertainty due to
imperfect models and model parameters are introduced into the foundation derivation.
The next chapter (part 2 of the report) derives various alternative design formats that stem from
the probability-based foundation developed in the former part of the report. Most of these formats
are analogous to Load and Resistance Factor Design (LRFD) procedures associated with static,
force-based structural engineering, e.g., the AISC LRFD Code, but due to the generalizations here
to a non-linear, dynamic displacement basis we refer to these new formats as DCFD (Demand
and Capacity Factor Design). The choice among these alternative formats must be made on
grounds such as familiarity, practicality, etc., because in many cases they are technically
equivalent.
2.3 Sources of Uncertainty in Engineering Problems
Sources of uncertainty in engineering safety problems are classified into two major groups
known, confusingly and unfortunately, by various pairs of words in the broader reliability
community, for example, randomness and statistical uncertainty, aleatory uncertainty and
epistemic uncertainty, frequency and probability, and simply Type I and Type II. Moreover, there
is still some inconsistency between the researchers regarding the nature of uncertainty each group
identifies. However, in the present work, the first term identifies the more familiar natural
variability such as the times and magnitudes of future earthquakes in a region, record-to -record
variability in acceleration time-history amplitudes and phases, etc. The second term of each pair
signifies the limited knowledge and data that the profession currently has about, for example the
modeling of structural systems in the highly non-linear range and exact numerical values of
parameters of physical and random (stochastic) models, e.g., the median value of the maximum
inter-story drift of a particular model frame under a population of future ground motions of
specified intensity. This second kind of uncertainty can be reduced by more data (larger sample
sizes) and/or by more research. In the following text we shall typically use the simple pair of
words randomness and uncertainty. Therefore we shall be using the second word in the more
restrictive sense of epistemic uncertainty and not in the broad, common sense, as we used it in the
title of this section. Occasionally, for example when precision is imperative, we shall use the
longer unambiguous terms aleatory uncertainty and epistemic uncertainty, which are now
quite common in seismic hazard analysis.
2.4 Document Map
This report contains a complete analytical background for the probability-based seismic design
procedure. For pedagogical reasons the development of the text follows a detailed stepwise
manner that makes it somewhat long. However, it is possible to bypass some sections without
losing the general picture. The document map below illustrates two possible routes the patient
reader can follow.
Source of Uncertainty: Limit State Probability

Randomness Only
Foundation Development
Sources of Uncertainty: Limit State Probbaility

Randomness and with Uncertainty
Uncertainty
Route 1
Route 2
DCFD Formats
Source of Uncertainty:
Randomness
Fragility/Hazard
Format Development
DCFD Format s
Sources of Uncertainty: with uncertainty
Randomness and
Uncertainty Fragility/Hazard
with uncertainty
Route 1 goes through the entire development of the framework taking in to account the
randomness source of variability only. Route 2 goes through a more generalized derivation that
considers uncertainty also.
2.5 Foundation Development
This part of the report is dedicated to developing an analytical foundation for the probability-
based framework. This foundation involves the entire endeavor that leads to the estimation of
probability of exceeding a specified limit state for a given structural system. In other words, the
final product of this section is the annual limit state probability that is calculated taking into
account the uncertainty in the various elements involved in the seismic design of the structural
system.
Estimation of the limit state probability (a short way of saying the probability of exceeding a
specified limit state) will be presented in two parts. In the first part, limit state probability is
derived considering only the variability due to randomness. In the second part, the more
generalized form of limit state probability is introduced which accounts for both randomness and
uncertainty.
2.5.1 Structural Limit States
Structural limit states are thresholds for structural behavior defined in different ways in various
codes. One of the most commonly used structural limit states is the global collapse limit state.
The foundation derivation represented in this text applies to virtually any limit state. However for
simplicity and clarity, this report focuses on the global collapse limit state C.
2.5.2 Structural Demand Variable (State Variable)
Demand or state variable is normally chosen as a displacement-based structural response

representative of structural dynamic and nonlinear behavior. The most common examples for
buildings include: roof displacement or inter-story drift.
For demonstration here we have chosen the maximum inter-story drift ratio (MIDR) as the
demand variable. (The maximum is the peak in the response time histories over all stories in the
building.) MIDR is particularly relevant to global collapse prediction (FEMA 350). Maximum
inter-story drift values may be obtained from the results of structural analyses for various ground
motion intensities.
We have chosen to refer to the maximum inter-story drift variable as D. This will keep the future
derivations general with respect to a generic demand variable D. It is also suggestive of the
displacement-based nature of the demand variable.
2.5.3 Structural Capacity Variable (Limit State Variable)
The capacity or limit state variable as the name suggests, is a limit (threshold) for acceptable
structural behavior. We have already introduced the demand (state) variable for describing the
structural behavior. The capacity (limit state) variable describes the limiting value for the demand
(state) variable. Obviously, it will be represented on the same basis as the structural demand
value; maximum inter-story drift ratio in this case. The capacity might be a pre-specified inter-
story drift ratio, e.g., 2% (which FEMA 350 uses for an onset of damage limit state), or
capacity with respect to connection failure that might be modeled as a random variable based on
test data. In this chapter, we shall focus on global (dynamic) collapse capacities determined for
IDA (Incremental Dynamic Analysis) curves corresponding to a set of ground motion records
applied to the structure (see Vamvatsikos and Cornell, 2001).
In order to keep the derivations general, we have used the generic notation C for the random
inter-story drift capacity. This will also be consistent with the demand variable denoted as D.
2.5.4 Limit State Probability PLS
The final product of the proposed probabilistic procedure is called the probability of exceeding a
structural limit state, where the limit state is the condition that, D > C. In order to be brief, we
will refer to it as the limit state probability PLS. For the case that we are mainly interested in, i.e.
the collapse limit state, it is also reasonable to call this quantity the failure probability. Therefore,
we seek:
PLS = P[ D > C ]
2.5.5 General Solution Strategy
In order to determine PLS = P[ D > C ] , we are going to decompose the problem into more tractable
pieces and then reassemble it. First, we introduce a ground motion intensity measure IM (such as
the spectral acceleration , Sa , at say 1 second period), because the level of ground motion is the
major determinant of the demand D and because this permits us to separate the problem into a
seismological part and a structural engineering part. To do this, we make use of a standard tool in
applied probability, The Total Probability Theorem (see Appendix B), TPT which permits the
following decomposition of PLS with respect to an interface variable (here, the spectral
acceleration):
PLS = P[ D > C ] = P[ D > C | S

all x
a = x] P[ S a = x] (2 - 1)
In Equation 2-1 we have introduced S a as the intensity measure. In simple words, the problem of
calculating the limit state probability has been decomposed into two problems that we already
know how to solve. The first problem is to calculate the term P[ S a = x] or the likelihood that the
spectral acceleration will equal a specified level, x. This likelihood is a number we can get from a
Probabilistic Seismic Hazard Analysis (PSHA) of the site. The second problem is to estimate the
term P[ D > C | S a = x] or the conditional limit state probability for a given level of ground motion
intensity, here represented by, S a = x . Estimating the conditional limit state probability, for a
given ground motion intensity, requires an understanding of, for example, response/demand
variability from record-to-record of the same intensity, which is an easier and purely structural
problem to resolve. The TPT simply tells us how to re-combine these two pieces of the problem
back into PLS . The solution strategy outlined above, calculating the limit state probability by
decomposing it with respect to spectral acceleration, shall be referred to as the IM-based

approach hereafter.
An alternative solution strategy (which is the main strategy employed in this chapter) consists of
decomposing the expression for the limit state probability in two steps and therefore employs two
interface variables. The first step is to decompose the limit state probability with respect to the
displacement-based demand (the first interface variable) using TPT:
PLS = P[ D > C ] = P[ D > C | D = d ] P[ D = d ]

all d
The second step is to decompose the term, P[ D = d ] , or the likelihood that the displacement-
based demand is equal to a value d, with respect to the spectral acceleration (the second interface
variable):
PLS = P[ D > C ] = P[ D > C | D = d ] P[ D = d | S

all d all x
a = x] P[ S a = x] (2 - 2)
This two-step solution strategy, which employs the displacement-based demand as one of the
interface variables, shall be referred to as the displacement-based approach. Equation 2-2 is a
special case of the framework equation used by Pacific Earthquake Engineering Research (PEER)
as a basis for probabilistic design and assessments.
It should be noted that the equations introduced in this section are valid for discrete interface
variables. However, here they solely serve as a schematic outline of the solution strategy. Later in
this chapter, we are going to present parallel equations for continuous interface variables.
2.5.6 Ground Motion Intensity Measure
The ground motion intensity measure, IM, implemented in the solution strategies outlined in the
previous section, serves as an interface between the seismicity characterization and structural
behavior assessment. Ideally, such a variable should contain sufficient information about the
ground motion to serve as an accurate and efficient predictor of structural response, and it should
preferably be a variable for which the PSHA results are available (or readily obtainable). This
problem has been studied by Shome et al. (1998) and by Luco and Cornell, (2003). It has been
demonstrated by Shome and Cornell (1999) that, for short and moderate-period structures, the
spectral acceleration at a period approximately equal to that of the fundamental mode of the
structure satisfies the criteria mentioned above. In fact, the study of such intensity measure is
the subject of significant current research by a variety of investigators within PEER. We shall use
this variable here for specificity, but the resulting derivations will not change if spectral
acceleration is replaced by any other scalar intensity measure, such as, for example, the inelastic
spectral acceleration (see Luco and Cornell, 2003).
2.5.7 Randomness: The Only Source of Variability
The probability-based seismic assessment and design procedure presented here aims to evaluate
the probability PLS that the limit state variable exceeds a limit state threshold LS. Our first
objective here is to derive the limit state probability assuming that randomness is the only source
of uncertainty in the design variables.
We will follow the displacement-based solution strategy discussed in Section 2.5.5. in order to
derive the limit state probability. The derivation are laid out in a step-by-step manner in order to
make them easier to follow. At the end of this section we are going to briefly present the IM-
based approach for deriving the limit state probability. We start by deriving the hazard values for
our adopted seismic intensity measure, which is the spectral acceleration of the first structural
mode. Then we use common probabilistic tools (e.g., TPT as explained previously) in order to
first derive the hazard values for the displacement-based response, (here, maximum inter-story
drift angle) and then to derive the limit state probability PLS .
2.5.7.1 Spectral Acceleration Hazard:
The hazard corresponding to a specific value of the ground motion intensity measure is defined as
the probability that the intensity measure for future ground motion events be greater than or equal
to this specific value. Spectral acceleration hazard curves are normally provided by seismologists
for a given site (e.g. The USGS website). Each curve provides the mean annual frequency of
exceeding a particular spectral acceleration for a given period and damping ratio. It is
advantageous to approximate such a curve in the region of interest by a power-law relationship
(see DOE, 1994, and, Luco and Cornell 1998):
H S a ( s a ) = P[ S a x] = k 0 x k (2 - 3)
k 0 and k are parameters defining the shape of the hazard curve.
Figure 2-1 shows a typical hazard curve for a Southern California site that corresponds to a period
of 1.8 seconds and damping ratio of 5%. As it can be seen from the figure, a line with slope k
and intercept k 0 is fit to the hazard curve in the logarithmic scales around the region of interest
(e.g., mean annual frequencies between 1/475 or 10% frequency of exceedance in 50 years and
1/2475 or 2% frequency of exceedance). Here, k =2.73 and k 0 =0.00012.
HSa(sa) =k0 (sa) -k
Figure 2-1 - A typical hazard curve for spectral acceleration. It corresponds to a damping ratio
equal to 5% and a structural fundamental period of 1.8 seconds
It is important to note that the hazard values are usually provided in terms of the mean rates of
exceedance over a certain time interval (usually a year) rather than the probabilities of
exceedance. Therefore, it is more appropriate to refer to the hazard function as, for example, the
mean annual frequency rather than the annual probability of exceeding a certain value.
Nonetheless, for very small probability values, that are for example derived from a Poisson
model, the average rate and the resulting probability value are almost the same. For simplicity, we
are going to drop the mean term before the frequency. However, in the part where epistemic
uncertainty is introduced into the problem, we will need to be more precise in how we refer to the
hazard function.
2.5.7.2 Median Relationship between Spectral Acceleration and

Inter-story Drift Demand
Observations of demand values are normally obtained from the result of structural time history
analyses performed for various ground motion intensity levels. Figure 2-2 shows such results, e.g.
maximum inter-story drift, D, versus Sa. For a given level of ground motion intensity, there will
be variability in the displacement-based demand results over any suite of ground motion records
applied to the structure. It is assumed here that this variability is a result of randomness in the
seismic phenomena as the discussed before (later in Section 2.5.7 we are also going to take into
account the epistemic uncertainties, such as modeling uncertainty, in addition to record-to-record
variability). It is convenient to introduce a functional relationship between the load intensity
measure and a central value, specifically the median D of the demand parameter based on the
data available from such time history analyses.
In general, for a spectral acceleration equal to x , the functional relationship will be:
D ( x) = g ( x )
This is called the conditional median of D given S a (more formally denoted as D|S a ( x) , but we
shall keep the simpler notation). We can construct a full conditional probabilistic model of the
variability displayed in Figure 2-2 by writing:
D = D ( x) = g ( x )
in which is a random variable with a median equal to unity and a probability distribution to be
discussed below. At this point we introduce a particular functional relationship that both
conforms to our perceptions of a structural performance curve and also helps simplifying the
future analytical efforts. We have used the linear regression in logarithmic space (i.e.
ln D ( xa ) = ln a + b ln x ) to fit a power-law curve, a S a b , to our collection of maximum inter story
drift ratio and first mode spectral acceleration data pairs.
D .e - D D = g(S a )
D .e D
Th is is a p rob ab ilistic m od el of th e
(con d ition al) d istrib u tion of d em an d
given an in ten sity level.
M axim um Inte r-s to ry D rift, D
Figure 2-2 - A set of spectral acceleration and demand data pairs, showing the model fit to these
data points.
It is not an objective here to describe the various ways D ( x) may be estimated. In design practice
it is likely to come from one or more structural analysis procedures, perhaps previously calibrated
to nonlinear dynamic results for similar structures (see FEMA 350, and Yun and Foutch, 2002).
In assessment practice or research it can be obtained through one or more schemes of selecting
and processing records and results (see Bazzurro et al, 1998, Luco and Cornell, 1998,
Vamvatsikos and Cornell, 2001, Chapters 3 and 4 of this thesis). We shall see below that the
number of required time history analyses may be quite small (e.g., of the order of 5 to 10 time
history analyses). For a set of drift demand and spectral acceleration data points, such a
regression in the logarithmic scale will result in the following relationship between spectral
acceleration and (median) inter-story drift response:
D ( x) = a x a b (2 - 4)
Figure 2-2 illustrates a typical power law relationship between the median maximum inter story
drift demand and the spectral acceleration for a three-story steel frame building located in Los
Angeles. In this case, b 1 which is consistent with the so-called equal displacement rule
(Veletsos and Newmark, 1960).
2.5.7.3 Annual frequency of exceeding demand Drift Hazard
We are going to break the displacement-based approach for deriving the limit state probability in
Equation 2-2 into two parts. The first part is to derive the probability that the displacement-based
demand exceeds a given value d, and the second part is to derive the probability that the
displacement-based demand exceeds the capacity or the limit state probability. In simple words,
the randomness in demand and capacity is modeled in two stages. This section describes the
derivation of a closed-form expression for the annual probability of exceeding a certain demand
value d, also known as the Drift Hazard, by modeling the randomness in the displacement-based
demand.
Recalling from last section, the median demand versus spectral acceleration relationship was
introduced as:
D ( x) = a x b (2 - 5)
As shown above, the demand can be written in terms of the product of its median value and a
lognormal random variable with the following characteristics:
D = D ( x) (2 - 6)
We assume (based on observation of data) that can be represented by a lognormal distribution,

in which case we define its parameters, the median and standard deviation of ln , to be:
= e mean(ln( )) = 1
ln( ) = D|S a (2 - 7)
where denotes the median value for . Note that what we call the dispersion, i.e., D|S a ,
will in general depend to some degree on the level of S a . Here for analytical tractability, we
assume it is constant; the value should be chosen for S a values in the range of primary interest. If
we replace D (x) with its corresponding value from Equation 2-4, the following expression for
drift demand as a function of spectral acceleration and lognormal random variable is obtained:
D = a xa b (2 - 8)
Since we have assumed that is a lognormal variable, we can also conclude that the
displacement-based demand D is also a random variable with the following statistical properties:
D| S a ( x ) = a x b
ln D|S a ( x) = D|S a (2 - 9)
where D|S a ( x) and ln D|S a ( x) are the conditional median and standard deviation of the natural
logarithm for the displacement-based demand given spectral acceleration. As mentioned above,
the conditional standard deviation of the natural logarithm ln D|S a ( x) or the conditional
fractional standard deviation D|S a ( x) of demand given spectral acceleration is assumed to be
constant. The conditional median demand for a given spectral acceleration D|S a ( x) (or more
briefly D (x) ) is approximated as a power-law function of the spectral acceleration level, x.

Figure 2-3 illustrates a graphical presentation of basic components for drift hazard evaluation in
which the median drift curve, the variability of the displacement-based response around it, and
the conditional lognormal distribution fit to the data (at any given S a ) are all plotted together
with the spectral acceleration hazard. The median drift times exp( D|S a ) is referred to as the
mean plus one sigma curve as it corresponds to the 84th percentile of the data for a lognormal
variable; this is illustrated in the figure as D ( x) exp( D|S a ) . In a similar manner, the median drift
times exp( D|S a ) is referred to as the mean minus one sigma curve as it corresponds to the 16th
percentile of the data (for a lognormal variable) that is illustrated in the figure as
D ( x) exp( D|S a ) . Our objective is to find the drift hazard curve, H D (d ) , the annual probability
of exceeding a certain drift angle d (e.g., P[ D d ] ). The strategy outlined below for finding the
PLS involves decomposition and re-composition via the TPT.
We have two random variables of interest D and S a , analogous to Equations 2-1 and 2-2
presented in Section 2.5.5. Under the general solution strategy, we can decompose the drift
hazard into conditional probability of exceeding drift value d for a given spectral acceleration
value x and the likelihood that the spectral acceleration is equal to the value x:
H D (d ) = P[ D > d ] = P[ D > d | S
all x
a = x] P[ S a = x ] (2 - 10)
mean minus sigma

median
mean plus sigma
D .e -D|Sa
D =g(x)
LN( D , D|Sa)
D .e D|Sa
x
H Sa(x) D (x)
H Sa(x) Maximum inter-sto ry drift angle (Demand), D
Figure 2-3 - Basic elements for drift hazard evaluation, H S s ( x) , and the distribution of D given Sa
characterized by D (x) and D|S a .

Where represents the (annual) rate of the occurrence of the events of interest (or rate of
seismicity for brevity), e.g., events with magnitude greater than 5 in surrounding 200km. Thus,
the drift hazard in Equation 2-10 is equal to P[ D > d ] times the rate of occurrence of the
earthquake events that interest us. Therefore, the drift hazard itself is expressed in terms of the
rate of exceedance or the mean annual frequency of exceedance (however, in this chapter we
may use the terms probability and frequency interchangeably).
We should note that the above expression involves discrete variables. However, since we are
using analytic parameter estimations, we are going to base our derivations on an equivalent
expression for the drift hazard derived for continuous variables:

H D (d ) = P[ D > d ] = P[ D > d | S a = x] f S a ( x) dx = P[ D > d | S a = x] dGS a ( x)
0 0
(2 - 11)
where f S a (x) is probability density function (PDF) at spectral acceleration value x, and, G S a (x)
is the complementary cumulative distribution function (CCDF) at S a = x . It should be noted that
the dG S a (x) term in Equation 2-11 is resulting from the following relationship between F ()
or the cumulative distribution function (CDF) and f () or the PDF for a random variable (e.g.,
spectral acceleration S a ):
P[ x S a x + x] P[ S a x + x] - P[ S a x] dFS a ( x) dG Sa ( x)
f S a ( x) = lim = lim = =
x 0 x x 0 x dx dx
(2 - 12)
The last equality is based on the fact that the CCDF is expressed in terms of the probability of
exceedance whereas the CDF is expressed in terms of the probability of being less than or equal
to. Therefore, their corresponding derivatives are equal in absolute values but will have opposite
signs.
It should be noted that the spectral acceleration hazard H S a ( x) is equal to the spectral
acceleration CCDF, G S a ( x) times the rate of seismicity :

H S a ( x) = G S a ( x) (2 - 13)
Therefore, we can re-write Equation 2-11 as a function of the spectral acceleration hazard:

H D (d ) = P[ D > d | S a = x] dG S a ( x) = P[ D > d | S a = x] dH S a ( x)
0 0
(2 - 14)
Since we have assumed that the displacement-based demand is a lognormal variable, we can
derive the term P[ D > d | S a = x] using the tables that provide the CDF of a standardized normal
variable (Rice 1995). In order to use the normal tables, we first need to transform the random
variable D into a standardized normal variable:
ln D mean ln D ln d mean ln D
P[ D > d | S a = x] = 1 P[ D d | S a = x] = 1 P[ | S a = x]
D| S a D| S a
d
ln D ln D|Sa ( x) ln d ln D|Sa ( x) ln( b
)
= 1 P[ ] = 1 a x (2 - 15)
D| S a D| S a D| S
a

where () is the standardized Gaussian CDF. The above equation is derived based on the
following property of a lognormal variable in which the mean of the logarithm is equal to the
logarithm of the median (Benjamin and Cornell 1970)
mean ln D = ln D
If we substitute the standardized Gaussian representation of P[ D > d | S a = x] in Equation 2-15

into Equation 2-14, the drift hazard will be expressed as:
d
ln(

b
)
a x } dH ( x)

H D (d ) = P[ D > d | S a = x] dH S a ( x) = {1
D|S Sa (2 - 16)
0 0 a

We are going to use integration by parts in order to re-arrange the above equation so that we can
integrate it analytically. We first need to calculate the derivative of the first term in the integrand:
d
ln( )
d b d ln d ln a x b b ln d ln a x b
{1 a x } = - = (2 - 17)
dx D| S dx D| S a x D| S D| S a
a
a

where () is the standardized Gaussian PDF which is equal to:
1
1 u2
(u ) = e 2 (2 - 18)
2
for any standardized normal variable u. The drift hazard in Equation 2-14 is re-arranged into the
following form after applying the integration by parts assuming that the term
P[ D > d | S a = x] H S a ( x) is close to zero for the integration limits, i.e., very small and very large
S a values. It should be noted that for a lognormal variable, the range of possible values vary from
0 to .

dP[ D > d | S a = x] b ln d ln a x b
H D (d ) = 0
dx
H S a ( x) dx = x
0 D|S a

D|Sa
H ( x) dx
Sa

(2 - 19)
Now we are going to replace the hazard term H S a (x) by its power-law approximation from
Equation 2-3 and also replace the Gaussian PDF by its analytical form in Equation 2-17:
2
b 1 1 ln d ln a b ln x
H D (d ) = x
0 D|S a

2
exp(
2
D|S a

) k 0 x k dx
(2 - 20)
In order to calculate the above integral analytically, we are going to form a square term (so that
we can form a Gaussian PDF) inside the integral. This way we can calculate the integral by using
the fact that the integral of a PDF function (over all possible values of the variable) is equal to
unity. We begin by some simple algebraic manipulations in order to make the equation is a bit
simplified:
2
d 1
ln( ) b ln x
b 1 1 1
H D (d ) = k 0 x exp( a ) exp( 2k ln x) dx
2 2 D|S a 2
0 D|S a
b

(2 - 21)
The next step is to form a full squared term inside the integral and also take all the constant terms
out of the integral:
H D (d ) =
1
D|Sa d
D|Sa 2 1 ln x {k ( ) 2 ln( ) b }
1 d b b 1 1 b a
x
2
k 0 exp( k ( ) ) exp(k ln( ) ) exp( ( ) 2 )dx
2 b a D|S a 2 2 D|Sa
0
b
(2 - 22)
We can notice that the term inside the integral is indeed the PDF for the standardized Gaussian
du
variable u with the derivative :
dx
1
D| S a d
ln x {k ( ) 2 ln( ) b }
u= b a
D| S a
b
du b
=
dx x D|S a
Therefore the expression for drift hazard can be also written as:
D|S a 2 d b
1
1 ln x {k ( ) ln( ) }
1 D|Sa 2 d d

H D (d ) = k 0 exp( k 2 ( ) ) exp(k ln( ) b ) { b a
2 b a dx D|Sa } dx
0
b

(2 - 23)
Noting that the integral of a normal PDF over all the possible values is equal to unity, the drift
hazard can be written in the following simplified form:
1 k 1 k2 2
1 D|Sa 2 d d D| S a
H D (d ) = k 0 exp( k 2 ( ) ) exp(k ln( ) b ) = k 0 ( ) b e 2 b2
2 b a a
(2 - 24)
In order to have a more condensed formulation of the drift hazard, we introduce the notation S ad
or spectral acceleration corresponding to drift angle d defined as:

1
d b
S ad =
a
This is also the solution of Equation 2-4 for a given value of d , i.e., if we read the corresponding
S a value4 from d = a S a b curve. The graphic interpretation of S ad can be seen from Figure 2-4.
In simple words, this means that for a given drift demand value d, we find the corresponding Sa
value from the median curve D ( S a ) = a S a b .
The derivation can be further simplified by making use of the hazard curve definition in Equation
2-3:
k 1 k2 2 1 k2 2
d
H D (d ) = k 0
b
e
D| S a
2 b2
= k0 ( )
S ad
k
e 2 b2
D| S a
a
1 k2 2
D| S a
H D (d ) = H S a ( S ad ) e 2 b2
(2 - 25)
S a at T1
b
= a Sa
D
d
Sa
d
Maximum Interstory drift Angle,max
Figure 2-4 - Spectral acceleration corresponding to the demand value, d.
It can be seen by inspection of Equation 2-25 that the hazard curve for the drift demand H D (d ) is
equal to the hazard function H S a (.) evaluated at the spectral acceleration corresponding to this
4
Note that S ad is not necessarily the median S a for a given value of drift angle d . It is just the
corresponding S a value found from the curve. In other words, the fact that the D - S a curve gives the
median drift d for a given value of S a does not mean that it will also provide the median S a for a given
value of drift d .
drift demand times a factor related to dispersion in the drift demand for a given spectral
acceleration. The first factor can be interpreted as a first order estimate; it is also the drift
hazard if the dispersion D|S a is zero. Experience suggests that the second factor may typically
have values in the order of 1.5 to 3. Note that in this form one can read the first factor directly
from a given hazard curve without actually making the approximating fit k0 xk . The log-log
slope k of the approximation is needed for the second factor however.
2.5.7.3.1 Numerical Example
We would like to derive the drift hazard curve for a 3-story (model) structure with brittle
connections located in Los Angeles. This structure is a typical 3-story steel moment resisting
frame building used in the SAC project (Luco and Cornell 1998). A set of nonlinear dynamic
analyses were conducted, and the resulting maximum inter-story drift ratios have been plotted
versus the first mode spectral acceleration as it is illustrated in Figure 2-5. The hazard curve
represented in Figure 2-5 corresponds to oscillators with a fundamental period around 1.0 sec and
located in Los Angeles, thus we have used it as the spectral acceleration hazard curve for our
model structure. In approximate analytical form it is:
3.03
H S a ( s a ) = P[ S a s a ] = 0.00124 s a
Note that the k value is nearly equal to 3.0. Our next step is to determine the median relationship
between spectral acceleration and drift. This is done in Figure 2-5, by fitting a line to the data
points in a log-log scale; which gives the following information:
D ( S a ) = 0.0325 S a 1.002
D|S a = 0.299 0.3
HSa(sa)=0.007
0.02
Sa =0.615
Figure 2-5 - Hazard curve for spectral acceleration values corresponding to an oscillator with a
period equal 1.0 and damping ratio of 2%.
Note that b 1 for this range of data, i.e., the median drift is approximately proportional to the
S a . It should be mentioned however that there may be a certain level of non-linearity (material or
geometric) in which b is not close to 1.0 anymore. Linear behavior is limited in this structure to
inter-story drifts less than about 0.01. We would like to evaluate the probability that the
maximum inter-story drift angle exceeds a specific value, say 2%, H D (0.02) . If we substitute
0.02 for d in Equation 2-25:
1 k2 2
2 D| S a
0.02
H D (0.02) = P[ D > 0.02] = H Sa ( S a ) e 2 b
1
d b
Recall that S ad is equal to S ad = per definition:
a
1
0.02
S a0.02 = = 0.615 [g]
0.0325
Equivalently we could simply have read this value from median line in Figure 2-3 by entering at a
drift value equal to 0.02. Now we look up for the value of H S a (0.615) for the spectral acceleration
hazard curve. As illustrated in Figure 2-5 it is equal to 0.007. Hence, H D (0.02) can be derived as:
1 32
( )( 2 )( 0.32 )
H D (0.02) = P[ D > 0.02] = 0.007 e 2 1
= 0.007 1.5 = 0.0105
1 32
Note that the factor exp( 0.3 2 ) is equal to 1.50.
2 12
We can repeat the above calculations for multiple drift values in order to obtain the drift hazard
curve or we can find an analytical expression for the drift hazard. In general we can compute the
drift hazard for a specified drift value, say d as follows:
1 k2 2
D| S a
H D (d ) = P[ D > d ] = H S a ( S ad )e 2 b2
1
1 32
Recalling that exp( 0.3 2 ) is equal to 1.5 and S ad = d b d the above equation
2 12 a
0 . 0325
becomes:
1 k2 2
2 D | S a d
H D (d ) = H Sa ( S ad )e2 b = 1.5 H Sa ( )
0.0325
Next we need to find the expression for the spectral acceleration hazard curve evaluated at
d
. This is:
0.0325
3.03
d d
H Sa ( ) = 0.00124 4.25 10 8 d 3
0.0325 0.0325
Finally the drift hazard for a specified value of drift, d is derived as:
H D (d ) 6.375 10 8 d 3
The above relationship is plotted in Figure 2-6.
1/100
d1/100
Figure 2-6 - Hazard curve derived for maximum inter-story drift values
One can use the above curve, to determine for example the 100 year return period drift, by setting
H D (d1 / 100 ) to 1/100 and solving for d1 / 100 :
1 / 100 = 6.375 10 8 (d1 / 100 ) 3
solving for d1 / 100 : d1 / 100 = 0.0185 . The same value can be also found simply from Figure 2-6.
2.5.7.4 Annual Frequency of Exceeding a Limit State
We have already derived the probability that the limit state variable exceeds a certain value. The
next step is to find the probability that this limit state variable exceeds a specified limit state
threshold or capacity, C. The difference in this case is due to the fact that the limit state threshold
can be a random variable itself. For example in the SAC project (FEMA 350) modern reduced
beam section (RBS) connections were concluded to have a median capacity of C = 0.07 (inter-
story drift ratio) with a dispersion of C = 0.2 reflecting specimen-to-specimen variability in
(hypothetical5) test results and even possible record-to-record variations in the drift failure due to
sequence effects in the low-cycle fatigue suffered by the connection. Beyond the drift capacity
the connection lost virtually all vertical load carrying (shear) capacity implying the potential
5
In fact no connections were able to be tested by the SAC project to such large drift ratios. The parameters
were estimated indirectly and are based on some level of expert opinion.
collapse of floor above. In this section we will derive the expression for the probability of limit
state, PLS = P[ D C ] , by introducing the variability in the limit state threshold. The basic
elements involved in the derivation are illustrated in Figure 2-7.
Once again we use the total probability theorem to sum the joint probabilities that limit state
variable exceeds the capacity variable for a given value of capacity, over the entire range of
possible values for the capacity variable:
PLS = P[ D C ] = P[ D c | C = c]P[C = c]
all c
(2 - 26)
We next assume that demand and capacity are (statistically) independent, i.e. that:
P[ D C | C = c] = P[ D c] (2 - 27)
C
sa
LN( D|Sa, D|Sa)
C LN( C , C )
H Sa(x)
D (x)
C
H Sa( s a )
Figure 2-7 - Basic elements for limit state probability evaluation, H S s (x) , distribution of drift
variable D given Sa characterized by D (x) and D|S a , distribution of capacity variable C
characterized by C and C
The annual frequency of exceeding the limit state, H LS , can be expressed as the limit state
probability PLS times the seismicity rate (as mentioned in the Section 2.5.7.4.):
H LS = PLS = P[ D C ] (2 - 28)
Since we are going to base our derivation on the expression for drift hazard, our calculations are
going to yield the mean annual frequency of exceedance (or limit state frequency in short), H LS 6,
as the end result. Therefore, the limit state frequency can be calculated by substituting Equations
2-26 and 2-27 into Equation 2-28:
H LS = PLS = P[ D C ] = P[ D c]P[C = c]
all c
(2 - 29)
The probability that drift demand exceeds drift capacity for a given value of drift capacity can be
readily determined from the drift hazard curve:
H D (c) = P[ D c] (2 - 30)
Substituting the term P[ D c] in Equations 2-29 by H D (c) from the above equation,
H LS = PLS = H
all c
D (c ) P[C = c] (2 - 31)
However, the above equation is valid for discrete variables; in the continuous form, the
summation is replaced by an integral and the probability term, P[C = c] , is replaced by a
probability density function term, f C (c) dc :

H LS = PLS = H D (c) f C (c) dc (2 - 32)
Substituting the drift hazard value for H D (c) from Equation 2-25 into Equation 2-32 results in:
1 k2 2
D|S a

c 2
H LS = H D (c) fC (c)dc = H S a ( Sa ) e 2 b fC (c)dc (2 - 33)
From Equations 2-3 and 2-4, H S a ( S ac ) is equal to c b

. Thus, the limit state frequency is
k0
a
obtained by performing the following integration:
6
In this chapter, we have used the notation H (.) in order to refer to the mean annual rate of exceedance.
However, in the next chapters we may use the notation (.) instead.
1 k2 2 k 1 k2 2
D| S a c D| S a

b
H LS = H S a ( S ad )e 2 b2
f C (c ) d c = k 0 e 2 b2
f C (c ) d c
a
(2 - 34)
For the above integral to be evaluated, the probability density function of the random variable C ,
f C (c ) , has to be known. Here for tractability, it is assumed that C is a lognormal random
variable with following characteristics:
median(C ) = C
ln(C ) = C
After some simple re-arrangements:
k 1 k2 2 k 1 k2 2 k
c D|S a 1 D|S a

b 2 b2 b 2 b2
H LS = k0 e f C (c ) dc = k 0 e c b f C (c ) dc (2 - 35)
a a
k

It can be seen that the term inside the last integral equals expectation of c b . It has been shown in
Appendix A that the expected value of log normal random variable Y (with median Y and
dispersion ln Y ), to the power of equals to:
1
a 2 ln2 Y
E (Y a ) = E (e a ln Y ) = (Y )a .e 2
Since C is assumed to be a lognormal variable, the above property can be used to solve the
integral in Equation 2-35 as follows (For further details regarding the integration scheme refer to
Appendix A):
k 1 k2 2 k k k 1 k2 2 1 k2 2
1 b D|S a 1 b D|S a C
H LS = k0 e 2 b2
E( c b ) = k0 C b e 2 b2
e 2 b2
a a
We conclude that:
k 1 k2 2 1 k2 2
b D|S a C
H LS = P[ D > C ] = k 0 C e 2 b2
e 2 b2
a
We can recognize in the above expression the spectral acceleration hazard from Equation 2-3
combined with the spectral acceleration-median drift relationship in Equation 2-4, k 0 ( C / a) k / b ,
which equals the hazard value for the spectral acceleration corresponding to median capacity,
S aC :
k
b
H S a ( S aC ) = k 0 C (2 - 36)
a
Thus:
k
1 k2 2 1 k2 2
b D| S a
C
C
k 0 C e 2 b2
= H Sa (S a ) e 2 b2
= H D ( C ) (2 - 37)
a
where the last equality is based on the expression for drift hazard H D (.) at median capacity, C ,
from Equation 2-25, and S aC , as mentioned before, is the spectral acceleration corresponding
to a drift value equal to C , i.e., S aC = ( C / a )b . Finally, the probability of limit state is derived
1
as:
1 k2 2 1 k2 2 1 k2 2
C D| S a C
H LS = P[ D > C ] = H D ( C ) e 2 b2
= H S a ( S aC ) e 2 b2
e 2 b2
(2 - 38)
It can be observed that the probability of limit state (or the probability of drift demand exceeding
the drift capacity) is equal to the hazard curve for the spectral acceleration corresponding to the
median drift capacity times two coefficients accounting for the randomness in drift demand for a
given spectral acceleration and the randomness in drift capacity itself. Again the first factor can
be seen as a first-order approximation to the limit state frequency, H LS .
2.5.7.4.1 Numerical Example:
Returning to our 3-story frame numerical example of the last section, now we assume that the
median drift capacity and its dispersion parameter are given as:
median(C ) = C 0.07
ln(C ) = C 0.20
We first need to find H S a ( S aC ) . We can do this graphically. S aC can be calculated as the spectral
acceleration corresponding to C = 0.07 from median-spectral acceleration curve in Figure 2-7
resulting in S a0.07 2.15 g (note that the capacity points in the figure are only for schematic
representation). We enter the hazard curve (Figure 2-8 below) with this value and find
H S a (2.15) 0.00012 .
0.007
HSa(Sa0.007)=0.00012 Sa =2.15
Figure 2-8 - The spectral acceleration hazard curve. The hazard value for a spectral acceleration
equal to, 2.15, is shown on the figure.
Alternately we can use the analytical approximations. Using Equation 2-36 we can calculate the
hazard for the spectral acceleration corresponding to the median drift capacity as:
3
0.07
H S a ( S aC ) = H S a ( S a0.07 ) = 0.00124 = 1.2 10 - 4
0.0325
1 k2 2 1 32
Also: exp( C ) = exp[ 0.2 2 ] = exp[0.18] = 1.19
2 b2 2 12
1 k2 2
This value together with the 1.50 value for the coefficient exp( D|S a ) calculated in Section
2 b2
2.5.7.3.1 are used to calculate the annual frequency of exceeding the limit state from Equation
2-38 as follows:
1 k2 2 1 k2 2
C D| S a
C
H LS = H S a ( S a ) e 2 b2
e 2 b2
= 1.2 10 - 4 1.50 1.19 = 2.2 10 - 4
It can be seen that in this example the randomness in the drift capacity and drift demand for a
given spectral acceleration causes the limit state frequency to increase about a factor of 2 over the
first order estimate of H S a ( S aC ) = 1.2 10 4 .
2.5.7.5 Annual Frequency of Exceeding a Limit State, the IM-based

approach
In this section we are going to derive the annual frequency of exceeding a limit state, H LS , by
following an IM-based approach. The total probability theorem (TPT) is used to decompose the
expression for the limit state frequency into (conditional) frequencies of exceeding the limit state
for a given spectral acceleration (the adopted IM), and composing the results by integration over
all spectral acceleration values:

H LS = P[ S a S a ,C ] = P[ S a S a ,C | S a = x] f Sa ( x) dx = P[ x S a ,C ] dH S a ( x)
(2 39)
where S a represents the IM-based demand, S a ,C represent the limit state capacity also expressed
in spectral acceleration terms, and represents the seismicity rate (the reason for including it in
the derivations is explained before for the displacement-based derivation). We have used
Equation 2-12 in order to express the PDF of spectral acceleration in terms of the increment in the
spectral acceleration hazard. We assume that the spectral acceleration capacity is a lognormal
variable with the following statistical parameters:
median( S a ,C ) = S a ,C
ln( S a ,C ) = S a ,C
We can observe that the first term in the integral P[ s a S a ,C ] can be also interpreted as the CDF
of the spectral acceleration capacity at S a = x :
FS a ,C ( x) = P[ x S a ,C ]
Since S a ,C is assumed to be a lognormal variable, the corresponding CDF can be expressed in
terms of the standardized normal CDF:

x
ln
Sa ,C
FS a ,C ( x ) = P[ x S a ,C ] = (2 - 40)
S a ,C

In order to be able to integrate Equation 2-39, we use integration by parts and transform the
equation into the following form:
x x
ln ln
S a ,C S a ,C
H LS = P[ S a S a ,C ] =
S a ,C

dH S a ( x) = d {
S a ,C
} H S a ( x)

(2 - 41)

Just as in Section 2.5.7.3., the derivative of the standard Normal CDF can be calculated as:
x
ln
d 1 ln x ln S a , C

S a ,C
{ } = (2 - 42)
dx S a ,C x Sa ,C S a ,C

After the derivative of the Normal CDF in Equation 2-42 is substituted in Equation 2-41, and the
hazard term is replaced by the power-law approximation from Equation 2-3:
k0 ln x ln S a ,C
H LS = P[ S a S a ,C ] = x x k dx (2 - 43)
S a ,C
S a ,C
If we substitute the expression for the Normal PDF in Equation 2-18 into the above equation:
2
k0 1 ln x ln S a ,C
x k dx
H LS = P[ S a S a ,C ] = 2 x S a ,C
exp
2

S a ,C

(2 - 44)
Similar to the derivation in Section 2.5.7.3, we transform the integrand into a complete square
term and take all the constant terms outside of the integrand:
H LS = P[ S a S a,C ] =
2
2
1 2 1 1 1 ln x {k Sa ,C ln S a ,C }

2
k 0 exp( k Sa ,C ) exp(k S a ,C ) exp( ) dx
2 x S a ,C 2 2 S a ,C
0
(2 - 45)
The term inside the integral is itself the derivative for a standard Normal CDF:
k S a ,C 2
1 d ln x ln Sa ,C e
H LS = P[ S a S a,C ] = k 0 exp( k 2 S a ,C 2 ) exp(k S a ,C )
2
dx
0
{(
S a ,C
)} dx
(2 - 46)
Noting that the integral is equal to unity, the limit state probability can be derived as:
1 1
H LS = P[ S a S a,C ] = k 0 exp( k 2 S a ,C 2 ) exp(k S a ,C ) = k 0 S a ,C k exp( k 2 S a ,C 2 )
2 2
(2 - 47)
We can observe the power-law term outside the exponential is equal to the frequency of
exceeding (i.e., hazard) a spectral acceleration equal to the median spectral acceleration capacity:
1
H LS = P[ S a S a,C ] = H S a ( S a ,C ) exp( k 2 S a ,C 2 )
2
(2 - 48)
We can argue that H Sa ( Sa ,C ) is a first-order approximation to the limit state probability and the
2
exponential term exp(k 2 Sa ,C / 2) is a magnifying factor that accounts for the sensitivity of the
limit state probability to the randomness in the spectral acceleration capacity. If we compare the
IM-based expression for the limit state frequency in Equation 2-48 to the displacement-based one
in Equation 2-38, we can observe that the exponential term accounting for the dispersion in
displacement-demand is missing. Also the slope parameter b that measures the gradient of the
displacement-based demand with regard to spectral acceleration is absent. This is because the IM-
based approach, when applying TPT to derive the limit state probability, does not employ the
displacement-based demand as (one of the) an intermediate variable(s).
2.5.8 Randomness and uncertainty as the sources of variability
In previous sections, a closed-form expression for annual probability of exceeding a limit state
(here, the collapse limit state) was derived. We saw that the hazard value for the load intensity
measure corresponding to median drift capacity (i.e., the annual probability of exceeding the load
intensity measure corresponding to median drift capacity) is a first-order approximation to the
limit state frequency. This first-order approximation is multiplied by two second-order estimate
coefficients accounting for the randomness in drift demand for a given spectral acceleration and
the randomness in drift capacity itself.
Our objective here is to derive the limit state frequency when there is both randomness and
uncertainty in the design variables such as spectral acceleration hazard, drift demand given
spectral acceleration and drift capacity. Our derivations are going to be based on the assumption
that to a first approximation we can represent all the epistemic uncertainty in variable X by the
uncertainty in its median. The model becomes:
X = X X (2 - 49)
where X is the current point estimate of the median of X , the unit-median random variable
represents the epistemic uncertainty as to the true value the median of X , and the unit-median
random variable X represents the aleatory randomness of X . We are also going to assume that
the deviation from median, , can be properly modeled by a lognormal distribution. In general,
of course, the epistemic uncertainty in X should also be taken into account. Also, the shape of
the distribution of X may not be lognormal.
As in the previous section, we start by deriving the hazard values for the load intensity variable,
spectral acceleration of the first structural mode. We use some probabilistic tools (e.g., TPT as
explained previously) to derive the hazard values for the limit state variable, maximum inter-story
drift, and then complete the derivation by obtaining the limit state probability PLS. Whenever
possible we will make use of the results obtained in the previous part and generalize them to the
case where there is both randomness and uncertainty in the design variables.
2.5.8.1 Spectral Acceleration Hazard:
The concept of hazard curves for the load intensity measure was introduced in the previous
section. Our focus was on the spectral acceleration hazard curves which are normally provided by
seismologists for a given site condition and its location with respect to a fault. The hazard curve
estimation involves many scientific assumptions (see Kramer 1996). In other words there is
uncertainty in the evaluation of a hazard curve. Thats why spectral acceleration hazard curves
are normally provided as mean and 84th percentile hazard curves (As shown in Figure 2-9). Here
we are going to take into account the uncertainty in the evaluation of the spectral acceleration
hazard.
In the previous sections, we found it advantageous to approximate the hazard curve by a power-
law relationship as proposed by Kennedy and Short (1994) and Luco and Cornell (1998):
H Sa (s a ) = k 0 x k
where k 0 and k are parameters defining the shape of the hazard curve. We are going to let an
equation of the same form as the one above represent the median estimate of the uncertain hazard
curve: H S a ( x ) = k 0 x k (2 - 50)
Further we introduce the random variable UH that represents the uncertainty in the spectral
acceleration hazard, so that we have: H S a ( x) = H S a ( x) UH (2 - 51) .
Here we have assumed that UH is a lognormal random variable whose statistical parameters
have the following characteristics:
median( UH ) = UH = e mean(ln( )) = 1
ln(UH ) = UH
(2 52)
where UH reflects the degree of uncertainty in the PSHA estimation. We recognize the spectral
~
acceleration hazard itself as an uncertain (random) variable, H S a ( x) , which can be represented as
the median (best) estimate times this uncertain deviation, ~UH :

~ )
H Sa ( x) = H Sa ( x) ~UH (2 - 53)
Note the use of a tilda to denote a random variable, when clarity is needed. Considering our
assumption about ~UH being log-normal, we can observe from the above equation that the hazard
for any value of S a itself is a lognormal random variable (i.e., instead of having a single
deterministic value assigned to it, it has a probability distribution). We can write the spectral
acceleration hazard as:
~ )
H S a ( x) = H S a ( x) UH = k 0 x k UH (2 - 54)
~ )
Where H S a ( x) is a lognormal random variable with its median equal to H S a (x) from Equation
2-50 and its dispersion measure (i.e., the standard deviation of the natural logarithm or fractional
standard deviation) equal to UH . The mean hazard curve can be written as:
1 2
) ) UH
H S a ( x) = H S a ( x) mean( UH ) = H S a ( x) e 2 (2 - 55)
This equation is based on a property of the lognormal variables; their expected value is equal to
the median times the exponential of half of the squared standard deviation (see Appendix A).
Figure 2-9 shows the 16th percentile, median, mean and 84th percentile hazard curves for a
California site that corresponds to a period of 1.8 seconds and damping ratio of 5%. Note that the
median curve in the figure is the same hazard curve we used in the previous section. The 84th
percentile is given by H S a ( x) e UH .
84th UH
HSa(x) =k0 (x)ke
2UH/2
HSa(x) =k0 (x)ke
HSa(x) =k0 (x)k
Figure 2-9 - 16th, median and 84th percentile spectral acceleration hazard corresponding to a
damping ratio equal to 5% and a structural fundamental period of 1.8 seconds.
Figure 2-10 shows the basic components of drift hazard evaluation when there is uncertainty (due
~
to limited knowledge and data) in the estimation of the spectral acceleration hazard H S a ( s a ) . The
probability density for uncertainty in hazard is plotted with solid black lines. The two hazard
curves on the graph correspond to the median estimate of hazard, H S a ( s a ) , and hazard curve for a
given value of deviation, UH , in the estimation of hazard curve, H S a ( s a ) UH , respectively.
2.5.8.2 Probability of Exceeding a Drift Demand value Drift

Hazard
Recalling from the previous section, the drift demand variable (given a specified S a level) was
introduced as the median demand value times a random variable representing the random
variation (e.g., record-to-record) around the median value. We assumed that has a lognormal
distribution:
D = D ( x) (2 56)
Randomness is assumed to be the only source of variability in the above expression. In general,
the median drift demand is also an uncertain quantity. The uncertainty in the median drift demand
is caused by the limited knowledge and data about modeling and analysis of the structural system
especially in the highly non-linear range and/or exact numerical values of the parameters of
structural model.
HSa(x) LN(D, D|Sa)

HSa(x).UH
LN(HSa(x), UH)
Hazard HSa(.) Maximum inter-story drift angle (Demand), D
Figure 2-10 - Basic components for the evaluation of drift hazard with uncertainty in the
~
estimation of spectral acceleration hazard, H S a ( x)
The uncertainty is also caused by using a finite number of non-linear analyses to estimate the
median value. The scatter of the displacement-based response in Figure 2-2 indicating record-to-
record variability implies that the estimate of the median, D ( x) , can depend on the particular
sample of records used and its size. In order to distinguish this type of uncertainty from the one
that we considered in the previous section, we refer to it as epistemic uncertainty. The median
inter-story drift can be expressed as the product of its median estimate, D ( x) and a random
variable UD (UD stands for the uncertainty in evaluation of D) representing the uncertainty
involved in the evaluation of D ( x) :
D ( x) = D ( x ) UD (2 - 57)
Replacing D ( x) in Equation 2-56 with its representation in Equation 2-57, the drift demand can
be written as:
D = D ( x) UD (2 58)
In order to be consistent with UD , we now subscript with RD, standing for the randomness
(aleatory uncertainty) in drift demand evaluation. Finally the drift demand is represented as:
D = D ( x) UD = D ( x) UD RD (2 - 59)
where RD and UD are assumed to be independent and to have log-normal distributions with
following characteristics:
RD = UD = e mean(ln( )) = 1
ln( RD ) = RD
ln(UD ) = UD
(2 60)
Our objective in this section is to derive the probability that the drift demand D exceeds a specific
value d. In order to minimize the calculation efforts, well make use of the drift demand hazard
that was derived in the previous section assuming that there was no variability due to uncertainty.
The drift hazard or the annual frequency that the drift demand exceeds a specific value was
derived from Equation 2-25 as:
1 k2 2 1 k2 2
H D (d ) = P[ D > d ] = k 0 ( )
s ad
k
e 2 b2
RD
= H S a ( s ad ) e
RD
2 b2
The spectral acceleration hazard for a given value of deviation in its evaluation, UH , can be
found from Equation 2-54 as:
) k
H Sa | UH ( x) = H Sa ( x) UH = k 0 x UH
Replacing the above value for spectral acceleration hazard in Equation 2-25, we obtain the drift
demand hazard for a given value of deviation in spectral acceleration hazard UD :
1 k2 2 1 k2 2
~ 2 RD 2 RD
H D|UH (d ) = P[ D > d ] = H S a | UH ( s ad ) e 2 b = H S a ( s ad ) UH e 2b
In the next step, we derive the drift hazard function for a given value of deviation in spectral
acceleration hazard, UH , and the given value of deviation of the median drift demand, UD :
1 k2 2
RD
H D|UH ,UD (d ) = P[ D > d | UD , UH ] = H S a |UD ( s ad ) UH e 2 b2
(2 - 61)
The term H S ( s ad ) must be interpreted as the median spectral acceleration hazard for the
a | UD
spectral acceleration that corresponds to drift d for a given value of deviation UD in the median
drift evaluation:
H S a |UD ( s ad ) = H S a ( s ad |UD ) (2 - 62)
In order to be able to calculate the above value, we need to find s ad |UD or the spectral acceleration
corresponding to drift d for a given value of deviation UD in drift evaluation. The median drift
demand for a given value of deviation UD can be found from Equation 2-57 as:
D ( x) = D ( x) UD
At this stage we assume that D ( x) has the same functional form as the one D ( x) had in the
previous part, stated in Equation 2-4, namely:
D ( x) = a x b (2 - 63)
Substituting the value for D ( x) from Equation 2-63 in Equation 2-57, the median drift demand
can be evaluated as:
D ( x) = a x b UD (2 64)
s ad |UD or the spectral acceleration corresponding to drift d for a given value of deviation UD in
drift evaluation can be evaluated by setting D ( x) in Equation 2-64 equal to d and solving for
s ad |UD . Hence, we can define s ad |UD as7:
1 1
d b d / UD b d
s ad |UD = =

= s a
UD
(2 65)
a UD a
7
In fact s ad |UD is nothing but the spectral acceleration corresponding to drift demand d which is being
calculated from the median curve a x b UD instead of a x b .
The graphic interpretation of s ad |UD can be seen from Figure 2-11. In simple words, this means
that we find the corresponding S a value from the median curve a x b UD :
Sa at T1
b
b
= a x UD
D = a Sa D
d/UD
Sa
d/UD d
Maximum Interstory Drift Angle, D
Figure 2-11 - Spectral acceleration corresponding to the inter-story drift ratio value d for a given
value of deviation of median drift, UD
Replacing the value for s ad |UD from Equation 2-65 in Equation 2-61:
1 k2 2 1 k2 2
2 RD d 2 RD
H D|UH ,UD (d ) = H S a |UD ( s ad ) UH e 2 b = H S a ( s a UD ) e 2b UH (2 - 66)
H ( s ad / UH ) can be calculated from Equations 2-65 and 2-50 as follows:
k
d d b
H S a (sa UD ) = k0
(2 - 67)
a UD
Substituting the value of H ( s ad / UH ) from Equation 2-67 in Equation 2-66 results in:
k
1 k2 2 1 k2
d b
2
d D| S a 2 D| S a
UD
H D|UH ,UD (d ) = H Sa (sa )e 2 b2
UH = k0 UH e 2 b

a UD
k 1 k2 2
2
d b D| S a k
= k0 e 2 b UH UD b (2 68)
a
In short, we have an expression for the drift hazard conditioned on the spectral acceleration
hazard and variables representing the uncertainty in drift, which is a simple analytical function of
UH and UD the random variable representation of those two uncertainties. Recalling from the
last section where the spectral acceleration hazard could be interpreted as an uncertain variable,
k
~ d b
H S a ( s ad ) = k 0 UH
a
~
we can interpret the drift hazard itself as an uncertain (random) variable H D (d ) , which is a
~
function of the uncertain spectral acceleration hazard H S a ( s a ) and uncertainty in drift
prediction, UD :
k 1 k2 2 1 k2
k 2 k
~ d RD ~
~ ~ b ~ b UD
b RD
H D (d ) = P[ D > d ] = k 0 2 b2
= H S a ( s ad ) e 2 b
2
e UH UD
a
(2 69)
The product of independent lognormal random variables raised to powers, such as k / b , is again a
lognormal random variable (Benjamin and Cornell, 1970). Therefore, we can conclude that the
drift hazard is also a log-normal random variable whose distribution parameters can be calculated
based on the information about the distribution characteristics of UH and UD from Equations
2-52 and 2-60:
1 k2 2 k 1 k2 2
~ RD RD
median( H D (d )) = H D (d ) = median( H S a ( s ad ) e 2 b UH H S a ( s ad
2
2 b2
b
UD ) = )e
k2
UH D = UH 2 + 2
UD 2
b
~
Therefore, the drift hazard H D (d ) is an uncertain quantity with median:
1 k2 2
RD
H D (d ) = H S a ( s ad ) e 2 b2
(2 70)
and fractional standard deviation:
k2
UH D = UH 2 + UD 2 (2 71)
b2
and, also mean value H D (d ) is equal to:

1 k2 1 k2 1 k2 2
2UH D 2 RD UH D
H D (d ) = H D (d ) e 2 b = H S a ( s ad ) e 2 b
2 2 2 b2
e (2 - 72)
After substituting UH D from Equation 2-69:
1 k2 2 1 2 1 k2 2 1 k2 2 1 k2 2
RD UH UD RD UD
H D (d ) = H S a ( s ad ) e 2 b2
e 2 e 2 b2
= H S a ( s ad ) e 2 b2
e 2 b2
(2 - 73)
Note that the uncertainty in the hazard curve can be dealt with simply by using the mean estimate
of the hazard curve. Similar to Figure 2-3, Figure 2-12 illustrates a graphical presentation of basic
components for drift hazard evaluation, but in this case there is uncertainty both in the estimation
~
of median drift curve ~D and spectral acceleration hazard H S a ( s a ) . In Figure 2-12, we have
plotted, the median estimate, D , of the uncertain median drift curve, the probability density
reflecting the uncertainty in D about that estimate, with a fractional standard deviation equal to,
UD , the median drift curve for a given value of uncertainty UD , a realization of median drift
curve D UD and the fractional standard deviation RD due to record to record variability in the
results of dynamic analyses around it. The median estimate for spectral acceleration hazard
H S a ( s a ) and the spectral acceleration hazard for a given value of uncertainty in hazard evaluation
H S a ( s a ) UH are shown in the same manner as in Figure 2-10.

D
D . UD
LN( D , UD ) LN( D . UD , RD )
H Sa(x). UH
H Sa(x)
LN(H Sa(x), UH )
H azard H Sa(.) M aximum inter-story drift angle (Demand), D

Figure 2-12 - Basic components for the evaluation of drift hazard with uncertainty in the
~
estimation of spectral acceleration hazard H S a ( s a ) and median drift ~D
Returning to our numerical example, we calculate the mean estimate of the drift hazard in the
case where there is uncertainty in the evaluation of the drift demand. We have the maximum
inter-story drift values resulting from 30 different nonlinear time history runs plotted in Figure
2-2. Fitting a line in log-log space to the data points gives us the following information about the
median inter-story drift and the dispersion around it:
D ( x) = 0.0325s a 1.002
RD = D|S a = 0.299 0.3
But strictly this is just the median estimate, D ( x) of the median drift curve. The uncertainty in
the median estimate for the inter-story drift can be due to modeling errors and other
approximations involved in the analysis procedure. Here we limit the consideration to the
statistical uncertainty in the median due to the finite sample size (nsample=30). The statistical
properties of the median inter-story drift can be estimated as (see Rice 1995 for the statistical
parameters for the mean estimate):
D ( x) = 0.0325 x 1.002
D| S a 0.3
UD = = 0.055
n sample 5.48
Analogous to the previous section, we would like to evaluate the mean estimate of the probability
that the maximum inter-story drift angle exceeds a specific value, say 2%, H D (0.02) . If we
substitute 0.02 for d in Equation 2-73:
1 k2 2 1 k2 2 1 k2 2
UH D RD UD
H D (0.02) = H D (0.02) e 2 b2
= H S a ( s a0.02 ) e 2 b2
e 2 b2
1
d b
Recall that s is equal to s = per definition:
d
a
d
a
a
1
0.02
s a0.02 = = 0.615 [g]
0.0325
The mean estimate for the spectral acceleration hazard can be evaluated as:
1
2UH
H S a ( s a ) = H S a ( s a ) e 2 .
UH here is assumed to be equal 0.50. With this assumption we can look up the value for
H S a (0.615) from the spectral acceleration hazard curve in Figure 2-5, which is equal to 0.007.
Hence,
1
( 0 .5 ) 2
H S a (0.615) = H Sa (0.615) e 2 = 0.007 1.13 = 0.079
H D (0.02) can be derived as:
2 2
1 32 1 32
x x 0.3 x x 0.055
H D (0.02) = P[ D > 0.02] = 0.0079 e 2 12
e 2 12
= 0.0079 1.5 1.014 = 0.012
Note that in the previous section where there was no uncertainty involved in the estimation of
inter-story drift demand H D (0.02) was equal to 0.0105. The net uncertainty here in the estimation
of H D (0.02) is equal to:
32
UH D = (0.5) 2 + 0.055 = 0.526
12
The primary contribution is from the PSHA, but this can change if UD increases to as large as
0.15. If non-linear dynamic modeling errors are considered, this value is likely to be considerably
larger than 0.15.
2.5.8.3 Annual Probability of Exceeding a Limit State
We would like to derive the probability that the drift demand exceeds the drift capacity or the
limit state probability, or simply the limit state frequency, in the face of epistemic uncertainty. In
the last section we derived the probability that the drift demand exceeds a specified value of drift
demand also known as drift hazard. The uncertainties in the derivation of the spectral acceleration
hazard and drift demand were considered in this derivation. Now we are interested in calculating
the probability that the drift demand exceeds a drift capacity, which is an uncertain quantity. In
the previous part the capacity was assumed to be a uncertain (random) variable whose variability
was caused by aleatory elements, e.g., connection-to-connection variability in a to-be-built
design, or record-to-record variability observed , as will be observed in Chapter 4 in global
instability capacity.
The drift capacity variable was introduced above as a median capacity value times a random
variable C representing the variation around the median value. We assumed that C has a
lognormal distribution:
C = C C (2 74)
In general, the median capacity is also an uncertain (random) variable. The uncertainty in the
median capacity is caused by the limited knowledge and data about the for example, untested
connection designs or the nonlinear structural modeling and/or analysis underlying global
stability prediction. The median capacity variable can be expressed as the product of its median
value, C and a random variable UC (UC stands for the uncertainty in evaluation of capacity, C)
representing the uncertainty involved in the evaluation of C :
C = C UC (2 75)
Finally, we subscript with RD, standing for the randomness in drift demand evaluation:
C = C UC = C UC RC (2 76)
RC and RC are assumed to be independent and to have log-normal distributions with
median( RC ) = median( UC ) = e mean(ln( )) = 1
ln( RC ) = RC
ln(UC ) = UC
(2 77)
Our objective in this section is to derive the probability that the drift demand D exceeds the drift
capacity C, recognizing all these uncertainty elements. Based on the expression derived for the
limit state frequency in previous part, Equation 2-38:
1 k2 2
RC
H LS = P[ D > C ] = H D ( C ) e 2 b2
(2 - 78)
We are going to use the above equation and combine it with our findings from the previous
section for the drift hazard in order to derive the limit state frequency, H LS . If capacity is a
specific deterministic value, c, (i.e., there is no randomness nor uncertainty in the evaluation of
capacity), the drift hazard function associated with exceeding drift level c (for a given value of
uncertainty in spectral acceleration hazard, UH , and uncertainty in drift demand, UD ) can be
derived based on the results of previous section from Equation 2-61, substituting c for d:
k 1 k2 2 k
c b RD
H D|UH ,UD (c) = P[ D > c | UH , UD ] = k 0 UH e 2 b2
UD
b (2 79)
a
We would like to first find the limit state frequency conditioned on the uncertainties in spectral
acceleration hazard, drift demand and drift capacity. The median capacity for a given deviation,
UC , of drift capacity median from its median estimator can be written as below based on
Equation 2-75:
C |UC = C UC (2 80)
If we substitute the median capacity associated with this given deviation, C|UC , from Equation
2-80 for c in Equation 2-79, the drift hazard conditioned on the uncertainties in spectral
acceleration hazard, drift demand and drift capacity will be found as:
k
1 k2 2 k
C |UC b RD
H D|UH ,UD ,UC ( C |UC ) = k 0 UH e 2 b2
b
UD (2 - 81)
a

Substituting the conditional drift hazard term at C |UC in the above equation into Equation 2-78,
the limit state frequency for a given value of uncertainty in spectral acceleration hazard, drift
demand and drift capacity can be found:
1 k2 2
RC
H LS |UH ,UD , = P[ D > C | UH , UD , UC ] = H D|UH ,UD ,UC ( C |UC ) e 2 b2
UC
k
1 k2 2 k 1 k2 2
b RD RC
= k 0 C UC UH e 2 b2
b
UD e 2 b2
a
k
1 k2 2 1 k2 2 k k
b RD RC
H LS |UH ,UD , = k 0 C e 2 b2
e 2 b2
UH UD
b b
UC (2 - 82)
UC
a
The above expression gives the limit state frequency conditioned on the spectral acceleration
hazard, drift demand and drift capacity as an analytical function of UH , UD and UC the
random variables representing the above-mentioned uncertainties. Similar to what we did in the
previous section for the drift hazard, here we can interpret the limit state frequency itself as an
~
uncertain (random) variable H LS . Recalling from Equation 2-54, the term,
k
b
k 0 C UH
a
~
is equal to H S A ( S aC ) . Hence, the limit state frequency can be introduced as a random quantity
~
that is a function of the uncertain hazard H S a (C ) , the uncertainty in drift demand
prediction, UD and the uncertainty in drift capacity prediction, UC :
1 k2 2 1 k2 k k
~ ~ RD RC
H LS = H S A ( S aC ) e 2 b

2 2 b2
e b
UD
b
UC (2 - 83)
It can be observed that the limit state frequency is also a log-normal random variable whose
distribution parameters can be calculated based on the information about the distribution
characteristics of UH , UD and UC from Equations 2-52, 2-60, and 2-77:
k k 1 k2 2 1 k2 2 1 k2 1 k2 2
~ ~ RD C 2 RD C
median( H LS ) = H LS = median( H S A ( s a C ) b UD ) = H S a ( C ) e 2 b
2 b2 2 b2 2 2 b2
b
UC e e e
k2 k2
H~ = UH 2 + UD 2 + UC 2
LS
b2 b2
~
Therefore, the uncertain limit state frequency H LS is an uncertain quantity with median,
1 k2 2 1 k2 2
RD C
H LS = H S a ( s a C
2 b2 2 b2
)e e (2 84)
and fractional standard deviation equal to:
k2 k2
H~ = UH 2 + UD 2 + UC 2 (2 85)
LS
b2 b2
and, also the mean limit state probability H LS is equal to:
1 1 1 k2 2 1 k2 2 1 k2 2 1 k2
2 PLS 2UH UD UC ( RD + 2UD ) ( 2 RC + 2UC )
C
H LS = H LS e 2 = H LS e 2 e 2 b2
e 2 b2
= H Sa (s a ) e 2 b2
e 2 b2
(2 86)
~
Note that in this final form, the mean estimate of H LS looks like H LS without uncertainty
~
(Equation 2-38) but now based on the mean estimate of H S a ( s a ) , and with increased 2 in the
capacity and demand (given S a ) exponential terms. Figure 2-13 illustrates a graphical
presentation of basic components for evaluation of the limit state probability; there is uncertainty
~
in the estimation of spectral acceleration hazard H S a ( s a ) , median drift demand curve ~D and
median drift capacity ~C . In Figure 2-13, we plot together the median estimate, C , of the
uncertain median drift capacity, the probability density reflecting the uncertainty in C about that
estimate, with dispersion, UC , median drift capacity for a given deviation, UC , in the estimation
of median drift capacity C UC , the probability density reflecting the randomness type of
uncertainty (for example, specimen-to-specimen variability the estimation of capacity) in
capacity C about the median drift capacity C UC , with dispersion, RC .
D
c
sa
D.UD
LN(D, UD)
H Sa(x).UH H Sa(x) LN(D.UD, RD)
LN(H Sa(x), UH)

LN(C, UC)
LN(C.UC, RC)
C C.UC
Hazard H Sa(.) Maximum inter-story drift angle (Demand), D
Figure 2-13 - Basic components for the derivation of the limit state probability when there is
~
uncertainty in the estimation of the spectral acceleration hazard H S a ( s a ) , median drift
demand D , and median drift capacity C .
For our 3-story frame numerical example, we would like to calculate the mean limit state
probability, when there is uncertainty both in the estimation of median drift demand and median
drift capacity. Recalling from last section, the median drift and the dispersion of drift for a given
level of spectral acceleration was estimated (by fitting a line in the log-log space to the data
points obtained by performing 30 different nonlinear time history analyses) as:
D ( x) = 0.0325 s a 1.002
RD = D|S a = 0.299 0.3
We have also estimated the statistical properties of the uncertain median drift demand as:
D ( x) = D ( x) = 0.0325 x 1.002
D| S a 0.3
UD = = 0.055
n sample 5.48
The median and dispersion for drift capacity were given before:
C = 0.07
RC = ln(C ) = C = 0.20
Note that the dispersion parameter C is due to the randomness type of uncertainty in drift
capacity. The statistical properties for the now uncertain median drift capacity we have estimated
here as (for an assumed sample of size 4 as the number of tests upon which the estimate of the
median connection capacity is based):
C 0.07
0.2
UC C = = 0.1
4 2
Equation 2-86 gives the mean limit state probability as:
1 1 k2 2 1 k2 2 1 k2 2 1 k2
2UH UD UC ( RD + 2UD ) ( 2 RC + 2UC )
C
H LS = H LS e 2 e 2 b2
e 2 b2
= H Sa (s a ) e 2 b2
e 2 b2
1 1
d d b b
We have defined s a equal to , therefore s a C can be calculated as C :
a a
1
0.07
s a C = s a0.07 =

= 2.15 [g]
0.0325
Our next step is to calculate the median spectral acceleration hazard at a spectral acceleration
corresponding to median drift capacity:
1 1 1
H S a ( s a C ) = H S a ( s a C ) e 2
2UH
( )
= k 0 s a C
k
e 2
2UH
= 0.00124 2.15 3.0 e 2
( 0.5 2 )
= 0.000124 1.13 = 0.00014

Assuming UH = 0.5 , same as in the previous sections. We can also look up the value for
H S a (2.15) from the spectral acceleration hazard curve from Figure 2-5, which is equal to 0.00012.
Hence, the mean estimate of the limit state probability H LS can be derived as:
1 32 1 32 1 32 1 32
( )( 0.32 ) ( )( 0.2 2 ) ( )( 0.12 ) ( )( 0.055 2 )
2 12 2 12 2 12 2 12
H LS = 0.00014 e e e e
= 0.00014 1.5 1.2 1.014 1.05 = 2.68 10 4
Note that in the first part of the document we calculated the limit state probability when the
epistemic uncertainty in the estimation of median demand and capacity was not taken into
account. The limit state probability in that case was equal to 2.2 10 4 , whereas the mean
estimate of the limit state probability calculated in the presence of uncertainty in the estimation of
hazard, demand and capacity is 2.68 10 4
0.007
HSa(Sa0.007)=0.00012 Sa =2.15
Figure 2-14 - The median estimate for the spectral acceleration hazard curve. The hazard value
for a spectral acceleration equal to 2.15 is shown on the figure.
2.5.8.4 Annual probability of exceeding a limit state: the IM-based

approach
The annual probability of exceeding a limit state following the IM-based was derived in Section
2.5.7.5, considering only the aleatory uncertainty (due to record-to-record variability) in demand
and capacity. In this section we are going to follow the same approach in order to derive the limit
state probability considering also the epistemic uncertainty. Similar to the previous sections, we
assume that the median capacity variable can be expressed as the product of its median
value, S a ,C and a random variable US a ,C representing the uncertainty involved in the evaluation
of S a ,C :
S a ,C = S a ,C US a ,C (2 87)
Therefore, similar to the previous sections, we can represent the spectral acceleration capacity as:
S a ,C = S a , C RS a , C = S a , C RS a , C US a ,C (2 88)
where RS a ,C and US a ,C are assumed to be independent and to have log-normal distributions with
median( RS a ,C ) = median( US a , C ) = 1
ln( RS ) = RS a ,C
a ,C
ln(US ) = US a , C
a ,C
(2 89)
where RS a ,C and US a ,C are fraction standard deviations representing the randomness and
uncertainty in the spectral acceleration capacity respectively. It can be shown (the procedure is
similar to the one described for the displacement-based approach in detail) that the limit state
~
frequency H LS is an uncertain quantity with median,
1
H LS = H S a ( S a ,C ) exp( k 2 RS a , C 2 ) (2 90)
2
and fractional standard deviation equal to:
H~ = UH 2 + k 2 US a ,C 2 (2 91)
LS
and, also the mean limit state probability H LS is equal to:
1 1 1 2 2 1 k2 2
2 PLS 2UH k US a , C ( RS a , C + 2US a , C )
C
H LS = H LS e 2 = H LS e 2 e 2 = H Sa (s a ) e 2 b2
(2 92)
As with the limit state probability derived following the displacement-based case, the mean
~
estimate of H LS looks like H LS without uncertainty (Equation 2-48) but based on the mean
~
estimate of H S a ( s a ) , and with increased 2 in the spectral acceleration capacity due to the
consideration of epistemic uncertainty.
2.6 Summary
A technical foundation for probabilistic seismic assessment of structures has been developed,
taking into account the randomness (aleatory uncertainty) and uncertainty (epistemic uncertainty)
in the seismic hazard, demand and capacity parameters. This foundation is based on a closed-
form analytical expression for the mean annual frequency of exceeding a limit state (limit state
frequency in short). Two different approaches were presented for deriving the limit state
frequency, namely, the displacement-based and the IM-based. Both approaches are based on
simplifying assumptions regarding the shape of the hazard curve and the probabilistic models
representing demand and capacity. This technical foundation forms an analytic basis upon which
alternative design and assessment formats can be developed. These formats are discussed in the
next chapter which is the second part to this report.
2.7 Acknowledgements
The authors would like to acknowledge the Pacific Earthquake Engineering Research Center,
PEER, funded by the U.S. National Science Foundation.
Chapter 3
Probability-based Demand and Capacity Factor Design
(DCFD) Seismic Formats
F. Jalayer and C. A. Cornell (2003). RMS Report No. 43- Part II Seismic Design Format
3.1 Introduction Format development
Chapter 2 was dedicated to developing an analytical foundation for the probability-based seismic
assessment of structures. The final product of this foundation development was the mean annual
frequency of exceeding a limit state or the limit state frequency in short. Limit state frequency
H LS was calculated taking into account the uncertainty in various elements involved in the
seismic assessment and design of the structural system. An analytical framework for calculating
the limit state probability is helpful for seismic assessment of the structures, e.g., calculating the
limit state probability for an existing structure and checking to see if its design falls within the
acceptable region. However, in a design problem, the actual structural members and connections
are not known beforehand. They are , rather, the end product of the design process. Conversely,
the performance objective for the design is usually set beforehand and can be expressed in terms
of the limit state probability. From a designers point of view, the structure can be designed to
satisfy a specified performance objective.
This part of the report addresses problems similar to the following: how to assess a proposed (or
existing) design for a structure for a known collapse limit state frequency of, say, 0.04 percent per
annum? or how to address the uncertainty (due to limited knowledge) involved in evaluation of
the design parameters?. This uncertainty is usually stated through questions such as, how to
design a structure for a known mean annual collapse limit state frequency of 0.04% with a
confidence interval level of 95%?.
We shall discuss various alternative design formats that stem from the probability-based
foundation developed in the first part of this document. These formats are in general suitable for
guidelines and code implementation. A major class of these formats are analogous in form to
(linear, static, force-based) Load and Resistance Factored Design (LRFD) procedures (AISC
LRFD code). However, these formats are based on generic, random (usually) displacement-based,
non-linear dynamic response variables: demand and capacity, and they are therefore referred
to as Demand and Capacity Factored Design (DCFD). The DCFD format can also be formulated
in terms of spectral acceleration-based (referred to as IM-based in Chapter 2) generic demand and
capacity variables. Fragility-hazard format is another IM-based format discussed in this chapter,
which is useful for designing/assessing the global behavior of a structure or a class of structures.
Unlike the foundation, which is unique, the formats are numerous. They are just various
representations of a common foundation. In other words, the choice among these alternative
formats is subjective. It is to be made on grounds such as familiarity, practicality, etc.
3.2 Randomness: The only source of uncertainty
Similar to the foundation development, the alternative design formats discussed in this chapter are
also presented in two parts. The probability-based foundations developed in this section are based
on the assumption that randomness is the only source of uncertainty, and hence, they are based on
the expression for limit state frequency derived in the first part of Chapter 2 (Equations 2-38 and
2-48).
Recalling from Equation 2-38 in Chapter 2, the limit state probability, expressed in displacement-
based terms, was derived as:
1 k2 2 1 k2 2 1 k2 2
C D| S a C
C
H LS = P[ D > C ] = H D ( C ) e 2 b2
= H Sa (S a ) e 2 b2
e 2 b2
(3 - 1)
where H () denotes the mean rate of exceedance in general8. We are going to re-arrange the above
equation into alternative forms, also known as DCFD design formats. The purpose for this re-
arrangement is to present this probability-based formulation in a way that is easy to be
implemented in the design practice. The basic components of Demand and Capacity Factored
Design Format (DCFD) are outlined in the following equation,
1 k 2 1 k
D | S a C2
D|P0 s e 2 b = C e 2 b (3 - 2)
a
8
Later in this thesis we shall use the notation (.) to represent mean annual rate of exceedance.
P0
where D|P0 s is the median drift demand for a given spectral acceleration, s a , corresponding to
a
hazard levels in the proximity of an acceptable limit state probability, P0 . C is the median drift
1 k 1 k
capacity, exp( 2 D|S a ) is the Demand Factor, and exp( 2 C ) is the Capacity Factor.
2 b 2 b
Equation 3-2 offers an alternative presentation of the formal foundation equation (Equation 3-1),
and is obtained by re-arranging Equation 3-1. We shall go through the re-arrangement step-by-
step below.
The Fragility-Hazard Format is another format discussed in this chapter. This format is derived
by re-arranging the S a based presentation of the foundation equation in Chapter 2 (Equation
2-48):
1 k 2 1 k
2 D| S a 2 C2
P0
sa e 2 b = s a C e 2 b
(3 - 3)
P0
where s a is the spectral acceleration with a Hazard level equal to an acceptable limit state

probability, P0 , and s a C is the spectral acceleration with a Fragility of 50%. Each format will be
developed and discussed in detail in the corresponding section. Before proceeding to the details
of the derivations, we are going to define or overview a few parameters (variables) that are going
to be very helpful in our future format presentations.
3.2.1 Spectral acceleration s ad corresponding to a displacement-

based demand equal to d
s ad or the spectral acceleration corresponding to displacement-based demand value, d , is defined

as the spectral acceleration corresponding to the value, d, from the median displacement-based
curve as a function of the spectral acceleration, i.e., as the inverse of that function:
s ad = D 1| S a ( d )
Recalling from the last chapter, the median displacement-based demand was approximated as a
b
power-law function of the spectral acceleration, D|Sa ( x) = a x . Based on this approximation,
s ad or the spectral acceleration corresponding to the displacement-based response d can be
expressed as:
1
d b
s ad = D 1|S a (d ) = (3 - 4)
a
Sa (T)
b
= a Sa
D
sad
d
Displacement-based demand, D
Figure 3-1 - Spectral acceleration corresponding to a displacement-based demand equal to d.
s ad is illustrated graphically in Figure 3-1. In simple words, s ad represents the spectral
acceleration value corresponding to a given demand value d from a median curve approximated
b
as, a S a .
P0
3.2.2. Spectral acceleration s a for a hazard level equal to P0
P0
s a is the spectral acceleration with a mean annual frequency of being exceeded (hazard,
defined in Chapter 2) equal to P0 :
1
P0
P0
s a = S1a ( P0 ) ( ) k (3 - 5)
k0
In which we make sue of the fact that (Chapter 2) the mean annual frequency of exceeding a
given spectral acceleration value (also known as the spectral acceleration hazard curve) can be
k
approximated (at least locally) by the power-law function, Sa ( x) = k0 x . Figure 3-2
illustrates the graphical presentation of P0

s a . The spectral acceleration s ad corresponding to a
drift demand equal to d is also plotted on the same figure.
Sa at T1
b
P0 = a Sa
Sa D
Sad
P0 d
Sa(sa) = [Sa > sa ] Displacement-based demand, D
Figure 3-2 - Spectral acceleration for a hazard level equal to P0
3.2.3 DCFD Format
DCFD format is analogous in form to the Load and Resistant Factored Design (LRFD)
procedures (see AISC design procedures, 1994). As the name suggests, this format is constituted
of demand and capacity multiplied by their respective factors. The same as LRFD procedures, the
DCFD format can be used to design a building against a certain factored demand by finding a
factored capacity. The probabilistic demand and capacity factors for the DCFD format are very
similar in concept to the partial safety factors applied to the load and resistance in LRFD design
procedures.
This format stems directly from the expression for limit state probability (Equation 3-1) after
some re-arrangements. It should be noted that the same simplifying assumptions that led to the
derivation of the closed-form foundation equation in the previous chapter are implicit here in the
derivation of the DCFD format. In order to develop a design format, we first need to set a design
criteria. One certain criteria is to design a structure so that the mean annual frequency of
exceeding a certain limit state (limit state frequency in short) is less than or equal to the allowable
annual probability P0 9:
H LS P0 (3 - 6)
where the equality holds at the onset of the limit state. Recalling from the previous chapter, the
limit state frequency can be expressed through a closed-form relationship (Equation 2-38 or 3-1).
This closed-form expression can be substituted for H LS in Equation 3-6:
1 k2 2 1 k2 2
D| S a C
C 2 b2 2 b2
H Sa (s a ) e e P0 (3 - 7)
where H S a ( s a C ) is the hazard value (mean annual frequency of exceedance) for a spectral

acceleration equal to s a C (i.e., the spectral acceleration corresponding to the median capacity
C ), and it can be derived from Equation 2-36 as:

k
C b
H Sa (s a ) = k 0 C (3 - 8)
a
If H S a ( s a C ) from the Equation 3-8 is replaced in Equation 3-7:
k
1 k2 2 1 k2 2
b D| S a C
k 0 C e 2 b2
e 2 b2
P0 (3 - 9)
a
After re-arranging the above equation in order to solve for median capacity, C , we get the
following expression for the median capacity required so that the limit state frequency H LS is
less than or equal to the allowable probability P0 :
b

1 1 k 2 1k 2
P k D| S a C
C a 0
e 2 b e2b (3 - 10)
k0

9
Note the mixing of the usage of the terms mean annual frequency and annual probability. Although
the more precise term to be used in this derivations is mean annual frequency, for the type of rare events
that we are interested, the corresponding values are virtually numerically identical.
The expression in the parenthesis (P0 k 0 )1 k is nothing but the spectral acceleration P0
s a , having
a hazard value equal to the allowable probability P0 as given in Equation (3-5). Substituting
(P0 k 0 )1 k with P0
s a in Equation 3-10 will make it look more simple:
1 k 2 1k
C a ( P0
sa )
b
e2
D| S a
b e2b
C2
(3 - 11)
where a ( P0
sa ) is in turn equal to the median drift demand
b
D| P0 S a
for a given spectral acceleration
of P0
s a (Equation 2-4). Thus, Equation 3-11 can be further simplified by replacing a ( P0
sa ) with
b
D|P0 S :
a
1 k 2 1 k 2
D|S a C
D|P0 S e2 b e2 b C
a
1 k
Finally we transfer the capacity-related exponential term, exp( 2 C ) , to the other side of the
2 b
equation changing the sign of the exponent:
1 k 2 1 k
D| S a C2
D|P0 S e2 b C e 2 b (3 - 12)
a
Equation 3-12 represents the DCFD format in its final form. The right-hand side or the capacity
side of the equation is called the factored capacity, parallel to LRFDs factored resistance. In a
similar manner, the left-hand side of the equation or the demand side is called the factored
demand for the allowable probability P0 , parallel to LRFDs factored load. It should be noted
that the factored demand (the equivalent to factored load) is a function of the allowable
probability level, P0 , whereas the factored capacity does not depend on P0 . In contrast to the two
factors in the AISC LRFD, where neither the demand nor capacity factor depends on P0 . The
DCFD format offers an alternative and equivalent statement for the design criterion, according to
which the factored demand for the allowable probability P0 should be less than or equal to the
factored capacity. This implies that at the onset of the limit state, the factored demand for the
allowable probability P0 is equal to the factored capacity. One of the main advantages of the
DCFD design format is that the probabilistic design criteria can be stated in terms of familiar
displacement-based response parameters. This makes the DCFD format compatible with existing
(deterministic) design procedures.
The following sections are going to discuss in more detail the components of the DCFD format
(Equation 3-12).
3.2.3.1 Displacement-based demand, D|P0 S

a
D|P0 S is the median displacement-based demand for a spectral acceleration equal to spectral
a
P0
acceleration, s a , (i.e., spectral acceleration with a frequency of exceedance equal to the
allowable probability, P0 ). We may also refer to it as the median demand for a given ground
P0
motion intensity s a in short. Adopting the analytical development above this median demand
was shown to be equal:
b

D| P0 s = a ( s)
P0
a
b P
= a 0
k
(3 - 13)
a
k0
But Figure 3-3 illustrates a graphical presentation of D|P0 s that demonstrates its more general
a
applicability. Looking at the figure, we can see that D|P0 s can be calculated in two simple steps.
a
P0
Step 1 is to find the spectral acceleration s a that has a frequency of exceedance (i.e., hazard)
equal to the allowable probability, P0 , from the hazard curve for the spectral acceleration. Step 2
is to find the displacement-based demand D|P0 s that corresponds to a spectral acceleration equal
a
P0
to s a from the median demand curve. Note that in application neither the hazard curve nor the
P0
median demand curve need to be in analytical form to evaluate s a and D|P0 s . This fact will be
a
exploited in applications that follow.

Sa at T1
P0
Sa
= (sa)
D D
P0
D| P0sa
Figure 3-3 - Graphical presentation of median demand D|P0 s for a spectral acceleration
a
equal to P0 s a
3.2.3.2 Displacement-based capacity, C
C is the median displacement-based capacity for the structure. Figure 3-4 illustrates the median
drift demand D|P0 s and capacity C on the same graph.
a
1 k
3.2.3.3 Demand factor exp( 2 D|S a )
2 b
1 k
exp( 2 D|S a ) or the displacement-based demand factor is a magnifying factor that takes into
2 b
account the randomness in the displacement-based demand. The randomness represented by the
factored demand is usually due to record-to-record variability. D|S , a dispersion measure for
a
the displacement-based demand, is equal to the standard deviation of the (natural) logarithm of
displacement-based demand for a given spectral acceleration. Typical values for D|S , in the
a
non-linear range, are about 0.30 to 0.60. In the special case (e.g., a linear SDOF oscillator) where
there is no dispersion in demand (given S a ), the demand factor will be equal to unity. k b can be
interpreted as the sensitivity of probability to a the change in the displacement-based demand;
which means that a factor of x change on the displacement scale will cause a factor of x k / b
change on the probability scale.
Sa at T1
P0
Sa
= ( sa)
D D
P0 C
D| P0sa
Figure 3-4 - Graphical presentation of median drift capacity C
1 k
Similar to the LRFD design procedures, the demand factor exp( 2 D|S a ) is also denoted by
2 b
. Clearly, is always greater than or equal unity (an exponential raised to a nonnegative
power). Thus, the design displacement-based demand is always greater than or equal to the
median demand due to the randomness-type of uncertainty in displacement-based demand.
1 k
3.2.3.4 Capacity factor exp( C2 )
2 b
1 k
exp( C2 ) or the factored capacity is a reduction factor that takes into account the
2 b
randomness type of uncertainty in the displacement-based capacity. It is an exponential term
raised to a non-positive power and hence is always smaller than one. Therefore, the design
capacity is always less than or equal to the median capacity due to the randomness-type of
uncertainty. The dispersion term in the exponential power, C , is the standard deviation of the
(natural) logarithm of the displacement-based capacity. Also k b is a factor reflecting the

sensitivity of the probability to a change in displacement-based capacity. Similar to the LRFD
1 k
design procedures, the capacity factor exp( C2 ) is also denoted by .
2 b
3.2.3.5 Factored demand and demand hazard
In the following, we are to present an alternative interpretation of factored demand. This

alternative derivation relates the factored demand to the demand hazard (mean frequency of
exceedance).In DCFD format, the factored demand (FD) was derived as:
1 k
FD = D| P0 s exp( 2 D|S a )
a 2 b
Replacing back the analytic expression for D|P0 s from Equation 3-13:
a
b

P k 1 k
FD = a 0
exp( 2 D|S a )
k0 2 b
We can solve the above equation for P0 :
k

FD b 1 k2
P0 = k 0 exp( 2 2 D|S a ) (3 - 14)
a 2 b
Realizing that (according to Equation 3-8) the term k 0 ( FD / a ) k / b is equal to the hazard value
for a spectral acceleration corresponding to a (median) demand value equal to FD:
k

FD b
H S a ( s aFD ) = k0 (3 - 15)
a
Replacing the term k 0 ( FD / a ) k / b in Equation 3-14 with its equivalent from Equation 3-15:
1 k2
P0 = H S a ( s aFD ) exp( 2 2 D|S a ) (3 - 16)
2 b
We can observe that right side of the above equality is equal to (demand) hazard for a demand
value equal to FD (Equation 2-25):
P0 = H D ( FD) (3 - 17)
Conversely, the factored demand can be written as the inverse of the hazard function at value P0 :
FD = H D1 ( P0 ) (3 - 18)
The above equation states that the factored demand for an allowable probability P0 is equal to the
(displacement-based) demand with a mean annual frequency of exceedance (hazard) equal to P0 .
This alternative interpretation for the factored demand is going to be helpful later in Chapters 4
and 5 where we need to evaluate the factored demand for more general cases (i.e., the analytic
assumptions underlying the derivation of DCFD have not been employed).
3.2.3.6 General Form for the DCFD design format:
We have already discussed the derivation of the closed-from for DCFD format (Equation 3-12),
which resulted from re-arranging the expression for limit state probability in Equation 3-1:
1 k 2 1k 2
D| S a C
D| LS S e 2 b C e 2b
a
However, it should be kept in mind that this format is based on the same simplifying assumptions
that we made for the foundation derivation in Chapter 2. The general form for the DCFD design
can be introduced based on the format we derived in Equation 3-12, but replacing D|PLS S with10
a
1 k 1 k
D, C with C, exp( 2 D|S a ) with , and exp( C2 ) with :
2 b 2 b
D C (3 - 19)
where D and C refer to a demand and capacity displacement-based parameters and and are
their corresponding factors. It can be noted that the DCFD format presented in its general form as
in Equation 3-19 looks similar to the LRFD format presentation. Another alternative general way
to present the DCFD format is by simply comparing the factored demand to factored capacity:
10
Despite their capital letter designation, D and C do not represent stochastic variables in this DCFD
context. (They do typically represent stochastic variables in Chapters 2and 3). Here, they are just referring
to some generic displacement-based demand and capacity parameter.
F.D. F .C. (3 - 20)
The benefit of this alternative representation is that factored demand and factored capacity can be
defined in a different manner from the DCFD format. A generalized definition for factored
demand is already discussed in the previous section. According to this definition the factored
demand is the demand value has a mean annual frequency of exceedance equal to the allowable
probability P0 . Generalized definitions for factored demand and factored capacity are going to be
discussed in Chapters 4 and 5.
3.2.3.7 Numerical Example: (Performance Evaluation for an

Existing Building)
Returning to the numerical example presented in Chapter 2, now we can assess the performance
of an existing 3-story frame for the collapse limit state for an allowable probability of
P0 = 4 10 4 (2% in 50 years). Based on the DCFD design format, we are going evaluate and
compare the factored demand for the allowable probability P0 = 4 10 4 and the factored capacity
for the collapse limit state.
Factored Demand D
Evaluation of the factored demand consists of two parts, a) calculating the median drift demand
D| 0.0004
Sa
for a spectral acceleration with a hazard equal to 4 10 4 , and b) calculating the demand
factor. The median demand D|0.0004 S itself can be calculated in two steps. The first step is to
a
0.0004
calculate the spectral acceleration s a with a hazard equal to 4 10 4 . This can be done either
by using Equation 3-5 or more generally by simply finding the spectral acceleration
corresponding to P0 = 4 10 4 from the hazard curve. The advantage of the second approach is
that the hazard curve does not necessarily need to be a power-law. This approach will be used
0.0004
extensively in Chapters 4 and 5. Here, we are going find s a both analytically and
0.0004
graphically. s a can be calculated from Equation 3-5 for P0 = 4 10 4 :
1
0.0004 0.0004 k
sa = ( )
k0
Recalling from the first part of the numerical example in the previous chapter, the parameter
estimates for k 0 and k were equal to:
k 0 = 0.00124
k = 3.03
0.0004
Finally, s a can be calculated as:
1
0.0004 0.0004 3.03
sa = ( ) = 1.458 [g]
0.00124
0.0004
Graphically, s a is estimated by the spectral acceleration corresponding to the hazard value
equal to 0.0004 from the spectral acceleration hazard curve. The hazard curve with parameters
k 0 and k (listed above) is plotted in Figure 3-5. It can be observed that a hazard value 0.0004
0.0004 0.0004
corresponds to s a equal to 1.45. After the s a is calculated (estimated), the next step is
to find the median displacement-based demand that corresponds to this spectral acceleration.
Again, the median demand can be either calculated from the power-law approximation,
D|Sa ( x) = a x b , or estimated graphically from the median displacement-based demand curve

that is plotted versus spectral acceleration.
Sa(sa)=0.0004
0.0004
Sa =1.45
Figure 3-5 - The hazard curve for S a (T = 1, = 2%)

0.0004
The median demand corresponding to a spectral acceleration equal to s a = 1.458[ g ] can be
calculated from the power-law relation as:
D|S a ( 0.0004 s a ) = a ( 0.0004

sa )
b
Recalling from the previous chapter, the parameter estimates for a and b were equal to:
a = 0.0325
b = 1.002
Finally, D|S a ( 0.0004 s a ) can be calculated as:
D|S a ( 0.0004 s a ) = 0.0325 ( 0.0004

sa )
1.002
= 0.0325 (1.458 )1.002 = 0.047
We can also estimate D ( 0.0004 s a ) graphically by finding the median demand value
0.0004
corresponding to a given spectral acceleration of s a or 1.45 [g] from the median demand
curve. In this example we have chosen the maximum inter-story drift angle (the absolute
maximum of the response time-history over all the stories in the structure) as the displacement-
based demand parameter. The maximum inter-story drift angle is plotted versus spectral
acceleration in Figure 3-6 below. In Chapter 4, we shall explore other ways to evaluate this
median drift, e.g., by simply making a small set of non-linear dynamic runs with all input records
0.0004
scaled to the single level, s a = 1.45 g .
D = D|Sa(sa)
0.0004
s a =1.45
D|Sa ( 0.0004 s a ) = 0.047
Maximum inter-story drift, D

Figure 3-6 - Spectral acceleration plotted versus maximum inter-story drift angle, and fitted
power-law relation (a line on the two-way logarithmic paper)
1 k
The next step is to calculate the demand factor or exp( 2 D|S a ) . As mentioned in Section
2 b
3.2.3.5, D|S is equal to the standard deviation in the (natural) logarithm of the demand given
a
spectral acceleration denoted by ln D|S . This can be approximated by the reported ln max |S on
a a
the graph in Figure 3-6 (it should be noted that the maximum inter-story drift angle, max , is in
fact the demand parameter used in this example) that is approximately equal to 0.3. The hazard
slope parameter k is reported on Figure 3-5 as 3.0. Also the median demand-spectral acceleration
slope factor b is equal to 1.0 (Figure 3-6). Now that are the parameter estimates are obtained, we
1 k
are ready to calculate the demand factor or exp( 2 D|S a ) :
2 b
1 k 2 1 3
D| S a ( )( )( 0.32 )
= e2 b =e 2 1 = 1.144
Finally, the factored demand is calculated by multiplying the median demand

1 k 2
D| S a
D|S a ( 0.0004 s a ) = 0.047 and the demand factor e 2 b = 1.144 :
1 k 2
D| S a
D = D|S a ( 0.0004 s a ) e 2b = (0.047)(1.144) = 0.0538 (3 - 21)
Factored Capacity C
As it was mentioned before, this numerical example demonstrates the assessment of structural
performance for the limit state of collapse, using maximum inter-story drift angle as the
displacement-based parameter. Therefore, the displacement-based capacity is represented by
maximum inter-story drift angle at the onset of the collapse limit state. Similar to factored
demand estimation, the estimation of factored capacity consists of two parts, a) calculating the
median capacity C for the collapse limit state, and b) calculating the capacity factor. Recalling
from the previous parts of this numerical example in Chapter 2, the median (drift) capacity C
for the collapse limit state and its dispersion parameter C (i.e., the standard deviation of the
natural logarithm of maximum inter-story drift angle capacity values) were estimated to be equal
to:
C 0.07
C = 0.20
1 k
Hence, the capacity factor or exp( C2 ) can be calculated as:
2 b
1 k 1 3
2C ( )( )( 0.2 2 )
=e 2 b =e 2 1 = 0.94
Finally, the factored capacity is calculated by multiplying the median capacity C = 0.07 by the
1 k
capacity factor exp( C2 ) = 0.94 :
2 b
1 k 2
C
C = C e 2 b = (0.07)(0.94) = 0.0658 (3 - 22)
Comparing the factored demand from Equation 3-21 and factored capacity from Equation 3-22,
we can observe that:
D = 0.0538 C = 0.0658
As we can see, the structure satisfies the design criteria in Equation 3-19 for an allowable
probability equal to 0.0004 (i.e., 10% in 50 years) corresponding to the collapse limit state. This
conclusion implies that the actual structures limit state frequency (probability of failure) is less
than 0.0004 per annum.
3.2.4 Fragility/Hazard Format An IM-based probabilistic format
In the previous sections, we outlined in detail the main components of DCFD format, which is a
displacement-based probabilistic design format. Here, we are going to discuss fragility/hazard
format, a design format that stems from the IM-based framework equation in Chapter 2. One
main advantage of an IM-based design format is that design and/or assessments are performed in
the level of spectral acceleration and do not involve explicitly the displacement-based response.
The fragility/hazard format (see e.g., DOE 1020, 1994, and Kennedy and Short, 1994) is a
graphical design format in which the design criterion involves comparing fragility curves to
hazard curves. The hazard curves represent the probabilistic ground motion intensity or in
general terms the loading characteristics, whereas the fragility curves represent the
probabilistic structural capacity or in general terms the structural resistance.
The same as the DCFD design format, the first step in developing a format is to set the design
criteria. In the case of the fragility/hazard format the IM-based design criterion can also be stated
as in Equation 3-6 for a given allowable annual probability, P0 : H LS , S a based P0
Where the limit state frequency is calculated from the IM-based expression for the limit state
probability in Chapter 2 (Equation 2-48):
1
k 2 2 S a , C
H LS , S a _ based = H S a ( S a , C ) e 2 (3 - 23)
The expression for the IM-based limit state frequency can be substituted in the design criterion
(Equation 3-6):
1
k 2 2 S a , C
H Sa ( Sa ,C ) e 2 P0 (3 - 24)
where S a ,C is the IM-based random variable representing the limit state capacity (or spectral
acceleration capacity in short) and k is the parameter reflecting the steepness of the hazard curve
for spectral acceleration. It will be demonstrated in the next chapter how to estimate the statistical
properties (i.e., median, S a ,C , and standard deviation of the natural logarithm, Sa,C ) of this
random variable using non-linear dynamic analysis procedures such as Incremental Dynamic
Analysis (see Vamvatsikos and Cornell, 2001). Recalling from Chapter 2, the spectral
acceleration hazard can be approximated (at least locally) by a power-law relationship:
Sa ( x) = k0 x k (3 - 25)
Therefore, S a ( Sa ,C ) , or the mean annual frequency of exceeding the median spectral
acceleration capacity can be calculated from the above expression and then substituted in
Equation 3-24:
1 2 2
k S a , C
k
k 0 S a ,C e2 P0 (3 - 26)
After some simple re-arrangements with the objective of separating the load and resistance
sides, Equation 3-26 would take the following form:
1
1
k 2 S a , C P k
S a ,C e 2 0
(3 - 27)
k0
Recalling from Equation 3-5, the right-hand side of the equation is in fact the indicator spectral
P0
acceleration s a for a hazard level equal to the allowable probability P0 (or indicator spectral
acceleration for allowable probability in short):
1
k 2 S a , C
P0
s a S a ,C e 2 (3 - 28)
This is the expression for the fragility/hazard format. A similar expression, with the exponential
P0
term (with a positive sign) applied to s a has been used in the current draft of ISO seismic
criteria for offshore structures (Banon et al. 2001). It will be shown below how this expression
relates to the fragility and hazard curves. Similar to the DCFD format, the left-hand side of the
expression represents the factored demand for the allowable annual probability P0 , and the
right-hand side represents the factored capacity. However, if compared to the expression for the
DCFD format in Equation 3-12, one can observe that the demand factor representing the
dispersion in displacement-based demand is missing in the demand side of the expression.
Nonetheless, the factored capacity looks similar to that of the DCFD format except for the fact
that the b value is missing from the capacity factor. This is to be expected since the b value
represents the (log) slope of the displacement-based demand parameter versus spectral
acceleration; and the fragility/hazard format does not explicitly involve the displacement-based
demand. Therefore, the design criterion based on the fragility/hazard format can be stated in
terms of the IM-based factored capacity being less than or equal to the IM-based factored demand
for a given allowable annual probability, P0 . The following sections are going to discuss fragility
and hazard curves and how they can be employed in order to make parameter estimates for the
fragility/hazard format in Equation 3-28.
3.2.4.1 Hazard curves
The hazard function, H S a ( s a ) , for a given spectral acceleration value, s a , can be defined as the
mean annual frequency of exceeding the spectral acceleration value, s a . The hazard function
H S a ( s a ) is discussed in more detail in chapter 2. Figure 3-7 illustrates a schematic hazard curve.
As it is mentioned before, the hazard curve is approximated by a power-law relation,
H Sa (sa ) = k0 sa k in the derivation of the framework equations in Chapter 2, where the
parameter k represents the steepness of the hazard curve. It is shown only schematically in Figure
3-7. In fact it is strictly the slope of the power-law hazard curve on a two-way logarithmic scale
graph. It will be discussed in Chapter 4 that the slope parameter k can be estimated as the local
slope of the hazard curve (Figure 3-7) in the region of hazard/spectral acceleration values that are
of interest.
In the context of the fragility/hazard format, the hazard curve represents the probabilistic
characteristics of load or demand. It is demonstrated in Figure 3-7 how the factored demand
P0
for the allowable probability P0 in the fragility/hazard format, s a , can be derived from the
P0
hazard curve. As it is shown in the figure, s a is the spectral acceleration with a mean annual
frequency of exceedance (hazard) equal to P0 .

3.2.4.2 Fragility curves
The structural fragility for a limit state is defined as the conditional probability of exceeding the
limit state capacity for a given level of ground motion intensity (conditional probability of failure
in short). If the ground motion intensity is represented in terms of the spectral acceleration, the
fragility can be expressed as:
FLS ( s a ) = P[ S a S a ,C | S a = s a ] = P[ S a ,C s a ] (3 29)
where FLS ( s a ) is the structural fragility at spectral acceleration s a for the limit state LS. It can be
observed from the above equation that the fragility is expressed as the probability that the random
variable S a ,C is less than or equal the value s a . Therefore, the fragility can also be stated as the
cumulative distribution function of the random capacity, S a,C . If it is assumed that the
probability distribution of the spectral acceleration capacity, S a,C , is lognormal with median,
S a ,C , and standard deviation of the natural logarithm, Sa,C , the fragility can be expressed in
terms of the widely tabulated standardized Gaussian distribution function:
sa
FLS ( s a ) = P[ S a ,C s a ] = (ln( ) S a ,C ) (3 - 30)
Sa ,C
It can be observed from the above equation that the structural fragility for a given limit state can
be plotted as a function of spectral acceleration. For a certain limit state, a monotonically
increasing fragility curve can be plotted. A schematic fragility curve is shown in Figure 3-7.
The median spectral acceleration capacity S a ,C is marked as the spectral acceleration that
corresponds to a fragility of 50%, since according to Equation 3-30:
S a ,C
FLS ( S a ,C ) = (ln( ) S a ,C ) = (0) = 0.50
S a ,C
Also the standard deviation of the (natural) logarithm of S a ,C is marked as the difference
between the spectral accelerations (on the logarithmic paper) corresponding to fragility values
16% and 50%, since again according to Equation 3-30:
S a ,C
S a ,C S a ,C e
FLS ( S a ,C e ) = (ln( ) S a ,C ) = (-1) = 0.16
Sa ,C
S a ,C
S a ,C e
S a ,C = ln = ln FLS 1 (0.50) ln FLS 1 (0.16)
S a ,C
The fragility curve for a certain limit state represents the probabilistic characteristics of structural
resistance or capacity for that limit state. Once the fragility curve is available for a limit state, the
1
factored capacity according to the fragility/hazard format, S a ,C exp( k 2 Sa ,C ) , can be
2
calculated based on the parameter estimates for, k , S a ,C , and Sa ,C from the hazard and
fragility curves (Figure 3-7).
Hazard Fragility
Sa(sa) = [Sa > sa ]
FSa,,C (sa) = P [Sa,C < sa ]

50%

Sa,C
16%
P0

P0 Sa
Sa,C
Figure 3-7 A schematic plot of the hazard and fragility curves. The main parameters of
hazard/fragility format are also shown on the figure.
3.2.4.3 The IM-based limit state frequency in terms of the fragility

and hazard functions
It was demonstrated in the previous sections that fragility and hazard curves are very helpful
graphic tools for estimating the IM-based factored demand and capacity. Moreover, it will be
shown in this sections that the fragility and hazard curves can also be used in order to calculate
the limit state frequency. The IM-based limit state frequency is derived from the following
expression (Equation 2-39):

H LS , S a based = P[ x S a ,C ] dH S a ( x) (3 31)
where the first term in the integrand is nothing but the fragility FLS ( s a ) at a spectral acceleration
equal to s a from Equation 3-29. Therefore, the limit state frequency in Equation 3-31 can also be
written as:

H LS , S a based = FLS (x ) dH S a ( x) (3 32)
Where the IM-based limit state frequency is derived in terms of fragility and hazard. This
equation states that the limit state frequency can be calculated as the area under the product of the
structural fragility curve for the limit state multiplied by the (absolute value of) the increment in
the hazard.
3.2.4.4 Numerical Example: (Performance Evaluation for an

Existing Building)
In the numerical example presented earlier for the DCFD format, we assessed the performance of
an existing 3-story frame for the collapse limit state for an allowable probability of
P0 = 4 10 4 (2% in 50 years). Here, we are going to use the same example in order to
demonstrate probabilistic assessments based on the fragility/hazard format. Based on the

fragility/hazard design format, we are going evaluate and compare the IM-based factored demand
for the allowable probability P0 = 4 10 4 and factored capacity for the collapse limit state.
Factored demand: The IM-based factored demand for a given probability P0 = 4 10 4 is equal
to P0 = 0 .0004
sa (Equation 3-28). Back in Section 3.2.3.7, the spectral acceleration with P0 = 4 10 4
P0 = 0 .0004
frequency of exceedance was found to be equal to s a = 1 . 45[ g ] . Therefore the factored
demand for an allowable probability P0 = 4 10 4 is equal to:
F .D .( P0 ) = P0 = 0.0004 s a = 1 .45 [ g ]
Factored capacity: In order to calculate the factored capacity for fragility/hazard format, we
assume that the (random variable) spectral acceleration capacity has median and (fractional)
standard deviation equal to:
Sa ,C 2.15 [ g ]
S a ,C 0.20
The slope parameter k for the spectral acceleration hazard curve is reported to be equal to 3 on
Figure 3-5. Now that we have the parameter estimates for estimates for, k , S a ,C , and S a ,C , we
can calculate the factored capacity:
1 1
F .C. = Sa ,C exp( k 2 Sa ,C ) = 2.15. exp( 3 0.2 2 ) = 2.15 x0.94 = 2.0 [ g ]
2 2
Comparing the factored demand for the allowable probability P0 = 4 10 4 that was calculated
above with the factored capacity, we can observe that:
F .D.( P0 ) = 1.45 [ g ] F .C. = 2 [ g ]
Therefore, we can conclude that the collapse limit state fragility/hazard design criterion is
satisfied for an allowable annual probability of P0 = 4 10 4 (i.e., 10% in 50 years). However, it
should be noted that the parameter estimates used in this section for the spectral acceleration
capacity are only for the sake of demonstration. In practical applications, non-linear dynamic
analysis procedures (see Chapter 4 or Vamvatsikos and Cornell, 2001) can be implemented in
order to build the structural fragility curve(s). Then the factored capacity can be calculated using
the structural fragility curve as it is shown in Section 3.2.4.2.
3.3 Randomness and uncertainty, the sources of uncertainty
The design/assessment formats introduced so far only considered the randomness (or aleatory)
type of uncertainty in the estimations for demand and capacity. This type of uncertainty is
normally caused by record-to-record variability in the demand and capacity. However, it is of
interest to include the uncertainty due to incomplete knowledge (epistemic uncertainty) in the
estimations for parameters related to hazard, demand, and capacity. As seen in Chapter 2,
consideration of the uncertainty due to incomplete knowledge affects the mean estimate of the
limit state frequency and/or the confidence statements that can be made about bounds on
estimates of that probability. confidence in the probabilistic assessments. Therefore, it is desirable
to measure the epistemic uncertainty involved in the estimation of the parameters, and also to
represent such uncertainty in the design or the assessments. One way to do this is to simply
replace H LS in the previous section (3.2) everywhere by its mean estimate, H LS . As per Equation
2-86 in Chapter 2:
1 k2 1 k2
( 2 RD + 2UD ) 2
( RC 2
+ UC )
C
H LS = P [ D > C ] = H S a ( S a ) e 2 b2
e 2 b2
(3 - 33)
In which H S a is the mean estimate of the hazard curve, RC 2 + UC 2 and RD 2 + RC 2 are the
total aleatory and epistemic uncertainty variances in demand and capacity , respectively. Then it
is clear comparing this to Equation 3-1, that both of the DCFD formats introduced in Section 3.2
can be up-graded to include epistemic uncertainty by simply replacing H LS by its mean
estimate H LS and the aleatory uncertainty variances, 2 for demand ( D | S a ) and capacity by
their total 2 s, i.e., the total aleatory and epistemic variances. For example, Equation 3-2
becomes:
1 k 2 2 1 k 2 2
( RD | S a + UD | S a ) ( RC + UC )
D|P0 s e 2 b = C e 2 b (3 - 34)
a
P0
where it is understood that the spectral acceleration s a is obtained from the mean estimate of
the hazard curve at P0 . It is implied that the tolerable limits state frequency is specified in terms
of its mean estimate. While not in this same format, DOE 1020 is based in using such a mean
estimate approach with combined or total (aleatory plus epistemic) variances used for the demand
and capacity. In this section we chose to outline a hybrid scheme. The DCFD format is extended
to account for the epistemic uncertainty in the demand and capacity parameter estimations by
associating a level of confidence with the frequency of exceeding a certain limit state. This format
has recently been implemented for performance evaluation of existing steel moment-resisting
structures in FEMA 351.
The DCFD format presented in this section mainly features consideration of epistemic uncertainty
in the structural demand and capacity parameter estimations. However, it does implicitly take into
account the epistemic uncertainty in the seismic hazard estimations by incorporating the mean
estimate for the hazard instead of the median estimate.
3.3.1 A confidence-based DCFD format
In Chapter 2, the frequency of exceeding a limit state was derived by taking into account the
uncertainty due to both aleatory and epistemic uncertainty. In such derivations, the limit state
frequency was not a single quantity but a range of possible values represented by a central value
(median) and a dispersion measure (standard deviation of the natural logarithm):
1 k2 2 1 k2 2
RD RC
H LS = H S a ( S aC ) e 2 b2
e 2 b2
k2 k2
H LS = H 2 + UD 2 + UC 2 (3 - 35)
b2 b2
Where H LS is the median estimate for the limit state frequency and H Sa (.) is the median estimate
for the spectral acceleration hazard. LS is the dispersion measure (standard deviation of the
natural logarithm) for the limit state frequency; and it contains the epistemic uncertainty-related
dispersion terms for hazard, demand and capacity. We note that these could be used to develop
one or more DCFD formats that treat epistemic uncertainty in hazard, demand, and capacity
uniformly. Here, however, we chose to develop the hybrid scheme introduced above. Suppose we
assume that there is no epistemic uncertainty in the estimation of the median spectral acceleration
hazard (i.e., H = 0 ), the dispersion term LS in Equation 3-35 would be simplified to:
k k
H LS = UD 2 + UC 2 = UT (3 - 36)
b b
where UT is the dispersion measure representing the total epistemic uncertainty in the
displacement-based demand and capacity parameters. In order to account for the epistemic
uncertainty in the estimation of hazard, we substitute the median estimate of the spectral
acceleration hazard, H LS , in Equation 3-35 by mean estimate of the spectral acceleration
1
hazard, H S A exp( H2 ) . The resulting median estimate for the limit state probability is denoted
2
by " H LS " (in order to distinguish it from the median hazard H LS in Equation 3-35:
1 k2 2 1 k2 2 1 k2 2
RD RC RT
C C
" H LS " = H S a ( S a ) e 2 b2
e 2 b2
= H Sa (S a ) e 2 b2
(3 - 37)
where the bar represents the mean estimate; and (parallel to UT ) RT is the dispersion
measure representing the total aleatory uncertainty in the displacement-based demand and
capacity parameters. It should be noted the hazard curves provided by the seismologists are
usually in terms of the mean estimates of the annual frequency of exceedance, or mean hazard
in short.
Now, we can build a confidence interval around the median estimate for the limit state
frequency reflecting the epistemic uncertainty in the estimation of the demand (given S a ) and
capacity parameters. The limit state frequency corresponding to the confidence level, x, denoted
x
by, H LS , can be expressed as:
K
x
H LS =" H LS "e x H LS (3 - 38)
where K x is the standard Gaussian variate associated with the probability x of not being
exceeded. K x can be found in standard probability tables under the Normal distribution as a
function of the number of standard deviations above or below the mean. Substituting the
median estimate for the frequency of exceeding the limit state from Equation 3-37 into
x
Equation 3-38, one obtains the upper x% confidence limit H LS on the limit state frequency:
1 k2 2
RT K x H LS
x C 2 b2
H LS = H Sa (S a ) e e (3 - 39)
Recalling from Chapter 2 and earlier sections in this chapter, the mean annual frequency of
exceeding (hazard) the spectral acceleration corresponding to median displacement-based
capacity can be estimated by power-law approximations:
k
C C k C
H Sa (S a ) = k 0 (S a ) = k 0 ( ) b (3 - 40)
a
Clearly this result represents a theoretically inconsistent treatment of the total epistemic
uncertainty, as that in hazard, H , is incorporated in H , while that in capacity and demand
K x H LS
(given S a ) is treated via the confidence factor e . The objective is to focus on the structural
epistemic uncertainties. More precisely, one should say that this represents an x% confidence
limit on H LS given the mean hazard curve.
After substituting the estimate for H S a ( S aC ) from Equation 3-40 into Equation 3-39, H LS
x
or the
limit state frequency corresponding to the confidence level x is derived as:
k
1 k2 2
x b RT K x LS
H LS = k 0 C e 2 b2
e (3 - 41)
a
Parallel to the derivation of the DCFD format in Section 3.2.3 (Equation 3-6), the design criterion
x
can be stated by comparing the limit state frequency H LS corresponding to a confidence level x
to an allowable probability, P0 :
x
H LS P0 (3 - 42)
x
Substituting the expression for H LS from Equation 3-41 into the design criterion in Equation 3-
42 above:
k
1 k2 2
b RT K x LS
k 0 C e 2 b2
e P0 (3 - 43)
a
Similar to the derivation of the DCFD format in section 3.2.3, we next make some re-
arrangements mainly in order to separate the demand and capacity sides:
b
1k 2 b
RT K x LS P k
C e 2b e k a 0
(3 - 44)
k0
Based on the relationship between the dispersion measure for the limit state frequency and the
one measuring the total epistemic uncertainty in Equation 3-36:
b
UT = LS (3 - 45)
k
Therefore, Equation 3-44 is simplified to:
b
1k 2
RT K x UT P k
C e 2b e a 0
(3 - 46)
k0
1k 2
Now, we multiply both sides of the inequality in exp( UT ) (in order to make this format
2b
look similar to DCFD):
b
1k 2 1k 2 1k 2
RT UT K x UT P k UT
C e 2b e 2b e a 0
e 2b (3 - 47)
k0
b

P k
After further simplifications noting that: D|S a ( P0 s a ) = a 0
, and also breaking up the total
k0
variance terms into the corresponding demand and capacity variances:
1k 2 1k 2 1k 2 1k 2 1k 2
RC UC RD UD UT K x UT
P0
C e 2b e 2b D| S a ( s a )e 2 b e 2 b e 2b e (3 - 48)
Equation 3-48 already resembles the DCFD format. Now, we can define the demand and capacity
factors as:
1k 2 1k 2
RD UD
= R U = e 2 b e 2 b
1k 2 1k 2
RC UC
= R U = e 2b e 2b (3 - 49)
Also the confidence factor corresponding to the confidence level, x, denoted by x can be defined
as:
1k
UT ( K x UT )
x = e 2b (3 - 50)
After the demand, capacity, and confidence factors are substituted from Equations 3-49 and 3-50
into Equation 3-48:
D|S a ( P0 s a ) C x (3 - 51)
or:
D|S a ( P0 s a )
x (3 - 52)
C
This is the final form for the DCFD format taking into account both aleatory and epistemic
uncertainty. This looks very similar to Equation 3-12 for DCFD format considering only the
aleatory; except for the confidence factor x and also the fact the demand and capacity factors in
Equation 3-51 also include the effect of epistemic uncertainty. As it was mentioned before, this
format is implemented in FEMA 351 for the purpose of the performance evaluation of the
existing steel moment-resisting structures. If Equation 3-51 is satisfied, one can say that the
probability of failure is less than P0 with confidence x%.
It is also interesting to re-arrange Equation 3-51 in terms of the ratio of the factored demand to
factored capacity related to aleatory uncertainty (according to Equation 3-49):
P
F .D.( P0 ) D|S a ( 0 s a ) R
= e K X UT (3 - 53)
F .C. C R
or:
F .D.( P0 ) e K X UT F .C. (3 - 54)
Which is the equivalent design criterion for the DCFD format taking into account the epistemic
uncertainty. It should be noted that the factored demand and capacity in Equation 3-54 take into
account only the aleatory uncertainty and are identical to those of Section 3.2.3.
The design criterion in Equation 3-53 can also be implemented in order to assess the level of
confidence in an existing design for an allowable probability P0 , by following the steps outlined
below:
1. Calculate the factored demand for an allowable probability P0 and also the factored
capacity from Equation 3-12 taking into account only the aleatory uncertainty.
2. Find the ratio of the calculated factored demand to factored capacity.
3. Estimate the dispersion measure UT = UD 2 + UC 2 accounting for the total
uncertainty in the estimation of the demand and capacity factors; examples appear in
DOE 1020, FEMA 351, etc.
4. Solve the equation,
F .D.( P0 )
= e K X UT (3 - 55)
F .C.
in order to find the corresponding Gaussian variate K x . Note that Equation 3-55 is a
special case of the design criteria in Equation 3-54 at the onset of the limit state.
5. Find the corresponding confidence level x for the existing design
3.3.1.1 Numerical Example
The procedure outlined above for finding the confidence level corresponding to an existing
design can be applied to the numerical example in Section 3.2.3.7 where the factored demand for
an allowable probability of P0 = 0.0004 and factored capacity for the collapse limit state where
calculated. The ratio of the factored demand to factored capacity is equal to:
F .D.(0.0004) 0.0538
= = 0.817
F .C. 0.0658
We have used the tables in FEMA 351 guidelines in order to estimate UT . For a 3-story (low-
rise) structure, the tables recommend the value UD = 0.15 representing the uncertainty involved
in the estimation of the displacement-based response using non-linear dynamic procedures for the
collapse prevention limit state. Also the guidelines recommend the value UC = 0.15 associated
with the uncertainty in the estimation of the global dynamic capacity for a low-rise structure.
Therefore, UT can be estimated as:
1
UT = (0.15 2 + 0.15 2 ) 2 = 0.212
The next step is to calculate the Guassian variate K x from Equation 3-55:
F .D.(0.0004)
= 0.817 = e K X 0.212
F .C.
which implies that: K x = 0.953
Hence, the corresponding confidence level for K x = 0.953 can be found from a normal table:
x = (0.953) = 0.83
Therefore, we can conclude that the confidence that associated with the existing design of this
structure is 83%. Precisely the same conclusion would be reached following the SAC-like format
with definitions in Equations 3-49 and 3-50. The factored demand and capacity would differ in
value as would their ratio, but the numerical confidence calculated via 3-50 would be the same.
3.4 Summary and conclusions
A probabilistic framework for the assessment of the performance of structures under seismic
excitations was developed in the previous chapter. This chapter discusses several of many
possible alternative design and assessment formats that stem from this probabilistic framework.
The design formats discussed can all be traced back to a general probabilistic design criterion,
which is satisfied when the frequency of exceeding a certain limit state is less than or equal to an
allowable probability P0 . A design format usually offers equivalent displacement-based or
spectral acceleration-based criteria parallel to the general design criterion. The advantage of these
equivalent criteria is that they are expressed in terms of structural response parameters and hence
the resulting format can be incorporated more easily into the existing design codes.
These formats can be categorized based on the types of uncertainty involved in the parameter
estimations. The first category takes into account the randomness, also known as the aleatory
uncertainty, in the assessment of the demand and capacity. The second category takes into
account both the randomness (aleatory uncertainty) inherent in the estimation of the demand and
capacity and also the uncertainty due to limited knowledge (epistemic uncertainty) in the
estimation of the hazard, demand and capacity parameters.
Within the first category, the Demand and Capacity Factored Design (DCFD) format was
discussed. This format is a (displacement-based) design format analogous to the LRFD
procedures that stems directly from the expression for the frequency of displacement-based
demand exceeding capacity for a certain limit state. The DCFD format offers a displacement-
based design criterion in which the factored (displacement-based) demand (representing load)
for the allowable probability P0 should be less than or equal to the factored (displacement-based)
capacity (representing resistance) for a certain limit state. Another format discussed under the
first category is a spectral acceleration-based format known as the Fragility/Hazard format, in
which the fragility curves represent the structural resistance and the hazard curves represent the
seismic load. This format offers a design criterion in which the spectral acceleration with a
hazard value (i.e., frequency of exceedance) equal to the allowable probability P0 , is less than or
equal to the factored capacity expressed in spectral acceleration terms. Each fragility curve is
specific to a certain limit state and provides the necessary parameter estimates for the calculation
of the factored capacity for that limit state. The fragility/hazard format has been implemented for
the design and evaluation of energy facilities (e.g., nuclear power-plants) in the DOE 1020
guidelines and for offshore structures in ISO guidelines.
Within the second category of the design formats, a more general form of the displacement-based
DCFD format is discussed. This format associates a level of confidence with the estimated
frequency of exceeding a limit state. This confidence level represents explicitly the epistemic
uncertainties involved in the estimation of the demand and capacity parameters and implicitly
(and approximately) the epistemic uncertainty in the hazard estimation. The displacement-based
design criterion offered by this format is similar to that of the DCFD with aleatory uncertainties
except for an additional factor that reflects the level of confidence in the estimation of the limit
state frequency. The DCFD format can be used for both designing a building with a certain level
of confidence and also determining the level of confidence associated to an existing design. This
format is implemented in the guidelines for the performance evaluation of existing and
earthquake damaged buildings in FEMA 351.
As a final note, it should be mentioned that there are numerous ways to transform the
probabilistic design criterion into design criteria that are suitable for code implementation. This
chapter only manages to discuss the most commonly-used of these formats. Nevertheless, the
fundamentals used for deriving these formats can be applied towards developing new alternative
formats.
Chapter 4
Alternative Nonlinear Demand Estimation Methods for
Probability-Based Seismic Assessment
Jalayer, F. and Cornell, C. A. (2003 a). Earthquake Spectra, (in preparation).
4.1 Abstract
Demand and capacity factored design (DCFD) is a probability-based LRFD-like format used for
seismic reliability assessment of structures. This format makes use of the results of nonlinear
dynamic analysis in order to make probabilistic statements about the seismic reliability of a
structure. This chapter discusses alternative methods for designing a program of nonlinear
analyses and for applying the results of dynamic analysis in the framework of this demand and
capacity factored design format, particularly as it relates to displacement-based demand
estimation. These alternative methods are demonstrated through a comprehensive case study of
an older reinforced concrete frame structure in Los Angeles. The dynamic analyses are performed
on a set of ground motion records selected from a catalog of California events recorded on stiff
soil. The spectral acceleration at the first mode period has been chosen as the representative
ground motion intensity measure for the analysis, although the proposed methods are general with
respect to the choice of the intensity measure. The methods are classified into narrow-range and
wide-range methods, based on the range of spectral acceleration and displacement values for
which they provide demand estimations. Single-stripe analysis and cloud analysis are two
narrow-range methods discussed in this chapter. Wide-range methods employ rather extensive
analysis efforts in order to provide estimations of displacement-based demand over a wider range
of spectral accelerations. Multiple-stripe analysis and incremental dynamic analysis are examples
of wide range methods. This chapter briefly addresses how incremental dynamic analysis method
can also be used for displacement-based collapse limit state (capacity) estimation. The estimated
displacement demand and capacity are compared -in the framework of DCFD in order to check
the safety of the structure for the limit state of collapse. The uncertainty in the estimation of
median displacement demand due to limited sample size, i.e., number of records, is also
discussed. The procedure stated here can be used in practical applications: they are fully
consistent with, for example, the FEMA/SAC (see FEMA 350-352) Guidelines although they
address only the aleatory component of the demand factor.
4.2 Keywords
Probability-based seismic assessment, reinforced concrete frame structures, displacement-based

demand, displacement-based capacity, ground motion intensity measure, spectral acceleration,
seismic hazard, randomness, uncertainty.
4.3 Introduction
This chapter discusses alternative methods for the estimation of nonlinear dynamic structural
displacement demand in the context of probabilistic seismic assessment. Demand estimation is an
essential part of the demand and capacity factored design format for seismic assessment of
structures. It is presumed here that a set of nonlinear dynamic analyses will be employed in order
to make this estimation. One of the objectives here is efficient demand estimation, namely,
finding procedures that estimate demand with acceptable accuracy and with minimum amount of
analysis effort. Most of these methods are based on a simplified closed-form expression for
demand and capacity factored design developed in Chapter 3 (Equation 3-12), which in turn
stems from a closed-form derivation of annual frequency of exceeding limit state capacity
developed in Chapter 2 (Equation 2-38).
The following expression represents the collective efforts leading to the calculation of the annual
frequency of exceeding a structural limit state capacity, LS 11:
LS = [capacity < demand ] =

P[capacity < demand | demand = y] P[demand = y | S
y x
a = x ] [ S a = x ] (4 - 1)
The first factor within the sum is the probability that the random capacity is less than some
displacement value y, the second factor is the probability that the displacement demand value is
equal to y for a ground motion intensity level (e.g., here the first mode spectral acceleration, S a )
11
Note that in the earlier chapters in this thesis we used the notation H (.) to refer to the annual frequency
of exceedance, whereas, in this chapter and also the ones to follow we are going to use the notation
(.) instead.
equal to x and the last factor is the annual frequency that the spectral acceleration is equal to x.
The limit state probability is obtained by summing the product of these three factors over all the
possible values of x and y. A corresponding integral representation will follow below.
Assessment of the three factors in the above integrand (Equation 4-1) can be classified into
capacity assessment, demand assessment and hazard assessment categories, respectively. This
chapter focuses on the techniques that are available for demand assessment, that is, the estimation
of probability distribution of demand for a given spectral acceleration.
Several design and assessment formats have been developed based on various manifestations of
the expression for annual frequency of exceeding the limit state capacity (Equation 4-1). The
formats, as such, enable the designer to design new structures and also to assess the seismic
performance of existing structures. One such format is the Demand and Capacity Factored Design
(DCFD) format developed in the previous chapter, analogous to typical equations for Load and
Resistance Factored Design (LRFD) specifications (see Galambos and Ravindra, 1978). This
format is expressed in the form, R L , where the factors and are applied to nominal
resistance, R, and load, L. The Demand and Capacity Factored Design (DCFD) Format is based
on the following representation of the expression for annual frequency of exceeding the limit state
capacity:
median capacity capacity factor median demand demand factor (4 - 2)
Where the median demand is associated with a spectral acceleration that has a mean annual
frequency of exceedance equal to P0 , an acceptable level for the limit state frequency (Figure
3-3). As it will be shown below, the demand factor in Equation 4-2 depends on a (second
moment) statistical parameter of the demand and also on a parameter reflecting the sensitivity of
demand to changes in the hazard level.
Once the probabilistic distribution parameters for demand (given spectral acceleration) are
estimated, one can calculate the Factored Demand, defined as the product of median demand and
the demand factor:
Factored Demand = median demand demand factor (4 - 3)

The formal equation for the factored demand will be discussed later in this chapter.
This chapter presents alternative methods for evaluating the factored demand. These methods will
be introduced in the order of the amount of analysis effort they involve.
As it is mentioned above, the median demand term in Equation 4-3 in fact a conditional median
for a given spectral acceleration level (Equation 3-13). Hence, the methods presented in this
chapter are classified based on the range of the spectral acceleration values they cover, in their
estimation of factored demand. The narrow-range methods estimate the factored demand over a
narrow range of spectral acceleration values, e.g., around the spectral acceleration corresponding
to the hazard level, P0 (based on the formal equation for factored demand Equation 3-12).
Single-stripe analysis and cloud analysis are two narrow-range methods that will be discussed in
this chapter. Narrow-range methods involve a relatively small amount of analysis effort, at the
price of providing only limited information about the factored demand. The wide-Range Methods
tend to involve extensive analysis efforts and estimate the factored demand over a wider range of
spectral acceleration values. These methods also permit limit state assessments over a wider range
of spectral acceleration and limit state probability values. Multiple-stripe analysis and incremental
dynamic analysis are the wide-range methods discussed in this chapter. This chapter will also use
numerical integration to test the validity the analytical assumptions that form the basis of factored
demand estimation.
4.4 A Comprehensive Assessment Example: An older reinforced concrete frame
We have selected an older reinforced concrete frame structure in Los Angeles, as a

comprehensive case study in probabilistic assessment of an earthquake-damaged existing
structure. This structure is serving as a test-bed for the activities of the PEER center. This chapter
is also an effort in summarizing and classifying the results of our analyses on a model frame from
this building.
For our probabilistic assessment purposes, the accuracy of the mathematical modeling of this
particular building was not imperative. Nevertheless, we were interested in analysis software that
would enable us to model the challenging issues of cyclic stiffness and strength degradation
behavior in reinforced concrete. We employed DRAIN2D-UW that is a modified version of
DRAIN2D that was produced by Professor Jose Pincheiras research team in the University of
Wisconsin (see Pincheira et al., 1999 and Dotiwala et al., 1998).
We modeled one of the transverse frames using DRAIN2D-UW. Figure 4-1 illustrates the
mathematical model of the frame:
105

105

105

105

105

106

157

241 241 241
Figure 4-1 A 7-story frame structure in LA: The modeling characteristics of a transverse frame
The members are modeled using a beam-column element. The degrading behavior is concentrated
in two rotational springs at the two ends (flexure) and a translation spring somewhere in the
middle of the element (shear). The hysteretic response of a beam-column element modeled by
DRAIN2D-UW is shown in Figure 4-2. The degrading behavior observed in the plot reflects the
degradation in both shear and flexure.
Figure 4-2 - Hysteretic response of a beam-column to cyclic loading.
The beam-column is modeled using DRAIN2D-UW
Onset of significant structural damage
Figure 4-3 - The Static Pushover Curve for the Structural Model.
The static pushover curve for the structure is plotted in Figure 4-3. Note that it is plotted with
respect, not to the customary roof rift, but to the response measure used here, maximum (over all
stories) of the inter-story drift angle. The onset of significant structural damage (inter-story drift
of 0.0075) in the frame is marked on the curve. It should be noted that the hysteretic model does
not take into account a mode of axial of failure prompted by large shear forces in the columns,
which is realistically expected to happen at drift values between 1-2% (see Moehle et al., 1994,
Elwood and Moehle, 2002). Therefore, the analysis results in this chapter should be considered
as, conditioned on no other failure modes. Meanwhile, consideration of other potential modes
of failure would be a challenging test of the accuracy of the methods presented in this chapter.
4.5 Record Selection
For this exercise a set of 30 ground motion records were selected from the Silva Catalog (Silva,
1998) for California Sites. These records were all California events recorded on stiff soil (Geo-
Matrix soils types C and D) and were selected from a moment magnitude- (closest) source-to-site
distance range of:
5.5 < M < 7.5

15 < R < 120 km
Addressing the current issues in ground motion selection is not the focus of this study. It is
generally presumed that one should select records representative of the events that dominate the
probabilistic hazard for the ground motion levels of interest, as determined informally or
formally, e.g., by disaggregation (see McGuire, 1995 and Bazzurro and Cornell,1998). This
magnitude range, 5.5 M 7.5 , is presumed to be representative of the events likely to cause
severe ground motions at this site. With the exception of potential near-source and directivity-
influenced records, the most important record characteristics (e.g., spectral shape), other than
general amplitude level, are comparatively insensitive to distance. In most cases in this study the
records will be scaled relative to their recorded values; therefore the recorded amplitude is not
directly relevant. Directivity issues (e.g., pulse-like records) are beyond the scope of this
particular study (see, e.g. Alavi and Krawinkler, 2000; Luco and Cornell, 2001). Issues of
nonlinear response sensitivity to magnitude and distance and record scaling are discussed by
Shome et al. (e.g., Shome, et al., 1998; Shome and Cornell, 1999). Nonetheless the subject of
site-specific record selection and modification for nonlinear demand estimation deserves, and is
the subject of, further research.
4.6 Probabilistic Framework for Design and Assessment of Structures
Presented below is the continuous version of Equation 4-1 (and also Equation 2-2), with
maximum inter-story drift angle denoted by max , as the displacement-based demand.. This
equation is based on the assumption that capacity C LS is statistically independent of S a and max
(see Section 2.5.7.4):
d S a ( x)
LS =
y x
FC LS ( y ) f max |S a ( y | x)
dx
dx (4 - 4 - a)
Where:
LS Mean annual frequency of exceeding a limit state, LS

C LS Capacity for the limit state, LS in displacement terms
max Demand in displacement terms
Sa Intensity measure; here it is the first mode spectral acceleration
FC LS Cumulative probability density function of capacity
f max |S a Probability distribution function of the demand for a given level of S a
Sa Spectral acceleration hazard, in terms of the mean annual frequency of
exceedance
x A given value of S a
y A given value of max
While Equation 4-1 is very general with respect to the choice of ground motion intensity measure
and demand variable, here, we have chosen to use the maximum inter-story drift angle within the
structure, max , as the demand variable in the design/assessment process. This is a particularly
suitable choice for moment resisting frame structures, since it relates the global response of the
structure quite directly to joint rotations where most of the inelastic behavior in the moment
resisting frames is concentrated. We have used the spectral acceleration with the period of the
fundamental mode of the structure, S a , as the intensity measure variable. Studies by Shome and
Cornell (e.g., Shome, et al., 1998; Shome and Cornell, 1999) show that for moment-resisting
frame structures with their first-mode period lying within the moderate range (e.g., around T=1.0
sec), the spectral acceleration of first mode is sufficient for relaying the primary ground motion
characteristics to the structural response.
4.7 Drift Hazard - the Annual Frequency of Exceeding Drift Demand
Equation 4-4-a can be re-written in the following manner:
d S a ( x) d max ( y )
LS =
y x
FC LS ( y ) f max |S a ( y | x)
dx
dx =
y
FC LS ( y )
dy
dy (4 - 4 - b)
where max ( y ) is the (mean) annual frequency that drift demand, max , exceeds the value, y.
max ( y ) can be calculated, using the same technique as the one used for the limit state frequency,
from the expression below:
d S a ( x)
max ( y ) = G
x
max | S a
( y | x)
dx
dx (4 - 5 - a)
Where G max | S a ( y | x) is the conditional complementary cumulative density function (CCDF) of
demand, max , for a given S a value. We can derive the expression for d max ( y ) dy in Equation
4-4-b by taking the derivatives of both sides of Equation 4-5-a with respect to y:
d max ( y ) d S a ( x)
dy
= f
x
max | S a
( y | x)
dx
dx (4 - 5 - b)
Where f max |S a ( y | x) is the conditional probability density function (PDF) of demand, max ,
for a given Sa value. The above derivation is based on the relationship between the PDF
and the CCDF of a probability distribution (Equation 2-12). There is a complete discussion
of drift hazard and its derivation in Chapter 2.
4.8 Demand and Capacity Factored Design Format
In order to derive a (codifiable) limit state safety checking format, we set the mean annual
frequency of exceeding the limit state, LS , in Equation 4-4 to less than or equal to a selected
tolerable level, P0 . The following expression in Equation 4-6 is an explicit presentation of
Equation 4-2, obtained upon making a series assumptions as discussed below. The derivation of
Equation 4-6 is explained in detail in Chapter 3.
1k 2 1 k 2
C LS max | S a
CLS e 2b P0 e2 b (4 - 6)
max | S a
Where:
max |S a Median demand max for a given spectral acceleration, S a (here in terms of the
maximum inter-story drift angle)

C LS Median capacity (here in terms of the maximum inter-story drift angle)
max | S a
Dispersion measure for demand given the spectral acceleration level, S a
C LS Dispersion measure for capacity
P0 Tolerable frequency of exceedance of the limit state

P0
Sa The spectral acceleration associated with the annual frequency of exceedance
equal to P0
P0
P0 Median value of max given a spectral acceleration equal to Sa
max | S a
k The (log) slope of the hazard curve if it is approximated by a power-law function

b The (log) slope of the median displacement-demand versus spectral acceleration
curve it is approximated by a power-law function
k /b A sensitivity factor reflecting the change in the probability with respect to the
change in the displacement-based demand
1 k 2
max | S a
| P0 S
e2 b :
max a
Factored demand for an allowable (mean) annual frequency of exceeding the

limit state capacity equal to P0 .
The exponential factors in Equation 4-6 represent the capacity and demand factors in Equation
4-2. A complete discussion of the assumptions made in the derivation of Equation 4-6 can be
found in Chapter 2. Briefly, probability density functions for limit state capacity and the demand
given spectral acceleration are assumed to be lognormal. The lognormal density function is
defined by the median (the term) and by the standard deviation of the natural logarithm
(denoted ), called here the dispersion measure or the fractional standard deviation. This
measure is approximately equal numerically to the coefficient of variation (COV), i.e., for values
of COV less than 0.2, one can assume that COV 2 2 with less than 2% error in 2 (Benjamin
and Cornell, 1970). These parameters will be estimated from the data in the examples to follow.
Thus, the probability distribution for the capacity is defined by LN ( C LS , C LS ) and the
probability distribution for demand given spectral acceleration is defined by LN ( max |Sa , max |Sa ) .
In order to carry out analytically the integrals presented above, and subsequently to derive
Equation 4-6, it is presumed that spectral acceleration hazard curve and median demand-spectral
acceleration relationship can be predicted, at least over a local region sufficiently wide to yield
reasonable numerical accuracy, by power-law type expressions in the form of S a ( x) = k 0 x k
and max |S a = a S a b , respectively. It should be noted that k and b are the slopes of ln S a versus
ln S a and ln max versus ln S a relationships.
We now focus our attention on the terms that constitute the expression for factored demand
(based on Equations. 4-3 and 4-6):
1 k 2
max | S a
Factored Demand = P0 e2 b (4 - 7)
max | S a
Verbally, the factored demand is equal to the product of the median demand, P0 , for the
max | S a
P0
motions within the given spectral acceleration level S a and the demand factor or the
exponential term, exp( 1 k 2 max |S a ) , which is a function of the sensitivity factor, k / b , and the
2 b
standard deviation of (the logarithm of) the demand for a given spectral acceleration (dispersion),
max |S a . Hence, the evaluation of factored demand consists of finding ways to estimate Po ,
max | S a
max |S a and k / b . Since the dispersion measure and the sensitivity factor are both in the power of
the exponential, they assume secondary roles as compared to the median.
Based on the same assumptions as discussed above, the drift hazard is derived as (the derivation
of the following analytic form for the drift hazard can be found in Section 2.5.7.3):
1 k2 2 1 k2 2
max | S a max | S a
max ( x) = S a ( S ax ) e 2 b2
= S a (max
1
| S a ( x )) e
2 b2
(4 - 8)
Where S ax is denoted as the spectral acceleration that corresponds to a drift equal to x. More
precisely, S ax is the inverse of x = x = a ( S ax ) b , i.e., S ax = 1 x = ( x a)1 b . This is the

max | S a max | S a
spectral acceleration value that has (approximately) a 50% chance of causing a drift response
equal to or greater than x. S a ( S ax ) is the spectral acceleration hazard value (the mean frequency
of exceeding) for S ax . It is interesting that the drift hazard in Equation 4-8 can be calculated as the
spectral acceleration hazard at a spectral acceleration value, S ax , times an exponential factor,
1 k2
exp( 2 2 max |Sa ) , closely resembling the one for the factored demand, both are functions of the
2 b
sensitivity factor, k / b , and the dispersion in demand for a given spectral acceleration, max |S a .
In order to calculate the drift hazard from Equation 4-8, we need to estimate max |S a , max |S a and
k / b and the spectral acceleration hazard curve S a for all spectral acceleration levels. Thus, the
alternative methods discussed in this chapter achieve a two-fold objective of estimating the
factored demand as well as the drift hazard curve.
Recalling from Section 3.2.3.5 in Chapter 3, factored demand happens to be the drift value that
corresponds to the hazard level P0 on the drift hazard curve. We will see later in this chapter that
this particular relationship is going to be very useful for the calculation of factored demand in
cases where there are no closed form solutions available for factored demand. One such example
is when numerical integration is used to calculate the integrals in Equations 4-4 and 4-5 for drift
hazard and the annual limit state frequency.
The methods discussed in this chapter estimate the three quantities above in alternative ways.
Being a characteristic of the hazard rather than the structural response, the k value is estimated in
a similar manner in all the methods. Hence, it suffices for each method to estimate Po ,
max | S a
max |S a and b .
4.9 Site Specific Hazard Curve: Estimation of k Value
The k value can be estimated from a site-specific or mapped regional basis spectral acceleration
hazard curve, provided by earth scientists. A spectral acceleration hazard curve provides the mean
annual frequency of exceeding a particular spectral acceleration for a given period and damping
ratio. In order to estimate k, we fit a power-law type of expression to the hazard curve in the
region of spectral acceleration values of interest as proposed by for example by development
underlying Department of Energy Standard 1020 (see Kennedy and Short, 1994, and Luco and
Cornell, 1998):
S a ( s a ) P[ S a s a ] = k 0 s a k (4 - 9)
Where k 0 and k are the fit parameters.
Figure 4-4 shows a firm soil site-specific hazard curve calculated for Van Nuys, CA, at T = 0.85
sec, T being the first natural period of the structure, plotted in a two-way logarithmic scale. The
figure also shows the line fitted to the hazard curve in the region of interest, i.e., a power-law
approximation in the arithmetic scale calculated around an annual frequency of about 1/1000. The
(absolute value of) slope of the fitted line is the k value in the region of interest.
4.10 Evaluation of Factored Demand by Numerical Integration
So far, we have talked about approximate methods for demand estimation. Inevitably, the
question arises as to how we can get a true estimate of the factored demand.
As mentioned before, the closed form solution for the factored demand is valid as long as the
underlying assumptions hold. However, we can evaluate the integrals in Equations 4-4 and 4-
5 by numerical integration even when those assumptions stop to be valid. Here, we have
calculated the expression for drift hazard (Equation 4-5) by numerical integration and have
plotted it versus the spectral acceleration in Figure 4-4.
k= 2.6
Figure 4-4 -Site Specific Hazard Curve, Van Nuys CA . Approximating the Hazard curve
with a line in the region of interest.
We estimated the conditional complementary cumulative density function of demand for a given
spectral acceleration, G max |S a ( y | x) in Equation 4-5-a, by the ratio of the demand values that
exceed the value, y, for a spectral acceleration equal to x (also known as Empirical Distribution,
see Efron and Tibshirani, 1993). This is also an approximation in that the sample size is limited.
The spectral acceleration hazard curve used is that in Figure 4-4, which is obtained by performing
conventional probabilistic seismic hazard analysis (PHSA).
As we have mentioned previously, the factored demand for a tolerable probability level, P0 , can
be calculated from the drift hazard curve. Therefore, we are going to find the factored demand
using the drift hazard calculated by numerical integration.
0
Annual Frequency of Exceeding Maximum Interstory Drift
10
1
10
max

120 Years Return Period

P0= 0.0084
2
10
F . D. = 0. 02
3
10
3 2 1
10 10 10

max
Figure 4-5 The drift hazard curve derived by numerical integration. The factored demand
corresponding to P0 = 0.0084 is also shown on the plot.
As illustrated in Figure 4-5, we can use this curve to estimate the factored demand for a tolerable
probability level, P0 = 0.0084 . This is a probability level that we believe is associated with the
spectral acceleration levels that drive the structure to the verge of collapse (this particular
probability level has been chosen merely for demonstration purposes). Therefore it is a severe test
of the analytical approximation, permitting us to demonstrate its limitations under certain
approaches. Later in the chapter, we are going to introduce another P0 level that is associated
with the onset of damage in structural members, P0 = 0.03 . At this level the analytical
approximations will be found to be quite robust. In order to be consistent, we have used the same
two values for P0 throughout the chapter.
For a tolerable probability level of P0 = 0.0084 , the drift hazard curve yields a factored demand
of, 0.02 (Figure 4-5), which corresponds to a ductility of 2 relative to the onset of significant
structural damage.
In this chapter, we are going to refer to this quantity as the true factored demand for the
tolerable probability P0 = 0.0084 . We are going to use this true factored demand as a measure
for gauging each methods accuracy in factored demand estimation. Recall that the true value is
in itself an estimate because a sample size of only 30 records has been used to estimate G max |S a
in Equation 4-5-a, and the hazard curve itself is an estimate.
4.11 Narrow Range Methods
Narrow range methods estimate the factored demand over a comparatively small neighborhood of
spectral acceleration values. They require a relatively small number of analysis runs, at least
compared to the methods that estimate the factored demand over wide ranges of S a values.
However, these methods have a limited scope of application. They are especially effective in
estimating probabilistic demand parameters that appear in the closed-form expression for the
DCFD design format, e.g., Equation 4-6, over a targeted range of probability, ground motions,
and displacements of interest.
4.11.1 Single-Stripe Analysis
The single-stripe analysis, as its name suggests, is a procedure that involves a structural dynamic
analyses for a set of records scaled to a common spectral acceleration value.
The output of the single-stripe analysis, which is scattered along a stripe of constant spectral
acceleration value, provides estimates for the median and dispersion of the inter-story drift
demand conditioned on the spectral acceleration level of the stripe. The first question is what the
appropriate spectral acceleration level for single-stripe analysis is. The answer depends on the
criteria that are going to be satisfied.
The results of the single-stripe analysis are particularly well suited for estimating the factored
P0
demand, which is expressed as a function of S a or the spectral acceleration corresponding to the
P0
tolerable probability level P0 . This makes S a a potential candidate for the spectral acceleration
P0
level to which the ground motion records are scaled. Figure 4-6 illustrates how S a is calculated
for an acceptable probability level of P0 =0.0084. The probability level, P0, is marked on the
hazard axis and its corresponding spectral acceleration value, P0 S a , is calculated by finding the
S a value that corresponds to a hazard level, P0 , via the hazard curve.
k= 2.6
P0= 0 . 0 0 8 4
P0
Sa=0.70 g
Figure 4-6 - The Spectral Acceleration Corresponding to an acceptable

Probability Level of 0.0034
P0
After choosing the spectral acceleration level of interest, in this case, S a = 0.70 g , the ground
motion records are all scaled to this value. The scaled ground motions are then applied to the
structure and the maximum inter-story drift demand values are calculated.
Single Stripe Analysis Holiday Inn Van Nuys
0
10
"First Mode" S (T=0.8,=0.05) [g] s

18
16
24 145 30
20 13
128
11
2 26 717 3238 922
12
6 29
4
10
21
15 25 19 27
P0
Sa= 0.70 g
a
1
10
2 1
10 10
Maximum Interstory Drift Angle
max
Figure 4-7 - Stripe response obtained by applying the selection of ground motion records scaled
P0
to the spectral acceleration of the stripe, S a = 0.70 g , to the model structure.
The stripe response of the model frame in Figure 4-1, subjected to the selection of ground
motion records, is plotted against spectral acceleration in the logarithmic scale (Figure 4-7).
The statistical parameters of the stripe response can now be used to estimate the median and
P0
dispersion at the spectral acceleration level, S a . Here we use the sample parameters denoted by
max |S a and max |S a called, respectively, the counted median and the counted standard
deviation of the stripe response for a given spectral acceleration level. In order to obtain the
counted statistical parameters of a data set, the data is first sorted in the ascending order. The
counted median is the 50th percentile of the sorted data. The counted standard deviation is
estimated by the average of ln( 84th 50th ) and ln( 50th 16th ) respectively. The counted median
P0
and standard deviation for a spectral acceleration level, S a = 0.70 g , are calculated as 0.0183
and 0.49, respectively. These quantities are shown in Figure 4-8 (again in the logarithmic scale).
We can observe that the displacement-based response is well into the strength degrading regime,
explaining the comparatively large value max |S a .
max |S a =0.70 = 0.49

0
10
"First Mode" S (T=0.8,=0.05) [g] s

18
16
24 145 30
20 13
128
11
2 26 717 3238 922
12
6 29
4
10
21
15 25 19 27
P0
Sa=0.70 g
max|Sa =0.70 = 0.0183
a
1
10
2 1
10 10
Maximum Interstory Drift Angle max
Figure 4-8 - Estimated Statistical Parameters for the Stripe Response
Once the statistical properties of the drift stripe response are calculated, they can be used as
estimates of D|S a and D|S a (the method of moments can often be used as well, but difficulties
will ensue if one encounters dynamic runs in which numerical convergence is not obtained,
and/or if very large, unrealistic displacements are computed.). The single stripe output does not
provide displacement versus spectral acceleration slope information, i.e., the b parameter
appearing in the expression for factored demand (Equation 4-7). As a preliminary estimate, we
assume that the b value is equal to 1, implying that D|S a and Sa have a proportional relationship
in arithmetic space. Elastic behavior and the equal displacement rule (see Veletsos and
Newmark, 1960) are special cases of this condition. The factored demand estimate using the
estimates of D|S a and D|S a obtained for single-stripe analysis method, is:
1 k 2 1 2.6
max | S a ( 0.49 ) 2
F .D. = P0 e2 b = 0.0183 e 2 1 = 0.0183 1.366 = 0.025
max | S a
We can see that the single-stripe prediction for factored demand is equal to 0.025, which is much
larger than the true factored demand, F.D.=0.02, in the previous section. In this case, because
D|S is relatively large, the demand factor is sensitive to the value of b and the single-stripe
a
method with a default value of b = 1 produces a very conservative estimate of the factored
demand. This brings up the question of how much extra analysis effort is required in order to
enhance the results of single-stripe analysis method.
Next, we consider the efficiency of a double-stripe method, because the minimum number of
stripes necessary for getting information about b value is two. As its name suggests the double-
stripe method consists of two separate single-stripe analyses, the original single stripe analysis
P0
plus an additional stripe somewhat close to the original stripe at S a = 0.70 g . In this example we
P0
place the second stripe at S a = 0.80 g . The b value is estimated as the slope of the line on a log-
log plot, that connects the medians of the two stripes:
max| S
a = 0.8
ln( )
max| S
b= a = 0.8
= 3.6
0.8
ln( )
0.7
It should be noted that the estimates for D|S a and D|S a are the same as those obtained for the
single stripe analysis above, i.e., the second stripe plays an auxiliary role for the purpose of
estimating the b value. The results of a double-stripe method are plotted in Figure 4-9. The
factored demand in this case is equal to:
1 2.6
( 0.49 ) 2
F .D. = 0.0183 e 2 3.6 = 0.0183 1.09 = 0.020
The answer is very close to the true factored demand estimated by numerical integration. We
should however note that the position of the second stripe, with respect to the original one- plays
a critical role in estimation of the b value. If the second stripe is too far, the estimated b value
may not be representative of the local slope around the original stripe. If it is too close, the
estimated b value may fail to represent the general trend in spectral acceleration versus demand
curve. Thus, the accuracy of a double-stripe analysis is dependant on the analysts judgment in
choosing the spacing between the stripes. A suggestion is to choose the second S a stripe to be
1 1
above the original value by a fraction (of that value) equal to or of D|S a (i.e.,
4 2
S a2 = S a1 + fraction D|S 1 S a1 ).
a
0
10
"First Mode" S (T=0.8,=0.05) [g] s
16
Sa=0.80 24 145 28
30
20 11132617 78
2 239 29
10
3 27
122
12
6425
21 15
18
19
24 14
5 30
20 13
128
11
2 26 717 323
8 922
12
6 29
410
21
1525 19 27
P0 16
Sa=0.70
18
b=3.6
a
1
10
3 2 1 0
10 10 10 10
max
Figure 4-9 - Estimation of factored demand using two stripes, max |S a =0.80 = 0.03
4.11.2 Cloud Analysis
The cloud analysis is a procedure in which a structure is subjected to a set of ground motion
records of different first-period S a values. In contrast to the single-stripe method, this method
provides a cloud rather than a stripe of response values as shown in Figure 4-10.
This cloud response method provides estimates of the two statistical parameters of demand
given the spectral acceleration, max |S a and max |S a , as well as the b value. In its simplest form,
the cloud response is obtained by applying original ground motions (as recorded) to the
structure. If the ground motion records are taken from a bin (like the record selection in this
chapter), they can represent an earthquake scenario defined by (Mbin, Rbin), the magnitude and
distance representative of the bin.
Once the ground motion records are selected, they are applied to the structure and the resulting
maximum inter-story drift angles are calculated. This provides a set of drift values that are the
basis for cloud-method calculations. The response of the model frame when subjected to the set
of (as recorded) ground motions is plotted in Figure 4-10.
Single Cloud Analysis Holiday Inn Van Nuys
0 17
10
24
30
5 28
20 11
25
"First Mode" Sa (T=0.80, =0.05) [g] s

14
10
12
13
829 7
19
2 18
1
23
9
26
22 6
15
21
1 16
10
3
27
2
10
4 3 2
10 10 10
Figure 4-10 The Cloud Response of the Structure Subjected to a set of Ground Motion
Records (plotted in the Logarithmic Scale).
In order to estimate the statistical properties of the cloud response, conventional linear
regression (using least squares) is applied to the cloud response in the natural
logarithmic scale, the standard basis for the underlying lognormal distribution model.
This is equivalent to fitting a power-law curve of the form, a S a b , to the cloud
response in the original (arithmetic) scale. This results in a curve that predicts the median
drift demand for a given level of structural acceleration:

D|Sa (sa ) = a sa b (4 - 10a) or
ln( D|Sa (sa )) = ln (a) + b ln (sa) (4 - 10b)
where ln(a ) and b are linear regression constants. Detailed information about the
estimation of the linear regression constants can be found in any applied statistics
reference, such as, Benjamin and Cornell, 1970, or Rice, 1995. The results of such
prediction can be used to estimate the median demand for a given spectral acceleration,
P0
D|S , for a specified spectral acceleration equal to
a
Sa . The standard deviation of
regression measuring the second moment of data points around the predicted curve can be used to
estimate the dispersion measure, D|S a :
D|Sa =
(ln(d ) ln(a s
i a,i
b 2
))
(4 -11)
n2
Where d i and s a ,i are the drift demand and the corresponding spectral acceleration for record
number i within the cloud response set and n is the number of records.
The standard deviation of regression, D|S a , as introduced in Equation (4-11), is presumed
constant with respect to spectral acceleration over the range of spectral accelerations in the cloud.
We will see later in this chapter that the dispersion measure, D|S a , tends to increase for the larger
values of spectral acceleration. Thus, assuming D|S a is constant, may be un-conservative in
certain cases. This stresses the importance of performing linear regressions locally, i.e., in a
region of spectral acceleration values of interest.
Figure 4-11 shows the median, D|S a , the dispersion, D|S a , and the b value, estimated (by fitting
the cloud response) by linear regression. We can estimate the b value by the slope of the fitted
line in the logarithmic scale. The estimated b value here is in fact somewhat less than one, a
phenomena commonly observed in the mild ductility region for moderate-period structures.
0 a = 0.015 P0 17
10
b = 0.8811
= 0.2679
Sa=0.70 24
30
ln( )|S 5 28
max a 20 11
14 25
"First Mode" Sa (T=0.80, =0.05) [g] s

10
12
13
829 7
19
2 18
1
23
26
9
max|Sa =0.70 = 0.27
22 6
15
21
1 16
10
3
b= 0.88
4
1
27
max |Sa =0.70 = 0.011
2
10
4 3 2
10 10 10
Figure 4-11 The statistical properties of the cloud response. The parameter estimates for the
median, D|S a , and dispersion measure, D|S a , are shown on the figure.
The estimated D|S a , D|S a and b values are finally used in to calculate the factored demand
P0
corresponding to P0 = 0.0084 or S a = 0.70 g :
1 k 2 1 2.6
max | S a ( 0.28) 2
F .D. = P0 e2 b = 0.011 e 2 0.88 = 0.011 1.12 = 0.012
max | S a
Recalling from the previous sections, the true factored demand was calculated as 0.02, which is
much larger than the prediction given by cloud analysis, 0.012. Thus, the cloud analysis is
underestimating the drift demand in this case. This is can be due to the fact that the ground
motion records used for the cloud analysis are not strong enough to properly represent the
P0
displacement-based response for spectral accelerations close to S a = 0.70 g .
4.11.3 Cloud Analysis Scaled
It is not easy to (a priori) find a selection of records that is representative of the displacement-
based response in the vicinity of a given spectral acceleration level. for the cloud analysis
method. Sometimes the records are not strong enough to represent the structural response in the
region of displacement demands that we are interested in. In this example, we are interested in
ground motions with a first-mode spectral acceleration around 0.70g, yet the set of ground motion
records, as recorded, have a median S a of only 0.27g. The median drift response, about 0.0045,
corresponds to a ductility of only one with respect to the onset of significant structural damage. In
order to get a more accurate estimate of the nonlinear response in the region of our interest, we
have scaled the same collection of records by a factor of 2. We anticipate that the scaled cloud
will more successfully capture the structural response in the range of S a values around
P0
S a = 0.70 g .
17
24
30
5 28
11 20
0 14 25
10
10
12
"First Mode" Sa (T=0.8, =0.05) [g] s
13 29
78
19
2 18
1
23
9
26
22 6
15
21
16
1
10
27
2
10
3 2 1
10 10 10
Figure 4-12 The scaled cloud response of the structure subjected to the selection of records,
scaled by a factor of two.
As we can see in Figure 4-12, scaling has moved the cloud further up and to the right. Now that
P0
that the cloud response has become nearly centered around S a = 0.70 g , we expect the
statistical information to be more representative of the demand level of interest we are estimating.
The estimated statistical parameters for demand given spectral acceleration are plotted in Figure
4-13. As before, we have used linear regression to estimate the statistical properties of demand
for a given spectral acceleration.
Single Cloud Analysis Holiday Inn Van Nuys Scaled by 2

17
24
30
5 28
11 20
0
10
P0 14 25
Sa=0.70 12
10
"First Mode" Sa (T=0.8, =0.05) [g] s
13 29
78
19
2 18
1
23
9
26
22 6
15
21
16
1
10
27
max | S a = 0.70 = 0.0166

max|Sa=0.94 =0.016
a = 0.0249
b = 1.131
ln(
= 0.5697
)|S
max a
2
10
3 2 1
10 10 10
max
Figure 4-13 The statistical properties of the scaled cloud response.
The estimated D|S a , D|S a and b values obtained from scaled cloud analysis - are incorporated
below in the formula for factored demand:
1 k 2 1 2.6
max | S a ( 0.57 ) 2
F .D. = P0 e2 b = 0.0166 e 2 1.13 = 0.0166 1.45 = 0.024
max | S a
We see that in this case the scaled cloud method is still not capable of providing a highly accurate
estimate for the factored demand for this S a level. The factored demand estimate provided by
this method is equal to 0.024, while the factored demand calculated by numerical integration is
equal to 0.020. It is clear that it has produced adequate estimates of D|S a , D|S a (compared to the
locally more accurate results obtain from the stripe analysis) but its estimate of b remains too low;
given the large value of D|S a the factor is in this case quite sensitive to b. As we shall see, b is
changing in this range of S a values and the regression can only provide a gross estimate.
Nevertheless, this does meet to our preliminary expectation that, since the scaled ground motions
P0
have spectral acceleration values closer to S a = 0.70 g , the scaled cloud would enhance the
factored demand estimation.
4.11.4 Narrow-Range Methods for a Lower Spectral Acceleration

P0
S a = 0.40 g
P0
We learned in the previous sections that, for S a = 0.70 g , the single-stripe analysis over
predicted factored demand by 25%. We also learned that for the same S a level of 0.70g, the
single cloud analysis under predicted factored demand by 40%. Although the two methods are
approximate, we know from closer inspection that their lack of accuracy here has been magnified
P0
by significant nonlinear behavior in S a levels as high as S a = 0.70 g and resulting irregularities
in the response. In the next chapter we shall apply a scheme, introduced by Shome et al., 1998, in
order to take into consideration the effect of having large displacements in the response.
In order to confirm this presumption, we have performed the two narrow-range methods, single-
stripe analysis and single cloud analysis, for a lower spectral acceleration level. Thus, we have
P0
chosen a spectral acceleration equal to S a = 0.40 g , corresponding to a mean frequency level of
P0 = 0.028 or a return period of 33 years from the hazard curve in Figure 4-4. We can estimate
the k value by the slope of the line fitted to the hazard curve in Figure 4-4, in the vicinity of
P0 = 0.028 This exercise yields a k value equal to 1.75.
Using the numerically integrated drift hazard curve in Figure 4-5, we can estimate the true
factored demand, corresponding to a mean frequency level P0 = 0.028 . The so-called true
factored demand is equal to 0.0080, which lies in the region of significant structural damage in
P0
the frame, with a ductility ratio around 1.0 (Figure 4-3). This roughly means that S a = 0.40 g
induces deformation and ductility levels that are approximately 40% of those of
P0
S a = 0.70 g investigated as the primary example.
P0
Single-Stripe Analysis on a Lower Stripe S a = 0.40 g - The results of single-stripe analysis
P0
for a spectral acceleration level equal to S a = 0.40 g are plotted in Figure 4-14. The statistical
parameters for the stripe response are also illustrated on the figure. We have plotted this lower
P0
stripe on the same plot as the higher stripe of S a = 0.70 g presented earlier. Assuming that the b
value is equal 1, and incorporating the estimated k value of 1.75, we have calculated the factored
P0
demand for S a = 0.40 g below:
1 k 2 1 1.75
max | S a ( 0.35) 2
F .D. = P0 e2 b = 0.007 e 2 1 = 0.007 1.113 = 0.0078
max | S a
0
10
"First Mode" S (T=0.8,=0.05) [g] s
18
24 14
5 30
20 13
128
11
2 26 717 323
8 922
12
6 29
4
10
21
15 25 19 27
16
P0
Sa=0.40 max |S a = 0 .40 = 0 .35
13128
24
428 12011
30
23
26
22
7 12
521
29
4 15
3 2510
627
16
17 9 19 18
a
max | S a = 0 . 40 = 0 . 007
1
10
3 2 1
10 10 10
Figure 4-14 Stripe response obtained by applying the selection of ground motion records
P0
scaled to the spectral acceleration of the stripe, S a , to the model Structure.
This is very close to the true factored demand by numerical integration, 0.0080. This confirms
our presumption that the narrow range methods provide more accurate estimates for lower
spectral acceleration and deformation levels. Among other factors the lower levels of D|S a (e.g.
0.35 versus 0.49) make the exponential term closer to unity and less sensitive to b, k and D|S a .
P0
Cloud Analysis for S a = 0.40 g - We can observe that the original cloud response in Figure
P0
4-10 is roughly centered on spectral acceleration, S a = 0.40 g . Thus, we are going to use the
original cloud response in Figure 4-10 in order to estimate the factored demand for a spectral
P0
acceleration of S a = 0.40 g . This provides the following estimates for D|S a =0.40 , D|S a and b:
D|S a =0.40 = 0.0067

D|S a = 0.27 (4 12)
b = 0.88
P0
These are the same b and D|S a values as for the S a = 0.70 g , since the b and D|S a estimates
provided by the single cloud analysis are constant for all spectral accelerations. We have
calculated the factored demand below:
1 k 2 1 1.75
max | S a ( 0.27 ) 2
F .D. = P0 e2 b = 0.0067 e 2 0.88 = 0.0067 1.075 == 0.0072
max | S a
This is within 10% of the true factored demand by numerical integration, F .D. = 0.0080 .
4.12 Wide-Range Methods
Estimation of factored demand parameters over a wide range of spectral acceleration values
requires relatively computation-intensive analysis efforts. This section presents a number of
methods capable of estimating probabilistic demand parameters over such a range of spectral
acceleration values. These methods have a larger scope of application compared to the narrow
range methods. The wide range methods can map out the behavior of the system with respect to
the range of possible future ground motions and a range of limit states probabilities. Thus, they
can be used not just to check whether the failure probability is less than a specific tolerable level,
Po , but further, as will be shown, for the calculation of the limit state probability in cases where
closed form solutions are not applicable.
Multiple-stripe analysis (MSA) is a collection of single-stripe analyses performed at multiple

levels of the spectral acceleration. This will provide estimates of D|S a , D|S a for each stripe as
well as local estimates of the b value. This method is also helpful for calculating the frequencies
of exceeding multiple structural limit states by numerical integration of Equation 4-4-a.
Incremental Dynamic Analysis is in essence the re-compilation of the results of multiple-stripe

analysis into a collection of random entities known as the IDA curves. In other words, each
ground motion record is successively scaled to multiple spectral acceleration levels and the
resulting maximum inter-story drift angles are calculated in each case. The IDA curve connects
the resulting inter-story drift angles corresponding to each ground motion record.
The IDA method also has the capability of extracting probabilistic information regarding the
global instability collapse limit state. This feature will be employed for the evaluation of factored
capacity of the collapse limit state, a step that completes the design/assessment process for this
limit state.
4.12.1 Multiple-Stripe Analysis (MSA)
Scattered along multiple stripes of constant spectral acceleration, the results of multiple-stripe
analysis can provide statistical information about the demand over a wide range of spectral
acceleration values. Multiple stripes of response are obtained by applying to the structure a set
of ground motion records that are scaled to multiple levels of spectral acceleration. The statistical
properties of each stripe are obtained in the same way as the single-stripe analysis.
Figure 4-15-a illustrates the result of multiple-stripe analysis when the model structure is
subjected to the same selection of the ground motion records mentioned before.
a) Plot in Arithmetic Scale b) Plot in Logarithmic Scale
Multiple Stripe Analysis Holiday Inn, Van Nuys Multiple Stripe Analysis Holiday Inn, Van Nuys
2
1.8
Spectral Acceleration of "first" mode, S [g] s
0
10
a
a
1.6
1.4
1.2
0.8
0.6
1
10
0.4
0.2
3 2 1
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 10 10 10
Maximum Interstory Drift Angle, max Maximum Interstory Drift Angle,
max
Figure 4-15 Multiple-Stripe Analysis

(The same selection of records described in the beginning of the chapter)
The lines connecting the (counted) 16th, 50th and 84th percentiles of the stripes are also shown in
Figure 4-15-a. Figure 4-15-b illustrates the same results in the logarithmic scale.
The median demand for a given spectral acceleration; D|S a , can be estimated by the 50th
percentile curve. Also, the dispersion measure, D|S a , can be estimated ( as a function of spectral
acceleration) with the average band width created by 16th, 50th and 84th percentiles in the
logarithmic plot. Finally, the b value can be estimated as the local/tangent slope of the (log-log)
median curve for a given spectral acceleration value.
The multiple-stripe method provides a much more complete picture as to how both the general
trend (median) and the dispersion of the response evolve under gradually increasing ground
motion levels. Between about 0.80g and 1.1g, the median of the response softens rapidly and
the dispersion increases markedly. The large dispersions beyond 1.0g imply that the median
cannot be well determined even with 30 records.
Figure 4-16 illustrates the application of multiple-stripe analysis in the estimation of factored
P0
demand for a spectral acceleration equal to S a = 0.70 g . The factored demand at a spectral
acceleration equal to P0 S a = 0.70 g , can be calculated using the estimates for D|S a , D|S a and b
values as shown in Figure 4-16:
1 k 2 1 2.6
max | S a ( 0.49) 2
F .D. = | P0 S
e2 b = 0.0183 e 2 2.70 = 0.0183 1.12 = 0.0205
max a
Comparing the result with the true factored demand calculated before, F .D. = 0.020 , it is clear
that MSA method estimates the factored demand very closely. The only difference between using
single stripe analysis and MSA here is local estimation of b value in the latter case. This
emphasizes the importance of estimating the slope information, b value, in the region of interest,
at least when D|S a is large, causing the factor to differ significantly from unity. It should be noted
that the estimate for b value here is more local than from the one obtained from the double-
stripe method.
It is possible to obtain accurate results more efficiently by performing MSA first with broader
spacing and then filling in extra stripes where the b value is found to be changing comparatively
rapidly. Also, as will be discussed further below, it would be advantageous to use more records
when the dispersion is large and fewer records when dispersion is small.
Multiple Stripe Analysis Holiday Inn, Van Nuys
Spectral Acceleration of "first" mode, Sa [g] s

0 P0
10
Sa=0.70
b=2.70
max |S a =0.70 = 0.49
= 0.0183
1
10
max|S a =0.70
3 2 1
10 10 10
Maximum Interstory Drift Angle,
max
Figure 4-16 - Evaluation of factored demand by multiple-stripe analysis.
We have used the multiple-stripe analysis and the analytical approximations for factored demand
to get the factored demand for different tolerable limit state frequencies.
The results of the multiple-stripe analysis and the analytical factored demand estimate for
multiple levels of spectral acceleration are plotted in Figure 4-17 below versus the probability
corresponding to these S a levels, yielding an estimate of the drift hazard curve. The figure also
illustrates the true factored demand as it is obtained by numerical integration. We can see that
for drift values less than or equal 0.03, the MSA method prediction of factored demand is very
close to the true factored demand calculated by numerical integration.
4.12.2 Incremental Dynamic Analysis
As described above, each IDA curve corresponds to a particular ground motion record and passes
through maximum inter-story drift results obtained for the record scaled to multiple levels of
spectral acceleration. Figure 4-18 illustrates individual IDA curves for two different records.
0
10
Factored demand by MSA

Facatored deman by numerical integration
1
10
max

2
10
3 2 1
10 10 10
max
Figure 4-17 - Factored Demand predicted by Multiple-stripe Analysis (MSA) for Multiple Limit
State Frequency Values, Po , and the Drift Hazard Curve Obtained by Numerical Integration.
As it can be seen from Figure 4-18 below, the IDA curve gives a richer picture of how structure
responds to different records, e.g., the benign nature of some records (Figure 4-18-b), and the
aggressive nature of others (Figure 4-18-a).
a) IDA curve with aggressive behavior b) IDA curve with benign behavior
Incremental Dynamic Analysis Holiday Inn, Van Nuys Incremental Dynamic Analysis Holiday Inn, Van Nuys
2 Loma Prieta Agnews State Hospital, 90 2 Loma Prieta, Waho, 90
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.02 0.04 0.06 0.08 0.1 0.12 0.14
Maximum Interstory Drift Angle, max Maximum Interstory Drift Angle, max
Figure 4-18- Individual IDA Curves
One can use curve fitting routines such as the spline fit in order to get smooth IDA curves (see
Vamvatsikos and Cornell, 2002). With the IDAs, the percentiles of response can now be
calculated as continuous curves, e.g., the median IDA curve or the 84th IDA curve (Figure 4-20).
In contrast to MSA method which is constructed by a collection of spectral acceleration stripes,
the IDA curve can be constructed by scaling the records to arbitrary spectral acceleration values.
This feature is useful when one is exploring different choices of Intensity Measures (IM) based
on the same analysis runs. On the other hand, one can use different record sets for different levels
in the MSA method, reflecting for example higher magnitudes at higher levels.
Incremental Dynamic Analysis Holiday Inn, Van Nuys

2
1.8

1.6
a
a
0
10
1.4
1.2
0.8
0.6
0.4
0.2
1
10
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Maximum Interstory Drift Angle, 3
10 10
2 1
10
max
Maximum Interstory Drift Angle, max
a) Plot in arithmetic scale b) Plot in logarithmic scale

Figure 4-19 Results of the Incremental Dynamic Analysis when the model structure is subjected
to the selection of ground motion records.
The application of IDA method in factored demand evaluation is illustrated in Figure 4-20.
Factored demand can be calculated using the estimated values for D|S a , D|S a and b value
showed in Figure 4-20:
1 k 2 1 2.6
max | S a ( 0.49 ) 2
F .D. = | P0 S
e2 b = 0.0183 e 2 2.70 = 0.0183 1.12 = 0.020
max a
As expected, the results are identical to those of MSA method. Similar to MSA analysis, IDA
method provides a good estimate of the factored demand.
Spectral Acceleration of "first" mode, Sa [g] s 0

10 P0
Sa=0. 70
slope b=2 70
max | S a = 0 . 70 = 0 . 49
max |S a = 0 . 70 = 0 . 0183
1
10
3 2 1
10 10 10
max
Figure 4-20 - Evaluation of the factored demand using the results of incremental dynamic
analysis.
4.12.2.1 Capacity Estimation Using the Results of Incremental

Dynamic Analysis-
The IDA method can also be used to obtain global instability (capacity) information about the
structure. By global stability we mean a mode of collapse that is indicated by global dynamic
instability in displacement response, i.e., the displacements increase an arbitrarily large amount
for arbitrarily small increases in ground motion intensity. This particular mode of collapse can be
captured from the results of an IDA. Following the definition in the FEMA/SAC Guidelines (see
FEMA 2000), we have implemented this notion by marking the onset of this failure mode below
(Figure 4-21) by a point where the local slope of the IDA curve decreases to a certain
percentage of the initial slope of the IDA curve in the elastic region. This certain percentage is a
more or less arbitrary value that represents the point where the IDA curve becomes flat enough;
here, as in the FEMA/SAC Guidelines, we have chosen it to be equal to 20%. We can clearly
observe the scatter in the capacity points for the different records in Figure 4-21.
2
1.8 30
Spectral Acceleration of "first" mode, Sa [g] s 1.6

14 20 17
5
24 7
1.4
11 28
8
1.2
23 4
26
1
13 29
9 10 25
0.8
1 226
12 2115
2 3 27
0.6
19
0.4 16 18
0.2
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055
max
Figure 4-21 - Collapse Points marked on the IDA curves.
The statistical properties of the so-called collapse points are shown in Figure 4-22. Similar to
factored demand estimation, this information can be used for estimating the factored capacity as:
1k 2
C LS
F .C. = CLS e 2b
Note that we have used the b value here equal to 4, which is the local slope of the median curve in
the vicinity of the drift capacity, cap = 0.0278 . Alternatively, the value associated with the level
P0
S a may be used.
1 k 1 2.6
C2 ( 0.41) 2
F .C. = C e 2 b = 0.0278 e 2 4 = 0.0278 0.95 = 0.026
Now that we have calculated the factored capacity, we can compare it with the factored demand
that is estimated by multiple-stripe analysis or incremental dynamic analysis. Multiple-stripe
analysis estimated the factored demand to be equal to, 0.020, which is smaller than the factored
capacity, 0.026. This, according to DCFD design format (Equation 4-6), means that the annual
frequency of collapse is less than the selected target value of 0.0084.
2
Spectral Acceleration of "first" mode, Sa [g] s 1.8 30

14 20 17
1.6 5 7
1.4
11 28
cao = 0.38 8
1.2
23 4
26
1
cao
= 0.39 13 29
9 10 25
0.8
12 3 22
126 2115
27
0.6
19
0.4 16 18 cao = 0.0278
0.2 cao = 0.41
0
0 0.01 0.02 0.03 0.04 0.05 0.06
max
Figure 4-22 - Evaluation of the factored capacity using the results of incremental dynamic
analysis (IDA)
4.12.3 Evaluation of Factored Capacity by Numerical Integration
For comparison we can also use numerical integration to estimate the factored capacity. First, we
calculate the annual frequency of exceeding the global collapse limit state, LS , by numerical
integration (Equation 4-4-a). We have estimated the cumulative distribution function of collapse
limit state, FCLS ( y ) , in Equation 4-4-a by the ratio of the capacity values that exceed the value, y,
in the sample of capacity values obtained in the previous section (i.e., by the empirical
distribution). The rest of the integration procedure is very similar to the calculation of drift hazard
by numerical integration in the beginning of this chapter.
Now we need to transform the annual frequency of exceeding the collapse limit state into factored
capacity format. According to the DCFD format, at the onset of collapse limit state, the factored
capacity is equal to factored demand, i.e. the inequality in Equation 4-2 becomes equality.
Therefore, the factored capacity is equal to the factored demand corresponding to a P0 level
equal to the limit state frequency, LS . This means finding the drift value that corresponds to,
LS , from the drift hazard curve in Figure 4-5.
The annual frequency of exceeding the collapse limit state, LS , which is calculated by numerical
integration, is equal to 0.0063. This corresponds to a (factored) drift value equal to, 0.026, from
the drift hazard curve in Figure 4-5. Thus, the true factored capacity is equal to factored
demand for P0 = 0.0063 from the drift hazard curve, 0.026. Similar to the factored demand
calculated by numerical integration, the accuracy of the numerical integration can be increased by
using larger number of records and by choosing broader integration limits. We can also recall that
the factored capacity estimated by incremental dynamic analysis method was equal to 0.026.
4.13 Number of Records
In this chapter we have investigated alternative methods for demand estimation. All the methods
were based on the original sample of 30 records. In this section, we discuss the effect of the
sample size, e.g., the number of records for the single-stripe method. The effect of sample size for
the other methods can be studied in the same manner.
We have conducted the single-stripe method with 30 ground motion records at

P0
S a = 0.70 g .
This selection of 30 records can also be regarded as a sample of 30 earthquake scenarios,
magnitude and distance pairs, that we have chosen out of the set of all possible earthquake
scenarios for a given site. Here, we have conducted a bootstrap procedure (Efron and Tibshirani,
1993) on the sample in order to estimate the uncertainty in the estimation of median demand for a
given spectral acceleration as a function of the sample size.
We generated numerous (n=500) replications, also known as boot strap replications, of sub-
samples of size n (n is less than or equal to 30) by re-sampling with replacement. Each replication
provides an estimate for the median drift for the given spectral acceleration level. The median and
standard deviation of these median estimates were calculated for all the bootstrap replications for
each sample size n. Figure 4-23-a illustrates the 500 observed median estimates from the
bootstrap replications for each sub-sample size n (number of records). We can observe that the
estimates of the median are more scattered (individually less reliable), as the number of records
gets smaller. This reflects the typical sample size concern.
The standard deviation of the median of the bootstrap replications measures the uncertainty in the
estimation of the median drift demand for a given spectral acceleration. Figure 4-23-a also
illustrates the 16th, 50th and 84th percentiles of estimate of the median drift for a given spectral
acceleration (obtained by bootstrap analysis) as a function of the number of records used in order
to estimate the median. The standard deviation of the median of the bootstrap replications can be
presented by the width of the band created by the 16th and 50th and 50th and 84th percentiles
(Figure 4-23-b).
As we mentioned above, the standard error in the estimation of the median is a measure of
uncertainty in median due to limited sample size, i.e., number of ground motion records. We can
observe that the standard error in the estimation of median drops with the increase in the sample
size. This increase is (roughly) inversely proportional to square root of the number of records, as
one would expect from simple statistical theory. Further, as a fraction of the median, the standard
D |S a P0
error of estimation is equal to (Rice, 1995). The results shown are for S a = 0.70 g when
n
D|S a is 0.49. We see that in this case to reduce the standard error of the estimate of median to less
than 20% will require a sample size of 12 or more.
Demand Estimation with Different Number of Records

0 Bootstrap Replications of the Stripe Response for Sa=0.70g, N=30
10
1
10
max
2
10
3
10
0 5 10 15 20 25 30
Number of Records
Figure 4-23-a) Median and standard error of the estimates of the median response as a function of
the sample size (number of records). The straight line is the median response for each sample
size. The 84th and 16th percentiles of the estimator of median response are also plotted in the
figure.
The Standard Error of Estimation for Median Displacement Response, Sa=0.70g
1
0.9
0.8
0.7
a
0.6
D|S

0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30
Number of Records
Figure 4-23-b) The (fractional) standard error of the estimates of median response as a function of
the sample size (number of records).
4.14 Summary and Conclusions
Estimation of the engineering demand parameter (here, the displacement-based demand) is an

essential part of the Demand and Capacity Factored Design (DCFD) procedure. This chapter
presents alternative methods for the estimation of the factored demand, a displacement-based
demand measure in the context of the DCFD format. These methods employ the results of non-
linear dynamic analysis for displacement demand estimation. The methods are presented through
a comprehensive case study of an existing reinforced concrete 7-story frame structure in Los
Angeles. This structure represents an older reinforced concrete structure with degrading behavior
in nonlinear range. The dynamic analyses are performed on a set of ground motion records
selected from a catalog of California events recorded on stiff soil. The spectral acceleration of the
first mode has been chosen as the representative ground motion intensity measure. The methods
were divided into narrow-range and wide-range, with respect to the range of spectral acceleration
values for which they provide estimations for the factored demand. The factored demand
calculated by numerical integration, also referred to as the true factored demand, was used as a
basis for comparison of the alternative methods, which are based on analytical integrations.
Single-stripe analysis and cloud analysis were the two narrow range methods presented in this
chapter.
Single-stripe analysis method provided accurate estimations of the median displacement demand
for a given spectral acceleration, Po , and the dispersion measure for the displacement for a
max | S a
given spectral acceleration, P0

. However, it was unable to provide an estimate of the b
max | S a
value. While this estimate was satisfactory at lower ductility levels it led to a significant error at
near collapse ductility levels. The factored demand was estimated assuming a default b value of
1. The estimated factored demand was much larger than the factored demand calculated by
numerical integration because both P0
and b are large in such cases.
max | S a
In order to enhance the factored demand estimation by single-stripe analysis, the double-stripe
analysis method was introduced. This method consisted of performing another single-stripe
analysis sufficiently close to the original stripe in order to provide a reasonable estimate of the
log-log slope b value. For the near capacity ground motion level ( P0 S a = 0.70 g ), the factored
demand that was calculated using the estimated b value became much closer to the factored
demand calculated by numerical integration. The improvement in the estimation of factored
demand can be totally attributed to a fairly accurate estimation of the b value in the double-stripe
method. At the higher levels of max |Sa associated with high ductility levels, the factored demand
is much more sensitive to any error in b. This stresses the importance of estimating the b value in
the region of interest. As seen in Figure 4-14-b the log-log slope changes at higher ground
motions.
Cloud analysis using the ground motions as recorded, while accurate at the lower deformation
level, provided an estimate for the factored demand that was much smaller then the factored
demand calculated by numerical integration for the higher Sa level. It failed to estimate the
median demand well and underestimated max | S a .
In order to make the cloud response nearly centered on the spectral acceleration of 0.70g, the
ground motion records were scaled by a factor of two. The cloud analysis using the scaled ground
motions provided conservative estimate of the factored demand. Nevertheless, scaling the ground
motions did enhance the factored demand estimation to some extent. The median, max|Sa and b
values are all closer to the more extreme values at S a = 0.70 g . Cloud analysis by its very nature
cannot provide good local estimates of those parameters when they are changing rapidly.
As wide-range methods, multiple-stripe analysis and incremental dynamic analysis were studied
in the chapter.
Multiple-stripe analysis provided an accurate estimate of the factored demand for the extreme
spectral acceleration of 0.70g. Comparing the multiple-stripe analysis and single-stripe analysis
procedures for estimating the factored demand, it can be observed that everything is similar
except for the estimation of the b value. Like the double-stripe analysis, the multiple-stripe
analysis estimated the b value equal to the slope of the line connecting the median values for two
adjacent stripes close to the spectral acceleration of 0.70g. Whereas, the single-stripe method
could not provide an estimate of the b value. An accurate estimation of the b value may be
essential for getting an accurate estimation of factored demand, especially if max | S a is large.
Multiple-stripe analysis was applied to estimate the factored demand for multiple limit state
levels, i.e., multiple spectral acceleration levels. Generally, multiple stripe analysis provided
accurate estimates of the factored demand. The resulting factored demand estimations drifted
away from the true factored demand when the (estimated) median demand became
(unrealistically) non-smooth. In the next chapter, we will address these issues more carefully.
Incremental dynamic analysis provided results virtually identical to that of the multiple-stripe
analysis. Compared to multiple-stripe analysis, the incremental dynamic analysis carries
additional information about individual ground motion records by means of the IDA curves. The
IDA curves follow the structural response to individual records as the spectral acceleration
increases. These curves can be used to mark the onset of global instability for each ground motion
record.
The onset of global dynamic instability (collapse) for each record, i.e., the collapse point, was
marked on the corresponding IDA curve. The criteria for determining this threshold may change
(see Vamvatsikos and Cornell, 2002).
The factored capacity, a displacement capacity measure in the context of the DCFD format, was
estimated using the collapse points extracted from the IDA curves. The estimated factored
capacity was found to be equal to the factored capacity as calculated from the numerical
integration. The estimated factored demand for a spectral acceleration of 0.70g was smaller than
the estimated factored capacity for global dynamic instability (collapse) limit state. Therefore,
according to DCFD format, annual failure probability is confirmed to be less than 0.0047.
In the final section of this chapter, the issue of uncertainty in the estimation of median drift due to
limited sample size, i.e., n=number of records, was briefly discussed. The Bootstrap procedure,
based on the original selection of records, was used to estimate the standard error in the
estimation of the median drift that was provided by the single-stripe analysis.
It was confirmed that in fractional terms the standard error in the estimation of median drift
provided by single-stripe analysis is approximately equal to max | S a / n . Therefore the required
sample size to achieve a pre-set accuracy in the median demand (and roughly, the factored
demand) is dependent on max | S a which in turn tends to increase with ductility level.
Chapter 5
Probability-Based Seismic Assessment: Non-linear
Dynamic Displacement Estimation in the Region of
Global Instability
Jalayer, F. and Cornell, C. A. (2003 b). Earthquake Spectra, (in preparation).
5.1 Abstract
The demand and capacity factored design procedures discussed in the previous chapter, may
break down in the region of global dynamic instability. This is the regime of large roof or other
displacements where such displacements grow rapidly even for small increments in ground
motion intensity. This chapter presents methods for determining the probability distribution for
displacement demand that explicitly account for demand values reaching global instability. This
probability distribution is usually conditioned on an appropriate seismic intensity measure, here,
the spectral acceleration at the first mode period of the structure. The analysis methods discussed
in the previous chapter can be modified for the estimation of the conditional probability
distribution of displacement demand (given intensity) in the range of instability or collapse
(e.g., collapse may be defined as global displacements beyond an arbitrary large value, or as
lack of numerical convergence). This conditional probability distribution is used then in the
implementation of probabilistic frameworks for seismic assessments. The mean annual frequency
of exceeding a given maximum inter-story drift angle or drift hazard can be derived in various
ways based on how the conditional probability distribution of drift demand given the spectral
acceleration and the spectral acceleration hazard (the mean annual probability of exceeding a
given spectral acceleration) are estimated. Factored demand and factored capacity were
introduced in Chapter 3, as the main components of demand and capacity factored design
(DCFD). A generalized definition of factored demand can be used to estimate the factored
demand based on the estimated drift hazard curve. Factored capacity can also be estimated based
on the estimated drift hazard and corresponding limit state probability. However, based on this
definition, the factored capacity for the limit state of global collapse lies at infinity. Therefore, the
modified DCFD design procedure taking into account the collapse information can be applied for
seismic assessments only for limit states other than collapse or (exogenous limit states such as
maximum local displacement limits). This study is presented through a comprehensive case study
of an existing reinforced concrete 7-story frame structure in Los Angeles, which is modeled with
degrading strength in the nonlinear range.
5.2 Introduction
This chapter discusses the formulation and estimation of nonlinear dynamic displacement demand
for probabilistic seismic assessments in the most severe ground motion regime, i.e., when the
displacement demand is in the region of global instability in the structure. Global instability
refers to a limiting case in structural response, namely, a mode of collapse when the global
(dynamic) displacement response in the structure grows rapidly, i.e., the displacements increase
an arbitrarily large amount for a (relatively) small increase in ground motion intensity. This
limiting case in structural behavior, for a particular ground motion record can be marked by a
point where the slope of the incremental dynamic analysis (IDA) curve decreases to a certain
small percentage of its initial slope in the elastic region. Included too in global instability are the
situations when the computer algorithm fails to converge. (See Chapter 4 or Vamvatsikos and
Cornell (2001) for the description of an IDA curve).
In the previous chapter, we studied alternative methods for estimating the conditional probability
distribution of displacement demand for a given intensity measure. We then calculated the
demand hazard by integrating the estimated conditional distribution of displacement demand
given the intensity measure and the hazard function for the intensity measure. An analytical
closed-form expression was derived in Chapter 2 for the demand hazard assuming a parametric
function (power-law) for the hazard and a lognormal distribution with constant dispersion for the
conditional distribution of displacement demand. A demand and capacity factored design (DCFD)
format was derived in Chapter 3 based on this analytical form. Factored demand, as introduced in
Chapter 3, essentially forms one half of the formulation; the other half is the factored
capacity. In chapter 4, we focused our attention on factored demand estimation. Building on a
prior development in Chapter 3, factored demand was shown to be the demand value that
corresponds to a hazard value equal to the allowable frequency of limit state (this demand value
can be found from the demand hazard curve.). This equivalence property will be retained and will
prove useful for the estimation of factored demand in this chapter.
Estimation of the conditional probability distribution for the displacement demand is pivotal in
the demand estimation process. In Chapter 4, the distribution was typically modeled by a
lognormal distribution, which was used for the derivation of the closed form DCFD format; it was
also estimated by a non-parametric empirical distribution which was used for the numerical
integration calculations. However, for large values of ground motion intensity level, when the
displacement demand approaches the region of global instability in the structure, a considerable
portion of structural analysis results used for calculating the non-parametric percentiles produce
arbitrarily large displacements. As a result, additional parameters may need to be introduced in
order to reflect the portion of data in which the displacement response has become too large to be
appropriately modeled by a common statistical model (e.g., log-normal). Shome and Cornell
(1999) introduced a three-parameter distribution model in order to model the conditional
probability distribution of displacement demand for a given intensity measure when cases of
global instability (due to large deformations) are observed in the structural response. In this
model, for a given intensity measure, the conditional probability distribution for non-collapse
cases is modeled with a two-parameter lognormal distribution and a third parameter is introduced
as the probability of collapse.
This chapter revisits some of the methods discussed in Chapter 4, this time considering explicitly
global instability or collapse cases observed in the dynamic displacement response of the
structure. The dynamic displacement response for the non-collapse portion of the sample of
responses is modeled using the same statistical models discussed in Chapter 4. The additional
variable accounting for the global instability is modeled both parametrically and non-
parametrically. Parametric modeling is helpful for deriving a closed-form expression for demand
hazard later in this chapter (similar to the DCFD presentation in Chapter 3), while non-parametric
models are going to be useful for demand hazard estimation by numerical integration.
In the context of this study, the spectral acceleration of the fundamental period of the structure,
S a , is chosen to represent the ground motion intensity measure and the maximum inter-story
drift, max , represents the dynamic displacement-based demand. The methods discussed are
equally applicable to other choices of the intensity measure and demand definitions.
5.3 Organization of this chapter
Chapter 4 was structured based on alternative analytical methods for the estimation of the
displacement-based demand (here, the maximum inter-story drift angle), such as, single-stripe
method or multiple-stripe method. This chapter, however is organized based on the quantities that
are being estimated, namely, the conditional drift demand, the drift hazard, and, the factored
demand.
This chapter starts with a brief overview of spectral acceleration (the intensity measure) hazard
estimation, as it will be the basis of future calculations. The next section discusses various ways
in which the drift demand (displacement-based demand) can be estimated. The main purpose for
demand estimation is to be able to estimate the conditional probability distribution of drift
demand for a given spectral acceleration. The discussion of alternative methods for drift hazard
estimation follows, as it is based on how the spectral acceleration hazard and the conditional
probability distribution of drift demand for a given spectral acceleration are estimated. The
demand and capacity factored design discussed in Chapter 4 is revisited, this time considering the
global instability information in its formulation. Within this context, the alternative options for
the estimation of the factored demand will be discussed.
Among the narrow-range and wide-range analysis methods discussed in Chapter 4, only single-
stripe analysis, multiple-stripe analysis (MSA) and incremental dynamic analysis (IDA) are going
to be used here as estimation tools.
Similar to the previous chapter, we have selected the transverse frame in a 7-story frame structure
in Los Angeles, CA, as the case study frame structure. Information about modeling of this
structure can be found in Chapter 4. Based on the static pushover results in Figure 4-3, we can
anticipate that when maximum inter-story drift levels reach 1 to 2% or more, global instability
may begin to be a threat. Further as we saw in the several figures there the curve of median
maximum drift versus (dynamic intensity) first-mode spectral acceleration begins to curve and
become much less stiff for drifts above about 1%, again implying that under some records (at
least) the drifts are increasing rapidly as they do when under incipient dynamic instability. The
nonlinear dynamic analyses are performed using the same selection of ground motion records
described in the previous chapter. The selection consists of a set of 30 ground motion records
selected from the Silva Catalog (Silva, 1998). These records were all California events recorded
on stiff soil and were selected from a moment magnitude-(closest) source-to-site distance range
of, 5.5 M 7.5 and 15 R 120 .
5.4 Global dynamic instability capacity (collapse) estimation
In order to conduct studies in the region of global instability in the structure, we need first to
estimate the drift demand values that represent the onset of global instability in the structure. For
brevity, we will refer to these points as the collapse cases.
We have used the results of the incremental dynamic analyses (see Chapter 4) in order to estimate
the onset of global dynamic instability in the structure for each ground motion record. In this
study, the onset of global instability is marked by the point where the local slope of the IDA
curve decreases to a certain percentage (here 16.6% or less12) of the initial slope of the IDA curve
in the elastic region (Figure 5-1-a and 5-1-b). Note that some of these maximum inter-story drift
values are as low as 1 to 2 % but the majority lie in the 2 to 4% range.
1-a 1-b
Incremental Dynamic Analysis Holiday Inn, Van Nuys Incremental Dynamic Analysis Holiday Inn, Van Nuys
2 2
1.8 30 1.8 30
14 20 17
14 20 17
1.6 1.6 5 7
5
24 7
1.4 1.4
11 28 11 28
8 8
1.2 1.2
23 4 23 4
26 26
1 1
13 29 13 29
9 10 25 9 10 25
0.8 0.8
1 226
12 2115 12 3 22
126 2115
2 3 27 27
0.6 0.6
19 19
0.4 16 18 0.4 16 18
0.2 0.2
0
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0 0.01 0.02 0.03 0.04 0.05 0.06
Maximum Interstory Drift Angle, Maximum Interstory Drift Angle,
max max
12
Note that this percentage (16.6% or 1/6 of the elastic slope ) is slightly less than the (20% or 1/5 of the
elastic slope) used in the previous chapter. By decreasing the percentage, we were able to catch the
corner point that precede a flat line in some of the IDA curves.
Figure 5-1 Estimation of collapse points from IDA curves. a) IDA curves with the collapse
points marked on them. b) Collapse Points plotted Separately.
Here we assume that for all spectral acceleration levels greater than the spectral acceleration
capacity (the corresponding ordinate in the plot in Figure 5-1-b for each record), the structure has
collapsed. Also, similar to FEMA 350 guidelines (SAC), we set an arbitrary upper bound (in
SAC, it is a drift equal to 10%, here we set it equal to 5%) beyond which the model is surely
invalid. It should be noted that while this program (virtually always converged, meaning that it
produced a number) some programs do not. Such non-converging cases are also decidedly
collapse cases.
5.5 Spectral acceleration hazard
The spectral acceleration hazard, i.e., the annual frequency of exceeding a given spectral
acceleration, S a () , can be represented both parametrically and non-parametrically. The hazard
curve is normally described non-parametrically, which is the natural outcome of performing a

site-specific probabilistic seismic hazard analysis (PSHA), as in the example in Chapter 4. A
particular parametric formulation for the spectral acceleration hazard was introduced in the
previous chapters (see also, Kennedy and Short, 1994, and, Luco and Cornell, 1998), in which the
non-parametric spectral acceleration hazard curve was estimated (at least locally) by a power-law
expression. Hence, at S a = x :
S a ( x) = k 0 x k (5 - 1)
5.6 Probability distribution of demand for a given spectral acceleration
As discussed above, we presented alternative methods for estimating the conditional probability
distribution of drift demand for a given spectral acceleration in the previous chapter. This chapter
focuses on demand estimations in the region of global instability in the structure. In order to
model the distribution of demand in this region, we disaggregate the reported displacement
response values into two groups: given non-collapse and given collapse groups. For example,
given a spectral acceleration level of 0.5g, 3 of the 30 records have led to collapse. Thus, 27
records represent the given non-collapse group and 3 records the given collapse group for
this ground motion level. All the alternative methods discussed in the previous chapter are still
applicable for modeling the distribution of demand given non-collapse. The non-collapse
estimations need to be supplemented with information about the conditional probability of
collapse for a given spectral acceleration. This chapter presents both parametric and non-
parametric methods for representing the conditional probability of collapse for a given spectral
acceleration. Combining alternative methods for demand estimation introduced in Chapter 4
(here, for the non-collapse part), and the available options for estimating the conditional
probability of collapse leads to a large variety of possibilities for estimating the distribution of
demand in the region of global instability. We are not going to discuss all the possible options,
however. We shall only elaborate on those options that will be employed later for the estimation
of drift hazard and factored demand.
For a given spectral acceleration, the conditional probability distribution function for drift
demand in the region of the global instability in the structure can be estimated by parametric,
semi-parametric and non-parametric methods, based on the methods used for modeling the
distribution of non-collapse drift demand and the conditional probability of collapse.
Most of the methods discussed in this section incorporate a parametric statistical model known as
the three-parameter distribution (Shome and Cornell, 1999) in their formulation. This model,
estimates the distribution of the drift demand for a given spectral acceleration given that global
instability is not happening (non-collapse cases), by a lognormal (two-parameter) distribution. It
employs an additional (third) parameter to account for the probability of global instability
(collapse) for a given spectral acceleration level. The conditional probability of collapse for a
given spectral acceleration level, x, is denoted as:
PC|S a ( x) = P[collapse | S a = x] (5 - 2)
Equivalently, we can define the limit state of collapse in spectral acceleration capacity terms, i.e.,
collapse occurs when the spectral acceleration demand exceeds the (random variable) spectral
acceleration capacity, S a ,C . Then, the conditional probability of collapse for a given spectral
acceleration demand can also be written as:
PC |S a ( x) = P[collapse | S a = x] = P[ S a ,C S a | S a = x] = FS a , C ( x) (5 - 3)
where FS a ,C ( x) is the cumulative distribution function (CDF) of the spectral acceleration capacity.
This function is also known as a fragility function. A fragility function refers to a conditional
probability of exceeding a limit state for a given intensity measure or demand parameter (for a
discussion of fragility functions see Chapter 3 or Kennedy and Ravindra, 1984).
We dedicate the next few paragraphs to introducing the three-parameter distribution. This will be
useful later when we discuss the alternative methods for estimating the conditional distribution of
demand for a given spectral acceleration. Most of these methods employ the proposed three-
parameter distribution for modeling the distribution of demand.
5.6.1 A Three-parameter distribution and its percentiles
Shome and Cornell (1999) expanded as follows the probability that drift demand, max , exceeds
a specified value, y, for a given spectral acceleration, x :
G max |S a ( y | x ) = P[ max > y | S a = x] = G max | NC , S a ( y | x) PNC |S a ( x) + G max |C , S a ( y | x) (1 PNC |S a ( x) ) (5 - 4)
where G () denotes the complementary cumulative distribution function (CCDF) and PNC |S a ( x)
(= 1 PC |S a ( x) ) is the probability of no collapse for a given spectral acceleration, x . Equation 5-4
is a special application of total probability theorem (see Appendix B or Benjamin and Cornell,
1970) expanding the probability of exceeding drift hazard based on whether the drift response has
reached collapse (denoted by C) or not (denoted by NC). Given collapse, the conditional
probability of exceeding any finite drift demand given collapse, G max |C , S a () , is assumed to be
equal to 1:
G max |C , S a ( y | x) = P[ max > y | Collapse, S a = x] = 1 (5 - 5)
Using the above assumption, Equation 5-4 can be re-written as follows:
G max |S a ( y | x) = G max | NC , S a ( y | x) PNC |S a ( x) + (1 PNC |S a ( x) ) (5 - 6)
where G max | NC , S a () is the CCDF of displacement demand, y, for a given spectral acceleration, x,
given that no collapse cases has happened. Assuming that the drift demand distribution of non-
collapse cases (for a given spectral acceleration) is lognormal, with median, max | NC, S a , and
fractional standard deviation (i.e., standard deviation of the natural logarithm or dispersion),
max | NC, S a . The CCDF, G max | NC , S a () , can be calculated as:
ln( y ) ln( max | NC , S a ( x))

G max | NC , S a ( y | x) = P[ max > y | No Collapse, S a = x] = Cmax|NC , S a ( ) (5 - 7)
max | NC , S a ( x)
Where C () is the standardized gaussian CCDF. Note that both the median and the dispersion
are in general functions of S a level, x. Substituting G D| NC () with expression above (Equation 5-7)
the CCDF for the drift demand given spectral acceleration, G max |S a () , is finally calculated as:
G max |S a ( y | x ) = P[ max > y | S a = x] =

Cmax|NC , S a ( ) PNC |S a ( x ) + (1 PNC |S a ( x) ) (5 - 8)
max | NC , S a ( x)
Equation 5-8 demonstrates that the conditional probability distribution of drift demand for a given
spectral acceleration is expressed by three parameters, namely, the conditional probability of no
collapse, PNC |S a (.) , the median drift demand, max | NC , S a (.) , and the fractional standard deviation
or dispersion of drift demand given no collapse, max | NC , S a (.) = ln max | NC , S a (.) , all functions of
ground motion intensity level.
The CDF of the three-parameter distribution can be calculated after some simple algebraic
manipulations and noting that G () = 1 F () :
F max |S a ( y | x) = P[ max y | S a = x] =
1 G max |S a ( y | x ) = ( ) PNC |S a ( x) (5 - 9)
max|NC , S a
max | NC , S a ( x)
where () is the gaussian CDF. Equation 5-9 can be used in order to calculate the drift demand
value, y p , corresponding to percentile p of the three-parameter distribution, by setting the left
side of the equation equal to p and solving it for y p :
p
y p (x) = Fmax
1 1
| S a ( p | x ) = max | NC , S a ( x ) exp( max | NC , S a ( x ) max | NC , S a ( )) (5 - 10)
PNC |S a ( x)
where y p (x) is the drift demand percentile for the spectral acceleration level x, i.e., the p% drift
value, assuming that the probability distribution of the displacement response is estimated by the
three-parameter distribution and 1 () is the inverse function of standardized normal
distribution. Note that for PNC |S a ( x) = 1 , this leads to familiar formulas for 16th, 50th, and 84
percentiles of a lognormal distribution:
y 0.16 ( x) = max | NC , S a ( x) exp( - max | NC , S a ( x))

y 0.50 ( x) = max | NC , S a ( x)
y 0.84 ( x) = max | NC , S a ( x) exp( max | NC , S a ( x)) (5 - 10 - a)
5.6.2 Methods incorporating the three-parameter distribution
In the section below, we discuss methods that incorporate the three-parameter distribution for
modeling the conditional probability distribution of demand for a given spectral acceleration.
These methods are differentiated based on how they estimate the (three) parameters of the three-
parameter distribution.
5.6.2.1 Fully parametric using power-law functions
In this method, the parameters of the three-parameter distribution are all represented by analytical
(parametric) functions that are assumed to apply at least locally in the region of interest. More
practically, this means that they are accurate enough to provide adequate numerical estimates of
the drift hazard, Equation 4-5-a, in some interval of drift levels of interest. The median drift
demand for a given spectral acceleration given non-collapse, max | NC , S a (.) , is assumed to follow a
power-law expression (e.g., Luco and Cornell, 1998, and, Cornell, Jalayer et al., 2002):
max | NC , S a ( x) = max | NC , S a = x = a x b (5 - 11)
The fractional standard deviation of drift for a given spectral acceleration given no-collapse is
assumed in this model to be constant with respect to spectral acceleration:
max | NC , S a ( x) = NC (5 - 12)
The parameters, a, b, and NC can be estimated for example, by simple regression applied to
cloud or set of stripes of response results, as in Chapter 4, but now using only non-collapse data
points. The conditional probability of non-collapse for a given spectral acceleration, PNC|S a , or
the spectral acceleration fragility, FS a ,C ( x) , can be modeled by a power-law distribution (see
Shome and Cornell (1999):
-C
x
PNC |S a ( x) = 1 FS a , C ( x) =
x > s a0 (5 - 13)
s a0
PNC |S a ( x) = 1 FS a , C ( x) = 1 x s a0
The collapse points in Figure 5-1 can be used to estimate PNC|S a empirically (e.g., by the ratio of
the collapse points with a spectral acceleration greater than a given value, x). The values of the
two parameters, s a 0 and c can be estimated, for example, by performing linear regression in the
logarithmic space on the empirical PNC|S a values (Figure 5-2 and Table 5-1). If we substitute the
above expressions in Equation 5-8, the CCDF for the drift demand given spectral acceleration,
G max |S a () will be written as:
G max |S a ( y | x) = P[ max y | S a = x] =
-C x -C
ln( y a x b ) x
Cmax|NC , S a ( )
+ 1

x s a0 (5 - 14)
NC s a0 s a0

and,
ln( y a x b )
G max |S a ( y | x) = P[ max y | S a = x] = Cmax|NC , S a ( ) x s a0 (5 - 15)
NC
Conditional Probability of NonCollapse given Spectral Acceleration
1
Powerlaw,Sa0=0.55g, c=2.3
0.9 Empirical sabased
0.8
0.7
0.6
a
PNC|S 0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Sa
Figure 5-2 - Conditional probability of non-collapse for a given spectral acceleration represented
both empirically and analytically.
sa0 C
0.55g 2.3
Table 5-1. Parameter estimates for s a 0 and c
For spectral acceleration values smaller than s a 0 , the conditional probability distribution of drift
demand for a given spectral acceleration is the same as the non-collapse conditional probability
distribution. This is because, according to the model, the spectral acceleration value s a 0 marks
the minimum spectral acceleration at which collapse cases can occur in the structure. In
application, as in this example, there may be cases, e.g., S a = 0.55 , where there are observed
collapses at S a levels less than the estimated value of s a 0 . Other methods of parameter
estimation, e.g., maximum likelihood, could be used to avoid this inconsistency if it should prove
unsatisfactory. The parametric model has the advantage that it can be estimated by results from as
few as two S a levels, implying savings in analysis runs. This property will be explored in a
following section. Equation 5-14 will be used later in this chapter for the derivation of a closed-
form result for the drift hazard considering now the global instability information.
5.6.2.2 Fully parametric using the IDA procedure
This method employs incremental dynamic analysis (IDA) to estimate the parameters of the
three-parameter distribution. For a set of ground motion records, the results of incremental
dynamic analysis of the structure are represented by the IDA curves. This new estimation method
is based on the typical observation that the individual IDA curves in the non-collapse regime may
each be relatively straight in a log-log representation but with quite widely varying slopes and
intercepts (see Figure 5-3). The underlying probabilistic model is that each such curve is a
straight line with random slope and intercept:
ln max| NC , S a ( x) = ln A + B ln x x S a ,c ( 5 - 16)
in which the intercept denoted by ln A (natural log of A) and slope b are assumed to be jointly
normally distributed random variables with means, ln A . and B , and variances, 2 ln A .and
2 B , and correlation coefficient, . The easily derived implications of this model are that
max| NC , S a ( x) is log-normally distributed with median:
max| NC , S ( x ) ( x) = A x B (5 - 17 )
a
in which the A is the median of A (i.e., exp[ ln A ] ), and fractional variance (i.e., variance of
ln ):
2 max | NC , S a ( x) = ln2 A + 2 ln x B ln A + (ln x) 2 2 B (5 - 18)
Note that unlike the model in previous section the fractional standard deviation is a function of
spectral acceleration, not a constant.
The statistical parameters of ln A and b can be estimated as follows. For each record, i, fit (by
regression) a line of the form, ln i ( x) = ln a i + bi ln x , to the nonlinear dynamic runs for all non-
collapse cases, i.e., for x S a,ci , where S a ,ci is the ith observed value of the S a capacity random
variable, i.e., the S a level at which collapse is observed under the ith record. Then the sample
values {ln a i , bi } for i=1,2,,n, can be simply processed for the sample means, standard
deviations and correlation coefficient, which can serve as estimates of the corresponding model
parameters. Note that we are ignoring the variability of individual data points about each IDA
line. Experience has shown that (in the non-collapse domain) this variability is quite small
relative to the record-to-record variability of the fitted lines themselves. Figure 5-3 shows a few
examples of such fitted lines. In fact, it is this small variability that implies that non-collapse
portion of this method could be accurately implemented by as few as two non-linear analyses per
record, which is a primary advantage it has.

1.8
1.6
a
1.4 Northridge, Hollywood Stor, 360
1.2
1
Kern County, Hollywood Stor, 90
0.8
0.6
0.4
Imperial Valley, Westmorland, 90
0.2
0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
max
Figure 5-3 - Some IDA curves in the non-collapse regime approximated by straight lines in the
log-log representation. Note that the points marked as circles in the end are the points at
which the global instability capacity was reached (Figure 5-1)
The expression for the CCDF of drift demand for a given spectral acceleration G max |S a () is
similar to Equations 5-14 and 5-15, where the median and fractional standard deviation are
calculated now from Equations 5-17 and 5-18. The conditional probability of no collapse
(Equation 5-13) is modeled and found as in the previous section. Figure 5-4 illustrates the 16th,
50th and 84th percentiles of the distribution of drift given spectral acceleration obtained by
substituting the values for max | NC , S a (.) and max | NC , S a (.) from Equations 5-17 and 5-18 into
Equation 5-10. The thin lines are the percentiles of the drift response (given spectral acceleration)
obtained using the fully parametric IDA curves given no-collapse but ignoring the likelihood of
collapse, i.e., by assessing s ao in Equation 5-13 equal to infinity (i.e., as if this IDA-based
method had been applied in Chapter 4 as it very well could have been). The thick lines are the
percentiles using the fully parametric IDA curves and incorporating the probability of collapse
information. The consideration of collapse cases has caused the percentiles to bend and diverge
from their corresponding no-collapse values. This approach and the information gathered from
the percentiles are going to be incorporated in the factored demand estimation later in this
chapter. The parameter estimates used in this example are given in Table 5-2.
A A B B
0.025 0.43 1.23 0.19 0.93
Table 5-2. Statistical parameters for ln A and B
Full Parametric IDA Holiday Inn, Van Nuys

0
10
th
50 percentile, fullyparametric IDA with no collapse
th
th
1
10 th
50 percentile, fullyparametric IDA with collapse
th
th
3 2
10 10
Figure 5-4 - Fully parametric 16th, 50th and 84th percentiles of the drift response given spectral
acceleration obtained using parametric IDA curves.
5.6.2.3 Semi-parametric using multiple-stripe analysis
This method uses the results of the multiple-stripe analysis (revisited in Chapter 4) in order to
estimate now the parameters of the three-parameter distribution and hence the probability
distribution of demand in the region of global instability. For a given set of records, the results of
the multiple-stripe analysis on the structure are represented by a series of stripes that are plotted
for multiple spectral acceleration levels. For each spectral acceleration level, x, the non-collapse
part of the stripe response is modeled by a lognormal (parametric) distribution. Also, the
proportion of the observed collapse cases in the stripe response is used to estimate the
(conditional) probability of collapse for the given spectral acceleration level, x. In contrast to the
previous sections of this chapter these three parameter distribution parameters are not represented
by parametric functions of x, however.
The results of the multiple-stripe analysis on the structures are plotted in Figure 5-5. These results
are the same as those obtained in the previous chapter. The collapse cases, i.e., those for which
the S a value exceeds S a ,ci of that record, are marked as stars inside the circles.

0
10
a
1
10 "collapse" cases
non"collapse" cases
3 2 1
10 10 10
max
Figure 5-5 - Response results of multiple-stripe analysis with cases identified as collapse marked.
In order to fit a lognormal distribution to the non-collapse results for each spectral acceleration
level, x, we need to estimate the median, max | NC , S a ( x) , and the fractional standard deviation,
max | NC , S a ( x) . Here, we have estimated the non-collapse median, max | NC , S a ( x) , by the (counted)
50th percentile of the non-collapse part of the stripe response. The non-collapse fractional
standard deviation, max | NC , S a ( x) , is estimated in the logarithmic scale by half the distance
between 16th and 84th percentiles of the non-collapse results. Note that the more conventional
method of moments might have been used as well in this case. (The counted method has been
used in the past largely to avoid the problems introduced by collapse-case data points, especially
when these are associated with the non-convergence for which no observed value of max is
available.) The conditional probability of non-collapse, PNC |S a ( x) , for the stripe is estimated
(empirically) by the ratio of the non-collapse cases to the total number response points on the
stripe.
Having estimated the three parameters of the three-parameter distribution, as explained in the
above paragraph, we can calculate the 16th, 50th and 84th percentiles of the marginal distribution in
Equation 5-9 for each spectral acceleration level, x. Figure 5-6 illustrates these percentiles, as a
function of the spectral acceleration, against the multiple-stripe analysis results in the
background. The heavy lines stop when the probability of collapse becomes so large as to leave
them undefined. The probability of collapse exceeds 50% above 0.7 g.

0
10
a
50th percentile, counted given nocollapse

th
1 50 percentile, threeparameter distribution
10 th
16 percentile, threeparameter distribution
th
84 percentile, threeparameter distribution
3 2 1
10 10 10
max
Figure 5-6 - The 16th, 50th and 84th percentiles of the semi-parametric three-parameter
distribution calculated from the results of the multiple-stripe analysis.
5.6.2.4 Pseudo-parametric method
An alternative way to consider the collapse cases in the estimations is to fit a two-parameter
lognormal distribution to the estimated percentiles. In this method, which we refer to as the
pseudo-parametric approach, the 16th, 50th and 84th percentiles of the three-parameter distribution
(calculated from Equation 5-10) are used to estimate the (two) parameters of the lognormal
distribution, namely, the conditional median and fractional standard deviation of the drift demand
given spectral acceleration. The conditional median, max |S a , is estimated by the 50th percentile
and the conditional factional standard deviation, max |S a , is estimated by half the distance
between counted 16th and counted 84th percentiles in the logarithmic scale. We have just
presented several methods for deriving the parameter values of the response based on the three-
parameter distribution. In this pseudo-parametric approach, the estimated percentiles using these
alternative methods are approximated by a lognormal distribution. This model should be used
with caution since it is trying to represent a sample including both extreme values (the collapse
cases) and moderate values (the non-collapse cases) by only two parameters. Nonetheless we are
going to use this method for factored demand estimation later in this chapter, if only because it
provides the opportunity to use the simple representation of factored demand in Chapter 4. We
shall compare its estimates with those obtained more rigorously.
5.6.2.5 Comparison of the fully-parametric IDA method to the

semi-parametric method with MSA
The counted 16th, 50th and 84th percentiles of the drift response given spectral acceleration
obtained semi-parametrically (using multiple-stripe analysis) and fully-parametrically (using
IDA) are compared in Figure 5-7.
We can observe that parametric percentiles are somewhat different from the semi-parametric
percentiles especially at lower S a values. Reasons for this difference include the fact that the
power-law curve in Figure 5-2 is not a very good approximation for the empirical PNC|S a , that
IDA fits are dominated here by the higher density of points at the higher S a levels, and that the
IDA method simply does not have the flexibility that the MSA method has. However, the
advantage of using the fully parametric method for estimating the probability distribution of drift
given spectral acceleration is that the resulting percentiles are smooth functions of spectral
acceleration.
Comparing the FullParametric (with IDA) and SemiParametric (with MSA) Methods
for Probabilistic Demand Estimation Holiday Inn, Van Nuys

0
10
th
50 percentile, semiparametric MSA
th
th
1 th
10 50 percentile, fullyparametric IDA
th
16 percentile, fullyparametric IDA
th
84 percentile, fullyparametric IDA
3 2
10 10
max
Figure 5-7 - The 16th, 50th and 84th percentiles of drift response given spectral acceleration
estimated semi-parametrically using multiple-stripe analysis (MSA) and fully parametrically
using the IDA procedure.
5.6.3 Non-parametric Methods
The methods that are discussed next do not explicitly use the three-parameter distribution for
modeling the conditional probability distribution of demand for a given spectral acceleration.
These methods simply estimate the counted percentiles of the drift demand given spectral
acceleration using the results of multiple-stripe analysis.
The calculation of the counted percentiles involves sorting the reported drift results in ascending
order. Therefore, the large drift values that correspond to collapse cases will move to the upper
end of the sorted data and, hence, they will not enter the calculations as long as the desired
percentile values exist (i.e., they have not yet reached the global dynamic instability limit).
5.6.3.1 Non-parametric considering the collapse cases
The non-parametric percentiles of the drift demand given the spectral acceleration can be
calculated by treating the collapse cases similar to non-converging results. In other words the
reported drift values that are beyond the drift capacity level (for each specific record) are
effectively set to infinity. When, for any spectral acceleration value, the number of such
considered collapse cases is large enough, the percentile value becomes effectively infinite as
well and we simply cut-off or stop reporting the percentiles at that point.
The non-parametric estimates can also be obtained for G max | NC , S a ( y | x) and p NC|S a ( x) . The CDF
for the non-collapse part of the stripe response is estimated empirically by the fraction of the non-
collapse results that are less than or equal to a given drift value. The probability of non-collapse is
estimated empirically by the ratio of the non-collapse cases observed in the stripe response. This
method will be applied later in this chapter for calculating the drift hazard by numerical
integration.
5.6.3.2 Non-parametric not considering the collapse cases
In this method the percentiles of the drift response for a given spectral acceleration are calculated
by counting as described above. The counted percentiles are calculated from the reported drift
values ignoring the fact that some may be beyond the collapse drift levels, i. e., collapse or non-
collapse is not considered or recognized. The counted percentiles obtained in the previous chapter
were calculated this way.
5.6.3.3 Comparison of the non-parametric methods considering

and not considering the collapse cases
Figure 5-8 illustrates the non-parametric (counted) percentiles of the drift response as a function
of spectral acceleration for the two methods discussed above, namely, the methods considering
and not considering the collapse cases. The two sets of percentiles are of course identical as long
as the counted percentiles considering the collapse cases exist. It can be observed that the counted
percentiles considering the collapse information are cut off when the percentage of non-collapse
cases is less than the corresponding percentile. For example, the 50th percentile curve considering
collapse is cut off at 0.7g. At this spectral acceleration level, some of the (relatively) smaller drift
values were recognized as collapse cases, and this has slightly shifted the 50th percentile curve
towards larger drift values. It is at higher S a and drift levels, near collapse that differences
emerge.
Counted Percentiles for MultipleStripe Analysis
Holiday Inn, Van Nuys
Spectral Acceleration of "First" Mode, S [g] s
a
0
10
counted 50th considering collapse infromation

th
counted 16 considering collapse information
th
counted 84 considering collapse information
th
counted 50 NOT considering collapse information
th
th
1
10
3 2 1
10 10 10
Figure 5-8 - Multiple-stripe analysis counted 16th, 50th and 84th percentiles of the response with
and without consideration of the collapse information
5.6.3.4 Comparison of the non-parametric method not considering

the collapse cases and semi-parametric method with multiple-
stripe analysis
The counted 16th, 50th and 84th percentiles of the drift response as a function of the spectral
acceleration (as in Chapter 4, not considering the collapse information) and the semi-parametric
estimation of parameters of the three-parameter distribution using multiple-stripe analysis
discussed in Section 5.6.2.3 are compared in Figure 5-9 below.
The median curves in the two cases are very close up to the cut-off point. The 16th and 84th
percentiles are fairly close for the two cases especially when the drift values are small. It can be
anticipated, however, that predictions associated with large drift values (beyond 2-3%)will be
strongly affected. We shall see in Section 5.7.2 that, indeed the drift hazard results diverge in this
(large drift) region.
5.6.3.5 Discussion
Several methods for estimating the conditional probability distribution of demand given spectral
acceleration have been discussed. These methods have been classified based on the degree to
which they use parametric estimations in their calculations. Obviously, not all the possible
options have been presented. The attention has been focused on the methods that are going to be
used later in this chapter for the estimation for the drift hazard and factored demand. For this
illustrative case at least, as shown in Figure 5-7, the percentiles for demand provided by the fully
parametric method based on IDA procedure, are not very close to those of the semi-parametric
method using MSA analysis. Nevertheless, the percentiles provided by the fully parametric
method can be obtained with a relatively small amount of analysis effort and yet they provide
smooth results.
MultipleStripe Analysis Holiday Inn, Van Nuys

0
10
a
50th percentile, semiparametric MSA

th
1 16 percentile, semiparametric MSA
10 th
th
50 counted percentile, NOT considering collapse cases
th
th
3 2 1
10 10 10
Figure 5-9 - The counted 16th, 50th and 84th percentiles not considering the collapse cases (as
per Chapter 4) and the semi-parametric 16th, 50th and 84th percentiles using multiple-stripe
analysis.
It is observed that the non-parametric counting method provides very good estimates at lower
spectral acceleration/drift levels where there are few (if any) collapse cases. But deviations at
large values may affect predictions, e.g., drift hazard at larger drift values in the collapse regime
(to be seen in the next section). Nonetheless, the non-parametric counting methods still provide
reasonable predictions in the near collapse regime, given their theoretical simplicity.
5.7 Drift Hazard Annual Frequency of Exceeding Maximum Inter-Story Drift

Demand
The drift hazard, or the mean annual frequency of exceeding maximum inter-story drift demand,
max () , when the drift demand is in the region of the global instability in the structure, can be
estimated by parametric, semi-parametric and non-parametric methods. Based on how the

spectral acceleration hazard and the probability distribution for demand given spectral
acceleration are estimated, there are various possibilities for calculating the drift hazard. Only
those that are more useful in the context of probabilistic seismic assessments in the region of
global instability in the structure will be discussed here.
5.7.1 Estimating drift hazard by disaggregating into collapse and

non-collapse parts
In the previous chapter, the drift hazard or the annual frequency of exceeding maximum inter-
story drift demand was expanded with respect to the spectral acceleration as follows:
d S a ( x )

max ( y ) = G max |S a ( y | x)
x
dx
dx (5 - 19)
Where () denotes the hazard function and G max |S a () denotes the CCDF of maximum inter-story
drift for a given spectral acceleration level.
While deriving the three-parameter distribution in the previous sections, we represented the
CCDF of maximum inter-story drift for a given spectral acceleration level, G max |S a () , in the
region of the global instability in the structure by Equation 5-6. Substituting that expression for
G max |S a () in the drift hazard equation in Equation 5-19, the drift hazard becomes:
d S a ( x)

max ( y ) = {G max | NC , S a ( y | x) PNC |S a ( x) + (1 PNC |S a ( x) }
x
dx
dx (5 - 20)
5.7.1.1 Drift hazard estimation in limiting cases
When the drift value, y, is very large, G max |S a , NC ( y | x) , the probability that the drift demand
exceeds y given non-collapse, approaches zero for all spectral acceleration values, x. In this case,
the drift hazard in Equation 5-20 becomes:
d S a ( x) d S a ( x)

max ( y ) = (1 PNC |S a ( x))
x
dx
dx =
x
PC |S a ( x)
dx
dx = LS (5 - 21)
The expression for drift hazard in Equation 5-21, which is independent of the value y, is equal to
the annual frequency of collapse limit state, LS . Therefore, for large drift values, the drift hazard
is asymptotically equal to the annual frequency of global instability limit state, LS (or frequency
of collapse).
When the drift value y is very small, PNC |S a ( y | x) , or the probability of non-collapse for a given
spectral acceleration, approaches to zero for all spectral acceleration values, x. In this case, the
drift hazard in Equation 5-20 becomes:
d S a ( x)

max ( y ) = [G max | NC , S a ( y | x)]
x
dx
dx = max | NC ( y ) (5 - 22)
Thus, for small drift values, the drift hazard is equal to the drift hazard assuming no collapse,
which we denote by, max | NC (.) .
5.7.2 Non-parametric calculation of drift hazard by numerical

integration
We can use the non-parametric method discussed in the previous section to estimate the CCDF of
drift for a given spectral acceleration and given non-collapse, G max |S a , NC (.) , and the conditional
probability of non-collapse for a given spectral acceleration, PNC |S a (.) . The spectral acceleration
hazard, S a (.) , can also be estimated non-parametrically as explained previously. Substituting the
non-parametric estimates for G max |S a , NC ( y | x) , PNC |S a (.) and S a (.) into Equation 5-20, we can
calculate the integral by numerical integration. This approach is parallel to the one used for the
calculation of drift hazard by numerical integration in Chapter 4.
The drift hazard derived by numerical integration of Equation 5-20 is plotted versus spectral
acceleration in Figure 5-10-a. In Figure 5-10-b, the drift hazard is plotted together with the drift
hazard curve calculated non-parametrically by numerical integration without taking into account
the collapse cases in the integration (as in Chapter 4). The two curves are identical for the low
drift values where the probability of collapse given spectral acceleration is small. This is expected
since the drift hazard given non-collapse is the upper limit for the drift hazard as discussed above.
The two curves start to diverge as the drift values become larger, when the drift hazard
considering the collapse cases draws close to its lower limiting value, the annual frequency of
collapse, LS .
The probability of the global instability limit state in the structure can be estimated by the lower
asymptotic limit of the drift hazard, as described in the previous section. Here, as the drift
demand becomes large, the drift hazard gradually drops to a limiting value of LS = 0.007. Thus,
the general form of this drift hazard was anticipated and it deviates from that which would have
been determined by methods presented in the previous chapter in the near collapse regime, here,
for drifts of 1-2% and higher.
Annual Frequency of Exceeding Maximum Interstory Drift Annual Frequency of Exceeding Maximum Interstory Drift
0 0
10 10
"collapse" cases considered (nonparametric distribution)

"collapse" cases not considered (nonparametric distribution)
1 1
10 10
max
max
2
10
LS=0.007 2
10
3
10 3
3 2
10 10 10 3 2
max 10 10
max
10-a 10-b
Figure 5-10 - Non-parametric estimation of drift hazard: a) collapse cases considered
b) Plot 10-a together with drift hazard when the collapse cases are not considered.
5.7.3 Semi-parametric calculation of drift hazard by numerical

integration using multiple-stripe analysis
The drift hazard can also be calculated by applying the semi-parametric method (using multiple-
stripe analysis) for demand estimation. Recall for each stripe, the non-collapse median,
max | NC, S a , and fractional standard deviation, max | NC, S a , are estimated in order to fit a lognormal
distribution to the non-collapse part of the stripe response. This provides G max |S a , NC (.) . The
conditional probability of non-collapse for the stripe, PNC|S a , and the spectral acceleration hazard,
S a (.) , are both estimated empirically, as in the previous section. The drift hazard is calculated by
numerical integration from Equation 5-20.
The semi-parametric drift hazard is plotted in Figure 5-11 together with the non-parametric drift
hazard calculated in the previous section. The two curves illustrated in Figure 5-11 are so close
that they may be used inter-changeably. This implies that the lognormal distribution is an
adequate representation of the drifts given non-collapse.
0
10
1
10
max

2
10
numerical integration considering collapse information

(lognormal distribution)
numerical integration considering collapse information

(nonparametric distribution)
3
10 3 2 1
10 10 10

max
Figure 5 -11 - Drift hazard calculated by numerical integration using the non-parametric and
semi-parametric (using the lognormal distribution for the non-collapse cases) methods.
5.7.4 Fully parametric (closed-form) evaluation of drift hazard
The fully parametric method for estimating the probability distribution of drift for a given spectral
acceleration was introduced in a previous section (Equations 5-13 to 5-15). We use this method
next to evaluate the drift hazard curve analytically. The conditional probability of non-collapse
for a given spectral acceleration, PNC |S a (.) , can be modeled parametrically by the power-law form
introduced in Equation 5-13. The spectral acceleration hazard, S a (.) , is also estimated
parametrically from Equation 5-1. After substituting these estimates for G max |S a , NC ( y | x) ,
PNC |S a (.) and S a (.) in Equation 5-20 and carrying out the integral, a closed form expression for
the drift hazard can be derived (for details of the derivation see Shome and Cornell, 1999):
max ( y ) = collapse term + non collapse term1 + non collapse term2 ( 5 - 23)
where:
C
collapse term = Sa (s a0 )
k + C
1 k2
1 k 2
NC
b NC ln(s / y
s ay )
2
NC 2 bNC
b ln( s e / s )

a
non collapse term1 = S a ( s a 0 ) a0 2 bNC 2 NC
+ S a ( s ay ) e a0
NC
NC

k b ln(s / s ay )
non collapse term2 = S a ( s a 0 ) NC a0 +
k + C
NC
1 k + C 2
1 (k + C )2 2 NC

-C
NC 2 bNC y
k s ay b ln( s e / s )
C
2 bNC 2 NC a
S a ( s ay ) e a0
k + C s a0 NC

Where s ay is the spectral acceleration corresponding to the drift value, y, from the median curve
(Equation 5-11), i.e.,
1
y b
s ay = 1 y ( y) or more specifically : s ay =
max | NC , s a
a
The other parameters have been defined previously. Recall that k is a (log) slope parameter for
the spectral acceleration hazard curve; b is a (log) slope parameter for the median drift demand
curve (Equation 5-11) and C is a (log) slope parameter for PNC|S a or spectral acceleration
fragility function (Equation 5-13). s a 0 is the fragility function parameter representing the
minimum spectral acceleration value in which collapse cases are presumed to occur in the
structure (Equation 5-13). NC is the constant fractional standard deviation of the non-collapse
drift demand given spectral acceleration (Equation 5-12).
The consideration of collapse information makes the closed-form solution significantly more
complicated than the DCFD formulation discussed in the previous chapter. It should also be noted
that although Equation 5-23 gives a closed-form solution for the drift hazard when collapse
information is being considered, it is not an analytical solution like the one in Chapter 4, due to
the (non-analytical) standardized Gaussian CDF, (.) , and its complement which appear in the
formulation.
When the drift level y is low enough that s ay is small relative to s a 0 , the drift hazard will be in this
limit equal to the drift hazard given non-collapse:
1 k2 2

2 NC
y
max ( y ) = max | NC ( y ) = Sa (s a ) e 2 b (5 - 24)
This is the same form as in the previous chapter. On the other hand when y is such that s ay is
large compared to s a 0 , the drift hazard will in this limit be equal to the limit state of collapse
frequency, LS :
C
max ( y ) = LS = S (s ) (5 - 25)
k + C a a0
0
FullParametric Drift Hazard Curve
10
max

Mean Annual Frequency of Exceedance,
1
10
2
10
Fullparametric drift hazard curve, given noncollapse:

k=3, k =0.003, b=1, a=0.015, beta =0.3, s =0.55g, beta =2.3
0 nc a0 c
3
10 Fullparametric drift hazard curve:
k=3, k =0.003, b=1, a=0.015, beta =0.3, s =0.55g, beta =2.3
0 NC a0 C
Fullparametric collapse limit state frequency:
k=3, k0=0.003, betaC=2.3, sa0=0.55g
4
10 3 2 1
10 10 10
Maximum Interstory Drift Ratio,
max
Figure 5-12 - Full-parametric drift hazard estimation using single-stripe analysis
Figure 5-12 illustrates a drift hazard calculated from Equation 5-23, using, s a 0 and C values
estimated in Figure 5-2 by fitting a power-law function to the empirical PNC|S a data points, the a
and b and max | NC values close to the estimates given in Figure 4-10 (cloud analysis for the un-
scaled records), and k and k 0 values by fitting a power-law to the hazard curve in the region of
interest (near S a = 0.70 g ). The figure also illustrates the limiting values for the drift hazard
curve, namely, the drift hazard given non-collapse, max | NC , and limit state frequency, LS .
5.7.5 Discussion
The drift hazard has been evaluated in this section by disaggregation of the drift response into
collapse and non-collapse parts. Various methods for drift hazard estimation stem from this
formulation, among which, several non-parametric, semi-parametric and fully parametric
methods haven been discussed. The non-parametric method uses numerical integration to
calculate the drift hazard. This method which relies on many runs at many closely spaced S a
stripes and hence many non-linear analysis runs, is the most reliable method for estimating the
drift hazard among the methods discussed in this chapter. Hence, its resulting drift hazard curve is
used in this chapter in order to measure the relative accuracy of the other methods, which in
practice can be implemented with fewer dynamic analyses.
It is demonstrated that the non-parametric drift hazard curve considering the collapse cases is
identical to the one that is calculated not considering the collapse cases until it starts to diverge
and approach its minimum asymptotic value, the annual frequency of exceeding the collapse limit
state (Figure 5-10). The semi-parametric method is very similar to the non-parametric method; it
differs only in that it estimates the non-collapse CCFD of demand given spectral acceleration by a
lognormal CCDF. As it is demonstrated in Figure 5-11, the resulting semi-parametric drift hazard
curve is nearly identical to the non-parametric curve, proving that the lognormal assumption for
the non-collapse CCDF of the response is a reasonable assumption in this case. A fully parametric
closed-form solution for the drift hazard is presented at the end of this section. This formulation is
parallel to the DCFD formulation presented in the previous chapter. Hence, it is going to be used
in the next section for factored demand estimation.
5.8 Factored Demand
Factored demand, F .D. , was introduced in Chapter 4 as a part of the DCFD formulation for
seismic assessment. When displacements are in the region of the global instability in the
structure, factored demand can be estimated by parametric, semi-parametric, pseudo-parametric
and non-parametric methods.
Recall that we also observed in the previous chapter under the assumptions made there that the
definition of factored demand for a given tolerable probability, p o , is just equal to the drift value
from the drift hazard curve corresponding to a hazard value equal to, p o . The assumptions
leading to the simple analytical DCFD formulation derived in Chapter 3 are not necessarily valid
in the region of global instability. Hence, we propose here to use the alternative interpretation of
factored demand, i.e., that which relates to drift hazard curve as a general definition of the
factored demand.
5.8.1 Factored demand: Non-parametric using the non-parametric
drift hazard
The non-parametric factored demand is calculated, using the general definition proposed above of
factored demand, from the non-parametric drift hazard curve (Figure 5-10). According to this
definition, the factored demand for an allowable limit state probability, p o , is equal to the drift
value corresponding to a drift hazard equal to p o . This implies that the factored demand
associated with a tolerable failure probability of p o is equal to the drift demand with a mean
annual frequency of exceedance of max = p 0 (recall annual probabilities and mean annual
frequencies are numerically equivalent in this range). Further this definition means that the non-
parametric factored demand curve plotted versus the acceptable probability values is the same as
the non-parametric drift hazard curve in Figure 5-10. It should be noted that the factored demand
in this case is read from the x-axis. Figure 5-14 is an example of such a curve re-labeled to
emphasize this difference.
5.8.2 Fully-parametric using IDA procedure with simple DCFD

formulation
The fully-parametric method for demand estimation was introduced in previous sections. In its
pseudo-parametric form, this method provides a simple two-parameter lognormal approximation
to the conditional distribution of demand given spectral acceleration even when there are collapse
cases in the data set. Although this approach is not potentially favorable (for reasons discussed
before), it is useful because it permits estimation of factored demand from the simple DCFD
formulation (Equation 4-7):
1 k 2
max | S a
Factored Demand = | P0 S
e2 b (5 - 26)
max a
As described in the pseudo-parametric demand estimation method, we begin by deriving the full
parametric 16th, 50th and 84th percentiles (Equation 5-10, the three-parameter distribution) of the
drift response given spectral acceleration obtained using parametric IDA curves (Section 5.6.2.2),
although other method such as MSA might also be used. Then, we use the full-parametric
percentiles in order to estimate the median drift for a given spectral acceleration, max |S a , the
conditional fractional standard deviation, max |S a , and the local b value. In order to estimate the
local k value and the spectral acceleration hazard, S a (.) , we use the non-parametric spectral
acceleration hazard curve.

0
10
Factored demand, using numerical integration
Factored demand, using parametric IDAs
Factored demand, using numerical integration (considering the collapse cases)
1
10
max

2 LS=0.0078
10
3
10 3 2
10 10
max
Figure 5-13 - Fully parametric factored demand estimation: Simple DCFD formulation using the
parametric IDA curves and non-parametric spectral acceleration hazard
Here, we have estimated pseudo-parametric factored demand as described above and have plotted
it in Figure 5-13. The non-parametric factored demand calculated by numerical integration is also
plotted in the figure. We can observe that the pseudo-parametric factored demand stops at a drift
value close to 0.024. This is due to the fact the parametric 50th percentile is only available up to a
spectral acceleration level of 0.7g. However, as long as the pseudo-parametric estimation is
available, there is close agreement between the two estimates for factored demand. This
agreement looks promising, considering that the pseudo-parametric factored demand can be
calculated with relatively little analysis effort compared to non-parametric factored demand
estimates. The relatively small analysis effort is attributed to the fact that the non-collapse parts of
the IDA curves in the logarithmic scale are approximated with lines by performing regression. It
is expected that the regression can result in reasonable estimates for the non-collapse IDA curves
using fewer points than the several stripes of multiple-stripe analysis (as it is used for non-
parametric factored demand estimation). Nonetheless, the method fails in the near-collapse
regime (drifts larger than 1.5%) of primary interest here.
5.8.3 Factored demand: Semi-parametric using parametric IDA
curves
We can employ the closed-form solution for the drift hazard considering the collapse cases in
Equation 5-23, in order to calculate the factored demand. Figure 5-4 illustrates the full-parametric
16th, 50th and 84th IDA percentiles given non-collapse in thin lines. These non-collapse IDA
percentiles are used to estimate the local median, max | NC, S a , fractional standard deviation of drift
given spectral acceleration, NC , and b NC value. The non-parametric spectral acceleration hazard
curve is used to estimate the local k value and the spectral acceleration hazard, S a (.) .
We have incorporated the above estimations in Equation 5-23 and have plotted the resulting drift
hazard in Figure 5-13. We have also plotted the non-parametric drift hazard curve for
comparison.
The semi-parametric drift hazard curve illustrated above shows good agreement with the non-
parametric drift hazard curve, especially for the drift values that are close to the range of global
dynamic instability. However, the semi-parametric factored demand curve has a lower right
asymptotic limit compared to the non-parametric factored demand. Similar to pseudo-parametric
factored demand estimation in the previous section, the semi-parametric method uses linear
regression to approximate the non-collapse portion of the IDA curves with lines in logarithmic
scale. However, the semi-parametric method provides slightly better estimates for the factored
demand compared to the pseudo-parametric method, since it is applying the (more elaborate)
closed-form solution for drift hazard considering the collapse cases (Equation 5-23) instead of the
simple DCFD formulation (Equation 5-24).
Factored Demand
0
10
Calculated from ClosedForm Drift Hazard using IDA
o
1
Allowable Probability, p
10
2
10
Factored demand, nonparametric (numerical integration)

Factored demand (drift hazard), using parametric IDAs
(closedform for drift hazard considering collapse
3
10 3 2 1
10 10 10
Factored Demand
Figure 5-14 - Semi-parametric factored demand estimation: Using the closed-form for drift
hazard in Equation 5-23 considering the collapse cases, with the parametric non-collapse
(linear) IDA curves and non-parametric spectral acceleration hazard
5.8.4 Factored demand: Semi-parametric using single and double-

stripe analysis
We have used the efficient single-stripe analysis in the previous chapter in order to estimate the
factored demand for single limit state reliability assessments. This was possible since the drift
hazard equation derived in the Chapter 2 (Equation 2-25) was invertible with respect to the drift
values, and that resulted in an explicit formulation for the factored demand as a function of the
allowable probability (Equation 5-26). Nevertheless, the single stripe analysis could not provide
an estimate for the slope parameter, b. It was observed that the single-stripe analysis could
provide accurate estimates of the factored demand provided that a reasonable value for the b-
value was assumed. In this chapter, the estimation of factored demand is further complicated by
the consideration of the collapse cases. The closed-form expression derived for the fully
parametric drift hazard evaluation, Equation 5-23, is not invertible with respect to the drift value,
y. Also, the number of parameters in the formulation increases. For example if we try to estimate
the factored demand using a single-stripe analysis, we would have to assume value for the b value
as well as two additional parameters that define the fragility function for the collapse capacity.
Nevertheless, provided with the required parametric estimates, one could calculate the drift
hazard curve from Equation 5-23 and then use the drift hazard curve to calculate factored demand
as the drift value corresponding to a hazard value equal to p o , i.e., factored demand for allowable
probability, p o . A double-stripe analysis (as per Chapter 3) may prove to be an adequate
estimation procedure, as with well chosen S a values it can provide at least local estimates of all
the six parameters necessary to calculate the drift hazard from Equation 5-23. In order for the
estimates to be local, they need to be in the vicinity of p o . Therefore, this approach may involve
an additional iteration, and hence an additional single-stripe analysis. This method is called semi-
parametric since it incorporates the non-parametric spectral acceleration hazard curve in the
parametric formulation for the drift hazard curve. We outline as a step-by-step procedure for
finding the factored demand using a double-stripe analysis.
o Determine the tolerable probability level, p o .
o Estimate the k value. The k value can be estimated by the local (log) slope of the hazard
curve in the vicinity of the allowable probability value, p o .
o Calculate the spectral acceleration corresponding to p o from the spectral acceleration
Po
hazard curve. Take the spectral acceleration of the first stripe equal to, sa .
Po
o Perform a single-stripe analysis at sa .
o Check to see if there are collapse cases in the response. Two cases can happen:
o There is no collapse cases observed in the response. calculate the drift hazard
from the parametric expression for drift hazard given non-collapse (Equation
5-24). This requires assumption of a value for the slope parameter, b. Since there
is no collapse case in the response, one may assume that the b value is equal to
one, as per the equal displacement rule. In order to estimate the drift hazard in the
vicinity of collapse one needs to repeat this first single-stripe analysis with a
larger spectral acceleration value. This value should be large enough so that
collapse cases are observed in its response.
o There are collapse cases observed in the response. Find the conditional median
and the fractional standard deviation of the stripe response given non-collapse,
max | NC , S a ( Po s a ) and max | NC , S a ( Po s a ) , respectively. Proceed to the next steps.
o Estimate PNC |S a ( Po s a ) , the probability of non-collapse at a spectral acceleration equal to

Po
s a , as the fraction of the non-collapse cases in the stripe, PNC |S a ( Po s a ) .
o Choose the spectral acceleration of the second stripe, s a2 . Since we are interested in
making estimations close to the collapse regime, the placement of the second stripe
depends on the number of collapse cases observed in the response of the original stripe. If
there are relatively few collapse cases, the spectral acceleration of the second stripe can
be larger than that of the first stripe. If there are many collapse cases among the
responses, it is preferable to take the spectral acceleration of the stripe smaller than the
first stripe. In both cases, it is useful to make sure that the two stripes are sufficiently
spaced from each other in order to prevent the parameter estimates from being affected
by possible local peculiarities. However, the spectral acceleration of the second stripe
should be large enough so that there is at least one collapse case observed in the response.
This is because the additional information provided by the second stripe is also going to
be used for estimating the parameter estimates for the conditional probability of collapse
for a given spectral acceleration, namely, s a 0 and C . It should be noted that this
procedure is implementing the double-stripe method outlined in the previous chapter.
o Perform the second single-stripe analysis at, s a2 .
o Find max | NC , S a ( s a2 ) and max | NC , S a ( s a2 ) .
o Estimate PNC |S a ( s a2 ) as the fraction of the non-collapse cases in the second stripe,
PNC |S a ( s a2 )
o Estimate the a NC and b NC (the later value as the slope of the line connecting the
conditional median of the two stripes given non-collapse versus spectral acceleration, in
the logarithmic space):
max | NC , S a ( Po s a )
log( )
max | NC , S a ( s a2 ) max | NC , S a ( Po s a )
b NC = and a NC =
log(
Po
sa
) ( Po
sa )
bNC
s a2
o If the parameters s a 0 and C were known, the conditional probability of non-collapse

for the two stripes could be calculated from the following equations:
-C
Po s a
PNC |S a ( s a ) =
Po for Po
s a s a0
s a0

-C
s2
PNC |S a ( s a2 ) = a for s a2 s a 0
s a0

Since both stripes have collapse cases in them (as it is ensured in the preceding steps), their
corresponding conditional probabilities of non-collapse are going to be described by the
above equations that are specific to spectral acceleration levels higher than s a 0 . Estimates of
the left hand sides of the two equations are available from prior steps. Therefore, s a 0 and C
can be calculated by solving the above equations, yielding:
PNC |S a ( Po s a )
log( )
PNC |S a ( s a2 ) log( PNC |S a ( Po s a )) + C log( Po s a )
C = and s ao = exp( )
Po
sa C
log( )
s a2
o Calculate the drift hazard from Equation 5-23 for a set of drift values close to
max | NC , S a ( Po s a ) and max | NC , S a ( s a2 ) and plot the resulting drift hazard curve.
o Find the drift value that corresponds to a drift hazard equal to p o . This is the factored
demand for an allowable probability equal to p o . This is only possible if the range of the
drift hazard values is wide enough to contain p o .
o In case the factored demand calculated is too far from the drift values, max | NC , S a ( Po s a )
and max | NC , S a ( s a2 ) , then another iteration or at least another single-stripe analysis may be
necessary.
As an example, we have used the procedure outlined above to calculate the factored demand for
an allowable probability of, p o = 0.0088 , which corresponds to a spectral acceleration equal to,
po = 0.0088
s a = 0.70 g . First, we have obtained the local parameter estimates in the expression for
drift hazard (Equation 5-23), following the steps above. The second stripe was chosen at
s a2 = 0.60 g . The factions of collapses in these two stripes were 0.4 and 0.13 (Figure 5-2),
respectively. Estimated parameter values are max | NC , S a ( Po s a ) and max | NC , S a ( s a2 ) . We have
plotted in Figure 5-15 the drift hazard for drift values near drift values, max | NC , S a ( Po s a ) and
max | NC , S a ( s a2 ) . The final step is to find the drift value that corresponds to a hazard value equal to
p o = 0.0088 . It is estimated to be 0.018 (Figure 5-15, heavy line). If we use the non-parametric
method to calculate the factored demand corresponding to an allowable probability of

p o = 0.0088 , we get a value equal to 0.025 (Figure 5-15, light line). We can enhance the former
estimate by selecting a new s a level higher than 0.7g. However, in this case study, the median
conditioned on no collapse displays abrupt changes at spectral accelerations larger than 0.70g
(Figure 5-6). This will cause the resulting drift hazard curves to be less reliable for intensity
levels larger than 0.70g.

0
10
Factored demand, nonparametric, considering collapse cases
Factored demand, semiparametric using doublestripe analysis

considering the collapse cases
1
Allowable Probability, p0
10
second stripe at Sa=0.60 g
first stripe at Sa=0.70 g
2
10
Po=0.0088
F.D.(po)=0.0183
3
10 3 2 1
10 10 10
Factored demand
Figure 5-15 - Double-stripe, semi-parametric factored demand estimation: Using the closed-form
drift hazard in Equation 5-23 considering the collapse cases, with results for S a = 0.6 g and
S a = 0.70 g and non-parametric spectral acceleration hazard.
It can be observed that the drift hazard estimated using double-stripe analysis is fairly close to the
non-parametric drift hazard for a range of drift values in the vicinity of the conditional median
values for the two stripes given non-collapse, max | NC , S a ( Po s a ) and max | NC , S a ( s a2 ) , e.g., drift
values as large as 1.6%. The drift hazard curve starts to diverge from the non-parametric drift
hazard curve, as the drift values get closer to the collapse regime. It should be mentioned that this
description has presented that collapse could be recognized from single-stripe analysis. This
would be true if collapse is defined by non-convergence or an excessive (unreasonable)
displacement. If, however, it is defined by the shape of the IDA, then more than one analysis of
each record is needed to define the collapse capacity. Clearly these runs can be of value in the
over-all drift hazard and/or factored demand analysis. While a systematic method to search
efficiently for IDA-based capacities exists (Vamvatsikos and Cornell, 2001), there are no
algorithms as yet that treat global capacity estimation and drift hazard/factored capacity
estimation together in a coordinated manner.
5.8.5 Discussion
Several methods for estimating the displacement-based demand (the engineering demand
parameter) in the collapse regime were discussed. These methods employ the proposed
generalized definition of the factored demand. In this definition, the factored demand for the
allowable probability, p o , is equal to the drift value that has an annual frequency of exceedance
(hazard), max , equal to p o . Based on this property, any drift hazard curve is also a factored
demand curve where the hazard axis is the same as the allowable probability axis and the drift
axis is the same as the factored demand axis. The non-parametric method for estimating the
factored demand makes use of the above property to estimate the factored demand from the non-
parametric drift hazard curve considering the collapse cases. This is the most accurate method for
demand estimation studied in this chapter. Other methods considered include the fully-parametric
method that uses the pseudo-parametric IDA percentiles to obtain the required parameter
estimates in order to calculate the factored demand from the simple DCFD format (see Chapter
4). Comparing the results of the fully parametric method with the non-parametric one, we can
observe that the fully-parametric method provides a smooth and accurate approximation to the
non-parametric results as long as a fully-parametric estimate for median is available (Figurer 5-
13). The relatively small computational effort involved in the fully parametric method together
with the simplicity of the DCFD formulation (Equation 5-26), makes this an efficient alternative
to use.
The next method studied in this section is the semi-parametric method using the fully parametric
IDA curves. This method is very similar to the one studied before apart from the fact that it
incorporates the local parameter estimates derived from the parametric IDA curves into the
formulation for the drift hazard considering the collapse cases (Equation 5-23). This method
better captures the close-to-collapse part of the drift hazard. Finally, a semi-parametric method
using a minimal number of stripes is discussed. This method is being presented in the form of a
step-by-step procedure for the local estimation of the drift hazard/factored demand. The
advantage of this method is that it uses few stripes (typically two) to obtain parameter estimates
for the drift hazard curve considering the collapse cases. This method is not very accurate in this
example.. The inaccuracy can be attributed to the less-than-smooth behavior of the given no
collapse median at intensity levels larger than 0.7g.
5.9 Factored capacity
Similar to factored demand, the factored capacity was defined in Chapter 4 in the context of the
demand and capacity factored design as:
factored capacity = median capacity the capacity factor
Alternatively, the factored capacity can be defined as the drift value with a frequency of
exceedance equal to the limit state frequency. In other words, the factored capacity is the drift
value corresponding to a frequency equal to the limit state frequency from the drift hazard curve.
For example, the factored capacity derived from the drift hazard considering the collapse cases
(e.g., Figure 5-10) lies at infinity. This is because the frequency of the collapse limit state is the
right asymptotic value for the drift hazard considering the collapse cases. The drift hazard
considering the collapse cases should however be used for the estimation of factored capacity
values that correspond to limit states other than collapse, called here the exogenous limit states,
e.g., the immediate occupancy limit state or a local member failure limits state short of collapse
due to global instability. It should be noted that the same information that is used to derive the
drift hazard considering the collapse cases is used to calculate the collapse limit state probability.
Thus, the drift hazard considering the collapse cases cannot be used to estimate the factored
capacity for the collapse limit state. It should be noted that this definition of factored capacity,
although it is theoretically sound, has very little practical use. However, there are many ways to
define the factored capacity as long as it is equal to factored demand at P0 = P f . Therefore, there
still is room for introducing more practical propositions for factored capacity.
5.10 Practical Applications
The methods proposed in this chapter can be used for reliability assessments for exogenous
limit states, (i.e., other than global instability induced collapse). The formulation presented in this
chapter already takes into account this collapse information yielding the probability of limit state
as an asymptote of the drift hazard curve (Figure 5-10). Reliability assessments for a single limit
state can be performed by obtaining local parameter estimates. By local, we mean a region that is
close to the onset of the limit state under investigation. In order to ensure that the estimates are
obtained close to the onset of the limit state under consideration, a re-iteration of the analyses
might become necessary especially if the structure is in the global collapse sensitive regime. It is
shown that in this regime these local estimates can be provided by means of performing a double-
stripe analysis. A procedure for both locating and performing the double-stripe analyses is
discussed. It was demonstrated in the previous chapter that the single-stripe analysis could be
used for reliability assessments for a single limit state, provided that a reasonable value for the
slope parameter b could be assumed. However, acquiring a (local) estimation of parameter b
required at least a second stripe. If the collapsing cases are considered in the modeling, there will
be more parameters to estimate compared to the simple DCFD formulation derived in Chapter 3.
Also, unlike in the previous chapter, the formulation of drift hazard considering the collapse
information is not invertible with respect to the drift values. These factors add to the difficulty of
reliability assessment based on the results of a single stripe analysis. Performing a double-stripe
analysis can be expected to enable us to estimate all the necessary parameters in order to estimate
the drift hazard or factored demand in the region of global instability.
Reliability assessment for multiple limit states requires the use of wide-range methods such as the
MSA or the IDA procedure (as presented in Chapter 4). In the case of IDA procedure, it is
demonstrated here that the analysis effort can be reduced by approximating the individual IDA
curves by parametric functions. Parametric approximations of IDA curves are obtained by
implementing the results of linear regressions performed on the individual IDA curves into the
three-parameter formulation. The number of points required for each IDA curve in order to get
these parametric estimates can be as low as 4 points. The first two points are needed in order to
estimate the non-collapse intercept and slope of the IDA curve in the logarithmic space. A
minimum of two additional points in needed in order to find the global stability limit state
capacity point according to a procedure proposed by Vamvatsikos and Cornell, 2002, for locating
the capacity points. This procedure uses spline fits to the IDA curves in order to predict the limit
state capacity point (see Vamvatsikos, 2002).
5.11 Summary and Conclusions
The conditional probability distribution of drift given spectral acceleration is developed by using
the total probability theorem in the range of drift values close to collapse (Equation 5-6). Collapse
is defined as the onset of global dynamic instability in the structure marked by the point where
the slope of the IDA curve (see Chapter 4, or Vamvatsikos and Cornell, 2001) is reduced to a
certain percentage of its initial slope. A proposed parametric representation of this conditional
probability distribution is called the three-parameter distribution (see Shome and Cornell, 1999)
in order to account for the fact that the probability distribution of drift given spectral acceleration
prior to the occurrence of collapse can be modeled by a lognormal (two-parameter) distribution
and the fraction of the collapse cases observed in the data-set can be defined by a third parameter.
Nonlinear dynamic procedures are used with the derived formulation in order to estimate the
distribution of the drift given spectral acceleration in the collapse range. The alternative methods
for the estimation of the probabilistic distribution of drift response given spectral acceleration can
be differentiated based on whether they are derived fully-parametrically, semi-parametrically or
non-parametrically. As a special application of the three-parameter distribution, parametric IDA
curves and the parametric IDA percentiles are derived for the drift values close to collapse.
Alternatively, the results of multiple-stripe analysis (see Chapter 4) are post-processed in order to
be implemented in the three-parameter distribution. The resulting semi-parametric percentiles
show good agreement with the counted percentiles that were derived in the previous chapter
without special consideration of collapse cases. The distribution of drift demand for a given
spectral acceleration can also be estimated empirically (non-parametrically) based on the results
of multiple-stripe analysis. The advantage of estimating the distribution non-parametrically is that
it avoids prior assumptions about the shape of the distribution.
The alternative drift demand estimation methods discussed above are then used in order to derive
the annual frequency of exceeding a certain drift value. The result is known as the drift hazard.
The non-parametric probability distribution for the drift demand is used to calculate the drift
hazard by numerical integration. The result is used as a standard by which the estimates provided
by other approaches are compared. Similar to the demand estimation procedures studied in this
chapter based on the three-parameter distribution, the alternative methods for the estimation of
the drift hazard can be differentiated based on whether they are derived fully parametrically,
semi-parametrically or non-parametrically. Parallel to the closed-form solution for the drift
hazard in the previous chapter, a fully parametric closed-form for the drift hazard is derived. In
practice it is desirable to avoid the close multiple stripes used in this research effort as they
require intensive computational effort. As shown in Chapter 4, the single-stripe analysis can be
used to estimate the drift hazard from the derived closed-form for a specific drift value, provided
that the displacement-based response is not in the region of global instability. It is demonstrated
here that utilizing the carefully designed double-stripe analysis in the closed-form solution for
drift hazard can provide fairly good local estimates of the drift hazard, in the large drift regime
when a large fraction of the records here cause collapse.
The factored demand for a given allowable probability, p o , can be defined as the drift value with
p o probability of exceedance (see Chapter 3). This definition is used to estimate the factored
demand from the estimated drift hazard curve(s). Therefore, factored demand can be estimated
non-parametrically from the non-parametric drift hazard curve derived by numerical integration.
The parametric IDA percentiles for drift values close to collapse can be used to estimate the
factored demand from the simple (two-parameter) closed-form formulation for DCFD format
derived in Chapter 3 (Equation 5-26). Factored demand calculated for multiple allowable
probability levels, p o , is compared to the non-parametric factored demand. The results show very
good agreement as long as the parametric IDA percentiles are still available. Alternatively, the
parametric IDA curves can be incorporated in the closed-form drift hazard derived based on the
three-parameter distribution (Equation 5-23). The resulting factored demand calculated for
multiple allowable probability levels have been compared to the corresponding non-parametric
estimates for factored demand and the results indicate good agreement. This agreement between
the parametric factored demand derived from the closed-form drift hazard curve and the non-
parametric results is particularly desirable since the parametric factored demand can be obtained
with considerably less analysis effort than the non-parametric factored demand. In an attempt to
estimate the drift hazard locally, a double-stripe method (see Chapter 4) was used. The double-
stripe method was used in order to provide parameter estimates for the drift hazard in Equation
5-23. The resulting estimate for drift hazard by following this approach was fairly close to the
non-parametric results. Nonetheless, this method needs to be tested for a variety of ductility
ranges in order to test its robustness.
Chapter 6
Non-linear Response Dependence on Ground Motion
Characteristics with Implications for Ground Motion
Record Selection
6.1 Abstract
Ground motion record selection is considered a critical problem in assessment of structural

response based on numerical dynamic analysis. The structural engineer expects the ground
motion specialist to select records whose magnitudes, distances, site conditions, and faulting style
are representative of the threat to the site. This may be accomplished with the use of
disaggregation (see McGuire, 1995). In some cases the records are selected to provide response
spectra that approximate the uniform hazard spectrum or other design response spectrum. The
properties (natural period, etc) of the structure may or may not be considered in the record
selection. The choice of the ground motions may be affected by the interface variable used to
measure the intensity of ground motion, known in PEER as the intensity measure IM. According
to criteria proposed by Luco and Cornell (2001), a preferred IM is both sufficient with respect
to the ground motion characteristics and also efficient. A sufficient intensity measure renders
the structural response conditionally statistically independent of other ground motion
characteristics such as event magnitude, while an efficient intensity measure predicts the
structural response with (relatively) small record-to-record variability. Theoretically, careful
record selection is not essential if the intensity measure is demonstrated to be sufficient with
respect to ground motion characteristics.
The spectral acceleration of the first-mode frequency (FMF) is commonly used as a scalar-based
intensity measure in probabilistic seismic assessments. Shome and Cornell (1999) have
demonstrated that spectral acceleration at FMF is sufficient with respect to the ground motion
characteristics for moment resisting frame structures with FMFs in the moderate period range. In
this study, the nonlinear dynamic displacement responses of two extreme cases of short-period
SDOF and long-period MDOF structural systems are presented. The short period system is a
generic SDOF bi-linear model with T=0.1 sec; and the long-period system is a 20 story moment-
resisting steel frame structure (LA 20 structure, see Luco, 1999) with T1=3.98 sec. For both cases,
it is demonstrated that the FMF spectral acceleration renders the maximum inter-story drift
response effectively conditionally independent of ground motion characteristics such as
magnitude, source-to-site distance, and normalized residual (epsilon) of the attenuation
relationship.
It is of interest to investigate the sufficiency of the FMF spectral acceleration over a range of
ductility levels and hence nonlinear behavior. Therefore, the yield strength level for the SDOF
system is adjusted in order to achieve structural response in a desired range of ductility. It is
demonstrated that the FMF spectral acceleration is also sufficient for the SDOF system with
modified yield strength and hence ductility level of the response. As an alternative to adjusting
the yield strength in the case of the SDOF system, the 20-story structure is subjected to ground
motion records that are scaled so that they render the structural response in a desired range of
ductility. It is also demonstrated that the FMF spectral acceleration is effectively sufficient based
on the response of the 20-story structure to the scaled ground motion records. These studies
suggest that the sufficiency of the FMF spectral acceleration is maintained at various levels of
linear and non-linear response and ductility.
For these two systems this sufficiency conclusion is not necessarily to be expected. Their non-
linear responses are expected to be particularly spectral shape sensitive and that shape is thought
to be magnitude dependent. Therefore, response is expected to be magnitude dependent. This
logic is studied step-by-step. For both example cases, dependence of the response, once it is
conditioned on the FMF spectral acceleration, on the shape of the elastic response spectrum is
investigated. The effect of the shape of the spectrum is measured by the spectral acceleration at
period T normalized to the FMF spectral acceleration, also known as the spectral shape factor at
period T. This is equivalent to measuring the relative efficiency of a vector-based intensity
measure consisting of FMF spectral acceleration and spectral shape factor at period T. In the case
of the 20 story structure (with long period), the response demonstrates significant dependence on
the spectral shape factor at a period close to the second mode frequency (SMF). In the case of the
SDOF system (with short period), the response is significantly dependent on the spectral shape
factor at a period larger than (about two to three times) the period at FMF. These observations are
confirmed for different ductility ranges by studying SDOF systems with varying yield strength
and by scaling the ground motion records in the case of the LA20 structure.
On the other hand, it is expected that the spectral shape factor demonstrate magnitude-
dependence. This hypothesis is investigated based both on the attenuation predictions for the
shape factor as a function of magnitude and also on the relation between magnitude and the
spectral shape factor of the ground motions used in this study. As stated above, if the shape-factor
is magnitude dependent, one would expect that the structural response conditioned on the FMF
spectral acceleration be also dependent on magnitude. This would explain why ground motion
magnitude is conceived as the most likely ground motion variable to affect the structural response
conditioned on the FMF spectral acceleration. However, the established sufficiency of the FMF
spectral acceleration with respect to magnitude seems to contradict such presumptions. It is
argued that the large variability in both the relation of response to shape factor and also the
relation of spectral shape to magnitude may explain the observed lack of conditional dependence
on ground motion magnitude.
The sufficiency found here will not necessarily carry over to all cases. Sufficiency depends on the
IM, the structure, the response, and the ground motion parameters. Ground motion parameters
such as site amplification and/or directivity may prove particularly troublesome because they may
imply strong sensitivity of spectral shape to certain ground motion parameters (impedance ratio,
X cos( ) ,M, etc.). If for a specific structure, the proposed intensity measure is not sufficient
with respect to ground motion characteristics, it will be argued that the results of probabilistic
seismic hazard analysis (specifically via site-specific seismic hazard disaggregation) can be used
to appropriately adjust the response statistics for a given selection of records to correct for
possible dependencies on ground motion characteristics, at least within certain limitations.
6.2 Keywords
Intensity measure, probabilistic seismic hazard analysis, moment magnitude, nonlinear dynamic
analysis, attenuation relationships, scaling, ground motion record selection
6.3 Introduction
The probabilistic seismic assessment of structures normally requires performing a set of time-
history analyses on a suite of ground motion records. This will provide an estimate of the central
value and the variability in structural response under seismic excitations. The estimated
distribution of structural response is dependent on the suite of ground motions used for the time-
history analyses. Ideally, these ground motions should represent the major earthquake scenarios
that can happen for a particular site condition. However, finding a selection of records that
represents all the possible earthquake scenarios and their relative likelihood for a site can be
extremely cumbersome; in fact it is not quite clear precisely what a representative record
selection is and in any case it may change with return period or IM level. Nevertheless,
depending on the performance level under consideration, only a subset of the possible earthquake
scenarios are going to be significant in terms of the probability of their occurrence. If the
significance of scenario earthquakes for a given performance level is measured by probabilistic
criteria, an efficient record selection would consist of choosing records representing those
scenario events with the largest relative contribution to the probability of exceedance (hazard)
corresponding to that performance level. This objective is approximated in some current practice
by disaggregation of spectral acceleration hazard with respect to M , R rup (see Somerville et al.,
1998 and Stewart et al. 2001). On the other hand, the statistical properties of structural response
to a given selection of records can be adjusted by weighting the response based on the actual
(relative) contribution of each record to the hazard corresponding to a particular performance
level.
The strength of ground motion is represented by a variable called the intensity measure in the
certain current earthquake engineering literature. The intensity measure IM serves as an interface
variable that implicitly represents the ground motion characteristics such as magnitude and
distance in the seismic assessment of a structure. Most common choices for the intensity measure
are the peak ground acceleration (PGA) and the spectral acceleration for a specified oscillator
frequency and damping. Luco and Cornell (2001) proposed probability-based and structure-
specific criteria for the selection of the ground motion intensity measure variable. These criteria
gauge the success of a candidate intensity measure in terms of how small the bias and variability
in the resulting structural response is going to be. According to these criteria, a preferred
candidate for the intensity is both efficient and sufficient with respect to the selected
structural response variable. An efficient intensity measure results in relatively small variability
in the structural response; while the sufficient intensity measure renders the response
conditionally statistically independent of the ground motion characteristics. It will be
demonstrated that the criteria for record selection will be significantly simplified when the
selected intensity measure is sufficient with respect to the structural response variable.
This chapter demonstrates how sufficiency of FMF spectral acceleration with respect to ground
motion characteristics can be established based on a suite of ground motion records. Shome and
Cornell (1999) have demonstrated that the FMF spectral acceleration is sufficient with respect to
ground motion characteristics for first-mode dominated moment resisting frame structures with
moderate FMFs. Hence, this chapter focuses on structures with extreme FMFs and on structures
where the second-mode is expected to be important. As example cases, the sufficiency of spectral
acceleration is investigated for two structures, a 20-story steel moment-resisting frame with
T1 = 3.98 , and a generic SDOF system with T = 0.1 . For these example cases, the sufficiency
of FMF spectral acceleration is investigated based on un-scaled ground motion records. In
addition to the ground motion characteristics, the dependence of the response on the shape of the
elastic acceleration spectrum is also investigated. It is demonstrated that the structural response
conditioned on the FMF spectral acceleration (also a measure of the amplitude of the spectrum at
FMF) is dependent on the shape of the spectrum. The effect of the shape of the spectrum is
measured by the spectral acceleration at period T normalized by the FMF spectral acceleration,
S a (T ) S a (T FM ) , also known as the spectral shape factor at period T. For the 20-story structure,
the displacement-based response of the structure is expected to exhibit second-mode dependence,

where as for the high-frequency SDOF system, the displacement-based non-linear response is
expected to exhibit dependence on the spectral shape at frequencies smaller than FMF. Hence for
both cases, the displacement-based response is expected to depend on the spectral shape factor at
a period T different from T FM .
It will also be demonstrated that the spectral shape factor at a period T may demonstrate some
degree of dependence on the moment magnitude. Such dependence is studied both from the point
of view of the attenuation predictions for spectral shape as a function of magnitude and also from
the observed relation between spectral shape and magnitude for the suite of records used in this
study. The dependence of the displacement-based response on the spectral shape factor together
with the dependence of the spectral shape factor on magnitude suggest that the displacement-
based response may in turn exhibit dependence on moment magnitude. This hypothesis is a basis
for expecting that the moment magnitude be the most likely ground motion characteristic variable
to be correlated to the structural displacement response conditioned on S a (TFM ) . Such hypothesis
is investigated in detail for both structural systems as part of the sufficiency investigations.
A suite of simple linear regressions together with tests of hypotheses is used for evaluating the
sufficiency of S a (TFM ) with respect to ground motion characteristics and also measuring the
dependence of the response on the shape of the spectrum. However, a more thorough
investigation may use multiple-variable regression for investigating the sufficiency of S a (TFM )
with respect to ground motion characteristic variables M and R rup , simultaneously.
In order to investigate the sufficiency of S a (TFM ) for higher ductility levels, the ground motion
records are scaled by a factor of three in the 20-story example case. For the SDOF structure, the
sufficiency of S a (TFM ) is investigated for different ductility levels, by varying the yield strength.
In both cases, it is observed that the sufficiency of S a (TFM ) is maintained over a broad range of
non-linear behavior and ductility. However, this conclusion may not always be true.
Once it is demonstrated how to establish sufficiency for a scalar IM, the advantages of a choosing
a sufficient intensity measure are discussed. It is argued that more-or-less arbitrary selection of
record is justified, if the intensity measure is sufficient with respect to those ground motion
characteristics that may affect the response.
What happens if sufficiency is not established? The last sections of the chapter focus on cases
where the intensity measure is not sufficient. This chapter demonstrates that the response
statistics can be adjusted in order to take into account possible dependencies on ground motion
characteristics. A procedure for post-processing the structural response based on available hazard
disaggregation (see Bazzurro and Cornell 1999) results is presented. This procedure uses the
disaggregation results as weights in order to re-adjust the structural response in order to take into
account dependencies on ground motion characteristics.
6.4 Records used in this chapter
In the study that follows 47 horizontal ground motion records have been extracted from the PEER
strong motion database. The records are all California earthquakes recorded on deep soil
(Geomatrix soil types C and D). There have been no particular criteria in choosing the records
apart from avoiding the recordings with potential near-sources effects (e.g., directivity) in them.
The selected records have magnitudes between 5 and 7.5 and the source-to-site distance from 15
km to 120 km. A complete list of the ground motion records and their characteristics can be found
in Appendix C. We will refer to this selection of records as combined selection, since it is in
fact the combination of the set of 30 ground motions used in Chapters 4 and 5 and another 17
ground motions. The reason for adding the extra 17 ground motions was to increase the range of
the corresponding magnitude values, and hence to make the selection more suitable for studying
possible dependencies on ground motion magnitude, i.e., for studying sufficiency.
6.5 Ground motion characteristics
The spectral acceleration at the period T, S a (T ) , can be uniquely determined in terms of what
shall be called the ground motion characteristics: magnitude M, source-to-site distance R rup , and
a normalized regression residual, , by an attenuation relationship:
ln S a (T ) = f ( M , R rup ) + ln S a (6 - 1)
where f ( M , R rup ) is a function that produces the median (more precisely, expected value of the
log) of all records for the scenario earthquake M and R rup and it referred to as the attenuation
law. It is typically obtained by regression. is the residual of the attenuation law regression
normalized with respect to the dispersion in the attenuation law prediction, ln S a . It should be
noted that the dispersion in the attenuation law may also depend on the ground motion
characteristics. This normalized residual is widely known as the epsilon of the attenuation law
in the literature. In this study, possible dependencies with respect to the following three ground
motion characteristics, magnitude M, closest (source-to-site) distance R rup , and the epsilon (the
normalized attenuation residual), , are going to be investigated.
6.6 The spectral acceleration and the spectral shape factor
It is going to be demonstrated in this study that given the FMF spectral acceleration, the
displacement-based response of the structure is dependent on the shape of the elastic acceleration
spectrum. The spectral acceleration at a period other than FMF is usually normalized with respect
to S a (TFM ) in order to reflect the shape of the spectrum. Hence, the spectral shape factor at the
period, T, is denoted by, R(T , TFM ) and is defined as the ratio of the spectral acceleration at
period T to the FMF spectral acceleration:
S a (T )
R (T , T FM ) =
S a (T FM )
The spectral shape factor can be used to demonstrate that the displacement-based response of the
structure for a given S a (TFM ) depends on the shape of the acceleration spectrum. This together
with the possible dependence of spectral shape on magnitude (discussed later in this chapter)
would seem to imply that the displacement-based response of the structure given S a (TFM ) is also
dependent on magnitude.
From a different point of view, it can be argued that the spectral acceleration values at periods
other than the FMF may also contain ground motion characteristic information that is not already
relayed by S a (TFM ) . Hence, if they are paired with S a (TFM ) , the ensemble may better describe
the record-to-record variability in the displacement response of the structure (the efficiency
criterion). Therefore, establishing the dependence of the structural response given S a (TFM ) on
the spectral shape factor is equivalent to establishing that the pair of S a (TFM ) and R(T , TFM ) as a
vector-valued intensity measure is more efficient than S a (TFM ) only.
6.7 Maximum inter-story drift angle: The displacement-based response parameter
In this study, the efficiency and sufficiency criteria are investigated based on the predictions for
the non-linear dynamic displacement response of the structure, provided by the candidate
intensity measure. Hence, establishing the efficiency and sufficiency criteria is dependent on the
choice of both the intensity measure and the structural response variable. Maximum inter-story
drift angle denoted by max is chosen as such dynamic displacement-based response variable. For
a moment-resisting frame structure, max is a global response parameter that may reflect the most
severe damage incurred by non-linear deformations in the moment connections. Since max is the
response parameter predicted by the intensity measure, it may also be referred to (later in this
chapter) as the dependent variable, wherever linear regression is used to make such predictions.
6.8 Residual-residual plot: A graphical method for measuring the significance of an
additional regression variable in predicting the structural response
Linear regression is a useful statistical tool for investigating efficiency and sufficiency criteria for
a candidate intensity measure (see Shome and Cornell 1999, and Luco, 2002). The displacement-
based response parameter, max , can be predicated as a function of the candidate intensity
measure, e.g., S a (TFM ) , by performing linear regression (usually in the logarithmic scale). The
efficiency of the candidate IM can be measured by the variability in the residuals of the
regression analysis. In order to establish sufficiency, the effectiveness of ground motion
characteristic variables as additional regression variables, i.e., in addition to S a (TFM ) , can be
investigated. In other words, ground motion characteristic variables cause very little improvement
in the regression prediction as regression variables in addition to a sufficient intensity measure.
This improvement may be judged by the reduction in the dispersion of the regression residuals
and/or the statistical significance of the regression coefficients corresponding to the ground
motion characteristic variables.
In this chapter, a simplified statistical approach based on regression is implemented for measuring
the effectiveness of ground motion characteristics as additional regression variables. This method
uses a graphical statistical tool known as the residual-residual plot. The main advantage of the
residual-residual plots is that they offer visual means for judging the improvement caused by an
additional regression variable.
Residual-residual plots are used in the applied statistics in order to investigate whether adding a
second variable to the regression improves the predictions compared to predictions based on the
regression on the first variable only. These plots are constructed by: a) performing regression of
the dependent variable (e.g., the displacement-based response, max ) versus the (first)
independent variable (e.g., S a (TFM ) ) b) performing regression of the second independent

variable (e.g., one of the ground motion characteristics) on also the first variable ( S a (TFM ) ) c)
plotting the residuals of the two regressions mentioned above against each other.
Roughly speaking, the two regressions on the first variable eliminate possible dependence of
both the dependent variable ( max ) and the second independent variable (one of the ground
motion characteristics) on the first independent variable ( S a (TFM ) ). This facilitates investigating
the potential dependence of the dependent variable on the second variable, by observing a
(statistically) significant trend, in the residual-residual plot, between the two sets of residuals
explained above. In order to simplify future references to these residuals, the residuals of the
regression of the dependent variable on the first variable are also called the response residuals,
and the residual of the regression of the second variable on the first variable are also called the
second-variable residuals. The possible trend between the two sets of residuals may be
observed by performing linear regression of response residuals on second variable residuals.
The significance of the trend measured by both the variability in the residuals of such regression
and/or testing the following hypothesis: The slope of the regression line is zero (i.e., test of
hypothesis, see Rice 1995). The significance of the slope is usually measured by a quantity
known as the p-value, assuming that the slope of the regression line is a random variable
described by Students t-distribution (see Rice 1995). The hypothesis is rejected (i.e., the slope is
non-zero) if the p-value is smaller than a certain (small) value, e.g., 0.01. If a two-variable
regression of the response on both the first and the second independent variable is performed, it
can be shown that the regression coefficient for the second variable is going to be the same as the
slope of the (simple) regression of the response residuals on second-variable residuals.
Moreover, the results of the hypothesis test are going to be identical.
As it is mentioned above, one way to measure the significance of an additional regression

variable is by measuring the amount of reduction it causes in the variability of data around the
regression prediction. The variability of data around the regression line can be estimated by the
(square) root of the mean of squares (RMS) of regression residuals, which is also referred to as
the sigma of the regression. It should be noted that the sigma of the residual-residual
regression is the same as that of the two-variable regression of response (dependent variable) on
the two independent variables, i.e., the two approaches are equivalent. The smaller is the ratio of
the sigma of the residual-residual regression (explained above) to the sigma of the regression of
dependent variable on the first variable (or original regression for brevity), the more significant
is the role of the second regression variable. In simple terms, an effective second regression
variable is going to explain part of the variability in the data that is not captured by the first
regression variable. In the context of sufficiency, this means that if the sigma of the residual-
residual regression is (significantly) smaller than the sigma of the original regression, this will
also confirm (in addition to the test of hypothesis) that the intensity measure is not sufficient with
respect to the ground motion characteristic in question. It also indicates that the two regression
variables together can provide a more efficient prediction of response, recalling that the
efficiency criterion is based on the variability of response for a given value of the intensity
measure.
The residual-residual plots and the test of hypothesis can also be used to as a visual way to
establish the sufficiency of S a (TFM ) with respect to ground motion characteristics. If the p-value
for the slope of the regression line on the residual-residual plot is near zero, it can be concluded
that S a (TFM ) is not sufficient with respect to the particular ground motion characteristic in
question. However, it should be noted that in this (simplified) approach the sufficiency of
S a (T FM ) is questioned for one ground motion characteristic at the time. This would ignore
possible interactions between the ground motion characteristics themselves. A more thorough
approach consists of performing a multi-variable regression of the displacement-based response
(the dependent variable) on S a (TFM ) and all of the ground motion characteristics in question, and
test the (joint) hypothesis of whether all the regression coefficients corresponding to the ground
motion characteristics are simultaneously zero. Nevertheless, it is believed that the residual-
residual plot approach is still effective in un-covering potential dependencies of response on
ground motion characteristics.
Residual-residual plots are also used in order to establish that the displacement-based response
given S a (TFM ) depends on the spectral shape factor R(T , TFM ) . In this case, R(T , TFM ) is the
additional regression independent variable whose effectiveness is studied. This is equivalent to
maintaining that the vector-valued intensity measure consisting of the two elements, S a (TFM )
and R(T , FM ) , is more efficient than the scalar intensity measure S a (TFM ) . The reduction in
the sigma of the residual-residual regression compared to the sigma of the original regression in
this case reflects the relative efficiency of the pair S a (TFM ) and R(T , TFM ) .
6.9 Establishing the sufficiency of S a (TFMF ) based on the un-scaled set of records
In this section, the residual-residual plots and hypothesis testing are used in order to investigate
the sufficiency of the scalar IM, S a (TFM ) with respect to ground motion characteristics.
The observations in this section are based on the ground motion records described in Section 6.4
and Appendix C. Two example cases are discussed. The first example is a 20-story steel moment
resisting frame with a long period ( T1 = 3.98 sec). The second example is a generic single-degree-
of-freedom system with a very short period ( T = 0.1 sec). These cases represent structural systems
with extreme FMF values, and hence with potential for demonstrating dependencies on ground
motion characteristics. For the 20-story structure, the displacement-based response of the
structure is expected to be second-mode dependent, where as for the high-frequency SDOF
system, the displacement based response is expected to exhibit dependence on the frequencies
smaller than FMF (e.g., due to period elongation). For both cases, the dependence of the response
on frequencies other than FMF is demonstrated by investigating the dependence of the response
on the spectral shape factor at a period T, after it is conditioned on S a (TFM ) . It is also
demonstrated that the shape of the acceleration spectrum depends on the moment magnitude of
the ground motion. This explains why the moment magnitude is a likely ground motion
characteristic variable to affect the displacement-based response after it is conditioned on
S a (T FM ) .
6.9.1 System with long first-mode period: LA 20 story Results
The 20 story ductile M1+building model used by Luco for near-source studies (Luco and Cornell,
2002) is a good example of a structure whose linear and non-linear dynamic behavior (as
measured by maximum inter-story drift ratio) is significantly influenced by higher modes.
Therefore spectral shape and magnitude can be expected to affect its response.
The 20-story steel moment resisting frame (SMRF) building was designed for Los Angeles
conditions as a part of the SAC steel project (Phase II), according to pre-Northridge earthquakes
provisions (see Luco, 2002). The structure is modeled using DRAIN-2DX and is denoted by
ductile M1+ in Lucos thesis (2002). The beam-columns are modeled by a bi-linear hysteretic
model (DRAIN-2DX, Element 02). This modeled is denoted by M1+ because the effect of the
interior gravity frames is considered by linking an equivalent gravity column to the two-
dimensional model of the perimeter frame. The p-delta effect is also considered in the analysis of
the frame.
Table 6-1 shows the periods corresponding to the first 5 mode shapes for the 20 story frame
structure. The spectral acceleration for T=3.9 seconds (close to T FM for the structure) and
damping of 5%, is chosen as the S a (TFM ) .
Mode Number 1 2 3 4 5
Period 3.96 1.35 0.78 0.55 0.41
Table 6-1- Modal periods for LA 20 M1+ model structure
The dynamic displacement response of the structure to ground motion records described in
Section 6.4, represented by maximum inter-story angle, max , is plotted in Figure 6-1 versus
S a (T FM ) on logarithmic paper. The linear regression prediction for ln max as a function of
ln S a (T FM ) is also plotted (the straight line in the two-way logarithmic plot). As mentioned in
Section 6.8, the regression of max on S a (TFM ) is labeled here as the original regression, as it
provides the basis for future comparisons when additional variables are used in order to make
predictions. (In this sentence and subsequently the fact that the (linear) regressions are on the
logarithm of the variables will be suppressed fro simplicity. The only exception is that magnitude
, not logarithm of magnitude is used in the linear regressions.)
It should be noted that the regression of ln max on ln S a (T FM ) in the logarithmic scale is

equivalent to fitting a power-law function in the form of, a S a b , to the maximum inter-story
drift angle data. It should be noted that b=1 corresponds to a linear relationship. This steel
moment-resisting frame is expected to behave linearly for maximum inter-story drifts less than
about %1. The value reported on the graph as, ln max |S a = 0.40 , is the (square) root of the mean
squares (RMS) of the residuals of the original regression (often called the standard error of
regression), also referred to as the sigma of the regression. This value, which is a measure of
the dispersion of the displacement response around the linear regression predictions (a straight
line in the logarithmic scale), is the basis of later comparisons when additional variables are
involved in the predictions. It should be noted that the displacement-based response is plotted on
the horizontal axis, as it is conventional in structural engineering. However the displacement-
based response is the dependent variable in the regression.
The residual-residual plots, introduced in Section 6.8, are used here in order to investigate the
sufficiency of the S a (T FM ) with respect to three ground motion characteristics, moment
magnitude M , rupture distance (horizontal distance to the projection of the rupture) Rrup , and the
epsilon (normalized residual of the attenuation relation), . For each ground motion
characteristic under consideration, the residuals of its regression on S a (T FM ) are going to be the
ordinates of points on the residual-residual plot. The abscissas, which are invariable with respect
to ground notion characteristic under consideration, are the residuals of the original regression
of the maximum inter-story drift (the displacement response) max on S a (T FM ) .
0
LA 20
10
47 Records with 5.0 < M <7.5 15< R <120
a = 0.0507
b = 0.6553
ln(
= 0.4009
)|S
max a 42
35
1
29 36
S at T =3.9 seconds [g]
10 46 34
22
33
45
37
127
26
18 15 47
6
2339
32
7 31
16 19
10 8 25
38
43 13
9 40
30
1224 17 4
2841
1
20
21 5
2
10 11 14
44
a
2 3
3
10
4 3 2
10 10 10
Maximum Story Drift Angle,
max
Figure 6-1 - The dynamic displacement response of LA20 model structure to the combined record
selection plotted versus the spectral acceleration for T=3.9 sec and %5 damping, S a (T FM )
If a (statistically) significant trend is observed between the two residuals on the residual-residual
plot, it can be concluded that S a (T FM ) is not sufficient with respect to the ground motion
characteristic under consideration. In order to measure the possible trend between the two sets of
residuals, a line is fit to the points on the plot using simple linear regression.. This is based on a
property of linear regression in which the regression prediction passes through the average values
of the two sets of horizontal and vertical components. As the two sets of components constructing
a residual-residual plot are themselves residuals of linear regressions, their corresponding average
values are always equal to zero, and hence the reason why the regression line always passes
through the origin. The slope of the regression line, if significantly different from zero,
indicates and measures the potential trend;. As mentioned in Section 6.8, the test of hypothesis
based on the Student-t Distribution is used to test if the slope is significantly different from
zero. It should be noted that there is a difference between the strength of the trend and the
significance of the trend. While the significance of a trend is measured by the p-value, the
strength of a trend is measured by the slope of the regression prediction line.
Figure 6-2 is an example of a residual-residual plot used to investigate the significance of rupture
distance Rrup as a second variable in the regression (consistent with the original regressions and
plots, however, a log scale is retained. Hence 10 0 corresponds to a log residual of zero). The a
value reported on the graph is the intercept of the fitted line. As it is explained above the a value
is always equal to zero or an infinitesimal number as it is reported on the graph. The b value is the
slope of the fitted line which is very close to zero in this case (Figure 6-2). The dispersion
measure denoted by ln residual|ln residual 2 is the (square) RMS of the regression residuals.
ln residual|ln residual 2 is a measure of the conditional dispersion of ln max given S a (T FM ) and the
second regression variable (here, the rupture distance, ln R rup ). Theoretically speaking, it should
be close to or smaller than the conditional dispersion of ln max given S a (T FM ) only. This means
the value for ln residual|ln residual 2 should be close to or smaller than the value reported for the
original regression in Figure 6-1 ( max |S a = 0.40 ). As will be observed in the following residual-
residual plots, the value for dispersion is very close to 0.4 in the cases where no trend is observed.
This confirms that the second variable is not significant in reducing the dispersion in the
regression or in simple words reducing the fraction of the variability in the structural response
that is not already explained by S a (T FM ) . On the contrary, when a significant trend is observed,
the reported regression value is (significantly) smaller than ln max |S a = 0.40 . The one-sided p-
value reported on the graph is a measure of the significance of the slope. It is called one-sided
because it is a probabilistic measure of how many standard deviations, b , the (absolute value of)
slope b is greater than zero.
1
LA 20
10
a = 6.96e16
Residual of the "original" regression on ln Sa(FMF)
b = 0.2296
ln residual|ln residual2 = 0.3869
= 0.125
b
4 onesided pvalue = 0.03638
41
3 28
17
47 13
24 19 34
12 36
30
0 25 3731 14
10 40 8 45
35
33 22 1143 46
27 1629
42
38 15
39 26
6 19
7 23
5 32 18 20 10
21
44
2
1
10
0 1
10 10
Residual of the resgression of distance ln Rrup on ln Sa
Figure 6-2 - Residual-residual plot for rupture distance, as a second independent variable for
predicting structural response for LA20 structure, combined record selection
The value for b is also reported on the graph. If the standard deviation b is very large relative
to the slope itself, the slope may not be statistically different from zero. The p-value is a
probabilistic way of quantifying the above statement. A p-value equal or close to zero indicates
that the slope is significantly different from zero. Looking at Figure 6-2, it can be observed that
there is a weak negative correlation between the two residuals. The p-value reported on the plot is
small (%3) but not infinitesimal, which confirms a the existence of a weak trend. However, The
dispersion value ln residual|ln residual 2 is equal to 0.39, which is very close to ln max |S a = 0.4 . The
negligible reduction in the dispersion indicates that S a (T FM ) is effectively sufficient with
respect to Rrup .
The residual-residual plot in Figure 6-3 investigates the sufficiency of ln S a (T FM ) with respect to
magnitude, M . As it is explained above, the residuals of the regression of ln max and of M on

ln S a (T FM ) are plotted against each other. The slope of the fitted line or the b value (b=-0.10) is
slightly less than zero indicating a potential weak dependence. But the p-value is not negligible,
which means that the reported slope is not significantly different from zero. Also the fact that the
reported dispersion, ln residual |residual 2 0.40 is very close to ln max |S a = 0.40 confirms that
magnitude is not significant as a potential second regression variable. As a result, S a is also
sufficient with respect to ground motion magnitude.
1
LA 20
10
Residual of the "original" regression on ln Sa (FMF) 4

41
3 28
17
1347
34 24
12 36 19 30
0 37 14 31 25
10 40 8 45 35
22 33
4216 43
11
29 46
27 1538
26 639 1 9
32 7 23
5 1020 18
21
44
2
a = 7.53e16
b = 0.1052
ln residual|residual2 = 0.3984
= 0.1405
b
1
onesided pvalue = 0.2289
10
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Residual of the regression of magnitude on ln S (FMF)
a
Figure 6-3 - Residual-residual plot for magnitude as a second independent variable for predicting
the structural response for LA20 structure, combined record selection
Figure 6-4 investigates the sufficiency of S a (T FM ) with respect to the normalized residual
(epsilon) of the attenuation law, denoted by . The values are calculated by finding the
attenuation law prediction for the natural logarithm (ln) of spectral acceleration values at T=4
seconds (Equation 6-1), which is close to the FMF for the structure with T1=3.96 seconds, and
subtracting the attenuation predictions from the corresponding ln(spectral acceleration) value for
each ground motion record. The attenuation epsilon accounts for any other ground motion
characteristic that is not explained by M and Rrup . There is no significant trend observed in the
residual-residual plot of Figure 6-3. As with the previous plots, the p-value is not significant and
the sigma of the residual-residual regression demonstrates virtually no reduction relative to the
sigma of the original regression. Therefore, it can be concluded that S a (T FM ) is sufficient with
respect to the attenuation epsilon.
1
LA 20
10
Residual of the "original" regression on ln S (FMF)

a
4
41
3 28
17
13 47
34 24
12 19 36 30
0
14 37 31 25
10 40 45
822 35
46 43
15 27 33 38
294216 11 9
1 26 6 39
23 7
5 10 18 3220
21
44
2
a = 1.15e15
b = 0.09411
= 0.1058
b
1
10
1.5 1 0.5 0 0.5 1 1.5
Residual of the regression of epsilon on ln Sa
Figure 6-4 - Residual-residual plot for the epsilon of the attenuation law, , as a second
independent variable for predicting the structural response for LA 20 structure, combined
record selection.
6.9.2 System with short period: Generic SDOF structure
The example case of a structure with long fundamental period was discussed in the previous
section. It was observed that the FMF spectral acceleration is sufficient with respect to each of the
ground motion characteristics, magnitude M , rupture distance Rrup , and the attenuation epsilon,
It is also of interest to investigate the sufficiency of S a (T FM ) for a system with short period. This
is because the non-linear displacement response of a high-frequency SDOF system is expected to
show dependence on frequencies lower than the FMF frequency, which can be due to the period-
elongation phenomenon (see Medina, 2002). The dependence of the response on periods longer
than that of FMFs, may demonstrate itself in terms of the dependence on spectral shape factor.
This dependence may be more pronounced in the high-frequency period range because of the
steep slope of the acceleration spectrum in this region.
As the second example case, a short-period (T=0.1) SDOF system is modeled having bilinear
hysteretic behavior with a strain hardening ratio equal to = 0.03 and a yield strength equal to
M P = 5000 k in . By definition, the displacement response of such an SDOF system is
proportional to the spectral acceleration in the linear-elastic range, meaning that the maximum
displacement response shows zero record-to-record variability for a given S a (T FM ) . However,
the displacement response in the non-linear range is not strictly proportional to the spectral
acceleration and demonstrates record-to-record variability for a given spectral acceleration level.
Hence, it is important to adjust the yield strength of this SDOF system in order to ensure that the
dynamic displacement response for this study is in the non-linear range. Studying possible
dependencies of response on ground motion characteristics decides whether they can (further)
explain the record-record variability in the data due to nonlinear behavior.
Figure 6-5 illustrates the static pushover curve for the bi-linear SDOF model structure. It can be
observed that the onset of inelastic behavior takes place at a drift angle equal to 0.0005. This is
going to be helpful in evaluating the ductility range for the dynamic non-linear response of this
SDOF system later in this section.
Static Pushover Analysis SDOF system with T=0.1sec and =0.03

0.35
0.3
0.25
Shear Force/ Mass [g]
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Drift Angle 3
x 10
Figure 6-5 - The static pushover curve for a bi-linear SDOF system with T=0.1sec, = 0.05 ,
(strain hardening ratio) = 0.03 , and M P = 5000 k in
In Figure 6-6, the maximum dynamic displacement response, max , of the SDOF model structure
with period, T=0.1, and damping ratio, = 0.05 , to the combined record selection, is plotted
versus S a (T FM ) . The straight line in the figure illustrates the result of the linear regression of
ln max on ln S a (T FM ) in the logarithmic scale, also referred to as the original regression. The
RMS of regression residuals or the sigma of the original regression, ln max |S a , is reported as 0.83
on the figure. Compared to the sigma of the original regression for the LA 20, the sigma of the
original regression for the SDOF system is about two times larger. It should also be noted that the
linear regression slope parameter (the b-value reported on the figure) is equal to 1.6, which is
much larger than the slope parameter reported for the LA20 (b=0.65). This b value implies the
softening that is anticipated to cause this system to be sensitive to ground motion frequency
content at frequencies less than its natural frequency. The maximum displacement response
values in the figure range between 0.0001 and 0.01, which corresponds to a ductility range of 1 to
20. The large ductility ratios indicate that the displacement response is in the non-linear range.
The large record-to-record variability in the response, indicated by the relatively large sigma
value, is also typical of an SDOF system with a short elastic period in the non-linear range
(Medina, 2002, Miranda, 2001, and Shome 1999).
SDOF, =3%, Mp=5000 kin

1
10
47 Records with 5.0 < M <7.5 15< R <120
47
0 41
10 13
27
28
23 46 43 17
19
36 302442
31 37 34
14 16 22
10 9 18 25 2935
38 26
3 45
5 3932 6 20
12
8 4
112 1
40 7
1 21 33
10 44
15
1
2
10
a
a = 0.00561
b = 1.571
= 0.8313
ln( )|S
max a
3
10
4 3 2
10 10 10
Maximum Story Drift Angle, max
Figure 6-6 - The dynamic displacement response of the bi-linear SDOF model structure with
T=0.1, damping = 5% , and Mp = 5000 k in to the combined record selection
The sufficiency of S a (T FM ) with respect to the magnitude in investigated in Figure 6-7. Figure
6-7 is a residual-residual plot for magnitude as a second regression variable. A slight negative
trend between the response residuals and the residuals of the regression of magnitude on S a is
observed. However, the p-value is too large for the trend to be significant.
1
SDOF, Mp=5000 kin
10
34

45
22 4
8
a
12 37
17 31 7
3 42 29
24
35 1
13 3025
0
10 18
6
2819
20
40 9 15
27 26 36
38
43 32
39 41
33
44 46
16 47
2 21
11 10
5 14 23
a = 1.68e15
b = 0.1359
= 0.2584
b
1
10
1.5 1 0.5 0 0.5 1 1.5
Residual of the regression of magnitude on ln Sa(FMF)
Figure 6-7 - Residual-residual plot for magnitude as a second independent variable for predicting
the response of an SDOF system, combined record selection
Also, the sigma of the residual-residual regression (reported on the figure) is equal to
ln residual |residual 2 0.83 , which is equal to the sigma of the original regression ( ln max |S a = 0.83 ).
This implies that the addition of magnitude as a second regression variable does not decrease the
dispersion of the response with respect to the regression prediction. Thus, it can be concluded that
S a (T FM ) is sufficient with respect to magnitude even for this very short-period system of large
ductilities.
Figure 6-8 is the residual-residual plot for the rupture distance, R rup , as a second regression
variable. It can be observed that the two ground motions with large rupture distances are mainly
responsible for the negative trend (judged by the seemingly large absolute value of the slope).
However, the large variability of response residuals around the regression line causes the p-value
to be very large. Hence, the observed slope is not in fact significantly different from zero. Also,
the rupture distance, R rup , does not reduce the dispersion of response with respect to regression
prediction ( ln residual|ln residual 2 = 0.82 ), as a second regression variable. Hence, S a (T FM ) is also
sufficient with respect to the rupture distance, R rup .

Figure 6-9 is the residual-residual plot for the normalized residual (epsilon) of the attenuation
law, , as a second regression variable. Judging from the p-value for the slope, it can be observed
that there is no significant trend between the two sets of residuals.
SDOF, M =5000 kin
p
1
10
34

45
22 4
8
12 37 a = 1.84e15
b = 0.3486
ln residual|ln residual2 = 0.8156
17 31 7
3 24 42 29
35 b = 0.2611
1
25
30 13 onesided pvalue = 0.09428
0
10 18
6 28 20
40 9 2719
26 36 15
38
32
39 41 43
33
44 46
47 16
21 2 10
11
5 14 23
1
10
0 1
10 10
Residual of the resgression of distance ln R on ln S
rup a
Figure 6-8 - Residual-residual plot for distance as a second independent variable for predicting
the maximum inter-story drift angle, SDOF system, combined selection.
SDOF, Mp=5000 kin

1
10
a = 1.85e15
b = 0.1531 34
45
= 0.3404 4 22
b
onesided pvalue = 0.3275 8
a
3712
31 177
3
29
35 42 24
25 1
13 30
10
0 18
20 40 28 19 6
15 27 36 9 26
3832
43 41 39
33
46 44
16 47
210 21
11
23 14 5
1
10
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Residual of the regression of "epsilon" of the attenuation on ln Sa
Figure 6-9 - Residual-residual plot for the normalized residual of the attenuation law, , as a
second independent variable for predicting the structural response for SDOF system,
combined record selection.
This observation is confirmed by the reported sigma of the residual-residual regression
( ln residual |ln residual 2 0.83 ), which is equal to the sigma of the original regression (reported on
Figure 6-6). Hence, S a is also sufficient with respect to the normalized residual (epsilon) of the
attenuation relation, .
6.10 Studying the dependence of the displacement-based response given S a (TFM )

on the spectral shape factor at period T, R(T , TFM )
In the previous section, the sufficiency of S a (T FM ) with respect to ground motion characteristics,
M , Rrup , and, , was established for two example cases with short and large FMFs. As was
discussed, this also implies that none of the ground motion characteristics studied above, as a
second regression variable, is going to (further) reduce the variability of the response around
linear regression prediction.
In this section, the dependence of the displacement-based response, once it is conditioned on the
FMF spectral acceleration, on the shape of the elastic acceleration spectrum is discussed. The
spectral shape factor at period T, R(T , T FM ) is the variable that reflects the effect of the spectral
shape. Recall that if this dependence is confirmed, and if the shape factor is also found to depend
on magnitude, it would be expected that S a (T FM ) would not be sufficient with respect to
magnitude, contrary to what was found in the previous section. Further, it can be argued that
establishing the dependence of the response given S a (T FM ) on the spectral shape factor is
equivalent to stating that the pair of S a (T FM ) and R(T , T FMF ) as a vector-valued intensity
measure is more efficient than S a (T FM ) only. In other words, the vector IM, { S a (T FM ) ,
R (T , T FM ) }, has S a (T FM ) representing spectral amplitude at FMF as its first element, and
R (T , T FM ) representing spectral shape at period T as the second element.
The procedure followed in this section for establishing the dependence of the displacement-based
response (given S a (T FM ) ), on R(T , T FM ) is similar to the one followed in order to investigate the
sufficiency of S a (T FM ) in the previous sections. In the same manner, the spectral shape factor
R (T , T FM ) is regarded as a second regression variable, and the residual-residual plots are used to
study the effect of shape factor in reducing the dispersion. It is important to note that this section
only presents a brief discussion of the role of spectral shape factor in increasing the efficiency in
response predictions, and does not delve into the complexities and limitations of implementing a
vector-based intensity measure. In summary, the main argument in this section is to relate the
shape-dependence of the response to possible dependence on magnitude and hence explain why
magnitude is generally believed to affect the response more than the other ground motion
characteristics.
Similar to the previous section, the discussion here is divided into two parts, with each part
focusing on one of the two example cases presented before.
6.10.1 System with large first-mode period: LA 20 story Results
The effect of spectral shape factor as a second regression variable is studied here for the LA 20
model structure. In the following, the residual-residual plot for spectral shape factor at its second
mode frequency (SMF) with T2=1.3 is presented and discussed in detail. The discussion is
followed by studying the effect of the spectral shape factor at a period, T, as the second
regression variable, from the point of view of the reduction in the dispersion around the
regression prediction. This leads to finding an optimal period at which the dispersion is minimum,
or in other words the pair { S a (T FM ) , R(T , T FM ) } has maximum efficiency.
Since LA 20 is not a first-mode dominated structure, the spectral shape factor R(TSF , T FM ) is a
potentially important variable to consider as the second regression variable. The residual-residual
plot for R(TSF , T FM ) as a second regression variable is plotted in Figure 6-10.
Unlike the residual-residual plots in section 6.9, the residuals in Figure 6-10 demonstrate a clear
and positive trend. The p-value is zero, which confirms that the slope of the line is significantly
different from zero. Also the reported sigma of the residual-residual regression, 0.25, is
significantly smaller than the sigma of the original regression, 0.40. This confirms the speculation
that spectral shapes or spectral ordinates at higher frequencies might be effective in predicting the
response for this tall building. Also, the significant reduction in the sigma of the regression
indicates that the vector { S a (T FM ) , R(TSM , T FM ) } is going to be (relatively) more efficient in
predicting the dynamic displacement response compared to the scalar S a (T FM ) (also see,
Vamvatsikos, 2002, and Shome 1999).
It is desirable to find the optimal period for the spectral shape factor as the second regression
variable. The plot in Figure 6-11, illustrates the values for the sigma (dispersion measure) in the
residual-residual regressions performed for a range of periods. For simplicity, the dispersions are
normalized with respect to the sigma of the regression on S a only. It can be observed that
choosing the spectral shape for T=1.1 seconds (close to the second mode period) leads to more
than 40% reduction in the dispersion compared to regression on S a only. The periods
corresponding to the first and second modes are marked on the figure. Also, the point marked by
PGA represents the T=0 data point.
1 LA 20
10
a = 1.12e15
b = 0.4937
ln(residual)|ln(residual2) = 0.2543
= 0.06039
b
onesided pvalue = 9.567e11
4
41
3
1728
13 47
34
36 19 24
30 1214
0 3725 31
10 4540
43 3546 8
15 33 38 16 229 2911 42 27
39 1 6 26
23
10 20 32187 5
21
44
2
1
10
1 0 1
10 10 10
Residual of the regression of ln R(T =1.3, T =3.9) on ln Sa(FMF)
SMF FMF
Figure 6-10 - residual-residual plot for spectral shape factor at SMF, as a second independent
variable for predicting the structural response for LA 20 structure, combined record selection.
Generally-speaking, it can be observed that the (relative) dispersions decrease for the spectral
shapes corresponding to the short-period range. This confirms the a priori speculations about the
dependence of response on the higher mode spectral shapes. It should also be noted that the drift
response of the LA 20 structure is generally smaller than 1%, indicating more-or-less elastic
behavior. Hence, the dependence of the response (given FMF spectral acceleration) on the
spectral shape at short periods seems to be consistent, as the effect of the higher modes on the
response is more pronounced in the linear range of response. Later in this chapter (Section 6.14),
the effect of the higher mode spectral shapes are going to be observed for the case where the
response is in the non-linear range.
LA 20
) 1
FMF
0.95
normalized by the "sigma" of the "original" regression
"sigma" of the residualresidual regression for R(T,T
PGA
0.9 FMF
0.85
0.8
0.75
0.7
0.65
0.6
0.55
SMF
0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
period, T seconds
Figure 6-11 - The reduction in the sigma of the regression when the spectral shape factor at
period T is used as a second independent variable in the regression, LA 20 structure,
combined record selection.
6.10.2 SDOF systems with short period
It was observed in the previous section that for the 20-story MDOF structure, the spectral shape
factor at frequencies higher than the FMF could explain part of the variability observed in the
residuals of the original regression on spectral acceleration. In this section, the dependence of the
non-linear response of the bi-linear SDOF system, visited earlier in this chapter, on the shape of
the spectrum is studied. Although in the case of an SDOF system higher modes do not exist, it is
expected that for the higher ductilies the system behaves as if it has an elongated effective period.
Thus, the spectral shapes for periods longer than T=0.1, may play a significant role in explaining
the variability observed in the response of the SDOF system.
In order to investigate the dependency of response of the SDOF system on the spectral shapes for
periods longer than T=0.1, the residuals of response against the residuals of the spectral shape for
T=0.3 are plotted in Figure 6-12. Judging from both the p-value and the reported sigma value, a
significant positive trend in the plot can be observed. The p-value is infinitesimal, which indicates
that the slope of the regression line is significantly different from zero; and the sigma is reduced
by 35% compared to the sigma of the original regression in Figure 6-6. This confirms the
speculations about the significance of the spectral shapes for periods longer than the first-mode
period in explaining the variability observed in the nonlinear response of the SDOF system with a
very short period.
SDOF, Mp=5000
1
10
a = 1.87e15
b = 1.872
= 0.5482 34
ln(residual)|ln(residual2)
45
= 0.2449 4 22
b
onesided pvalue = 5.657e10 8
37 12
7
17
31
3 24
29
42 35
25130
13
0
10 18
620
4028
19
9
27
36 26 15
32 38
39 43 41
33
44 46
16 47
2 21
10 11
14 23
5
1
10
1 0 1
10 10 10
Residual of the regression of ln R(T=0.3, TFMF=0.1) on ln Sa(FMF)
Figure 6-12 - The residual-residual plot for the shape factor at T=0.3 as a second independent
variable for predicting the structural response for SDOF system with T=0.1,
combined record set
The same steps explained above (for the spectral shape factor at T=0.3), are repeated over a range
periods and the resulting relative dispersions are plotted versus period in Figure 6-13. Similar to
the previous section, the sigma values (the sigma of the residual-residual regression) for each
period are normalized with respect to the sigma of the original regression on spectral acceleration
at T=0.1. It can be observed that the spectral shape factors for periods between 0.5 and 1 second
lead to reductions up to 40% in the dispersion. This is consistent with the earlier observations for
the spectral shape at T=0.3. In summary, the displacement-based response of this high-frequency
bi-linear SDOF system is strongly dependent on spectral shapes at periods longer than that of the
FMF. This observation is consistent with the elongation of the effective period in the high-
ductility range.
SDOF, Mp=5000 kin

1
"sigma" of the residualresidual regression for R(T,TFMF)

normalized by the "sigma" of the "original" regression
0.95
0.9
0.85
0.8
FMF
0.75
T=0.3
0.7
0.65
0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
period, T seconds
Figure 6-13 - The normalized sigma of the residual-residual plot for the spectral shape factor at
period T with respect to the sigma of the original regression on spectral acceleration at T=0.1.
6.11 Discussion: The correlation between spectral-shape factor and magnitude

It was observed in the previous sections that for a given value of S a (T FM ) , the displacement-
based response of the two example structures demonstrate statistically significant dependence on
the spectral shape factor. The dependence on the spectral shape factor may be observed for
structures with a wide range of first-mode periods. For example, a first-mode dominated structure
might undergo considerable damage when subjected to a ground motion excitation. This will
cause the structure to behave as a softer system with a longer effective period; hence, the spectral
shape for a longer period might be effective as a second regression variable. For a tall structure
with long first-mode period, the displacement-based response is most likely to be dictated by
higher modes. In this case the spectral shapes for periods shorter than that of the FMF may be
more effective as second regression variables.
On the other hand, since they are physically limited in the production of longer period waves,
smaller magnitude records are expected to have comparatively stronger short-period (high-
frequency) energy content. Thus, one would expect to observe a negative correlation between the
spectral shape factor for second-mode period and the magnitude. This hypothesis is tested for the
suite of records used in this chapter. The spectral shape factors for the suite of records at a period
equal to T=1.3 seconds (corresponding to second mode frequency for the 20-story building),
normalized to the FMF spectral acceleration at a period to T= 3.9 seconds are plotted versus
magnitude in Figure 6-14. The anticipated inverse trend can be observed from the data points and
also the negative slope of the line fitted to the data. The slope of the fitted line indicates roughly a
factor of 50% reduction in this particular shape-factor for a magnitude increase of one unit. The
reported p-value for the slope of the line in Figure 6-14 indicates that there is an effectively
significant correlation between the spectral shape and magnitude.
2
Combined selection
10
=3.9)
1
10
FM
=1.3, T
SM
shape factor(T
0
10
a = 4.64
b = 0.483
lnR|M = 0.6767
= 0.4869
M
= 0.2027
b
1
10
5 5.5 6 6.5 7 7.5
magnitude
Figure 6-14 - Combined selection, shape factor for T=1.3 sec and T=3.9 sec plotted against
moment magnitude
Alternatively, one can use the attenuation relation predictions for the spectral shape versus
magnitude in order to test the hypothetical correlation between the spectral shape factor and
magnitude. The attenuation relation prediction for the log of the spectral acceleration at period T
can be calculated from Equation 6-1. The mean (of the log) spectral shape factor, R (T1 , T2 ) , can
be predicted by subtracting the mean (of the log) attenuation predictions for spectral accelerations
at periods T1 and T2 respectively (Equation 6-2a):
E[ln R (T1 , T2 )] = E[(ln S a (T1 ) ln S a (T2 ))] = f ( M , R rup ) (6 - 2a)
2 ln R (T1 ,T2 ) = 2 ln S a (T1 ) + 2 ln S a (T2 ) 2 ln S a (T1 ),ln S a (T2 ) ln S a (T1 ) ln S a (T2 ) (6 - 2b)
Also the squared standard deviation (variance) for the shape factor can be calculated by
combining the attenuation predictions for the standard deviation of the natural logarithm of
spectral acceleration at periods T1 and T2 (Equation 6-2b). It should be noted that expression for
standard deviation in Equation (6-2b) includes the correlation coefficient between the (natural log
of) the spectral acceleration values at periods T1 and T2 . Figure 6-15 illustrates the results of
attenuation predictions for the spectral shape factors at T1 = 1 and T2 = 4 (approximately the
second-mode and the first-mode periods for the 20-story building).
2
Attenuation predictions, T =1, T =4
10 1 2
1
shape factor(T =1, T =4)
10
2
1
0
10
Abrahamson and Silva (1997) Attenuation Law

16th, median, and 84th percentile predictions
Deep soil, horizontal, strikeslip faulting, Rrup=25km
1
=0.60 (correlation coefficient between the spectral accelerations at T=1 and T=4)
10
5 5.5 6 6.5 7 7.5 8
magnitude
Figure 6-15 - Attenuation prediction for shape factor at T=1 sec ( TSM ) and T=4 sec ( T FM )
plotted against magnitude (thick lines). The thin is the linear regression prediction based on
the combined ground motion record set.
The spectral shape factors are predicted using Abrahamson and Silva (1997) attenuation relations
for the horizontal component of the ground motion. It has been assumed that the style of
faulting is strike-slip and the sites soil type is deep soil. In order to calculate the standard
deviation of the predicted shape factor from the Equation 6-2b, the correlation coefficient
between the spectral acceleration values at the two periods under consideration should be
estimated. For the two periods, 1 and 4, a 60% correlation between the (log of the)
corresponding spectral acceleration values is assumed (based on the observed correlation for
the ground motions used in this chapter). Figure 6-15 illustrates the 16th, median and 84th
percentiles of the predicted shape factors against the shape factors calculated from the suite of
records.
It can be observed that most of the spectral shape factors for the suite of records (the combined
selection) fall between the 16% and 50% attenuation law predictions. The slope of the predicted
median of the attenuation shape factor is roughly the same as the slope of the line fitted to the
data points. This also confirms a weak correlation between the spectral shape factor and ground
motion magnitude for T1 = 1 and T2 = 4 . It is also interesting to study the dependence of the
spectral shape factor at very short periods on ground motion magnitude. The conclusion may be
useful in understanding the dependence of the response of the high-frequency SDOF system on
magnitude.
1
Combined selection
10
shape factor(T=0.3, T =0.1)
FM
0
10
a = 0.444
b = 0.04494
lnR|M = 0.3979
= 0.4869
M
= 0.1192
b
1
10
5 5.5 6 6.5 7 7.5
magnitude
Figure 6-16 - Combined selection, shape factor for T=0.3 sec and T=0.1 sec plotted against
moment magnitude
Therefore, the spectral shape factors at periods T1 = 0.1 and T2 = 0.3 are plotted in Figure 6-16
for the suite of records used in this chapter. It can be observed that there is no particular trend in
the data; this is also confirmed by the close-to-zero slope of the line fitted to the data and the
large p-value for the slope. Nonetheless, the attenuation relation prediction for the shape factor at
T1 = 0.1 and T2 = 0.3 in Figure 6-17 indicates a slightly positive trend with regard to magnitude.
Nevertheless, since the two periods, T1 and T2 are both in the high-frequency range, the ratio
between their corresponding spectral acceleration values may not be strongly dependent on
magnitude.
Attenuation predictions for T=0.3, T =0.1

1
10
FM
=0.1)
FM
shape factor(T=0.3, T
0
10
Abrahamson and Silva (1997) Attenuation Law

16th, median, and 84th percentile predictions
Deep soil, horizontal, strikeslip faulting, Rrup=25km
=0.80 (correlation coefficient between the spectral accelerations at T=0.3 and T=0.1)
1
10
5 5.5 6 6.5 7 7.5
magnitude
Figure 6-17 - Combined selection, shape factor for TFM=0.1 sec and T=0.3 sec plotted against
magnitude
6.12 Discussion: Does magnitude matter?
The ground motion magnitude is probably the most likely ground motion characteristic variable
to affect the structural response. This section is going to discuss the effect of magnitude on the
structural response from the point of view of its relation to the shape of the acceleration spectrum.
It was established in Section 6.10 that, given the FMF spectral acceleration, the structural
response is positively correlated to spectral shape factor at a period different than FMF. For the
case of the 20-story structure, the response is most strongly correlated to spectral shape factors at
periods shorter than FMF and close to the second-mode frequency (SMF). For the high-frequency
SDOF system, the response is most strongly correlated to shape factors at periods longer than
FMF. Consistent with the above observations, the relation between shape factor and magnitude
was investigated for the spectral shape factors at T=1.3 (SMF period for 20-story structure) and
T=0.3 (3 times the FMF period for the SDOF system). For the 20 story building, the spectral
shape factor at SMF period was negatively correlated to magnitude with a relatively large
dispersion around the line fitted to the data. However, for the high-frequency SDOF system no
significant correlation was observed between the shape factor at T=1.3.
The dependence of response on the shape factor and also the negative correlation between the
shape factor and magnitude in the case of a long-period structure like the 20-story structure may
suggest that the response is in turn (negatively) related to magnitude. However, this prediction
seems to be contradicted by the results drawn from the residual-residual plot for magnitude in
Figure 6-3. These results indicate that given the FMF spectral acceleration, the displacement-
based response for the 20-story structure is not significantly dependent on magnitude. A closer
look at Figure 6-3 reveals that given the FMF spectral acceleration the response is indeed
negatively correlated with magnitude, but the slope of the residual-residual regression line is very
small and its corresponding p-value is quite high. Nevertheless, this may be explained by the
(relatively large) dispersion around the line fitted to the shape factor-magnitude data points in
Figure 6-14 and the (non-infinitesimal) p-value for the slope of this line. In summary it can be
argued that, although the response may suggest some minor dependence on magnitude for the 20-
story structure, the large variability in the data-points causes the observed dependence not to be
(statistically) significant. Further, regression analyses that show dependence of response (given
FMF spectral acceleration) on shape factor and also shape factor on magnitude, do not necessarily
imply that there must be a positive (linear) correlation between response (given FMF spectral
acceleration) and magnitude. Hence, it can be concluded that despite the reasonable expectations
that (given FMF spectral acceleration) magnitude might provide some additional predictive
power, there is no evidence to this effect in the data for either the very long or very short period
systems studied here. This conclusion is consistent with the conclusions others have reached for
moderate-period, first-mode dominated structures (e.g., Bazzurro 1994, Shome 1999, Luco,
2002).
6.13 Observations for the SDOF system with a different yield strength
In the previous sections, the sufficiency of S a (T FM ) with respect to ground motion characteristics
such as magnitude, distance and epsilon of the attenuation law, was established for a short-
period SDOF system. It was also observed that given the FMF spectral acceleration, the
displacement-based response of such SDOF system demonstrates a strong dependence on spectral
shape factors at longer periods (due to the elongation of effective period for high-ductility range).
It is important to investigate whether these observations will still hold for a different degree of
non-linearity in the system. Thus the displacement-based response of the bi-linear SDOF system
is studied for different ductility levels by changing the yield strength of the system. In Section
6.9.2, the dynamic displacement response of an SDOF system with a yield strength equal to
M P = 5000 k in was studied. According to the range of ductility observed in the response
(Figures 6-5 and 6-6), the system is already in a highly non-linear range. In this section, the yield
strength is scaled up to M P = 15000 and M P = 10000 in order to achieve lower ductility
levels in the displacement response. In Figure 6-18, the response of the SDOF system with
M P = 15000 (three times the original strength) to the combined set of records is plotted. It is
apparent that response is in a lower ductility region since the yield drift angle shown in the static
pushover curve for M P = 5000 k in in Figure 6-5 is also going to be scaled up; and also the
displacement-based response in Figure 6-18 is generally smaller than those of Figure 6-6. In
contrast to the similar plot in Figure 6-6 corresponding to the yield strength of M P = 5000 k in ,
an more-or-less linear portion can be observed in the displacement-based response in Figure 6-18
SDOF combined selection, Mp=15000 kin

47 Records with 5.0 < M <7.5 15< R <120
47
41
0
10
27 13
28
2343 46
19 17
36
4237
31
3024 34
14 22
16
353
929
45
18
10
26
25
38
20
632
539
12
284
111
740
1
21
33
10 44
1
a = 0.00119
b = 1.307
ln( )|S = 0.3564
a
max a
15
2
10
4 3 2
10 10 10

max
Figure 6-18 - The SDOF system with M P = 15000 subjected to the combined record selection
The same steps described in Sections 6.9.2 and 6.10.2 are followed here in order to investigate the
sufficiency of S a (T FM ) with respect to ground motion characteristics and also to investigate the
dependence of response on spectral shape factors at periods larger than T=0.1, when the yield
strength is scaled up by factor of 2 and 3. The results are tabulated for brevity. Table 6-2 lists the
Regression on Sa a b ln max |S a
Mp=5000 k-in 0.0056 1.571 0.8313

Mp=10000 k-in 0.0023 1.562 0.6091
Mp=15000 k-in 0.0012 1.307 0.3564
parameters of the regression of response on the spectral acceleration. It can be observed that the
dispersion increases with decreasing yield strength. This is expected as the dispersion increases
for displacement response in higher ductilities. Also for higher yield strength levels a larger
number of response data points are still behaving elastically, which for an SDOF systems implies
zero dispersion locally. The same trend is observed for the b value (the slope parameter in the
log-log scale) that is increasing with decreasing strength. This is also consistent, as a larger b
value is associated with softening behavior in a non-linear system.
Table 6-2 - The results of the regression of response on S a for different yield strengths
Tables 6-3 and 6-4 outline the parameters reflecting the importance of ground motion
characteristics, M , R rup , and , and spectral shape factors for T=0.3 and T=0.5, as a second
regression variable in addition to S a (T FM ) (for two different yield strength values). The same as
Section 6.9.2, the sufficiency of S a (T FM ) with respect to the three ground motion characteristics,
M , R rup , and , is established for higher yield strength values (10000 k-in and 15000 k-in).
This is observed through significant p-values as well as negligible reduction in dispersion for both
strength levels. This confirms the sufficiency of S a (T FM ) for different strength levels and hence
different ductility levels. The dependence of the displacement-based response on spectral shapes
at longer periods can also be observed from Tables 6-3 and 6-4 for both strength levels. Although
such dependence is more pronounced (i.e., smaller p-values, more reduction in the dispersion, and
steeper slopes) for the case with smaller strength ( M P = 10000 ). This can be explained by the
fact that for the case with smaller strength, the response lies in a larger ductility range, which
implies more drastic elongation of the effective period.
It should also be noted that for the largest strength level, when the ductitilies are the mildest, there
is little benefit to the shape factor. This could be expected since in the linear range there is no
benefit in adding the spectral information at other periods, as the spectral acceleration is the
perfect predictor.
residual | residual 2
M=10000 k-in b ln max |ln S a p-value

M -0.013 1 0.472
Rrup -0.03 1 0.484
0.013 1 0.455
R(T=0.3, TFMF=0.1) 1.065 0.81 0
R(T=0.5, TFMF=0.1) 0.693 0.78 0
Table 6-3 - The importance of ground motion characteristics, M , R rup , and , and two different
spectral shape factors as the second regression variable, SDOF system M P = 10000
(moderate ductilities)
M=15000 k-in b ln max |ln S a p-value

M 0.057 0.997 0.3
Rrup -0.09 0.993 0.48
0.019 0.999 0.39
R(T=0.3,TFMF=0.1) 0.343 0.947 0.013
R(T=0.5,TFMF=0.1) 0.267 0.91 0.002
Table 6-4 - The importance of ground motion characteristics, M , R rup , and , and two different
spectral shape factors as the second regression variable, SDOF system M P = 15000
(low ductilities)
6.14 Observations for LA 20 story building when the ground motion records are
scaled
It has been established in Section 6.9.1 that for a long-period 20-story structure, S a (T FM ) is
sufficient with respect to the ground motion characteristics, and that given S a (T FM ) the
displacement-based response is dependent on spectral shape factors at periods shorter than FMF
period. In this section, the validity of such conclusions are investigated for the displacement-base
response of the 20-story structure in a different ductility range.
In order to study the dynamic response of the 20-story building at a different ductility level, the
ground motion records have been scaled by a factor of 3 so that the structure is pushed further
into the non-linear regime.
In Figure 6-19, the displacement-based response of the LA 20 structure to the scaled ground
motion records is plotted versus FMF spectral acceleration. It can be observed that the drift angles
are larger than those in Figure 6-1 in Section 6.9.1 for the un-scaled records. The median drift is
equal to 0.017 whereas in Figure 6-1 the median drift is equal to 0.0051. Also in ductility terms,
the response to the scaled records is on average in a ductility range of 2 compared to the response
in the un-scaled case where the response is effectively in the elastic range. The ductilities are
calculated based on a yield drift angle of 1% obtained from static pushover results for the LA 20
building. (courtesy of Vamvatsikos, 2002).
L20 combined selection, =0.05, Scaled by 3.0

47 Records with 5.0 < M <7.5 15< R <120
0
10
42
35
36
46 34 29
22 33
45
37
27 1
264715
18
1 23 39 6
31
10 32
716 19
10 43 25
38 8 13
9 40 30
24
12 17 4
20 41
28
21 5
3.0*S (T =3.9)
11 14
44
1
2
a
10 2 3
3
10
a = 0.0764
b = 0.6871
ln(
= 0.4333
)|S
max a
4
10
3 2 1
10 10 10
max
Figure 6-19 - Maximum inter-story drift angles for LA20 structure, combined selection scaled by
a factor of 3
Table 6-5 lists the regression parameters for LA 20 building when subjected to the scaled ground
motion records. It can observed that the intercept is larger compared to the one for the un-scaled
response in Section 6.9.1 (Figure 6-1). The dispersion and the b value (log-log slope parameter)
are also slightly larger in this case.
Regression on Sa a b ln max |S a
scaled by 3 0.0764 0.6871 0.433

not scaled 0.0507 0.6553 0.401
Table 6-5 The results of the regression of response on S a (T FM ) for the case where the ground
motion accelerations are scaled by a factor 3, LA 20 story building
Table 6-6 outlines the parameters reflecting the significance of M , R rup , and , and second-
mode shape factor as second regression variables in addition to the FMF spectral acceleration.
This case of sufficiency with respect to magnitude is not as clear-cut as for the un-scaled records,
however, Figure 6-20 illustrates the residual-residual plot for magnitude as a second variable.
Visually, there is a mild trend but a single (magnitude 5) data point (number 3) is largely
responsible for the appearance and for the marginally low p-value (0.068). (If this point is
removed the p-value increases to 0.211). It can be observed that magnitude provides only a 2%
reduction in standard deviation (0.42 versus 0.43). Finally, even if estimate of b is accepted as
statistically significant (i.e., different from zero), its value (0.22) implies that the strength of
dependence of response on magnitude is not great in practical terms. The value implies that
change in magnitude of one-half magnitude unit implies only a 10% change in predicted
response. Taken together with the clear lack of significance at lower ductility levels (when one
would a priori expect a stronger dependence on magnitude), the conclusion that magnitude plays
a minor, if any, role cannot be avoided.
scale factor= 3 b ln max |ln S a p-value

M -0.2237 0.976 0.068
Rrup -0.299 0.948 0.455
-0.15 0.98 0.0866
R(TSMF=1.3,TFMF=3.9) 0.3886 0.826 0
Table 6-6 - The importance of ground motion characteristics and spectral shape factor as the
second regression variable, LA 20 building S .F . = 3.0
It can also be observed from Table 6-6 that given the FMF spectral acceleration, the
displacement-based response is positively correlated with the SMF spectral shape (at T=1.3). This
is due to the zero p-value and also due to significant reduction in dispersion (17%). However, the
reduction in dispersion is smaller (by about half). As compared to the un-scaled case (when it was
more than 35% in that case).
L20, combined selection scaled by a factor of 3.0

Residual of original regression on Sa (T = 3.9, =0.05)
29
3
34
4
28 41
45 17 33 13
24
12 31 19 30
14
8 6
0 40
10 1511
36 25
1
35
38 47
22 16 39 43 9
26
37 27 7
5 32 20 46
10 23
42 21
18
244
a = 4.66e16
b = 0.2237
residual|residual2 = 0.4229
= 0.1475
b
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Residual of regression of magnitude on ln Sa
Figure 6-20- The residual-residual plot for magnitude as a second regression variable, the
combined record set scaled by a factor of 3.
This may be explained by the stronger higher mode dependence for the response in the low
ductility range. Figure 6-21 illustrates (square) RMS of the residual-residual regressions for the
spectral shape factors at different periods normalized by the (square) RMS of the original
regression on FMF spectral acceleration for the 20-story structure subjected to the combined set
of records scaled by a factor of 3. It can be observed the maximum reductions in (square) RMS
occur at period shorter or around the SMF period.
L20 Relative (square) RMS of the reisdualresidual regression for
spectral shape factor
regression with respect to the original regression

Relative (square) RMS of the residualresidual
0.95
0.9
0.85
0.8
0.75
0.7
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
period, T seconds
Figure 6-21 - Comparing the (square) RMS of the residual-residual regression for the shape factor
at different periods with the (square) RMS of the original regression on FMF spectral
acceleration, LA 20 structure, the combined record set scaled by a factor of 3.
6.15 Sufficiency and ground motion record selection
It was demonstrated in the previous sections how to establish the sufficiency of S a (T FM ) with
respect to ground motion characteristics, magnitude M, (source-to-site) distance Rrup , and the
epsilon, . This section is going to discuss how establishing the sufficiency criterion is going to
affect ground motion record selection.
To address this problem, it can be recalled first that the sample average and the sample variance
of the relevant displacement-based response can be calculated as:
NT
1
max = E [ max ] =
NT
i
(i )
max (6 - 3)
s 2 max = E [ max 2 ] - ( max ) 2 (6 - 4)

where E[.] denotes the expected value (mean) estimated by the sample average. The
above equations can also be applied to the natural logarithm of the relevant displacement-
based response for a given S a (TFM ) or other intensity measure equals level, x:
NT
1
ln max |S a (T ) ( x) E [ln max | S a (T )] =
NT ln
i
(i )
max (6 - 5)
2 max 2 ln max E [ln max 2 | S a (T )] - ( E [ln max | S a (T )]) 2 (6 - 6)
Where the (log of) median has been approximated by the sample average of the logarithm of the
response. However, for a lognormal variable the mean of the logarithm is equal to the logarithm
of the median. Also, the squared standard deviation of the logarithm of response has been
approximated by the sample variance of logarithm of the response. The next step is to expand the
estimated median and the standard deviation (of the logarithm) can be expanded with respect to
ground motion characteristics, M, R rup , and :
N M N R rup N
ln max| S
a (T )
( x) = E [ln
i =1
max ( x, m i , r j , k ) | S a , M , R rup , ] . p M , Rrup , |S a (T ) ( mi , r j , k | x )
j =1 k =1
(6 7)
where (.) denotes the estimated median. For a selection of records consisting of N T ground
motion records the above expression can be expanded by dividing the chosen records into
N M N R rup N bins, in which each bin is represented by values, mi , r j , k .
average(ln max ( x, m i , ri , i )) is the estimated conditional mean (i.e., sample average) for the
natural log of response in each bin, represented by values, mi , r j , k , for a given spectral
acceleration value, x and p M , Rrup , |S a (T ) (mi , r j , k | x) is the ratio of the number of ground motion
the bin to the total number of ground motions, N T , i.e., the fraction of records in bin (i,j,k).
In a similar manner, the conditional variance of the natural logarithm of response for a given
spectral acceleration , x, can be expanded with respect to ground motion characteristics, M, R rup ,
and :
ln2 max |S a (T ) ( x) =
N M N Rrup N
E[(ln
i =1 j =1 k =1
max ( x, mi , r j , k ))
2
| S a (T ), M , Rrup , ] . p M , Rrup , |S a (T ) (mi , r j , k | x) (ln max |S a (T ) ( x)) 2
(6 8)
Here, E [(ln max ( x, mi , r j , k )) 2 | S a (T ), M , R rup , ] denotes the expected value estimated by the
conditional sample average of the squares of the natural logarithm (i.e., the estimated second
moment) of structural response to ground motions in each bin represented by values mi , r j , k ,
(for a given spectral acceleration value, x). In order to maintain the generality of the above
equations, the spectral acceleration is presented by spectral acceleration, S a (T ) , evaluated at the
period, T. Therefore, S a (T FM ) is a special case of S a (T ) evaluated at the period corresponding to
FMF.
As it was stated in the beginning of this chapter, an intensity measure (e.g., S a (T FM ) ) is

sufficient if it renders the displacement-based response conditionally independent of the ground
motion characteristics, for a given intensity level (e.g., S a ( FM ) = x ). The statistical equivalent to
this statement is to establish that the conditional probability distribution for the displacement-
based response of the structure for a given intensity level is independent of the ground motion
characteristics, namely:
f max |S a (T ) ( x) = f max |S a (T ), M , Rrup , ( x, mi , r j , k ) (6 - 9)
for any mi , r j , k value.
The sufficiency criterion can be also approximated in terms of the (conditional) statistical
moments of the response being independent of the ground motion characteristics. For example, a
first-order measure of the sufficiency criterion can be obtained by establishing that the first
(conditional) moment of the response for a given spectral acceleration level, x, is independent of
the ground motion characteristics:
E[ln max ( x ) | S a (T )] = E[ln max ( x, mi , r j , k ) | S a (T ), M , R rup , ] for all mi , r j and k

(6 - 10)
For a lognormal random variable, the above equation can be written as the equality of the
medians:
max| S ( x ) E [ln max | ( x, mi , r j , k )] (6 - 11)

a (T )
If a ground motion is sufficient with respect to the ground motion characteristics, it can be
demonstrated that the two sides of Equation 6-7 will be always equal (by substituting
max | S a (T ), M , Rrup ,
( x, mi , ri , i ) by max |S a (T ) ( x) in the equation):
N M N Rrup N
ln max |S
a (T )
( x) ln
i =1 j =1 k =1
max | S a ( T )
( x ) . p M , Rrup , |S a (T ) (mi , r j , k | x )
N M N Rrup N
= ln max | S a ( T )
( x) p
i =1 j =1 k =1
M , Rrup , |S a (T ) ( m i , r j , k | x ) = ln max | S
a (T )
( x)
The above conclusion is based on the fact that the sum of the fractions
p M , Rrup , |S a (T ) (mi , ri , i | x) is equal to unity. A second order measure of sufficiency criterion in
Equation 6-9 can be expressed in terms of the (conditional) second moment of the response:
E [(ln max ( x)) 2 | S a (T )] = E [(ln max ( x, mi , r j , k )) 2 | S a (T ), M , R rup , ] for all mi , r j and k

(6 - 12)
Similarly, it can be demonstrated that the two sides of Equation 6-8 will always be equal if the
second order approximation to sufficiency is established.
It is demonstrated above that if the candidate intensity measure is sufficient with respect to
ground motion characteristics, careful selection of ground motion records is not essential, as the
displacement response for a given intensity level is independent of the ground motion
characteristics. It was also discussed that a first-order approximation to the sufficiency criterion
results in the (conditional) median of the displacement response for a given intensity level being
independent of the ground motion characteristics. A second-order approximation to the
sufficiency criterion results in the conditional fractional standard deviation (i.e., the standard
deviation of the natural logarithm) of the displacement-based response the being independent of
the ground motion characteristics.)
It was described how to establish sufficiency by using the residual-residual plots earlier in this
chapter. Now, one might question how that method for establishing sufficiency relates to the
discussions above. The residual-residual plots can reveal possible dependence of the response on
a ground motion characteristic variable once they are both normalized with respect to the
candidate intensity measure. It can be argued that this is approximately equivalent to establishing
the first-order approximation to the sufficiency criterion stated above, because it is based on
determining whether there is linear dependence of the (conditional) mean of the (log) of response
on magnitude (for example), i.e., whether the coefficient 2 in a linear regression
E[ln max | ln S a , M ] = + 1 ln S a + 2 M is zero or not. Therefore, showing that 2 is not
statistically different from zero is a necessary but not sufficient condition to prove sufficiency.
6.16 Investigating the sufficiency of PGA as the intensity measure with respect to
magnitude
The previous studies showed S a (T FM ) to be (at least linear first order) sufficient with respect to
the ground motion characteristics, magnitude M , rupture distance Rrup , and the attenuation
epsilon, . What happens if the intensity measure is not sufficient with respect to the ground
motion characteristic variables? The peak ground acceleration (PGA) is commonly perceived as
an insufficient intensity measure for moderate to long-period buildings. In this section the
sufficiency of PGA as the candidate intensity measure is studied with respect to M , Rrup , and
, for the LA 20 building.

1
L20 combined selection, scaled to 3, linear regression
10
a= 0.02
b=0.39
=0.69
|S
max a 47
13
0 41
28 34
10 17
27
31
3 3024 37 19
43 25 42 35 29
4 36
20 9184012 826
38
716
PGA
14
10 32 39 16
5
22 45
2 11 33
21 23
1
44 46
10
15
2
10
3 2 1
10 10 10
max
Figure 6-22 - The original regression of the displacement-based response on PGA, LA 20

building, the records scaled by a factor of 3.
For reasons discussed in Sections 6.11 and 6.12, the moment magnitude is the most likely ground
motion variable to affect the displacement-based response. Hence, it would be interesting to
investigate the sufficiency of PGA with respect to magnitude. The displacement-based response
of the LA 20 structure to the records scaled by a factor of 3 are plotted versus PGA in Figure
6-22. The original regression prediction of the response as a function of PGA (i.e., the fitted
line) is also plotted in the figure. The residual-residual plot for magnitude as a second regression
variable is illustrated in Figure 6-23.
1
L20, combined selection, scaled by a factor of 3.0
10
29
Residual of the "original" regression on PGA

33
45 34
15
46
36
22 35
6 42 1
4
8 26 19
31
0 39
37 23
10 13
12 38
16 27 24
40 17 32 28 4125
30 7
43 10 18
47
9
14 11
5 21
3 20
44
a = 5.79e16
b = 0.4349
b = 0.1959
2
1
10
2 1.5 1 0.5 0 0.5 1 1.5 2
Residual of regression of magnitude on ln Sa
Figure 6-23 - The residual-residual plot for magnitude as a second regression variable, LA 20
structure, combined record set scaled by a factor of 3.
It can be observed that the residuals of the original regression of ln max on PGA demonstrate a
positive trend with respect to the residuals of the regression of magnitude on PGA. As discussed
before in Section 6.11, large-magnitude events have a relatively strong frequency content in the
short frequency range. Hence, the positive trend observed in Figure 6-23 seems reasonable. The
trend observed in the residual-residual plot in Figure 6-23 is quite strong as indicated by the slope
parameter b=0.435. Also, the p-value=0.0157 suggests that the observed trend is also quite
significant. However, the small reduction in the sigma (i.e., 0.65 in the residual-residual
regression versus 0.69 of the original regression) implies that M is not very helpful in
increasing the efficiency, compared to PGA alone. Therefore, one may conclude that PGA in this
case is not sufficient with respect to magnitude, but the addition of magnitude as a second
regression variable may not (practically) increase the efficiency.
6.17 Adjusting for possible dependencies using weighted regression
It was concluded in the previous section that PGA is not sufficient with respect to the ground
motion magnitude. Using a weighted regression scheme (Weisberg, 1985) may help in reducing
the dependence of the residuals (of the original regression on PGA) on magnitude. Shome
(1999) implemented the weighted regression scheme in order to take into account the effect of the
SMF shape factor in predicting the response.
It should be recalled that regression analysis works by minimizing the sum of the squared errors
(residuals) between the observed drift and the predicted drift. The weighted regression scheme
weights each error term (residual) proportional to its corresponding variance. A similar weighting
scheme is implemented in this section in order to adjust the residuals (of the original regression
of ln max on PGA) with regard to their dependence on magnitude
1
L20 combined selection, scaled to 3, weighted regression
10
a=0.02
b=0.57
theta |S
=0.6775
max a 47
13
0 41
28 34
10 17
27
31
3024 3719
3 43 25 42 35
36
29
20 12 826 4
91840
7 38
16
PGA
14
10 32 39 16
5 22 45
2 11 33
21 23
44 46
1
10
15
2
10
3 2 1
10 10 10
max
Figure 6-24 - The weighted regression of the displacement-based response on PGA, LA 20

structure, the combined set of record scaled by a factor of 3.
It can be argued that the variance of each error term and hence the corresponding weight may be
positively related (ideally, proportional) to the following ratio:
p M |PGA (mi | x) disaggregation

wi (6 - 13)
p M |PGA (mi | x) data
where p M |PGA (m | x) is the is the fraction of the ground motions with magnitude m for a given
PGA equal to x. For the suite of record used in this chapter, p M |PGA (mi | x) data is equal to 1 N T
(i.e., no two records in this set have the exact same magnitude and PGA), where N T is the total
number of records. p M | PGA (mi | x) disaggregation is the fraction of records with magnitude mi for a
given PGA equal to x, estimated from the disaggregation of hazard for a PGA equal to x (see
Bazzurro and Cornell, 1997). This weighted regression scheme is applied to the displacement-
based response of the 20-story structure to the combined set of records scaled by a factor of 3.
Figure 6-24 illustrates the resulting weighted regression prediction plotted against the same data
points in Figure 6-22, plotted by stars with areas proportional to the corresponding weighting
factor for the residuals.
Comparing the weighted regression parameters in Figure 6-24 to those of the original
regression in Figure 6-22, it can be observed that the sigma is somewhat reduced from 0.69 to
0.6775, and also the b-value is increased from 0.39 to 0.57. In order to see how the weighting
affects the sufficiency with respect to magnitude, the weighted residual-residual plot for
magnitude is plotted in Figure 6-25. It should be noted that the residual-residual regression line in
Figure 6-25 has a non-zero intercept. This is because the weighted residuals have a non-zero
expected value. Also it can be observed that in general the smaller residuals have received larger
weights (judging by the radii of the circles).
1
L20, combined selection, scaled by a factor of 3.0
10
33 29
45
15
34 46
Residual for regression on PGA
22 36
35
6 1
42
4 23
26
39
8
31 19
0
10 12 16 3738
32 7
40 13 25
24
17 27 10 30
18
41
14 11 28 9
43 21
47
5
20
3 44
a = 0.177
b = 0.4349
residual|residual2 = 0.6651
2
= 0.1993
b
1
10
2 1.5 1 0.5 0 0.5 1 1.5 2
Residual of the regression of magnitude on ln Sa
Figure 6-25 - The weighted residual-residual plot for magnitude, LA 20 structure, the combined
record set scaled by a factor of 3.
0
10
IM: PGA
IM: PGA, weighted regression
1
10
max
2

10
3
10
4
10
3 2 1
10 10 10
max
Figure 6-26 -The hazard curves for the displacement-based response with PGA as the intensity
measure. The hazard curve in dashed lines is based on weighted regression.
However, it would be interesting to study how the probabilistic assessments for the displacement-
based response are going to be affected by the weighed regression scheme. Figure 6-26 illustrates
the hazard curves for displacement-based response with PGA as the intensity measure. The
hazard curves are calculated by numerical integration of Equation 6-14 below (Similar to
Equation 4-5-a with spectral acceleration as the IM):
max ( y ) = P[
x
max > y | PGA = x] d PGA ( x) (6 - 14)
where (.) is the hazard function or in more detail, the mean annual probability of exceedance.
The term P[ max > y | PGA = x] or the probability of exceeding the displacement-based response
value y for a given value of PGA can be calculated by implementing the regression parameters
discussed previously assuming that the conditional probability distribution for the displacement-
based response is lognormal. The solid line in Figure 6-26 is based on the original regression
parameters from Figure 6-22 and the dashed line is based on the weighted regression
parameters from Figure 6-24. The term PGA ( x ) or the hazard function for PGA can be estimated
by the site-specific PGA hazard curve for a site in LA based on Abrahamson and Silvas
attenuation law for horizontal motion on soil (see Abrahamson and Silva 1997), and calculated
using a probabilistic seismic hazard software developed by Norman Abrahamson (see
Abrahamson 2001, hazard code version 30).
In order to be able to judge if the weighted regression is helpful in adjusting for the dependence
on magnitude, the hazard curve for the displacement-based response has been calculated with the
pair [PGA, M] as the intensity measure. Similar to Section 6.15, the hazard curve can be
calculated by expanding the term P[ max > y | PGA = x] in the expression for the hazard curve in
Equation 6-14 with respect to magnitude (see Luco, 2001):
max ( y ) = P[ max > y|PGA = x,M = mi ] p M|PGA (mi|x) dPGA ( x)

i x
(6 - 15)
In a similar way, the term P[ max > y | PGA = x, M = mi ] can be calculated by implementing the
regression parameters for the multi-variable regression of the displacement-based response on
PGA and M. The parameters for this regression are outlined in Table 6-7. Also the term
p M |PGA (m i | x) can be estimated from the results of the disaggregation of hazard for a PGA equal
to x.
Regression on PGA and M a bPGA bM ln max |PGA, M
scaled by 3 0.0012 0.376 0.4349 0.6536
Table 6-7 The results of the linear multi-variable regression of ln max on ln PGA and M for
the case where the ground motion accelerations are scaled by a factor 3, LA 20 story building
Figure 6-27 illustrates the hazard curve calculated from Equation 6-15 together with the hazard
curve shown in Figure 6-26. The thick line represents the hazard curve using the pair [PGA, M]
as the intensity measure, the thin line represents the hazard curve using PGA as the intensity
measure, and the dashed line represents the hazard curve using PGA as the intensity measure but
adjusting for the dependence on magnitude by weighted regression. It can be observed that the
hazard curve calculated using the weighted regression is almost identical to the one calculated
using the pair [PGA, M] as the intensity measure. This indicates that the weighting scheme is
effective for taking into account the magnitude dependence in the prediction of hazard for the
displacement-based response.
0
10
IM: PGA
IM: PGA, weighted regression with respect to M
IM: [PGA, M]
1
10
max
2

10
3
10
4
10
3 2 1
10 10 10
max
Figure 6-27 - The hazard curve for the displacement-based response with the pair [PGA,M] as the
intensity measure (thick line) compared to the curves in Figure 6-26.
It is argued that, theoretically speaking, careful record selection is not essential if the candidate
intensity measure (IM) is demonstrated to be sufficient with respect to the ground motion
characteristic variables. Sufficiency (see Luco and Cornell, 2001) is a probabilistic criterion for
a preferred IM, in which the structural (displacement-based) response for a given IM level is
conditionally statistically independent of the ground motion characteristics variables. A first-
order approximation to the sufficiency criterion states that the conditional expected value of the
structural response for a given intensity level is conditionally statistically independent of the
ground motion characteristic variables. A suite of linear regressions was used as a statistical tool
for investigating the first-order sufficiency of the candidate IM. In the cases where the first-
order sufficiency for the candidate IM could not be established, a weighted linear regression
scheme based on the results of the seismic hazard disaggregation with respect to the ground
motion characteristic variable(s) was implemented in order to adjust for the observed
dependencies.
In this chapter, the spectral acceleration at the first-mode frequency (FMF) was selected as the
candidate IM. Also the maximum inter-story drift angle max represented the displacement-based
structural response parameter. A suite of (more-or-less arbitrarily selected) ground motion
recordings on stiff soil was used in this chapter for dynamic time-history analyses. In this chapter,
the non-linear dynamic displacement-based responses of two extreme cases of short-period and
long-period structural systems were studied. The short-period system is a generic SDOF bi-linear
model with T=0.1; and the long period system is a 20-story moment-resisting steel frame
structure (LA 20 structure) with T1 = 3.98 . It was demonstrated for both structural systems that
the FMF spectral acceleration is effectively sufficient with respect to ground motion
characteristic variables, magnitude M , distance R rup , and the epsilon of the attenuation
relation. It was also demonstrated that the structural response for a given FMF spectral
acceleration is (strongly) dependent on the shape of the elastic acceleration spectrum. The
spectral shape factor, defined as the spectral acceleration at a period T normalized by FMF
spectral acceleration, was the parameter used to represent the shape of the spectrum. For the high-
frequency bi-linear SDOF system, the response was demonstrated to be positively related to the
spectral shape factors at period longer than the FMF. This observation could be explained by the
elongation of the effective period in the high ductility region. For the long-period 20-story
structure, the response was shown to be positively related to the spectral shape factors at shorter
periods close to that of the second-mode frequency (SMF). This confirmed the a priori
expectation that the response of the tall long-period moment-resisting frame be dependent on the
higher modes.
Among the ground motion characteristic variables, the moment magnitude is generally perceived
to be the most likely one to affect the structural response once it is conditioned on FMF spectral
acceleration. The validity of such perception is tested in this chapter by studying the correlation
between the spectral shape factor and magnitude based on the attenuation relation predictions and
also on the suite of records used in this chapter. It is demonstrated that high-frequency spectral
shape factors (i.e., periods close to or smaller than TSM ) are negatively correlated with
magnitude. This could be explained by the fact that the small-magnitude events are (physically)
less capable of generating long-period (low-frequency) waves. This together with the observed
positive correlation between the response of the long-period 20-story structure and the high
frequency spectral shape factors leads one to expect that the structural response will be negatively
correlated with magnitude. This seems to contradict previous observations indicating that the
response for a given FMF spectral acceleration does not display a statistically significant
dependence on magnitude. However, a closer look at the relations between magnitude and
spectral shape factor and between the response and the shape factor indicates that there is
considerable uncertainty (i.e., uncertainty due to record-to-record variability) in these relations
(judged by the large p-values). On the other hand, the weak negative correlation observed
between the (conditional) response and magnitude is categorized as statistically insignificant due
to the large uncertainties involved in the predictions.
Similar observations have been made for different ductility levels. The yield strength level for the
high-frequency bi-linear SDOF system is adjusted so that different levels of ductility in the
response are achieved. It was demonstrated, based on the response of the SDOF system with
modified yield strength, that the FMF spectral acceleration is effectively sufficient with respect
to ground motion characteristic variables. For the 20-story long-period structure the ground
motion records are scaled by a factor of 3 in order to achieve a higher level of ductility in the
response (the response to the un-scaled records was effectively in the linear range). Based on the
response of the 20-story structure to the suite of records scaled by a factor 3, it was again
established that FMF spectral acceleration is sufficient with respect to ground motion
characteristic variables.
In order to study a case in which the candidate IM is not sufficient with respect to ground motion
characteristic variables, the peak ground acceleration (PGA) was adopted as the candidate IM.
Based on the response of the 20-story structure to the suite of records scaled by a factor 3, it was
observed that PGA is not sufficient with respect to ground motion magnitude. This again
indicated that the uncertainty due to record-to-record variability in the correlation between the
conditional response and magnitude very large and hence the observed dependence is not very
strong. A weighted regression scheme based on the results of the disaggregation of the hazard for
PGA with respect to magnitude was used to adjust for the dependence on magnitude. The
estimated hazard curve for the displacement-based response based on weighted regression was
very close to the one derived by adopting the pair [PGA, M] as the intensity measure. Hence, it
was concluded that the weighted regression enhances the hazard estimations for the displacement-
based response of the 20-story structure.
6.18.1 Limitations
It is important to note that the observations made in this chapter are subject to certain limitations.
The suite of records used in this chapter excludes near-source records that usually have more
propensity for causing near-source effects such as directivity in the response. Also the
observations made for the long-period structure are specific to the 20-story structure studied in
this chapter. The ground motions were scaled in some cases in order to achieve higher degrees of
ductility. It was implicitly assumed in doing so the scaled ground motions are able to affect the
response more-or-less (more accurate?) in the same way as the ground motions that are
(naturally) stronger. This is supported by the demonstrated sufficiency which implies that larger
M and or smaller R rup (which would imply strong records) would not affect the responses
(given spectral acceleration) Finally the sufficiency of the FMF spectral acceleration was
established using a suite of simple mono-variable regressions; this ignored possible correlations
between the ground motion characteristic variables. A more thorough approach would implement
multi-variable regression in order to take into account possible cross-correlations between ground
motion characteristic variables.
6.19 Acknowledgments
PEER, funded by the U.S. National Science Foundation (Project No. 3202000). They would also
like to thank Dr. Nicolas Luco for providing the MDOF structural model.
Chapter 7
Conclusions
In this work, an analytic foundation, using basic probabilistic concepts, for design and assessment
of moment-resisting frame structures under seismic loads is developed (Chapter 2). This
foundation forms the theoretical basis to alternative formats suitable for implementation in design
and assessment guidelines (Chapter 3). Examples of how to organize ground motion records,
choose intensity levels, perform non-linear time history analyses, and, process the results, in order
to obtain parameter estimates for the analytic framework, are demonstrated (Chapters 4,5,6).
7.1.1 Chapter 2: A technical framework for probability-based design

and assessments
The probabilistic foundation developed, yields an analytic closed-form expression for the annual
frequency of exceeding specified structural performance levels, or more briefly limit state
frequency, is derived based on simplifying assumptions. The limit state frequency is derived by
assuming that the parameters involved in the assessments have a stochastic nature, which is
modeled by considering two different types of uncertainty. The first type identifies the more
familiar natural variability in the parameters, and is referred to as randomness or more
precisely the aleatory uncertainty. The second type addresses limited knowledge and data and is
referred to as uncertainty or epistemic uncertainty. This second kind of uncertainty can be
reduced by more acquiring data (larger sample sizes) and/or by performing more thorough
research.
The derivation of the limit state frequency employs a probabilistic tool known as the total
probability theorem (TPT) in order to decompose the derivations into smaller, and less complex
parts. Therefore, the process of evaluating the limit state frequency involves additional
interface variables. Two distinct approaches for deriving the expression for limit state
frequency are presented, namely, the displacement-based approach, and the ground motion
intensity-based approach. The displacement-based approach evaluates the limit state frequency as
the frequency that a displacement-based demand variable exceeds the corresponding limit state
capacity. The derivations in this case are performed in two steps: 1) Evaluating the frequency that
the displacement-based demand exceeds a given value by decomposing it with respect to the
ground motion intensity level and then composing the results by integration over all possible
intensity levels. This first step is done by employing the total probability theorem and an interface
variable representing the ground motion intensity. This variable is referred to as the intensity
measure (IM). The assumptions made in this step of the derivation include, approximating the
frequency that the IM exceeds a certain level, also know as the hazard for the IM, by a power-
law function, modeling the probability distribution of the displacement-based demand for a given
level of ground motion intensity by a lognormal distribution, and assuming that this lognormal
distribution is defined by a median (central value) that is itself a power-law function of the
ground motion IM and a (log) standard deviation (dispersion measure) that is invariant with
respect to the ground motion intensity, 2) The second (final) step is to evaluate the frequency that
the displacement-based demand exceeds capacity by decomposing it into (conditional)
frequencies of exceeding given values for the limit state capacity and then composing these
frequencies by integration over all possible values of capacity. In this step it is assumed that
probabilistic distribution of the (displacement-based) capacity can be modeled with a lognormal
distribution with constant median and standard deviation and also that the capacity and demand
are uncorrelated. The second or ground motion intensity-based approach evaluates the mean
annual frequency that the IM variable exceeds the corresponding limit state capacity IM or more
briefly the IM capacity for a specific limit state (also called limit state frequency). The derivation
involves decomposing the limit state frequency into conditional limit state frequencies that the IM
exceeds IM capacity for a given intensity measure and integrating the conditional limit state
frequencies over all levels of ground motion intensity.
7.1.2 Chapter 3: Probability-based Demand and Capacity Factor

Design (DCFD) formats
The closed-form analytic expression(s) derived for the limit state frequency can be formed into
alternative formats. These formats re-shape the probabilistic expression of exceeding a certain
limit state into alternative displacement-based or ground motion intensity-based
design/assessment criteria. These criteria, being expressed in common engineering terms rather
than the more abstract probabilistic ones, can be implemented in existing design and assessment
procedures and guidelines.
Demand and capacity factored design (DCFD) represents a family of displacement-based design
formats that are distinguished with regard to the type of uncertainties considered in the
formulation of the limit state frequency. This format has been already implemented in, FEMA
350 for the design of new steel moment resisting frames, FEMA 351 for the assessment of the
existing steel moment resisting frames, and ISO guidelines for the design of offshore structures.
Fragility/hazard format represents a ground motion intensity-based family of design formats, also
capable of considering both types of uncertainty. The fragility/hazard has a graphic representation
based on fragility and hazard curves and forms of it have been implemented in the DOE 1020
seismic criteria (Kennedy and Short, 1994). The consideration of the epistemic uncertainty in the
development of these formats may result in designing the structure with certain degree of
confidence or assess the level of confidence in the design of an existing structure.
7.1.3 Chapter 4: Alternative non-linear demand estimation methods
Non-linear displacement demand estimation methods can be used in the context of the DCFD
format in order to design and assess the performance of the structures under seismic loads. It is
demonstrated how alternative non-linear dynamic analysis procedures can be implemented in
order to assess the performance of an existing (older) reinforced concrete frame for a specific
limit state, using DCFDs displacement-based criteria. These non-linear procedures involve,
selection of intensity levels, performing non-linear time history analyses, and post-processing the
results. Special attention is given to the estimation of the displacement-based demand parameter,
which is compared to the limit state capacity in the framework of DCFD assessment criterion.
The demand estimations obtained using the analytic DCFD format (based on the analytic
assumptions outlined above), are compared to the results obtained by numerical integration
(based on empirical probabilistic representations of the parameters). In order to be able to
compare the accuracy of the alternative demand estimation methods, the DCFD displacement-
based design/assessment criterion is extended to a more general equivalent form. Hence these
methods are compared in the light of the equivalent probabilistic demand estimations they
provide.
It was desirable to lay out the methods based on the range of ground motion intensity levels
included in the dynamic analyses performed. Therefore, these methods are discussed in two
categories, namely, the wide-range and narrow-range demand estimation methods. In general
the non-linear procedures discussed are employed in order to obtain local demand parameter
estimates for the DCFD format. The narrow-range methods, provide parameter estimations with
relatively small number of non-linear dynamic analyses. It is shown that the accuracy of these
methods is significantly improved when the range of the ground motion intensity levels are
slightly increased to accommodate the estimation of a demand parameter related to the gradient of
displacement-based demand with respect to ground motion intensity level. The wide-range
methods involve more extensive analysis efforts in order to provide local parameter estimations
as a function of the ground motion intensity level. Although the derivation of the DCFD design
format is based on certain analytic assumptions about the parameters, (e.g., constant standard
deviation and power-law median for the displacement demand), it is shown that incorporating the
local estimates, provided by the wide-range methods, into the DCFD formulation provides
accurate results. However, the non-linear nature of the displacement-based demand at higher
intensity levels make the resulting parameter estimates subject to potential (local) irregularities in
displacement response in the non-linear range. Hence, the resulting (local) parameter estimates
should be treated with care and with intuition of the actual physical behavior at the corresponding
intensity level. The wide-range methods can also be used in order to obtain parameter estimates
related to limit state capacity. An application of such methods to the estimation of the global
dynamic instability limit state is discussed.
7.1.4 Chapter 5: Probabilistic seismic assessments in the region of

global dynamic instability
One of the main assumptions leading to the derivation of the DCFD format is that the
displacement-based response of the structure for a given ground motion intensity level can be
adequately modeled by a lognormal distribution. However, this assumption may not be true for
displacements close to the onset of global dynamic instability, also referred to as collapse, in
the structure. The three-parameter distribution is an alternative probability distribution that is
defined by using the total probability theorem (TPT) to decompose the displacement-based
demand into two mutually exclusive states of collapse and no collapse, where the lognormal
distribution is still is applied only in order to model the no-collapse part. Alternative methods
for incorporating the three-parameter distribution or parallel probabilistic models based on
empirical distributions, other than the lognormal distribution, into the demand assessments are
discussed. As a special case, an alternative formulation for the DCFD format, representing the
displacement-based demand (for a given intensity measure) by the three-parameter distribution, is
derived. It is important to note that the alternative DCFD formulation is only applicable to
assessments for limit states other than the so-called collapse limit state (also called the
exogenous limit states). It is demonstrated that wide-range methods can also be used to provide
parameter estimates (as a function of the intensity level) for the alternative DCFD formulation in
the range of collapse. Alternatively, a pseudo lognormal distribution can be constructed based
on the three-parameter distribution. This pseudo lognormal distribution can be used to make local
parameter estimations in the range of global dynamic instability in the original DCFD
formulation.
7.1.5 Chapter 6: Non-linear response dependence on ground motion

characteristics with implications for ground motion record
selection
The closed form expression for the limit state probability was derived by adopting an interface
variable referred to as IM in order to represent the ground motion intensity. It was implicitly
assumed that the adopted IM can provide an adequate presentation of the ground motion
characteristics, such as magnitude and source-to-site distance. On the other hand, the non-linear
dynamic analysis procedures implemented for making probabilistic assessments involve selecting
a suite of ground motion records. Inevitably, it is of interest to know what a representative ground
motion selection is, what are the criteria for creating such selection, and whether it depends on the
ground motion intensity level. It is demonstrated that the criteria for selecting a representative
suite of ground motion records are related to the criteria measuring the adequacy of an adopted
IM in representing the ground motion characteristics. Adequacy of an IM is measured through its
sufficiency and efficiency. A sufficient IM is one that renders the structural response (here
the displacement-based response) conditionally statistically independent of the ground motion
characteristics, such as moment magnitude and closest source-to-site distance, for any given
intensity level. An efficient intensity measure predicts the structural response with relatively
small record-to-record variability (i.e., minimizes the dispersion due to randomness).
Theoretically speaking, careful selection of the ground motion records is not essential if the
adopted IM is shown to be sufficient with respect to the ground motion characteristic variables. A
first-order approximation to the sufficiency criterion states that the expected value of the response
is functionally independent of the ground motion characteristics for any given intensity level. In
this context, a suite of linear regressions can be used to establish (linear) first order sufficiency of
the adopted IM for various intensity levels. In the cases where sufficiency is not established, a
weighted linear regression scheme based on the results of seismic hazard disaggregation may be
used to enhance the probabilistic demand estimations.
It was demonstrated that, for the structure studied when maximum inter-story drift angle is
adopted as the displacement-based demand, the spectral acceleration at the first-mode frequency
(FMF) is a sufficient IM with respect to the ground motion magnitude, source-to-site distance,
and the epsilon of the attenuation law. The sufficiency was established for different ductility
levels through the case study of two structures with very short and long first-mode periods,
subjected to a suite of more-or-less arbitrarily selected ordinary (i.e., no near-source recordings)
ground motion recordings on stiff soil. It was also demonstrated for both cases that the
displacement-based response (for a given FMF spectral acceleration) depends on the shape of the
elastic response spectrum. The short-period SDOF system demonstrated dependence on the shape
of the spectrum at longer periods, whereas for the long-period MDOF system the response was
dependent on the shape of the spectrum at shorter periods (i.e., higher frequencies). This could be
explained by the elongation of the natural period for the short-period system in the nonlinear
range and the effect of higher modes for the long-period system. The observed sufficiency with
respect to ground motion characteristics may contradict the common perception that the ground
motion magnitude is likely to affect the structural response specially for systems with very long
natural periods. This perception seems in tune with the observed correlation between the spectral
shape at higher frequencies and magnitude, together with the demonstrated dependence of the
response of the long-period structure on spectral shape. It can be argued that the although these
two pair-wise correlations exist (as measured by the slopes of the fitted lines), they are not strong
enough to induce (statistically) significant correlation between response and magnitude (in part
due to large variability of data points around the fitted line).
The peak ground acceleration (PGA) is shown to be insufficient with respect to the ground
motion magnitude for the long-period system in the non-linear range of response. A weighted
regression scheme based on the results of the disaggregation of seismic hazard for PGA with
respect to magnitude is used in order to take into account the magnitude-dependence. The
displacement-based demand estimations (demand hazard) based on the weighted regression show
very good agreement with those based on adopting the pair [PGA, M] as a vector-valued intensity
measure.
7.2 Limitations and future work
This work presents an analytic framework for probabilistic seismic assessments of moment-
resisting frame structures. Examples of how to organize ground motion records, intensity levels,
perform non-linear dynamic analyses, and process the results in order to perform probabilistic
seismic assessments are also demonstrated. At every step, certain assumptions were made. These
assumptions define the limitations and the scope of this work as well as possible schemes for
future research.
The assumptions underlying the derivation of an analytic closed-form framework for probabilistic
assessments were outlined above. They consist of assuming that 1) the IM hazard can be modeled
as a power-law function of the IM, 2) the conditional distribution of the displacement-based
demand can be modeled by a lognormal distribution with its median approximated as a power-
law function of the IM and constant standard deviation, 3) the distribution of the displacement
capacity can be modeled by a lognormal distribution with constant median and standard
deviation, 4) the displacement-based demand and capacity are independent, and 5) the epistemic
uncertainties in hazard, demand and capacity were represented by lognormal distributions.
Some of these assumption were relaxed later by obtaining local estimates of the IM hazard and/or
by implementing nonlinear dynamic methods in order to obtain local estimates of the conditional
median and standard deviation of the displacement-based demand. It was demonstrated that the
closed-form framework was still able to provide reasonable estimates even though the underlying
assumptions were not strictly correct. It is important to note that the accuracy of the closed-form
framework for seismic assessments was compared to numerical solutions based on non-
parametric probabilistic distributions. These numerical solutions were in turn obtained based on
certain assumptions, such as, 1) the suite of ground motions used were the same for all different
IM levels considered. 2) these records did not include near-source recordings, 3) it was assumed
that scaling is justified 4) a limited number of ground motion records were used, 5) the IM hazard
calculations did not include directivity parameters, and, 6) the numerical integration was carried
out assuming that the adopted IM is sufficient with respect to ground motion characteristics.
Finally, all of these studies in Chapters 4 and 5 used a single structural model. While it was
intentionally chosen to include characteristics such as cyclic strength degradation, it would be
desirable to extend these studies to a broader population of structural models to seek any limits to
the conclusions drawn here.
It was demonstrated how to measure the epistemic uncertainty associated with limited sample
size, i.e., number of ground motion records. However, future work should demonstrate how the
uncertainties associated with specific parameter estimations can be propagated into the
derivations. Also, the validity of assuming a lognormal presentation for the epistemic uncertainty
was not investigated.
The presented framework is based on the mean annual frequency of exceeding limiting states that
are defined in terms of structural response parameters . It is also important to derive analytic
frameworks based on life-cycle costs, since the probabilistic performance objectives are ideally
based on life cycle considerations.
As it was mentioned before, the sufficiency investigations for FMF spectral acceleration were
done by using a suite of scalar linear regressions. Therefore, the sufficiency of FMF spectral
acceleration was established for one ground motion variable at a time. Ideally, sufficiency should
be established for all the effective ground motion variables simultaneously, meaning that multi-
variable regressions are more appropriate for this purpose. However, it is believed that the
method presented in this work is capable of distinguishing any major dependence on ground
motion characteristics. Also the linear regression scheme used can only establish linear (first-
order) sufficiency for the range of intensity that the ground motion records cover. Also, the
conclusions are specific to the suite of ground motion records used, which do not include near
source recordings. It is important to investigate the sufficiency with respect to the ground motion
variables that include directivity variables and based on a suite of records that does include near-
source recordings. In order to establish the sufficiency of FMF spectral for the long-period system
for different ductility levels, the ground motion records were scaled. A thorough investigation
that would confirm scaling for long-period structures is needed.
As a general note, the structural systems used in this work are intended to serve as case studies
and the accuracy of the modeling assumptions and their limitations were not the focus of this
thesis. Furthermore, while the structures in Chapter 6 were specifically chosen to be extreme
cases with respect to the periods and higher-mode effects, the limits of applicability of the
conclusions made here have not been established.
Appendix A
The expected value of Y where Y is a log normal random variable:
Assume lnY is a normal random variable (i.e., Y is lognormal) with mean m and standard
deviation . One can always write the following relationship for Y raised to a power, :
Y = e ln Y
lnY can be transformed into a standard normal variable U,
ln Y m
U= (A - I)

for which, the standard normal probability density function (PDF) at U=u is equal to:
1
1 u2
(u ) = e 2 (A - II)
2
Based on the linear relation between lnY and U (Equation A-I), and the standard normal PDF for
U (Equation A-II), the PDF for normal random variable lnY at lnY=x can be obtained as:
1 1 ln Y m 2
1 1 u2 1 ( )
f ln Y ( x) = (u ) = e 2 = e 2 (A - III)
2 2
where f(.) denotes the PDF function. The expected value for a function g(.) of a continuous
random variable Z can calculated as:

E[ g ( Z )] =

g ( z) f Z ( z ) dz (A - IV)
Therefore, the expected value of g (Y ) = Y can be written as (using Equations A-II, A-III and A-
IV):
+ + 1 xm 2
1 ( )
e
ln Y x x
E[Y ] = E[e ]= f ln Y ( x) dx = e e 2 dx

2
After some algebraic operations, which involves adding and subtracting some(necessary) square
terms the following equation is obtained:
+ ( x m)2 1 2 2 + ( x m 2 ) 2
1 1

ln Y x 2 2 m 2 2
E (e )= e e dx = e e 2 e dx

2
2
We can recognize that the term inside the integral is nothing but the PDF for a normal variable
with a mean equal to m + 2 , and a standard deviation equal to, . Therefore, the resulting
integral (from to ) is equal to unity. Hence, the expected value of Y is simplified to the
product of the following two terms:
1
2 2
E[Y ] = E[e ln Y ] = e m .e 2
For a lognormal random variable, the mean of the logarithm of the variable is equal to the
logarithm of the median of the variable (see Benjamin and Cornell, 1970):
ln Y = E[ln Y ]
where (.) denotes the median. Hence, for normal random variable lnY with mean m and standard
deviation , the expected values of Y can be written as:
1 2 2 1 2 2 1 2 2

ln Y m ln Y
E[Y ] = E[e ]=e e2 =e e2 = (Y ) e2 (A - V)
Thus, the expected value of a lognormal random variable raised to a power can be expressed as
the product of the median value raised to the power times a magnification factor, which is an
1 2
exponential function of the variance of lnY times .
2
Appendix B
Statement of the Total Probability Theorem
Given a set of mutually exclusive and collectively exhaustive events, B1, B2 , , Bn, the
probability P[A] of another event A can always be expanded in terms of the following joint
probabilities (see Benjamin and Cornell 1970):
n
P[ A] = P[ B1 A] + P[ B 2 A] + . . . + P[ Bn A] = P[ B A]
i
i (B - I)
Appendix C
The combined record selection used in Chapters 4 and 5(part I):
N. Ear thquake Station C om ponent M w R (km )

1 C ape M endocino, 1992 Fotuna Blvd. 0 7.1 23.6
2 C halfant V alley , 1986 M c. Gee Greek 270 6.2 36.3
3 C oalinga, 05/09/1983 8 Sub (tem p) 90 5 14.5*
4 C oalinga, 05/02/1983 Park field, Zone 3 90 6.4 36.4
5 H ollister, 01/26/1986 H ollister, Diff A rray #1 125.5 5.4 16.9*
6 Im perial V alley , 10/15/1979 El C enntro A rray #12 230 6.5 18.2
7 L anders, 06/28/1992 Barstow 0 7.3 36.1
8 L iverm ore, 01/24/1980 San R am on, Eastm an Kodak180 5.8 17.6
9 L om a Prieta, 10/18/1989 C oy ote L ak e Dam , DO W N 195 6.9 22.3
10 San Fernando, 02/09/1971 W hittier N arrow s Dam 233 6.6 45.1
11 M organ H ill, 04/24/1984 H ollister, Diff A rray #3 165 6.2 28.3
12 M t .Lew is, 03/31/1986 H alls V alley 90 5.6 15.5*
13 N orthridge, 01/17/1994 L A , Saturn St. 20 6.7 30
14 N . Palm Springs, 07/08/1986 Indio 315 6 39.6
15 San Fernando, 02/09/1971 T erm inal Island 249 6.6 69.2
16 Superstition H ills(A ), 11/24/1W ildlife L iquefaction A rray 360 6.3 24.7
17 W hittier N arrow s, 10/01/1987C astlegate St. 0 6 16.9
18 L om a Prieta, 10/18/1989 A gnew s State H ospital 90 6.9 28.2
19 N orthridge, 01/17/1994 24303 L A - H olly w ood 90 6.7 25.5
20 Im perial V alley , 10/15/1979 C om puertas 285 6.5 32.6
21 Im perial V alley , 10/15/1979 Plaster C ity 135 6.5 31.7
22 M organ H ill, 04/24/1984 H ollister, Diff A rray 255 6.2 28.3
23 Kern C ounty , 07/21/1952 L A , H olly w ood Stor L ot 180 7.4 120.5
24 L om a Prieta, 10/18/1989 A nderson Dam 270 6.9 21.4
25 L om a Prieta, 10/18/1989 C oy ote L ak e Dam 285 6.9 22.3
26 Im perial V alley , 10/15/1979 El C entro A rray #12 140 6.5 18.2
27 Im perial V alley , 10/15/1979 C ucapah 85 6.5 23.6
28 N orthridge, 01/17/1994 L A , H olly w ood Stor FF 360 6.7 25.5
29 L om a Prieta, 10/18/1989 Sunny vale, C olton A ve. 270 6.9 28.8
30 L om a Prieta, 10/18/1989 A nderson Dam 360 6.9 21.4
31 Im perial V alley , 10/15/1979 C hihuahua 12 6.5 28.7
33 Im perial V alley , 10/15/1979 W estm orland Fire Station 90 6.5 15.1
34 L om a Prieta, 10/18/1989 H ollister South & Pine 0 6.9 28.8
35 L om a Prieta, 10/18/1989 Sunny vale, C olton A ve. 360 6.9 28.8
36 Superstition H ills, 11/24/1987W ildlife L iquefaction A rray 90 6.7 24.4
37 Im perial V alley , 10/15/1979 C hihuahua 282 6.5 28.7
39 Im perial V alley , 10/15/1979 W estm orland Fire Station 180 6.5 15.1
40 M t .Lew is, 03/31/1986 H alls V alley 90 5.6 15.5*
41 L om a Prieta, 10/18/1989 W aho 0 6.9 16.9
42 Superstition H ills, 11/24/1987W ildlife L iquefaction A rray 360 6.7 24.4
43 Im perial V alley , 10/15/1979 C om puertas 15 6.5 32.6
44 Im perial V alley , 10/15/1979 Plaster C ity 45 6.5 31.7
45 M organ H ill, 04/24/1984 H ollister, Diff A rray 165 6.2 28.3
46 Kern C ounty , 07/21/1952 H oly w ood Stor Lot 90 7.4 120.5
47 L om a Prieta, 10/18/1989 W aho 90 6.9 16.9
*Hypocentral distance instead of closest distance.
The Combined Record Selection (part II)
N. PGA (g) PGV (cm/sec) PGD (cm) Mechanism HP (HLP (Hz) Site Soi Hanging W all
1 0.116 30 27.59 reverse 0.07 23 B unknown
2 0.078 2.3 0.65 strike-slip 0.1 50 _ NA
3 0.216 12.2 1.42 reverse 0.1 40 _ unknown
4 0.164 24.5 4.94 reverse-oblique 0.1 22 _ unknown
5 0.101 9.3 1.95 strike-slip 0.1 45 _ NA
6 0.116 21.8 12.06 strike-slip 0.1 40 C NA
7 0.132 21.9 20.59 strike-slip 0.07 23 B unknown
8 0.154 18.9 6.13 strike-slip 0.08 20 C NA
9 0.16 13 6.11 reverse-oblique 0.1 30 IHD-B Foot Wall
10 0.107 9.8 5.04 reverse 0.1 20 IQD unkown
11 0.078 7.2 1.47 strike-slip 0.1 30 IQD NA
12 0.159 18.4 4.4 unknown 0.1 20 _ unknown
13 0.474 34.6 6.55 reverse 0.1 30 C unknown
16 0.134 13.4 5.2 strike-slip 0.2 50 IQD NA
19 0.231 18.3 4.81 reverse 0.2 23 IPD unknown
20 0.147 9.5 2.49 strike-slip 0.2 _ _ NA
21 0.057 5.4 1.94 strike-slip 0.1 40 _ NA
22 0.088 11.9 1.89 strike-slip 0.2 23 _ NA
24 0.244 20.3 7.73 reverse-oblique 0.2 41 B unknown
25 0.179 22.6 13.2 reverse-oblique 0.1 29 IHD Footwall
26 0.143 17.6 11.3 strike-slip 0.1 40 C NA
27 0.309 36.3 10.44 strike-slip 0.05 _ C NA
28 0.358 27.5 3.04 reverse 0.2 23 _ unknown
29 0.207 37.3 19.11 reverse-oblique 0.1 40 C unknown
31 0.27 24.9 9.08 strike-slip 0.05 _ C NA
32 0.117 14.7 7.33 strike-slip 0.2 40 C NA
33 0.074 21.2 16.59 strike-slip 0.1 40 C NA
35 0.209 36 16.9 reverse-oblique 0.1 32 _ unknown
36 0.181 29.9 19.9 strike-slip 0.1 50 _ NA
37 0.254 30.1 12.89 strike-slip 0.05 _ _ NA
38 0.139 13 5.84 strike-slip 0.2 40 _ NA
39 0.11 21.9 10 strike-slip 0.1 40 _ NA
40 0.159 18.4 4.4 unknown 0.1 20 _ unknown
41 0.37 27.2 3.84 reverse-oblique 0.1 _ _ unknown
42 0.207 34.5 21 strike-slip 0.1 40 _ NA
43 0.186 13.9 2.92 strike-slip 0.2 _ C NA
44 0.042 3.2 1.34 strike-slip 0.1 40 C NA
45 0.089 8.7 1.72 strike-slip 0.2 29 _ NA
46 0.042 7.5 4.79 reverse-oblique 0.2 15 C NA
47 0.638 38 5.85 reverse-oblique 0.1 _ _ unknown
Bibliography
Abrahamson, N. A.; Silva, W. J., (1997), Empirical response spectral attenuation relations for
shallow crustal earthquakes, Seismological Research Letters, 68, 1, Jan.-Feb. 1997, pages 94-
127.
AISC. (1994). Manual of steel construction load and resistance factor design, Second Edition,
American Institute of Steel Construction Inc., Chicago, IL.
Alavi, Babak and Krawinkler, Helmut (2000). Consideration of near-fault ground motion Effects
in seismic design, Proceedings of 12th World Conference on Earthquake Engineering, New
Zealand Society for Earthquake Engineering, Upper Hutt, New Zealand.
ATC-40 (1996), Methodology for seismic evaluation and retrofit of existing concrete buildings,
Applied Technology Council, Redwood City, CA.
Banon, H., Cornell, C.A., Crouse, C.B., Marshall, P., Nadim, F., and Younan, A. H., (2001) ISO
seismic design guidelines for offshore platforms, OMAE2001/S&R-2114, Proceedings, 20th
OMAE, Rio de Janiero, June, 200l
Bazzurro, P.; Cornell, C. A., (1994), Seismic hazard analysis of nonlinear structures. I:
Methodology, Journal of Structural Engineering, 120, 11, Nov. 1994, pages 3320-3344.
Bazzurro, Paolo; Cornell, C. Allin., (1998), Disaggregation of seismic hazard, Bulletin of the
Seismological Society of America, 89, 2, Apr. 1999, pages 501-520.
Bazzurro, Paolo; et al. (1998), Three proposals for characterizing MDOF nonlinear seismic
response, Journal of Structural Engineering, 124, 11, Nov. 1998, pages 1281-1289.
Bazzurro, P.; Cornell, C. A.; Pelli, F. ,(1999), Site- and soil-specific PSHA for nonlinear soil
sites, Advances in Earthquake Engineering Vol. 4, Earthquake Resistant Engineering Structures
II, WIT Press, Southampton, United Kingdom and Boston, 1999, pages 331-341.
Beck, J. L. (1999); et al., Multi-criteria optimal structural design under uncertainty, Earthquake
Engineering & Structural Dynamics, 28, 7, pages 741-761
Benjamin, Jack R. and Cornell, C. Allin (1970). Probability, statistics and decision for civil
engineers, McGraw-Hill.
Collins, K. R., Wen, Y.K., and Foutch, D.A., (1996), Dual-level design: A reliability-based
methodology, Earthquake Engineering and Structural Dynamics, Vol. 25, No.12, pp 1433-1467.
Cornell, C.A., and Krawinkler, H. (2000). Progress and Challenges in Seismic Performance
Assessment, PEER newsletter, spring 2000.
Cornell, C. Allin, Jalayer, Fatemeh, Hamburger, Ronald O., and Foutch, Douglas A. (2002). The
probabilistic basis for the 2000 SAC/FEMA steel moment frame guidelines, ASCE Journal of
Structural Engineering, Vol. 128, No.4, pp 526-533.
DOE (1994). Natural phenomena hazards design and evaluation criteria for Department of
Energy Facilities, DOE-STD-1020-94, U. S. Dept. of Energy, Washington, D. C.
Dotiwala, Faridun S., DSouza, Jonathan T. and Pincheira, Jos A, (1998). Drain2D-UW users
manual, Department of Civil and Environmental Engineering, University of Wisconsin -
Madison.
Efron, Bradley and Tibshirani, Robert J. (1993). An introduction to bootstrap, Chapman &
Hall/CRC.
Elwood, Kenneth J.; Moehle, Jack P., (2002), Shake table tests on the gravity load collapse of
reinforced concrete frames, Seventh U.S. National Conference on Earthquake Engineering
(7NCEE): Theme: Urban Earthquake Risk, Earthquake Engineering Research Institute, Oakland,
California, 2002.
FEMA 273 (1996), NEHRP Guidelines for the seismic rehabilitation of buildings, building
seismic safety council, Washington, D.C.
FEMA 350 (2000). Recommended seismic design criteria for new steel moment-frame
buildings. Report No. FEMA-350, SAC Joint Venture, Federal Emergency Management
Agency, Washington, DC.
FEMA 351 (2000). Recommended seismic evaluation and upgrade criteria for Existing Welded
Steel Moment-Frame Buildings. Report No. FEMA-351, SAC Joint Venture, Federal
Emergency Management Agency, Washington, DC.
FEMA 352 (2000). Recommended post-earthquake evaluation and repair criteria for welded
steel moment-frame buildings. Report No. FEMA-352, SAC Joint Venture, Federal Emergency
Management Agency, Washington, DC.
FEMA 356 (2000), Pre-standard and Commentary for the Seismic Rehabilitation of Buildings,
Federal Emergency Management Agency, Washington, DC.
Galambos, Theodor V. and Ravindra, M. K. (1978). Proposed Criteria for Load and Resistance
Factored Design, Engineering Journal ASCE, Vol.15, No.1, pp.8-17
Goel, Rakesh K., et al. (2000). Evaluating current procedures and modeling for seismic
performance of reinforced concrete buildings, Proceedings of 12th World Conference on
Earthquake Engineering, New Zealand Society for Earthquake Engineering, Upper Hutt, New
Zealand, 2000, Paper No. 2060.
Han, S.W., and Y.K. Wen (1997), Method of reliability-based seismic design. I: Equivalent non-
linear systems. II: Calibration of code parameters, Journal of Structural Engineering, Seattle,
WA.
Jalayer, F. and Cornell, A. (2003). A technical framework for probability-based demand and
capacity factor (DCFD) seismic formats. RMS Technical Report No.43 to the PEER Center,
Department of Civil and Environmental Engineering, Stanford University, Stanford, CA.
Jalayer, Fatemeh, and, Cornell, C. Allin (2003). Alternative nonlinear demand estimation
methods for probability-based seismic assessment to be submitted to Earthquake Spectra (The
companion paper)
Jalayer, Fatemeh, and, Cornell, C. Allin (2003). Probability-Based Seismic Assessment: Non-
linear Dynamic Displacement Estimation in the Region of Global Instability to be submitted to
Earthquake Spectra.
Kennedy, R. P., and, Ravindra, M. K. (1984). Seismic fragilities for nuclear power plant risk
studies, Nuclear Engineering and Design, 79, 1, May 1984, pages 47-68.
Kennedy, R. P., and Short, S. A. (1994). Basis for seismic provisions of DOE-STD-1020,
UCRL-CR-111478 and BNL-52418, Lawrence Livermore National Laboratory and Brookhaven
National Laboratory.
Kramer, S. L., (1996), Geotechnical earthquake engineering, Prentice Hall.
Krawinkler, H., (1996) A few basic concepts for performance based design, Eleventh World
Conference on Earthquake Engineering, Pergamon, Elsevier Science Ltd., Paper No. 1133.
Krawinkler, H. (1999). Advancing performance-based engineering, PEER newsletter, Vol. 2

No. 1 January 1999.
Luco, N.; Cornell, C. A. (1998) Seismic drift demands for two SMRF structures with brittle
connections, Structural Engineering World Wide 1998, Elsevier Science Ltd., Oxford, England,
1998, Paper T158-3.
Luco, N. and Cornell, C. A. (2000) Effects of connection fractures on SMRF seismic drift
demands, ASCE Journal of Structural Engineering, Vol. 126, No. 1, pp.127-136.
Luco, Nicolas and Cornell, C. Allin.(2001). Structure-specific scalar intensity measures for near-
source and ordinary earthquake ground ,motions, Submitted for Publication: Earthquake
Spectra, April 2001.
McGuire, R. K., (1995), Scenario earthquakes for loss studies based on risk analysis,
Proceedings of the Fifth International Conference on Seismic Zonation, October 17-18-19, 1995,
Nice, France, Ouest Editions, Nantes, France, Vol. II, 1995, pages 1325-1333
McGuire, R. K. (1995). Probabilistic seismic hazard analysis and design earthquakes: closing the
loop, Bulletin of Seismic Society of America, Vol. 85, No.5, pp.1275-1284.
Medina, R., (2002), Seismic demands for non-deteriorating frame structures and their
dependence on ground motions, Ph.D. Thesis, Department of Civil and Environmental
Engineering, Stanford, CA.
Miranda, Eduardo, (2001), Estimation of inelastic deformation demands of SDOF systems,

Journal of Structural Engineering, 127, 9, Sept. 2001, pages 1005-1012.
Moehle, Jack P., et al. (1994), Performance assessment for a reinforced concrete frame
building, Proceedings of the NEHRP Conference and Workshop on Research on the Northridge,
California Earthquake of January 17, 1994, California Universities for Research in Earthquake
Engineering (CUREe), Richmond, California, Vol. III-A, 1998, pages III-140 -- III-156.
Pincheira, Jose A., Dotiwala, Faridun S. and D'Souza, Jonathan T (1999). Seismic analysis of
older reinforced concrete columns, Earthquake Spectra, 15, 2, May 1999, pages 245-272.
Rice, John A. (1995). Mathematical statistics and data analysis, Duxbury Press, Second
Edition.
SEAOC VISION 2000 (1995), A framework for performance-based design Volumes, I, II, and
III, Structural Engineer Association of California (SEAOC), Sacramento , CA.
Shome, N., Cornell, C.A., Bazzurro, P., and Carballo, J.E. (1998). Earthquakes, records and
nonlinear responses, Earthquake Spectra, Vol. 14, No. 3, pp. 469-500.
Shome, N., and Cornell, C.A. (1999). Probabilistic seismic demand analysis of non-linear
structures. Report No. RMS-35, Stanford University, Stanford, CA.
Shome, Nilesh, (1999), Probabilistic Seismic Demand Analysis of Nonlinear Structures, Ph.D.
Thesis, Department of Civil and Environmental Engineering, Stanford, CA.
Silva, Walter J. (1998). Pacific Engineering and Analysis Strong-Motion Catalog, Pacific
Engineering and Analysis, Pacific Engineering & Analysis, 311 Pomona Ave. El Cerrito, CA
94530, or, http://peer.berkeley.edu/smcat/.
Somerville, Paul; et al., (1998), Generation of ground motion time histories for performance-
based seismic engineering Call No: CF 10, Proceedings, Sixth U.S. National Conference on
Earthquake Engineering, Earthquake Engineering Research Inst., Oakland, California.
Stewart, Jonathan P.; Chiou, Shyh-Jeng; Bray, Jonathan D.; Graves, Robert W.; Somerville, Paul
G.; Abrahamson, Norman A., (2001), Ground motion evaluation procedures for performance-
based design, PEER-2001/09, Berkeley: Pacific Earthquake Engineering Research Center,
University of California, Sept. 2001, 229 pages.
USGS, (2002) USGS national seismic hazard maps, United States Geological Survey
http://geohazards.cr.usgs.gov/eq//.
Vamvatsikos, D. and Cornell, C. A. (2001). Incremental dynamic analysis, Earthquake

Engineering and Structural Dynamics. 31(3), pp 491-514.
Vamvatsikos, D. (2002). Seismic performance, capacity and reliability of structures as seen

through incremental dynamic analysis, Ph.D. Dissertation, Department of Civil and
Environmental Engineering, Stanford University, July, 2002.
Veletsos, A. S.; Newmark, N. M, (1960). Effect of inelastic behavior on the response of simple
systems to earthquake motions, Proceedings of the Second World Conference on Earthquake
Engineering, Japan, Tokyo, Vol. II, pp 895-912.
Weisberg, S., (1985), Applied linear regression John Wiley & Sons., Second Edition.
Wen, Y. K., (1995), Building reliability and code calibration, Earthquake Spectra, 11, 2, May
1995, pages 269-296.
Wen, Y.K., (2001), Reliability and performance-based design, Structural Safety, 23.
Yun, S.Y., Hamburger, R. O., Cornell, C. A., and Foutch, D. A. (2002). Seismic performance
evaluation for steel moment frames. ASCE Journal of Structural Engineering, Vol. 128, No.4, pp
534-545.
View publication stats

Direct Probabilistic Seismic Analysis Dissertation

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

Direct Probabilistic Seismic Analysis Dissertation

Uploaded by

Copyright:

Available Formats

DIRECT PROBABILISTIC SEISMIC ANAYSIS:

IMPLEMENTING NON-LINEAR DYNAMIC ASSESSMENTS

Approved for the University Committee on Graduate Studies:

Ground motion record selection is potentially significant in implementing a program of nonlinear

2 A Technical Framework for Probability-based (DCFD) Seismic Formats 8

3 Probability-based Design (DCFD) Seismic Formats 57

3.2.3 DCFD Format 61

5 Probability-Based Seismic Assessment: Non-linear Dynamic Displacement

Table 6-1- Modal periods for LA 20 M1+ model structure 183

given Sa characterized by D (x) and D|S a . 20

Figure 2-4 - Spectral acceleration corresponding to the demand value, d 25

drift variable D given Sa characterized by D (x) and D|S a , distribution

of capacity variable C characterized by C and C . 29

Figure 3-1 - Spectral acceleration corresponding to a displacement-based demand equal

Figure 3-2 - Spectral acceleration for a hazard level equal to P0 . 61

Figure 3-4 - Graphical presentation of median drift capacity. 65

Figure 3-5 - The hazard curve for S a (T = 1, = 2%) . 70

Figure 4-6 - The Spectral Acceleration Corresponding to a given acceptable probability

Figure 4-9 - Estimation of factored demand using two stripes 107

Figure 4-16 - Evaluation of factored demand by multiple-stripe analysis. 118

Figure 4-18- Individual IDA Curves 119

Figure 4-21 - Collapse Points marked on the IDA curves. 122

Figure 5-2 - Conditional probability of non-collapse for a given spectral acceleration

Figure 5-10 - Non-parametric estimation of drift hazard: a) collapse cases considered

Figure 5 -11 - Drift hazard calculated by numerical integration using the

Figure 5-15 - Double-stripe, semi-parametric factored demand estimation: Using the

Figure 6-2 - Residual-residual plot for rupture distance, as a second independent

Figure 6-22 - The original regression of the displacement-based response on PGA,

Figure 6-24 - The weighted regression of the displacement-based response on PGA,

1.1 General overview

This dissertation resembles a journey through probabilistic performance-based design equipped

1.2 This thesis in the context of probabilistic PBD

2.3 Sources of Uncertainty in Engineering Problems

2.4 Document Map

Source of Uncertainty: Limit State Probability

Sources of Uncertainty: Limit State Probbaility

2.5 Foundation Development

2.5.1 Structural Limit States

2.5.2 Structural Demand Variable (State Variable)

Demand or state variable is normally chosen as a displacement-based structural response

2.5.4 Limit State Probability PLS

2.5.5 General Solution Strategy

PLS = P[ D > C ] = P[ D > C | S

decomposing it with respect to spectral acceleration, shall be referred to as the IM-based

PLS = P[ D > C ] = P[ D > C | D = d ] P[ D = d ]

PLS = P[ D > C ] = P[ D > C | D = d ] P[ D = d | S

2.5.6 Ground Motion Intensity Measure

2.5.7 Randomness: The Only Source of Variability

2.5.7.1 Spectral Acceleration Hazard:

k 0 and k are parameters defining the shape of the hazard curve.

2.5.7.2 Median Relationship between Spectral Acceleration and

drift ratio and first mode spectral acceleration data pairs.

M axim um Inte r-s to ry D rift, D

2.5.7.3 Annual frequency of exceeding demand Drift Hazard

We assume (based on observation of data) that can be represented by a lognormal distribution,

fractional standard deviation D|S a ( x) of demand given spectral acceleration is assumed to be

briefly D (x) ) is approximated as a power-law function of the spectral acceleration level, x.

mean minus sigma

H Sa(x) Maximum inter-sto ry drift angle (Demand), D

characterized by D (x) and D|S a .

is the complementary cumulative distribution function (CCDF) at S a = x . It should be noted that

acceleration CCDF, G S a ( x) times the rate of seismicity :