IMPLEMENTATION OF EUROCODES

HANDBOOK 2

RELIABILITY BACKGROUNDS
Probability density E(x), R(x)
Load effect E

0,06

Resistance R
0,04

dx
0,02

0,00

40

60

x x+dx

100

120

140

Random variable X

Guide to the basis of structural reliability and risk engineering

related to Eurocodes, supplemented by practical examples

LEONARDO DA VINCI PILOT PROJECT CZ/02/B/F/PP-134007

DEVELOPMENT OF SKILLS FACILITATING IMPLEMENTATION OF

EUROCODES
Leonardo da Vinci Pilot Project CZ/02/B/F/PP-134007

DEVELOPMENT OF SKILLS FACILITATING IMPLEMENTATION OF

EUROCODES

HANDBOOK 2

RELIABILITY BACKGROUNDS

PARTNERSHIP:
The Klokner Institute of the Czech Technical University in Prague (KI CTU),
convener, Prof. Milan Holick, Czech Republic
The Czech Chamber of Certified Engineers and Technicians Engaged in Construction
(CKAIT), Prof. Alois Materna, Czech Republic,
The Institute of Steel Constructions of the University of Technology Aachen (RWTH),
Prof. Gerhard Sedlacek, Germany
The Spanish Organisation for Scientific Research (IET), Spain, Dr. Angel Arteaga
The University of Pisa (UOP), Prof. Luca Sanpaolesi, Italy
The Netherlands Organisation for Applied Scientific Research (TNO), Prof. Ton
Vrouwenvelder, The Netherlands
The Institute for Metal Constructions (IMK), Dr. Igor Kovse, Slovenia
The Building Research Establishment (BRE), Prof. Haig Gulvanessian, The United
Kingdom

Prague 10.2005
Leonardo da Vinci Pilot Project CZ/02/B/F/PP-134007

DEVELOPMENT OF SKILLS FACILITATING IMPLEMENTATION OF

EUROCODES
HANDBOOK 2

BASIS OF STRUCTURAL RELIABILITY AND RISK ENGINEERING

I

BASIC CONCEPTS OF STRUCTURAL RELIABILITY (Pages I-1 to I-10)

Summary
I-1
1
INTRODUCTION
I-1
1.1
Background materials
I-1
1.2
General principles
I-1
2
UNCERTAINTIES
I-2
2.1
Classification of uncertainties
I-2
2.2
Available tools to describe uncertainties
I-2
3
RELIABILITY
I-3
3.1
General
I-3
3.2
Definition of reliability
I-3
3.3
Probability of failure
I-4
3.4
Reliability index
I-5
3.5
Time variance of failure probability
I-5
4
DESIGN TARGETS
I-5
4.1
Indicative values of design working life
I-5
4.2
Target reliability level
I-6
5
DESIGN METHODS IN PRACTICE
I-6
5.1
General
I-6
5.2
Permissible stresses
I-7
5.3
Global safety factor
I-7
5.4
Partial factor method
I-7
5.5
Probabilistic methods
I-8
6
DESIGN ASSISTED BY TESTING
I-8
7
CONCLUDING REMARKS
I-9
REFERENCES
I-10
ATTACHMENTS
I-10
1. MATHCAD sheet Beta-Time
I-11

II

ELEMENTARY METHODS OF STRUCTURAL RELIABILITY I

(Pages II-1 to II-17)
Summary
1
INTRODUCTION
1.1
Background materials
1.2
General principles
2
FUNDAMENTAL CASES OF STRUCTURAL RELIBILITY
2.1
General
2.2
Fundamental cases of one random variable
2.3
Fundamental case of two random variables
3
EXACT SOLUTION FOR TWO RANDOM VARIABLES
4
CONCLUDING REMARKS
REFERENCES
ATTACHMENTS

II-1
II-1
II-1
II-1
II-1
II-1
II-3
II-6
II-8
II-10
II-11
II-11
4

1. MATHCAD sheet SteelRod.mcd

2. MATHCAD sheet DesVRod.mcd
3. MATHCAD sheet PrLnLn.mcd
III

IV

II-12
II-14
II-16

RELIABILITY DIFFERENTIATION (Pages III-1 to III-12)

Summary
1
INTRODUCTION
1.1
Background documents
1.2
General principles
2
BASIC RELIABILITY ELEMENTS
3
DESIGN WORKING LIFE AND RELIABILITY
4
VARIATION OF FAILURE PROBABILITY WITH TIME
5
PARTIAL FACTOR OF A MATERIAL PROPERTY
6
PARTIAL FACTORS OF SELF-WEIGHT
7
CLIMATIC ACTIONS AND IMPOSED LOADS
8
EXAMPLES
9
CONCLUDING REMARKS
REFERENCES
ATTACHMENTS
1. MATHCAD sheet GammaRG.mcd
2. MATHCAD sheet PSI0.mcd
3. MATHCAD sheet PSI12.mcd

III-1
III-1
III-1
III-1
III-2
III-3
III-4
III-5
III-6
III-7
III-9
III-10
III-10
III-11
III-12
III-13
III-15

DESIGN ASSISTED BY TESTING (PAGES IV-1 TO IV-24)

Summary
1
INTRODUCTION
1.1
Background documents
1.2
General principles
1.3
Preliminary statistical concepts
2
STATISTICAL DETERMINATION OF A SINGLE PROPERTY
2.1
General principles
2.2
Assessment via the characteristic value
2.3
Direct assessment of the design value
2.4
Approximation of the factors kn and kd,n
3
STATISTICAL DETERMINATION OF RESISTANCE MODELS
REFERENCES
APPENDIX A - THE DERIVATION OF THE EQUATION (50)

IV-1
IV-1
IV-1
IV-1
IV-2
IV-4
IV-4
IV-6
IV-8
IV-9
IV-9
IV-14
IV-15

ATTACHMENTS
1. EXCEL worksheet from the workbook dast.xls
2. Source file dast.c to the program dast.exe
3. Input file dast.i1 for the program dast.exe
4. Output file dast.o1 produced by dast.exe from dast.o1
5. Input file dast.i2 for the program dast.exe
6. Output file dast.o2 produced by dast.exe from dast.o2
7. EXCEL chart showing the data from dast.o2

IV-16
IV-17
IV-18
IV-21
IV-22
IV-23
IV-24
IV-24

ASSESSMENT OF EXISTING STRUCTURES (PAGES V-1 TO V-15)

Summary

V-1

INTRODUCTION
1.1
Background documents
1.2
General principles
2
GENERAL FRAMEWORK OF ASSESSMENT
2.1
Reasons for assessment
2.2
Common rules
2.3
General procedure
3
INVESTIGATION
3.1
Purpose
3.2
Statement
4
BASIC VARIABLES
4.1
General
4.2
Characteristic values
5
EVALUATION OF INSPECTION RESULTS
5.1
Updating in general
5.2
Updating of probability distribution
5.3
Updating of failure probability
5.4
Updating of characteristic and design values
6
STRUCTURAL ANALYSIS
7
VERIFICATION
8
ASSESSMENT IN THE CASE OF DAMAGE
9
FINAL REPORT AND DECISION
10
CONCLUDING REMARKS
REFERENCES
APPENDIX A - GENERAL FLOW OF ASSESSMENT OF EXISTING
STRUCTURES
APPENDIX B - PROBABILITY UPDATING
APPENDIX C - BAYESIAN METHOD FOR FRACTILE ESTIMATION
ATTACHMENTS
1. MATHCAD sheet Update.mcd
2. MATHCAD sheet BayesFract.mcd
VI

PRINCIPLES OF RISK ASSESSMENT (PAGES VI-1 TO VI-12)

Summary
1
INTRODUCTION
1.1
Background documents
1.2
General principles
2
HAZARD IDENTIFICATION
3
DEFINITION AND MODELLING OF RELEVANT SCENARIOS
4
ESTIMATION OF PROBABILITIES
5
ESTIMATION OF CONSEQUENCES
6
ESTIMATION OF RISK
7
LOGIC TREES
8
BAYESIAN NETWORK
9
DECISION-MAKING
10
THE IMPLIED COST OF AVERTING A FATALITY
11
CONCLUDING REMARKS
REFERENCES
NOTATION

V-1
V-1
V-1
V-2
V-2
V-2
V-2
V-3
V-3
V-3
V-4
V-4
V-4
V-5
V-5
V-5
V-6
V-6
V-7
V-7
V-8
V-9
V-9
V-10
V-11
V-12
V-14
V-16
V-17
V-18
VI-1
VI-1
VI-1
VI-1
VI-2
VI-3
VI-4
VI-4
VI-5
VI-5
VI-8
VI-10
VI-10
VI-11
VI-11
VI-12

ANNEX A - BASIC STATISTICAL CONCEPTS AND TECHNIQUES

(PAGES A-1 TO A-42)
Summary
1
INTRODUCTION
1.1
Background materials
1.2
General principles
2
POPULATION AND SAMPLES
2.1
General
2.2
Sample characteristics
2.3
Distribution function
2.4
Population parameters
3

SELECTED MODELS OF RANDOM VARIABLES

3.1
Normal distribution
3.2
Lognormal distribution
3.3
Gamma distribution
3.4
Beta distribution
3.5
Gumbel and other distributions of extreme values
3.6
Function of random variables
4
ESTIMATION OF FRACTILES
4.1
Fractile of a theoretical model
4.2
Coverage method of fractile estimation
4.3
Prediction method of fractile estimation
4.4
Coefficients of the coverage and prediction methods
4.5
Bayes' method of fractile estimation
4.6
Estimation of fractiles according to Eurocodes
REFERENCES
APPENDIX 1 - PROBABILISTIC MODELS OF BASIC VARIABLES
APPENDIX 2 - STATISTICAL PARAMETERS OF FUNCTIONS OF
RANDOM VARIABLES
APPENDIX 3 - FRACTILE OF A RANDOM VARIABLE
ATTACHMENTS
1. MATHCAD sheet DistFract.mcd
2. MATHCAD sheet SampFract.mcd
3. MATHCAD sheet Mod_est.mcd
ANNEX B - ELEMENTARY METHODS OF STRUCTURAL RELIABILITY II
(PAGES B-1 TO B-42)
Summary
1
INTRODUCTION
1.1
Background materials
1.2
General principles
2
DESIGN POINT
3
PARTIAL FACTORS
3.1
Material properties
3.2
Permanent load
3.3
Variable load
4
THE GENERAL CASE OF RELIABILITY ANALYSIS
4.1
General

A-1
A-1
A-1
A-1
A-1
A-1
A-2
A-3
A-4
A-6
A-6
A-6
A-9
A-11
A-13
A-17
A-17
A-17
A-21
A-22
A-23
A-26
A-28
A-30
A-31
A-32
A-33
A-34
A-35
A-37
A-39

B-1
B-1
B-1
B-1
B-1
B-4
B-4
B-6
B-7
B-9
B-9

4.2
Basic variables
4.3
Tail sensitivity problem
5
AN EXAMPLE OF REINFORCED CONCRETE SLAB
5.1
General
5.2
A reinforced concrete slab
5.3
Design of a slab
5.4
Reliability consideration
6
ASSESMENT OF THE FAILURE PROBABILITY IN
GENERAL CASE
6.1
General
6.2
First and Second Order Reliability Methods (FORM and SORM)
7
SYSTEM RELIABILITY
7.1
General
7.2
Parallel systems
7.3
Series systems
8
CONCLUDING REMARKS
REFERENCES

B-15
B-15
B-15
B-19
B-19
B-20
B-21
B-21
B-21

ATTACHMENTS
1. MATHCAD sheet GammaRGQ.mcd
2. MATHCAD sheet Prindex.mcd
3. MATHCAD sheet RCBeam.mcd
4.M A TH CA D sheetRelRCB.m cd
5. EXCEL sheet RCBeam.xls
6. MATHEMATICA notebook Fit_distribution.nb
7. MATHEMATICA notebook FORM.nb
8. MATHEMATICA package Level2.m
9. MATHCAD sheet FORM2.mcd
10. MATHCAD sheet FORM7.mcd
11. EXCEL sheet FORM7.xls
12. MATLAB package FORM7.m
13. MATLAB function Lndens (x,mu,sigma,sk)
14. MATLAB function Lndist (x,mu,sigma,sk)
15. MATLAB function Ndens (x, mu,sigma)
16. MATLAB function "Ndinv (p)"
17. MATLAB function "Ndist (x, mu,sigma)"

B-22
B-24
B-26
B-27
B-29
B-32
B-34
B-40
B-41
B-44
B-45
B-47
B-48
B-50
B-52
B-53
B-54
B-55

ANNEX C - CALIBRATION PROCEDURE (PAGES C-1 TO C-42)

Summary
1
INTRODUCTION
1.1
Background materials
1.2
General principles
2
FUNDAMENTAL LOAD COMBINATIONS
3
GENERIC STRUCTURAL MEMBER
4
PRINCIPLES OF RELIABILITY ANALYSIS
4.1
Limit state function
4.2
Probabilistic models of basic variables
4.3
Reliability measures
4.4
Sensitivity factors

B-10
B-11
B-11
B-11
B-12
B-12
B-14

C-1
C-1
C-1
C-1
C-2
C-4
C-5
C-5
C-6
C-7
C-8

RESULTS FOR THE GENERIC CROSS-SECTION

5.1

One variable action

5.2
Two variable actions
6
CONCLUDING REMARKS
REFERENCES
APPENDIX A - DIRECT COMPARISON OF LOAD EFFECTS
APPENDIX B - EFFECT OF THE RESISTANCE VARIABILITY
APPENDIX C NOTATION
ATTACHMENTS
1. MATLAB function generic(gR,wr,k)
2. MATLAB function "Action3(mr,sr,skr,Rd,k)
3. MATLAB function Lnpf(mr,sr,skr,me,se,ske)
4. MATLAB function Lndens(x,mu,sigma,sk)
5. MATLAB function Lndist(x,mu,sigma,sk)
6. MATLAB function Ndinv(p)
7. MATHCAD sheet Generic.mcd
8. MATHCAD sheet Load effect.mcd

C-8
C-8

C-10
C-13
C-13
C-15
C-19
C-20
C-21
C-23
C-24
C-29
C-31
C-32
C-33
C-34
C-39

FOREWORD
The Leonardo da Vinci Pilot Project CZ/02/B/F/PP-134007, Development of Skills
Facilitating Implementation of Structural Eurocodes addresses the urgent need to implement
the new system of European documents related to design of construction works and products.
These documents, called Eurocodes, are systematically based on recently developed Council
Directive 89/106/EEC The Construction Products Directive and its Interpretative Documents
ID1 and ID2. Implementation of Eurocodes in each Member State is a demanding task as each
country has its own long-term tradition in design and construction.
The project should enable an effective implementation and application of the new
methods for designing and verification of buildings and civil engineering works in all the
partner countries (CZ, DE, ES, IT, NL, SI, UK) and in other Member States. The need to
explain and effectively use the latest principles specified in European standards is apparent
from various enterprises, undertakings and public national authorities involved in construction
industry and also from universities and colleges. Training materials, manuals and software
programmes for education are urgently required.
The submitted Handbook 2 is one of 5 upcoming handbooks intended to provide
required manuals and software products for training, education and effective implementation
of Eurocodes:
Handbook 1: Basis of Structural Design
Handbook 2: Reliability Backgrounds
Handbook 3: Load Effect for Buildings
Handbook 4: Load Effect for Bridges
Handbook 5: Design of Buildings for Fire Situation
It is expected that the Handbooks will address the following intents in further
harmonisation of European construction industry:
- reliability improvement and unification of the process of design;
- development of the single market for products and for construction services;
- new opportunities for the trained primary target groups in the labour market.
The Handbook 2 is focused on the basis of structural reliability and risk engineering
related to Eurocodes. The following topics are treated in particular:
- basic concepts of structural reliability;
- elementary methods of the reliability theory;
- reliability differentiation and design working life;
- design assisted by testing;
- assessment of existing structures;
- basis of risk assessment.
Annex A to the Handbook 2 provides a review of "Basic Statistical Concepts and
Techniques" frequently used in the text. Annex B provides an extension of elementary
methods of structural reliability and annex C describes calibration procedures that may be
used for specification of reliability elements. The Handbook 2 is written in a user-friendly way
employing only basic mathematical tools. Attached software products accompanying a number
of examples enable applications of general rules in practice.
A wide range of potential users of the Handbooks and other training materials includes
practising engineers, designers, technicians, experts of public authorities, young people - high
school and university students. The target groups come from all territorial regions of the
partner countries. However, the dissemination of the project results is foreseen to be spread
into all Member States of CEN and other interested countries.
Prague 05/2005

10

Chapter I - Basic concepts of structural reliability

CHAPTER I - BASIC CONCEPTS OF STRUCTURAL RELIABILITY

Milan Holick1 and Ton Vrouwenvelder2
1

Klokner Institute, Czech Technical University in Prague, Czech Republic

2
Delft University of Technology, TNO BOUW, The Netherlands

Summary
Uncertainties affecting structural performance can never be entirely eliminated and
must be taken into account when designing any construction work. Various design methods
and operational techniques for verification of structural reliability have been developed and
worldwide accepted in the past. The most advanced operational method of partial factors is
based on probabilistic concepts of structural reliability and available experience. General
principles of structural reliability can be used to specify and further calibrate partial factors
and other reliability elements. Moreover, developed calculation procedures and convenient
software products can be used directly for verification of structural reliability using
probabilistic concepts and available experimental data.
1

INTRODUCTION

1.1

Background materials
Basic concepts of structural reliability are codified in a number of national standards,
in the new European document EN 1990 [1] and the International Standard ISO 2394 [2].
Additional information may be found in the background document developed by JCSS [3]
and in recently published handbook to EN 1990 [4]. Guidance for application of probabilistic
methods of structural reliability may be found in working materials provided by JCSS [5] and
in relevant literature listed in [4 and [5]. Elementary methods of the theory of reliability are
described in Chapter II and III in this Handbook 2.
1.2

General principles
General principles of structural reliability are described in both the international
documents EN 1990 [1] and ISO 2394 [2]. Basic requirements on structures are specified in
Section 2 of EN 1990 [1]: a structure shall be designed and executed in such a way that it will,
during its intended life, with appropriate degrees of reliability and in an economic way
-

sustain all actions and influences likely to occur during execution and use;
remain fit for the use for which it is required.

It should be noted that two aspects are explicitly mentioned: reliability and economy
(see also Handbook 1). However, in this Handbook 2 we shall be primarily concern with
reliability of structures, which include
-

structural resistance;
serviceability;
durability.

I-1

Chapter I - Basic concepts of structural reliability

Additional requirements may concern fire safety of structures (see Handbook 5) or
other accidental design situations. In particular it is required by EN 1990 [1] that in the case
of fire, the structural resistance shall be adequate for the required period of time.
To verify all the aspects of structural reliability implied by the above-mentioned basic
requirements, an appropriate design lifetime, design situations and limit states should be
considered (as described in Handbook 1). Note that the basic lifetime for a common building
is 50 years and that, in general, four design situations are identified: permanent, transient,
accidental and seismic. Two types of limit states are normally verified: ultimate limit states
and serviceability limit states. Detail guidance is provided in Handbook 1.
2

UNCERTAINTIES

2.1

Classification of uncertainties
It is well recognised that construction works are complicated technical systems
suffering from a number of significant uncertainties in all stages of execution and use.
Depending on the nature of a structure, environmental conditions and applied actions, various
types of uncertainties become more significant than the others. The following types of
uncertainties can be identified in general:
- natural randomness of actions, material properties and geometric data;
- statistical uncertainties due to a limited size of available data;
- uncertainties of the resistance and load effect models due to simplifications of actual
conditions;
- vagueness due to inaccurate definitions of performance requirements;
- gross errors in design, during execution and use;
- lack of knowledge concerning behaviour of new materials and actions in actual
conditions.
The order of the listed uncertainties corresponds approximately to the decreasing level
of current knowledge and available theoretical tools for their description and consideration in
design (see following sections). It should be emphasized that most of the above listed
uncertainties (randomness, statistical and model uncertainties) can never be eliminated
absolutely and must be taken into account when designing any construction work.
2.2

Available tools to describe uncertainties

Natural randomness and statistical uncertainties may be relatively well described by
available methods provided by the theory of probability and mathematical statistics. In fact
the EN 1990 [1] gives some guidance on available techniques. However, lack of credible
experimental data (e.g. for new materials, some actions including environmental influences
and also for some geometrical properties) causes significant problems. In some cases the
available data are inhomogeneous, obtained under different conditions (e.g. for material
resistance, imposed loads, environmental influences, for inner dimensions of reinforced
concrete cross-sections). Then it may be difficult, if not impossible, to analyse and use them
in design.
The uncertainties of computational models may be to a certain extent assessed on the
basis of theoretical and experimental research. EN 1990 [1] and materials of JCSS [5] provide
some guidance. The vaguenesses caused by inaccurate definitions (in particular of
serviceability and other performance requirements) may be partially described by the means
of the theory of fuzzy sets. However, these methods have a little practical significance, as

I-2

Chapter I - Basic concepts of structural reliability

suitable experimental data are rarely available. The knowledge of the behaviour of new
materials and structures may be gradually increased through theoretical analyses verified by
experimental research.
The lack of available theoretical tools is obvious in the case of gross errors and lack of
knowledge, which are nevertheless often the decisive causes of structural failures. To limit
gross errors due to human activity, a quality management system including the methods of
statistical inspection and control may be effectively applied.
Various design methods and operational techniques, which take these uncertainties
into account, have been developed and worldwide used. The theory of structural reliability
provides background concept techniques and theoretical bases for description and analysis of
the above-mentioned uncertainties concerning structural reliability.
3

RELIABILITY

3.1

General
The term "reliability" is often used very vaguely and deserves some clarification.
Often the concept of reliability is conceived in an absolute (black and white) way the
structure either is or isnt reliable. In accordance with this approach the positive statement is
understood in the sense that a failure of the structure will never occur. This interpretation is
unfortunately an oversimplification. Although it may be unpleasant and for many people
perhaps unacceptable, the hypothetical area of absolute reliability for most structures (apart
from exceptional cases) simply does not exist. Generally speaking, any structure may fail
(although with a small or negligible probability) even when it is declared as reliable.
The interpretation of the complementary (negative) statement is usually understood
more correctly: failures are accepted as a part of the real world and the probability or
frequency of their occurrence is then discussed. In fact in the design it is necessary to admit a
certain small probability that a failure may occur within the intended life of the structure.
Otherwise designing of civil structures would not be possible at all. What is then the correct
interpretation of the keyword reliability and what sense does the generally used statement
the structure is reliable or safe have?
3.2

Definition of reliability
A number of definitions of the term reliability are used in literature and in national
and international documents. ISO 2394 [2] provides a definition of reliability, which is similar
to the approach of national standards used in some European countries: reliability is the
ability of a structure to comply with given requirements under specified conditions during the
intended life, for which it was designed. In quantitative sense reliability may be defined as the
complement of the probability of failure.
Note that the above definition of reliability includes four important elements:
- given (performance) requirements definition of the structural failure,
- time period assessment of the required service-life T,
- reliability level assessment of the probability of failure Pf,
- conditions of use limiting input uncertainties.
An accurate determination of performance requirements and thus an accurate
specification of the term failure are of uttermost importance. In many cases, when considering
the requirements for stability and collapse of a structure, the specification of the failure is not

I-3

Chapter I - Basic concepts of structural reliability

very complicated. In many other cases, in particular when dealing with various requirements
of occupants comfort, appearance and characteristics of the environment, the appropriate
definitions of failure are dependent on several vaguenesses and inaccuracies. The
transformation of these occupants' requirements into appropriate technical quantities and
precise criteria is very hard and often leads to considerably different conditions.
In the following the term failure is being used in a very general sense denoting simply
any undesirable state of a structure (e.g. collapse or excessive deformation), which is
unambiguously given by structural conditions.
The same definition as in ISO 2394 is provided in Eurocode EN 1990 [1] including
note that the reliability covers the load-bearing capacity, serviceability as well as the
durability of a structure. Fundamental requirements include the statement (as already
mentioned) that a structure shall be designed and executed in such a way that it will, during
its intended life with appropriate degrees of reliability and in an economic way sustain all
actions and influences likely to occur during execution and use, and remain fit for the use for
which it is required. Generally a different level of reliability for load-bearing capacity and
for serviceability may be accepted for a structure or its parts. In the documents [1] and [2] the
probability of failure Pf (and reliability index ) are indicated with regard to failure
consequences (see Handbook 1).
3.3

Probability of failure
The most important term used above (and in the theory of structural reliability) is
evidently the probability of failure Pf. In order to defined Pf properly it is assumed that
structural behaviour may be described by a set of basic variables X = [X1, X2, ... , Xn]
characterizing actions, mechanical properties, geometrical data and model uncertainties.
Furthermore it is assumed that the limit state (ultimate, serviceability, durability or fatigue) of
a structure is defined by the limit state function (or the performance function), usually written
in an implicit form as
Z(X) = 0

(1)

The limit state function Z(X) should be defined in such a way that for a favourable (safe) state
of a structure the function is positive, Z(X) 0, and for a unfavourable state (failure) of the
structure the limit state function is negative, Z(X) < 0 (a more detailed explanation is given in
the following Chapters of this Handbook 2).
For most limit states (ultimate, serviceability, durability and fatigue) the probability of
failure can be expressed as
Pf = P{Z(X) < 0}

(2)

The failure probability Pf can be assessed if basic variables X = [X1, X2, ... , Xn] are
described by appropriate probabilistic (numerical or analytical) models. Assuming that the
basic variables X = [X1, X2, ... , Xn] are described by time independent joint probability density
function X(x) then the probability Pf can be determined using the integral
Pf =

X ( x )dx

(3)

Z( X ) < 0

More complicated procedures need to be used when some of the basic variables are
time-dependent. Some details concerning theoretical models for time-dependent quantities
(mainly actions) and their use for the structural reliability analysis are given in other Chapters
of this Handbook 2. However, in many cases the problem may be transformed to a timeindependent one, for example by considering in equation (2) or (3) a minimum of the function
Z(X) over the reference period T.

I-4

Chapter I - Basic concepts of structural reliability

Note that a number of different methods [2] and software products [7, 8, 10] are
available to calculate failure probability Pf defined by equation (2) or (3).
3.4

Reliability index
An equivalent term to the failure probability is the reliability index , formally defined
as a negative value of a standardized normal variable corresponding to the probability of
failure Pf. Thus, the following relationship may be considered as a definition

= U1 ( Pf )

(4)

Here u1 ( p f ) denotes the inverse standardised normal distribution function. At present the
reliability index defined by equation (4) is a commonly used measure of structural
reliability in several international documents [1], [2], [5].
It should be emphasized that the failure probability Pf and the reliability index
represent fully equivalent reliability measures with one to one mutual correspondence given
by equation (4) and numerically illustrated in Table 1.

Pf

Table 1. Relationship between the failure probability Pf and the reliability index .
101
102
103
104
105
106
107
1,3
2,3
3,1
3,7
4,2
4,7
5,2

In EN 1990 [1] and ISO 2394 [2] the basic recommendation concerning a required
reliability level is often formulated in terms of the reliability index related to a certain
design working life.
3.5

Time variance of failure probability

When the vector of basic variables X = X1, X2, ... , Xm is time variant, then the failure
probability p is also time variant and should be always related to a certain reference period T,
which may be generally different from the design working life Td. Considering a structure of a
given reliability level, the design failure probability pd = pn related to a general reference
period Tn = n T1 can be derived from the alternative probability pa = p1 corresponding to Ta =
T1 (to simplify notation note that the previously used subscript "d" corresponds now to "n"
and subscript "a" to "1"). Detail description of this transformation is provided in Chapter III.
4

DESIGN TARGETS

4.1

Indicative values of design working life

Design working life Td is an assumed period of time for which a structure or part of it
is to be used for its intended purpose with anticipated maintenance but without major repair
being necessary. In the recent documents CEN [1] and ISO [2] indicative values of Td are
provided for five categories of structures as shown in Table 2.
A more detailed specification of structural categories and design working lives may be
found in some national standards. In general the design working lives may be greater (in some
cases by 100 %) than those given in Table 2. For example the design working life for
temporary structures may be 15 years, for agricultural structures 50 years, for apartment and
office buildings 100 years, and for railways structures, dams, tunnels and other underground
engineering works 120 years or more.

I-5

Chapter I - Basic concepts of structural reliability

Table 2. Indicative design working life Td.
Category Design working life Td (years)
1
10
2
10 to 25
3
15 to 30
4
50
5
100 and more

Examples
Temporary structures
Replaceable structural parts, bearings, girders
Agricultural and similar structures
Building structures and common structures
Monumental building or civil structures, bridges

4.2

Target reliability level

Design failure probabilities pd are usually indicated in relation to the expected social
and economical consequences. EN 1990 [1] provides the classification of target reliability
levels into three classes of consequences (high, normal, low) and indicates the adequate
reliability indexes for two reference periods T (1 year and 50 years). No explicit link to the
design working life Td. is given Similar -values may be found also in some national
standards and international standards ISO [2]. Detail description of the target is given in
Chapter III in this Handbook 2.
It should be underlined that the couple of values (a and d) recommended in [1] for
each reliability class (for 1 year and 50 years) correspond to the same reliability level.
Practical application of these values depend on the reference period Ta considered in the
verification, which may be connected with available information concerning time variant
vector of basic variables X = X1, X2, ..., Xn. For example, if the reliability class 2 and 50 years
design working period is considered, then the reliability index d = 3,8 should be used in the
verification of structural reliability. The same reliability level corresponding to the class 2 is
achieved when the time period Ta = 1 years and a = 4,7 are considered. Thus, various
reference periods Ta, in general different from the design working life Td, may be used for
achieving a certain reliability level.
5

DESIGN METHODS IN PRACTICE

5.1

General
During their historical development the design methods have been closely linked to
the available empirical, experimental as well as theoretical knowledge of mechanics and the
theory of probability. The development of various empirical methods for structural design
gradually crystallized in the twentieth century in three generally used methods, which are, in
various modifications, still applied in standards for structural design until today: the
permissible stresses method, the global factor and partial factor methods. All these methods
are often discussed and sometimes reviewed or updated.
The following short review of historical development illustrates general formats of
above mentioned design methods, indicate relevant measures that are applied to take into
account various uncertainties of basic variables and to control resulting structural reliability.
In addition a short description of probabilistic methods of structural reliability and their role
in further development of design procedures is provided. Detailed description of probabilistic
methods of structural reliability is given in Chapter II, Chapter III and in Annex B of this
Handbook 2.

I-6

Chapter I - Basic concepts of structural reliability

5.2

Permissible stresses
The first of the worldwide-accepted design methods for structural design is the method
of permissible stresses that is based on linear elasticity theory. The basic design condition of
this method can be written in the form

max < per, where per = crit / k

(5)

The coefficient k (greater than 1) is the only explicit measure supposed to take into account all
types of uncertainties (some implicit measures may be hidden). Moreover, only a local effect
(a stress) max is compared with the permissible stress per and, therefore, a local (elastic)
behaviour of a structure is used to guarantee its reliability. No proper way is provided for
treating geometric non-linearity, stress distribution and ductility of structural materials and
members. For that reasons the permissible stress method leads usually to conservative and
uneconomical design.
However, the main insufficiency of the permissible stress method is lack of possibility
to consider uncertainties of individual basic variables and computational models used to
assess load effects and structural resistances. Consequently, reliability level of structures
exposed to different actions and made of different material may be not only conservative
(uneconomical) but also considerably different.
5.3

Global safety factor

The second widespread method of structural design is the method of global safety
factor. Essentially it is based on a condition relating the standard or nominal values of the
structural resistance R and load effect E. It may be written as
s = R / E > s0

(6)

Thus the calculated safety factor s must be greater than its specified value s0 (for example s0 =
1,9 is commonly required for bending resistance of reinforced concrete members). The global
safety factor method attempts to take into account realistic assumptions concerning structural
behaviour of members and their cross-sections, geometric non-linearity, stress distribution
and ductility; in particular through the resulting quantities of structural resistance R and action
effect E.
However, as in the case of the permissible stresses method the main insufficiency of
this method remains a lack of possibility to consider the uncertainties of particular basic
quantities and theoretical models. The probability of failure can, again, be controlled by one
explicit quantity only, by the global safety factor s. Obviously harmonisation of reliability
degree of different structural members made of different materials is limited.
5.4

Partial factor method

At present, the most advanced operational method of structural design [1, 2] accepts
the partial factor format (sometimes incorrectly called the limit states method) usually applied
in conjunction with the concept of limit states (ultimate, serviceability or fatigue). This
method can be generally characterised by the inequality
Ed (Fd, fd, ad, d) < Rd (Fd, fd, ad, d)

(7)

where the design values of action effect Ed and structural resistance Rd are assessed
considering the design values of basic variables describing the actions Fd = F Fk, material
properties fd = fk /m, dimensions ad + a and model uncertainties d. The design values of
these quantities are determined (taking into account various uncertainties) using their
characteristic values (Fk, fk, ak, k), partial factors , reduction factors and other measures of
reliability [1, 2, 3, 4], Thus the whole system of partial factors and other reliability elements

I-7

Chapter I - Basic concepts of structural reliability

may be used to control the level of structural reliability. Detailed description of the partial
factor methods used in Eurocodes method is provided in Handbook 1.
Compared with previous design methods the partial factor format obviously offers the
greatest possibility to harmonise reliability of various types of structures made of different
materials. Note, however, that in any of the above listed design methods the failure
probability is not applied directly. Consequently, the failure probability of different structures
made of different materials may still considerably vary even though sophisticated calibration
procedures were applied. Further desired calibrations of reliability elements on probabilistic
bases are needed; it can be done using the guidance provided in the International standard ISO
2394 [2] and European document EN 1990 [1].
5.5

Probabilistic methods
The probabilistic design methods introduced in the International Standard [2] are
based on a requirement that during the service life of a structure T the probability of failure Pf
does not exceed the design value pd or the reliability index is greater than its design value d
Pf Pd or > d

(8)

In EN 1990 [1] the basic recommended reliability index for ultimate limit states d =
3,8 corresponds to the design failure probability Pd = 7,2 10-5, for serviceability limit states
d = 1,5 corresponds to Pd = 6,7 10-2. These values are related to the design working life of
50 years that is considered for building structures and common structures. In general greater
- values should be used when a short reference period (one or five years) will be used for
verification of structural reliability.
It should be mentioned that probabilistic methods are not yet commonly used in
design praxis. However, the developed calculation procedures and software products (for
example [7, 8] and [10]) already enable the direct verification of structural reliability using
probabilistic concepts and available experimental data. Recently developed software product
CodeCal [10] is primarily intended for calibration of codes based on the partial factor method.
In Chapter II of this Handbook 2 numerical examples will be presented to illustrate the
methods discussed above.
6

DESIGN ASSISTED BY TESTING

In some cases there is a need to base the design on a combination of tests and
calculations, for instance if no adequate calculation model is available. The tests may vary
from wind tunnel tests to prototype testing of new structural materials, elements or
assemblies. Tests may also be carried out during or after execution to confirm the design
assumptions. The extreme example is a proof load. For design by testing the following types
of tests can be distinguished:
a) tests to establish directly the resistance for given loading conditions
b) tests to obtain specific material properties
c) tests to reduce model uncertainties in loads, load effects or resistance models
Test should be set up and evaluated in such a way that the usual required level of
reliability is achieved. The derivation of a characteristic or design value should take into
account the scatter of test data, statistical uncertainty associated with the number of tests and
prior statistical knowledge. If the response of the structure or structural member or the

I-8

Chapter I - Basic concepts of structural reliability

resistance of the material depends on influences not sufficiently covered by the tests such as
duration or scale effects, corrections should be made.
When evaluating test results, the behaviour of test specimens and failure modes should
be compared with theoretical predictions. When significant deviations from a prediction
occur, an explanation should be sought: this might involve additional testing, perhaps under
different conditions, or modification of the theoretical model.
The evaluation of test results should be based on statistical methods. In Eurocodes both
Bayesian and classical frequentistic methods are used. When frequentistic methods are used a
confidence level has to be chosen. The level of the confidence interval may influence the final
value. On the average, a confidence level of 0.75 leads to the same result as the Bayesian
methods. For this reason 0.75 is chosen in most cases, however also other numbers are used (eg.
0.85 in EN 1995).
In Basis of Design the preference is given for Bayesian methods, which generally is
believed to be more consistent with modern reliability theory then frequentistic methods.
Moreover, Bayesian methods provide a formal framework for the use of prior knowledge, which
is essential especially in the case of small samples and quality control methods. In most
Eurocodes special rules for small samples are presented, but in general without formal
background.
Rules for execution and evaluation of design by testing are presented in Annex D of
EN 1990 Basis of Design. For detailed background and worked examples, the reader is
referred to Annex A of this Handbook 2.
7

CONCLUDING REMARKS

The basic concepts of the probabilistic theory of reliability are characterized by two
equivalent terms, the probability of failure Pf and the reliability index . Although they
provide limited information on the actual frequency of failures, they remain the most
important and commonly used measures of structural reliability. Using these measures the
theory of structural reliability may be effectively applied for further harmonisation of
reliability elements and for extensions of the general methodology for new, innovative
structures and materials.
Historical review of the design methods worldwide accepted for verification of
structural members indicates different approaches to considering uncertainties of basic
variables and computational models. The permissible stresses method proves to be rather
conservative (and uneconomical). The global safety factor and partial factor methods lead to
similar results. Obviously, the partial factor method, accepted in the recent EN documents,
represents the most advanced design format leading to a suitable reliability level that is
relatively close to the level recommended in EN 1990 ( = 3,8). The most important
advantage of the partial factor method is the possibility to take into account uncertainty of
individual basic variables by adjusting (calibrating) the relevant partial factors and other
reliability elements.
Various reliability measures (characteristic values, partial and reduction factors) in the
new structural design codes using the partial factor format are partly based on probabilistic
methods of structural reliability, partly (to a great extent) on past empirical experiences.
Obviously the past experience depends on local conditions concerning climatic actions and
traditionally used construction materials. These aspects may be considerably different in
different countries. That is why a number of reliability elements and parameters in the present
suite of European standards are open for national choice.

I-9

Chapter I - Basic concepts of structural reliability

It appears that further harmonisation of current design methods will be based on
calibration procedures, optimisation methods and other rational approaches including the use
of methods of the theory of probability, mathematical statistics and the theory of reliability.
The probabilistic methods of structural reliability provide the most important tool for gradual
improvement and harmonisation of the partial factor method for various structures from
different materials. Moreover, developed software products enable direct application of
reliability methods for verification of structures using probabilistic concepts and available
data.
Design assisted by testing may be used when there is a need to base the design on a
combination of tests and calculations. The tests may vary from wind tunnel tests to prototype
testing of new structural materials, elements or assemblies. Tests may also be carried out
during or after execution to confirm the design assumptions. An operational technique
recommended in [1] is described in Chapter IV of this Handbook 2.
REFERENCES
[1] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[2] ISO 2394 General principles on reliability for structures, ISO 1998.
[3] JCSS: Background documentation, Part 1 of EC 1 Basis of design, 1996.
[4] Gulvanessian, H. Calgaro, J.-A. Holick, M.: Designer's Guide to EN 1990, Eurocode:
Basis of Structural Design; Thomas Telford, London, 2002, ISBN: 07277 3011 8, 192 pp.
[5] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/, 2001.
[6] EN 1991-1-1 Eurocode 1 Actions on structures. Part 1-1 General actions. Densities, selfweight, imposed loads for buildings, CEN 2002
[7] VaP, Variable Processor, version 1.6, ETH Zurich, 1997.
[8] COMREL, version 7.10, Reliability Consulting Programs, RCP MUNICH, 1999.
[9] ISO 13822. Basis for design of structures - Assessment of existing structures, ISO 2001.
[10] CodeCal, Excel sheet developed by JCSS, http://www.jcss.ethz.ch/.
ATTACHMENTS
1. MATHCAD sheet Beta-Time.mcd
Mathcad sheet "Beta-Time" is intended for transformation of probability and
reliability index Beta" for different reference periods.

I - 10

Chapter I - Basic concepts of structural reliability

Attachment 1 - MATHCAD sheet Beta-Time.mcd

Mathcad sheet Beta -Time

Mathcad sheet "Beta-Time" is intended for transformation of probability and reliability index
Beta" for different reference periods"
1 Input data

n := 10, 20.. 100

1 := 0 , 1 .. 6

2 Probability
pn ( 1 , n ) := 1 ( 1 p1( 1) )

p1( 1) := pnorm( 1 , 0 , 1)

3 Reliability index
n ( 1 , n ) := qnorm1 ( 1 pnorm( 1 , 0 , 1) ) , 0 , 1
n

4 Numerical results

5 Graphical results

1 =

5 .10
pn ( 1 , 50)

p1( 1) =

pn ( 1 , 50) =

0.5

0.159

0.023

0.684

1.3510-3

0.065

3.16710-5

1.58210-3

2.86710-7

1.43310-5

9.86610 -10

4.93310-8

1 := 0 , 0.1.. 6
1 .10

n ( 4.7, 50) = 3.826

nt := 3.8

pnt := p1( nt )

pnt = 7.235 10

pn ( 1 , 1)
pn ( 1 , 5)

pnt

6
n( 1 , 1)
n( 1 , 5)

n( 1 , 50)
nt

4
1

I - 11

Chapter II - Elementary methods of structural reliability I

CHAPTER II - ELEMENTARY METHODS OF STRUCTURAL

RELIABILITY I
Milan Holick1 and Ton Vrouwenvelder2
1

Klokner Institute, Czech Technical University in Prague, Czech Republic

2
Delft University of Technology, TNO BOUW, The Netherlands

Summary
Elementary methods of structural reliability are described considering a fundamental
case of two random variables when the limit state function is formulated as a difference
between the resulting structural resistance and load effect. The initial assumption of normal
distribution of both resulting variables is generalised to any type of probability distribution.
The described computational procedures are illustrated by a number of numerical examples,
which are supplemented by MATHCAD and EXCEL sheets. An extension of the elementary
methods of structural reliability is presented in Annex B.
1

INTRODUCTION

1.1

Background materials
Fundamental concepts and procedures of structural reliability are well described in a
number of national standards, in the new European document EN 1990 [1] and International
Standard ISO 2394 [2]. Additional information may be found in the background document
developed by JCSS [3] and in recently published handbook to EN 1990 [4]. Guidance on
application of the probabilistic methods of structural reliability may be found in publications
and working materials developed by JCSS [5] and in relevant literature listed in [4] and [5].
1.2

General principles
The theory of structural reliability considers all basic variables as random quantities
having appropriate types of probability distribution. Different types of distributions should be
considered for actions, material properties and geometrical data. In addition, model
uncertainties of actions and resistance models should be taken into account. Prior theoretical
models of basic variables and procedures for probabilistic analysis are indicated in JCSS
documents [5].
2

FUNDAMENTAL CASES OF STRUCTURAL RELIBILITY

2.1

General
The fundamental task of the theory of structural reliability concerns a basic
requirement for the relation between the action effect E and the structural resistance R written
in the form of inequality
E<R

II-1

(1)

Chapter II - Elementary methods of structural reliability I

Condition (1) describes a desirable (satisfactory, safe) state of a considered structural
component. It is assumed that structural failure occurs when the condition (1) is not satisfied.
Thus, an assumed sharp (unambiguous) distinction between a desirable (safe) and undesirable
(failure) state of the structure is given as
RE=0

(2)

Equation (2) represents a fundamental form of the failure boundary called the limit
state (performance) function (see also Chapter I of this Handbook 2). It should be noted that
for some structural members and materials the assumption of sharp failure boundary might be
rather artificial and can be accepted as an approximation only. Such a case is indicated in the
following Example 1.
Example 1.
A steel rod indicated in Figure 1 has a tensile resistance R
= d2 fy / 4, where d denotes the diameter of the rod and fy the
yield point. The rod is loaded by a weight E = V, where
V denotes the volume and the bulk weight density of the load.
Thus the inequality (1) has the form

V < d2 fy / 4
The limit state function (2) can be then written as
d2 fy / 4 - V = 0
In this example, the limit state is defined as the state when
the stress in the rod reaches the yield point fy. This simplification
is accepted in many common cases, but (depending on a type of
E
structural steel) it may not correspond to the actual failure of the
rod. In particular when structural steel with significant ductility
and strain hardening is used, then a failure (rupture) will occur Figure 1. A rod.
when the stress reaches the ultimate strength of the steel, which is
a considerably greater than the yield point.
Attached MATHCAD sheets SteelRod.mcd, DesVRod.mcd may be used to make all
numerical calculations.
Both the variables E and R are generally random variables and the validity of
inequality (1) cannot be guaranteed absolutely, i.e. with the probability equal to 1 (the total
certainty). Therefore, it is necessary to accept the fact that the limit state described by
equation (2) may be exceeded and failure may occur with a certain small probability. The
essential objective of the reliability theory is to assess the probability of failure Pf and to find
the necessary conditions for its limited magnitude. For the simple condition in the form of
inequality (1), the probability of failure may be formally written as
Pf = P(E > R)

(3)

The random character of the action effect E and the resistance R, both expressed in
terms of a suitable variable (performance indicator) X (i.e. stress, force, bending moment,
deflection) is usually described by appropriate distribution function, i.e. by distribution
functions E(x), R(x) and by corresponding probability density functions E(x), R(x), where
x denotes a general point of the considered variable X used to express both the variables E and
R. Distributions of variables E and R further depend on appropriate parameters, e.g. on

II-2

Chapter II - Elementary methods of structural reliability I

moment parameters E, E, E, R, R and R. Let us further assume that E and R are
mutually independent (which may be provided by appropriate transformation).
Figure 2 shows an example of the probability density functions of both the variables E
and R and their mutual location. Types of distribution and their parameters shown in Figure 2
are just indicative information. In particular, the moment parameters (the means and standard
deviations) may be considered as relative values given as a percentage of the resistance mean
R (i.e. normalised by R).
Note, that the probability density functions E(x) and R(x) shown in Figure 2 overlap
each other and, therefore, it is clear that unfavourable realizations of variables E and R,
denoted by small letters e and r, may occur in such a way that e > r, i.e. the load effect is
greater than the resistance and failure will occur. Obviously in order to keep the failure
probability Pf = P(E>R) within an acceptable limits, the parameters of variables E and R must
satisfy certain conditions (concerning the mutual position and variances of both distributions)
depending on the types of distribution.
Probability density E(x), R(x)
0.06

Load effect E,
Gamma distribution,
E = 70, E = 7

Resistance R
log-normal distribution,
R = 100, R = 10

0.04

0.02

0.00

40

60

80

100

120

140

Random variable X
Figure 2. Action effect E and resistance R as random variables.
The desired conditions will certainly include the trivial inequality E < R (see Figure
2). Obviously, this requirement for mutual position of both distributions is not sufficient to
ensure specified failure probability Pf. The correct conditions should certainly include also
conditions for variances of both variables. This will be clarified by the following discussion of
fundamental cases of structural reliability.
2.2

Fundamental cases of one random variable

First, consider a special case when one of the variables E and R, say the action effect
E, has a very low (negligible) variability comparing to the variability of resistance R. Then E
may be considered as non-random (deterministic) variable, i.e. such a variable that attains a
certain fixed value E0 (E = E0) in its every realization. This assumption may certainly be
considered as an approximation of some practical cases. One of these cases is the loaded steel
II-3

Chapter II - Elementary methods of structural reliability I

rod from Example 1, where the weight of the suspended mass can be determined with
sufficient accuracy (i.e. without any significant uncertainty). This special case is illustrated in
Figure 3, where the action effect is indicated by a fixed value e0 = 80 (E = 80, E = 0) and
the resistance by the lognormal distribution having the mean R = 100, R = 10 (all numerical
data being normalised to dimensionless quantities).
The probability of failure Pf for the special case of deterministic load effect of actions
shown in Figure 3 may be assessed directly from the distribution function R(x) similarly as
in the case of a fractile. The value e0 may be simply considered as the fractile of the resistance
R for which the probability Pf may be calculated using equation
Pf = P(R < e0) = R(e0)

(4)

The value of distribution function R(E0) is usually assessed from tables for a
standardized random variable U, for which the value u0 corresponding to E0 is computed. It
follows from the general transformation formula
u0 = (e0 - R) / R

(5)

Probability density R(x)

0,06

Deterministic load effect E:

e0 = 80 (E = 80, E = 0)

Random resistance R:
Log-nomal distribution,
R = 100, R = 10

0,04

0,02

0,00

-u0R

40

60

80

100

120

140

Random variable X
Figure 3. Deterministic effect of actions E and random resistance R.
The probability of failure is then given as
Pf = P(R < e0) = R(e0) = U(u0)

(6)

where U(u0) is the value of distribution function of a standardized random variable of the
appropriate distribution (e.g. normal or log-normal).
Note that the value -u0 is the distance of the fixed value E0 of action effect E from the
mean R of resistance R expressed in the units of standard deviation R. If the distribution of
resistance R is normal, then the defined distance is called the reliability index

= (R e0) / R

II-4

(7)

Chapter II - Elementary methods of structural reliability I

and the probability of failure may be expressed by the relation
Pf = P(R < e0) = U()

(8)

In general the reliability index is defined as the negative value of a standardized

normal variable corresponding to the probability of failure Pf. Thus, the following relationship
is accepted as a definition (see Chapter I in this Handbok)

= U1 ( p f )

(9)

where u1 ( p f ) denotes the inverse standardised normal distribution function. At present the
reliability index defined by equation (9) is a commonly used measure of structural
reliability in several national and international documents (see also previous Chapter I of this
Handbook 2). Note, however, that the probability distribution of the resistance R may differ
from the normal distribution.
Example 2.
Consider that resistance R has the mean R = 100 (expressed in dimensionless units),
standard deviation R =10 (the coefficient of variation is VR = 0,10). For the deterministic
action effect it holds that e0 = 80 (see Figure 3). If R has normal distribution, then the
reliability index follows directly from equation (7)

= (100 - 80) / 10 = 2
and probability of failure follows from relation (8)
Pf = P(R < 80) = u(-2) = 0,023
where u(-2) is the value of the distribution function of the standardized normal distribution
for u = 2. However, if the distribution of R is not normal but lognormal with the lower limit
at zero (skewness R = 3VR + V R 3= 0,301 [9]), then it follows from equation (5)
u0 = (80 100) / 10 = 2
The probability of failure Pf is then given as
Pf = P(R < 80) = LN,U(2) = 0,014
where LN,U(-2) is the distribution function of the standardized random variable U with lognormal distribution having the lower bound at zero (the skewness = 0,301). The resulting
probabilities do not much differ but their values are rather high.
If the fixed value of the action effect decreases to e0 = 70, then for normal distribution
of resistance R the reliability index is = 3 and probability of failure is
Pf = P(R < 70) = u(-3) = 0,00135
If the distribution of resistance R is log-normal with the lower limit at zero, then
Pf = P(R < 70) = LN,U(-3) = 0,00021
The reliability index defined by equation (9) is then = U1 (0,00021) = 3,53 , i.e.
greater than the value 3, which holds if normal distribution of resistance R is assumed.
Obviously, when the load effect is only e0 = 70 the resulting failure probabilities are
remarkably lower than in the case when e0 = 80. Furthermore, the numerical example also
shows that the assumption concerning the type of distribution plays an important role and may
be, in some cases, decisive.

II-5

Chapter II - Elementary methods of structural reliability I

2.3

Fundamental case of two random variables

Assume that both basic variables, the action effect E and the resistance R are random
variables. Then it is generally more complicated to assess the probability of failure defined by
equation (3). A simple solution can be obtained assuming a normal distribution for both E and
R. Then also the difference
Z=R-E

(10)

called the safety margin, has the normal distribution with parameters

Z = R - E

(11)

2
2
2 2
Z2 = R + E + 2 RE R E

(12)

where RE is the coefficient of correlation of R and E. It is often assumed that R and E are
mutually independent and RE = 0. Equation (3) for the probability of failure Pf can now be
modified to
Pf = P(E > R) = P(Z < 0) = Z(0)

(13)

and the whole problem is reduced to determining the distribution function Z(z) for z = 0,
which leads to the probabilities of the safety margin Z being negative. The distribution
function Z(0) is usually determined by transformation of the variable Z to standardised
random variable U. Using this equation, the value u0 corresponding to the value g = 0 is
u0 = (0 - Z) /Z = - Z /Z

(14)

The probability of failure is then given as

Pf = P(R < E) = Z(0) = U(u0)

(15)

The probability density function Z(z) of the safety margin Z is shown in Figure 4,
where the grey area under the curve Z(z) corresponds to the failure probability Pf.
Probability density Z(g)
0,04

0,03

1 Pf

0,02

u0 Z
0,01

Pf

0,00
10

10

20

30

Safety margin Z
Figure 4. Distribution of the safety margin Z.

II-6

40

50

Chapter II - Elementary methods of structural reliability I

Assuming that Z has a normal distribution, the value u0 is called the reliability index,
which is commonly denoted by the symbol . In case of a normal distribution of the safety
margin Z, it follows from equations (11), (12) and (14) that the reliability index is given by
a simple relationship

= Z / Z =

R E
R2 + E2 + 2 RE R2 E2

(16)

If the quantities R and E are mutually independent, then the coefficient of correlation
RE vanishes (RE=0). Thus, the reliability index is the distance of the mean Z of the safety
margin Z to the origin, given in the units of the standard deviation Z.
Example 3.
Consider again the Example 2, in which the resistance R and the load effect E are
mutually independent random variables (RE=0) having normal distribution. The resistance R
has the mean R = 100, variance R = 10 (coefficient of variation is therefore only w = 0,10),
and the effect of actions E has the mean E = 80 and E = 8 (all expressed in dimensionless
units). It follows from equation (11) and (12) that

Z = 100 80 = 20
2
2Z = 102 + 82 = 12,81

As both the basic variables R and E have normal distributions, the reliability index
follows directly from equation (16)

= 20 / 12,81 = 1,56
and the probability of failure follows from relation (8)
Pf = P(Z < 0) = U(-1,56) = 0,059
If the variables E and R are not normal, then the distribution of the safety margin G is
not normal either and then the above-described procedure has to be modified. In a general
case, numerical integration or transformation of both variables into variables with normal
distribution can be used. The transformation into a normal distribution is primarily used in
software products.
An approximate simple procedure can be used for a first assessment of the failure
probability Pf. The safety margin Z may be approximated by a three-parameter lognormal
distribution. Assume that the distributions of E and R depend on the moment parameters E,
E, E, R, R and R. The mean and variance of the safety margin Z may be assessed from
the previous equations (11) and (12), which hold for variables with an arbitrary distribution.
Assuming mutual independence of E and R, the skewness Z of the safety Z may be estimated
using the approximate formula (see Annex A - Basic statistical concepts and techniques in
this Handbook 2)

Z =

R3 R E3 E
( R2 + R2 ) 3 / 2

(17)

Then it is assumed that the safety margin Z can be described with sufficient accuracy
by a log-normal distribution with determined moment parameters Z, Z and Z (equations
(11), (12) and (17)). It shows that this approximation offers satisfactory results if the
probability of failure is not too small.
II-7

Chapter II - Elementary methods of structural reliability I

Example 4.
Consider a tie rod having a resistance R under a suspended load of weight E. Let R be
a log-normal variable with origin at zero having the parameters (expressed again in relative
dimensionless units) R = 100 and R = 10 (and therefore R = 0,301), E has Gumbel
distribution with moment parameters E = 50 and E = 10 (for Gumbel distribution [9] has the
positive skewness E = 1,14).
The moment parameters of the safety margin are assessed according to equations (11),
(12) and (17)

Z = R - E = 100 50 = 50
G2 = 2R + 2E = 102 + 102 = 14,14
Z =

3
3
R R E E

2
R

+ 2E )

3/ 2

3
3
10 0,301 10 1,14
= 0,30
(102 + 102 )3 / 2

For a standardized random variable it follows from equation (14) that

u0 = - Z / Z = - 50 / 14,14 = - 3,54
For a log-normal distribution having the skewness Z = - 0,30 it holds that
Pf = P(R < E) = LN,U(-3,54) = 0,00101
which corresponds to the reliability index = 3,09. A more precise result obtained by
application of the software VaP [7] is Pf = 0,00189.
However, when skewness is not taken into account in the assessment of failure
probability and the normal distribution is assumed, it follows that
Pf = P(R < E) = U(-3,54) = 0,00020
which differs significantly from the result when the log-normal distribution was assumed.
Attached MATHCAD sheets StRod.mcd, DesVRod.mcd may be used to make all
numerical calculations.
3

EXACT SOLUTION FOR TWO RANDOM VARIABLES

In the case of two random variables E and R having any distribution, the exact
determination of the failure probability Pf, defined by equation (3), may be obtained by
probability integration. Figure 5 is used to explain this procedure. Let the event A denote the
occurrence of the action effect E in the differential interval <x, x+dx>. Probability of the event
A is given as
P(A) = P(x < E < x+dx) = E(x) dx

(18)

Let us denote B as the event that resistance R occurs within the interval <-, x>.
Probability of the event B is [9] given as
P(B) = P(R < x) = R(x)

(19)

The differential increment of failure probability dPf corresponding to the occurrence

of the variable E in the interval <x, x+dx> is given by the probability of the simultaneous
occurrence of the events A and B, i.e. by the probability of their intersection A B. According
to the principle of multiplication of probabilities [10], it holds that
dPf = P(A B) = P(A) P(B) = P(x < E < x+dx) P(R < x) = R(x) E(x) dx
II-8

(20)

Chapter II - Elementary methods of structural reliability I

The above-mentioned assumption of mutual independence of the variables E and R,
and thus also of the events A and B, is applied here.
Probability density E(x), R(x)
Load effect E

0,06

Resistance R
0,04

dx
0,02

0,00

40

60

x x+dx

100

120

140

Random variable X
Figure 5. Distribution of variables E and R.
The integration of the differential relationship (20) over the interval in which both
variables E and R occur simultaneously (generally the interval <-, >) leads to the relation

p f = R ( x ) E ( x ) dx

(21)

the integration of the relation (21) usually has to be carried out numerically or using the
simulation methods (e.g. direct Monte Carlo methods).
Attached MATHCAD sheet PrLnLn.mcd offers a simple programme that can be used
to evaluate the numerical integration of relation (21) assuming that both the variables E and R
can be described (at least approximately) by the general (three-parameter) lognormal
distribution.
Example 5.
The action effect E and the resistance R are described by a log-normal distribution
with the same parameters as in Example 4 (the Gumbel distribution for E was simply
substituted by the log-normal distribution having the same parameters). The approximate
solution in Example 4, based on log-normal distribution with the lower bound at zero, leads to
the probability of failure Pf = P(R < E) = LN,U(-3,54) = 0,00101. The numerical integration
according to relation (21) using the programme MATHCAD leads to a solution Pf = P(R < E)
= 0,000792, the programme VaP suggests a solution Pf = P(R < E) = 0,000707, which can be
considered as a very good approximation.
The probability of failure Pf assessed by the direct integration may be determined
using MATHCAD sheet PrLnLn.mcd for the given parameters of variables E and R (R =

II-9

Chapter II - Elementary methods of structural reliability I

100, R = 10, E = 50 and E = 10). Variation of probability of failure Pf with the coefficients
of skewness E and R is shown in Figure 6.
It follows from Figure 6 that the probability of failure Pf depends greatly on the
skewnesses E and R (therefore on assumed theoretical models), and in practical conditions,
can differ by several orders of magnitude, even when the means and standard deviations of
the variables E and R remain the same.
Skew ness E
0
1,00E -02
1,00E -03

0,5
-2

1,5

-1

Skew ness R

1
1,00E -04

1,00E -05
pf
1,00E -06

Figure 6. Probability of failure Pf versus the coefficients of skewness E and R for R

= 100, R = 10, E = 50 and E = 10.
It appears that the determination of the failure probability in the case of a simple
example described by inequality (1), where only two random variables E and R are involved,
is easy only when both variables are normally distributed. If they have other distributions, the
exact solution is more complicated and the resulting values depend significantly on the
assumed types of distributions. The approximate solution assuming for E and R a general
(three parameter) lognormal distribution provides a good first estimate of the failure
probability. The obtained values should be, however, verified by more exact procedures
considering appropriate theoretical models of E and R.
4

CONCLUDING REMARKS

Elementary methods of the structural reliability can be used to assess the reliability of
fundamental cases of two random variables when the limit state function is formulated as the
difference between the resulting structural resistance and load effect.
Basic principles of the reliability theory provide operational techniques that can be
used for estimating the partial factors of basic variables. The assessment of various reliability
measures in the new structural design codes is, however, partly based on historical and
empirical experiences. Obviously the past experience depends on local conditions including
climatic actions and traditionally used construction materials and, consequently, might be in
different countries considerably diverse. That is why number of reliability elements and
parameters in the present suite of European standards are open for national choice.

II-10

Chapter II - Elementary methods of structural reliability I

REFERENCES
[1] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[2] ISO 2394 General principles on reliability for structures, ISO 1998.
[3] JCSS: Background documentation, Part 1 of EC 1 Basis of design, 1996.
[4] Gulvanessian, H. Calgaro, J.-A. Holick, M.: Designer's Guide to EN 1990, Eurocode:
Basis of Structural Design; Thomas Telford, London, 2002, ISBN: 0 7277 3011 8, 192 p.
[5] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/, 2001.
[6] EN 1991-1-1 Eurocode 1 Actions on structures. Part 1-1 General actions. Densities, selfweight, imposed loads for buildings, CEN 2002
[7] VaP, Variable Processor, version 1.6, ETH Zurich, 1997.
[8] COMREL, version 7.10, Reliability Consulting Programs, RCP MUNICH, 1999.
[9] Melchers R.E.: Structural Reliability Analysis and Prediction. John Wiley & Sons,
Chichester, 437 p, 1999.
[10] Ang A.H-S. & Tang: W.H. 1975. Probabilistic Concepts in Engineering Planning and
Design. Volume I - Basic principles. John Wiley, London, 1975, 409p.

ATTACHMENTS
1. MATHCAD sheet SteelRod.mcd
Mathcad sheet SteelRod is intended to investigate an effect of the partial factor G on
reliability of a steel rod exposed to permanent load G.
2. MATHCAD sheet DesVRod.mcd
Mathcad sheet DesVRod is intended to investigate of sensitivity factor E and R and
design values Ed and Rd.
3. MATHCAD sheet PrLnLn.mcd
Mathcad sheet PrLnLn is intended for calculation of the failure probability Pf =
P{E>R}based on approximation of E and R by three parameter lognormal distribution.

II-11

Chapter II - Elementary methods of structural reliability I

Attachment 1 - MATHCAD sheet SteelRod.mcd

A steel rod under a permanent loadG - parameter study of G

1 Design of a rod cross section area A = Gd / fd
Design input data:

Gk := 1

G := 1.0 , 1.05.. 1.6 ( parameter)

Design of the cross section area

A ( G) :=

fk := 235

( Gk G)

Check:

fd

m := 1.10

fd :=

fk
m
3

A ( 1.35) = 6.32 10

2 Parameters of basic varibles G and f

Parameters of G and f: G := Gk
Model uncertainty:

280

vG := 0.1 G := vG G :=

XS := 1 XS := 0

XR := 1

f := fk vf := 0.08 f := vf f

235

XR := 0.00

vXR:=

XR

vXS:=

XR

XS
XS

3 Parameters of the resistance R and load effect E

The mean of R and E
CoV: vR :=

R ( G) := f XR A ( G) E := G XS

vXR + vXR vf + vf

vE :=

R ( 1.35) = 1.77
2 Check:
vR = 0.08

vXS + vXS vG + vG

Skewness of R for lognormal and E for gamma distribution:

R := 3 vR + vR

E = 1
vE = 0.1
E := 2 vE

4 Parameters of the reliability margin g = R - E

g ( G) := R ( G) E
g ( G) :=

R ( G) := vR R ( G)

( R ( G) ) + ( E )
3

g ( G) :=

R ( 1.35) = 0.14

g ( 1.35) = 0.77

R R ( G) E E
g ( G)

E := vE E

g ( 1.35) = 0.17

g ( 1.35) = 0.09

5 Reliability assessment without integration

Reliability index assuming normal distribution of g (a first estimate)
0 ( G) :=

g ( G)
g ( G)

Check:

Pf0 ( G) := pnorm( 0 ( G) , 0 , 1)

0 ( 1.35) = 4.44

Reliability index assuming three parameter lognormal distribution of g (a refine estimate)

Parameter C of three
parametr lognormal
distributon of g:

C ( G) :=

g ( G) + 4 + g ( G) g ( G) + 4 g ( G)
1

Parameters of transformed variable:

sg ( G) :=

ln 1 + C ( G)

mg ( G) := ln ( C ( G)

x0 ( G) := g ( G)

1
C ( G)

) + ln ( g ( G) ) ( 0.5) ln (1 + C ( G) 2)

g ( G)

Check:

Pf1 ( G) := plnorm( 0 x0 ( G) , mg ( G) , sg ( G) ) 1 ( G) := qnorm( Pf1 ( G) , 0 , 1)

II-12

x0 ( 1.35) = 4.85
1 ( 1.35) = 4.76

Chapter II - Elementary methods of structural reliability I

6 Reliability assessment using integration
Assuming normal distribution for E:

En ( x ) := dnorm( x , E , E)

E
k :=

Assuming gamma distribution for E:

E Eg ( x ) := dgamma( x , k)
2
E

:=

Assuming lognormal distribution of R having the lower limit at a (0 default):

a ( G) := R ( G) 0.0

C ( G) :=

R ( G)

( R ( G) a ( G) )

m ( G) := ln ( R ( G) ) ln ( C ( G) ) ( 0.5) ln 1 + C ( G)
Probability lognormal distribution of R

aR ( G) := C ( G) + 3 C ( G)
s ( G) :=

ln 1 + C ( G)

Rln ( x , G) := plnorm [ ( x a ( G) ) , m ( G) , s ( G) ]

Failure probability Prob{R<E} and reliability index

t := 3.8

E has normal, R lognormal distribution

Pfn ( G) :=

En ( x ) Rln ( x , G) dx

n ( G) := qnorm( Pfn ( G) , 0 , 1)

E has gamma, R lognormal distribution

Pfg ( G) :=

Eg ( x ) Rln ( x , G) dx

g ( G) := qnorm( Pfg ( G) , 0 , 1)

7 Parametric study of G

Check:

Pf0 ( 1.35) = 4.51 10

0 ( G)

7
6
7

Pfn ( 1.35) = 6.24 10

0 ( 1.35) = 4.44

Reliability index

Pf1 ( 1.35) = 9.68 10

Pfg ( 1.35) = 1.98 10

Effect of the partial factor of G

1 ( G)

g ( G)
n ( G)
t

1 ( 1.35) = 4.76
g ( 1.35) = 4.61

n ( 1.35) = 4.85

1.2

1.4

1.6

Note: Reliability assessment assuming normal distribution for E and R seems to be on a safe side
(leads to a lower bound for ], while assessment assuming three parameter distribution for the
reliability margin g seems to provide a more realistic estimate.

II-13

Chapter II - Elementary methods of structural reliability I

Attachment 2 - MATHCAD sheet DesVRod.mcd

A steel rod under a permanent load G - sensitivity factors E

and R
1 Design of a hanger cross section area A = Gd / fd
Design input data:

Gk := 1

G := 1.0 , 1.05.. 1.6 ( parameter)

Design of the cross section area

A ( G) :=

fk := 235

( Gk G)

fk
m
3

Check:

fd

fd :=

m := 1.10

A ( 1.35) = 6.32 10

2 Parameters of basic variables G and f

Parameters of G and f: G := Gk
Model uncertainty:

vG := 0.1 G := vG G :=

XS := 1 XS := 0

XR := 1

280

f := fk vf := 0.08 f := vf f

235

XR := 0.00

vXR:=

XR

vXS:=

XR

XS
XS

3 Parameters of the resistance R and load effect E

The mean of R and E

R ( G) := f XR A ( G) E := G XS

CoV: vR :=

vXR + vXR vf + vf

vE :=

R ( 1.35) = 1.77
2 Check:
vR = 0.08

vXS + vXS vG + vG

Skewness of R for lognormal and E for gamma distribution:

R := 3 vR + vR

E = 1
vE = 0.1
E := 2 vE

4 Parameters of the reliability marging = R - E

g ( G) := R ( G) E

R ( G) := vR R ( G)

g ( G) :=

( R ( G) ) + ( E)

g ( G) :=

R ( 1.35) = 0.14

g ( 1.35) = 0.77

R R ( G) E E
g ( G)

E := vE E

g ( 1.35) = 0.17

g ( 1.35) = 0.09

5 Sensitivity coefficients E and R

E ( G) :=

E
g ( G)

R ( G) :=

R ( G)
g ( G)

E( G)
R( G)

0.8
0.6
0.4

1.2

1.4

1.6

Note that the sensitivity factor E is shown with the oposite sign (as a positive quantity).

II-14

Chapter II - Elementary methods of structural reliability I

6 Design values Ed and Rd

EC 1990 recommendation:

:= 3.8

E0 := 0.7

R0 := 0.8

Ed ( G) := E E ( G) E

Rd ( G) := R ( G) R ( G) R ( G)

Ed0 ( G) := E E0 E

Rd0( G) := R ( G) R0 R ( G)

Rd0ln( G) := R ( G) exp ( R0 vR)

Check:

Ed ( 1.35) = 1.22
Ed0 ( 1.35) = 1.27

1.8
Ed ( G)
Rd ( G)

Rd ( 1.35) = 1.33
Rd0( 1.35) = 1.34
Rd0ln( 1.35) = 1.39

1.6
1.4

Ed0 ( G)
Rd0 ( G)
Rd0ln ( G)

1.2
1
0.8

1.2

1.4

1.6

Notes: 1) Figure shows that the partial factor G should be greater than about 1,25 otherwise the
design value of the load effect E d would be greater than the design value of the resistance Rd.
2) The design value of the resistance Rd determined assuming lognormal distribution with
the lower bound at zero is greater than Rd determined assuming the normal distribution.

II-15

Chapter II - Elementary methods of structural reliability I

Attachment 3 - MATHCAD sheet PrLnLn.mcd

Failure probability pf=P{E>R}

for log-normal distribution LN(, , ) of E and R
1. Input parameters for E and R : E := 50

E := 10. E := 0 , 0.1.. 2 x := 0 , 0.1E.. 3E wE :=

E
E

R := 100 R := 10. R := 1 , 0.9.. 2 x := 0 , 0.1R .. 3R wR :=

Distribution parameter C
given by the skewness :

E2 + 4 + E

CE( E) :=
3

x0E( E) := E

Distribution bound x0E:

R2 + 4 + R

CR( R) :=
3

Distribution bound x0R:

x0R( R) := R

x0E( 0) =

Check:

R + 4 R

CR( 0.3) =

if R 0

CR( R)

CE( 0.301) =

E 6 E otherwise

Distribution parameter CR
given by the skewness R:

Check:

E + 4 E

if E 0

CE( E)

R 6 R otherwise

x0R( 0.000) =

2. Integration bounds assuming E>0, R arbitrary:

x0( E, R) :=

max( x0E( E) , x0R( R) ) if R 0

x0E( E)

x1( E, R) :=

x0( 0.608, 0.301) =

otherwise

R + 6 R if R 0

x0( 0.608, 1) =

x0R( R) otherwise

x1( 1 , 1) =

3. Transformation to the standardised normal distribution (u) (for any ):

Standardised variable E:

uuE( x, E) :=

uE( x) :=

ln uE( x) +

CE( E)

(x E)

Transformed standardised variable E:

+ ln CE( E) 1 + CE( E)

if E 0

sign ( E) ln 1 + CE( E)

uE( x) otherwise
E( x, E) :=

dnorm( uuE( x, E) , 0 , 1)
E uE( x) +

CE( E)

dnorm( uuE( x, E) , 0 , 1)
E
Standardised variable R:

ln 1 + CE( E)

uR ( x) :=

if CE( E) 0

E( 50, 0) =
E( 50, 0.0001) =

otherwise

( x R)
R

Transformed standardised variable R:

II-16

Chapter II - Elementary methods of structural reliability I

ln uR ( x) +

uuR ( x, R) :=

+ ln CR( R) 1 + CR( R) 2

CR( R)
if R 0
1

sign ( R) ln 1 + CR( R)

uR ( x) otherwise
Distribution function LN,X(x) = LN,U(u) = (uu):
R( x, R) := pnorm( uuR ( x, R) , 0 , 1)

12

uuR ( 50, 0) = 5

R( 0 , 0.3) = 1.237 10

4. Failure probability pf using transformation to normal distribution (for E>0, R arbitrary :

pf ( E, R) :=

x1( E , R)

E( x, E) R( x, R) dx

pf ( 0.608, 0.0001) = 8.745 10

x0( E , R)

0.01

1 .10

1 .10

1 .10

0.5

1.5

Figure 1. Failure probability pf=P{E>R} versus E.

5. Alternative procedure for determination of failure probability using built-in distribution
function for log-normal distributiopn LN,X(x) (for positive only):
mE( E) := ln( CE( E)

) + ln( E) ( 0.5) ln( 1 + CE( E) 2)

sE( E) := ln 1 + CE( E)

Probability density of E: E( x, E) := dlnorm( x x0E( E) , mE( E) , sE( E) )

mR( R) := ln( CR( R)

) + ln( R) ( 0.5) ln( 1 + CR( R) 2)

Distribution function of R:

pf ( E, R) :=

E( 50, 0.0001) = 0.04

sR ( R) := ln 1 + CR( R)

R( x, R) := plnorm( x x0R( R) , mR( R) , sR ( R) )

Failure probability pf (for positive only):

x1( E , R)

Check:
4

E( x, E) R( x, R) dx

pf ( 0.608, 0.0001) = 8.745 10

x0( E , R)

II-17

Chapter III - Reliability differentiation

CHAPTER III - RELIABILITY DIFFERENTIATION

Milan Holicky1) and Jana Markov1)
1)

Czech Technical University in Prague, Czech Republic

Summary
Basic reliability elements specified in current standards for structural design
commonly include failure probability related to a certain reference period T. Required
reliability level of buildings and other civil engineering works is usually specified by the
design (target) failure probability pd or by appropriate reliability index d corresponding to a
specified design working life Td (for example 50 years). In reliability verification the design
values d and Td are sometimes replaced with an alternative reliability index a derived from
the design values d and Td for a convenient reference period Ta (for example 1 year).
Submitted study clarifies relationships between the alternative elements a, Ta and
design values d, Td, and indicates relevant procedures for reliability verification when
alternative reference period Ta is considered. It is emphasised that verification based on a, Ta
should be distinguished from verification of temporary or auxiliary structures when the design
working life Td itself is short. Theoretical consideration and numerical examples show that
characteristic values and partial factors of basic variables describing material properties and
self-weight are significantly dependent on the relevant reference period.
1

INTRODUCTION

1.1

Background documents
Recent documents [1], national [2], [3] and international documents ([4] to [7])
provide general principles and guidance for application of probabilistic methods to structural
designs. The latest European document [5] and international standards [6] and [7] also
indicate a theoretical basis of the so called partial factor method and procedures for
determination of partial factors of material properties and actions using probabilistic
principles.
The basic reliability elements considered in these procedures include probability of
failure p (or equivalent reliability index ) corresponding to a certain reference period T used
in verification of structural reliability. The reference period T used in verification may or may
not coincide with the design working life Td, which is the time period during which a structure
is required to perform adequately. When the reference period used in reliability verification is
different from Td then it is called an alternative period and denoted in this study Ta.
1.2

General Principles
Basic probabilistic methods are used to analyse principles of reliability differentiation.
Similarly as in Chapter I in this Handbook two essentially different cases are distinguished in
the following:
- an alternative reference period Ta (for example 1 or 5 years), which is different from
the design working life Td (for example 50 years), is considered; this case is applicable when
probabilistic models related to the period Ta are more credible than those related to Td;
III - 1

Chapter III - Reliability differentiation

- the design working life Td itself is short (for example 2, 5 or 10 years); this is the
case of temporary or auxiliary structures and structures under a transient design situation
(during execution or repair).
In the following the principles of reliability differentiation specified in current
international documents [5,6,7] and related procedures for determining reliability measures to
be applied in verification cases considering various design-working lives are discussed.
Appropriate reliability elements (characteristic values and partial factors) are derived for
material properties, self-weight and climatic actions (temperature, snow and wind) taking into
account time dependence of failure probability and the reliability index.
2

BASIC RELIABILITY ELEMENTS

The basic reliability measures include the probability of failure and reliability index as
introduced in Chapter I and II in this Handbook. The probability of structural failure Pf can be
generally defined as
Pf = P{Z(X) < 0}
(1)
The limit state (performance) function Z(X) is formulated in such a way that the
reliable (safe) domain of a vector of basic variables X = X1, X2, ... , Xn corresponds to the
inequality Z(X) > 0 while the failure domain to the complementary inequality Z(X) < 0. A
simple example of Z(X) describes the basic relationship between the resulting load effect E
and resistance R
Z(X) = Z = R E
(2)
The random variable Z in equation (2) is often called the reliability (safety) margin; its
mean Z, standard deviation Z and skewness Z may be derived from corresponding
characteristics of resulting variables R and E as indicated in Chapter II.
Instead of the failure probability Pf, the reliability index is frequently used in
reliability consideration as an equivalent quantity to Pf. The reliability index is related to the
failure probability Pf as already indicated in Chapter I
Pf = ()

(3)

In this equation, ( ) denotes the distribution function of standardised normal

distribution. Note that, if the safety margin Z has normal distribution, then the reliability index
may be determined simply as the ratio of Z and Z, thus = Z /Z (in this case denotes the
distance of the mean Z from the origin taking the standard deviation Z as a unit). Chapter I
shows the numerical relationship of both quantities. It should be emphasized that the failure
probability Pf and the reliability index represents fully the equivalent reliability measures
with one to one mutual correspondence given by equation (3).
In the recent European document [5] a design working life for common structures is
considered as Td = 50 years, the reliability index for ultimate limit states d = 3,8 corresponds
to the design failure probability Pd = 7,2 10-5, for serviceability limit states d = 1,5 and pd =
6,7 10-2 (a more appropriate term is the target probabilities used in ISO documents [6]
and [7]). These quantities are recommended as reasonable minimum requirements and it is
emphasized that Pd and d are formal conventional quantities only and may not correspond to
actual frequency of failures.
In design analysis of a structure it is generally required that
Pf Pd
or equivalently in terms of reliability index
III - 2

(4)

Chapter III - Reliability differentiation

(5)

where pd denotes specified design (target) failure probability corresponding to the target
reliability index d.
Conditions (4) or (5) have to be used by designers when probabilistic methods are
applied for verification of structural reliability. Indicative target values pd and d are declared
in some national standards (e.g. [2] and [3]) and recently also specified in international
documents (e.g. [4] to [7]) for various design conditions (limit states, failure consequences
and economic aspects).
3

DESIGN WORKING LIFE AND RELIABILITY

Design working life Td is an assumed period of time for which a structure or part of it
is to be used for its intended purpose with anticipated maintenance but without major repair
being necessary. In recent documents of CEN [5] and ISO [6] indicative values of Td are
provided for five categories of structures as shown in Chapter I of this Handbook.
More detailed specification of structural categories and design working lives is
available in the ISO documents [6, 7]. In general the design working lives indicated in [2] are
greater (in some cases by 100 %) than those given in Chapter I. For example the design
working life for temporary structures indicated in [2] is 15 years, for agricultural structures 50
years, for apartment and office buildings 100 years, and for railway structures, dams, tunnels
and other underground engineering works 120 years.
Design failure probabilities pd are usually indicated in relation to the expected social
and economical consequences. Table 1 shows classification of target reliability levels
provided in EN 1990 [5]. Reliability indexes are given for two reference periods T (1 year
and 50 years) only, without any explicit link to the design working life Td. Similar -values as
in Table 1 are given in [3] for the ultimate limit states, for which, however, the design
working life Td = 80 years (for building structures) is considered.
It should be underlined that a couple of values (a and d) specified in Table 1 for
each reliability class (for 1 year and 50 years) correspond to the same reliability level.
Practical application of these values depends on the time period Ta considered in the
verification, which may be connected with available information concerning time variant
vector of basic variables X = X1, X2, ..., Xn. For example, if the reliability class 2 and 50 years
design working period is considered, then the reliability index d = 3,8 should be used in the
verification of structural reliability. The same reliability level corresponding to class 2 is
achieved when the time period Ta = 1 year and a = 4,7 is used. Thus, various reference
periods Ta, in general different from the design working life Td, may be used for achieving a
certain reliability level.
Table 1. Reliability classification in accordance with CEN [5]
Reliability Consequences for loss of
Reliability index
classes
human life, economical,
a for Ta= d for Td=
social and environmental
1 year
50 years
consequences
3 high
High
5,2
4,3
2 normal Medium
4,7
3,8
1 low

Low

4,2

III - 3

3,3

Examples of buildings and

civil engineering works
Bridges, public buildings
Residential and office
buildings
Agricultural buildings,
greenhouses

Chapter III - Reliability differentiation

Similar target d values are provided in ISO 2394 [6] for the design working life Td
(called in ISO life time) without specification of any particular value of Td. As indicated in
Table 2, two factors are considered for reliability differentiation in [6]: relative costs of safety
measures and consequences of failure.
Table 2. Target reliability index d for the design working life Td given in ISO 2394 [6]
Relative costs of safety
Consequences of failure
measures
small
some
moderate
great
High
0
1,5
2,3
3,1
Moderate
1,3
2,3
3,1
3,8
Low
2,3
3,1
3,8
4,3
It appears that available documents do not provide an explicit guidance on how to take
into account the design working life Td. Both international documents CEN [5] and ISO [6]
give the target value d for specific reference periods T, however, no explicit rule is offered
for adjustment of target value d to different working design lives Td recommended for
various types of construction works.
Nevertheless, some indication is provided in another ISO document [7] for assessment
of existing structures where it is recommended that reliability levels for any residual lifetime
could be similar to those considered for the design working life Td in the case of a new
structure. Consequently, similar reliability levels (expressed in terms of probability pd or
reliability index d) may be considered when designing structures for different design
working lives Td, for example for Td = 50 and Td = 25 years.
4

VARIATION OF FAILURE PROBABILITY WITH TIME

When the vector of basic variables X = X1, X2, ... , Xm is time variant, then failure
probability p is also time variant and should always be related to a certain reference period T,
which may be generally different from the design working life Td. Considering a structure of a
given reliability level, the design failure probability pd = pn related to a reference period Tn = n
T1 can be derived from the alternative probability pa = p1 corresponding to Ta = T1 (to simplify
notation note that previously used subscript "d" corresponds now to "n" and subscript "a" to
"1") using approximate relationship given in [6], [7]
Pn = 1 (1 P1)n

(6)

For very small probabilities, this relationship could be further simplified as pn = p1 Tn /

T1. Time periods T1 and Tn may have an arbitrary length and n = Tn / T1 may not be an integer;
T1 is, however, often one year. Probability pn increases (almost linearly) with Tn.
It follows from equation (6) that reliability indexes 1 = a and n = d, given in
accordance to equation (3) as p1 = (1) and pn = (n) are related as follows [5]
(n) = [(1)]n

(7)

Here (.) denotes the distribution function of standardised normal distribution. Figure
1 shows variation of n with 1 for n = 5, 25, 50 and 100. Note that, if the reference period T1
is one year, then n indicates the number of years of the reference period Tn (n = Tn).

III - 4

Chapter III - Reliability differentiation

Figure 1 confirms data indicated in Table 1. For example, if the target reliability level
of a structure is specified by 50 = 3,8 for the design working life Td = Tn = 50 years, then it
could be verified using reference period Ta = T1 =1 year and a = 1 =4,7. When, however, the
same reliability index 3,8 is specified for a structure having a design working life Tn = 25
years only, thus 25 = 3,8, then the reliability of this structure could be verified using an
alternative reference period T1 = 1 year and reliability index 1 = 4,5, similarly when 5 = 3,8
then 1 = 4,2 (see Figure 1).

5
4,5 3,8 n = 5
25
4
50
3,5
100
3
2,5
2
1,5
1
0,5
0
2,5
3
3,5

4,7
4

4,5

5,5

1
Figure 1. Variation of n with 1 for n = 5, 25, 50 and 100
Note that, if 1-year period would be used for specification of the target reliability level
of a structure, then Figure 1 provides information on the resulting failure probability
corresponding to a given working life Tn. For example, if the target reliability level is
specified by the reliability index 1 = 4,7 (corresponding to the probability p1 = 1,3 10-6),
then (as already mentioned) the reliability level of a structure having a working life, Tn = 50
years is characterised by 50 = 3,8. Similarly when a period Tn = 5 years is used, then 5 = 4,3
or when Tn = 100 years, then 100 = 3,6.
So, the reliability level of a structure can be specified using different time periods T,
which may not necessarily coincide with the design working life Td. This may be useful when
experimental data concerning time variant basic variables are available for a specific
reference period T (for example 1 or 5 years) that is different from the design working life Td.
In such a case, however, all the basic variables (including those that are time independent)
should be considered by appropriate design values related to the same reference period T.
The following simple example indicates the effect of using a reference period T different from
the design working life considering a resistance variable (strength) having lognormal
distribution.
5

PARTIAL FACTOR OF A MATERIAL PROPERTY

Consider a resistance variable R (strength) having lognormal distribution. When an

alternative reference period Ta instead of the design working life Td is used in reliability
verification of a structure, then the design value of R should be determined for Ta instead of
Td. It is assumed that the characteristic value Rk of R is defined as its 5% fractile [5], [6] a [7].

III - 5

Chapter III - Reliability differentiation

Then of the resistance variable R, the characteristic value Rk and design value Rd are defined
as [4], [5]
Rk = R exp ( 1,645 VR)

(8)

Rd = R exp (R a VR)

(9)

Taking into account equations (8) and (9) it follows that the partial factor is given as

R = Rk / Rd = exp ( 1,645 VR) / exp (R a VR)

(10)

Considering selected values of the coefficient of variation VR, Figure 2 shows the
partial factor R for lognormal distribution of R (equation (10)).
It follows from Figure 2 that when reliability of a structure is verified using a short
alternative reference period Ta (for example for example for Ta = 1 year when a = 4,7), the
partial factor R should generally be greater than in the case when the whole design working
life Td (for example for Td = 50 when d = 3,8) is considered. It may be noted that the partial
factor R of material property R increases with the increasing value of the reliability index a.
Similar conclusions can be expected for partial factors of other basic variables, in
particular for partial factors of permanent actions.

2,5

R
2

VR = 0,20
0,15
0,10
0,05

1,5

0,5
0

2 a

Figure 2. Variation of R with a for selected coefficients of variation (R lognormal)

6

PARTIAL FACTORS OF SELF-WEIGHT

Consider a self-weight G having normal distribution. Similarly as in the case of

material property, when an alternative reference period Ta instead of the design working life
Td is used in reliability verification of a structure, then the design value of G should be
determined for Ta instead of Td. The characteristic value Gk of G is defined as the mean G
[5], [6] and [7]:
Gk = G
The design value Gd is given as [4], [5]
III - 6

(11)

Chapter III - Reliability differentiation

Gd = G G G = G + 0,7 a G = G(1 + 0,7 a VG)

(12)

In equation (11) and (12) G denotes the mean, G the standard deviation, VG the coefficient
of variation and G = 0,7 the sensitivity factor of G. The partial factor G of G is defined as
[5], [6] a [7]

G = Gd / Gk

(13)

Taking into account equations (11) and (12) it follows from (13) that

G = (1 + 0,7 a VG)

(14)

Figure 3 shows variation of the partial factor G with the reliability index a for
selected values of the coefficient of variation VG = 0,05; 0,10; 0,15 and 0,20. Note that G =
1,35 (recommended in EN 1990 [5]) corresponds approximately to the reliability index a =
3,8 if the coefficient of variation is about 0,1 (the value in EN 1990 [5] was increased by 5%
to take into account model uncertainty).

2,5

G
2

VG = 0,20
0,15
0,10
0,05

1,5

0,5
0

2 a

Figure 3. Variation of G with a and coefficient of variation VG (G normal).

Assuming the coefficient of variation 0,1 for both the resistance R and the self weight
G Figures 2 and 3 indicate that the partial factor of self-weight G varies slightly more
significantly with a - values than with the partial factor R of resistance variable R. This
finding is, however, dependent on the distributions assumed for both variables.
7

CLIMATIC ACTIONS AND IMPOSED LOADS

Drafts of European documents for climatic actions due to temperature [8], snow [9]
and wind [10] indicate possible reduction of characteristic values Qk for temperature, snow
load and wind speed in case of shorter reference (return) period (for example 5 years) than 50
years considered in normal cases. Such a reduction may be applied in transient design
situations (for example during execution).
III - 7

Chapter III - Reliability differentiation

The following relationships for thermal, snow and wind actions, respectively, are
recommended in relevant Parts of Eurocode EN 1991:
(a) In accordance with EN 1991-1-5 [8] Thermal actions, the maximum and minimum
shade air temperature Tmax,50/Tmin,50 for 50-year return period may be reduced to Tmax,n/ Tmin,n
for n-year return period using the following formulae
Tmax,n = k Tmax,50 , for k = {k1 k2 ln[-ln(1-1/n)]}

(15)

Tmin,n = k Tmin,50, for k = {k3 + k4 ln[-ln(1-1/n)]}

(16)

where Tmax,n/Tmin,n is the maximum/minimum, and the coefficients k1 = 0,781, k2 = 0,056, k3 =

0,393, k4 = -0,156 might be used (based on data of UK [11]),
(b) In accordance with EN 1991-1-3 [9] Snow actions the characteristic value of snow
action sk,n corresponding to the return period of n years is given using Gumbel distribution as

sk,n = k sk,50, where k =

1 Vs
1 Vs

[ln( ln(1 p)) + 0,57722]

(17)

[ln( ln(0,98)) + 0,57722]

where sk,50 is the characteristic snow load on the ground for 50-year return period and sk,n for
n-year return period, p denotes here the probability of sk,n being exceeded corresponding to n
years of return period and Vs is the coefficient of variation of annual maximum snow load,
(c) In accordance with EN 1991-1-4 [10] the basic wind speed vb,n having the return
period n years may be assessed using semi-empirical expressions
1 K ln( ln(1 p ))
vb,n = k vb,50, where k =

1 K ln( ln(0,98))

0 ,5

(18)

where vb,50 is the basic wind velocity for 50-year return period and vb,n for n-year return period
and p denotes here the probability of vb,n being exceeded corresponding to n years of return
period. The constant K in equation (18) follows from Gumbel distribution as K = Vv6/,
where Vv denotes coefficient of variation of annual wind speed. An approximate value K = 0,2
(which corresponds to the coefficient of variation Vv= 0,26) is used in the following
comparison of reduction coefficients k for considered climatic actions.
Table 3 shows reduction coefficients k for climatic actions (applied in a general
relationship Qk,n = k Qk,50) for selected return periods of n - years.
Table 3. Reduction coefficient k for climatic actions (Qk,n = k Qk,50) for different return
periods of n - years.
Return period
p
Reduction coefficient k for
of n-years

Tmax,n

Tmin,n

sn,n

vb,n

2 years

0,5

0,8

0,45

0,64

0,77

5 years

0,2

0,86

0,63

0,75

0,85

10 years

0,1

0,91

0,74

0,83

0,90

50 years

0,02

III - 8

Chapter III - Reliability differentiation

It follows from Table 3 that the characteristic value of climatic actions may be
considerably reduced if shorter reference period is considered in the design. For example for
5-year return period of action due to snow or wind reduces to 75 or 85 % of the characteristic
values for 50-year return period, similarly the characteristic value of the maximum shade air
temperature to 86%, the minimum shade air temperature even to 63%. Note that in
verification of bridge decks during execution phases the characteristic values of uniform
temperature components are derived from shade air temperature [8].
It should be noted that no reduction of partial factors for load is indicated in
documents [8], [9] and [10]. Thus, the same reliability level as for 50-year design working life
described by pd = 7,2 10 5 (d = 3,8) may be considered also for the reference period T = n
years. Certainly, a different reliability level (for example reduced to d < 3,8) can be chosen
taking into account economic and other aspects in accordance to the principles of reliability
differentiation discussed above.
Imposed load could be possibly also reduced when short reference time is considered
similarly as climatic actions. Some statistical data are available in documents of JCSS [12].
However, a variety of random properties of different types of imposed loads make it very
difficult to formulate general rules. Unless convincing data are available the characteristic
values specified in current documents may be accepted without any reduction.
8

EXAMPLES

Consider a steel structure having the design working life Td = 50 years, for which the
target failure probability is specified as pd = 7,2 10 5 (d = 3,8). Failure probability p for the
alternative reference period Ta = 1 year, which is considered in design due to data concerning
actions, will be lower than the target failure probability pd (p < pd and > d); from equation
(6):
pa = 1 - (1-7,2 10 5 )1/ 50 = 1,44 10 6
When the reference period Ta = T1 = 1 year is considered in design verification, then
the reliability index follows from equation (7) as

1 = -1(1,44 10 6) = 4,7
Reliability index 1 is greater than the target value d = 3,8 specified for the design working
life Td = 50 years.
Using equation (10) the partial safety factor R for Ta = T1 = 1 year assuming the
coefficient of variation VR = 0,08 (corresponding to the common variability of strength of
structural steel) the partial safety factor is given as (see also Figure 2)

R = exp ( 1,645 0,08) / exp ( 0,8 4,7 0,08) = 1,18

Note that when the design working life Td = 50 is considered in reliability verification then:

R = exp ( 1,645 0,08) / exp ( 0,8 3,8 0,08) = 1,12

Obviously, the partial factor R increases with the decreasing reference period Ta.
The partial factor of self-weight G is given by equation (14). Assume again, that the
specified reliability level for 50-year design working life is given by d = 3,8. Assuming the
coefficient of variation VG = 0,1 and considering the one year time period for reliability
verification (1 = a = 4,7), then the partial factor G that should be used is

G = (1 + 0,7 4,7 0,1) = 1,33

III - 9

Chapter III - Reliability differentiation

If the verification period is equal to the design working life (d = a = 50 = 3,8), then

G = (1 + 0,7 3,8 0,1) = 1,27

Thus, the variation in G is less significant than the variation in R (see also Figure 3).
A different task is reliability verification of an agricultural structure having the design
working life Td = 25 years, for which the target reliability index can be decreased to d = 3,3
(see Table 1). It follows from equation (10) that the partial factor R for Td = 25 is

R = exp ( 1,645 0,08) / exp ( 0,8 3,3 0,08) = 1,08

The partial factor R may, therefore, be decreased from 1,15 to about 1,1. However it
should be emphasized that this reduction of R is due to a reduced target reliability index d =
3,3 and not due to a shorter design working life Td = 25 instead of the usual Td = 50 years.
Annex A includes MATHCAD Sheet GammaRG that can be used to make
numerical calculations.
9

CONCLUDING REMARKS

(1) In present international documents the target values of failure are related to economic
aspects of safety measures and consequences of structural failure only vaguely, without
any explicit relation to various design working lives Td for different types of structures.
(2) When alternative failure probability pa is derived for a suitable reference period Ta from
the target failure probability pd and design working life Td, partial factors and
characteristic values of variable actions for pa and Ta should also be specified.
(3) For temporary structures, with a short design working life Td, the target failure probability
pd can be specified in accordance with the general principles of reliability differentiation;
reliability elements for basic variables should be derived for specified pd and Td.
(4) The partial factors derived for an alternative reference period Ta different from Td may
vary considerably from the values corresponding to the design working life Td depending
on Ta and distributions of relevant basic variables.
(5) The partial factor of self-weight G corresponding to an alternative reference period Ta
varies with a-values less significantly than the partial factor of material property R.
(6) Partial factors R derived for an alternative reference period Ta of one-year may be
considerably greater than R specified for the design working life Td.
(7) Following recommendations of Eurocodes, the characteristic value for climatic actions
due to snow corresponding to 5-year return (reference) period may be reduced to 75 % of
the characteristic values for 50-year return period, similarly the characteristic value of
wind speed may be reduced to 85 %, the maximum temperature to 86%, the minimum
temperature to 63%.
REFERENCES
[1]
[2]

Gulvanessian, H. Calgaro, J.-A. Holick, M.: Designer's Guide to EN 1990,

Eurocode: Basis of Structural Design; Thomas Telford, London, 2002, ISBN: 07277
3011 8, 192 pp.
SN 730031 Structural reliability. Basic requirements for design. (Spolehlivost
stavebnch konstrukc a zkladovch pd. Zkladn ustanoven pro vpoet), SNI
1990.

III - 10

Chapter III - Reliability differentiation

[3]

SN 731401 Design of steel structures. (Navrhovn ocelovch konstrukc), SNI

1998.
[4] ENV 1991-1 Basis of design and actions on structures. Part 1: Basis of design. CEN
1994.
[5] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[6] ISO 2394 General principles on reliability for structures, ISO 1998.
[7] ISO 13822. Basis for design of structures - Assessment of existing structures, ISO 2001.
[8] EN 1991-1-5 Eurocode 1 Actions on structures. Part 1-5: Thermal actions. CEN,
06/2002.
[9] EN 1991-1-3 Eurocode 1 Actions on structures. Part 1-3: Snow actions. European
Committee for Standardisation, 06/2002.
[10] EN 1991 Actions on structures. Part 1-4: Wind load. European Committee for
Standardisation, 06/2002.
[11] G. Knig et al: New European Code for Thermal Actions, Background document,
Report N.6, University of Pisa, 1999.
[12] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/,
2001.

ATTACHMENTS
1. MATHCAD sheet GammaRG.mcd
MATHCAD sheet Gamma is intended for determination of the partial factor R of the
resistance R and the partial factor G of the permanent load G.
2. MATHCAD sheet PSI0.mcd
MATHCAD sheet PSI0 is intended for determination of the Combination factor 0
for accompanying action.
3. MATHCAD sheet PSI12.mcd
MATHCAD sheet PSI12 is intended for determination of the combination factor 12
for accompanying action.

III - 11

Chapter III - Reliability differentiation

Attachment 1 - MATHCAD sheet Gammarg.mcd

GammaR, gammaG for a theoretical model

MATHCAD sheet for determination of the charactreistic, design values and partial factors R and G.
Coeficients of fractile estimation given in EN 1990
5% fractile V uknown

k := 1.65

0,1 % fractile V unknown

d := 3.09

Sensitivity factors:
Prameters:

R := 0.8

:= 0 , 0.1.. 5

R( ) := R

E := 0.7

E( ) := E

V := 0 , 0.1.. 0.5

Characteristic and design values (relative values related to the mean

Normal distribution

kn( V) := ( 1 k V)

Normal distribution

kln( V) :=

xk=k *x, xd= dsx

dn ( , V) := ( 1 R( ) V)

exp( k) ln 1 + V

dln( V) :=

dn ( 3.8, 0.1) = 0.696

1+ V
dln ( , V) :=

1+ V

exp( R( ) ) ln 1 + V

kn( 0.1) = 0.835

dln( 3.8, 0.1) = 0.735

1+ V
Rn( , V) :=

GammaR

kn( V)

dn ( , V)

Rln( , V) :=

1.5

Rn( , 0.1)

kln( V)

dln( , V)

Rn( 3.8, 0.1) = 1.2

1.5

Rn( , 0.05)

Rln( , 0.05)
Rln( , 0.1)

Rn( , 0.15)

Rln( , 0.15)

0.5

0.5

GammaG for permanent load assuming normal distribution

Gn( , V) := 1 E( ) V

Gn( , 0.05)

1.4

Gn( , 0.15)

Rn( 3.8, 0.1) = 1.2

Gn( 3.3 , V)

Gn( , 0.1)

Gn( 3.8 , V)
1.2

Gn( 4.3 , V)

exp( d ) ln 1 + V

0.2

0.4
V

III - 12

Chapter III - Reliability differentiation

Attachment 2 - MATHCAD sheet PSI0.MCD

MATHCAD sheet "PSI0"

for calculating PSI 0 assuming theoretical models
Combination factor 0 for accompanying action
1 Input data V is coefficient of variability of the accompanying action related to the reference period T
(50 years), r = T/T1 where T1 is the greater of the of basic periods actions to be combined
(for example 5, 7, 10, 50)
Range variables
r := 1 .. 50
:= 3.8 (reliability index)
V := 0.0 , 0.05 .. 1.0
2 Factor 0 for normal distribution:
Formula following Turkstra's rule 0 = F-1( (0,4*0,7 )r)/F-1( (0,7 )):
0n( V, r) :=

1 + qnorm pnorm( 0.28 , 0 , 1) , 0 , 1 V

Check:

1 + 0.7 V

0n( 0.15, 7) = 0.67

Approximation in EC 1990
0na( V, r) :=

1 + ( 0.28 0.7 ln( r) ) V

0na( 0.15, 7) = 0.683

1 + 0.7 V

3 Factor 0 for Gumbel distribution:

0g( V, r) :=

1 0.78 V ( 0.58 + ln( ln( pnorm( 0.28 , 0 , 1) ) ) + ln( r) )

4 General 0 = F -1( (0,4*0,7 c)r)/F-1( (0,7 c)r):

0d( V, r) :=

0g( 0.15, 7) = 0.584

1 0.78 V ( 0.58 + ln( ln( pnorm( 0.7 , 0 , 1) ) ) )

qgamma pnorm( 0.4 c( r) , 0 , 1) , V

r

qgamma( pnorm( c( r) , 0 , 1) , V )
r

pnorm( 0.7 , 0 , 1) , 0 , 1

c( r) := qnorm

0d( 0.3, 7) = 0.488

c( 7) = 3.259

0g( V, 7)
0n( V, 7)
0.5
0na( V , 7)
0d( V, 7)

0.2

0.4

0.6

0.8

Figure 1. Variation of 0 with V for selected distributions.

Check:

V := 0.1 , 0.2 .. 0.5

Note. Gumbel distribution

leads to the lowest 0

V=

0n( V, 10) =

0na( V, 10) = 0g( V, 10) =

0.1

0.727

0.747

0.664

0.2

0.548

0.581

0.474

0.3

0.423

0.465

0.352

0.4

0.33

0.378

0.268

0.5

0.258

0.312

0.205

III - 13

Chapter III - Reliability differentiation

4. Combination factor 0 , for time sensitivity factor T < 1, V refers to period T (50 years)
N := 10
0g( , V, N , T) :=

T := 0.5

V := 0 , 0.1.. 1

1 0.78 V ( 0.58 + ln( ln( pnorm( 0.28 , 0 , 1) ) ) + T ln( N) )

1 0.78 V ( 0.58 + ln( ln( pnorm( 0.7 , 0 , 1) ) ) )

0.8
0g( 3.8 , V , N , 0.5 )
0g( 3.8 , V , N , 0.6 )0.6
0g( 3.8 , V , N , 0.7 )
0.4
0g( 3.8 , V , N , 1 )
0.2

0.2

0.4

0.6

0.8

Figure 2. Variation of 0 with V for Gumbel distribution and factors T.

N := 7

Variation of Q with T

.
T := 0 , 0.1 .. 1

V := 0.3

0.8
0g( 3.3 , V , N , T )0.6
0g( 3.8 , V , N , T )
0g( 4.3 , V , N , T )0.4

0.2

0.2

0.4

0.6

0.8

Figure 3. Variation of 0 with T for Gumbel distribution and reliability indeces .

III - 14

Chapter III - Reliability differentiation

Attachment 3 - MATHCAD sheet PSI12.MCD
MATHCAD SHEET PSI12

Combination factor 12 for accompanying action

1 Input data V is coefficient of variability of the accompanying action related to annual extremes.
Coefficient of variation referred to point in time distribution

w := 0 , 0.1.. 1.1

Probability = 1 - / q for determining Q1

:= 0 , 0.02.. 1.01

Rreliability index

:= 3.8

2. Factor 12 for normal distribution

12( w , ) :=

1 + qnorm( , 0 , 1) w

12( 0.5, 0.5) = 0.493

1 + qnorm( 0.98, 0 , 1) w

3 Factor 12 for Gumbel distribution:

12g( w , ) :=

1 0.78 w ( 0.58 + ln( ln( ) ) )

12g( 0.5, 0.5) = 0.399

1 0.78 w ( 0.58 + ln( ln( 0.98) ) )

4 Comparison of 12 for normal and Gumbel distribution:

0.8

0.6
12g( w , 0.5)
12( w , 0.5)

0.4

0.2

.
0

0.2

0.4

0.6

0.8

Figure 1. Variation of 12 factor with the coefficient of variation V assuming the

normal and Gumbel distribution

The Gumbel distribution leads to a lower 12 factor than the normal distribution

III - 15

Chapter III - Reliability differentiation

5 Variation of 12 with the probability assuming Gumbel distribution:

0.8

0.6

0.4

0.2

0.2

0.4

0.6

0.8

Figure 2. Variation of 12 with the probability for selected coefficients of variation V

assuming Gumbel distribution.

Examples of probability = 1 - / q where is fraction of the refence period (0.01 or 0.5) during which
Q1 and Q2 are exceeded, p probabilioty of Q being non zero
- imposed w=1.1,
short term 1: on 18 days a year = 1 - / q = 1- 0.01/0.05 ~ 0.8
~ 0.5
2: almost always on = 1 - / q = 1- 0.5/1

long term

- wind w=0.5, 1: 10x8 hours a year

- snow w=0.7, on 5 days a year

= 1 - / q = 1- 0.01/0.009 ~ 0.1
= 1 - / q = 1- 0.5/0.009 ~ NA >> 0

2 ~ 0.0

1~ 0.2
= 1 - / q = 1- 0.01/0.014 ~ 0.3
=
1
/
q
=
10.5/0.014
~
NA
>>
0

2 0.0

5 Variation of 12 with the probability and q assuming Gumbel distribution:

Probability of Q being non zero

12g( w , , q ) :=

q := 0 , 0.01.. 1

1 0.78 w 0.58 + ln ln 1

1 0.78 w ( 0.58 + ln( ln( 0.98) ) )

III - 16

12g( 0.5, 0.5, 0.50001) = 0.067

Chapter III - Reliability differentiation

0.8
12g( 1.1 , 0.01 , q)
0.6
12g( 0.5 , 0.01 , q)
12g( 1.1 , 0.5 , q)
12g( 0.5 , 0.5 , q)

0.4

0.2

0.2

0.4

0.6

0.8

Figure 3. Variation of 12 with the probability q for selected coefficients of variation V

and fraction assuming Gumbel distribution.

III - 17

Chapter IV - Design assisted by testing

CHAPTER IV - DESIGN ASSISTED BY TESTING

Igor Kove1
1

Institute for Metal Constructions, Ljubljana, Slovenia

Summary
Under particular circumstances it may be favorable or necessary to carry out tests in
order to obtain certain design parameters. Typical parameters determined from the tests are
actions on the structure, resistance of the structure or structural component and material
properties. Tests can be performed also to calibrate parameters in the theoretical model of
resistance. The design value of the parameter is obtained from the test results as the estimated
value of a certain fractile of the parameter in question. The procedures are explained for the
determination of a single property and for the determination of a probabilistic model of
resistance.
1

INTRODUCTION

1.1

Background materials
The section 5.2 of the European standard EN 1990 [1] mentions briefly the most
general principles of the design assisted by testing and refers to Annex D of the same
standard, where the procedures are dealt with in detail. The international standard ISO 2394
[2] also explains the procedures of the design assisted by testing in its Annex D. The two
standards differ in a number of details regarding this subject. Some insight in the derivation of
the statistical formulations is given in the international standard ISO 12491 [3] and in
statistical literature (e.g. [5]).
1.2

General principles
Under particular circumstances it may be favorable or necessary to carry out the tests
in order to obtain certain design parameters. The examples of possible such circumstances
include:
calculation models are lacking or are inadequate;
large number of similar components will be used;
cases when the calculation model leads to very conservative results;
derivation of new design formulae;
confirmation of assumptions made in the design.
The unknown quantities which are evaluated as a result of the tests may be
actions on the structure (e.g. wind loads);
structural response under loading or accidental effect;
strength or stiffness of the structure or structural element.
The level of reliability of a structure designed by testing should be at least the same as
for structures designed only by calculation models.
IV - 1

Chapter IV - Design assisted by testing

The evaluation of test results should be based on statistical methods. Test results
should in principle include probability distribution of unknown quantities, including the
statistical uncertainties. This distribution is the base for obtaining the design values and partial
factors. The classical statistical interpretation is possible if a large series of tests is performed,
or a smaller series of tests is carried out in order to calibrate a model with one or more
parameters. When only a small number of tests are performed, no classical statistical
interpretation is possible. With the prior information about the distribution of the investigated
quantities it is possible to interpret the test results as statistical using Bayesian procedures.
The design values for a material property, a model parameter or resistance should be
derived from the tests either by (a) assessing the characteristic value and applying the
appropriate partial and conversion factors, or by (b) direct determination of the design value
implicitly or explicitly accounting for reliability required and conversion of results.
The derivation of the characteristic value should take into account the scatter of test
data, statistical uncertainty associated with the number of tests and prior statistical knowledge.
The partial factors should be taken from the appropriate Eurocode. The calculation model
should take into account for differences between test arrangement and real behavior.
In case when method (b) is used, the relevant limit states and the required level or
reliability should be accounted for.
1.3

Preliminary statistical concepts

The basic idea behind the expressions given for the determination of design values
from the tests is the following: the values x1,x2,...,xn of the sample (for example the values
obtained in n realizations of the test) can be regarded as n observed values of the same
random variable X. However, we may equally well regard these n values as a single
observation of n random variables X1,X2,...,Xn (a random vector X) that have the same
distribution (the distribution of X) and are independent, since sample values are assumed to be
independent.
Example 1.
When testing n samples we obtained values x1,x2,...,xn for the parameter X. If we
assume that the parameter X is random variable with the mean and the standard deviation ,
what is the mean m and the standard deviation s of the average value of the sample?
To answer this, we will make use of one of the theorems in probability theory: if we
have n independent random variables X1,X2,...,Xn and ai are arbitrary numbers, then the
following expressions hold:
E((aiXi))=aiE(Xi)

(1)

D((aiXi))=ai2D(Xi)

(2)

where E(X) and D(X) represent the mean value (expectation) and the variance (dispersion) of
the random variable X. The statement about the mean value holds even if the variables XI are
not independent. Note also, that X is distributed arbitrarily.
We can look at the sample x1,x2,...,xn as a realization of the random variables
X1,X2,...,Xn. Then M=(Xi)/n is also a random variable whose realization is the average value
of the sample. If we put ai=1/n in the above equations, we get the expressions for the mean m
and standard deviation s of the sample average value as follows:
m = E(M) = E((Xi)/n) = (E(Xi))/n = n /n =

(3)

s2=D(M) = D((Xi)/n) = (D(Xi))/n2 = n2 /n2 =2/n

(4)

IV - 2

Chapter IV - Design assisted by testing

We see that that if the number of samples (tests) increases, the expected value of the mean of
the average value stays the same (and equal to the mean of the parameter we are testing), but
the standard deviation decreases with the square root of the number of samples. The
coefficient of variation of the average value of the sample is
VM = s/m = /(n) = VX/n

(5)

and also decreases with the number of samples. The ratio VM/VX is shown on the Figure 1.
1
0,9
0,8

VM /Vx

0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
0

10

20

30

40

50

60

70

80

90

100

number of samples n

Figure 1. Ratio VM /VX as a function of the number of samples. VM is the coefficient of

variation of the average value of the sample and VX is coefficient of variation of the measured
parameter.
As has been already mentioned, no statement has been given about the distribution of
the parameter X. The above formulae are valid for arbitrary distribution. If the random
variable X is distributed normally, then the sum Xi is also distributed normally and so is the
mean M. This follows from the fact that a linear combination of several normal random
variables is also a normal random variable (even if these variables are not independent, see
e.g. [5]). The similar also holds for a log-normal distribution: if the random variable X is
distributed log-normally, then the mean M is also distributed log-normally. This is a direct
consequence of the definition: if X is normal random variable, then Y=ln(X) is log-normal
random variable.
Example 2.
In tensile tests the yield stress Y and the tensile strength m of N = 45 specimens of
same (steel) material have been measured. Specimens were cylindrical with diameter d. The
Table 1 gives the measured values (units are mm and MPa). For each of the random parameter
Y , m and d compute the coefficient of variation Vn of the average value of samples
n=1,2,,N and, assuming that V=VN is the coefficient of variation of the particular parameter,
plot a ratio Vn/V as a function of the number of samples n. Compare with the equation (5).

IV - 3

Chapter IV - Design assisted by testing

Table 1. The results of the tensile tests.
n
d
Y
m
Y m n

1
2
3
4
5
6
7
8
9
10
11
12

7,98
8
8
7,99
7,98
7,95
7,98
7,99
7,96
7,99
7,97
7,98

816
845
832
830
846
821
826
822
841
807
831
830

924
944
948
925
969
937
928
934
956
946
942
926

13
14
15
16
17
18
19
20
21
22
23
24

7,98
7,98
7,97
8
8
7,98
8
7,98
7,98
7,98
7,98
8

832
811
840
839
855
830
833
826
836
843
840
847

949
932
937
934
943
928
934
934
942
948
937
928

25
26
27
28
29
30
31
32
33
34
35
36

d
8
7,98
7,99
7,97
7,98
7,99
7,98
7,98
7,97
7,99
7,98
8

818
828
817
851
855
847
822
836
830
830
845
829

907
940
941
959
970
947
921
925
945
938
945
934

37
38
39
40
41
42
43
44
45

7,98
7,98
7,96
7,99
7,98
8
7,98
7,98
8

829
810
832
823
826
829
815
826
823

930
904
925
916
957
931
910
942
932

First we compute the successive average values m1=x1, m2=(x1+x2)/2, , mN=(x1+x2

++xN)/N . Then we treat mi as a realization of the random variable and we find the mean,
standard deviation and the coefficient of variation of successive samples (m1), (m1,m2), ,
(m1,m2,,mN). We repeat this procedure for the three random variables X : the diameter of the
specimen, tensile strength and yield stress. The results are shown in Figure 2 as a plot of the
coefficient of variation Vm against the number of samples. The difference between analytical
and test results is also due to the fact that we dont know the true coefficient of variation Vx.
Instead we took a value N/N, computed from the sample mean and sample standard
deviation of all the N values for a particular parameter in Table 1.
1,2

Vm/Vx

1
analytical
yield stress
tensile strength
diameter

0,8
0,6
0,4
0,2
0
0 5 10 15 20 25 30 35 40 45 50
number of samples n

Figure 2. Coefficient of variation Vm of the average value as a function of the number

of samples. Comparison of analytical result with results from the tests.
2

STATISTICAL DETERMINATION OF A SINGLE PROPERTY

2.1

General principles
This section gives expressions for deriving design values of the ultimate resistance or
serviceability parameters of a structure or a component and for deriving the design values of
material properties.
It is assumed that all variables follow normal or lognormal distribution and that there
is no prior knowledge about the value of the mean. Two cases are considered regarding the
IV - 4

Chapter IV - Design assisted by testing

knowledge of the standard deviation, namely X known and X unknown In EN 1990 [1]
the assumptions of the knowledge of X are replaced with assumption of knowledge of the
coefficient of variation VX (see also remarks in the section 4.6 of the Appendix 1 of Handbook
H1). In practice it is often preferable to use VX unknown together with a conservative
estimate for VX rather then the assumption VX known.
In Eurocode 1990 [1] the determination of characteristic value Xk or design value Xd of
the parameter X in question (a material property, resistance or a model) is based on the
prediction method of fractile estimation (see section 4.3, Appendix 1 to the Handbook H1 for
more details). Similar results are obtained by using the coverage method of fractile estimation
with the confidence level 0,75. In the following two examples this method is explained.
Example 3.
In a test we obtained values x1,x2,...,xn for the parameter X, which we assume to be
distributed normally. Assuming that the standard deviation x of the population is known, find
the value k such that we will have the probability (confidence level) that the true mean x of
the parameter X will be greater then the sample mean mx, according to equation:
Xk = mx - k x x

(6)

In other words, we are solving the equation:

P (mx - k x x) =

(7)

As we have shown (equations (3) and (5)) the sample mean mx is a random variable
with the mean X and the standard deviation X/n. If we multiply the above inequality by -1,
add a value mx and divide by X/n , we get
P ((mx - x)n /x kn = u) =

(8)

where the expression on the left side of the inequality is standardized normal variable and u
is a fractile of the standardized normal distribution corresponding to the probability . The
value k is therefore:
k = u /n

(9)

Example 4.
We have the same situation as in the Example 3, but now the standard deviation x is
unknown. We are searching k such that we will have the probability that the true mean x of
the parameter X will be greater then the sample mean mx, according to equation:
Xk = mx k sx x

(10)

We make use of the following theorem: if all the terms are the same as in Example 3
and sx is the sample standard deviation:
sx = (xi - mx)2/(n-1)

(11)

(mx - x)n / sx

(12)

then the random variable

has a t-distribution with the degree of freedom n-1.
Now we proceed exactly as in the Example 3 and obtain the value k:
k = t /n

(13)

with t being the fractile of the t-distribution of degree n-1 corresponding to the probability .
IV - 5

Chapter IV - Design assisted by testing

2.2

Assessment via the characteristic value

When determining the design value Xd of the parameter X from the assessed
characteristic value Xk , we use the following equation according to Eurocode 1990 [1]:
Xd = d Xk /m

(14)

where the characteristic value Xk is given by

Xk = mx(1 kn Vx)

(15)

This equation is equal to equation (6) with the coefficient of variation Vx given by:
Vx= x/mx

(16)

The m is the partial factor for the parameter X and it should be taken from the appropriate
Eurocode EN 1992 to EN 1998. The d is the design value of the conversion factor. This
factor covers the differences between the laboratory test conditions and conditions during the
actual use. The value kn is obtained from the prediction method of fractile estimation and is:
kn = -up (1/n + 1)1/2

(17)

When the coefficient of variation Vx is known, then up is taken as a fractile of the

standardized normal distribution corresponding to probability p. For characteristic values of
the single property the probability p=0,05 is used, so that up = -1.645. When the coefficient of
variation Vx is unknown, then a fractile tp of the t-distribution with the degree of freedom n 1
corresponding to probability p is used instead of up. The factor kn is dependent on the number
of samples n and is given in Table 2 for two cases, Vx known and Vx unknown and for
probability p=0,05.
Table 2. Values of kn for the 5% characteristic value.
n
1
2
3
4
5
6
8
Vx
2,31 2,01 1,89 1,83 1,80 1,77 1,74
known
Vx
3,37 2,63 2,33 2,18 2,00
unknown

10
1,72

20
1,68

30
1,67

1,64

1,92

1,76

1,73

1,64

The numbers in the Table are actually based on assumptions x known and x
unknown. In EN 1990 [1] these assumptions are replaced with assumption of knowledge of
the coefficient of variation VX. If the standard deviation x is known then Vx should be
computed from the equation (16). If x or Vx is unknown, then Vx is calculated from the
sample standard deviation sx (equation (11)) as
Vx = sx /mx

(18)

The cases above assume that the variable X is distributed normally. What if X is
distributed log-normally?
When the parameter X is distributed log-normally, we use the transformation ln X=Y
to obtain the variable Y which is distributed normally with N(Y,Y2) (see also Appendix A to
the handbook H1). The relationships between the mean and variance of both variables are

Y = ln(x2)/(x2+x2)

(19)

Y 2= ln(1+x2/x2)=ln(1+Vx2)

(20)

IV - 6

Chapter IV - Design assisted by testing

If the parameter X is distributed log-normally, then we proceed as follows. We
transform all the test results according to the equation
yi = ln(xi)

(21)

and compute the sample mean mY from the values yi :

mY = (yi)/n

(22)

Then, if the coefficient of variation Vx is known, we compute Y according to equation (20),

VY from the equation (16), kn from Table 2 and then we calculate Yk from equation (15) using
mY and VY instead of mx and Vx. If the coefficient of variation Vx is unknown, we compute the
sample variance:
sY 2 = (yi - mY )2/(n-1)

(23)

VY from the equation (18), kn from Table 2 and then we calculate Yk from equation (15) using
mY and VY instead of mx and Vx.
Finally we transform the computed characteristic value Yk = mY- k Y (or mY- k sY ) of
the variable Y to the characteristic value Xk of the original variable X:
Xk = exp(mY- k Y)

(24)

Xk = exp(mY- k sY)

(25)

in case if Vx is known and

in case if Vx is unknown. The design value Xd is then calculated using equation (14).
Example 5.
Take the test data from the example 2 and factors m.=1,1 and d =0,8. Calculate the
design value of the tensile strength via the 5% characteristic value for the first 5 test values.
Assume both cases, Vx.=0,05 (known) and Vx.unknown. Assume also both types of
distribution: normal and lognormal.
We assume first the normal distribution of the tensile strength. The sample mean of
the first n=5 test values is m =942 MPa and the sample standard deviation is s= 18.59 MPa.
The coefficient of variation is Vx = 18.59/942=0,01973. For the case Vx. known we have
from the Table 2 kn=1,80 and the design value is
m(d) = 0,8 942(1 1,800,05)/1,1=623,4 MPa.
For the case Vx. unknown we have from the Table 2 kn=2,33 and the design value is
m(d) = 0,8 942(1 2,330,01973)/1,1=653,6 MPa.
Next we assume that the tensile strength is distributed log-normally. We transform the values
of tensile strength using equation (21) and calculate the sample mean mY=6,85 ln(MPa) and
standard deviation sY= 0,01967 ln(MPa). For the case Vx. known we have
VY = (ln(1+0,052))1/2/6,85=0,00729 and sY= 0,007296,85=0,04997 ln(MPa)
from the Table 2 kn=1,80, the value
Yk= 6,85-1,80,04997=6,76 ln(MPa)
and the design value is
m(d) = 0,8 exp(6,76)/1,1= 627,4 MPa.
For the case Vx. unknown we have from the Table 2 kn=2,33, the value
Yk= 6,85-2,330,01967=6,80 ln(MPa)
and the design value is
m(d) = 0,8 exp(6,80)/1,1=655,7 MPa.
IV - 7

Chapter IV - Design assisted by testing

We see that for both types of distribution the design value is greater for the case Vx.
unknown then for the case Vx. known. The reason for this is that we have assumed much
greater Vx.in the case Vx. known that it really is. If we had assumed Vx.=0,015, as obtained
from all the test results, then we would get:
Normal distribution, Vx. known:
m(d) = 0,8 942(1 1,800,015)/1,1=666,6 MPa.
Log-normal distribution, Vx. known, sY = (ln(1+0,0152))1/2=0,015
Yk= 6,85-1,80,015=6,82 ln(MPa)
m(d) = 0,8 exp(6,82)/1,1= 668,2 MPa.
2.3

Direct assessment of the design value

When determining the design value Xd directly, we should use the following formula:
Xd = d mx(1 kd,n Vx)

(26)

In case this method is used, the relevant limit states and the required level or reliability should
be accounted for. The conversion factor d should cover all uncertainties not covered by the
test. The factor kd,n is obtained from the prediction method of fractile estimation with the
lower value of about 0,1 % (the probability p = 0,001). When the coefficient of variation Vx is
known, then parameter X is assumed to be distributed normally and value kd,n is
kd,n = -up (1/n + 1)1/2

(27)

with the 0,001 fractile of the standardized normal distribution up = -3,09 (the value 3,04 is
used in Eurocode 1990 [1]). When the coefficient of variation Vx is unknown, then a fractile tp
of the t-distribution with the degree of freedom n-1 corresponding to probability p=0,001 is
used instead of up. The factor kd,n is dependent on the number of samples n and is given in
Table 3:
Table 3. Values of kd,n for the direct assessment of the design value.
n
1
2
3
4
5
6
8
10
20
Vx
4,36 3,77 3,56 3,44
3,37 3,33 3,27 3,23 3,16
known
Vx
11,40 7,85 6,36 5,07 4,51 3,64
unknown

30
3,13

3,04

3,44

3,04

Example 6.
With the data from the example 5 calculate the design value of tensile strength using
the direct method.
We assume first the normal distribution of the tensile strength. For the case Vx.
known (Vx.=0,015), we have from the Table 3 kd,n=3,37 and the design value is
m(d) = 0,8 942(1 3,370,015)=715,5 MPa.
For the case Vx. unknown we have from the Table 2 kn=7,85 and the design value is
m(d) = 0,8 942(1 7,850,01973)=636,9 MPa
When log-normal distribution is assumed, for the case Vx. known (Vx.=0,015), we
have
Yk= 6,85-3,370,015=6,80 ln(MPa)
m(d) = 0,8 exp(6,80)= 717,9 MPa.
And for the case Vx. unknown
Yk= 6,85-7,850,01967= 6,70 ln(MPa)
IV - 8

Chapter IV - Design assisted by testing

m(d) = 0,8 exp(6,70)=647,1 MPa.

We obtained higher values with the direct method of calculation. This is because we
used the same conversion factor d in both cases.
2.4

Approximation of the factors kn and kd,n

Factors kn and kd,n can be calculated by interpolation from the values in Tables 2 and 3
or, alternatively, by approximation functions:
kn = 1,655+0,672/n ,
p=0,05, Vx known
p=0,05, Vx unknown
kn = n/(-0,950+0,614n) ,
kn = 3,099+1,294/n ,
p=0,001, Vx known
kn = n/(-0,986+0,323n) ,
p=0,001, Vx unknown
Figure 3 shows how these formulas approximate the data from the Tables 2 and 3. The
error when using these formulas is typically less then 1%.

Figure 3. Approximation functions for the factors kn and kd,n .

3

STATISTICAL DETERMINATION OF RESISTANCE MODELS

The procedures given in this section are intended for the calibration of resistance
models and for the derivation of design values from the tests undertaken to reduce
uncertainties in parameters of the resistance model.
Based on observations and theoretical considerations, a design model of the resistance
is developed. The statistical interpretation of the test results should then be used to validate
and adjust the model, until sufficient correlation between test and theoretical data is achieved.
As in the previous section two methods are considered: (a) by assessing the characteristic
value of resistance and (b) by directly assessing the design value of resistance. We start with
the method (a).
The following assumptions are made: the resistance function is a function of
statistically independent variables X=(X1,..,Xj), which are either normally or log-normally

IV - 9

Chapter IV - Design assisted by testing

distributed; a sufficient number of tests is carried out; all relevant material and geometrical
data are measured.
The first step is to develop a design model for the theoretical resistance rt
rt = grt(X)

(28)

The model should include all relevant basic variables Xi that affect the resistance. We then
compare the theoretical model with experimental results. Theoretical values rti are calculated
by substituting actual measured properties of the sample i in the theoretical model. This is to
be compared with measured resistance values rei . We plot the points (rti , rei) in a twodimensional diagram with rti on the abscissa and rti on the ordinate, as shown in Figure 4. If
the theoretical model is accurate, all the points should lie on the diagonal of the first quadrant.
Some scatter will always be present in realistic situations, but if any considerable deviation
from that line occurs, further investigation of the experimental procedures and theoretical
model should be done.

re
re = b rt

rt
Figure 4. The diagram of the experimental to theoretical resistance.
Next, represent the probabilistic model of the resistance as:
rp = b grt(X)

(29)

where b is the slope of the least-squares best fit line, given by

b = (re rt ) / (rt 2)

(30)

is the error term, which represents the model uncertainty:

= r e / rp

(31)

In absence of other information it is assumed that >0 and that it is distributed log-normally.
It means that =ln() is distributed normally. For each test value i we calculate

i= rei / (brti)

(32)

i= ln(i)

(33)

and
The estimated mean and variance s for are
= (i)/n

(34)

s2 = (i- )2 / (n-1)

(35)

IV - 10

Chapter IV - Design assisted by testing

For the estimated value of the coefficient of variation a value
V = exp( s 2 ) 1

(36)

may be used (derived from the equation (20)).

The coefficients of variation VXi of the basic variables should now be determined.
These can be obtained from test data, if it can be shown that the tests are fully representative
of the actual population. Since this is not generally the case, the coefficients of variation VXi
will be established based on prior knowledge or assumptions.
The coefficient of variation Vr of the resistance function is then computed as follows.
If the resistance function is a product of the basic variables Xi
X = X1X2X3Xj

(37)

which are considered to be independent and distributed normally, then by taking logarithm of
the above expression
ln(X) = ln(X1) + ln(X2) + ln(X3) ++ ln(Xj )

(38)

we have a sum of log-normally distributed variables. The variance of the sum of independent
variables is (see equation (2))

ln(X)2 = ln(Xi)2= (ln(1+VXi2))

(39)

where we used the equation (20). Taking the anti-logarithm, we have

1+Vrt2=exp(ln(X)2)=(1+ VX12)(1+ VX22)...(1+ VXj2)

(40)

When VXi are small, the above equation can be simplified to

Vrt2 = (VXi2)

(41)

If the resistance function is more complex function, so that it cannot be expressed as a product
of basic variables, then Vrt is computed from the equation
Vrt2 = D[grt(X)]/grt2(Xm) =

g
1
( rt i )2
g (Xm)
X i
2
rt

(42)

where Xm are the mean values of the basic variables.

The coefficient of the variation Vr of the resistance function is then computed as
Vr2 = (1+V 2)(1+Vrt2) -1

(43)

Qrt = ln(X) = ln(1+Vrt2)

(44)

Q = ln() = ln(1+V2)

(45)

Q = ln(r) = ln(1+Vr2)

(46)

rt = Qrt/Q

(47)

= Q /Q

(48)

the standard deviations

and the weighting factors

The mean value of the resistance function is obtained from the theoretical model using
the mean values Xm of the basic variables:
rm = b grt(Xm)
IV - 11

(49)

Chapter IV - Design assisted by testing

The characteristic resistance rk is finally obtained from the equation (the derivation of this
equation is given in the appendix A):
rk = rm exp(-k rt Qrt- kn Q -0,5 Q2)

(50)

The factor kn is taken from the Table 2 for the case "Vx unknown" and k is the value
of kn for large n (k =1,64). When only one variable is present in the resistance model, only
the term - kn Q from the equation (50) is taken into account.
When the design value of the resistance is assessed directly, rather that from the
characteristic value of resistance, then the procedure is the same with the only modification
that the values k and kn are replaced with the values k,d and kd,n, taken from the Table 3 for
the case " Vx unknown".
Example 7.
Consider the resistance model in the form F=Am /1000. As the basic variables we
take the section of a rod A and the tensile strength m . Let the mean and coefficient of
variation of the basic variables be E(A)=269,76 mm2, V(A)=0,00295, E(m)=936.5 MPa,
V(m)=0,01509. The theoretical Ft and experimental Fe data for various values of the basic
variables are given in the Table 4. Compute the characteristic value of the resistance at the
mean values of basic variables.
n
1
2
3
4
5
6

Table 4. Theoretical Ft and experimental Fe data of the resistance in [kN].

Ft
Fe
n
Ft
Fe
n
Ft
Fe
214,9
215,9
7
243,6
245,2 13
269,3
270,8
219,9
222,4
8
248,1
253,2 14
273,4
274,2
224,9
224,7
9
252,5
261,3 15
277,4
279,6
229,7
228,5 10
256,8
260,1 16
281,3
281,7
234,4
228 11
261
256,5 17
285,2
282,9
239,1
239,7 12
265,2
261,7 18
289
278,7

Fe

The diagram in Figure 5 shows the data from the Table 4 with the regression line
Ft=bFe. The value b is obtained from equation (30): b=0.9995.
300
290
280
270
260
250
240
230
220
210
200
200 210 220 230 240 250 260 270 280 290 300
Ft

Figure 5. Experimental Fe vs. theoretical Ft resistance.

IV - 12

Chapter IV - Design assisted by testing

The mean and sample standard deviation of the logarithm of the error term are given
by =0,000567, s2=0,000277 , the coefficients of variation are calculated from equations
(36), (40) and (43) and are V 2=0,000277, Vrt 2 =0,000236, Vr 2 = 0,000513. The factor
kn=1.78054, is obtained using the approximation formula. The theoretical value of the
resistance function at the mean values of basic variables is rm=252,5 kN. Finally, the
characteristic value of the resistance is computed:
rk=252,5exp(-1,640,67850,01537-1,78050,73460,0166 -0,50.02262)=242.8 kN
We have repeated the above procedure for 10 values of V(A) and V(m) equally spaced
in the range from 0 to 0,4. The Figure 6 shows the dependence of the characteristic value of
the resistance on both coefficients of variation.

Figure 6. The characteristic value of the resistance for various coefficients of variation
of the basic variables A and m .
If the design value is calculated directly, then k,d =3,04 and kn,d = 3,7223 by the
approximation formula, and the design value is
rd=252,5exp(-3,040,67850,01537-3,72230,73460,0166 -0,50.02262)=233,7 kN
The partial factor m for the model resistance in this case is:

m = rm. /rd = 242,8/233,7 = 1,039

(51)

Two software products for the calculation of the characteristic resistance, design
resistance and partial factors are given in the attachments. These products include the
following.

IV - 13

Chapter IV - Design assisted by testing

1. The attached Excel workbook dast.xls can be used to calculate the values rk, rd, and
m for different theoretical and experimental data. The user fills in the yellow input fields with
the theoretical and experimental values of the resistance, the mean and the coefficient of
variation of the basic variables and the mean value of resistance. The worksheet automatically
calculates the characteristic and design values of resistance and the partial factor (shown in
blue fields) and draws the data in a chart.
2. The computer program dast.exe can be used to produce tables of values of these
quantities using different values of the indexes of variation of the basic variables. The source
file dast.c to this computer program is also attached and can be used to produce the dast.exe
if compiled with a C or C++ compiler.
The program dast.exe works in command mode only, i.e. the user must enter the
command mode (MSDos or Command Prompt in the Windows system) to use this program.
The program reads input data from an input file provided by the user (the program and input
files must be at the same, working directory). The format of this input file is as follows. In the
first line the word resistance is entered. The following lines contain the values of theoretical
and experimental values of resistance, two numbers per line. Then a word variations is
entered, followed by the values of the coefficients of variation VX for each basic variable, one
number per line. If the user requires the values rk , rd and m only for the input values of VX,
then the word end completes the input file. If the user requires the values rk , rd or m for a
range of the coefficients of variations VX , then a word calculate is entered followed by the
line containing one of the words characteristic, design or partial, respectfully. Next the
word table is entered followed with the line containing eight numbers. These numbers
indicate the indices of two (arbitrarily chosen) basic variables, the minimum and maximum
value of the coefficient of variation for the first basic variable, the minimum and maximum
value of the coefficient of variation for the second basic variable, and the number of tabulated
values for the first and the second basic variable. The word end completes the input.
Comment lines starting with ; are ignored. The input format is also indicated in more detail
in the source file dast.c.
An example of the use is provided with two input files, dast.i1 and dast.i2, which
contain data from the example 7. The output files dast.o1 and dast.o2 that were produced by
the command-mode commands dast dast.i1 > dast.o1 and dast dast.i2 > dast.o2 are also
attached. As shown in the attachment, the second file dast.o2 can be directly used in Excel to
produce three dimensional chart of the calculated quantity in dependence of the indexes of
variation of two (arbitrarily chosen) basic variables.
REFERENCES
[1] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[2] ISO 2394 General principles on reliability for structures, ISO 1998.
[3] ISO 12491:1997(E): Statistical methods for quality control of building materials and
components, ISO 1997.
[4] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/, 2000
[5] Kreyszig, E.: Advanced Engineering Mathematics, John Wiley & sons, New York,
Chichester, Brisbane, Toronto, Singapore, 1993.

IV - 14

Chapter IV - Design assisted by testing

APPENDIX A. The derivation of the equation (50)
Let X be lognormaly distributed. Then the distribution ln(X) is normal with mean
lnX=(lnX) and standard deviation lnX. The characteristic value of ln(X) can be written:
ln(X)k = lnX kn lnX

(A.1)

Xk = exp(lnX kn lnX)

(A.2)

or
Since the mean (X) of X can be expressed with the mean lnX and standard deviation lnX of
ln(X) by the relationship:

(X)=exp(lnX+lnX2/2)

(A.3)

Xk=(X) exp(- knlnX - lnX2/2)

(A.4)

from there we have:

If X=Y Z is a product of two influences, Y and Z, then
ln(X) = ln(Y) + ln(Z)

(A.5)

lnX = (ln(Y))+(ln(Z))

(A.6)

and the mean of ln(X) is:

The standard deviation lnX of ln(X) can be expressed using FORM factors Y and Z as:

lnX = Y lnY + Z lnZ

(A.7)

If we now combine equations (A.7) and (A.4), then the characteristic value of X as a
combination of Y and Z is:
Xk=(X) exp(-kn Y lnY kn Z lnZ lnX2/2)

(A.8)

If we consider the resistance function r as a variable X, the theoretical resistance function rt as

variable Y and error term as variable Z in the above equation, we rewrite the equation (A.8)
with the notation from section 3:
rk= rm exp(-kn rt Qrt kn Q Q2/2)

(A.9)

Finally, since we assume that there is no statistical uncertainty for theoretical resistance
function rt with respect to the number of samples n, we can substitute k for kn and we obtain
the equation (50).

IV - 15

Chapter IV - Design assisted by testing

ATTACHMENTS
1. EXCEL worksheet from the workbook dast.xls
2. Source file dast.c to the program dast.exe
3. Input file dast.i1 for the program dast.exe
4. Output file dast.o1 produced by dast.exe from dast.o1
5. Input file dast.i2 for the program dast.exe
6. Output file dast.o2 produced by dast.exe from dast.o2
7. EXCEL chart showing the data from dast.o2

IV - 16

Chapter IV - Design assisted by testing

Attachment 1 - EXCEL worksheet from the workbook dast.xls

IV - 17

Chapter IV - Design assisted by testing

Attachment 2 - Source file dast.c to the program dast.exe
/*----------------------------------------------------------------------------Model resistance calculations according to EN 1990, Anex D, 8.2.2 and 8.2.3.
-----------------------------------------------------------------------------*/
#define N_n 100
#define N_j 10
#include <stdio.h>
#include <math.h>
double Rtheor[N_n], Rexper[N_n], Vx[N_j], Va1, Va2, Vb1, Vb2;
int n, nj, na, nb, Xa, Xb, table, calc;
enum { characteristic, design, partial };
char *scalc[]={"characteristic","design","partial"};
/*----------------------------------------------------------------------------Model resistance calculations according to EN 1990, Anex D, 8.2.2 and 8.2.3.
Calculates rk/rm, rd/rm or rk/rd (depending on variable 'calc')
using D.17a and D.21 in EN 1990, Anex D.
If table data Xa, Xb, ... are given (see read_data() below), computes a table
of na*nb values, where the index of variation for the variable Xa is changing
from Va1 to Va2 and the index of variation for the variable Xb is changing
from Vb1 to Vb2.
-----------------------------------------------------------------------------*/
main(argc,argv) int argc; char **argv; { int i,j; extern double Rmodel();
if ( !argv[1] ) printf("USAGE: dast inputfile\n"), exit(1);
read_data(argv[1]);
if ( table ) {
printf("%s values for Vx%d=[%g,%g] (|) and Vx%d=[%g,%g] (-->)\n\n",
scalc[calc],Xa,Va1,Va2,Xb,Vb1,Vb2);
printf("%7s ","");
for ( j=0; j<nb; j++ )
printf("%7.4g ",Vb1+(Vb2-Vb1)/(nb-1)*j);
for ( i=0; i<na; i++ ) {
printf("\n%7.4g ",Vx[Xa-1]=Va1+(Va2-Va1)/(na-1)*i);
for ( j=0; j<nb; j++ ) {
Vx[Xb-1]=Vb1+(Vb2-Vb1)/(nb-1)*j;
printf("%7.4g ",Rmodel(n,Rtheor,Rexper,nj,Vx,1,calc));
}
}
}
else Rmodel(n,Rtheor,Rexper,nj,Vx,0,calc);
}
/*----------------------------------------------------------------------------Reads numerical input data:
Rtheor[]:
Rexper[]:
Vx[]:
Xa, Xb:
Va1, Va2:
Vb1, Vb2:
na, na:

theoretical values of resistance model

experimental values of resistance model
indexes od variation of basic variables
indexes of the first and second basic variable to be tabulated
the range of the index of variation for the first basic variable
the range of the index of variation for the second basic variable
number of tabulated values for the firs and second basic variable

Data are input through a file in the following format:

;comment: the line starting with ';' (or a blank line) is ignored.
resistance
Rtheor[1] Rexeper[1]
Rtheor[2] Rexeper[2]
...
Rtheor[n] Rexeper[n]

IV - 18

Chapter IV - Design assisted by testing

variations
Vx[1]
Vx[2]
...
Vx[j]
table
Xa Xb Va1 Va2 Vb1 Vb2 na nb
calc
characteristic OR design OR partial
end
-----------------------------------------------------------------------------*/
read_data(s) char *s; { FILE *fd; char buf[501];
if ( (fd=fopen(s,"r"))==0 ) printf("can't open %s!",s), exit(1);
for ( ; fgets(buf,500,fd); ) {
LINE:
if ( buf[0]==';' || buf[0]=='\n' ) continue;
else if ( !strncmp("resistance",buf,10) ) {
for ( ; fgets(buf,500,fd); ) {
if ( n>=N_n ) printf("too many points!"), exit(1);
if ( sscanf(buf,"%lg %lg",Rtheor+n,Rexper+n)!=2 ) goto LINE;
n++;
}
}
else if ( !strncmp("variations",buf,10) ) {
for ( ; fgets(buf,500,fd); ) {
if ( nj>=N_j ) printf("too many variables!"), exit(1);
if ( sscanf(buf,"%lg",Vx+nj)!=1 ) goto LINE;
nj++;
}
}
else if ( !strncmp("table",buf,5) ) {
fgets(buf,500,fd);
if ( sscanf(buf,"%d %d %lg %lg %lg %lg %d %d",
&Xa,&Xb,&Va1,&Va2,&Vb1,&Vb2,&na,&nb)!=8 )
printf("error in 'table'!"), exit(1);
table=1;
}
else if ( !strncmp("calculate",buf,9) ) {
fgets(buf,500,fd);
if
( !strncmp("design",buf,6) ) calc=design;
else if ( !strncmp("partial",buf,7) ) calc=partial;
else if ( !strncmp("characteristic",buf,14) ) calc=characteristic;
else goto ERROR;
}
else if ( !strncmp("end",buf,3) ) break;
else ERROR: printf("error in file: '%s'",buf), exit(1);
}
fclose(fd);
if ( n<=0 || nj<=0 || nj>n ) printf("error in n/nj!"), exit(1);
if ( table && (Xa<=0 || Xb<=0 || Xa>nj || Xb>nj) )
printf("error in table!"), exit(1);
}
/*----------------------------------------------------------------------------Computes the factor of a value of model resistance according to EN 1990,
Anex D, 8.2.2 and 8.2.3. Depending on the value of calc (characteristic,
design, partial) returns the exponent from equation D.17a, D.21 and the ratio
of the two, respectfully.
If table!=0 prints intermediate and final calculatuions.
-----------------------------------------------------------------------------*/

IV - 19

Chapter IV - Design assisted by testing

double Rmodel(n,Rtheor,Rexper,j,Vx,table,calc)
int n,j,table; double Rtheor[],Rexper[],Vx[]; {
double rt,ret,b,m,s2,t,Vd2,Vrt2,Vr2,Qrt,Qd,Q,art,ad,kn,knd,fd,fk,gm; int i;
extern double VX2();
for ( rt=ret=i=0; i<n; i++ ) {
rt+= Rtheor[i]*Rtheor[i];
ret+= Rtheor[i]*Rexper[i];
}
b=ret/rt;
for ( s2=m=i=0; i<n; i++ ) {
m+= (t=log(Rexper[i]/(b*Rtheor[i])));
s2+= t*t;
}
m/=n;
s2= n>1? (s2-m*m*n)/(n-1):0.0;
Vd2=exp(s2)-1;
Vrt2=VX2(nj,Vx);
Vr2=(1+Vd2)*(1+Vrt2)-1;
Qrt=sqrt(log(1+Vrt2));
Qd=sqrt(log(1+Vd2));
Q=sqrt(log(1+Vr2));
art=Qrt/Q;
ad=Qd/Q;
kn=n/(-0.95045+0.61443*n);
knd=n/(-0.98623+0.32344*n);
fk=exp(-1.64*art*Qrt-kn*ad*Qd-0.5*Q*Q);
fd=exp(-3.04*art*Qrt-knd*ad*Qd-0.5*Q*Q);
gm=fk/fd;
if ( !table ) {
printf("rt=%g\n",rt);
printf("ret=%g\n",ret);
printf("b=%g\n",b);
printf("m=%g\n",m);
printf("s=%g\n",sqrt(s2));
printf("Vd=%g\n",sqrt(Vd2));
printf("Vrt=%g\n",sqrt(Vrt2));
printf("Vr=%g\n",sqrt(Vr2));
printf("Qrt=%g\n",Qrt);
printf("Qd=%g\n",Qd);
printf("Q=%g\n",Q);
printf("art=%g\n",art);
printf("ad=%g\n",ad);
printf("kn=%g\n",kn);
printf("knd=%g\n",knd);
printf("fk=%g\n",fk);
printf("fd=%g\n",fd);
printf("gm=%g\n",gm);
}
return calc==partial? gm : (calc==design? fd :fk);
}
/*----------------------------------------------------------------------------Computes Vx^2 as a product of (1+Vxi^2)-1.
-----------------------------------------------------------------------------*/
double VX2(j,Vx) int j; double Vx[]; { double s; int i;
for ( s=1, i=0; i<j; i++ ) s*= (Vx[i]*Vx[i]+1);
return s-1;
}

IV - 20

Chapter IV - Design assisted by testing

Attachment 3 - Input file dast.i1 for the program dast.exe
;Example 7, calculation with fixed indexes of variation
resistance
214.9
215.9
219.9
222.4
224.9
224.7
229.7
228.5
234.4
228
239.1
239.7
243.6
245.2
248.1
253.2
252.5
261.3
256.8
260.1
261
256.5
265.2
261.7
269.3
270.8
273.4
274.2
277.4
279.6
281.3
281.7
285.2
282.9
289
278.7
variations
0.00295
0.01509
end

IV - 21

Chapter IV - Design assisted by testing

Attachment 4 - Output file dast.o1 produced by dast.exe dast.i1 >dast.o1
rt=1167250
ret=1166670
b=0.999505
m=0.000567298
s=0.0166434
Vd=0.0166445
Vrt=0.0153757
Vr=0.022661
Qrt=0.0153748
Qd=0.0166434
Q=0.022658
art=0.678558
ad=0.734547
kn=1.78054
knd=3.72232
fk=0.961622
fd=0.925447
gm=1.03909

IV - 22

Chapter IV - Design assisted by testing

Attachment 5 - Input file dast.i2 for the program dast.exe
;Example 7, calculation of the partial factor (gm) in a table
resistance
214.9
215.9
219.9
222.4
224.9
224.7
229.7
228.5
234.4
228
239.1
239.7
243.6
245.2
248.1
253.2
252.5
261.3
256.8
260.1
261
256.5
265.2
261.7
269.3
270.8
273.4
274.2
277.4
279.6
281.3
281.7
285.2
282.9
289
278.7
variations
0.00295
0.01509
calculate
partial
table
1 2 0 0.5 0 0.5 11 11

IV - 23

Chapter IV - Design assisted by testing

Attachment 6 - Output file dast.o2 produced by dast.exe dast.i2 >dast.o2

partial values for Vx1=[0,0.5] (|) and Vx2=[0,0.5] (-->)
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5

0
1.033
1.08
1.154
1.235
1.322
1.414
1.51
1.611
1.716
1.826
1.939

0.05
1.08
1.109
1.173
1.249
1.333
1.423
1.519
1.619
1.724
1.833
1.946

0.1
1.154
1.173
1.221
1.288
1.366
1.452
1.545
1.644
1.747
1.855
1.967

0.15
1.235
1.249
1.288
1.346
1.417
1.498
1.587
1.683
1.785
1.891
2.002

0.2
1.322
1.333
1.366
1.417
1.482
1.558
1.643
1.736
1.835
1.94
2.05

0.25
1.414
1.423
1.452
1.498
1.558
1.63
1.711
1.801
1.898
2.001
2.109

0.3
1.51
1.519
1.545
1.587
1.643
1.711
1.79
1.877
1.972
2.073
2.18

0.35
1.611
1.619
1.644
1.683
1.736
1.801
1.877
1.962
2.054
2.154
2.26

7. Excel chart from the above data

2,6
2,4
2,2
2
1,8
1,6
1,4
0,4

1,2

0,4

0,3

0,5

Vx1

0,2

0,2
0,1

1
0

IV - 24

Vx2

0.4
1.716
1.724
1.747
1.785
1.835
1.898
1.972
2.054
2.146
2.244
2.349

0.45
1.826
1.833
1.855
1.891
1.94
2.001
2.073
2.154
2.244
2.341
2.446

0.5
1.939
1.946
1.967
2.002
2.05
2.109
2.18
2.26
2.349
2.446
2.549

Chapter V - Assessment of existing structures

CHAPTER V - ASSESSMENT OF EXISTING STRUCTURES

Milan Holick1
1

Klokner Institute, Czech Technical University in Prague, Czech Republic

Summary
The approach to assessment of an existing structure is in many aspects different from
that taken in designing the structure of a newly proposed building. The effects of the
construction process and subsequent life of the structure, during which it may have undergone
alteration, deterioration, misuse, and other changes to its as-built (as-designed) state, must be
taken into account. To assess existing structures general principles and rules of the Eurocode
EN 1990 Basis of structural design must be supplemented by specific procedures provided by
International Standards ISO.
1

INTRODUCTION

1.1

Background documents
Background documents related to assessment of existing structures are limited to few
national codes and three International Standards ISO 2394 [1], ISO 13822 [2] and ISO 12491
[3]. General principles and rules of the Eurocode EN 1990 Basis of structural design [4] must
be supplemented by specific procedures provided in the above mentioned International
Standards ISO [1,2,3] that are primarily used in this contribution. Additional information
concerning assessment of existing structures may be found in scientific papers and
publications, for example in the publications [5], [6] and [7].
1.2

General principles
Assessment of existing structures is becoming a more and more important and
frequent engineering task. Continued use of existing structures is of a great significance due to
environmental, economic and socio-political assets, growing larger every year. These aspects
are particularly relevant to buildings that always constitute a great social and economic value.
General principles of sustainable development regularly lead to the need for extension
of the life of a structure, in majority of practical cases in conjunction with severe economic
constraints. That is why assessment of existing structures often requires application of
sophisticated methods, as a rule beyond the scope of traditional design codes.
The approach to assessment of an existing structure is in many aspects different from
that taken in designing the structure of a newly proposed building. The effects of the
construction process and subsequent life of the structure, during which it may have undergone
alteration, deterioration, misuse, and other changes to its as-built (as-designed) state, must be
taken into account. However, even though the existing building may be investigated several
times, some uncertainty in behaviour of the basic variables shall always remain. Therefore,
similarly as in design of new structures, actual variation in the basic variables describing
actions, material properties, geometric data and model uncertainties are taken into account by
partial factors or other code provisions.

V-1

Chapter V - Assessment of existing structures

2

GENERAL FRAMEWORK OF ASSESSMENT

2.1

Reasons for assessment

In general, an existing structure may be subjected to the assessment of its actual
reliability in case of:
- rehabilitation of an existing constructed facility during which new structural members
are added to the existing load-carrying system;
- adequacy checking in order to establish whether the existing structure can resist loads
associated with the anticipated change in use of the facility, operational changes or
extension of its design working life;
- repair of an existing structure, which has deteriorated due to time dependent
environmental effects or which has suffered damage from accidental actions, for
example, earthquake;
- doubts concerning actual reliability of the structure.
In some circumstances assessments may also be required by authorities, insurance
companies or owners or may be demanded by a maintenance plan.
2.2

Common rules
Two common rules are usually accepted when assessing existing structures:
- Currently valid codes for verification of structural reliability should be considered,
historic codes valid in the period when the structure was designed should be used as
guidance documents only.
- Actual characteristics of structural materials, actions, geometric data and structural
behaviour should be considered, the original design documentation including drawings
should be used as guidance documents only.
The first rule should be applied in order to achieve similar reliability level as in case of
newly designed structures. The second principle should avoid negligence of any structural
condition that may affect actual reliability (in favourable or unfavourable way) of a given
structure.
Most of the current codes are developed assuming the concept of limit states in
conjunction with the partial factor method. In accordance with this method, which is mostly
considered here, basic variables are specified by characteristic or representative values. The
design values of the basic variables are determined on the basis of the characteristic
(representative) values and appropriate partial factors.
It follows from the second principle that a visual inspection of the assessed structure
should be made whenever possible. Practical experience shows that inspection of the site is
also useful to obtain a good feel for actual situation and state of the structure.
As a rule the assessment need not to be performed for those parts of the existing
structure that will not be affected by structural changes, rehabilitation, repair, change in use or
which are not obviously damaged or are not suspected of having insufficient reliability.
2.3

General procedure
In general, the assessment procedure consists of the following steps (see the flow chart
in Annex A to this Chapter):
- specification of the assessment objectives required by the client or authority;
- scenarios related to structural conditions and actions;
- preliminary assessment:
- study of available documentation;
- preliminary inspection;
- preliminary checks;

V-2

Chapter V - Assessment of existing structures

- decision on immediate actions;
- recommendation for detailed assessment;
- detailed assessment:
- detailed documentary search;
- detailed inspection;
- material testing and determination of actions;
- determination of structural properties;
- structural analysis;
- verification of structural reliability;
- report including proposal for construction intervention;
- repeat the sequence if necessary.
When the preliminary assessment indicates that the structure is reliable for its intended
use over the remaining life a detailed assessment may not be required. Conversely if the
structure seems to be in dangerous or uncertain condition immediate interventions and
detailed assessment may be necessary.
3

INVESTIGATION

3.1

Purpose
Investigation of an existing structure is intended to verify and update the knowledge
about the present condition (state) of a structure with respect to a number of aspects. Often,
the first impression of the structural condition will be based on visual qualitative
investigation. The description of possible damage of the structure may be presented in verbal
terms like: 'unknown, none, minor, moderate, severe, destructive'. Very often the decision
based on such an observation will be made by experts in purely intuitive way.
A better judgement of the structural condition can be made on the basis of
(subsequent) quantitative inspections. Typically, assessment of existing structures is a cyclic
process when the first inspection is supplemented by subsequent investigations. The purpose
of the subsequent investigations is to obtain a better feel for the actual structural condition
(particularly in the case of damage) and to verify information required for determination of
the characteristic and representative values of all basic variables. For all inspection
techniques, information on the probability of detecting damages if present, and the accuracy
of the results should be given.
3.2

Statement
Results of an investigation should be included in the statement that usually contains
the data describing
- actual state of the structure;
- types of structural materials and soils;
- observed damages;
- actions including environmental effects;
- available design documentation.
A proof loading is a special type of investigation. Based on such tests one may draw
conclusions with respect to:
- the bearing capacity of the tested member under the test load condition;
- other members;
- other load conditions;
- the behaviour of the system.

V-3

Chapter V - Assessment of existing structures

The inference in the first case is relatively easy; the probability density function of the
load bearing capacity is simply cut off at the value of the proof load. The inference from the
other conclusions is more complex. Note that the number of proof load tests needs not to be
restricted to one. Proof testing may concern one element under various loading conditions
and/or a sample of structural elements. In order to avoid an unnecessary damage to the
structure due to the proof load, it is recommended to increase the load gradually and to
measure the deformations. Measurements may also give a better insight into the behaviour of
the system. In general proof loads can hardly address long-term or time-dependent effects.
These effects should be compensated by calculation.
4

BASIC VARIABLES

4.1

General
In accordance with the above-mentioned general principles and rules, characteristic
and representative values of all basic variables shall be determined taking into account the
actual situation and state of the structure. Available design documentation is used as a
guidance material only. Actual state of the structure should be verified by its inspection to an
adequate extent. If appropriate, destructive or non-destructive inspections should be
performed and evaluated using statistical methods.
4.2

Characteristic values
For verification of the structural reliability using partial factor method, the
characteristic and representative values of basic variables shall be considered as follows:
(a) Dimensions of the structural elements shall be determined on the basis of
adequate measurements. However, when the original design documentation is
available and no changes in dimensions have taken place, the nominal dimensions
given in the documentation may be used in the analysis.
(b) Load characteristics shall be introduced with the values corresponding with the
actual situation verified by destructive or non-destructive inspections. When some
loads have been reduced or removed completely, the representative values can be
reduced or appropriate partial factors can be adjusted. When overloading has been
observed in the past it may be appropriate to increase adequately representative
values.
(c) Material properties shall be considered according to the actual state of the
structure verified by destructive or non-destructive inspections. When the original
design documentation is available and no serious deterioration, design errors or
construction errors are suspected, the characteristic values given in original design
may be used.
(d) Model uncertainties shall be considered in the same way as in design stage unless
previous structural behaviour (especially damage) indicates otherwise. In some
cases model factors, coefficients and other design assumptions may be established
from measurements on the existing structure (e.g. wind pressure coefficient,
effective width values, etc.).
Thus reliability verification of an existing structure should be backed up by inspection
of the structure including collection of appropriate data. Evaluation of prior information and
its updating using newly obtained measurements is one of the most important steps of the
assessment.

V-4

Chapter V - Assessment of existing structures

5

EVALUATION OF INSPECTION RESULTS

5.1

Updating in general
Using results of an investigation (qualitative inspection, calculations, quantitative
inspection, proof loading) the properties and reliability estimates of the structure may be
updated. Two different procedures can be distinguished:
(1) Updating of the structural failure probability.
(2) Updating of the probability distributions of basic variables.
Direct updating of the structural reliability (procedure (1)) can be formally carried out
using the following basic formula of probability theory:
P(F|I) =

P( F I )
P( I )

(1)

where P denotes probability, F local or global failure, I inspection information, and

intersection of two events. The inspection information I may consist of the observation that
the crack width at the beam B is smaller than at the beam A. An example of probability
updating using equation (1) is presented in Annex B to this Chapter.
5.2

Updating of probability distribution

The updating procedure of a univariate or multivariate probability distribution
(procedure (2)) is given formally as:
fX(x|I) = C P(I|x) fX(x)

(2)

where fX(x|I) denotes the updated probability density function of X, fX(x) denotes the
probability density function of X before updating, X a basic variable or statistical parameter, I
inspection information, C normalising constant, and P(I|x) likelihood function. An illustration
of equation (2) is presented in Figure 1.
fX(x), fX(x|I)

updated distribution fX(x|I)

prior distribution fX(x)

X
prior xd

updated xd

Figure 1. Updating of probability density function for an expected variable X.

V-5

Chapter V - Assessment of existing structures

In the example shown in Figure 1 updating leads to a more favourable distribution
with a greater design value xd than the prior design value xd. In general, however, the updated
distribution might be also less favourable than the prior distribution.
5.3

Updating of failure probability

Once the updated distributions for the basic variables fX(x) have been found, the
updated failure probability P(F|I) (procedure (1)) may be determined by performing a
probabilistic analysis using common method of structural reliability for new structures.
Symbolically it can be written

P(F|I) =

(3)

( x | I )dx

g( x <0)

where fX(x|I) denotes the updated probability density function and g(x) < 0 denotes the failure
domain (g(x) being the limit state function). It should be proved that the probability P(F|I),
given the design values for its basic variables, does not exceed a specified target value.
5.4

Updating of characteristic and design values

The updating procedure (2) can be used to derive updated characteristic and
representative values (fractiles of appropriate distributions) of basic variables to be used in the
partial factor method. The Bayesian method for fractile updating is described in Annex C to
this Chapter. More information on updating may be found in ISO 12491 [3].
A more practical procedure is to determine directly updated design values for each
basic variable. For a resistance parameter X, the design value can be obtained using
operational formula of ISO 2394 [1]. For normal and lognormal random variable it holds
xd = (1 V )

xd = exp 0,5

(4)
(5)

where xd is the updated design value for X, updated mean value, probabilistic influence
coefficient, target reliability index, V updated coefficient of variation, and 2 = ln(1+V2).
The value of the target reliability index is discussed in ISO/CD 13822 [2], the values
of can be taken equal to those commonly used for new structures (0,7 for the dominating
load parameter, 0,8 for the dominating resistance parameter and 0,3 for non-dominating
variables according to ISO 2394 [1]).
Alternatively one might determine the characteristic value xk first and then calculate
the design value xd by applying the appropriate partial factor m:
xd = xk /m

(6)

For normal and lognormal random variable X the characteristic value xk then follows as
xk = (1 kV )

xk = exp k 0,5

(7)
(8)

where k = 1,64 (5% fractile of the standardised normal distribution) is usually used. It may be
helpful to consider both methods and to use the most conservative result.

V-6

Chapter V - Assessment of existing structures

This procedure may be applied for all basic variables. However, for geomechanical properties
and variable loads other distributions apart from the normal and lognormal distribution may
be more suitable.
Note that a lower acceptable reliability level can be specified by reducing - values
for probabilistic design and reducing - values in the partial factor method. For a material
property X described by a normal distribution the partial factor m may be estimated using
equation

m =

xk
k
=
xd

(9)

which follows from general relationship (6). All the symbols used in (9) are defined above (k
= 1,64 is usually used for the characteristic strength). Similar relationships between m and
may be derived for lognormal or other distributions.
6

STRUCTURAL ANALYSIS

Structural behaviour should be analysed using models that describe actual situation
and state of an existing structure. Generally the structure should be analysed for ultimate limit
states and serviceability limit states using basic variables and taking into account relevant
deterioration processes.
All basic variables describing actions, material properties, load and model
uncertainties should be considered as mentioned above. The uncertainty associated with the
validity and accuracy of the models should be considered during assessment, either by
adopting appropriate factors in deterministic verifications or by introducing probabilistic
model factors in reliability analysis.
When an existing structure is analysed, conversion factors reflecting the influence of
shape and size effect of specimens, temperature, moisture, duration-of-load effect, etc., should
be taken into account. The level of knowledge about the condition of components should be
also considered. This can be achieved by adjusting the assumed variability in either the load
carrying capacity of the components or the dimensions of their cross sections, depending on
the type of structure.
When deterioration of an existing structure is observed, the deterioration mechanisms
shall be identified and a deterioration model predicting the future performance of the structure
shall be determined on the basis of theoretical or experimental investigation, inspection, and
experience.
7

VERIFICATION

Reliability verification of an existing structure shall be made using valid codes of

practice, as a rule based on the limit state concept. Attention should be paid to both the
ultimate and serviceability limit states. Verification may be carried out using partial safety
factor or structural reliability methods with consideration of structural system and ductility of
components. The reliability assessment shall be made taking into account the remaining
working life of a structure, the reference period, and changes in the environment of a structure
associated with an anticipated change in use.
The conclusion from the assessment shall withstand a plausibility check. In particular,
discrepancies between the results of structural analysis (e.g. insufficient safety) and the real
V-7

Chapter V - Assessment of existing structures

structural condition (e.g. no sign of distress or failure, satisfactory structural performance)
must be explained. It should be kept in mind that many engineering models are conservative
and cannot be always used directly to explain an actual situation.
The target reliability level used for verification can be taken as the level of reliability
implied by acceptance criteria defined in proved and accepted design codes. The target
reliability level shall be stated together with clearly defined limit state functions and specific
models of the basic variables.
The target reliability level can also be established taking into account the required
performance level for the structure, the reference period and possible failure consequences. In
accordance with ISO 2394 [1] the performance requirements for assessment of existing
structures are the same as for design of a new structure. Lower reliability targets for existing
structures may be used if they can be justified on the basis of economical, social and
sustainable consideration (see Annex F to ISO 13822 [2]).
An adequate value of the reliability index should be in general determined [2]
considering appropriate reference period. For serviceability and fatigue the reference period
equals the remaining working life, while for the ultimate limit states the reference period is in
principle the same as the design working life specified for new structures (50 years for
buildings). This general approach should be in specific cases supplemented by detailed
consideration of the character of serviceability limit states (reversible, irreversible), fatigue
(inspectable, not inspectable) and consequences of ultimate limit states (economic
consequences, number of endangered people).
8

ASSESSMENT IN THE CASE OF DAMAGE

For an assessment of a damaged structure the following stepwise procedure is

recommended:
1)
Visual inspection
It is always useful to make an initial visual inspection of the structure to get a feel for
its condition. Major defects should be reasonably evident to the experienced eye. In the case
of very severe damage, immediate measures (like abandonment of the structure) may be
taken.
2)
Explanation of observed phenomena
In order to be able to understand the present condition of the structure, one should
simulate the damage or the observed behaviour, using a model of the structure and the
estimated intensity of various loads or physical/chemical agencies. It is important to have
available the documentation with respect to design, analysis and construction. If there is a
discrepancy between calculations and observations, it might be worthwhile to look for design
errors, errors in construction, etc.
3)
Reliability assessment
Given the structure in its present state and given the present information, the reliability
of the structure is estimated, either by means of a failure probability or by means of partial
factors. Note that the model of the present structure may be different from the original model.
If the reliability is sufficient (i.e. better than commonly accepted in design) one might be
satisfied and no further action is required.

V-8

Chapter V - Assessment of existing structures

4)
Additional information
If the reliability according to step 3 is insufficient, one may look for additional
information from more advanced structural models, additional inspections and measurements
or actual load assessment. The updating techniques about how to use this information have
been discussed in section 5.
5)
Final decision
If the degree of reliability is still too low, one might decide to:
- accept the present situation for economical reasons;
- reduce the load on the structure;
- repair the building;
- start demolition of the structure.
The first decision may be motivated by the fact that the cost for additional reliability is
much higher for existing structure than for a new structure. Those who claim that a higher
reliability should be generally required for a new structure than for an existing one sometimes
use this argument. However, if human safety is involved, economical optimisation has a
limited significance.
9

FINAL REPORT AND DECISION

The final report on structural assessment and possible interim reports (if required)
should include clear conclusions with regard to the objective of the assessment based on
careful reliability assessment and cost of repair or upgrading. The report shall be concise and
clear. A recommended report format is indicated in Annex G to ISO 13822 [2].
If the reliability of an existing structure is sufficient, no action is required. If an
assessment shows that the reliability of a structure is insufficient, appropriate interventions
should be proposed. Temporary intervention may be recommended and proposed by the
engineer if required immediately. The engineer should indicate a preferred solution as a
logical follow-up to the whole assessment in every case.
It should be noted that the client in collaboration with the relevant authority should
make the final decision on possible interventions, based on engineering assessment and
recommendations. The engineer performing the assessment might have, however, the legal
duty to inform the relevant authority if the client does not respond in a reasonable time.
10

CONCLUDING REMARKS

Assessment of existing structures is usually based on two common rules:

- Currently valid codes for verification of structural reliability are considered, historic
codes valid in the period when the structure was designed, should be used only as guidance
documents;
- Actual characteristics of structural material, action, geometric data and structural
behaviour should be considered; the original design documentation including drawing should
be used as guidance material only.

V-9

Chapter V - Assessment of existing structures

The most important step of the whole assessment procedure is evaluation of inspection
data and updating of prior information concerning strength and structural reliability. It appears
that a Bayesian approach can provide an effective tool.
Typically, assessment of existing structures is a cyclic process in which the first
preliminary assessment is often supplemented by subsequent detailed investigations and
assessment. A report on structural assessment prepared by an engineer should include a
recommendation on possible intervention. However, the client in collaboration with the
relevant authority should make the final decision concerning possible interventions.
REFERENCES
[1] ISO 2394 (1998) General principles on reliability of structures. ISO, Geneva,
Switzerland.
[2] ISO 13822 (2001) Basis for design of structures -Assessment of existing structures. ISO,
Geneva, Switzerland.
[3] ISO 12491 (1998) Statistical methods for quality control of building materials and
components. ISO, Geneva, Switzerland.
[4] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[5] R.E. Melchers (2001) Structural reliability analysis and prediction. John Wiley & Sons.
[6] Ellingwood B.R. (1996) Reliability-based condition assessment and LRFD for existing
structures. Structural Safety, 18 (2+3), 67-80.
[7] JCSS publication edited by D. Diamantidis (2001) Probabilistic Assessment of Existing
Structures. RILEM. ENS France.

V - 10

Chapter V - Assessment of existing structures

APPENDIX A - GENERAL FLOW OF ASSESSMENT OF EXISTING

STRUCTURES
Requests/Needs
Specification of the assessment objectives and plan
Scenarios
Preliminary assessment
- Study of documents and other evidence
- Preliminary inspection
- Preliminary checks
- Decisions on immediate actions
- Recommendation for detailed assessment
No
Detailed assessment?
Yes

Detailed documentary search and review

Detailed inspection
Material testing and determination of actions
Determination of properties of the structure
Structural analysis
Verification of structural reliability
Yes
Further inspection?
No
-

Reporting results of assessment

Judgement and decision

Sufficient reliability?

Yes

No
Intervention
Construction
- Repair
- Upgrading
- Demolition

Operation
- Monitoring
- Change in use

V - 11

Periodical inspection
Maintenance

Chapter V - Assessment of existing structures

APPENDIX B - PROBABILITY UPDATING
This example of probability updating is adopted from [4] and [5]. Consider the limit
state function G(X), where X is a vector of basic variables, and the failure F is described by
the inequality G(X) < 0. If the result of an inspection of the structure I is an event described
by the inequality H > 0 then using equation (1) in the main text the updated probability of
failure P(F| I) may be written as
P(F| I) = P(G(X) < 0| H > 0) =

P(G ( X ) < 0 H > 0)

P( H > 0)

(B.1)

For example consider a simply supported steel beam of the span L exposed to
permanent uniform load g and variable load q. The beam has the plastic section modulus W
and the steel strength fy.
Using the partial factor method the design condition Rd Sd > 0 between the design
value Rd of the resistance R and design value Sd of the load effect S may be written as
W fyk /m - (g gkL2/8 + q qkL2/8) > 0

(B.2)

where fyk denotes the characteristic strength, gk the characteristic (nominal) value of
permanent load g, qk the characteristic (nominal) value of variable load q, m partial factor of
the steel strength, g the partial factor of permanent load and q the partial factor of variable
load. By analogy to (B.2) the limit state function G(X) follows as
G(X) = R - S = W fy - (gL2/8 + qL2/8)

(B.3)

where all the basic variables are generally considered as random variables described by
appropriate probabilistic models.
To verify its reliability the beam has been investigated and a proof loading up to the
level qtest is carried out. It is assumed that gact is the actual value of the permanent load g. If
the beam resistance is sufficient the information I obtained is described as
I = {H > 0} = {W fy - (gact L2/8 + qtest L2/8) > 0 }

(B.4)

where fy is the actual steel strength, gact the actual permanent load assuming it has been
determined (using non-destructive methods) reasonably accurately.
To determine the updated probability of failure P(F| I) using equation (B.1) it is
necessary to assess the following two probabilities:
P(G(X) < 0 H > 0)=P(W fy -(gL2/8 + qL2/8) < 0 W fy-(gact L2/8 + qtest L2/8) > 0) (B.5)
P(H > 0) = P(W fy - (gact L2/8 + qtest L2/8) > 0)

(B.6)

Additional assumptions concerning the basic variables are needed. Having the results of (B.5)
and (B.6) the updated probability of failure P(G(X) < 0| H > 0) follows from (B.1).
Alternatively, considering results of the proof test, the probability density function
fR(r) of the beam resistance R = Wfy may be truncated below the proof load as indicated in
Figure B.1.

V - 12

Chapter V - Assessment of existing structures

fR(r)

updated resistance R

prior resistance R
r
resistance adequate to proof load
Figure B.1 Truncated effect of proof loading on structural resistance.
Obviously, the truncation of structural resistance R decreases the updated probability
of structural failure defined as
Pf = P(R S < 0)
and increase, therefore, the updated value of structural reliability.

V - 13

(B.7)

Chapter V - Assessment of existing structures

APPENDIX C - BAYESIAN METHOD FOR FRACTILE ESTIMATION
Fractiles of basic variables can be effectively updated using the Bayesian approach
described in ISO 12491 [3]. This procedure is limited here to a normal variable X only for
which the prior distribution function (,) of and is given as
1
1+ + ( n ) )
2
2
exp
(C.1)
( , ) = C (

( s ) + n ( m )
2 2

where C is the normalising constant, (n') = 0 for n' = 0 and (n') = 1 otherwise. The prior
parameters m', s', n', ' are parameters asymptotically given as
1
s
E() = m', E() = s', V() =
, V ( ) =
(C.2)
2
m n
while the parameters n' and ' are independent and may be chosen arbitrarily (it does not hold
that ' = n' 1). In equation (C.2) E(.) denotes the expectation and V(.) the coefficient of
variation of the variable in brackets. Equations (C.2) may be used to make estimates for
unknown parameters n' and ' provided the values V() and V() are estimated using
experimental data or available experience.
The posterior distribution function "(,) of and is of the same type as the prior
distribution function, but with parameters m", s", n" and ", given as

n" = n' + n
" = ' + + ( n )
m"n"= n'm' + nm
"(s")2 + n"(m")2 = '(s')2 + n'(m')2 + s2 + nm2

(C.3)

where m and s are the sample mean and standard deviation, n is the size of the observed
sample and = n 1. The predictive value xp,pred of a fractile xp is then
x p , Bayes = m + t p s 1 + 1 / n

(C.4)

where tp is the fractile of the t-distribution (see Table C.1) with " degrees of freedom. If no
prior information is available, then n'= '= 0 and the characteristics m", n", s", " equal the
sample characteristics m, n, s, . Then equation (C.4) formally reduces to so called prediction
estimates of the fractile given as
x p ,pred = m + t p s 1 + 1 / n

(C.5)

where tp denotes again the fractile of the t-distribution (Table C.1) with degrees of freedom.
Furthermore, if the standard deviation is known (from the past experience), then = and
s shall be replaced by .
Example
A sample of n = 5 concrete strength measurements having the mean m = 29,2 MPa and standard
deviation s = 4,6 MPa is to be used to assess the characteristic value of the concrete strength fck
= xp, where p = 0,05. If no prior information is available, then n'= '= 0 and the
characteristics m", n", s", " equal the sample characteristics m, n, s, . The predictive value
of xp then follows from (C.5) as

V - 14

Chapter V - Assessment of existing structures

f cu,Bayes = 23,9 - 1,8 4,3

1
+ 1 = 15,5 MPa
6

where the value tp = 2.13 is taken from Table C.1 for 1 p = 0.95 and = 5 1 = 4. When
information from previous production is available the Bayesian approach can be effectively used.
Assume the following prior information
m = 30,1 MPa, V(m) = 0,50, s = 4,4 MPa, V(s) = 0,28
It follows from equation (C.2)
2

1 1
4,6 1
n =
6
<1 , =
2 0,282
30,1 0,50
The following characteristics are therefore considered: n' = 0 and ' = 6. Taking into account that

= n - 1 = 4, equations (C.3) yield

n = 6, = 11, x = 23,9 MPa , s = 4,3 Mpa

and finally it follows from equation (C.4)
f cu,Bayes = 23,9 - 1,8 4,3

1
+ 1 = 15,5 MPa
6

where the value tp = 1.81 is taken from Table C.1 for 1 p = 0.95 and = 10.
In this example the resulting characteristic strength is greater (by about 10 %) than the
value obtained by prediction method without using prior information. Thus, when previous
information is available the Bayesian approach may improve (not always) the fractile estimate,
particularly in the case of a great variance of the variable. In any case, however, due caution
should be paid to the origin of the prior information with regard to the nature of considered
variable.
Table C.1 - Fractiles tp of the t-distribution with
1p

0,90 0,95 0,975 0,99 0,995

3 1,64 2,35 3,18
4,54 5,84
12
4 1,53 2,13 2,78
3,75 4,60
14
5 1,48 2,02 2,57
3,37 4,03
16
6 1,44 1,94 2,45
3,14 3,71
18
7 1,42 1,89 2,36
3,00 3,50
20
8 1,40 1,86 2,31
2,90 3,36
25
9 1,38 1,83 2,26
2,82 3,25
30
10 1,37 1,81 2,23
2,76 3,17

V - 15

degrees of freedom
1p
0,90 0,95 0,975
1,36 1,78 2,18
1,35 1,76 2,14
1,34 1,75 2,12
1,33 1,73 2,10
1,32 1,72 2,09
1,32 1,71 2,06
1,31 1,70 2,04
1,28 1,64 1,96

0,99
2,68
2,62
2,58
2,55
2,53
2,49
2,46
2,33

0,995
3,06
2,98
2,92
2,88
2,85
2,79
2,75
2,58

Chapter V - Assessment of existing structures

ATTACHMENTS
1. MATHCAD sheet Update.mcd
MATHCAD sheet Update is intended for determination of updated probability using
Bayes formula. Prior probabilities and likelihoods are taken from file update.prn
2. MATHCAD sheet BayesFract.mcd
MATHCAD sheet BayesFract intended for determination of the characteristic and
design values and material partial factor M using test data in accordance to EN 1990,
Annex D.

V - 16

Chapter V - Assessment of existing structures

Attachment 1 - MATHCAD sheet Update.mcd

"Update.mcd" is MATHCAD sheet for probability updating

Probability updating using Bayes formula
or

where p' ~ P(Bi ) denotes prior probabilities, l~ P(A|B i ) likelihoods and p''~ P(Bi ) updated
(posterior) probabilities.
1 Reading data for apriory probabilties and likelihoods from file
Check vaue

DATA := READPRN( "Update.prn" )

Apriory probabilities

p := DATA

Likelihoods

l := DATA

2 Updated (posterior) probabilities p'' i

n := length ( p )

n=2

pp :=

( p l)

pp =

p l

0.889

0.111

3 Alternative specification of input data using directly this sheet

T

Prior probabilities pi '

p := ( 0.8 0.2 0 0 0 0 0 0 0 0 )

Likelihoods l i

l := ( 1. 0.5 1 1 1 1 1 1 1 1 )

4 Updated (posterior) probabilities p'' i

pp :=

( p l)
p l

5 Listing of the updated probabilities

V - 17

0.889

0.111

pp = 4

pp = 0.889
0

pp = 0.111
1

Chapter V - Assessment of existing structures

Attachment 2 - MATHCAD sheet BayesFract.mcd

MATHCAD sheet "BayesFract" for estimation of sample fractile

MATHCAD sheet for determination of the characteristic and design values and material
partial factor M using test data in accordance to EN 1990, Annex D.
1. Coefficients of variations of prior mean and stadard deviation
V := .01, 0.02.. 0.8

V := .01, 0.02.. 0.8

2. Prior and current observations

Prior obcervations:

m1 := 30.1

s1 := 4.4

Current observation:

m := 29.2

s := 4.6

n := 5

:= n 1

3. Estimates of prior n' and ' assuming V and V

m1 V

n1( V) := floor

Size n1

n1( 0.5) = 0

2
2 V

Degrees of freedom 1
Updated size n2
Updated 2

s1

2( V , V) :=

1( V) := floor

1( 0.28) = 6

n2( V) := n + n1( V)

n2( 0.5) = 5

+ 1( V) 1 if n1( V) 1

2( 0.5, 0.28) = 10

+ 1( V) otherwise
4. Updated means and standard deviations

m2( V) :=

m n + m1 n1( V)

s2 ( V , V) :=

m2( 0.5) = 29.2

n2( V)

s + 1( V) s1 + n m + n1( V) m1 n2( V) m2( V)

2

s2 ( 0.5, 0.3) = 4.49

2( V , V)

5. Coeficients of fractile estimates for probability select the probability

p fractile V unknown
qt inverse Student's
distribution
p fractile V known
qnorm is inverse
normal distribution

ks ( V , V) := qt ( p , 2( V , V) ) 1 +

k( V) := qnorm( p , 0 , 1) 1 +

V - 18

n2( V)

n2( V)

p := 0.05

Chapter V - Assessment of existing structures

6. Fractile estimates
Standard deviation s2 unknown

xp( V , V) := m2( V) + ks ( V , V) s2 ( V , V)

xp( 0.5, 0.28) = 20.303

24

22
xp( 0.5 , V)
20
xp( 0.1 , V)
18

16
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Vxp with V for selected V .

Figure 1. Variation of the frictile

Standard deviation s2 is known, for example s2=s2(0.5,0.28)

xp ( V) := m2( V) + k( V) s2 ( 0.5, 0.28)

xp ( 0.5) = 21.126

24

22
xp( V , 0.28)
20
xp( V , 0.8)
18

16
0

0.1

0.2

0.3

0.4
V

0.5

0.6

Figure 2. Variation of the frictile xp with V for selected V .

V - 19

0.7

0.8

Chapter V - Assessment of existing structures

n1( 0.5) = 0
n1( 0.1) = 2
20

15
n1( V)
n2( V)

10

2( V , 0.28)

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

m2( 0.5) = 29.2

m2( 0.1) = 29.457
10

s2 ( 0.5 , V)

s2 ( 0.1 , V)
4

0.1

0.2

0.3

0.4
V

V - 20

0.5

0.6

0.7

0.8

Chapter VI - Principles of risk assessment

CHAPTER VI - PRINCIPLES OF RISK ASSESSMENT

Milan Holick1
1

Klokner Institute, Czech Technical University in Prague, Czech Republic

Summary
Traditional methods for designing of civil engineering structures and other engineering
systems are frequently based on the concept of target probability of failure. However, this
fundamental quantity is usually specified on the basis of comparative studies and past
experience only. Moreover, probabilistic design methods suffer from several deficiencies,
including lack of consideration for accidental and other hazard situations and their
consequences. Both of these extreme conditions are more and more frequently becoming
causes of serious failures and other adverse events. Available experience clearly indicates that
probabilistic design procedures may be efficiently supplemented by a risk analysis and
assessment, which can take into account various consequences of unfavourable events. It is
therefore anticipated that in addition to traditional probabilistic concepts the methods of
advanced engineering design will also commonly include criteria for acceptable risks.
1

INTRODUCTION

1.1

Background documents
Background documents of the risk analysis of civil engineering systems considered in
this contribution consist of a number of national and international documents [1] to [9]. It
should be noted that Eurocode EN 1990 [10] for design of civil structures is based on the
concept of the target probability of failure pd. However, it is well recognised that the reliability
of structures and other engineering systems suffers from a number of uncertainties that can
hardly be analysed and well described by probabilistic methods [11, 12]. Moreover, traditional
probabilistic concepts consider the significance of failure and other adverse events only very
vaguely [10]. That is why probabilistic methods are often supplemented by recently developing
methods of risk assessment [12]. In some countries, risk assessment even becomes compulsory
by law in the case of complex technical systems (power stations, tunnel routes).
1.2

General principles
General principles of the risk analysis and the common tools applied for investigating
civil engineering systems considered in this contribution follow the basic concepts presented in
documents [1] to [9]. The risk analysis is an important part of the risk assessment and the entire
risk management of a system as indicated in Figure 1 (adopted from [2]).
The risk analysis of a system consists of the use of all available information to estimate
the risk to individuals or populations, property or the environment, from identified hazards.
Risk assessment further includes risk evaluation (acceptance or treatment) as indicated in
Figure 1 (adopted from [2]). The whole procedure of the risk assessment is typically an
iterative process as indicated in Figure 2 (adopted from [9]). The first step in the risk analysis
involves the context (scope) definition related to the system and the subsequent identification
of hazards.

VI - 1

Chapter VI - Principles of risk assessment

Risk management
Risk control

Risk assessment
Risk
analysis

Hazard
identification

Risk evaluation

Risk
estimation

Decisionmaking

Risk
acceptance

Monitoring

Risk
treatment

Figure 1. A framework for risk management (adopted from [2]).

The system is understood [2] as a bounded group of interrelated, interdependent or
interacting elements forming an entity that achieves in its environment a defined objective
through interaction of its parts. In the case of technological hazards related to civil engineering
works, a system is normally formed from a physical subsystem, a human subsystem, their
management, and the environment. Note that the risk analysis of civil engineering systems
(similarly as analysis of most systems) involves usually several interdependent components
(e.g. human life, injuries, economic loss).
Any technical system may be exposed to a multitude of possible hazard situations. In
the case of civil engineering facilities, hazard situations may include both, environmental
effects (wind, temperature, snow, avalanches, rock falls, ground effects, water and ground
water, chemical or physical attacks, etc.) and human activities (usage, chemical or physical
attacks, fire, explosion, etc.). As a rule hazard situations due to human errors are more
significant than hazards due to environmental effects.
2

HAZARD IDENTIFICATION

A hazard is a set of circumstances, possibly occurring within a given system, with the
potential for causing events with undesirable consequences. For instance the hazard of a civil
engineering system may be a set of circumstances with the potential to an abnormal action (e.g.
fire, explosion) or environmental influence (flooding, tornado) and/or insufficient strength or
resistance or excessive deviation from intended dimensions. In the case of a chemical
substance, the hazard may be a set of circumstances likely to cause its exposure [2].
Hazard identification and modelling is a process to recognize the hazard and to define
its characteristics in time and space. In the case of civil engineering systems the hazards Hi may
be linked to various design situations of the building (as defined in [7]) including persistent,
transient and accidental design situation. As a rule Hi are mutually exclusive situations (e.g.
persistent and accidental design situations of a building). Then if the situation Hi occurs with
the probability P{Hi}, it holds P{Hi} = 1. If the situations Hi are not mutually exclusive, then
the analysis becomes more complicated.

VI - 2

Chapter VI - Principles of risk assessment

Start

Hazard identification

Concequence analysis

Risk estimation

Risk assessment

Probabilty analysis

Risk analysis

Definition of the system

Risk evaluation
No

Risk treatment

Acceptable risk?
Yes

Stop

Figure 2. Flowchart of iterative procedure for the risk assessment (adopted from [9]).
Note that in some documents (for example in the recent European document EN 1990
[10]) the hazard is defined as an event, while in risk analysis [2] it is usually considered as a
condition with the potential for causing event, thus as a synonym to danger.
3

DEFINITION AND MODELLING OF RELEVANT SCENARIOS

Hazard scenario is a sequence of possible events for a given hazard leading to undesired
consequences. To identify what might go wrong with the system or its subsystem is the crucial
task to risk analysis. It requires detail examination and understanding of the system [6].
Nevertheless, a given system is often a part of a larger system. Consequently, modelling and
subsequent analysis of the system is a conditional analysis.
The modelling of relevant scenarios may be dependent on specific characteristics of the
system. For this reason a variety of techniques have been developed for the identification of
hazards (e.g. PHA HAZOP) and for the modelling of relevant scenarios (fault tree, event
tree/decision trees, causal networks). Detail description of these techniques is beyond the scope
of this contribution, may be however found in [6, 9] and other literature.

VI - 3

Chapter VI - Principles of risk assessment

4

ESTIMATION OF PROBABILITIES

Probability is generally the likelihood or degree of certainty of a particular event

occurring during a specified period of time. In particular, reliability of a structure is often
expressed as probability related to a specific requirement and a given period of time, for
example 50 years [3,10].
Assuming that a system may be found in mutually exclusive situations Hi, and the
failure F of the system (e.g. of the structure or its element) given a particular situation Hi
occurs with the conditional probability P{F|Hi}, then the total probability of failure pF is given
by the law of total probability (see for example [11]) as:
pF =

P{H }P{F | H }
i

(1)

Equation (1) can be used for the modification of the partial probabilities P{Hi}P{F|Hi}
(appropriate to the situations Hi) with the aim to comply with the design condition pF < pt,
where pt is a specified target probability of failure. The target value pt may be determined using
the probabilistic optimisation of an objective function describing, for example, the total cost.
The conditional probabilities P{F|Hi} must be determined by a detail probabilistic
analysis of the respective situations Hi under relevant scenarios. The traditional reliability
methods [8] assume that the failure F of the system can be well defined in the domain of the
vector of basic variables X. For example, it is assumed that a system failure may be defined by
the inequality g(x) < 0, where g(x) is the so called limit state function, where x is a realisation
of the vector X. Note that g(x) = 0 describes the boundary of the limit state, and the inequality
g(x) > 0 the safe state of a structure.
If the joint probability density fX(x|Hi) of basic variables X given situation Hi is known,
the conditional probability of failure P{F|Hi} can be then determined [6] using the integral
P{F|Hi} =

X
g ( x )<0

( x | H i )dx

(2)

It should be mentioned that the probability P{F|Hi} calculated using equation (2) suffers
generally from two essential deficiencies:
- uncertainty in the definition of the limit state function g(x),
- uncertainty in the theoretical model for the density function fX(x|Hi) of basic variables
X [8].
These deficiencies are most likely the causes of the observed discrepancy between the
determined probability pF and actual frequency of failures; this problem is particularly
disturbing in case of fire. Yet, the probability requirement pF < pt is generally accepted as a
basic criterion for design of structures.
In a risk analysis we need to know not only probability of the structural failure F but
probabilities of all events having unfavourable consequences. In general, the situations Hi may
cause a number of events Eij (e.g. excessive deformations, full development of the fire). The
required conditional probabilities P{Eij|Hi} must be estimated by a separate analysis using
various methods, for example the fault tree method or causal networks.
5

ESTIMATION OF CONSEQUENCES

Consequences are possible outcomes of a desired or undesired event that may be

expressed verbally or numerically to define the extent of human fatalities and injuries or
environmental damage and economic loss [1]. A systematic procedure to describe and/or
calculate consequences is called consequence analysis. Obviously, consequences are generally
VI - 4

Chapter VI - Principles of risk assessment

not one-dimensional. However in specific cases they may be simplified and described by
several components only, e.g. by human fatalities, environmental damage and costs. At present
various costs are usually included only. It is assumed that adverse consequences of the events
Eij can be normally expressed by several components Cij,k, where the subscript k denotes the
individual components (for example the number of lost lives, number of human injuries and
damage expressed in a certain currency).
6

ESTIMATION OF RISK

Risk is a measure of the danger that undesired events represent for humans,
environment or economic values. Risk is commonly expressed in the probability and
consequences of the undesired events. It is often estimated by the mathematical expectation of
the consequences of an undesired event. Then it is the product "probability consequences".
However, a more general interpretation of the risk involves probability and consequences in a
non-product form. This presentation is sometimes useful, particularly when a spectrum of
consequences, with each magnitude having its own probability of occurrence, is considered [2].
The estimation of risk is the process used to produce an estimate of a measure of risk.
As already stated above the risk estimation is based on the hazard identification and generally
contains the following steps: scope definition, frequency analysis, consequence analysis, and
their integration [2]. If there is one-to-one mapping between the consequences Cij,k and the
events Eij, then the risk component Rk related to the considered situations Hi is the sum
Rk =

ij , k

P{Eij | H i }P{H i }

(3)

ij

If the dependence of consequences on events is more complicated than just one-to-one

mapping, then equation (3) will have to be modified. A practical example of equation (3) can
be found in [10], where an attempt to estimate the risk due to persistent and fire design
situation is presented.
In some cases it is possible to deal with one-component risk R only. Then the subscript
k in equation (2.3) may be omitted. Moreover, probability of undesired events may depend on
the vector of basic variables X. Then the total risk R may be formally written as
R = C (x)f X (x)dx

(4)

where R(x) denotes the degree of risk as a function of basic variables X, and fX(x) denotes joint
probability density function X.
7

LOGIC TREES

A number of different logic (decision) trees (fault tree, event tree, cause/consequence
chart) have been developed to analyse the risk of a system [11] to [13]. Applications of logic
trees significantly improve the completeness and clarity of the engineering work. The use of
this kind of tool is widespread in risk analysis and implies some important advantages.
Influences of the environment and of human activities can easily be considered simultaneously.
Logic trees can also enable the detection of the most effective countermeasures. Furthermore,
they can be easily understood by inexperienced persons and therefore can provide very
effective communication means between experts and public authorities.
The fault tree can be defined as a logical diagram for the representation of combinations
of influences that can lead to an undesired event. When establishing a fault tree, the undesired
VI - 5

Chapter VI - Principles of risk assessment

event constitutes the starting point. Going out from this event, possible causes are to be
identified. The possible causes and consequences are to be linked in a logic way, without
introducing any loops. Every event that is not a consequence of the previous event has to be
considered as an independent variable.
An example of the fault tree shown in Figure 3 describes the failure of a plane frame
(indicated at the bottom of Figure 3).
Structural failure

OR

H
M1

M2

Sway

Vertical

Combined

AND

AND

AND

2
M2

M4
Sway
2b

M5

4
G+Q

M3
Vertical

M4 H

G+Q

M4

M3
M1

Combined

M5

Figure 3. Fault tree describing the failure of a plane frame.

Fault trees can be used to clarify the causes of failures in case that they are unknown.
The most common application, however, consists in detecting possible causes of undesirable
events before they can occur. Since the fault trees also show the possible consequences of
events, they are very useful for the establishment of the most accurate measures for prevention
of these events.
An event tree identifies possible subsequent events starting from an initial event. Each
path consists of a sequence of events and ends up at the consequence level (for example at
structural failure, see Figure 4). The aim of the event tree analysis is to identify possible
consequences of an initial event and to calculate probabilities of the occurrence of these
consequences corresponding to a different sequence of events.
Simple examples of an event tree describing the collapse of a structure under persistent
and fire (accidental) design situation is shown in Figure 4. The probabilities indicated in Figure
4 are illustrative values only (correspond approximately to a 50-year period of an
administrative building having the fire compartment area 250 m2 with a protected steel
structure and without sprinklers).

VI - 6

Chapter VI - Principles of risk assessment

Fire starts

Fire flashover

Collapse
Yes
0,01

Yes
0,05

Frequency per 50 years

0,00005

Yes

No
0,99

0,00495

0,1

Yes

0,00000095

0,00001

No
0,95

No

0,09499905

0,99999
Design situation
Yes

0,000009

0,00001

No
0,9

No

0,899991

0,99999

Figure 4. Event tree describing the collapse of a structure under persistent and fire design
situation (all data are approximately related to a 50-year period of an administrative building
having the fire compartment area 250 m2 without sprinklers and with a protected steel
structure).
Logic trees may be supplemented by the consequences of events; graphical
representation of such a tree is called the cause/consequence-chart. The consequence chart
corresponds to an event tree with a suitable representation of expected consequences. For
example Figure 4 may include consequences linked to each failure probability (frequency per
year) of the structural collapse under given conditions. Then the tree may be used for the
cause/consequence or risk (utility) analysis.
The simplest form of the cause/consequence consideration is the so-called prior-analysis
of the risk (utility) when the basic statistical and probabilistic information is available prior to
any decision or activity. The prior analysis is an assessment of the risk associated with different
decisions; commonly used for comparing the risks corresponding to different decisions. The
posterior decision analysis differs from the prior analysis by considering possible changes in
the branching probabilities and/or the consequences due to risk reducing measures, risk
mitigating measures and the collection of additional information. The posterior decision
analysis may be used to evaluate different additional activities affecting the total risk.
Other important modification of logic trees is known as the pre-posterior decision
analysis. The aim of the pre-posterior decision analysis is to identify the optimal decisions with
regard to activities that may be performed in the future, e.g. planning of risk reducing activities
and/or the collection of new information. An important pre-requisite for the pre-posterior
decision analysis is the consideration of future actions that may be applied taking into account
the results of the planned activities.

VI - 7

Chapter VI - Principles of risk assessment

BAYESIAN NETWORK

Another promising tool for the risk analysis seem to be Bayesian (believe) causal
networks [13,14]. A simple example of the causal network is shown in Figure 5. The network
containing only four chance nodes describes the structural failure under persistent and fire
design situation similarly as the event tree in Figure 4. Compared with the event tree shown in
Figure 4 the network in Figure 5 includes also the effect of sprinklers (node B). Note that the
directional arrows in Figure 5 indicate the causal links between interconnected chance nodes.
B - Sprinkler

D-Flashover

A-Situation

E-Collapse

Figure 5. The causal network describing the structural failure under persistent and fire design
situation.
The collapse of a structure depends on the probability of persistent and fire situation
and on the conditional probabilities of full development of fire, that depends on the ability of
sprinklers and on conditional probability of the structural collapse under the conditions given
by parents nodes (for example when fire is fully developed - fire flash over). Obviously the
causal network representation seems to be much more effective than the event tree version.
Moreover each node may have several states. Consequently, the input data are not indicated
directly in the graphical representation of the network but are given in the tables of conditional
probabilities.
The basic principle of probability calculation used in the Bayesian networks may be
illustrated considering the nodes A, B and D of the network in Figure 5. One child node D (Fire
flashover) is dependent on two parent nodes: A (Design situation) and B (Sprinklers). If the
parents' nodes A and B have the discrete states Ai and Bj, then the probability of the event Dk (a
particular state of the node D) is given by the formula
P(Dk) = P(Dk| AiBj)P(Ai) P(Bj)

(5)

Equation (5) represents the fundamental theoretical tool for analysing the Bayesian
network. The input data consist of the probabilities P(Ai) and P(Bj), and the conditional
probabilities P(Dk|AiBj). These extensive data are based on available statistical evidence,
probabilistic analysis or expert assessment (judgement) and are transparently summarised in
the tables of conditional probabilities.
Bayesian networks supplemented by decision and utility nodes called influence
diagrams [13,14] provide a powerful tool for the risk estimation. In fact the influence diagram
is a generalisation of the cause/consequence-chart discussed above. The main features of this
tool are illustrated by the example shown in Figure 6, which is an extension of the fundamental
task indicated in Figure 5. Figure 6 shows a simplified influence diagram developed recently
[15,16] for the risk analysis of buildings under persistent and fire design situation.

VI - 8

Chapter VI - Principles of risk assessment

The network consists of seven chance nodes numbered 1, 2, 3, 4, 5, 12 and 14, four
decision nodes 6, 7, 15 and 16, and six utility nodes 8, 9, 10, 11, 13 and 17. The utility nodes
represent the costs of various fire safety measures (nodes 8, 10, 17), damage to the building
(nodes 9, 11), and injuries (node 13).
Directional arrows indicating the causal links between the parent and children nodes
interconnect the chance, decision and utility nodes. All the causal links must be described by
appropriate input data (conditional probabilities or utility units) linked to assumed states of the
nodes. For example the utility nodes (except the utility node 13) are directly dependent on the
size of the building (node 15). The utility node 13, describing the cost of injury, is affected by
the size of the building through the number of endangered persons represented by chance node
14. These data are sometimes difficult to specify, and an expert assessment has often to be
often.

1-Situation

8-C8

6-Sprinklers

10-C10

7-Protection

17-C17

2-Sprinklers

4-Protection

3-Flashover

5-Collapse

15-Size

12-Smoke

16-Escape r.

14-Number

9-C9

11-C11

13-C13

Figure 6. The Bayesian network describing a structure under normal and fire design situations.
The chance nodes 1, 2, 3, 4, 5, 12 and 14 represent alternative random variables having
two or more states. The node 1-Situation describes the probability of fire start pfi,s = P(H2) and
the complementary probability 1 pfi,s of normal situation H1. The chance node 2-Sprinklers
describes the functioning of sprinklers provided that the decision (node 6) is positive; the
probability of the active state of the sprinklers given the fire start is assumed to be very high,
for example 0,999. The chance node 3-Flashover has two states: the design situation H3 (fire
design situation without flashover) and H4 (fire design situation with flashover when the fire is
fully developed).
When sprinklers are installed, the flashover in a compartment of 250 m2 has the positive
state with the conditional probability 0,002; if sprinklers are not installed then P{H4|H2} =
0,066 [15,16]. It is assumed that with the probabilities equal to the squares of the above values
the fire will flash over the whole building, thus the values 0,000004 and 0,0044 are considered
for the chance node 3. The chance node 4-Protection (introduced for formal computational
reasons) has identical states as the decision node 7-Protection. The chance node 5-Collapse
represents the structural failure that is described by the probability distribution linked to three
children nodes (1,3,4). This situation can hardly be modelled using a decision tree. Note that
the probability of collapse in the case of fire but not flashover may be smaller than in a
persistent situation, due to the lower imposed load.

VI - 9

Chapter VI - Principles of risk assessment

9

DECISION-MAKING

The decision-making is generally based on the process of the risk acceptance and option
analysis (see Figure 1) that is sometimes referred to as the risk evaluation. The risk acceptance
is based on various criteria of risk that are reference points against which the results of the risk
analysis are to be assessed. The criteria are generally based on regulations, standards,
experience, and/or theoretical knowledge used as a basis for the decision about the acceptable
risk. Acceptance criteria and the criteria of risk may be sometimes distinguished [1]. Various
aspects may be considered, including cultural, social, psychological, economical and other
aspect [6], [17], [18] and [19]. Generally the acceptance criteria may be expressed verbally or
numerically [6].
Assuming for example that the acceptance limits Ck,d for the components Ck are
specified, then it is possible to design the structure on the basis of acceptable risks using the
criterion Ck < Ck,d, which may supplement the probability requirement pF < pt.
It should be noted that various levels of risk might be recognized, for example
acceptable risk, tolerable risk, and objective risk [6] (see the definitions of theses terms in [2]).
It is a remarkable fact that the public seems to be generally better prepared to accept certain
risks than to stand for specified probabilities of failure [17].
10

THE IMPLIED COST OF AVERTING A FATALITY

The consequences may generally include economic as well as social and environmental
costs [17,18,19]. An example is provided by the influence diagram shown in Figure 6 used to
assess the risk of a building due to fire. Thus, in order to compare all possible consequences it
is necessary to express all consequences in terms of a single unit. This seems to be an
extremely difficult task. One of the possible approaches is represented by the concept of the
Implied Cost of Averting a Fatality ICAF or Life Quality index LQI [19]. Table 1 shows values
of the cost ICAF for selected countries adopted from [19]. It appears that the cost ICAF may be
estimated to about 1 to 3 million of USD.
Table 1. The Implied Cost of Averting a Fatality ICAF(e)), financial data in PPP US$ (1999)
obtained from UN-HDR 2001, World Bank.
Country
US
Japan
Germany
UK
Czech Republic
Mexico
South Africa
Colombia
China
India
Nigeria

g- annual
income

e- life
time

34000
26000
25000
22000
8000
8800
9100
5900
3900
2400
800

VI - 10

77
81
77
77
75
72
55
70
70
63
47

2 wICAF(e)
working
[ 106]
part of e
0.15
2.6
0.15
2.1
0.125
1.9
0.125
1.7
0.15
0.6
0.15
0.6
0.15
0.5
0.15
0.4
0.15
0.3
0.15
0.1
0.18
0.04

Chapter VI - Principles of risk assessment

The Implied Cost of Averting a Fatality ICAF can be expressed as
1

e 1 w

(6)
ICAF (e) = g 1 (1 + ) e
e

where symbols g, e and w are defined in Table 1. However, the concept of the Implied
Cost of Averting a Fatality described by equation (6) is just one of possible approaches to the
complex problem of evaluating social consequences. At present further intensive investigation
is expected.

11

CONCLUDING REMARKS

Risk is commonly estimated by the mathematical expectation of the consequences of an

undesired event that often leads to the product "probability consequences". As a rule the risk
of civil engineering systems is a multidimensional quantity having several components.
The risk analysis is based on the hazard identification and generally contains the
following steps: the scope definition, hazard identification, definition and modelling of hazard
scenarios, estimation of probabilities, estimation of consequences, estimation of risk and
decision-making.
The most important contribution of the risk analysis and assessment consists in the
systematic consideration of various consequences. Several techniques are available at present:
the decision trees, the Bayesian belief networks and influence diagrams. Available experience
indicates that the Bayesian belief networks provide a transparent, logical and effective tool for
analysing engineering systems. It should however be underlined that any analysis of an
engineering system is always dependent on the assumed input data, often of a very uncertain
nature. The input data should be estimated with due regard to the specific technological and
economic conditions of a given system. In particular, the economic, social and environmental
consequences of adverse events should be further investigated.
It appears that the methods of risk analysis and assessment may significantly contribute
to further improvement of current engineering design. The remarkable fact that the public is
better prepared to accept certain risks than to stand for specified probabilities of failure will
make the application of the risk assessment easier. It is therefore anticipated that in the near
future probabilistic methods in engineering will be supplemented by criteria for acceptable
risks.
REFERENCES
[1] NS 5814, Requirements for risk analysis. 1991.
[2] CAN/CSA - Q634-91 Risk analysis requirements and guidelines.1991.
[3] ISO 2394 General principles on reliability for structures. 1998.
[4] ISO/DIS 8930 General principles on reliability of structures - List of equivalent terms.
1999.
[5] TNO report, 96-CON-R1599 Proposal for a framework in behalf of developing terminology
with regard to the process of the probabilistic design and/or assessment of building and
civil engineering structures with reference to ISO 8930. 1996.
[6] CIB TG 32, Report 259 Risk assessment and risk communication in civil engineering, CIB
secretariat 2001.
[7] ISO/IEC Guide 73: Risk management Vocabulary - Guidelines for use in standards, 2002.
[8] ISO 9000: Quality management systems Fundamentals and vocabulary, 2000.

VI - 11

Chapter VI - Principles of risk assessment

[9] ISO/IEC Guide 51: Safety aspects Guidelines for their inclusion in standards, 1999.
[10] EN 1990: Basis of design. CEN TC 250, Draft, April 2002.
[11] Schneider, J.: Introduction to safety and reliability of structures. Structural Engineering
Documents n 5, International Association for Bridge and Structural Engineering, IABSE,
Zurich, 1997.
[12] M.G.Steward and R.E.Melchers, Probabilistic risk assessment of engineering system.
Chapman & Hall, London 1997.
[13] Jensen Finn V. (1999) Introduction to Bayesian networks. Aalborg University, Denmark,
1996.
[14] Hugin system: Version 5.7, professional. Hugin Expert A/S, Niels Jernes Vej 10, DK-9220
Aalborg, Denmark, 2001.
[15] Holick M. & Schleich J.-B.: Modelling of a Structure under Permanent and Fire Design
Situation. Proc. of Safety, Risk and Reliability - Trends in Engineering. International
Conference, Malta, 21/23.3.01, A.A. Balkema, Rotterdam, 2001, pp. 789-794.
[16] Holick M.: Risk assessment of steel buildings and occupants under fire situation. ICASP
9, Berkeley, 2003, pp. 163- 168.
[17] Lewis R.: The Public Perception of Risk, RSA Journal, November 1995, pp 52-63.
[18] Schneider J.: Safety - A Matter of Risk, Cost and Consensus. Structural Engineering
International. No. 4, November 2000, pp. 266-269.
[19] Rackwitz R.: New LQI-developments, JCSS Workshop on Reliability Based Calibration,
Zrich, 2002.
NOTATION
Ai
States of node A
Bj
States of node B
Dk States of node D
Cij Consequences of events Eij (utility, cost, damage, injuries)
R
The total expected risk
Rk The risk component
Eij Events
Hi Hazard situation i.
P(F/Hi) Probability of failure F given situation Hi
e
Expected life-time
g
Annual income
w
Working part of e
g(x) Performance (limit state) function.
pF Probability of failure F.
Target probability of failure.
pd
pf
Probability P(F|H2) of structural failure during fire.
pfi,s Probability of fire start P(H2).
x
Generic point of the vector of basic variables.
X
Vector of basic variables.

Reliability index.
X(x) Probability density function of the vector of basic variables X.
N1 ( p F ) Inverse distribution function of a standardized normal variable.

VI - 12

Annex A - Basic statistical concepts and techniques

ANNEX A - BASIC STATISTICAL CONCEPTS AND TECHNIQUES

Milan Holick
Klokner Institute, Czech Technical University in Prague, Czech Republic

Summary
Elementary concepts and techniques of the theory of probability and mathematical
statistics required for understanding of basic reliability methods are reviewed and illustrated
by a number of numerical examples. Computational procedures for determination of sample
characteristics, fractiles of common theoretical models and estimates for fractiles based on
small samples can be applied using the attached MATHCAD sheets.
1

INTRODUCTION

1.1

Background materials
Elementary concepts and techniques of the theory of probability and mathematical
statistics applicable to civil engineering are available in a number of standards [1 to 5],
background materials [6, 7, 8], software products [9, 10, 11] and books [12 to 24]. Additional
information may be found in the extensive literature listed in the books [12, 13] and others. In
particular, documents developed by JCSS [6, 7] and recently published handbook [8] are
closely related to the statistical techniques described in this text.
1.2

General principles
The theory of structural reliability is based on a general principle that all the basic
variables are considered as random variables having appropriate type of probability
distribution. Different types of distributions should be used for description of actions, material
properties and geometric data. Prior theoretical models of basic variables and procedures for
probabilistic analysis are indicated in JCSS documents. Sample characteristics are used as
estimates of population parameters. In addition the population fractiles must be often assessed
using small samples. MATHCAD sheets that supplement described computational procedures
can be effectively used in practical applications.
2

POPULATION AND SAMPLES

2.1

General
Actions, mechanical properties and geometric data are generally described by random
variables (mainly by continuous variables). A random variable X, (e.g. concrete strength), is
such a variable, which may take each of the values of a specified set of values (e.g. any value
from a given interval), with a known or estimated probability. As a rule, only a limited
number of observations, constituting a random sample x1, x2, x3,..., xn of size n taken from a
population, is available for a variable X. Population is a general statistical term used for the
totality of units under consideration, e.g. for all concrete produced under specified conditions
within a certain period of time. The aim of statistical methods is to make decisions concerning

A-1

Annex A - Basic statistical concepts and techniques

the properties of the population using the information derived from one or more random
samples.
2.2

Sample characteristics
A sample characteristic is a quantity used to describe the basic properties of a sample.
The three basic sample characteristics, which are most commonly used in practical
applications, are:
- the mean m representing the basic measure of central tendency;
- the variance s2 describing the basic measure of dispersion; and
- the coefficient of skewness giving the basic measure of asymmetry.
The sample mean m (an estimate of the population mean) is defined as the sum
m = ( xi) / n

(2.1)

with the summation being extended over all the n values of xi.
The sample variance s2 (an estimate of the population variance), is defined as:
s2 = ( (xi - m)2) / (n - 1)

(2.2)

the summation being again extended over all values xi. Sample standard deviation s is the
positive square root of the variance s2.
The sample coefficient of skewness (an estimate of the population skewness)
characterising asymmetry of the distribution is defined as

= [n ( (xi - m)3) / (n-1) / (n-2)] / s3

(2.3)

Thus, the coefficient of skewness is derived from the central moment of order 3
divided by s3. If the sample has more distant values to the right from the mean than to the left,
the distribution is said to be skewed to the right or to have a positive skewness. If the reverse
is true, it is said to be skewed to the left or to have a negative skewness.
In some cases two different samples may be taken from one population and their
combination is needed. If the original data are not available, then the characteristics of
combined sample may be determined using the characteristics of both samples. If the sample
sizes are n1, n2, the means m1, m2, standard deviations s1, s2 and skewnesses 1, 2, then the
combined sample of the size n = n1+n2 has the characteristics
m=
s2 =

n1m1 + n2 m2
n

n1s12 + n2 s22 n1n2

+ 2 (m1 m2 ) 2
n
n

(2.4)

1 n1s131 + n2 s23 2 3n1n2 (m1 m2 )( s13 s23 ) n1n2 (n1 n2 )(m1 m2 )3

+

s 3
n
n2
n3

Another important characteristic describing the relative dispersion of a sample is the

coefficient of variation v, defined as the ratio of standard deviation s to the mean m
v=s/m

(2.5)

The coefficient of variation v can be effectively used only if the mean m differs from
zero. When the mean is much less than the standard deviation, then the standard deviation
rather then the coefficient of variation should be considered as a measure of the dispersion.
The coefficient of variation v is often used as a measure of production quality; for concrete

A-2

Annex A - Basic statistical concepts and techniques

strength may be expected within a broad range from 0,05 up to 0,20, for structural steel from
0,07 to 0,10.
2.3

Distribution function
Probability distribution is a term generally used for any function giving the probability
that a variable X belongs to a given set of values. The basic theoretical models used to
describe the probability distribution of a random variable may be obtained from a random
sample by increasing the sample size or by smoothing either the frequency distribution or the
cumulative frequency polygon.
An idealisation of a cumulative frequency polygon is the distribution function (x)
giving, for each value x, the probability that the variable X is less than or equal to x:
(x) = P (X x)

(2.6)

A probability density function (x) is an idealisation of a relative frequency

distribution. It is formally defined as the derivative (when it exists) of the distribution
function:
(x) = d(x) / dx

(2.7)

Note that Appendix 1 to this Chapter provides a review of selected theoretical models
of continuous random variables that are most frequently used in reliability analysis of civil
structures.
Example 2.1.
A continuous random variable, which may attain equally likely any point x within a
two-sided interval <a, b> (each point x has the same probability density (x)) is described by
a so-called uniform distribution shown in Figure 2.1.
1
(x) = (x-a)/(b-a)
Distribution function

x
a

(x) = 1/(b-a)

Probability density
function
x

Figure 2.1. Uniform distribution

The uniform distribution is a basic type of distribution used not only in simulation
procedures but also in theoretical modelling of some actions and geometric data. Shapes of
the distribution function (x) and probability density function (x) for the uniform
distribution are shown in Figure 2.1. We can easily observe that it is a general property of the

A-3

Annex A - Basic statistical concepts and techniques

probability density function that the probability of a set of all values of any random variable is
equal to 1

(x)dx =

(x)dx = 1

(2.8)

Thus, the surface bounded by the horizontal axis x and the curve of the density function (x)
has the area equal to unity.
2.4

Population parameters
The population parameters are quantities used in describing the distribution of a
random variable, as estimated from one or more samples. As in the case of random samples,
three basic population parameters are commonly used in practical applications:
- the mean representing the basic measure of central tendency;
- the variance 2 as the basic measure of dispersion; and
- the coefficient of skewness .giving the degree of asymmetry.
The population mean , for a continuous variable X having the probability density
(x), is defined as
=

x (x)dx

(2.9)

the integral being extended over the interval of variation of the variable X. The population
variance 2, for a continuous variable X having the probability density function (x), is the
mean of the squared deviation of the variable from its mean:

2 = (x - )2 (x)dx

(2.10)

The population standard deviation is the positive square root of the population
variance 2.
The population coefficient of skewness, characterising asymmetry of the distribution,
is defined as

= (x - )3 (x)dx / 3

(2.11)

Another population parameter based on the fourth order moment is called kurtosis .

= (x - )4 (x)dx / 4 - 3

(2.12)

Note that for normal distribution (described in Section 3.1) the kurtosis defined by
equation (2.12) is zero. However, this parameter is used mainly in theoretical consideration.
Another important parameter of the population is the coefficient of variation V defined
similarly as the sample coefficient of variation
V=/

(2.13)

The same restriction on the practical use of V applies as in the case of samples.
Geometrically is actually the x coordinate of the centre of gravity of the area
bounded by the horizontal axis x and the curve of density function (x). Figure 2.2 shows an
example of probability density function of lognormal distribution illustrating the geometric
interpretation of the mean and standard deviation .
The measure of dispersion of a random variable X relative to the mean is given by
the central moment of the second order (moment of inertia) of the area, and standard deviation

A-4

Annex A - Basic statistical concepts and techniques

is therefore the centroidal radius of gyration around the mean of the area bounded by the

horizontal axis x and the curve of probability density function (x).

A very important population characteristic is the fractile xp. If X is a continuous
variable and p is a probability (a real number between 0 and 1), the p-fractile xp is the value of
the variable X for which the probability that the variable X is less than or equal to xp is p, and
hence, for which the distribution function (xp) is equal to p. Thus,
P (X xp) = (xp) = p

(2.14)

In civil engineering the probabilities p = 0,001; 0,01; 0,05 and 0,10 are used most
frequently. The probability p is often written as a percentage (e.g. p = 0,1 %; 1 %; 5 %; 10
%). If this is done, then xp is called a percentile, for example the 5th percentile is used when p
= 5 %. If p=50 %, then xp is called the median. More details about the fractiles of continuous
variables are given in the following sections.
Probability density (x)
2.5

2.0

1.5

=1.0
1.0

0.5

0.0
0.4

0.6

0.8

1.0

1.2

1.4

1.6

x
Figure 2.2. Geometric illustration of the mean and standard deviation

1.8

Example 2.2.
Parameters of the uniform distribution from example 2.1 may be derived using
equations (2.9) to (2.13) as

= (a+b)/2, = (b-a)/ 12 , = 0, = - 2,96, V = (b-a)/((a+b) 3 )

The skewness of a uniform distribution is zero, kurtosis is negative (independent of the
bounds a and b). Obviously the distribution is symmetric as the values of the random variable
are distributed uniformly. If the lower bound of the distribution is zero, a = 0 (which is
sometimes assumed in practical applications), then

= 0,5 b, = 0,289 b, = 0, = - 2,96, V = 0,577

Let us note that the coefficient of variation V in this case (when a = 0) is independent
of b and its value is relatively high (V = 0,577).

A-5

Annex A - Basic statistical concepts and techniques

3

SELECTED MODELS OF RANDOM VARIABLES

3.1

Normal distribution
Most frequently used models of continuous random variables that are applied in
reliability analysis of civil structures are reviewed in Appendix 1 of this Chapter. From a
practical and theoretical point of view the most important type of distribution of a continuous
random variable is the normal (Laplace-Gauss) distribution. Symmetric normal distribution of
a variable X is defined on an unlimited interval - < x < (which can be undesirable in some
practical applications) and depends on two parameters only on the mean and on the
standard deviation . Symbolically it is often denoted as N(,).
The normal distribution is frequently used as a theoretical model of various types of
random variables describing some loads (self-weight), mechanical properties (strengths) and
geometrical properties (outer dimensions). It is convenient for symmetric random variable
with a relatively low variance (coefficient of variation V < 0,3). It fails when used for
asymmetric variables with great variance and skewness > 0,5.
The probability density function of a normal random variable X with a mean and
standard deviation is given by the exponential expression
1 x 2
1
(x) =
exp

2
2

(3.1)

Skewness and kurtosis are zero for a normal distribution.

Tables for normal distributions are commonly available [12, 13] for probability
density function (u) and distribution function (u) of a standardized variable U, which is
defined by a general transformation relation (used for any type of distribution)
U=

(3.2)

The standardized random variable U has a zero mean and variance (standard
deviation) equal to one; symbolically it is often denoted as N(0, 1).
The probability density function of the standardized random variable U is then given
as a function of u
(u) =

1
2
exp u
2
2

(3.3)

The probability density function of a normal and lognormal distribution with a

coefficient of skewness = 1,0 (described in the next section 3.2) of the standardized random
variable u is shown in Figure 3.1.
Note that the probability density function of the standardized normal distribution is
plotted in Figure 3.1 for u in the interval <3,+3>, which covers the standardised variable U
with a high probability of 0,9973 (in engineering practice this interval is often called interval
3).
3.2

Lognormal distribution
Generally one-sided limited asymmetric lognormal distribution is defined on a limited
interval x0 < x < or - < x < x0. Therefore it eliminates one of the undesirable properties of
the normal distribution. A lognormal distribution is generally dependent on three parameters.
Commonly the moment parameters are used: mean X, standard deviation X and skewness
x. If the skewness x is unknown or uncertain, the lower or upper bound x0 is used.
A-6

Annex A - Basic statistical concepts and techniques

Random variable X has a lognormal (general three-parametric) distribution if the
transformed random variable
Y = ln |X x0|

(3.4)

has a normal distribution. In this relation x0 denotes the lower or upper limit of distribution of
a variable X, which depends on skewness x. If the variable has a mean x and standard
deviation x, then the lower or upper limit can be expressed as
x0 = X - X /c

(3.5)

where the coefficient c is given by the value of skewness X according to the relation

X = c3 + 3c3

(3.6)

from which follows an explicit relation for c

c = 2X + 4 + x

) (
13

2
X

+ 4 X

13

1 3
2

(3.7)

Probability density (x)

0.5

Lognormal distribution, = 1,0

0.4
0.3

Normal distribution N(0,1)

0.2
0.1
0.0
-3

-2

-1

Standardized random variable u

Figure 3.1. Normal and lognormal distribution (skewness = 1,0)
The dependence of the limit x0 on the coefficient c is obvious from Table 3.1 in which
the lower bound u0= 1/c of the standardised random variable U=(X- X)/X are given for
selected values of the coefficient of skewness X 0. For X 0 values of u0 with an inverse
sign (i.e. positive) are considered. A lognormal distribution with the skewness x = 0 becomes
a normal distribution (u0= 1/c ).
Table 3.1. The lower limit u0= 1/c for selected values of coefficient of skewness X 0.
X

0,5

1,0

1,5

2,0

u0= 1/c

-6,05

-3,10

-2,14

-1,68

A-7

Annex A - Basic statistical concepts and techniques

When creating a theoretical model it is therefore possible to consider, besides the

mean X and standard deviation X, the skewness X or alternatively the lower or upper bound
of distribution x0. Generally the former possibility is preferred because more credible
information is available about the coefficient of skewness, which better characterises the
overall distribution of the population (particularly of large populations) compared to the lower
or upper bounds.
The probability density function and distribution function of the general three
parameter lognormal distribution may be obtained from well known normal distribution using
modified (transformed) standardised variable u obtained from the original standardised
random variable u = (x-X)/X as

1
ln u + + ln c 1 + c 2
c
u =
ln(1 + c 2 )

)
(3.8)

where (as above) u = (x-X)/X denotes the original standardised variable. The probability
density function LN,U(u) and the distribution function LN,U(u) = LN,X(x) of the lognormal
distribution are then given as
LN,U(u) =

(u )

1
u + ln(1 + c 2 )
c

LN,X(x) = LN,U(u) = (u)

(3.9)

(3.10)

where (u) and (u) denote the probability density and distribution function of the
standardised normal variable.
A special case is the popular lognormal distribution with a lower bound at zero (x0 =
0), which like the normal distribution, depends on two parameters only the mean X and the
standard deviation X (symbolically it is denoted LN(, )). In such a case it follows from
equations (3.5) that the coefficient c is equal to the coefficient of variation VX. It further
follows from equation (3.6) that the skewness X of the lognormal distribution with a lower
bound at zero is given by the value of the coefficient of variation VX as

X = 3V X + V 3X

(3.11)

Thus the lognormal distribution with the lower bound at zero (x0 = 0) always has a
positive skewness, which may have relatively high value (greater than 0,5); e.g. for the
coefficient of variation equal to 0,30 a coefficient of skewness Vx = 0,927 obtained from
relation (3.11). Applications of the lognormal distribution with the lower limit at zero (xo = 0)
can thus lead to unrealistic theoretical models (usually underestimating the occurrence of
negative and overestimating the occurrence of positive deviations from the mean), particularly
for higher values of coefficient of variation VX. Although the occurrence of negative values
can also be undesirable (unrealistic for most mechanical quantities), it is usually negligible
from a practical point of view.
Example 3.1.
Reinforcement cover layer of a reinforced concrete cross-section X has a mean =
25 mm and standard deviation = 10 mm. The probability density function (x) for a normal
distribution and for a lognormal distribution with a lower limit at zero is shown in Figure 3.2.
A-8

Annex A - Basic statistical concepts and techniques

It follows from Figure 3.2 that the normal distribution leads to occurrence of negative
values of the reinforcement cover layer, which obviously does not correspond to reality. On
the other hand, the lognormal distribution with lower limit at zero overestimates the
occurrence of positive deviations of the cover layer, which may not be realistic either and can
further lead to unfavourable influences on the strength of the cross-section. The
overestimation of occurrence of extreme positive deviations corresponds to a high skewness
= 1,36 of the lognormal distribution, which follows from equation (3.11). The available
experimental data on the concrete cover indicate that the skewness of the distribution is
around 0,5, in most cases < 1,0.
Probability density (x)
0.05

Log-normal distribution LN(25,10)

0.04
0.03

Normal distribution N(25,10)

0.02
0.01
0.00
-10

10

20

30

40

50

60

70

80

90

Concrete cover X [mm]

Figure 3.2. Probability density function for the concrete cover

The lognormal distribution is widely applied in the theory of reliability. It is used as a
model for various types of random variables describing some loads (self-weight of some
materials), mechanical properties (strengths) as well as geometrical data (inner and outer
dimensions of cross-sections). It can be used for general asymmetric random variables with
both positive and negative skewness. The lognormal distribution with lower limit at zero (x0 =
0) is very often used for description of mechanical properties (strengths) of various materials
(concrete, steel, masonry).
3.3

Gamma distribution
Another popular type of one-side limited distribution is the type III Pearson
distribution. Its detailed description is e.g. in the book [13]. A special case of the type III
Pearson distribution with lower limit at zero is the gamma distribution. The probability
density function of this important distribution is dependent on two parameters only: on the
mean and standard deviation . To simplify the notation two auxiliary parameters and
k are often used
( x) = x
k

exp(x)

, = 2 ,k =
(k )

k 1

A-9

(3.12)

Annex A - Basic statistical concepts and techniques

(k ) is the gamma function of parameter k. For the moment parameters of the gamma
distribution it holds that

, =

, =

2
2
=
= 2V

(3.13)

The curve is bell shaped for k > 1, i.e. for skewness < 2 (in the inverse case it is a
decreasing function of x). For k , the gamma distribution approaches the normal
distribution with parameters and .
The gamma distribution is applied similarly as the lognormal distribution with lower
bound at zero. However, it varies from the lognormal distribution by its skewness, which is
equal to twice the coefficient of variation ( = 2V) and is thus lower than the skewness of
lognormal distribution, which is more than 50% higher (according to equation (3.11) it is
X = 3VX + VX3 ). That is the reason why the gamma distribution is more convenient for
describing some geometrical quantities and variable action that do not have a great skewness.
Probability density (x)
0.05
n = 157
m = 26,8
s = 11,1
v = 0,42
a = 0,40

0.04

Log-normal distribution
Gamma distribution
Normal distribution
Beta distribution

0.03

0.02

0.01

0.00

10

20
30
40
Concrete cover [mm]

50

60

70

Figure 3.3. Histogram and theoretical models for concrete cover of reinforcement
Example 3.2.
A sample of the size n = 157 experimental results of concrete cover of reinforcement
measurements has these characteristics: m = 26,8 mm, s = 11,1 mm and v = 0,42. It is a
relatively large sample, which can be used for the assessed skewness (furthermore a long-term
experience is available). A histogram of the obtained values and theoretical models of normal
distribution, lognormal distribution with origin at zero, gamma distribution and beta
distribution are shown in Figure 3.3, with help of which the appropriateness of the individual
models can be considered.

A - 10

Annex A - Basic statistical concepts and techniques

According to Figure 3.3 it seems that the gamma distribution describes the histogram
of obtained results better than the normal and lognormal distribution. But also the both-side
limited beta distribution (described in the following Section 3.4) seems to be an appropriate
model. However, to choose an appropriate theoretical model for describing variables of
interest is a complicated task, which can be treated in theoretical way. Information about
some methods of mathematical statistics (about the so-called goodness of fit tests) can be
found in the textbook [4] and in specialised literature [12, 13]. In this book
some practical aspects and procedures will be indicated only.
3.4

Beta distribution
An interesting type of distribution is the so-called beta distribution (also called
Pearsons type I curve), which is defined on a both-side limited interval <a, b > (but this
interval can be arbitrarily extended and the distribution then approaches the normal
distribution). Generally it is dependent on four parameters and it is used mainly in those cases
when it is evident that the domain of the random variable is limited on both sides (some
actions and geometrical data, e.g. weight of a subway car, fire load intensity, concrete cover
of reinforcement in a reinforced concrete cross-section). The principal difficulty in practical
application is the need to estimate all the four parameters, for which credible data may not be
available.
The beta distribution is usually written in the form
( x) =

( x a )c 1 ( x b) d 1
B(c, d )(b a )c + d 1

(3.14)

For the lower and upper limit of distribution it holds

a = - c g , b = + d g , g =

c + d +1
cd

(3.15)

where g is an auxiliary parameter. From equations (3.15), relations for parameters c and d can
be derived
c=

a ( a )(b )
b ( a )(b )
1 , d =
1

2
ba
ba

(3.16)

For the moment parameters of the beta distribution it holds that

a + (b a )c
(b a)
,=
(c + d )
(cg + dg )

(3.17)

2 g (d c)
3g 2 (2(c + d ) 2 + cd (c + d 6))
,=
3
(c + d + 2)
(c + d + 2)(c + d + 3)

(3.18)

=
=

Note that skewness and kurtosis are dependent only on the parameters c and d
(they are independent of the limits a and b). That is why the parameters c and d are called
shape parameters. In practical applications the distribution is used for c > 1 and d > 1
(otherwise the curve is J or U shaped), for c = d = 1 it becomes a uniform distribution, for c =
d = 2 it is the so-called parabolic distribution on the interval <a, b >. When c = d, the curve is
symmetric around the mean. When d , the curve becomes the type III Pearson
distribution (see Section 3.3). If c = d , it approaches the normal distribution. Depending

A - 11

Annex A - Basic statistical concepts and techniques

on the shape parameters c and d the beta distribution thus covers various special types of
distributions. The location of the distribution is given by parameters a and b.
The beta distribution can be defined in various ways. If the parameters a, b, c and d
are given, then the moment parameters , , and can be assess using equations (3.17) to
(3.18). In practical applications however, two other combinations of input parameters are
often applied:
1.
The input parameters are , , a and b. The remaining parameters c and d can
be assessed from equations (3.16), the parameters and from equations (3.18).
2.
The input parameters are , , and one of the limits a (for > 0) or b (for
< 0); the parameters b (or a), c and d can be assessed using equations (3.16) to (3.17).
In practical applications the distribution with lower limit a = 0 is often used. It can be
shown that in such a case the beta distribution is defined if

2V

(3.19)

where the coefficient of variation V = / . For = 2V the curve becomes the type III
Pearson distribution (see Section 3.3). Therefore if the input parameters are the mean ,
standard deviation and skewness 2V, the beta distribution with the lower limit at zero (a
= 0) is fully described. The upper limit b of the beta distribution with the lower limit at zero
follows from the relation (3.15)
b=

(c + d )
c

(1 + V (2 + V ))
(2V )

(3.20)

In equation (3.20) the parameters c and d are substituted by the following expressions
c=

d=

(2V ) 2 (4 + 2 )
2V (V + 2) 2 (4 + 2 )

(2V ) 2 (4 + 2 ) 2 + V
2 (V + 2) 2 (4 + 2 ) 2V

(3.21)

(3.22)

which follow from general equations (3.16) to (3.18) for a = 0.

Example 3.3.
Given the mean = 25 mm, standard deviation 10 mm (V = 0,40) and skewness
= 0,5, assess the parameters of a beta distribution with the lower bound at zero (a = 0) for a
reinforcement cover layer. Equation (3.19) is satisfied (0,5 < 2 0,4). From equations (3.21)
and (3.22) it follows that
c=

d=

0,5 (2 0,4 0,5) 2 (4 + 0,52 )

= 4,407
2 0,4 (0,4 0,5 + 2) 2 (4 + 0,52 )

0,5 (2 0,4 0,5) 2 (4 + 0,52 ) 2 + 0,5 0,4

= 12,927
2 (0,4 0,5 + 2) 2 (4 + 0,52 ) 0,5 2 0,4

The upper bound of the distribution b follows from equation (3.20) that
b=

25 (4,407 + 12,927)
= 98,325
4,407

A - 12

Annex A - Basic statistical concepts and techniques

The beta distribution having the assessed parameters is shown in Figure 3.4 together
with a corresponding normal, lognormal and Gamma distribution with the lower bound at
zero and the same mean and standard deviation .
Figure 3.4 further shows that the normal distribution (skewness = 0) leads to the
occurrence of negative values, which may not correspond to the real conditions for the
reinforcement cover layer. According to equation (3.11) the lognormal distribution with lower
limit at zero has skewness = 1,264, which does not correspond to experimental results and
leads to an overestimation of the occurrence of positive deviations (which may further lead to
unfavourable consequences in the reliability analysis of the reinforced concrete element).
The gamma distribution has, according to equation (3.13), a skewness = 2V = 0,8,
which is closer to the experimental value 0,5. The most convenient seems to be the beta
distribution having the skewness = 0,5 corresponding exactly to the experimental results.

Probability density (x)

0,05

Log-normal distribution LN(25;10)

a = 0, = 1,264;

0,04

Normal distribution N(25;10), = 0

Gamma distribution Gamma(25;10)
a = 0, = 0,8;

0,03

Beta distribution Beta(25;10)

a = 0, b = 98,3, = 0,5;

0,02
0,01
0,00
-10

10

20

30

40

50

60

70

80

Concrete cover x [mm]

Figure 3.4. Normal, Lognormal, Gamma and Beta distributions for the concrete cover layer of
reinforcement in a reinforced concrete element
It should be mentioned that mathematical statistics offers a number of goodness of fit
tests for evaluation of fitness of a distribution as a theoretical model for obtained
experimental results (see for example documents [4, 12, 13] and a number recently developed
ISO standards). The above discussion can therefore be supplemented by statistical tests. On
the other hand it is essential to remark that goodness of fit tests very often fail and do not lead
to an unambiguous result. In such a case the selection of a convenient model depends on the
character of the basic variable, on available experience and on common experience.
3.5

Gumbel and other distributions of extreme values

The extreme values (maximal or minimal) in a population of a certain size are random
variables and their distribution is very important in the theory of structural reliability. Three
types of extreme values distribution denoted as types I, II and III are usually covered in the
A - 13

Annex A - Basic statistical concepts and techniques

specialised literature. Each of the types has two versions one for the distribution of minimal
values, the second for maximal values distribution. All these types of distribution have a
simple exponential shape and are convenient to work with. We will describe in detail the type
I extreme value distribution, which is commonly called the Gumbel distribution. Description
of other types of distribution can be found in textbook [12,13] and in specialised literature
[15, 16, 17, 18, 19].
The distribution function for the type I maximal values distribution (Gumbel
distribution of maximum values) has the form
(x) = exp(-exp(-c(x - xmod)))

(3.23)

It is a distribution defined on an infinite interval, which depends on two parameters:

on mode xmod and parameter c > 0. By differentiating the distribution function we obtain the
probability density function in the form
(x) = c exp(-c (x - xmod) - exp(-c(x - xmod)))

(3.24)

Both the parameters xmod, c of the Gumbel distribution can be assessed from the mean
and standard deviation
xmod = 0,577
c=

(3.25)

(3.26)

Skewness and kurtosis of the distribution are constant: = 1,14, = 2,4.

An important characteristic of the Gumbel distribution is the simple transformation of
the distribution function (x) of the original distribution to the distribution function N(x)
describing the maxima of populations that are N times greater than the original population
with mean and standard deviation . If the individual multiples of the original population
are mutually independent, then it holds for the distribution function N(x)
N(x) = ((x))N

(3.27)

By substitution of equation (3.23) into equation (3.27) we obtain the distribution

function N(x) as
N(x) = exp(-exp(-c(x - xmod ln N/c)))

(3.28)

so the mean N and standard deviation N of maxima of populations that are N times greater
than the original population are

N = + ln N/c = + 0,78 ln N , N =

(3.29)

Thus the standard deviation N of the greater population is equal to the standard
deviation of the original population, N = , but the mean N is greater than the original value
by 0.78 lnN/c.
Example 3.4.
One-year maxima of wind pressure are described by Gumbel distribution with a mean
1 = 0,35 kN/m2, 1 = 0,06 kN/m2. The corresponding parameters of the 50-year maximum
value distribution, i.e. parameters 50 and 50, follow from equation (3.29)

50 = 0,35 + 0,78 ln (50 0,06) = 0,53 kN/m2, 50 = 0,06 kN/m2

A - 14

Annex A - Basic statistical concepts and techniques

Figure 3.5 shows both distributions of one-year and fifty-year maxima of wind
pressure described by the Gumbel distribution.

Probability density N(x)

8

1(x)

50(x)

6
5
4
3
2
1
0

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Wind pressure x

Figure 3.5. Distribution of maximum wind pressure over the periods of 1 year and 50 years.
The distribution function of type I minimal values distribution (Gumbel distribution of
minimum values) has the form
(x) = 1 - exp(-exp(-c(xmod - x)))

(3.30)

This distribution is symmetrical to the distribution of maximal values given by

equation (3.23). It is therefore also defined on an open interval and depends on two
parameters: on mode xmod and parameter c > 0. By differentiating the distribution function we
obtain the probability density function in the form
(x) = c exp(-c (xmod - x) - exp(-c (xmod - x)))

(3.31)

Both these parameters can be assessed from the mean and standard deviation
xmod = + 0,577
c=

(3.32)

(3.33)

The probability density function of the minimum values is symmetrical to the shape of
maximal values relative to mode xmod, as it is apparent from Figure 3.6.

A - 15

Annex A - Basic statistical concepts and techniques

Probability density (x)
0,4

0,3

Distribution of the
minimum values

0,2

Distribution of the
maximum values

0,1

0,0

10

12

13

15

16

Variable x

Figure 3.6. The Gumbel distribution of the minimum and maximum values.
In a similar way the type II distribution, the so-called Frchet distribution, and type III
distribution, the so-called Weibull distribution, are defined. All the three types of distribution
complement each other with respect to the possible values of skewness . Each type covers a
certain area of skewnesses, as indicated in Figure 3.7.

Distribution of the maximum values

type III

type I

0
Distribution of the minimum values
type II type I

-1,14

1,14

type II

type III

Figure 3.7. Types of distribution of extreme values versus the skewness .

The extreme values distributions of the type I and II are often used to describe random
variables depending on the maximal values of populations (for example climatic actions). The
type II is particularly convenient for variables with high skewness > 1,14 (for example for
flood discharge that have ~ 2). The extreme values distribution of the type III is usually
applied for random variables depending on the minimal values of populations (e.g. strength
and other material properties) assuming that > 1,14.

A - 16

Annex A - Basic statistical concepts and techniques

3.6

Function of random variables

In general many variables entering reliability analysis of structures may be considered
as a function of basic variables X = [X1, X2, ... , Xn]. For example resistance R or load effect E
may be given as a function
Z = F(X)

(3.34)

where X = [X1, X2, ... , Xn] denotes a vector of basic variables. Then the resulting variable Z is
a random variable and its characteristics may be derived from relevant characteristics of basic
variables X = [X1, X2, ... , Xn]. Usually three moment parameters, the mean , standard
deviation and skewness , are used for a first assessment of the resulting variable Z.
Experience shows that using derived moment parameters (, and ) three parameter
lognormal distribution provides satisfactory approximation of Z. However, the software VaP
[9] applies a more accurate approximation based on four moment parameters (, , and
kurtosis ).
Appendix 2 of this Chapter provides approximate expressions for fundamental
functions of two basic variables that can be used in assessment of failure probability in case of
small number of basic variables.
4

ESTIMATION OF FRACTILES

4.1

Fractile of a theoretical model

One of the most important keywords of the theory of structural reliability is the term
fractile of a random variable X (or of its probability distribution), sometimes called also
quantile. Appendix 3 to this Chapter provides a review of formulas for determining fractiles
of most important theoretical models of continuous random variables.
Let us recall the definition of the fractile. For a given probability p, the p-fractile xp
denotes such a value of the random variable X, that values less than or equal to xp occur just
with the probability p. If (x) is distribution function of the random variable X, then it follows
from equation (2.6) that the value (xp) of the distribution function (x) at the point xp is
equal to the probability p
P(X xp) = (xp) = p

(4.1)

The same definition holds also for standardised random variable U (given by
transformation equation (3.2)) when in equation (4.1) U is substituted for X and up is
substituted for xp. Fractiles up of standardised random variables U are commonly available in
tables. Figure 4.1 illustrates the definition of the fractile described by equation (4.1) for
standardised random variable U; it shows distribution function (u), probability density
function (u), probability p (approximately equal to 0,05) and fractile up for normal
standardised distribution U.
In general fractile xp of the original random variable X may be calculated using tables
for up available for standardised random variables U with a relevant type of distribution. It
follows from transformation (3.2) that the fractile xp may be determined from the fractile of
the standardised random variable up (found in available tables) using relationship
xp = + up = (1 + up V)

(4.2)

where denotes the mean, the standard deviation and V the coefficient of variation
of the observed variable X.
If the probability p < 0,5, then the value xp is called the lower fractile, for p > 0,5 the
xp is called the upper fractile. Figure 4.2 shows the lower and upper fractiles up of a
A - 17

Annex A - Basic statistical concepts and techniques

standardized random variable U with normal distribution for probabilities p = 0,05 and 0,95,
and thus denoted u0,05 and u0,95.
The fractile corresponding to the probability p = 0,05, is usually applied for an
assessment of the characteristic value of material properties (strength of concrete, yield point
of steel, masonry strength). However, the design values of dominant variables are fractiles,
which correspond to a lower probability (p 0,001), design values of variables which are not
dominant are fractiles corresponding to a greater probability (p 0,10).
Distribution function (u)
1.0

0.8

0.6

0.4

0.2

up

p
0.0
-3.0

-2.5

-2.0

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

2.0

2.5

3.0

0.0

0.5

1.0

1.5

2.0

2.5

3.0

u
Probability density (u)
0.5

0.4

0.3

0.2

0.1

p
up

0.0
-3.0

-2.5

-2.0

-1.5

-1.0

-0.5

u
Figure 4.1. Definition of the fractile for the standardised random variable U.
The values up of the lower fractile of a standardized random variable U having normal
distribution for selected probabilities p are given in Table 4.1. Considering the symmetry of
the normal distribution, the values up of the upper fractile can be assessed from Table 4.1 by
substituting of p by 1 - p and by changing the sign of values up (from negative to positive).
A - 18

Annex A - Basic statistical concepts and techniques

Detailed tables can be found e.g. in books [12, 13], in the International Standard ISO 12491
[4] and in specialised literature [18, 19, 20].
For a standardized random variable with a general three-parametric lognormal
distribution the value up of the standardized random variable is dependent on skewness . The
values up for selected skewnesses and probabilities p are given in Table 4.2.
Table 4.1. Fractile up of a standardized random variable with normal distribution
10-6
10-5
10-4
0,001 0,010 0,050 0,100 0,200
p
10-7
up

0,500

-5,199 -4,753 -4,265 -3,719 -3,091 -2,327 -1,645 -1,282 -0,841 0,000

Probability density (u)

0,4

0,3

0,2

0,1

1- p = 0,05

p = 0,05
u0,05 = -1,645

0,0
-3,5

-2,5

-1,5

u0,95 = 1,645

-0,5

0,5

2,5

1,5

3,5

Standardised random variable U having normal distribution

Figure 4.2. The lower and upper fractiles of a standardized random variable U having normal
distribution
In the case of a lognormal distribution with lower limit at zero, which is described in
section 3.2, it is possible to calculate the fractile from the value of fractile of a standardized
random variable with normal distribution using the relation
xp =

1+V

exp unorm , p ln(1 + V 2 )

(4.3)

where unorm,p is the fractile of a standardized random variable with normal distribution, is
the mean and V the coefficient of variation of the variable X. An approximation of relation
(4.3) is often applied in the form
xp exp (unorm,p V)

(4.4)

whose accuracy is fully satisfactory for the coefficient of variations V < 0,2, but it is
commonly used for greater V as well.
A - 19

Annex A - Basic statistical concepts and techniques

Table 4.2. Fractile up of a standardized random variable having lognormal distribution

Probability p
0,20

0,50

0,80

0,90

0,95

0,99 1-10-3 1-10-4

-2,0 -9,52 -6,24 -3,52 -1,89 -1,24 -0,61

0,24

0,77

0,97

1,89

1,28

1,42

1,49

-1,5 -7,97 -5,51 -3,31 -1,89 -1,29 -0,68

0,20

0,81

1,04

1,21

1,45

1,65

1,77

-1,0 -6,40 -4,70 -3,03 -1,85 -1,32 -0,74

0,15

0,84

1,13

1,34

1,68

1,99

2,19

-0,5 -4,94 -3,86 -2,70 -1,77 -1,32 -0,80

0,08

0,85

1,21

1,49

1,98

2,46

2,81

0,0 -3,72 -3,09 -2,33 -1,65 -1,28 -0,84

0,00

0,84

1,28

1,65

2,33

3,09

3,72

0,5 -2,81 -2,46 -1,98 -1,49 -1,21 -0,85 -0,08

0,80

1,32

1,77

2,70

3,86

4,94

1,0 -2,19 -1,99 -1,68 -1,34 -1,13 -0,84 -0,15

0.74

1,32

1,85

3,03

4,70

6,40

1,5 -1,77 -1,65 -1,45 -1,21 -1,04 -0,81 -0,20

0,68

1,29

1,89

3,31

5,51

7,97

2,0 -1,49 -1,42 -1,28 -1,89 -0,97 -0,77 -0,24

0,61

1,24

1,89

3,52

6,24

9,52

-4

10

-3

10

0,01

0,05

0,10

Example 4.1.
Let us assess the fractile xp of a normal and lognormal distribution with lower limit at
zero for p = 0,001; 0,01; 0,05 and 0,10, if V = 0,3. We know that the lognormal distribution
with lower limit at zero has, in this case, a positive skewness = 0,927 (according to equation
(3.11)), which needs to be known for interpolation in Table 4.2. The resultant values xp are
given in the following table in the form of dimensionless coefficients xp/ (expressing the
ratio of the fractile to the mean), which were for normal and for lognormal distribution
assessed by different ways.
Table of coefficients xp/.
Probability p
Coefficient xp/ for

0,001

0,010 0,050 0,100

normal distribution, equation (4.2) and Table 4.1

0,073

0,302 0,506 0,615

lognormal distribution, equation (4.2) and Table 4.2

0,385

0,483

0,591

0,658

lognormal distribution, equation (4.3) and Table 4.1

0,387

0,484

0,591

0,657

lognormal distribution, equation (4.4) and Table 4.1

0,396

0,496

0,610

0,681

Table of coefficients xp/ shows the expected difference between the fractiles of
normal and of lognormal distributions. The lower fractile of normal distribution is
significantly lower than the corresponding fractile of lognormal distribution particularly for
small probabilities p. The table also shows that the approximate formula (4.4) provides
satisfactory results for computation of fractile of lognormal distribution (the error will
decrease with decreasing coefficient of variation V).
The fractile of gamma distribution can be calculated from the available tables for type
III Pearson distribution [12, 13]. To calculate the fractile of beta distribution, the available
tables of incomplete beta function may be used or it can be assessed by integration of
probability density function according to definition (4.1). However, when it is needed (and

A - 20

Annex A - Basic statistical concepts and techniques

neither appropriate tables nor software product are available), the fractile of beta distribution,
which is bell shaped (for shape parameters it holds that c > 2 and d > 2), may be assessed
approximately from equation (4.2) using table values of up for a standardized lognormal
distribution, having the same skewness as the beta distribution. Analogical procedure may
be used for other types of distribution, too.
The fractile xp can be easily assessed for Gumbel distribution. From equation (3.23)
and definition (4.1) follows an explicit relation for xp directly dependent on the probability p
1
x p = xmod ln( ln( p)) (0,45 + 0,78 ln( ln( p)))
c

(4.5)

where mode xmod and parameter c are substituted by relations (3.25) and (3.26).
Example 4.2.
Let us determine the upper fractile of wind pressure from Example 3.4 described by
Gumbel distribution when probability p = 0,98 is considered. It is known from Example 3.5
that for the one-year maximum 1 = 0,35 kN/m2, 1 = 0,06 kN/m2. The fractile x0,98 for such
parameters follows from equation (4.5)
x0,98 = 0,35 (0,45 + 0,78 ln(-ln(0,98))) 0,06 = 0,51 kN/m2
The corresponding fractile of the maximum for a period of 50 years (as shown in
Example 3.4 that 50 = 0,53 kN/m2, 50 = 0,06 kN/m2) is
x0,98 = 0,53 (0,45 + 0,78 ln(-ln(0,98))) 0,06 = 0,69 kN/m2
Simple mathematical procedures, including the computation of fractile, are some of
the reasons of the wide popularity of Gumbel, distribution which is frequently used for
random variables describing climatic and other variable actions that are defined by maximal
values for a given period (e.g. during one year).
However, theoretical models are not always known in practical applications. In civil
engineering, the fractile of a random variable (e.g. strength of a new or unknown material) has
to be assessed from a limited sample, the size n of which may be very small (sometimes less
than 10). Furthermore, considered random variables may have a high variability (the
coefficient of variation is sometimes greater than 0,30). Assessment of the fractile of a
population from a sample is then a serious problem, which is in mathematical statistics solved
by various methods of estimation theory. In the following three basic methods are shortly
described: the coverage method, the prediction method and the Bayesian method for
estimation of the population fractile.
4.2

Coverage method of fractile estimation

The keyword of the coverage method for the fractile estimation from a sample of a
limited size n is the confidence , i.e. the probability (usually 0,75, 0,90 or 0,95) that the
estimated value covers the population fractile (that is why the method is called coverage
method). The estimator xp,cover of the lower fractile xp is determined by the coverage method in
such a way that
P(xp,cover < xp) =

(4.6)

Thus, the estimator xp,cover is lower (on the safe side of the lower fractile) than the
unknown fractile xp with the probability (confidence) .

A - 21

Annex A - Basic statistical concepts and techniques

In the following summary practical formulas are given without being derived,
assuming that the population has a general three-parameter distribution characterized by
skewness , known from previous experience. In addition it is assumed that the mean of the
population is never known in advance and the estimation is based by the average m obtained
from a sample. The standard deviation of the population is assumed to be either known (and
then it is used) or unknown (and then the sample standard deviation s or the coefficient of
variation V is used instead of ).
If the standard deviation of the population is known from previous experience, the
estimator xp,cover of the lower p-fractile is given as
xp,cover = m p

(4.7)

If the standard deviation of the population is unknown, then the sample standard
deviation s is considered
xp,cover = m kp s

(4.8)

Coefficients of estimation p = (, p, , n) and kp = k (, p, , n) depend on

skewness , on probability p corresponding to the fractile xp that is estimated, on confidence
and on the size n of the population. The knowledge of confidence that the estimator xp,cover
will be on the safe side of the real value is the greatest advantage of the classic coverage
method. In documents [1, 2] the confidence is recommended by the value 0,75. In the cases
of increased reliability demands when a detailed reliability analysis is required, a higher value
of confidence, say of 0,95, may be more appropriate [4].
4.3

Prediction method of fractile estimation

According to the prediction method [4] the lower p-fractile xp is estimated by the socalled prediction limit xp,pred for which it holds that a new value xn+1 randomly drawn from the
population will be lower than the estimator xp,pred only with the probability p, i.e. it holds that
P(xn+1 < xp,pred) = p

(4.9)

It can be shown that for growing n the estimator xp,pred defined in this way is
asymptotically approaching the unknown fractile xp. It can be also shown that the estimator
xp,pred corresponds approximately to the estimator obtained by the coverage method xp,cover for
confidence = 0,75 [4].
If the standard deviation of the population is known, then the lower p-fractile is
estimated by the value xp,pred according to the relation
xp,pred = m + up (1/n + 1)1/2

(4.10)

where up = u(, p) is the p-fractile of a standardized lognormal distribution having the

skewness . If the distribution of the variable X is normal then up is the p-fractile of
standardised normal distribution.
If, however, the standard deviation of population is unknown, then the sample
standard deviation s must be considered instead of
xp,pred = m + tp (1/n + 1)1/2 s

(4.11)

where tp = t(, p, ) is the p-fractile of the generalized Students t-distribution for = n 1

degrees of freedom, which has a skewness (information about Students distribution and
about the number of degrees of freedom may be obtained from the textbook [12,13] and from
other specialised sources [18, 19]). If the distribution of the variable X is normal, then up is the
p-fractile of standardised normal distribution Students t-distribution for = n 1 degrees of
freedom.
A - 22

Annex A - Basic statistical concepts and techniques

4.4

Coefficients of the coverage and prediction methods

The coverage and prediction methods represent two basic procedures of estimation of
the populations fractile from the available sample of a limited size n. If the standard
deviation of the population is known, then equations (4.7) and (4.10) are applied in which
two analogical coefficients p and up(1/n + 1)1/2 appear. Both of these coefficients depend on
the sample size n, coefficient p of the coverage method depends furthermore on the
confidence . Table 4.3 shows the coefficients p and up(1/n + 1)1/2 for p = 0,05 and selected
values of n and when normal distribution of the population is assumed.
Table 4.3. Coefficients p and up(1/n + 1)1/2 from equations (4.7) and (4.10) for p = 0,05 and
normal distribution of the population (when is known).
Sample size n
Coefficient
3
4
5
6
8
10
20
30

2,03 1,98 1,95 1,92 1,88 1,86 1,79 1,77 1,64

= 0,75
2,39 2,29 2,22 2,17 2,10 2,05 1,93 1,88 1,64
p
= 0,90
2,60 2,47 2,38 2,32 2,23 2,17 2,01 1,95 1,64
= 0,95
1/2
1,89 1,83 1,80 1,77 1,74 1,72 1,68 1,67 1,64
up(1/n+1)
It is evident from Table 4.3 that with the growing sample size n both the coefficients
approach the value 1,64, which holds for a theoretical model of the normal distribution (see
Table 4.1). The coefficient p of the coverage method increases with increasing confidence .
Note that for confidence = 0,75 it holds that p up(1/n + 1)1/2. Thus, for = 0,75 the
coverage method leads approximately to the same estimator as the prediction method, xp,cover
xp,pred (for greater confidence > 0,75 the xp,cover < xp,pred).
If the standard deviation of the population is unknown, equations (4.8) and (4.11)
are applied in which two analogical coefficients kp and tp(1/n + 1)1/2 appear. Both of these
coefficients depend again on the sample size n, coefficient kp of the coverage method depends
furthermore on the confidence . Table 4.4 and Figure 4.3 show the values of coefficients kp
and tp(1/n + 1)1/2 for p = 0,05 and selected values of n and when normal distribution of the
population is assumed.
Table 4.4. Coefficients kp and tp(1/n + 1)1/2 from equations (4.8) and (4.11) for p = 0,05 and
normal distribution of the population (when is unknown).
Coefficient
Sample size n
3
4
5
6
8
10
20
30

3,15 2,68 2,46 2,34 2,19 2,10 1,93 1,87 1,64

= 0,75
kp
5,31 3,96 3,40 3,09 2,75 2,57 2,21 2,08 1,64
= 0,90
7,66 5,14 4,20 3,71 3,19 2,91 2,40 2,22 1,64
= 0,95
1/2
3,37 2,63 2,33 2,18 2,00 1,92 1,76 1,73 1,64
tp(1/n+1)
It is obvious from Table 4.4 and Figure 4.3 that with increasing sample size n both the
coefficients kp and tp(1/n + 1)1/2 approach the value 1,64, which is valid for a theoretical
model of the normal distribution (see Table 4.1). In case of the coverage method, the
coefficient kp increases with increasing confidence and the relevant estimators xp,cover of the
lower fractile are decreases (on the safe side). Note, that as in the case of known standard
deviation both coefficients are approximately equal, kp tp(1/n + 1)1/2 and for confidence
A - 23

Annex A - Basic statistical concepts and techniques

= 0,75 the coverage method leads to approximately the same estimator, xp,cover xp,pred, as the
prediction method.
10

Coefficients kp and tp(1/n+1)1/2

kp for = 0,95
kp for = 0,90

kp for = 0,75

1,64

tp(1/n+1)1/2

0
0

10

15

20

Figure 4.3. Coefficients kp and -tp(1/n + 1)1/2 for p = 0,05 and normal distribution of the
population (when is unknown).
Also the skewness (asymmetry) of the population may affect significantly the
estimator of the populations fractile. Tables 4.5 and 4.6 show the coefficients kp from
equation (4.8) for three value of the skewness = 1,0, 0,0 and 1,0, probability p = 0,05 and
confidence = 0,75 (Table 4.5) and = 0,95 (Table 4.6). Values of the coefficients from
Table 4.6 are shown in Figure 4.4.
Table 4.5. Coefficient kp from equation (4.8) for p = 0,05, = 0,75 and lognormal distribution
having skewness (when is not known).
Sample size n
Skewness
3
4
5
6
8
10
20
30

4,31 3,58 3,22 3,00 2,76 2,63 2,33 2,23 1,85

= 1,00
3,15 2,68 2,46 2,34 2,19 2,10 1,93 1,87 1,64
= 0,00
2,46 2,12 1,95 1,86 1,75 1,68 1,56 1,51 1,34
= 1,00
Table 4.6. Coefficient kp from equation (4.8) for p = 0,05, = 0,95 and lognormal distribution
having the skewness (when is not known).
Sample size n
Skewness
3
4
5
6
8
10
20
30
10,9 7,00 5,83 5,03 4,32 3,73 3,05 2,79 1,85
= 1,00
7,66 5,14 4,20 3,71 3,19 2,91 2,40 2,22 1,64
= 0,00
5,88 3,91 3,18 2,82 2,44 2,25 1,88 1,77 1,34
= 1,00
It is evident from Tables 4.5 and 4.6 that as the sample size n increases, the
coefficients kp approach the values of up, which are valid for theoretical model of lognormal
distribution (see Table 4.2). Thus, the influence of the skewness does not disappear when n
A - 24

Annex A - Basic statistical concepts and techniques

, and it is especially significant for small samples and greater confidence = 0,95 (see
Figure 4.4).
10
kp

= 1,00
= 0,00
= +1,00

1,64

n
0
0

10

15

20

Figure 4.4. Coefficient kp for p = 0,05 and confidence = 0,95 (when is unknown).
A similar dependence on skewness may be observed in the case of the generalized
Students t-distribution for which the fractiles tp are given in Table 4.7. These values tp are
applied in the prediction method using formula (4.11) and further in the Bayes method. That
is why Table 4.7 gives directly the values of fractiles tp depending on the number of degrees
of freedom . Similarly as in Tables 4.5 and 4.6 the probability p = 0,05 and three skewnesses
= 1,0; 0,0 and 1,0 are considered.
Table 4.7. Coefficient tp from equation (4.11) for p = 0,05 and lognormal distribution with
skewness (when is unknown).
Coefficient tp for = n 1 degrees of freedom
Skewness
3
4
5
6
8
10
20
30

2,65 2,40 2,27 2,19 2,19 2,04 1,94 1,91 1,85

= 1,00
2,35 2,13 2,02 1,94 1,86 1,81 1,72 1,70 1,64
= 0,00
1,92 1,74 1,64 1,59 1,52 1,48 1,41 1,38 1,34
= 1,00
It follows from Table 4.7 that as the size of the sample n increases, the values of tp
approach the theoretical values of up, which are valid for a model of lognormal distribution
with the appropriate skewness and are given in Table 4.2. Therefore, the influence of the
skewness again (as in the case of kp) does not disappear for n , but it is especially
significant for small samples (it increases with decreasing sample size n).
Example 4.3.
A sample of the size n = 5 measurements of strength of concrete has the mean m =
29,2 MPa and the standard deviation s = 4,6 MPa. We assume that the population is normal
and that its standard deviation is unknown. The characteristic strength fck = xp, for p = 0,05
is firstly assessed by the coverage method. If the confidence is = 0,75, then it follows from
equation (4.8) and Table 4.4 that

A - 25

Annex A - Basic statistical concepts and techniques

xp,cover = 29,2 2,46 4,6 = 17,9 MPa
If the higher confidence = 0,95 is required, then
xp,cover = 29,2 4,20 4,6 = 9,9 MPa
If the prediction method is used, then it follows from equation (4.11) and Table 4.4 that
xp,pred = 29,2 2,33 4,6 = 18,5 MPa
The characteristic strength obtained by the prediction method is only a little greater
than the value according to the coverage method with confidence = 0,75. However, if a
higher confidence = 0,95 is required, then the prediction method leads to a value which is
almost twice greater than the value obtained by the coverage method.
If the sample comes from a population with lognormal distribution and a positive
skewness = 1, then the coverage method with the confidence = 0,75 (Table 4.5) gives an
estimator
xp,cover = 29,2 1,95 4,6 = 20,2 MPa
which is a value by 13% greater than when the skewness is zero.
Similarly it follows for the prediction method from equation (4.11) and Table 4.7 that
x p , pred = 29,2 1,74

1
+ 1 4,6 = 20,4 MPa
5

where the value tp = -1,74 is given in Table 4.7 for = 1,0 and = 5 - 1= 4. The resulting
strength is in this case by 10% greater than the value, which corresponds to the normal
distribution ( = 0).
4.5

Bayes' method of fractile estimation

If previous experience is available for a random variable (e.g. in the case of a long
term production) it is possible to use so-called Bayes` method, which generally follows the
idea of updating of probabilities described in section 2.5. The Bayes method of fractile
estimation is described here without deriving any important relations. More detailed
description is given in documents ISO [3, 4] and other specialised literature [12, 13].
Assume that a sample of size n with an average m and standard deviation s is
available. Note that degrees of freedom = n 1. Besides an average m and sample standard
deviation s' assessed from an unknown sample (of an unknown size n and degrees of freedom
) are available from previous experience. It is, however, assumed that both the samples
come from the same population having the mean and the standard deviation . If this
important assumption is valid, then the two samples may be combined. This could be a simple
task if the individual values of the previous set were known, but that is not the case. However,
the Bayes` method must be used.
Parameters of the combined sample are generally given by relations [3, 4]
n'' = n + n'

'' = + ' 1 if n' 1, '' = + ' if n' = 0

m'' = (mn + m'n') / n''
s'' 2 = ( s2 + 's' 2 + nm2 + n'm' 2 n''m'' 2) / ''

A - 26

(4.12)

Annex A - Basic statistical concepts and techniques

The unknown sample size n' may be assessed using the relations for coefficients of
variation of the mean and standard deviation V() and V(), (parameters and are
considered as random variables in the Bayes concept) for which it holds [3, 4]
n' = [s' / (m' V())]2, ' = 1 / (2 V()2)

(4.13)

Both the unknown variables n' and ' may be assessed independently (generally ' n'
- 1), depending on previous experience with the degree of uncertainty of estimator of the
mean and standard deviation of the population.
The next step of the procedure applies the prediction method of fractile estimation.
The Bayes estimator xp,Bayes of the fractile is given by relationship similar to equation (4.11)
for prediction estimator, assuming that the standard deviation of the population is not
known
1/ 2
(4.14)
x p,Bayes = m' ' +t p (1 / n' ' +1) s' '
where t' p' = t'p' ( , p, ' ' ) is a fractile of the generalised Students t-distribution having an
appropriate skewness for '' degrees of freedom (that is generally different from the value
n'' 1).
If the Bayes method is applied for an assessment of material strength, the advantage
may be taken of the fact that the long-term variability is constant. Then the uncertainty of an
assessment of and the value v() are relatively small, variables ' assessed according to
equation (4.13) and '' assessed according to equation (4.12) are relatively high. This factor
may lead to a favourable decrease of the value t'p' and to augmentation of the estimator of the
lower fractile of xp according to equation (4.14). On the other hand, uncertainties in
assessment of the mean and the variable v() are usually great and previous information
may not affect significantly the resulting values n'' and m''.
If no previous information is available, then n' = ' = 0 and the resulting characteristics
m'', n'', s'', '' equal the sample characteristics m, n, s, . In this case the Bayes method is
reduced to the prediction method and equation (4.14) becomes equation (4.11); if is known
equation (4.10) is used. This particular form of the Bayes method, when no previous
information is available, is considered in Eurocode EN 1990 [1] and international standards
ISO [2, 3].
Example 4.4.
If previous experience was available for Example 4.3, the Bayes method could be
used. Suppose that the information is m = 30,1 MPa, V() = 0,50, s = 4,4 MPa, V() = 0,28.
It follows from equation (4.13) that
2

1
4,4 1
n' =
6
< 1 , ' =
2 0,282
30,1 0,50
Further on these values are thus considered: n' = 0 and ' = 6. Because = n 1 = 4, it
follows from equation (4.12)
n'' = 5, '' = 10, m'' = 29,2 MPa, s'' = 4,5 MPa.
From equation (4.14) the fractile estimate follows as
x p , Bayes = 29,2 1,81

1
+ 1 4,5 = 20,3 MPa
5

A - 27

Annex A - Basic statistical concepts and techniques

where the value t 'p' = 1,81 is given in Table 4.7 for = 0 and '' = 10. The resulting strength is
thus greater (by 10%) than the value obtained by the prediction method.
If the population has lognormal distribution with skewness = 1, then it follows from
equation (4.14) considering the value t'p' = 1,48 given in Table 4.7 that
x p , Bayes = 29,2 1,48

1
+ 1 4,5 = 21,9 MPa
5

which is a value by 8% greater than the Bayes estimator for = 0.

Examples 4.3 and 4.4 clearly showed that the estimator of characteristic strength
(fractile with probability p = 0,05) assessed from one sample may be expected within a broad
range (in Examples 4.3 and 4.4 from 9,9 MPa to 21,9 MPa), depending on the applied
method, required confidence, previous information and on assumptions concerning the
population. Besides the alternatives considered in Examples 4.3 and 4.4, knowledge of the
standard deviation of the population and assumption of the negative skewness (in the case
of some materials of high strength) may be applied as well.
Even more significant differences in the resulting values may occur when design
values of strength are being estimated, i.e. when fractiles corresponding to a small probability
(p 0,001) are considered. However, a direct estimation of such fractiles from a limited
sample of the population is recommended only in such cases when a sufficient amount of
information on the relevant random variable is available. In such cases, it is necessary to
proceed carefully and, if possible, in co-operation with experts in the field of mathematical
statistics.
4.6

Estimation of fractiles according to Eurocodes

Eurocode EN 1990 [1] gives in tables the coefficients for estimation of a fractile of a
random variable with normal distribution (asymmetric distributions thus are not considered
for the fractile estimation) from a sample for three probabilities p = 0,05 (for characteristic
value xk), p = 0,001 (for design value xd of the dominant variable) and for p = 0,10 (for design
value xd of the non-dominating variable). As already mentioned above, the characteristic
values xk and design values xd are defined as fractiles xp, which correspond to a given
probability p (application of these variables in structural design is explained in the following
chapters).
For characteristic values of material properties a fractile corresponding to probability p
= 0,05 is usually considered (however, for variables which describe variable loads the
probability p is usually less than that), i.e. it holds
P(X < xk) = 0,05

(4.15)

For design values xd of dominating variables it holds approximately that p = 0,001 (or
another value close to this one), i.e. it holds
P(X < xd) = 0,001

(4.16)

Finally, for design values xd of non-dominant variables it holds approximately that p =

0,1, i.e. it holds that
P(X < xd) = 0,1

(4.17)

A more detailed description of the dominating and non-dominating variables is given

in Handbook 1.

A - 28

Annex A - Basic statistical concepts and techniques

The following Tables 4.8 and 4.10, which give the required coefficients for estimation
of variables xk and xd according to equations (4.15) to (4.17), are adopted from the document
[1] in its original version, even though the first Table 4.8 partially overlaps with the precedent
Tables 4.3 and 4.4. Tables 4.8 and 4.9 are taken from the final version of EN 1990 [1]. Let us
remark that all the coefficients in [1] are denoted by the symbol kn, which is used also in the
following tables.
Table 4.8. Coefficients kn for a 5% characteristic value (see Tables 4.4 and 4.3).
Sample size n
Coefficient
1
2
3
4
5
6
8
10 20 30

1/2
up(1/n+1) , known 2,31 2,01 1,89 1,83 1,80 1,77 1,74 1,72 1,68 1,67 1,64
- 3,37 2,63 2,33 2,18 2,00 1,92 1,76 1,73 1,64
tp(1/n+1)1/2, unknown Table 4.9. Coefficients kn for a design value xd of a dominating variable, P(X < xd) = 0,001.
Sample size n
Coefficient
1
2
3
4
5
6
8
10 20 30

1/2
4,36
3,77
3,56
3,44
3,37
3,33
3,27
3,23
3,16
3,13
3,09
up(1/n+1) , known
- 11,4 7,85 6,36 5,07 4,51 3,64 3,44 3,09
tp(1/n+1)1/2, unknown Table 4.10. Coefficients kn for a design value xd of a non-dominating variable, P(X < xd) = 0,1.
Sample size n
Coefficient
1
2
3
4
5
6
8
10 20 30

1/2
up(1/n+1) , known 1,81 1,57 1,48 1,43 1,40 1,38 1,36 1,34 1,31 1,30 1,28
tp(1/n+1)1/2, unknown - 3,77 2,18 1,83 1,68 1,56 1,51 1,45 1,36 1,33 1,28
The assumption concerning knowledge of the standard deviation is replaced
(inaccurately) in the document by the assumption that the coefficient of variation V is known.
The original version of Table 4.9 [1] gives for the sample size of a wrong value of 3,04 for
the coefficients (correct is 3,09). Let us also note that when knowledge of the standard
deviation is assumed, Tables 4.8 to 4.10 give values of coefficients already for the sample
size n = 1. Application of these values is, however, associated with significant statistical
uncertainties and therefore a minimum sample size n = 3 is recommended here. Note, that
Table 4.10 (for 0,1 fractile) is included only in the prestandard ENV 1991-1 and not in the
final document EN 1990 [1].
Statistical methods for determining the characteristic and design values of resistance
variables are provided in Annex D Design assisted by testing of EN 1990 [1]. Relevant
basic variables describing structural resistance are described by lognormal distribution. The
whole procedure is described in detail in the Annex D. Attached MATHCAD sheet
Mod_est.mcd can be used to evaluate experimental data using the whole procedure. It is
provided with explanatory notes and needs no additional information.
In order to simplify computational procedure the assessment coefficients given in
Tables 4.8 and 4.9 are in the attached MATHCAD sheet Mod_est.mcd expressed using
built-in distribution function of normal and Student t- distribution. In accordance with the
principles of Annex D in [1] single variable and model representation of a resistance variable
R are distinguished. The results shown in the attached sheet indicates that both approaches
lead to similar results.

A - 29

Annex A - Basic statistical concepts and techniques

REFERENCES
[1] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[2] EN 1991-1-1 Eurocode 1 Actions on structures. Part 1-1 General actions. Densities, selfweight, imposed loads for buildings, CEN 2002.
[3] ISO 2394 General principles on reliability for structures, ISO Geneve, Switzerland 1998.
[4] ISO 12491 Statistical methods for quality control of building materials and components.
Geneve, Switzerland 1997.
[5] ISO 13822. Basis for design of structures - Assessment of existing structures, ISO 2003.
[6] JCSS: Background documentation, Part 1 of EC 1 Basis of design, 1996.
[7] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/, 2001.
[8] Gulvanessian, H. Calgaro, J.-A. Holick, M.: Designer's Guide to EN 1990,
Eurocode: Basis of Structural Design. Thomas Telford, London, 2002, ISBN: 07277 3011
8, 192 pp.
[9] VaP, Variable Processor, version 1.6. ETH Zurich, 1997.
[10] COMREL, version 7.10, Reliability Consulting Programs, RCP MUNICH, 1999.
[11] CodeCal, Excel sheet developed by JCSS, http://www.jcss.ethz.ch/, 2001.
[12] J.R. Benjamin and C.A. Cornell, Probability concept and decision for civil engineers.
McGraw Hill New York, 1970.
[13] A.H.-S. Ang and W.H. Tang, Probabilistic concepts in engineering planning and design.
John Wiley and Sons, New York, 1975.
[14] CIRIA Report 63, Rationalization of safety and serviceability factors in structural codes.
London,1975.
[15] P.T. Christensen, M. Baker, Structural Reliability Theory and its Applications. SpringerVerlag, Berlin 1982.
[16] H.O. Madsen, S. Krenk and N.C. Lind, Methods of Structural Safety. Prentice-Hall, Inc.,
Englewood Cliffs, 1986.
[17] R.E. Melchers, Structural Reliability. Analysis and Prediction. John Wiley & Sons, New
York 1987.
[18] M.E. Haar, Reliability Based Design in Civil Engineering. McGraw-Hill, 1987
[19] M. Tich, Applied methods of structural reliability. Kluwer academic publishers 1993.
[20] O. Dietlevsen, H.O. Madsen, Structural Reliability Methods. John Wiley & Sons, New
York 1996.
[21] H. Gulvanessian, M. Holick, Designers' Handbook to Eurocode 1, Thomas Telford,
London, 1996.
[22] J. Schneider, Introduction to Safety and Reliability of Structures. IABSE, Zrich, 1997.

A - 30

Annex A - Basic statistical concepts and techniques

Appendix 1 - Probabilistic models of basic variables
Distribution,
notation
Rectangular
R(a,b)
Normal
N(,)
Lognormal,
general
LN(,,)
LN(,,x0)
Lognormal,
zero origin
LN(,)
Gamma
Gam(,)
Beta,
general
Beta(,,,b)
Beta(,,a,b)
Beta,
zero origin
Beta(,,)
Beta(,,b)
Gumbel
Gum(,)

Probability density function

Domain
of X

Parameters

Mean

Standard
deviation

Skewness

1/(b a)

axb

a
b>a

(a + b)/2

(b a)/12

1 x 2
exp

2
2

x0 + c

3c+c3

3V+V3

k/

k/

2 /k

| x x 0 || c | 1 + c 2

exp ln

|x x 0| ln(1 + c 2 ) 2

x 1+V 2

exp ln

x ln(1 + V 2 ) 2

(2 ln(1 + c 2 ))

(2 ln(1 + V 2 ))

k xk-1 exp(-x) / (k)

c 1

d 1

( x a ) (b x)
B(c, d )(b a ) c + d 1

c 1

x0 x < x0 = c
pro > 0,

c
<xx0
pro < 0
0x<

V = /
0x<

= /2
k = ( /)2

axb

a
b >a
c 1
d 1

0xb

d 1

( x) (b x)
B(c, d ) b c + d 1

c exp(c(xxmod) exp( c(x xmod)))

A - 31

b >0
c 1
d 1

x < xmod =
0,5776 /

c = /(6)

2 g (d c)
(b a )
a+
,
,
cg
+
dg
c
+
d
+
2
(b a ) c
+
c + d +1 g = c + d +1
c+d
g=
cd
cd

b c
c+d

b
,
cg + dg
g=

xmod +
0,577/c

2 g (d c)
,
c+d +2

c + d +1 g = c + d +1
cd
cd

/(6c)

1,14

Annex A - Basic statistical concepts and techniques

Appendix 2 - Statistical parameters of functions of random variables

Function Z
aX+b
X2

1
X

*)

*)

aX + bY + c

The mean Z

Standard deviation Z

Skewness Z

aX + b

| a | X

X pro > 0, - X pro < 0

X2 + X2
1 + V X2 V X3 X

X
aX + bY + c

2 X X2 + X X X

(V

2
X

2V X3 X

(a
2

X
2
X

+ b 2 Y2

X+Y

X + Y

2
X

+ Y2

XY

X Y

2
X

+ Y2

XY

X
Y

*)

X Y

*)

X (1 + VY2 VY3 Y )
Y

1/ 2

Z3

6 V X4 V X3 X

a 3 X3 X + b 3 Y3 Y

1/ 2

X3 Z3

1/ 2

Z3

1/ 2

X3 X + Y3 Y
Z3

1/ 2

X3 X Y3 Y
Z3

X Y (V X2 + VY2 + V X2 VY2 )1 / 2
X (V X2 + VY2 2 VY3 Y )
Y

1/ 2

*) Expressions for parameters of marked functions are approximations only.

A - 32

8 X3 X3 ( X + 3 V X )

X3 Y3 (V X3 X + VY3 Y + 6V X2 VY2 )
Z3
X3 (V X3 X VY3 Y + 6VY4 + 6V X2 VY2 )
Y3 Z3

Annex A - Basic statistical concepts and techniques

Appendix 3 - Fractile of a random variable xp, P( X x p ) = ( x p ) = p
Distribution,
notation

Domain of
X

Fractile xp of the theoretical model

xp =

Rectangular
R(a,b)
Normal
N(,)

axb

a + p (b a)

+ up = (1+ up V)

m p
p from
Table 4.3

Lognormal,
general
LN(,,)
LN(,,x0)

Estimate using coverage

method
known
unknown

up from Table 4.1

x0 x <

1
pro > 0, 1
exp sign ( )u p ln(1 + c 2 )

2
c
1+ c
< x x0

pro < 0
+ x0
= x0 +
exp sign ( )u p ln(1 + c 2 )
2
1+ c

Estimate using prediction method

known

unknown

m kp s
kp from
Table 4.4

m+up(1/n+1)1/2
up from Table 4.1

m+tp(1/n+1)1/2 s
tp(1/n+1)1/2 from
Table 4.4

m kp s

m+up(1/n+1)1/2

m+tp(1/n+1)1/2 s

kp from 4.5
and 4.6

up from Table 4.2

tp from Table 4.7

m p

m kp s

m+up(1/n+1)1/2

m+tp(1/n+1)1/2 s

p not given

kp from
Table 4.5
and 4.6

up from Table 4.2

tp from Table 4.7

m p

p not given

up for normal distribution or

+ up = (1+ up V)
Lognormal,
zero origin
LN(,)

0x<

up for lognormal distribution from Table 4.2

1+V 2

exp u p ln(1 + V )
2

exp(u p V ) for V < 0,2

up for normal distribution or

+ up = (1+ up V)
and up for lognormal distribution from Table 4.2
Gumbel
Gum(,)

x <

Fractile can be estimated using the parameter lognormal distribution

as an approximation

1
x mod ln( ln( p ))
c
(0,45 + 0,78 ln( ln( p )))
A - 33

Annex A - Basic statistical concepts and techniques

ATTACHEMENTS
1. MATHCAD sheet DistFract.mcd Fractiles of basic types of distributions.
MATHCAD sheet DistFract.mcd is intended for determination of fractiles of selected
theoretical models.
2. MATHCAD sheet SampFract.mcd Estimation of sample fractile.
MATHCAD sheet SampFract.mcd is intended for determination of fractiles using
limited samples.
3. MATHCAD sheet Mod_est.mcd Estimation of models.
MATHCAD sheet Mod_est.mcd is intended for determination of fractiles using limited
samples taking into account model uncertainties.

A - 34

Annex A - Basic statistical concepts and techniques

Attachment 1 - MATHCAD sheet DistFract.mcd

MATHCAD sheet "DistFract"

Fractiles of basic types of distributions
Definition of the fractileXP : P = Prob (X<XP), relative value P = XP / X
1 Input data for a random variableX
Basic characteristics :

:= 1

V := 0.10

:= V

An example of the design value for a resistance variable:

:= 0.8

Range for the probability P considered below:

check:

:= 3.8 P := pnorm ( , 0 , 1)

P = 1.183 10

p := 0.001, 0.005.. 0.999

Standardised normal fractile given by inverse distribution function:

u ( p) := qnorm( p , 0 , 1)

2 Fractiles of the normal distribution n(p) = XP / X

n ( p) := 1 + u ( p) V

3 Fractiles of the two parameter lognormal distribution ln(p) = XP / X

ln ( p) :=

exp u ( p) ln 1 + V
1+ V

lna ( p) := exp( u ( p) V )

Correct formula for any V

Common approximation for V < 0,2

4 Fractiles of a general three parameter lognormal distribution

Skewness a as a range variable
Parameter C of three
parameter lognormal
distribution of g:

a := 1 , 0.5 .. 1
1

2 + 4 + a a2 + 4 a
a

C ( a) :=
1

Parameters of transformed variable: mg ( a) := ln ( C ( a)

sg ( a) :=

ln 1 + C ( a)

x0 ( a) :=

1
C ( a)

) + ln ( ) ( 0.5) ln (1 + C ( a) 2)
Check:

2
exp sign ( a) u ( p) ln 1 + C ( a)

lng ( p , a) := 1
1
C ( a)
2
1 + C ( a)

A - 35

x0 ( 1) = 0.69

Annex A - Basic statistical concepts and techniques

5 Fractiles of the gamma distribution

k :=

Parameters of gamma distribution:

:=

Transformed variable u = x, shape factor s = k

No explicit formula is available

gam ( p) :=

qgamma ( p , k )

6 Fractiles of the Gumbel distribution

Explicit formula:

gum ( p) := 1 V ( 0.45 + 0.78ln ( ln ( p) ) )

7 Relative values of fractiles P = xP/ X versus probability P

The coefficient of variation:

Lower fractiles

V = 0.1

Check:
n ( 0.001) = 0.691
ln ( 0.001) = 0.731
lna ( 0.001) = 0.734
lng ( 0.001 , 1) = 0.801
lng ( 0.001 , 1) = 0.53
lng ( 0.001, 0.4) = 0.743
gam ( 0.001) = 0.719

n ( p )

0.9

ln ( p)
lng ( p , 1)

0.8

lng ( p , 1)
gam ( p )
gum ( p )

0.7

0.6

gum ( 0.001) = 0.804

0.5 3
1 .10

0.01

0.1

Notes. 1) It follows from Figure

that the skewness of the
distribution may have significant
effect on assessment of the
design value (0,001 fractile).
2) Approximate formula
for two parameter lognormal
distribution yields sufficiently
accurate results for the
coefficient of variation V < 0,2.
3) Gamma and Gumbel
distribution can be well
approximated by three
parameter lognormal
distribution having skewness
equal to = 2*V and = 1,14
respectively.

Upper fractiles
1.8
n ( p )
ln ( p)

1.6

lng ( p , 1)
lng ( p , 1)

1.4

gam ( p )
gum ( p )

1.2

0.9

0.92

0.94

0.96
p

A - 36

0.98

Annex A - Basic statistical concepts and techniques

Attachment 2 - MATHCAD sheet SampFract.mcd

MATHCAD sheet "SampFract" for estimation of sample fractile

MATHCAD sheet for determination of the characteristic and design values and material
partial factor M using test data in accordance to EN 1990, Annex D.
1. Analytic expressions for coefficients of fractile estimation given in EN 1990, Annex D
5% fractile V unknown

ks ( n ) := qt ( 0.95, n 1) 1 +

5% fractile V known

1
n
1

k( n ) := qnorm( 0.95, 0 , 1) 1 +

0,1 % fractile V unknown

ds ( n ) := qt ( 0.999, n 1) 1 +

d( n ) := qnorm( 0.999, 0 , 1) 1 +

0,1 % fractile V known

1
n

2. Characteristic and design values (relative values related to the mean)

Single variable, V unknown: xk= ks* x, xd= ds x
ks ( n , V) :=

exp( ks ( n ) ) ln 1 + V
2

ds ( n , V) :=

1+ V
Single variable, V known: xk= k *x, xd= d*x
k( n , V) :=

exp( k( n ) ) ln 1 + V
2

d( n , V) :=

exp( ds ( n ) ) ln 1 + V
1+ V

1+ V

exp( d( n ) ) ln 1 + V
2

1+ V

Model, V unknown: xk= ks* x,xd= ds x, weighting factors:

2
( 2
)
2
2
dsmod ( n , V) := exp( 3.09rt + ds ( n ) + 0.5 V) V

rt := 0.707 := 0.707

ksmod ( n , V) := exp 1.65rt + ks ( n ) + 0.5 V V

Model, V unknown: xk= k *x, xd= dx

2
( 2
)
2
2
dmod( n , V) := exp( 3.09rt + d( n ) + 0.5 V) V

kmod( n , V) := exp 1.65rt + k( n ) + 0.5 V V

3. Partial factor M

Check values:

Single variable, V unknown

Ms ( n , V) :=

Single variable, V known

M( n , V) :=

Model, V unknown
Model, V known

ks ( n , V)
ds ( n , V)
k( n , V)

d( n , V)
ksmod ( n , V)
Msmod ( n , V) :=
dsmod ( n , V)
kmod( n , V)
Mmod( n , V) :=
dmod( n , V)

A - 37

Ms ( 100, 0.162) = 1.278

M( 100, 0.162) = 1.263
Msmod ( 100, 0.162) = 1.271
Mmod( 100, 0.162) = 1.264

Annex A - Basic statistical concepts and techniques

4. Graphs for characteristic and design values

V := 0.01, 0.02.. .4

0.9

ks( 10 , V)

0.8

ds( 10 , V)
ksmod( 10 , V)
dsmod( 10 , V)

0.7

0.6

0.5

0.4

0.1

0.2

0.3

0.4

Figure 1. Characteristic and design values versus coefficient of variation V for n = 100
and weighting factors rt= 0,301, = 0,955
5. Graphs for M factors

1.8
Ms ( 10 , V)
M ( 10 , V)

1.6

Msmod ( 10 , V)
M mod( 10 , V)

1.4

1.2

0.1

0.2

0.3

0.4

Figure 2. Partial factor M versus coefficient of variation V for n=100 and weighting factors
rt= 0,301, = 0,955

A - 38

Annex A - Basic statistical concepts and techniques

Attachment 3 - MATHCAD sheet Mod_est.mcd

MATHCAD sheet "Mod_est" Estimation of Models, characteristic,

design and M values determined from test data
MATHCAD sheet for estimation of models, the characteristic and design
values of resistance variable R and material partial factor M using test data
(file "rdata.prn") in accordance to EN 1990, Annex D.
Test data to run the sheet without data
recorded in the file "rdata.prn":

1. Experimental data

rt := ( 1 2 3 4 5 6 7 8 9 10 )

re := ( 1 3 4 4 5 5 6 8 10 9 )

Reading experimental data from the file "rdata.prn" located in the same directory
Check values:

DATA := READPRN( "rdata.prn" )

0
rt := DATA

1
re := DATA

The means of rt and re

nr := length ( rt)

nr = 21

mre := mean( re) mrt := mean( rt)

The least square fit for y=a+bx b := slope ( rt , re) a := intercept ( rt , re)
The least square fit for y=bx, a = 0

b = 0.975

a = 0.032

b = 1.014

rt rt

1
re
b rt

The error terms i

mrt = 0.705

1.5

2. Check of experimental
data taken from the file
"rdata.prn"
and check of the least
square fit - slope b

0.5

:= ln( )

re
:=

b rt

re rt

b :=

mre = 0.719

Figure 1. Experimental data and

the line re = b rt.

0.5

1.5

rt

m := mean( )

Characteristics of :

s2 :=

nr var ( )
nr 1

Check values:

V := exp( s2) 1

Coeff. of variation of error terms V

V = 0.107

Coefficient of variation of the model variables X1, X2, ...:

For example:

VX1 := 0.08 VX2 := 0.05

Model coeff. of variation: Vrt := VX12 + VX22

Vrt = 0.094
2

The total coefficient of variation

Vr := V + Vrt

Vr = 0.143

Standard deviations

Qrt := ln Vrt + 1
Weighting factors:

Q := ln V + 1
rt :=

Qrt
Q

:=

Q := ln Vr + 1
Q

Q = 0.142
rt = 0.663

A - 39

= 0.752

Annex A - Basic statistical concepts and techniques

3. Coefficients of fractile estimation given in EN 1990, Annex D

5% fractile V unknown

ks ( n ) := qt ( 0.95, n 1) 1 +

5% fractile V known, appr.

Check values:

k( n ) := qnorm( 0.95, 0 , 1) 1 +

0,1 % fractile V unknown ds ( n ) := qt ( 0.999, n 1) 1 +

0,1 % fractile V known, appr.

n := 3 , 3.5.. 30

ks ( 10) = 1.923

1
n2

k( 10) = 1.745

1
ds ( 10) = 4.507

n
1

d( n ) := qnorm( 0.999, 0 , 1) 1 +

d( 10) = 3.278

n2
4. Characteristic and design values (relative values related to the mean

V := 0.0, 0.001.. .4

Note that the range variable V is generally used for the coefficient of variation of a single variable
V and for a model investigation Vr, these may be different, for example V = 0,12 and Vr = 0,142.
Single variable, V unknown: rk=ks*mr, rd=dsmr
ks ( n , V) :=

exp( ks ( n ) ) ln 1 + V

ds ( n , V) :=

1+ V

1+ V

Single variable, V known: rk= k*mr, rd= d*mr

k( n , V) :=

exp( ds ( n ) ) ln 1 + V

exp( k( n ) ) ln 1 + V

d( n , V) :=

1+ V

exp( d( n) ) ln 1 + V
2

1+ V

Model, V unknown: rk= ks*rx, xd=dsmr, weighting factors rt and taken from the
above experimental data re and rt (for V < 0,4 approximately Q =~ V):

ksmod ( n , V) := exp 1.65rt + ks ( n ) + 0.5 V V

2

dsmod ( n , V) := exp 3.09rt + ds ( n ) + 0.5 V V

2

Check values:
ksmod ( nr , Vr) = 0.774
dsmod ( nr , Vr) = 0.608

Model, V known: rk=k *mx, rd= dmr, (for V < 0,4 approximately Q =~ V):
2
( 2
)
2
2
dmod( n , V) := exp( 3.09rt + d( n ) + 0.5 V) V

kmod( n , V) := exp 1.65rt + k( n ) + 0.5 V V

dmod( 10, 0.2) = 0.516

Check values:

5. Estimates of the partial factors M

Single variable, V unknown Ms ( n , V) :=

kmod( 10, 0.2) = 0.696

ks ( n , V)
ds ( n , V)
k( n , V)

Single variable, V known

M( n , V) :=

Model, V unknown

Msmod ( n , V) :=

Model, V known

Mmod( n , V) :=

d( n , V)
ksmod ( n , V)
dsmod ( n , V)
kmod( n , V)
dmod( n , V)

A - 40

Ms ( 10, 0.1) = 1.294

M( 10, 0.1) = 1.165
Msmod ( nr , Vr) = 1.273
Mmod( nr , Vr) = 1.234

Annex A - Basic statistical concepts and techniques

6. The relative characteristic and design values, the model values estimated using
weighting factors determined above from experimental data given in the file "rdata.prn"
1

0.9

ks ( 10 , V)
ds ( 10 , V)

0.8

k( 10 , V)
d( 10 , V)

0.7

ksmod ( 10 , V)
dsmod ( 10 , V)

0.6

kmod( 10 , V)
dmod( 10 , V)

0.5

0.4

0.1

0.2

0.3

0.4

Figure 2. Variation of the characteristic and design values with coefficient of variation V for n = 10.
The char. value of X:

Example

rks( n , V) := b mrt ks ( n , V)

rks( 21, 0.12) = 0.575

The model char. value of X: rksmod( n , V) := b mrt ksmod ( n , V) From data: rksmod( nr , Vr) = 0.553
7. Partial factor M, the model values estimated using weighting factors determined
above from experimental data given in the file "rdata.prn"
2

1.8
Ms ( 10 , V)
M ( 10 , V)

1.6

Msmod ( 10 , V)
M mod( 10 , V)

1.4

1.2

0.1

0.2
V

0.3

0.4

Figure 3. Variation of the partial factor M versus coefficient of variation V for n=10.

A - 41

Annex A - Basic statistical concepts and techniques

8. Variation of ks and Ms with n and V for the model values estimated using weighting
factors determined above from experimental data given in the file "rdata.prn".

ks ( 3 , 0.0) = 1

ks
Figure 4. Variation of the characteristic values ks with n and V
.
Ms ( 20, 0.1) = 1.208

Ms
Figure 5. Variation of the partial factor M with n and V .

A - 42

Annex B Elementary methods of structural reliability II

ANNEX B ELEMENTARY METHODS OF STRUCTURAL

RELIABILITY II
Milan Holick1 , Ton Vrouwenvelder2 and Angel Arteaga3
1

Klokner Institute, Czech Technical University in Prague, Czech Republic

2
Delft University of Technology, TNO BOUW, The Netherlands
3
Institute of Construction Sciences E. Torroja, CSIC, Madrid, Spain

Summary
Using basic principles of the reliability theory described in Chapter II Elementary
methods of structural reliability I, the operational techniques for estimating partial factors of
basic variables are derived and applied to common permanent and variable loads. The
described computational procedures are illustrated by a number of numerical examples, which
are supplemented by MATHCAD, EXCEL and MATHEMATICA sheets.
1

INTRODUCTION

1.1

Background materials
Fundamental concepts and procedures of structural reliability are well described in a
number of national standards, in the new European document EN 1990 [1] and International
Standard ISO 2394 [2]. Additional information may be found in the background document
developed by JCSS [3] and in recently published handbook to EN 1990 [4]. Guidance on
application of the probabilistic methods of structural reliability may be found in publications
and working materials developed by JCSS [5] and in relevant literature listed in [4] and [5].
Guidance on structural systems and time dependent reliability may be found in [6] and [7].
1.2

General principles
The theory of structural reliability considers all basic variables as random quantities
having appropriate types of probability distribution. Different types of distributions should be
considered for actions, material properties and geometrical data. In addition, model
uncertainties of actions and resistance models should be taken into account. Prior theoretical
models of basic variables and procedures for probabilistic analysis are indicated in JCSS
documents [5].
This appendix is a direct extension of Chapter II "Elementary methods of structural
reliability" of the main text to which reference is frequently made.
2

DESIGN POINT

Theoretical principles can be utilized in estimation of essential reliability elements

(partial factors, reduction factors, combination rules, etc.) used in operational standards
including Eurocodes that are based on a partial factor method. To explain how theoretical
findings are transmitted into design recommendations a graphical representation of the

B-1

Annex B Elementary methods of structural reliability II

random variables E and R and the corresponding limit state function (2) of Chapter II R - E =
0 is used similarly as in Annex C of EN 1990 [1].
Figure 1 (adopted from [1]) shows the random variables E and R in a two-dimensional
diagram, where the horizontal axis indicates the fraction R/R, the vertical axis the fraction
E/E. It is assumed that E and R are mutually independent variables having a normal
distribution. As indicated in Examples 5 and 6 of Chapter II such an assumption might not be
entirely realistic and should be considered as an approximation only. However, generally any
distribution may be transformed to the normal distribution (at least in some domain) and,
therefore, the random variables E and R in Figure 1 may be considered as transformed
variables having originally other type of distribution.
Figure 1 also shows the limit state function (failure boundary) R - E = 0, which
corresponds to expression (2) of Chapter II transformed to the coordinates used in Figure 1.
Note that the failure boundary would be the diagonal of the main axes if the standard
deviations of R and E have the same magnitude, R = E. The safe (desirable) domain of the
variables R and E, where condition (1) of Chapter II is satisfied, is located under the failure
boundary, the failure (undesirable) domain lies above the boundary R E = 0.

Limit state function R E = 0

Design point
(Rd /R, Ed /E)

E / E

R / R
R

Figure 1. Design point Rd, Ed.

In verification of structural reliability any point on the failure boundary R E = 0
could be considered as the critical (design) point (as apparent from the historical development
of design methods described in previous chapter of this Handbook 2). However, it has been
proved (see for example [5]) that the best option (assuring consistency and invariance of the
solution for a different formulation of the limit state function and different basic variables) is
the point (Rd, Ed) closest to the mean (E, R), which is indicated in Figure 1. Accepting this
finding, it follows directly from Figure 1 that the design point coordinates (Rd, Ed) may be
written in the form
Rd = R RR

(1)

Ed = E EE

(2)

Here E and R denote the so-called sensitivity factors of variables E and R. The minus
signs in equations (1) and (2) are conventionally accepted in the documents CEN [1] and ISO
[2].

B-2

Annex B Elementary methods of structural reliability II

It follows from Figure 1 that the sensitivity factors E and R (direction cosines of the
failure boundary) can be written as

E = E / E + R

(3)

R = R/ E + R

(4)

In Eurocodes an approximation of these sensitivity factors by fixed values is further accepted

2
2
R = R / E + R = 0,8

(5)

2
2
E = E / E + R = 0,7

(6)

The validity of such an approximation is delimited by means of a condition for the ratio of the
standard deviations in the form
0,16 < E / R < 7,6

(7)

When this condition is not satisfied, then the sensitivity factor = 1,0 is
recommended to be used for the variable having a greater standard deviation. Let us remark
that this simplification is on the safe side as the sum of squared direction cosines should be
equal to one.
The design values Ed and Rd of the variables E and R are thus defined as fractiles of
normal distribution
P(E > Ed) = U(+E) = U(-0,7)

(8)

P(R < Rd) = U(-R) = U(-0,8)

(9)

where U(u) denotes a standardized normal distribution. If = 3,8, then the design values ed
and rd are fractiles corresponding approximately to probabilities 0,999 and 0,001. Note that in
equation (9) the use of the symmetry of normal distribution is taken into account, i.e. of the
relationship 1 - U(+E) = u(-E).
When the load or resistance model contains several basic variables (other loads, more
materials, geometrical data), equations (8) and (9) holds only for the leading variables (the
most significant for the studied condition of reliability). For other (accompanying) variables
the requirements on design values are reduced and it holds
P(E > Ed) = U(+0,4E) = U(-0,28)

(10)

P(R < Rd) = U(-0,4R) = U(-0,32)

(11)

When = 3,8 the design values of accompanying (non-leading) variables are fractiles
approximately corresponding to probabilities 0,9 and 0,1.
Design values are the upper fractiles (for actions) or the lower fractiles (for
resistance), corresponding to certain probabilities of being exceeded (actions) or not reached
(resistance). For leading variables, the probabilities are given by the distribution function of
the normal standardized distribution for values u = +E and -R, in the case of non-leading
variables for reduced values u = +0,4E and 0,4R. These probabilities (for the lower
fractile approximately 0,001 for leading and 0,1 for accompanying variables) then serve to
determine the design values even for those variables, which do not have normal distribution.
Let us note that according to the general principles, it is necessary in the case of upper
fractiles (actions) to consider the complementary probabilities (close to the value 1).

B-3

Annex B Elementary methods of structural reliability II

Example 1.
The design values Ed and Rd of variables E and R from Example 4 will be assessed
assuming that the reliability index = 3,8, E = -0,7 and R = 0,8. According to equation (8),
it holds for E that
P(E > ed) = U(E) = U(-2,66) = 0,0039
The complementary probability is therefore 0,9961 and we obtain from equation
ed = - (0,45 + 0,78ln(-ln(p))) = 50 (0,45 + 0,78ln(-ln(0,9961)))10 = 88,75
Note that when the normal distribution is assumed, it is given
ed = + up = 50 + 2,66 10 = 76,6
According to equation (9), it holds for R
P(R < Rd) = U(-R) = U(-3,04) = 0,0012
For the lognormal distribution with the mean of 100 (units) and standard deviation of
10 (units) it follows
Rd exp(unorm,p V) = 100 exp(-3,040,10) = 73,79
For normal distribution we obtain
Rp = + up = 50 3,04 10 = 69,6
Obviously, it holds for the coordinates of the design point that ed > rd and the tie rod
does not satisfy the condition (1) of Chapter II (we know from the Example 4 of Chapter II
that is only 3,09). In order to satisfy the condition for a reliability index of 3,8, the
parameters of variables E and R would have to be modified.
The attached MATHCAD sheets StRod.mcd, DesVRod.mcd may be used to make all
numerical calculations.
3

PARTIAL FACTORS

3.1

Material properties
The above described reliability concepts may be used to assess the partial factors. The
attached MATHCAD sheets GammaRGQ.mcd cover all computational procedures described
below and may be used to make additional numerical calculations.
In accordance to EN 1990 [1] or ISO 2394 [2] the partial factor R of material
resistance R is defined as the fraction of its characteristic value Rk and the design value Rd,
thus

R = Rk / Rd

(12)

It is further assumed that the characteristic value Rk of a resistance variable R is

defined as its 5% fractile [1], [2] and [5].
If a resistance variable R (strength) has a normal distribution, then the characteristic
value Rk is given as
Rk = R 1,645 R = R(1 1,645 VR)

B-4

(13)

Annex B Elementary methods of structural reliability II

The design value Rd of R can be estimated using the above derived equation (22) (see
also documents [1] and [2]), thus
Rd = R R R = R 0,8 R = R(1 0,8 VR)

(14)

In equation (13) and (14) R denotes the mean, R the standard deviation, VR the coefficient of
variation and R = 0,8 the sensitivity factor of R.
Taking into account equations (13) and (14) it follows from (12) that the partial factor
R for a normal distribution of R can be assessed as

R = (1 1,645 VR) / (1 0,8 VR)

(15)

Assuming a lognormal distribution of R its characteristic value Rk can be determined

[1], [2] using approximate equation
Rk = R exp ( 1,645 VR)

(16)

Similarly the design value Rd is approximated [1], [2] as

Rd = R exp (R VR)

(17)

Taking into account equations (16) and (17) it follows from (12) that the partial factor R for a
lognormal distribution can be assessed as

R = exp ( 1,645 VR) / exp (R VR)

(18)

Figures 2 and 3 show the variation of the partial factor R of the material property R
with the reliability index for selected values of the coefficient of variation VR given for
normal distribution by equation (15) (Figure 2) and lognormal distribution by equation (18)
(Figure 3).

2,5

2,0

VR = 0,20
0,15
0,10
0,05

1,5
1,0
0,5
0

Figure 2. Variation of R with for selected coefficients of variation VR = 0,05; 0,10; 0,15
and 0,20, and for normal distribution of R.
Generally the partial factor R increases with increasing . The increase of R is
considerably greater in the case of normal distribution (Figure 2) than in the case of lognormal
distribution (Figure 3). The effect of the type of distribution is particularly obvious for

B-5

Annex B Elementary methods of structural reliability II

coefficients of variation VR greater than 0,10. A considerable effect of the type of distribution
on the theoretical value of partial factors can be expected also for other basic variables, in
particular for actions.
2,5

R
2

VR = 0,20
0,15
0,10
0,05

1,5

0,5
0

Figure 3. Variation of R with for selected coefficients of variation VR = 0,05; 0,10; 0,15 and
0,20, and for lognormal distribution of R.

3.2

Permanent load
Consider a self-weight G having a normal distribution. Similarly as in the case of
material property when a reference period T instead of the design working life Td is used in
the reliability verification of a structure, then the design value of G should be determined for
T instead of Td. It is assumed that the characteristic value Gk of G is defined as the mean G
[1], [2] and [5]:
Gk = G

(19)

The design value Gd is given by equation (23) of Chapter II (see also documents [1], [2]) as
Gd = G G G = G + 0,7 G = G(1 + 0,7 VG)

(20)

In equation (20) G denotes the mean, G the standard deviation, VG the coefficient of
variation and G = 0,7 the sensitivity factor of G.
The partial factor G of G is given as [1], [2]

G = Gd / Gk

(21)

Taking into account equations (19) and (20) it follows from (21) that

G = (1 +0,7 VG)

(22)

Figure 4 shows the variation of the partial factor G with the reliability index for
selected values of the coefficient of variation VG = 0,05; 0,10; 0,15 and 0,20. Note that G =
1,35 (recommended in EN 1990 [1]) corresponds approximately to the reliability index =
3,8 if the coefficient of variation is about 0,1 (the value recommended in EN 1990 [1] was
further increased by 5% to take into account model uncertainty).

B-6

Annex B Elementary methods of structural reliability II

2,5

G
2

VG = 0,20
0,15
0,10
0,05

1,5

0,5
0

Figure 4. Variation of G with for selected coefficients of variation VG = 0,05; 0,10; 0,15 and
0,20, and for normal distribution of G.
It follows from Figures 2, 3 and 4 that less significant variation with -values should
be generally expected for the partial factor of self-weight G than for the partial factor of
material property R..
3.3

Variable load
A similar procedure as in the case of permanent load G can be used for estimation of
the partial factors Q for variable loads Q. Assuming Gumbel distribution the characteristic
value (0,98 fractile) is given as
Qk = Q (1 VQ (0,45 + 0,78 ln(ln(0,98))))

(23)

The design value Qd is given as

Qd = Q (1 VQ (0,45 + 0,78 ln(ln(-1(E))))

(24)

In equation (23) and (24) G denotes the mean, VQ the coefficient of variation of annual
extremes of Q and G = 0,7 the sensitivity factor of Q.
The partial factor Q of Q is given as [1], [2]

Q = Qd / Qk

(25)

Figure 5 shows the variation of Q with the coefficients of variation VQ for selected
values of assuming Gumbel distribution of Q. It appears that in case of a variable action Q
the reliability index has a significant effect on the partial factor Q. The following Figure 6
shows the variation of Q with the reliability index for selected values of the coefficients of
variation VQ assuming again Gumbel distribution of Q.
It follows from Figures 5 and 6 that for the reliability index = 3,8 and the coefficient
of variation VQ up to 0,5, the partial factor Q is less than 1,3. However the coefficient of
variation may be also greater than 0,5 and other distribution may be more adequate (see other
Chapters of this Handbook). That is why a conservative value Q = 1,5 is recommended in
EN 1990 [1].
B-7

Annex B Elementary methods of structural reliability II

1.5

= 4,3

1.4

1.3

= 3,8

1.2

= 3,3
1.1

0.9

0.1

0.2

0.3

0.4

0.5

VQ
Figure 5. Variation of Q with the coefficients of variation VQ for selected values of
assuming Gumbel distribution of Q.
2

VQ = 0,4

VQ = 0,3

1.5

VQ = 0,2
1

0.5

2.5

3.5

4.5

Figure 6. Variation of Q with the reliability index for selected values of the coefficients of
variation VQ assuming Gumbel distribution of Q.

B-8

Annex B Elementary methods of structural reliability II

It follows from Figure 6 that for the reliability index 3 the theoretical value of the
partial factor Q is about 1. This is due to the fact that the characteristic value Qk is defined as
0,98 fractile, and when 3 then Qk, is approximately equal to Qd (for more details see
equations (23) and (24)). Additional numerical calculations can be easily obtained using the
attached MATHCAD sheets GammaRGQ.mcd, which include all the above described
computational procedures.
4

THE GENERAL CASE OF RELIABILITY ANALYSIS

4.1

General
In the Chapter II Elementary methods of structural reliability I of this handbook it
has been presented what is called the fundamental case of structural reliability. It is the case in
which the limit state function can be represented by only two random independent variables,
the effect of the actions and the resistance. This fundamental case represents a very interesting
case for introducing the reliability concepts due to its very intuitive reasoning, and that we are
used to this concepts of global action effect and resistance. Also, due to the fact that in two
dimensions only is easy render graphic simple representations.
But, unfortunately, only in a few cases the limit state function may be represented by
the fundamental case, reducing the structural reliability problem to a simple resistance versus
action effect formulation assuming their independents. In most cases, at least, a few more
variables will be needed. In general the resistance is a function of the material(s) properties
and the dimensions of the structure or element and the action effects depend on the various
applied loads and densities and dimensions of the structure. Even, not always the resistance
and action effects can be considered independent: for instance, some dimensions affect both
the actions and resistances. Even more, the action effects may depend on the response of the
structure as a whole (e.g., the dynamic actions).
The limit state function can be written as:
Z(X1, X2, X3, ...) = 0 ;

(26)

Z(X) = 0 ;

(27)

or in vector form:
where X = {X1, X2, X3, ...} is the vector of the random variables depending on time defining
the limit state function.
In this case also, Z(X) > 0 represents the safe region and Z(X) < 0 the unsafe region.
The probability of failure, in this case is obtained as:
Pf = P [Z(X) 0] =

...

fZ(x) dx;

(28)

Z ( X )< 0

where fZ(x) is the joint density probability distribution of the vector of variables X.
In the case of all the basic variables are independent, in a lot of cases we can adopt this
hypothesis as a good approximation, the joint density probability distribution of X is the
product of the marginal density probability distribution of each variable. Therefore, is
possible to put the equation (18) in the form:
Pf = P [Z(X) < 0] =

...

fX1(x1) fX2(x2) ...fXn(xn) dx1dx2...dxn.

Z ( X )< 0

B-9

(29)

Annex B Elementary methods of structural reliability II

4.2

Basic variables
Basic variables are those variables needed to characterize and define the behaviour of
a structure, and its safety, relating to a determined limit state.
The designer has some degree of freedom in order to choose the basic variables: In
general, those used in the conventional design are chosen: dimensions, weights, loads,
material strengths, etc. In general, independent basic variables are considered as any
dependence adds complexity and it is difficult to define the degree of dependence. However,
some variables as the tension, compression strength and the modulus of elasticity of the
material are known to be dependent, but one can, generally, deal with them as they were
independent.
In order to analyse structural reliability it is necessary to characterize basic variables
statistically. That is, to define, at least, their distribution functions, parameters and correlation
matrix. The parameters of distributions can be estimated on the basis of data by means of the
usual statistical techniques: maximum likelihood, methods of moments, etc. The data should
be carefully investigated in order to exclude the outliers, to analyse trends, and so on. A
graphical representation of data and of the model adopted is generally useful.
Guidance concerning distribution functions and their parameters for the models of
common actions and resistances in structural reliability are given in reference [5].
Example 1:
Consider a basic variable and the following experimental data:
data={1.3,3.2,4.3,1.3,5.4,3.7,3.8,4.0,2.9,3.2,4.5,4.0,3.4,2.4,1.8,1.7,2.2,4.1,2.6,4.1,3.3,
3.5,3.7,2.4,2.8,2.5,3.,3.3,2.6,2.9,2.4,2.6,2.9,2.8,3.1,3.2,3.4,3.5,3.2,3.3,3.};
In the following the results obtained with the attached MATHEMATICA package
distribution-fit are used. The following statistical characteristics were obtained:
Number of data
41

Mean
3,105

Variance
0,700

Standard variation
0,837

Coefficient of variation
0,270

We try to fit the normal, lognormal and gamma distributions to data. The parameters
of the distributions are obtained by the method of moments. Assuming the following
intervals: {1.3-2.3, 2.3-2.7, 2.7-3.3, -3.7, 3.7-4.1, 4.1-6}, the Chi square test is performed.
In the figure are represented the histogram and the density distribution function (in
this case the normal distribution). The results obtained for each distribution are shown in the
following table:
normal

0.5
0.4

Distribution Parameters

Estimator
W
Normal
3,103- 0,847 0,612
Gamma
13,40 - 0,231 1,001
Lognormal 1,096 0,268 1,310

Confidence
Level %
96,0 - 99,6
91,0 98,6
86,0 97,1

0.3
0.2
0.1
2

The range of the confidence level is due to the different hypothesis. The moments are
known a priori, then the degrees of freedom are the number of intervals minus one; or

B - 10

Annex B Elementary methods of structural reliability II

obtained from data, then the degrees of freedom are the number of intervals minus three.
Therefore to the same estimator correspond two confidence levels.
It appears that all three distributions fit well the data (normal distribution seems to be
the best). Therefore, it is not clear what distribution to choose; plotting the cumulative
distribution functions and the point values in direct representation and with the ordinates in
double logarithmic scale, then it is possible to have a new insight in the problem.
4.3

Tail sensitivity problem

In statistics, the decisions are, generally, not taken after a mathematical proof. A
hypothesis is accepted if evidence for its rejection is not found. The probability distribution
assigned to any variable can have an important influence on the estimated probabilities of
failure. Besides, when assigning distribution functions, the data available are mainly, of
course, of the central part of the distribution where the election of one or other distribution
function is no very significant. This problem is called "tail sensitivity problem".
This problem is show graphically in Figure 7 where three distribution functions with
the same mean and standard deviation, corresponding to the case of the Example 1: normal,
gamma and log-normal, are represented both at normal scale and with double logarithmic
ordinates. (i.e. z = Log(-Log(y)) ), that is in Gumbels paper, that increases the scale of the
upper tail). At normal scale there are no significant difference among the distributions, but
looking at the Gumbels paper, the differences are little for the 0,95 fractile (i.e. characteristic
value), while the differences are appreciable for the 0,999 fractile (i.e. design value). Also, on
the Gumbels paper may be seen that all distributions fit adequately to the points distribution,
assuming for the ordered points a probability of i/n +1, but in order to estimate the design
value (to say, fractile 0,999) it is necessary to extrapolate, and thus the difference among
distributions is important. The higher point of the data corresponds to approximately to 0,975
fractile (1-1/42).

.999

0.8

.99
.95
.9

0.6
0.4

.5
.1
.01
.001

0.2

Figure 7. Cumulative distribution functions at normal scale and in Gumbels paper.

5

AN EXAMPLE OF REINFORCED CONCRETE SLAB

5.1

General
Various design concepts mentioned above may be illustrated considering a simple
example of a reinforced concrete slab in an office building. The example shows how different
design methods (permissible stresses, global safety factor, partial factor method) treat

B - 11

Annex B Elementary methods of structural reliability II

uncertainties of basic variables by choosing different input (design) values. The example also
indicates significance of the reliability theory in structural design and advantages of the
reliability based partial factor method compared to the other design formats.
Some basic terms (e.g. characteristic strength) and calculation procedures used in this
Section will be properly defined in various Chapters of this Handbook 2. Nevertheless the
following text can be understood at least intuitively.
5.2

A reinforced concrete slab

A simply supported slab having the span of 6 m is exposed to a permanent load (selfweight of the slab and other fixed parts of the building), which is estimated by the
characteristic value (equal to the mean value) gk=7 kN/m2. In accordance with the EN 1991-11 [8] the characteristic value of the imposed load in an office area qk=3 kN/m2 may be
assumed. It is, however, well known that the mean value of this load is considerably lower,
about 0,8 kN/m2.
Further, the concrete C20/25 having the characteristic strength fck=20 MPa (the mean
30 MPa) and reinforcement bars S 500 having the characteristic strength fyk=500 MPa (the
mean 560 MPa) are to be used. Using previous experience, the total height of the slab 0,25 m
(effective depth about 0,25 0,03 = 0,22 m) was specified in advance. Given the above data
verification of the preliminary specifications and estimation of the necessary reinforcement
area of the slab should be done.
5.3

Design of a slab
Consider first a simple drawing of the slab cross-section including simplified stress
distribution diagrams in the compressive concrete zone (rectangular and triangular) as
indicated in Figure 8.
fc

fc

0,8bxfc

0,5bxfc
0,8x

x
d M

As

Asfy

Figure 8. Stress distribution in a reinforced concrete slab.

When the rectangular stress diagram is assumed the following equilibrium conditions
can be written (see Figure 8):
0,8 fc x b = As fy

(30)

As fy(d 0,4 x) = M

(31)

The basic variables used in equations (30) and (31) are evident from Figure 8:
d denotes the effective depth, x the depth of the neutral axis, b the width of the slab
(considered as 1 m), As the area of the reinforcement, fc the concrete strength and fy the
reinforcement strength (yield point). The bending moment in equations (31) is given as

B - 12

Annex B Elementary methods of structural reliability II

M=

( g + q ) L2
8

(32)

L denotes the span of the simply supported beam.

Using equilibrium conditions (29) and (30) the following formula for the
reinforcement area As can be derived:
As
f
2M
= c 1 1
bd f y
f cbd 2

(33)

Without going into technical details note, that equation (33) may be used
approximately for the global safety factor methods and partial safety factor method. Attached
EXCEL sheet RCBeam and MATHCAD sheet RCBeam may be applied to make necessary
calculations.
The classical permissible stresses method assumes a triangular compressive stress
block in the compressive concrete zone, which is also indicated in Figure 8, and linear stress
strain diagram. The reinforcement area As can be found using the attached MATHCAD sheet
RCBeam.mcd, which includes further calculation details.
Note that up to now all the basic variables have been considered as deterministic
quantities without considering any uncertainties that may potentially affect their actual values.
However, it is well recognised that some of the basic variables entering equation (31) might
have considerable scatter (particularly the load effect (bending moment) M, concrete strength
fc, and reinforcement strength fy). On the other hand the geometric data As, b and d seem to be
much less uncertain (almost fixed or deterministic).
To get a first estimate of the area As, one may take the mean (average) values of all the
basic variables involved. Intuitively it is clear that this might be not safe enough and instead
of the mean values somehow conservative (safe) values should be applied. Table 1 indicates
the mean values together with values of basic variables as used in design in accordance with
the rules of above-mentioned design methods.
Table 1. Input design (characteristic) values of loads and material strengths used in design
calculation using different design methods.
Design method
Basic variable
The mean
Permissible Global
Partial
value
stresses
safety factor factor
Permanent load g [kN/m2]
7
7
7
9,45
2
Imposed load q [kN/m ]
0,8
3
3
4,5
Concrete compressive strength fc [MPa] 30
5,5
20
13,3
Reinforcement tensile strength fy [MPa] 560
275
500
435
Table 1 clearly indicates the differences between considered design methods. For
example, the input values of the permanent load g used in design calculation in accordance
with the permissible stresses method and the global safety factor method are equal to the
mean value (7 kN/m2), while the design value in the partial factor method (9,45 kN/m2) is
obtained as a product of the characteristic value and the partial factor G = 1,35. The design
value of imposed load q used in design calculation in accordance with the permissible stresses
method and the global safety factor method are equal to the characteristic value (3 kN/m2),
while the design value (4,5 kN/m2) in the partial factor method is product of the characteristic
value and partial factor Q = 1,5.

B - 13

Annex B Elementary methods of structural reliability II

The input calculation values for the compressive concrete strength fc used by different
design methods are considerably different from its mean (30 MPa). The same is true for the
tensile strength of reinforcement fy. Permissible stresses indicated in Table 1 can be found in
design standards. Obviously they are much lower than the mean value as they are assumed to
take into account all uncertainties including loads. Input strengths in the global factor methods
are equal to the characteristic values. In the partial factor method the design strength of
concrete 13,3 MPa is obtained by dividing the characteristic value 20 MPa by the partial
factor M = 1,5, similarly the design strength of steel 435 MPa is obtained as a fraction of the
characteristic value 500 MPa and the partial factor M = 1,15. Note that the factor 1,9 is used
to enhance the load effect (bending moment M) when the global factor method is used to
specify the reinforcement area As.
Resulting reinforcement areas obtained for all the design methods mentioned above
are indicated in Table 2.
Table 2. A simply supported reinforced concrete slab designed using different historical
methods assuming the span L = 6 m, the height h = 0,25 m (d = 0,22 m) and the loads gk=
7 kN/m2, qk= 3 kN/m2 (the mean 0,8 kN/m2), C20/25 (fck = 20MPa, the mean 30 MPa),
fyk=500 MPa (the mean 560 MPa).
Design method
M [kNm]
The mean value (absurd)
35,1
The permissible stresses
45,0
The global safety factor (s0 =1,9)
45,0
The partial factors method (CEN)
62,8

As [m2]
0,00038
0,00204
0,00082
0,00069

MR [kNm]
35,1
228,9
97,4
82,4

0
8.0
5,0
4,2

pf
0,5
0,4410-16
0,3210-7
0,1210-5

The attached EXCEL sheet RCBeam and MATHCAD sheet RCBeam may be used to
check the resulting data indicated in Table 2. These sheets also indicate further details of
applied computational procedures.
It follows from Table 2 that resulting reinforcement areas vary within a broad range
from As = 0,00038 m2 (the mean value estimate) up to As = 0,00204 m2 (the permissible stress
method). The most economic procedure appears to be provided by the partial factor methods
that leads to the lowest reinforcement area As = 0,00069 m2.
5.4

Reliability consideration
The rectangular stress distribution (indicated in Figure 8), assumed in design
calculation according to the global or the partial factor method, is accepted for reliability
investigation of the slab designed by any of the design methods illustrated in Table 2
(including the permissible stress method, which assumes a triangular stress diagram). Thus,
taking into account equilibrium conditions (30) and (31) the limit state function (16) may be
written as
Z ( X ) = As f y (d

As f y
2bf c

) M = As f y (d

As f y
2bf c

( g + q ) L2
8

(34)

Assuming the limit state function (34) the reliability indexes and the failure
probabilities pf can be determined using commercially available software VaP [9] or
COMREL [10] or the MATHEMATICA notebook FORM.nb. An approximate value of the
reliability index and the failure probability pf (with an acceptable accuracy) may be
obtained using the attached MATHCAD sheet RelRCB, which is based on a simple procedure
of numerical integration. The self-content MATHCAD sheet RelRCB is provided with
B - 14

Annex B Elementary methods of structural reliability II

explanatory notes (including information on input theoretical models of basic variables) and
might be intuitively understood without any additional information. Detailed description of
the applied numerical procedure is given in the other chapters of this Handbook 2.
Resulting reliability indexes and the failure probabilities pf are indicated in Table 2.
The partial factor method leads to the reliability level described by = 4,2 (failure probability
1,2 10-5) that is very close to the value = 3,8 (failure probability 7,2 10-5) recommended in
EN 1990 [1]. Slightly more conservative design ( = 5) is provided by the global safety factor
methods (see Table 2). However, the permissible stresses method seems to lead to a rather
uneconomical design ( = 8). Obviously, "the mean value method" proves to be unacceptable
as it yields the lowest reinforcement area As = 0,00038 m2 (reinforcement ratio 0,0022 only)
corresponding to = 0 and high probability of failure of Pf = 0,5.
6

ASSESMENT OF THE FAILURE PROBABILITY IN GENERAL CASE

6.1

General
There are several different procedures for assessing failure probability in a general
case of more variables:

Analytical: Only in a few, very simple cases it is possible to find the analytical correct
solution. Depends on the variables vector, all must be independent and normally
distributed, and on the limit state region, it must be defined by hyper-planes. It cannot
be considered as a general solution.
Numerical: It is an exact solution in the sense that we can get, in principle, all the
precision we need. The simple trapezoidal rule of integration gives, in general good
results if there are not too many variables (4 or 5). The complexity of integration
increases exponentially with the number of variables.
Monte Carlo methods: The Monte Carlo simulation techniques are bases on the
random sampling of the variables and carry a long number of artificial experiments.
The use of this approach is increasing notably nowadays, due to bigger power and
speed of the computers. The crude application of the method lead to the same
difficulties explained above. Variance reduction and importance methods are
employed to avoid those difficulties.
First and Second Order Reliability Methods (FORM and SORM): These approximate
methods give iterative algorithms that allow to obtain the reliability index, using a
linear or quadratic approximation to the limit state surface in the point of minimum
distance to the mean point of the variables.

6.2

First and Second Order Reliability Methods (FORM and SORM)

Hassofer and Lind developed an algorithm invariant for the formulation of the limit
state function. To obtain the reliability index the following steps have to be followed:
1
2
3
4

Define the limit state function.

Characterize statistically the basic variables; i.e.: mean, standard deviation,
distribution function and correlation matrix.
Transform the set of basic variables into a set of independent variables (by means
of the Rosenblatts transformation, for instance).
Standardize the set of basic variables by the transformation XU, such that
E (U) = 0 , and CoV[U,UT] = 1 .
B - 15

(35)

Annex B Elementary methods of structural reliability II

Obtain the length of the minimum distance vector from the new origin point to the
tangent hyper-plane of the limit state surface, referred to the new variables, in the
intersection point of that vector with the limit state surface.
6 Obtain the design point, (X1d, X2d,...,Xnd) and the sensitivity coefficients, 1,
2,...,n, the unitary cosines of that vector.
Figure 1 represents the space of the fundamental case, with only two variables: the
action effect and the resistance. The origin of the vector corresponds to the mean point of
the variables X, i.e.: (X1, X2,...Xn). The minimum distance vector from this point to the limit
state surface is perpendicular to the hyper-plane (a straight line in the two variables case)
tangent to the limit state surface in the design point (xd1,xd2,...xdn). The sensitivity coefficients
represent the influence of the corresponding variable in the failure probabilities.
The following equations hold for the sensitivity coefficients
i 1;

i2 = 1

(36)

Usually these values i, are taken as positives if the correspond to a resistance variable
and negatives in the case of effects of action variables.
When the limit state surface is highly non-linear, the error of substituting the surface
by the tangent hyper-plane in that point can be important. In those cases a minor error is
obtained if the limit state surface is substituted by the tangent quadratic surface. That is:
taking the square term in the Taylor's series. In this case the method is called Second Order
Reliability Method or SORM.
As it is said, the reliability index and the probability of failure pf are related by the
formula:
Pf = (-);

=-1(1- Pf );

(37)

-1

where () stands for the standard normal distribution and () for its inverse.
Example 2
a) Design of a steel beam
Simply supported beam: IPE 240 S235
g

Span
Cross section area:
Resistance modulus,
Yield stress

L = 6,0 m
A = 39,12 10-4 m2
W = 3 243 10-6 m3
fy = 235 MPa

Actions:
Permanent load: gk = 7,0 kN/m
Variable load: qk = 3,0 kN/m
Limit State function
Z(X) = 1 Wfy - 2 (g + q) L2 /8
In the following table all the variables are described by the mean, standard deviation
and the distribution function, see also [5]:

B - 16

Annex B Elementary methods of structural reliability II

Variable
Span
Resistance modulus
Resistance model
Yield strength
Action effects model
Permanent load
Variable load

Symbol
L [m]
W [m3]
1 [-]
fy [MPa]
2 [-]
g [kN/m2]
q [kN/m2]

Mean
6
3 243 10-6
1
280
1
0.007
0 .0008

Standard deviation
0
0
0,1
19,6
0,2
0.007 0,1
0. 00080,6

Distribution function
Deterministic
Deterministic
Lognormal
Lognormal
Lognormal
Normal
Gumbel

The following results are obtained using the attached MATHEMATICA notebook
Form.nb and package Level II.m.
Case a: Results
Reliability index: = 3,82 ; Probability of failure: (-) = 6,6710-5
Variable
Sensitivity coefficient

1
-0,392

fy
-0,275

2
0,778

g
0,304

q
0,270

From the results can be seen that the reliability index is well adjusted, but also that the
limit state is very sensitive to the actions model uncertainty (influence coefficient 0,778).
Case b:
Imagine that we can study more deeply the actions and we arrive to the conclusion
that the mean and variances of the actions are the same as before, but now we have reduced
the uncertainty model and we have now a coefficient of variation of 10%.
Action effects model

2 [-]

0,1

Lognormal

Performing a new study with the new values we obtain:

Reliability index: = 5,04 ; Probability of failure: (-) = 2,3710-7
Variable

fy

Sensitivity coefficient

-0,505

-0,354

0,505

0,332

0,503

We can see that the influence coefficients are more equilibrated, but now the
reliability index is quite high.
Case c:
Perhaps we can reduce the steel profile. We can try with a smaller profile:
New steel profile: IPE 220, W= 252 cm3
Now the results are the following:
Reliability index: = 3,74; Probability of failure: (-) = 9,2910-5

B - 17

Annex B Elementary methods of structural reliability II

Variable

Sensitivity coefficient -0,524

fy

-0,367

0,524

0,383

0,411

Example 3
Consider the slab of the example in Section 5 (see Figure 8). From the equilibrium
consideration was obtained:
0,8 fc x b = As fy
As fy(d 0,4 x) = M
Using equilibrium conditions (30) and (31), and considering uncertainty of the models
of resistances, 1, and actions, 2, the following formula for the limit state function can be
derived:

As f y

g(X) = 1 A s f y d
2 fcb

2 ( g + q )L 2 / 8

The following table shows the variables that are considered in the study using the
MATHEMATICA notebook FORM.nb:
Variable

Symbol

Resistance model
Reinforcement tensile strength
Reinforcement area
Effective height
Concrete compressive strength
Load model
Permanent load
Imposed load
Span

Mean value

1 [-]
1
fy [MPa]
560
As [m2] 0,00069
d [m]
0,23
fc [MPa]
30
2 [-]
1
7
g [kN/m2]
2
q [kN/m ]
0,8
L [m]
6

Standard
deviation
0,1
30
0,0000345
0,01
5,5
0,2
0.7
0,48
-

Distribution
function
Lognormal
Lognormal
Normal
Normal
Lognormal
Lognormal
Normal
Gamma
Deterministic

Results:
Reliability index: = 3,56; Probability of failure: (-) = 1,8710-4
Variable
1
As
fy
d
fc
2
g
q
Sensitivity coefficient -0,383 -0,193 -0,201 -0,177 -0,018 0,761 0,300 0,274
From the results it is possible to draw the following conclusions: the reliability index
is a little bit lower; the coefficient of influence of the concrete resistance, fc, is almost zero, it
could be possible to consider it as deterministic without influencing the results; and the
uncertainty in the actions has the maximum influence.
Case b:
The new study is performed considering reduced uncertainty in the load model as
follows:
Load model

2 [-]

B - 18

0,1

Lognormal

Annex B Elementary methods of structural reliability II

Results:
Reliability index: = 4,63; Probability of failure: (-) = 1,8210-6
Variable
Sensitivity coefficient

1
-0.500

As
-0,259

fy
-0,263

d
-0,235

fc
-0,023

2
0,500

g
0,347

q
0,431

Case c:
Now the reliability index is a bit higher. Reducing the cross section of the reinforcing
steel bars:
Reinforcement area
Load model

As [m2]
2 [-]

0,00059
1

0,000590,05
0,1

Normal
Lognormal

Results:
Reliability index: = 3,86; Probability of failure: (-) = 5,6110-5
Variable
1
Sensitivity coefficient -0,504

As
-0,260

fy
-0,266

d
-0,234

fc
-0,020

2
0,504

g
0,368

q
0,401

From both the examples can be seen that reducing the uncertainty in the model of
actions, is possible to reduce the steel section without reducing (even increasing) the overall
reliability.
7

SYSTEM RELIABILITY

7.1

General
Even in the simplest case of one structural element: a beam or column, more than one
ultimate limit state function has to be considered: failures at positive or negative moments or
shear of the beam. In most cases, the structure has multiple elements what is called a
'structural system'.
The reliability of the system depends on the reliability of its elements: the effects of
actions in each element depend on applied loads, loads and resistances may not be
independent, there may be a correlation between the properties of the elements in different
parts of the structure. Furthermore, there are limit states for the structure as a whole like the
overall deflection or foundation settlement.
When all the different failure modes are identified, a 'fault-tree' or an 'event-tree of all
the failures modes can be established.
Example 4
Consider the simple portal frame of the following Figure 4 submitted to the horizontal
and vertical loads Q. Assuming a plastic behaviour, the frame has three possible modes of
collapse. In each path different plastic hinges will be formed:
a)
b)
c)

Sway mode: Plastic hinges formed at sections 1 and 3;

Beam mode: Plastic hinges at sections 1, 2 and 3;
Combined mode: Plastic hinges at sections 2 and 3.

B - 19

Annex B Elementary methods of structural reliability II

Q
1

The failure in any path implies the failure of the structure and the event 'structural
failure' will be the union of the all n failure modes (n = 3, in the example case). And therefore
the probability of failure of the structure will be:
Pf = P(FS) = P(F1 F2 ... Fn)

(38)

where Fi is the event failure in mode i. For each mode m nodes or structural elements have to
fail. So the probability of failure of each mode, Fi, will be:
Pf(Fi) = P(F1i F2i ... Fmi)

(39)

where Fji is the event failure of the j-th element or node in the i-th failure mode. To reach the
collapse of the structure in the mode i, m elements or nodes must fail.
In general the structure is idealized as a parallel system, a series system or a
combination of both.
7.2

Parallel systems
In the parallel system are the elements so interconnected that the failure of one or
more of the elements does not mean the failure of the whole structure. Such a structure is
called a redundant structure. This redundancy can be active, when the redundant members can
activate before the limit state of any member was reached, or passive, when the redundant
members only act when a limit state is reached in some member.
It should be taken into account that any hyperstatic structure is not necessary a parallel
system: if the elements are brittle, the failure of any of them can lead to a new distribution of
stresses such that new failures are reached immediately.
The failure of a pure parallel system with m components is given by:

Pf sys = P

I
1

Fj = P I (Zj < 0)

(40)

where Fj is event of the failure of the j-th component with Zj limit state function. Thus:
P(Fj) =P(Zj < 0) (-)

(41)

The probability of failure of the system by FORM is given by:

Pf sys = m(- ; )

(42)

where m is the multi-dimensional standard normal distribution, is the vector of the

component reliability indices and the mm correlation matrix of the reliability indices,
given by
jk =

ij ik ;

with j, k = 1, 2, ... , m

(43)

and ij is the sensitivity factor of the i-th variable in the j-th component limit state function.

B - 20

Annex B Elementary methods of structural reliability II

It is a demanding task to calculate the probability of failure of the system. An upper
bound of this probability can be obtained as:
m

pf sys = Min P(Fj

j , k =1

IF )]

(44)

A simple approximation for only two elements:

pf sys = (- 1) (- *2); with *2 = (2 - 1) /

(45)

7.3

Series systems
A series system is that the failure of any element leads to the collapse of the structure.
It is called a 'weakest link'. Any statically determined structure is a series system. In a similar
manner that was shown in the previous section, the failure probability in a pure series system
with m components is given by:

C = P

U
1

Fj

(46)

The probability of failure of the system given by FORM is:

Pf sys = 1- m( ; )

(47)

Simple bound to this probability are given by:

m
m
Max P( F j ) Pf sys Min P( F j ),1 .
1
1

(48)

These bounds are usually too wide. A method to find more precise bounds can be
found in [11].
8

CONCLUDING REMARKS

Elementary methods of structural reliability can be used to assess the reliability of

fundamental cases of two random variables when the limit state function is formulated as the
difference between the resulting structural resistance and load effect. Software products,
usually based on the methods FORM and SORM, must be used in a general case of more
basic variables.
Basic principles of the reliability theory provide operational techniques that can be
used for estimating the partial factors of basic variables. The assessment of various reliability
measures in the new structural design codes is, however, partly based on historical and
empirical experiences. Obviously the past experience depends on local conditions including
climatic actions and traditionally used construction materials and, consequently, might be in
different countries considerably diverse. That is why number of reliability elements and
parameters in the present suite of European standards are open for national choice.
MATHCAD sheets, EXCEL sheets and MATHEMATICA notebooks supplement the
numerical examples and make it possible to recalculate similar examples of basic structural
members exposed to common permanent and variable loads.

B - 21

Annex B Elementary methods of structural reliability II

REFERENCES
[1] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[2] ISO 2394 General principles on reliability for structures, ISO 1998.
[3] JCSS: Background documentation, Part 1 of EC 1 Basis of design, 1996.
[4] Gulvanessian, H. Calgaro, J.-A. Holick, M.: Designer's Guide to EN 1990, Eurocode:
Basis of Structural Design; Thomas Telford, London, 2002, ISBN: 0 7277 3011 8, 192
pp.
[5] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/, 2001.
[6] Melchers R.E.: Structural Reliability Analysis and Prediction. 2nd ed. John Wiley & Sons,
Chichester, 437 p, 1999.
[7] Ditlevsden, O. and Madsen, H.O. 1966. Structural Reliability Methods. John Wiley,
Chichester, 372 p
[8] EN 1991-1-1 Eurocode 1 Actions on structures. Part 1-1 General actions. Densities, selfweight, imposed loads for buildings, CEN 2002
[9] VaP, Variable Processor, version 1.6, ETH Zurich, 1997.
[10] COMREL, version 7.10, Reliability Consulting Programs, RCP MUNICH, 1999.
ATTACHMENTS
1. MATHCAD SHEET "GammaRGQ.mcd"
MATHCAD sheet "GammaRGQ.mcd" is intended for calculation of the partial
factorsR, R and R assuming selected theoretical models.
2. MATHCAD SHEET "Prindex.mcd"
MATHCAD sheet "Prindex.mcd" is intended for determining the reliability index
from the failure probability.
3. MATHCAD sheet "RCBeam.mcd"
MATHCAD sheet "RCBeam.mcd" is intended for design of a reinforced concrete
beam exposed to permanent and variable loads.
4. MATHCAD sheet "RelRCB.mcd"
MATHCAD sheet "RelRCB.mcd" is intended for reliability analysis of a reinforced
concrete beam exposed to permanent and variable loads.
5. EXCEL sheet "RCBeam.xls"
EXCEL sheet "RCBeam.xls" is intended for design of a reinforced concrete beam
exposed to permanent and variable loads.
6. MATHEMATICA notebook "Fit_distribution.nb"
MATHEMATICA notebook "Fit_distribution.nb" is intended for fit selected
theoretical models to experimental data.
7. MATHEMATICA notebook "FORM.nb"
MATHEMATICA notebook "FORM.nb" is intended for reliability analysis of a
structural member.
8.

MATLAB package "Level2.m"

MATLAB package "Level2.m" is intended for determining the reliability index using
FORM method.
B - 22

Annex B Elementary methods of structural reliability II

9.

MATHCAD sheet "FORM2.mcd"

MATHCAD sheet "FORM2.mcd" is intended for calculation of the reliability index
and failure probability assuming function g(X)= R - E = 0 assuming general three parameter
lognormal distribution LN(,,) of E and R.
10.

MATDCAD sheet "FORM7.mcd"

MATHCAD package "FORM7.mcd" is intended for calculation of the reliability index
and failure probability assuming a non-linear limit state function g(X) and general three
parameter lognormal distribution LN(,,) for basic variables.
11.

EXCEL sheet "FORM7.xls"

MATHCAD package "FORM7.mcd" is intended for calculation of the reliability index
and failure probability assuming a non-linear limit state function g(X) and general three
parameter lognormal distribution LN(,,) of basic variables

12.

MATLAB function "FORM7.m"

MATLAB package "FORM7.m" is intended for determining the probability of failure
pf, assuming a pre-defined limit state function of seven basic variables.
13.

MATLAB function "Lnden (x, mu, sigma,sk)"

MATLAB function "Lndens" is intended for calculation of the probability density
function of three-parameter lognormal distribution. The function is called by the function
FORM7 using command Lndens(ske,me,se), and returns the value of probability density
function.
15.

MATLAB function " Lndist (x, mu, sigma,sk)"

MATLAB function "Lndist" is intended for calculation of the distribution function of
three-parameter lognormal distribution. The function is called by the function FORM7 using
command Lndist(skr,mr,sr), and returns the value of the distribution function.
15.

MATLAB function Ndens (x, mu, sigma)

MATLAB function "Ndens" evaluates the one-dimensional normal density function.
The function is called by the function FORM7 using command Ndens(x) (or
Ndens(x,mu,sigma) or Ndens(x,mu)), and returns the value of the inverse distribution
function.
16.

MATLAB function "Ndinv (p, mu, sigma)"

MATLAB function "Ndinv" calculates the inverse distribution function of the normal
distribution (determining the reliability index beta). The function is called by the function
FORM7 using command Ndinv(p) (or Ndinv(p,mu,sigma) or Ndinv(p,mu)), and returns
the value of the inverse distribution function.
17.

MATLAB function "Ndist (x, mu, sigma)"

MATLAB function "Ndist" evaluates the one-dimensional normal distribution
function. The function is called by the function FORM7 using command Ndist(x) (or
Ndist(x,mu,sigma) or Ndist(x,mu)), and returns the value of the inverse distribution
function.

B - 23

Annex B Elementary methods of structural reliability II

Attachment 1 - MATHCAD sheet GammaRGQ.mcd

GammaR, gammaG and gammaQ

assuming theoretical model
MATHCAD sheet for determination of partial factors R, G and Q.
Study parameters: reliability index and coefficient of variation: := 0 , 0.1.. 5 V := 0 , 0.1.. 0.5

1 Coeficients and factors used in EN 1990

Coefficient for 5% fractile

k := 1.65

Standardised normal variable

Coefficient for 0,1 % fractile

d := 3.09

Not dicrectly used in this sheet

Sensitivity factors:

R := 0.8

E := 0.7

Reduced values:

R( ) := R

E( ) := E

FORM factors assumed in EN 1990

Basic reliability index = 3.8

2 Characteristic values (relative values related to the mean, k = xk / x)

Normal distribution

kn( V) := ( 1 k V)

Lognormal distribution

kln( V) :=

Check:

exp( k) ln 1 + V

kn( 0.1) = 0.835

kln( 0.1) = 0.844

1+ V
Gumbel distribution

kgum( V) := 1 V ( 0.45 + 0.78ln( ln( 0.98) ) )

kgum( 0.35) = 1.908

3 Design values (relative values related to the mean, d = xd / x )

Normal distribution

dn ( , V) := ( 1 R( ) V)

Lognormal distribution

dln( , V) :=

dn ( 3.8, 0.1) = 0.696

exp( R( ) ) ln 1 + V

dln( 3.8, 0.1) = 0.735

1+ V
dgum ( , V) := 1 V ( 0.45 + 0.78ln( ln( 1 pnorm( E( ) , 0 , 1) ) ) )

Gumbel distribution

dgum ( 3.8, 0.1) = 1.387

4 GammaR for resistance assuming normal and lognormal distribution

Rn( , V) :=

kn( V)

Rln( , V) :=

dn ( , V)

kln( V)

1.5

1.5

Rn( , 0.05)
Rn( , 0.1)

Rn( 3.8, 0.1) = 1.2

dln ( , V)

Rln( , 0.05)
Rln( , 0.1)

Rn( , 0.15)

Rln( , 0.15)

0.5

0.5

B - 24

Annex B Elementary methods of structural reliability II

5 GammaG for permanent load assuming normal distribution
Gn( , V) := 1 E( ) V
Rn( 3.8, 0.1) = 1.2

Gn( , 0.05)

1.4

Gn( 3.3 , V)

Gn( , 0.1)
Gn( , 0.15)

Gn( 3.8 , V)
1.2

Gn( 4.3 , V)

0.2

0.4
V

6 GammaQ for variable load assumingGumbel distribution

dgum( , V)

kgum( V)

Qgum( , V) :=

Partial factor Q:

dgum( 3.8, 0.5) = 2.937

Qgum( 3.8, 0.5) = 1.279

2
Qgum( , 0.2)

Qgum( 3.3 , V)

Qgum( , 0.3)

Qgum( 3.8 , V)

Qgum( , 0.4)

Qgum( 4.3 , V)

1.4

1.2

0.2

0.4
V

Note. Calculation procedures applied in this sheet for determination of Gamma

factors do not take into account model uncertainties of relevant variables. An
additional magnifying factor of a magnitude from 1.05 to 1.10 is considered in
Eurocode recommendations.

MH, 21.2.2003

B - 25

Annex B Elementary methods of structural reliability II

Attachment 2 - MATHCAD sheet PrIndex..mcd

B - 26

Annex B Elementary methods of structural reliability II

Attachment 3 - MATHCAD sheet RCBeam.mcd

B - 27

Annex B Elementary methods of structural reliability II

B - 28

Annex B Elementary methods of structural reliability II

Attachment 4.

MATHCAD sheet RelRCB.mcd

B - 29

Annex B Elementary methods of structural reliability II

B - 30

Annex B Elementary methods of structural reliability II

B - 31

Annex B Elementary methods of structural reliability II

Attachment 5. EXCEL sheet RCBEAM.XLS

REINFORCED CONCRETE BEAM OR SLAB

Partial factor or global safety factor only
Bending moment
L[m]=
gk[kN/m]=
qk[kN/m]=
ME[kNm]=

6,00
7,00 gammaG=
3,00 gammaQ=
62,78

1,35
1,5
L

Rectangular cross-section
Dimensions
Factors
Concrete
Rebars
Reratio
=x/d<

Estimate

b= 1
d= 0,17
c= 1,5
cc= 1
s= 1,15
fcd=acc*fck/c= 13,3
fctm= 2,2
fck [MPa] = 20
fyd=fyk/= 434,8
fyk [MPa] = 500
min [%] =
0,130
max
0,45
x
z~
0,9 d
As~Md/(z*fyd)=
0,000944
A[m^2]= 0,000933
As
[%] =
0,55
>min ? PRAVDA
b
=x/d=
0,22
<max?
1,10
PRAVDA max[%] =

General Table
m = Md/
/(b*d^2*fcd)
0,00
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40

=
1,8

Reratio
[%]
0,00000
0,15737
0,32376
0,50091
0,69124
0,89821
1,12714
1,38698
1,69521

As
f
= cd
f yd
bd

=x/d z/d = 1 - 0,4*

0,00
0,06
0,13
0,20
0,28
0,37
0,46
0,57
0,69

1,00
0,97
0,95
0,92
0,89
0,85
0,82
0,77
0,72

s
[%]
5,11
2,30
1,36
0,89
0,61
0,41
0,27
0,16

As
m^2
0,00000
0,00027
0,00055
0,00085
0,00118
0,00153
0,00192
0,00236
0,00288

c=0,35%
d

Md
[kNm]
0,0
19,3
38,5
57,8
77,1
96,3
115,6
134,9
154,1

1 1 2 M E , min 0,0013 , max 0,011

f cd bd 2

1,6
1,4
1,2
1,0
0,8
0,6
0,4
0,2
0,0
0,00

0,05

0,10

0,15

0,20

0,25

M/(fcd*b*d^2)

B - 32

0,30

0,35

0,40

Annex B Elementary methods of structural reliability II

REINFORCED CONCRETE BEAM OR SLAB

Partil factor or global safety factor method
Bending moment
L[m]=
gk[kN/m]=
qk[kN/m]=
ME[kNm]=

1.35
1.5
L

Rectangular cross-section
Dimensions
Factors
Concrete
Rebars
Reratio
=x/d<

Estimate

6.00
7.00 gammaG=
3.00 gammaQ=
62.78

b= 1
d= 0.17
c= 1.5
cc= 1
s= 1.15
acc*fck/c= 13.3
fck [MPa] = 20
fctm= 2.2
yd=fyk/= 434.8
fyk [MPa] = 500
min [%] =
0.130
max
0.45
x
z~
0,9 d
As~Md/(z*fyd)=
0.000944
A[m^2]= 0.000933
As
[%] =
0.55
>min ? PRAVDA
b
=x/d=
0.22
<max?
PRAVDA max[%] =
1.10

c=0,35%
d

General Table
m = Md/
/(b*d^2*fcd)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40

Reratio
[%]
0.00000
0.15737
0.32376
0.50091
0.69124
0.89821
1.12714
1.38698
1.69521

As
f
= cd
bd
f yd

1.8

=x/d

/d = 1 - 0,4*

0.00
0.06
0.13
0.20
0.28
0.37
0.46
0.57
0.69

1.00
0.97
0.95
0.92
0.89
0.85
0.82
0.77
0.72

s
[%]
5.11
2.30
1.36
0.89
0.61
0.41
0.27
0.16

As
m^2
0.00000
0.00027
0.00055
0.00085
0.00118
0.00153
0.00192
0.00236
0.00288

Md
[kNm]
0.0
19.3
38.5
57.8
77.1
96.3
115.6
134.9
154.1

1 1 2 M E , min 0,0013 , max 0,011

f cd bd 2

1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.00

0.05

0.10

0.15

0.20

0.25

M/(fcd*b*d^2)

B - 33

0.30

0.35

0.40

Annex B Elementary methods of structural reliability II

Attachment 6 MATHEMATICA notebook Fit_distribution.nb

Fit data to distribution functions

Statistical characterization of data. It takes the following steps:
1- Obtain the characteristics of datas: number of elements, mean, variance, standard deviation and coefficeint of
variation
2- By the method of moments obtain the indicated distribution parameters
3-Manually the values of the steps in the hystogram are introduced, and then perform the Chi square test of goodness
of fit and draw the hystogram and the adjusted density distribution function. Gives two values of Confidence level, the
first one corresponding to a degrees of freedom equal number of intervals minus 1 and the second one number of
intervals -3.
4-Draw the data points asignig a probability of p =m/(n+1). m = number of order, n= number of data and the cumula tive distribution function, both in real scale and double logarithmic scale (i.e., z = -Log(-Log x))
Distributions defined: normal or gauss, lognormal or lognor, gamma, weibull, and extrems or gumbel
Last updated 2 May 2004

Data
Packages needed
$TextStyle

= 9FontSize 14, FontFamily

.> "Arial " <;

Data
data = {1.3, 4.3, 1.3, 5.4, 3.7, 3.8, 4., 2.9, 3.2, 4.5, 4., 3.4, 2.4,
1.8, 1.7, 2.2, 4.1, 2.6, 4.1, 3.3, 3.5, 3.7, 2.4, 2.8, 2.5, 3., 3.3, 2.6,
2.9, 2.4, 2.6, 2.9, 2.8, 3.1, 3.2, 3.4, 3.5, 3.2, 3.3, 3.};
distributions = 9"normal", "lognormal", "gamma"<;
numdistributions = LengthAdistributionsD;
Descriptivestatistic
ndata = LengthAdataD
mean = MeanAdataD
variance = VarianceAdataD
deviation = StandardDeviationAdataD
deviation
coefvariation =
mean
40
3.1025
0.718199
0.847466
0.273156

Distribution functions
Parameters of distribution

B - 34

Annex B Elementary methods of structural reliability II

DoBIfBdistributionsPiT === "normal" distributionsPiT === "gauss",
par1 = mean;
par2 = deviation;
distAiD = NormalDistributionApar1, par2D,
IfBdistributionsPiT === "lognormal" distributionsPiT === "lognor",
mean2 + deviation2
w = ;
mean2
mean2
par1 = .5LogB F;
w
!!!!!!!!!!!!!!!!
par2 = LogAwD ;
distAiD = LogNormalDistributionApar1, par2D,
IfBdistributionsPiT === "gamma",
mean2
par1 = ;
deviation2
deviation2
par2 = ;
mean
distAiD = GammaDistributionApar1, par2D,
IfBdistributionsPiT === "weibull",
par1 = a .
GammaA1 + 1aD
mean2
FindRootB 2 == 2, 9a, .1<
deviation
GammaA1 + 2aD . GammaA1 + 1
D
a
mean
par2 =
;
GammaB1 + par1
1F
distAiD = WeibullDistributionApar1, par2D,
IfBdistributionsPiT === "extrems" distribucionesPiT === "gumbel",
!!!!
6 EulerGamma deviation
par1 = media . NB F;

!!!
deviation 6
par2 = NB F;

distAiD = ExtremeValueDistributionApar1, par2D,

PrintA"Error: the distribution ", i, " H", ditributionsPiT, "is not defined
BreakADFFFFF, 9i, numdistributions<F

Quantiles
DoA
PrintAdistAiDD;
PrintA
NAQuantileAdistAiD, 0.5D, 4D, " ",
NAQuantileAdistAiD, 0.75D, 4D, " ",
NAQuantileAdistAiD, 0.9D, 4D, " ",
NAQuantileAdistAiD, 0.95D, 4D, " ",
NAQuantileAdistAiD, 0.99D, 4D, " ",
NAQuantileAdistAiD, 0.999D, 4D, " ",
NAQuantileAdistAiD, 0.9999D, 4DD;
PrintA" "D,
9i, 1, numdistributions<D
N o r m a l D i s t r i b u t i o nA3 . 1 0 2 5, 0 . 8 4 7 4 6 6D

3 . 1 02 5 3 . 6 7 4 1 1 4 . 1 88 5 7 4 . 4 9 6 4 6 5 . 0 7 4 5 . 7 2 1 3 7 6 . 2 54 2 4

L o g No r m a l D i s t r i b ut io nA1 . 0 9 6 2 3, 0 . 2 6 8 2 5 7D
2 . 9 92 8 5 3 . 5 8 6 4 5 4 . 2 20 7 5 4 . 6 52 8 1 5 . 58 6 1 2 6 . 8 5 6 5 4 8 . 1 1 6 33

B - 35

Annex B Elementary methods of structural reliability II

G a m m a D i s t r i b u t i o nA1 3 . 4 0 2 3, 0 . 2 3 1 4 9D

3 . 0 25 6 9 3 . 6 2 4 5 4 4 . 2 26 4 . 6 1 5 1 8 5 . 4 0 5 9 6 . 3 8 8 9 4 7 . 2 76 9 8

Goodness of Fit Test

xmax = 1.5MaxAdataD;
xmin = .9 MinAdataD;

Defining the intervals

Adjust the intervals length in order to obtain more or less the same number of samples in each interval (minimum
samples)
intervals = 92.3, 2.7, 3, 3.3, 3.7, 4.1, xmax<;
count = RangeCountsAdata, intervalsD
9 5, 7, 5, 6, 7, 5, 5, 0<
frecuencies = DropAcount, . 1D;
numintervals = LengthAintervalsD;
PrependToAintervals, xminD;

total = Plus @@ frecuencies;

hystogr = 99intervalsP1T, 0.<<;

Draws the histogram

frecuenciesPiT
DoBAppendToBhystogr, 9intervalsPiT, =F;
total HintervalsPi + 1T . intervalsPiTM
frecuenciesPiT
AppendToBhystogr, 9intervalsPi + 1T, =F,
total HintervalsPi + 1T . intervalsPiTM
9i, numintervals<F
AppendToAhystogr, 9intervalsPnumintervals + 1T, 0.<D;
hystogram = ListPlotAhystogr, PlotJoined True, DisplayFunction IdentityD;

Performs of Chi square test and draw the density function

DoBestadW = 0.;
DoBnj = frecuenciesPjT;
ej = HCDFAdistAiD, intervalsPj + 1TD . CDFAdistAiD, intervalsPjTDM total;
Hnj . ejM2.
estadW = estadW + ,
ej
9j, numintervals<F;
alfa = H1 . CDFAChiSquareDistributionAnumintervals . 1D, estadWDM 100.;
alfa1 = H1 . CDFAChiSquareDistributionAnumintervals . 3D, estadWDM 100.;
PrintAdistAiDD;
PrintA"Estimator W = ", estadW, "
Degrees of freedom = ", numintervals . 1D;
PrintA"Confidence level between ", alfa1, " and ", alfa, " %"D;
PrintA" "D;
fdensity = PlotAPDFAdistAiD, xD, 9x, intervalsP1T + .001, intervalsPnumintervals +
DisplayFunction IdentityD;
ShowA9hystogram , fdensity<,
PlotLabel distributionsAAiDD,
DisplayFunction $DisplayFunctionD,
9i, numdistributions<F;
NormalDistributionA3.1025, 0.847466D
Estimator W = 0.635298

Degrees of freedom = 6

C o nf i d e n c e le v e l bet w e e n 9 5 . 9 0 6 2 a n d 9 9 . 5 7 8 2 %

B - 36

Annex B Elementary methods of structural reliability II

normal
0.5
0.4
0.3
0.2
0.1
2

L o g N o r m a l D i s t r i b u t i o nA1 . 0 9 6 2 3, 0 . 2 6 8 2 5 7D

E s ti m a t o r W = 1 . 3 1 8 28

D e g r e e s o f f re e d o m = 6

C o n f i de n c e l e v e l b et w e e n 8 5 . 8 2 6 7 a n d 9 7 . 0 6 4 %

lognormal
0.5
0.4
0.3
0.2
0.1
2

GammaDistributionA13.4023, 0.23149D
Estimator W = 1.00414

Degrees of freedom = 6

Confidence level between 90.9168 and 98.5455 %

gamma
0.5
0.4
0.3
0.2
0.1
2

Distribution functions
Distribution of the data points
datosord = SortAdataD;
i
probab = TableB , 9i, 1, ndata<F;
ndata + 1.
loglogprob = .LogA. LogAprobabDD;

B - 37

Annex B Elementary methods of structural reliability II

points1 = 99datosordP1T, probabP1T<<;
points2 = 99datosordP1T, .LogA. LogAprobabP1TDD<<;
DoApoints1 = AppendToApoints1, 9datosordPiT, probabPiT<D, 9i, 2, ndata<D;
DoApoints2 = AppendToApoints2, 9datosordPiT, loglogprobPiT<D, 9i, 2, ndata<D;
p1 = ListPlotApoints1,
Prolog 9PointSizeA.02D<,
DisplayFunction Identity, PlotRange AllD;
pp3 = ListPlotApoints2,
Prolog 9PointSizeA.02D<,
DisplayFunction Identity,
PlotRange AllD;
p3 = ShowApp3,
Prolog 9PointSizeA.02D<,
DisplayFunction IdentityH,
AxesOrigin 90.,.1.5<,
Ticks 9Automatic,
99..834,".1"<,9.366,".5"<,92.25,".9"<,
92.97,".95"<,94.6,".99"<<<M
D;
Draws in normal and double-logarthmic scale
DoAfdistrib1AiD = PlotACDFAdistAiD, xD, 9x, xmin, xmax<,
PlotStyle 9DashingA9.008 i1.5 , .02i<D<,
DisplayFunction IdentityD,
9i, 1, numdistributions<D;
DoAfdistrib2AiD = LogPlotACDFAdistAiD, xD, 9x, xmin, xmax<,
PlotStyle 9DashingA9.008 i1.5 , 0.02i<D<,
DisplayFunction IdentityD,
9i, 1, numdistributions<D;
DoAfdistrib3AiD = PlotAEvaluateA.LogA. LogACDFAdistAiD, xDDDD, 9x, xmin, xmax<,
PlotStyle 9DashingA9.008 i1.5 , 0.02i<D<,
DisplayFunction IdentityD,
9i, 1, numdistributions<D;
ShowA9p1, fdistrib1A1D, fdistrib1A2D, fdistrib1A3D<,

AspectRatio 1.2,
GridLines 9None, 90.95<<,
DisplayFunction $DisplayFunctionD;

B - 38

Annex B Elementary methods of structural reliability II

1

0.8

0.6

0.4

0.2

ShowA9p3, fdistrib3A1D, fdistrib3A2D, fdistrib3A3D<,

AspectRatio 1.2,
AxesOrigin 9xmin, .3<,
PlotRange 99xmin, xmax<, 9. 3, 7<<,
GridLines 9None, 92.97, 6.9<<,
Ticks 9Automatic,
99.1.93, ".001"<, 9. 1.527, ".01"<, 9. .834, ".1"<, 9.366, ".5"<,
92.25, ".9"<, 92.97, ".95"<, 94.6, ".99"<, 96.9, ".999"<
<<, H 99.21,".9999"<M
DisplayFunction $DisplayFunctionD;

.999

.99
.95
.9

.5
.1
.01
.001
2

B - 39

Annex B Elementary methods of structural reliability II

Attachment 7 MATHEMATICA notebook FORM.nb

FORM
This notebook compute the reliability index, failure probability and influence factors in level II, using the package
"Reliability`Level2". In this package those variables are determined through the algorithm "Normal Tail Approxima
tion" as is explained in the book of Madsen et al.: Methods of Structural Safety, pp. 94 and following.
The failure function of the limit state must be defined and, also, the independent basic variables given by a matrix
with a row for each variable with the kind of distribution function assumed, the mean and the standard deviation.
The following names of distribution functions are implemented:
"normal" or "gauss",
"lognormal" or "lognor",
"gamma",
"logistic",
"uniform",
"weibull", and
"extremes" or "gumbel".
The iterations defined in the loop of the algorithm is stopped when the differences between two consecutive values i
less than the error defined.

Needs["Reliability`Level2`"]
LImit State functio
Define the limit state function
gAz_D :=
zAA1DD HzAA2DD zAA3DDM HzAA4DD . HHzAA2DD zAA3DDM H2 zAA5DD bMMM .
zAA6DD HzAA7DD + zAA8DDM L2 8

Deterministic
variable
b = 1;
L = 6.;

Ramdomvariable
Define the variables matrix: distribution function, mean and standard deviation
m = {{"lognormal",
1,.1},
{"normal", 0.00082 ,0.00082*0.05},
{"lognormal",
560,560*0.054},
{"normal",0.22,0.01}
,
{"lognormal",
30,30*0.167},
{"lognormal",
1.2,.18},
{"normal", 0.007,0.007*.1},
{"gumbel", .0008,.0008*0.6
}
};

Failprob[m,g]
Obtain the reliability index beta ; the failure probabilities; and the sensitivity factor for each variable
b e t a = 3 . 8 0 0 0 4, P h iH .b e t aM = 7 . 2 3 3 6 7 1 0 .5 H 5 i t e r a t i o n sM
a l p h a = 9 .0 . 4 4 2 3 8 8, . 0 . 2 2 3 4 5 5, .0 . 2 3 0 8 9 8, . 0 . 2 1 6 8 0 7, .0 . 0 2 5 8 5 8 5, 0 . 6 6 1 5 3 5, 0 . 3 3 4 9 9 6, 0 . 3 2 1 7 4 7<

B - 40

Annex B Elementary methods of structural reliability II

Attachment 8 MATLAB package Level2.m
BeginPackage["Reliability`Level2`"]
Needs["Statistics`ContinuousDistributions`"]
Failprob::usage =
"Work out the Hassofer-Lind reliability index and the
coefficients of sensitivity using the algorithm Norman Tail
Approximation given in STRUCTURAL SAFETY by
Madsen et al., p94 . It admits a maximum of 10 variables "
Characterization::usage =
" Characterize the distribution functions given by a matrix "
Factores::usage =
" Gives the sensitivity factors"
Begin["`Private`"]
Off[Part::partw]
Grad[s_, var_List] := D[s, #]& /@ var
invnormal[y_]:=Sqrt[2.] InverseErf[2y-1];

(* Definition of Grad*)

Failprob[matrix_,g_]:=
Module[
{
n=First[Dimensions[matrix]],
znew=muiter=Transpose[matrix][[2]],
sigmaiter=Transpose[matrix][[3]],
distnor=NormalDistribution[0.,1.],
error= 10.^(-3),
(* error admited *)
betanterior=0,
ind=0,
z,aux,gradiente,gradpart, ceta,muzeta,sigmazeta,alpha,beta,prob },
z=Array[a,n];
gradiente=Grad[g[z],z] ;
Characterization[matrix]; (* Characterization of the variables *)
Do[
(* iterative cycle *)
ziter=znew;
partic=Take[{z[[1]]->ziter[[1]],z[[2]]->ziter[[2]],z[[3]]->ziter[[3]],
z[[4]]->ziter[[4]],z[[5]]->ziter[[5]],z[[6]]->ziter[[6]],
z[[7]]->ziter[[7]],z[[8]]->ziter[[8]],z[[9]]->ziter[[9]],
z[[10]]->ziter[[10]]},n];
(* iteration values *)
Do[
aux = invnormal[CDF[dist[i],ziter[[i]]]];
sigmaiter[[i]] = PDF[distnor,aux]/PDF[dist[i], ziter[[i]] ];
muiter[[i]] = ziter[[i]] - aux sigmaiter[[i]],

B - 41

Annex B Elementary methods of structural reliability II

{i,n}];
gradpart=gradiente/.partic;
ceta=gradpart . ziter;
muzeta=gradpart . muiter;
sigmazeta=Sqrt[(gradpart^2).(sigmaiter^2)];
alpha = -(gradpart*sigmaiter)/sigmazeta;
beta = -(ceta-g[ziter]-muzeta)/sigmazeta;
znew = muiter + beta(alpha * sigmaiter);
prob=CDF[distnor,-beta];
ind++
If[Abs[beta-betanterior] < error,Break[],betanterior=beta],
{10}];
Print["beta = ",N[beta,5], ", Phi(-beta) = ",ScientificForm[prob]," ( \
",ind," iterations)" ];
Print["alpha = ",N[alpha,3] ]
]
Factores:=
Module[
{alpha,znew,muiter,sigmaiter},
Print["alpha = ",N[alpha,4]];
Print["znew = ",N[znew,4]];
x=(znew-muiter)1./sigmaiter;
Print["x = ",N[x,4]]
]
Characterization[matrix_]:=
Module[
{w,par1,par2,a,
n=First[Dimensions[matrix]]
},
Do[
If[matrix[[i,1]]==="normal"||matrix[[i,1]]==="gauss",
par1=matrix[[i,2]];
par2=matrix[[i,3]];
dist[i]=NormalDistribution[par1,par2],
If[matrix[[i,1]]==="lognormal"||matrix[[i,1]]==="lognor",
w=(matrix[[i,2]]^2 + matrix[[i,3]]^2)/matrix[[i,2]]^2;
par1=.5Log[matrix[[i,2]]^2/w];
par2=Sqrt[Log[w]];
dist[i]=LogNormalDistribution[par1,par2],
If[matrix[[i,1]]==="gamma",
par1=matrix[[i,2]]^2/matrix[[i,3]]^2;
par2=matrix[[i,3]]^2/matrix[[i,2]];
dist[i]=GammaDistribution[par1,par2],
If[matrix[[i,1]]==="logistic",
par1=matrix[[i,2]];
par2=matrix[[i,3]] Sqrt[3.]/Pi;
dist[i]=LogisticDistribution[par1,par2],

B - 42

Annex B Elementary methods of structural reliability II

If[matrix[[i,1]]==="uniform",
par1=matrix[[i,2]]-Sqrt[3.] matrix[[i,3]] ;
par2=matrix[[i,2]]+Sqrt[3.] matrix[[i,3]] ;
dist[i]=UniformDistribution[par1,par2],
If[matrix[[i,1]]==="weibull",
par1=a /. FindRoot[ matrix[[i,2]]^2/ matrix[[i,3]]^2==
Gamma[1+1/a]/(Gamma[1+2/a]Gamma[1+1/a]^2),{a,.1}];
par2=matrix[[i,2]]/Gamma[1+1/par1];
dist[i]=WeibullDistribution[par1,par2],
If[matrix[[i,1]] === "extremes" || matrix[[i,1]] ==="gumbel",
par1 = matrix[[i,2]] - N[(Sqrt[6]*EulerGamma*matrix[[i,3]])/Pi];
par2 = N[matrix[[i,3]]*Sqrt[6]/Pi];
dist[i] = ExtremeValueDistribution[par1, par2],
Print["Error: the distribution ", i ," (",matrix[[i,1]],
") is not defined. "];
Return[]]]]]]]],
{i,n}]
]
End[]
EndPackage[]\.1a

B - 43

Annex B Elementary methods of structural reliability II

Attachment 9 MATHCAD sheet FORM2.mcd
MATHCAD sheet "FORM2.mcd", FORM for g(X)= R - E = 0 assuming
general three parameter lognormal distribution LN(,,) of E and R
1. Parameters for E and R: Highlighted regions

E := 50

wE := 0.2

Default distribution of R - two parameter lognormal R := 100 wR := 0.1

2. Parameter C and skewness :
Distribution parameter C
given by the skewness :
Distribution bound x0
( - 6 for zero ):

C( ) :=

x0( , , ) :=

2
+ 4+
3

C( )

E := 1.14
R := 3 wR + wR

R := wR R

Check:

2
+ 4

C( 0) = 0

14

if 0

x0( R , R , R) = 8.527 10
x0( E , E , E) = 22.522

6 otherwise

Transformation of parametersm( , ) := ln( C( )

E := wE E
3

) + ln( ) ( 0.5) ln( 1 + C( ) 2)

s ( ) :=

ln 1 + C( )

3. Probability density of E and R (for any ):

E( x) :=

dlnorm[ sign( E) ( x x0( E , E , E) ) , m( E , E) , s( E) ] if E 0

dnorm( x, E , E) if E = 0

R( x) :=

dlnorm[ sign( R) ( x x0( R , R , R) ) , m( R , R) , s( R) ] if R 0

dnorm( x, R , R) if R = 0

4. Distribution function of E and R (for any ) :

E( x) := 0.5 ( 1 sign( E) ) + sign( E) plnorm [ sign( E) ( x x0( E , E , E) ) , m( E , E) , s ( E) ] if E 0
pnorm ( x, E , E) if E = 0
R( x) :=

0.5 ( 1 sign( R) ) + sign( R) plnorm [ sign( R) ( x x0( R , R , R) ) , m( R , R) , s ( R) ] if R 0

pnorm ( x, R , R) if R = 0
xR := xE Number of iterations n := 1.. 4

5. FORM iteration process: Guess values:xE := E + 0.005 ( R E)

n :=

n =

xR xR
xE xE

3.304

for i 1 .. n

2.904

Re

dnorm( qnorm( R( xR) , 0 , 1) , 0 , 1)

2.904

R( xR)

2.904

pf := pnorm ( , 0, 1)

Re xR Re qnorm( R( xR) , 0 , 1)
dnorm( qnorm( E( xE) , 0 , 1) , 0 , 1)

Ee

pf n =

E( xE)

4.7710-4

Ee xE Ee qnorm( E( xE) , 0, 1)

aR

aE

1.84310-3

Re Ee

1.84410-3

( Re2 + Ee2)

0.5

1.84410-3

Re

2
2
Re + Ee

3.4

0.5

Ee

( Re2 + Ee2) 0.5

3.2

xR Re + aR Re
2.8

xE Ee + aE Ee

3
n

B - 44

Annex B Elementary methods of structural reliability II

Attachment 10 MATHCAD sheet FORM7.mcd

MATHCAD sheet "FORM7" for calculation of the reliability index and

failure probability assuming a non-linear limit state function
g(X) = a0 + a1*X1( a2*X2 + a3*X3) + a4*X4 (a5*X5+a6X6+a7X7)
, , ) of basic
and general three parameter lognormal distribution LN(
variables X1, X2, X3, X4, X5, X6 and X7
A General three-parameter lognormal distribution for any
1. Parameter C and skewness :
3

Distribution parameter C
given by the skewness :

+4+

C( ) :=

x0( , , ) :=

Distribution bound x0
( - 6 for zero ):

Check:

+ 4

C( 0) = 0

C( )

if 0

x0( 0 , 1 , 1) = 3.1038

6 otherwise

2. Probability density and distribution function (for any ):

u ( x, , ) :=

Standardised variable:

ln u ( x, , ) +

uu ( x, , , ) :=

Transformed standardised variable:

+ ln C( ) 1 + C( ) 2

if 0

C( )

sign ( ) ln 1 + C( )

u ( x, , ) otherwise

uu ( 50, 50, 10, 0) = 0

Density probability function:

dnorm ( uu ( x, , , ) , 0 , 1)

( x, , , ) :=

u ( x, , ) +

C( )

2
ln( 1 + C( ) )

dnorm( uu ( x, , , ) , 0 , 1)

if 0

( 50, 50, 1 , 0) = 0.3989

otherwise

( x, , , ) := pnorm( uu ( x, , , ) , 0 , 1)

Distribution function:

( 100, 100, 10, 0) = 0.5

B FORM method for determination of the reliability index and probability pf

Coefficients a0, a1, a2, a3, a4, a5, a6 and a7 of the limit state functions and
Input parameters for basic variables {X}={X1, X2, X3, X4, X5,X6 and X7}
0
0.
1
1
Three parameter


lognormal distribution
2
1
LN(, ,) for any , if

1
3
= 0 then the normal
distribution is used . Index := 4 a := 1


When X1 and X4 are
5
1
model uncertainties

1
then LN(,) is used.
6

7

Check of the bounds

The check of the
initial gues values

15

1
0.05

10
1
:=

0.05
10
5

0.01
4

x0 = 8.8818 10

x0 = 8.3195 10

x := i
i

0
1.

100
10
:=

1
50
10

0.01
4

x :=
1

x0 = 0
5

4 4( 5 5
6 6
a (a x + a x )
1 2 2
3 3

0
0.15

0.301
0.301
:=

0.15
0.608
1.14

0
x0 = 3.7388
7 7)

a + a x a x + a x + a x

B - 45

i := 1 .. 7

x0 := x0 i , i , i
i

x0 = 8.3195 10
1

14

x0 = 8.5265 10
2

x0 = 0.05
7

x = 0.5455
1

187

Annex B Elementary methods of structural reliability II

4. FORM iteration procedure: Probability of failure pf is determined from the reliability index
Number of iterations
The initial guess
values of X
The value x2 is
calculated
from g (X)= 0

n :=

n := 1 .. 5

x
x
1

(5

a + a x a x + a x + a x

4 4

(2

a a x + a x
1

7 7

6 6

3 3

for j 1 .. n
for i 1 .. 7

( (i
) ) )
( x , i , i , i)
i
ei x ei qnorm ( ( x , i , i , i) , 0 , 1)
i
i

Equivalent
normal
distributions of
the basic
variables X.

x ei
i

u
i

Standardised variables

dnorm qnorm x , i , i , i , 0 , 1 , 0 , 1

ei

ei

(2

g a a x + a x e1
1

Derivatives of g(X):

3 3

g a a x e2
2

1 2 1

g a a x e3
3

1 3 1

(5

g a a x + a x e4
4

6 6

g a a x e5
5

4 5 4

g a a x e6
6

4 6 4

g a a x e7
7

Reliability index:

4 7 4

( g u )

Sensitivity factors:

(gg)

0.5

for i 1 .. 7
g
aa
i

New design point

to be used in the
next iteration, go
back to the
section 5 and use
this data in a new
run

Iteration of the
reliability index

Probability of
failure pf

(gg)

0.5

x ei aa ei
i
i
x
1

(5

a + a x a x + a x + a x

4 4

(2

6 6

a a x + a x
1

3 3

7 7

0
2.9022

2.9303

=
2.9292
2.929

2.9289

B - 46

pf := pnorm 5 , 0, 1

pf = 1.7006 10

Annex B Elementary methods of structural reliability II

Attachment 11 EXCEL sheet FORM7.xls
EXCEL sheet RORM7.xls - FORM iterative computation of the reliability index beta
Limit state function g(X)= a0+a1*X1*( a2*X2+a3*X3)+a4*X4*( a5*X5+a6*X6+a7*X7)

A
i
0
1
2
3
4
5
6
7

0
1
2
3
4
5
6
7

Basic variables Xi are approximated by general three-parameter lognormal distribution

LN(mu,sigma, skew), which becomes automatically normal distribution when skew = 0.
Note that:
skew = 3sigma/mu for two-parameter lognormal distribution
skew = 1.14 for Gumbel distribution
skew = 2sigma/mu for Gamma distribution
B C
D
E
F
G
H
I
J
K
L
Input data
The initial x1 is automatically calculated, do not change x1
ai
Xi
Guess of x2
80,00
Lognormal of Xi
0
mu
sigma skew
C
x0 Design p.
u
uu
phi
PHI
1
1,000
0,050 0,150 0,050 0,00
0,963 -0,74 -0,73 6,3395 0,2319
1 100,000 10,000 0,301 0,100 0,00
65,451 -3,45 -4,20 0,0000 0,0000
1
0,010
0,010 0,301 0,100 -0,09
0,009 -0,07 -0,02 40,2671 0,4918
-1
1,000
0,050 0,150 0,050 0,00
1,035 0,71 0,72 5,9511 0,7642
1 30,000 10,000 0,000 0,000 infinity 54,089 2,41 2,41 0,0022 0,9920
1
5,000
2,000 1,140 0,364 -0,496
5,727 0,36 0,53 0,1581 0,7015
1
1,000
1,000 1,140 0,364 -1,748
1,061 0,06 0,24 0,3914 0,5943
Requirement for the design point, skewi>0 then x0i<xi, when skewi<0 then x0i>xi
Iteration of the FORM method - enter the new x manually instead of the initial x
SensitivityNew point
Partial factors
ai
xi
Equivalent normalDeriv.
x/mu
0 mue
sigmae ui
gi
ui*gi alpha
New x
1
0,998
0,048 -0,73 3,15 -2,308
0,240 0,963
0,963
1 75,000
6,529 -1,46 6,286 -9,194
0,479 65,451
Enter the 0,655
1
0,010
0,010 -0,02 0,01 -2E-04
0,001 0,009
new xi
0,948
-1
0,998
0,052
0,72 -3,09 -2,228
-0,236 1,035
to x2
1,035
1 30,000 10,000
2,41 -10,4 -24,94
-0,788 54,089
to x7
1,803
1
4,567
2,195
0,53 -2,27 -1,201
-0,173 5,727
in H16
1,145
1
0,825
0,991
0,24 -1,03 -0,245
-0,078 1,061
to H21
1,061
Sum
13,13 -40,12
1,000
1 beta in iter. n 3,0551 Macro procedure
Number of iter.
Required acc.
0,001 beta in iter. n-1 3,0551 1,8E-05 Accuracy reached
Probability of failure pf=PHI(-beta) - command =NORMDIST(-G34;0;1;1)
1,1E-03
The whole iteration procedure may be performed manually following instruction given
in the cells K27 to K32 or using Macro "ITERATION".
Note that the initial gueas of variables X2 to X7 should be enteredto the cells H16 to H21.

B - 47

Annex B Elementary methods of structural reliability II

Attachment 12 MATLAB package FORM7.m
function pf = FORM7
% DESCRIPTION, 28.02.2005
% FORM7a evaluates the probability of failure pf considering the limit state function
%
%
g(X)=A0 + A1*X1*(A2*X2+A3*X3) + A4*X4*(A5*X5+A6*X6+A7+A7)
%
% FUNCTIONS USED
% LNDENS(x,ske,me,se), LNDIST(x,skr,mr,sr), NDENS(x), NDIST(x) and NDINV(p)
% INPUT
% Input data (except A0) are loaded from the files A.dat (coefficients A), and X.dat
(parameters
% of the basic variables X). All the basic variables are characterised by the mean m,
standard deviation s
% and skewness sk (arbitrary). The FORM procedure approximates the basic variables by
general
% lognormal distribution LN(m,s,sk), including normal distribution (for sk=0).
%
% OUPUT
% val : failure probability pf
% VERSION
% MH, Czech Technical University in Prague, Klokner Institute, 28.2.2005
% Initialization
% %loading external data files
load A.dat, load X.dat, %The matrix X can be also defined in the comand window
A0=0;% additive constant (not included in the data file A.dot), default value A0=0. When A0
is different from 0,
% numrical probles may arise. Then a new alternative initial point (for example modifying
resistance) may be choosen.
for i=1:1:7;
x(i)=X(i,1); % Initial guess value of basic variables
end
x(1)=-(A0+A(4)*x(4)*(A(5)*x(5)+A(6)*x(6)+A(7)*x(7)))/(A(1)*(A(2)*x(2)+A(3)*x(3))); %
Initial guess value of x1
% FORM iterations
acc=0.001;delta=1;betap=0;j=0;%required accuracy (may be modified if required)initial
iteration parameters
while delta > acc %for j=1:1:5 %Iteration loop for a given accuracy (5 cycles are usually
sufficient for acc = 0.001)
j=j+1; % The indicator of the number of cycles
for i=1:1:7; % Loop for transformation of original distributions to equivalent normal
ditributions
se(i)=NDENS(norminv(LNDIST(x(i),X(i,1),X(i,2),X(i,3))))/LNDENS(x(i),X(i,1),X(i,2),X(i,3
));
me(i)=x(i)-se(i)*norminv(LNDIST(x(i),X(i,1),X(i,2),X(i,3)));
u(i)=(x(i)-me(i))/se(i); % Standardized variables
end
% Derivatives of g(X)

B - 48

Annex B Elementary methods of structural reliability II

g(1)=A(1)*(A(2)*x(2)+A(3)*x(3))*se(1);g(2)= A(1)*A(2)*x(1)*se(2);g(3)=
A(1)*A(3)*x(1)*se(3);
g(4)=A(4)*(A(5)*x(5)+A(6)*x(6))*se(4); g(5)= A(4)*A(5)*x(4)*se(5);g(6)=
A(4)*A(6)*x(4)*se(6);
g(7)= A(4)*A(7)*x(4)*se(7);
% Auxiliary quantities
gg=sqrt((g(1)*g(1)+g(2)*g(2)+g(3)*g(3)+g(4)*g(4)+g(5)*g(5)+g(6)*g(6)+g(7)*g(7)));
gu=g(1)*u(1)+g(2)*u(2)+g(3)*u(3)+g(4)*u(4)+g(5)*u(5)+g(6)*u(6)+g(7)*u(7);
% Reliability index
beta =-gu/gg;
for i=1:1:7; % Loop for determining sensitivity factors and a new design point
aa(i)=g(i)./gg; % sensitivity factors
x(i)=me(i)-beta.*aa(i).*se(i); % New design point
end
x(1)=(A0+A(4)*x(4)*(A(5)*x(5)+A(6)*x(6)+A(7)*x(7)))/(A(1)*(A(2)*x(2)+A(3)*x(3))); % Initial
guess value of x1
delta=abs(beta-betap); % Difference of beats of two last cycles
betap=beta; % Saving the current beta
end
% Outputs
Number_of_iterations_snd_achieved_accuracy=[j,delta],
Alphas=[aa(1),aa(2),aa(3),aa(4),aa(5),aa(6),aa(7)],
Design_points=[x(1),x(2),x(3),x(4),x(5),x(6),x(7)],
Design_points_over_means=[x(1)/X(1,1),x(2)/X(2,1),x(3)/X(3,1),x(4)/X(4,1),x(5)/X(5,1),x(6
)/X(6,1),x(7)/X(7,1)],
beta, %a1=aa(1),a2=aa(2),a3=aa(3),a4=aa(4),a5=aa(5),a6=aa(6),a7=aa(7),% To be printed if
needed
pf=NDIST(-beta); % Answer of the function FORM7
% The end of the function FORM7

B - 49

Annex B Elementary methods of structural reliability II

Attachment 13 MATLAB function Lndens (x,mu,sigma,sk)

function val = Lndens (x,mu,sigma,sk)
% DESCRIPTION
% NDENS evaluates the one-dimensional normal density function.
% CALL
% val = Lndens (x,sk);
% val = Lndens (x,sk,mu,sigma);
% INPUT
% x : real vector of arguments
% sk : coefficient of skewness (must be given)
% mu : mean value; optional; default = 0.0 (i.e. standard)
% sigma : std. dev. > 0; optional; default = 1.0 (i.e. standard)
% OUPUT
% val : vector of normal density values for the x's
% VERSION
% Milan Holicky, Czech Technical University in Prague, Klokner Institute
% 18.09.1999
% Initialization
if nargin < 3
mu = 0.0;
sigma = 1.0;
end
% Evaluate
x = (x-mu)/sigma;
% normalize
c=(0.5*sk+(sk^2/4+1)^0.5)^(1/3)-(-0.5*sk+(sk^2/4+1)^0.5)^(1/3);%constant of lognormal
if c==0; % sk=0
x0=10^10;
else
x0=-1/c; %bound of the distribution
end
if sk>0; %check of x range
if x0>x
error ('x out of range')
else
end
else
if x0<x
error('x out of range')
else
end
end
%
if abs(c)>0.0001; %if for c=0 (sk=0)
tt=sign(sk)*(log(abs(x+1/c))+log(abs(c))+0.5*log(1+c^2))/((log(1+c^2))^0.5);
else
tt=x;
end
if abs(c)>0.0001; %if for c=0 (sk=0)

B - 50

Annex B Elementary methods of structural reliability II

val = exp(-0.5*tt.^2)/(sqrt(2*pi)*sigma*abs(x+1/c)*(log(1+c^2))^0.5);
else
val = exp(-0.5*tt.^2)/(sigma*sqrt(2*pi));
end

B - 51

Annex B Elementary methods of structural reliability II

Attachment 14 MATLAB function Lndist (x,mu,sigma,sk)
function val = Lndist (x,mu,sigma,sk)
% DESCRIPTION, 18.09.1999
% LNDIST evaluates the one-dimensional lognormal distribution function.
% CALL
% val = lndist (x,sk)
% val = lndist (x,sk,mu,sigma)
% INPUT
% x : real vector of arguments.
% sk : coefficient of skewness; default = 0.0 (i.e. normal distribution)
% mu : mean value; optional; default = 0.0 (i.e. standard).
% sigma : std. dev. > 0; optional; default = 1.0 (i.e. standard).
% OUPUT
% val : vector of the lognormal distribution evaluated at the x's.
% VERSION
% Milan Holicky, Czech Technical University in Prague, Klokner Institute
% Initialization
if nargin < 3
mu = 0.0;
sigma = 1.0;
end
x = (x-mu)/sigma;
% standardize
c=(0.5*sk+(sk^2/4+1)^0.5)^(1/3)-(-0.5*sk+(sk^2/4+1)^0.5)^(1/3);%constant of lognormal
if c==0; % sk=0
x0=10^10;
else
x0=-1/c; %bound of the distribution
end
if sk>0; %check of x range
if x0>x
error ('x out of range')
else
end
else
if x0<x
error('x out of range')
else
end
end
if abs(c)>0.01; %if for c=0 (sk=0)
tt=sign(sk)*(log(abs(x+1/c))+log(abs(c))+0.5*log(1+c^2))/((log(1+c^2))^0.5);
else
tt=x;
end
val = normcdf(tt);%(1+erf(tt/sqrt(2)))/2; % transformed error function

B - 52

Annex B Elementary methods of structural reliability II

Attachment 15 MATLAB function Ndens (x, mu,sigma)
function val = ndens (x, mu,sigma)
%
% DESCRIPTION
% NDENS evaluates the one-dimensional normal density function.
%
% CALL
% val = ndens (x);
% val = ndens (x, mu,sigma);
%
% INPUT
% x : real vector of arguments
% mu : mean value; optional; default = 0.0 (i.e. standard)
% sigma : std. dev. > 0; optional; default = 1.0 (i.e. standard)
%
% OUPUT
% val : vector of normal density values for the x's
%
% VERSION
% Niels Jacob Tarp-Johansen
% Department of Structural Engineering and Materials
% Technical University of Denmark
% 15.06.1999
%
% Initialization
if nargin < 2
mu = 0.0;
sigma = 1.0;
end
% Evaluate
x = (x-mu)/sigma;
% normalize
val = exp(-0.5*x.^2)/(sigma*sqrt(2*pi));

B - 53

Annex B Elementary methods of structural reliability II

Attachment 16 MATLAB function Ndinv (p)
function [x] = norminv(p,mu,sigma)
%NORMINV Inverse of the normal cumulative distribution function (cdf).
% X = NORMINV(P,MU,SIGMA) returns the inverse cdf for the normal
% distribution with mean MU and standard deviation SIGMA, evaluated at
% the values in P.
% Default values for MU and SIGMA are 0 and 1, respectively.
%
% MH, Klokner Institute, CTU Prague 4.8.2003
%
if nargin < 2
mu = 0;
end
if nargin < 3
sigma = 1;
end
% Return NaN for out of range parameters or probabilities.
sigma(sigma <= 0) = NaN;
p(p < 0 | 1 < p) = NaN;
x0 = -sqrt(2).*erfcinv(2*p);
x = sigma.*x0 + mu;

B - 54

Annex B Elementary methods of structural reliability II

Attachment 17 - MATLAB function Ndist (x, mu,sigma)

function val = ndist (x, mu,sigma)
%
% DESCRIPTION
% NDIST evaluates the one-dimensional normal distribution function.
%
% CALL
% val = ndist (x)
% val = ndist (x, mu,sigma)
%
% INPUT
% x : real vector of arguments.
% mu : mean value; optional; default = 0.0 (i.e. standard).
% sigma : std. dev. > 0; optional; default = 1.0 (i.e. standard).
%
% OUPUT
% val : vector of the normal distribution evaluated at the x's.
%
% VERSION
% Niels Jacob Tarp-Johansen
% Department of Structural Engineering and Materials
% Technical University of Denmark
% 13.06.1999
% Initialization
if nargin < 2
mu = 0.0;
sigma = 1.0;
end
% Evaluation
x = (x-mu)/sigma;
% standardize
val = (1+erf(x/sqrt(2)))/2; % transform error function

B - 55

Annex C - Calibration procedure

ANNEX C - CALIBRATION PROCEDURE

Haig Gulvanessian1 and Milan Holick2
1

BRE Watford, United Kingdom

Klokner Institute, Czech Technical University in Prague, Czech Republic

Summary
The basic European standard for design of buildings and other engineering works,
EN 1990 Basis of structural design, provides alternative design procedures, for which
national choice is allowed. One of the most important questions concerns three fundamental
combinations of actions for persistent and transient design situations in the Ultimate limit
states. Simple example of generic structural member shows, that the alternative load
combinations may lead to considerably different reliability levels. Probabilistic methods of
structural reliability theory are used to identify characteristic features of each combination and
to formulate general recommendations. It appears that further calibration studies concerning
structures made of different materials are needed during the examination period of EN 1990
in order to analyse all possible consequences of national choice.
1

INTRODUCTION

1.1

Background materials
Each part of Eurocodes, including basic document EN 1990 [1], contains a number of
the Nationally Determined Parameters (NDP) for which national choice is allowed. In
accordance with the Guidance paper L concerning the Construction Products Directive an
important two years period after date of availability of each Eurocode Part is allowed to fix
the NDPs. However it is expected that calibration will continue during the coexistence period,
which starts at the end of the National calibration period and lasts up to three years after the
national publication of the last Part of a Package.
1.2

General principles
Basic concepts of code calibration are mentioned in Annex C of EN 1990 [1], in the
International Standard ISO 2394 [2] and ISO 13822 [3]. Additional information may be found
in the background document developed by JCSS [4] and in recently published handbook [5]
to EN 1990 [1]. Guidance for application of probabilistic methods of structural reliability may
be also found in working materials provided by JCSS [6] and in relevant literature listed in [5]
and [6].
In general NDPs may be calibrated either by direct comparison or by probabilistic
methods. Results of both approaches are usually combined with judgement (as mentioned in
ISO 2394 [2]). In this study probabilistic approach is applied mainly, a direct comparison of
load effects is shortly described in Appendix A to this contribution. Note that for the
probabilistic calibration software products [7,8,9] can be used. In particular the programme
[9] is intended for calibration purposes. Special purpose MATLAB functions and
MATHCAD sheet attached to this Annex may be also used for calibration studies.

C-1

Annex C - Calibration procedure

It should be noted that two aspects of calibration might be explicitly considered:
reliability and economy (see also Handbook 1). However, the following text shall be primarily
concerned with reliability of structures with respect to ultimate limit states. Additional
calibration aspects may concern fire safety of structures (see Handbook 5) or other accidental
design situations. In particular EN 1990 [1] requires that in the case of fire, the structural
resistance shall be adequate for the required period of time.
To consider all the above-mentioned aspects of structural reliability, an appropriate
design lifetime, design situations and limit states should be considered (as described in
Handbook 1). Note that the basic lifetime for a common building is 50 years and that, in
general, four design situations are identified: persistent, transient, accidental and seismic.
2

FUNDAMENTAL LOAD COMBINATIONS

In the following, the combination of three actions is considered: permanent action G,

imposed load Q (leading) and wind W (accompanying). EN 1990 [1] for the fundamental
combination of these loads in persistent and transient design situations introduces three
alternative procedures denoted here A, B and C. The loads (actions) G, Q and W and their
characteristic values Gk, Qk and Wk denote generally load effects (for example internal
bending moments) of appropriate loads (actions) and should be distinguished from the
original loads (actions) themselves. However, when mutual proportions of loads (actions) and
load effects are the same, then the distinction between load and load effects is not needed.
Design value of action effect Ed is obtained using the characteristic values Gk, Qk and
Wk and appropriate partial factors G, Q, W and reduction factors , Q and W as follows.
A. Considering the formula (6.10) in EN 1990 [1], the design value of action effect Ed
is given as
Ed = G Gk + Q Qk + W W Wk

(1)

B. An alternative procedure is provided in EN 1990 [1] by twin expressions (6.10a)

and (6.10b)
Ed = G Gk + Q Q Qk + W W Wk

(2)

Ed = G Gk + Q Qk + W W Wk

(3)

The less favourable action effect from (2) and (3) should be considered. In equation
(3) is a reduction factor for unfavourable permanent actions G. Note that in equations (1) to
(3) + generally implies to be combined with.
C. In addition EN 1990 [1] allows further modification of alternative B, simplifying
equation (2) by considering permanent loads only, thus the load effect is then
Ed = G Gk

(4)

The less favourable action effect resulting from (3) and (4) is then considered. In
addition to the combinations A, B, C provided in EN 1990 [1] (for recommended values G =
1,35, Q = 1,5) an additional combination may be also considered in the analysis to illustrate
the sensitivity of the resulting reliability level to partial factors, and the possible effect of their
reduction.
If the leading action is wind W, then in equations (1) and (2) instead of reducing wind
action W by factor W, the imposed load Q should be reduced by the appropriate factor Q.
C-2

Annex C - Calibration procedure

Factors G, Q and W denote the partial factors of actions G, Q and W (the partial factors for
both variable actions are equal, Q = W).
To investigate resulting load effects under various intensities of variable actions, the
characteristic values of Gk, Qk and Wk are related using quantities given as the ratio of
variable actions Qk+Wk to total load Gk+Qk+Wk, and ratio k of accompanying action Wk to
the main action Qk

= (Qk+Wk)/(Gk+Qk+Wk), k = Wk/Qk

(5)

Note that a realistic range of is from 0,1 to 0,6. However in some cases the load ratio
may be very low if not zero (e.g. underground garages).
For a given design value of the load effect Ed the characteristic values of individual
actions Gk, Qk, Wk can be expressed using variables and k as follows
Gk =

Ed

(( Q ) Q + k ( W ) W )
( ) G +
(1 + k )(1 )

, Qk =

Gk
, Wk = k Qk
(1 + k )(1 - )

(6)

The factors , G and Q indicated in the first relationship of (6) in brackets are applied
in the same way (either yes or no) as in equations (1) to (4) for alternative combination rules
A, B and C.
For alternative A, equation (1) is valid in the whole range 0 1, whereas using
alternative B, equation (2) is valid in the interval 0 lim,B and equation (3) in the interval
lim,B 1. Correspondingly, for alternative C equation (4) is valid in the interval 0
lim,C and equation (3) in the interval lim,C 1. The limiting values lim,B and lim,C can be
derived from equations (2) to (5) as follows

lim,B=

G (1 )(1 + k )
G (1 )(1 + k ) + Q (a Q ) + W k (b W )

(7)

G (1 )(1 + k )
G (1 )(1 + k ) + Q a + W kb

(8)

lim,C=

where the auxiliary variable a = 1 and b = W when for k (1Q)/(1W) (imposed

load Q is the leading action) and a = Q and b = 1 when k > (1Q)/(1W) (action W is the
leading action).
EN 1990 allows through the National Annex, which will be published by national
standardisation institution

Which of the combination expression to use, and

The specification of appropriate safety factors

Thus, the National Annexes should include the recommendation of one of the
alternatives indicated in EN 1990 [1] for a fundamental combination of actions in the
Ultimate limit states and partial factors G and Q for permanent and variable actions.
Considering a generic structural member it will be shown that the choice of these nationally
determined parameters may significantly affect the resulting reliability level. Partial and
C-3

Annex C - Calibration procedure

reduction factors , and recommended in EN 1990 [1] and used in this paper are
summarized in Table 1.
Table 1. Partial and reduction factors.
Action
Partial factors

Permanent G
Imposed Q
Climatic W

Combination factor

Reduction factor

1,0
0,7
0,6

0.85
-

1,35
1,5
1,5

In addition to the factors indicated in Table 1 other values will be used to make comparison of
Eurocode procedures with some national rules.
3

GENERIC STRUCTURAL MEMBER

In case of generic structural member it is assumed that the characteristic value Rk of

the resistance R may be defined as the 5% fractile of R and the design value Rd as
Rd = Rk/R

(9)

where R denotes the global resistance factor (commonly expected to be within the
range from 1 to 1,2). The significance of both values Rk and Rd is obvious from Figure 2,
where the random variable R is described by the probability density function R (R), and the
design value Rd is indicated as a particular value of R corresponding to a certain small
probability p of being violated.

R(R)

5%

Rd

Rk

Figure 2. Random variable R, the characteristic value Rk and design value Rd.
In design calculation of a structural member the design value Rd of the resistance R is
normally obtained by substituting design values Xdi for the random variables Xi, thus
Rd = R(Xdi)
C-4

(10)

Annex C - Calibration procedure

This expression is also used in design of members for the generic structural member
and for the different materials.
Table 2 shows the assumed values for the global resistance factor R and the
coefficient of variation VR, used in the following reliability analysis for a generic crosssection.
Table 2 Global resistance factor R, the coefficient of variation VR and the mean factor .
EN - global safety factor R = Rk /Rd
BSI - global safety factor R = Rk /Rd, BSI
The coefficient of variation VR
The mean factor = R/Rk

Middle value
1,15
1,10
0,15
1,28

Range
1,0 1,3
1,0 1,20
0,10 0,25
1,10 1,40

Note 1. The coefficient of variation VR includes the variability of the model uncertainty assumed to have the
coefficient of variability 0,05.
Note 2. The values are different for BSI codes reflecting the fact that lower values of R are used for particular
materials.

A middle values for the global safety factor R = 1,15, for the coefficient of variation
VR = 0,15 and for the mean ratio = R/Rk = 1,28 are considered in the following example of
a code condition.
4

PRINCIPLES OF RELIABILITY ANALYSIS

4.1

Limit state function

The most important step in reliability analysis is definition of a limit state function
(reliability margin) Z(X) separating safe and unsafe domain of basic variables X. In this report
the limit state function Z(X) is considered as in a simple form as a difference between the
resistance R(X) and the load effect E(X)
Z(X) = R(X) E(X) = R R0(X) E E0(X)

(11)

where factor R represents uncertainties of the resistance model R0(X) and factor E represents
uncertainties of the load effect model E0(X). The vector X denotes all the basic variables
entering the expressions for the resistance R(X) and the load effect E(X). Taking into account
general expressions (1) to (4) the load effect E0(X) may be written as
E(X) = E (G0 + Q0 + W0)

(12)

Considering the limit state function given by equation (11) and expression (12) giving
the load effect, it follows that basic variables R, G, Q, and W covering effects of model
uncertainties are defined as follows
R = R R0(X), G= E G0, Q= E Q0, W= E W0

(13)

Taking into account equation (13), the limit state function (11) may be written in a
simple form as
Z(X) = R (G + Q+ W)

C-5

(14)

Annex C - Calibration procedure

Note that the cumulative basic variables R, G, Q, W in equation (14) include effects of the
factors R and E (see equation (13)).
4.2

Probabilistic models of basic variables

It is assumed that structural members are designed economically, which means that the
design value of the resistance Rd(X) equals the design value of the load effect Ed(X)
Rd(X) = Ed(X)

(15)

It should be noted that normally (due to several reasons) the design resistance Rd(X) is greater
than the design load effect Ed(X), which may provide additional safety margin not considered
here.
Assuming a certain set of partial and combination factors , , and , the design
expression (15) can be used to specify the characteristic values Xk of each basic variable X.
The probabilistic characteristics (the mean, standard deviation) of each basic variable X can
be then related to its characteristic value Xk as indicated in Table 3.
Table 3. Probabilistic models of basic variables for time invariant reliability analysis using
Turkstra's rule (combination of 50-year maximum of leading action and an annual maximum
of accompanying action).
St.dev.
No. Category of
Name of basic
Sym. Dim- DistriMean
variables
variables
X ension bution
X
X
1
Actions
Permanent
G0
kN
N
Gk
0,1X
2
Imposed - 5 years
Q0 kN/m2
GU
0,2Qk
1,1X
GU
0,6Qk
2
Imposed - 50 y.
Q0 kN/m2
0,35X
3
Wind - 1 year
W0 kN/m2
GU
0,3Wk
0,5X
4
Wind - 50 year
W0 kN/m2
GU
0,7Wk
0,35X
5
Resistance
Resistance
R
kN/m2
LN Rk +1.65R 0,15X
6
Uncertainty
Uncertainty
LN
1
0.05
E
Probabilistic models indicated in Table 3 are based on data available in the
recommendation of JCSS [4,6] and literature [11,12,13,14]. As mentioned above the
probabilistic characteristics indicated in Table 3 represent just conventional models that might
be slightly conservative.
Note that the mean of a resistance R indicated in Table 3 in terms of the characteristic
value Rk and the standard deviation R may be assessed assuming a given coefficient of
variation VR using relationship

R = Rk exp(1,65 VR)

(16)

Under this assumption the mean resistance factor considered in Table 2 is given as

= R / Rk = exp(1,65 VR)

(17)

Considering the coefficient of variation VR = 0.15, the mean resistance factor becomes
= 1.28 as indicated in Table 2.
It should be emphasised that the probabilistic models of basic variables indicated in
Table 3 are primarily intended as "conventional models" in time invariant reliability analysis

C-6

Annex C - Calibration procedure

of structural members using Turkstra's combination rule [10] (explained also in [2]) for the
probabilistic calibration of the rules for combination of actions.
Conventional models indicated in Table 3 should enable the objective comparison of
results of various reliability studies expected in the near future in connection with
implementation of the present suite of Eurocodes into the national systems of design codes.
However, when the reliability of different types of structural members under particular
conditions is assessed, the proposed models in Table 3 may have to be adjusted to the
concrete conditions of the analysed structural member.
4.3

Reliability measures
The probability of failure Pf is the basic reliability measure used in this study. It can be
expressed on the basis of a limit state (performance) function Z(X) defined in such a way that
a structure is considered to survive if Z(X)> 0 and to fail if Z(X) 0. An example of the
function Z(X) is given by equation (14). In a general case the failure probability Pf can be
determined using the integral

( X )dX

Pf = Prob(Z 0) =

(18)

g
g ( X ) 0

where g(X) denotes joint probability density distribution of the basic variable X,
which may not be , however, available.
Assume that both the resistance R(X) and the load effect E(X) represent a single
variable X used to analyse structural performance (e.g. axial force or bending moment that is
represented by R(X) and E(X)). Then the integration indicated in expression (18) may be
simplified and the probability Pf can then be expressed as:

Pf = Prob(Z(x) 0) = E ( x ) R ( x )dx

(19)

where E(x) denotes the probability density function of E(X), R(x) the distribution of R(X).
To use equation (19) both the probability density function E(Z) and the distribution function
R(x) must be known (at least in an approximate form). Simplified procedure based on
expression (19) is used in this study.
Note that there are commercially available software products (e.g. VaP, COMREL),
which can be used to determine the failure probability Pf in more complicated cases than
considered here (when expression (19) cannot be used). These software products were used in
this study to check results obtained by numerical integration based on expression (19).
In Annex C of EN 1990 an alternative measure of reliability is conventionally defined
by the reliability index , which is related to Pf as
Pf = ( )

(20)

where is the cumulative distribution function of the standardised normal distribution. The
relation between Pf and is indicated in Table 4.
Table 4. Relation between and Pf.
Pf

10-1
1,28

10-2
2,32

10-3
3,09

10-4
3,72

C-7

10-5
4,27

10-6
4,75

10-7
5,20

Annex C - Calibration procedure

Table C2 of EN 1990 recommends for the ultimate limit state of buildings over a fifty
year design working life a target value of reliability index t = 3,8. If one year period is
considered in reliability verification, then t = 4,7. Both the equivalent reliability measures,
the failure probability Pf and the reliability index , are used in this study.
4.4

Sensitivity factors
Sensitivity factors of the First Order Reliability Methods (FORM) are normally used
[1,2] to calibrate design values of basic variables and partial safety factors. Considering the
limit state function Z(X) (reliability margin) given by equation (11), the sensitivity factors for
the four cumulative variables R, G, Q, W can be defined in terms of their standard deviations
R, G, Q and W as follows

R =

, G = G , Q = Q , W = W
g
g
g
g

(21)

where g denotes the standard deviation of Z(X) given as

g = R2 + G2 + Q2 + W2

(22)

In the following investigation the sensitivity factors R, G, Q and W defined by

equation (21) are considered together with the failure probability Pf and the reliability index
. It should be underlined that R, G, Q and W defined by (21) refer to cumulative
variables R, G, Q, W, which include effects of the factors of model uncertainties R and E
(see equation (13)).
5

RESULTS FOR THE GENERIC CROSS-SECTION

5.1

One variable action

Results of the reliability analyses are presented in graphical form that indicates
variation of the reliability index , failure probability Pf, and sensitivity factors R, E, G, Q
and W with the load ratio . In particular Figure 3 shows results of a simple case of one
variable action only (the main variable action Q); Figure 3 indicates the variation of
-

the reliability index ,

failure probability Pf, and
for expression 6.10 of EN 1990 sensitivity factors R, E, and partial sensitivity
factors G, Q and W

with the load ratio .

For the analysis it has been assumed that a single variable action, the imposed load Q
having the characteristic given in Table 3 is acting on the generic element only (i.e. k = 0.0).
A middle value for the global safety factor R = 1,15 and for coefficient of variation VR = 0,15
have been considered.

C-8

Annex C - Calibration procedure

A
B

C
B

B
A

R
W
Q

Figure 3. Variation of the reliability index , the failure probability Pf and the sensitivity
factors R, E, G, Q and W with the load ration for k = 0, for a generic cross-section
assuming R = 1,15 and the coefficient of variation VR = 0,15.
It follows from Figure 3 that for the assumed higher coefficient of variation VR = 0,15
only the combination rule A (i.e. expression (6.10) of EN 1990) [1] seems to be fully
acceptable ( > 3.8 and Pf.< 7,2310-5) in the interval 0 < < 0.8, however the reliability
level considerably varies with . In some cases the alternative A might lead to an uneconomic
design.
The alternative B (i.e. expression (6.10a) and (6.10b) of EN 1990) is acceptable in a
slightly shorter range of , 0 < < 0.7 than the variant A but provides obviously much more
uniform distribution of reliability level with . Obviously it would lead to a more economic
C-9

Annex C - Calibration procedure

design than the alternative A. Alternative C (i.e. modified expression (6.10a) and (6.10b) of
EN 1990 [1]) is providing rather low reliability level particularly for the interval 0 < < 0.3
and should not be used unless partial factors are changed.
Similar results were obtained in previous studies [11,12,13,14] of structural elements
made of different materials (concrete and steel elements). These studies differ from the
presented results primarily by the value of the partial factor R and the coefficient of variation
VR (and also by the asymmetry of the distribution of R). Just the conclusions formulated above
seem to be supported by a number of different material oriented examples.
Note that the sensitivity factor R increases to about R ~ 0,9 while the factor E
decreases, E > - 0,5, indicating that the resistance gives a greater contribution to safety than
intended by EN 1990. However this conclusion is strongly dependent on assumed coefficient
of variation VR. With increasing VR the sensitivity factor R increases. It is interesting to note
that than the sensitivity factors are very close to the values recommended in EN 1990 [1], i.e.
R ~ 0,9 E ~ -0,7.
5.2

Two variable actions

A more general case when two variable actions (a leading imposed load Q, together
with an accompanying action W) are acting is shown in Figure 4, which (similarly as Figure
3) shows the variation of the
reliability index ,
failure probability Pf, and
for expression 6.10 of EN 1990 sensitivity factors R, E, G, Q and W
with the load ratio for k = 0,75 and the coefficient of variation VR = 0,15.
The case considered for Figures 3 (i.e. k = 0, with a single imposed load Q acting) is
extended so that a more detailed insight of the effect for the reliability parameters considered
can be obtained. However Previous investigations [11,12] clearly show that reliability in case
of two variable actions is considerably greater than reliability in case of one variable action.

C - 10

Annex C - Calibration procedure

A
B

B
C

C
B

B
A

G
E

Figure 4. Variation of the reliability index , the failure probability Pf and the sensitivity
factors R, E, G, Q and W with the load ratio for k = 0.75, for a generic cross section
assuming R = 1,15 and the coefficient of variation VR = 0,15.
It follows from Figure 4 that for the assumed coefficient of variation VR = 0,15 and the
consideration of two variable actions the reliability of the generic cross-section exposed to
two variable actions is considerably greater than the reliability of the same cross-section
exposed to one variable action only. This finding also indicates that the factor W may be
rather high. Note that the sensitivity factors R seems to be slightly greater than the values R
= 0,8 considered in EN 1990 [1] and E in absolute value is less than E = - 0,7 recommended
in [1]. This finding depends on assumed variability of basic variables.
C - 11

Annex C - Calibration procedure

Figure 5. Variation of the reliability index with the load ratio and partial factor for
resistance R and k = 0 (i.e. imposed load Q is the only variable action), for the generic-cross
section assuming partial safety factors G = 1,35 and Q = 1,5, and the coefficient of variation
VR = 0,15.
It follows from Figure 5 that for the assumed variables the acceptable domain of the
load ratio and the coefficient of variation VR is limited by the contour line determined as an
intersection of the surface and the plain = 3,8 in Figure 5. Obviously with increasing R
reliability index increases, R = 1.15 would be satisfactory for most of the practical range of
the load ratio (for the load ratio .< 0,8).

C - 12

Annex C - Calibration procedure

CONCLUDING REMARKS

The newly available EN 1990 provides alternative design procedures and parameters
that should be unambiguously specified in the National Annexes of Member States of CEN.
These alternative design procedures lead in some cases to significantly different reliability
levels. Preparation of National Annexes is therefore a complicated task for each Member
State. Furthermore, the Eurocode standards recognise the responsibility of the regulatory
authorities in each Member State and safeguard their right to determine values related to
regulatory safety matters at national level.
Simple examples of a generic structural member confirm the results of the earlier
studies that the reliability of structures, designed according to the alternative combination
rules provided in EN 1990 by expressions (6.10), (6.10a) and (6.10b), may vary considerably.
Expression (6.10) leads to the most reliable but in some cases uneconomical structures. Twin
expressions (6.10a) and (6.10b) provide a lower but comparatively most uniform reliability
level for all load ratios. Moreover, they seem to fully comply with EN recommendations
(reliability index 3,8 for a 50-year time period). The lowest reliability is obtained from the
third alternative, given by modified expression (6.10a) and expression (6.10b). This
alternative seems to lead to a rather low reliability level, particularly for structures exposed
mainly to a permanent load.
In order to make an unambiguous recommendation for National Annexes to EN 1990,
further investigations are urgently needed. Obviously more complicated structural elements,
made of various materials, should be analysed and compared. Such a calibration activity
should preferably be organised on an international level. The short-term objective of these
activities should be to develop the necessary background materials for preparation of the
National Annexes. The long-term objective should be to further harmonization of the
alternative design procedures considered during the next revision of the present generation of
Eurocodes.
It is expected that further calibration studies concerning structures made of different
materials will be needed during the examination period of EN 1990 (next few years) in order
to analyse all possible consequences of national choice.
REFERENCES
[1] EN 1990 Eurocode - Basis of structural design. CEN 2002.
[2] ISO 2394 General principles on reliability for structures, ISO 1998.
[3] ISO 13822. Basis for design of structures - Assessment of existing structures, ISO 2001.
[4] JCSS: Probabilistic model code. JCSS working materials, http://www.jcss.ethz.ch/, 2001.
[5] Gulvanessian, H. Calgaro, J.-A. Holick, M.: Designer's Guide to EN 1990, Eurocode:
Basis of Structural Design; Thomas Telford, London, 2002, ISBN: 07277 3011 8, 192 p.
[6] JCSS: Background documentation, Part 1 of EC 1 Basis of design, 1996.
[7] VaP, Variable Processor, version 1.6, ETH Zurich, 1997.
[8] COMREL, version 7.10, Reliability Consulting Programs, RCP MUNICH, 1999.
[9] CodeCal, Excel sheet developed by JCSS, http://www.jcss.ethz.ch/.
[10] Turkstra, C.J.: Application of Bayesian Decision Theory. Study No. 3 Structural
Reliability and Codified Design. Solid Mechanics Division, University of Waterloo,
Ontario, Canada, 1970.
[11] Holick M., Markov J.: Verification of load factors for concrete components by
reliability and optimization analysis: Background documents for implementing

C - 13

Annex C - Calibration procedure

Eurocodes. Progress in Structural Engineering and Materials, Vol. 2 No. 4, 2000, pp.
502-507.
[12] Gulvanessian, H. Holick, M. Markov, J.: Calibration of Eurocode Reliability
Elements Considering Steel Members; In: Proceedings of Third European Conference
on Steel Structures, Volume II; Antonio Lamas and Luis Simoes da Silva, CMM
Associacao Portuguesa de Construcao Metalica e Mista, Guimaraes, Portugal; ISBN:
972-98376-3-5; pp. 1511-1520
[13] Holick, M. Markov, J.: Reliability of Concrete Elements Designed for Alternative
Load Combinations Provided in Eurocodes. Acta polytechnica, 2003/1.
[14] SAKO; Joint Committee of NKB and INSTA-B. Basis of Design of Structures.
Proposal for Modification of Partial Safety Factors in Eurocodes. 1999, 55 p.

C - 14

Annex C - Calibration procedure

APPENDIX A. DIRECT COMPARISON OF LOAD EFFECTS

Deterministic approach the global load factor
Concept of the global load factor is sometimes used to compare various alternatives
for load combination with no regard to a resistance of a structure. The deterministic global
load factor E follows directly from codified combination rules and given partial factors
without any probabilistic consideration; it is simply expressed as

E = Ed / (Gk + Qk + Wk)

(A.1)

where the design load effect Ed is given by one of equations (1) to (4) depending on the
combination rule considered (for example combination rules A, B or C). It follows from
equation (1) to (4) and (A.1) that in general deterministic E may be expressed as

E = (1 - ) G () + (Q (Q) + k W (W)) / (1+ k)

(A.2)

where the factors in brackets (), (Q) and (W) are applied in accordance with the principles
of appropriate combination rule. For example assuming that Q is the leading variable load and
W is accompanying load, the combination rule A based on expression (6.10) of EN 1990 [1]
the global factor E follows from (1) and (A.2) as

E = (1 - ) G + (Q + k W W) / (1+ k)

(A.3)

Similarly the global factors E of other combination rules B and C may be obtained from
general expression (A.2). It follows from (2) that equation (A.2) becomes

E = (1 - ) G + (Q Q + k W W) / (1+ k)

(A.4)

When equation (3) is applied, then equation (A.2) becomes

E = (1 - ) G + (Q + k W W) / (1+ k)

(A.5)

When equation (4) is applied, then equation (A.2) becomes

E = (1 - ) G

(A.6)

Thus combination rule A is described by equation (A.3), combination B by equations

(A.4) and (A.5), combination rule C by equations (A.4) and (A.6).
Figure A.1 shows the global factor E for all three-combination rules A, B and C
assuming the load factor k = 0 (two loads G and Q are considered only). It is interesting to
note that the global load factor E is strongly dependent on the load factor k. Figure A.2 shows
the case of three variable actions G, Q and W assuming k = 0,75. Similar results may be
obtained for any load ratio k. However, it is well recognised that decisive requirements
(compare Figures 3 and 4) for calibration of reliability elements follow from combinations of
two actions only (G and Q). Figure A.2 just illustrates variation of the global factor with the
load ratio in the case of two variable actions.

C - 15

Annex C - Calibration procedure

E
1.5

A
1.4

B
1.3

1.2

1.1

lim,C
1

lim,B

0.2

0.4

0.6

0.8

Figure A.1. The global load factor E for the combination rules A, B and C assuming k = 0.

E
1.5

1.4

A
1.3

B
C

1.2

C
1.1

lim,C
1

0.2

lim,B
0.4

0.6

0.8

Figure A.2. The global load factor E for the combination rules A, B and C assuming k = 0,75.

C - 16

Annex C - Calibration procedure

Probabilistic approach the theoretical global load factor
Probabilistic approach to comparison of load effects E considers the probability p of E
exceeding Ed = E (Gk + Qk + Wk), thus the probability
p = P(E > Ed)

(A.7)

The probability p is obviously dependent on the global load factor E determining the load
effect Ed as follows from equation (A.1). When probabilistic models of actions G, Q and W
are known, then for a given value of E the probability p may be determined. Let us remind
that in accordance with the principles of EN 1990 [1] (considering E = -0.7 and = 3.8) the
recommended value of the probability p given by equation (A.4) is
pE = (E ) = (- 0,7 3,8) = 0,004

(A.8)

Assuming probabilistic models of actions G, Q and W considered above in accordance

with Table 3, variation of the theoretical load factors E with the load factor for selected
probabilities p is shown in Figure A.3 together with the deterministic load factors E described
above. In Figure A.3 full lines indicate the theoretical (probabilistic) load factors E, the
dashed lines indicate the deterministic load factors E.

E
p=0.001

1.5

A
B

1.4

p=0.002

1.3

p=0.004
C
1.2

C
1.1

lim,C
0

0.2

lim,B
0.4

0.6

0.8

Figure A.3. The global load factor E for the combination rules A, B and C assuming k = 0
and theoretical values of E corresponding to selected probabilities of E exceeding Ed

C - 17

Annex C - Calibration procedure

Figure A.3 clearly indicates differences between the theoretical global factors E
determined using probabilistic approach and corresponding deterministic values described by
equation (A.2). It follows from Figure A.3 that the deterministic values are greater than the
theoretical values of E corresponding to the probability p = 0,004 indicated in equation (A.5).
Thus, the Eurocode combination rules seem to be on a safe side, in particular the combination
rules A and B. Note that for small load ratios the combination rule C provides lower values
of the global factor than the theoretical E and, therefore, seems to be unsatisfactory.
Figure A.3 further indicates that the theoretical E is better followed up by E
corresponding to the combination rule B than those corresponding to the combination rules A
or C. In that sense direct comparison of load effects confirms conclusions of previous studies
when both the load effect and resistance are taken into account.

1 p

Figure A.4. Variation of the theoretical and deterministic global factor E with the load ratio
and the probability p assuming the combination rule A.
Variation of the global factor E with the load ratio and the probability p clearly
indicates that the combination rule A, represented in Figure A.4 by a plane, is rather safe (and
perhaps uneconomic) substitution of the theoretical (probabilistic) values.

C - 18

Annex C - Calibration procedure

APPENDIX B. EFFECT OF THE RESISTANCE VARIABILITY
In reliability analysis of a generic cross section the coefficient of variability VR = 0.15
and the partial factor R = 1,15 are assumed as an example of a code condition. However,
resistance of various structural members made of different materials may have different
variability and the partial factor. The coefficient of variability VR can be expected within a
broad range from 0,05 up to almost 0,50 (including uncertainty resistance model). This should
be reflected by appropriate value of the partial factor R. Assuming lognormal distribution of
R, the partial factor R corresponding to the coefficient of variation VR can be expressed as

R = exp(1,65 VR)/exp(E VR) = exp(1,65 VR)/exp( 3,04 VR)

(B.1)

where E = 0,8 3,8 = 3,04 as recommended in EN 1990 [1]. Note that for VR = 0,10
equation B.1 yields the partial factor R = 1,15.
Figure B.1 shows the variation of the partial factor R with the coefficient of variability
VR together with corresponding reliability index determined taking into account the partial
factor R as a function of VR given by equation (B.1). In Figure B.1 the combination rule A
and a generic cross section are considered only.
5

=0.3
=0.6

4.5
4

=0.0

3.5
3
2.5

2
1.5
1

wR
0

0.1

0.2

0.3

0.4

0.5

Figure B.1. Variation of the reliability index of a generic cross section with the coefficient
of variability VR for selected load ratios assuming the partial factor R as a function of VR.
Figure B.1 indicates that if the partial factor R is considered as a function of the
coefficient of variability VR, the effect of resistance variability is not essential. Considering a
realistic range of resistance variability 0,05 < VR < 0,25, differences in values seem to be
about 0,5. This finding justifies the concept of a generic cross section used in reliability
analysis of alternative load combinations.

C - 19

Annex C - Calibration procedure

APPENDIX C. NOTATION
E
E0
Ed
Ek
Ed
G
G0
Gd
Gk
k
Pf
Q
Q0
Qd
Qk
R
Rd
Rk
VR
W0
W
Wd
Wk
X
Z(X)

R
E
G
Q
W

()

G
Q
W
Q
W

load effect including model uncertainty

load effect without model uncertainty
design value of the load effect E
characteristic value of the load effect E
design value of the load effect E
permanent load including model uncertainty, G = G0
permanent load without load uncertainty
design value of the resistance G, Gd=G Gk
characteristic value of the permanent load G
load ratio, k = Wk/Qk
failure probability
main (dominant) variable load including model uncertainty, Q = Q0
main (dominant) variable load without model uncertainty
design value of the variable load Q, Qd=Q Qk
characteristic value of the variable load G
resistance including model uncertainty
design value of the resistance R, Rd=R Rk
characteristic value of the resistance R
coefficient of variation
main (dominant) variable action without model uncertainty
accompanying (non dominant) variable action including model uncertainty, W = W0
design value of the variable load W, Wd=W Wk
characteristic value of the variable load W
vector of basic variables
limit state function
sensitivity factors of R
sensitivity factors of E
sensitivity factors of G
sensitivity factors of Q
sensitivity factors of W
reliability index, Pf = ( )
probability density function
load ratio, =(Qk + Wk )/ (Gk + Qk + Wk )
partial factor for unfavourable permanent actions G
partial factor for unfavourable variable actions Q
partial factor for unfavourable variable actions W
reduction factor for unfavourable permanent actions Q
reduction factor for unfavourable permanent actions W
reduction factor for unfavourable permanent actions G
coefficient of model uncertainty
distribution function of standardised normal distribution

C - 20

Annex C - Calibration procedure

ATTACHMENTS

1. MATLAB function "generic(gR,wr,k)".

Matlab function "Generic" is intended for investigation of the combination rules
provided in EN 1990. A general structural member of the resistance R (including model
uncertainty) is considered.
Function "Generic" calls function
Action3(mr,sr,skr, Rd,k), which further calls functions
Lnpf (mr,sr,skr,me,se,ske), which calls
Lndens(x,ske,me,se)
Lndist(x,skr,mr,sr)
Ndinv(p)
2. MATLAB function "Action3(mr,sr,skr,Rd,k)"
MATLAB function "Action3" is intended for determining statistical characteristics of
the load effect of different combinations of three actions.
Function "Generic" calls function
Lnpf (mr,sr,skr,me,se,ske), which further calls functions
Lndens(x,ske,me,se)
Lndist(x,skr,mr,sr)
Ndinv(p)
3. MATLAB function "Lnpf(mr,sr,skr,me,se,ske)"
MATLAB function "Lnpf" calculates the failure probability using three parameter
lognormal distribution for approximation of the load effect and resistance.
Function "Lnpf" further calls functions
Lndens(x,ske,me,se)
Lndist(x,skr,mr,sr)
Ndinv(p)
4. MATLAB function "Lndens(x,mu,sigma,sk)"
MATLAB function "Lndens" is intended for calculation of the probability density
function of three-parameter lognormal distribution. The function is called by the function
LNPF using command Lndens(ske,me,se), and returns the value of probability density
function.
5. MATLAB function "Lndist(x,mu,sigma,sk)"
MATLAB function "Lndist" is intended for calculation of the distribution function of
three-parameter lognormal distribution. The function is called by the function LNPF using
command Lndist(skr,mr,sr), and returns the value of the distribution function.
6. MATLAB function "Ndinv(p)"
MATLAB function "Ndinv" calculates the inverse distribution function of the normal
distribution (determining the reliability index beta). The function is called by the function
Action3 or Action3i using command Ndinv(p) (or Ndinv(p,mu,sigma) or Ndinv(p,mu)),
and returns the value of the inverse distribution function.

C - 21

Annex C - Calibration procedure

7. MATHCAD sheet"Generic.mcd"
MATHCAD function "Generic" is intended for investigation of the combination rules
provided in EN 1990. A general structural member of the resistance r (including model
uncertainty) is considered.
8. MATHCAD sheet"Load effect.mcd"
MATHCAD sheet "LoadEffect" is intended for investigation of combination rules
provided in EN 1990 by expressions (6.10), (6.10a) and (6.10b) considering three loads: G, Q
and W. Turkstra's rule ( 50-year extremes of a leading and annual extremes of an
accompanying action) is applied.

C - 22

Annex C - Calibration procedure

Attachment 1 MATLAB function "generic(gR,wr,k)"
function Gener=Gener(gR,wr,k)
% Program "Generic" is intended for investigation of the combination rules
% provided in EN 1990.
% A generic structural member of the resistance R (including model uncertainty)
% is considered.
% Function "Generic" calls function
% Action3i(mr,sr,skr, Rd,k,i), which further calls functions
% Ndinv(p), Lnpf (mr,sr,skr,me,se,ske), which calls
%
Lndens(x,ske,me,se), Lndist(x,skr,mr,sr) and Ndinv(p)
%
% INPUT data describing random variable R: Rd, gR, wR
%
% VERSION
% MH, Klokner Institute, Czech Technical University in Prague, 1.08.2003
%
% Input load ratio k = Wk/Qk used by the function "Action3i",
% k=0.00; % Input parameter that may be changed
% Characteristic of the resistance R
Rd=1; % may be chosen arbitrary
Rk=Rd*gR; % gR=1.15; gR given by a fixed value not related to wr and beta
% an alternative is indicated below in the first line of the loop for wr
%
% Statistical parameters of R (having lognormal distribution) determined
% in the following loop for selected coefficients of variation wr
%for i= 3:3 % Range of the loop that may be adjusted.
%wr=0.05+(i-1)*0.05; betat = 3.8;
% gR= exp(0.7*betat*wr)/exp(1.65*wr); Rk=Rd*gR; % alternatively
Kr=1; wKr=0.05; % lognormal distribution
mr=Kr*Rk*exp(1.645*wr); wr=(wr^2+wKr^2+wr^2*wKr^2)^0.5;
sr=mr*wr; skr=3*wr+wr^3;
% alternatively the mean mr=Rk/(1-1.645*wr)
Action3(mr,sr,skr,Rd,k) % Call function Action3i
%end

C - 23

Annex C - Calibration procedure

Attachment 2 MATLAB function "Action3 generic(mr,sr,skr,Rd,k)"
function Action3 = action3(mr,sr,skr,Rd,k)
% CALL
% Lnpf (mr,sr,skr,me,se,ske), which further calls LNDENS and LNDIST
% INPUT
% R: mr,sr,skr; E:me,se,ske; vector of real arguments
% OUPUT
% beta: vector of beta values
% VERSION
% Klokner Institute, Czech Technical University in Prague, 24.04.2002
% Laod factors and parameters:
gG=1.35; gQ=1.5; gW=1.5; psi1=0.7; psi2=0.6; wG=0.1; %psi1=psiQ, psi2=psiW
% Characteristics of variable loads Q and W for k<=(1-psiQ)/(1-psiW)
if k<=(1-psi1)/(1-psi2);
mmQ=0.49; wQ=0.4; skQ=1.14; mmW=0.3; wW=0.5; skW=1.14;
else %Characteristics of variable loads Q and W for k>(1-psiW)/(1-psiQ)
mmQ=0.2; wQ=1.1; skQ=1.14; mmW=0.7; wW=0.35; skW=1.14;
end
% Model uncertainties of actions
Ke=1; wKe=0.00; sKe=wKe*Ke;skKe=3*wKe+wKe^3;
% Parameters k=Wk/Qk given in the function that calls Action3
% Combination factors for expression (6.10) if k<=(1-psiQ)/(1-psiW)=0,75 or k>(1psiW)/(1-psiQ)=0,75
if k<=(1-psi1)/(1-psi2);
ksi=1; psiQ=1; psiW=psi2;
else
ksi=1; psiQ=psi1; psiW=1;
end
%!!!!!!!!!!!!!!!!!!!!!!!!!Case A, (6.10)
%Effect of the load ratio CHI for expression (6.10) for the load ratio
CHI=(Qk+Wk/(Gk+Qk+Wk))
for n=1:21 %loop for CHI in the interval <0,1>
CHI(n)=0+(n-1)*0.0499;
mG(n)=Rd/(ksi*gG+(CHI(n).*(psiQ*gQ+k*psiW*gW))./((1-CHI(n))*(1+k)));
sG(n)=mG(n)*wG;
Qk(n)=CHI(n).*mG(n)./((1-CHI(n))*(1+k)); mQ(n)=Qk(n)*mmQ; sQ(n)=mQ(n)*wQ;
Wk(n)=Qk(n)*k; mW(n)=Wk(n)*mmW; sW(n)=mW(n)*wW;
me0(n)=mG(n)+mQ(n)+mW(n); se0(n)=sqrt(sG(n)^2+sQ(n)^2+sW(n)^2);
we0(n)=se0(n)./me0(n);
ske0(n)=(sQ(n)^3*skQ+sW(n)^3*skW)./se0(n)^3;
me(n)=(mG(n)+mQ(n)+mW(n))*Ke;
se(n)=me0(n)*Ke.*sqrt(wKe^2+we0(n)^2+we0(n)^2.*wKe^2);
ske(n)=me0(n)^3.*Ke^3.*(wKe^3*skKe+we0(n)^3.*ske0(n)+we0(n)^2*wKe^2*6)./se(n)^3;
%mG(n).*gG+Qk(n).*gQ+Wk(n).*psiW.*gW
pf(n)=Lnpf(mr,sr,skr,me(n),se(n),ske(n)); beta(n)= -ndinv(pf(n)); pft(n)=0.0000723;
betat(n)=3.8;

C - 24

Annex C - Calibration procedure

sg(n)=sqrt(sr^2+se(n)^2);
alG(n)=-mG(n)*sqrt(wG^2+wKe^2+wG^2*wKe^2)./sg(n);
alQ(n)=-mQ(n)*sqrt(wQ^2+wKe^2+wQ^2*wKe^2)./sg(n);
alW(n)=-mW(n)*sqrt(wW^2+wKe^2+wW^2*wKe^2)./sg(n);
ale(n)=-se(n)./sg(n);
alr(n)=sr./sg(n);
end% end of the loop
% Check selected values
beta(1), beta(21), %The first and the last Beta
sg(1);
alr(1); ale(1); %The first and alr and alr
sr; se(1);
%!!!!!!!!!!!!!!!!!!!!!!!!!Case B, (6.10a)
% Combination factors for expression (6.10a) for any k<=>(1-psiW)/(1-psiQ)=0,75
if k<=(1-psi1)/(1-psi2);
ksi=1; psiQ=psi1; psiW=psi2;
else
ksi=1; psiQ=psi1; psiW=psi2;
end
for n=1:12 %loop for CHI in the interval <-1,1>
CHIa(n)=0+(n-1)*0.0499;
mG(n)=Rd/(ksi*gG+(CHIa(n).*(psiQ*gQ+k*psiW*gW))./((1-CHIa(n))*(1+k)));
sG(n)=mG(n)*wG;
Qk(n)=CHIa(n).*mG(n)./((1-CHIa(n))*(1+k)); mQ(n)=Qk(n)*mmQ; sQ(n)=mQ(n)*wQ;
Wk(n)=Qk(n)*k; mW(n)=Wk(n)*mmW; sW(n)=mW(n)*wW;
me0(n)=mG(n)+mQ(n)+mW(n); se0(n)=sqrt(sG(n)^2+sQ(n)^2+sW(n)^2);
we0(n)=se0(n)./me0(n);
ske0(n)=(sQ(n)^3*skQ+sW(n)^3*skW)./se0(n)^3;
me(n)=(mG(n)+mQ(n)+mW(n))*Ke;
se(n)=me0(n)*Ke.*sqrt(wKe^2+we0(n)^2+we0(n)^2.*wKe^2);
ske(n)=me0(n)^3.*Ke^3.*(wKe^3*skKe+we0(n)^3.*ske0(n)+we0(n)^2*wKe^2*6)./se(n)^3;
%mG(n).*gG+Qk(n).*gQ+Wk(n).*psiW.*gW
pfa(n)=Lnpf(mr,sr,skr,me(n),se(n),ske(n)); betaa(n)= -ndinv(pfa(n));
end% end of the loop
%!!!!!!!!!!!!!!!!!!!!!!!!!Case B, (6.10b)
if k<=(1-psi1)/(1-psi2);
ksi=0.85; psiQ=1; psiW=psi2;
else
ksi=0.85; psiQ=psi1; psiW=1;
end
for n=1:19 %loop for CHI in the interval <-1,1> for expression (6.10b)
CHIb(n)=0.1+(n-1)*0.0499;
mG(n)=Rd/(ksi*gG+(CHIb(n).*(psiQ*gQ+k*psiW*gW))./((1-CHIb(n))*(1+k)));
sG(n)=mG(n)*wG;
Qk(n)=CHIb(n).*mG(n)./((1-CHIb(n))*(1+k)); mQ(n)=Qk(n)*mmQ; sQ(n)=mQ(n)*wQ;

C - 25

Annex C - Calibration procedure

Wk(n)=Qk(n)*k; mW(n)=Wk(n)*mmW; sW(n)=mW(n)*wW;
me0(n)=mG(n)+mQ(n)+mW(n); se0(n)=sqrt(sG(n)^2+sQ(n)^2+sW(n)^2);
we0(n)=se0(n)./me0(n);
ske0(n)=(sQ(n)^3*skQ+sW(n)^3*skW)./se0(n)^3;
me(n)=(mG(n)+mQ(n)+mW(n))*Ke;
se(n)=me0(n)*Ke.*sqrt(wKe^2+we0(n)^2+we0(n)^2.*wKe^2);
ske(n)=me0(n)^3.*Ke^3.*(wKe^3*skKe+we0(n)^3.*ske0(n)+we0(n)^2*wKe^2*6)./se(n)^3;
%mG(n).*gG+Qk(n).*gQ+Wk(n).*psiW.*gW
pfb(n)=Lnpf(mr,sr,skr,me(n),se(n),ske(n)); betab(n)= -ndinv(pfb(n));
end% end of the loop
%!!!!!!!!!!!!!!!!!!!!!!!!!Case C, (6.10amod)
ksi=1; psiQ=0.0; psiW=0.0;
%Effect of the load ratio CHI for expression (6.10amod)=(6.10c)
for n=1:4 %loop for CHI in the interval <-1,1>
CHIc(n)=0+(n-1)*0.0499;
mG(n)=Rd/(ksi*gG+(CHIc(n).*(psiQ*gQ+k*psiW*gW))./((1-CHIc(n))*(1+k)));
sG(n)=mG(n)*wG;
Qk(n)=CHIc(n).*mG(n)./((1-CHIc(n))*(1+k)); mQ(n)=Qk(n)*mmQ; sQ(n)=mQ(n)*wQ;
Wk(n)=Qk(n)*k; mW(n)=Wk(n)*mmW; sW(n)=mW(n)*wW;
me0(n)=mG(n)+mQ(n)+mW(n); se0(n)=sqrt(sG(n)^2+sQ(n)^2+sW(n)^2);
we0(n)=se0(n)./me0(n);
ske0(n)=(sQ(n)^3*skQ+sW(n)^3*skW)./se0(n)^3;
me(n)=(mG(n)+mQ(n)+mW(n))*Ke;
se(n)=me0(n)*Ke.*sqrt(wKe^2+we0(n)^2+we0(n)^2.*wKe^2);
ske(n)=me0(n)^3.*Ke^3.*(wKe^3*skKe+we0(n)^3.*ske0(n)+we0(n)^2*wKe^2*6)./se(n)^3;
%mG(n).*gG+Qk(n).*gQ+Wk(n).*psiW.*gW
pfc(n)=Lnpf(mr,sr,skr,me(n),se(n),ske(n)); betac(n)= -ndinv(pfc(n));
end% end of the loop
%!!!!!!!!!!!!!!!!!!!!!!!!!Case D CSN (6.10)
gG=1.2; gQ=1.4; gW=1.4; %alternative values og gamma
if k<=(1-psi1)/(1-psi2);
ksi=1; psiQ=1; psiW=psi2;
else
ksi=1; psiQ=psi1; psiW=1;
end
%Effect of the load ratio CHI for expression (6.10)
for n=1:21 %loop for CHI in the interval <0,1>
CHI(n)=0+(n-1)*0.0499;
mG(n)=Rd/(ksi*gG+(CHI(n).*(psiQ*gQ+k*psiW*gW))./((1-CHI(n))*(1+k)));
sG(n)=mG(n)*wG;
Qk(n)=CHI(n).*mG(n)./((1-CHI(n))*(1+k)); mQ(n)=Qk(n)*mmQ; sQ(n)=mQ(n)*wQ;
Wk(n)=Qk(n)*k; mW(n)=Wk(n)*mmW; sW(n)=mW(n)*wW;
me0(n)=mG(n)+mQ(n)+mW(n); se0(n)=sqrt(sG(n)^2+sQ(n)^2+sW(n)^2);
we0(n)=se0(n)./me0(n);

C - 26

Annex C - Calibration procedure

ske0(n)=(sQ(n)^3*skQ+sW(n)^3*skW)./se0(n)^3;
me(n)=(mG(n)+mQ(n)+mW(n))*Ke;
se(n)=me0(n)*Ke.*sqrt(wKe^2+we0(n)^2+we0(n)^2.*wKe^2);
ske(n)=me0(n)^3.*Ke^3.*(wKe^3*skKe+we0(n)^3.*ske0(n)+we0(n)^2*wKe^2*6)./se(n)^3;
%mG(n).*gG+Qk(n).*gQ+Wk(n).*psiW.*gW
pfd(n)=Lnpf(mr,sr,skr,me(n),se(n),ske(n)); betad(n)= -ndinv(pfd(n));
end% end of the loop
%!!!!!!!!!!!!!!!!!!!!!!!!!Case E: BSI (6.10)
ksi=1; psiQ=1; psiW=1; %reduction factor not considered
% Combination factors for expression (6.10) for one k=0, or two k>0
Rd=Rd*1.15/1.1; % adjustement for different for gm=1.10;
if k<=0.001;
gG=1.4; gQ=1.6; gW=1.4;
else
gG=1.2; gQ=1.2; gW=1.2;
end
%Effect of the load ratio CHI for expression (6.10)
for n=1:21 %loop for CHI in the interval <0,1>
CHI(n)=0+(n-1)*0.0499;
mG(n)=Rd/(ksi*gG+(CHI(n).*(psiQ*gQ+k*psiW*gW))./((1-CHI(n))*(1+k)));
sG(n)=mG(n)*wG;
Qk(n)=CHI(n).*mG(n)./((1-CHI(n))*(1+k)); mQ(n)=Qk(n)*mmQ; sQ(n)=mQ(n)*wQ;
Wk(n)=Qk(n)*k; mW(n)=Wk(n)*mmW; sW(n)=mW(n)*wW;
me0(n)=mG(n)+mQ(n)+mW(n); se0(n)=sqrt(sG(n)^2+sQ(n)^2+sW(n)^2);
we0(n)=se0(n)./me0(n);
ske0(n)=(sQ(n)^3*skQ+sW(n)^3*skW)./se0(n)^3;
me(n)=(mG(n)+mQ(n)+mW(n))*Ke;
se(n)=me0(n)*Ke.*sqrt(wKe^2+we0(n)^2+we0(n)^2.*wKe^2);
ske(n)=me0(n)^3.*Ke^3.*(wKe^3*skKe+we0(n)^3.*ske0(n)+we0(n)^2*wKe^2*6)./se(n)^3;
%mG(n).*gG+Qk(n).*gQ+Wk(n).*psiW.*gW
pfe(n)=Lnpf(mr,sr,skr,me(n),se(n),ske(n)); betae(n)= -ndinv(pfe(n));
end% end of the loop
% Ploting beta and pf and alpha versus CHI
ksi=0.85; psiQ=psi1; psiW=psi2;gG=1.35;gQ=1.5;gW=1.5;
if k<=(1-psi1)/(1-psi2);
a=1; b=psi2;
else
a=psi1; b=1;
end
CHIlim=gG*(1-ksi)*(1+k)/(gG*(1-ksi)*(1+k)+gQ*(a-psiQ)+gW*k*(b-psiW))
CHIlimc=gG*(1-ksi)*(1+k)/(gG*(1-ksi)*(1+k)+gQ*a+gW*k*b)
for n=1:14
B(n)=3+(n-1)*0.15;
C(n)=10^(-2-(n-1)*0.35);

C - 27

Annex C - Calibration procedure

end
maxbeta=ceil(max(beta));
maxpf=ceil(100*max(pf))/100;
h=figure(1);
% plot beta versus CHI
subplot(2,1,1)%CSN Beta: ,CHI,betad,'r.', BSI beta CHI,betae,'k--'
plot(CHI,beta,'r',CHI,betae,'k--',CHIa,betaa,'b',CHIb,betab,'b',CHIc,betac,'g--',CHI,betat,'k.',CHIlimc,B,'g.',CHIlim,B,'b.','LineWidth',1.5,'MarkerSize',5) %betaa,CHI,betab,CHI,betat,
grid,axis([0,1,3,6]) % maxbeta
ylabel('Index Beta')
subplot(2,1,2)
% plot pf versus CHI
semilogy(CHI,pf,'r',CHI,pfe,'k--',CHIa,pfa,'b',CHIb,pfb,'b',CHIc,pfc,'g--',CHI,pft,'k.',CHIlimc,C,'g.',CHIlim,C,'b.','LineWidth',1.5,'MarkerSize',5)
%plt using semilogarthmical scale, ,pfa,CHI,pfb,CHI
grid,axis([0,1,1e-8,maxpf])
ylabel('Probability Pf')
%subplot(3,1,3) % not generally active
% plot Alphas versus CHI
%plot(CHI,alG,CHI,alQ,CHI,alW,CHI,alr,CHI,ale,
%'LineWidth',1.5) %Alphasplt scale, ,pfa,CHI,pfb,CHI
%grid,axis([0,1,-1,1])
%xlabel('Load ratio CHI=(Qk+Wk)/(Gk+Qk+Wk)')
%ylabel('Alphas for A - (6.10)')

C - 28

Annex C - Calibration procedure

Attachment 3 MATLAB function "Lnpf(mr,sr,skr,me,se,ske)"
function pf = Lnpf (mr,sr,skr,me,se,ske)
% DESCRIPTION, 19.09.1999
% Lnpf evaluates the probability of failure pf considering the fundamental
% limit state function G = R - E.
% CALL
% val = Lnpf (mr,sr,skr,me,se,ske);
% FUNCTIONS USED
% lndens(x,ske,me,se).*Lndist(x,skr,mr,sr)
% INPUT
% mr : the mean of R
% sr : standard deviation of R
% skr : coefficient of skewness of R (must be given)
% me : the mean of E
% se : standard deviation of E
% ske : coefficient of skewness of E (must be given)
% OUPUT
% val : failure probability pf
% VERSION
% MH, Czech Technical University in Prague, Klokner Institute, 2.8.2003
% Initialization
cr=(0.5*skr+(skr^2/4+1)^0.5)^(1/3)-(-0.5*skr+(skr^2/4+1)^0.5)^(1/3);%constant of
lognormal R
ce=(0.5*ske+(ske^2/4+1)^0.5)^(1/3)-(-0.5*ske+(ske^2/4+1)^0.5)^(1/3);%constant of
lognormal E
if cr==0
r0=10^10;
else
r0=mr-sr/cr;
%bound of lognormal distribution of R
end
if ce==0
e0=10^10;
else
e0=me-se/ce;
%bound of lognormal distribution of E
end
% Determination of integration interval
k=10; % Coefficient of standard deviation
if skr>0;
%R positive
if ske>0;
%R and E positive
a=max(r0,e0);
b=min(mr+k*sr,me+k*se); % limit 6*sr
else
if ce==0
%R pos, E sym
a=max(r0,me-k*se);
b=min(mr+k*sr,me+k*se);
else
% R pos E neg
a=max(r0,me-k*se);
b=min(e0,mr+k*sr);
end

C - 29

Annex C - Calibration procedure

end
else
% R sym or neg
if cr==0;
% R sym c=0 (sk=0)
if ce==0
% both R and E sym
a=max(mr-k*sr,me-k*se);
b=min(mr+k*sr,me+k*se);
else
if ske>0 %R sym, E pos
a=max(mr-k*sr,e0);
b=min(mr+k*sr,me+k*se);
else
% R sym, E neg
a=max(mr-k*sr,me-k*se);
b=min(mr-k*sr,e0);
end
end
else
% R neg
if ske>0; % R neg, E pos
a=max(mr-k*sr,e0);
b=min(r0,me+k*se);
else
% E sym or neg
if ce==0 % R neg, E sym
a=max(mr-k*sr,me-k*se);
b=min(r0,me+k*se);
else
% R neg, E neg
a=max(mr-k*sr,me-k*se);
b=min(e0,r0);
end
end
end
end
% Integration interval
if a<0
a=0;
end
delta=b-a;
a=a+0.00000001*delta; % adjusted integration limits to avoid singularity
b=b-0.00000001*delta;
if a>b
' error in input data, a>b, increase the lower limit of the design parameter'
pause
else
n=20; inc=(b-a)/n; Y=0; % parameters of integration, n may adjusted if needed
for x=a:inc:b; % Integration of failure probability by trapezoidal rule
e= Lndens(x,me,se,ske); % call Lndens
r= Lndist(x,mr,sr,skr); % call Lndist
Y=Y+ e*r; % sum of e*r
end
pf=inc*Y;
end

C - 30

Annex C - Calibration procedure

Attachment 4 MATLAB function"Lndens(x,mu,sigma,sk)"
function val = Lndens (x,mu,sigma,sk)
% DESCRIPTION
% NDENS evaluates the one-dimensional normal density function.
% CALL
% val = Lndens (x,sk);
% val = Lndens (x,sk,mu,sigma);
% INPUT
% x : real vector of arguments
% sk : coefficient of skewness (must be given)
% mu : mean value; optional; default = 0.0 (i.e. standard)
% sigma : std. dev. > 0; optional; default = 1.0 (i.e. standard)
% OUPUT % val: vector of normal density values for the x's
% VERSION % Milan Holicky, Czech Technical University in Prague, Klokner Institute
% 18.09.1999
if nargin < 3
mu = 0.0;
sigma = 1.0;
end
% Evaluate
x = (x-mu)/sigma;
% normalize
c=(0.5*sk+(sk^2/4+1)^0.5)^(1/3)-(-0.5*sk+(sk^2/4+1)^0.5)^(1/3);%constant of lognormal
if c==0; % sk=0
x0=10^10;
else
x0=-1/c; %bound of the distribution
end
if sk>0; %check of x range
if x0>x
error ('x out of range')
else
end
else
if x0<x
error('x out of range')
else
end
end
%
if abs(c)>0.0001; %if for c=0 (sk=0)
tt=sign(sk)*(log(abs(x+1/c))+log(abs(c))+0.5*log(1+c^2))/((log(1+c^2))^0.5);
else
tt=x;
end
if abs(c)>0.0001; %if for c=0 (sk=0)
val = exp(-0.5*tt.^2)/(sqrt(2*pi)*sigma*abs(x+1/c)*(log(1+c^2))^0.5);
else
val = exp(-0.5*tt.^2)/(sigma*sqrt(2*pi));
end

C - 31

Annex C - Calibration procedure

Attachment 5 MATLAB function"Lndist(x,mu,sigma,sk)"
function val = Lndist (x,mu,sigma,sk)
% DESCRIPTION, 18.09.1999
% LNDIST evaluates the one-dimensional lognormal distribution function.
% CALL
% val = lndist (x,sk)
% val = lndist (x,sk,mu,sigma)
% INPUT
% x : real vector of arguments.
% sk : coefficient of skewness; default = 0.0 (i.e. normal distribution)
% mu : mean value; optional; default = 0.0 (i.e. standard).
% sigma : std. dev. > 0; optional; default = 1.0 (i.e. standard).
% OUPUT
% val : vector of the lognormal distribution evaluated at the x's.
% VERSION
% Milan Holicky, Czech Technical University in Prague, Klokner Institute
% Initialization
if nargin < 3
mu = 0.0;
sigma = 1.0;
end
x = (x-mu)/sigma;
% standardize
c=(0.5*sk+(sk^2/4+1)^0.5)^(1/3)-(-0.5*sk+(sk^2/4+1)^0.5)^(1/3);%constant of lognormal
if c==0; % sk=0
x0=10^10;
else
x0=-1/c; %bound of the distribution
end
if sk>0; %check of x range
if x0>x
error ('x out of range')
else
end
else
if x0<x
error('x out of range')
else
end
end
if abs(c)>0.01; %if for c=0 (sk=0)
tt=sign(sk)*(log(abs(x+1/c))+log(abs(c))+0.5*log(1+c^2))/((log(1+c^2))^0.5);
else
tt=x;
end
val = (1+erf(tt/sqrt(2)))/2; % transformed error function

C - 32

Annex C - Calibration procedure

Attachment 6 MATLAB function"Ndinv(x,mu,sigma)"

function [x] = norminv(p,mu,sigma)
%NORMINV Inverse of the normal cumulative distribution function (cdf).
% X = NORMINV(P,MU,SIGMA) returns the inverse cdf for the normal
% distribution with mean MU and standard deviation SIGMA, evaluated at
% the values in P.
% Default values for MU and SIGMA are 0 and 1, respectively.
%
% MH, Klokner Institute, CTU Prague 4.8.2003
%
if nargin < 2
mu = 0;
end
if nargin < 3
sigma = 1;
end
% Return NaN for out of range parameters or probabilities.
sigma(sigma <= 0) = NaN;
p(p < 0 | 1 < p) = NaN;
x0 = -sqrt(2).*erfcinv(2*p);
x = sigma.*x0 + mu;

C - 33

Annex C - Calibration procedure

Attachment 7 MATHCAD sheet"Generic.Mcd

Mathcad sheet "Generic" is intended to investigate combination rules provided in EN

1990 by expressions (6.10), (6.10a) and (6.10b) considering a generic structural memeber.
Turkstra's rule ( 50 years extremes of a leading and annual extremes for accompnying
action) is applied for the reference period of 50 years. MH, August 2002.
Design expression:
R0k / = () G*Gk + (Q)Q*Qk +(W)* W*Wk
Limit state function:
g(X) = R0 * (G + Q + W), R0=K*fu
Resistance of an element R = * R0 is described by two parameter lognormal
distribution LN( R, R,), basic variables R0 by LN(Rk, 0.06R), =1/(1 2*wR0) and by
LN(1.1, 0.05).
Load effect E = * (G + Q + W) is discribed by athree parameter lognormal distribution
LN ( E, E, E), permanent load G by N(Gk, 0.1*Gk), 50 yearsimposed load Q by GUM(0.6
Qk, 0.35 Q), annual wind load W by GUM(0.3Wk, 0.5W), uncertainty by LN(1.0, 0.05).
Parameters: =( Qk+Wk)/(Gk+Qk+Wk), k=Wk/Qk, factorsm, G, Q,W, , Q, W.
1 Input data:

Constants:

:= 0.85 Q := 0.70

Range variables: := 0 , 0.09.. 0.99

Load parameters:

m := 1.0, 1.05.. 1.5

W := 0.60

G := 1.1, 1.15.. 1.5

Normal distribution of G:
Gumbel distribution of Q:

k := 0.0

Q = mQ*Qk

Rd := 1

Q := 1.2, 1.25.. 1.6 W := 1.2, 1.25.. 1.6

G = Gk

wG := 0.1

mQ := 0.6

wQ := 0.35

Gumbel distribution of W: W = mW*Wk

mW := 0.3 wW := 0.5

Model uncertainty , Lognormal distribution:

:= 1.0

Resistance varables: R0 = *R0k:

w := 0.05

wR0 := 0.0, 0.12.. 0.35 ( wR0) := exp( 1.65 wR0)

Model uncertainty , Lognormal distribution:

:= 1.0

Q := 1.14
W := 1.14
:= 3 w + w

3
3

R0( wR0) := wR0 3 + wR0

w := 0.05

:= 3 w + w

2 Determination of the load variablesG, Q and W for a given resistance Rd:

Characteristic values Gk= G, Qk and Wk determined assuming Ed=Rd. Thus G*Gk+ Q*Qk+ W*Wk=Rd
G( , G, Q , W ) :=

Qk( , G, Q , W ) :=

Rd

G + ( Q + k W )

( 1 + k) ( 1 )

G( , G, Q , W )
( 1 + k) ( 1 )

Check:

Wk( , G, Q , W ) := k Qk( , G, Q , W )

G( 0.5, 1.35, 1.5, 1.5) = 0.351

Qk( 0.5, 1.35, 1.5, 1.5) = 0.351

Ed( , G, Q , W ) := G G( , G, Q , W ) + Q Qk( , G, Q , W ) + W Wk( , G, Q , W ) Wk( 0.5, 1.35, 1.5, 1.5) = 0

Normal distribution of G:

G( , G, Q , W ) := wG G( , G, Q , W )

Gumbel distribution of Q:

Q( , G, Q , W ) := mQ Qk( , G, Q , W )

Gumbel distribution of W

W ( , G, Q , W ) := mW Wk( , G, Q , W ) W ( , G, Q , W ) := wW W ( , G, Q , W )

Ed( 0.5, 1.35, 1.5, 1.5) = 1

Q( , G, Q , W ) := wQ Q( , G, Q , W )

3 Load effect E = (G+Q+W) = *E0:

The mean and st. deviation of E0: E0( , G, Q , W ) := G( , G, Q , W ) + Q( , G, Q , W ) + W ( , G, Q , W )
E0( , G, Q , W ) :=

wG G( , G, Q , W ) + wQ Q( , G, Q , W ) + wW W ( , G, Q , W )
2

C - 34

Annex C - Calibration procedure

E0( 0.4, 1.35, 1.5, 1.5) = 0.596

The coefficient of variation of E0 (without model uncertainty ):

wE0( , G, Q , W ) :=

E0( , G, Q , W )

E( , G, Q , W ) := E0( , G, Q , W )

E0( , G, Q , W )

The coefficient of variation of E:

wE( , G, Q , W ) :=

wE0( , G, Q , W ) + w + wE0( , G, Q , W ) w

E( , G, Q , W ) := E( , G, Q , W ) wE( , G, Q , W )

The standard deviation of E:

4 Three parameter lognormal distribution of E:

Skewness of E0:

wE0( 0.4, 1.35, 1.5, 1.5) = 0.123

E0( , G, Q , W ) :=

wE( 0.4, 1.35, 1.5, 1.5) = 0.133

3

Q( , G, Q , W ) Q + W ( , G, Q , W ) W
E0( , G, Q , W )

E0( 0.4, 1.35, 1.5, 1.5) = 0.614

Skewness of E:
E( , G, Q , W ) :=

wE0( , G, Q , W ) E0( , G, Q , W ) + 6 w wE0( , G, Q , W ) + w

wE( , G, Q , W )

E( 0.4, 1.35, 1.5, 1.5) = 0.591

Parameter C:
1

2
2

E( , G, Q , W ) + 4 + E( , G, Q , W ) E( , G, Q , W ) + 4 E( , G, Q , W )
C( , G, Q , W ) :=
1

Parameters of transformed variable:

mE( , G, Q , W ) := ln( C( , G, Q , W )

sE( , G, Q , W ) := ln 1 + C( , G, Q , W )

) + ln( E( , G, Q , W )) ( 0.5) ln(1 + C( , G, Q, W )2)

) x0( , G, Q, W) := E( , G, Q, W )

Probability density of E, approximation by three parameter lognormal distribution:

C( , G, Q , W )

E( , G, Q , W )

x0( 0.4, 1.35, 1.5, 1.5) = 0.189

Eln( x, , G, Q , W ) := dlnorm( x x0( , G, Q , W ) , mE( , G, Q , W ) , sE( , G, Q , W ) )

R( m, wR0) := Rd m ( wR0)

5 Resistance variables R = *R 0:
wR( wR0) :=

wR0 + w + wR0 w

Check:

R( m, wR0) := wR( wR0) R( m)

R ( 1.15, 0.15) = 1.473

wR( 0.15) = 0.158

6 Two parameter lognormal distribution of R:

Transformed variable: mR( m, wR0) := ln( R( m, wR0) ) ( 0.5) ln 1 + wR( wR0) 2

Distribution function Rln( x, m, wR0) := plnorm( x, mR( m, wR0) , sR ( m, wR0) )

C - 35

sR ( m, wR0) := ln 1 + wR( wR0)

mR( 1 , 015) = 22.038 sR ( 1 , 0.15) = 0.157

Annex C - Calibration procedure

7 Failure probability and reliability index : d ( , G, Q , W ) := if( x0( , G, Q , W ) 0 , 0 , x0( , G, Q , W ))

pf50( , m, wR0, G, Q , W ) :=

Eln( x, , G, Q , W ) Rln( x, m, wR0) dx

d( , G , Q , W)

pf50( 0.0, 1.15, 0.15, 1.35, 1.5, W 1.5) = 1.645 10

ln( , m, wR0, G, Q , W ) := qnorm( pf50( , m, wR0, G, Q , W ) , 0 , 1)

8 Reliability index versus ratio : limit for dominant action:

a := if( k k0, 1 , Q)

b := if( k > k0, 1 , W )

Limit value of for (6.10a) and (6.10b):

t := 3.8 Auxiliary:

a := 0 , 0.05.. l + 0.05

( G, Q , W ) :=

wG = 0.1

k0 :=

1 Q

Check:

1 W

b = 0.6

G ( 1 ) ( 1 + k)

G ( 1 ) ( 1 + k)

G ( 1 ) ( 1 + k) + ( Q a + W k b )

la := a( 1.35, 1.5, 1.5)

l := ( 1.35, 1.5, 1.5)

0 := 3 , 3.5.. 5

b := la 0.01, la + 0.04.. 0.999

mQ = 0.6

k0 = 0.75

G ( 1 ) ( 1 + k) + Q ( a Q) + W k ( b W )

c := 0 , 0.05.. la + 0.04 ln( 0 , 1.15, 0.15, 1.35, a 1.5, b 1.5) = 3.591

Turkstra's for 50 years:

ln( 0.0, 1.15, 0.15, 1.35, 1.5, W 1.5) = 3.591

a=1

Limit of for (6.10a-mod) and (6.10b): a( G, Q , W ) :=

Target probability

d ( 0.4, 1.35, 1.5, 1.5) = 0.189

am := 0 , 0.05.. la + 0.04

Check:

ln( 0.999, 1.15, 0.15, 1.35, a 1.5, b 1.5) = 3.583

wQ = 0.35 mW = 0.3 wW = 0.5

k=0

l = 0.31

la = 0.119

5.5

4.5

3.5

.
3

0.2

0.4

0.6

0.8

Figure 1: Reliability index corresponding to equation (6.10), (6.10a) and (6.10b).

Note that expression (6.10) is represented by solid red line, expressions (6.10a) and
(6.10b) by solid blue line,expression (6.10a,mod) by dashed green line, BSI
combination by dashed brown line, CSN combination by red dashed line, target
reliability index by horizontal black dashed line. The twin expressions (6.10a) and
6.10b) provide the most uniform distribution of the reliability index against the load
ratio with reliability index greater than the target value = 3,8 for majority of .

C - 36

Annex C - Calibration procedure

9 Reliability index versus G a Q:

1( G, Q) := ln( 0.4, 1.15, 0.15, G, a Q , b Q)

1t( G, Q) := 3.8

2( , m) := ln( , m, 0.10, 1.2, a 1.4, b 1.4)

2t( , m) := 3.8

Figure 2: Reliability index

corresponding to expression
(6.10) versus partial factors
G a Q.
Parameters : k = 0
mQ = 0.6 wQ = 0.35 mW = 0.3 wW = 0.5

1 , 1t

10 Reliability index versus m a :

Figure 3: Reliability index
corresponding to expression
(6.10) versus partial factor m
and parameter .
Parameters : k = 0
mQ = 0.6 wQ = 0.35 mW = 0.3 wW = 0.5

2 , 2t

C - 37

Annex C - Calibration procedure

11 Reliability index versus m a :

2( , wR0) := ln( , 1.15, wR0, 1.35, a 1.5, b 1.5) 2t( , wR0) := 3.8

Figure 4: Reliability index

corresponding to expression
(6.10) versus parameter and
the coefficient of variationwR.
Parameters : k = 0
mQ = 0.6 wQ = 0.35 mW = 0.3 wW = 0.5

2 , 2t

12 Reliability index versus m a wR: 4( m, wR0) := ln(0.4, m, wR0, 1.2, a 1.4, b 1.4) 4t(m, wR0) := 3.8
Figure 5: Reliability index
corresponding to expression
(6.10) versus parameter R
and the coefficient of variation
wR.
Parameters : k = 0
mQ = 0.6 wQ = 0.35 mW = 0.3 wW = 0.5

4 , 4t

C - 38

Annex C - Calibration procedure

Attachment 8 MATHCAD sheet"LoadEffect.mcd"

MATHCADsheet "LoadEfect"
Mathcad sheet "LoadEffect" is intended for investigation of combination rules
provided in EN 1990 by expressions (6.10), (6.10a) and (6.10b) considering three loads: G,
Q and W. Turkstra's rule ( 50 years extremes of a leading and annual extremes of an
accompnying action) is applied. MH, May 2003.
Design Load effect:
Ed
= () G*Gk + (Q)Q*Qk +(W
Stochastic model:
E=
* (G + Q + W)
Load effect E = * (G + Q + W) is discribed by athree parameter lognormal distribution
LN ( E, E, E), permanent load G by N(Gk, 0.1*Gk), 50 yearsimposed load Q by GUM(0.6
Qk, 0.35 Q), annual wind load W by GUM(0.3Wk, 0.5W), uncertainty by LN(1.0, 0.05).
Parameters: =( Qk+Wk)/(Gk+Qk+Wk), k=Wk/Qk, factorsG, Q,W, , Q, W.
1 Input data:

Normailised load effect

Range variables: := 0.001, 0.01.. 0.99

Reduction factors:

Ed := 1

G := 1., 1.1.. 1.9

:= 0.85 Q := 0.70

W := 0.60

Q := 1.2, 1.25.. 1.6

Loasd ratio:

W := 1.2, 1.25.. 1.6

k := 0.0

2 Deterministic glabal factor:

( , G, Q , W ) := G ( 1 ) + ( Q + k W )

Global load factor according to EN 1990

( 1 + k)

3 Probabilistic models
Permanent load G:

Normal distribution of G:

Variable load Q:

Gumbel distribution of Q:

Variable load W:

Gumbel distribution of W: W = mW*Wk

mW := 0.3 wW := 0.5

Model uncertainty

Model uncertainty , Lognormal distribution:

:= 1.0 w := 0.05

Q = mQ*Qk

G = Gk

wG := 0.1

mQ := 0.6

wQ := 0.35

Q := 1.14
W := 1.14
:= 3 w + w

4 Determination of the load variables G, Q and W for a given load effect Ed = Rd:
Characteristic values Gk= G, Qk and Wk determined assuming Ed=Rd. Thus G*Gk+ Q*Qk+ W*Wk=Rd
G( , G, Q , W ) :=

Qk( , G, Q , W ) :=

Ed

G +

( Q + k W )

( 1 + k) ( 1 )

G( , G, Q , W )
( 1 + k) ( 1 )

Check:

Wk( , G, Q , W ) := k Qk( , G, Q , W )

G( 0.4, 1.35, 1.5, 1.5) = 0.426

Qk( 0.40, 1.35, 1.5, 1.5) = 0.284

Ed( , G, Q , W ) := G G( , G, Q , W ) + Q Qk( , G, Q , W ) + W Wk ( , G, Q , W ) Wk( 0.5, 1.35, 1.5, 1.5) = 0

Normal distribution of G:

G( , G, Q , W ) := wG G( , G, Q , W )

Gumbel distribution of Q:

Q( , G, Q , W ) := mQ Qk( , G, Q , W )

Gumbel distribution of W

W ( , G, Q , W ) := mW Wk( , G, Q , W ) W ( , G, Q , W ) := wW W ( , G, Q , W )

Ed( 0.5, 1.35, 1.5, 1.5) = 1

Q( , G, Q , W ) := wQ Q( , G, Q , W )

5 Load effect E = (G+Q+W) =*E0:

The mean and st. deviation of E0: E0( , G, Q , W ) := G( , G, Q , W ) + Q( , G, Q , W ) + W ( , G, Q , W )
E0( , G, Q , W ) :=

wG G( , G, Q , W ) + wQ Q( , G, Q , W ) + wW W ( , G, Q , W )
2

C - 39

Annex C - Calibration procedure

The coefficient of variation of E0 (without model uncertainty):

wE0( , G, Q , W ) :=

E0( , G, Q , W )

E0( 0.4, 1.35, 1.5, 1.5) = 0.596

E( , G, Q , W ) := E0( , G, Q , W )

E0( , G, Q , W )

The coefficient of variation of E: wE( , G, Q , W ) :=

wE0( , G, Q , W ) + w + wE0( , G, Q , W ) w

E( , G, Q , W ) := E( , G, Q , W ) wE( , G, Q , W )

The standard deviation of E:

6 Three parameter lognormal distribution of E:

Skewness of E0:

wE0( 0.4, 1.35, 1.5, 1.5) = 0.123

wE( 0.4, 1.35, 1.5, 1.5) = 0.133

Q( , G, Q , W )
E0( , G, Q , W ) :=

Q + W ( , G, Q , W ) W

E0( , G, Q , W )

E0( 0.4, 1.35, 1.5, 1.5) = 0.614

Skewness of E:
E( , G, Q , W ) :=

wE0( , G, Q , W ) E0( , G, Q , W ) + 6 w wE0( , G, Q , W ) + w

wE( , G, Q , W )

E( 0.4, 1.35, 1.5, 1.5) = 0.591

Parameter C:
1

E( , G, Q , W ) 2 + 4 + E( , G, Q , W ) E( , G, Q , W ) 2 + 4 E( , G, Q , W )

C( , G, Q , W ) :=
1

Parameters of transformed variable:

mE( , G, Q , W ) := ln( C( , G, Q , W )

sE( , G, Q , W ) := ln 1 + C( , G, Q , W )

) + ln( E( , G, Q , W ) ) ( 0.5) ln( 1 + C( , G, Q, W ) 2)

) x0( , G, Q, W) := E( , G, Q, W)

Probability density of E, approximation by three parameter lognormal distribution:

C( , G, Q , W )

E( , G, Q , W )

x0( 0.4, 1.35, 1.5, 1.5) = 0.189

Eln( x, , G, Q , W ) := dlnorm( x x0( , G, Q , W ) , mE( , G, Q , W ) , sE( , G, Q , W ) )

7 Theoretical value of the global factor for a given exceedance probability p of Ed:
p = P{ E > Ed }

p := 0.001, 0.0011.. 0.006

Probability considered in EN:

1 pnorm( 0.7 3.8, 0 , 1) = 3.907 10

pp ( Ed , , G, Q , W ) := plnorm( Ed x0( , G, Q , W ) , mE( , G, Q , W ) , sE( , G, Q , W ) )

Ed( p , , G, Q , W ) := x0( , G, Q , W ) + qlnorm( 1 p , mE( , G, Q , W ) , sE( , G, Q , W ) )
Ek( , G, Q , W ) := G( , G, Q , W ) + Qk( , G, Q , W ) + Wk( , G, Q , W )
p ( p , , G, Q , W ) :=

pp ( 0.8, 0.4, 1.35, 1.5, 1.5) = 0.986

Ed( 0.002, 0.4, 1.35, 1.5, 1.5) = 0.884

Ed( p , , G, Q , W )
Ek( , G, Q , W )

p ( 0.002, 0.3, 1.35, 1.5, 1.5) = 1.228

p ( 0.002, 0.3, 1.5, 1.5, 1.5) = 1.228

C - 40

Annex C - Calibration procedure

8 The global load factor versus ratio : limit for dominant action: k0 :=
a := if( k k0, 1 , Q)

Auxiliary quantities:
Limit value of for (6.10a) and (6.10b):

( G, Q , W ) :=

Limit of for (6.10a-mod) and (6.10b): a( G, Q , W ) :=

Target probability

t := 3.8 Auxiliary:

a := 0 , 0.05.. l + 0.05
Turkstra's for 50 years:

1 W

b := if( k > k0, 1 , W )

a=1

k0 = 0.75
b = 0.6

G ( 1 ) ( 1 + k) + Q ( a Q) + W k ( b W )
la := a( 1.4, 1.6, 1.6)
l := ( 1.35, 1.5, 1.5)

0 := 1 , 1.1.. 1.3

mQ = 0.6

Check:

G ( 1 ) ( 1 + k)

G ( 1 ) ( 1 + k)
(
)
G 1 ( 1 + k) + ( Q a + W k b )

b := la 0.01, la + 0.04.. 0.999

wG = 0.1

1 Q

c := 0 , 0.05.. la + 0.04

wQ = 0.35 mW = 0.3 wW = 0.5

Check:
k=0

l = 0.31
la = 0.116

1.5

1.4

1.3

1.2

1.1

0.2

0.4

0.6

0.8

Figure 1: Deterministic global load factorE versus , code values corresponding

to equation (6.10), (6.10a) and (6.10b).

The global load factor E corresponding to expression (6.10) - combination A is

represented by solid red line, factor corresponding to expressions (6.10a) and (6.10b) combination B by blue line and factor corresponding toexpression (6.10a,mod) combination C by dashed green line. Vertical dashed lines indicate boundaries for
validity of expressions (6.10a) and (6.10b) orexpressions (6.10a,mod) and (6.10b).

C - 41

Annex C - Calibration procedure

p ( 0.001 , , 1.35 , 1.5 , 1.5)

1.5

p ( 0.002 , , 1.35 , 1.5 , 1.5)

p ( 0.004 , , 1.35 , 1.5 , 1.5)

1.4

( , 1.35 , 1.5 , W 1.5)

( a , 1.35 , Q 1.5 , W 1.5) 1.3
( b , 1.35 , 1.5 , W 1.5)
( c , 1.35 , 0 , 0)

1.2

0
0

1.1

0.2

0.4

0.6

, , , , a , b , c , l , la ,

0.8
.

Figure 2: Global load factor versus , theoretical and deterministic values

corresponding to equation (6.10), (6.10a) and (6.10b).
9 Global index versus and Q

G := 1.35 Q := 1.5

W := 1.5

( p , ) := p( p , , G, a Q , b W )

( p , ) := ( , 1 1.35, 1 1.5, W 1.5)

Parameters : k = 0
mQ = 0.6 wQ = 0.35
mW = 0.3 wW = 0.5
a=1

b = 0.6

( 0.004, 0.35) = 1.198

Figure 3: Global factor versus

probability p and load ratio
for load combination given by
equation (6.10).

C - 42