Reliability Based Code Calibration
Reliability Based Code Calibration
la=da
Paper for the Joint Committee on Structural Safety
Draft, March, 2002
M.H. Faber
Swiss Federal Institute of Technology, Zürich, Switzerland.
J.D. Sørensen
Aalborg University, Aalborg, Denmark.
Keywords: Code calibration, stochastic models, target reliabilities, partial safety factors,
reliability analysis, decision theory
Summary
The present paper addresses fundamental concepts of reliability based code calibration. First ba-
sic principles of structural reliability theory are introduced and it is shown how the results of
FORM based reliability analysis may be related to the partial safety factors and characteristic
values. Thereafter the code calibration problem is presented in its principal decision theoretical
form and it is discussed how acceptable levels of failure probability (or target reliabilities) may
be established. Furthermore suggested values for acceptable annual failure probabilities are
given for the ultimate and the serviceability limit states. Finally the paper describes a procedure
for the practical implementation of reliability based code calibration of LRFD based design
codes. In conclusion a number of important aspects are discussed relating to the difficulties en-
countered in practice when design codes are calibrated across a broad variety of types of struc-
tures, materials and loading conditions.
1. Introduction
Ultimately structural design codes are established for the purpose of providing a simple, safe
and economically efficient basis for the design of ordinary structures under normal loading, op-
erational and environmental conditions. Design codes thereby not only greatly facilitate the daily
work of structural engineers but also provide the vehicle to ensure a certain standardization
within the structural engineering profession which in the end enhances an optimal use of the re-
sources of society for the benefit of the individual.
Traditionally design codes take basis in design equations from which the reliability verification
of a given design may be easily performed by a simple comparison of resistances and loads
and/or load effects. Due to the fact that loads and resistances are subject to uncertainties, design
values for resistances and load effects are introduced in the design equations to ensure that the
design is associated with an adequate level of reliability. Design values for resistances are intro-
duced as a characteristic value of the resistance divided by a partial safety factor (typically larger
than 1) and design values for load effects are introduced as characteristic values multiplied by a
partial safety factor (typically larger than 1). Furthermore in order to take into account the effect
of simultaneously occurring variable load effects so-called load combination factors (smaller
than 1) are multiplied on one or more of the variable loads.
1
Paper for the Joint Committee on Structural Safety
Draft, March 2002
Over the years different approaches for establishing design values for resistances and loads
have been applied in different countries. Within the last decade, however, almost all design
codes have adopted the Load and Resistance Factor Design format (LRFD). Different versions
of the LRFD format exist see e.g. CIRIA [1], CEB [2] and [3], Eurocodes [4], AHSTO [5] and
OHBDC [6], but they are essentially based on the same principles.
The structural engineering profession has an exceptionally long tradition going several thou-
sand years back. During these years experience and expertise has been collected to some extent
by trial end error. The design of new types of structures, with new materials or subject to new
loading and environmental conditions had to be performed in an adaptive manner based on care-
ful and/or “conservative” extrapolations of existing experience. The results were not always sat-
isfactorily and some iteration has in general been necessary. In fact one may consider the present
structural engineering traditions as being the accumulated experience and knowledge collected
over this long period. This applies not least to the level of inherent safety with which the present
engineering structures are being designed.
The development of structural reliability methods during the last 3 to 4 decades have provided
a more rational basis for the design of structures in the sense that these methods facilitate a con-
sistent basis for comparison between the reliability of well tested structural design and the
reliability of new types of structures. For this reason the methods of structural reliability have
been applied increasingly in connection with the development of new design codes over the last
decades.
By means of structural reliability methods the safety formats of the design codes i.e. the design
equations, characteristic values and partial safety factors may be chosen such that the level of
reliability of all structures designed according to the design codes is homogeneous and inde-
pendent of the choice of material and the prevailing loading, operational and environmental con-
ditions. This process including the choice of the desired level of reliability or “target reliability”
is commonly understood as “code calibration”. Reliability based code calibration has been for-
mulated by several researchers, see e.g. Ravindra and Galambos [7], Ellingwood et al. [8] and
Rosenblueth and Esteva [9] and has also been implemented in several codes, see e.g. OHBDC
[6], NBCC [10] and more resent the Eurocodes [4].
The present paper attempts to give an overview of the methodology applied in reliability based
code calibration. First a short description of the LRFD safety format is given. Secondly the fun-
damental principles of structural reliability theory and the relation between reliability analysis
results and the LRFD safety format are explained. Thereafter a decision theoretical formulation
of the code calibration problem is formulated, the issue concerning the choice of target reliabil-
ities is discussed and guidelines are given for the rational treatment of this problem. Finally a
practically applicable approach for reliability based code calibration is proposed.
2
M. H. Faber & J.D. Sorensen
ing of these uncertainties and provides methods for the quantification of the probability that the
structures do not fulfil the performance criteria.
F = {g ( x) ≤ 0} (1)
where the components of the vector x are realizations of the so-called basic random variables X
representing all the relevant uncertainties influencing the probability of failure. The basic ran-
dom variables must be able to represent all types of uncertainties that are included in the analy-
sis.
Having established probabilistic models for the uncertain variables the problem remains to
evaluate the probability of failure corresponding to a specified reference period. However, also
other non-failure states of the considered component or system may be of interest, such as exces-
sive damage, unavailability, etc. In general any state, which may be associated with conse-
quences in terms of costs, loss of lives and impact to the environment are of interest. In the fol-
lowing, however, for simplicity these states are not differentiated.
For a structural component for which the uncertain resistance R is modeled by a random vari-
able with probability density function f R (r ) subjected to the load s the limit state function is
simply
g ( x) = R − s (2)
PF = P( R ≤ s ) = FR ( s ) = P( R / s ≤ 1) (3)
3
Paper for the Joint Committee on Structural Safety
Draft, March 2002
In case also the load s is uncertain and modeled by the random variable S with probability den-
sity function f S (s ) the probability of failure is
∞
PF = P ( R ≤ S ) = P( R − S ≤ 0) = ∫ FR ( x) f S ( x)dx (4)
−∞
assuming that the load and the resistance variables are statistically independent. This case is
called the fundamental case. The integration in (4) is illustrated in Figure 1. The contributions to
the probability integral in (4) are illustrated. Note that the probability of failure is not determined
through the overlap of the two curves.
f R ( r ), f S ( s )
Load S
Resistance R
dx x
Figure 1 Illustration of the integration in (4).
There exists no general closed form solution to the integral in (4) but for a number of special
cases solutions may be derived. One case is when both the resistance variable R and the load
variable S are normally distributed. In this case the failure probability may be assessed directly
by considering the random variable M often referred to as the safety margin
M = R−S (5)
PF = P( R − S ≤ 0) = P( M ≤ 0) (6)
where M is also being normal distributed with parameters µ M = µ R − µ S and standard devia-
tion σ M = σ R2 + σ S2 .
The failure probability may now be determined by use of the standard normal distribution func-
tion as
0 − µM
PF = Φ = Φ (− β ) (7)
σM
where µ M / σ M = β is called the safety index. The geometrical interpretation of the safety in-
dex is illustrated in Figure 2.
4
M. H. Faber & J.D. Sorensen
f M ((m
m)
Failure Safe
σM σM
m
µM
Figure 2 Illustration of the probability density function for the normally
distributed safety margin M.
From Figure 2 it is seen that the safety index β may be interpreted as the number of standard
deviation by which the mean value of the safety margin M exceeds zero, or equivalently the dis-
tance from the mean value of the safety margin to the most likely failure point.
As indicated previously closed form solutions may also be obtained for other special cases.
However, as numerical methods have been developed for the purpose of solving (4) we will not
consider these in the following.
In the general case the resistance and the load cannot be described by only two random vari-
ables but rather by functions of random variables. Then the more failure probability may be as-
sessed through the following integral
PF = ∫ f (x)dx
X (8)
g ( x ) ≤0
Joint probability
density function
Resistance
Load
Safe Failure
5
Paper for the Joint Committee on Structural Safety
Draft, March 2002
The first developments of First and Second Order Reliability Methods (FORM/SORM) took
place almost 30 years ago with pioneering work performed by Basler [14], Cornell [15] and
Hasofer & Lind [16]. Since then these methods together with advanced Monte Carlo simulation
techniques have been refined and extended significantly and by now they form the most impor-
tant methods for reliability evaluations in structural reliability theory. For the most common
practical purposes the problem of estimating probabilities may be considered as solved. Several
commercial computer codes have been developed for FORM/SORM and simulation analysis and
the methods are widely used in practical engineering problems and not least for code calibration
purposes, see e.g. STRUREL [17] and Proban [18].
Evaluation of (9) is generally very difficult, and approximations are used in practical applica-
tions. Often an upper bound of the probability of failure in the time interval [0, T ] is used:
T
PT (T ) ≤ ∫ν + (t , ξ )dt (10)
0
(
G = zRc / γ m − γ G Gc + γ Q QC = 0
a
) (11)
where
RC is the characteristic value for the resistance
z is a design variable (e.g. the cross sectional area of the steel rod considered previously)
GC is a characteristic value for the permanent load
QC is a characteristic value for the variable load
γm is the partial safety factor for the resistance
γG is the partial safety factor for the permanent load
γQ is the partial safety factor for the variable load
6
M. H. Faber & J.D. Sorensen
In the codes different partial safety factors are specified for different materials and for different
types of loads. Furthermore when more than one variable load is acting load combination factors
are multiplied on one ore more of the variable load components to take into account the fact that
it is unlikely that all variable loads are acting with extreme values at the same time.
The partial safety factors together with the characteristic values are introduced in order to en-
sure a certain minimum reliability level for the structural components designed according to the
code. As different materials have different uncertainties associated with their material parame-
ters the partial safety factors are in general different for the different materials. The principle is
illustrated in Figure 4 for the simple r-s case.
f R (r ), f S ( s )
S R
r,s
sd rc
γQ = γm =
sc SC RC rd
S d , Rd
Figure 4 Illustration of the relation between design values, characteristic
values and partial safety factors.
In accordance with a given design equation such as e.g. (9) a reliability analysis may be made
with a limit state function of the same form as the design equation but where the characteristic
values for the resistance and load variables are now replaced by basic random variables, i.e.
g = zR − (G + Q) = 0 (12)
For given probabilistic models for the basis random variables R, G and Q and with a given re-
quirement to the maximum allowable failure probability it is now possible to determine the value
of the design variable z which corresponds to this failure probability. Such a design could be in-
terpreted as being an optimal design because it exactly fulfils the given requirements to struc-
tural reliability.
Having determined the optimal design z we may also calculate the corresponding design point
in the original space, i.e. x d for the basic random variables. This point may be interpreted as the
most likely failure point, i.e. the most likely combination of the outcomes of the basic random
variables leading to failure. Now partial safety factors may be derived from the design point for
the various resistance variables as
7
Paper for the Joint Committee on Structural Safety
Draft, March 2002
xc
γm = (13)
xd
xd
γQ = (14)
xc
where xd is the design point for the considered design variable and xc the corresponding charac-
teristic value.
For time-variant reliability problems a similar procedure can be used to determine partial safety
factors.
L
[
max W (γ ) = ∑ w j B j − C Ij (γ ) − C Rj (γ ) − C Fj PFj (γ ) ]
γ j =1 (15)
s.t. γ il ≤ γ i ≤ γ iu , i = 1,..., m
where γ = (γ 1 ,..., γ m ) are the m partial safety factors to be calibrated. Load combination factors
will in general also be calibrated/optimized, therefore γ = (γ 1 ,..., γ m ) can e assumed also to con-
tain those load combination factors to be calibrated. γ 1l ,..., γ ml and γ 1u ,..., γ mu are lower and upper
bounds on the partial safety factors. L is the number of different failure modes / limit states used
to cover the application area considered. w j is a factor indicating the relative frequency of fail-
ure mode j . B j is the expected benefits (in general for the society, but in some cases the bene-
fits can be related to the owner of the structures considered), C Ij is the initial (or construction)
costs, C Rj is the repair/maintenance costs during the design life time and C Fj is the cost of fail-
ure. C Fj is assumed to be independent of the partial safety factors. PFj is the probability of fail-
ure for failure mode j if the structure is designed using given partial safety factors.
The formulation in (15) is based on single failure modes and corresponds to the single failure
mode checking format used in structural codes of practice. A similar systems approach can be
formulated where the probability of failure of the system can be determined assuming system
failure if one of the single failure modes fails (series system model) and where systems related
costs are introduced. However, the corresponding deterministic systems reliability measures (ro-
8
M. H. Faber & J.D. Sorensen
bustness measures) are difficult to identify and are generally not used in structural codes. In the
following the single failure mode checking format is assumed to be used.
The limit state functions related to the failure modes considered are written:
g j (x, p j , z ) = 0 (16)
where p j is a vector with deterministic parameters and z = ( z1 ,..., z N ) are the design variables.
The application area for the code is described by the set I of L different vectors p j , j = 1,..., L .
The set I may e.g. contain different geometrical forms of the structure, different parameters for
the stochastic variables and different statistical models for the stochastic variables.
The deterministic design equation related to the limit state equation in (16) is written
G j (x c , p j , z, γ ) ≥ 0 (17)
where x c = ( xc1 ,..., xcn ) are characteristic values and γ = (γ 1 ,..., γ m ) are the partial safety factors.
C Ij (γ ) , C Rj (γ ) and PFj (γ ) can be determined on the basis of the solution of the following de-
terministic optimization problem where the optimal design z is determined using the design
equations and given partial safety factors:
min C Ij (z )
γ
s.t. ( )
G j x c , p j , z, γ ≥ 0 (18)
z li ≤ z i ≤ z iu , i = 1,..., N
The objective function in (18) is the construction costs, and the constraints are related to the
design equations. Using the limit state equation in (16) the probability of failure of the structure
PFj and the expected repair/maintenance costs C Rj to be used in (15) are determined at the opti-
mum design point z * . In cases where more than one failure mode is used to design a structure
included in the code calibration, the relevant design equations all have to be satisfied for the op-
timal design z * . The objective function in (18) can be extended also to include the repair / main-
tenance costs and the benefits.
It is noted that when the partial safety factors are determined from (15) they will in general not
be independent. In the simplest case with only a resistance partial safety factor and a load partial
safety factor only the product of the two partial safety factors is determined.
9
Paper for the Joint Committee on Structural Safety
Draft, March 2002
is thus not immediately possible to judge whether the estimated reliability is sufficiently high
without first establishing a more formalized reference for comparison.
Such a reference may be established by the definition of an optimal or best practice structure.
The idea behind the "best practice" reference is that if the structure of consideration has been
designed according to the "best practice" then the reliability of the structure is "optimal" accord-
ing to agreed conventions for the target reliability. Typical values for the corresponding target
annual failure probability are in the range of 10-6 to 10-7 depending on the type of structure and
the characteristics of the considered failure mode. Using this approach the target reliability is
determined as the reliability of the "best practice" design as assessed with the given probabilistic
model.
The determination of the "best practice" design can be performed in different ways. The sim-
plest approach is to use the existing codes of practice for design as a basis for the identification
of "best practice" design. Alternatively the "best practice design" may be determined by consul-
tation of a panel of recognized experts.
In case where the probabilistic modeling does not rest on subjective assessments the most ra-
tional approach is to establish the optimal design on the basis of the economic decision theory.
By considering the expected total benefit E [B ] associated with the considered structure
E [B ] = I ⋅ (1 − PF (C D )) − C D − C F ⋅ PF (C D ) = I − C D − ( I + C F ) PF (C D ) (19)
where I is the expected benefit from the structure, C F is the cost consequence in case of failure,
C D is the cost of some risk reducing measure, e.g. an increase of a dimension, and where the
probability of failure is a function of the costs invested in the risk reduction we have that the op-
timal investment in risk reducing measures may be determined from the following optimality
criterion.
∂E [B ] ∂P (C )
= −1 − ( I + C F ) ⋅ F D = 0 (20)
∂C D ∂C D
from which the cost efficient level of risk reducing measures may be determined. Having deter-
mined these we may, by application of (17) assess the feasibility of the considered structure by
recognizing that the total expected benefit of the structure shall be larger than zero.
Without going in to the details prevailing the derivations it is interesting to notice that it is pos-
sible, based on recent research work by Nathwani and Lind [20] and Rackwitz [21] to establish
optimal values for risk reduction costs when also the consequences of loss of human lives are
considered by means of the Life Quality Index. The Life Quality Index, L is a compound social
indicator defined as
L = g w e1− w (21)
where g is the gross domestic product per year per person, e is the life expectancy at birth and w
is the proportion of life spent in economic activity. In developed countries it may be assumed
that w=1/8. g lies in the interval of $US 2600-14000 being average numbers ranging from poor
to well developed countries. The life expectancy at birth e being 56 years in poorly developed
10
M. H. Faber & J.D. Sorensen
countries, 67 years in medium developed countries and 73 years in highly developed countries,
see e.g. Skjong and Ronold [22]. The LQI implies that a risk reducing measure is feasible if
∆e ∆g w
≥− (22)
e g 1− w
which may be obtained from (21) as explained in Nathwani and Lind [20]. From (22) the optimal
risk reducing measure for saving the life of a person may be identified by considering the case of
equality. Then we obtain
g 1− w g 1− w
∆g max
= ∆e = (23)
e w 2 w
which may be interpreted as the optimal acceptable costs per life year saved and where it has
been assumed that number of life years saved by saving one individual ∆e in average equals
e
∆e = .
2
From (21) we may now readily calculate to optimal costs of saving the life of one individual,
also called the optimum acceptable implied cost of averting a fatality (ICAF) from
ge 1 − w
ICAF = (24)
4 w
from which it may be found that optimum values of ICAF lies in the range of $US 2 – 3 x 106.
These costs may be included in (19) when the optimal investments into safety are considered
and thus treated within the same framework as any other asset loss. It should be noticed that as a
consequence hereof the acceptable failure probability associated with a specific project or struc-
ture depends on its specific characteristic, i.e. the monetary consequences in case of failure to-
gether with the expected benefits of the activity.
In Tables 1 - 2 target failure probabilities and corresponding target reliability indexes are given
for ultimate limit states and serviceability limit states, respectively based on the recommenda-
tions of JCSS [12]. Note that the values given correspond to a year reference period and the sto-
chastic models recommended in JCSS [12].
Table 1: Tentative target reliability indices β (and associated target failure probabilities) re-
lated to a one-year reference period and ultimate limit states
Relative cost of Minor consequences Moderate consequences Large consequences
safety measure of failure of failure of failure
-3
High β=3.1 ( PF ≈10 ) β=3.3 ( PF ≈5 10-4) β=3.7 ( PF ≈10-4)
Normal β=3.7 ( PF ≈10-4) β=4.2 ( PF ≈10-5) β=4.4 ( PF ≈5 10-5)
Low β=4.2 ( PF ≈10-5) β=4.4 ( PF ≈10-5) β=4.7 ( PF ≈10-6)
Table 2: Tentative target reliability indices (and associated probabilities) related to a one-year
reference period and irreversible serviceability limit states
11
Paper for the Joint Committee on Structural Safety
Draft, March 2002
Ad 1.
The class of structures and the type of relevant failure modes to be considered are defined.
Ad 2.
The code objective may be defined using target reliability indices or target probability of fail-
ures. Those can for example be based on the target reliabilities indicated in recommendations
(e.g. Eurocodes [4] and ISO [24]), see section 4 or on reliabilities obtained by reliability analyses
of structures designed by old, well-proven and accepted structural codes of practice. It is impor-
tant to note that the target reliabilities are linked closely to the stochastic models used for the un-
certain variables and the applied limit states.
Ad 3.
The code format includes:
- how many partial safety factors and load combination factors to be used
- should load partial safety factors be material independent, and
- should material partial safety factors be load type independent
- how to use the partial safety factors in the design equations
- rules for load combinations
In general for practical use the partial safety factors should be as few and general as possible.
On the other hand a large number of partial safety factors is needed to obtain economically and
safe structures for a wide range of different types of structures.
Ad 4.
Within the class of structures considered typical failure modes are identified. Limit state equa-
tions and design equations are formulated and stochastic models for the parameters in the limit
12
M. H. Faber & J.D. Sorensen
state equations are selected. Also the frequency at which each type of safety check is performed
is determined.
The stochastic model for the uncertain parameters should be selected very careful. Guidelines
for the selection can be found in JCSS [12]. Also in the Eurocodes [4] and ISO [24] some guide-
lines can be found. In general the following main recommendations can be made.
Strength / resistance parameters are often modeled by Lognormal distributions. This avoids the
possibility of negative realizations. In some cases it can be relevant also to consider a Weibull
distribution for a material parameter. This is especially the case if the strength is governed by
brittleness, size effects and material defects. The coefficient of variation varies with the material
type considered. Typical values are 5 % for steel and reinforcement, 15 % for the concrete com-
pression strength and 15-20 % for the bending strength of structural timber. The characteristic
value is generally chosen as the 5 % quantile.
Variable loads (imposed and environmental) can be modeled in different ways, see JCSS [12].
The simplest model is to use a stochastic variable modeling the largest load within the reference
period (often one year). This variable is typically modeled by an extreme distribution such as the
Gumbel distribution. The coefficient of variation is typically in the range 20-40 % and the char-
acteristic value is chosen as the 98 % quantile in the distribution function for the annual maxi-
mum load.
Permanent loads are typically modeled by a Normal distribution since it can be considered as
obtained from many different contributions. The coefficient of variation is typically 5-10 % and
the characteristic value is chosen as the 50 % quantile.
Model uncertainties are in many cases modeled by a Lognormal distributions if the they are
introduced as multiplicative stochastic variables and by Normal distributions if the they are mod-
eled by additive stochastic variables. Typical values for the coefficient of variation are 3-15%
but should be chosen very carefully. The characteristic value is generally chosen as the 50 %
quantile.
Ad 5.
The partial safety factors γ are calibrated such that the reliability indices corresponding to L
different vectors p j are as close as possible to a target probability of failure PFt or equivalently
( )
a target reliability index β t = −Φ −1 PFt . This can be formulated by the following optimization
problem
min W (γ ) = ∑ w j (β j (γ ) − β t )
L 2
(25)
γ j =1
where w j , j = 1,..., L are factors ( ∑ Lj=1 w j = 1 ) indicating the relative frequency of appearance /
importance of the different design situations. Instead of using the reliability indices in (25) to
measure the deviation from the target, for example the probabilities of failure can be used:
13
Paper for the Joint Committee on Structural Safety
Draft, March 2002
L
(
min W ' (γ ) = ∑ w j PFj (γ ) − PF
γ j =1
t
)
2
(26)
where PFt is the target probability of failure in the reference period considered. Also, a nonlinear
objective function giving relatively more weight to reliability indices smaller than the target
compared to those larger than the target can be used.
The above formulations can easily be extended to include a lower bound on the reliability or
probability of failure for each failure mode.
Ad 6.
The optimal partial safety factors are obtained by numerical solution of the optimization prob-
lem in step 5. The reliability index β j (γ ) for combination j given the partial safety factors γ is
obtained as follows. First, for given partial safety factors γ the optimal design is determined.
If the number of design variables is N = 1 then the design z * can be determined from the de-
sign equation, see (17)
G j (x c , p j , z, γ ) ≥ 0 (27)
If the number of design variables is N > 1 then a design optimization problem can be formu-
lated:
min C (z )
s.t. ci (x c , p j , z, γ ) = 0 , i = 1,..., me
(28)
ci (x c , p j , z, γ ) ≥ 0 , i = me + 1,..., m
z il ≤ z i ≤ z iu , i = 1,..., N
C is the objective function and ci , i = 1,..., m are the constraints. The objective function C is
often chosen as the weight of the structure. The me equality constraints in (28) can be used to
model design requirements (e.g. constraints on the geometrical quantities) and to relate the load
on the structure to the response (e.g. finite element equations). Often equality constraints can be
avoided because the structural analysis is incorporated directly in the formulation of the inequal-
ity constraints. The inequality constraints in (28) ensure that response characteristics such as
displacements and stresses do not exceed codified critical values as expressed by the design (27).
The inequality constraints may also include general design requirements for the design variables.
The lower and upper bounds, z il and z iu , to z i in (28) are simple bounds. Generally, the optimi-
zation problem (28) is non-linear and non-convex.
Ad 8.
As discussed above a first guess of the partial safety factors is obtained by solving these opti-
14
M. H. Faber & J.D. Sorensen
mization problems. Next, the final partial safety factors are determined taking into account cur-
rent engineering judgment and tradition.
Examples of reliability-based code calibration can be found in Nowak [25], Sørensen et al. [26]
and SAKO [27].
6. Discussion
Code calibration using principles of structural reliability and/or taking basis in decision theo-
retical principles highly facilitates the development of rational and cost efficient design basis in
consistency with the prevailing uncertainties. The formal basis for reliability based code calibra-
tion in fact exists and has so for more than a decade. Since then various additional developments
have taken place of which maybe the most important are those concerned with the assessment of
target failure probabilities. However, the basic technique has essentially remained the same.
Code calibration in practice, however, is not always a straightforward task. The main reason for
this being that in order to obtain a consistent code calibration the assessment of design values or
more specifically, partial safety factors and load combination factors needs to be performed
jointly in essentially one process. This necessity in turn poses strong requirements to the organ-
izational aspects of the code calibration process, which may be manageable on a national scale
but which has proven to be rather difficult on an international scale. Due to the above-
mentioned organizational problems reliability based code calibration may in effect despite the
numerous theoretical and conceptual advantages lead to situations where the resulting design
codes become inconsistent.
In order to benefit from reliability based code calibration it is of utmost importance that consis-
tency is achieved and maintained between all steps in the probabilistic representation of
uncertainties and limit state functions across the application domain of the code in terms of types
of structures, load conditions and materials. In addition to this a strict consistency must be main-
tained between acceptance criteria and the applied probabilistic modeling. These aspects often
pose problems in practice when e.g. for reasons of tradition some partial safety factors are se-
lected by choice in accordance with previous codes and afterwards combined with partial safety
factors derived on the basis of reliability analysis. This in turn also has serious implications
when attempting to establish a consistent basis for design supported by experiments.
7. References
[1]CIRIA (1977). Rationalisation of Safety and Serviceability Factors in Structural Codes.
CIRIA Report No. 63, London.
[2]CEB (1976a). First Order Concepts for Design Codes. CEB Bulletin No. 112, Munich.
15
Paper for the Joint Committee on Structural Safety
Draft, March 2002
[3]CEB (1976b). Common Unified Rules for Different Types of Construction and Material, Vol.
1. CEB Bulletin No. 116, Paris.
[4]Eurocode 0 (2001). Basis of Structural Eurocodes, Eurocode 0, EN 1990.
[5]ASHTO (1994). ASHTO LRFD Bridge Design Specifications. American Association of State
Highway and Transportation Officials, Washington, DC.
[6]OHBDC (1983). OHBDC (Ontario Highway Bridge Design Code), Ontario Ministry of
Transportation and Communication, Ontario.
[7] Ravindra & Galambos (1978). Ravindra, M.K. & T.V. Galambos: Load and Resistance
Factor Design for Steel. ASCE, Journal of the Structural Division, Vol. 104, N0. ST9,
1978, pp. 1337-1353.
[8]Ellingwood et al. (1982). Ellingwood, B., J.G. MacGregor, T.V. Galambos & C.A. Cornell:
Probability Based Load Criteria: Load Factors and Load Combinations. ASCE, Journal of
the Structural Division, Vol. 108, N0. ST5, 1982, pp. 978-997.
[9]Rosenblueth & Esteva (1972). Rosenblueth, E. & L. Esteva : Reliability Basis for Some
Mexican Codes. ACI Publication SP-31, 1972, pp. 1-41.
[10] NBCC (1980). NBCC (National Building Code of Canada), National Research Council of
Canada.
[11] Melchers (1999). Melchers, R.E.: Structural Reliability, Analysis and Prediction. John
Wiley.
[12] JCSS (2001). JCSS: Probabilistic model code. The Joint Committee on Structural Safety.
[13] Madsen et al. (1986). Madsen, H.O., S. Krenk & N.C. Lind: Methods of Structural Safety.
Prentice-Hall.
[14] Basler (1961). Basler, B.: Untersuchungen über den Sicherheitsbegriff von Bauwerken,
Schweiz Arch. Vol. 27, No. 4, 1961, pp. 133-160.
[15] Cornell (1969). Cornell, C.A.: A Probability-Based Structural Code. ACI-Journal, Vol. 66,
1969, pp. 974-985.
[16] Hasofer & Lind (1974). Hasofer, A.M. and N.C. Lind: An Exact and Invariant First Order
Reliability Format. ASCE, J. Eng. Mech. Div., 1974, pp. 111-121.
[17] STRUREL (1998). STRUREL, Version 6.1, Theoretical, Technical and Users manual,
RCP-GmbH, Munich.
[18] Proban (1993). Proban, Version 4, Theory manual. DNV Research Report no. 93-2056.
[19] Sørensen et al. (1994). Sørensen, J.D., I.B. Kroon and M.H. Faber: Optimal Reliability-
Based Code Calibration. Structural Safety, Vol. 14, 1994, pp. 197-208.
[20] Nathwani et al. (1997). Nathwani, J.S., N.C. Lind and M.D. Pandey: Affordable Safety by
Choice: The Life Quality Methods. Institute for Risk Research, University of Waterloo,
Canada.
[21] Rackwitz (2001). Rackwitz, R.: A new Approach for setting Target Reliabilities. Proc
IABSE Conf. Safety, Risk and Reliability – Trends in Engineering, Malta, 2001, IABSE,
Zürich, 2001, pp. 531-536.
[22] Skjong & Ronold (1998). Skjong, R. and K. Ronold: Social Indicators and Risk Accep-
tance. Proc. OMAE 1998, Lisbon, Portugal.
[23] Thoft-Christensen & Baker (1982). Thoft-Christensen, P. & M.B. Baker: Structural Reli-
ability Theory and Its Applications. Springer Verlag.
[24] ISO 2394 (1998). ISO 2394. General Principles on Reliability for Structures.
[25] Nowak (1989). Nowak, A.S.: Probabilistic Basis for Bridge Design Codes. Proc. ICOS-
SAR'89, 1989, pp. 2019-2026.
[26] Sørensen et al. (2001). Sørensen, J.D., S.O. Hansen & T. Arnbjerg Nielsen. Calibration of
Partial Safety Factors for Danish Structural Codes. Proc. IABSE Conf. Safety, Risk and
Reliability – Trends in Engineering, Malta, 2001, IABSE, Zürich, pp. 179-184.
[27] SAKO (1999). SAKO: Probabilistic Calibration of Partial Safety Factors in the Eurocodes.
16
M. H. Faber & J.D. Sorensen
17