A second comparison index is proposed to account for the influence that
the quality of the experimental results has on the
reliability to be assessed to the
"class A" predictions.
The index is based on the measure of the half length "b" of
the confidence interval centered at the predicted value "p" which
contains the estimated true mean value "
" of the experiment
with the probability
.
In other words,
is the likelihood that the true
value
lies within the confidence interval of length 2b,
centered at the predicted value p:
This "b" index can only be evaluated when more than one experimental result is available for comparisons with "class A" predictions. The methodology is briefly described in the followings.
Before a series of N experiments are performed, their results
are random variables with expected values
and standard deviations
,
.
The following assumptions are made:
Under the above assumptions, the results
are identically distributed
random variables, i.e.
, for
all
(assumption 1), the expected outcome equals the
true value,
(assumption 2), and the average
outcome
is a normally distributed random variable, even for small N
(assumption 3).
The mean and standard deviation of
are
[12]:
and
, respectively.
It follows that the random variable:
is standard normal.
Since the real standard deviation of the test results is not known,
it is replaced by its estimate - the sample standard
deviation
, evaluated after the experiments have been
performed.
The new random variable:
can be shown to follow a Student's t distribution with N - 1 degrees
of freedom [12].
Denoting by
the respective cumulative distribution, and
replacing
in (5) using (6), one gets:
with:
After the experiments have been performed and the results
-
realizations of
- are known at every time instant t,
the random variable
in (8) is replaced by its realization
,
and, given a value for
,
the equation (7) can be numerically solved for
.
The meaning of the size b of the
confidence interval is illustrated
in the above figure.
Before the experiments are performed, the random variable
has
the probability distribution function
, shown in
the figure.
If the standard deviation of the test results
is known,
is derived from the normal distribution, and it comes
from the Student's t distribution if
is
approximated by the sample standard deviation
.
After the experiments,
is replaced by its realization
,
but the true value
remains unknown.
Under the "random sampling" assumption, the uncertainty in locating
the true value
relative to
can be expressed
by the same probability distribution function.
Finally, given the predicted value p, the
confidence interval
has a size 2b, corresponding to an area which equals the
likelihood
(hatched area).
An overall measure of how close the predicted time history is to the
experimental measurements is obtained by averaging the value
,
computed at every time instant t, over the analysis time interval
.
Finally, to get a nondimensional index, the average value is normalized with
respect to the initial effective vertical stress
:
The quantity
, evaluated for a certain
value,
will be
refered to as "size of
confidence interval".