2. GENERALIZATION OF RELIABILITY CONCEPT
The originally defind reliability only describes the ability of a
photogrammetric system to detect gross and systematic errors with
the aid of a statistical test. Now, in order that the reliability
concept can be Spoken about both to a system and to a result,and
not only in the case of using data-snooping but also in that of
using robust methodS, we need to generalize the concept of
reliability as follows:
Reliability is concerned with the bias, or the difference of the
Statistical expectation of an estimator from the true value of the
estimated quantity. The reliability of a System describes the
ability of the system to avoid biases. While the reliability of a
result describes the biases contained in the concrete result. The
reliability referring to the unnecessary unknowns, Such as the
additional parameters, is called internal reliability and that to
the necessary unknowns or their required functions called external
reliability.
This generalization of the reliability concept is rather natural.
No contradiction to the original meaning of reliability has been
made. The original reliability is just the reliability of a system.
While the measure to be proposed below iS a measure of the
reliability of a result. In the context gross and systematic
errors are both considered as the errors in the functional model
only, that is, as the biases of estimators.only.
3. POSTERIORI RELIABILITY IN THE CASE ADOPTING DATA-SNOOPING
Suppose that the following linear functional model and stochastical
model are true for observation vector 1 (nx1):
E(1)
Ax. Hs (1)
D(1)
g2% | (2)
where x is a vector of main unknowns (m,x1),
S is a vector of additional parameters (mox1),
A is a design matrix (nxm,) with the rank being m, ,
H is a coefficient matrix (nxm, , I^ £ n-m4).
Because estimation using the complete model (1) does not necessarily
lead to good estimators of the main unknowns in the sense of mean
Square error, some procedure of parameter selection is usually
carried out and the complete model is revised thereby for estimation
/^/. However biases are thus induced in the result of adjustment.
Suppose that the complete model (1) is revised by deleting
additional parameter 8; . Then the least Squares estimators of the
unknowns by the revised model are biased if the true value of sj
is not equal to zero. In fact, the deleted parameter S, takes zero
as its estimator. So the bias for s, is just the true value of S4
itself. This bias is passed to the estimators of other parameters.
Thus the biases for the parameterS are all relative to the true
value of S, . From section 2 we know evaluating those biases is
required to assess the reliability of the result.
- 368 -