test
V
The power g(r) is an increasing function of A, so A should be as large as possible.
A is maximal if N,, = 0 (Grayhill /9/), i.e. if dz is orthogonal to dx, dx", dt
Thus a statistical advantage of the orthogonal additional parameter concept be-
comes evident: It increases the power of this test.
Beside this it is clear that the power of a test depends mainly on the alternative
hypothesis and on the type I error size o.
If He is rejected, i.e. Af a significant systematic influence is apparent, the
detection of individual significant components becomes necessary.
Therefore the set of hypotheses
wii . dz, = dz, ; (A= 10505. 0p) (39)
Oo
is commonly used to test the individual parameters on significance.
Hence we get the test criterions
T "e —— : dziz4 7 variance of the ith (40)
99 (0212 additional parameter
Under pit) the TU). values are distributed as Student's t with r degrees of
freedom. Thus the more-dimensional T-test (31) is reduced to an one-dimensional
t-test.
In the case of independence of the individual parameters the type I error size a
of the individual test is related to the type I error size o of the global test
as
]el oi ei (411 e 8) ; (41)
œ =
"Ie
The main problem in testing subsets or even single parameters estimated in multi-
dimensional models arises with the dependence of these subsets (single parameters)
on the other model parameters. In the case of significant correlations the pro-
bability a = prse, | HO) of rejection of a true null-hypothesis Hit).
dz; = dz; of a single event (parameter) is no further independent. It should be
noticed that the application of the one-dimensional t-test with the usual limits
and confidence intervals leads the more to wrong decisions the more the correla-
tions do increase. Thus we have another important argument for using the concept
of orthogonal additional parameter sets.
Whereas the orthogonal concepts are regarded as very useful, the practical geo-
metrical conditions however often do not provide for sufficient orthogonality.
Then we have to set up simultaneous confidence intervals for the single events.
This can be done by the a-posteriori orthogonalization of the additional parame-
ter set, i.e. the transformation of the additional parameter vector (or, .if
necessary, even the vector of all unknowns of the bundle system (1)) into ortho-
gonal components, which can be tested independently (see Roy, Bose /19/, Pelzer
/16/).
Sometimes high correlations do appear only within a subset of the additional pa-
rameters. Then these parameters can be tested together on common significance.
The corresponding test criterion may analogously be derived from the general li-