aC-
vue
Hay leads, to
~ 2 -. ^ 2
Ele, /Han) = Ble [WJ + Won (15a)
2 X 2
Yoke = Sef (15b)
Ap depends on the type I error size o, on the type II error size B(Ap)
(B(Xp) * 1-P(II), P(II) = probability of accepting a false hypothesis) and on the
degrees of freedom r, =, For a given a ,8(Ap); r we can find Ap by using Baarda's
nomograms in /3/, pp. 21-23.
If HA is true, then the estimators x and 6 are no more unbiased. If the effect
of a gross error vector vl, in the observations upon the residual vector v has to
p
be computed then we obtain
E(v/HR 197-9 EOv2[/030€ veu 1 (16)
p
vv. 7 71
p yP Vlp ° (17)
V
Although in practical projects we normally don't know anything about the vector
ep?- it indicates the relationship of the gross errors - equation (17) enables
us to study the effect of an a-priori supposed gross error vector vi, upon the
residual vector v. From (14), (15), (17) and from the idempotency of the matrix
product Qn9r it follows for small gross errors
An 7 N e
I
p pp 5 Np = cpPQ, ,PCp . (18)
The vector vl; contains all assumed gross errors of the pth alternative hypothesis.
If we suppose only one gross error in the observation i we get
Tr
>
+
«1
-—À
Pt
~~
—
—
"
e cl" ; (19)
and with c1 s (0,...0,1,0,...,0)
Hay: MC e 827) : (20)
combining (18) and (20) we obtain
for diegonal P:
m. d
Hay: Vl O6 N1 , Ni = pit quivi ; (21)
qyiv; = 1th diagonal
element of Qyy
Te pr) is the minimal gross error which can be detected with the probaility
s{x17). Thus for a constant 34 for all possible i vr, (1) represents a suitable in-
dividual reliability indicator ("measure for internal reliability" in Baarda /3/).
Furthermore, Baarda /3/ has developed a test criterion for testing the residuals
of an adjustment ("data-snooping"). If Ho is rejected, we get
T
-CpPv
= : 22
"o^ s Np (22)
With one of the possible alternative hypotheses (615. Cp) s i from (14) and by
using a weight matrix of diagonal form we obtain