0% found this document useful (0 votes)
10 views10 pagesHandout-4b-Simultaneous CI
Statistics multivariate methods Simultaneous CI
Uploaded by
kamwebazemarkkCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
10 views10 pagesHandout-4b-Simultaneous CI
Statistics multivariate methods Simultaneous CI
Uploaded by
kamwebazemarkkCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 10
220 Chapter S Inferences about a Mean Vector
Result 5.2. When the sample size n is large, under the null hypothesis Hy,
max L(0) 4
-2mA = — = -
nA = 2191 ax £(0) :
m8 .
is, approximately, a random variable, Here the degrees of freedom are v ~ y» :
~ = (dimension of @) — (dimension of @o). 2
Statistical tests are compared on the basis of their power, which is defined as the ~
curve or surface whose height is P[test rejects Hy|@], evaluated at each parameter :
vector @. Power measures the ability of a test to reject Hp when it is not true. In the
rare situation where @ = Opis completely specified under Hy and the alternative H, ;
consists of the single specified value 8 = @;, the likelihood ratio test has the highest
power among all tests with the same significance level a = P(test rejects Ho|@ =p). °
In many single-parameter cases (@ has one component), the likelihood ratio test is
uniformly most powerful against all alternatives to one side of Ho: @ = 6. In other
cases, this property holds approximately for large samples.
We shall not give the technical details required for discussing the optimal prop-
erties of likelihood ratio tests in the multivariate situation. The general import of
these properties, for our purposes, is that they have the highest possible (average)
power when the sample size is large.
5.4 Confidence Regions and Simultaneous Comparisons
of Component Means
To obtain our primary method for making inferences from a sample, we need to ex-
tend the concept of a univariate confidence interval to a multivariate confidence re-
gion. Let 8 be a vector of unknown population parameters and @ be the set of all
possible values of @. A confidence region is a region of likely 8 values. This region is
determined by the data, and for the moment, we shall denote it by R(X), where
X = |X), X,..., Xp]' is the data matrix.
‘The region R(X) is said to be a 100(1 ~ «)% confidence region if, before the
sample is selected,
P(R(X) will cover the true @] = 1 — a 6-17)
This probability is calculated under the true, but unknown, value of 8.
‘The confidence region for the mean # of a p-dimensional normal population is
available from (5-6). Before the sample is selected,
x tg ¢
~ p)'S(X - p) <
pfx BYS'Y(X-p) = (n
whatever the values of the unknown x and ¥. In words, X will be within
[0 ~ 1)pFpn-plad/(n ~ p) PP
of 1, with probability 1 — a, provided that distance is defined in terms of nS”.
For a particular sample, ¥ and $ can be computed, and the inequality Confidence Regions and Simultaneous Comparisons of Component Means 221
n(& — p)'S(X ~ w) = (n — 1)pFy_-p(a)/(n ~ p) will define a region R(X)
within the space of all possible parameter values. In this case, the region will be an
ellipsoid centered at x. This ellipsoid is the 100(1 — @)% confidence region for yz.
A 100(1 — @)% confidence region for the mean of a p-dimensional normal
distribution is the ellipsoid determined by all j such that
~1
n(X — pw)'SUR - pw) won, mC) (5-18)
eae ee : oF
where = aes ma %) (xj — ¥)! and xy,x2,..-4%, are
the sample observations,
To determine whether any jo lies within the confidence region (is a
plausible value for 4), we need to compute the generalized squared distance
n(X — pto)'S"(% — pro) and compare it with [p(n ~ 1)/(n — p)]Fp,q-p(a). If the
squared distance is larger than [p(m — 1)/(n ~ p)|Fp,a~p(@), #o is not in the confi-
dence region. Since this is analogous to testing Ho: # = po versus Hy: ft # po [see
(5-7)], we see that the confidence region of (5-18) consists of all sto vectors for which
the 7?-test would not reject Ho in favor of H; at significance level a.
For p = 4, we cannot graph the joint confidence region for 4. However, we can
calculate the axes of the confidence ellipsoid and their relative lengths. These are
determined from the eigenvalues A; and eigenvectors e; of S. As in (4-7), the direc-
tions and lengths of the axes of
¥- py'S1K - a ete)
MR ~ WYSMR = B) SP = To Fano
are determined by going
Viie/ Vai = VX V p(n — 1) Fon-pla)/n(n — P)
units along the eigenvectors e;. Beginning at the center i, the axes of the confidence
ellipsoid are
‘The ratios of the A,’s will help identify relative amounts of elongation along pairs
of axes.
Example 5.3 (Constructing a confidence ellipse for yz) Data for radiation from
microwave ovens were introduced in Examples 4.10 and 4.17. Let
x, = Wmeasured radiation with door closed
and
x2 = W measured radiation with door open 222 Chapter 5 Inferences about a Mean Vector
For the = 42 pairs of transformed observations, we find that
x= | 36 s =| om ou7
603 |” 0117 0146 |’
203.018 ~163391
163.391 200.228
= st=
The eigenvalue and eigenvector pairs for $ are
A = 026, e} = [.704, 710]
Ao = 002, e) = [-.710, 704]
‘The 95% confidence ellipse for 4 consists of all values (12, #1) satisfying
203.018 ~163. a [ - |
let ay etal [io 200.228 || 603 — 2
2(41)
=~] Fl.05)
or, since Fy,¢q(.05) = 323,
42(203.018) (.564 — 44)? + 42(200.228) (.603 — py)”
~ 84(163.391) (564 — 4) (603 — wr) < 6.62
To see whether 42’ = [.562, .589] is in the confidence region, we compute
42(203.018) (.564 — .562)? + 42(200.228) (.603 - .589)?
— 84(163.391) (564 — .562)(,603 — .589) = 1.30 = 6.62
We conclude that yx’ = [.562, 589] is in the region, Equivalently, a test of Hg:
62
39 at the a = .05 level
co = | Rould mot be rejected in favor of Fw # IF
of significance.
The joint confidence ellipsoid is plotted in Figure 5.1. The center is at
’ = [.564, 603), and the half-lengths of the major and minor axes are given by
VI OO Baap) = VI | ED (323) = 068
1) = 241) fe
Vig Fn p(a) = V 002 Bean (223) = O18
respectively. The axes lie along ej = [.704, .710] and e; = [—.710, .704] when these
vectors are plotted with % as the origin. An indication of the elongation of the confi-
dence ellipse is provided by the ratio of the lengths of the major and minor axes.
This ratio is
and Confidence Regions and Simultaneous Comparisons of Component Means 223
Figure 5.1 A 95% confidence
ellipse for 1 based on microwave-
x, radiation data,
‘The length of the major axis is 3.6 times the length of the minor axis. .
Simultaneous Confidence Statements
While the confidence region n(x — x)'S"\(x - ) < c?, for ca constant, correctly
assesses the joint knowledge concerning plausible values of 2, any summary of con-
clusions ordinarily includes confidence statements about the individual component
means. In so doing, we adopt the attitude that all of the separate confidence state-
ments should hold simultaneously with a specified high probability. It is the guaran-
tee of a specified probability against any statement being incorrect that motivates
the term simultaneous confidence intervals. We begin by considering simultaneous
confidence statements which are intimately related to the joint confidence region
based on the 7*-statistic.
Let X have an N,(#, £) distribution and form the linear combination
Z = aX, + aX) +--+ a,X, = aX
From (2-43),
wz = E(Z)=a'p
and
o} = Var(Z) = a'ka
Moreover, by Result 4.2, Z has an N(a’,a’Za) distribution. If a random sample
X1,Xz,...,Xq from the Np(4, 2) population is available, a corresponding sample
of Z’s can be created by taking linear combinations. Thus,
Zj = a, Xp + apXjz +--+ apXjp = a'X;
‘The sample mean and variance of the observed values 21, Z2,..., Z, are, by (3-36), 224 Chapter 5 Inferences about a Mean Vector
and
si = a'Sa
where ¥ and S are the sample mean vector and covariance matrix of the x/s,
respectively.
Simultaneous confidence intervals can be developed from a consideration of con. ~
fidence intervals for a’ w for various choices of a. The argument proceeds as follows.
Fora fixed and o% unknown, a 100(1 ~ a)% confidence interval for uz = a’ |
is based on student's (-ratio .
uz Va(a'k-a’'p)
6-20) ©
IVa
and leads to the statement
= ln (a2) S pz S 2 + t,-1(0/2)
or
aK — ty-1(@/2) “7 S 8m = aR + t-3(a/2) vase (5-21)
where f,_1(a/2) is the upper 100(a/2)th percentile of a r-distribution with n ~ 1 df.
Inequality (5-21) can be interpreted asa statement about the components of the
‘mean vector #2. For example, with a’ = [1,0,...,0},a’s = wx, and (5-21) becomes
the usual confidence interval for a normal population mean. (Note, in this case, that
a’Sa = 51,.) Clearly, we could make several confidence statements about the com-
ponents of 2, each with associated confidence coefficient 1 ~ a, by choosing differ-
ent coefficient vectors a. However, the confidence associated with all of the
statements taken together is not 1 — a.
Intuitively, it would be desirable to associate a “collective” confidence coeffi-
cient of 1 — a With the confidence intervals that can be generated by all choices of
a. However, a price must be paid for the convenience of a large simultaneous confi-
dence coefficient: intervals that are wider (less precise) than the interval of (5-21)
for a specific choice of a.
Given a data set X,%2,...,%, and @ particular a, the confidence interval in
(5-21)is that setof a’ values for w!
or, equivalently,
= A(a/2) (622)
A simultaneous confidence region is given by the set of a’ » values such that ¢” is rel-
atively small for all choices of a. It seems reasonable to expect that the constant
#2. 4(a/2) in (5-22) will be replaced by a larger value,c”, when statements are devel-
oped for many choices of a. Confidence Regions and Simultaneous Comparisons of Component Means 225
Considering the values of a for which 1? = c?, we are naturally led to the deter-
mination of
z 2
n(a’(x — #))
as
max (? = max
. . Sa
Using the maximization lemma (2-50) with x = = ), and B = S, we get
MOT. Lm EERO) epee w= 7 629
2 a'Sa
with the maximum occurring for a proportional to S'(x — p).
Result 5.3. Let X,, X2,...,X,, be a random sample from an N,(4, %) population
with positive definite. Then, simultaneously for all a, the interval
x [pea . x. |p) "
(: Ro Voie py Freelada’sa, aK + Fano) s:)
will contain a‘ with probability 1 —
Proof. From (5-23),
T? = n( — w)'S"(% — w) = C2 implies
i iz Sa
Salp < az tc)
for every a. Choosing c? = p(n — 1)Fp.n-p(a)/( — p) {see (5-6)] gives intervals
that will contain a’ p for all a, with probability 1 — a = P[T? = ¢"). .
for every a, or
It is convenient to refer to the simultaneous intervals of, Result 5.3 as
T?-intervals, since the coverage probability is determined by the distribution of T?.
The successive choices a’ = [1,0,...,0], a’ = (0,1,...,0], and so on through
a’ = [0,0,..., 1] for the 7?-intervals allow us to conclude that
5 FO pasta fe ay 38, + [LO pega) f
5, -./p@=) S22 . p(n ~ 1) (saa
ao \Goo) ie) [Esme se We =p Fool) 2 a
, . pea D oe =1) 5
Bp apy Fool) (
