4.

Inferences about a Mean Vector

4.1 The Plausibility of µ0 as a Value for a Normal
Population Mean

• The univariate theory for determining whether a specific value of µ0 is a

plausible value for the population mean µ.
– Test of the competing hypotheses

H 0 : µ = µ0 vs H1 : µ 6= µ0

Here H0 is the null hypothesis and H1 is the (two-side) alternative

hypothesis.

1
– Rejecting H0 when |t| is large is equivalent to rejecting H0 in favor of H1,
at significance level α if

2 (X̄ − µ0)2 2 −1 2
t = = n( X̄ − µ 0 )(s ) (X̄ − µ 0 ) > tn−1 (α/2)
s2/n

– If H0 is not rejected, we conclude that µ0 is a plausible value for the

normal population mean. From the well-known correspondence between
acceptance region for test of H0 : µ = µ0 versus H1 : µ 6= µ0 and
confidence interval for µ we have
 
x̄ − µ0
Do not reject H0 : µ = µ0 at level α or √ ≤ tn−1(α/2)
s/ n

is equivalent to
 
s
µ0 lies in the 100(1 − α)% confidence interval x̄ ± tn−1(α/2) √
n
2
• A natural generalization of the squared distance above is its multivariate
analog
 −1
1
T 2 = (X̂ − µ0)0 S (X̂ − µ0) = n(X̂ − µ0)0S−1(X̄ − µ0)
n

The statistics T 2 is called Hotelling’s T 2.

• If the observed statistical distance T 2 is too large— that is, if x̂ is “too far”
from µ0—the hypothesis H0 : µ = µ0 is rejected.

•
(n − 1)p
T 2 is distributed as Fp,n−p
n−p
where Fp,n−p denotes a random variable with an F-distribution with p and
n − p degree of freedom.

3
To summarize, we have the following:

Let X 1, X 2, . . . , X n be a random sample from an Np(µ, Σ) population. Then

n n
1 1
(X j − X̄)(X j − X̄)0,
P P
with X̄ = n X j and S = n−1
j=1 j=1

 
2 (n − 1)p
α = P T > Fp,n−p(α)
n−p
 
0 −1 (n − 1)p
= P n(X̄ − µ) S (X̄ − µ) > Fp,n−p(α)
n−p

whatever the true µ and Σ. Here Fp,n−p(α) is the upper (100α)th percentile
of the Fp,n−p distribution.

4
Example 4.1(Evaluating T 2) Let the data matrix for a random sample of size
n = 3 from a bivariate normal population be
 
6 9
X =  10 6 
8 3

Evaluate the observed T 2 for µ00 = [9, 5]. What is the sampling distribution of
T 2 in this case ?

Example 4.2 (Testing a multivariate mean vector with T 2) Perspiration

from 20 healthy females was analyzed. Three components, X1 =sweat rate,
X2 =sodium content, and X3 =potassium content, were measured, and the
results, which we call the sweat data, are presented in Table 5.1.

Test the hypothesis H0 : µ0 = [4, 50, 10] against H1 : µ0 6= [4, 50, 10] at
level of significance α = .10.

5
6
4.2 Confidence Regions and Simultaneous Comparisons of
Component Means

• The region R(X) is said to be a 100(1 − α)% confidence region if, before
the sample is selected,

P [(X) will cover the true θ] = 1 − α

This probability is calculated under the true, but unknown value of θ.

• The confidence region for the mean µ of a p-dimension normal population is

available. Before the sample is selected,
 
(n − 1)p
P n(X̄ − µ)0S−1(X̄ − µ) ≤ Fp,n−p(α) = 1 − α
(n − p)

whatever the values of the unknown µ and Σ. In words, X̄ will be within

[(n − 1)pFp,n−p(α)/(n − p)]1/2 of µ, with probability 1 − α, provided that
distance is defined in terms of nS−1. 7
• For P ≥ 4,we cannot graph the joint confidence region for µ. However we
can calculate the axes of the confidence ellipsoid and their relative lengths.

• These are determined from the eigenvalues λi and eigenvectors ei of S. The

direction and lengths of the axes of
0 −1 2 p(n − 1)
n(x̄ − µ) S (x̄ − µ) ≤ c = Fp,n−p(α)
n−p
are determined by going
√ √ p q
λc/ n = λi p(n − 1)Fp,n−p(α)/n(n − p)
units along the eigenvectors ei. Beginning at the center x̄, the axes of the
confidence ellipsoid are
s
p p(n − 1)
± λi Fp,n−p(α)ei where Sei = λiei, i = 1, 2, . . . , p
n(n − p)
The ratio of the λi’s will help identify relative amount of elongation along
pair of axes
8
Example 4.3 (Constructing a confidence ellipse for µ) Data for radiation
from microwave oven were introduced in Example 3.10 and 3.17. Let

1
x1 = (measured ration with door closed) 4

and
1
x2 = (measured ration with door open) 4

Construct the confidence ellipse for µ = [µ1, µ2].

9
 Simultaneous Confidence Statements

• While the confidence region n(x̄ − µ)0S−1(x̄ − µ) ≤ c2, for c a constant,

correctly assesses the joint knowledge concerning plausible value of µ, any
summary of conclusions ordinarily includes confidence statement about the
individual component means.

• In so doing, we adopt the attitude that all of the separate confidence

statements should hold simultaneously with specified high probability.

• It is the guarantee of a specified probability against any statement being

incorrect that motivates the term simultaneous confidence intervals

10
• Let X have an Np(µ, Σ) distribution and form the linear combination
Z = a1X1 + a2X2 + · · · + apXp = a0X
Hence
µZ = E(Z) = a0µ and σZ
2
= Var(Z) = a0Σa

• Moreover, by Result 3.2, Z has an N (a0µ, a0Σa) distribution. If a random

sample X 1, X 2, . . . , X n from the Np(µ, Σ) and Zj = a0X j , j = 1, 2, . . . , n.
Then the sample mean and variance of the observed values z1, z2, . . . , zn are
z̄ = a0x̄ and s2z = a0Sa
where x̂ and S are the sample mean vector and covariance matrix of the xj ’s
respectively.

2
• For a fixed and σZ unknown, a 100(1 − α)% confidence interval for µZ = a0µ
is based on student’s t-ratio
√
z̄ − µZ n(a0x̄ − a0µ)
t= √ = √
sZ / n a0Sa
11
• and leads to the statement
sz sz
z̄ − tn−1(α/2) √ ≤ µZ ≤ z̄ + tn−1(α/2) √
n n
√ √
or 0
a Sa a0Sa
0 0 0
a x̄ − tn−1(α/2) √ ≤ a µ ≤ a x̄ + tn−1(α/2) √
n n
where tn−1(α/2) is the upper 100(α/2)th percentile of a t-distribution with
n − 1 d.f.

• Clearly, we could make serval confidence statements about the components

of µ, each with associated confidence coefficient 1 − α, by choosing different
coefficient vector a. However, the confidence associated with all of the
statements taken together is not 1 − α.

• Intuitively, it would be desirable to associate a “collective” confidence

coefficient of 1 − α 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 confidence coefficient: intervals that are wider than the
interval for a specific choice of a. 12
• Given a data set x1, x2, . . . , xn and a particular a, the confidence interval is
that set of a0µ values for which
√
n(a0x̄ − a0µ)
|t| = √ ≤ tn−1(α/2)
0
a Sa

• A simultaneous confidence region is given by the set a0µ values such that t2
is relatively small for all choice of a. It seems reasonable to expect that the
constant t2(α/2) will be replaced by a large value c2, when statements are
developed for many choices of a.

• Considering the values of a for which t2 ≤ c2, we are naturally led to the
determination of

2 n(a0(x̄ − µ))2 0 −1 2
max t = max = n( x̄ − µ) S (x̄ − µ) = T
a a a0Sa

with the maximum occurring for a proportional to S−1(x̄ − µ).

13
Result 4.3. Let X 1, X 2, . . . , X n be a random sample from an Np(µ, Σ)
population with Σ positive definite. Then simultaneously for all a, the interval
 s s 
a0X̄ − p(n − 1) p(n − 1)
Fp,n−p(α)a0Sa, a0X̄ + Fp,n−p(α)a0Sa
n(n − p) n(n − p)

will contain a0µ with probability 1 − α.

Example 4.4 (Simultaneous confidence intervals as shadows of the

confidence ellipsoid In Example 4.3, we obtain the 95% confidence ellipse
for the means of the four roots of the door-closed and door-open microwave
radiation measurements. Obtain the 95% simultaneous T 2 intervals for the two
component means.

Example 4.5 (Constructing simultaneous confidence intervals and ellipse)

The scores obtained by n = 87 college students on the College Level Examination
Program (CLEP) subtest X1, and the College Qualification Test (CQT) subtests
X2 and X3 are given in Table 5.3 for X1 =social science and history, X2 =verbal,
and X3 =science. Construct simultaneous confidence intervals for µ1, µ2 and
µ3 . 14
15
16
 4.4 Large Sample Inference about a Population Mean
Vector
Result 4.4. Let X 1, X 2, . . . , X n be a random sample from a population with
mean µ and positive definite covariance matrix Σ. When n − p is large, the
hypothesis H0 : µ = µ0 is rejected in favor of H1 : µ 6= µ0, at a level of
significance approximately α, if the observed

n(x̄ − µ0)0S−1(x̄ − µ0) > χ2p(α)

Here χ2p(α) is the upper (100α)th percentile of a chi-square distribution with

p.d.f .

Result 4.5. Let X 1, X 2, . . . , X n be a random sample from a population with

mean µ and positive definite covariance matrix Σ. If n − p is large,
r
0
q
2
a0Sa
a X̄ ± χp(α)
n

will contain a0µ, for every a, with probability approximately 1 − α. 17

Example 4.6 (Constructing large sample simultaneous confidence
intervals) A music educator tested thousands of Finnish students on their
native musical ability in order to set national norms in Finland. Summary
statistics for part of the data set are given in Table 5.5. These statistics are
based on a sample of n = 96 Finnish 12th graders. Construct 90% simultaneous
confidence intervals for individual mean components µi, i = 1, 2, . . . , 7.

18
 4.5 Paired Comparisons
Paired Comparisons
In the single response (univariate) case, let Xj1 denote the response to treatment
1, and let Xj2 denote the response to treatment 2 for the jth trial. That is,
(Xj1, Xj2) are measurements recorded on the jth unit or jth pair of like units.
By design, the n differences

Dj = Xj1 − Xj2, j = 1, 2, . . . , n
Should reflect only the differences Dj represent independent observations from
an N (δ, σd2) distribution. the variable

D̄ − δ
t= √
sd / n
n n
1 2 1
(Dj − D̄)2 has a t-distribution with n − 1
P P
where D̄ = n Dj and sd = n−1
j=1 j=1
d.f.

19
• An α-level test of
H0 : δ = 0 vs H1 : δ 6= 0
may be conducted by comparing |t| with tn−1(α/2)-the upper 100(α/2)th
percentile of a t-distribution with n-1 d.f.

• A 100(1 − α)% confidence interval for the mean difference δ = E(Xj1 − Xj2)
is provided the statement

sd sd
D̄ − tn−1(α/2) √ ≤ δ ≤ D̄ + tn−1(α/2) √
n n

• Multivariate extension of the paired-comparison procedure to distinguish

between p response,two treatments, and n experimental units. The p paired-
difference random variables become
Dj1 = X1j1 − X2j1
Dj2 = X1j2 − X2j2
.. ..

Djp = X1jp − X2jp 20

Let D 0j = [Dj1, Dj2, . . . , Djp], and assume for j = 1, 2, . . . , n that

E(D j ) = δ = [δ1, δ2, · · · , δp]0 and Cov(D j ) = Σd

If, in addition, D 1, D 2, . . . , D n are independent Np(δ, Σd) random vectors,

inference about the vector of mean differences δ can based upon a T 2 statistics

T 2 = n(D̄ − δ)0S−1
d (D̄ − δ)

n n
1 1
(D j − D̄)(D j − D̄)0.
P P
where D̄ = n D j and Sd = n−1
j=1 j=1

Result 4.6 Let the differences D 1, D 2, . . . , D n be a random sample from

Np(δ, Σd) population. Then

T 2 = n(D̄ − δ)0S−1
d (D̄ − δ)

is distributed as an [(n − p)p/(n − p)Fp,n−p] random variable, whatever the true

δ and Σd.
21
 Given the observed differences D 0j = [Dj1, Dj2, . . . , Djp], j = 1, 2, . . . , n,
• an α-level test of
H0 : δ = 0 versus H1 : δ = 0
for an Np(δ, Σd) population rejects H0 if the observed
0 (n − p)p
2
T = nD̄ S−1
d D̄ > Fp,n−p(α)
n−p
where Fp,n−p(α) is the upper (100α)%th percentile of an F-distribution with
p and n − p d.f.

• A 100(1 − α)% confidence region for δ consists of all δ such that

(n − p)p
(D̄ − δ)0S−1
d (D̄ − δ) > Fp,n−p
n(n − p)
Also 100(1 − α)% simultaneous confidence intervals for the individual mean
differences δi are given by
s s
(n − 1)p s2Di
δi : D̄ i ± Fp,n−p(α)
(n − p) n
where D̄ i is the ith element of D̄ and s2Di is the ith diagonal element of Sd22.
Example 4.7 (Checking for a mean difference with paired observations)
Muicipal wasterwater treatment plants are required by law to monitor their
discharges into rivers and streams on a regular basis. Concern about the
reliability of data from one of these self-monitoring programs led to a study in
which samples of effluent were divided and sent to two laboratories for testing.
One-half was sent to a private commercial laboratory routinely used in the
monitoring program. Measurements of biochemical oxygen demand (BOD) and
suspended solid (SS) were obtained, for n = 11 sample splits, from the two
laboratories . The data are displayed in Table 6.1.
Do the two laboratories’s chemical analyses agree ? If differences exist, what
is their nature?

23
 4.6 Comparing Mean Vectors from Two Populations
Sample Summary statistics
(Population 1)
n1 n1
1 1
(x1j − x̄1)(x1j − x̄1)0
P P
x11, x12, . . . , x1n1 x̄1 = n1 x1j S1 = n1 −1
j=1 j=1
(Population 2)
n2 n2
1 1
(x2j − x̄2)(x2j − x̄2)0
P P
x21, x22, . . . , x2n12 x̄2 = n2 x2j S2 = n2 −1
j=1 j=1

Assumptions Concerning the Structure of the Data

1. The sample X 11, X 12, . . . , X 1n1 is a random sample of size n1 from a

p-variate population with mean µ1 and covariance matrix Σ1.

2. The sample X 21, X 22, . . . , X 2n2 is a random sample of size n1 from a

p-variate population with mean µ2 and covariance matrix Σ2.

3. Also, X 11, X 12, . . . , X 1n1 are independent of X 21, X 22, . . . , X 2n2 .

24
 H0 : µ1 = µ2 = 0 vs H1 : µ1 − µ2 6= 0
Further Assumption When n1 and n2 Are Small
1. Both populations are multivariate normal.
2. Also, Σ1 = Σ2 (same covariance matrix)

Set the estimate of Σ as

n1 n2
0
(x2j − x̄2)(x2j − x̄2)0
P P
(x1j − x̄1)(x1j − x̄1) +
j=1 j=1
Spooled =
n1 + n2 − 2
n1 − 1 n2 − 1
= S1 + S2
n1 + n2 − 2 n 1 + n2 − 2

Then
1 1 1 1
Cov(X̄ 1 − X̄ 2) = Cov(X 1) + Cov(X 2) = Σ + Σ = ( + )Σ
n1 n2 n1 n2
 
Hence 1 1
+ Σ
n1 n2
25
is an estimator of Cov(X̄ 1 − X̄ 2).
Result 4.7 If X 11, X 12, . . . , X 1n1 is a random sample of size n1 from
Np(µ1, Σ) and X 21, X 22, . . . , X 2n2 is an independent random sample size
n2 from Np(µ2, Σ), then
  −1
1 1
T 2 = [X̄ 1 − X̄ 2 − (µ1 − µ2)]0 + Spooled [X̄ 1 − X̄ 2 − (µ1 − µ2)]
n1 n2
is distributed as
(n1 + n2 − 2)p
Fp,n1+n2−p−1
n1 + n2 − p − 1
Consequently,
P (T 2 ≤ c2) = 1 − α
where
(n1 + n2 − 2)p
c2 = Fp,n1+n2−p−1(α).
n1 + n2 − p − 1

26
Example 4.8 (Constructing a confidence region for the difference of two
mean vectors) Fifty bars of soap are manufactured in each of two ways. Two
characteristics, X1 =lather and X2 =mildness, are measured. The summary
statistics for bars produced by method 1 and 2 are

x1 = [8.3 4.1]0, x2 = [10.2 3.9]0

   
2 1 2 1
S1 = , S2 =
1 6 1 4
Obtain a 95% confidence region for µ1 − µ2.

27
Simultaneous Confidence Intervals
Result 4.8 Let c2 = [(n1 + n2 − 2)p/(n1 + n2 − p − 1)]Fp,n1+n2−p−1(α). With
probability 1 − α.
s  
1 1
a0(x̄1 − x̄2) ± c a0 + Spooleda
n1 n 2

will cover a0(µ1 − µ2) for all a. In particular µ1i − µ2i will be covered by
s 
1 1
(X̄ 1i − X̄ 2i) ± c + sii,pooled for i = 1, 2, . . . , p
n1 n2

28
The Two-Sample Situation When Σ1 6= Σ2

Result 4.9 Let the sample sizes be such that n1 − p and n2 − p are large. Then,
an approximate 100(1 − α)% confidence ellipsoid for µ1 − µ2 is given by all
µ1 − µ2 satisfying
  −1
1 1
T 2 = [X̄ 1−X̄ 2−(µ1−µ2)]0 + Spooled [X̄ 1−X̄ 2−(µ1−µ2)] ≤ χ2p(α)
n1 n2
where χ2p(α) is the upper (100α)th percentile of a chi-square distribution
with p d.f. Also 100(1 − α)% simultaneous confidence interval for all linear
combinations a0(µ1 − µ2) are provided by
s  
q 1 1
a0(µ1 − µ2) belongs to a0(x̄1 − x̄2) ± χ2p(α) a0 S1 + S2 a.
n1 n2

29
 4.7 Testing for Equality of Covariance Matrics
With g populations, the null hypothesis is

H0 : Σ1 = Σ2 = · · · = Σg = Σ

where Σl is the covariance matrix for the lth population, l = 1, 2, . . . , g, and Σ

is the presumed common covariance matrix. The alternative hypothesis is that
at least two of the covariance matrices are not equal.

30
• Assuming multivariate normal populations, a likelihood ratio statistic for
testing above is given by

Y (nl−1)/2
|Sl|
Γ=
|Spooled|
l

Here nl is the sample size for the lth group, Sl is the lth group sample
covariance matrix and Spool is the pooled sample covariance matrix given by

1
Spool = P {(n1 − 1)S1 + (n2 − 1)S2 + · · · + (ng − 1)Sg }
(nl − 1)
l

• Box’s test is based on his χ2 approximation to the sampling distribution of

−2 ln Γ. Setting −2 ln Γ = M (Box’s M statistics) gives
" #
X X
M= (nl − 1) ln |Spooled| − [(nl − 1) ln |Sl].
31
l l
 Box’s Test for Equailty of Covariance Matrices
Set  
2
 
X 1 1  2p + 3p − 1
u= −P
(nl − 1) (nl − 1) 6(p + 1)(g − 1)

l
l
where p is the number of variables and g is the number of groups. Then
(" # )
X X
C = (1 − u)M = (1 − u) (nl − 1) ln |Spooled| − [(nl − 1) ln |Sl]
l l

has an approximate χ2 distribution with

1 1 1
ν = g p(p + 1) − p(p + 1) = p(p + 1)(g − 1)
2 2 2

degrees of freedom. At significance level α, reject H0 if C > χ2p(p+1)(g−1)/2(α).

32
Example 4.9 (Testing equality of covariance matrices—nursing homes)
The Wisconsin Department of Health and Social Services reimburse nursing
homes in the state for the services provided. The department develops a set of
formulas for the rates for each facility, based on factors such as level of care,
mean wage rate, and average wage rate in the state.

Nursing homes can be classified on the basis of ownership (private party,

nonprofit organization, and government) and certification (skilled nursing facility,
intermediate care facility, or combination of the two).

One purpose of a recent study was to investigate the effects of ownership

or certification (or both) on cost.s. Four costs, computed on a per-patient-day
basis and measured in hours per patient day, were selected for analysis X1 =
cost of nursing labor, X2 =cost dietary labor, X3 =cost of plant operation and
maintenance labor, and X4 =cost of housekeeping and laundry labor. A total of
n = 516 observations on each of the p = 4 cost variables were initially separated
according to ownership. Summary statistics for each of the g = 3 groups are
given in the following table.

33
 Group Number of observations Sample mean vectors
l = 1 (private) n1 = 271 x¯1 = [2.066 .480 .082 .360]0
l = 2 (nonprofit) n2 = 138 x¯2 = [2.167 .596 .124 .418]0
l = 3(government) n3 = 107 x¯3 = [2.273 .521 .125 .283]0

Sample covariance matrices

   
.291 .561
 −.001 .011   −.011 .025 
S1 = 
 ; S2 = 
 ;
.002 .000 .001  .001 .004 .005 
.010 .003 .000 .010 .037 .007 .002 .019
 
.261
 −.030 .017 
S3 =  ;
 .003 −.000 .004 
.018 .006 .001 .013
Assuming multivariate normal data, test hypothesis H0 : Σ1 = Σ2 = Σ3 = Σ.

34