Confidence Regions

Simultaneous Confidence Statements

Large Sample Inferences

Inferences about a mean vector - first part

Hauptseminar
"Multivariate Statistics and Copulas"

Patrick Follmann

29.10.2009

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

Overview

1 Confidence Regions
Univariate Case
Multivariate Case

2 Simultaneous Confidence Statements

One-at-a-Time Intervals
T 2 -intervals
Bonferroni Intervals

3 Large Sample Inferences

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

Motivation

x2

x1

Source: [Link]

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

Data set
n = 45 female birds
Values of interest: x1 =wing length and x2 = tail length (in
mm)
=> dimension p = 2
Bird data
x1 x2
(Tail (Wing
length) length)
191 284
198 300
188 280
210 283
179 257
.. ..
. .
Source: Data courtesy of S. Temple.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

Tail length: min x1 = 173 , max x1 = 216 , x 1 = 193.6222

Wing length: min x2 = 245 , max x2 = 310 , x 2 = 279.7778

Correlation: ρ = 0.7711697

310
210

300
290
200

280
190

270
260
180

250

wing length tail length

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

Task

Find the 95% confidence ellipse for the population means µ1 and µ2
Suppose it is known that µ1 = 190mm and µ2 = 275mm for male
birds. Are these plausible values for the female birds?
Construct the 95%T 2 - and Bonferroni intervals for µ1 and µ2 .
Is the bivariate normal distribution a viable population model?
Explain with reference to Q-Q plots and a scatter diagram

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Confidence Regions - General

Let θ be a vector of unknown population parameters and Θ be

the set of all possible values of θ.
A confidence region is a region of likely θ values,
denoted by R(X), where X = [X1 , X2 , . . . , Xn ]0 is the data matrix.
Note: Before the data is collected R(X) is a random region!

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Confidence Regions - General

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

sample is selected,

P[R(X) will cover the true θ] = 1 − α

calculated under the true, but unknown, value of θ.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Univariate Case

Is the specific value µ0 plausible for the population mean µ?

Hypothesis test:

H0 : µ = µ0 and H1 : µ 6= µ0

X1 , X2 , ..., Xn random sample from a normal population.

Appropriate test statistic:
(X̄ −µ
√ 0) ,
t= s/ n
t ∼ tn−1

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Do not reject H0 at significance level α if

 
|t| =  x̄s/−µ
√ 0  ≤ tn−1 (α/2) or
 
n

µ0 lies in the 100(1 − α)% confidence interval:

h i
µ0 ∈ x̄ − tn−1 (α/2) √sn ; x̄ + tn−1 (α/2) √sn

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Note: Rejecting H0 when |t| > tn−1 (α/2) is equivalent to

rejecting it when

(x̄ − µ0 )2
t2 = = n(x̄ − µ0 )(s 2 )−1 (x̄ − µ0 ) > tn−1
2
(α/2)
s 2 /n
where t 2 is the square of the distance from the sample mean x̄ to
the test value µ0 , expressed in terms of estimated standard
deviations of x̄ .

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Excercise

R Code:

x1 <- c(191,...,188)
x2 <- c(284,...,258)
n <- 45
p <- 2
m1 <- mean(x1)
m2 <- mean(x2)
s1 <- sd(x1)
s2 <- sd(x2)

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

R Code:

t <- qt(0.025, n-1,[Link] = FALSE)

(u1 <- m1 - t * s1 / sqrt(n))

[1] 190.3216
(o1 <- m1 + t * s1 / sqrt(n))
[1] 196.9228
(u2 <- m2 - t * s2 / sqrt(n))
[1] 275.4392
(o2 <- m2 + t * s2 / sqrt(n))
[1] 284.1163

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

So we have:

Do not reject H01 : µ01 = µ1 at level 0.05 if

µ01 ∈ [190.32; 196.92]

and do not reject H02 : µ20 = µ2 at level 0.05 if

µ02 ∈ [275.44; 284.17]

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Exercise Result Univariate Case

Here:

µ01 = 190mm ∈
6 [190.32mm; 196.92mm]
µ02 = 275mm ∈6 [275.44mm; 284.17mm]

So we have to reject both H01 and H02 at level α = 0.05, i.e.

µ01 = 190mm and µ02 = 275mm are no plausible values for the
mean tail and wing length for the female birds if we look at them
one at a time.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Multivariate Case

Is a given p × 1 vector µ0 a plausible value for the mean of a

multivariate normal population?

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Generalization of the squared distance:

1
T 2 = (X − µ0 )0 ( S)−1 (X − µ0 ) = n(X − µ0 )0 S−1 (X − µ0 )
n
where
 
µ10
n n µ20 
1X 1 X
X = Xj , S = (Xj −X)(Xj −X)0 , and µ0 =  . 
 
(p×1) n (p×p) n−1 (p×1)  .. 
j=1 j=1
µp0

Note: The statistic T 2 is called Hotelling’s T 2 .

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

(n−1)p
One can show: T 2 ∼ (n−p) Fp,n−p

Reject H0 : µ = µ0 against H1 : µ 6= µ0 at significance level α if

(n − 1)p
T 2 = n(x − µ0 )0 S−1 (x − µ0 ) > Fp,n−p (α)
(n − p)
Note: Here Fp,n−p (α) denotes the upper (100α)th percentile of the Fp,n−p

distribution.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Equivalently before the sample is selected:

(n − 1)p
 
0 −1
P n(X − µ) S (X − µ) 6 Fp,n−p (α) = 1 − α
(n − p)

independent of the unknown µ and Σ.

That means X will be within
1
(n − 1)p

2
Fp,n−p (α)
(n − p)

of µ, with probability 1 − α.
Distance is defined in terms of nS−1 .

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Multivariate Case - Confidence Region

A 100(1 − α)% confidence region for the mean of a p-dimensional

normal distribution is the ellipsoid determined by all µ such that

p(n − 1)
n(x − µ)0 S−1 (x − µ) 6 Fp,n−p (α)
(n − p)

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Multivariate Case - Exercise

R Code:
m <- c(m1,m2)
v1 <- var(x1)
v2 <- var(x2)
c <- cov(x1,x2)
S <- matrix(c(v1, c, c, v2),nrow=2,ncol=2)
Sinv <- solve(S)
mu0 <- c(190,275)

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

To see whether µ0 is in the confidence region we calculate

n(x − µ0 )0 S−1 (x − µ0 ) and compare it with p(n−1)
(n−p) Fp,n−p (α).
R Code:
n*(m-mu0) %*% Sinv %*% (m-mu0)
[, 1]
[1, ] 5.54313
(F <- (n-1)*p/(n-p)*qf(0.95,p,n))
[1] 6.578471
p(n−1)
We see: n(x − µ0 )0 S−1 (x − µ0 ) 6 (n−p) Fp,n−p (α).
=> µ0 lies inside the confidence region.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Graphical solution using R (package ellipse needed):

>e <- ellipse(S/n,centre = m,t = F)
>plot(e,type=’l’,main=’confidence
ellipse’,xlab=’tail length in mm’,ylab=’wing
lengths in mm’,col="red")
>points(mu0[1],mu0[2], col = "blue")
>text(190,272.5, labels = expression(mu0), col =
"blue")

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

confidence ellipse

284
282
wing lengths in mm

280
278
276

mu0
274

190 192 194 196 198

tail length in mm

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Univariate Case
Simultaneous Confidence Statements
Multivariate Case
Large Sample Inferences

Excercise Result

We can conclude that the mean tail- and wing length of the male
birds are plausible values for the mean tail- and wing length of the
female birds as well.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Simultaneous Confidence Statements

On the one hand:

(n−1)p
With T 2 > (n−p) Fp,n−p (α) we reject
   
µ1 µ10
µ2  µ20 
   
H0 : µ =  .  =  . 
  
 = µ0 ,
 ..   .. 
µp µp0

but we don’t know which component means differ too much!

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

On the other hand:

(n−1)p
With T 2 6 (n−p) Fp,n−p (α) we do not reject
   
µ1 µ10
µ2  µ20 
   
H0 : µ = 
 ..   ..  = µ0 ,
 =  
 .   . 
µp µp0

at level α.
I.e. all individual confidence statements should hold simultaneously
with probability 1 − α.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Let X ∼ Np (µ, Σ) and

Z = a1 X1 + a2 X2 + · · · + ap Xp = a0 X

Then: µZ = E (Z ) = a0 µ and σZ2 = Var (Z ) = a0 Σa

Moreover: Z ∼ N(a0 µ, a0 Σa) and

z = a0 x , sz2 = a0 Sa

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

For a fixed and σZ2 unknown, a 100(1 − α)% confidence interval for
µZ = a0 µ is based on student’s t-ratio (compare: univariate case):
√ 0
z − µZ n(a x − a0 µ)
t= √ = √ ∼ tn−1
sz / n a0 Sa

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

And we have:
sz sz
z − tn−1 (α/2) √ 6 µZ 6 z + tn−1 (α/2) √
n n

or equivalent:
√ √
0 a0 Sa 0 0 a0 Sa
a x − tn−1 (α/2) √ 6 a µ 6 a x + tn−1 (α/2) √
n n

For fixed a, the last inequality becomes the usual confidence

interval for a normal population mean.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

By considering confidence intervals for a0 µ for various choices of a

we develop the simultaneous confidence intervals:

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

One-at-a-Time Intervals

One approach to the construction of simultaneous confidence

intervals is to consider the components µi one at a time.

Therefore in the last inequality set a0 = [0, . . . , 0, ai , 0, . . . , 0]

where ai = 1.

Note: This approach ignores the covariance structure of the p

variables.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

This leads to the intervals

r r
s11 s11
x 1 − tn−1 (α/2) 6 µ1 6 x 1 + tn−1 (α/2)
n n
r r
s22 s22
x 2 − tn−1 (α/2) 6 µ2 6 x 2 + tn−1 (α/2)
n n
.. .. ..
.r . . r
spp spp
x p − tn−1 (α/2) 6 µp 6 x p + tn−1 (α/2)
n n

where the ith interval has probability 1 − α of covering µi .

Note: Simultaneous intervals of this type are also called t-intervals.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Problem: The probability of all intervals containing their respective

µi ’s simultaneously is not 1 − α!

For example consider the case where the p variables are

independent. Then:
P[all t-intervals contain the µi ’s] = (1 − α)(1 − α) · · · (1 − α) = (1 − α)p

Due to (1 − α) < 1, we have (1 − α)p < (1 − α).

That means to get a simultaneous confidence statement with

probability 1 − α the individual confidence intervals have to be
wider!

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

T 2 -intervals

Given observations x1 , x2 , . . . , xn and a fixed a, the confidence

interval is that set of a0 µ values for which
√ 
 n(a0 x − a0 µ) 
|t| = 

√  6 tn−1 (α/2)

 a0 Sa 

or equivalently,

n(a0 x − a0 µ)2 n(a0 (x − µ))2

t2 = = 2
6 tn−1 (α/2)
a0 Sa a0 Sa

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

A simultaneous confidence region is given by the set of a0 µ values

such that t 2 is relatively small for all choices of a. As we consider
statements for many choices of a, naturally the constant tn−1 2 will
be replaced by a larger value, c . 2

This leads to the determination of

n(a0 (x − µ))2
max t 2 = max
a a a0 Sa

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Using the maximation lemma (see cpt. 2 of [1]) we get

n(a0 (x − µ))2 (a0 (x − µ))2

 
max = n max = n(x − µ)0 S−1 (x − µ) = T 2
a a0 Sa a a0 Sa

Note: Here the maximum occurs for a proportional to S−1 (x − µ).

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

So we have:
n(a0 (x − µ))2
T 2 6 c 2 implies t 2 = 6 c2
a0 Sa
for every a, or
√ √
0 a0 Sa 0 0 a0 Sa
ax−c √ 6aµ6ax+c √
n n

Choosing c 2 = p(n−1) 0
(n−p) Fp,n−p (α) leads to intervals that contain a µ
for all a with probability 1 − α = P[T 2 6 c 2 ].

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Result:

Let X1 , X2 , . . . , Xn be a random sample from an Np (µ, Σ)

population (Σ pos. def.). Then, simultaneously for all a, the
interval
s s !
0 p(n − 1) p(n − 1)
ax − Fp,n−p (α)a0 Sa , a0 x + Fp,n−p (α)a0 Sa
n(n − p) n(n − p)

will contain a0 µ with probability 1 − α.

Note: Simultaneous intervals of this type are also called T 2 -intervals.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Choosing a0 = [1, 0, . . . , 0], a0 = [0, 1, 0, . . . , 0], and so on through

a0 = [0, . . . , 0, 1] successively allows us to conclude that
s r s r
p(n − 1) s11 p(n − 1) s11
x1 − Fp,n−p (α) 6 µ1 6 x1 + Fp,n−p (α)
(n − p) n (n − p) n
s r s r
p(n − 1) s22 p(n − 1) s22
x2 − Fp,n−p (α) 6 µ2 6 x2 + Fp,n−p (α)
(n − p) n (n − p) n
.. .. ..
s . . s.
r r
p(n − 1) spp p(n − 1) spp
xp − Fp,n−p (α) 6 µp 6 xp + Fp,n−p (α)
(n − p) n (n − p) n

all hold simultaneously with confidence coefficient 1 − α.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Excercise - T 2 -intervals

R Code:
F <- (n-1)*p/(n-p)*qf(0.05,2,43,[Link]=FALSE)

(U1 <- m1 - sqrt(F*v1/n))

[1] 189.4217
(O1 <- m1 + sqrt(F*v1/n))
[1] 197.8227
(U2 <- m2 - sqrt(F*v2/n))
[1] 274.2564
(O2 <- m2 + sqrt(F*v2/n))
[1] 285.2992

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

confidence ellipse with T^2-intervals

284
282
wing lengths in mm

280
278
276

mu0
274

190 192 194 196 198

tail length in mm

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Note:

The simultaneous T 2 confidence intervals for the individual

components of a mean vector (see last slide) are just the shadows,
or projections, of the confidence ellipsoid on the component axes.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Bonferroni Intervals

Suppose that, before we collect the data, we are interested in

confidence statements about m linear combinations
a01 µ, a02 µ, . . . , a0m µ.

Let Ci denote a confidence statement about the value of a0i µ with

P[Ci true] = 1 − αi , i = 1, 2, . . . , m.
Then we have:

P[all Ci true] = 1 − P[at least one Ci false]

m
X m
X
> 1− P(Ci false) = 1 − (1 − P(Ci true))
i=1 i=1
= 1 − (α1 + α2 + · · · + αm )

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

We can now control the overall error rate α1 + α2 + · · · + αm .

For example we can develop simultaneous interval estimates for the

components µi of µ. If we set αi = α/m for all i, we get the
t-intervals
r
α sii

x i ± tn−1 i = 1, 2, . . . , m
2m n
With the result from the last slide we have:
  α  rs  α α α
ii
P X i ± tn−1 contains µi , all i > 1− + + ··· +
2m n m m m
= 1−α

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Hence with a probability greater than or equal to 1 − α the

following m = p inequalities hold:
 r  r
α s11 α s11
x 1 − tn−1 6 µ1 6 x 1 + tn−1
2p n 2p n
 r  r
α s22 α s22
x 2 − tn−1 6 µ2 6 x 2 + tn−1
2p n 2p n
.. .. ..
. . .
 r  r
α spp α spp
x p − tn−1 6 µp 6 x p + tn−1
2p n 2p n

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Excercise - Bonferroni intervals

R Code:
(B11 <- m1 - tb*sqrt(v1/n))
[1] 189.8216
(B12 <- m1 + tb*sqrt(v1/n))
[1] 197.4229
(B21 <- m2 - tb*sqrt(v2/n))
[1] 274.7819
(B22 <- m2 + tb*sqrt(v2/n))
[1] 284.7736

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

confidence ellipse with T^2 & Bonferroni-intervals

284
282
wing lengths in mm

280
278
276

mu0
274

190 192 194 196 198

tail length in mm

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

T 2 vs. Bonferroni

In general both intervals have the same structure:

s
a0 Sa
a0 X ± (critical value)
n

Consequently, with αi = α/m,

Length of Bonferroni interval tn−1 (α/2m)

2
=r
Length ofT −interval p(n−1)
(n−p) Fp,n−p (α)

I.e. for small m, the Bonferroni intervals are shorter, i.e. more
precise, than the T 2 -intervals.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Advantage of the T 2 -intervals over the Bonferroni intervals:

Independence of m.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Q-Q-plots
Q-Q-plot x1 Q-Q-plot x2

310
210
Sample Quantiles

Sample Quantiles

290
200
190

270
180

250
-2 -1 0 1 2 -2 -1 0 1 2

Theoretical Quantiles Theoretical Quantiles

Q-Q-plot unique(x1) Q-Q-plot unique(x2)

310
210
Sample Quantiles

Sample Quantiles

290
200
190

270
180

250

-2 -1 0 1 2 -2 -1 0 1 2

Theoretical Quantiles Theoretical Quantiles

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions One-at-a-Time Intervals
Simultaneous Confidence Statements T 2 -intervals
Large Sample Inferences Bonferroni Intervals

Scatterplot

Scatterplot

310
300
290
280
x2

270
260
250

180 190 200 210

x1

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

Large Sample Inferences

Large sample => Assumption of a normal population not needed.

We then have:

(X − µ)0 (n−1 S)−1 (X − µ) = n(X − µ)0 S−1 (X − µ)

is approximately χ2p , i.e.

P[n(X − µ)0 S−1 (X − µ) 6 χ2p (α)] = 1 − α

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

q q
a0 Sa
T 2 -intervals: a0 X ± χ2p (α) n
q
a0 Sa
Bonferroni: a0 X ± z( 2m
α
) n
q
a0 Sa
One at a time: X i ± z( α2 ) n

Here z( α2 ) denotes the upper 100(α/2)th percentile of the standard

normal distribution.

Patrick Follmann Inferences about a mean vector - first part

 Confidence Regions
Simultaneous Confidence Statements
Large Sample Inferences

References

[1] Johnson, R. A., Wichern, D. W. Applied Multivariate Statistical

Analysis, 6th ed., New Jersey: Pearson Prentice Hall, 2007

Patrick Follmann Inferences about a mean vector - first part