Technical University of Denmark Page 1 of 45 pages.

Written examination: 17. December 2017

Course name and number: Introduction to Statistics (02402)

Aids and facilities allowed: All

The questions were answered by

(student number) (signature) (table number)

There are 30 questions of the ”multiple choice” type included in this exam divided on 18
exercises. To answer the questions you need to fill in the prepared 30-question multiple choice
form (on 6 seperate pages) in CampusNet.

5 points are given for a correct answer and −1 point is given for a wrong answer. ONLY the
following 5 answer options are valid: 1, 2, 3, 4 or 5. If a question is left blank or another answer
is given, then it does not count (i.e. ”0 points”). Hence, if more than one answer option is
given to a single question, which in fact is technically possible in the online system, it will not
count (i.e. ”0 points”). The number of points corresponding to specific marks or needed to
pass the examination is ultimately determined during censoring.

The final answers should be given in the exam module in CampusNet. The
table sheet here is ONLY to be used as an ”emergency” alternative (remember
to provide your study number if you hand in the sheet).

Exercise I.1 II.1 II.2 III.1 III.2 IV.1 IV.2 V.1 V.2 V.3
Question (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Answer
5 2 4 5 3 3 4 5 3 3

Exercise VI.1 VI.2 VII.1 VIII.1 VIII.2 IX.1 IX.2 IX.3 X.1 XI.1
Question (11) (12) (13) (14) (15) (16) (17) (18) (19) (20)
Answer
3 2 5 2 1 2 2 2 5 4

Exercise XII.1 XIII.1 XIII.2 XIV.1 XV.1 XVI.1 XVI.2 XVII.1 XVII.2 XVIII.1
Question (21) (22) (23) (24) (25) (26) (27) (28) (29) (30)
Answer
4 1 2 3 5 4 2 1 5 5

The questionnaire contains 45 pages.

Continues on page 2

1
 Multiple choice questions: Note that not all the suggested answers are necessarily mean-
ingful. In fact, some of them are very wrong but under all circumstances there is one and only
one correct answer to each question.

Exercise I

We consider pairwise measurements of 2 stochastic variables, X og Y . Both variables can be

assumed normally distributed.

x 38 35 47 38 42 41 48 35
y 25 21 26 23 28 27 29 18

Data can be loaded into R using the following command:

x <- c(38, 35, 47, 38, 42, 41, 48, 35)

y <- c(25, 21, 26, 23, 28, 27, 29, 18)

Question I.1 (1)

Provide an estimate for the correlation coefficient, ρ, between X and Y :

1  ρ̂ = 0.12

2  ρ̂ = 0.22

3  ρ̂ = 0.64

4  ρ̂ = 0.73

5*  ρ̂ = 0.82

----------------------------------- FACIT-BEGIN -----------------------------------

See Definition 1.19. The easiest way to do this is to use R. We copy into R to read in the data
two vectors

x <- c(38, 35, 47, 38, 42, 41, 48, 35)

y <- c(25, 21, 26, 23, 28, 27, 29, 18)

and the we calculate the estimate of the correlation as the sample correlation

2
cor(x,y)

## [1] 0.8237548

------------------------------------ FACIT-END ------------------------------------

Continues on page 4

3
 Exercise II

A biologist is evaluating the effect of 3 different diets on weight change in mice. In the experi-
ment 4 different strains (genetically different types) of mice are included, as strain is expected
to have an influence on weight change. Thus, 4 different strains of mice are exposed to 3 dif-
ferent diets, i.e. a total of 12 mice are included. The weight change is measured after 5 weeks
for each diet. The weight change is denoted Yij (in grams). The weight change can be assumed
normally distributed and thus the following model has been applied

Yij = µ + αi + βj + εij .

In this model αi denote the effect of diet i (i = 1, 2, 3) and βj denotes the effect of mouse
strain j (j = 1, 2, 3, 4). µ is the overall mean and εij are the errors, assumed independent and
normally distributed with mean 0 and constant standard deviation σε .

Question II.1 (2)

State the critical value when you want to test whether the mean weight change is the same for
the 3 diets and the significance level is α = 0.05.

1  12.20

2*  5.14

3  1.96

4  3.81

5  4.35

----------------------------------- FACIT-BEGIN -----------------------------------

The test we need to carry out is the F -test for a two-way ANOVA, in this case for the effect
of diet which is the treatment following the book (Theorem 8.22). Hence, we have to find the
two degrees of freedom and look up the 1 − α quantile. The degrees of freedom are:

• df1 = k − 1 = 2, since the number of levels for the treatment (number of diets) k = 3

• and df2 = (k − 1)(l − 1) = 2 · 3 = 6

The answer is then found by:

qf(p=0.95, df1=2, df2=6)

## [1] 5.143253

4
 Note, that the results doesn’t change if blocks and treatments are switched, i.e. if the diets are
thought of as blocks and strains as treatments.

------------------------------------ FACIT-END ------------------------------------

Question II.2 (3)

Assume that we have estimated the model parameters using R and concluded that both diet
and type of mouse strain are statistically significant. Also assume that the model residuals,
ε̂ij , are stored in the vector resi, and that we will use R to further analyze these. Which of
the following claims is not correct?

1  The command qqnorm(resi) plots normal scores for ε̂ij

2  The command plot(ecdf(resi)) plots the cumulative distribution for ε̂ij

3  The command sum(resi*resi)/6 gives an estimate of the variance of εij

4*  The command qnorm(resi) gives a test for normality of εij

5  The command sum(resi)/length(resi) gives an estimate for the mean of εij

----------------------------------- FACIT-BEGIN -----------------------------------

Lets go through the answers one by one:

1. TRUE statement. The residuals are sorted and, where i ∈ (1, 2, . . . , n) denotes the i’th
element in sorted order, εˆi is plotted versus the (i−0.5)/n quantile in the standard normal
distribution

2. TRUE statement

3. TRUE statement. We know that MSE is an estimate of the error variance. And this can
SSE
be found as (k−1)(l−1) .

4. FALSE statement. The command qnorm(resi)returns the quantiles in the standard

normal distribution of the values in resi

5. TRUE statement. It is the sample mean, which is used as an estimate of the mean

------------------------------------ FACIT-END ------------------------------------

Continues on page 6

5
 Exercise III

In a study 605 test persons, all with a record of previous heart disease, were randomized to
one of two possible diets (A or B), in order to study the effect of diet on health. After an
observation period of 4 years the test persons were classified according to health status: (I)
dead, (II) cancer, (III) other disease, (IV) well.

Health status
I II III IV Total
Diet A 15 24 25 239 303
Diet B 7 14 8 273 302
Total 22 38 33 512 605

The null hypothesis in the study was that there is no association between diet and health.

Question III.1 (4)

State the distribution of the usual test statistics, when assuming that the null hypothesis is
true:

1  The usual test statistics follows a χ2 -distribution with 8 degrees of freedom

2  The usual test statistics follows a F -distribution with (1, 603) degrees of freedom
3  The usual test statistics follows a t-distribution with 4 degrees of freedom
4  The usual test statistics follows a t-distribution with 302 degrees of freedom
5*  The usual test statistics follows a χ2 -distribution with 3 degrees of freedom

----------------------------------- FACIT-BEGIN -----------------------------------

The setup of the data is a multi-sample proportion setup (chapter 7.4). We must test the
hypothesis, that the proportions in each group is equal
H0 : P1 = p2 = p3 = p4 .
and under this hypothesis the test statistic follows a χ2 -distribution with c − 1 degrees of
freedom, and there are 4 groups, so 3 degrees of freedom (Method 7.20).

------------------------------------ FACIT-END ------------------------------------

Question III.2 (5)

We now only consider the proportion of test persons who are healthy at the end of the 4 year
period. We want to estimate at 95% confidence interval for the difference in proportions of test

6
 persons who are healthy for each of the 2 diets. Which of the suggestions below is the correct
code in R to achieve this?

1  prop.test(x=c(512), n=c(605), correct=FALSE)

2  prop.test(x=c(303,302), n=c(512,512), correct=FALSE)

3*  prop.test(x=c(239,273), n=c(303,302), correct=FALSE)

4  prop.test(x=c(239,273), n=c(605,605), correct=FALSE)

5  prop.test(x=c(239,273), n=c(512,512), correct=FALSE)

----------------------------------- FACIT-BEGIN -----------------------------------

Here we are working with proportions in two populations as described in Chapter 7.3. We need
the observed proportion which are well for each diet. So on Diet A 239 out of 303 are well and
for Diet B 273 out of 302 are well, and these numbers are passed to prop.test, which then
prints out the estimated confidence interval (same as Example 7.19).

------------------------------------ FACIT-END ------------------------------------

Continues on page 8

7
 Exercise IV

In the production of a consumer product 3 subprocesses are involved, denoted A, B and C. The
time (in hours) it takes to complete each subprocess is represented with a random variable,
which we denote XA , XB and XC , respectively. It can be assumed, that XA , XB and XC
are all independent and normally distributed given by XA ∼ N (12, 22 ), XB ∼ N (25, 32 ) and
XC ∼ N (42, 42 ).

The total production time, Y, is now defined by

Y = XA + XB + X C .

Question IV.1 (6)

State the probability that the total production time, Y , exceeds 85 hours:

1  0.0081
2  0.1080
3*  0.1326
4  0.4180
5  0.6301

----------------------------------- FACIT-BEGIN -----------------------------------

We need to find the mean and variance of Y , which we know is normal distributed, since a
linear function of normal distributed random variables is also normal distributed (Theorem
2.56).

We use the identities in Theorem 2.56 to get

µY = E(Y ) = E(XA + XB + XC ) = E(XA ) + E(XB ) + E(XC ) = 12 + 25 + 42 = 79,
and
σY2 = V(Y ) = V(XA + XB + XC ) = V(XA ) + V(XB ) + V(XC ) = 4 + 9 + 16 = 29.

Alternatively we could also have simulated the variance in R.

k <- 1000000
X_a <- rnorm(k, 12, 2)
X_b <- rnorm(k, 25, 3)
X_c <- rnorm(k, 42, 4)
Y <-X_a + X_b + X_c
var(Y)

8
 ## [1] 29.01567

This we use to look up the probability P (Y > 85) = 1 − P (Y ≤ 85) in R by:

1 - pnorm(q=85, mean=79, sd=sqrt(29))

## [1] 0.1326027

------------------------------------ FACIT-END ------------------------------------

Question IV.2 (7)

An engineer is now able to perform some optimization of the process, so that the improved
process time Y ∗ , becomes

Y ∗ = 0.9 · XA + 0.8 · XB + XC ,

where XA , XB and XC are defined as in the previous question.

State the variance of Y ∗ :

1  V(Y ∗ ) = (0.9 + 0.8 + 1) · (22 + 32 + 42 )

2  V(Y ∗ ) = (0.92 + 0.82 + 12 ) · (22 + 32 + 42 )

3  V(Y ∗ ) = 0.9 · 22 + 0.8 · 32 + 1 · 42

4*  V(Y ∗ ) = 0.92 · 22 + 0.82 · 32 + 12 · 42

5  V(Y ∗ ) = 22 + 32 + 42

----------------------------------- FACIT-BEGIN -----------------------------------

Again the identities in Theorem 2.56 to get

σY2 ∗ = V(Y ∗ ) = V(0.9 · XA + 0.8 · XB + XC , )

= 0.92 V(XA ) + 0.82 V(XB ) + V(XC )
= 0.92 · 22 + 0.82 · 32 + 12 · 42

------------------------------------ FACIT-END ------------------------------------

Continues on page 10

9
 Exercise V

The yearly rainfall has been registered within a region for the last 100 years. It can be assumed
that the rainfall is independent from year to year. The cumulative distribution for the yearly
rainfall is shown in the figure below:

Cumulative distribution

1.0

0.8

0.6
Fn(x)

0.4

0.2

0.0

660 680 700 720 740 760

rainfall [mm]

The following summary of the data has been conducted by the use of R, where the yearly
rainfall measurements are stored in the variable rainfall:

> var(rainfall)
[1] 412.7042
> summary(rainfall)
Min. 1st Qu. Median Mean 3rd Qu. Max.
652.8 686.6 701.9 701.3 714.9 749.1

Continues on page 11

10
 Question V.1 (8)

Which of the following statements is not correct?

√
412.7042
1  The estimate of the standard deviation of the mean σ̂X̄ , becomes 10
mm

2  The 50% quantile for the 100 observations is 701.9 mm

√
3  The standard deviation of the sample, s, for the 100 measurements is 412.7042 mm

4  50% of the 100 observations are between 686.6 and 714.9 mm

412.7042
5*  The estimated coefficient of variation for the 100 observations becomes 701.9

----------------------------------- FACIT-BEGIN -----------------------------------

Lets go through them one by one:

1. TRUE statement. The formula for the estimate is √s (also called the standard error of
n
the mean). See Definition 3.7

2. TRUE statement. Seen from the summary() call

3. TRUE statement. Standard deviation is the square root of the variance

4. TRUE statement. 686.6 is the first quartile (25% quantile) and 714.9 is the third quartile
(75% quantile), and certainly 50% of the observations lies between the 25% and 75%
quantile
√
s 412.7042
5. FALSE statement. The estimated coefficient of variation is V̂ = x̄
= 701.3
. See
Definition 1.12.

------------------------------------ FACIT-END ------------------------------------

Question V.2 (9)

Provide a 95% confidence interval for the variance of the rainfall based on the 100 observations,
still assumed to be normally distributed:

2 2
1  [ 20.3151299·134.6416 ; 20.3151299·69.22989 ]
2
·99 20.31512 ·992
2  [ 20.31512
134.6416
; 69.22989 ]

3*  [ 412.7042·99
128.422
; 412.7042·99
73.36108
]

4  [ 412.7042·99
123.2252
; 412.7042·99
77.04633
]

11
5  [ 20.31512·99
123.2252
; 20.31512·99
77.04633
]

----------------------------------- FACIT-BEGIN -----------------------------------

We find the formula for a 1 − α confidence interval for the variance of a normal distributed
population in Method 3.19 and insert the values
" #
s2 (n − 1) s2 (n − 1)
,
χ21−α/2 χ2α/2

The chi-square quantiles are found in R as

qchisq(c(0.025, 0.975), 99)

## [1] 73.36108 128.42199

------------------------------------ FACIT-END ------------------------------------

Continues on page 13

12
 Question V.3 (10)

We continue with the exercise from the previous page. The following code in R has now been
run:

k = 10^5
Q5 = function(x){ quantile(x, 0.95) }
samples = replicate(k, sample(rainfall, replace = TRUE))
simvalues = apply(samples, 2, Q5)
interval = quantile(simvalues, c(0.025,0.975))

which gives the result:

> interval
2.5% 97.5%
728.9515 742.0814

What has been calculated in the vector interval?

1  A 95% confidence interval for the mean of the yearly rainfall (parametric bootstrap)
2  A 95% confidence interval for the 5% quantile of the yearly rainfall (parametric bootstrap)
3*  A 95% confidence interval for the 95% quantile of the yearly rainfall (non-parametric
bootstrap)
4  A 95% confidence interval for the 2.5% and 97.5% quantile of the yearly rainfall (non-
parametric bootstrap)
5  A 95% confidence interval for the 2.5% and 97.5% quantile of the yearly rainfall (para-
metric bootstrap)

----------------------------------- FACIT-BEGIN -----------------------------------

We look at the R code and see that it is a bootstrapping is carried out by simulating the sample
100000 times, and not assuming any distribution (since the sample function is used), therefore
it is non-parametric.

The statistic calculated for each simulated sample is the 95% quantile and since the quantiles
taken for these values are the 2.5% and the 97.5%, then the results is a 95% confidence interval
for the 95% quantile.

------------------------------------ FACIT-END ------------------------------------

Continues on page 14

13
 Exercise VI

We consider an experiment that can result in one of two possible outcomes, here denoted A
or B. The probability of outcome A is denoted P (A). By defintion we get the probability of
outcome B as P (B) = 1 − P (A).

Question VI.1 (11)

Assume that we observe a random variable, X, which counts the number of times that we
observe the outcome A out of n = 300 independent trials of the experiment. If we assume that
P (A) = 0.40 in a single trial, what is then the expected number E(X) and variance V(X)?

1  E(X) = 300 · 0.4 · (1 − 0.4) and V(X) = 3002 · 0.4

2  E(X) = 300 · 0.4 and V(X) = 3002 · 0.4 · 0.6

3*  E(X) = 300 · 0.4 and V(X) = 300 · 0.4 · 0.6

4  E(X) = 300 · 0.4 · 0.6 and V(X) = 3002 · 0.42 · 0.62

5  E(X) = 300 · 0.4 · 0.6 and V(X) = 300 · 0.42 · 0.62

----------------------------------- FACIT-BEGIN -----------------------------------

X follows a Binomial distribution with p = 0.4 and we have a formula for the mean and variance
defined in Theorem 2.21, which we use to get

µ = E(X) = np = 300 · 0.4,

σ 2 = V(X) = np(1 − p) = 300 · 0.4 · 0.6.

------------------------------------ FACIT-END ------------------------------------

Question VI.2 (12)

Regardless of your answer to the previous question we now want to estimate the probability
P (A) based on the n = 300 trials. From the n = 300 trials we count that in 120 of these the
outcome was A and in the remaining 180 trials the outcome was B. Provide a 95% confidence
interval for the probability P (A):

1  [0.33, 0.48]

2*  [0.35, 0.46]

3  [0.35, 0.42]

14
4  [0.31, 0.53]

5  [0.29, 0.54]

----------------------------------- FACIT-BEGIN -----------------------------------

Using the inbuilt function in R the results is

prop.test(120, 300, correct=FALSE)

##
## 1-sample proportions test without continuity correction
##
## data: 120 out of 300, null probability 0.5
## X-squared = 12, df = 1, p-value = 0.000532
## alternative hypothesis: true p is not equal to 0.5
## 95 percent confidence interval:
## 0.3461652 0.4563634
## sample estimates:
## p
## 0.4

whereas using the formula in method 7.3 gives a slightly different result is obtained

n <- 300
x <- 120
phat <- x/n
phat + c(-1,1) * qnorm(p=0.975) * sqrt(phat*(1-phat)/n)

## [1] 0.3445638 0.4554362

This is due to a numerical rounding by R and can occur sometimes. The answer is in any case
closest to the answer marked correct [0.35, 0.46].

------------------------------------ FACIT-END ------------------------------------

Continues on page 16

15
 Exercise VII

An engineer is examining the quality in a batch of raw materials. The quality demand is that
the purity of the raw material is at least 90%. The engineer takes a sample of 10 independent
measurements from the batch and saves the measured values (in %) of the purity in a vector x.

He then runs the following code in R

> x <- c(90.6, 90.3, 88.9, 87.5, 87.6, 88.1, 87.5, 88, 88, 89.6)
> n <- length(x)
> tobs <- (mean(x) - 90) / (sd(x) / sqrt(n))
> pt(tobs, df=n-1)

Which yields the following output

[1] 0.002279236

Question VII.1 (13)

Based on the calculations listed above, and assuming that the measurements of the purity
are normally distributed and applying a significance level of α = 0.05, what can the engineer
conclude?

1  The engineer can conclude that the purity of the raw material is at least 88.6%
2  The engineer can conclude that the mean purity of the raw material is at most 88.6%
3  The engineer has with probability 99.7% shown that the mean purity of the raw material
is 90%
4  The engineer can assume that the mean purity of the raw material is 90%
5*  The engineer can reject that the mean purity of the raw material is 90%

----------------------------------- FACIT-BEGIN -----------------------------------

We can see from the way that tobs is calculated that the null hypothesis is that the µ = 90
(See Method 3.23) Since the p-value is 2*pt(tobs, df=n-1)=0.0046 and thus much lower than
α = 0.05. This leads to the conclusion that the null hypothesis, that the mean purity is 90%,
must be rejected.

------------------------------------ FACIT-END ------------------------------------

Continues on page 17

16
 Exercise VIII
2
(w−30)
We consider a random variable W with density function f (w) = √1 e− 162 .
9 2π

The density function is shown in the figure below, where the probability
P (30 < W < 33) is shown as the shaded area.
0.04
0.03
f(w)

0.02
0.01
0.00

0 10 20 30 40 50 60

Question VIII.1 (14)

Calculate the probability P (30 < W < 33):

1  0.09

2*  0.13

3  0.24

4  0.34

5  0.84

----------------------------------- FACIT-BEGIN -----------------------------------

The answer is obtained from recognizing the the formula for the probability density function
(pdf) for the normal distribution in definition 2.37
1 (w−µ)2
f (w) = √ e− 2·σ2
σ 2π

17
and thus to find the mean µ = 30 and variance σ = 9. These are then used to obtain
P (30 < W < 33) = P (X < 33) − P (X < 30)
in R

pnorm(33, mean=30, sd=9) - pnorm(30, mean=30, sd=9)

## [1] 0.1305587

SINCE in the original exam the plot was which indeed was wrong, it was of the normal distri-
0.12
0.10
0.08
f(w)

0.06
0.04
0.02
0.00

15 20 25 30 35 40 45

bution with mean µ = 30 and variance σ = 3

pnorm(33, mean=30, sd=3) - pnorm(30, mean=30, sd=3)

## [1] 0.3413447

then the Answer 4 is also counted as correct!

------------------------------------ FACIT-END ------------------------------------

18
 Continues on page 19

19
 Question VIII.2 (15)

We consider a situation where we take 3 different samples denoted A, B, and C. All three
(w−30)2
samples are from the population characterized by the density f (w) = √1 e− 162 as in the
9 2π
previous question.

Sample A is of size nA = 10 and the estimated mean is denoted µ̂A . Sample B is of size
nB = 30 and the estimated mean is denoted µ̂B . Sample C is of size nC = 100 and the
estimated mean is denoted µ̂C .

The question is now whether the sample mean will exceed the value 33, even when the popu-
lation mean is equal to 30.

Which statement is correct?

1*  P (µ̂A ≥ 33) > P (µ̂B ≥ 33)

2  P (µ̂C ≥ 33) > P (µ̂A ≥ 33)
3  P (µ̂C ≥ 33) = P (µ̂B ≥ 33)
4  P (µ̂A ≥ 33) = P (µ̂B ≥ 33) · P (µ̂C ≥ 33)
5  P (µ̂A ≥ 33) = 12 P (µ̂B ≥ 33)

----------------------------------- FACIT-BEGIN -----------------------------------

Lets go through the statements:

1. TRUE statement. The µ̂ is the sample mean, which we know follow the distribution
µ̂ ∼ N (µ, σ 2 /n) (Theorem 3.3), so we get the following
µ̂A ∼ N (30, 81/10)
µ̂B ∼ N (30, 81/30)
µ̂C ∼ N (30, 81/100)
and we can actually then realize, that the probability of getting a an outcome above the
same value, must be higher for XA than the two others, since its pdf has higher variance
than the others. In R we can check it by:

## P(X_A >= 33)

(1-pnorm(q=33, mean=30, sd=sqrt(81/10)))
## [1] 0.1459203
## P(X_B >= 33)
(1-pnorm(q=33, mean=30, sd=sqrt(81/30)))
## [1] 0.03394458

20
 2. FALSE statement. Following same argument as above

3. FALSE statement. Since the variance is different, then they are not equal

4. FALSE statement. Be sure by checking the product in R:

(1-pnorm(q=33, mean=30, sd=sqrt(81/30))) *

(1-pnorm(q=33, mean=30, sd=sqrt(81/100)))

## [1] 1.456427e-05

5. FALSE statement. Be sure by checking the product in R:

0.5 * (1-pnorm(q=33, mean=30, sd=sqrt(81/30)))

## [1] 0.01697229

------------------------------------ FACIT-END ------------------------------------

Continues on page 22

21
 Exercise IX

The yield from a chemical process, Yi , is assumed to depend linearly on the temperature, ti ,
measured in degrees. In order to achieve insight about this relation, an experiment has been
conducted where n = 50 pairwise measurements of Yi and ti has been taken. It is assumed that
the following model can give a reasonable description of the relation

Yi = β0 + β1 · ti + εi .

The residuals in this model are assumed independent and normally distributed with constant
variance, i.e. εi ∼ N (0, σε2 ). Relevant output from the analysis in R is given below:

Call:
lm(formula = y ~ t)

Residuals:
Min 1Q Median 3Q Max
-5.0816 -1.4994 -0.2493 1.5175 4.8506

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 65.4919 2.7757 23.595 <2e-16 ***
t 0.1637 0.1103 1.485 0.144
---

Residual standard error: 2.296 on 48 degrees of freedom

Multiple R-squared: 0.04392,Adjusted R-squared: 0.024
F-statistic: 2.205 on 1 and 48 DF, p-value: 0.1441

Question IX.1 (16)

Which of the following statements is correct when the significance level α = 0.05 is applied?

1  The yield increases by 16.37% when the temperature increase one degree

2*  There is no significant linear relation between temperature and yield

3  The test statistics for no effect of temperature on yield (i.e. the null hypothesis H0 : β1 =
0) is 23.595

4  A 95% confidence interval for the effect of temperature, β1 , is [-0.132027, 0.4594821]

5  The correlation between temperature and yield is 0.04392

22
----------------------------------- FACIT-BEGIN -----------------------------------

Lets go through the answers one by one:

1. FALSE statement. The yield is estimated to increase 0.1637 units (we are not informed
about the units) per degree, which is not the same as 16.37% (increasing some proportion
per degree, would also lead to an exponential relation, not linear)

2. TRUE statement. The test of the null hypothesis

H0 : β1 = 0

leads to a p-value of 0.144, which is not below the significance level α = 0.05 and since
this is equivalent to testing for correlation equal to zero

H0 : ρ = 0

there is not found a significant linear relation between the yield and the temperature

3. FALSE statement. Since, the test statistic for no effect is 1.485

4. FALSE statement. The lower limit of the CI is 0.1637 − 1.96 ∗ 0.1103 = −0.052 and the
upper is 0.1637 + 1.96 ∗ 0.1103 = 0.380
√ √
5. FALSE statement. The correlation is r2 = 0.04392 = 0.21

------------------------------------ FACIT-END ------------------------------------

Continues on page 24

23
 Question IX.2 (17)

We continue with the exercise from the previous page. It turns out that the pH of the process
may influence the yield, and since pH has been measured, it is decided to include it into the
model, which in its extended form becomes:

Yi = β0 + β1 · ti + β2 · pH i + εi .

Estimation of the model parameters gives the following output in R:

Call:
lm(formula = y ~ t + pH)

Residuals:
Min 1Q Median 3Q Max
-3.7253 -1.2818 -0.2978 1.0724 4.4488

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 49.46756 4.09799 12.071 5.25e-16 ***
t 0.24113 0.09315 2.589 0.0128 *
pH 2.37090 0.50097 4.733 2.06e-05 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 1.91 on 47 degrees of freedom

Multiple R-squared: 0.3525,Adjusted R-squared: 0.3249
F-statistic: 12.79 on 2 and 47 DF, p-value: 3.667e-05

Give estimates for the model parameters, i.e. β0 , β1 , β2 and σε2

1  (β̂0 , β̂1 , β̂2 , σ̂ε2 ) = (4.09799, 0.24113, 2.37090, 0.3525)

2*  (β̂0 , β̂1 , β̂2 , σ̂ε2 ) = (49.46756, 0.24113, 2.37090, 1.912 )

3  (β̂0 , β̂1 , β̂2 , σ̂ε2 ) = (49.46756, 0.24113, 2.37090, 1.91 · 47)

4  (β̂0 , β̂1 , β̂2 , σ̂ε2 ) = (4.09799, 0.09315, 0.50097, 1.91 · 47)

5  (β̂0 , β̂1 , β̂2 , σ̂ε2 ) = (2.37090, 0.50097, 4.733 , 1.91)

----------------------------------- FACIT-BEGIN -----------------------------------

The estimates are read directly from the printed output. See Example 6.3

24
------------------------------------ FACIT-END ------------------------------------

Continues on page 25

25
 Question IX.3 (18)

We continue with the exercise from the previous page and the model

Yi = β0 + β1 · ti + β2 · pHi + εi .

Provide a 95% confidence interval for the effect on yield when pH increases one unit:

1  0.24113 ± 2.01174 · 0.09315

2*  2.37090 ± 2.01174 · 0.50097

3  (49.46756 + 0.24113 + 2.37090) ± 2.01174 · (4.09799 + 0.09315 + 0.50097)

4  2.37090 ± 0.509920 · 0.50097

5  (49.46756 + 0.24113 + 2.37090) ± 0.509920 · 0.50097

----------------------------------- FACIT-BEGIN -----------------------------------

See Method 6.5. The confidence interval for the effect of pH is found inserting the printed
values into

β̂2 ± t1−α/2 · σ̂β2

using the t-distribution with n − (p + 1) = 47 degrees of freedom to find the quantile t1−α/2 :

qt(p=0.975, df=47)

## [1] 2.011741

------------------------------------ FACIT-END ------------------------------------

Exercise X

Assume there exists a dice with 10 sides and where the probability for each of the 10 outcomes,
1, 2, . . . , 10, is the same. Consider the discrete random variable X with density f (x) = 0.1 for
x ∈ (1, 2, . . . , 10).

Question X.1 (19)

Give the mean value of X:

26
 1
P10
1 (10−1) i=1 xi = 6.11
1
P10
2 (10−6.11) i=1 |xi − 6.11| = 6.48
1
P10
3 (10) i=1 (xi − 6.11)2 = 8.62
P10 10−1
4 i=1 10
xi · 0.1 = 4.95
P10
5*  i=1 xi · 0.1 = 5.50

----------------------------------- FACIT-BEGIN -----------------------------------

See Definition 2.13. We use the formula for calculating the mean value of a discrete random
variable
n
X
xi f (xi )
i=1

and insert the values. In R:

sum(1:10*0.1)

## [1] 5.5

------------------------------------ FACIT-END ------------------------------------

Continues on page 28

27
 Exercise XI

The yield of a process is µ = 60 mg/l. Certain changes to the process are being planed and it
is desirable to be able to prove an effect on the mean yield if the change is at least 5 mg/l (i.e.
a two-sided test).

An engineer is now going to plan an experiment to evaluate the effect of the process changes.
He wants to decide how large a sample is needed. The sample size has to be large enough to
detect the relevant effect (5 mg/l) with a power of 0.8 when applying a significance level of
α = 0.05. It can be assumed that the standard deviation is σ = 10 mg/l.

Question XI.1 (20)

Based on the information above, and by applying the function power.t.test in R, one con-
cludes that, if an equal number of measurements are taken, then the minimum number of
measurements n needed becomes:

1  n ' 256 measurements

2  n ' 128 measurements

3  n ' 64 measurements

4*  n ' 34 measurements

5  n ' 27 measurements

----------------------------------- FACIT-BEGIN -----------------------------------

Based on the given information the planned test is a one-sample test, since it is not stated that
a sample should be taken before the change, only that the yield before is µ = 60 mg/l. See
Example 3.67.

power.t.test(delta=5, sd=10, sig.level=0.05, power=0.8, type="one.sample")

##
## One-sample t test power calculation
##
## n = 33.3672
## delta = 5
## sd = 10
## sig.level = 0.05
## power = 0.8
## alternative = two.sided

Rounding up to n ' 34 measurements.

28
Since, it is not completely clear, that the it should not be a two-sample setup – one could argue
that a nothing in the information given prevents it from being a two-sample test – then Answer
3 is also taken as correct, since:

power.t.test(delta=5, sd=10, sig.level=0.05, power=0.8, type="two.sample")

##
## Two-sample t test power calculation
##
## n = 63.76576
## delta = 5
## sd = 10
## sig.level = 0.05
## power = 0.8
## alternative = two.sided
##
## NOTE: n is number in *each* group

Further, since it is also not specified that n is the number of measurements is in each group
(and not the total), then Answer 2 is also taken as correct.

------------------------------------ FACIT-END ------------------------------------

Continues on page 30

29
 Exercise XII

In a study the aim is to investigate the possible cholesterol lowering effect of a product. 9 test
persons had their cholesterol level measured (denoted x1). After 3 months, while using the
product, the same 9 test persons had their cholesterol level measured again (denoted x2). Data
is shown in the table below:

Person 1 2 3 4 5 6 7 8 9
x1 63.5 66.7 59.2 57.4 63.9 63.2 60.7 62.6 63.3
x2 51.3 51.9 57.8 50.2 54.6 43.3 51.2 40.4 52.2

The following code is now run in R, in order to test whether the change over time can be
assumed to be zero (H0 : δ = 0):

x1 <- c(63.5, 66.7, 59.2, 57.4, 63.9, 63.2, 60.7, 62.6, 63.3)
x2 <- c(51.3, 51.9, 57.8, 50.2, 54.6, 43.3, 51.2, 40.4, 52.2)

The output from the standard statistical analysis is given below. Please note that some numbers
in the standard output have been replaced by the letters A, B and C.

t = -5.6354, df = A, p-value = B
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
-16.847799 C
sample estimates:
mean of the differences
-11.95556

Question XII.1 (21)

What conclusion can be made when applying a significance level of α = 0.05?

1  We can show an effect since µD = −11.95556

2  We can not show an effect since the upper limit of the confidence interval is 7.063312

3  We can not show an effect since the lower limit of the confidence interval is -7.063312

4*  We can show an effect since the p-value is 4.897 · 10−4

5  We can show an effect since the p-value is 2.394 · 10−4

30
----------------------------------- FACIT-BEGIN -----------------------------------

The standard statistical test for this setup is a paired two-sample t-test. The R output is from
t.test(), and the easiest way to solve this is by copying and running

x1 <- c(63.5, 66.7, 59.2, 57.4, 63.9, 63.2, 60.7, 62.6, 63.3)
x2 <- c(51.3, 51.9, 57.8, 50.2, 54.6, 43.3, 51.2, 40.4, 52.2)
## The call is then either "t.test(x2, x1, paired=TRUE)" or
t.test(x2-x1)

##
## One Sample t-test
##
## data: x2 - x1
## t = -5.6354, df = 8, p-value = 0.0004897
## alternative hypothesis: true mean is not equal to 0
## 95 percent confidence interval:
## -16.847799 -7.063312
## sample estimates:
## mean of x
## -11.95556

and from the p-value we can find the correct answer. See section 3.1.7 for more examples.

------------------------------------ FACIT-END ------------------------------------

Continues on page 32

31
 Exercise XIII

A biologist is interested in examining the effect of 4 different growth inhibitors, denoted V1 ,

V2 , V3 og V4 . The 4 growth inhibitors are added to samples from the same cell line and growth
after one week is measured Yij (number of cells per cm2 ). 8 replicates are made for each growth
inhibitor, i.e. we have a total of 32 measurements. As the measurements can be assumed
normally distributed, it is chosen to apply the following analysis of variance model
Yij = µ + αi + εij .
In this model αi denotes the effect of growth inhibitor i (i = 1, 2, 3, 4), µ is the overall average
εij are the errors, assumed independent and normally distributed with mean zero and standard
deviation σε .

An analysis of variance is performed for the above model and the output is given below. Please
note that the output is incomplete as some numbers are replaced by the symbols A, B and C.

Analysis of Variance Table

Response: growth
Df Sum Sq Mean Sq F value Pr(>F)
treatment A 281.07 B C 0.0001409 ***
Residuals 28 268.46 9.588

Question XIII.1 (22)

Provide the usual test statistics (denoted by C) in order to test for equal mean effect of the 4
growth inhibitors

1*  9.77
2  7.23
3  2.95
4  4.57
5  16.11

----------------------------------- FACIT-BEGIN -----------------------------------

As stated in Theorem 8.6, we can calculate the observed test statistic by

SS (Tr )/(k − 1) 281.07/(4 − 1)
Fobs = = = 9.77,
SSE /(n − k) 268.46/(32 − 4)
where

32
 • SS (Tr ) is the variance explained by the effect of the treatment

• SSE is the variance remaining after the model (sum of squared error)

• n is the total number of observations

• k is the number of groups

------------------------------------ FACIT-END ------------------------------------

Continues on page 34

33
 Question XIII.2 (23)

We now want to calculate a post hoc 95% confidence interval for a difference in mean between
growth inhibitor V1 and V2 , here denoted I0.95 (V1 − V2 ). From the experiment it is known that
the estimated mean difference between V1 and V2 is 4.5. State the interval I0.95 (V1 − V2 ):

9.588
√
1  I0.95 (V1 − V2 ) = 4.5 ± 2.048 · 12
28 ·
√ p
2*  I0.95 (V1 − V2 ) = 4.5 ± 2.048 · 9.588 · 2/8
√
9.588
3  I0.95 (V1 − V2 ) = 4.5 ± 2.306 · √
12
p
4  I0.95 (V1 − V2 ) = 4.5 ± 2.306 · 9.5882 · 1/8
9.588
5  I0.95 (V1 − V2 ) = 4.5 ± 1.960 · √
8

----------------------------------- FACIT-BEGIN -----------------------------------

See method 8.9. The post hoc confidence interval for the difference is
s  
SSE 1 1
ȳi − ȳj ± t1−α/2 + .
n − k ni nj

So we use the t-distribution with n − k = 32 − 4 = 28 degrees of freedom

qt(p=0.975, df=28)

## [1] 2.048407

and insert the values

s  
268.46 1 1
4.5 ± 2.048 · + ,
28 8 8

which we cannot directly find among the answers, so we shorten it

s  
2
4.5 ± 2.048 · 9.588 ,
8

and finally find the answer

√ p
4.5 ± 2.048 · 9.588 · 2/8.

------------------------------------ FACIT-END ------------------------------------

34
 Exercise XIV

We consider a continuous random variable random, where the well-known cumulative distribu-
tion function F (x) is given by P (X ≤ x) = 1 − e−x/2 , where x > 0.

Question XIV.1 (24)

Provide the mean of X:

1
1 2

2 1

3*  2
3
4 2

5 4

----------------------------------- FACIT-BEGIN -----------------------------------

It is recognized as the cdf of the exponential distribution (Definition 2.48), which is verified by
Z x
x
λeλy dy = −e−λy + c 0 = −e−λx + e0 = 1 − e−λx

0

and it can be seen that λ = 12 . Using the formula for the mean of an exponential distribution
(Theorem 2.49)
1
µ= = 2.
λ

------------------------------------ FACIT-END ------------------------------------

Continues on page 36

35
 Exercise XV

A biologist is examining the bio-diversity within an area and has measured the number of
different type of plants per 10 m2 in different places in the area. She has obtained a total of
30 independent measurements, yi , and these are in in the vector Yobs in R.

Question XV.1 (25)

The biologist would like to estimate a 95% confidence interval for the coefficient of variation for
the bio-diversity (number of different type of plants per 10 m2 ) by applying the non-parametric
bootstrap. Which of the following suggestions in R is most suitable to achieve this?

1  samples = replicate(10000,rnorm(30,mean(Yobs),sd(Yobs))
results = apply(samples,2,sd)/apply(samples,2,mean)
quantile(results, c(0.025,0.975))
2  samples = replicate(10000,sample(Yobs,replace=TRUE))
results = apply(samples,2,var)/apply(samples,2,sd)
quantile(results, c(0.025,0.975))
3  samples = replicate(10000,rnorm(30,mean(Yobs),sd(Yobs))
results = apply(samples,2,var)/apply(samples,2,median)
quantile(results, c(0.025,0.975))
4  samples = replicate(10000,sample(Yobs,replace=FALSE))
results = apply(samples,2,sd)/apply(samples,2,mean)
quantile(results, c(0.025,0.975))
5*  samples = replicate(10000,sample(Yobs,replace=TRUE))
results = apply(samples,2,sd)/apply(samples,2,mean)
quantile(results, c(0.025,0.975))

----------------------------------- FACIT-BEGIN -----------------------------------

In the code in Answer 1 and 3 the samples are simulated using rnorm(), hence a normal
distribution is assumed and it is not non-parametric bootstrapping (but parametric).

In Answer 2 it is not the coefficient of variation which is calculated by

apply(samples,2,var)/apply(samples,2,sd),
which it is in Answer 4 and 5 by
apply(samples,2,sd)/apply(samples,2,mean).
The difference between 4 and 5 is that in Answer 4 the samples are drawn without replacement
sample(Yobs,replace=FALSE),
which is wrong, where in Answer 5 the samples are drawn correctly with replacement
sample(Yobs,replace=TRUE). See Chapter 4.3 for more on non-parametric bootstrap.
------------------------------------ FACIT-END ------------------------------------
Continues on page 37

36
 Exercise XVI

In a study 178 men and 180 women were asked to answer whom of 2 political candidates, A or
B, they preferred. Alternatively, they could answer ”none of the two”. The distribution of the
answers is shown in the figure below.

85

80
78

60
60

55

choice
Candidate A
count

40
40 40 Candidate B
None of the two

20

men women
sex

Continues on page 38

37
 Question XVI.1 (26)

It is seen from the figure that we observe that 85 out of the 180 women prefer Candidate B. If
we can assume the same distribution of answers by gender, how many women out of the 180
would we expect to prefer Candidate B?

163 95
1 358
· 358
· 358
100 223
2 358
· 358
· 358
95 190
3 358
· 358
· 358
163 180
4*  358
· 358
· 358
95 180
5 358
· 358
· 358

----------------------------------- FACIT-BEGIN -----------------------------------

See chapter 7.2. The total number of respondents are n = 180 + 178 = 358 and if we assume
the same distribution of answers by gender, i.e. the under the hypothesis that the proportion
of men and women prefering B is equal

H0 : pmen,B = pwomen,B = p,

then
”Total number for B” 78 + 85 163
p= = = .
”Total number” 358 358
It is then simply this fraction we expect out of the total number of women
163
· 180,
358
which is then expressed a little longer by
163 180
· · 358.
358 358

------------------------------------ FACIT-END ------------------------------------

Question XVI.2 (27)

Provide the usual test statistics when you want to conduct the test of whether the distribution
of answers is the same for men and women:

1  χ2obs = 5.9915

38
2*  χ2obs = 6.6581

3  χ2obs = 16.212

4  χ2obs = 8.3836

5  χ2obs = 4.5067

----------------------------------- FACIT-BEGIN -----------------------------------

Maybe the easiest is to copy example 7.21 from the book of testing multiple proportions

prop <- matrix(c(60, 78, 40, 40, 85, 55), ncol = 3, byrow = TRUE)
rownames(prop) <- c("Men", "Women")
colnames(prop) <- c("A", "B", "None")
prop

## A B None
## Men 60 78 40
## Women 40 85 55

chisq.test(prop, correct=FALSE)

##
## Pearson's Chi-squared test
##
## data: prop
## X-squared = 6.6581, df = 2, p-value = 0.03583

------------------------------------ FACIT-END ------------------------------------

Continues on page 40

39
 Exercise XVII

Cloud seeding is a form of weather modification that can be used to increase the amount of
precipitation that falls from the clouds, by dispersing substances (small particles) e.g. alumini-
umoxid into the clouds to modify their development.

In an experiment the aim was to study the effect of cloud seeding by using a new type of
particles. The amount of precipitation (mm precipitation per day) for 35 days with cloud
seeding using the new particles is denoted Xi , (i = 1, 2, . . . , 35). This was compared to the
amount of precipitation on 30 days without cloud seeding, denoted Yj , (j = 1, 2, . . . , 30).
Measurements were only taken on days where there was sufficient humidity in the air to make
the experiment relevant. Data from the experiment is shown in the figure below.

16
●

12 ●
●
Precipitation (mm)

X (particle) Y (control)
Experiment

Continues on page 41

40
 We now want to analyze the data described on the previous page using R. Data xi is stored in
the vector x and data yj is stored in the vector y, and the following code has been run:

k <- 10^4
resultX <- replicate(k, sample(x, replace = TRUE))
resultY <- replicate(k, sample(y, replace = TRUE))
result <- apply(resultX, 2, median) - apply(resultY, 2, median)
quantile(result, c(0.5, 0.025,0.975))

Which gives the result

50% 2.5% 97.5%

1.6283069 0.2843492 2.4233546

Question XVII.1 (28)

If we apply a significance level of α = 0.05 what can then be concluded?

1*  The median for X is significantly higher than the median for Y

2  The median for X is 62.8% higher than the median for Y
3  Precipitation for X is between 28.4% and 142.3% higher than precipitation for Y
4  The mean precipitation can be assumed equal for the two methods
5  The medianen for Y is [0.28; 2.42] higher than the median for X

----------------------------------- FACIT-BEGIN -----------------------------------

In the R code a 95% non-parametric bootstrap confidence interval for the difference in median
is calculated, and since 0 is not contained in the interval, then the hypothesis

H0 : q0.5,X = q0.5,Y

must be rejected on significance level α = 0.05, thus concluded that

H1 : q0.5,X 6= q0.5,Y

and further, since X − Y was calculated and the interval is on the positive side, then it can be
concluded that q0.5,X > q0.5,Y .

------------------------------------ FACIT-END ------------------------------------

Continues on page 42

41
 Question XVII.2 (29)

In a different experiment using cloud seeding a different kind of particles were examined. Also
in this experiment the amount of precipitation was compared when the particles were used to a
situation with no use of particles. In this study, however, it was decided to log transform (the
natural logarithm) the data before comparing the groups. By transforming the data it can be
assumed that data in the two groups follows a normal distribution. The data is summarized in
the table below (unit is log mm precipitation).

Particles, X Control, Y
(log mm precipitation) (log mm precipitation)
Estimated mean µ̂X = 1.573 µ̂Y = 1.314
2
Estimated variance σ̂X = 0.333 σ̂Y2 = 0.171
Number of observations nX = 35 nY = 30

We now want to test whether the means of the 2 groups can be assumed equal, i.e.

H0 : µX = µY
H1 : µX 6= µY

It is given that the usual test statistics assuming the null hypothesis becomes 2.0958 with 61.19
degrees of freedom. State the p-value and conclusion when a significance level of α = 0.05 is
applied:

1  p-value ' 0.82 i.e. H0 is accepted

2  p-value ' 0.41 i.e. H0 is rejected

3  p-value ' 0.21 i.e. H0 is accepted

4  p-value ' 0.10 i.e. H0 is rejected

5*  p-value < 0.05 i.e. H0 is rejected

----------------------------------- FACIT-BEGIN -----------------------------------

This is a two-sample t-test and we get the information we need from tobs = 2.0958 and degrees
of freedom is 61.19, so the p-value is calculated by

2 * (1-pt(abs(2.0958), df=61.19))

## [1] 0.04024393

which is lower than 0.05, so we reject the null hypothesis.

42
------------------------------------ FACIT-END ------------------------------------

Continues on page 43

43
 Exercise XVIII

At a Christmas marked there is a lottery. 24 balls are placed in bowl. On each of 4 balls there
is a picture of a star. On each of the remaining 20 balls there is a picture of an elf. The lottery
is now played so that 2 balls are drawn without replacement from the bowl. If both balls show
a picture of a star then you have won a prize!

Question XVIII.1 (30)

You participate in the game once. Provide the probability of winning a prize:

80
1 276
56
2 276
40
3 276
16
4 276
6
5*  276

----------------------------------- FACIT-BEGIN -----------------------------------

This is drawing without replacement, hence we must use the hypergeometric distribution (Chap-
ter 2.3.2). However, to get most easily to the answer in the presented form, we can use the
basic definition of probability
x
P (success) = ,
n
where x is the number of successes in a population of size n. We need possible successful
combinations, where a ball with a star is drawn. In the first draw one out of the four must be
drawn and in the second draw one out of the three remaining must be drawn, thus

x = 4 · 3 = 12.

The number of elements in the population (of possible draws) is

n = 24 · 23 = 552,

since in the first draw there are 24 balls and in the second there are one less. Put together this
gives
12 6
= .
552 276
Alternatively, the x number of successful combinations could be calculated by

44
dhyper(x=2, m=4, n=20, k=2)

## [1] 0.02173913

which multiplied with the population size gives x

dhyper(x=2, m=4, n=20, k=2) * (24*23)

## [1] 12

------------------------------------ FACIT-END ------------------------------------

The exam is finished. Have a great Christmas vacation!

45