Data Analysis and Decision Making

Alois Geyer

VIENNA UNIVERSITY OF ECONOMICS AND BUSINESS

+43 1 31336 4559
alois.geyer@wu.ac.at
http://www.wu.ac.at/~geyer
Contents
1 Introduction 1

2 Describing data - Descriptive statistics 2

2.1 Types of data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Measures of location – mean, median and mode . . . . . . . . . . . . 5
2.3 Measures of dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Describing the distribution of data . . . . . . . . . . . . . . . . . . . 12
2.4.1 Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.2 Skewness and kurtosis . . . . . . . . . . . . . . . . . . . . . . 13
2.4.3 Rules of thumb . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.4 Empirical quantiles . . . . . . . . . . . . . . . . . . . . . . . . 15

3 How likely is . . . ? – Some theoretical foundations 17

3.1 Random variables and probability . . . . . . . . . . . . . . . . . . . 17
3.2 Conditional probabilities and independence . . . . . . . . . . . . . . 17
3.3 Expected value, variance and covariance of random variables . . . . 18
3.4 Properties of the sum of random variables . . . . . . . . . . . . . . . 20

4 How likely is . . . ? – Some applications 22

4.1 The normal distribution . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 How likely is a value less than or equal to y∗? . . . . . . . . . . . . . 23
4.3 Which value of y is exceeded with probability 1−α? . . . . . . . . . 24
4.4 Which interval contains a pre-specified percentage of cases? . . . . . 25
4.5 Estimating the duration of a project . . . . . . . . . . . . . . . . . . 28
4.6 The lognormal distribution . . . . . . . . . . . . . . . . . . . . . . . 30
4.7 The lognormal distribution . . . . . . . . . . . . . . . . . . . . . . . 30
4.8 Value-at-Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.9 Value-at-Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.10 The binomial distribution . . . . . . . . . . . . . . . . . . . . . . . . 34

5 How accurate is an estimate? 37

5.1 Samples and confidence intervals . . . . . . . . . . . . . . . . . . . . 37
5.2 Sampling procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

i
6 Describing relationships 51
6.1 Covariance and correlation . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 Simple linear regression . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.3 Regression coefficients and significance tests . . . . . . . . . . . . . . 58
6.4 Goodness of fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.5 Estimating the CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.6 Multiple regression analysis . . . . . . . . . . . . . . . . . . . . . . . 63

7 Decision analysis 68
7.1 Elements of a decision model . . . . . . . . . . . . . . . . . . . . . . 69
7.2 Decisions under uncertainty . . . . . . . . . . . . . . . . . . . . . . . 71
7.3 Decisions under risk . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.4 Decision trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

8 References 79

9 Exercises 80

10 Cases 83
10.1 Olson Diversified Marketing Services – Analysis of Receivables . . . 83
10.2 Production System Inc. – Development of a Salary Model . . . . . . 86
10.3 American Motors – Predicting Fuel Economy and Price . . . . . . . 89
10.4 ACME Marketing Strategy – A Multistage Decision Problem . . . . 92

11 Symbols and short definitions 94

ii
 1

1 Introduction
The term statistics often refers to quantitative information about particular sub-
jects or objects (e.g. unemployment rate, income distribution, . . . ). In this text the
term statistics is understood to deal with the collection, the description and the
analysis of data. The objective of the text is to explain the basics of descriptive
and analytical statistics.
The purpose of descriptive statistics is to describe observed data using graphics,
tables and indicators (mainly averages). It is frequently necessary to prepare or
transform the raw data before it can be analyzed. The purpose of analytical statistics
is to draw conclusions about the population on the basis of the sample. This
is mainly done using statistical estimation procedures and hypothesis tests. The
population consists of all those elements (e.g. people, companies, . . . ) which share
a feature of interest (e.g. income, age, height, stock price, . . . ). A sample from the
population is drawn if the observation of all elements is impossible or too expensive.
The sample is used to draw conclusions about the properties of that feature in the
population. Such conclusions may be used to prepare and support decisions.

Excel contains a number of statistical functions and analysis tools. This text
includes short descriptions of selected Excel-functions1 .
The menu ’Tools/Data Analysis’2 contains the item ’Descriptive Statistics’3 .
Upon activating ’Summary Statistics’4 a number of important sample statis-
tics are computed. All results can be obtained using individual functions,
too.
If the entry ’Data Analysis’ is not available, use the add-in manager (available
under ’Tools’) to activate ’Data Analysis’.

Many examples in this text are taken from the book ”Managerial Statistics” by
Albright, Winston and Zappe (AWZ) (www.cengage.com). The title of the third
edition is ”Data Analysis and Decision Making”. This book can be recommended as
a source of reference and for further study. It covers the main areas of (introductory)
statistics, it includes a large variety of (practically relevant) examples and cases, and
is strongly tied to using Excel.
1
Descriptions of the functions are provided in English. Function names will be specified in
English and German.
2
’Extras/Analyse-Funktionen’
3
’Populationskenngrößen’
4
’Statistische Kenngrößen’
2.1 Types of data 2

Figure 1: Summary statistics for the variable ’Salary’.

Salary

Mean 52,263
Standard Error 2,098
Median 50,800
Mode 62,000
Standard Deviation 11,493
Sample Variance 132,081,023
Kurtosis 3.56
Skewness 0.64
Range 50,400
Minimum 31,000
Maximum 81,400
Sum 1,567,900
Number of Observations 30

2 Describing data - Descriptive statistics

2.1 Types of data
Example 15 : The sheet ’coding’ represents responses from a question-
naire concerning environmental policies. The data set includes data on
30 people who responded to the questionnaire. As an example Figure 1
contains summary statistics for the variable ’Salary’ which will be de-
scribed below.

A sample usually consists of variables (e.g. age, gender, state, children, salary, opin-
ion) and observations (the record for each person asked). Samples can be categorized
either as cross-sectional data or time series data. Cross-sectional data is collected
at a particular point of time for a set of units (e.g. people, companies, countries,
etc.). Time series data is collected at different points in time (in chronological order)
as, for instance, monthly sales of one or several products.
Important categories of variables are numerical and categorical. Numerical (car-
dinal or metric) data such as age and salary can be subject to arithmetics. Numerical
variables can be subdivided into two types – discrete and continuous. Discrete
data (e.g. the number of children in a household) arises from counts whereas con-
tinuous data arises from continuous measurements (e.g. salary, temperature).
It does not make sense to do arithmetics on categorical variables such as gender,
state and opinion. The opinion variable is expressed numerically on a so-called
Likert scale. The numbers 1–5 are only codes for the categories ’strongly disagree’,
’disagree’, ’neutral’, ’agree’, and ’strongly agree’. However, the data on opinion
5
Example 2.1 on page 29 in AWZ.
2.1 Types of data 3

implies a general ordering of categories that does not exist for the variables ’Gender’
and ’State’. Thus opinion is called an ordinal variable. If there is no natural
ordering variables are classified as nominal (e.g. gender or state). Both ordinal and
nominal variables are categorical.
Some categorical variables can be coded numerically (e.g. male=0, female=1). For
some types of analyses recoding may be very useful (e.g. the mean of 0-1 data on
gender is equal to the percentage of women in the sample).
A special type of data are returns which are mainly used in the context of financial
economics.6 There are several possibilities to compute returns from stock or bond
prices (or indices). Log returns are computed on the basis of changes in the
logarithm of prices or indices:

pt
log return: yt = ln pt − ln pt−1 = ln .
pt−1

’ln’ is the natural logarithm and pt is the price or the value of the index at time t.
This definition corresponds to continuous compounding.
Simple returns are computed on the basis of relative price changes:

pt − pt−1 pt
simple return: rt = = − 1.
pt−1 pt−1

This definition corresponds to discrete compounding. It implies that interest accrues

at distinct points in time.
To explain the relation between simple and log returns we consider an investment
with initial value of p0 =100 which increases to p1 =105 within one year. The simple
return is r=5%=0.05. We now decompose this time interval into two steps (m=2),
and look for the simple (annual) return which has to be applied two times to yield
the same terminal value. This return can be computed from

" 1/2 # " 1/2 #

p1 105
r2∗ =2· −1 =2· − 1 = 0.04939.
p0 100

If this return is used to obtain the value of the investment after one year – but
interest accrues twice a year – we obtain

p1 = p0 · (1 + r2∗ /2) · (1 + r2∗ /2) = p0 · (1 + r2∗ /2)2 = 100 · (1 + 0.04939/2)2 = 105.

6
The rest of this subsection is only relevant for banking, finance or similar courses.
2.1 Types of data 4

Using six steps within a year, we obtain r6∗ =0.048989. In general, the implied simple
return for compounding m times within a year is given by

" 1/m #
∗ p1
rm =m −1 .
p0

As m goes to infinity (i.e. considering an infinite number of time steps within a

∗ converges to the log return y (which is 0.04879 in the present example).
year), rm
Using the log return the value of investment grows according to

p1 = p0 · exp{y} = 100 · exp{0.04879} = 105.

Table 1 shows the results of computing log and simple returns using one year from
the sample7 . The log return from December 1994 to January 1995 is computed from

ln 2021.27 − ln 2106.58 = 7.611481 − 7.652821 = −0.04133975 = −4.133975%.

The simple return is computed from

2021.27 − 2106.58 −85.31

= = −0.0404969 = −4.049692%.
2106.58 2106.58
Table 1 shows that yt and rt differ only slightly. However, there is a systematic dis-
crepancy between the mean of the two returns that will be analyzed more thoroughly
in the next section.
The simple return of a portfolio of m assets is a weighted average of the simple
returns of individual assets:

m
X
rpt = wi · rit
i=1

where wi is the weight of asset i in the portfolio. For log returns this relation only
holds approximately:

m
X
ypt ≈ wi · yit .
i=1
7
Returns will be expressed in percentage terms. Therefore some statistics based on returns will
be interpreted as percentage or percentage points. However, the percentage sign will typically be
omitted in the rest of the text.
2.2 Measures of location – mean, median and mode 5

Table 1: Comparing log and simple returns of the DAX.

Date t DAX (pt ) ln pt pt –pt−1 yt (%) rt (%)
12.94 0 2106.58 7.652821 . . .
01.95 1 2021.27 7.611481 –85.31 –4.133975 –4.049692
02.95 2 2102.18 7.650730 80.91 3.924887 4.002929
03.95 3 1922.59 7.561429 –179.59 –8.930167 –8.543036
04.95 4 2015.94 7.608841 93.35 4.741235 4.855429
05.95 5 2092.17 7.645957 76.23 3.711622 3.781363
06.95 6 2083.93 7.642011 –8.24 –0.394627 –0.393849
07.95 7 2218.74 7.704695 134.81 6.268393 6.469027
08.95 8 2238.31 7.713476 19.57 0.878165 0.882032
09.95 9 2187.04 7.690304 –51.27 –2.317209 –2.290567
10.95 10 2167.91 7.681519 –19.13 –0.878546 –0.874698
11.95 11 2242.83 7.715494 74.92 3.397489 3.455863
12.95 12 2253.88 7.720408 11.05 0.491471 0.492681
arithmetic mean: 0.563228 0.648957
geometric mean: 0.564817

2.2 Measures of location – mean, median and mode

The most important statistical measure is the (arithmetic) mean. Given n obser-
vations y1 , . . . , yn the mean is defined by

n
1X
arithmetic mean: ȳ = yt .
n
t=1

ȳ is the average of the data. In statistics ȳ is called estimate. It is estimated from

the sample y1 , . . . , yn . This terminology applies to all statistics introduced below
which are ’computed’ from observed data.

The arithmetic mean can be computed using the function AVERAGE(data

range)8 .

The mean is only meaningful for numerical data. In example 1 the average salary ȳ
equals $52,263.
The median is the value in the middle of a sorted sequence of data.9 Therefore 50%
of the cases are less than (or greater than) the median. The median can be used for
8
MITTELWERT(data range)
9
If there is an even number of cases the median is the mean of the two values in the middle of
the sequence.
2.2 Measures of location – mean, median and mode 6

numerical or ordinal data. The median is not affected by extreme values (outliers)
in the data. For instance, the sequence 1, 3, 5, 9, 11 has the same median as –11, 3,
5, 9, 11. The means of these two samples differ strongly, however.

The median can be computed using the function MEDIAN(data range).

In example 1 the median is $50,800. Half of the respondents earn more than this
number, and the other half earns less than that. The mean and the median salaries
are very similar in this example. Therefore we conclude that salaries are distributed
symmetrically around the center of the data. Since the median is slightly less than
the mean we conclude, however, that a few salaries are relatively high.
The mode is the most frequent value in a sample. Similar to the median, the mode
is not affected by extreme values. It can be interpreted as a ’typical’ salary under
’normal’ conditions.
The mode is typically applied to recoded nominal data or discrete data. For example,
if each state is coded using a different number, the mode identifies the most frequent
state. If the variable is continuous (e.g. temperature) the mode may not be defined.
In very small samples or when the data is measured very precisely it may be that no
value occurs more than once10 . Such is the case with salaries in the present example.
This can happen because the sample is too small or the accuracy of coding is too
high. This problem may be overcome by computing the mode of rounded values.
The mode of rounded salaries equals $62,000.

The mode can be computed using the function MODE(data range)11 . The
function returns #NV if the data range contains not a single number, that
appears more than once. This can be avoided by using rounded values.

Example 212 : Consider the data on sheet ’Shoes’ – the shoe sizes pur-
chased at a shoe store. We seek to find the best-selling shoe size at this
store. Shoe sizes come in discrete increments, rather than a continuum.
Therefore it makes sense to find the mode, the size that is requested
most often. In this example it turns out that the best-selling shoe size
is 11.

Example 3: This example is based on the study ”Growth in a Time

of Debt” by Reinhart and Rogoff (American Economic Review, 2010:
Papers&Proceedings, 100). This study had considerable impact on the
debate and on political decisions made during the global debt crisis after
10
In Excel this case is indicated by #NV.
11
MODALWERT(Datenbereich)
12
Example 3.2 on page 76 in AWZ.
2.2 Measures of location – mean, median and mode 7

2008. It has received even more attention after the student Thomas Hern-
don had found several mistakes13 in the Excel sheet used by Reinhart
and Rogoff (RR). This example only focuses on one particular aspect:
the potentially misleading results of computing ”means of means”. RR
have excluded some data for some countries in some years without justi-
fication. The impact of this exclusion is ignored for the present purpose
and the reduced dataset is used. The frequently mentioned ”Excel cod-
ing error” in the media refers to the use of a wrong cell range in their
spreadsheet, whereby five countries have been excluded. This mistake is
not evaluated here.

RR have tried to analyze the relation between national debt levels and
GDP growth rates. For that purpose they have classified debt (the ratio
of public debt to GDP) into four categories. They first compute the aver-
age GDP growth rates for each country in each category. Subsequently,
they compute the growth rate for each category by averaging averages
across countries. Thereby, they identify a sharp drop in growth rates
to 0.3% for debt ratios above 90% compared to 3-4% growth rates for
lower debt ratios (see the copy of Figure 2 from RR’s study in Figure 3).
This piece of evidence had a key impact on the policy recommendation
derived from their study.

However, it should be noted that computing the mean of averages

may have unintended side-effects, because of the implicit weighting asso-
ciated with this way to proceed. For example, the average 0.3% for the
category above 90% debt is based on the average growth of UK based
on 19 years in which UK’s debt ratio was above 90%, as well as on the
average growth of the US, whose debt ratio has been assigned to this cat-
egory in four years. However, for the average of averages each of those
means has the same weight, which implies that the four years for the
US are equally important as the 19 years of the UK. If the growth rate
of each country and year enters with the same weight, the average of in
this debt category is 1.9%, substantially above the 0.3% obtained and
reported by RR. This implicit weighting scheme may or may not agree
with the purpose and intentions of the analysis. In any case, conclusions
derived from computing averages of averages should be interpreted with
care and awareness of such a fact.
13
http://www.peri.umass.edu/fileadmin/pdf/working_papers/working_papers_301-350/
WP322.pdf provides a detailed account of all mistakes documented by Herndon and his thesis
advisors.
2.2 Measures of location – mean, median and mode 8

Figure 2: Key results as reported in the study by Reinhart and Rogoff (2010).

We now consider simple and log returns of the DAX in more detail.14 The arithmetic
mean of DAX log returns using the entire sample is 0.56. This implies that an
investment in the DAX yields – on average – a monthly interest rate of slightly more
than one half percent. The mean using all simple returns in the sample is 0.73,
which implies a much higher average interest rate.
We use the twelve log and simple returns from Table 1 to analyze this discrepancy
more thoroughly. The mean log return is

12
1 X
yt = 0.563228.
12
t=1

The average simple return is much larger:

12
1 X
rt = 0.648957.
12
t=1

This discrepancy not only holds in the present example but holds in general.
14
The rest of this subsection is only relevant for banking, finance or similar courses.
2.2 Measures of location – mean, median and mode 9

Which average is correct? The mean return over a period should reflect the actual
growth rate of the stock or index. According to financial calculus the internal rate
of return – assuming discrete compounding – can be computed from

 1/t
∗ pt
i = − 1.
p0

In case of continuous compounding the internal rate of return is given by

i∗ = (ln pt − ln p0 )/t.

Over the twelve-month period in this example the rate of return is either given by

 1/12
2253.88
− 1 = 0.564817
2106.58

if discrete compounding is assumed or

ln 2253.88 − ln 2106.58
= 0.563228
12
if continuous compounding is assumed. The (minor) difference between these two
values is due to the different assumptions about compounding.
This shows that the average of log returns correctly reflects the change in value,
whereas the average of simple returns systematically overstates the actual change in
value. We conclude that the arithmetic mean of log returns is an unbiased measure
of the average, whereas the arithmetic mean r̄ of simple returns (obtained from the
same price series) is biased upwards. To obtain a correct measure for the mean of
simple returns one needs to calculate the geometric mean:

geometric mean: [(1 + r1 ) · (1 + r2 ) · · · (1 + rn )]1/n − 1.

Using the twelve simple returns from Table 1 one obtains:

12
!1/12
Y
(1 + rt ) − 1 = 0.564817
t=1

which agrees with financial calculus.

2.3 Measures of dispersion 10

The geometric mean can be computed using the function15 GEOMEAN(data

range)–1. Note that the data range must contain the gross simple returns
1+rt .

2.3 Measures of dispersion

Example 416 : Suppose that Otis Elevator is going to stop manufac-
turing elevator rails. Instead, it is going to buy them from an outside
supplier. Two suppliers are considered. Otis has obtained samples of
ten elevator rails from each supplier which should have a diameter of
2.5cm. Because of unavoidable, random variations in the production
process this request cannot be fulfilled in each case. But the rails should
deviate as little as possible from 2.5cm. The sheet ’otis’ lists the data
from both suppliers and should be used to support the choice among the
two suppliers?

As it turns out the mean, median and mode of both suppliers are identical to 2.5
cm. Based on these measures, the two suppliers are equally good and right on
the mark. Thus we require an additional measure for reliability or variability that
allows Otis to distinguish among the suppliers. A look at the data shows that the
variability of diameters from supplier 2 around the 2.5cm mean is greater than that
of supplier 1. This visual impression can be expressed in statistical terms using
measures of dispersion (around the mean).
The mean (or other measures of location) is insufficient to describe the sample,
since it must be taken into account, that individual observations may deviate more
or less strongly from the mean. The degree of dispersion can be measured with the
standard deviation s . The standard deviation is based on the variance s2 which
is computed as follows:

n
1 X
variance: s2 = (yt − ȳ)2 .
n−1
t=1

The essential feature of this formula is the focus on deviations from the mean. Taking
squares avoids that positive and negative deviations from the mean cancel out (the
sum or average of deviations from the mean is always zero!).
The standard deviation is a measure for the (average) dispersion around the mean.
The advantage of using the standard deviation rather than the variance is the fol-
lowing: s has the same units of measurement as yt and can therefore be more easily
15
GEOMITTEL(Datenbereich)
16
Example 3.3 on page 78 in AWZ.
2.3 Measures of dispersion 11

Table 2: Computing variance and standard deviation for Supplier 2.

yt yt − ȳ (yt − ȳ)2
2.400 –0.100 0.010000
2.625 0.125 0.015625
2.500 0.000 0.000000
2.425 –0.075 0.005625
2.500 0.000 0.000000
2.575 0.075 0.005625
2.450 –0.050 0.002500
2.550 0.050 0.002500
2.375 –0.125 0.015625
2.600 0.100 0.010000
sum 25.000 0.0 0.067500
mean 2.500 0.0 0.006750
variance: 0.0075
standard deviation: 0.0866

interpreted. The squared units of variance inhibit a simple and straightforward

interpretation as a measure of dispersion.

Variance and standard deviation can be computed using the functions VAR(data
range) and STDEV(data range)17 .

Table 2 shows the computation of variance and standard deviation using data from
supplier 2. The variance is given by 0.0075. This number cannot be easily inter-
preted 2
√ since it is measured in squared units of y (cm ). The standard deviation
s= 0.0075=0.0866 can be interpreted as the average dispersion of yt around its
mean measured in cm. Note however, that this is not a simple average. Because
of the square in the definition of the variance, large deviations from the mean are
weighed more strongly than small deviations.
The coefficient of variation g=s/ȳ – the ratio of standard deviation and mean – is
a standardized measure of dispersion. It is used to compare different samples. The
coefficient of variation is frequently interpreted as a percentage. For the variable
’salary’ in example 1 g=11,493/52,263=0.22: on average, salaries deviate from the
mean by 22%.
To obtain a complete picture of the dispersion of the data it is useful to compute
the minimum, the maximum and the range – the difference between minimum and
maximum. The range for supplier 2 is given by 0.25 which is much larger than the
17
VARIANZ(Datenbereich); STABW(Datenbereich)
2.4 Describing the distribution of data 12

0.1 range of supplier 1. The range, the minimum and maximum again show that
the deliveries of supplier 2 are less reliable.

2.4 Describing the distribution of data

Example 5: Consider monthly prices and returns of the DAX from
January 1986 through December 1996 (sheet ’DAX’). The return is the
monthly percentage change in the index. We want to describe the dis-
tribution of the returns.

2.4.1 Histogram

A histogram is used to draw conclusions about the distribution of observed data.

In particular, the purpose is to find out whether the data can – at least roughly – be
described by a normal distribution. A normal distribution is assumed in many appli-
cations and in many statistical tests. Further details about the normal distribution
are explained in section 4.1.
Suppose the observations in the sample are assigned to a set of prespecified categories
(intervals). A good choice are about 10–25 categories plus a possible open-ended
category at either end of the range. The number of cases in each interval is di-
vided by the total number of observations in the sample. This ratio is the relative
frequency. The bar chart of relative frequencies is the so-called histogram.

The menu ’Tools/Data analysis’18 contains the item Histogram19 . The in-
tervals are automatically selected if the field ’Bin Range’20 is left empty.
Note that the function computes absolute rather than relative frequen-
cies! Absolute frequencies can also be computed using the function FRE-
QUENCY(data array;bins array)21 .

Example 6: The histogram in Figure 3 shows that the range from -2.5
to 0.0 contains 18.3% and the interval [2.5,5.0] contains 19.1% of monthly
returns. 11.5% of the returns are less than –5.0. 39.8% of all returns
are negative. This percentage is obtained by summing up the relative
frequencies in all intervals from –25.0 to 0.0.

18
’Extras/Analyse-Funktionen’
19
Histogramm
20
’Klassenbereich’
21
HÄUFIGKEIT(Datenbereich;Klassenbereich)
2.4 Describing the distribution of data 13

Figure 3: Histogram of monthly DAX returns and normal density.

25.0%

20.0%
empirical normal

15.0%

10.0%

5.0%

0.0%

and greater
-22.5

-17.5

-12.5

-7.5

-2.5
-25

-20

-15

-10

-5

2.5

7.5

12.5

17.5

22.5
0

10

15

20

25
2.4.2 Skewness and kurtosis

As already mentioned the normal distribution plays an important role in statis-

tics and various applications. The following two measures can be used to indicate
deviations from normality.
The skewness

n
1 X (yt − ȳ)3
skewness:
n s3
t=1

is a measure of the histogram’s symmetry. It is an indicator and has no units of

measurement. A normal distribution is symmetrical and has a skewness of zero. If
the skewness is negative, the left tail of the histogram is flatter (or longer) than the
right tail. A distribution with negative (positive) skewness is said to be skewed to the
left (right). Simply speaking, when the skewness is negative there are more negative
extremes than positive extremes (more precisely: extremely large negative deviations
from the mean are more frequent and/or more pronounced than the positive ones).
If a distribution is skewed, mean, median and mode are not identical. It is possible,
however, to say something about their order. If the skewness is positive the mode
is less than the median, and the median is less than the mean. The converse is true
in case of a negative skewness.
A second important measure for the shape of the histogram is the kurtosis
2.4 Describing the distribution of data 14

n
1 X (yt − ȳ)4
kurtosis: .
n s4
t=1

The kurtosis is an indicator and has no units of measurement. The kurtosis of a

normal – bell-shaped – distribution equals three. Thus a kurtosis different from 3
indicates a deviation from a ’normal’ shape. The data is said to have a leptokurtic
distribution if it is strongly concentrated around the mean and there is a relatively
high probability to observe extreme values on either side (so-called fait tails). This
property holds when the kurtosis is greater than 3. A kurtosis less than 3 indicates
a platykurtic distribution which is not strongly concentrated around the mean.

The skewness can be computed using the function SKEW(data range)22 .

The kurtosis can be computed using the function 3+KURT(data range).
Adding the value 3 is necessary to obtain results that agree with the formula
above.

Example 7: The sample skewness of DAX returns equals –1.0 which

indicates that negative extremes are more likely than positive extremes.
This agrees with the histogram in Figure 3. The sample kurtosis of
monthly DAX returns equals 6.2 which strongly indicates that DAX
returns are not normally distributed but leptokurtic.

Example 823 : The sheet ’arrival’ lists the time between customer ar-
rivals – called interarrival times – for all customers in a bank on a given
day. The skewness of interarrival times is given by 2.2. This indicates
a distribution which is positively skewed, or skewed to the right. The
skewed distribution can also be seen from a histogram of the data. Most
interarrival times are in the range from 2 to 10 minutes but some are
considerably larger. The median (2.8) is not affected by extremely large
values. Consequently, it is lower than the mean (4.2).

2.4.3 Rules of thumb

The distribution of many data sets can be described by the following ”rules of
thumb”.

1. Approximately two thirds of the observations are in a range plus/minus one

standard deviation around the mean.
22
SCHIEFE(Datenbereich)
23
Example 2.4 on page 37 in AWZ.
2.4 Describing the distribution of data 15

2. Approximately 95% of the observations are in a range plus/minus two standard

deviations around the mean.

3. Almost all observations are in a range plus/minus three standard deviations

around the mean.

Example 9: Applying the rules of thumb to the DAX returns (see Fig-
ure 4) shows that only the second rule seems to work. The empirical
(relative) frequencies and the probabilities based on the normal distri-
bution are very close. The discrepancies observed for the first and third
rule may be explained by the leptokurtosis of returns.

Figure 4: Rules of thumb and relative frequencies of DAX returns.

rules of thumb
absolute relative
from to frequency frequency
more than 3 std.dev's below mean -3 −∞ -16.93 2 1.5%
between 2 and 3 std.dev's below mean -2 -16.93 -11.10 2 1.5%
between 1 and 2 std.dev's below mean -1 -11.10 -5.27 10 7.6%
between mean and 1 std.dev below mean 0 -5.27 0.56 46 35.1%
between mean and 1 std.dev above mean 1 0.56 6.39 54 41.2%
between 1 and 2 std.dev's above mean 2 6.39 12.22 14 10.7%
between 2 and 3 std.dev's above mean 3 12.22 18.05 3 2.3%
more than 3 std.dev's above mean 18.05 +∞ 0 0.0%

rule of thumb exact actual

1. within one std.dev around the mean 66.7% 68.3% 76.3%
2. within two std.dev's around the mean 95% 95.4% 94.7%
3. within three std.dev's around the mean ~100% 99.7% 98.5%

2.4.4 Empirical quantiles

In order to compute an empirical quantile (or percentile) a relative frequency α

is chosen. The α-quantile divides the data set such that α percent of the observa-
tions are lower than the α-quantile and (1−α) percent are larger than the quantile.
The median is the 50%-quantile. The quantile need not correspond to an actually
observed value in the sample. However, it has the same units of measurements as
the observed data.

Empirical quantiles can be computed using the function PERCENTILE(data

range; α)24 .
24
QUANTIL(Datenbereich; α)
2.4 Describing the distribution of data 16

Example 10: The empirical 1%-quantile of DAX returns equals –18.9;

i.e. one percent of the returns are less than –18.9. The 5%-quantile equals
–8.9. Quantiles for small values of α can be used as measures of risk.

Example 11: Consider the variable ’Salary’ from example 1 again. The
empirical 25%-quantile of salaries is given by $44,675. In other words,
25% of the respondents earn less than $44,675. The 75%-quantile is
$59,675, so 25% of the respondents earn more than $59,675.
3.1 Random variables and probability 17

3 How likely is . . . ? – Some theoretical foundations

3.1 Random variables and probability
In statistics it is usually assumed that observed values are realizations of random
variables. This term is based on the view that there are so called random exper-
iments with specific outcomes. It is uncertain which of the possible outcomes will
take place. The randomness is due to the fact that the outcome cannot be predicted.
A frequently used example is the experiment of throwing dice. On which side the
die will fall – the outcome – is random, or is assumed to be random.
A random variable Y assigns real numbers y to each outcome of a (random) exper-
iment. The number y is a realization of the random variable. In the dice throwing
example there are six possible realizations: (y1 =1),. . . ,(y6 =6).
Probability is a measure for the (un)certainty of an outcome. The probability that
a random variable equals a specific value yi is denoted by P[yi ]=pi . Probabilities
have to satisfy two conditions: they must not be negative and the sum over all
possible realizations must be equal to 1.
The law (or function) that defines probabilities is the probability distribution
of a random variable. Probability distributions can be based on (a) theoretical
(objective) considerations, (b) a large number of experiments, or (c) subjective as-
sumptions. In the example of the die, the first theoretical considerations lead to
pi =1/6 for each of the possible realizations yi . The second, experimental foundation
involves throwing dice e.g. 100 times and to count each of the six possible outcomes.
The resulting probabilities are given by pi =ni /100, where ni is the number of cases
where yi =i. Subjective probabilities are based on intuition or experience.

3.2 Conditional probabilities and independence

It is important to distinguish unconditional from conditional probabilities (and
probability distributions). The former make statements about experimental out-
comes irrespective of any conditions that (may) affect the results of the experiment.
Conditional probabilities P[y|x] take into account the condition x under which the
experiment is carried out.
The need to distinguish unconditional and conditional probabilities depends on
the case at hand. For instance, if the probability to find a person with a job of
type A is different for men and women, the unconditional probability P[job=’A’]
is rather meaningless whereas the conditional probabilities P[job=’A’|man] and
P[job=’A’|woman] are clearly more informative. On the other hand, in the dice
rolling experiment the conditional probability to observe a particular outcome un-
der the condition that a particular outcome was observed in the previous experiment
should (theoretically) not differ from the unconditional probability: P[yt =i|yt−1 ] =
3.3 Expected value, variance and covariance of random variables 18

P[yt =i]. A conditional viewpoint does not appear to be necessary in this or simi-
lar cases. An empirical analysis may be used to find out, whether conditional and
unconditional probabilities differ.
The relation between unconditional and conditional probability is used to define
independence. The two random variables Y and X are said to be independent if
P[Y |X]=P[Y ].

3.3 Expected value, variance and covariance of random variables

The expected value of the random variable Y is given by

n
X
expected value: µ = E[Y ] = pi · yi ,
i=1

where n is the number of possible realizations.

The expected value for throwing dice is given by (1/6)·1+(1/6)·2+· · · +(1/6)·6)=3.5.
If a fair die is thrown a very large number of times the sample average should be
close to 3.5.
The variance of Y is given by

n
X
2 2
variance: σ = var[Y ] = E[(Y − µ) ] = pi · (yi − µ)2 .
i=1

As another example we consider two investments where profits are assumed to de-
pend on the so-called ’state of the world’ (or economy). For each of the possible
states (’bad’, ’medium’ and ’good’) a probability and a profit/loss can be specified:

investment 1 investment 2
squared squared
state of deviation deviation deviation deviation
'the world' pi profit/loss from µ from µ profit/loss from µ from µ
bad 0.2 -180 -209 43681 -10 -25.5 650.25
medium 0.5 10 -19 361 5 -10.5 110.25
good 0.3 200 171 29241 50 34.5 1190.25
exp.value µ 29 15.5
variance σ² 17689.0 542.3
std.dev. 133.0 23.3

The expected value (expected profit) of investment 1 can be computed as follows:

µ1 = −180 · 0.2 + 10 · 0.5 + 200 · 0.3 = 29.

3.3 Expected value, variance and covariance of random variables 19

The variance25 is based on the squared deviations from the expected value:

σ12 = (−180 − 29)2 · 0.2 + (10 − 29)2 · 0.5 + (200 − 29)2 · 0.3 = 17689.

The covariance between two random variables Y and X is given by:

n
X
covariance: cov[Y, X] = E[(Y − µY ) · (X − µX )] = pi · (yi − µY ) · (xi − µX ),
i=1

where pi is the (joint) probability that Y = yi and X = xi . The correlation

between Y and X is given by the ratio of the covariance and the product of the
standard deviations:

cov[Y, X]
correlation: corr[Y, X] = .
σY σX

The correlation is bounded between –1 and +1. Mean and (co)variance are also
called first and second moments of random variables.
Consider throwing a pair of dice. There are 36 possible realizations which are all
equally likely: [y1 =1,x1 =1], [y1 =1,x2 =2],. . . , [y6 =6,x6 =6]. As expected, the covari-
ance between the resulting numbers is zero (pi is a constant equal to 1/36):

1
[(1 − 3.5)(1 − 3.5) + (1 − 3.5)(2 − 3.5) + · · · + (6 − 3.5)(6 − 3.5)] = 0.
36

If a pair of dice is thrown very often the empirical covariance (or correlation) between
the observed pairs of numbers should be close to zero.
If two random variables are normally distributed and their covariance is zero, the
two variables are said to be independent. For general distributions a covariance of
zero merely implies that the variables are uncorrelated. It is possible, however, that
(nonlinear) dependence prevails between the two variables.
The concept of conditional probability extends to the definition of conditional expec-
tation and (co)variance, by using conditional (rather than unconditional) probabili-
ties in the definitions above. For instance, if the conditional expected value E[Y |X]
is assumed to be a linear function of X it can be shown that E[Y |X] is given by:
25
√ Note that the variance is measured in units of squared profits. The standard deviation
17689=133 is measured in original (monetary) units.
3.4 Properties of the sum of random variables 20

cov[Y, X]
conditional expectation: E[Y |X] = E[Y ] + · (X − E[X]).
var[X]

This shows that a conditional viewpoint is necessary when the covariance between
Y and X differs from zero. In a regression analysis (see section 6) a sample is used
to determine if there is a difference between the conditional and the unconditional
expected value, and if it is necessary to take more than one conditions into account.

3.4 Properties of the sum of random variables

The expected value of the sum of two random variables X and Y is given by

E[X + Y ] = E[X] + E[Y ].

The expected value of a weighted sum is given by

E[a · X + b · Y ] = a · E[X] + b · E[Y ].

The expected value of the sum of n random variables is Y1 , . . . , Yn the sum of their
expectations:

E[Y1 + Y2 + · · · + Yn−1 + Yn ] = E[Y1 ] + E[Y2 ] + · · · + E[Yn−1 ] + E[Yn ].

The expected value of the sum of n random variables with identical mean µ equals
n · µ:

E[Y1 + Y2 + · · · + Yn−1 + Yn ] = n · µ if E[Yi ] = µ (i = 1, . . . , n).

The variance of the sum of two uncorrelated random variables X and Y is the sum
of their variances:

var[X + Y ] = var[X] + var[Y ].

The variance of the sum of n uncorrelated random variables is the sum of their
variances:
3.4 Properties of the sum of random variables 21

var[Y1 + Y2 + · · · + Yn−1 + Yn ] = var[Y1 ] + var[Y2 ] + · · · + var[Yn−1 ] + var[Yn ].

The variance of the sum of n uncorrelated random variables with identical variance
σ 2 is given by n · σ 2 :

var[Y1 + Y2 + · · · + Yn−1 + Yn ] = n · σ 2 if var[Yi ] = σ 2 (i = 1, . . . , n).

The variance of the sum of two correlated random variables is given by

var[X + Y ] = E[{(X − µX ) + (Y − µY )}2 ] = var[X] + var[Y ] + 2cov[XY ].

The variance of the sum of n correlated random variables is given by:

n
X n X
X
var[Y1 + Y2 + · · · + Yn−1 + Yn ] = var[Yi ] + cov[Yi , Yj ].
i=1 i=1 i6=j

As an example we assume that both investments mentioned above are realized and
we consider the sum of profit/loss in each state of the world. The covariance between
the two investments is given by

(–180–29)·(–10–15.5)·0.2+(10–29)·(5–15.5)·0.5+(200–29)·(40–15.5)·0.3=2935.5.

Since the covariance is not zero, the sum of the variances of the two investments is
not equal to the variance of the sums as shown in the following table:

computing
covariance properties of the sum
&correlation of both investments
product of squared
state of deviations profit/loss deviation
'the world' pi from µ inv1+inv2 from µ
bad 0.2 5329.5 -190 54990.25
medium 0.5 199.5 15 870.25
good 0.3 5899.5 250 42230.25
µ 44.5 24102 <=variance of the sum
covariance 2935.5 18231 <=sum of the variances
correlation 0.948 24102 <=sum of the variances+2xcovariance

If we deal with a weighted sum we have to make use of the following fundamental
property:
4.1 The normal distribution 22

var[a + Y ] = var[Y ] var[a · Y ] = a2 · var[Y ].

The variance of a weighted sum of uncorrelated random variables is given by

var[a · X + b · Y ] = a2 · var[X] + b2 · var[Y ].

The variance of a weighted sum of two correlated random variables is given by

var[a · X + b · Y ] = a2 · var[X] + b2 · var[Y ] + 2 · a · b · cov[XY ].

For any constant a (not a random variable) and random variables W, X, Y, Z the
following relations hold:

if Y = a · Z: cov[X, Y ] = a · cov[X, Z].

if Y = W + Z: cov[X, Y ] = cov[X, W ] + cov[X, Z].

cov[Y, a] = 0.

4 How likely is . . . ? – Some applications

4.1 The normal distribution
Many applications are based on the assumption of a normal distribution. The
shape of the normal distribution is determined by two parameters: mean µ and
variance σ 2 . Given values of µ and σ 2 the normal density (or density function of
a normal distribution) can be computed:

−(y − µ)2
 
1
f (y) = √ exp − ∞ ≤ y ≤ ∞.
σ 2π 2σ 2

For a particular range of values – e.g. between y1 and y2 – the area underneath the
density equals the probability to observe values within that range (see Figure 5).
Usually the normal distribution of a random variable Y is denoted by Y ∼N(µ, σ 2 ).
Ψα on the y-axis is the α-quantile under normality. It has the following property:

P[y ≤ Ψα ] = α,
4.2 How likely is a value less than or equal to y ∗ ? 23

Figure 5: Normal density curve.

.....................
.... ....
.... ...
... ...
.... ...
... ...
... ...
... ...
.. ...
. ...
... .......
... ... ...
... ... ....
... ... ....
P[y ≤ Ψα ] = α ... .
.
..
.
.. ... ...
... .... P[y1 ≤ y ≤ y2 ] ...
... .
.
. ... ... .......
... .
. ... ... .......
.
. ... ... .......
...
... .
.
... .
.
...
. .....
. .......
.. ....
... ....
...
... .. ... ............
... ... ... ......
... ... .........
...
... ... ..... ... .......... ....
.
...
... ... .... ... .........
... .....
.. ... ... ......
... ........... ... ... .......
..
...... . . .
. .
. .....
......
........... .. .... .... .........
......... ..... .... ... ... .............
............................................. ...
..
... ... ...
...................................... y
Ψα ȳ y1 y2

where P[ ] is the probability for the event in brackets. The area to the left of Ψα
equals α – the probability to observe values less than Ψα . This implies that Ψα is
exceeded with probability 1−α.
Assuming a normal distribution for a variable y having mean ȳ and standard de-
viation s allows to answer some interesting questions, as shown in the following
subsections.

Example 12: Assuming a normal distribution for monthly DAX returns

allows to approximate the histogram in Figure 3 on page 13. The dashed
line is the empirical normal density. Its shape is based on using the sam-
ple mean 0.56 (ȳ) and standard deviation 5.8 (s). Comparing the normal
density and the shape of the histogram shows whether the assumption of
a normal distribution is justified. In the present case the histogram can-
not be approximated very well. This confirms the discrepancies observed
by applying the rules of thumb, which are based on the normal distribu-
tion (see below). Returns close to the mean and at the tails are (much)
more frequent than expected under the normal distribution. The kurtosis
of monthly DAX returns was found to be 6.2. This discrepancy shows the
leptokurtic distribution of observed DAX returns. Despite this discrep-
ancy the normal assumption is frequently maintained, mainly because
of the simplifications that result in various applications (e.g. portfolio
theory and option pricing) and tests.

4.2 How likely is a value less than or equal to y ∗ ?

Example 1326 : ZTel’s personnel department is reconsidering their hir-
ing policy. Currently all applicants take a test and their hire or no-hire
26
Example 6.3 on page 254 in AWZ.
4.3 Which value of y is exceeded with probability 1−α? 24

decision depends partly on the results of the exam. The applicants scored
have been examined closely. They are normally distributed with a mean
of 525 and standard deviation of 55 (see sheet ’personnel’).
The hiring policy occurs in two phases: The first phase separates all
applicants into three categories: automatic accepts (exam score≥600),
automatic rejects (exam score≤425), and ”maybes”. The second phase
takes all the ”maybes” and uses their previous job experience, special
talents and other factors as hiring criteria.
ZTel’s personnel manager wants to calculate the percentage of applicants
who are automatic accepts and rejects, given the current policy.

ZTel’s question can be answered as follows. The percentage of rejected applicants

is the probability to observe scores less than or equal to 425. This probability
corresponds to the area under the normal density to the left of a prespecified value
y ∗ . As it turns out 3.5% of applicants are automatically rejected.

The function NORMDIST(y ∗ ; mean ȳ; standard deviation s;1)27 computes

the probability to observe values of a normal variable y (with mean ȳ and
standard deviation s) that are less than or equal to y ∗ .

To compute the percentage of accepted applications we need to find the probability

for scores above 600. We proceed by first computing the probability to observe
scores below 600 and then subtract this number from 100%. We find that 8.6% of
all applicants are accepted.

4.3 Which value of y is exceeded with probability 1−α?

Example 1428 : ZTel’s personnel manager also wants to know how to
change the standards in order to automatically reject 10% of all appli-
cants and automatically accept 15% of all applicants. How should the
scores be determined to achieve this goal?

Now the manager takes a reversed viewpoint. Rather than computing probabilities
he wants to pre-specify a probability and work out the corresponding threshold score
that is exceeded with that probability. These questions can be answered using the
α-quantile of a normal variable.

The function NORMINV(probability α; mean ȳ; standard deviation s) com-

putes the α-quantile Ψα of a normal variable y∼N(ȳ, s2 ).
27
NORMVERT(y ∗ ; mean ȳ; standard deviation s;1)
28
Example 6.3 on page 254 in AWZ.
4.4 Which interval contains a pre-specified percentage of cases? 25

Table 3: Selected quantiles of the standard normal distribution.

α (%) 0.1 0.5 1.0 2.5 5.0 10.0 50.0 90.0 95.0 97.5 99.0 99.9
zα –3.090 –2.576 –2.326 –1.960 –1.645 –1.282 0.0 1.282 1.645 1.960 2.326 3.090

The 10%-quantile is given by 455. This score is exceed with 90% probability. 10% of
the scores are below this score. To achieve a 15% acceptance rate we need to know
the 85%-quantile. This quantile is equal to 582 points and is exceeded in 15% of all
cases.

4.4 Which interval contains a pre-specified percentage of cases?

The computation of intervals is based on the quantiles of a standard normal
distribution – this is a normal distribution with mean 0 and variance 1. Some
frequently used quantiles of the standard normal distribution are given in Table 3.
These numbers can be used to make probability statements about a standard normal
variable z∼N(0, 1). For example, there is a probability of 2.5% to observe a value
of z which is less than –1.96. This is expressed as follows:

P[z ≤ −1.96] = 0.025 = 2.5%,

where P[ ] is the probability of the term in brackets. In general

P[z ≤ zα ] = α,

where zα is the α-quantile of the standard normal distribution.

The quantiles zα of standard normal distribution are computed with the

function NORMSINV(probability α)29

The standard normal quantiles can be used to compute quantiles and intervals for a
normal variable y having mean ȳ and variance s2 . The α-quantile of y is given by30

Ψα = ȳ + zα · s.
29
STANDNORMINV(Wahrscheinlichkeit α)
30
Ψα can be computed directly using the function NORMINV.
4.4 Which interval contains a pre-specified percentage of cases? 26

Figure 6: Standard normal distribution and 95% interval of z.

..................
..... ....
.... ...
... ...
.... ...
..
. ...
.... ...
...
... ...
... ...
... ...
... ...
...
... ...
... ...
.. ...
2.5% ... .
.
..
.
.. ...
...
...
2.5% ..
... .
.. ...
... ...
... ..
. ... ...
... ... ... ...
... .. ... ...
... .. ...
...
... ... ... ..
... .... ... ..
... .
. ..
. ..
... .. .
. ..
... .... ..... ... .. ...
... .... ...... ...
... .... ....
....... ....
... ............ .
. .
.. .
.
........ .. ... .
........... .. .... ....................
......... ..... .... . ... .. .......................
............................................. .. .... ... ..............................
z
−1.96 0 1.96

Example 15: The monthly DAX returns have mean ȳ=0.56 and stan-
dard deviation s=5.8. To get some idea about the magnitude of ex-
tremely negative returns one may want to compute the 1%-quantile. As-
suming that returns are normally distributed and using the 1%-quantile
of the standard normal distribution (−2.326) yields

0.56 − 2.326 · 5.8 = −12.9.

Thus, there is a 1% probability to observe returns which are less than

–12.9.
The 1%-quantile of DAX returns assuming a normal distribution is much
larger than the empirical 1%-quantile –18.9 (see page 16). This corre-
sponds to the discrepancy between the histogram and the normal density
(see Figure 3). In case of α=0.05 the empirical and the normal quantile
are much closer (–8.9 and –8.98).
The question ”which return is exceeded with a probability of 5%” can
be answered using

0.56 + 1.645 · 5.8 = 10.1,

where 1.645 is the 95%-quantile of the standard normal distribution. 95%

of the returns are smaller than 10.1 and 5% of the returns are greater
than 10.1.

Because of the symmetry of the standard normal (e.g. 1.96 for a 95%-interval) the
absolute value of the α/2-quantile is sufficient. The formula for computing the 95%-
interval for y∼N(ȳ, s2 ) is given by:

ȳ ± 1.96 · s,
4.4 Which interval contains a pre-specified percentage of cases? 27

Figure 7: Normal distribution and 95% interval of y.

..................
..... ....
.... ...
... ...
.... ...
..
. ...
.... ...
...
... ...
... ...
... ...
... ...
...
... ...
... ...
.. ...
2.5% ... .
.
..
.
.. ...
...
...
2.5% ..
... .
.. ...
... ...
... ..
. ... ...
... ... ... ...
... .. ... ...
... .. ...
...
... ... ... ..
... .... ... ..
... .
. ..
. ..
... .. .
. ..
... .... ..... ... .. ...
... .... ...... ...
... .... ....
....... ....
... ............ .
. .
.. .
.
........ .. ... .
........... .. .... ....................
......... ..... .... ... .. .......................
............................................. ..
.
.... ... .............................. y
ȳ − 1.96s ȳ ȳ + 1.96s

or, in general, for a 1−α interval:

ȳ ± |zα/2 | · s.

The quantiles of the standard normal distribution are the basis of the rules of thumb
mentioned in section 2.4.3:

1. Approximately two thirds of the observations are in a range plus/minus one

standard deviation around the mean. This rule is based on z0.1587 =−1, which
implies that 68.3% (1−2 · 0.1587) are within one standard deviation.

2. Approximately 95% of the observations are in a range plus/minus two standard

deviations around the mean. In this case the 2.5% quantile 1.96 is rounded up
to 2.0. The resulting interval covers 95.45%.

3. Almost all observations are in a range plus/minus three standard deviations

around the mean. Here the 0.1% quantile 3.09 is rounded down to 3.0 and the
corresponding interval covers 99.73%.

Example 16: Consider the Dax returns again. Assuming a normal dis-
tribution we want to compute an interval for returns that contains 95%
of the data.
Under the normal assumption the mean and standard deviation of the
returns are sufficient to compute a 95% interval. Using ȳ=0.56 and s=5.8
95% of all returns can be found in the interval

[0.56 − 1.96 · 5.8, 0.56 + 1.96 · 5.8] = [−10.8, 11.9].

4.5 Estimating the duration of a project 28

4.5 Estimating the duration of a project

A project can typically be decomposed into many single activities or tasks. If his-
torical data about the duration of such tasks is available, mean and standard devi-
ation for each activity can be estimated (see sheet ’project duration’). Probability
statements about the duration of the entire project are particularly interesting for
planning purposes. This requires to consider the statistical properties of the sum
over all tasks.31 According to the central limit theorem the sum of a large number
(more than 30) of random variables can be described by a normal distribution, if
the components of the sum are independent (in case of a normal distribution this
is equivalent to uncorrelated components). If a small number of activities is consid-
ered, or the durations are not independent of each other, normality of the sum only
holds if the duration of each activity is approximately normal.
The mean of the total duration of m tasks is the sum of the means of all individual
tasks:

ȳt = ȳ1 + ȳ2 + · · · + ȳm .

The standard deviation of the entire duration is based on the variance of the sum
of all individual tasks:

s2t = s21 + s22 + · · · + s2m .

This sum is only correct if the durations of the individual tasks are independent/uncorrelated
among each other. If this is not the case, the covariance among activities must be
taken into account as follows:

m
X m X
X
var[y1 + y2 + · · · + ym−1 + ym ] = s2t = s2i + cov[yi , yj ].
i=1 i=1 i6=j

The standard deviation of the total duration of the project st is the square root of
s2t (the variance of the sum). In other words, it is not appropriate to sum up the
standard deviations of individual tasks.
In practice, it may be questionable to describe the durations of individual activities
by a normal distribution. If only a small number of activities is considered, the
sum of durations cannot be assumed to be normal. Similarly, it may be difficult to
31
We consider (the sum of) activities on the so-called ”critical path”. Any delay in the completion
of such tasks leads to a delayed start of all subsequent activities, and leads to an increase in the
overall duration of the project.
4.5 Estimating the duration of a project 29

provide or estimate the means and standard deviations of activities. It may be easier
for the management to summarize activity durations by specifying the minimum,
maximum and most likely (i.e. mode) duration times. In project management, the
beta distribution is widely used as an alternative to the normal, whereby the
following approximations32 are typically used:

min + 4·mode + max

mean =
6

max − min
standard deviation = .
6

The two parameters of the beta distribution α and β are related to mean and variance
as follows:

 
(mean − min) (mean − min) · (max − mean)
α= ·
(max − min) variance

(mean − min)
β =α· .
(max − min)

The function BETADIST(y ∗ ; α; β; min; max)33 computes the probability to

observe values of a beta distributed variable min≤y≤max with parameters
α and β that are less than or equal to y ∗ .

Example 17: For the data on the sheet ’project duration’ we obtain
mean ȳt =55 and standard deviation st =5.2. Assuming uncorrelated ac-
tivities and using a normal distribution we find that the probability to fin-
ish the project in less than 60 weeks is 83.3%. Using the beta distribution
the corresponding probability is 81.5%. If the correlations/covariances
among activities are taken into account, the standard deviation of the
sum is 8.6 weeks, and the (normal) probability drops to ≈72%.
32
These
√ approximations
√ can be derived by choosing the parameters of the beta distribution to be
α=3+ 2 and β=3− 2.
33
BETAVERT(y ∗ ; α; β; min; max)
4.6 The lognormal distribution 30

4.6 The lognormal distribution

A random variable X has a lognormal distribution if the log (natural loga-
rithm) of X (e.g. Y =ln X) is normally distributed. Conversely, if Y is normal (i.e.
Y ∼N(µ, σ 2 )) then the exponential of Y (i.e. X=exp{Y }) is lognormal. The density
function of a lognormal random variable X is given by

−(ln x − µ)2
 
1
f (x) = √ exp x ≥ 0,
xσ 2π 2σ 2

where µ and σ 2 are mean and variance of ln X, respectively. Mean and variance of
X are given by

E[X] = E[exp{Y }] = exp{µ + 0.5σ 2 } var[X] = exp{2µ + σ 2 }[exp{σ 2 } − 1].

Assuming a lognormal distribution has the following implications:

1. A lognormal variable can never become negative.

2. A lognormal distribution is positively skewed.

As such, the lognormal assumption is suitable for phenomena which are usually
positive (e.g. time intervals or amounts).

The function LOGNORMDIST can be used to compute probabilities assum-

ing a lognormal distribution. Prior to applying this function the log of the
data which is assumed to be lognormally distributed should be computed
(i.e. Y=LN(X)). Using the mean ȳ and the standard deviation sy of the log-
arithm of X, the probability to observe values less than x∗ can be computed
using LOGNORMDIST(x∗ ,ȳ,sy )34 .

4.7 The lognormal distribution35

A random variable X has a lognormal distribution if Y =ln X is normally dis-
tributed. Conversely, if Y ∼N(µ, σ 2 ) then X=exp{Y } is lognormal. The density
function of a lognormal random variable X is given by

−(ln x − µ)2
 
1
f (x) = √ exp x ≥ 0,
xσ 2π 2σ 2
34
LOGNORMVERT(x∗ ,ȳ,sy )
35
Only relevant for banking, finance or similar courses.
4.8 Value-at-Risk 31

where µ and σ 2 are mean and variance of ln X, respectively. Mean and variance of
X are given by

E[X] = E[exp{Y }] = exp{µ + 0.5σ 2 } var[X] = exp{2µ + σ 2 }[exp{σ 2 } − 1].

We now consider the log return in t and treat it as a random variable (denoted by Yt ;
yt is the corresponding sample value or realization). µ and σ 2 are mean and variance
of the underlying population of log returns. Assuming that log returns are normal
random variables with Yt ∼N(µ, σ 2 ) implies that (1+Rt )=exp{Yt }, the simple, gross
returns are lognormal random variables with

E[Rt ] = exp{µ + 0.5σ 2 } − 1 and var[Rt ] = exp{2µ + σ 2 }[exp{σ 2 } − 1].

If the simple, gross return is lognormal (1+Rt )∼LN(1+m, v), mean and variance of
the corresponding log return are given by

 
v
E[Yt ] = ln(1 + m) − 0.5σY2 σY2 = var[Yt ] = ln 1 + . (1)
(1 + m)2
What are the implications for the corresponding prices? Normality of Yt implies
that prices given by Pt =exp{Yt }Pt−1 or Pt =(1+Rt )Pt−1 are lognormal (for given,
non-random Pt−1 ). Thus, prices can never become negative if log returns are normal.
Another attractive feature of normal log returns is their behavior under temporal
aggregation. If single-period log returns are normally distributed Yt ∼N(µ, σ 2 ), the
multi-period log returns are also normal with Yt (h)∼N(hµ, hσ 2 ). This property is
called stability (under addition). It does not hold for simple returns.
Many financial theories and models assume that simple returns are normal. There
are a number of conceptual difficulties associated with this assumption. First, sim-
ple returns have a lower bound of −1, whereas the normal distribution extends to
−∞. Second, multi-period returns are not normal even if single-period (simple) re-
turns are normal. Third, a normal distribution for simple returns implies a normal
distribution for prices, since Pt =(1+Rt )Pt−1 . Thus, a non-zero probability may be
assigned to negative prices which is, in general, not acceptable. These drawbacks
can be overcome by using log returns rather than simple returns. However, empirical
properties usually indicate strong deviations from normality for both simple and log
returns.

4.8 Value-at-Risk
The calculation of Value-at-Risk (VaR) is an important application of the quantiles
of a standard normal distribution. Value-at-Risk is the expected loss in the market
4.9 Value-at-Risk 32

value pt of a risky asset or portfolio that will be exceeded with a given probability
α. Usually α is chosen to be very small (e.g. α=0.01 or α=0.05). Therefore the VaR
refers to a loss over a time-period, typically one day or one week. In statistical terms
VaR is the α-quantile of the distribution of the change in market value pt+1 −pt .
A simplified approach VaR computations is based on assuming that returns of the
asset or portfolio are normally distributed (a highly questionable assumption). It
only relies on the current market value pt , the standard normal quantile zα , and the
standard deviation of returns s. The mean of returns is assumed to be zero, which
makes the resulting measure of risk only relevant in the short term. The VaR is
given by36

VaR(α) = −pt · zα · s.

Example 18: The daily VaR of a DAX investment with a current market
value of 500 units can be derived from the monthly sample statistics.
Assuming 21 trading
√ days per month the daily standard deviation is
given by s=5.8/ 21=1.266.37 Using α=0.05 the 5%-VaR is given by

VaR(0.05) = −500 · −1.645 · 0.01266 = 10.41.

Interpretation: on the next trading day there is a 5% probability for a

loss in the market value of 10.41 or more. The expected loss that will
be exceeded with only 1% probability is based on the 1%-quantile and
is given by 14.72.

4.9 Value-at-Risk38
The calculation of Value-at-Risk (VaR) is an important application of the quantiles
of a standard normal distribution. Value-at-Risk is the expected loss in the market
value of a risky asset or portfolio that will be exceeded with a given probability α:

P[(pt+1 − pt ) ≤ −VaR] = α.

pt is the market value of the asset at time t. Since α is usually very small (e.g.
α=0.01 or α=0.05) the VaR refers to a loss over a time-period, typically one day. In
statistical terms VaR is the α-quantile of the distribution of the change in market
value pt+1 −pt .
36
Note that s in this formula has to be a decimal number.
37
Important note: When returns have been measured in percentage terms, as in the present case,
s has to be divided by 100!
38
Only relevant for banking, finance or similar courses.
4.9 Value-at-Risk 33

If log returns are assumed to be normally distributed VaR may be calculated on

the basis of mean and standard deviation of historical returns of the asset. For log
returns and a probability of 5% the VaR is given by

VaR(0.05) = −pt [exp{ȳ + 0.5s2 − 1.645s} − 1].

ȳ und s are mean and standard deviation of the returns of the asset, –1.645 is the
5%-quantile of the standard normal distribution, and exp{x}=2.718x . The term
0.5s2 in the exponent is due to the properties of the lognormal distribution.39 In
general

VaR(α) = −pt [exp{ȳ + 0.5s2 + zα s} − 1].

RiskMetrics uses a simplified approach for value-at-risk computations. The mean is

ignored and the VaR is given by40

VaRRM (α) = −pt · zα · s.

Example 19: The daily VaR of a DAX investment with a current market
value of 500 units can be derived from the monthly sample statistics. As-
suming 21 trading days per month the daily√ mean and standard deviation
are given by ȳ=0.56/21=0.027 and s=5.8/ 21=1.266. In computations
involving exp{} decimal numbers have to be used in the exponent! When
returns have been measured in percentage terms – as in the present case
– ȳ and s have to be divided by 100! Using α=0.05 the 5%-VaR is given
by

Ψ0.05 = 0.00027 + 0.5 · 0.012662 − 1.645 · 0.01266 = −0.02047

VaR(0.05) = −500[exp{−0.02047} − 1] = 10.13.

Interpretation: on the next trading day there is a 5% probability for a
loss in the market value of 10.13 or more. The expected loss that will
be exceeded with only 1% probability is based on the 1%-quantile and
is given by

VaR(0.01) = −500[exp{−0.029097} − 1] = 14.34.

39
X=exp{Y } is lognormal if Y =ln X is normal. If Y ∼N(µ, σ 2 ) the expected value of X is given
by exp{µ+0.5σ 2 }.
40
Note that s in this formula has to be a decimal number.
4.10 The binomial distribution 34

4.10 The binomial distribution

The binomial distribution is used to compute probabilities for outcomes of a random
experiment with the following properties:

1. the experiment involves n independent trials.

2. each trial has two possible outcomes. These are usually called success or
failure.

3. the probability of success p is the same in each trial.

The probability of y successes in n trials is given by

 
n y
f (y) = p (1 − p)(n−y) ,
y

where

 
n n!
= n! = 1 · 2 · · · n, 0! = 1
y y!(n − y)!

is the binomial coefficient.

If the number of trials in a binomial experiment is large, the binomial distribution
can be replaced by the normal distribution with mean np and variance np(1−p). As
a rule of thumb np≥5 and n(1−p)≥5 must hold.

The probability of y or less successes in n trials of a binomial experiment

can be computed with the function BINOMDIST(y; n; p; 1)41 .

Example 2042 : This example presents a simplified version of the calcu-

lations used by airlines when they overbook flights. Airlines know that a
certain percentage of customers cancel at the last minute. Thus to avoid
empty seats they sell more tickets than there are seats. We will assume
the no-show rate is 10%. That is, we are assuming that each customer,
independently of others, shows up with probability 0.9 and cancels with
probability 0.1.
The sheet ’overbooking’ contains the calculations to determine the fol-
lowing probabilities for a flight with 200 available seats: the probability
41
BINOMVERT(y; n; p; 1)
42
Example 6.10 on page 273 in AWZ.
4.10 The binomial distribution 35

that (a) more than 205 passengers will show up, (b) more than 200 pas-
sengers will show up, (c) at least 195 seats will be filled, and (d) at least
190 seats will be filled. The first two of these are ”bad” events for the
airlines while the last two are ”good” events.
In order to answer the questions in this example we consider individual
customers. For each ticket sold we carry out a binomial experiment. We
recall the three necessary conditions for a binomial experiment:

1. two possible outcomes: in each trial we distinguish two cases – i.e.

a customer may show up or not. We will treat ’show-up’ as success.
2. independence: whether a customer shows up or not does not depend
upon any other customer. This assumption may not be justified for
groups of customers (e.g. one family member gets sick and his or
her partner does not show up either).
3. constant probability: for each customer the probability p for success
is the same. Again this may be considered a strong simplification.
Note that this probability must be for the event defined to be a
success. In other words, if success was defined to be ’no-show’, p
would have to be defined differently.

The number of trials is given by the number of tickets sold. Considering

the ”bad” events we are interested in the probability to observe more
than 205 (i.e. 206, 207, . . . ) customers to show up. Thus we subtract
the probability to see 205 or less customers from 100%. The resulting
probability is 0.1%. Note that the number of available seats does not
affect the computation of probabilities.
To consider the ”good” events we compute the probability that at least
195 seats (i.e. 195, 196, . . . ) will be filled. This can be obtained by
computing one minus the probability to see 194 or less passengers. This
probability is given by 42.1%.

Example 21: An important financial application of binomial experi-

ments is option pricing using binomial trees. At each point in time
only two possibilities are distinguished: the prices may either increase
(success) or decrease. In the next period, for each of the two possible
outcomes, another binomial experiment is carried out. This leads to a
so called binomial tree that covers a certain time interval. Each state at
the end of the time interval corresponds to a specific sequence of success
and failure. The binomial distribution can be used to compute the prob-
ability for each possible final state.
4.10 The binomial distribution 36

Consider the following example. The probability for a stock price in-
crease (u) is 0.6 and 0.4 for a decrease (d). Assume that successive price
changes are independent. After three periods the following sequences are
possible: (u, u, u), (u, u, d), (u, d, u), (u, d, d), (d, u, u), (d, u, d), (d, d, u),
(d, d, d). The probability for three consecutive increases (u, u, u) is given
by (y=3, n=3, p=0.6):
 
3
0.63 (1 − 0.6)0 = 0.216.
3

The probability for two increases in three periods – (u, u, d), (u, d, u) or
(d, u, u) – is given by (y=2, n=3, p=0.6):
 
3
0.62 (1 − 0.6)1 = 0.432.
2

Note that the probability for one decrease in three periods is not equal
to (1−0.432) but 0.432, too! Considering decreases implies changing the
meaning of y. Now a decrease is treated as success with probability
p=0.4. Thus the probability for one decrease in three periods is given by
(y=1, n=3, p=0.4)
 
3
0.41 (1 − 0.4)2 = 0.432.
1

Now suppose that more than one event is considered – e.g. one or two
decreases. This requires to compute the probability of each event and to
sum these probabilities. The probability for one or no decrease in three
periods is given by
   
3 0 3 3
0.4 (1 − 0.4) + 0.41 (1 − 0.4)2 = 0.216 + 0.432 = 0.648.
0 1

This is equal to one minus the probability for two or three decreases in
three periods.
5.1 Samples and confidence intervals 37

5 How accurate is an estimate?

There are two major ways to describe an interesting phenomenon in statistical terms
(using mean, variance, . . . ): one can use the population43 or use a sample from the
population. Samples are mainly used for economic reasons, or to save time. It is
important to draw a random sample, i.e. each element of the population must have
the same chance to be drawn.
Descriptive statistics are used to describe the statistical properties of samples. Fre-
quently the sample statistics are used to support various decisions. In applications,
the estimated mean is treated as if it was the true mean of the population. Since
the mean has been derived from a sample it has to be taken into account that the
estimated mean is subject to an estimation error. Another sample would have a
different mean. One of the major objectives of statistics is to use samples to draw
conclusions about the properties of the population. This is done in the context of
computing confidence intervals and hypotheses tests.

Example 22: We consider the data from example 1 and focus on the
average salaries of respondents. The purpose of the analysis is three-
fold. First, we want to assess the effects of sampling errors. Second, we
ask whether the average of the sample is compatible with a population
mean of $47500 or strongly deviates from this reference. Third, the av-
erage salaries of females and males will be compared to see whether they
deviate significantly from each other.44

5.1 Samples and confidence intervals

The mean ȳ is computed from the n observations y1 , . . . , yn of a sample using

n
1X
ȳ = yt .
n
t=1

The means is said to be estimated from the sample, and it is a so-called estimate.
The estimate is a random variable – using a different sample results in a different
estimate ȳ. It should be distinguished from the population mean µ – which is also
called expected value.45 The symbols µ and σ 2 are used to denote the population
43
The population consists of all elements which have the feature of interest.
44
This rather loose terminology will subsequently be changed, whereby questions and answers
will be formulated in a statistically more precise way.
45
The contents of sections 5.1 and 5.3 is explained in terms of the mean of a random sample.
Similar considerations apply to other statistical measures.
5.1 Samples and confidence intervals 38

mean and variance. The expected value µ can be considered to be the limit of the
empirical mean, if the number of observations tends to infinity:

n
1X
µ = E[Y ] = lim yt .
n−→∞ n
t=1

Usually ȳ will differ from the true population value µ. However, it is possible to
compute a confidence interval that specifies a range which contains the unknown
parameter µ with a given probability α. When a confidence interval is derived one
has to take into account the sample dependence and randomness of ȳ. In other
words, the sample mean is a random variable and has a corresponding (probability)
distribution.
The distribution of possible estimates ȳ is called sampling distribution. For large
samples the central limit theorem states that the sample mean ȳ is normally
distributed with expected value µ and variance46 s2 /n: ȳ∼N(µ, s2 /n). The theo-
rem holds for arbitrary distributions of the population provided the sample is large
enough (n>30); if the population is normal it holds for any n.47
Using the properties of the normal distribution a confidence interval which contains
the true mean µ with (1−α) probability can be derived. More precisely, (1−α) per-
cent of all samples (randomly drawn from the same population) will contain µ. In
general, the (1−α) confidence interval of µ is given by

√
ȳ ± |zα/2 | · s/ n.

For example, the 95% confidence interval of µ is given by

√
ȳ ± 1.96 · s/ n.
√
The function CONFIDENCE(α; s; n)48 computes the value |zα/2 | · s/ n.

From the sample we obtain the following estimates: ȳ=52263, s=11493,

n=30. A 95% confidence interval for the population mean µ is given by
46
To simplify the exposition the sample variance s2 is assumed to be the same in each sample
and equal to the population variance σ 2 . Therefore, on the following pages, the standard normal
distribution can be used instead of the t-distribution, which theoretically applies if s2 is used. If n
is large the t-distribution is very similar to the standard normal distribution.
47
The applet on http://onlinestatbook.com/stat_sim/sampling_dist/index.html illustrates
this theorem.
48
KONFIDENZ(α; s; n)
5.1 Samples and confidence intervals 39

√
52263 ± 1.96 · 11493/ 30 = 52263 ± 4113 = [48151, 56376].

Based on the sample, we conclude that the actual average µ can be found
in the interval [48151, 56376] with 95% probability. Note that this is not
an interval for the data, but an interval for the mean of the population.

Average salary
mean 52263
standard deviation 11493
number of observations 30

standard error 2098

confidence interval for the mean of the population

lower upper
α quantile bound bound
0.05 1.960 48151 56376

If ȳ is used instead of µ there will be an estimation error =µ−ȳ. The expected

value of the estimation error equals zero since µ is the expected value of ȳ. The 95%
confidence interval for the estimation error  is given by

√ √
[−1.96 · s/ n, +1.96 · s/ n]

and the (1−α) confidence interval for  is given by

√
±|zα/2 | · s/ n.
√
s/ n is also called standard error (standard deviation of the estimation error).
This formula is valid if the population has infinitepsize. If the size of the population
√
is known to be N the standard error is given by (N −n)/(N −1)s/ n.
The boundaries of the interval can be used to make statements about the magnitude
of the absolute estimation error. Using α=0.05 the boundaries of the interval in this
example are given by

√
±1.96 · 11493/ 30 = ±4113.

In words: there is a 95% probability that the absolute estimation error for the
average salary in the population is less than $4113.
5.2 Sampling procedures 40

The confidence interval for the estimation error can be used as a starting point to
derive the required sample size.49 For that purpose it is necessary to fix an acceptable
magnitude of the (absolute) error; more specifically the absolute error which can be
exceeded with probability α. This value α corresponds to the boundaries of the
(1−α) confidence interval for the estimation error :

√
|zα/2 | · s/ n = α .

This expression can be rewritten to obtain a formula for the corresponding sample
size:

2
zα/2 · s

n= .
α

Suppose that a precision of α =$500 is required and α=0.05 is used. This means
that the (absolute) error in the estimation of the mean is accepted to be more than
$500 in five percent of the samples. In this case the required sample size is given by

 2
1.96 · 11493
n= ≈ 2030.
500

5.2 Sampling procedures

Example 2350 : The objective of a study is (among others) to inves-
tigate the volume (page numbers) of master theses. Using a sample
of theses the objective is to compute a 95% confidence interval for the
average number of pages in the population.

Drawing a sample from a population can be done on the basis of several princi-
ples. We consider three possibilities: random, stratified and clustered sampling.
Random sampling – which has been assumed in previous sections of the text – col-
lects observations from the population (without replacement) according to a random
mechanism. Each element of the population has the same chance of entering the
sample. The objective of alternative sampling methods is to reduce the standard
49
Note: These considerations are based on the assumption, that the standard deviation of the
data s is known, before the sample has been drawn.
50
Bortz J. and Döring N. (1995): Forschungsmethoden und Evaluation, 2. Auflage, Springer,
p.390.
5.2 Sampling procedures 41

errors compared to random sampling and to obtain smaller confidence intervals. Al-
ternative methods are chosen because they can be more efficient or cheaper (e.g.
clustered sampling).
A random sample can be obtained by assigning a uniform random number to each
of the N elements of the population. The required sample size51 n determines the
percentage α=n/N . The sample is drawn by selecting all those elements whose
associated random number is less than α. The number of actually selected elements
will be close to the required n if N is large. Exactly n elements are obtained if the
selection is based on the α-quantile of the random numbers as shown on the sheet
’random sampling’.
Stratified sampling is based on separating the population into strata (or groups)
according to specific attributes. Typical attributes are age, gender, or geographical
criteria (e.g. regions). Random samples are drawn from each stratum. Stratified
sampling is used to ascertain that the representation of specific attributes in the
sample corresponds (or is similar) to the population. If the distribution of an at-
tribute in the population is known (e.g. the proportion of age groups or provinces in
the population), the sample can be defined accordingly (e.g. each age group appears
in the sample with about the same frequency as in the population).
In the present example stratified sampling can be based on the type of a thesis
(empirical, theoretical, etc.) or the field of study (law, economics, engineering, etc.).
Stratified sampling is particularly important in relatively small samples to avoid
that specific attributes (e.g. fields of study) do not appear at all, or are incorrectly
represented (too few or too many cases). The subject of the analysis (number of
pages) should be related to the stratification criterion (type of thesis).
The ratio of the number of observations nj in stratum j and the sample size n
defines weights wj =nj /n (j is one out of m strata; n is the sum of all nj ). If the
proportions of the attributes in the population are known (e.g. the percentage of
empirical theses in the population), the weights wj should be determined such that
the proportions of the sub-samples correspond exactly to the proportions of the
attributes in the population. If such information is not available and the sample is
large, the proportions in a random sample will approximate those in the population.
The (overall) mean of a stratified sample is the weighted average of the means of
each stratum ȳj :

m
X
ȳ = wj · ȳj .
j=1

This mean is equal to the mean obtained from all observations in the sample. If
51
As shown in section 5.1 n can be chosen on the basis of the required precision (or absolute
estimation error).
5.2 Sampling procedures 42

nj ·wj >10 the mean is approximately normally distributed.

The standard error (of the mean ȳ) is based on a weighted average of the standard
errors of each stratum sȳj :

m
X
s2ȳ = wj2 · s2ȳj .
j=1

If the weights deviate from those in the population, the standard error cannot be
reduced, or can even increase, compared to random sampling. At the same time, the
mean ȳ will be biased and will deviate from the mean of the population and from
random sampling.
If the dispersion in each stratum is rather small (i.e. individual strata are rather
homogeneous), the standard error can be lower compared to a random sample. This
will be the case if the stratification criterion is correlated with the subject of the
analysis (e.g. if the distribution of the number of pages depends on the type of the
thesis). For example, to analyze the intensity of internet usage, strata could be
defined on the basis of age groups. If the dispersion in sub-samples is about the
same as in the overall sample, or the means in each stratum are rather similar, there
is no need for stratification (or, another attribute has to be considered).
In the present example on the sheet ’stratified sampling’ two strata based on the
type of a thesis are used. A sample of n1 =34 empirical and n2 =16 theoretical theses
is drawn from a population consisting of 136 and 64 theses, respectively; i.e. the
proportions in the sample correspond to those in the population. The means and
standard deviations of the two strata are given by ȳ1 =68, sȳ1 =2.8 and ȳ2 =123,
sȳ2 =12. Empirical theses have less pages and less dispersion than theoretical theses.
The stratified mean is given by 0.68·68+0.32·123≈86. Its standard error is given by

p
sȳ = 0.682 · 2.8 + 0.322 · 122 = 4.3.

The boundaries of the 95% confidence interval are 86±1.96·4.3=[77.1;93.9]. Com-

pared to the interval [79.0;102.2] obtained from a random sample, the mean of the
population can be estimated more accurately with stratified sampling.
Clustered sampling divides the population into clusters (e.g. schools, cities, com-
panies). A cluster can be viewed as a minimized version of the population and
should be characterized by as many aspects of the population as possible. Since this
will (usually) not be the case, several clusters are selected. As opposed to stratified
sampling all elements of a cluster are included in the sample. In the present example
15 supervisors from the entire set of 450 supervisors are randomly selected. All mas-
ter theses supervised by the selected professors are contained in the sample. When
5.2 Sampling procedures 43

computing the standard error, the number of theses supervised by each professor is
taken into account.
Clustered sampling can be more easily administered than other sampling procedures.
For example, the analysis of grades is based on only a few schools rather than
choosing students from many schools all over the country. The random element in
the sampling procedure is the choice of clusters. The procedure only requires a list
of all schools rather than a list of all students from the population. A list of all
students is only required for each school.
The ratio of the number of elements nj in each cluster (e.g. the number of theses
supervised by each professor) and the sample size n defines the weights wj =nj /n (j
is one of m clusters; n is the sum over all nj ). The coefficient of variation of n̄, the
mean over all nj , should be smaller than 0.2. The means ȳj of each cluster (e.g. of
each supervisor) are treated as the ”data”. The mean across all observations ȳ is
the weighted average of the cluster means ȳj :

m
X
ȳ = wj · ȳj .
j=1

This mean is equal to the mean obtained from all n observations.

The standard error (of the mean ȳ) is based on the weighted sum of squared devia-
tions between ȳj and ȳ:

m
X
s2ȳ = wj2 · (ȳj − ȳ)2 .
j=1

Since all observations of a cluster are sampled (which does not imply any estimation
error) the standard error only depends on the differences among clusters. There-
fore clusters should be rather similar whereas the dispersion within clusters can be
relatively large.
The computation of the standard error can be based on a more exact formula which
takes the ratio of selected clusters m and available clusters M into account (in the
present example 15/450):

m   
X m m
s2ȳ = 1− · · wj2 · (ȳj − ȳ)2 .
M m−1
j=1

Figure 8 shows data and results for the present example. Compared to stratified
sampling the confidence interval can be substantially reduced.
5.3 Hypothesis tests 44

Figure 8: Clustered sampling.

Number of theses in the sample 100
Number of supervisors in the sample 15
Total number of supervisors (population) 450

number of mean page weighted sq.

supervisor theses number mean dev.
1 8 90 0.0256 mean 92
2 2 105 0.0676 std.error 0.985
3 10 95 0.09 95%-CI (lower bound) 90
4 9 93 0.0081 95%-CI (upper bound) 94
5 7 94 0.0196
6 9 91 0.0081
7 6 92 0
8 1 124 0.1024
9 7 88 0.0784
10 11 86 0.4356
11 5 91 0.0025
12 9 89 0.0729
13 3 97 0.0225
14 6 95 0.0324
15 7 93 0.0049

5.3 Hypothesis tests

A hypothesis refers to the value of an unknown parameter of the population (e.g.
the mean). The purpose of the test is to draw conclusions about the validity of a
hypothesis based on the estimated parameter and its sampling distribution.
For this purpose the so-called null hypothesis H0 is formulated. For instance, the
H0 : µ=µ0 states that the unknown mean is equal to µ0 . Every null hypothesis has a
corresponding alternative hypothesis, e.g., Ha : µ6=µ0 . Acceptance of H0 implies
the rejection of Ha and vice versa.
To test a (null) hypothesis one proceeds as follows: ȳ is estimated from a sample of
n observations. In general, the sample estimate ȳ will differ from µ0 . The question
is whether the difference is large enough to assume that the sample comes from a
population with another mean µ6=µ0 .
The hypotheses test is used to determine whether the difference between ȳ and µ0
is statistically significant52 or random. The test uses critical values to determine
the significance.
52
The issue of the economic relevance of a deviation is not the purpose of the statistical analysis,
but should be treated nonetheless.
5.3 Hypothesis tests 45

Figure 9: Acceptance region and critical values for H0 : µ=µ0 using α=5%.
..................
..... ....
.... ...
... ...
.... ...
..
. ...
.... ...
...
... ...
... ... ... ...
... ... ... ...
... ... ...
... ...
... ... ... ...
√ ...
...
.
..
.. .
... ...
...
√
µ0 <ȳ−1.96·s/ n ... .
..
.
..
...
.
... µ >ȳ+1.96·s/ n
... 0
... . .
... ...
............................................................. ... ... .............................................................
... ... .
... .
.
... ... ... ....
reject H0 ...
...
..
..
...
...
...
...
reject H0
...
...
..... .... ..
..................... acceptance region ...........................
.... ..... ... ...
... ..
....... ......
.....
.
...... .
..... .. .... ..........
...... .... ... ...........
............... .
. ... ...........
.. .
............................................. .... .... ...
..........................................
ȳ, µ0
√ √
ȳ−1.96·s/ n ȳ ȳ+1.96·s/ n

Two-sided tests based on the confidence interval

One possible decision rule is based on the confidence interval. For (1−α) percent of
all samples the population mean µ lies within the bounds of the confidence interval.
If the mean under the null hypothesis µ0 lies outside the confidence interval, the null
hypothesis is rejected (see Figure 9). In this case it is too unlikely that the sample
at hand comes from a population with mean µ0 . If µ0 lies outside the confidence
interval, the estimated mean ȳ is said to be significant (or significantly different
from µ0 ) at a significance level of α.
The test is based on critical values – these are the boundaries of the (1−α) confidence
interval – which are given by

√
ȳ ± |zα/2 | · s/ n,

where s ist the estimated standard deviation from the sample. Then µ0 is compared
to the critical values. Using α=0.05 the null hypothesis is rejected, if µ0 is less
√ √
than ȳ−1.96·s/ n or greater than ȳ+1.96·s/ n (see Figure 9). If µ0 lies in the
acceptance region, H0 is not rejected. In a two-sided test H0 is rejected, if µ0
is above or below the critical values. In one-sided tests53 , only one of the two
critical values is relevant.

In example 22, the objective is to find out whether the sample mean is
consistent with the target average $47500. For that purpose a two-sided
test is appropriate since the kind of deviations (ȳ is above or below µ0 )
is not relevant. The 95% confidence interval is given by [48151,56376].
Using a significance level of 5% the null hypothesis is rejected, since
53
Details on one-sided tests can be found in section 10.2.2 of AWZ, 3rd edition.
5.3 Hypothesis tests 46

the target average $47500 is outside the confidence interval. The data
does not support the assumption that the sample has been drawn from
a population with an expected value of $47500. In other words: the
sample supports the notion that the average salary of respondents differs
significantly from the target mean.

standard.
two-sided test lower upper test critical
µ0 α bound bound statistic value p-value
H0 47500 0.05 48151 56376 reject 2.270 1.960 reject 0.023 reject
47500 0.01 46859 57668 accept 2.270 2.576 accept 0.023 accept

Standardized test statistic

Instead of determining the bounds of the confidence intervals and comparing the
critical values to µ0 , the standardized test statistic

ȳ − µ0
t= √
s/ n

can be used. In this formula the difference between ȳ and µ0 is treated relative to the
√
standard error s/ n. When the null hypothesis is true, there is a (1−α) probability
to find the standardized test statistic within ±|zα/2 | (in a two-sided test). The null
hypothesis is rejected when the difference between ȳ and µ0 is too high relative to
the standard error. A decision is based on comparing the absolute value of t to the
absolute value of the standard normal α/2-quantile (see Figure 10).
The decision rule in a two-sided test is: If |t| is greater (less) than |zα/2 | the null
hypothesis is rejected (accepted).

Using the data from example 22 the standardized test statistic is given
by

52263 − 47500
t= √ = 2.27.
11493/ 30

For a 5% significance level the critical value is 1.96. In the present

example H0 is rejected at the α=5% significance level since |2.27| is
greater than the absolute value of the α/2-quantile. The conclusion
based on the standardized test statistic must always be identical to the
conclusion based on the confidence interval.

The steps involved in a two-sided test can be summarized as follows:

5.3 Hypothesis tests 47

Figure 10: Acceptance region and critical values for H0 : µ=µ0 using a standardized
test statistic.
...................
..... ....
.... ...
... ...
...
. ...
.... ...
...
... ...
... ...
.... ... ...
... ....
... ... ... ...
... ... ... ...
... ... .
... ...
... ... ...
t < zα/2 ...
... .
.
..
..
.
...
...
.
...
...
...
t > z1−α/2
.
. ... ... ...
............................................................. ... .
... ............................................................
.... ... .... ....
... ..
. ...
reject H0 ...
... .
.
...
.
.
....
...
.
...
reject H0
...
...
........................ .
..........................
.
... . ....
... ...
acceptance region ... ..
... ..
.....
. ......
..... .....
.... .. .. .....
..... .. ... .........
.........
.
...... .... ... ........
........... ȳ − µ0
....................................... .......... .
.
...
.
...
...
..
..........................................
t= √
zα/2 0 z1−α/2 s/ n

1. formulate the null hypothesis and fix the level of significance:

H0 : µ0 = 47500; α = 0.05

2. estimate the sample mean and standard deviation:

ȳ = 52263, s = 11493

3. compute the standardized test statistic:

52263 − 47500
√ = 2.27
11493/ 30

4. obtain the critical value for the level of significance α=0.05:

|zα/2 | = |z0.025 | = 1.96

5. compare the absolute value of the test statistic to the critical value and draw
a conclusion:

|2.27| > 1.96 H0 is rejected

5.3 Hypothesis tests 48

Figure 11: t-statistic and p-value in a two-sided test.

.....................
.... ....
.... ...
... ...
.... ...
... ...
.. ...
−1.96 .
... ...
...
+1.96
.... .... ... ....
... ... ... ...
... ... ... ...
... ... ...
... ...
... ... ... ...
..
t < −1.96 ...
... .
.
..
.
...
...
...
t > +1.96
...
.
. ... .
... ...
.............................................................. ... .
... .............................................................
... ... ... ...
... ... . ...
reject H0 ...
... .
.
.
.
..
.
...
....
.
...
reject H0 ...
...
........................ ..........................
.
.. . ... accept H0 ... ..
....
... ....
.....
.
... ..
......
......
p-value: this area×2
....
...
....
..
.......... . ....... ....
. .
.....
.
. . ... .
...... ... ..
.... .........................
.......
......... .... .. ... ... ..... ..................................................
............................................. ... ... ... ...
0 t=2.27

Errors of type I and significance level

The acceptance region of a hypothesis depends on the specified significance level.

Therefore, by choosing a small enough value for α, the acceptance region can always
be made large enough to make any value of ȳ consistent with H0 . This is not a
very informative test, however. The specification of too large values for α is equally
problematic, since H0 will be rejected almost certainly.
In order to find a reasonable value for α the following aspect must be taken into
account: α is the probability that the unknown mean is actually outside the accep-
tance region. If a null hypothesis is rejected, there is a probability of α that a wrong
decision has been made. This is said to be a type I error: the null hypothesis
is rejected although it is correct.54 Therefore the value of α should be determined
with regard to the consequences associated with a type I error. The more important
(or the ’more unpleasant’) the consequences of an unjustified rejection of the null
hypothesis are, the lower α should be chosen. However, α should not be too small
because one may exclude the possibility to reject a wrong null hypothesis. Typical
values for α are 0.01, 0.05 or 0.1.

P-value

For a given value of the test statistic the chosen significance level α determines
whether the null hypothesis is accepted or rejected. Changing α may lead to a
change in the acceptance/rejection decision. The p-value (or prob-value) of a test
is the probability of observing values of the test statistic that are larger (in absolute
54
A type II error occurs, if a null hypothesis is not rejected, although it is false. This type of
error and the aspect of the power of a test are not covered in this text.
5.3 Hypothesis tests 49

terms) than the value of the test statistic at hand if the null hypothesis is true (see
Figure 11). The more the standardized test statistic differs from zero the smaller
the p-value. The p-value can also be viewed as that level of α for which there is
indifference between accepting or rejecting the null hypothesis. The significance level
α is the accepted probability to make a type I error. H0 is rejected, if the p-value is
less than the pre-specified significance level α.55

The p-value for a standardized test statistic in a two-sided test can be

√
computed from 2*(NORMDIST(µ0 ;ȳ;s/ n;1)56 (provided that µ0 is less
than ȳ) or 2*(1−NORMSDIST(ABS(t)))57 .

Decision rule: if the p-value is less (greater) than the pre-specified significance level
α the null hypothesis is rejected (accepted).

The value of the standardized test statistic based on the sample (ȳ=52263,
s=11493) is given by 2.27. The associated p-value is 0.023. Rejecting
the null hypothesis in this case implies a probability of 2.3% to com-
mit a type I error. Given a significance level of 5% this probability is
sufficiently small and H0 is rejected.

Testing the difference between means

Frequently, there is an interest to test whether two means differ significantly from
each other. Examples are differences between treatment and control groups in med-
ical tests, or differences between features of females and males. Two situations can
be distinguished: (a) a paired test applies when measurements are obtained for the
same observational units (e.g. the blood pressure of individuals before and after
a certain treatment); (b) the observational units are not identical (e.g. salaries of
females and males); this is referred to as independent samples.
In a paired test the difference between the two available observations for each element
of the sample is computed. The mean of the differences is subsequently tested
against a null hypothesis in the same way as described above. For example, the
effectiveness of a drug can be tested by measuring the difference between medical
parameter values before and after the drug has been applied. If the mean of the
differences is significantly different from zero, whereby a one-sided test will usually
be appropriate, the drug is considered to be effective.
If data has been collected for two different groups, the summary statistics for the two
groups will differ, and the number of observations may differ. It is usually assumed,
55
The conclusions based on the three approaches to test a hypothesis must always coincide.
56
NORMVERT
57
STANDNORMVERT
5.3 Hypothesis tests 50

that the elements of each sample are drawn independently from each other.58 Sup-
pose the means of the two groups are denoted by ȳ1 and ȳ2 , the standard deviations
for each group are s1 and s2 , and the sample size of each group are n1 and n2 .
For the null hypothesis that the difference between the means in the population is
µ1 −µ2 the standardized test statistic is given by

(ȳ1 − ȳ2 ) − (µ1 − µ2 )

t= s .
s21 s22
+
n1 n2

The test statistic is compared to |zα/2 |, as described in the context of the standard-
ized test statistic.

Using data from example 1 we want to test whether average salaries of

females and males are statistically different. This situation calls for an
independent samples test. The null hypothesis states that the average
salaries of females and males in the population are identical: µf =µm .
The sample means of females and males are $48033 and $55083, respec-
tively. Using the sample sizes and standard errors for each group the
corresponding test statistic is given by

(55083 − 48033) − 0
t= r = 1.5636.
119722 121822
+
18 12
This test statistic is less than the 5% critical value 1.96 and the p-value
is 11.8%. Although the difference between $48033 and $55083 is rather
large, it is not statistically significant (different from zero). Thus, the
sample provides insufficient evidence to claim that the salaries of females
and males are different. This can be explained by the small sample, but
also by that fact that other determinants of salaries are ignored.

58
The independence assumption does not hold in case of a paired test situation. Therefore, the
subsequently derived test statistic cannot be applied in this case.
6.1 Covariance and correlation 51

6 Describing relationships – Correlation and regression

Consider a sample of observations from two random variables Y and X (yt , xt ,
t=1, . . . , n) which are supposed to be related. Correlation is a measure for the
strength and direction of the relationship between Y and X. However, if the cor-
relation coefficient is found to be significantly59 different from zero (e.g. age and
weight of children) cannot be used to compute the expected value of one of the two
variables based on specific values of the other variable (e.g. expected weight in kg for
an age of five years). it makes sense to run a regression analysis. For that purpose
a regression analysis is required. A regression model allows to draw conclusions
about the expected value (or mean) of one of the two variables based on specific
values of the other variable.

Example 24: We consider the data from example 1 and focus on the
relation between salaries and age of respondents. The purpose of the
analysis is to estimate the average increase in salaries over the lifetime
of an individual.

6.1 Covariance and correlation

Correlation is a measure for the common variation of two variables. The correlation
coefficient indicates the strength and the direction of the relation between the two
variables. In portfolio theory, the correlation between the returns of assets has key
importance, because it determines the extent of the diversification effect.
Consider a sample of observations of two variables (yt , xt , t=1, . . . , n) as shown
in the scatter diagram in Figure 12. Each dot corresponds to the (simultaneous)
observation of two values. Correlation is mainly determined by deviations from the
mean (see below). Therefore the position of the axes in Figure 12 is defined by the
means of the two variables.
The correlation is negative if there is a tendency to observe positive (negative)
deviations from the mean of one variable, and negative (positive) mean-deviations
of the other variable (e.g. price and quantity sold of some products). In other words,
the observations of the two variables tend to be located on opposite sides of their
means. Positive correlation prevails if there is a tendency to observe deviations from
the mean with the same sign (e.g. income and consumption). In this case the values
of the two variables tend to be located on the same side of their means.
The correlation coefficient ranges from −1 to +1. The correlation is an indicator
– it has no units of measurement. The strength of the relationship is measured by
the absolute value of the correlation. A strong relationship holds if there are hardly
59
The significance of a correlation coefficient can be tested using a hypothesis test along the lines
√
described in section 5.3. The standard error required for this test is given by 1/ n.
6.1 Covariance and correlation 52

Figure 12: Scatter diagram.

Correlation

Age Salary $90,000

58 $65,400
48 $62,000 $80,000
58 $63,200
56 $52,000
$70,000
68 $81,400
60 $46,300 Salary (y)
$60,000
28 $49,600
48 $45,900
$50,000
53 $47,700 25 30 35 40 45 50 55 60 65 70
61 $59,900
36 $48,100 $40,000

55 $58,100
48 $56,000 $30,000
51 $53,400 Age (x)
39 $39,000
45 $61,500 correlation coefficient 0.59
43 $37,700

any exceptions to the tendencies described above. This is indicated by a correlation

coefficient close to ±1. The correlation is close to zero if none of the two tendencies
prevails. In this case the absence of a relationship is inferred. The correlation of the
data in Figure 12 equals 0.59.
The correlation coefficient between yt and xt is computed from

syx
correlation: ryx = .
sy sx

syx is the covariance which is computed from yt und xt as follows:

n
1 X
covariance: syx = (yt − ȳ) · (xt − x̄).
n−1
t=1

ȳ (x̄) and sy (sx ) are mean and standard deviation of yt (xt ).

Correlation and covariance can be computed with the functions CORREL(data

range of y; data range of x) and COVAR(data range of y; data range of
x)60 .

Note that the correlations (and covariances) are symmetrical: the correlation be-
tween y and x (ryx ) is identical to the correlation between x and y (rxy ).
60
KORREL; KOVAR
6.1 Covariance and correlation 53

Table 4: Computing the correlation coefficient.

y x y-mean(y) x-mean(x) product rank y rank x difference
65400 58 8060 4.2 33852 2 4 -2
62000 48 4660 -5.8 -27028 4 8 -4
63200 58 5860 4.2 24612 3 4 -1
52000 56 -5340 2.2 -11748 6 6 0
81400 68 24060 14.2 341652 1 1 0
46300 60 -11040 6.2 -68448 9 3 6
49600 28 -7740 -25.8 199692 7 10 -3
45900 48 -11440 -5.8 66352 10 8 2
47700 53 -9640 -0.8 7712 8 7 1
59900 61 2560 7.2 18432 5 2 3
mean 57340 53.8 sum 585080
std.dev 11257.4 10.9 covariance 65009
correlation 0.53 rank correlation 0.52

Table 4 illustrates the computation of the correlation coefficient using data from 10
regions (xt denotes ’age’ and yt denotes ’salary’). First, the means of the data are
estimated. Next, the means are subtracted from the observations and the product
of the resulting deviations from the mean is calculated. Dividing the sum of these
products (585080) by 9 (=n−1) yields the covariance syx =65009. The correlation ryx
is computed by dividing the covariance by the product of the standard deviations:
ryx =65009/(11257.4·10.9)=0.53. The covariance is measured in [units of y]×[units
of x]. The correlation coefficient has no dimension. The the correlation coefficient
using all available data in the present example is 0.59.
The correlation coefficient can also be computed from standardized values or scores.
The standardization

yt0 = (yt − ȳ)/s.

transforms the original values such that yt0 has mean zero and variance one. The
covariance between yt0 and x0t is equal to the correlation between yt and xt .
If more than two variables are considered, the covariances and correlations among
all pairs of variables are summarized in matrices. For example, the variance-
covariance matrix C and the correlation matrix R for three variables yt , xt
and zt have the following structure:

s2y syx syz

   
1 ryx ryz
C =  sxy s2x sxz  R =  rxy 1 rxz  .
szy szx s2z rzy rzx 1
If the observations are not normally distributed it may happen that the correlation
coefficient indicates no relation although, in fact, there is a nonlinear relation be-
tween yt and xt . In this case the rank correlation can be used. The rank of each
6.2 Simple linear regression 54

observation in the sorted sequence of yt and xt is determined (see Table 4). The
rank correlation is computed using the differences among ranks dt :

n
6 X
rr = 1 − d2t .
n(n2 − 1)
t=1

If the rank of both variables are identical rr =1. If the ranks are exactly inverse
rr =−1. In the present case the rank correlation hardly differs from the ’regular’
(linear) correlation.

6.2 Simple linear regression

A significant correlation coefficient between two variables (e.g. age and weight of
children) does not allow to draw conclusions about the expected value of one of the
two variables based on specific values of the other variable (e.g. expected weight in
kg for an age of five years). The positive correlation between age and salaries from
Figure 12 is not sufficient to compute the expected (average) salary for a specific
age of an individual. To answer this kind of questions requires a regression model.
The following distinction is made for that purpose. One variable (Y ) – the variable
of main interest – is considered to be the dependent variable. The other variable
(X) is assumed to affect Y . The regression model allows to make statements about
the mean of Y conditional on observing a specific value of the explanatory (or
independent) variable X. If there is, in fact, a dependence on X the conditional
mean will differ from the unconditional mean ȳ which results without taking X into
account. A forecast of Y , which is based on a specific – assumed or observed –
value of X is called a conditional mean.
To illustrate the analysis, every observed pair (yt ,xt ) is represented in a scatter
diagram (see Figure 13). The diagram can be used to draw conclusions about
the strength of the relationship between Y and X which can be measured by the
correlation coefficient. However, the correlation between yt and xt is not sufficient
to obtain a specific value for the conditional mean. For that purpose the scatter of
data pairs can be approximated by a straight line. This corresponds to condensing
the information contained in individual cases. If the straight line is a permissible
and suitable simplification, it can be used to make statements about Y on the basis
of X. The simplification is not without cost, however, since not every observation
yt can be predicted exactly. On the other hand, without this simplification, only a
set of individual cases is available that does not allow general conclusions.
Approximating the scatter of points by a straight line is based on the assumption
that yt can be described (or explained) using xt in a simple linear regression
model
6.2 Simple linear regression 55

Figure 13: Data points and simple linear regression model.

y
...
......
◦ ......
◦..........................
ŷt
........ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...................
..
......
..
........ ..
.
.
...... ..
......
......
......
..
..
et
yt ◦ ...
......
...
....... ◦ ..
........ . . . . . . . . . . . . . . . . . . . . . . . . . . ............ . . . . . . . . ..........
◦ ......
..........
.
◦ .
.. .
...
. .
...... ..
.
...
....... . .

◦ .
......
......
.◦
.
......
. .
.
.
. ◦ .
.
.
...
......
...
. .
.
.
.
b .
.
...
..... .
........
. .
.
. .
.
............ ....... ....... ....... ....... ................ .
◦ ◦ .
...
......
...
....... ..
1 .
.
.
...
.....
..
......
. .
.
......
...
......
...
....... ◦ .
.
...... .
.
...
.... .
...........
...
......
.
...
.
◦ .
.
.
.
.
.
c .
.
.
..xt
..
x
0

simple linear regression: yt = c + b · xt + et = ŷt + et (t = 1, . . . , n).

ŷt is the fitted value (or the fit) and depends on xt . et is the error or residual and
is equal to the difference between the observation yt and the corresponding value on
the line ŷt =c+b·xt .
The coefficients c and b determine the level and slope of the line (see Figure 13).
A large number of similar straight lines can approximate the scatter of points. The
least-squares principle (LS) can be used to fix the exact position of the line.
This principle selects a ’plausible’ approximation. The LS criterion states that the
coefficients c and b are determined such that the sum of squared errors is minimized:

n
X
least-squares principle: e2t −→ min .
t=1

Using this principle it can be shown that the slope estimate is based on the covariance
between yt and xt and the variance of xt and can also be computed using the
correlation coefficient:

sy syx
slope: b = ryx = 2 .
sx sx
6.2 Simple linear regression 56

The coefficient b can be interpreted as follows: if xt changes by ∆x units, ŷt – the

conditional mean of Y – changes by b·∆x units. Note that the change in ŷt does not
depend on the initial level of xt .
The definition of the slope implies that its dimension is given by [units of yt ] per
[unit of xt ]. This property distinguishes the regression coefficient from the correlation
coefficient, which has no dimension.
The intercept61 (or constant term) c depends on the means of the variables and
on b:

intercept: c = ȳ − b · x̄.

This definition guarantees that the average error equals zero. c has the same dimen-
sion as yt .
Errors et =yt −ŷt occur for the following reasons (among others): (a) X is not the only
variable that affects Y . If more that one variable affects Y a multiple regression
analysis is required. (b) A straight line is only one out of many possible functions
and can be less suitable than other functions.
The coefficients c and b can be used to determine the conditional mean ŷ under the
condition that a particular value of xt is given:

conditional mean (fit): ŷt = c + b · xt .

ŷt replaces the (unconditional) mean ȳ, which does not depend on X. In other words,
only the mean ȳ is available if X is ignored in the forecast of Y . Using the mean
ȳ corresponds to approximating the scatter of points by a horizontal line. If the
regression model turns out to be adequate – if X is a suitable explanatory variable
and a straight line is a suitable function – the horizontal line ȳ is replaced by the
sloping line ŷt =c+b·xt .

Least-squares estimates of a regression model can be obtained from ’Tools/Data

Analysis/Regression’62 .
The required input is the data range of the dependent variable (’Input Y
Range’) and the explanatory variable(s) (’Input X Range’)63 It is useful to
include the variable name in the first row of the data range. In this case the
field ’Labels’64 must be activated.
61
Schnittpunkt
62
’Extras/Analyse-Funktionen/Regression’
63
’Y-Eingabebereich’; ’X-Eingabebereich’
64
’Beschriftungen’
6.2 Simple linear regression 57

Example 25: We consider the data from example 1 and run a simple
regression using ’salary’ as the dependent variable and ’age’ as the ex-
planatory variable. The scatter of observations and the regression line
are shown in Figure 14. The regression line results from a least-squares
estimation of the regression coefficients. Estimation can be done with
suitable software. The results in Figure 15 are obtained with Excel.
The resulting output contains a lot of information which will now be
explained using the results from this example.

Figure 14: Scatter diagram of ’age’ versus ’salary’ and regression line.
90000

80000 y^ = 634x + 20610

R2 = 0.34
70000
salary (y)

60000

50000

40000

30000
25 30 35 40 45 50 55 60 65 70

age (x)

Figure 15: Estimation results for the simple regression model.

Regression Statistics
Multiple R 0.59
R Square 0.34
Adjusted R Sq. 0.32
Standard Error 9480
Observations 30

Standard
Coefficients Error t Stat P-value Lower 95% Upper 95%
Intercept 20610 8457 2.44 0.021 3286 37933
Age 634 166 3.82 0.001 294 974
6.3 Regression coefficients and significance tests 58

6.3 Regression coefficients and significance tests

Estimated coefficients

The estimated coefficients are 634 (b) and 20610 (c). In order to interpret these
values assume that the current age of a respondent is 58 (this is the first observation
in the sample). The estimated regression equation

yt = 20610 + 634 · xt + et

can be used to compute the conditional mean of salaries for this age: 20610+634·58=
57377. This value is located on the line in Figure 14. The observed salary for
this person is 65400. The error et =yt −ŷt is given by 65400−57377=8023; it is the
difference between the observed salary (yt ) and the (conditional) expected salary
(ŷt ). The discrepancy is due to the fact that the regression equation represents an
average across the sample. In addition, it is due to other explanatory variables which
are not (or cannot be) accounted for.
If age increases by one year the salary increases on average by $634 (or, the condi-
tional expected salary increases by $634). If we consider a person who is five years
older, the conditional mean increases to 20610+634·(58+5)=60547; i.e. its value in-
creases by 634·5=3170. Thus, the slope b determines the change in the conditional
mean. If xt (age) changes by ∆x units, the conditional mean increases by b·∆x
units. Note that the (initial) level of xt (or yt ) is irrelevant for the computed change
in ŷ.
The intercept (or constant) c is equal to the conditional mean of yt if xt = 0. The
estimate for c in the present example is 20610 which corresponds to the expected
salary at birth (i.e. at an age of zero). This interpretation is not very meaningful if
the X-variable cannot attain or hardly ever attains a value of zero. It may not be
meaningful either, if the observed values of the X-variable in the sample are too far
away from zero, and thus provide no representative basis for this interpretation.
The role of the intercept can be derived from its definition c=ȳ−b·x̄. This implies
that the conditional expected value ŷt is equal to the unconditional mean of yt if xt is
equal to its unconditional mean. The sample means of yt and xt are 52263 and 49.9,
respectively, which agrees with the regression equation: 20610+634·49.9=52263.

Standard error of coefficients and significance tests

If the sample mean ȳ is used instead of the population mean µy an estimation error
results. For the same reason the position of the regression line is subject to an error,
since c and b are estimated coefficients. If data from a different sample was used,
the estimates c and b would change. The standard errors (the standard deviation of
6.3 Regression coefficients and significance tests 59

estimated coefficients) take into account that the coefficients are estimated from a
sample.
√
When the mean is estimated from a sample the standard error is given by s/ n.
In a regression model the standard error of a coefficients decreases as the standard
deviation of residuals se (see below) decreases and the standard deviation of xt
increases. The standard error of the slope b in a simple regression is given by

s
sb = √e .
sx n − 1
The standard errors of b and c are 166 and 8457 (see Figure 15). These standard
errors can be used to compute confidence intervals for the values of the constant
term and the slope in the population. The 95% confidence interval for the slope is
given by

b ± 1.96 · sb .

This range contains the slope of the population β with 95% probability (given the
estimates derived from the sample). The confidence interval can be used for testing
the significance of the estimated coefficients. Usually the null hypothesis is β0 =0;
i.e the coefficient associated with the explanatory variable in the population is zero.
If the confidence interval does not include zero the null hypothesis is rejected and
the coefficient is considered to be significant (significantly different from zero).
The boundaries of the 95% confidence interval for b are both above zero. Therefore
the null hypothesis for b is rejected and the slope is said to be significantly different
from zero. This means that age has a statistically significant impact on salaries.
The constant term is also significant because zero is not included in the confidence
interval. Note that the mean of residuals equals zero if a constant term is included
in the model. Therefore the constant term is usually kept in a regression model even
if it is insignificant.
If the explanatory variable in a simple regression model has no significant impact
on yt (i.e. the slope is not significantly different from zero), there is no significant
difference between conditional and unconditional mean (ŷt and ȳ). If that was the
case one would need to look for another suitable explanatory variables.
Significance tests can also be based on the t-statistic

b − β0
t= .
sb
The t-statistic corresponds to the standardized test statistic in section 5.3. The null
hypothesis is rejected, if t is ’large enough’, i.e. if it is beyond the critical value at
6.4 Goodness of fit 60

the specified significance level. The critical values at the 5% level are ±1.96 for large
samples.
The t-statistics for b is 3.82. The null hypothesis for b is rejected and the slope
is significantly different from zero. The constant term c is significant, too. These
conclusions have to agree with those derived from confidence intervals.
Significance tests can be based on p-values, too.65 As explained in section 5.3 the
p-value is the probability of making a type I error if the null is rejected. For a given
significance level, conclusions based on the t-statistic and the p-value are identical.
For example, if a 5% significance level is used, the null is rejected if the p-value is
less than 0.05.
In the present case the p-values of the slope coefficient is almost zero. In other
words, if the null hypothesis ”the coefficient equals zero” is rejected, there is a very
small probability to make a type I error. Therefore the null hypothesis is rejected
and the explanatory variable is kept in the model.
If the null hypothesis was rejected for the constant term the probability for a type I
error would equal 2.1%. Since this is less than α the null hypothesis is rejected and
the constant term is considered to be significant.

6.4 Goodness of fit

Standard error of regression

Approximating the observations of yt with a regression equation implies errors

et =yt −ŷt . The information in xt is not sufficient to match the value of yt in each
and every case. Therefore the regression model only explains a part of the variance
in yt . A measure for the unexplained part is the variance of residuals:

n
1 X
s2e = e2t ,
n−k−1
t=1

where k is the number of explanatory variables. se is usually called standard error

(of the regression) and must not be confused with the standard error of a coefficient.
se has the same units of measurement as yt and et . It can be compared to sy , the
standard deviation of the dependent variable. sy is based on the deviations of yt
from the (unconditional) mean ȳ. If se is small compared to sy , the conditional
mean ŷ provides a much better explanation for yt than ȳ. If se is almost equal to sy ,
there is hardly any difference between the unconditional and the conditional mean.
In other words, the regression model does not explain much more than the sample
65
The p-values in the regression output are based on the t-distribution rather than the standard
normal distribution. If n is large (above 120) the two distributions are almost identical.
6.5 Estimating the CAPM 61

mean. Thus a comparison of se and sy allows conclusions about the explanatory

power of the model. This comparison has the advantage that both statistics have
the same units of measurement as the dependent variable.
In the present example the standard error (of the regression) se is 9480. The standard
deviation of yt (of salaries) is 11493. This difference is not very large. If expected
salaries are computed using the age of respondents, the associated errors are not
much less than using the (unconditional) mean salaries (ȳ).

Multiple correlation coefficient

The multiple correlation coefficient measures the correlation between the observed
value yt and the conditional mean (the fit) ŷt . The multiple correlation coefficient
approaches one as the fit improves. The number 0.59 (see Figure 15) indicates an
acceptable, although not very high explanatory power of the model.

Coefficient of determination R2

The coefficient of determination66 R2 is another measure for the goodness of fit of

the model. R2 measures the percentage of variance of yt that is explained by the
X-variable. It compares the variance of errors and data:

(n − k − 1) s2e
R2 = 1 − 0 ≤ R2 ≤ 1.
(n − 1) s2y

R2 ranges from zero (the errors variance is equal to the variance of yt ) and one (error
variance is zero). The number 0.34 (see Figure 15) shows that 34% of the variance
in salaries can be explained by the variance in age.
Note, however, that high values of the multiple correlation coefficient and R2 do not
necessarily indicate that the regression model is adequate. There exist further crite-
ria to judge the adequacy of a model, which are not treated in this text, however.

6.5 Estimating the CAPM67

According to the Capital Asset Pricing Model (CAPM) the following relation holds
between the expected return of an asset (µi ), the risk free return rf and the expected
return of the market portfolio (µm ):

µi = rf + βi (µm − rf ).
66
Bestimmtheitsmaß
67
Only relevant for banking, finance or similar courses.
6.5 Estimating the CAPM 62

In the CAPM the so called beta-coefficient βi is the appropriate measure of risk

of the asset rather than the variance of the asset’s returns. To estimate βi or test
the CAPM one can use observed returns of the asset (yti ) and the market portfolio
(ytm ). Under the assumption that observed returns deviate from expected returns
we obtain

yti − µi = uit ytm − µm = um

t ,

where uit and um

t are mean zero random variables. When we replace expected values
in the CAPM by observed returns we obtain the so called market model

yti = αi + βi ytm + et ,

where αi =(1−βi )rf and et =uit −βi um

t . The coefficients αi and βi in this equation
can be estimated using a linear regression analysis. The beta-coefficient βi can also
be estimated using the formula

sim
βi = ,
s2m

where sim is the (sample) covariance between the returns of asset i and the market
return and s2m is the (sample) variance of the market return. This formula can also
be derived on the basis of financial theory.
As an example we use 80 monthly observations of ATX index values and stock prices
of Bank-Austria (BA) from January 1986 until August 1992. Using log returns the
estimated regression equation is (standard errors in parentheses):

ŷti = 0.076 + 0.883 ytm R2 = 0.693 se = 4.62 si = 8.3.

(0.52) (0.067)

The estimated coefficient βi is 0.883. The estimate is highly significant. The stan-
dard error of the coefficient equals 0.067 which implies a t-statistic of 13.2.
The coefficient 0.883 can be interpreted as follows. A change in the market return
by 10 percentage points implies a change in the asset’s expected return by 8.83
percentage points. However, if the market return equals 10% the expected return of
the asset is given by 8.91% (0.076+0.883 · 10.0=8.91).
R2 equals 0.693 which implies that almost 70% of the variance in the asset’s return
can be explained by – or is due to – the market return. Based on the market model
6.6 Multiple regression analysis 63

the total variance of an asset can be split into market- and firm-specific variance as
follows:

s2i = βi2 s2m + s2e .

Since R2 can also be written as (βi2 s2m )/s2i it measures the proportion of the market-
specific variance to total variance. The R2 of the example shows that about 70% of
the asset’s total variance is systematic or market-specific (not diversifiable risk) and
about 30% are unsystematic or firm-specific (diversifiable risk).

Value-at-risk and the CAPM

The value-at-risk of an asset can be computed on the basis of its beta-coefficient.
In section 4.8 we have used the total variance s2y as a measure of risk. However,
VaR attempts to measure the market risk of an asset or portfolio. According to the
CAPM the market-specific risk is given by β 2 s2m . Thus the VaR formula becomes

VaR(α) = −pt [exp{ȳ + zα βsm } − 1].

In the simplified RiskMetrics approach the formula is given by

VaRRM (α) = −pt · zα · β · sm .

Using the data from the present example the 5%-VaR of the BA stock can be com-
puted from

VaR = −500[exp{0.0025 − 1.645 · 0.883 · 0.078} − 1] = 52.69

or

VaRRM (α) = −500 · −1.645 · 0.883 · 0.078 = 56.67.

6.6 Multiple regression analysis

Frequently an appropriate description and explanation of a variable of interest re-
quires to use several explanatory variables. In this case it is necessary to carry
out a multiple regression analysis. If observations for k explanatory variables
x1 , . . . , xk are available the coefficients c, b1 , . . . , bk of the regression equation
6.6 Multiple regression analysis 64

y = c + b1 · x1 + · · · + bk · xk + e

can be estimated using the least-squares principle.

The interpretation of the coefficients b1 , . . . , bk is different from a simple regression.
bk measures the change in ŷt if the k-th X-variable changes by one unit and all other
X-variables stay constant (ceteris paribus (c.p.) condition). In general the change
in ŷt as a result of changes in several explanatory variables ∆xi units is given by

∆ŷ = b1 · ∆x1 + · · · + bk · ∆xk .

The intercept c is the fitted value for yt if all X-variables are equal to zero. At the
same time it is the difference between the mean of yt and the value of ŷt that results
if all X-variables are equal to their respective means:

c = ȳ − (b1 · x̄1 + · · · + bk · x̄k ).

The coefficients from simple and multiple regressions differ when the explanatory
variables are correlated. A coefficient from a multiple regression measures the effect
of a variable by holding all other variables in the model constant (c.p. condition).
Thus, by taking into account the simultaneous variation of all other explanatory
variables, the multiple regression measures the ’net effect’ of each variable. The
effect of variables which do not appear in the model cannot be taken into account
in this sense. A simple regression ignores the effects of all other (ignored) variables
and assigns their joint impact on the single variable in the model. Therefore the
estimated coefficient (slope) in a simple regression is generally too small or too large.

Example 26: Obviously, a person’s salary not only depends upon age,
but also on factors like ability and qualifications. This aspect can be
measured (at least roughly) by the education time (schooling). A multi-
ple regression will now be used to assess the relative importance of age
and schooling for salaries.

The results of the multiple regression between salary and the explanatory variables
’age’ and ’schooling’ are summarized in Figure 16. By judging from the p-values we
conclude that both explanatory variables have a significant effect on salaries.
An increase in schooling by one year leads to an increase in expected salaries by
$1501, if age is held constant (ceteris paribus; i.e. for individuals with the same
age). The coefficient 723 for ’age’ can be interpreted as the expected increase in
6.6 Multiple regression analysis 65

Figure 16: Estimation results for the multiple regression model.

Regression Statistics
Multiple R 0.73
R Square 0.53
Adjusted R Sq. 0.50
Standard Error 8150
Observations 30

Standard
Coefficients Error t Stat P-value Lower 95% Upper 95%
Intercept -2350 10064 -0.23 0.817 -23000 18299
Age 723 145 4.98 0.000 425 1021
School 1501 455 3.30 0.003 567 2434

salaries induced by getting older by one year, if education does not change (i.e. for
people with the same education duration). Note that this effect is stronger than
estimated in the simple regression (see Figure 15). The estimate 723 in the current
regression can be interpreted as the net-effect of one additional year on expected
salaries accounting for schooling. If salaries are only related to age (as done in the
simple regression) the effects of schooling on salaries are erroneously assigned to the
variable ’age’ (since it is the only explanatory variable in the model).
To measure the joint effects from several variables we use the general formula

∆ŷ = b1 · ∆x1 + · · · + bk · ∆xk .

For example, comparing two individuals with different age (10 years) and schooling
(2 years) shows that expected salaries differ by 723·10+1501·2=10232.

Example 27: Frequently, concerns are raised about gender discrimina-

tion. This may show in significantly lower salaries of women compared
to men. The data from example 1 can be used to test these concerns.
The variable ’G01’ is added to the regression which is assigned a value
of 1 for women and 0 for men.

The variable ’G01’ is a dummy-variable. The 0-1 coding allows for a meaningful
application and interpretation in a regression. Adding the variable G01 to the
multiple regression equation and estimating the coefficients yields the results in
Figure 17. The coefficient of G01 is −5601. It shows that women earn on average
$5601 less than men, holding everything else constant (i.e. compared to men with
the same age and education). This negative effect is not as significant as the effects
of age and schooling as indicated by the p-value 6.6%. Using a significance level of
5% the gender-specific difference in salaries is not statistically significant.
6.6 Multiple regression analysis 66

Figure 17: Estimation results for the multiple regression model including a dummy
variable ’G01’ for gender.
Regression Statistics
Multiple R 0.77
R Square 0.59
Adjusted R Sq. 0.54
Standard Error 7771
Observations 30

Standard
Coefficients Error t Stat P-value Lower 95% Upper 95%
Intercept 1367 9790 0.14 0.890 -18756 21490
Age 694 139 4.98 0.000 408 980
School 1500 434 3.46 0.002 609 2392
G01 -5601 2915 -1.92 0.066 -11592 390

Adding the third explanatory variable to the regression equation leads to a reduction
in the standard error from 9480 in the simple regression to 7771. Accordingly, the
coefficient of determination increases from R2 =0.34 to R2 =0.59. Note however, that
R2 always increases when additional explanatory variables are included. In order to
compare models with a different number of X-variables the adjusted coefficient
of determination R̄2 should be used:

s2e
R̄2 = 1 − .
s2y

Out of several estimated multiple regression models the one with the maximum
adjusted R2 can be selected. Note, however, that there is a large number of other
criteria available to select among competing models.

Model specification and variable selection

When a coefficient in a regression model is found to be insignificant the correspond-
ing variable can be eliminated from the equation. Eliminating variables usually
affects the coefficients of the remaining variables. The same is true for including
additional variables which affects the coefficients of the original variables. This can
be explained as follows. Assume that several X-variables actually affect Y , but only
some of these X-variables are included in the model. Thus, a part of the effect of
the omitted variables is assigned to the included variables. The coefficients of the
included variables do not only have to carry their own effect on Y , but also the ef-
fect of omitted variables. This bias in the estimated coefficients results whenever the
included and omitted variables are correlated. In general the omission of relevant
variables has more severe disadvantages than the inclusion of irrelevant variables.
6.6 Multiple regression analysis 67

Given that a regression model contains some insignificant coefficients the following
guidelines can be used to select variables.

1. The selection of variables must not be based on simple correlations between

the dependent variable and potential regressors. Because of the bias associated
with omitted variables any selection should be done in the context of estimating
multiple regressions.

2. Coefficients having a p-value above the pre-specified significance level indicate

variables to be excluded. If several variables are insignificant it is recommended
to eliminate one variable at a time. One can start with the variable having
the largest p-value, re-estimate the model and check the p-values again (and
possibly eliminate further variables).

3. If the p-value indicates elimination but the associated variable is considered to

be of key importance theoretically, the variable should be kept in the model
(in particular if the p-value is not far above the significance level). A failure to
find significant coefficients may be due to insufficient data or a random sample
effect (bad luck).

4. Statistical significance alone is not sufficient. There should be a very good

reason for a regressor to be included in a model and its coefficient should have
the expected sign.

5. Adding a regressor will always lead to an increase of R2 . Thus, R2 is not

a useful selection criterion. If a variable with a t-statistic less than one is
eliminated, the standard error of the regression (se ) drops and R̄2 increases.
This criterion is suitable when the primary goal of the analysis is to find a well
fitting model (rather than to search for significant relationships).
6.6 Multiple regression analysis 68

7 Decision analysis
Example 28: NEWDEC has developed and carefully tested a new elec-
tronic device. The marketing department is currently discussing a suit-
able marketing strategy. NEWDEC is aware of the fact that a selected
strategy need not necessarily achieve the desired results. Consumer
tastes and short-terms fashions are hard to predict. In addition, the
main competitor – known to set prices well below NEWDEC – may also
be about to introduce a new device.
NEWDEC considers a range of marketing activities which may be char-
acterized in terms of three strategies:
An aggressive strategy entails substantial advertising expenditures and
aggressive pricing. This strategy includes hiring additional staff and
investing in additional production facilities to cope with the increased
demand associated with a successful marketing campaign.
In the basic strategy advertising levels will be moderately increased for
a few weeks. This will be supported by reduced prices during the intro-
ductory phase. Existing production facilities are planned to be modified
and slightly expanded. Only a limited number of additional staff is re-
quired in this case.
A cautious strategy is mainly based on the use of existing production
facilities and does not require to hire new personnel. Advertising would
be mainly done by local sales representatives.
The current market situation – which refers to the actual but unknown
state of the market – is unknown to NEWDEC. However, to facilitate the
search for a suitable marketing strategy, the following three possibilities
are considered: the readiness of the market to accept the new product
is considered to be high, medium or low. These categories are mainly
based on sales forecasts whereupon probabilities can be assigned to each
category.
The management at NEWDEC carefully evaluates each possible case and
determines its monetary consequences (in terms of present values of the
coming two years). Expected payoffs (positive or negative cash-flows)
are summarized in Table 5.

This problem – the optimal choice of a marketing strategy given uncertainty about
the market conditions – can be solved using a decision analysis. One out of m
mutually exclusive alternatives (Ai , i=1, . . . , m) is chosen by taking into account
n uncertain, exogenous states (Zj , j=1, . . . , n). For each pair Ai -Zj the monetary
consequences (payoffs) must be specified. The choice is based on a suitable decision
criterion.
Note the sequence of steps associated with this approach. The decision is made (e.g.
7.1 Elements of a decision model 69

Table 5: Payoff-matrix.
acceptance level
high medium low
strategy Z1 Z2 Z3
aggressive A1 120 50 –40
basic A2 80 60 20
cautious A3 30 35 40

Table 6: Elements of a decision model.

states
Z1 Z2 ... Zn
probabilities p1 p2 ... pn
decisions A1 C11 C12 . . . C1n
A2 C21 C22 . . . C2n
.. .. .. .. ..
. . . . .
Am Cm1 Cm2 ... Cmn

alternative A3 is chosen). Then one of the anticipated states is actually taking place
(e.g. Z2 ). Finally, the monetary consequences associated with the combination A3
and Z2 are realized. If the choice has an effect on the states of nature or the number
of possible states, the decision problem can be solved using a decision tree.

7.1 Elements of a decision model

A decision model is based on the following elements (summarized in Table 6):

1. Several, uniquely defined decisions or alternatives Ai , i=1, . . . , m (e.g. mar-

keting strategies or investment projects). The decisions must be under the
control of the decision maker.
2. Several states (of nature) Zj , j=1, . . . , n. These may not be under the control
of the decision maker (e.g. the business cycle). If the decisions have a (partial)
impact on the states and probabilities (see next item), a decision tree can be
used to describe the problem.
3. Probabilities pj (j=1, . . . , n) associated which each state (e.g. the probability
for a recession is 30% whereas the probability for a recovering economy is
70%). Probabilities can be based on theory (e.g. rolling dice), historical data
or on subjective experience (intuition).
From historical data the relative frequency of an event can be derived. If the
number of observations is large enough it is possible to make statements like
7.1 Elements of a decision model 70

Figure 18: Two probability distributions, representing different levels of risk.

(a) high risk (b) low risk
p6 p6

Z1 Z2 Z3 Z1 Z2 Z3

”Out of the previous 120 months sales dropped in 36 months.” The relative
frequency 36/120=0.3 can be used to set the probability of the state ’sales
down’ at 30%.
The probabilities in the present example are based on historical data: 25% for
state ”low”, 40% for state ”medium”, and 35% for state ”high”.
Probabilities describe the degree of available information of a decision maker
and can be used to characterize the decision problem as follows:

(a) Decisions under certainty: The probability for one of the states is one
and zero for all others.
(b) Decisions under uncertainty: It is not possible to specify probabilities at
all. Such cases are solved by referring to decision rules.
(c) Decisions under risk : Probabilities (different from one) can be assigned
to each state. The riskiness of a decision problem can be characterized on
the basis of the probability distribution. If all probabilities are about the
same – as in Figure 18(a) – the involved risk is larger than in a situation
where one of the states has a relatively high probability (e.g. state Z2 in
Figure 18(b)).

4. A decision model assigns outcomes and payoffs to each pair (Ai , Zj ). As a

starting point the outcomes are described in words (e.g. if alternative Ai is
chosen and the economy is, in fact, recovering (state Zj ) sales will go up). To
derive an optimal solution it is necessary, however, to quantify the monetary
consequences in terms of a cash-flow (or payoff) Cij . The payoffs are defined
as the difference between cash inflows and outflows associated with the i-th
decision and the j-th state. The payoff matrix has to be complete (i.e. all
entries must be filled with numbers).

5. A criterion to select the optimal decision (e.g. maximizing expected wealth).

Commonly used criteria will be discussed in the following sections.
7.2 Decisions under uncertainty 71

Table 7: Payoff matrix.

state
decision Z1 Z2 Z3 Z4 Z5 Z6 Z7 Z8 Z9 Z10
A1 100 100 100 100 9 100 100 100 100 100
A2 10 10 10 10 10 10 10 10 10 10

7.2 Decisions under uncertainty

Decisions under uncertainty can be solved with decision rules. Out of a large
number of suggested rules we consider some of the most frequently used.
The maximin-criterion proceeds in two steps. First, the decision with the mini-
mum payoff in each state (the worst case across states) is determined. Second, the
maximum out of these payoffs points at the optimal decision. This is a pessimistic
criterion choosing ’the best out of the worst’. In formal terms this is given by

max min Cij .

i j

According to the maximax-criterion the maximum payoff (across states) for each
alternative is determined first. The optimal decision is the one which points at the
maximum out of these. That is why this criterion is considered to be optimistic.
The choice can be formalized by

max max Cij .

i j

The Laplace-criterion assigns the same importance to each state. The decision
with the maximum (unweighted) average payoff is chosen:

n
1X
max Cij .
i n
j=1

These criteria have been criticized by using examples which lead to unacceptable
solutions. Consider the payoff-matrix in Table 7. According to the maximin-criterion
alternative A2 is optimal. However, it is very plausible that even pessimists would
chose A1 because it dominates A2 in almost every case. In addition, the payoff
associated with A1 in state Z5 (in which A2 is preferable) is not much worse. Given
that a criterion leads to questionable results in such simple cases it is hard to justify
its use in more complex situations.
7.3 Decisions under risk 72

7.3 Decisions under risk

Given the potential difficulties associated with decision rules it may be worth while
trying to assign probabilities to states. Decisions under risk are characterized by
probabilities assigned to states. Such problems can be solved on the basis of ex-
pected values and the decision with the maximum expected value is chosen.
The expected value of an alternative Ai is the weighted sum of payoffs across states
using the probabilities pj as weights:

n
X
µi = pj Cij .
j=1

One important aspect is ignored if decisions are made on the basis of expected
values: the variability of outcomes across states. In the present example the payoffs
of the cautious strategy are rather similar across states. The aggressive strategy is
characterized by a substantial variation of payoffs.
This fact can be accounted for by the variance (or standard deviation) of payoffs.
The variance of alternative Ai is defined as

n
X
σi2 = pj (Cij − µi )2 .
j=1

A frequently used decision criterion can be defined in terms of mean and variance
of payoffs. The optimal decision is based on the value of

µi − lσi2 .

l is a factor which determines the trade-off between expectation and variance, and
depends on the risk aversion of the decision maker. More risk aversion is taken into
account by higher values of l. Thereby the subjective risk attitude of the decision
maker is taken into account.
The value assigned to l may be derived from interviewing the decision maker. The
questions are formulated in such a way that the personal trade-off between µ and σ 2
can be estimated. A frequently used approach is to present the decision maker with a
game. For example, he may win 30 or 60 units with 50% probability. The magnitude
of these amounts should correspond to the relevant magnitudes in the current, actual
decision problem. The decision maker is asked to specify a certain amount which
generates the same utility as playing the game (with uncertain outcomes). Suppose
7.3 Decisions under risk 73

Figure 19: Results of the decision analysis to select a marketing strategy.

Payoff matrix

states> high medium low

p(state) 35% 40% 25%
strategies µ σ min max Laplace µ−σ
aggressive 120.0 50.0 -40.0 52.0 61.1 -40.0 120.0 43.3 -31.0
basic 80.0 60.0 20.0 57.0 23.0 20.0 80.0 53.3 45.2
cautious 30.0 35.0 40.0 34.5 3.8 30.0 40.0 35.0 34.2
max 120.0 60.0 40.0 76.0
λ 0.022
criterion optimal decision value
maximin cautious 30.0
maximax aggressive 120.0
Laplace basic 53.3
expected value basic 57.0
µ−σ basic 45.2

the decision maker states an amount equal to 40. This amount is the so-called
certainty equivalent. The expected value of the game is given by 45 (0.5 · 30+0.5 ·
60=45). It is higher than the certainty equivalent and the difference is due to the
risk associated with the game. l can be derived from the equation µ−lσ 2 =40. Since
the game implies µ=45 and σ=15 we obtain l=0.02̇.
Figure 19 presents the results from applying various criteria to NEWDEC’s decision
problem. According to the pessimistic maximin-criterion the cautious strategy is
chosen, the optimistic maximax-criterion chooses the aggressive marketing strategy,
and applying the Laplace-criterion yields the basic strategy.
If probabilities are taken into account, decisions can be based on maximizing the
expected payoff µ. It turns out that the basic strategy has the maximum expected
value of 57. If NEWDEC considers this criterion as appropriate, choosing the basic
strategy can be viewed as being indifferent between a certain amount of 57, and
receiving uncertain outcomes of 80, 60 or 20 (with the associated probabilities). The
expected value of the aggressive strategy has similar order of magnitude, whereas
the expectation of the cautious strategy is much lower.
Maximizing the expected value does not take into account the risk aversion of a
decision maker. In fact, it can be shown that this criterion is only appropriate if she
or he is risk neutral. Accounting for mean and variance (and thereby for the risk
attitude) overcomes this deficiency. In the present case, the basic strategy is chosen
according to the µ-σ criterion. The cautious strategy would only be chosen for a
higher degree of risk aversion (e.g. λ=0.05).
A second possibility to account for the risk attitude of a decision maker is to max-
imize expected utility. According to the Bernoulli-criterion all payoffs are
replaced by utility units as shown below. Utility can be viewed as a measure of the
7.3 Decisions under risk 74

Figure 20: Utility function in case of risk aversion.

U
1.00 ...
............
......... ...
......... .... ... ...
.........
......... .......
......... ......
...
...
......... .
...... ......
.
........
........ ......
........
0.75 . . . . . . . . . . . . . . . . . . . . . . . . . . . . ................ . . . . ....... ◦.
.
....... .
.
...... .
....... ......
....... . .
....... .......
.
...
........ .
.
......
. ...... . .
.
...
....... ....
... . .
...
...... ....
..
..... ... . .
0.50 ......
.
.......
.
.
.
.
......
.
.
..... .
.
.
.
...
....... .
... . .
..
...... .
..
..... ..... . .
...
...
..... .
...... . .
..... ......
. . .
.
..
...... . . .
.... .....
0.25 ..
.. .
...
.
.. .
.....
.....
..
..... .
.
.
.
..... .. . .
..
.....
. ......
. . .
..
..... ....... . .
..
.... ......
..
.... .. . .
... .... . .
....
.......
0.00 . .

−40 −20 0 20 40 60 80 100 120

attractiveness of an outcome.
A utility function reflects the risk attitude of a decision maker. It can be derived in
a similar way as the trade-off parameter λ. The units of measurement are arbitrary.
It is useful to assign a value of one to the maximum payoff and zero to the minimum
payoff. From Table 5 we find U (120)=1 and U (−40)=0 (see Figure 20). Consider the
utility of the payoff C22 =60, for example. The decision maker is asked to specify a
probability p (between zero and one) such that he is indifferent between the following
alternatives: receive a certain payment 60, or play a game which yields either 120
with probability p or −40 with probability 1−p.
Suppose the decision maker is indifferent if p=0.75. In this case the expected value
of the game is 80 (0.75·120+0.25·−40=80). This amount is larger than the certain68
payoff 60. The decision maker requires a compensation for playing the risky game.
The probability p=0.75 is the utility assigned to the payoff C22 : U (60)=0.75. Using
the same procedure, utilities can be assigned to each payoff from Table 5. Based
on the utility function in Figure 20, the maximum expected utility is found for the
basic strategy (see sheet ’decision analysis’).
The concept of a utility function introduced above can also be applied by using
specific mathematical functions. One example is the exponential utility function

U (W ) = − exp{−W/T }.

W is the monetary outcome, typically the profit or wealth associated with a decision
problem. It is a random variable which is affected by the decisions and the uncertain
68
60 can be viewed as the certainty equivalent of a game with expectation 80.
7.3 Decisions under risk 75

outcomes. T is a coefficient which reflects the risk aversion of the decision maker. In
fact, it is a parameter of risk tolerance, which is inversely related to risk aversion.
As a rough guideline for the choice of T we can refer to empirical evidence (see AWZ,
p.359). Companies were found to choose T approximately as 6% of net sales, 120%
of net income, and 16% of equity.
A decision criterion based on this (or any other) utility function is to maximize the
expected utility (of wealth). This expectation depends on the statistical properties
of W . If W is assumed to be normally distributed it can be shown that maximizing
the expected value of the exponential utility function is equivalent to maximizing

E[W ] − 0.5V[W ]/T.

This corresponds to the µ-σ criterion defined above, with λ=0.5/T .

Example 29: We consider an investor who has an amount w0 avail-

able for investment. She considers investing an amount X into a fund
which yields an annual return R. Returns are assumed to be normally
distributed with µ=0.07 and σ=0.2. The invested amount X can range
from 0 to w0 . The remaining wealth (w0 −X) will be put into a safe
bank account which yields a risk-free rate of rf =0.02. If X=0 (no risky
investment), the resulting wealth W after one year is given by the cer-
tain amount w0 (1+rf ). If X>0 the resulting wealth W depends on the
amount X and the uncertain return R. In general, the (uncertain) wealth
after one year is given by

W = X(1 + R) + (w0 − X)(1 + rf ).

The sheet ”expected utility” contains a set of (simulated) returns, and

the resulting wealth W for a given choice X. The choice of X deter-
mines the distribution of wealth (for given properties of returns). Small
values of X imply a rather narrow distribution; increasing X makes the
distribution of wealth more and more dispersed.
The utility function effectively assigns weights to the different levels of
wealth. Low levels of wealths (or losses) receive a (very) low weight (are
not attractive). High levels of wealth (or gains) receive a higher weight
(are attractive). Since the utility function is concave, the additional util-
ity of increasing gains is comparatively less than drop in utility associated
with increasing losses (of the same size). For high levels of risk aversion
the utility function is rather flat, and starts do decrease sharply if wealth
is below a certain level. Thus, maximizing expected utility can be viewed
as judging the attractiveness of different distributions of wealth.
7.4 Decision trees 76

It can be shown that the optimal choice of X based on maximizing ex-

pected exponential utility (and normally distributed returns) is given
by
µ − rf
X∗ = .
σ 2 /T

For the current situation and a risk tolerance of T =0.125, the investor
should invest X ∗ =0.156. Using the numerical example with simulated
returns we obtain X ∗ =0.16.

7.4 Decision trees

Decision trees provide a convenient and flexible way of describing a decision problem.
Other than the matrix-based approach in the preceding sections, decision trees offer
the possibility to assign different probabilities for different strategies. For example,
in the NEWDEC case it may be appropriate to assume that the aggressive strategy
leads to a higher probability for the ”high” state and a lower probability for the
’low” state. Decision trees are also very useful to describe multi-stage decisions,
where the possibility of revising decisions, or responding to uncertain outcomes is
taken into account.
Decision trees are composed of nodes and branches (see sheet ”SciTool” which rep-
resents the data from example 30). The nodes represent either decisions (squares)
or events (circles). Event (or probability) nodes shows the outcomes when the re-
sult of an uncertain event becomes known. An end node (a triangle) indicates that
the problem is completed – all decisions have been made, all uncertainty have been
resolved and all payoffs have been incurred.
The solution procedure associated with decision trees is called ”folding back on the
tree”. Starting at the right on the tree and working back to the left, the procedure
consists of two types of calculations. At each probability node we calculate the
expected value of payoffs (or utility). At each decision node we find the maximum
of the expected value (or expected utility). Folding back is completed when we have
arrived at the left-most decision node. The maximum expected value (or utility)
indicates the decision to be taken.

Example 3069 : SciTools Inc. specializes in scientific instruments and

has been invited to make a bid on a government contract. The contract
calls for a specific number of these instruments to be delivered during the
coming year. SciTools estimates that it will cost $5000 to prepare a bid
and $95,000 to supply the instruments. On the basis of past contracts,
69
Example 7.1 on page 302 in AWZ.
7.4 Decision trees 77

SciTools believes that the possible bids from the competition (if there is
competition) and the associated probabilities are:

bid probability
less than $115,000 20%
between $115,000 and $120,000 40%
between $120,000 and $125,000 30%
greater than $125,000 10%

There are three elements to SciTools’ problem. The first element is that
they have two basic strategies – submit a bid or do not submit a bid. If
they decide to submit a bid they must determine how much they should
bid. The bid must be greater than $100,000 for SciTools to make a profit.
Given the data on past bids, and to simplify the subsequent calculations,
SciTools considers to bid either $115,000, $120,000, or $125,000.
The next element involves the uncertain outcomes and their probabili-
ties. The only source of uncertainty is the behavior of the competitors
– will they bid and, if so, how much? From past experience SciTools is
able to predict competitor behavior, thus arriving at an estimated 30%
for the probability of no competing bids.
The last element of the problem is the value model that transforms deci-
sions and outcomes into monetary values for SciTools. The value model
in this example is straightforward. If SciTools decides right now not to
bid, then its monetary values is $0 – no gain, no loss. If they make a
bid and are underbid by a competitor, then they lose $5000, the cost of
preparing the bid. If they bid B dollars and win the contract, then they
make a profit of B minus $100,000; that is, B dollars for winning the
bid, less $5000 for preparing the bid, less $95,000 for supplying the in-
struments. The monetary values are summarized in the following payoff
table (all entries in $1000):

competitors’ bid
strategy no bid <115 115–120 120–125 >125
no bid 0 0 0 0 0
bid 115 15 –5 15 15 15
bid 120 20 –5 –5 20 20
bid 125 25 –5 –5 –5 25

Using a decision tree, all possible sequence of steps based on distinguish-

ing decisions and consequences – to bid or not, competitors bid or not,
the bid is won or lost – can be identified. Subsequently, the payoffs for
each possible sequence and the associated probabilities for winning or
7.4 Decision trees 78

losing can be derived. In the present example, the maximum expected

payoff of $12,200 is obtained for bidding $115,000 (see sheet ’SciTools’).
If maximizing expected payoff is considered to be the relevant criterion,
SciTools should be indifferent between a certain amount of $12,200 and
bidding $115,000 – with the associated risk of winning $15,000 or losing
$5,000.
8 REFERENCES 79

8 References
(more recent editions may be available)

comprehensive, many examples, uses Excel: Albright S.C., Winston W.L., und
Zappe C.J. (2002): Managerial Statistics, 1st edition, Wadsworth; the title of
the third edition is Data Analysis and Decision Making.

comprehensive, many examples: Anderson D.R., Sweeney D.J., William T.A., Free-
man J., und Shoesmith E. (2007): Statistics for Business and Economics,
Thomson.

good and simple starting point: Morris C. (2000): Quantitative Approaches in

Business Studies, 5th Edition, Prentice Hall. (an online Excel tutorial can be
found on the companion website www.booksites.net/morris)

a classic (but old-fashioned) textbook: Bleymüller J., Gehlert G. und Gülicher H.

(2002): Statistik für Wirtschaftswissenschaftler, 13. Auflage, Vahlen.

not too technical; for social sciences and psychology: Bortz J. und Döring N. (1995):
Forschungsmethoden und Evaluation, 2. Auflage, Springer.

covers sampling procedures on a basic level and (very) advanced methods; many
examples: Lohr S.L. (2010): Sampling: Design and Analysis, 2nd edition.

I like this one: Neufeld J.L. (2001): Learning Business Statistics with MS Excel,
Prentice Hall.

rather mathematical, but many examples: Brannath W. und Futschik A. (2001):

Statistik für Wirtschaftswissenschaftler, UTB: Wien.

introductory econometrics textbook with many examples: Wooldridge J.M. (2003):

Introductory Econometrics, 2nd Edition, Thomson.

introductory econometrics textbook with many examples: Studenmund A.H. (2001):

Using Econometrics, 4th Edition, Addison Wesley Longman: Boston.

advanced textbook on econometrics and forecasting: Pindyck R.S. und Rubin-

feld D.L. (1991): Econometric Models and Economic Forecasts, 3rd Edition,
McGraw-Hill: New York.

provides a very applied approach to multivariate statistics (including regression

and factor analysis): Hair, Anderson, Tatham, Black (1995): Multivariate Data
Analysis with Readings, Prentice-Hall: New Jersey.
9 EXERCISES 80

9 Exercises
The data for these exercises can be found in the file ’exercises.xls’.

Exercise 170
The Spring Mills Company produces and distributes a wide variety of manufactured
goods. Due to its variety, it has a a large number of customers. Spring Mills
classifies these customers as small, medium and large, depending on the volume of
business each does with them. Recently they have noticed a problem with accounts
receivable. They are not getting paid by their customers in as timely a manner as
they would like. This obviously costs them money.
Spring Mills has gathered data on 280 customer accounts (see sheet ’receive’). For
each of these accounts the data set lists three variables: ’Size’, the size of the cus-
tomer (coded 1 for small, 2 for medium, 3 for large); ’Days’, the number of days
since the customer was billed; ’Amount’, the amount the customer owes.
Consider the variables ’Days’ and ’Amount’, carry out the following calculations and
describe your findings. You may want to distinguish your analysis by the variable
’Size’.

1. Compute all important statistical measures.

2. Compute the histogram (relative frequencies) and compare it to the normal
distribution. You do not need to draw a diagram!
3. Compute the empirical 25%- and 75%-quantiles.
4. Compute the 25%- and 75%-quantiles based on a normal distribution.
5. Compute the 95% interval assuming a normal distribution.
6. What is the probability to observe seven or less days assuming a normal dis-
tribution?
7. What is the probability to observe an amount greater than $750 assuming a
normal distribution?

Exercise 2
A team of physiotherapists wants to test the effectiveness of a new treatment. For
that purpose, a sample of 34 people is randomly selected. A key parameter, which
should respond to the treatment, is measured immediately before and one hour after
treatment.
70
Example 3.9 on page 95 in AWZ.
9 EXERCISES 81

1. Compute the difference if the measurements before and after (before minus
after).

2. Compute all summary statistics for the difference.

3. Use the sample to compute a 95% confidence intervals for the average difference
in the population!

4. Use the sample to test the null hypotheses µ0 =0 (no effect) using the signifi-
cance level α=0.05. For that purpose use

• the confidence interval approach,

• the standardized test statistic, and
• the p-value!

Explain your reasoning in each case!

Exercise 371
The sheet ’costs’ lists the number of items produced (items) and the total cost (costs)
of producing these items.

1. Estimate a simple regression model to analyze the relationship between ’costs’

and ’items’.

2. Compute the fitted values (expected costs) and the residuals (errors).

3. Interpret the coefficient of ’items’.

4. Comment on the goodness of fit of this model. What can you say about the
magnitude of errors?

Exercise 472
The sheet ’car’ contains annual data (1970-1987) on domestic auto sales in the
United States. The variables are defined as ’quantity’: annual domestic auto sales
(in thousand units); ’price’: real price index of new cars; ’income’: real disposable
income; ’interest’: prime rate of interest (in %).

1. Estimate a multiple regression model for ’quantity’ using ’price’, ’income’ and
’interest’ as explanatory variables.
71
Example 13.4 on page 696 in AWZ.
72
Example 13.5 on page 703 in AWZ (slightly modified).
9 EXERCISES 82

2. Compute the fitted values (expected cars sold) and the residuals (errors).

3. Interpret the coefficients of all explanatory variables.

4. Comment on the goodness of fit of this model. What can you say about the
magnitude of errors?

5. Test the significance of all explanatory variables (choose a significance level!).

6. Consider the following scenario: price index=235, income=2690, interest=8.5.

Compute the expected quantity for this scenario!

7. What is the probability to exceed a quantity of 7800 in this scenario, assuming

a normal distribution?

8. Suppose income drops by 100 units and the interest rate increases by two per-
centage points. What is the required change in the price index to compensate
for the associated effect on the expected car sales?

Exercise 5
O’Drill Inc. plans to drill for oil in a promising area. O’Drill is using a new drilling
station which has a new drilling head already built in. A drilling head has to be
replaced after drilling for 2000 meters. O’Drill does not know how deep it has to
drill until oil is found. According to the estimates, there is a 30% probability to find
oil between 0m and 2000m. The probability to find oil between 2000m and 4000m is
considered to be 50%. 20% is the probability to find oil between 4000m and 6000m.
O’Drill rules out the case of finding oil below 6000m. Given this uncertainty it is
unclear how many additional drilling heads – in addition to the one already built in
– are required.
Drilling heads can be ordered from two different suppliers. Supplier A charges 60
for each head ordered right now (a special deal). If ordered at a later date, supplier
A charges 100 for an additional head. Drilling heads which have not been used can
be sold back to supplier A for 20. Installing an additional head costs 40.
Supplier B offers an all-inclusive contract, and charges 120 for delivering and in-
stalling any additional head.

1. Determine the states of nature and the set of decisions (strategies) O’Drill can
choose from.

2. Which costs are associated with each pair of decisions and states of nature?

3. Which strategy should O’Drill choose? Use a suitable criterion to support its
decision!
10 CASES 83

10 Cases
The following three cases are taken from the book ”Business Statistics: For Con-
temporary Decision Making” by Ken Black (Wiley). The fourth case is taken from
AWZ (Example 7.5).

10.1 Olson Diversified Marketing Services – Analysis of Receivables

Background

Olson Diversified Marketing Services is a medium-sized business service organiza-

tion that provides sales and promotion services to various medium to large-size
companies. Olson’s services include sampling of trial sizes of new products to po-
tential customers, management and administration of product promotion contests,
and planning and operation of national sales meetings for major corporations. For
every project contracted, Olson assigns a project manager and various staff profes-
sionals. When the project is completed, Olson prepares a complete billing, including
overhead and a management service fee. After the bill has been received, the client
is required to submit payment within 15 days or pay contracted interest payments.
Olson Diversified Marketing Services has established a target of 45 days after the
end of the project to complete its account of accumulated charges and to submit a
bill to the client.
The most recent audit by outside auditors has revealed that accounts receivable
(money owed to Olson by clients) has grown considerably. Auditors have demon-
strated that the receivables constitute too large a percentage of Olson’s sales and
have advised the company to develop a plan to substantially reduce receivables.
Company management believes that reducing the fifteen-day period for client pay-
ment would seriously damage its relationships with present and future clients. Thus
any improvement in the receivables problem must result from reducing the time
required to prepare clients’ bills.
You have been assigned to analyze the problem and recommend a procedure to
substantially reduce the receivables. First, you are to study the pattern of times
currently required to prepare clients’ bills and the bill amounts. Next, you must
determine the potential savings from reducing the number of days spent preparing
the bills. Prior to beginning your analysis, you meet with George Reale, manager of
billing and receivables, in the accounting department. He has been concerned about
this problem for some time, but heretofore his concerns have not been given serious
consideration. Thus he is pleased that you are conducting the study and offers to
help in any way he can. He provides you with a data set containing information on
the 445 projects completed and billed during the previous calendar year.
To reduce the time needed to prepare clients’ bills, George has suggested implement-
10.1 Olson Diversified Marketing Services – Analysis of Receivables 84

ing a new computerized billing system. This system would require project managers
to enter all purchase orders and staff hours as soon as the services are completed.
Suppliers would be asked to submit their charges electronically, and these charges
would be added to the charges data file as soon as received. At the end of the
project, a list of outstanding charges from suppliers would be prepared, and the
project manager would direct an effort to obtain all remaining supplier charges.
Previously, senior management argued that developing such a system would be too
expensive and that operating it would necessitate too much extra work. After your
meeting, you and George agree that you will focus on the potential cost savings
and then determine whether these are sufficient to justify the computerized system
proposed by George.

Technical background

Accounts receivable are funds owed by clients for goods or services provided. In ac-
counting terms, accounts receivable are recognized when the services are performed.
In addition to the actual charges, the receivables include a markup to cover profit,
management, and allocated overhead. Receivables impose a cost on Olson because
the company must borrow funds to cover payments to its suppliers before payment
is received from the client. Olson maintains a line of credit with a local bank at
an annual interest rate of 10 percent to cover these receivables. Thus, every $1,000
dollars held as receivables for the year results in an interest charge of $100. If Ol-
son can reduce the time for a $1,000,000 client charge by one day, the net savings is
$274 [0.10·$1,000,000/365], and a savings of ten days is worth $2,740. In general, the
amount of savings obtainable by reducing the billing time for each client contract is
equal to the dollar amount of the client bill multiplied by the 0.10 bank interest rate,
with the result then multiplied by the reduction in billing time (in days) divided by
365.

Definition of analysis project

You have been asked to study the pattern of bill preparation times cross-referenced
by the size of client bills. From this research, you are to indicate the potential savings
that could result from reducing billing time. You might, for example, recommend
hat efforts be concentrated on bills for amounts above a certain level. Or you might
recommend that efforts be applied uniformly to all bills. Generally, larger bills
entail substantially more cost items and thus necessitate more work to accumulate
all charges.
The data file supplied by George Reale contains 445 observations on two variables.
The first column indicates the total charges billed for the project, and the second
column indicates the number of days required to prepare and submit the bill to the
10.1 Olson Diversified Marketing Services – Analysis of Receivables 85

client. The data are stored in a file named ”olson.xlsx”. Your case project includes
the following tasks.

1. Describe the statistical properties of the number of days and the amount of
the client’s bill.

2. What are the chances (i.e. percentage of cases) that the company reaches its
goal of preparing the bill within 45 days?

3. What is the probability of meeting this target assuming a normal distribution?

What is the probability assuming a lognormal distribution?

4. Now consider only bills with an amount in the top 25 percent.

(a) What is the average value of these bills?

(b) What is the average number of days required for preparation of these
bills?
(c) How much money could be saved if the average bill preparation time
could be reduced by 20 percent for these bills?
10.2 Production System Inc. – Development of a Salary Model 86

10.2 Production System Inc. – Development of a Salary Model

Introduction

Sally Parsons, president of Production System Inc., has asked you to assist in ana-
lyzing the company’s salary data. She has recently received a series of complaints
that women employees are receiving lower wages for comparable jobs. The com-
plaints came as a surprise to Sally because she thought salary increases were based
on experience and Performance. Sally was aware that the average wages of women
staff were less overall than those of men. However, she also knew that women staff
had less experience and thus would be expected to have lower wages. Given the
complaints Sally understands that she must have objective information.
Production Systems Inc. is a regional computer systems development company hat
specializes in work with banks and insurance companies. The company started as
a service department for a regional accounting company in the 1960s. A few of the
current employees actually came from the accounting company. In the early 1970s,
Production Systems Inc. became an independent company. The company has been
quite successful, experiencing steady growth over its entire life. Employment has
not grown as rapidly as have the company’s total billings because of productivity
innovations introduced during the past ten years.
The company has tended to hire experienced professionals with masters degrees in
technical fields including business, management science, computer science, engineer-
ing, economics, and mathematics. Most of the employees have come with significant
experience. The youngest employee in the professional group is 29, and the most
senior is 65. Experience with Production Systems varies over a wide range, with
one-fourth having less than seven years and one-fourth having more than 22 years.
Most of the women employees have less experience with the company.
The professional staff has only three levels: systems analyst, team systems analyst,
and project systems analyst. Salary ranges are quite wide within each of the levels.
Project systems analyst is the highest level, followed by team systems analyst. Pro-
motion to the higher ranks is awarded by a committee of project systems analysts,
and advancement to each level usually requires a minimum of six years’ experience
and significant project work. In general, persons at the higher levels are more pro-
ductive and tend to direct projects. It is possible, however, for a group that includes
several project systems analysts to be directed by a team systems analyst. For ev-
ery project under contract, a team of the best available people is created to carry
out the work. It is also well known that some persons at the highest level are less
productive than others at lower levels. Thus higher status and salary is a reward
for past performance and not a reliable measure of present contribution.
Salary adjustments are sometimes made to recognize certain specialty skills that
demand a high price in the labor market. Persons who work in database systems
10.2 Production System Inc. – Development of a Salary Model 87

programming have unique skills that are highly sought by other companies. Another
special category is technical systems developers – people who prepare specialized
high-performance software for key parts of large systems. People with either of
these skills are in great demand, so they must be paid a premium if they are to
be retained. Such specialists work at all three professional staff levels, depending
on their experience with the company, but the company has not provided premium
salaries merely by promoting the specialists. The personnel policy has been to base
promotions on a wide range of work and project management skills. Special skills
are compensated by a separate adjustment. Because promotions to higher levels
are related to experience, the company has sought to avoid confusing levels and
specialized skills that have a market premium.

Problem analysis

Your project analysis begins with a series of meetings you have with Sally Parsons
and the director of human resources, Gilbert Chatfield. Both Sally and Gilbert in-
dicate that wages tend to increase with experience in the company. The managers
conduct an annual employee review, which relies heavily on input from project lead-
ers who are directing teams at various remote locations. Project leaders shift as
projects are completed and new teams are assigned. Thus, obtaining consistent
information to provide the basis for a high-quality employee evaluation is difficult.
Most of the employees at Production Systems enjoy their independence and challeng-
ing work; salary levels have not been a major concern for most employees. Certain
people are recognized as strong performers, and their increases and promotions are
generally accepted by the professional staff.
In recent years, however, concerns have been raised about the fairness of the system
of awarding salary increases. The complaint by women employees is the most serious,
but other complaints have been made over the past several years. In view of these
concerns, you recommend that a salary regression model be developed. This model
would use data based on the current salaries paid to professional staff and important
variables that define the experience and skill levels of the staff. Such a model would
indicate the effect of various factors that contribute to salary level, and it would
identify persons whose salaries are above and below the expected salary. The model
could also be used to determine whether an employee’s gender indicates a salary
that is higher or lower than would be expected on the basis of experience and
qualifications.
After some discussion, Sally and Gilbert agree that this model should be developed.
It would be useful for answering the present complaint, and it would provide a
tool for reviewing the complete professional staff salary structure. After reviewing
the employee data records, you select a candidate set of variables for the model
development. These variables, which are contained in a file named ”prodsys.xlsx”,
10.2 Production System Inc. – Development of a Salary Model 88

are described in the table below. To protect the confidentiality of each employee’s
salary record, there is no variable to identify individual employees in this file. At
the completion of the study, you will provide Gilbert with a list of employees who
are substantially below or above the standard predicted by the model. Since he
has the identification key for each employee, and has access to other performance
information, he can decide whether certain persons’ salaries should actually be above
or below the standard.

file: prodsys.xlsx
variable name description
age age of the employee
yearsexp years of experience
yearslv2 number of years as a team systems analyst
yearslv3 number of years as a project systems analyst
gender gender: 1=female; 0=male
spec1 specialty: 1=database systems development skill; 0=else
spec2 specialty: 1=technical systems development skill; 0=else
salary annual salary (in dollars)

Your final discussion concludes with your agreeing to include the following tasks in
your study.

1. Estimate a multiple regression model describing the salaries of Production

System’s professional staff.

2. Does the average compensation paid to women differ from that paid to men?
Answer that question

(a) without using the results from the regression and

(b) using the results from the regression.

In each case carry out statistical tests in order to judge whether the observed
difference is statistically significant! Discuss the reasons for obtaining (poten-
tially) different results from (a) and (b) regarding the difference in average
salary.

3. Evaluate the salary effects associated with specialty skills. Derive an estimate
of the premium that has to be paid for these skills, and judge the (statisti-
cal) significance of the premium for specialty skills. Compare the results you
obtain from the complete regression to another one which ignores all other
characteristics (i.e. only includes the dummy variables for specialty skills).
10.3 American Motors – Predicting Fuel Economy and Price 89

10.3 American Motors – Predicting Fuel Economy and Price

Introduction and background

American Motors Inc. is a major national manufacturer of automobiles and light

trucks. It competes with one other large domestic company and two smaller firms.
Significant competition also comes from European and Japanese companies. The
industry has undergone major changes, including an emphasis on fuel economy man-
dated by the national government standards and market demand from an increasing
number of consumers. Engineering and production staff have developed numerous
improvements in response to market demands. Marketing research has continued to
expand the company’s understanding of customer needs and desires.
Foreign competition resulted in major sales reductions for all domestic U.S. man-
ufacturers. When it experienced those reductions, American Motors was forced to
close a number of facilities and reduce employment. In addition, new design and
production innovations had to be implemented to obtain the same cost and quality
parameters as those of the foreign manufacturers.
American Motors has committed considerable resources to new facilities for product
development, quality control, and manufacturing. As a result, quality has improved
and productivity has increased. However, the costs of these improvements has added
large debt service charges to annual costs. Thus sales levels must be maintained and
increased if possible.

Long-range planning analysis

After years of declining sales and facility downsizing American Motors has begun
to experience increased demand for its vehicles. Design, performance, and quality
measures indicate that its vehicles compare favorably with those produced by other
foreign and domestic companies.
Reflecting on the years of major business problems, senior management realized that
during the 1970s it had not been sensitive to the business environment. Consumer
tastes changed, and foreign competitors produced cars that responded better to
consumer needs. In addition American Motor’s quality had decreased so that per-
formance was substandard for the industry. To prevent future problems, American
developed a long-range market planning group responsible for monitoring the mar-
ket environment and recommending change to help maintain American’s competitive
position.
The long-range market planning group prepares alternative strategies for developing
future cars, and it monitors American’s performance compared to the rest of the in-
dustry. Each company in the industry has various automobile models that resemble
those in American’s product mix. However, the number of units sold for each of var-
10.3 American Motors – Predicting Fuel Economy and Price 90

ious models differs for each company. Therefore, it is difficult to compare companies
by the product mix. Instead the market planning group has developed compar-
isons based on important vehicle performance characteristics, including number of
cylinders, horsepower, acceleration, engine displacement, and vehicle weight.
You have been asked to develop mathematical models to describe fuel economy
and vehicle selling price as a function of these performance characteristics. These
models will be used to estimate which performance characteristics have the greatest
effect on purchasing decisions. In addition, the effect of various combinations of
characteristics will be estimated. A representative sample of the various automobiles
in the present national vehicle mix has been obtained for your analysis.
The analysis will determine which performance variables are significant factors and
will estimate the importance of those variables. Before beginning the analysis, you
meet with a number of experienced engineers to learn more about the vehicle per-
formance characteristics and their relationship to fuel economy. You also talk with
a number of experienced members of the marketing staff to learn how various con-
sumer groups rate the importance of the different performance characteristics.
The fuel economy variable has become increasingly important for vehicle market-
ing. Initially, fuel economy represented an important national policy objective after
major producer countries restricted supply and increased price. The national policy
contained a number of energy conservation measures, including fuel economy (miles
per gallon) minimums, imposed on each manufacturer. A manufacturer can meet
these standards either by selling a greater proportion of small fuel-efficient cars or
by improving the fuel economy of its larger and higher-priced cars. The company
would prefer the latter strategy, because larger cars provide higher revenue and a
larger contribution to overhead and profit. Fuel economy improvements can be ob-
tained by reducing vehicle size and weight, by reducing engine size and horsepower,
or by providing a combination of weight and horsepower reductions. Having smaller
size and weight usually reduces comfort, while the lower horsepower reduces vehicle
performance. Fuel economy can also be improved by instituting better engine design
and engine operating control. For example, the combination of sensors and a com-
puter processor control can provide the ideal fuel/air mixture for different vehicle
load conditions. The present overall vehicle mix results from a number of consumer
choices. Thus, analysis of the present mix offers a way of measuring consumer pref-
erences.
After meeting to define specific objectives for the study, the planning staff decides
that the study should identify key factors that affect fuel economy and key factors
that affect vehicle price. These factors could then be used as planning parameters
for developing new vehicle designs. American Motors also wants to know how it
compares with the rest of the industry as gauged by the driving factors, and it
wants to know the relationship of these factors to fuel economy and vehicle price.
10.3 American Motors – Predicting Fuel Economy and Price 91

The planning group has asked you to perform an analysis to answer these questions.
The variables in the file ”motors.xlsx” are described in the table below. Your analysis
should include the following steps.

1. Run a regression to describe fuel economy (miles per gallon) using all available
explanatory variables (don’t use the columns ’country’ and ’company’ which
are only included in the sheet to derive the regressors ’US’ and ’AmMo’).

2. Identify which factors have a significant effect on fuel economy using a signif-
icance level of 10%.

3. Determine whether the fuel economy for American Motors vehicles is signifi-
cantly above or below – or roughly matches – the fuel economy of its competi-
tors.

4. Run a regression to describe the vehicle selling price using all available ex-
planatory variables.

5. Indicate what factors have a significant effect on the selling price (using a
significance level of 10%).

6. Determine whether the selling price for American Motors vehicles is signifi-
cantly above or below – or roughly matches – the industry level.

7. Determine whether the selling price of cars built in Europe or Japan differs
significantly or roughly matches the price of U.S. cars.

file: motors.xlsx
variable name description
price price in dollars
milpgal miles per gallon (measure for fuel economy)
cylinder number of cylinders in the car
displace cubic inches of engine displacement
horspwr horsepower generated by the engine
accel acceleration (in seconds) to 60 miles per hour
year model year of the vehicle
weight vehicle weight (in pounds)
country country of origin (1=U.S., 2=Europe, 3=Japan)
company manufacturer (1=American Motors; 2=Apex Motors;
3=Smith Motors; 4=other)
10.4 ACME Marketing Strategy – A Multistage Decision Problem 92

10.4 ACME Marketing Strategy – A Multistage Decision Problem

Background Information

ACME Company is trying to decide whether to market a new product. As in many

new-product situations, there is much uncertainty about whether the product will
”catch-on”. ACME believes that it would be prudent to introduce the product to
a test market first. Thus the first decision is whether to conduct the test market.
ACME has determined that the fixed cost of the test market is $3 million. If they
proceed with the test, they must then wait for the results to decide if they will
market the product nationally at a fixed cost of $90 million. If the decision is not to
conduct the test market, then the product can be marketed nationally with no delay.
ACME’s unit margin, the difference between its selling price and its unit variable
cost, is $18 in both markets.
ACME classifies the results in either market as great, fair or awful. Each of these
has a forecasted total units sold as (in 1000s of units) 200, 100 and 30 in the test
market and 6000, 3000 and 900 for the national market. Based on previous test
markets for similar products, it estimates that probabilities of the three test market
outcomes are 0.3, 0.6 and 0.1. Then, based on historical data on products that
were tested and subsequently marketed nationally, it assesses the probabilities of
the national market outcomes given each test market outcome. If the test market
is great, the probabilities for the national market are 0.8, 0.15, and 0.05. If the test
market is fair, then the probabilities are 0.3, 0.5, 0.2. If the test market is awful,
then the probabilities are 0.05, 0.25, and 0.7. Note how the probabilities of the
national market mirrors those of the test market.

Elements of Decision Problem

The three basic elements of this decision problem are: the possible strategies, the
possible outcomes and their probabilities, and the value model. The possible strate-
gies are clear: ACME must first decide whether to conduct the test market. Then
it must decide whether to introduce the product nationally. If ACME decides to
conduct a test market they will base the decision to market nationally on the test
market results. In this case its final strategy will be a contingency plan, where it
conducts the test market, then introduces the product nationally if it receives suffi-
ciently positive test market results, and abandons the product if it receives negative
test market results. The optimal strategies from many multistage decision problems
involve similar contingency plans.
10.4 ACME Marketing Strategy – A Multistage Decision Problem 93

Conditional Probabilities

Approaching this problem requires to know the probabilities of the test market
outcomes and the conditional probabilities of national market outcomes given
the test market outcomes. However, suppose ACME decides not to run a test
market and then decides to market nationally. Then what are the probabilities of
the national market outcomes? You cannot simply assess three new probabilities.
These probabilities are implied by the given probabilities. This follows from the rule
of conditional probability. Let T1 , T2 , and T3 be the test market outcomes and N be
any national market outcome, then the probability for a particular national market
outcome is given by

P(N ) = P(N and T1 ) + P(N and T2 ) + P(N and T3 ) =

= P(N |T1 ) · P(T1 ) + P(N |T2 ) · P(T2 ) + P(N |T3 ) · P(T3 ).

This is called the Law of Total Probabilities.

The monetary values associated with this problem are the fixed costs of test mar-
keting or marketing nationally and these are incurred as soon as the ”go ahead”
decisions are made. The file ”acme.xlsx” contains the inputs for the decision tree.

1. Describe ACME’s decision problem using a decision tree.

2. Which strategy should ACME choose? Describe ACME’s contingency plan,

and use a suitable criterion to support its decision!
11 SYMBOLS AND SHORT DEFINITIONS 94

11 Symbols and short definitions

α ... in the context of a quantile: the probability to observe a value
less than or equal to the α-quantile
... in the context of an interval: the probability to observe a value
outside the (1−α) interval
... significance level (maximum probability for a type I error)
 ... estimation error
µ, E[ ] ... mean of the population
σ2 ... variance of the population
X n
yt ... the sum of all values yt from t=1 to t=n (y1 +y2 +· · · +yn )
t=1
zα ... α-quantile of the standard normal distribution
Ψα ... α-quantile of y∼N(ȳ, s2 )
bj ... coefficient of (or slope with respect to) the explanatory
variable xj in a regression equation
c ... intercept (constant term) in a regression equation
et ... residual (error) in a regression model (et =yt −ŷt )
g ... coefficient of variation (g=s/ȳ)
n ... the number of observations in the sample
... in the context of the binomial distribution: the number of trials
P[ ] ... the probability of the event in brackets
ryx ... sample correlation coefficient between y and x
R2 ... coefficient of determination in a regression model (proportion of
explained variance of yt )
R̄2 ... adjusted R2 in a regression model
s ... sample standard deviation; square root of the variance
s2 ... sample variance; average squared deviation from the mean
se ... standard error of regression (standard deviation of residuals
in a regression model)
syx ... sample covariance between y and x
t ... standardized test statistic; t-statistic of regression coefficients
y∼N(µ, σ 2 ) ... the random variable y has a normal distribution
... with mean µ and variance σ 2
yt ... sample observation for unit or time t
ȳ ... (arithmetic) mean or average of yt
ŷt ... conditional expected value (fit) of yt
z ... standardized variable z=(y−ȳ)/sy