Data reduction and descriptive statistics

Business Mathematics (IBA)

October 2023

1 Introduction
Economics, marketing, finance, and most sciences in general, are based on em-
pirical data, on facts so to say. We make observations on prices, sales, and many
more phenomena that we wish to study. Let us take an example: the sales of
ice cream of a particular vendor near the beach. During summer, he has his
business for 72 days. Eager to know how his business is flourishing, he carefully
writes down his sales (in euro) every day. At the end of the summer, he has a
record of 72 numbers, looking like 320, 410, 176, etc. These are the data. But
data make no real knowledge. If we ask, how were your sales, he can do no
better than recite the set of 72 numbers. Impatiently, we will interrupt him,
and insist: “yeah, I mean, on the average”. Or: “in total”.
This is an important observation: there are situations in which we want to
condense data into more meaningful summaries that are easier to communicate
or understand. For obvious reasons, the process of constructing such (numerical)
summaries is often referred to as data reduction. It yields summaries that are
known as descriptive statistics.
The average (or mean) is the most important of such descriptive statistics.
It expresses the central tendency of the data set. There are, however, other
measures of central tendency, such as the median value. All such statistics are
examples of measures of location.
Further, a central value is not always enough to describe the phenomenon
under study. If we want to convey more information about the data, we could,
e.g., construct a measure of spread or variation (variability). How constant
are the sales? Are they fluctuating wildly, or only mildly? Thus, we may be
interested in a second descriptive statistic, one that reflects the degree of varia-
tion. Such indicators are often referred to as measures of dispersion. The most
important members of this family are the variance and the standard deviation.
All such summary indicators are examples of descriptive statistics: measures
that describe a data set in concise terms. Whenever people speak of “statistics”,

1
they often have this in mind. Indeed, the national and supranational statistical
bureaus, such as the CBS in The Netherlands, Eurostat in the EU, and the
United Nations Statistics Division collect statistics in this descriptive, data-
summarizing sense.
The term statistics, however, has more meanings. Later in your study, you
will do a course on statistics, but then, we will address the subject of inferential
statistics, which studies how to infer properties of a large or even infinite popu-
lation on the basis of knowledge of only a sample of data. In the present course,
we will not address inferential statistics, but only discuss descriptive statistics.

2 Measures of location
We first discuss the measures of location, also known as the measures of central
tendency. Such measures are often used to give a one-number summary of
“how big”. For instance, if we ask how many hamburgers are sold daily by
McDonald’s, we want an indication like 6.5 million, acknowledging that the
number will vary day by day.

2.1 The mean

The most well-known measure of location is the average, also known as the
mean, or more precisely, the arithmetic mean. For a dataset of length n with
observed numerical values x1 , x2 , . . . , xn , the mean x̄ is calculated by
1
x̄ = (x1 + x2 + · · · + xn ) (1)
n
Using the summation notation, this can be written concisely as
n
1X
x̄ = xi (2)
n i=1

The symbol x̄ is pronounced “x bar”.

The book by Sydsæter et al. (fifth edition) introduces the mean briefly on
page 57. It uses a slightly different notation:
T
1X
µx = xi (3)
T i=1

The individual data point is in both notations written as xi , but we could

obviously also have used xj or zk . The number of data points is written here as n,
and by the book as T . This is also just a choice, and not in any way fundamental.
The last difference is the symbol used at the lefthand-side of the definition. We
use x̄, the book uses µx . This is not merely an unimportant difference. In
our later course on statistics, we will use both notations to represent different
but related concepts, x̄ referring to the sample mean, and µx to the population
mean. Basically, whenever no context is given, we do not know if the data set
represents a sample or a population, so both notations are right, and we just use
the word “mean” to refer to it. In our example on Hamburger sales, suppose
that we have collected data on sales on 34 days randomly selected over the year.

2
In that case, the data set x1 , x2 , . . . , x34 is a sample, so the notation x̄ is more
appropriate. In business and economics, we will often have samples: on sales,
investments, employment, prices, etc. If we want to find out the average price of
a beer in Dutch bars, we will make a sample of n prices p1 , . . . , pn , and calculate
the sample mean p̄ by the formula given, which should in that case be rephrased
as
n
1X
p̄ = pi (4)
n i=1

2.2 Mean versus expected value

The discussion on sample versus population touches on another distinction, be-
tween mean and expected value. In fact, the expected value is in some imprecise
texts considered to be identical to the mean, but there is a crucial difference.
The expected value is defined on probability distributions, and it is equal to the
population mean. The sample mean is instead defined for the data observed in
a sample.
In practice, the word expectation is unfortunately used by some books to
refer to the mean of a set of observed values. This is not correct, for theoretical
reasons (the expectation is defined for probability distributions, not for data),
in a numerical sense (the expected value of a die is 3.5, whereas the mean after
1000 observations may well be 3.52), and even for linguistic reasons (the 3.52
was not expected, but calculated from observations).

2.3 Properties of the mean

The mean is a numeric property of a data set. It can also be conceived as a
function of the data set to a single number:
1
f : (x1 , x2 , . . . , xn ) → (x1 + x2 + · · · + xn ) (5)
n
or in vector-notation
n
1X
f :x→ xi (6)
n i=1
The mean satisfies some convenient properties. Most importantly, it is a
linear function of its arguments. This can be formulated mathematically as
follows:
f (ax) = af (x) (7)
and
f (x + y) = f (x) + f (y) (8)
In words, if we scale all arguments up or down by a factor a, the mean will also
scale up or down by a factor a. It implies that the mean can be calculated in
an arbitrary unit, because changes of unit will rescale the value of the mean. If
the average price of beer is 2.02 euro, it is also 202 cent.
The mean also satisfies the property of additivity. That means that the sum
of the means of two data sets is equal to the mean of the sum of the data points.
If the average price of a glass of beer is 2.02 euro and the average price of a

3
glass of wine 3.11 euro, the average price of one glass of beer and one glass of
wine is 5.13 euro.
Another property of the mean is that it has a unit of measurement (cm, sec,
dollar) that is equal to that of the data points. So, if x is in euro/year, x̄ is
in euro/year as well. A simple consequence is that we can only compute the
average of numbers with the same unit. We can’t compute the average of 2 euro
and 4 kg.

2.4 Other measures of central tendency

Besides the (arithmetic) mean, additional measures of central tendency are used.
The median is an important one, as is the geometric mean. We will postpone
the discussion of such alternative measures of central tendency to the course on
statistics.

3 Measures of dispersion
Next, we discuss the measures of dispersion. Dispersion refers to how scattered
the data points are around their mean. If all data points have the same value,
there is no dispersion. Dispersion is thus measured around the mean (or another
central value). A natural starting point therefore is to consider the deviation of
each data point from the mean, so xi − x̄, and to calculate the mean value of
this series of deviations. However, that will not work. The reason is that the
mean deviation is always 0. See the exercises for a proof.
The problem is caused by the fact that positive deviations (xi > x̄) cancel
the effect of negative deviations (xi < x̄), while we of course need non-cancelling
deviations. There are two main ways to make sure this does not happen:
• by looking at the absolute value of each deviation;
• by looking at the square of each deviation.
The first option yields the mean absolute deviation (MAD), which we will not
further discuss. The second option yields the variance, which we will discuss
below.

3.1 The variance

The variance s2x of a data set of length n > 1 is defined as
n
1 X 2
s2x = (xi − x̄) (9)
n − 1 i=1

It is clearly a formula that aims to calculate a sort of mean squared deviation

from the mean.
There is one subtle issue: while we would expect to divide by n to find
the mean square deviation, we actually divide by n − 1. There are theoretical
reasons for this that will be discussed in the later course on statistics. Further,
in some cases, a division by n is actually performed. If you use software (or a
calculator with a σ-button), always check what happens. Some programs (or
calculators) offer both options, indicated for instance as σn−1 and σn .

4
3.2 Properties of the variance
What happens when we scale the data set, for instance, if we move from meter
to centimeter? Such a change of data can be represented as y = ax, with a the
scaling factor. Under this change, the variance will transform as
n
1 X 2
s2y = (yi − ȳ)
n − 1 i=1
n
1 X 2
= (axi − ax̄)
n − 1 i=1 (10)
2 n
a X 2
= (xi − x̄)
n−1 i=1
= a2 s2x

In words, if we change all data by a factor 100 (as we do when moving from
meter to centimeter), the variance will change by a factor 1002 = 10, 000, while
the mean will move by a factor 100.
This is, of course, related to the fact that the variance has a unit that is the
square of the unit of the data. If x is in meter, s2x is in square meter.
Above, we already discussed the issue of x̄ versus µ. For the variance, a
similar phenomenon can be observed. Here the book discusses the variance as
σx2 (or σx,x ; see below), while we use s2x (or perhaps sx,x ). For now, we keep s2x ,
but come back to this in the course on statistics.

3.3 The standard deviation

Because the variance has a unit that is the square of that of the data, one often
uses the square root of the variance, also known as the standard deviation sx :
v
u n
p u 1 X 2
2
sx = sx = t (xi − x̄) (11)
n − 1 i=1

The standard deviation has simpler scaling properties: it just scales, like the
mean, linearly with the data.

3.4 The coefficient of variation

In some cases it is convenient to have a measure of dispersion that is independent
of the scale of the data. The coefficient of variation (or variation coefficient) is
a measure that satisfies this criterion. It is defined as
sx
CVx = (12)
x̄

It can be interpreted as a relative standard deviation, relative to the mean of the

data. It has no unit, and if we change the unit of measurement, the coefficient

5
of variation will not change. It is easy to prove this, again using y = ax:
sy
CVy =
ȳ
asx
=
ax̄ (13)
sx
=
x̄
= CVx

Notice that the coefficient of variation is only defined when x̄ ̸= 0.

3.5 Other measures of dispersion

Again, there is a wide choice of alternative measures of dispersion. We do not
discuss these here, but will mention a few in the statistics course later on.

4 Measures of association
Central tendency and dispersion apply to single data vectors. In some cases, it
is informative to summarize to which extent one data vector varies with another
data vector. For instance, if we have observations on house prices and floor areas,
we can compute the mean and the variance price and the mean and the variance
floor area, but probably are also interested to what extent big houses are more
expensive. Probably they are, but there are exceptions, so the association is not
perfect. Can we develop an indicator of the degree of association?
We mention two widely-used indicators of association: the covariance and
the correlation coefficient.
These indicators are only meaningful for related (or paired) data vectors.
We cannot compute the covariance of the prices of beer in 50 bars in the UK
and 50 bars in the US, because there is no sense of “pairing” between a bar
in the UK and a bar in the US. But we can study the covariance of the price
of beer in a certain bar in the UK in 50 years and in a certain bar in the US
in the same 50 years. The observation in the UK in 1971 can then be paired
with that in the US is the same year. Thus, covariance will indicate to what
extent price increases in one country will be associated with prices increases in
the other country. Obviously, the idea of pairing means that the two data sets
must have the same length.

4.1 The covariance

The covariance is an indicator of two data vectors, say x and y. We therefore
add a double subscript to it and write sx,y :

n
1 X
sx,y = (xi − x̄) (yi − ȳ) (14)
n − 1 i=1

The covariance measures to which extent the two data sets “run together”.
Above, we discussed that the variance (and the standard deviation) can be
calculated with a denominator n or n − 1, and that calculators and software

6
Figure 1: Scatterplot of the house prices and floor areas of a sample of 71 houses
in Amstelveen in 2014.

sometimes offer both options, but also that it is not always clear which formula
is implemented. The same remark applies for the covariance. We define it for
this moment with n − 1, but we will come back in the statistics course on the
form with n.
The covariance has a rather complicated unit: the product of the units of the
two underlying data vectors. In the example of house prices in euro and floor
areas in square meter, the covariance is measured in euro times square meter.
Do not confuse this with euro per square meter!
The similarity of the term “covariance” to “variance”, the similarity of the
formulas, and the similarity of the symbols suggest that we can try to calculate
the covariance of a data set with itself, so sx,x . We find
n
1 X
sx,x = (xi − x̄) (xi − x̄) = s2x (15)
n − 1 i=1

Thus the covariance of a data set with itself is simply the variance of that data
set. Of course, that one is never negative. The covariance of two different data
sets, however, can be positive, negative, or zero. If it is positive, the two run
together. If it is negative, the two run in opposite way. And if it is zero, the
two are said to be unrelated.

7
4.2 The correlation coefficient
A final data summary indicator is the “relative covariance”: the correlation
coefficient. It is defined as
sx,y
rx,y = (16)
sx sy

The correlation coefficient is scale-free: it has no unit, and it is bounded between

−1 and 1. A value of 1 means that the two data sets are perfectly linearly
correlated. For instance, if x is a series of temperature data in Celsius. and y the
same series in Fahrenheit, the correlation coefficient will be 1. A value between 0
and 1 means that there is some correlation, but not a perfect linear correlation.
For instance, there is quite a good linear correlation (r = 0.9) between the floor
area of a sample of houses and the prices of theses houses. However, the floor
area is not the only thing that matters, because other factors (age, location,
presence of a garage) have an influence as well. A correlation coefficient smaller
than 1 can also indicate a non-linear relationship. For instance, if y depends
on x in a quadratic way, the data points will lie on a perfect parabola, but the
correlation coefficient will be less than 1 because a parabola is not a straight
line.
The correlation coefficient of a data vector with itself is 1. A value of −1
indicates a perfect linear relationship in the opposite way. A value of 0 indicates
that the variables are not related. Note well that the value of the correlation
coefficient does not indicate an empirically important relation.

4.3 Other measures of association

A final word: there are other indicators for association besides those mentioned.
All such indicators have special properties, which make them superior in cer-
tain cases and inferior in other cases. There is no one-size-fits-all indicator for
centrality, dispersion, and association.