Zeszyty Naukowe WWSI, No 27, Vol. 16, 2022, pp.

57-82
DOI: 10.26348/znwwsi.27.57

On the Analysis of Correlation

Between Nominal Data and Numerical Data

Zenon Gniazdowski*

Warsaw School of Computer Science

Abstract
The article investigates the possibility of measuring the strength of a linear corre-
lation relationship between nominal data and numerical data. Correlation coeffi-
cients for variables coded with real numbers as well as for variables coded with
complex numbers were studied. For variables coded with real numbers, unam-
biguous measures of real linear correlation were obtained. In the case of complex
coding, it has been observed that the obtained complex correlation coefficients
change with the permutation of the phases in the complex numbers used to code
classes of elements with equal cardinalities. It was found that a necessary condi-
tion for linear correlation is the possibility of linear ordering of a set with data.
Since linear order is not possible in the set of complex numbers, complex correla-
tion coefficients cannot be used as a measure of linear correlation. In the event of
such a situation, a substitute action was suggested that would prevent equal cardi-
nality of classes of identical elements contained in the set with nominal data. This
action would consist in the correction of data, analogous to the correction during
preprocessing or cleaning of data containing missing or outlier values.

Keywords — nominal data, numerical data, numerical coding of nominal data, complex ran-
dom variable, correlation coefficient, complex correlation, complex least squares method

1 Introduction

In classical statistics, the χ2 test is used to test the correlation between nominal variables. A
contingency table is created for nominal data. On its basis, the χ2 statistic can be used to assess
the significance of the correlation between two nominal variables. The V-Cramer coefficient,
also estimated using χ2 statistics, can be used to measure the strength of this relationship [1].
On the other hand, for variables measured in at least an ordinal scale, the Pearson correlation
coefficient or the Spearman rank correlation coefficient [1][2] are examined.
* E-mail: zgniazdowski@wwsi.edu.pl

Manuscript received November 7, 2022 57

 Zenon Gniazdowski

In [3] a different approach to the problem was proposed. In this approach, nominal data is
given a numerical interpretation. Since a random variable measured on a nominal scale takes
k different values, each of the k subsets of identical values will be called a class later in this
article. Depending on the cardinality of different classes, the elements of a given class will be
coded with real numbers or complex numbers. If there are no classes with equal cardinality,
then a given class will be assigned a real number which is a function of the cardinality of the
class. If there are m classes of equal cardinality in the considered set of values, then each of
these classes is assigned a complex number whose modulus is a function of the cardinality of a
given class, and the phase is equal to the phase of one of the m roots of unity. The phases are
chosen arbitrarily. However, it is required that the phases for each of the classes with identical
cardinalities be different.
The properties of numerical coding make it possible to cluster or classify nominal data using
algorithms specific to numerical data [3]. It remains an open question whether the numerical
coding of nominal data mentioned here can be used to assess the level of correlation between
coded nominal random variables. This article will attempt to answer this question. For this
purpose, the possibility of measuring the strength of the correlation between variables measured
in the nominal scale and variables measured in the nominal scale or at least in the ordinal scale
will be analyzed. The starting point for this analysis will be two observations:
• The correlation coefficient between two random variables has an interpretation of the
cosine of the angle between the vectors containing the random components of these
variables [4].
• Nominal data can be coded with real or complex numbers. Since in a real and complex
vector space it is possible to define the scalar product and the Euclidean norm, it is also
possible to calculate the cosine of the angle between the vectors [3].
The examples will show the influence of coding proposed in the paper [3] on the possibility of
identifying linear correlation between nominal variables. First, examples of correlation anal-
ysis for nominal variables described by real numbers will be shown. Next , the possibility of
correlation analysis for nominal variables described by complex numbers will be examined and
discussed.
The analysis will be carried out on artificial data, specially prepared for the purposes of this
work1 . Sample input data for the analysis will be provided in the form of a contingency table.
A contingency table provides a compact representation of data consisting of many records.
Since the aim of this article is to attempt to assess the possibility of measuring the strength
of correlation between a variable measured on a nominal scale and a variable measured on a
nominal or stronger scale, the analyzed data should be selected in such a way that at least one
variable can be interpreted as if it had been measured on a scale stronger than nominal, and
therefore at least on an ordinal scale. For this reason, an additional restriction is imposed on the
contingency table. It is required that after numerical coding of the nominal variables, the second
variable may be treated as if it were measured on a scale stronger than the nominal scale:
1
To present the discussed problems, the author of this article uses artificially prepared data, because so
far he has not encountered non-trivial nominal data sets that would contain classes of identical elements
with equal cardinalities.

58
 On the Analysis of Correlation Between Nominal Data and Numerical Data

• The first variable will be constructed in such a way that it can be coded using real num-
bers or complex numbers. In the case of coding with complex numbers, the first variable
will contain at least two classes with identical cardinalities. This will be manifested
by the fact that at least two rows in the contingency table will have the same sums of
elements.
• The second variable will always be constructed in such a way that it can only be coded
with real numbers. This will only be possible if the variable does not contain classes
with equal cardinalities. This will be manifested by the fact that in the columns of the
contingency table the sums of elements will always be different in pairs. This approach
will allow the second variable to be treated as if it were measured on a scale with a
possible linear order relation, i.e. on a scale stronger than the nominal scale.
Since the second variable will be coded with real numbers with a well-defined linear order
relation, the analyzed problem will become equivalent to the problem of examining the strength
of the linear correlation relationship between a random variable measured in a nominal scale
and a random variable measured in at least an ordinal scale. Thanks to this, the conclusions
resulting from the study of correlations for two nominal variables will also be appropriate for
the case of correlation between a variable measured in a nominal scale and a variable measured
in one of the stronger scales: ordinal, interval or ratio.
Finally, there is a comment on the accepted designations. In some mathematical formulas,
a line will appear over the letter denoting the variable. If a given subsection concerns the
geometric interpretation of correlations between random variables, the line will be placed over
the capital letter and will represent the average value of the random variable. On the other
hand, when a subsection deals with the definition of a scalar product for vectors containing
complex numbers, or concerns correlation for complex random variables, then the line will be
over lowercase letters and will indicate the conjugate of a complex number.

2 Preliminaries
The preliminaries will present some ideas to which the author will refer later in this article.
Here, the following concepts will be introduced: binary relation, measurement scales, statistics
χ2 , strength of the correlation relationship for nominal data, numerical coding of nominal data,
geometric interpretation of a correlation, as well as the least squares complex method.

2.1 Binary relation

For given sets X and Y , the binary relation ρ defined over sets X and Y is any subset of the
Cartesian product X × Y . If X = Y , then the relation over the set X is a subset of the Cartesian
product X × X [5]. Here, only relations over the n-element set X will be considered.
A binary relation can have specific properties. Among others, a binary relation can be re-
flexive, irreflexive, symmetric, antisymmetric, connected, and transitive (Table 1). The proper-
ties of a binary relation determine its type. Several types of relations in particular are important
from the point of view of data analysis:

59
 Zenon Gniazdowski

Table 1: Properties of a binary relation

The name of property The essence of a given property

Reflexivity ∀x ∈ X : xρx
Irreflexivity ∀x ∈ X : ¬(xρx).
Symmetry ∀x, y ∈ X : xρy ⇒ yρx
Antisymmetry ∀x, y ∈ X : xρy ∧ yρx ⇒ x = y
Connecttivity ∀x, y ∈ X : xρy ∨ yρx ∨ x = y
Transitivity ∀x, y, z ∈ X : xρy ∧ yρz ⇒ xρz

• If a relation is reflexive, symmetric, and transitive, it is an equivalence relation. An

equivalence relation divides a set into disjoint equivalence classes.
• If the relation is reflexive and symmetric, then it is a similarity relation. The similarity
relation divides a set into similarity classes.
• If the relation is reflexive, antisymmetric and transitive, then it is a partial order relation.
In a set with a partial order relation, ordering of a set (or sorting in the sense of a given
relation) is possible only within certain subsets, and a set with a partial order relation is
called a partially ordered set.
• If a relation is a partial order relation and at the same time is connected, then it is a linear
order relation (or total order relation). A linear order is a stronger property than a partial
order. In a set with a linear order relation, it is possible to fully ordering of a set.
In the set of real numbers R, the following linear order relation is possible:

∀m, n ∈ R mρn ⇔ m ≥ n. (1)

In the set of complex numbers C, it is not possible to define the relation in the above way. Here
it is possible to define the weaker relation, that is, a partial order relation:

∀m, n ∈ C mρn ⇔ (m = n) ∨ (|m| > |n|). (2)

2.2 Scales of measurement

Measurement (or data collection) assumes the existence of four measurement scales [6]:
• Nominal scale,
• Ordinal scale,
• Interval scale,
• Ratio scale.

60
 On the Analysis of Correlation Between Nominal Data and Numerical Data

Table 2: Sample contingency table

X Y Total
A 3 0 3
B 2 2 4
C 0 2 2
Total 5 4 9

The nominal scale assumes the classification of data into different classes. For data measured
on a nominal scale, it can be said that the two measured values are equivalent or different. It
cannot be said that one value precedes another. This means that nominal data cannot be sorted
in any way.
Measurement on the ordinal scale is more precise than measurement on the nominal scale.
The ordinal scale allows you to order a set according to the degree to which the elements of the
set have certain features, but does not give information about the magnitude of the differences
between these elements.
The interval scale makes it possible not only to order objects in terms of the degree of pos-
sessing a certain feature, but also gives the ability to determine the distance between objects.
An interval scale is a continuous numerical scale that has no absolute zero. Zero on the mea-
surement scale is set arbitrarily. An example of interval scale measurement is a temperature
measurement in degrees Celsius or degrees Fahrenheit. It is known how much one measure-
ment result is greater than another. However, it is impossible to say how many times one result
is greater than the other. For example, for temperature measurement on the Celsius scale, it
can be said that the temperature of 27 degrees is 18 degrees higher than the temperature of 9
degrees. On the other hand, it cannot be said that a temperature of 27 degrees is three times
higher than a temperature of 9 degrees.
The ratio scale has all the features of an interval scale. Its additional feature is that this
scale has absolute zero. Therefore, with regard to the measurement results on a ratio scale, it
can be said how many times one measurement result is greater than another. An example of a
ratio scale is the temperature measurement scale in degrees Kelvin.
The concept of measurement scales is closely related to the concept of two types of rela-
tions. For elements of a set measured at a nominal scale, an equivalence relation can be defined.
For elements of a set measured on an ordinal, interval and ratio scale, a linear order relation
can be defined. Thanks to the linear order relation, data measured on ordinal, interval, and ratio
scales can also be ranked.

2.3 The study of the relationship between nominal data in classical statistics
An example of a contingency table for nominal data (Table 2) is given. The table describes the
interdependence of two nominal random variables V1 and V2 . The first variable (V1 ) takes three
values: {A, B, C}. The second variable (V2 ) takes two values: {X, Y }. Table 3 shows the
random variables V1 and V2 reconstructed from the contingency table (Table 2).

61
 Zenon Gniazdowski

Table 3: Random variables V1 and V2 identified from Table 2

V1 V2
A X
A X
A X
B X
B X
B Y
B Y
C Y
C Y

Table 4: Table of expected frequencies estimated from Table 2

X Y Total
A 1.667 1.333 3
B 2.222 1.778 4
C 1.111 0.889 2
Total 5 4 9

2.3.1 Statistics χ2
Individual cells of the contingency table (Oi,j ) count the observed frequencies of pairs of dif-
ferent nominal values. Based on the observed frequencies, the relevant elements (Eij ) in the
expected frequency table (Table 4) can be estimated [1]. The Eij element
P is equal to the product
of the sum of the elements in the i-th row of the contingency
P table ( j Oij ) and the sum of the
elements in the j-th column
P of the contingency table ( O
i ij ), divided by sum of all elements
of the contingency table ( ij Oij ):
P P
j Oij · i Oij
Eij = P . (3)
ij Oij

The estimated table of expected frequencies is shown in Table 4. Now, based on the contents of
the table of observed frequencies and the table of expected frequencies, the value of the statistic
χ2 can be estimated [1]:
X (Oij − Eij )2
χ2 = . (4)
i.j
Eij

For the contingency table considered here (Table 2), the statistic χ2 is 4.95. In classical statis-
tics, the value of the χ2 statistic can be used to test the significance of the correlation between

62
 On the Analysis of Correlation Between Nominal Data and Numerical Data

the two nominal variables. To do this, a null hypothesis should be formulated and then checked
whether the null hypothesis can be rejected.

2.3.2 Testing the null hypothesis

The null hypothesis assumes that the examined variables are independent of each other, i.e.
there is no correlation between the variables V1 and V2 . The negation of the null hypothesis is
the alternative hypothesis, which says that the variables V1 and V2 are significantly correlated.
If the null hypothesis is rejected, the alternative hypothesis should be accepted instead.
Denoting by r and c respectively the number of rows and the number of columns in the
contingency table, the number of degrees of freedom of the test is obtained:

df = (r − 1) × (c − 1) . (5)

For the example considered here, df = (3 − 1) (2 − 1) = 2. The probability p is estimated as

the integral of the distribution function χ2 :
Z ∞
p= f (x, df ) dx. (6)
χ2

If the probability p is small enough, the null hypothesis can be rejected. Assuming that the
significance level α = 0.1, the null hypothesis can be rejected when the obtained probability p
is less than α. For the example in Table 2, the estimated probability p is 0.0841. This means
that at the significance level α = 0.1, the null hypothesis should be rejected. At the same time,
the alternative hypothesis, which says that both variables are significantly correlated, should be
accepted.
The χ2 test also has its limitations. When using this test, it is required that the expected
frequencies be not less than five [1]. Some sources state that expected frequencies should not be
less than ten [2][7]. If the expected numbers are too small, Yates’ correction or Fisher’s exact
test [1][2][7] is used for contingency tables of size 2 × 2. For tables of larger sizes, Fisher’s
exact test is a combinatorial problem of high computational complexity. For such a case, a
graph algorithm was proposed that extends the feasibility limits of Fisher’s exact test [8].

2.3.3 The strength of the correlation relationship

Regardless of the significance assessment, the χ2 statistic can also be used to estimate the
strength of the correlation relationship. For this, you can use the V-Cramer coefficient, which
takes values in the range [0, 1]. The square of the V-Cramer coefficient can be calculated using
the formula:
χ2
V2 = . (7)
N · min (r − 1, c − 1)
In formula (7), N is the number of observations. In the considered case N = 9 and the calcu-
lated value of the coefficient V 2 is 0.550.

63
 Zenon Gniazdowski

2.4 Numerical coding of nominal data

In statistics, in some cases, the numerical values of the observation results are replaced by their
ranks. The ranking procedure consists in the fact that first the data set is sorted in non-decreasing
order, and then its successive elements are assigned a rank equal to its position in the sorted set
[9].
It may happen that in a sorted set there will be elements with identical values. In this case,
the ranks assigned to identical values should be the same. When a sorted set contains elements
with identical values, these elements are assigned a rank equal to the average value of their
positions in the sorted set [9]. In this case, we talk about tied ranks.
It should be emphasized that the value of assigned ranks depends on two factors. The first
factor is the numerical value of the element in the set to be sorted. The second factor is the
number of elements with equal values.
In the case of nominal data, it is not possible to establish a linear order relation. A potential
method of ranking nominal data could not use element sorting. On the other hand, intuition
suggests that in a statistical collectivity, both in the case of numerical data and nominal data, an
element with a higher cardinality is more important than an element with a lower cardinality.
Therefore, a proposal arises to rank nominal variables, starting from the cardinality of their
individual values.

2.4.1 The case of classes with different cardinality

A specific feature of nominal data is that it cannot be sorted by values, but can be clustered
by identical values. In a given class containing n identical elements, these elements can be
numbered with numbers from 1 to n. By analogy with tied ranks, each of the identical elements
can be assigned a rank equal to the average value of the assigned numbers:

(n + 1)
R= . (8)
2
The rank of elements in the class with higher cardinality will be greater than the rank of elements
belonging to the class of lower cardinality. Hence, it can be seen that in the set with ranks
defined by the formula (8) a linear order relation (1) is possible.

2.4.2 The case of classes with equal cardinality

For classes with identical cardinalities, method (8) would give identical ranks. This would lead
to a situation where, after coding, elements belonging to classes with identical cardinalities
would be indistinguishable. To prevent this, the method should be modified. If the nominal
variable contains k different classes with identical cardinality, then the elements of the j-th class
(j = 0, 1, . . . , k − 1) can be coded with k successive complex roots of unity:
√
 
k i· 2πj 2πj 2πj
Rj = R · 1 = R · e k = R cos + i sin . (9)
k k

64
 On the Analysis of Correlation Between Nominal Data and Numerical Data

√
In the above expression i = −1 , φ = 2πj/k is the phase arbitrarily assigned to the successive
(j-th) class, and R is the rank calculated according to the formula (8) - the same as in the case
of classes with different cardinalities. In the presented concept, R is a module of complex
rank, depending on the cardinality of a given nominal value in the set. This approach assigns
identical rank modules to classes with identical cardinalities, and distinguishes classes with
identical cardinalities by their different phases.
Unfortunately, unlike in a set with real ranks, a linear order relation is not possible in a set
with complex ranks. Here only the partial order relation (2) is possible.

2.4.3 Properties of numerical coding

Data coded with numbers (real or complex) acquire additional properties [3]:
• Within a given variable, all classes are mutually distinguishable.
• The module of the number assigned to a given class contains information about the car-
dinality of the class, and thus about its statistical strength.
• In the phase of the complex number assigned to a given class, information about the
number of classes with identical cardinality is contained.
Random variables gain the properties of vectors in numerical space (real or complex), on which
arithmetic operations such as sum, difference, product, quotient, as well as exponentiation and
root can be performed. In a complex vector space, it is also possible to define the scalar product:
n
X
(x, y) = xi yi . (10)
i=0

Based on the above definition of the scalar product, the Euclidean norm can also be defined:
p
||x|| = (x, x). (11)

If a norm is defined, then a metric can also be consistently defined:

ρ (x, y) = ||y − x||. (12)

Since the real number is a special case of a complex number, the above formulas for the scalar
product, the Euclidean norm and the metric are also valid when coding nominal variables with
real numbers. And the ability to define a metric allows the use of numerical coding for the
purposes of clustering and classification [3].

2.5 Geometric interpretation of the correlation coefficient

The measure of the relationship between two random variables is their covariance:

Cov(X, Y ) = E[(X − X)(Y − Y )]. (13)

65
 Zenon Gniazdowski

The covariance normalized to unity is called the Pearson correlation coefficient:

Cov(X, Y ) E[(X − X)(Y − Y )]

R(X, Y ) = =q q . (14)
sX sY
E[(X − X)2 ] E[(Y − Y )2 ]

This expression can be further transformed to the form:

Pn
[(Xi − X)(Yi − Y )]
R(X, Y ) = qP i=1 qP . (15)
n 2 n 2
i=1 (Xi − X) i=1 (Yi − Y )

Denoting the random components of both variables X and Y as x = X − X and y = Y − Y ,

the formula for the correlation coefficient can be transformed into:
Pn
xi yi
R(X, Y ) = pPn i=1 2
pPn
2
. (16)
x
i=1 i i=1 yi

In the numerator of the above formula there is the scalar product of the vectors x and y, and
in the denominator there is the product of the lengths of these vectors. This quotient expresses
the cosine of the angle between the two vectors. This means that the correlation coefficient is
identical to the cosine of the angle between the random components of random variables [4]:
Pn
xi yi x·y
R(X, Y ) = Pn i=1
p pPn = = cos(∠x, y). (17)
i=1 xi
2 2
i=1 yi
||x|| · ||y||

2.5.1 Linear transformation of a random variable

A random variable X can be represented as the sum of its mean value X and its random com-
ponent x:
X = X + x. (18)
It can be seen that the linear transformation (a + bX) of a random variable X does not change
the direction of the vector representing its random component:

a + bX = a + b(X + x) = a + bX + bx. (19)

The constant a + bX represents the mean value of variable X after a linear transformation.
The random component x becomes a new random component bx after the transformation of
the variable X. Since the vector bx is parallel to the vector x, and the correlation coefficient
is formally identical to the cosine of the angle between the vectors representing the random
components of two random variables, this linear transformation does not change the modulus
of the correlation coefficient.

66
 On the Analysis of Correlation Between Nominal Data and Numerical Data

2.6 Pearson correlation coefficient for complex random variables

By analogy, a method of calculating the correlation coefficient for complex random variables
can be proposed. If x and y are the random components of the complex random variables X
and Y , then the correlation coefficient found as the cosine of the angle between the random
components has the following form:
Pn
Cov(X, Y ) E[xy] i=1 xi yi
R= p =p = pPn pPn . (20)
var(X)var(Y ) E[xx]E[yy] i=1 i xi
x i=1 yi yi

As noted in the introduction, in the formula above, x and y are conjugates of the complex
numbers x and y.

2.7 Least squares complex method

A linear model of yb of a complex variable y with respect to complex variables x1 , x2 , . . . , xs is
T
considered. The vector y = [y1 , y2 . . . yn ] is the dependent variable of the model. The vector
T
yb = [by1 , yb2 . . . ybn ] is the estimate of the variable y. For successive realizations of variables
x1 , x2 , . . . , xs successive approximations of the vector y are obtained in the form of calculated
values of yb:
yb1 = b0 + b1 x11 + . . . + bs x1s
yb2 = b0 + b1 x21 + . . . + bs x2s
.. . (21)
.
ybn = b0 + b1 xn1 + . . . + bs xns
The above system of equations can be represented in matrix notation:
yb = Xb. (22)
In the above equation, the following notations were assumed:
• The matrix X represents the independent variables of the model:
 
1 x11 . . . x1s
X =  ... .. .. ..  . (23)

. . . 
1 xn1 ... xns
T
• The vector b = [b0 , b1 . . . bs ] is the vector of the model parameters.
T
• The vector e = [e1 , e2 . . . en ] = y − yb is the error vector of the model.
The measure of model error is the quantity Q, which is the square of the Euclidean norm of the
error vector e. Denoting the Hermitian transposition2 of a vector (or matrix) M as M H , the
measure of error of the model takes the form:
Q (b) = ||e||2 = eH · e = (y − yb)H · (y − yb) = (y − Xb)H · (y − Xb). (24)
2
The Hermitian transposition (or conjugate transposition) of a complex vector or complex matrix is
understood as the composition of the transposition and conjugate of that vector or matrix, respectively.

67
 Zenon Gniazdowski

Table 5: Nominal data from Table 3 coded with numbers

V1 V2
2 3
2 3
2 3
2.5 3
2.5 3
2.5 2.5
2.5 2.5
1.5 2.5
1.5 2.5

After multiplication, the error measure takes the form:

H
Q (b) = (y − Xb) (y − Xb) = y H y − y H Xb − bH X H y + bH X H Xb. (25)
It can be seen that the error measure Q (the square of the Euclidean norm) is the sum of the
four real terms. On the other hand, y H Xb is a Hermitian transposition of bH X H y. Since the
Hermitian transposition of a real number does not change its value, the equality holds:
y H Xb = bH X H y. (26)
Therefore, the expression for Q (b) can be simplified a bit:
Q (b) = y H y − 2bH X H y + bH X H Xb. (27)
The model error measure depending on model parameters b should be minimal. Hence:
∂Q
= −2X H y + 2X H Xb = 0. (28)
∂b
After transformation, a system of normal equations is obtained [10][11]:
X H Xb = X H y. (29)

3 Study of linear correlation between nominal random variables coded

with numbers
Nominal data can be coded using real or complex numbers. On the one hand, the correlation
coefficient between real random vectors has an interpretation of the cosine of the angle between
their random components [4]. On the other hand, for vectors in complex space, one can define
the scalar product and the Euclidean norm, so one can also calculate the cosine of the angle
between their random components. Similarly to the real space, this cosine can be interpreted as
a correlation coefficient between complex random variables.

68
 On the Analysis of Correlation Between Nominal Data and Numerical Data

Table 6: Contingency table for nominal data with a random variable V1 containing two
classes (A and B) with identical cardinalities

X Y Total
A 3 1 4
B 2 2 4
Total 5 3 8

Table 7: Observation table identified from Table 6

V1 V2
A X
A X
A X
A Y
B X
B X
B Y
B Y

3.1 Example No. 1 - Nominal data coded with real numbers

After coding the data from Table 2, their numerical representation was obtained as presented in
Table 5. For the data from Table 5, the correlation coefficient R (V1 , V2 ) = 0.253 was obtained.
The problem was also investigated for many other datasets where both random variables
could be coded with real numbers. Coding with real numbers introduced a linear order to the
set of nominal data. In all analyzed cases, unambiguous real measures of linear correlation were
obtained.

3.2 Nominal data coded with complex numbers

It can be seen that in Example No. 1 (Subsection 3.1), for a given variable, its different classes
had different cardinalities. That is, they could be coded with different real numbers. It should be
investigated whether the existence of classes with equal cardinality also allows the estimation
of correlations between variables.

3.2.1 Example No. 2 – nominal variable with two classes with identical cardinal-
ities
Table 6 shows an example of a contingency table in which the first nominal variable contains
two classes with identical cardinalities. These are classes containing values A and B. Table 7

69
 Zenon Gniazdowski

Table 8: Data from Table 7 after complex coding of variables V1 and V2 . In the columns
V11 and V12 two phase permutations are taken into account

V11 V12 V2
−2.5 + 0i 2.5 + 0i 3
−2.5 + 0i 2.5 + 0i 3
−2.5 + 0i 2.5 + 0i 3
−2.5 + 0i 2.5 + 0i 2
2.5 + 0i −2.5 + 0i 3
2.5 + 0i −2.5 + 0i 3
2.5 + 0i −2.5 + 0i 2
2.5 + 0i −2.5 + 0i 2

Table 9: Contingency table for nominal data with a random variable containing three
classes (A, B and C) with identical cardinalities

X Y Total
A 1 4 5
B 2 3 5
C 3 2 5
D 1 2 3
Total 7 11 18

shows the reconstructed random variables V1 and V2 . Since the variable V1 takes the value of A
four times and also takes the value of B four times, the modulus of numbers used to code these
values is 2.5. On the other hand, the arbitrarily adopted phase allowing to distinguish the values
of A and B after their complex coding is φ = 00 and φ = 1800 , respectively.
Since the phase for complex coding is chosen arbitrarily, the question arises about the in-
fluence of the selected phase on the value of the correlation coefficient. For two classes with
identical cardinalities, the phases can be chosen in 2! = 2 ways. Both ways of coding a random
variable V1 are shown in the first two columns of Table 8 (V11 and V12 ). The third column
shows the variable V2 after coding. For the data from Table 8, correlation coefficients have been
estimated, which are R (V1 , V2 ) = −0.258 and R (V1 , V2 ) = 0.258, respectively.
The correlation coefficients for different phase permutations are shown in Figure 1. Com-
paring both values of the estimated correlation coefficients, it should be noted that their values
result from the arbitrarily assigned phase in the coding of the nominal variable with two classes
with identical cardinalities. It can be seen that both values of the correlation coefficient are
located on the real axis symmetrically with respect to the center of the coordinate system. This
means that the obtained solution is not unambiguous.

70
 On the Analysis of Correlation Between Nominal Data and Numerical Data

Figure 1: The correlation coefficients between variables V1 and V2 presented in the

complex plane for the coded data from Table 7, obtained for 2! = 2 different phase
permutations

3.2.2 Example No. 3 - nominal variable with three classes with identical cardi-
nalities
Another example of data is presented in Table 9. Variable V1 contains three classes of elements
with equal cardinality. These are classes A, B and C. There is also a class D with cardinality
other than classes A, B and C. Classes A, B and C contain 5 elements each. Class D contains
3 elements. The nominal value of D was coded with the real number 2 = (3 + 1) /2. The
nominal values A, B and C are assigned the modulus 3 = (5 + 1)/2. An arbitrary phase
allowing to distinguish the coded values of A, B and C can take one of three values φ = 00 ,
φ = 1200 and φ = 2400 , respectively. In table 10, the first six columns show the coded prime
variable for all 6 = 3! phase permutations. The second index digit in the variable name V1 is
used to distinguish permutations. The last column of Table 10 also shows the variable V2 after
coding.
For the data in Table 10, six complex correlation coefficients between the V1 and V2 vari-
ables were estimated, for different phase permutations attributed to different values of the V1
variable belonging to classes having identical cardinalities. The obtained results are presented
in the complex plane in Figure 2. It can be seen that the obtained correlation coefficients are
distributed in the complex plane symmetrically with respect to the point with the coordinate
0.013 lying on the real axis. In contrast to the previous example, the center of symmetry for
all obtained correlation coefficients is outside the central point of the complex plane. This is
because in the current example, the variable V1 , in addition to the three equal-cardinality classes
containing the values A, B, and C, also has a class D with a cardinality different from classes
A, B, and C. As in the previous example, the solution obtained is not unambiguous.

71
 Zenon Gniazdowski

Table 10: Data from the Table 9 after coding variables V1 and V2 . The first six columns
show the representations of the V1 variable obtained for different phase permutations

V11 V12 V13 V14 V15 V16 V2

3 -1.5+2.6i -1.5-2.6i 3 -1.5+2.6i -1.5-2.6i 4
3 -1.5+2.6i -1.5-2.6i 3 -1.5+2.6i -1.5-2.6i 6
3 -1.5+2.6i -1.5-2.6i 3 -1.5+2.6i -1.5-2.6i 6
3 -1.5+2.6i -1.5-2.6i 3 -1.5+2.6i -1.5-2.6i 6
3 -1.5+2.6i -1.5-2.6i 3 -1.5+2.6i -1.5-2.6i 6
-1.5+2.6i 3 3 -1.5-2.6i -1.5-2.6i -1.5+2.6i 4
-1.5+2.6i 3 3 -1.5-2.6i -1.5-2.6i -1.5+2.6i 4
-1.5+2.6i 3 3 -1.5-2.6i -1.5-2.6i -1.5+2.6i 6
-1.5+2.6i 3 3 -1.5-2.6i -1.5-2.6i -1.5+2.6i 6
-1.5+2.6i 3 3 -1.5-2.6i -1.5-2.6i -1.5+2.6i 6
-1.5-2.6i -1.5-2.6i -1.5+2.6i -1.5+2.6i 3 3 4
-1.5-2.6i -1.5-2.6i -1.5+2.6i -1.5+2.6i 3 3 4
-1.5-2.6i -1.5-2.6i -1.5+2.6i -1.5+2.6i 3 3 4
-1.5-2.6i -1.5-2.6i -1.5+2.6i -1.5+2.6i 3 3 6
-1.5-2.6i -1.5-2.6i -1.5+2.6i -1.5+2.6i 3 3 6
2 2 2 2 2 2 4
2 2 2 2 2 2 6
2 2 2 2 2 2 6

3.2.3 Example No. 4 - nominal variable with four classes with identical cardinal-
ities
Another example of data is shown in Table 11. The variable V1 contains four classes with equal
cardinality, containing the values of U , W , X and Y . This variable also contains a class of
Z values whose cardinality is different from the cardinality of classes U , W , X and Y . The
variable V1 was coded using complex numbers. Since the variable V1 contains four classes
with the same cardinality, 4! = 24 different phase permutations are possible. For all these
permutations, Figure 3 shows all possible values of correlation coefficients between variable V1
and variable V2 in the complex plane. Since the variable V1 contains a class of Z values with a
cardinality different from the cardinality of classes U , W , X and Y , therefore also this time the
estimated correlation coefficients are not symmetric with respect to the center of the coordinate
system. The results are distributed symmetrically around a point with the coordinate 0.542 on
the real axis. In this case, too, no unambiguous solution was obtained.

72
 On the Analysis of Correlation Between Nominal Data and Numerical Data

Figure 2: The correlation coefficients between variables V1 and V2 presented in the

complex plane for the coded data from Table 10, obtained for 3! = 6 different phase
permutations

3.2.4 Example No. 5 – nominal variable with five classes with identical cardinal-
ities
Another example of nominal data is shown in Table 12. In the considered example, the first
variable takes 450 times each of the values from the set {A, B.C, D, E}. For five classes
with equal cardinality, the number of different codings resulting from the phase permutations
is 5! = 120. For all these cases, complex correlation coefficients were estimated between
both variables. 120 different correlation coefficients were obtained, which are presented in
the complex plane in Figure 4. Since all classes contained in the set of values of the first
variable have equal cardinality, therefore the obtained correlation coefficients are distributed
symmetrically with respect to the center of the coordinate system.

4 Linear correlation ambiguity problem with complex coding of nominal

data
The examples show that for nominal data coded with complex numbers, the problem of finding
the linear correlation coefficient is ambiguous. Complex coding of k equal-cardinality classes
assumes that each of these classes is coded with a complex number of equal modulus and with an

73
 Zenon Gniazdowski

Table 11: Contingency table for data with random variable V1 containing four classes
(U , W , X and Y ) with the same cardinality

A B C D Total
U 45 20 13 12 90
W 15 45 20 10 90
X 5 8 50 27 90
Y 8 10 17 55 90
Z 10 30 50 450 540
Total 83 113 150 554 900

Table 12: Contingency table for data with random variable V1 containing five classes
(A, B, C, D and E) with the same cardinality

X Y Z Total
A 200 120 130 450
B 250 100 100 450
C 50 80 320 450
D 170 130 150 450
E 300 100 50 450
Total 970 530 750 2250

arbitrarily chosen phase equal to the phase of one of the k roots of unity. Therefore, k classes of
equal cardinality can be assigned phases in k! ways. This means that a nominal random variable
can be represented in k! ways. For these k! possible representations of the first random variable
and a unique representation of the second random variable, k! different complex correlation
coefficients are obtained. In the complex plane, these correlation coefficients are symmetrical
about some point on the real axis.
In real space, the correlation coefficient is identical to the cosine of the angle between the
random components of two random variables. Calculating the cosine of an angle in a complex
space and treating it as a correlation coefficient, due to ambiguity, does not lead to a satisfactory
result.

4.1 Searching for invariants with respect to phase permutations in complex cod-
ing of nominal data
Since the measure of linear correlation cannot be the cosine of the angle between the vectors
representing the random components of complex random variables, an attempt was also made to
find alternative correlation measures that would be invariant with respect to phase permutations
between classes with equal cardinalities. For this purpose, a linear correlation between the V2

74
 On the Analysis of Correlation Between Nominal Data and Numerical Data

Figure 3: The correlation coefficients between variables V1 and V2 presented in the

complex plane for the coded data from Table 11, obtained for 4! = 24 different phase
permutations

variable and its model Vc2 = f (V1 ) was considered. First, tests were carried out for the linear
model f (V1 ), and then for the non-linear model.

4.1.1 Correlation between variable V2 and linear model V

c2 = f (V1 )
In the considered cases, the data were selected in such a way that the first (only the first) ran-
dom variable had classes with identical cardinality. For these classes, all phase permutations
were taken into account in their complex coding. For many pairs of random variables, the least
squares method was used to identify linear models of variable V2 with respect to the complex
variable V1 . To solve the complex system of normal equations (29, the Gaussian elimination
algorithm with partial pivoting was used [12][13]. In all cases, complex correlations between
the V2 variable and its V c2 model were estimated. These correlations depended on phase per-
mutations, and the correlation modulus was identical to the linear correlation modulus between
variables V1 and V2 .
Table 13 shows an example of the data from Table 9. The table compares the correlation
coefficientsR (V
 1 , V2 ) obtained for all phase permutations from Table 10 with the correlation
coefficients R V2 , V c2 obtained for the same phase permutations.

75
 Zenon Gniazdowski

Figure 4: The correlation coefficients between variables V1 and V2 presented in the

complex plane for the coded data from Table 12, obtained for 5! = 120 different phase
permutations

As a result of the research, it was found that the correlation between the variable and its
linear model is not invariant with respect to phase permutations. Therefore, it cannot be consid-
ered as a practical measure of the correlation between a nominal random variable and a random
variable measured on a nominal scale or on a stronger measurement scale.

4.1.2 Correlation between variable V2 and non-linear model V

c2 = f (V1 )
Due to the resulting ambiguities of linear correlation, another attempt was made to find a mea-
sure used to assess the strength of the correlation relationship. This time the non-linear correla-
tion was studied.
If the nominal variable V1 contains k classes, then after numerical coding this variable
takes k of different complex or real values. Each value of the variable V1 corresponds to some
(real) values representing the variable V2 . The relationship Vc2 = f (V1 ) can be described by a
polynomial of degree k − 1. This relationship can be identified by the method of least squares.
This is symbolically illustrated by the examples in Figure 5. For k = 2, the relationship is
approximated by a straight line (Figure 5(a)), and for k = 4 it is a polynomial of the third
degree (Figure 5(b)).
By finding the correlation coefficient between the values calculated using the V c2 = f (V1 )
model and the numerical values representing the V2 variable, the non-linear correlation between
the V1 variable and the V2 variable can be estimated. If among all k different values of the vari-
able V1 there are m classes with equal cardinality (m ≤ k), then these classes are coded with
complex numbers. For all these classes, the moduli of the complex numbers are identical. To
distinguish between classes of equal cardinality, modules were arbitrarily assigned phases of

76
 On the Analysis of Correlation Between Nominal Data and Numerical Data

 
Table 13: Comparison between the correlation coefficients R (V1 , V2 ) and R V2 , Vc2
for the coded data from Table 10, obtained for 3! = 6 different phase permutations
 
No. R (V1 , V2 ) |R (V1 , V2 ) | V
c2 = b0 + b1 V1 R V2 , V
c2

1 0.194+0.104i 0.220 (5.200 + 0.012i) + (0.067 + 0.036i) V1 0.220

2 0.013+0.209i 0.209 (5.221 + 0.024i) + (0.005 − 0.072i) V1 0.209
3 0.013-0.209i 0.209 (5.221 − 0.024i) + (0.005 + 0.072i) V1 0.209
4 0.194-0.104i 0.220 (5.200 − 0.012i) + (0.067 + 0.036i) V1 0.220
5 -0.167+0.104i 0.197 (5.241 + 0.012i) − (0.058 + 0.036i) V1 0.197
6 -0.167-0.104i 0.197 (5.241 − 0.012i) − (0.058 − 0.036i) V1 0.197

(a) (b)

Figure 5: Approximation of the relationship V c2 = f (V1 ) for k of different values

contained in the variable V1 : (a) straight line for k = 2; (b) a polynomial of the third
degree for k = 4

successive roots of degree m of unity. For different phase permutations, different coefficients
in the identified polynomial will also be obtained. The question remains whether the correla-
tion coefficients between the values calculated using the V c2 model and the numerical values
representing the V2 variable change for different phase permutations.
The problem posed in this way was tested for many different data sets in which there were
classes with equal cardinality. Variables with two, with three, with four and also with five
classes with the same cardinalities were tested. For each data set, for all phase permutations,
non-linear models V c2 = f (V1 ) were identified using the least squares method. For two, three,
four and five identical cardinalities, 2! = 2, 3! = 6, 4! = 24 and 5! = 120 polynomials were
identified, respectively. In all tested cases, regardless of the phase permutations for the coded
nominal values of the random variable, the linear correlation coefficient R(V2 , V c2 ) between
the non-linear polynomial V2 = f (V1 ), and the variable V2 did not change with successive
c
permutations. This means that the measure obtained here is invariant with respect to phase

77
 Zenon Gniazdowski

Table 14: Comparison of correlation coefficients between variable V2 and its non-linear
model Vc2 = f (V1 ) for the coded data from Table 10, obtained for 3! = 6 different
phase permutations
 
No. c2 = b0 + b1 V1 + b2 V 2 + b3 V 3
V R V2 , V
c2
1 1

1 (5.074 + 0.036i) + (0.067 − 0.038i) V1 + (0.022 + 0.013i) V12 + (0.005 − 0.001i) V13 0.31
2 (5.389 + 0.073i) + (0.000 − 0.077i) V1 + (0.000 + 0.026i) V12 − (0.007 + 0.003i) V13 0.31
3 (5.389 − 0.073i) + (0.000 + 0.077i) V1 + (0.000 − 0.026i) V12 − (0.007 − 0.003i) V13 0.31
4 (5.074 − 0.036i) + (0.067 + 0.038i) V1 + (0.022 − 0.013i) V12 + (0.005 + 0.001i) V13 0.31
5 (5.705 + 0.036i) − (0.067 + 0.038i) V1 − (0.022 − 0.013i) V12 − (0.019 + 0.001i) V13 0.31
6 (5.705 − 0.036i) − (0.067 − 0.038i) V1 − (0.022 + 0.013i) V12 − (0.019 − 0.001i) V13 0.31

permutations in the complex coding of a random variable. Unfortunately, the found measure
measures the non-linear correlation between the variables V1 and V2 . For this reason, in those
cases where a linear measure of correlation is expected, the non-linear correlation coefficient
above could not be used.
Table 14 shows examples of results obtained for the data in Table 9. In the dataset, variable
V1 contained four different classes, three of which had equal cardinality. Hence, the variable
V2 could be modeled using a third degree polynomial. For three classes with equal cardinality,
3! = 6 different polynomials were obtained.

5 Conclusions
The aim of the work was to examine the possibility of measuring the strength of the linear cor-
relation relationship between two random variables measured on a nominal scale, coded with
real numbers or complex numbers. The research was conducted with the assumption that the
second random variable will always be coded with real numbers. This assumption caused the
analyzed problem to become equivalent to the problem of testing the strength of a linear cor-
relation relationship between a random variable measured on a nominal scale and a random
variable measured at least on an ordinal scale. The calculations made use of the fact that the
correlation coefficient has an interpretation of the cosine of the angle between the vectors con-
taining the random components of the analyzed variables. Since for vectors of numbers (both
real and complex) Euclidean norms and the scalar product can be calculated, it is also possible
to calculate the cosine of the angle between vectors, and consequently also to estimate the cor-
relation coefficient between them. The conducted research allowed to draw several conclusions
that may potentially be useful in the analysis of linear correlation for nominal data, and which
will be presented in more detail in the following subsections.

78
 On the Analysis of Correlation Between Nominal Data and Numerical Data

5.1 Study of the correlation relationship between nominal data coded with real
numbers
A set of data measured on a nominal scale may be coded using real numbers if there are no
classes with identical cardinalities among the classes of identical elements contained in this set.
In this case, correlations were studied for two variables, each of which could be coded with
real numbers. Coding with real numbers introduced a linear order to the nominal data sets. As
a consequence, unambiguous real measures of linear correlation were obtained in all analyzed
cases.

5.2 Study of the correlation relationship between nominal data coded with com-
plex numbers
According to the adopted assumptions, correlations between two random variables were inves-
tigated, one of which was coded with real numbers, and the other had to be coded with complex
numbers. A set of data measured on a nominal scale cannot be coded with real numbers if there
are at least two classes with identical cardinalities among the classes of identical elements con-
tained in this set. For two random variables, one of which contained complex numbers and the
other contained real numbers, complex correlation coefficients were obtained, which changed
with the permutation of phases in complex numbers coding classes of elements with equal car-
dinality. The solution to the problem of finding linear correlation turned out to be ambiguous.

5.3 Necessary condition for linear correlation

The correlation coefficient can only be estimated for real random variables, that is, variables
that have been measured on an ordinal, interval or ratio scale. The conducted research shows
that the correlation coefficient can also be estimated for nominal data without classes of equal
cardinality, which will be coded with real numbers. On the other hand, for data coded with
complex numbers, it was not possible to unambiguously estimate the correlation coefficient.
A common feature of the above-mentioned sets of real numbers, for which the linear corre-
lation coefficient can be estimated, is that these sets can be sorted and, consequently, they can
also be assigned ranks3 . Thus, these are sets for which a linear order relation can be established.
For sets containing complex numbers, linear order relations cannot be defined. The strongest
relation possible in the set of complex numbers is the partial order relation.
It can therefore be said that a necessary condition that should be met in order to be able to
estimate the linear correlation coefficient is the possibility of establishing a linear order relation
in both data sets representing the analyzed random variables.
3
For ranks, a correlation coefficient can be estimated. This is the so-called Spearman rank correlation
coefficient.

79
 Zenon Gniazdowski

5.4 Non-linear correlation

Due to the dependence of correlation on phase permutations in complex coding, an attempt was
also made to search for a possible measure of non-linear correlation that would be invariant
with respect to phase permutations in complex coding. For different permutations of phases
in a complex variable V1 , the correlation coefficient between a real variable V2 and its non-
linear approximation V c2 = f (V1 ) was investigated. For k different values of the variable V1
containing m classes with equal cardinalities (m ≤ k), m! polynomials of degree k − 1 were
identified. In all cases, identical correlation coefficients were obtained. The obtained result is
only partially satisfactory. Instead of the expected measure of linear correlation, this coefficient
measures non-linear correlation.
Regardless of this result, an unsolved problem beyond the scope of this article is the math-
ematical proof of the hypothesis that the linear correlation coefficient R(V2 , V c2 ) between the
variable V2 and the polynomial V2 = f (V1 ) does not change with successive phase permuta-
c
tions.

5.5 Alternative solutions for classes with equal cardinalities

The analysis described in the article was conducted using artificial data, specially prepared for
the purposes of this work. This was due to the fact that the author of the article did not find non-
trivial nominal data sets that would contain classes of identical elements with equal cardinality.
Thus, it cannot be excluded that the intuition that suggests that in large data sets it is unlikely
that a given variable will contain two classes with equal cardinalities is true. It is even less likely
that there could be more such classes.
However, even if this is true, it is necessary to be able to deal with such potential problems.
It seems that the methods used for pre-processing of missing or outlying data can be used here
[14][15][16]. For example, if the rejection of a single object causes two classes to no longer have
equal cardinality, then the need for complex coding of the nominal variable also disappears. On
the other hand, if the analyzed sets have high cardinalities, then one can also reasonably hope
that such an action will have a negligible effect on the statistics in the big dataset.
The proposal presented here is not elegant. This proposal is a substitute. It can be used
temporarily until further research provides a satisfactory solution to the problem presented here.

5.6 Possible simplification in nominal data coding

In subsection 2.4.1, the coding method is analogous to that used for ranking with tied ranks.
A class of identical elements whose cardinality is equal to n is assigned the rank (8). This
approach was used in the calculations described in Section 3. On the other hand, Subsection
2.5.1 noted that a linear transformation of a random variable does not change the direction of the
vector representing its random component. Thanks to this, the coding method can be simplified
for the purposes of correlation studies. In this case, it is enough to replace the ranks (8) with
cardinalities of the elements. The above change will give an unchanged value of the correlation
coefficient, and at the same time slightly simplify the calculations performed during the coding
procedure for the purpose of testing the strength of the correlation relationship.

80
 On the Analysis of Correlation Between Nominal Data and Numerical Data

5.7 Possible further directions of research

The usefulness of the proposed method for assessing the linear correlation between nominal
random variables coded with numbers requires further investigation. Regardless of the research
carried out in this work, preliminary studies were carried out on the possibility of clustering
numerically coded nominal variables into similarity classes (see Subsect. 2.1) based on the
correlation matrix. As a result, promising results have been obtained, which are a positive
incentive for further work in this area.
In subsection 4.1.2 it is noted that the linear correlation R(V2 , V
c2 ) between the variable V2
and the polynomial V2 = f (V1 ) does not change with successive phase permutations at com-
c
plex coding of the nominal variable V1 . The above result is only partially satisfactory, because
instead of the expected measure of linear correlation, it gives a measure of non-linear correla-
tion. Regardless of this, it would be interesting to solve the theoretical problem of proving or
rejecting the hypothesis that these correlation coefficients are invariant with respect to phase
permutations in the complex coding of nominal data.

Acknowledgments
The author would like to express his gratitude to Dr. Leszek Rudak for valuable comments that
helped to improve the content of the article.

References
[1] H. M. Blalock, Social Statistics. McGraw-Hill, 1960.
[2] P. Francuz and R. Mackiewicz, Liczby nie wiedza,˛ skad ˛ pochodza.˛ Przewodnik po
metodologii i statystyce nie tylko dla psychologów. Lublin: Wydawnictwo KUL, 2007.
[3] Z. Gniazdowski and M. Grabowski, “Numerical Coding of Nominal Data,” Zeszyty
Naukowe WWSI, vol. 9, no. 12, pp. 53–61, 2015. [Online]. Available: http:
//doi.org/10.26348/znwwsi.12.53
[4] Z. Gniazdowski, “Geometric interpretation of a correlation,” Zeszyty Naukowe WWSI,
vol. 7, no. 9, pp. 27–35, 2013. [Online]. Available: http://doi.org/10.26348/znwwsi.9.27
[5] ——, “O relacjach i algorytmach,” in Zbiór wykładów wszechnicy popołudniowej:
Algorytmika i programowanie. Zastosowania informatyki. Warszawska Wyższa Szkoła
Informatyki, 2011, pp. 265–286. [Online]. Available: http://akademickaseriawwsi.wwsi.
edu.pl/ksiazki/5/O_relacjachi_algorytmach.pdf
[6] S. S. Stevens, “On the theory of scales of measurement,” Science, vol. 103, no. 2684,
pp. 677–680, 1946. [Online]. Available: https://psychology.okstate.edu/faculty/jgrice/
psyc3120/Stevens_FourScales_1946.pdf
[7] StatSoft, “Elektroniczny podr˛ecznik statystyki,” 2011. [Online]. Available: https:
//www.statsoft.pl/textbook/stbasic.html

81
 Zenon Gniazdowski

[8] C. R. Mehta and N. R. Patel, “A network algorithm for performing fisher’s

exact test in r × c contingency tables,” Journal of the American Statistical
Association, vol. 78, no. 382, pp. 427–434, 1983. [Online]. Available: https:
//doi.org/10.1080/01621459.1983.10477989
[9] F. Wilcoxon, “Individual comparisons by ranking methods,” Biometrics Bulletin, vol. 1,
no. 6, pp. 80–83, 1945.
[10] T. Mok and H. B. Iz, “Vector regression introduced,” Journal of geodetic science, vol. 4,
no. 1, pp. 57–64, 2014. [Online]. Available: https://core.ac.uk/download/pdf/61032201.
pdf
[11] whuber, “Analysis with complex data, anything different?” 2013. [Online]. Available:
https://stats.stackexchange.com/q/66268
[12] M. Dryja, J. Jankowska, and M. Jankowski, Przeglad
˛ metod i algorytmów numerycznych,
cz˛eść 2. Warszawa: Wydawnictwa Naukowo-Techniczne, 1982.
[13] A. Kiełbasiński and H. Schwetlick, Numeryczna algebra liniowa: wprowadzenie do
obliczeń zautomatyzowanych. Wydawnictwa Naukowo-Techniczne, 1992.
[14] O. Maimon and L. Rokach, Eds., Data mining and knowledge discovery handbook.
Springer, 2010. [Online]. Available: https://link.springer.com/content/pdf/10.1007/
b107408.pdf
[15] D. J. Hand, H. Mannila, and P. Smyth, Principles of data mining. MIT press, 2001.
[16] M. Berthold and D. J. Hand, Intelligent data analysis. An introduction. Springer, 2007.

82