University of Veterinary and Pharmaceutical Sciences Brno

Basics of Statistics
for Students of Veterinary Medicine

Doc. RNDr. Iveta Bedáňová, Ph.D.

Brno, 2019
2
Contents

Preface ..………………………………………………………………………………….. 5

1 Basic Concepts of Statistics …………………………………………………………... 7

1.1 Statistics – Importance and Use in Medicine and Biology ………………………… 7
1.2 Types of Biological Data …………………………………………………………... 9
1.3 Statistical Sets ……………………………………………………………………… 10
1.4 Characteristics of Variables ………………………………………………………... 12
1.4.1 Probability Distribution ………………………………………………………. 14
1.4.2 Shapes of Probability Distributions ………………………………………….. 15
1.4.3 Portions of Distribution ………………………………………………………. 17

2 Descriptive Characteristics of Statistical Sets ………………………………………. 18

2.1. Measures of Central Tendency ……………………………………………………. 19
2.1.1 The Arithmetic Mean ………………………………………………………… 19
2.1.2 The Median …………………………………………………………………... 20
2.1.3 The Mode …………………………………………………………………….. 20
2.2 Measures of Variability (dispersion) ………………………………………………. 21
2.2.1 The Range ……………………………………………………………………. 21
2.2.2 The Variance …………………………………………………………………. 22
2.2.3 The Standard Deviation ……………………………………………………… 23
2.2.4 The Coefficient of Variability ………………………………………………... 24
2.2.5 The Standard Error of the Mean ……………………………………………... 24

3 Distributions Commonly Used in Statistics (Continuous data) ……………………. 26

3.1 Distributions for Population ………………………………………………………... 26
3.1.1 Gaussian Normal Distribution ……………………………………………….. 26
3.1.2 Standard Normal Distribution ………………………………………………... 28
3.1.3 Non-normal Distribution ……………………………………………………... 29
3.2 Distributions for Samples …………………………………………………………... 29
3.2.1 t-distribution (Student’s) ……………………………………………………... 29
3.2.2 Chi-square (2 ) Distribution (Pearson´s) …………………………………….. 31
3.2.3 F-distribution (Fisher-Snedecor´s) …………………………………………… 32

4 Estimation of population parameters (Confidence intervals) ……………………… 35

4.1 Normal Distribution – Estimation of  and  ……………………………………… 35
4.1.1 Confidence Interval for the Mean Value  …………………………………... 36
4.1.2 Confidence Interval for the SD () …………………………………………... 37
4.2 Non-normal Distribution – Estimation of the Median ……………………………... 39
4.2.1 Confidence Interval for the Median ………………………………………….. 39

5 Statistical Hypotheses Testing ………………………………………………………... 40

5.1 Statistical Hypothesis ………………………………………………………………. 40
5.2 Statistical Tests .……………………………………………………………………. 42
5.3 Classifications of Statistical Tests for Different Types of Data ……………………. 45

3
6 Parametric Tests ……………………………………………………………………… 48
6.1 F-test (Variance ratio Test) ………………………………………………………… 48
6.2 t-test (Student’s) ……………………………………………………………………. 51
6.2.1 Population vs. Sample Comparison (One-sample t-test) …………………….. 52
6.2.2 Samples comparison (Two-sample t-test) ……………………………………. 54

7 Non-Parametric Tests ………………………………………………………………… 59

7.1 Mann-Whitney U-Test (Rank-Sum Test) ………………………………………….. 60
7.2 Wilcoxon Signed-Rank Test ……………………………………………………….. 62

8 Relationship Between 2 Data Sets ……………………………………………………. 65

8.1 Functional vs. Statistical Relationship ……………………………………………... 65
8.2 Linear Correlative Relationship ……………………………………………………. 69
8.2.1 Regression Analysis ………………………………………………………….. 71
8.2.2 Correlation Analysis …………………………………………………………. 72
8.2.3 Significance of the Correlation Coefficient ………………………………….. 73
8.3 Non-linear Correlative Relationship ……………………………………………….. 74
8.3.1 Spearman Rank Correlation Coefficient ……………………………………... 74

9 Categorical Data ………………………………………………………………………. 77

9.1 Analysis of Categorical Data ………………………………………………………. 79
9.2 Test for Difference between Empirical and Theoretical Counts …………………… 81
9.3 Test for Difference between 2(or more) Empirical Counts ………………………... 82
9.4 Contingency Tables ………………………………………………………………… 85
9.4.1 Contingency table r x c ………………………………………………………. 85
9.4.2 Contingency Table 2 x 2 ……………………………………………………... 86

Appendix – Statistical Tables …………………………………………………………... 89

4
 Preface

There is no doubt that Statistics is a discipline which is a necessary part of education in

every biological and medical (both human and veterinary) science. Its importance results from
principle of obtaining, analysing, presentation and interpretation of data in medical sciences. Our
knowledge and experience in the area of medicine would be full of many errors and mistakes if we
would not have statistics. We would not be able to properly interpret the knowledge and experience
we gained in the area of medicine without use of statistical methods. From this point of view the
statistics is among basic subjects of veterinary medicine education. Contents of the discipline
Statistics for students of veterinary medicine predominantly consists of statistical methods for
description of data sets and testing of hypotheses for quantitative and qualitative data with
specialization on data and hypotheses used in the area of biology and veterinary medicine.
Statistics has a practical impact especially in the area of research and development in
biological and medical sciences. It may also be important in the area of clinical veterinary practice
and in the sphere of hygiene and ecology in case of inspections regarding the animal provenance
food safety. Statistics education is particularly focused on the ability to solve specific problems of
veterinary medicine with the use of biostatistical methods.
This textbook is designed especially for English speaking students of veterinary faculties at
the University of Veterinary and Pharmaceutical Sciences Brno but it can also be useful for other
users of statistics in the area of biological and medical sciences. It represents a basic collection of
data-analysis techniques and it may serve as an introductory textbook of biostatistics, assuming no
prior knowledge of statistics and mathematics. The text is intended to be a brief guide to choosing
and applying the appropriate statistical tests and methods for analysis of biological and medical
data. The textbook is compact enough to be used in the laboratory or in the field while providing
sufficient details to give the user some knowledge of the theoretical basis for the methods covered.
Because of the limited time that can be devoted to this discipline at the University of Veterinary and
Pharmaceutical Sciences Brno, the statistical techniques described in this textbook are sometimes
simplified in order to be understandable also to non-mathematically specialised students of
veterinary medicine as well as other biological disciplines.
We hope that this textbook, in spite of its concise form, will serve students of veterinary
medicine as well as other users of statistics in the sphere of biosciences as a useful and handy tool
covering basic statistical techniques and analyses.

Author

Brno, February 2007

5
6
 Chapter 1
Basic concepts of statistics

1.1 Statistics – Importance and Use in Medicine and Biology

Statistics is the science that allows us to formulate and describe complex data in a short
form, easily understood by all professionals. It allows us to compare data (numerical facts resulting
from observations in some investigative monitoring) and gives us probabilities of the likelihood of
studied events. The term “statistics” is often encountered as a synonym for “data”: statistics of
sickness rate during the last month (how many patients, number of cured patients), labour statistics
(number of workers unemployed, number employed in various occupations), election statistics
(number of votes in different regions, parties), etc. Hereafter, this use of the word “statistics” will
not appear in this textbook. Instead, “statistics” will be used in its other common manner: to refer to
the analysis and interpretation of data with a view toward objective evaluation of the reliability of
the conclusions based on the data.
Statistics are predominantly needed in more probabilistic and less predictive sciences such
as biology and applied biology (medicine). In a predictive science such as physics, to find out how
fast a 300 g stone will reach the ground if dropped from a height of 30 m, one has only apply the
data in the appropriate formula to obtain an accurate answer. In art, on the other hand, the
evaluation of a given piece depends to a great extend on subjective criteria. Medicine falls
somewhere in between. There are numerous and complex physical/chemical events occurring
simultaneously which cannot be evaluated separately. For instance, if one wants to determine the
time of induction, or return, of a given reflex of the specific tendon, the issue is more complicated
than it appears initially. In this case we are dealing with transmission of electrical potential
difference across many nerves and transmission to muscles, making relevant calculations more
tedious and specific data less well known. Furthermore, the specific functions are affected by
several other components of internal milieu such as the level of hormones. To complicate matters,
this will represent just one out of many concomitant functions of a total inhomogeneous system,
living body. It is, therefore, easy to appreciate why biological sciences in general, and applied
biology such as medicine, in particular, are probabilistic in nature. As a result, a good grasp of
statistics is essential for one to be effective in this field.
Statistics applied to biological problems is simply called biostatistics or, sometimes,
biometry (the latter term literally meaning “biological measurements”). As biological entities are
counted or measured, it becomes apparent that some objective methods are necessary to aid the
investigator in presenting and analysing research data. Although the field of statistics has roots
extending back hundreds of years, its development began in earnest in the late nineteenth century,
and a major impetus from early in this development has been the need to examine biological and
medical data.

7
 Nowadays the statistical methods are common and more and more important in all
biological and medical sciences. Biostatistics is a necessary part of every biology and medicine
(both human and veterinary) education. When dealing with living organisms, we must always keep
in mind, that every individual is unique and there is a high level of insecurity regarding its
reactions. Therefore all data obtained from biological objects may be very different and variable.
This results from vast genetic variability in living organisms and also from other aspects (ambient
environment, adaptability, etc.).
This large variability of biological data causes problems and difficulties in monitoring,
measurements and data acquisition in animals and other living organisms. These problems can
partially be solved by means of statistics, because only statistical methods are able to take into
account this great variability of biological data, evaluate them and give correct inferences
concerning studied biological objects. Statistics handles variability in two ways. First it provides
precise ways to describe and measure the extent of variability in our measured data. Secondly it
provides us with methods for using those measures of variability to determine a probability of the
correctness of any conclusions we draw from our data.
Before data can be analysed, they must be collected, and statistical considerations can aid in
the design of experiments and in the setting up of hypotheses to be tested. Many biologists attempt
the analysis of their research data only to find that too few data were collected to enable reliable
conclusions to be drawn, or that much extra effort was expended in collecting data that cannot be of
ready aid in the analysis of the experiment. Thus, knowledge of basis statistical principles and
procedures is important even before an experiment is begun.
Once the data have been obtained, we may organize and summarize them in such a way as
to arrive at their orderly and informative presentation. Such procedures are often termed descriptive
statistics. For example, tabulation might be made of the heights of all students of the Faculty of
veterinary medicine, indicating an average height for each sex, or for each age. However, it might
be desired to make some generalizations from these data. We might, for example, wish to make
reasonable estimate of the heights of all students in the university. Or we might wish to conclude
whether the males in the university are on the average taller than the females. The ability to make
such generalized conclusions, inferring characteristics of the whole from characteristics of its parts,
lies within the realm of inferential statistics.

Summary: Use of Biostatistics in the Area of Veterinary Medicine:

1. Research – analysis of data measured during experiments, e.g. in the course of
examination of effects of new drugs, remedies, feed mixtures, medical treatments, methods, etc.
We can confirm or reject hypotheses that are investigated in experiments by means of special
statistical methods (statistical hypotheses testing - inferential statistics).
2. Clinical practice – evaluation and generalization of observation results in clinical practice
– i.e. monitoring and comparison of disease incidence in different groups of animals, in regions,
time periods, etc. (descriptive statistics). We can for example compare the sickness rate in
animals observed in a stable to the statistical sickness rate (known in the whole population from
long-term monitoring). Statistical methods (inferential statistics) help us to decide, whether the
possible increase in sickness rate in the stable is only random, or whether it is caused by some
ambient causes (e.g. bad treatment, feeding stuff or conditions in the stable).

8
 1.2 Types of Biological Data
In biological and medical sciences, we analyse biological properties of living organisms that
are described on the basis of selected biological characters. These biological characters can be
measured usually by some means. Their values differ from one entity to another – therefore they are
called variables in statistics. Variables describe studied biological characters (properties of living
organisms usually) and they can quantify (more or less) these biological properties. Different kinds
of variables may be encountered by biologists, and it is desirable to be able to distinguish among
them. Variables can be quantitative or qualitative. Quantitative variables record the amount of
something (ordinal data and numerical data); qualitative variables describe the category to which
the data ca be assigned and are therefore sometimes referred to as categorical data.
Exactness of those biological variables can differ in their values – according to the exactness
we can distinguish between 3 types of biological data in statistics:

A. Nominal Data (Categorical)

Sometimes the variable under study is classified by some quality it possesses rather than by
a numerical measurement. In such cases the variable may be called an attribute, and we are said to
be using a nominal scale of measurement (categories, classes). On a nominal scale (“nominal” is
from the Latin word for “name”), animals might be classified as male or female, as left- or right-
handed, as ill or healthy, with or without horns, as vaccinated or not vaccinated, as alive or dead,
etc. Such variables describe only some quality in a living organism that is not measurable – there
are no values (it is the lowest level of quantification in variables). We can only distinguish between
2 possibilities (situations): the quality is either present or not present in each individual observed.
Sometimes data from an ordinal or numerical scale of measurement (see below) may be recorded in
nominal-scale categories. For example, heights may be recorded as tall or short, or performance on
an examination as pass or fail, etc.

B. Ordinal Data (Rank-Order Data)

They are represented by an ordering (up or down) of observations based on subjective scale
given by an evaluator (experimenter). These data are arranged into an ascending or descending row
and may be a record only of the fact that one individual has lower intensity of a studied biological
character than the other (with no indication of how much more). Differences between various
degrees on the scale are different and dependent on a measure given by the evaluator (e.g.
classification using grades in school, points in breeding competitions, marks for animal behaviour
in experiment, etc.). Thus, we are dealing with relative differences rather than with quantitative
differences. Such data that consists of an ordering or ranking of measurements are said to be on an
ordinal scale of measurement (“ordinal” being from the Latin word for “order”). One may speak of
one biological entity being shorter, darker, faster, or more active than other; the sizes of five cell
types might be labelled 1, 2, 3, 4, and 5, to denote their magnitudes relative to each other; or success
in learning may be recorded as A, B, C, D, E, or points gained in breeder competitions, etc.

C. Numerical Data
They are represented by exact numeric values obtained by means of some objective
measurement (meter, thermometer, scale, measuring device etc.). Differences between various
degrees on the scale are uniform – the numerical scale consists of the same intervals. There is the

9
highest level of quantification in statistical data – they are most often used for statistical evaluation.
Numerical variables allow us to record an amount for each observation and to compare the
magnitude of differences between them. They can be either continuous or discrete.
- Discrete Data (discontinuous)
Variables that can take only specific available values – most often integer numbers. For
example the number of bacterial colonies on a Petri dish can only be a positive integer
value; there can be 24 colonies, but never 24.5 colonies or -24 colonies. Similarly also
number of laid eggs, puppies in a litter, animals in a stable, patients, cells, etc.
- Continuous Data
These variables can take on any conceivable value in our infinite spectrum of real
numbers - within any observed real range (height, length, weight, volume, body
temperature, concentration of enzyme, etc.)

Different categories of statistical data have their own specific statistical method used for
their examination. These methods are differently exact according to the exactness of data category.
Statistical methods used for numerical or ordinal data are more exact and generally they are not
applicable to nominal data (since nominal data contain only little information for exact methods). It
is possible reversely: the less exact methods intended for nominal (or ordinal) data are useful also
for numerical data. In this case we can purposely use these not very precise methods e.g. for
preliminary analyses that must be performed quickly.
Sometimes the distinction between different types of data is not very obvious, e.g. categories
fall into a natural order and it is not reasonable to distinguish between ranked and categorical data.
Values of heights in students are continuous data. Of course, they may also be ranked: smallest,
next smallest, ……., highest. If the height is categorized into three groups, <160, 161-180, >181,
the values are still ordered, but we have lost a lot of information. There are so many ties that
analysis methods will be sensitive only to three ranks, one for each category. Although we could
analyse them as categories A, B, and C, the methods treating them as ranks first, second, and third
are still “stronger” methods. When rendering data into categories, one should note whether the
categories fall into a natural order. If they do, treat them as ranks. For example, categories of ethnic
groups do not fall into a natural order, but the pain categories severe, moderate, small, and absent
do.
Note that we can always change data from higher to lower level of quantification, that is,
continuous to discrete to ranked to categorical, but not the other way. Thus, it is always preferable
to record data as high up on this sequence as possible; it can always be dropped lower.

1.3 Statistical Sets

Basic to statistical analysis is the desire to draw conclusions about a group of measurements
of a variable being studied. Statistical variables in biological and medical sciences may be studied
in groups (statistical sets) of living individuals – animals, plants, cells, bacteria, etc. (in general
entities or items). We can distinguish between 2 types of statistical sets in statistics:
 The Population (the Universe) : N =  (number of members)

10
 By the word population, we denote the entire set of subject about whom we want
information. Thus, the population means also “all items” (individuals) that could show studied
variable. If we were to take our measurements on all individuals in the population, descriptive
statistics would give us exact information about the population. Populations are often very large
sets; number of individuals in the population is considered to be “endless” (for statistical purposes
and calculations). In practice, the number of members in the population can be literally “endless” –
especially from the time viewpoint: e.g. body weights in all cattle in CR, dogs in Europe, etc. –
number of individuals is not fixed (it fluctuates since new members are born and others die).
Occasionally populations of interest may be relatively small, such as the ages of men who
have travelled to the moon or the heights of women who have swum the English Channel. If the
population under study is very small, it might be practical to obtain all the measurements in the
population. Generally, however, populations of interest are so large as to render the obtaining of all
the measurements unfeasible (it’s time-consuming, expensive, etc.). We are not able to obtain all
possible measurements from the population in practice – for example, we could not reasonably
expect to determine the body weight of each dog in Europe. What can be done in such cases is to
obtain a subset of all measurements in the population. This subset of measurements comprises
a sample, and from the characteristics of samples conclusions can be drawn about the characteristics
of the populations from which the samples come.
Often one samples a population that does not physically exist. Suppose an experiment is
performed in which a food supplement is administered to thirty piglets, and the sample data consist
of the growth rates of these thirty animals. Then the population about which conclusions might be
drawn is the growth rates of all the piglets that conceivably might have been administered the same
food supplement under identical conditions. Such a population is said to be “imaginary” and is also
referred to as “hypothetical” or “potential”.

 The Sample (the Subset) : n (number of members)

By the word sample, we denote definite number (marked as n) of individuals selected from
population. We perform measurements in this sample in practice calculate descriptive statistics
from the sample and use them to estimate the true characteristics of the population. The definite
(often quite small) number of members in samples measured implies some inaccuracies in
examinations and calculations performed on the basis of these small subsets in comparison to the
whole population.
To reach the most valid conclusions about a population, the sample must be a representative
subset of the population the sample must fully represent the population in its characteristics. For
example, male dogs tend to be heavier than females because they tend to be bigger. We could be led
into making wrong decisions on the basis of weight if we generalized about dogs from a sample
containing only males. We would say this sample is biased. To avoid bias, our sample should
contain the same proportion of males as the dog population contains male dogs.
Samples from populations can be obtained in a number of ways; however, to reach valid
conclusions about populations by induction from samples, statistical procedures typically assume
that the samples are obtained in a random fashion. To sample a population randomly requires that
each member of the population has an equal and independent chance of being selected. That is, not
only must each measurement in the population have an equal chance of being chosen as a member

11
of the sample, but the selection of any member of the population must in no way influence the
selection of any other member.
It is sometimes possible to assign each member of a population a unique number and to
draw a sample by choosing a set of such numbers at random. This is equivalent to having all
members of a population in a hat and drawing a sample from them while blindfolded. We may not
to do a subjective choice generating a random sample; we can use e.g. table of random digits from
statistical literature (E.g. Zar: Biostatistical Analyses), drawing lots for registration numbers of
animals in a stable, etc.
Another requirement for a dependable generalization about certain characteristics is an
appropriate size of sample. The pattern of sample values (as well as sample descriptive
characteristics) gets closer in nature to the pattern of population values as the sample size gets
closer to the population size. It means that the bigger is our sample, the better, but there are
practicable limits in practice - not enough time, money, etc. Thus, we must do compromises often in
practice; in general the sample size above 30 members is considered to be an appropriate sample
size to give us results of calculations, which are comparable to population. However, samples that
consist e.g. only of 10 individuals may be sufficient in some cases in practice.

1.4 Characteristics of Variables

Data obtained through measurements in a random sample represent a random variable

(discrete or continuous), that can be described by means of some specific terms used in statistics:
Variant Sequence – a listing (rank) of all the observed values (variants) measured in
a sample, that are arranged up (in an ascending row) or down (descending row).
E.g.: 2, 3, 4, 4, 5, 5, 5, 6, 6, 7, 7, 8 (discrete data – number of pups in a litter).
Note that some values are repeated in the row – this repetition (multiplication) of certain
value is called frequency of this variant.
Frequency of the Variant – how many times each value (variant) is observed (repeated)
in the sample (e.g.: frequency of values 2, value 3, and value 8 is 1 (these values occur only
once in the row), frequency of value 5 is 3 (value 5 is repeated 3 times in the row).
Frequency Distribution – a graphically presented distribution of all the observed
frequencies in the sample (among the various values – variants). It is the way the data are
distributed along the scale (or axis) of the variable of interest.
The frequency distribution is a concept through which most of the essential elements of
biological and medical statistics can be assessed. The graphical presentation of frequency
distribution can have various forms that may be slightly different in discrete data and
continuous data. Bar charts with separate strokes are most often used for presentation of
frequency distribution of discrete variables (see Fig.1.1 Frequency distribution for number of
pups in a litter) and histograms (column charts) are common in continuous variables.

12
 Fig. 1.1 Frequency Distribution – Discrete Data (Bar Graph)

When measuring continuous data we create classes i.e. equivalent intervals (categories) of
data to simplify the situation. Number of classes should be appropriate according to the sample size:

Up to 100 items in the sample: we usually create 6 - 9 classes,

Up to 500 items in the sample: we usually create 10 -15 classes,
Above 500 items in the sample: we usually create 16 – 20 classes.

Then a certain number of values fall into each defined interval (class). All the data in this
interval get the same value – so called midpoint (mean value) of the class. These values replace the
original values measured in all individuals in the sample monitored. In this way, we are able to
obtain a frequency of the class i.e. number of items (individuals) in the appropriate interval to draw
a chart of frequency distribution for continuous data.

Fig.1.2 represents the frequency distribution of continuous data (e.g. body weights).

13
 Fig. 1.2 Frequency Distribution – Continuous Data (Histogram)

As you can see from the chart above, it is possible to use one more type of graphical
presentation of frequency distribution in the sample - a polygon. Polygon is represented by a broken
line that joins tops of columns in the midpoints of the classes. The shape of the polygon is specific
for the unique sample that was used for our measurements i.e. the shape of polygon varies from
sample to sample. Therefore we can use the term “empirical curve” for polygon as well (from Latin
word empiria = experience).

Most of biological variables (both discrete and continuous) possess the characteristic
property – frequencies in the middle of the sample (around the mean value) are the highest and
frequencies of extremely small and large values in the sample are the lowest.

1.4.1 Probability distribution

When repeatedly measuring the same variable in various sample(s) selected from one
population we get different shape(s) of empirical curve(s) – this results from genetic variability of
individuals in the sample. Nobody (empirical curve) will have the same shape. Empirical curves for
different samples (obtained from one population) are located along only one theoretical curve
(continuous) that describes probability distribution of the variable in the population. In practice
we can’t measure all data for constructing of theoretical curve – its shape is only hypothetical
(theoretically). However we can estimate its shape on the basis of empirical curves of samples
selected from the population (Fig.1.3).

14
 Fig. 1.3 Frequency (Probability) Distribution

The frequency distribution for the whole population is a statistical distribution that
determines the probability of occurrence of values in studied variable; therefore we use the term
probability P(x) instead of frequency on the axis y. The term frequency used with samples
represents an absolute scale (that is possible only in samples - they have definite number of
entities), whereas the term probability represents a relative scale (proportion of cases) that is
necessary to be used in populations, where the number of entities is infinite.

1.4.2 Shapes of Probability Distributions

The probability distribution gives us the fundamental piece of statistical information of

studied variables. It contains all the basic information we need for our statistical methods. From
distribution, we can learn what is typical or characteristic of variable (e.g. where the majority of
values are located in population, how much the values are variable etc.). Various biological
characters can have different shapes of probability distribution.
Most variables in biological sciences follow normal (Gaussian) probability distribution that
is symmetrical – most values in population are located around the mean value (Fig. 1.4). But some

15
variables in biology and medicine can behave in another way – they can follow different probability
distributions, which have other shapes of their curves: asymmetrical - up to extreme (Fig. 1.5, 1.6)
or non-normal, irregular (Fig.1.7):

Fig. 1.4 Normal (Gaussian) Distribution - symmetric bell curve

Fig. 1.5 Asymmetric (right-skewed, left-skewed) Distribution

y y

x x

Fig. 1.6 Extreme (decreasing, increasing)

y y

x x

Fig. 1.7 Non-normal (unknown, irregular, 2 and more peaks)

y y

x x

16
 1.4.3 Portions of Distribution

For each distribution we can define measures (quantiles) that divide a group of data -
population (displayed as the area under the curve of distribution) into 2 parts (portions):
- Values which are smaller than quantile,
- Values which are larger than quantile.
There are specific quantiles used for description of distributions in statistics:
50% quantile – x0.5 (called median) divides a group of data into 2 equal halves (Fig. 1.8):

Fig. 1.8 Median (50% quantile) in normal and non-normal distribution

y y

x x

Quartiles – divide a group of data into four equal parts,

Deciles – divide a group of data into 10 equal parts,
Percentiles – divide a group of data into 100 equal parts.

Important quantiles and their corresponding proportions of the most common distribution
curves are tabulated in statistical tables and used as critical values in statistical hypotheses testing
(see Chapter 5) or as coefficients in calculations (see Chapter 4: confidence intervals of statistical
parameters – e.g. mean value , standard deviation ).

17
 Chapter 2
Descriptive Characteristics of Statistical Sets

The aim of statistical data evaluation: to get an image of monitored biological characters in
the whole population on the basis of data samples. At first we usually classify observed sample data
according to the measured values, arrange the variant sequences and draw graphs of frequency
distributions. These arranged data us give basic information about the sample and offer source
material for further statistical methods of data evaluation for monitored biological characters.
The deeper analysis follows, when we try to resume data information into one or several
numbers by means of specific exactly defined parameters (statistical characteristics). We can’t
really determine the exact values of these parameters at the level of the whole population, therefore
we select a sample (or several samples) from the studied population and we calculate the so-called
statistics from this sample data. It serves as estimation of the exact population parameters.
Several measures help to describe or characterize a population. For example, generally
a preponderance of measurements occurs somewhere around the middle of the range of a population
of measurements. Thus, some indication of a population “average” would express a useful bit of
descriptive information. Such information is called a measure of central tendency, and several such
measures (e.g. the mean and the median) will be discussed below. It is also important to describe
how dispersed the measurements are around the “average”. That is, we can ask whether there is
a wide spread of values in the population or whether the values are rather concentrated around the
middle. Such a descriptive property is called a measure of dispersion (e.g. the range, the standard
deviation, the variance etc.).
A quantity such as a measure of central tendency or a measure of dispersion is called
a parameter when it describes or characterizes a population, and we shall be very interested in
discussing parameters and drawing conclusions about them when studying a biological character in
the population. However, one seldom has data for entire populations, but nearly always has to rely
on samples to arrive at conclusions about populations. Thus, as mentioned above, one rarely is able
to calculate the true exact parameters. However, by random sampling of populations, parameters
can be estimated very well by means of special statistical methods (see the chapter Estimation of
population parameters). Due to the statistical methods, we can determine so-called confidence
intervals for population parameters or to calculate the estimates of population parameters called
statistics (on the basis of sample data). It is statistical convention to represent the true population
parameters by Greek letters and sample statistics by Latin letters.
Among the most often used measures of central tendency belong: the mean, the median, and
the mode. Among the most often used measures of dispersion and variability of statistical sets
belong: the range, the variance, the standard deviation, the coefficient of variability, and the
standard error of mean.

18
 2.1 Measures of Central Tendency

In samples, as well as in populations, one generally finds a preponderance of values

somewhere around the middle of the range of observed values. The description of this concentration
near the middle is an average, or a measure of central tendency to the statistician. It is also termed a
measure of location, for it indicates where, along the measurement scale, the sample or population
is located. Various measures of central tendency are useful parameters, in that they describe
a property of populations. The characteristics of the most often used parameters and the sample
statistics that are good estimates of them are described below.

2.1.1 The Arithmetic Mean (average - AVG)

Notation:  (Population), x (Sample)

The most widely used measure of central tendency is the arithmetic mean, usually referred
to simply as the mean, which is the measure most commonly called an “average” (the term
“average” is used predominantly for sample statistic, the term “mean” is used for population exact
parameter most often).
Each measurement in a population may be referred to as xi value. The subscript i might be
any integer value up through N, the total number of values X in the population.
The calculation of the population mean  (the theoretical exact parameter):

x i

 i 1

The most efficient and unbiased estimate of the population mean , is the sample mean,
denoted as x (read as “x bar”). Whereas the size of the population (which we generally do not
know) is denoted as N, the size of a sample is indicated by n (definite number of members in
a specific sample used for measurements).
The calculation of the sample average x :
n

x i

x i 1

n
Properties of the mean:
- Mean is affected by extreme values in the set (when changing one value xi arithmetic mean
change as well). The average is the correct measure of the central tendency of a sample only if
the sample is homogenous enough in its values (it should be used in homogenous regular
distributions (Gaussian) only). Otherwise, especially in small samples, the average can be

19
 biased through possible extreme values in the sample, and does not represent the correct
measure of centre in sample (in irregular distributions as well).
- It has the same units of measurement as the individual observations.
-  x i  x   0 - the sum of all deviations from the mean will be always 0.

2.1.2 The Median

Notation: ~ (Population), ~
x (Sample)

The median is typically defined as the middle measurement in an ordered set of data
(ordered in an ascending or descending row). That is, there are just as many values larger than the
median as there are smaller ones. The sample median is the best estimate of the population median.
In a symmetrical distribution the sample is also an unbiased estimate of , but it is not as efficient
a statistic as x , and should not be used as a substitute for x . If the frequency distribution is
asymmetrical, the median is poor estimate of the mean.
The median of a sample of data may be found by first arranging the measurements in order
of magnitude. Then the middle value of this row is the median.
In larger samples we can find out, which datum in a ranked sample data is median by means
n 1
of the calculation of rank of this figure: (it can be applied for centre of any row of n values
2
in math generally).

- If the sample size (n) is odd  there is only 1 middle value (rank will be an integer) and
indicates, which datum in ordered sample is the median.
- If n is even  rank of the median is a half-integer and it indicates that there are two middle
values, and the median is a midpoint (mean) between them.

Properties of median:
- Median is not affected by extreme values in the sample;
- Median = 50% quantile (divides the sample data into 2 halves: values that are smaller then
median and values that are larger than median);
- It may be used in irregular (asymmetric) distributions – in this case median is better
characteristic for the middle of the set than the average.

2.1.3 The Mode

Notation: ̂ (Population), x̂ (Sample)

The mode is commonly defined as the most frequently occurring measurement in a set of
data (the value with the highest frequency). Mode always indicates the top of the distribution curve.
A distribution with two modes (two tops) is said to be bimodal and may indicate a combination of
two distributions with different modes (e.g. heights of men and women). The sample mode is the

20
best estimate of the population mode. When we sample a symmetrical unimodal population, the
mode is an unbiased estimate of the mean and median as well as, but it is relatively inefficient and
should not be so used.
The mode is a somewhat simple measure of the central tendency and it is not often used in
biological and medical research, although it is often interesting to report the number of modes
detected in a population, if there are more than one.
Properties of the mode:
- Mode is not affected by extreme values in the sample.
- It is not a very exact measure of the middle of the set.

2.2. Measures of Variability (dispersion)

Mean value indicates only the centre of sets but does not indicate the dispersion of values
around the centre (how much the values are scattered). For this purpose we use measures of
variability that describe this dispersion of values around the centre of the set, and determine also
the reliability of mean value of the set - reliability will be larger in samples that have similar
(homogenous) values and no extreme values. Measures of variability of a population are exact
parameters of the population, and the sample measures of variability that estimate them are
statistics.

2. 2. 1 The Range
The difference between the highest and the lowest value in the set of data is termed the
range. If sample measurements are arranged in increasing order of magnitude, as if the median were
about to be determined, then sample range R is calculated:

R = xmax – xmin

Properties of the range:

- It is dependent on 2 extreme values of data
- It is a relatively rough measure of variability – it does not take into account any
measurements between the highest and lowest value.
Furthermore, as it is unlikely that the sample will contain both the highest and the lowest
values in the population, the sample range usually underestimates the population range.
Nonetheless, it is considered useful by some to present the sample range as an estimate (although
a poor one) of the population range. Whenever the range is specified in reporting data, however, it
is usually a good practice to report another measure of variability as well.
It is evident that the range conveys no information about how clustered about the middle of
the distribution the measurements are. As the mean is so useful a measure of central tendency, one
might express dispersion in terms of deviations from the mean.

21
 If we want to express variability in terms of deviations from the mean, there will be
a difficulty: as the sum of all deviations from the mean, i.e.,  ( xi  x ) , will always equal zero,
summation would be useless as a measure of dispersion and variability. Summing the absolute
values of the deviations from the mean results in a quantity that is an expression of dispersion about
the mean. Dividing this quantity by n yields a measure known as the mean deviation (or the mean
absolute deviation) of the sample. This measure has the same units as do the data, but it is not very
often used as a measure of dispersion and variability in practice.
Another method of eliminating the signs of the deviations from the mean is to square the
deviations. The sum of squares of the deviations from the mean is called the sum of squares,
abbreviated SS, and is defined as follows:

population SS   ( xi   ) 2

sample SS   ( xi  x ) 2

By means of term “SS” we can define other measures of variability in the set of data:

2.2.2 The Variance

Notation:  2 (Population), s 2 (Sample)

The variance is defined as the mean sum of squares about the mean value of data.
Sometimes this measure is called also mean square – short for mean squared deviation.

 x   
2
i
2  i 1 Population variance
N

The best estimate of the population variance, 2, is the sample variance, s2:
n

 x  x 
2
i
s2  i 1 “Estimated” variance (used for samples)
n 1

If, in equation of population variance, we replace  by x and N by n, the result is a quantity

that is a biased estimate of 2 and this would be not a correct measure of variability for sample data
(especially in the case of small sample sizes). The dividing of the sample sum of squares by n-1
(called degrees of freedom, abbreviated DF or df) rather than by n, yields an unbiased estimate, and
it is equation for “Estimated” variance that should be used to calculate the sample variance.
If all measurements are equal, then there is no variability and s2 = 0. And, s2 becomes
increasingly large as the amount of variability, or dispersion, increases. Because s2 is a mean sum of
squares, it can never be a negative quantity.

22
 The variance expresses the same type of information as does the mean deviation, but it has
certain very important properties relative to probability and hypothesis testing that make it distinctly
superior. Thus, the mean deviation is very seldom encountered in biostatistical analysis.
The calculation of s2 can be tedious for large samples, but it can be facilitated by the use of
the equality:

 x  2

x  n
2 i
i
s2 
n 1

This formula is much simpler to work with in practice, therefore is often referred to as
a “working formula”, or “machine formula” for sample variance. There are, in fact, two major
advantages in calculating s2 by this equation rather than by equation of previous (original) formula
for “estimated variance”. First, fewer computational steps are involved, a fact that decreases chance
of error. On many calculators the summed quantities, xi and xi2, can both be obtained with only
one pass through the data, whereas the original formula for “estimated variance” requires one pass
through the data to calculate x and at least one more pass to calculate and sum the squares of the
deviations, xi - x . Second, there may be a good deal of rounding error in calculating each deviation
xi - x , a situation that leads to decreased accuracy in computation, but which is avoided by the use
of the latter formula above.
Variance has the square units as the original measurements. If measurements are in grams,
their variance will be in grams squared, or if the measurements are in cubic centimetres, their
variance will be in terms of cubic centimetres squared, even though such squared units have no
physical interpretation.

2.2.3 The Standard Deviation ( SD)

Notation:  (Population), s (Sample)

The standard deviation is the positive square-root of the variance; therefore, it has the same
units as the original measurements.
Thus, the formula for a population SD is:
N
 x  2

  xi    x
2

2 i
i
 i 1
or  N
N N

and for a sample SD:

  xi  x 
2  x  2

x 
2 i
i
s i 1
or s n
n 1 n 1

23
 Remember that the standard deviation is, by definition, always a nonnegative quantity. The
standard deviation and the mean shall be reported with the same number of decimal places.

2.2.4 The Coefficient of Variability („Relative SD“)

The coefficient of variability is a relative measure of variability (expressed in % most often).
It is not dependent on measurement units like the standard deviation, therefore it can be used for
comparison of the variability in data sets with different magnitude of their units (e.g. body weight in
mice and cows).
Calculation formulas for sample coefficient of variability:

s s  100
V  or V  [%]
x x

The coefficient of variability expresses sample variability relative to the mean of the sample;
because s and x have identical units, V has no units at all, a fact emphasizing that it is a relative
measure, divorced from the actual magnitude or units of measurements of data.
The coefficient of variability of a sample, namely V, is an estimate of the coefficient of
variability of the population from which the sample came (i.e., an estimate of /).

2.2.5 The Standard Error of the Mean (SEM, SE)

Notation:  x (Population), s x (Sample)

The population standard error of the mean is the theoretical standard deviation of all sample
means of size n that could be drawn from a population. The sample standard error of the mean can
be used as a measure of the precision with which the sample mean x estimates the true population
mean .
We don’t know what the true mean value in the population is. We can only estimate it by
means of the sample average. But we don’t know how precise our calculation is and what’s the
difference between our calculated sample AVG and the true population mean .. SEM may serve as
a measure of precision of the calculated sample mean.
Value of the standard error of the mean depends on both the population variance (2) and the
sample size (n):

x 
n
Since in general we don’t know the population variance, the best estimate for the population
standard error of the mean in practice (sample standard error of the mean) is calculated as:

s
sx 
n

24
 The sample standard error of the mean is useful for construction of the confidence interval
for the mean. The true mean value of population will lie within the interval sample average AVG 
SEM (see Chapter 4 for details).

25
 Chapter 3
Distributions Commonly Used in Statistics
(Continuous Data)

In Chapter 1, we saw how frequency distributions arise from sample data and that the
population distribution, arising from sampling the entire population, becomes the probability
distribution. This probability distribution is used in the process of making statistical inferences
about population characteristics on the basis of sample information. There are, of course, endless
types of probability distributions possible. However, luckily, the great majority of statistical
methods for continuous data use only several probability distributions.
The probability distributions most often used in statistics for continuous data are the normal,
the non-normal, Student’s t, Pearson’s 2 (chi-square), and Fisher-Snedecor’s F distribution. Rank–
order methods depend on distributions of ranks rather than continuous data, but several of them use
the normal or chi-square. Categorical data depend mostly on the chi-square, with larger samples
transformed to normal. We need to become familiar with these commonly used distributions to
understand most of the methods given in this text. The following paragraphs describe these
distributions and some of their properties needed to use and interpret statistical methods.
Some of these distributions are useful with the population variables and some of
distributions are useful with the sample variables.

3.1 Distributions for Population

3.1.1 Gaussian Normal Distribution

The normal distribution, called Gaussian by some users, is the perfect case of the famous
symmetrical bell curve. In biology, a great many of data sets naturally follow Gaussian normal
distribution, at least approximately.
The graphical description of the Gaussian normal distribution is in the Figure 3.1.
Axis x: values of monitored biological character,
Axis y: probability of occurrence of these values in population.

26
 Fig. 3.1 Graph of Gaussian Normal Distribution

The curve is bell-shaped and symmetrical; majority of values is located around the mean
(centre of symmetry) with progressively fewer observations toward the extreme values. In
extremes, the curve is not terminated – the curve theoretically gets near the axis x in infinities (both
+ and - infinity).
The shape of the normal curve is fully described by means of two parameters -  and :
 (Mean value) – “Parameter of location” – it describes the centre of symmetry and also the
location of the curve on axis x.
 (Standard deviation) – “Parameter of dispersion (variability)”. It describes the spread of
the curve in the inflexion point (where the flexure changes from convex to concave).
Spread of the curve determines the variability of biological character monitored in the
population.
The whole area under the curve represents all individuals in the population (100%); then:
Within the range   1: there are 68.3 % of all values (individuals) in the population,
Within the range   2: there are 95.5 % of all values (individuals) in the population,
Within the range   3: there are 99.7 % of all values (individuals) in the population.

The occurrence of remaining values (0.3%) in both extreme ends of axis x is so highly
improbable, that such extreme values are considered as an error of measurement in terms of
statistics.

27
 3.1.2 Standard Normal Distribution

The normal (Gaussian) variable (X) can be standardized, transforming it to Z by subtracting

the mean and dividing by the standard deviation, e.g.,

X 
Z

The normal distribution then becomes the standard normal, which has mean in the value 0
and standard deviation always equal to 1. Units of the new standardized variable Z on the axis x
express the number of standard deviations away from the mean (zero value), positive for above the
mean and negative for below the mean. It means that standardized variable Z represents
a dimensionless quantity, it is a relative measure, divorced from the actual magnitude or units of
measurements of data.
This transformation usually is made in practice as the probability tables available usually
are of the standard normal curve. Table Appendix 1, in the back of the textbook, contains selected
value of z with four areas that often are used: (a) the area under the curve in the positive tail for
given z, i.e., one-tailed α; (b) the area under all except that tail, i.e., 1- α; (c) the areas combined for
both positive and negative tails, i.e., two-tailed α; and (d) the area under all except the two tails,
i.e., 1- α. (See Statistical tables Appendix 1)
The graphical description of the Standard normal distribution is in the Figure 3.2.
Axis x: values of standardized variable Z,
Axis y: probability of values Z.

Fig. 3.2 Graph of Standard Normal Distribution

28
 3.1.3 Non-normal Distribution

Some of variables monitored in biological and medical sciences don’t follow Gaussian
normal distribution – then they usually have variously irregular shape of the probability distribution
curve, often asymmetrical or with 2 and more peaks. Such curves are most often called “non-
normal” or “unknown”; because for their irregularity it is not possible to describe the shape of the
curve in a very exact way.
The example of graphical description of the non-normal distribution is in the Figure 3.3.
Axis x: values of monitored biological character,
Axis y: probability of occurrence of these values in population.

Fig. 3.3 Graph of Non-normal Distribution

The curves of the non-normal probability distribution have the shapes that can be variously
irregular; therefore it is impossible to use some exact parameters that would determine centre and
spread of data (like it was possible in Gaussian normal distribution). Only one descriptive
characteristics, the median, is usually used in such non-normal distributions. The median is
considered as the centre of such irregular curve. Since the median is defined as 50% quantile, it
divides the whole area under the curve into 2 equal halves regardless of the shape of the probability
distribution. It is not possible to determine the spread of the curve (variability of monitored
character) for its irregularity.

3.2 Distributions for Samples

3.2.1 t-distribution (Student’s)

This distribution is defined for description of a theoretical variable t that is calculated from
mean and standard deviation in a sample by means of various formulas used in different situations
in statistics (first of all in statistical hypotheses testing). Common form of this formula for
calculation of variable t is t  X   . This variable is used in testing for differences between 2 means
s
of statistical sets particularly.

29
(t-distribution was published in 1908 by English chemist W. S. Gosset under the pseudonym
“Student”, because the policy of his employer, Guinness Brewery, forbade the publication.)
X 
The t looks like the standard normal curve for variable Z  , however, it is a little

fatter because it uses s instead of accurate . Whereas the normal is a single distribution, t is
a family of curves. In the figure 3.4, two t distributions are superposed on a standard normal
distribution. The particular member of the t family depends on the sample size or, more exactly, on
the degrees of freedom (see below).

The graphical description of the t-distribution is in the Figure 3.4.

Axis x: values of variable t,

Axis y: probability of values t.

Fig. 3.4 Graph of t-Distribution

t-distribution reflects an error of samples (when compared to the population sampled) that is
evident in all statistical calculations performed on the basis of such samples. This error of the
sample is caused by small numbers of members in the sample, and generally we can say that the
smaller is the sample, the more erroneous are calculations performed on the basis of this sample.
The shape of t-distribution curve is similar to the standard Normal distribution (bell-shaped,
symmetry above 0 value), but the spread of the curve is specific for different samples according to

30
the sample size – or more exactly according to the Degrees of Freedom (DF, ) of the sample
monitored:  = n-1.
It is obvious from the graph of t-distribution figured above that:
- The smaller is the sample size, the broader and lower is the curve,
- The larger is the sample size, the narrower and higher is the curve.

In case of the endless expansion of a sample in the extreme: n =  the curve joins the
Normal distribution that describes all population (such sample will have no error in statistical
calculations in comparison with population).
In small samples (that have large error in comparison with the population) the shape of the
curve is also much different from the shape of Normal distribution (used for population).
We can define the exact spread of the curve for t-distribution by ratio: /-2.
Values of t-distribution are tabulated in statistical tables (See the statistical tables -
Appendix 2: Critical values for Student’s t-distribution) and they can be used in statistical
calculations as e.g.:
- Critical values in testing for difference between two means (see Chapter 6: Student t-test),
- Coefficients in calculations of confidence intervals for mean values (see Estimation of
population parameters).
We search the critical values in the tables of t-distribution according to the degree of
freedom calculated for our sample ( = n- 1) and also according to an error  chosen to specify the
exactness of our calculations in statistics - for biological data is commonly used  = 5% or 1%
(when we need more precise calculations).

3.2.2 Chi-square (2 ) Distribution (Pearson’s)

Chi-square distribution is defined for description of a theoretical variable 2 (in practice, we

calculate it from data frequencies in samples) and we use it for calculations in testing for differences
between frequencies in samples - e.g. when we need to compare sickness rate in different groups of
animals (see Chapter 9 Categorical Data).
The graphical description of the chi-square distribution is in the Figure 3.5.
Axis x: values of variable 2,
Axis y: probability of values 2.

31
 Fig. 3.5 Graph of Chi-square (2) distribution.

Chi-square distribution has an asymmetrical curve – right skewed (it rises from 0 rapidly to
a mode and then tails off slowly in a skew to the right) and it has different shapes for different
sample sizes ( = n-1 determines the shape).
It is obvious from the graph of 2 distribution figured above that:
- the smaller the sample size is the higher and more asymmetrical is the curve shape
- the larger the sample size is the lower and more symmetrical is the curve shape
Values of chi-square distribution are tabulated in statistical tables (See tables of the 2
distribution: Appendix 3, 4) and they can be used in statistical calculations e.g.:
- Critical values in testing for difference between frequencies in samples (See 2-test),
- Coefficients used in calculations of confidence intervals for standard deviation (see
Estimation of population parameters).

3.2.3 F-distribution (Fisher-Snedecor’s)

F-distribution is defined for description of a theoretical variable F and we use it in
calculations in statistics e.g. in testing for differences between 2 variances in two groups of data. In
practice, the theoretical variable F is calculated in the so called F-test, when we test the two sample
variances s12 and s22 to determine whether, in the populations being sampled, one is greater in

32
probability. To do this, we use their ratio, dividing the bigger by the smaller. The probability
distribution for this ratio is called F, named (by George Snedecor) after Sir Ronald Fisher, the
greatest producer of practical theories in the field of statistics.

The F-distribution looks very much like the chi-square distribution (indeed it is the ratio of
two independent chi-square-distributed variables), as can be seen in figure 3.6. F distribution has
one more complication than chi-square: because it involves two samples, the degrees of freedom for
each sample must be used. Then the shape of the F-distribution curve is determined by the degree of
freedom of 2 samples (1 and 2) that are used in testing (e.g. when we want to test differences
between variability of monitored biological character in two groups of animals: see Chapter 6
F test for details).
The graphical description of the F-distribution is in the Figure 3.6.
Axis x: values of variable F,
Axis y: probability of values F.

Fig. 3.6 Graph of F-distribution.

As it follows from the figure, the curve is asymmetrical, it rises from 0 value, shortly runs
up and then tails off slowly in a skew to the right toward higher values on the axis x. The shape of
the curve is changing according to sample sizes (more exactly according to 1 and 2).
It is obvious from the graph of F-distribution figured above that:

33
 - The smaller the sample size is the lower and more asymmetrical is the curve shape,
- The larger the sample size is the higher and more symmetrical is the curve shape.
Values of F-distribution are tabulated in statistical tables (See tables of the F-distribution:
Appendix 5) and they are most often used as critical values in testing for differences between 2
sample variances (see F-test).

34
 Chapter 4
Estimation of population parameters
(Confidence Intervals)

Having obtained a random sample from a population of interest, we are ready to use
information from that sample to estimate the characteristics of the underlying population. If you are
willing to assume that the sample was drawn from a normal distribution, summarize data with the
sample mean and sample standard deviation, the best estimates of the population are mean and
population standard deviation, because these two parameters completely define the normal
distribution. When there is evidence that the population under study does not follow a normal
distribution, summarize data with the median as the only descriptive characteristics used for the
non-normal distribution.
Although sample statistics are the best estimates of true population parameters, they are still
only estimates. Therefore, it is appropriate to determine confidence intervals for true population
parameters that allow us to express the precision of the estimates based on the sample data.
A confidence interval is an interval about an estimate, based on its probability distribution, that
expresses the confidence, or probability, that that interval contains the true population parameter
being estimated.

4.1 Normal Distribution – Estimation of  and 

Populations with Gaussian normal distribution are fully described by means of mean value
and standard deviation (SD); true exact parameters can’t be calculated in practice, so we use
imprecise sample statistics as estimations of true parameters. The mean value  of the population is
estimated by means of the sample average x (AVG); and standard deviation  is estimated by
means of the sample s.
Although sample average x is the best estimate of population , and sample standard
deviation s is the best estimate of population , they are still only estimates. Therefore, it is useful
to calculate also the confidence intervals for  and  that allow us to express the precision of the
estimates.

35
 4.1.1 Confidence Interval for the Mean Value 

The calculation of confidence interval for the mean consists of determination of confidence
limits L1 (lower) and L2 (upper). Limits will be symmetrical around the sample average x , true
mean value will lie within the interval restricted by the limits L1, L2.
For determination of the limits L1, L2 we need to know the standard error of the mean
s x (SEM, SE) = a measure of the precision with which a sample average x estimates the true
population mean .
If we try to estimate the true mean value  by means of several sample averages – we will
see that every AVG is slightly different (caused by variability of individuals in samples) – but all of
them will estimate the same true  . The question is: what AVG is the best one? We need some
measure to specify its precision = SEM. This statistic quantifies the certainty with which the mean
computed from a random sample estimates the true mean of the population from which the sample
was drawn.
Calculation formula for standard error of the mean:

s
SEM  s x 
n

SEM is dependent on standard deviation (directly) and sample size (indirectly).

• If the sample size increases -> SEM decreases
(Precision with which we estimate the true mean increases),
• If the sample is more variable (SD increases) -> SEM increases
(Precision with which we estimate the true mean decreases).

SEM is used for calculation of confidence interval:

L1, 2  x  sx . t ,

x - sample mean
s x - standard error of the mean
t, - confidence coefficient = critical value of t-distribution (Appendix 2. Critical values for
Student’s t-distribution) - determined according to the selected error  and DF:  = n-1.

In the course of calculation we can determine a specific precision of the calculation by

selecting the error . For biological data this error  = 0.05 or 0.01 is usually used. When
referring to the selected  in the calculation of confidence interval, we call the quantity 1-
(namely, 1 - 0.05 = 0.95 or 1 – 0.01 = 0.99) the confidence level (95% or 99% confidence level).

36
 Example:
Calculate confidence intervals for the mean of body weights of piglets at the 95% and 99%
confidence level.
Body weights (in kg) of the sample of 25 piglets:
xi : 25.8, 24.6, 26.1, 22.9, 25.1, 27.3, 24.0, 24.5, 23.9, 26.2, 24.3, 24.6, 23.3, 25.5, 28.1, 24.8, 23.5,
26.3, 25.4, 25.5, 23.9, 27.0, 24.8, 22.9, 25.4.

Method:
1) Calculation of the sample mean and SEM:
Mean: x  25.0 kg
SEM: sx  0.27 kg

2) Calculation of the confidence intervals:

At the 95% confidence level: =24 t0.05, 24 = 2.064
x  s x . t (0.05, 24)  25.0  0.27 . 2.064  25.0  0.56 kg

At the 99% confidence level: =24 t0.01, 24 = 2.797

x  s x . t (0.01, 24)  25.0  0.27 . 2.797  25.0  0.76 kg

3) Conclusion:
The true mean value for population of body weights in piglets lies within the confidence
intervals: 25.0  0.56 kg (at the 95% confidence level),
25.0  0.76 kg (at the 95% confidence level).

4.1.2 Confidence Interval for the SD ()

The calculation of confidence interval for the population standard deviation consists of
determination of confidence limits L1 (lower) and L2 (upper).
Confidence limits L1, L2 for determination of confidence interval above estimate of standard
deviation (s) must be calculated separately, since the interval is not symmetrical; that is, the
distance from L1 to s is not the same as the distance from s to L2.

(n  1) s 2 (n  1) s 2
L1  L2 
 21 / 2 ( )  2 / 2 ( )

37
 s2 – sample variance
n – sample size
Confidence coefficients ( =critical values of 2distribution):
Quantile  2/2 – left tail value  upper limit of the interval
Quantile  21-/2 – right tail value  lower limit of the interval
(See Appendix 3 and 4: Critical values for 2 distribution, Right tail, Left tail)

Appropriate quantiles of 2 distribution used in the calculation of limits L1 and L2 for

confidence interval of standard deviation are presented in the Figure 4.1.

Fig. 4.1 Left tail value and right tail value of  2-distribution (quantiles  2/2 ,  21-/2 )

Example:
Calculate confidence intervals for the standard deviation of body temperatures of twenty
five intertidal crabs placed in air at 24.3°C at the 95% confidence level.
Body temperatures (measured in °C): 25.8, 24,6, 26.1, 22.9, 25.1, 27.3, 24.0, 24.5, 23.9,
26.2, 24.3, 24.6, 23.3, 25.5, 28.1, 24.8, 23.5, 26.3, 25.4, 25.5, 23.9, 27.0, 24.8, 22.9, 25.4.
Method:
1) Calculation of sample variance:
s2 = 1.80 (°C)2
2) Calculation of the confidence intervals:
At the 95% confidence level: =24 20.975, 24 = 12.40, 20.025, 24 = 39.36
(See Appendix 3 and 4: Critical values for 2 distribution, Right tail, Left tail)

38
 L1 
25  1.1.80  43.20
 1.10  1.05 C
39.36 39.36

L2 
25  1.1.80  43.20
 3.48  1.87C
12.40 12.40

3) Conclusion:
The true standard deviation for population of body temperatures of intertidal crabs
placed in air at 24.3°C lies within the confidence interval that is restricted by limits: L1 = 1.05°C
and L2 = 1.87°C (calculated at the 95% confidence level).
(Note that the confidence limits are not symmetrical around s; that is, the distance
from L1 to s is not the same as the distance from s to L2).

4.2 Non-normal Distribution – Estimation of the Median

When the population under study does not follow the Gaussian normal distribution, then the
only one descriptive characteristic, the median, can be used for definition of such non-normal
distribution. For irregularity of the distribution curve, it is not possible to determine the spread of
the distribution, i.e. the variability of monitored data set.
The true exact median of the population under study can’t be calculated in practice, so we
use imprecise sample median as estimation of the true parameter. The population median ~ is
estimated by means of the sample median ~ x.
Although sample median ~ x is the best estimate of population ~ , it is still only estimate.
Therefore, it is useful to calculate also the confidence intervals for ~ that allows us to express the
precision of the estimate.

4.2.1 Confidence Interval for the Median

The calculation of confidence interval for the population median ~ consists of
determination of confidence limits L1 (lower) and L2 (upper).
Confidence limits L1, L2 are values derived from the statistical tables (see Table 1 below).
According to the sample size n and selected  we find in statistical tables ranks for L1, L2.
Then we replace these ranks with the actual values from the variant sequence (arranged
order of measured data, ascending or descending).

Example:

Calculate the confidence interval for the true population median of body weights (kg) in the
sample of 14 dogs of a particular breed:

39
 Measured values (in kg): 14.1, 16.4, 16.8, 14.3, 12.3, 14.9, 15.3, 12.8, 15.6, 13.5, 16.0, 16.2,
17.1, 17.0
Method:
1) According to the sample size and selected  we find in statistical tables ranks for L1, L2:
n = 14
 = 0.05

Tab. 4.1 Ranks for Confidence Limits for the Median (Part of the table,  = 0.05)

n Lower Limit Upper Limit

8 1 8
9 2 8
10 2 9
11 2 10
12 3 10
13 3 11
14 3 12
: : :
: : :
: : :
100 40 61

2) We arrange the measured values in an ascending row (variant sequence).

3) We replace found ranks (3 and 12) with the actual values from the variant sequence of
measured values to determine confidence limits L1 and L2 (See Table 2):

Tab. 4.2 Sample data (n = 14) with marked confidence limits for the median
x1 12.3
x2 12.8
x3 13.5
x4 14.1
x5 14.3
x6 14.9
x7 15.3
x8 15.6
x9 16.0
x10 16.2
x11 16.4
x12 16.8
x13 17.0
x14 17.1

3) Conclusion:

Confidence interval for the true median of body weights in dogs of the particular breed
will lie within limits: L1 = 13.5 and L2 = 16.8 (calculated at the 95% confidence level).

40
 Chapter 5
Statistical Hypotheses Testing

5.1 Statistical Hypothesis

Testing of statistical hypotheses is one of the most important parts of statistics in regard to
the practical use of biostatistics (first of all it can be used for evaluation of experimental data) and it
helps us to make conclusions from experiments performed with animals or some individuals in
general. The major goal of statistical analyses is to draw conclusions regarding the whole
population by examining a sample (or more samples) from that population – on the basis of this
sample data we can decide on acceptance of some hypothesis regarding the population that we are
interested in. A very common example of this is the desire to draw conclusions about one or more
population means, as the most important statistical characteristics, or conclusions regarding
variability of two populations, sometimes also conclusions regarding distribution of variables
monitored in the population.
We begin by making a concise statement about the population – a specific hypothesis.
Hypothesis can be any statement about a population characteristic: its distribution or
parameters (mean, SD).
For example: A population matches Gaussian normal distribution
2 populations have the same mean
2 populations have the same variance

This hypothesis is called a null hypothesis (H0) – it expresses the concept of “no
difference”.
For example:
H0 :  = const. (e.g. the population mean is equal to certain value known about the studied
population – e.g. physiological values of some biochemical indices)
1 = 2 (2 populations have the same mean value)
12 = 22 (2 populations have the same variance)
If it is concluded (through a statistical test – see below) that it is likely that a null hypothesis
is false, then an alternate hypothesis (abbreviated HA) is assumed to be true. HA denies H0, so for
examples above:
HA:   const.
1  2
12  22

41
 One states a null hypothesis and an alternate hypothesis for each statistical test performed,
and all possible outcomes are accounted for this pair of hypotheses.
It must be emphasized that statistical hypotheses are to be stated before data are collected to
test them. A statement of hypotheses after examination of data can devalue a statistical test. One
may, however, legitimately formulate hypotheses after inspecting data if a new set of data is then
collected with which to test the hypotheses.

The use in practice: The experimental data evaluation.

E.g.: we need to find out if a vitamin supplement in food causes the increase of body weight
in piglets.
We set up an Experiment:
Group1 of piglets (Test sample) gets the vitamin supplement in food
Group2 of piglets (Control sample) gets standard food
After some period we measure the body weight in both groups of animals and we can find
out e.g. that test sample has a mean x1 which is higher than the mean of the control group: x 2 . We
have to decide (through a statistical test), whether the difference between the sample means is only
random (caused by variability of animals) – or whether it is big enough to conclude that population
means are different as well. It would mean that the difference was caused by our experimental
activity (we can say that this experimental activity is generally effective).
In this case we can reject the null hypothesis (1 = 2) and it means that the alternate
hypothesis is true: (1  2).

Conclusion in practice (for this particular experiment): the statement that “the increase of
body weight is caused by the vitamin supplement” is generally true (the increase of the body
weight is not a random effect).

Probability and significance

To draw conclusions from experimental data we need first to set arbitrary critical thresholds
of probability (P-values). The occurrence of an event whose estimated probability is less than
a critical threshold is regarded as a statistically significant outcome. The usual thresholds of
probabilities (P-values) chosen for biological and medical data are P = 0.05, i.e. significant; P =
0.01, i.e. highly significant. The procedure for deciding if an outcome is significant is called a
statistical test.

5.2 Statistical Tests

The statistical test is used as a decision rule about the acceptance (or rejecting) of the null
hypothesis verified in an experiment. The objective of a statistical test is to obtain (from
experimental data) a single number called a test statistic (calculated variable, whose probability
distribution is known).

42
 Different variables are used as the test statistics in various statistical tests e.g.:
t – Testing for difference between 2 means (t-test),
F – Testing for difference between 2 variances (F-test),
2 – Testing for difference between 2 frequencies (2-test).

The procedure of each of the statistical tests consists in a test statistic calculation – then we
determine if the calculated value of the test statistic exceeds some critical value of the test statistic.
When the calculated test statistic (in absolute value) exceeds the critical value, then the null
hypothesis is rejected. Otherwise, the null hypothesis is accepted. Fig. 5.1 shows an example of
critical value for the test statistic t (t distribution).

Fig. 5.1 Critical value for the test statistic t

t – Test statistic t (for details, see Chapter 6: t-test)

P(t) – Probability of t-values

This critical value is associated with a particular probability threshold (P-value) that is used
as a criterion for rejection of H0 in the test and that is called the significance level, denoted by α. In
fact, the critical value is usually the 1-α/2 quantile of the appropriate distribution used as the test
statistic. As explained above, a probability of 5% or less is commonly used as the criterion for
rejection of H0 in biological and medical data testing. It means that the calculated test statistic (in
absolute value) has to exceed the critical value at the α = 0.05 level of significance to obtain
a statistically significant outcome of the test (denoted as P < 0.05 usually). When the calculated test
statistic exceeds the critical value at the α = 0.01 level of significance, we obtain a statistically
highly significant outcome of the test (denoted as P < 0.01 usually). In the case, when the calculated

43
test statistic does not exceed the critical value at the α = 0.05 level of significance, we obtain
a statistically not significant outcome of the test (denoted as P > 0.05 usually). Following Figure 5.2
demonstrates the aforesaid possible outcomes of the test statistic t.

Fig. 5.2 Test statistic t – possible outcomes of the test

t – Test statistic t (for details, see Chapter 6: t-test)

P(t) – Probability of t-values

Types of Errors in Hypotheses Testing

It is very important to realize that a true null hypothesis occasionally will be rejected, which
of course means that we have committed an error in drawing a conclusion about the sampled
population. Moreover, this error will be committed with a probability of α. That is, if H0 is in fact
a true statement about a statistical population, it will be concluded (erroneously) to be false 5% of
the time. The rejection of a null hypothesis when it is in fact true is what is known as a Type I error
(“Type 1 error”, also called an alpha error or an “error of the first kind”). On the other hand, if H0 is
in fact false, a statistical test will sometimes not detect this fact, and we shall thus reach an
erroneous conclusion by not rejecting H0. The probability of committing this error, of not rejecting
the null hypothesis when it is in fact false, is represented by . This error is referred to as a Type II
error (“Type 2 error”, also called a beta error, or an “error of the second kind”). The power of a
statistical test is defined as 1- ; i.e., power is the probability of rejecting the null hypothesis when
it is in fact false and should be rejected.
Whereas the probability of committing a Type I error is α, the specified significance level,
the probability of committing a Type II error is , a value that generally we neither specify nor

44
know. What we do know is that for a given sample size, n, the value of α is related inversely to the
value of . That is, lower probabilities of committing a Type I error are associated with higher
probabilities of committing a Type II error. Both types of error may be reduced simultaneously by
increasing n. Thus, for a given α, larger samples will result in statistical testing with greater power
(1 - ). Table 5.1 summarizes these two types of statistical errors.

Tab. 5.1 The Two Types of errors in Hypotheses Testing

DECISION
REJECTING NOT REJECTING
H0 H0
REALITY

H0 IS TRUE Type I error NO ERROR

 1- 

H0 IS FALSE NO ERROR Type II error

1-  (power of test) 

Since, for a given n, one cannot minimize both of types of errors, it is appropriate to ask
what the acceptance combination of the two might be. In terms of veterinary medicine: “Not to treat
an ill animal (statistically evaluated as a healthy one – Type I error) is a more serious mistake than
to treat a healthy animal statistically evaluated as the ill one (Type II error)”. Therefore, statistical
tests used in medicine are set up to achieve a minimal Type I. error α. By experience, and hence by
convention, an α of 0.05 is usually considered to be a “small enough” chance of committing a Type
I error, while not being so small as to result in “too large a chance” of a Type II error. But there is
nothing sacred about the 0.05 level. Although it is the most widely used significance level,
researchers may decide for themselves whether it is more important to minimize one type of error or
the other. In some situations, for example, a 5% chance of an incorrect rejection of H0 may be felt to
be unacceptably high, so the 1% level of significance is sometimes employed.
It is necessary, of course, to state the significance level used when reporting the results of
a statistical test. Indeed, rather than simply stating whether the null hypothesis is rejected, it is good
practice to state also the test statistic itself and the best estimate of its exact probability (calculated
by means of a statistical software). In this way, readers of the research results may draw their own
conclusions, even if their choice of significance level is different from author’s.
Bear in mind, however, that the choice of α is to be made before even seeing the data.
Otherwise there is a great risk of having the choice influenced by examination of the data,
introducing bias instead of objectivity into proceedings. The best practice generally is to decide on
the null hypothesis, alternate hypothesis, and significance level before commencing with data
collection.
As we already know, it is conventional to refer to rejection of H0 at the 5% significance
level as denoting a “significant” difference (e.g. between compared population means) and rejection
at the 1% level as indicating a “highly significant difference”. As the significance level selected is

45
somewhat arbitrary, if test results are very near that level (e.g. between 0.04 and 0.06 if α = 0.05 is
used), then it may be wiser to repeat the analysis with additional data than to declare emphatically
that the null hypothesis is or is not a reasonable statement about the sampled population.

5.3 Classifications of Statistical Tests for Different Types of Data

It is important to choose the appropriate statistical test for a specific type of data. This is not
always straightforward and sometimes, more than one test can be used indeed. We have noted in the
chapter 1 that there are three types of data: categorical, rank-order, and numerical. The categorical
and rank-order data are discrete by their nature; numerical data may be continuous or discrete. Each
type of data requires its own form of statistical testing.

A) Categorical Form of Testing

To compare two variables using categorical data, we compare counts (frequencies) in two
samples – e.g. number of ill animals in a stable, number of vaccinated dogs, number of dead born
piglets etc. We form two-way tables of counts with one variable representing rows and the other
variable representing columns. We test the proposition that knowledge of the counts in one
variable’s categories tells us something about the counts in the other variable’s categories, i.e. that
the two variables are not independent. Analyses of such dependences in categorical data, as well as
analyses of differences between counts in categorical data are performed by means of Chi-square
tests (See Chapter 9: Categorical data, Contingency tables).

B) Rank-Order Form of Testing

To compare two groups that are in rank order, we attach ranks to the data combined over the
two groups and then add the rank values for each group separately, forming rank sums. If the group
rankings are not much different, the ranking from the two groups will be interleaved and the rank
sums will not be much different. If one group has most of its members preceding the other in rank,
one rank sum will be larger and the other small. Probabilities of rank sums have been tabulated, so
that the associated P-value can be looked up in the table and the decision about the group difference
made.
Rank-order form of statistical testing is represented by Non-parametric tests that are used
for testing of hypotheses with rank-order data, discrete numerical data and numerical continuous
data those are not assumed to come from a normally distributed population. (See Chapter 7: Non-
parametric tests).
Non-parametric Tests – a summary:
- They are used for data sets following the non-normal distribution especially,
- Hypotheses concerning common characteristics of statistical sets are tested in the non-
parametric tests (e.g. two sets have the same shape of distribution),
- Calculations in these tests are based on ranks of measured values.
.

46
 C) Continuous Form of Testing
Whether a difference between means exists most often is the focus in comparing two groups
with data in continuous form. Our first inclination is to look at the difference between means.
However, this difference depends on the scale. The offset distance of a broken femur appears larger
if measured in centimetres than in inches. The distance must be standardized into units of data
variability. We divide the distance between means by a measure of variability and achieve a statistic
t (if the population variability is estimated by small samples: see t-test). The risk of concluding
a difference when there is none (the P-value) is looked up in a table and the decision about the
group difference is made.
Continuous form of statistical testing is represented by Parametric tests that may be used for
testing of hypotheses with numerical data those are assumed to come from a normally distributed
population (See Chapter 6: Parametric tests).
Parametric Tests – a summary:
 They are used in testing for differences between data sets that follow Gaussian normal
distribution,
 Hypotheses concerning parameters (µ,) of this distribution are tested,
 Calculations in these tests are based on sample statistics ( x , s).

47
 Chapter 6
Parametric Tests

Among the most commonly employed biostatistical procedures is the comparison of two
samples to infer whether differences exist between the two populations sampled. In parametric tests,
we consider hypotheses concerning population parameters µ (mean value) and 2 (variance) of
Gaussian normal distribution.
As the mean is the most important characteristic of a population, the basic question asked in
parametric test most often is whether two samples have the same mean or whether a sample mean is
the same as a population mean. Questions concerning two variances (or standard deviations) are
also considered in parametric tests in some instances. The question is answered by testing the null
hypothesis that the means (or variances) are equal and then accepting or rejecting this hypothesis.
Tests of means and variances were developed under the assumption that the sample was
drawn from a normal distribution. Whereas usually not truly normal, a distribution that is roughly
normal in shape is adequate for a valid test. That is because the test is moderately robust.
Robustness is an important concept. A robust test is one that is affected little by deviations from
underlying assumptions. If a small-to-moderate sample is too far from normal in shape, the
calculation of error probabilities, based on the assumed distribution, will lead to erroneous
decisions; then non-parametric (rank-order) methods should be used preferably (see chapter Non-
parametric tests). In particular, tests of means are only moderately robust and they are especially
sensitive to outliers (extremely high or low values), whereas tests of variance are much more robust.
Student’s t-test (used in testing for difference between two means) and Snedecor’s F-test
(used in testing for difference between two variances) belong to the group of parametric tests.

6.1 F-test (Variance ratio Test)

We can decide by means of this F-test whether some treatment (activity used in an
experiment) influences the variability (variance - 2) of some biological character studied in
2 2
a population. A null hypothesis H0: 1 = 2 is verified by examining sample variances – s12 and
s2 2 .

We select 2 samples from the population monitored:

Sample 1 (n1 individuals)

Sample 2 (n2 individuals)

48
 We apply a tested treatment (e.g. a new medical preparation) to one of these samples, the
second sample (without any treatment) will serve as a control group for comparison.

Method:
2 2
1) We calculate sample variances s1 and s2 :

 x  2
 x  2

x x
i i
i
2
 i
2

n1 n2
s 2  s 2
1 n1  1 2 n2  1

2 2
higher variance (out of s1 , s2 )
2) We calculate a test statistic: F 2 2
lower variance (out of s1 , s2 )
3) We specify degree of freedom of Numerator (of the F ratio) and Denominator (of the F
ratio), appropriate to the size of sample1 and sample2:
2
DF Numerator: N = n1(2)- 1 (for the higher out of s1 , s22)
2
DFDenominator: D = n1(2) - 1 (for the lower out of s1 , s22)

4) We find out critical value F(α, N, D) in statistical tables for F-distribution (Appendix 5)
according to the chosen error  (0.05), and degree of freedom (DF) for numerator and denominator
of our ratio for the test statistic F.

5) We will compare the calculated test statistic F with table critical value Fcrit. to make
a conclusion about population variances and about an effect of the treatment used in the experiment
on variability of studied biological character:
2 2
If calculated F > F(crit.)  We reject H0, then alternate hypothesis is true: HA: 1  2 .
i.e. There is a significant difference between variances – it means
that variability of 2 populations sampled is not equal (at the 
level).
Conclusion: the treatment used in experiment has influenced the
variability of the studied biological character.
2 2
If calculated F  F(crit.)  H0 is true: 1 = 2 .
i.e. There is an insignificant difference between variances – it
means that variability of 2 populations sampled is equal (at the 
level).
Conclusion: the treatment used in experiment has not influenced the
variability of the studied biological character.

49
 Example:
The effect of a new veterinary preparation on AST (aspartate aminotransferase) level in
blood serum in dairy cows has been monitored. In 10 dairy cows (control group), to which the
preparation was not applied, the following AST activities in blood serum have been found (in
µmol.l-1 ):
0.409, 0.345, 0.392, 0.377, 0.398, 0.381, 0.400, 0.405, 0.302, 0.337
In 10 dairy cows (test group), to which the preparation was applied. the following AST
activities in blood serum have been found (µmoll-1):
0.341, 0.302, 0.504, 0.452, 0.309, 0.375, 0.479, 0.423, 0.311, 0.333
Does the preparation influence the variance of AST activity in blood serum of dairy cows?

Method:
2 2
1) We calculate sample variances s1 and s2 :
Control: s12 = 0.00125
Preparation: s22 = 0.00575

2) We calculate a test statistic F:

2
s 0.00575
F  22   4.618
s1 0.00125

3) We specify degree of freedom of the numerator and denominator of the ratio:

2
vN = n2- 1 = 9 ( for s2 )
2
vD = n1- 1 = 9 ( for s1 )

4) Critical value Fcrit (0.05, 9, 9) = 4.026

5) F > F(crit.).  statistically significant difference was found between variances (An
alternate hypothesis HA: 12  22 is true) at the level α = 5%.

Conclusion: As difference between variances of control and test groups is statistically

significant (P < 0.05), the preparation tested influences the variance of the activity of AST in blood
serum in dairy cows.

50
 6.2 t-test (Student’s)

Student’s t-test is used for testing for difference between 2 population means  (in general).
However, there are several different variants of t-test in practice according to the data sets that are
available for comparison (see below). The t-test is the most common statistical procedure in the
medical and biological literature; you can expect it to appear in more than half the papers you read
in the general medical literature. In addition to being used to compare two group of means, it is
widely applied incorrectly to compare multiple groups, by doing all the pairwise comparisons, for
example, by comparing more than one intervention (treatment) with a control condition.
Student’s t-test is especially useful for testing for significant differences between results
obtained under two experimental conditions (treatments). First, we hypothesise that the means of
our two populations are not different (null hypothesis, e.g. H0: 1 = 2). We then determine the
probability (our P-value) that the difference in our samples´ means could have arisen by chance.
This P-value is thus a measure of the compatibility between our experimental observations and our
null hypothesis. A low P-value), say <0.05, is typically regarded as statistical evidence to reject the
null hypothesis and conclude that there is significant difference in the result obtained from the two
experimental conditions (treatments).
The null hypothesis states that the mean of the population from which the sample is drawn is
not different from a theoretical mean or from the population mean of another sample drawn from
the same population. We also must choose the alternate hypothesis, which will select a two-tailed
or one-tailed test. We should decide this before seeing the data so that our choice will not be
influenced by the outcome. The two-tailed t-test is used to test against the alternative hypothesis
that 1  2. It is sometimes the case that before the data are collected there is only one reasonable
way in which the 2 means could differ, and the alternative hypothesis would be for example 1 > 2
(or 1 < 2). It is the appropriate to carry out a one-tailed t-test. We often expect the result to lie
toward one tail, but expectation is not enough. If we are sure the other tail is impossible, such as for
physical or psychological reasons, we unquestionably use a one-tailed test. Surgery to sever
adhesions and return motion to joint frozen by long casting will allow only a positive increase in
angle of motion; a negative angle physically is not possible. An one-tailed test is appropriate.
There are cases in which an outcome in either tail is possible, but a one-tailed test is
appropriate. When making a decision about a medical treatment, i.e., whether we will alter
treatment depending on the outcome of the test, the possibility requirement applies to the alteration
in treatment, not the physical outcome. If we will alter treatment only for significance in the
positive tail and it will in no way be altered for significance in the negative tail, a one-tailed test is
appropriate. However, very often we also need to test an alternate hypothesis that there is any
difference (whichever difference: toward positive or negative tail) between treatments used in
experiment – in these cases the two-tailed test is appropriate.
The difference between one-tailed and two-tailed tests is also reflected in the size of the
critical value; generally we can say that critical values used in one-tailed tests are lower than critical
values used in two-tailed tests (at the same α level of significance). In the table of critical values of
t-distribution (Appendix 2), we can see that one-tailed critical values (marked as α(1)) at the
specific α level are the same as two-tailed critical values (marked as α(2)) at the double α level of
significance for given .

51
 The procedure of Student’s t-test consists in the calculation of a test statistic t that results
from the estimation of parameters  and  in samples: x and s. Calculated test statistic is compared
with the tabulated critical value tα, that we can find out in tables of t-distribution (Appendix 2)
according to the chosen error  (our probability level for acceptance of significant difference it is
typically set at 0.05 by most researchers in the biological and medical sciences) and  (DF - degree
of freedom calculated by means of a specific formula for each of the variants of t-test).
If calculated t > tα, => we reject the null hypothesis, it means that there is a significant
difference between means of populations sampled at the α level of significance. We accept the
alternate hypothesis that our two experimental groups (typically one sample with the treatment and
the second one – control without treatment) produced statistically significant results (i.e. samples
was not drawn from the same population).
If calculated t  tα, => we accept the null hypothesis, it means that there is an insignificant
difference between means of populations sampled at the α level of significance, i.e. our two
experimental groups (typically one sample with the treatment and the second one – control without
treatment) produced insignificant results (i.e. samples was drawn from the same population).

6.2.1 Population vs. Sample Comparison (One-sample t-test)

This variant of t-test is used for evaluation of data in experiments, where a population
parameter  is known. It may be e.g. physiological value of a biochemical indicator – this value is
considered as a constant. Then in the experiment, we verify a null hypothesis whether the test
sample (under a treatment) comes from a population with this known parameter  (H0:  = const.).
An alternate hypothesis is HA:   const.

Method:
1) We calculate sample mean and variance

2) We calculate a test statistic:

x μ
t
s2
n
x -sample mean, - population mean, s – sample SD, n–number of items in sample

3) We specify degree of freedom for the test:  = n-1

4) We compare calculated t with the tabulated critical value t(,) , where  = n-1 and  = 0.05
(or 0.01).

 If t  t(,)  we reject H0:  = const. There is a statistically significant difference

between tested means at the  = 0.05 level (P<0.05)
or highly significant difference at the  = 0.01 level (P<0.01).

52
 It means that the treatment has been effective – it caused a change of
the mean in comparison with the known population mean = const., i.e.
the tested sample comes from another population with  const.

 If t  t(,)  we accept H0:  = const. (i.e. H0 is true). There is a statistically

insignificant difference between tested means at the specific  level
(P>0.05).
It means that “the treatment has been ineffective” – it did not cause
a change of the mean in comparison with the known population mean
= const., i.e. the tested sample comes from the another population with
 const.

Example:
In a population of dairy cows the mean value of glucose in blood serum is  = 3.1 mmol.l-1.
After applying an energy preparation glucose level in serum in 10 cows selected at random was
measured:
3.1, 2.7, 3.3, 3.1, 3.1, 3.2, 3.0, 2.8, 2.9, 2.7.
Does the preparation influence the glucose level in serum?

Method:

1) We calculate sample statistics:

x  3.0  9
s  0.21
s 2  0.044
2) Mean value known for the whole population:  = 3.1 mmoll-1.

3) We calculate test statistic t:

x μ 3.0  3.1
t   1.58
s2 0.212
n 10

4) Critical value found in statistical tables of the t-distribution: tcrit.(0.05; 9) = 2.262

5) We compare the calculated test statistic with the critical value:

t < tcrit.  there is a statistically insignificant difference between means (P>0.05).

(H0 is not rejected; the sample comes from the population with =3.1).

6) Conclusion:

Preparation used in experiment is not effective (to change the glucose level in dairy cows).

53
 6.2.2 Samples comparison (Two-sample t-test)
This variant of t-test is used for evaluation of data in experiments, where a population
parameter  is not known. We compare data of 2 samples that comes either from one group of
subjects measured twice (typically before and after treatment – “paired experiment”, “dependent
samples”) or from two different random groups of subjects (typically treated test group and
untreated control group) – “unpaired experiment”.
A) Paired t-test (paired experiment, dependent samples)
Data evaluated in this variant of t-test come from paired subjects; it means that the same
subjects are submitted to two measurements (both test and control treatments are performed in one
group of subjects). Such a situation represents for instance the weight before and after 2 weeks
since the beginning of a new diuretic medication. In that case the outcome we are interested in, i.e.
the weight, is evaluated before and after the medication of the diuretic using the same individual.
Therefore we get “matched” data from these repeated measurements.
Note that paired experiments are, in principle, more robust than unpaired experiments. For
example, if each individual taking a diuretic loses 2 kg in 2 weeks then you can feel comfortable
that this represents an effect of the diuretic. In contrast, in two random groups, if one receives
placebo (control) and the other the diuretic, a 2 kg weight reduction in the diuretic-treated group is
less strong evidence of effectiveness of the medication. This is because the variation of weight in
each group is larger than 2 kg, so it is unclear whether the difference between the two groups is
a real effect of the diuretic or simply a random small difference in mean weight between the two
groups. It follows that, when possible, it is better to evaluate a given manipulation in the same
subject.
The first step in the procedure of paired t-test consists in calculation of differences between
paired (matched) values: xi = xtest – xcontrol. Then we calculate the sample mean x and SD
(standard deviation) of the differences xi. We test a hypothesis that population mean  of the
measurements before and after the treatment are equal (i.e. mean value  of the differences xi
between matched measurements is equal to 0). Then the null hypothesis is H0: differ.=0 and an
alternate hypothesis is HA: differ.0).
Method:
1) We calculate differences between paired values, mean and standard deviation of the
differences.
2) We calculate test statistic for paired t-test:
x
t
s2
n
x - mean of differences between paired values, s – variance of differences, n – number of
2

pairs
3) We specify degree of freedom for the test:  = n-1

54
 4) We compare the calculated t with the tabulated critical value t(,), where  = n-1 and  =
0.05 (or 0.01):
 If t  t(,)  H0 is rejected, i.e. difference between means is statistically
significant (at  = 0.05)
or highly significant (at  = 0.01)
It means that the treatment has been effective: mean  after the
treatment is different from mean  before the treatment.

 If t  t(,)  H0: differ.= 0 is true, i.e. difference between means of values before
and after the treatment is statistically insignificant (at the specific 
level).
It means that the treatment has not been effective: mean  after the
treatment is the same as the mean  before the treatment.

Example:
Determine a weight change of twelve rats after being subjected to a regimen of forced
exercise. Each weight change (in g) is the weight after exercise minus the weight before:
0.2, -0.5, -1.3, -1.6, -0.7, 0.4, -0.1, 0.0, -0.6, -1.1, -1.2, -0.8. Does the exercise cause any significant
change in rat weight?

Method:
1) We calculate sample statistics:
x  0.6 g   11
s  0.40 g
2 2

2) We calculate test statistic for the paired t test:

x  0.6
t   3.39
s2 0.40
n 12

3) Critical values found in statistical tables of the t-distribution: tcrit.(0.05; 11) = 2.201
tcrit.(0.01; 11) = 3.106
4) We compare the calculated test statistic with critical value:
t > tcrit.(0.01)  H0 is rejected, there is a statistically highly significant difference
between mean before and after exercise (P<0.01).

5) Conclusion:
The exercise causes a highly significant weight loss in rats.

55
 B) Unpaired t-test (unpaired experiment, independent samples)

Data evaluated in this variant of t-test come from two independent groups if individuals - we
deal with “unmatched” data. Typically there is one test group, to which we apply some tested
treatment, and one control group without any treatment. We can also compare two groups with
different treatments in experiment, when we are interested in the possibility, whether there is any
difference between effects of these two treatments.
Unpaired t-test is then used to determine whether the means of two independent samples are
different enough to conclude that they were drawn from different populations. We test the null
hypothesis, whether a population mean value 1 in the test group (treated) is the same as the
population mean value 2 in the control group: H0: 1 = 2. A two-tailed alternate hypothesis is
HA: 1  2.
The populations sampled can have different variability – this variability affects the
calculation of t-test. Therefore at first we have to determine the difference between variances
(through F-test) to specify what type of calculation formula we need to use for the following t-test.
Therefore the first step of the procedure of unpaired t-test consists in calculation of sample statistics
(estimated mean, standard deviation and variance for both samples compared):
Sample 1 (n1): we calculate x 1, s21
Sample 2 (n2): we calculate x 2, s22
In the following step, we determine by means of F-test whether there is any difference
between population variances:
F-test:
2 2
higher ( s1 , s2 )
F 2 2
lower ( s1 , s2 )

Degree of freedom: for numerator N = n –1

for denominator D = n –1

We compare the calculated F with the tabulated critical value of F-distribution that we find
out according to the chosen  and degrees of freedom: N (DF of numerator) and D (DF of
denominator).

According to the F-test result:

 If F  F(N, D)  populations compared have the same variability (12 = 22), we use
the following formula for the unpaired t-test:
a) 12 = 22:

x1  x2
Test statistic: t DF:  = n1+ n2 -2
n1  1  s12  n2  1  s2 2 n1  n2
n1  n2  2 n1  n2

56
 In a special case of equal sizes of samples compared:

x1 x2
For n1 = n2 = n: t DF:  = (n-1) .2
2 2
s1 s2
n

 If F > F(N, D)  populations compared have different variances (12  22),
we use the following formula for the unpaired t-test:

b) 12  22 :
x1  x 2
Test statistic: t
s1 2 s2 2

n1 n2

2
 s1 2 s2 2 
  
 n1 n2 
DF:  2 2 (For n1, n2 > 30:  =  )
 s1 2   s2 2 
   
 n1   n2 

n1  1 n2  1

x1 x2
(For n1 = n2 = n : t )
2 2
s1 s2
n

Conclusion:

 If t  t(,)  statistically significant difference between 1 and 2 (at  = 0.05) or

highly significant difference (at  = 0.01)
i.e. samples was not drawn from the same population (it means that
the “experiment has been effective” and caused a change of mean
value in the treated group compared to the control: 12 ).
Therefore H0: 1= 2 is rejected.

 If t  t(,)  statistically insignificant difference between 1 and 2 at the specific .

H0: 1= 2 is true; i.e. samples was drawn from the same population.
(it means, that “the treatment has not been effective”)

57
 Example:
Determine a drug effect on the change in blood-clotting times (in minutes) in pigs.
Times of individuals treated with drug (T): 9.9, 9.0, 11.1, 9.6, 8.7, 10.4, 9.5.
Times of untreated control individuals (C): 8.8, 8.4, 7.9, 8.7, 9.1, 9.6, 8.7.

Method:
1) We calculate statistics of both samples:

x1  9.7 min n1  7
Test sample:
s1  0.67 min
2 2
1  6

Control sample: x2  8.7 min n2  7

s 2  0.28 min 2
2
2  6
0.67
2) We calculate test statistic for the F test: F   2.37
0.28

4) Critical value for the F test: Fcrit. (0.05;6,6) = 5.820

F<Fcrit.  12 = 22

5) We calculate test statistic for the unpaired t-test for equal variances :

x1  x 2 9.7  8.7
t   3.83
s1  s 2 0.26
2 2

6) Degree of freedom for the t-test:  = (n – 1).2 = 12

7) Critical values for the t test: tcrit. (0.05;12) = 2.179
tcrit. (0.01;12) = 3.055

8) We compare the calculated test statistic with critical value:

t > tcrit. (0.01)  there is a statistically highly significant difference between tested means
(P<0.01).
(H0 is rejected).

9) Conclusion:
Drug administration causes a highly significant longer blood-clotting time in pigs.

58
 Chapter 7
Non-Parametric Tests

Non-parametric tests belong to special statistical methods that comprise procedures not
requiring the estimation of the population parameters (mean and variance) and not stating
hypotheses about parameters. As these methods also typically do not make assumptions about the
nature of the distribution (e.g., normality) of the sampled populations (although they might assume
that the sampled populations have the same dispersion or shape), they are sometimes referred to as
distribution-free tests. Non-parametric tests are often called “rank tests“ as their calculations are
based on sum of ranks of values measured in experiment.
Non-parametric tests may be applied in any situation where we would be justified in
employing a parametric test, such as two-sample t test, as well as in instances when the assumptions
of the latter are untenable. However, if either the parametric or non-parametric approach is
applicable, then the latter will always be less powerful than the former (difference between tested
data sets must be considerable to achieve a statistical significance).
Non-parametric tests are especially employed when dealing with ordinal scale data (data that
consist of ranks) and numerical scale data when normality is not assumed, but they may also be
employed when dealing with numerical data that follow normal distribution (for preliminary
analyses especially, as calculations of non-parametric tests are often more quick and simpler then
parametric ones). Non-parametric tests may also be useful in instances when dealing with small
sample sizes – then sample frequency distribution is insufficient to tell us whether the assumption
of normality may be confirmed.
In non-parametric tests, the hypothesis is verified that data from both samples were drawn
from the same population, i.e. that they have the same dispersion or shape of the distribution curve
(null hypothesis). We then determine (similarly to parametric tests) the probability (P-value) that
the observed difference in our samples could have arisen by chance. A low P-value (P<0.05), is
often regarded as statistical evidence to reject the null hypothesis and conclude that there is
significant difference in the results obtained from our two experimental conditions.

7.1 Mann-Whitney U-Test (Rank-Sum Test)

(Two-Sample Rank Testing)

For this test, as for many other non-parametric procedures, the actual measurements are not
employed, but we use instead the ranks of the measurements.. The data of both samples compared
are arranged into one (mixed) sample and may be ranked either from the highest to lowest or from

59
the lowest to the highest values. The samples compared can consist of equal or unequal number of
the observations. E.g. if data in samples are arranged from the highest to the lowest, then the highest
value in either of the two samples compared is given rank 1, the second highest value is assigned
rank 2, and so on, with the lowest value being assigned rank N, where
N = n1 + n2.

We calculate the Mann-Whitney test statistics:

n1 * (n1  1)
U  n1 * n2   R1 ,
2

n2 * (n2  1)
U ´ n1 * n2   R2 ,
2
where n1 and n2 are the number of observations in samples,
R1 is the sum of the ranks of the observations in Sample1,
R2 is the sum of the ranks of the observations in Sample2.

When U is already calculated U´ can be also found more quickly as

U´= n1* n2 – U

The larger of the two calculated test statistics U and U´ is compared to the critical value Uα,
n1, n2, found in Appendix 6 (Critical Values for Mann-Whitney U-test). This table assumes that
n1 > n2; if n2 > n1, simply use Uα, n2, n1 as the critical value.
Then:
If the larger from U and U´ > Uα, n1, n2 => we reject H0 at the α level of significance (i.e.
samples tested were not drawn from the same population, there is a significant difference between
populations sampled - they don’t have the same shape of the distribution curve). It means that “the
treatment used in the experiment was effective“ at the α.level of significance.
If the larger from U and U´ < Uα, n1, n2 => we accept H0 at the α level of significance (i.e.
samples tested were drawn from the same population, there is an insignificant difference between
populations sampled - they have the same shape of the distribution curve). It means that “the
treatment used in the experiment was not effective“ at the α.level of significance.

Note that neither parameters nor parameter estimates are employed in the statistical
hypotheses or in the calculations of test statistics U or U´.
We may assign ranks either from large to small data, or from small to large, calling the
smallest datum rank 1, the next smallest rank 2, and so on. The value of U obtained using one
ranking procedure will be the same as the value of U´ using the other procedure.

60
 Example:
By means of Mann-Whitney test for non-parametric testing find out whether there is some
difference between the heights of male and female students.

Method:
1) We arrange data from the highest to the lowest and find out ranks of male and female
heights in this arranged (mixed) sample:
193 >188 > 185 > 183 > 180 > 178 > 175 >173 > 170 > 168 > 165 > 163

Heights of males Heights of females Ranks of Ranks of

(cm) (cm) male heights Female heights
193 175 1 7
188 173 2 8
185 168 3 10
183 165 4 11
180 163 5 12
178 6
170 9
n1 = 7 n2 = 5 R1 = 30 R2 = 48

2) We calculate test statistics:

n1 * (n1  1) 7 *8
U  n1 * n2   R1  7 * 5   30  33
2 2

U´ = n1*n2 – U = 7*5 - 33 = 2

3) Critical value for the U-test: Uα, 7,5 = 30

4) We compare the calculated test statistic with critical value:

U > U0.05, 7,5 => we reject the null hypothesis, i.e. there is statistically significant
difference between data sets (P < 0.05).

5) Conclusion:
There is a statistically significant difference between male and female heights at the
significance level α = 0.05.

61
 7.2 Wilcoxon Signed-Rank Test
(A non-parametric test for paired samples)

The Wilcoxon paired-sample test is a non-parametric analogue to the paired-sample t test,

just as the Mann-Whitney test is a non-parametric procedure analogous to the two-sample t test.
Whenever the paired-sample t test is applicable, the Wilcoxon paired-sample test is also applicable.
But there are instances when the Wilcoxon paired-sample test is applicable and the parametric
paired-sample t test is not, as when one can not assume that data are from a normal distribution.
Since the two groups of measurements to be compared are obtained from the same sample of
subjects tested twice (before and after treatment), instead of analysing the raw data from each group
separately we look only at the differences between the pre-treatment and post-treatment values for
each subject. By subtracting the first value from the second for each subject and then analysing only
these paired differences we exclude all the variation in our data that results from the differing initial
values of individual subjects.
The testing procedure involves the calculation of differences, as does the paired-sample
t test. Then we rank the absolute values of the differences, from low to high, and affixes the sign of
each difference to the corresponding rank. If there are some equal differences (in absolute value),
then they will get so called “average rank“ (e.g. if the first and second difference has the same
value, then they both will get the 1.5). If there is some difference that is equal to 0 (di = 0), then the
ith pair is omitted from the analysis.
Then we sum the ranks with a plus sign (we shall call this sum W+) and the ranks with
a minus sign (calling this sum W-). Having calculated either W+ or W-, the other can be determined
also as:
n * ( n  1)
W   W
2
or
n * ( n  1)
W   W
2

The smaller of the two calculated W+ and W- is compared to the critical value Wα, n from
the Tables of critical values for Wilcoxon signed rank test (Appendix 7):
If the smaller from W+ and W- < Wα, n => we reject H0, i.e. difference between
measurements before and after treatment is statistically significant at the α.level („The treatment
used in the experiment was effective“).
If the smaller from W+ and W- > Wα, n => we accept H0, i.e. difference between
measurements before and after treatment is statistically insignificant at the α.level („The treatment
used in the experiment was not effective“).

62
 Example:

By means of Wilcoxon paired-sample test for non-parametric testing find out whether there
is some difference between the lengths of hind- and forelegs in deer.

Method:
1) We calculate differences di between paired values:

Hind leg length Foreleg length Difference

Deer
(cm) (cm) (di)
1 142 138 4
2 140 136 4
3 144 147 -3
4 144 139 5
5 142 143 -1
6 146 141 5
7 149 143 6
8 150 145 5
9 142 136 6
10 148 146 2

2) We arrange the absolute values of differences into an ascending row:

|-1| < 2 < |-3| < 4 = 4 < 5 = 5 =5 < 6 = 6
Note that there are some equal values of differences di.

3) We determine the ranks of differences and apply appropriate sign to the rank according to
the difference. Note that there are several “average ranks“ used for equal differences (E.g. there are
three differences di = 5 in the ascending row above, therefore all of them will get the rank 7 instead
of original ranks 6, 7, 8):

Difference
Deer Ranks of |di| Signed ranks of |di|
(di)
1 4 4.5 4.5
2 4 4.5 4.5
3 -3 3 -3
4 5 7 7
5 -1 1 -1
6 5 7 7
7 6 9.5 9.5
8 5 7 7
9 6 9.5 9.5
10 2 2 2

63
4) We calculate sums of plus and minus ranks :
W+ = 4.5 + 4.5 + 7 + 7 +9.5 +7 +9.5 + 2 = 51
W- = 3 + 1 = 4

5) Critical value W0.05, 10 = 8

6) Since W- < W0.05, 10 => we reject H0 at the 5% level of significance.

7) Conclusion:

Deer hind leg lengths are not the same as the foreleg lengths (P < 0.05).

64
 Chapter 8
Relationship Between 2 Data Sets
(Quantitative Data)

2-dimensional statistics is used for evaluation of relation between 2 variables (biological

characters) in a data set. Two variables are related if their values correspond to each other in some
systematic way. For example, tall people are usually heavier than short people; therefore height and
weight are related variables. Two-dimensional statistics helps us to establish the nature of the
relations between variables. In particular we wish to know how strong is the relationship we have
observed in biological data and how reliable is our observation of a relationship.
We try to qualify and describe the relationship between 2 variables monitored: one being an
independent and one being a dependent variable. The dependent variable can be predicted by the
independent variable (if we know values of the independent variable, we can calculate the values of
the dependent variable).

8.1 Functional vs. Statistical Relationship

We can distinguish between 2 basic groups of relations between variables in general:

A) Functional Relationship (typical for relations in mathematics and physics):
The magnitude of one of the variables (the dependent variable) is assumed to be
determined by (i.e. is a function of) the magnitude of the second variable independent variable,
whereas the reverse is not true. Each value of the independent variable (xi) corresponds to one
exact value of the dependent variable ( yi) in this type of relation. Sometimes, the independent
variable is called the “predictor”, or “regressor”, variable and the dependent variable is called
the “response”, or “criterion”, variable.
Such a relation can be described by means of an exact equation (formula): e.g., relation
between circle radius (r) and its circumference (y=2r) or surface (y=r2).
The functional relation is an example of the strictly causal relationship - it is not affected
by random. Such a dependent relationship is termed a regression; the term simple regression
refers to the fact that only two variables are being considered.
It is very convenient to use a graph in order to describe this functional relation, using the
ordinate (Y axis) for the dependent variable (conventionally termed Y) and the abscissa (X axis)
for the independent variable (X). An example for the chart of functional relation is shown in the
Figure 8.1; it is a linear function in the case of relation between circle radius and its
circumference.

65
 Fig. 8.1 Linear function – relation between circle radius (X) and its circumference (Y)

B. Statistical Relationship (Correlative)

This relationship is typical for relations between data in biology and medicine: there is no
exact functional relationship, because most of biological characters are very changeful and unstable;
therefore relations result from this variability and they are relative (statistical, correlative) only.
These relations in biology and medicine are very complicated – there are many different causes
including random effects that we are not able to exclude during our monitoring.
This relation is more or less free - the magnitude of one of the variables probably changes
as the magnitude of the second variable changes. Each value of xi corresponds to several random
values of yi and also the reverse is possible (“variables are correlated”). In such a case it is not very
often reasonable to consider that there is an independent and dependent variable (e.g. fore- and hind
leg lengths in animals, human height and weight, arm and leg lengths, etc.). It might be found that
an individual with long arms will in general possess long legs, so a relationship may be describable;
but there is no justification in stating that the length of the limb is dependent upon the length of the
other. In such situations, correlation, rather than regression, analyses are called for, and both
variables are theoretically to be random-effects factors.
We use so called correlation chart (“scatter diagram”, “dot plot”) for the graphical
description of such statistical relationship – each point represents a pair of X and Y values measured
in one member of sample under study. One pair of X and Y data may be denoted as (x1, y1), another

66
as (x2, y2), another as (x3, y3), etc. These corresponding values on the axis x and y for one point in
the scattered plot are called “correlation pairs” (xi, yi).
An example for the scattered diagram of correlative relation is shown in the Figure 8.2;
it is a correlation between height and weight in men.

Fig. 8.2 Dot plot for correlative relationship between height and weight in men

If the points in the scattered diagram are clustered in some direction – it means that there is
some relation between biological characters monitored; correlation may be positive – “direct
relation” (Fig. 8.3) or negative – “inverse relation” (Fig. 8.4).

Fig. 8.3 Positive correlation Fig. 8.4 Negative correlation

y y

x x

If the points in the scattered diagram are irregularly scattered in the area – it means that there
is no correlation between biological characters monitored (Fig. 8.5).

67
 Fig. 8.5 No correlation

x
If we wish to know the strength of any relationship we have observed, and how reliable this
observation is, we need to employ statistical techniques called correlation and regression analysis.
Correlation is a measure of the relationship between two (or more) variables that helps us determine
whether the variables really are related and the degree to which they vary together. Regression is
a statistical tool for determining the mathematical relationship between one or several independent
or predictor variables and a single dependent or criterion variable, allowing us to calculate the
value of one variable given known values of the other variables. If we need to evaluate and describe
the statistical relation in a graphical presentation - we have to estimate the best-fit function that can
be used for description of this relationship, and to determine its equation (to calculate coefficients
for this equation – either linear or nonlinear).
According to the allocation of points in the scatter diagram we can distinguish between two
types of correlative relation: Linear or Non-linear correlation (Fig. 8.6). These two types of
correlative relationships differ in the way of their statistical evaluation and analysis.

Fig. 8.6 Linear (a) and Non-linear (b) correlation

x x

68
 8.2 Linear Correlative Relationship

The linear function is the most frequently used equation that we can use for estimation and
description of some correlative relationship between two variables (biological characters) monitored
in biology and related sciences. We need to employ the simplest case of regression analysis, the
simple linear regression, in this situation. Data amenable to simple regression analysis will consist
of a dependent variable that is a random-effect factor and an independent variable that is either
a fixed-effect or a random-effect factor. The data can be visualised in a scatter-plot and analysed by
fitting the best straight line to the points. The simplest and most commonly used fitting technique
of this sort is named least squares. The name comes from minimizing the sum of squared vertical
distances from the data points to the proposed line.

The analysis and description of the statistical relation is usually performed in the following
steps:
1) The construction of an empirical curve that describes the relation in a sample (estimates
the supposed theoretical line for the whole population):
We measure several values yi for the same value xi (e.g. in several men that have the same
height (xi) we measure their weights; we obtain several random values yi). We calculate an average
from these values yi in the appropriate xi, and then we join these averages in order to construct the
empirical curve that describes the relation in the particular sample monitored in our study. This
empirical curve can serve as the estimation of the best-fit linear function.
An example for the empirical curve in the case of relationship between height and weight in
men is presented in the Figure 8.7.

Fig. 8.7 Empirical curve for the correlative relation

69
 2) The construction of a regression line (i.e. calculation of equation for this best-fit line)
that can be used for description of the relation in the whole population.
We need to calculate coefficients of the best-fit regression equation using regression
analysis: y = a + bx
Coefficients a and b in the regression equation determine properties of the line:
a (called intercept) – represents the intercept point on the axis y for x=0,
b (slope, regression) = tg  (α - an angle that is formed by the line and the axis x).

We need always keep in mind that coefficients a and b are only the best estimates of the true
coefficients denoted α and  of the theoretical regression line that would uniquely describe the
functional relationship existing in the whole population.
Figures 8.8 and 8.9 demonstrate properties of the regression line that are determined by
coefficients a and b in linear equation.

Fig. 8.8 Coefficient a represents the intercept point on the axis y

If the coefficient a is a positive value, then the line intersects the axis y above the value 0, if
the coefficient a is a negative value, the line intersects the axis y below the value 0.

Fig. 8.9 Coefficient b represents the slope of the line

70
 If the coefficient b is a positive value, then the line is ascending (it indicates a direct relation
between x and y variables), if the coefficient b is a negative value, then the line is descending (it
indicates an inverse relation between x and y).

8.2.1 Regression Analysis

Regression analysis is a statistical technique that calculates the coefficients (parameters) of
the linear function: y = a + bx. The calculation results from sample data - correlation pairs (xi, yi),
measured for each of n number of individuals in the sample under study.
Calculation formula for regression coefficient b (slope):

n x y x  y
 

b i i i i
2
n  x 2    x 

i  i

The coefficient a (intercept) is derived from calculated b through the formula:

a
 y b
i   xi
n

After calculation of the equation for regression line, we need to determine two points for
construction of the theoretical regression line. We can choose any value x1 and calculate its
appropriate value y1 (according to the calculated regression line equation), and then choose another
x2 and calculate appropriate y2: y1 = a + bx1
y2 = a + bx2

Figure 8.10 demonstrates construction of the best-fit regression line for the scattered
diagram in the case of direct correlative relationship.

Fig. 8.10 Construction of the theoretical regression line

71
 Knowing the parameter estimates a and b for the linear regression equation, we can predict
the value of the dependent variable expected at a stated value xi. A word of caution is in order
concerning predicting yi values from regression equation. Generally, it is an unsafe procedure to
extrapolate from regression equations – that is, to predict yi values for xi values outside the observed
range of xi. What the linear regression actually describes is Y as a function of X within the range of
observed values of X. For values of X above or below this range, the function may not be the same
(i.e., α and/or  may be different); indeed, the relationship may not even be linear in such ranges,
even though it is linear within the observed range. If there is good reason to believe that the
described function holds for X values outside the range of those observed, then we may cautiously
extrapolate. Otherwise, beware.

8.2.2 Correlation Analysis

Correlation analysis is the statistical technique used for determination of association level
between variables in the analysis of the correlative relation monitored. In simple linear correlation,
we consider the linear relationship between two variables X and Y, whereas neither is assumed to be
functionally dependent upon the other. An example of a correlation situation is the relationship
between the wing length and tail length of a particular species of bird.
We calculate a correlation coefficient r that determines tightness (closeness) of the relation
between variables X and Y (and also determines the measure of dispersion of points around the
theoretical regression line in the scatter diagram). The calculation results from sample data -
correlation pairs (xi, yi), measured for each of individuals in the sample under study.
Calculation formula for correlation coefficient r:

r
 x  x  yi  y 
i 

 x  x    y  y 
2 2

i i

Correlation coefficient r is called “Parametric” (or Pearson’s) correlation coefficient, as we

need parameters (means of variables X and Y) for its calculation. Therefore it should be used in data
that follow Gaussian normal distribution only.

Values of correlation coefficient r are located within the interval -1 ; +1. The larger is the
absolute value of r, the closer is the correlation between X and Y variable. A positive correlation
coefficient implies that for an increase in the value of one of the variables, the other variable also
increases in value; a negative correlation coefficient indicates that an increase in value of one of the
variables is accompanied by a decrease in value of the other variable. If the correlation coefficient r
= 0, and one has a zero correlation, denoting that there is no linear association between the
magnitudes of two variables; that is, a change in magnitude of one does not imply a change in
magnitude of the other. Correlation coefficient r = +1 represents total (functional) direct relation

72
(ascending line), correlation coefficient r = -1 represents total (functional) inverse relation
(descendent line). Figure 8.11 presents these considerations graphically.

Fig. 8.11 Dot plots of correlations with different correlation coefficients

r=0 r>0 r<0

No correlation Direct (positive) correlation Inverse (negative) correlation
(X, Y increase together) (When X increases, Y decreases)
r = +1 r = -1
(total relation: ascending line) (total relation: descending line)

8.2.3 Significance of the Correlation Coefficient

The correlation coefficient r that we calculate from a sample is only an estimate of an actual
correlation coefficient in the population (denoted ). If we need to know whether the correlation in
the population really exists, we have to test a hypothesis of the independence (H0: =0) using t-test:

r
Test statistic: t
sr

Where sr is the standard error of the correlation coefficient r and is calculated using the
following formula:

1 r 2
sr 
n2

Degree of freedom needed for tabulated critical value:  = n-2

73
 We compare the calculated t with the critical value (Appendix 2: Critical values for Student’s
t-distribution) according to the chosen  and given v = n-2:

If t > t() => H0 is not true, the correlation between X,Y really exists in the population
sampled (r is significant),

If t  t() => H0 is true, the correlation between X,Y really does not exist in the population
sampled (r is insignificant).

8.3 Non-linear Correlative Relationship

There are many difficulties in calculations of non-linear regression equations; therefore it is

very convenient to use a computer in case of such non-linear relations. A statistical software with
options for so called polynomial regression is useful. In such cases, we get different regression
models (curves) computed by means of this polynomial regression. The most common non-linear
regression is the quadratic equation: y = a +b1 x + b2 x2 (“second-order polynom”). In this case, we
need to calculate regression coefficients a, b1, b2 for this equation by means of a computer in order
to find out the best-fit curve (parabola).
The calculation of a Spearman rank correlation coefficient represents another method for the
analysis of the non-linear relation between variables in biology.

8.3.1 Spearman Rank Correlation Coefficient

If we deals with a non-linear relation between two variables or if we have data obtained
from a bivariate population that is far from normal, then the correlation procedures discussed in the
chapter 8.2 are generally not applicable. Instead, we may operate with the ranks of the
measurements for each variable studied in these situations.
Calculation of the Spearman rank correlation coefficient is a non-parametric method, since
we don’t need parameters (means of variables X and Y) for calculation. This method may also be
used for data sets that don’t follow Gaussian normal distribution, and it can be used more generally
– in both linear and non-linear correlations. This method may also be used in normal data sets, but
non-parametric correlation coefficient is less forceful (less effective) than the parametric one.
Therefore, the non-parametric correlation coefficient is mostly used only for preliminary
calculations in normal data.
In the course of the calculation of Spearman rank correlation coefficient rs, we use only the
ranks of values instead of the actually measured values xi, yi,. The calculation results from the
number of individuals (n) in the sample and correlation pairs (xi, yi ).

74
 Method:

First, we arrange observed values of variables X and Y separately in two variant

sequences (ascending or descending rows). Then we assign appropriate ranks to the values in these
variant sequences:
e.g.: x2 < x4 < x1 < x5 < x3 < x8 < x6 < x7 ..............
Rank: 1 2 3 4 5 6 7 8 n
y3 < y1 < y5 < y2 < y4 < y8 < y7 < y9 .............
Rank: 1 2 3 4 5 6 7 8 n

If there are some equal values in the row, they get so called “average ranks”: e.g., if x4 and
x1 are equal then both get the rank 2.5 (calculated as = (2+3)/2)

Then we calculate Di: differences between ranks of corresponding xi and yi values:

e.g.: D1=3-2, D2=1-4, D3=5-1, D4=2-5 .........

Calculation formula for Spearman rank correlation coefficient:

6 Di
2

rS  1 
n.(n 2  1)
:

Where:
Di - differences between ranks of corresponding xi and yi values
n - number of members in the sample

After calculation, we compare computed rs to the critical rank coefficient found in the
statistical tables (Appendix 8: Tables of Spearman rank correlation) according to the chosen  and
given n:
If |rs| > rcrit. => There is a significant correlation between X and Y variables (relation
really exists in the population sampled),
If |rs|  rcrit. => There is an insignificant correlation between X and Y variables (relation
does not really exist in the population sampled).

Example:

Calculate the Spearman rank correlation coefficient for the relation between wing and tail
lengths among birds of a particular species:

75
 Method:

1) We arrange observed values of variable X and Y in an ascending variant sequence and

assign appropriate ranks to the values in these variant sequences.
2) We found appropriate values xi and yi and assign their ranks (see columns 1 - 4 in the
following table).
3) We calculate differences Di between ranks of corresponding xi and yi values (see columns
5 in the following table):

Wing length Tail length Difference

Rank of X Rank of Y Di2
(X) [cm] (Y) [cm] Di
10.2 1.5 7.1 1 0.5 0.25
10.2 1.5 7.2 2.5 -1 1
10.3 3 7.4 5 -2 4
10.4 4 7.4 5 -1 1
10.5 5 7.2 2.5 2.5 6.25
10.6 6 7.8 9.5 -3.5 12.25
10.7 7 7.4 5. 2 4
10.8 8.5 7.6 7 1.5 2.25
10.8 8.5 7.8 9.5 -1 1
11.1 10 7.9 11 -1 1
11.2 11 7.7 8 3 9
11.4 12 8.3 12 0 0

4) We calculate sum of squared differences : Di2 = 42.00 (number of pairs: n =12)

5) We calculate Spearman rank correlation coefficient:

6 Di
2
6(42.00)
rS  1   1  0.853
n(n  1)
2
1716

6) Critical rs (for =0.05, n=12) = 0.587  Spearman rank correlation coefficient is

statistically significant (at the level =0.05).
7) Conclusion:

Correlation between wing and tail lengths among birds of a particular species is statistically
significant (it really exists in the population).

76
 Chapter 9
Categorical Data
(Qualitative Data: On a Nominal Scale)

Sometimes an observed variable in biology can be described as nominal data (“qualitative”

data) – they can have only 2 levels of their “quality”: 0-1, yes-no, true-false, etc. We can not
measure any values in these biological characters.
We can determine only a presence (or absence) of this quality in an individual in the
samples under study.
E.g. when we evaluate survival rate (alive or dead), presence or absence of a disease
(healthy or ill), anatomic anomaly (yes or no), incidence of parasites (positive or negative), gender
(male or female), vaccinated or not vaccinated, etc.
Data of this sort are placed into named categories - so called “qualitative classes” (as
opposed to being measured as a point on a scale or ranked in order); therefore they usually are
referred to as categorical data. Categories of such data represent different variants of the observed
biological character. In some biological characters, there can be only two categories or, in other
biological character monitored, there may be more categories, e.g.:
• 2 categories: healthy-ill, alive-dead, male-female,

• More categories: eye colour – blue, brown, green, grey

hair colour, hair type - long, short, medium, smooth, curly, etc.
(each of these categories has 2 levels – “qualities”: yes – no).

Categorical groups are formed in a natural way in most qualitative biological characters.
However, sometimes, categorical groups may be formed also artificially, by dividing the scale upon
which continuous data occur. If we were to categorize age by decade (50-59, 60-69, and 70-79
years), we would have age groupings, which we could name 1, 2, and 3. These groups could be
considered as categories and categorical methods used. However, they fall into a natural rank order,
as group 1 clearly comes before group 2, etc. Rank methods give better results than categorical
methods, since rank methods are more powerful than categorical methods, which are not very
powerful. When ranking is a natural option, rank methods should be used.

77
 When dealing with categorical data, the basic statistic, count = frequency, is obtained by
counting the number of “events” per category (individuals that possess the appropriate “quality”) in
a sample of total number of n individuals.
The symbol for number of “events” used in calculation formulas is:

fi – frequency (count) in class i, and

n – total number of “events” (individuals) summed over all categories in a sample.

Another important statistic obtained from categorical data is the proportion of data in
a category, which is the count in the category divided by the total number of n individuals in the
sample under study. Proportion of data is a relative measure (in contrast to absolute frequencies)
and gives us probability of data in the category. Multiplication by 100 yields percent (denoted %).
Percent is useful in that most of the public is used to thinking in terms of percent, but statistical
methods have been developed for proportions. A symbol for a sample proportion could be P:

fi
P
n

E.g.: If 5 animals out of 50 have a disease  (Proportion) Probability of the disease in this
sample is P=5/50=0.1 (incidence of disease is 10%)

In statistical methods intended and used for categorical data, we can distinguish between
empirical and theoretical counts (in the terms of the sample and population):
f i – Empirical count (frequency) – observed in a sample (actually found)
fˆi – Theoretical count (frequency) – theoretically expected in a population sampled (this
theoretical count may be obtained in various ways in particular statistical methods for categorical
data, e.g. according to some literature sources, from a long-term monitoring of the “event” under
study in the past or by means of calculation from tables of empirical counts).

In the terms of a sample and population, we can also distinguish between:

Empirical proportion (probability) – in the sample: fi

P
n fˆi
Theoretical proportion (exact probability) in the population (for N=): π 
N
Theoretical (population) proportion is symbolized , in keeping with the convention of using
Greek letters for population values and Latin letters for sample values. Population proportion  is
a theoretical value, as we cannot calculate it in practice. We can only estimate its value by sample
proportion P. The exactness of the estimate is dependent (direct relation) on the sample size used
for the estimate calculations; it holds generally that the larger is n, the better is our estimation, i.e.
sample estimate P is close to the true population ).

78
 9.1 Analysis of Categorical Data

For nominal (categorical) data we can only use categorical methods that are based on
frequencies (counts) or proportions of “events” in statistical sets (we can’t use any measured
values and parameters like in methods used for numerical data). When dealing with counts in
statistical methods for categorical data, we usually arrange them into tables of counts that allow us
a better technique for all calculations used in the course of analysis of nominal data.
Nominal data analyses give us a possibility to assess:

A) Difference between counts in statistical sets:

- Sample vs. Population comparison,

- Sample1 vs. Sample 2 comparison.

B) Relationships between categorical data – “Contingency tables”

It is frequently desired to obtain a sample of nominal data and to infer whether the
population from which it came conforms to a specified theoretical distribution. For example, a plant
geneticist may rise 100 progeny from a cross that is hypothesized to result in a 3:1 phenotypic ratio
of yellow-flowered to green-flowered plants. Perhaps a ratio of 84 yellow : 16 green is observed,
although out of this total of 100 plants, the geneticist’s hypothesis would predict a ratio of 75
yellow : 25 green. The question to be asked, then, is whether the observed frequencies (84 and 16)
deviate significantly from the frequencies expected if the hypothesis were true (75 and 25).

The statistical procedure for attacking the question first involves the concise statement of the
hypothesis to be tested. The hypothesis in this case is that the population which was sampled has
a 3 : 1 ratio of yellow-flowered to green-flowered plants. This is referred to as a null hypothesis
(abbreviated H0), because it is a statement of “no difference”; in this instance, we are hypothesizing
that the population flower colour ratio is not different from 3 : 1. If it is concluded that H0 is false,
then an alternate hypothesis (abbreviated HA) will be assumed to be true. In this case, HA would be
that the population sampled has a flower-colour ratio which is not 3 yellow : 1 green. Recall that we
state a null hypothesis and an alternate hypothesis for every statistical test performed, and all
possible outcomes are accounted for by the two hypotheses.

The following calculation of a statistic called Chi-Square is used as a measure of how far
a sample distribution deviates from a theoretical distribution. This Chi-Square analysis represents
2
the basis of all calculations and techniques used for nominal data – we calculate test statistic of  -
test for differences between observed counts (in a sample) and those that would be expected
theoretically (in all population).

The null hypothesis used in this Chi-Square test is H0: observed counts = expected counts
(“no difference”).

79
 Calculation of test statistic Chi-square:

 
2
m
f i  fi 
ˆ
2

or 
m 2
fi
n
i 1 fˆi i 1 fˆi
Where:
f i - observed frequencies (in the sample class i),

fˆi - expected frequencies (in the whole population class i), i.e. frequencies expected in
class i if the null hypothesis is true.
m – number of classes (categories) in the sample or population.
n – total nimber members in the sample

The summation is performed over all m categories of data; in the example with the flower-
coloured ratio, there are two categories of data (i.e. m = 2): yellow-flowered plants and green-
flowered plants. The expected frequency, fˆi , of each class is calculated by multiplying the total
number of observations, n, by the proportion of the total that the null hypothesis predicts for the
3 1
class. Therefore, for the two classes in the example, fˆ1  100 .  75 and fˆ2  100 .  25 .
4 4

In the calculation of Chi-square test:

If calculated  2 = 0 then the observed and theoretical frequencies are exactly identical.
The bigger is the value of calculated statistic  2 the bigger is the difference between
observed and theoretical frequencies. Thus, this type of calculation is referred to as a measure of
goodness of fit (“goodness of fit test”).
If we compare calculated  2 with the critical value  2 for the specific  and DF:  = m-1
from the statistical tables , then:
 If  2 >  2,  difference between the observed and expected counts is significant (at the
level ). The null hypothesis is not true, i.e., sample distribution (empirical frequencies)
deviates from a theoretical distribution (theoretical frequencies).
 If  2   2,  difference between the observed and expected counts is not significant (at
the level ). The null hypothesis is true, i.e., sample distribution (empirical frequencies)
does not deviate from a theoretical distribution (theoretical frequencies).

In practice, Chi-Square test is usually used in testing for:

- difference between observed frequencies of patients in a sample and the statistical
probability of a disease in the population,
- difference between observed frequencies of patients in 2 (or more) samples (e.g. groups or
herds of animals),
- contingency tables (analysis of relations between categorical data).

80
 9.2 Test for Difference between Empirical and Theoretical Counts
(Sample vs. Population)

In practice, this test is usually used in the situations when the theoretical probability for
a studied “event” is known (e.g. predicted ratios in genetics, probabilities for the incidence of
a particular disease according to literature sources, from a long-term monitoring of the “event”
under study in the past, etc.). In the chi-square analysis we compare the expected frequencies
(calculated from theoretical probabilities) with the empirical frequencies observed in the sample
under our study in order to asses the statistical significance of differences between these counts.

Example:
From the total number of 146 calves in a sample 13 have enteritis. In the whole population
the probability of this disease is 4.5%. Is the enteritis occurrence in the sample different from the
whole population?
Method:
We can distinguish between 2 categories in the sample and population:
- ill animals
- healthy animals

Sample: n = 146 Population: =0.045 (4.5%)

Enteritis: 13

Empirical counts (observed in the sample): Theoretical counts (calculated):

f1 13 (ill) fˆ1 : .n = (0.045).146= 6.57 (ill)

f2 133 (healthy) fˆ2 : 146 - 6.57=139.43 (healthy)

Calculation of Chi-square test statistic:

 2  (13  6.57)  (133  139.43)  6.5895

2 2

6.57 139.43

Degree of freedom: = m-1=1

Critical values from the tables of chi-square distribution:

2crit.0.05 = 3.84

2crit.0.01 = 6.63

81
 2 > 2crit.0.05  difference between empirical and theoretical counts is significant (at the
α = 0.05 level of significance).

Conclusion: There is a significantly higher proportion of ill animals in the sample than in
the population (P<0.05).
(Empirical frequency of ill animals is 13, but theoretically it should be only 6.57 (to have the
same probability like in the whole population).

9. 3 Test for Difference between 2(or more) Empirical Counts

(Sample1 vs. Sample2)

In most of situations in practice, we don’t know theoretical probabilities or expected

frequencies for the whole population – more often we have to compare 2 or more groups of
empirical frequencies known from samples monitored, and have to decide whether these samples
differ in their empirical frequencies.
In the Chi-Square analysis we work with 2 or more groups having several qualitative
classes (unlike in the previous case, where there was only one group with 2 classes).
We mark: number of groups as r and frequencies in groups as fi
number of classes as c and empirical counts in classes as fj
In the course of the Chi-Square analysis, the empirical frequencies are usually arranged into
a table. By means of the double subscript, fij refers to the frequency observed in row i (group) and
column j (class); see the following example.

Example:
The number of live- and dead born piglets was observed in 3 farms (A, B, C) in a region.
We have to decide whether the frequencies of dead born piglets differ in the farms monitored.
Frequencies obtained from 10 litters in each farm are summarized in the following table:
3 groups - rows (A, B, C) – in general r, (i)
2 classes - columns (live, dead) – in general c, (j)

82
 c
Live Dead
r
A 96 25

B 121 22

C 89 16

Empirical frequencies in the table - fij

For the Chi-Square analysis we also need theoretical (expected) counts – we are able to
calculate them from the sums in rows and columns in the table; thus the next step in the analysis is
to sum empirical frequencies in rows and columns:

c Row 
Live Dead
r (Ri)
96 25
A 121
(100.34) (20.66)

121 22
B 143
(118.59) (24.41)

89 16
C 105
(87.07) (17.93)

Col. 
306 63 369 (n)
(Cj)

ˆ
Then, we can calculate theoretical frequencies f i j for each cell in the table.

Calculation formula for the theoretical frequencies in each table cell (row i, column j):

Ri . Cj
fˆij 
n

Where :
Ri = sum of empirical counts in row i
Cj = sum of empirical counts in column j

83
 E.g. calculation of the theoretical frequency for the cell in the first row and the first column:

121 .306
fˆ11   100 ,34
369

Calculation of the Chi-square test statistic:

2
6 f ij 96 2 25 2 1212 22 2 89 2 16 2
2   ˆ
i , j 1 f ij
 n      
100,34 20,66 118,59 24,41 87,07 17,93
 369  1,637

Degrees of freedom for the test (needed for the critical value  2 from the statistical tables)
are calculated according the following formula:

Degree of freedom: v = (r-1).(c-1) = 2

We compare the calculated Chi-square statistic with the critical value:

 2 crit. 0,05 = 5.99

Conclusion:
 2 <  2crit. 0.05  Difference between the empirical and theoretical counts is insignificant
(P>0.05).
It means that the farms monitored don’t differ in the mortality of born piglets; i.e.
frequencies of live - and dead born piglets don’t differ among farms A, B, and C .

9. 4 Contingency Tables
(Analysis of relations between categorical data)

In many situations, nominal data for two variables may be tested for a hypothesis H0:
frequencies in categories of one variable are independent on the frequencies in the second variable.
E.g. whether the incidence (frequency) of a parasitic infection in dogs is the same in vaccinated as
in the non-vaccinated individuals (= Does the incidence of parasites depend on the vaccination? Is
the vaccination effective? )

84
 Observed data are arranged in a contingency table:
- number of rows – r (categories of variable1: incidence of parasites)
- number of columns – c (categories of variable2: vaccination)

The null hypothesis H0 for this contingency table: frequencies in columns are independent
on the frequencies in rows.

According to the number of rows and columns we can distinguish between:

 Contingency table r x c (“r by c”)

 Contingency table 2 x 2 (special case of the table r x c for 2 categories in each variable)

9.4.1 Contingency table r x c

Method for the analysis of the contingency table r x c is the same as in the previous testing
(see the test for differences between empirical counts) by means of Chi-square statistic calculation.
We test the null hypothesis (independence of variables) through the testing for difference between
empirical and theoretical counts in this contingency table r x c.

E.g. it can be a situation, when we are concerned with the question whether some diseases
are associated with special breeds of cattle:
Variable 1 – Breeds of cattle (A, B, C)
Variable 2 – Diseases (1, 2)

Method:
1)We create the contingency table 3 x 2:

c
Disease 1 Disease 2
r
Breed A f11 f12

Breed B f21 f22

Breed C f31 f32

fij - Empirical frequencies

85
 2) We calculate theoretical frequencies fˆi j for all table cells (according to the known
formula):

Ri . C j
fˆij 
n
Ri – sums in rows
Cj – sums in columns

3) We calculate test statistic 2 (according to the known formula):

 
2
m
f i  fˆi 
2
or  
2
2
f ij
n
i 1 fˆi i, j fˆij

4) Degrees of freedom:  =(r-1) . (c-1)

5) Conclusion:

If the calculated Chi-square statistic is small, there is a little dependence between the
variables.
A large statistic indicates the positive dependence. If the critical value is exceeded, then the
H0 (independence) is rejected and dependence of variables monitored is statistically proved (In this
case it would mean that there is some dependence between breeds (A, B, C) and monitored
diseases).

9.4.2 Contingency Table 2 x 2

Contingency table 2 x 2 is a special case of the table r x c for 2 categories in each variable
only. We can solve such table either in the same way like the previous table r x c or by means of the
special (shortened) method.
Following situation can be solved by contingency table 2 x 2:
E.g.: Does the vaccination affect the incidence of parasites?
(Does the incidence of parasites depend on vaccination?)
Variable A – vaccine application
Variable B – incidence of parasites

86
 We test the null hypothesis H0: the incidence of parasites is not dependent on the
vaccination.

Variable 1 – vaccine application (A- yes, A´-no)

Variable 2 – incidence of parasites (B-yes, B´-no)

Method:
1) We create the contingency table 2 x 2:

B B´ Row 

A a b
a+b

A´ c d
c+d

Column  a+c b+d n= a + b + c + d

a – Frequency of animals that have A and B (vaccinated animals that have parasites)
b – Frequency of animals that have A and B´ (vaccinated animals that have no parasites)
c – Frequency of animals that have A´ and B (not vaccinated animals that have parasites)
d – Frequency of animals that have A´ and B´ (not vaccinated animals that have no
parasites)
(a, b, c, and d represent the empirical frequencies)

n – Total number of animals in the experiment

2) We calculate the test statistic 2 according to the formula:

n.(a.d  b.c) 2
2 
(a  b).(c  d ).(a  c).(b  d )

3) Degree of freedom:  =(r-1) . (c-1)=1

4) We compare the calculated Chi-square statistic with the critical value:

If  2 > 2crit.  H0 (independence of A and B) is rejected
If  2   2crit.  H0 (independence of A and B) is true

87
Example:
In a sample of 50 dogs: 25 dogs got an experimental anti-parasitic substance
25 dogs did not get the substance
Does the substance affect the incidence of parasites in dogs?

Without sub. With sub. Total

With parasites 15 9 24
Without parasites 10 16 26

Total 25 25 50

n.(15.16  9.10) 2
Test statistic:    2.885
2

(15  9).(10  16).(15  10).(9  16)

 =1
 2crit. 0.05 = 3.84

Conclusion:
 2 <  2crit.  H0 is not rejected i.e. frequencies in columns are independent on the
frequencies in rows (P>0.05).
(It means that the tested substance does not affect the incidence of parasites in dogs.)

88
 Appendix

Statistical Tables
Source:
Riffenburgh, R. H.: Statistics in medicine. ACADEMIC PRESS, San Diego, USA 1999
Zar, J. H.: Biostatistical Analysis. Prentice Hall, Upper Saddle River, N.Y. 1999

List of Tables:

Appendix 1 Normal Distribution

Appendix 2 Critical values for Student’s t-distribution
Appendix 3 Critical values for 2 distribution, Right tail
Appendix 4 Critical values for 2 distribution, Left tail
Appendix 5 Critical values for Snedecor’s F – test
Appendix 6 Critical Values for Mann-Whitney U-test
Appendix 7 Critical values for Wilcoxon signed rank test
Appendix 8 Critical Values for the Spearman’s Rank Correlation Coefficient rs

89
Appendix 1 Normal Distribution 1)

One-tailed applications Two-tailed applications

One-tailed α 1- α (area except Two-tailed α (area 1- α (area except
z
(area in right tail) right tail) in both tails) both tails)
0 .500 .500 1.000 .000
.10 .460 .540 .920 .820
.20 .421 .579 .842 .158
.30 .382 .618 .764 .236
.40 .345 .655 .690 .310
.50 .308 .692 .619 .381
.60 274 .726 .548 .452
.70 .242 .758 .484 .516
.80 .212 .788 .424 .576
.90 .184 .816 .368 .632
1.00 .159 .841 .318 .682
1.10 .136 .864 .272 .728
1.20 .115 .885 .230 .770
1.281 .100 .900 .200 .800
1.30 .097 .903 .194 .806
1.40 .081 .919 .162 .838
1.50 .067 .933 .134 .866
1.60 .055 .945 .110 .890
1.645 .050 .950 .100 .900
1.70 .045 .955 .090 .910
1.80 .036 .964 .072 .928
1.90 .029 .971 .054 .946
1.960 .025 .975 .050 .950
2.00 .023 .977 .046 .934
2.10 .018 .982 .036 .964
2.20 .014 .986 .028 .972
2.30 .011 .989 .022 .978
2.326 .010 .990 .020 .980
2.40 .008 .992 .016 .084
2.50 .006 .994 .012 .088
2.576 .005 .995 .010 .990
2.60 .0047 .9953 .0094 .9906
2.70 .0035 .9965 .0070 .9930
2.80 .0026 .9974 .0052 .9948
2.90 .0019 .9981 .0038 .9962
3.00 .0013 .9987 .0026 .9974

1) For selected distances (z) to the right of the mean are given (a) one-tailed α, the area under the curve in the
positive tail; (b) one-tailed 1- α, the area under all except the tail; (c) two-tailed α, the areas combined for both
positive and negative tails; and (d) two-tailed 1- α, the area under all except the two tails. Entries for the most
commonly used areas are italicized.

90
Appendix 2 Critical values for Student´s t-distribution.

α (1): 0.05 0.025 0.01 0.005 0.0025 0.001

DF = 
α (2): 0.10 0.05 0.02 0.01 0.005 0.002
1 6.314 12.706 31.821 63.657 127.321 318.309
2 2.920 4.303 6.965 9.925 14.089 22.327
3 2.353 3.182 4.541 5.841 7.453 10.215
4 2.132 2.776 3.747 4.604 5.598 7.173
5 2.015 2.571 3.365 4.032 4.773 5.893
6 1.943 2.447 3.143 3.707 4.317 5.208
7 1.895 2.365 2.998 3.499 4.029 4.785
8 1.860 2.306 2.896 3.355 3.833 4.501
9 1.833 2.262 2.821 3.250 3.690 4.297
10 1.812 2.228 2.764 3.169 3.581 4.144
11 1.796 2.201 2.718 3.106 3.497 4.025
12 1.782 2.179 2.681 3.055 3.428 3.930
13 1.771 2.160 2.650 3.012 3.372 3.852
14 1.761 2.145 2.624 2.977 3.326 3.787
15 1.753 2.131 2.602 2.947 3.286 3.733
16 1.746 2.120 2.583 2.921 3.252 3.686
17 1.740 2.110 2.567 2.898 3.222 3.646
18 1.734 2.101 2.552 2.878 3.197 3.610
19 1.729 2.093 2.539 2.861 3.174 3.579
20 1.725 2.086 2.528 2.845 3.153 3.552
21 1.721 2.080 2.518 2.831 3.135 3.527
22 1.717 2.074 2.508 2.819 3.119 3.505
23 1.714 2.069 2.500 2.807 3.104 3.485
24 1.711 2.064 2.492 2.797 3.091 3.467
25 1.708 2.060 2.485 2.787 3.078 3.450
26 1.706 2.056 2.479 2.779 3.067 3.435
27 1.703 2.052 2.473 2.771 3.057 3.421
28 1.701 2.048 2.467 2.763 3.047 3.408
29 1.699 2.045 2.462 2.756 3.038 3.396
30 1.697 2.042 2.457 2.750 3.030 3.385
31 1.696 2.040 2.453 2.744 3.022 3.375
32 1.694 2.037 2.449 2.738 3.015 3.365
33 1.692 2.035 2.445 2.733 3.008 3.356
34 1.691 2.032 2.441 2.728 3.002 3.348
35 1.690 2.030 2.438 2.724 2.996 3.340
36 1.688 2.028 2.434 2.719 2.990 3.333
37 1.687 2.026 2.431 2.715 2.985 3.326
38 1.686 2.024 2.429 2.712 2.980 3.319
39 1.685 2.023 2.426 2.708 2.976 3.313
40 1.684 2.021 2.423 2.704 2.971 3.307
41 1.683 2.020 2.421 2.701 2.967 3.301
42 1.682 2.018 2.418 2.698 2.963 3.296
43 1.681 2.017 2.416 2.695 2.959 3.291
44 1.680 2.015 2.414 2.692 2.956 3.286
45 1.679 2.014 2.412 2.690 2.952 3.281
46 1.679 2.013 2.410 2.687 2.949 3.277
47 1.678 2.012 2.408 2.685 2.946 3.273
48 1.677 2.011 2.407 2.682 2.943 3.269
49 1.677 2.010 2.405 2.680 2.940 3.265
50 1.676 2.009 2.403 2.678 2.937 3.261
52 1.675 2.007 2.400 2.674 2.932 3.255
54 1.674 2.005 2.397 2.670 2.927 3.248
56 1.673 2.003 2.395 2.667 2.923 3.242
58 1.672 2.002 2.392 2.663 2.918 3.237
60 1.671 2.000 2.390 2.660 2.915 3.232

91
Appendix 3 Critical values for 2 distribution, Right tail

α (area in right tail):

DF =
0.05 0.025 0.01 0.005 0.001
1 3.84 5.02 6.63 7.88 10.81
2 5.99 7.38 9.21 10.60 13.80
3 7.81 9.35 11.34 12.84 16.26
4 9.49 11.14 13.28 14.86 18.46
5 11.07 12.83 15.08 16.75 20.52
6 12.59 14.45 16.81 18.54 22.46
7 14.07 16.01 18.47 20.28 24.35
8 15.51 17.53 20.09 21.95 26.10
9 16.92 19.02 21.67 23.59 27.86
10 19.31 20.48 23.21 25.19 29.58
11 19.68 21.92 24.72 26.75 31.29
12 21.03 23.34 26.22 28.30 32.92
13 22.36 24.74 27.69 29.82 34.54
14 23.69 26.12 29.14 31.32 36.12
15 25.00 27.49 30.57 32.81 37.71
16 26.30 28.84 32.00 34.27 39.24
17 27.59 30.19 33.41 35.72 40.78
18 28.87 31.53 34.80 37.16 42.32
19 30.14 32.85 36.19 38.58 43.81
20 31.41 34.17 37.57 39.99 45.31
21 32.67 35.48 38.94 41.40 46.80
22 33.92 36.78 40.29 42.80 48.25
23 35.17 38.08 41.64 44.19 49.75
24 36.41 39.36 42.97 45.56 51.15
25 37.65 40.65 44.31 46.93 52.65
26 38.88 41.92 45.64 48.30 54.05
27 40.11 43.20 46.97 49.65 55.46
28 41.34 44.46 48.28 51.00 56.87
29 42.56 45.72 49.59 52.34 58.27
30 43.77 46.98 50.89 53.68 59.68
35 49.80 53.20 57.34 60.27 66.62
40 55.76 59.34 63.69 66.76 73.39
50 67.51 71.42 76.16 79.50 86.66
60 79.08 83.30 88.38 91.96 99.58
70 90.53 95.02 100.43 104.22 112.32
80 101.88 106.63 112.32 116.32 124.80
100 124.34 129.56 135.81 140.16 149.41

92
Appendix 4 Critical values for 2 distribution, Left tail

α (area in left tail):

DF =
0.001 0.005 0.01 0.025 0.05
1 .000016 .000039 .00016 .00098 .0039
2 .0020 .010 .020 .051 .10
3 .024 .072 .12 .22 .35
4 .091 .21 .30 .48 .71
5 .21 .41 .55 .83 1.15
6 .38 .68 .87 1.24 1.64
7 .60 .99 1.24 1.69 2.17
8 .86 1.34 1.65 2.18 2.73
9 1.15 1.73 2.09 2.70 3.33
10 1.48 2.16 2.56 3.25 3.94
11 1.83 2.60 3.05 3.82 4.57
12 2.21 3.07 3.57 4.40 5.23
13 2.61 3.57 4.11 5.01 5.89
14 3.04 4.07 4.66 5.63 6.57
15 3.48 4.60 5.23 6.26 7.26
16 3.94 5.14 5.81 6.91 7.96
17 4.42 5.70 6.41 7.56 8.67
18 4.90 6.26 7.01 8.23 9.39
19 5.41 6.84 7.63 8.91 10.12
20 5.92 7.43 8.26 9.59 10.85
21 6.45 8.03 8.90 10.28 11.59
22 6.99 8.64 9.54 10.98 12.34
23 7.54 9.26 10.20 11.69 13.09
24 809 9.89 10.86 12.40 13.85
25 8.66 10.52 11.52 13.12 14.61
26 9.23 11.16 12.20 13.84 15.38
27 9.80 11.81 12.88 14.57 16.15
28 10.39 12.46 13.57 15.31 16.93
29 10.99 13.13 14.25 16.05 17.71
30 11.58 13.79 14.95 16.79 18.49
35 14.68 17.19 18.51 20.57 22.46
40 17.93 20.71 22.16 24.43 26.51
50 24.68 27.99 29.71 32.36 34.76
60 31.73 35.53 37.49 40.48 43.19
70 39.02 43.28 45.44 48.76 51.74
80 46.49 51.17 53.54 57.15 60.39
100 61.92 67.32 70.07 74.22 77.93

93
Appendix 5 Critical values for Snedecor’s F – test (two-tailed, α = 0.05)

Numerator DF
1 2 3 4 5 6 7 8 9
1 647.79 799.50 864.16 899.58 921.85 937.11 948.22 956.66 963.28
2 38.506 39.000 39.165 39.248 39.298 39.331 39.355 39.373 39.387
3 17.443 16.044 15.439 15.101 14.885 14.735 14.624 14.540 14.473
4 12.218 10.649 9.979 9.605 9.365 9.197 9.074 8.980 8.905
5 10.007 8.434 7.764 7.388 7.146 6.978 6.853 6.757 6.681
6 8.813 7.260 6.599 6.227 5.988 5.820 5.696 5.600 5.523
7 8.073 6.542 5.890 5.523 5.285 5.119 4.995 4.899 4.823
8 7.571 6.060 5.416 5.053 4.817 4.652 4.529 4.433 4.357
9 7.209 5.715 5.078 4.718 4.484 4.320 4.197 4.102 4.026
10 6.937 5.456 4.826 4.468 4.236 4.072 3.950 3.855 3.779
11 6.724 5.256 4.630 4.275 4.044 3.881 3.759 3.664 3.588
12 6.554 5.096 4.474 4.121 3.891 3.728 3.607 3.512 3.436
13 6.414 4.965 4.347 3.996 3.767 3.604 3.483 3.388 3.312
14 6.298 4.857 4.242 3.892 3.663 3.501 3.380 3.285 3.209
Denominator DF

15 6.200 4.765 4.153 3.804 3.576 3.415 3.293 3.199 3.123

16 6.115 4.687 4.077 3.729 3.502 3.341 3.219 3.125 3.049
17 6.042 4.619 4.011 3.665 3.438 3.277 3.156 3.061 2.985
18 5.978 4.560 3.954 3.608 3.382 3.221 3.100 3.005 2.929
19 5.922 4.508 3.903 3.559 3.333 3.172 3.051 2.956 2.880
20 5.872 4.461 3.859 3.515 3.289 3.128 3.007 2.913 2.837
21 5.827 4.420 3.819 3.475 3.250 3.090 2.969 2.874 2.798
22 5.786 4.383 3.783 3.440 3.215 3.055 2.934 2.839 2.763
23 5.750 4.349 3.751 3.408 3.184 3.023 2.902 2.808 2.731
24 5.717 4.319 3.721 3.379 3.155 2.995 2.874 2.779 2.703
25 5.686 4.291 3.694 3.353 3.129 2.969 2.848 2.753 2.677
26 5.659 4.266 3.670 3.329 3.105 2.945 2.824 2.729 2.653
27 5.633 4.242 3.647 3.307 3.083 2.923 2.802 2.707 2.631
28 5.610 4.221 3.626 3.286 3.063 2.903 2.782 2.687 2.611
29 5.588 4.201 3.607 3.267 3.044 2.884 2.763 2.669 2.592
30 5.568 4.182 3.589 3.250 3.027 2.867 2.746 2.651 2.575
40 5.424 4.051 3.463 3.126 2.904 2.744 2.624 2.529 2.452
60 5.286 3.925 3.343 3.008 2.786 2.627 2.507 2.412 2.334
120 5.152 3.805 3.227 2.894 2.674 2.515 2.395 2.299 2.222
 5.024 3.689 3.116 2.786 2.567 2.408 2.288 2.192 2.114

94
Appendix 6 Critical Values for Mann-Whitney U-test (2-tailed, α = 0.05)

n1
n2 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

2 16 18 20 22 23 25 27 29 31 32 34 36 38
3 15 17 20 22 25 27 30 32 35 37 40 37 45 47 50 52
4 16 19 22 25 28 32 35 38 41 44 47 50 53 57 60 63 67
5 23 27 30 34 37 42 46 49 53 57 61 65 68 72 76 80
6 31 36 39 44 49 53 58 62 67 71 75 80 84 89 93
7 41 46 51 56 61 66 71 76 81 86 91 96 101 106
8 51 57 63 69 74 80 86 91 97 102 108 114 119
9 64 70 76 82 89 95 101 107 114 120 126 132
10 77 84 91 97 104 111 118 125 132 138 145
11 91 99 106 114 121 129 136 143 151 158
12 107 115 123 131 139 147 155 163 171
13 124 132 141 149 158 167 175 184
14 141 151 160 169 178 188 197
15 161 170 180 190 200 210
16 181 191 202 212 222
17 202 213 224 235
18 225 236 248
19 248 261
20 273

95
Appendix 7 Critical values for Wilcoxon signed rank test

n 0.05 0.01 0.001

6 1 - -
7 2 - -
8 4 0 -
9 6 2 -
10 8 3 -
11 11 5 0
12 14 7 1
13 17 10 2
14 21 13 4
15 25 16 6
16 30 19 8
17 35 23 11
18 40 28 14
19 46 32 18
20 52 37 21
21 59 43 25
22 66 49 30
23 73 55 35
24 81 61 40
25 90 68 45
30 137 109 78
35 195 160 120
40 264 221 172
45 344 292 232
50 434 373 304

96
Appendix 8 Critical Values for the Spearman’s Rank Correlation Coefficient rS

n α: 0.20 0.10 0.05 0.02 0.01 0.005 0.002

5 0.800 0.900 1.000 1.000
6 0.657 0.829 0.886 0.943 1.000 1.000
7 0.571 0.714 0.786 0.893 0.929 0.964 1.000
8 0.524 0.643 0.738 0.833 0.881 0.905 0.952
9 0.483 0.600 0.700 0.783 0.833 0.867 0.917
10 0.455 0.564 0.648 0.745 0.794 0.830 0.879
11 0.427 0.536 0.618 0.709 0.755 0.800 0.845
12 0.406 0.503 0.587 0.678 0.727 0.769 0.818
13 0.385 0.484 0.560 0.648 0.703 0.747 0.791
14 0.367 0.464 0.538 0.626 0.679 0.723 0.771
15 0.354 0.446 0.521 0.604 0.654 0.700 0.750
16 0.341 0.429 0.503 0.582 0.635 0.679 0.729
17 0.328 0.414 0.485 0.566 0.615 0.662 0.713
18 0.317 0.401 0.472 0.550 0.600 0.643 0.695
19 0.309 0.391 0.460 0.535 0.584 0.628 0.677
20 0.299 0.380 0.447 0.520 0.570 0.612 0.662
21 0.292 0.370 0.435 0.508 0.556 0.599 0.648
22 0.284 0.361 0.425 0.496 0.544 0.586 0.634
23 0.278 0.353 0.415 0.486 0.532 0.573 0.622
24 0.271 0.344 0.406 0.476 0.521 0.562 0.610
25 0.265 0.337 0.398 0.466 0.511 0.551 0.598
26 0.259 0.331 0.390 0.457 0.501 0.541 0.587
27 0.255 0.324 0.382 0.448 0.491 0.531 0.577
28 0.250 0.317 0.375 0.440 0.483 0.522 0.567
29 0.245 0.312 0.368 0.433 0.475 0.513 0.558
30 0.240 0.306 0.362 0.425 0.467 0.504 0.549
31 0.236 0.301 0.356 0.418 0.459 0.496 0.541
32 0.232 0.296 0.350 0.412 0.452 0.489 0.533
33 0.229 0.291 0.345 0.405 0.446 0.482 0.525
34 0.225 0.287 0.340 0.399 0.439 0.475 0.517
35 0.222 0.283 0.335 0.394 0.433 0.468 0.510
36 0.219 0.279 0.330 0.388 0.427 0.462 0.504
37 0.216 0.275 0.325 0.383 0.421 0.456 0.497
38 0.212 0.271 0.321 0.378 0.415 0.450 0.491
39 0.210 0.267 0.317 0.373 0.410 0.444 0.485
40 0.207 0.264 0.313 0.368 0.405 0.439 0.479
41 0.204 0.261 0.309 0.364 0.400 0.433 0.473
42 0.202 0.257 0.305 0.359 0.395 0.428 0.468
43 0.199 0.254 0.301 0.355 0.391 0.423 0.463
44 0.197 0.251 0.298 0.351 0.386 0.419 0.458
45 0.194 0.248 0.294 0.347 0.382 0.414 0.453
46 0.192 0.246 0.291 0.343 0.378 0.410 0.448
47 0.190 0.243 0.288 0.340 0.374 0.405 0.443
48 0.188 0.240 0.285 0.336 0.370 0.401 0.439
49 0.186 0.238 0.282 0.333 0.366 0.397 0.434
50 0.184 0.235 0.279 0.329 0.363 0.393 0.430
51 0.182 0.233 0.276 0.326 0.359 0.390 0.426
52 0.180 0.231 0.274 0.323 0.356 0.386 0.422
53 0.179 0.228 0.271 0.320 0.352 0.382 0.418
54 0.177 0.226 0.268 0.317 0.349 0.379 0.414
55 0.175 0.224 0.266 0.314 0.346 0.375 0.411

97
 References

Armitage, P., Berry, G., Matthews, J.N.S.: Statistical methods in medical research. Blackwell
Publishing, Oxford UK 2002, 814 p.

Ashcroft, S., Pereira, Ch.: Practical Statistics for the Biological Sciences. PALGRAVE
MACMILLAN, New York, N.Y. 2003, 167 p.

Carvounis, Ch.: Handbook of Biostatistics. PARTHENON PUBLISHING, New York, USA 2000,
103 p.

Everitt, B. S.: Medical Statistics from A to Z. CAMBRIDGE University Press, Cambridge, UK

2003, 230 p.

Glantz, S. A.: Primer of Biostatistics. McGRAW-HILL, N.Y. USA 2005, 520 p.

Riffenburgh, R. H.: Statistics in medicine. ACADEMIC PRESS, San Diego, California USA 1999,
581 p.

Zar, J. H.: Biostatistical Analysis. Prentice Hall, Upper Saddle River, N.Y. 1999, 663 p.

98