Quantitative Methods for Communication

Students1

Petro K. Poutanen2

10.2.2014

1
This material can be seen at http://blogs.helsinki.fi/quantitative-communication/.
Note that all the materials are licensed under the
Attribution-NonCommercial-ShareAlike 2.5 Generic.
2
I want to thank Sam Kingsley for correcting the language, and Heikki Hyhkö and
Olli Parviainen for their comments to some sections of this work. These contributions
improved the quality of this work. However, any errors are mine, and feedback and
comments about this material can be sent to me at petro.k.poutanen@gmail.com.
Contents

1 Introduction 3

2 Quantitative Research 5
2.1 Starting to think statistically . . . . . . . . . . . . . . . . . . . . 5
2.2 Theories and models . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Research design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Research process . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 Further resources . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Methods 11
3.1 Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Content analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Data 14
4.1 Data and variables . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Getting the data . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.4 Preparing the data . . . . . . . . . . . . . . . . . . . . . . . . . . 18

5 Analysis 20
5.1 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.1.1 Statistics for single variables . . . . . . . . . . . . . . . . 20
5.1.2 Statistics for two variables . . . . . . . . . . . . . . . . . . 23
5.2 Inferential statistics . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.1 Confidence intervals . . . . . . . . . . . . . . . . . . . . . 29
5.2.2 Testing for a difference between means . . . . . . . . . . . 31
5.3 Correlation and regression . . . . . . . . . . . . . . . . . . . . . . 33

1
CONTENTS 2

5.3.1 Linear regression . . . . . . . . . . . . . . . . . . . . . . . 35

Chapter 1

Introduction

This document consists of introductory material on quantitative methods. It is

aimed at students of communication but is suitable for the other social scientific
fields as well. The emphasis will, though, be on observational studies, with a
focus on surveys and content analysis. Experimental research settings are not
discussed in this document, although they form an essential part of the scientific
study of communication. This limitation is due to framing issues, since this
material has been prepared for a specific communication methods university
course.
The examples presented in this paper are not from real studies, but specif-
ically calculated for the purposes of the discussed themes, so they should not
be taken as real research results. I have tried to avoid using any formulas and
presented mathematical expressions only when I think it is absolutely necessary
for the understanding of the explained matter. However, even then I have done
so informally, with a focus on getting an intuitive grasp of the matter.
What will follow in this document is mainly based on my own experiences in
teaching quantitative methods courses for communication students during three
semesters. Hence, the choices made on what is included and excluded from the
presentations of various methods are based on my comprehension of how much
information is necessary for students whose background knowledge of statistics
and research methods is not that strong yet. I have added a lot of links and
options for further readings for those who would like to deepen their knowledge
on some issues.
The presentations of issues and methods is not by any means meant to be
exhaustive; rather the purpose has been to provide an easy and self-contained

3
CHAPTER 1. INTRODUCTION 4

introduction to various options that are available if one chooses to conduct

quantitative research. The focus will be heavily on the hands-on approach that
I have adopted in my own studies of statistics and quantitative methods.
Chapter 2

Quantitative Research

This section covers basic protocols and starting points for developing a research
design for quantitative inquiries in communication.

2.1 Starting to think statistically

An important first step is to try to tune one’s brain in to seeing things through
the prism of statistics. This is not easy though, since human beings are not
statistical thinkers. In general, statistical reasoning is based on absolute and
relative frequencies and relations between two or more phenomenon. A second
step is to try to avoid anxiety and accept the fact that learning statistical
methods will take some time and effort. Especially when one is inexperienced,
everything feels disconnected and thus incomprehensible. But step by step,
things will begin to come together, and some insights and a-ha moments will
follow.
In statistics everything is connected. To understand regression analysis one
needs to understand the concept of correlation. To understand correlation one
needs to know what continuous variables, distributions, and linearity mean.
Once the knowledge is built gradually and the relevant steps are taken, it will
be much easier to begin to understand more complicated methods, such as
regression analysis.
First of all, it is relevant to note that the kind of statistical research done in
the social sciences is empirical in its nature, i.e. we are not doing mathemat-
ics with numbers, but using the statistical toolbox (i.e. methods) for making

5
CHAPTER 2. QUANTITATIVE RESEARCH 6

observations and inferences about empirical phenomena in a rigorous way. Sec-

ondly, all the information that is collected should be in a measurable form, i.e.
transferable into numbers. By empirical I did not mean the kind of empiricism
without any theoretical concepts that was conducted by the early positivists
in the 19th century. Theories have an imperative role in scientific empirical
research in formulating research hypotheses and interpreting results.

2.2 Theories and models

Theories and models form the basis for empirical inquiries. In empirical research
we are not only interested in variations in our data (what we observe happening
in the world), but we also want to test whether the data fits to our model and
a theory. We also want to develop theories or even build new ones based on the
previous ones and/or new empirical evidence.
Social scientists usually ask questions concerning humans’ social behavior,
attitudes, and beliefs. How do people behave in certain social circumstances?
Are there any regular patterns of behavior that can be repeatedly observed? For
example, what attitudes are associated with a high degree of online participa-
tion? What makes people vote for certain candidates in presidential elections?
What is the effect of media coverage? How do socioeconomic factors predict
and influence media consumption? And so on.
In general, there are two strategies to follow when conducting quantitative
research. The first, in which researchers formulate hypotheses on the basis of the
previous research and test them against empirical data, is called confirmatory
research. Confirmatory research confirms (or rejects) hypotheses. The other,
possibly supplementary, strategy is exploratory research. Exploratory research is
by definition exploration, a kind of adventure into the data. This method starts
with the data and exploring in order to formulate hypotheses and theories based
on the understanding derived from the data.
Some data analyzers may suggest that exploratory data analysis could be
followed by confirmatory analysis: first the researcher explores the data, finds
some interesting associations, then finds theory to support the observed varia-
tions, and finally performs a statistical test to verify what has just been found.
However, there is a danger in this reasoning. The data can include correlations
which do not really exist but are in the data just by chance. In theory, one can
find arbitrary evidence from the data and develop a theory around it to “con-
CHAPTER 2. QUANTITATIVE RESEARCH 7

firm” the observations. This is how human brains mostly work: we see patterns
(“evidences”) around us and try to explain them. However, patterns are also
seen where there is actually nothing at all going on!
This confirmation bias can be harmful in scientific research, and statistical
methods are exactly the way of avoiding succumbing to it. Therefore, good
scientific research is based on some type of theoretical reasoning, either taken
as given and tested against empirical data or developed over the course of data
exploration and carefully linked with an existing body of empirical research and
theories.

2.3 Research design

Research design consists of the different options and choices to be made when
conducting empirical research. A study comprises a sequence of choices, which
all influence the validity and plausibility of the research. Among things to
consider are research questions and objectives, operationalization of variables,
reliability and validity, and data-gathering methods. Some of these areas will
be covered in the further sections. Here, we will briefly make a bold separation
between two common research designs in social and communication research-
experimental and non-experimental designs – and briefly discuss validity and
reliability.
Experimental design is based on a well-prepared and framed experiment in
which some particular causal relationship is tested under controlled conditions.
In experiments people are usually randomly divided into separate groups thus
controlling the possible bias caused by the variables that are not studied. Then
the interesting variable is manipulated and the possible effect is observed and
measured. Good experimental studies test real causal mechanisms.
Non-experimental research design refers to observational studies, such as
a survey or a content analysis. In observational studies a researcher collects
observations using a research instrument, such as a questionnaire in surveys,
and then performs statistical tests on the data. Good sampling techniques can
make observational studies reliable and generalizable to the population within
certain limits. No causal mechanisms can, however, be tested as reliably as in
controlled trials.
There are two important criteria that should always be considered with
respect to a given research design. The first of them is validity – is the study
CHAPTER 2. QUANTITATIVE RESEARCH 8

really measuring what it claims to be measuring? For example, if the study

design is a survey that intends to measure people’s attitude towards social media,
we can evaluate how well the theoretical constructs are operationalized into
survey questions and how well the items used describe attitudes towards social
media. Validity will in the end determine how well the study can predict the
behaviors or attitudes it measures.
Reliability is the other criteria, which could be called “repeatability”. It de-
scribes how accurate the measurement is, i.e. if the same study were conducted
again, would similar results be drawn? It is useful to start thinking about reli-
ability by asking how consistent the measurable constructs would be from one
sample to another. There are many ways to test reliability, such as test-retest
(how much two different samples correlate together for the tested questions) or
reliability estimates, such as Cronbach’s alpha.

2.4 Research process

In textbooks (and guides like this) one can often find an idealistic and unrealistic
model for how the research process should proceed. These process models are,
however, useful in describing what different stages there are in conducting a piece
of research, and especially when planning one. Since we are focusing here on non-
experimental research designs, a process illustrating a non-experimental research
process is presented. The following outline of a research process includes 8 steps
of which steps 1–3 are related to the preparation of the study, 4 to the data
collection, 5–7 to the actual analysis, and 8 to presenting the results.

1. Exploring theory (i.e. reading books and scientific articles)

2. Determining the research questions and formulating hypotheses

3. Defining what/who the observational units are (i.e. people), the variables
(what is to be measured/asked) and the sample (how the observational
units will be selected, i.e. the criteria for inclusion and how many there
will be of them)

4. Collecting data and preparing it for analysis

5. Describing the data through frequencies and relations between different

variables

6. Conducting statistical tests and inferences

CHAPTER 2. QUANTITATIVE RESEARCH 9

7. Making any required additional inquiries to the data and testing cause
and effect relations

8. Interpreting the results and writing the research report.

2.5 Further resources

A good place to start training in statistical thinking and reasoning is of course
any text book. Good "open textbooks" can be found online, such as OpenIntro Statistics.
Plenty of great online introductory courses on statistics, probability, data
analysis, etc. are also available. For example, Coursera provides high-quality
massive open online courses (MOOC’s) on topics such as data analysis, com-
putational analysis, R (statistical programming language), and network analy-
sis. To begin with, take a look at Scott E. Page’s Model Thinking (Univer-
sity of Michigan), Andrew Conway’s Statistics One (Princeton University),
or Data Analysis and Statistical Inference (Duke University) taught by Mine
Çetinkaya-Rundel.
Khan Academy’s free tutorials on statistics and probability are of great qual-
ity and interest for those who want to learn more about the mathematics of
statistical reasoning and methods. Khan Academy is an educational website
hosted by Salman Khan, consisting of 4,000 of his micro lectures on various
topics.
Good online sources for familiarizing oneself with communication theories
and models are also available. Communication theories can serve as testable
theories or models for building up own one’s own theoretical framework.

• Communication theoriesis a comprehensive resource of different commu-

nication theories maintained by Twente University. It is a well-organized
catalog of different theories arranged according to their level and purpose.
Also the most important references and practical examples are presented
for some theories.

• Theory for communicationis a "work in progress" web page by the Univer-

sity of Colorado Boulder and it comprises a massive amount of information
about communication theories. There are plenty links and lectures con-
cerning important theorists, different research paradigms, and so forth.

• A First look at Communication Theoryis a web site for a communication

theory text book by Em Griffin including an overview of many of the
CHAPTER 2. QUANTITATIVE RESEARCH 10

important theories in the area of communication studies.

Chapter 3

Methods

This section introduces three types of different methodological approaches for

quantitative research in communication: survey, content analysis, and network
analysis.

3.1 Survey
A survey is one of the most popular data collection and research methods across
the social sciences and market research. So much so, in fact, that it has become
difficult to get people to answer survey questionnaires. Moreover, online surveys
have arrived and changed the game a little, making data collection easier and
cheaper, but at the same time subject to many biases.
In essence, a survey is a research instrument that usually takes the form of
a structured questionnaire. The questionnaire data can be collected through
face-to-face or phone interviews or by sending an invitation to answer to the
questions by post or via email. In research methodology jargon, survey design
is called a cross-sectional study if it is conducted only once, or a longitudinal
study if data is collected more than once over time.

3.2 Content analysis

Content analysis is a widely used technique in communication sciences. It is
a method by which some observable contents, such as texts, images, objects,
etc. are transferred through coding into a measurable and verifiable form. The

11
CHAPTER 3. METHODS 12

process of conducting a content analysis follows the general research process

(see section 2.4) of quantitative inquiries. Perhaps the most distinctive part is
the construction of a coding scheme, a detailed instruction for identifying and
classifying the units of analysis.
The unit of analysis refers to what is counted, what is the "analytical whole"
taken as a separate unit. Originally, content analysis was introduced to the
communication sciences to bring analytical rigor and objectivity. Therefore,
the "content" was originally defined as manifest, i.e. observable and as un-
ambiguously counted as possible. Examples of such units are the lengths of
newspaper articles as measured in inches or the numbers of different sexes in
different scenes of a soap opera. Content can also be visual, non-verbal or even
audible. Manifest content is easy to determine beforehand and to identify from
the data.1
The identification of contents can be subject to the subjective biases of the
coder. More objectivity can be brought to the analysis by using more than one
coder and calculating an inter-coder coefficient. In this way, the reproducibility
of the results by other researchers can be used as a measure of reliability. There
are many ways to do the calculation, and the whole data does not need to
be double-coded, but a fraction of it, such as 10–20%. One of the most-used
reliability measures in this context is Cohen’s Kappa.

3.3 Network analysis

Network analysis, and especially social network analysis, has always been a
popular method in communication sciences, although recently it has become
even more popular. Nowadays, the idea of social networks is so widespread that
almost everyone has some notion of what a social network is. For example, one of
the most well-known ideas about social networks is that every two people are, on
average, six steps away from each other. This six degrees of separation thesis was
famously tested empirically by sociologist Stanley Milgram in the late sixties.
In the age of Facebook, the virtual length between every two Facebook users
has shrunk to 4.7 steps. This illustrates the "small-world structure" of social
networks, meaning that social networks comprise highly connected clusters of
nodes (people), which have several "weak ties" between them holding the whole
1 Units of analysis can be less manifest though, i.e. based on the subjective interpretation

of the coder. An example of such unit of analysis could be attractiveness or an emotional tone
of voice, positivity or negativity of a speech act, etc.
CHAPTER 3. METHODS 13

network together and making it possible to reach almost anyone quickly.

A network is comprised of nodes (e.g. people) and edges (e.g. social connec-
tions). When planning to conduct a network analysis the researcher will need
to consider what the units of analysis are, i.e. who are the actors. They can
be people in a small workplace or in a community, but they can also be nation
states or power plants. One also needs to consider the sample: what is the
population of actors that comprises the network? The other crucial thing to
consider is to determine the form of the tie between the two units of analysis:
what makes the relation? Usually the connection takes the form of some type
interaction between the actors, which can be directional or mutual.
For beginners, many great examples and easy instructions for conducting a
social network analysis with Facebook or Twitter data can be found atthis blog,
for example. Network analysis of one’s own Facebook friends provides both fun
and excitement and is a good step towards conducting serious research using
network analysis. For these purposes this blog provides a lot of good resources
to begin with.
Chapter 4

Data

This section concerns issues related to data gathering and sources of data.

4.1 Data and variables

Data refers to a set of values, which are usually organized by variables (what is
being measured) and observational units (members of the sample/population).
An example of data is a data matrix in a spreadsheet program, such as Excel or
SPSS (see figure 4.1). Along the upper horizontal line there are the variables
(e.g. survey questions) and down the first vertical line there are the observations
(e.g. people). In each cell there is a value that is the given observational unit’s
value of a given variable.
Variables are of different types and can be classified in many ways, for exam-
ple as numerical and categorical variables. Numerical variables are measured
by some (usually existing) measures, whereas categorical variables are qualita-
tive, and not necessarily more or less, or bigger and smaller than one another.
Another way of classifying variables is according to their measurement scale.
The continuous variable is numerical and it can take, in theory, an infinite
amount of values. An example of such a variable is length in centimeters or
inches. The discrete variable is also numerical, but differs from a continuous
variable in that it takes a finite number of values. An example of a discrete
variable is a performance score from 1 to 10. Each value is, in theory, equally
far from the subsequent value, so that 4 is exactly the same increase from 3 as
9 is from 8.

14
CHAPTER 4. DATA 15

Figure 4.1: An example of a data matrix in SPSS program. The figure shows a
data matrix of World Economic Freedom data. The countries on each row are
observational units or "cases", and the columns of different freedom indexes,
such as business freedom, are variables. Each country then has its individual
cell scores for the variables.
CHAPTER 4. DATA 16

If a numerical variable has an absolute zero, the variable can be measured

on a ratio scale. A typical example is weight. It can be zero, but definitely not
less than that. Therefore we can say that one observation is twice as heavy as
the other one. If the variable has no zero point, it has an interval scale, meaning
that the distances between different values are same (e.g. from 10 to 20 and
from 40 to 50), but the zero point is arbitrary. An example of an interval scale
is temperature. A centigrade thermometer can show "0", but the quality itself,
temperature, does not cancel out. The zero point is arbitrary. Therefore we
cannot really say that it is "twice as cold" as yesterday.
Categorical variables, in turn, can be nominal, in which case there is no order
at all: each category has its unique meaning ("What domestic pets do you like
most: 1 = cats, 2 = dogs, 3 = hamsters, 4 = bunnies?"). If there is a sense of
order there, the variables are called ordinal. A Likert-type scale represents an
ordinal measurement: 1 = “not like me at all”, 2 = “not like me”, 3 = “not sure”,
4 = “somewhat like me”, 5 = “very much like me”. A special type of variable is
the dichotomous one. It can have only two values (e.g. gender) and it can be
interpreted both as numerical and categorical.

4.2 Sampling
It is relatively easy to observe people’s behaviors or ask them what they think
about something. However, it is another thing completely to claim that these
observations (e.g. ticks on a survey form) apply to other people than those who
we have asked as well. The problem of how general our results are can be solved
by sampling methods. The size and representativeness of a sample define the
limits for inferences that are made on the basis of the sample data. In general,
researchers collect from a few hundred to thousands or even tens of thousands
of observations. But even with a sample as small as 30 cases some conclusions
can be drawn.
One of the most critical issues in data analysis is to understand what the
difference between the sample and the population is. Basically, a sample is
subset of a population. It is not just any subset, however, as a researcher
usually has some kind of idea of what kind of subset it is, i.e. how the sample
has been selected. In the ideal case, 1) each member of the population has the
same probability of being selected for the sample and 2) they are independent
with regard to the measurable characteristic. Then we refer to a random sample.
CHAPTER 4. DATA 17

This is an ideal case and we can define the exact limits within which the results
drawn from a random sample can represent the whole population.
The fact that a random subset of a larger population can be representative
is based on the central limit theorem (CLT), a finding belonging to the area of
probability theory. The central idea of CLT is that given the sample size and the
variation in population, a sufficiently large number of samples (or their mean
values) summed up will yield a bell curve -shape normal distribution. As we
know what the properties of this distribution are, we can assume, under certain
conditions, that the mean of the sample we have, is near to most of the other
samples’ means, meaning that we don’t need to take hundreds of samples but
only one, randomly selected. This finding forms a basis for statistical estimation
and testing, which are covered in the analysis section 5.
In reality, a random sample is not easily achievable, and there are other
sampling strategies too. For example, in a convenience sample the members of
the population are selected on the basis of their availability. Results drawn from
studies based on this type of sample are not generalizable to any population. If
the study is conducted for a sufficiently small population, such as a firm with
500 workers, sampling need not be used. All the workers can be sent an email
inquiry to participate in the study. Researchers need to evaluate then how
well those who participated represent the whole firm (for example, are all the
departments included) and whether there are any possible sources for systematic
bias (such as people without computers, who only rarely read their emails).

4.3 Getting the data

In addition to traditional data-gathering methods such as interviewing and ob-
serving, the Internet, digitalization, and the recent development of computa-
tional techniques have opened up new possibilities. The use of mobile phones,
credit cards, customer-loyalty cards, social media, web browsing – all leave
behind digital traces that can be used as research material on our social behav-
ior. The massive accumulation of this type of data has led some to talk about
big data and its possibilities for changing our society. However, although not
just anyone can get access to such data files, there are many new intriguing ways
for gathering data from the Internet that are accessible to all. For example, data
from social media forms an important part of communication research today.
We will not cover the techniques for computational data mining here, but
CHAPTER 4. DATA 18

refer to some examples of collecting digital data easily. For example, data on
real social networks can be collected from social network services such as Twitter
and Facebook. For example, an Excel plug-in called NodeXL enables users with
direct connections to social networks, such as Twitter, Facebook, YouTube,
Flickr, Wikis, and emails to mine data (take look at these handouts on how to
use NodeXL). An easy option for mining network data from Facebook is also
provided by an app called Nettvizz. Many more tools and techniques for data
collection can be found at DIRT: Digital Research Tools Wiki that has gathered
together hundreds of tools for conducting research in the digital environment.
It lists tools from data mining to reference management, from quantitative to
qualitative data, and from commercial to open software.
If one is interested in extracting information from websites, web scraping is
the order of the day. OutWit Hub is an example of a tool for scraping. By this
tool a researcher can collect data from web pages directly and automatically
without painful copy-pasting. Scrapers can be used directly via web browsers.
OutWit Hub is available for Mozilla Firefox, and for Google Chrome there is
another option that can be used. Software requiring programming skills, such as
R, offers several possibilities for gathering different types of online content. Take
a look at this nice example of gathering data on consumer attitudes towards an
airline company through Twitter.
Last but not least, it goes without saying that there is tons of open and free
data online that has already been gathered by someone and is made available for
anyone to analyze. Data collected by governments, NGOs, and think tanks is
readily available. Sources for open data sets can be found through governmental
data pages and, for example, through a platform held by the Open knowledge
foundation, an organizations promoting for the idea of open data.

4.4 Preparing the data

Before doing any analysis it is necessary to prepare the data so that it is as error-
free as possible and organized so that it is suitable for spreadsheet programs to
read it. There are many sources of error, such as coding errors, that should be
checked before starting any analysis. Otherwise the risk of drawing incorrect
conclusions because of biased data remains. The first thing to do is to browse
through the data and try to find the possible errors manually. All the variables
should also be checked for their range of possible values - for example, there
CHAPTER 4. DATA 19

are no 999-years-old people - so checking the minimum and maximum values

for all variables is useful. The convention for coding "missing values" (i.e. non-
responses) is to mark them as 9999, N/A (standing for "no answer"), or blank.
You should check whether the program you are using interprets missing values
correctly.
Another type of data-preparation problem can occur if the data has been
gathered by hand or scraped from a website, discussion boards, Facebook, etc.
Data that comprises text, such as names of people or companies, can include
spelling errors, missing letters, etc. Usually such errors have to be dealt with by
hand, but nowadays there are good computational resources for the purposes
of data cleaning, such as Google’s Open refine or Fusion Tables. Open refine
is easy to use, and there are many good tutorials on its main functions. Open
refine includes algorithms that can, for instance, help to identify spelling errors
and unify textual values. Once ready, the data can be transferred into a CSV
file and uploaded into any spreadsheet program.
Chapter 5

Analysis

This section discusses some of the most common statistics and statistical anal-
ysis methods.

5.1 Descriptive statistics

Broadly speaking there are two types of research strategies: exploratory and
confirmatory (see: Theories and models 2.2). In statistical data analysis, the
descriptive part of the analysis can be seen as a kind of exploration. Its purpose
is to get a general view of the data and the distributions of the variables by
diagrams, tables, and basic statistics, such as mean and standard deviation. The
descriptive analysis is a necessary part of the research and is always conducted
before doing any statistical tests or more complicated modeling. This part
presents some common techniques for descriptive data analysis, while the next
section Inferential statistics 5.2 focuses on statistical testing and modeling.

5.1.1 Statistics for single variables

Visual presentations are indispensable means for exploring the distributions of
different variables. A variable’s distribution is basically a set of all the different
values that the variable can take and their observed frequencies. For visualizing
the distribution of a numerical variable a histogram is the usual choice (5.1).
The histogram in the figure 5.1 shows the frequencies of the observed salaries.1
1 Salaries lower than 1238€ representing the minimum wage, and bigger than 5300€, rep-

resenting the top 10% of the observed values, are cut off for making the variable’s distribution
easier to analyze for the purposes of these examples.

20
CHAPTER 5. ANALYSIS 21

Figure 5.1: A histogram of salaries with the mean (red line) and median (green
line) and normal curve.
CHAPTER 5. ANALYSIS 22

On the horizontal line are salaries and on the vertical line the corresponding fre-
quencies (how many times a certain value has been observed). Not all observed
values are shown in a histogram, but adjacent values are grouped together into
several bins that are represented as the bars in the histogram. The important
thing to consider is how symmetrical the distribution is. The more symmetrical
it is, the easier it will be to describe using simple statistics. The distribution
of salaries is clearly skewed to the right, as there are relatively very few people
who earn a lot (the tail of the distribution goes right) and the bulk of the obser-
vations are located near to the mean (red vertical line) and the median (green
vertical line) a litle bit left from the center of the distribution.
Among the simplest possible descriptive statistics are the measures of central
tendency and dispersion. The measurement scale (see section 4.1) and distri-
bution of the variable generally determine which statistics would best describe
the variable. For numerical variables we can calculate an arithmetic mean. The
symbol for population mean is µ (the Greek letter “mu”) and for the sample
mean x̄. The sample mean is what we can calculate on the basis of our data,
and it serves as an estimate for the population mean, which is a theoretical “real”
value in the population. The sample mean is calculated by summing all the in-
dividual values together and dividing that sum by the number of observations:
(x1 + x2 + . . . + xn ) /n.
For the distribution of salaries above the mean is about 2782, which is located
slightly to the right of the peak of the frequencies, since the distribution is skewed
to the right (the extreme values “attract” the mean). It is easy to see that the
mean value does not capture the skewness of the distribution. Actually, the
mean best describes the central tendency of a variable when the variable is fully
symmetrical. However, if it is not, we can use the median (Md), which is the
middle value of the all values ranked on a scale from minimum to maximum (or a
mean of the two middle values in case of an even number of different values). In
this case, the median is 2505 and probably a better measure of central tendency,
since it is robust to the effect of extreme values.
The most basic measure for dispersion is called standard deviation. The
symbol for population standard deviation is σ (the Greek letter “sigma”) and
for sample it is s. Standard deviation describes how far the values are on average
from their mean value. The standard deviation of the salaries is 913, meaning
that the values are dispersed on average 913 euros away from the mean value.
Symmetrical distributions are useful, because if we know the mean and the
standard deviation, we know quite a lot of other things about the distribution
CHAPTER 5. ANALYSIS 23

Figure 5.2: A symmetrical curve with standard deviations and corresponding

percentage coverage of all the values.

as well. In the case of a symmetrical distribution about 68% of the all values
are within one standard deviation from the mean, and almost 95 % of all the
values are two standard deviations away from the mean (see figure 5.2).
For describing the distribution of a categorical variable a bar chart is used
instead of a histogram. The bar chart (see figure 5.3) displays different cate-
gories of a variable measuring how often people say they use the Internet for
personal matters. The variable varies from non-users to those who spend time
on the Internet daily. The measurement scale is ordinal, since categories can
be meaningfully ordered, but the intervals are not necessarily equal. The first
category “no access” is different from the other categories in the way that it
could be omitted from the analysis or handled independently.
For categorical variables central tendency measurements are not usually cal-
culated, except mode (Mo). Mode gives us the frequency of the biggest category.
In this case it is the category "Every day", which has 1012 observations in it.
The bar chart could also be drawn with relative frequencies on the vertical axis,
in which case each category would have a value in percentages instead of ab-
solute frequencies. If the distribution is fully symmetrical, all the measures of
central tendency – mean, median and mode – are situated in the middle of the
distribution where the peak of the frequencies are.

5.1.2 Statistics for two variables

When describing distributions of more than one variable at the same time, the
interest is usually in the relationship between the two variables. For example,
CHAPTER 5. ANALYSIS 24

Figure 5.3: A bar chart of personal use of the Internet.

CHAPTER 5. ANALYSIS 25

Figure 5.4: A scatterplot of age and salary.

one could be interested to find out if age and salary are somehow related. Do
the values of age get higher, for example, as salary increases? In the case of two
continuous variables we can draw a scatterplot (5.4) figure and study how the
values of the two variables vary together, i.e. whether there is any recognizable
pattern. However, it is important to note that no causal claims (such as X
causes Y: X → Y ) should be made on the basis of this kind analysis.
In the scatterplot (see figure 5.4), the association seems to be relatively weak
over the different ages. Salary hovers around its median value (2505) for all the
ages from 25 onwards. Some higher salaries can be seen across all ages. These
exceptional observations are grouped together in the tail that was pointing right
in the histogram (5.1). A very slight ascending pattern could be observed.
A scatterplot is the standard method for exploring the relationship between
two variables, but it works only for numerical continuous variables. For explor-
ing the relationships between two categorical variables we need a contingency
table or a special kind of bar chart. A contingency table is a table with absolute
CHAPTER 5. ANALYSIS 26

Table 5.1: A contingency table of gender and the personal use of the Internet
with absolute and relative column frequencies.

or relative frequencies and it can be used for 2–3 categorical variables. It depicts
the frequencies of one variable across all the classes of another variable. Usually
of the interest are the relative frequencies (i.e. percentages), since the sample
sizes of the classes that we are interested in comparing can vary.
An example will clarify this. Let us take “gender” as the first variable and
“personal use of the Internet” as the second variable. Now, we are interested
in finding out whether the values of males and females (that is, the classes
of the first variable “gender”) vary similarly across the different values of the
personal use of the Internet. To do this we need to use relative frequencies, as
the absolute frequencies are not comparable due to different amounts of males
and females in the sample. The relative frequencies are usually presented so
that they are summed up at the bottom of the columns, i.e. for each of the
classes individually that are compared together (see table 5.1).
Relative frequencies can also be depicted visually by a bar stacked chart (see
figure 5.5). In this case, bars represent different gender classes, and Internet use
is presented as relative areas within the bars, summing up to 100% for each
bar. On the basis of the table or the bar chart, we have good reason to suspect
that there are slightly more daily Internet users in the male category than
CHAPTER 5. ANALYSIS 27

Figure 5.5: A bar chart of gender and the personal use of internet.

the female, and slightly more those who have no access at all among females.
This relationship could, however, diminish or change if a third variable, such as
different age groups, is introduced to the table. Therefore it is often not enough
to only study the relationship between two variables, rather several variables
together with different combinations should be explored. The results should
also be tested statistically to verify that the observed differences are arguably
more than just random variation, which we will discuss in section 5.2.
There is a third form of visual presentation, the box-plot, that is useful when
a categorical variable needs to be studied against a numerical variable. The
box-plot has the categorical variable on the horizontal axis and the numerical
variable on the vertical axis. Each category of the categorical variable gets its
own box with two whiskers. The boxes and whiskers illustrate the distributions
of each class of categorical variable across the values of the numerical variable.
The ends of the whiskers represent maximum and minimum values, the stars
and points are outliers, the box contains 50 % of all the observations, and the
CHAPTER 5. ANALYSIS 28

Figure 5.6: A box plot of gender and salary.

black line within the box is the median value. In the figure 5.6 salaries of males
and females are explored by plotting the categorical gender variable against the
continuous salary variable.
This plot tells us that the median salaries of males (3000 euros) are higher
and that there is more dispersal in the salaries of males (the box including 50 %
of the observations is taller). We can also spot numerous outliers (exceptional
observations) in the high end of females.

5.2 Inferential statistics

Whereas descriptive statistics are necessary for making sense of the data and
exploring it, inferential statistics, especially significance testing is what statistics
is all about. The idea of doing statistical tests is related to the discussion on
sample: if we have a representative sub-set of a population as our sample, we
can make statistical inferences concerning the population on the basis of that
CHAPTER 5. ANALYSIS 29

sample. These inferences are then generalizable within particular limits, called
confidence intervals, that need to be specified.

5.2.1 Confidence intervals

Let us start with a concrete example. In the descriptive statistics section5.1 we
calculated a mean value for a cropped income variable and got 2782 euros as
the overall mean of salaries. The number of observations (n) was 641. If we
randomly take one of the cases from the data and check whether that person’s
salary is close to the mean, it is likely that it is not exactly the same. There
is some distance between the mean and the individual value, either up or down
from the mean value. In order to get a picture of all the differences in our
sample, we usually calculate the standard deviation, which describes the average
distance of data points from the mean. The standard deviation (s) of our salary
variable is 913 euros.
Using this information it is possible to calculate the standard error (SE) of
the mean, which is a measure of how far the sample means deviate from the
true population mean on average from sample to sample. SE is the standard
deviation for the theoretical sampling distribution containing all the possible
means from all the different samples. According to the central limit theorem
(see section 4.2), the sampling distribution is nearly normal and centered at
√
the population mean. The equation for SE is: s/ n, i.e. the sample standard
deviation divided by the square root of the sample size. This information helps
us to define the confidence intervals (CI) for the mean, i.e. the interval within
which the true population mean most likely lies.
Let us calculate an example in which we want to be 95% sure about the exact
interval within which the true population mean lies. For the upper boundary
we add SE multiplied by 1.96 to the sample mean. For the lower boundary we
subtract the same value from the sample mean, i.e: CI = ±(1.96 × SE). Where
does the multiplier “1.96” come from? This value is based on the idea that
the sampling distribution is symmetrical, and the value describes the critical
limit at both tails of a standardized normal distribution, telling us that there
are no more than 2.5% of observations above/below these values. Therefore, the
multiplier times SE gives us the limits for mean values that 95% of any sample
we could possibly take from the population would yield.
√
The SE for our sample mean is 913/ 641 ≈ 36. Let us calculate the 95%
CI for our sample mean: 2782 ± (1.96 × 36) ≈ ±80.7. This allows us to say
CHAPTER 5. ANALYSIS 30

Figure 5.7: A plot of mean salaries for males and females with 95% confidence
intervals.

that in 95 of 100 samples the real population mean value lies between 2711 and
2853 euros, and yet we haven’t done more than one sample! The 95% confidence
intervals for the salaries of males and females (these groups have different means,
standard deviations, and sample sizes, and hence different standard errors) are
depicted in the figure 5.7.
In figure 5.7 the T-bars show the ranges within which the population means
for both males and females will fall. The bars show the observed sample means.
An important observation is that the error bars are not overlapping, i.e. they
do not cover same range of values. This observation makes us more confident in
saying that the population means of males and females are different from each
other and that the difference is not due to random deviations between different
samples.
CHAPTER 5. ANALYSIS 31

5.2.2 Testing for a difference between means

In the theories and models section 2.2 we briefly touched on the topic of hypoth-
esis testing, when making the distinction between confirmatory and exploratory
research. Statistical testing is a method by which hypothesis testing is con-
ducted. Usually we assume the null hypothesis H0 and the alternative hypothe-
sis H1 . The alternative hypothesis H1 is what we are interested in. Hypotheses
are the research questions that take the form of a "claim". The null hypothesis
is what we test and try to falsify, i.e. to refute by empirical evidence, that is,
the data. The statistical test gives us the assurance for saying that the null
hypothesis is probably not true, but it also gives us the probability for being
wrong (i.e. "risk") when making such a claim. When we have a reason to sus-
pect the claim of the null hypothesis, then we say that we have found evidence
for the alternative hypothesis (if not, however, “proved” it to be true). Most
typically, H0 takes the following form: "there is no difference between A and
B" or "A and B are equal". In turn, H1 can be expressed in the following way:
"A is bigger/smaller than B" or "A is not zero in the population".
Let us perform a test by which we can be sure that the mean salaries of
males and females are different from each other, as the figure 5.7 suggested.
We begin by setting the null hypothesis H0 : there is no difference between the
mean salaries of males and females. The alternative hypothesis H1 would hence
be mean salaries differ from each other. Since we are not testing whether one
or the other group has bigger or lower salary, but simply if they differ from
each other, the test is called a two-tailed test. The two “tails” refer to the 2.5%
tails of the sampling distribution. If we test whether one group’s mean salary is
bigger or smaller than the other one’s, we call it a one-tailed test, in which case
we would have one 5% tail. The logic of hypothesis testing is illustrated in the
figure 5.8, in which there is a sampling distribution with a one and two-tailed
tests and regions for rejection and acceptance of the null hypothesis.
To compare two means from independent samples we can use a statistical test
called a t-test. The logic is the same as above when we calculated confidence
intervals for the sample mean, but this time we use a different formula for
calculating how probable it would be to observe a particular difference between
the means of the two samples (males and females) given the null hypothesis is
true. In our sample the difference happens to be 529.9 euros. The results of the
t-test can be seen in the table 5.2.
The test starts from the assumption made in H0 that there is no difference
CHAPTER 5. ANALYSIS 32

Figure 5.8: A sampling distribution with regions for acceptance and rejection
of the null hypothesis.

Table 5.2: t-test statistics.

between the salaries, and tests how likely it would be to get 529.9 as the dif-
ference given the null hypothesis. It appears that given the sample sizes and
standard deviations, it would be nearly impossible to observe such a difference
by chance. This is indicated by the t-test value which is 7.670, being way up
from the 95% critical value. According to the p-value ("Sig. (2-tailed)” column),
less than 0.001% of the all possible samples could generate such a difference.
The last columns of the table gives the 95% confidence intervals of the differ-
ence, indicating that out of 100 samples 95 will yield a difference between 394.2
and 665.6. Since almost every sample, if properly conducted, would yield such a
difference, we have a very good reason to suspect the null hypothesis and hence
get support for the alternative hypothesis.
It is good to note that the t-test and the comparison of means is just one
type of statistical test. There are many possibilities available, of which the most
important are the t-test, F-test and χ2 (“chi-sqare”) -test. For example, when
doing a contingency table, the test used for examining the statistical significance
of the observed differences is the chi-square test. Tests may look different and be
calculated differently but the basic idea behind statistical testing is as presented
CHAPTER 5. ANALYSIS 33

here. The Decision Tree for Statistics is a helpful aid when finding out what
statistical test should be used given the data and measurement levels of the
variables.

5.3 Correlation and regression

In the Descriptive statistics section 5.1 we used a scatter plot to draw two con-
tinuous variables, age and salary, against each other. On the basis of the picture
we were not able to determine if there was any association between the variables.
For studying the linear relationship between two continuous variables a mea-
sure called the Pearson product-moment correlation coefficient (“correlation” in
short) can be used. It is a measure of a linear relationship, not a curvilinear
one. It does not take a stand on any kind of causality (such as X is a cause of
Y), but describes the strength of the association.
The symbol of the Pearson correlation is ρ or r and in reality it is a standard-
ized measure of covariance ("co-variation"). The standardized values can vary
between -1 and +1, where 1 indicates perfect positive (linear) relationships, -1 a
perfect negative (liner) relationship, and 0 stands for no correlation at all. For
the correlation to be considered “weak” or “strong” depends on the study and
data, but usually more than ± 0.3 to 0.5 is expected. For example, correlation
between age and salary in our example is 0.124, which is, although statistically
significant (p<.05), very weak.
The correlation coefficient is easily calculated by any statistical package,
but the results are not meaningful unless there is a linear relationship between
the variables. The distributions of the variables should also be approximately
symmetrical for correlation to be meaningful. For example, the presence of
outliers would severely harm the results of a correlation coefficient as seen in
the classical example by Anscombe2 in figure 5.9.
It is good to note how misleading correlation coefficients (or any other sin-
gle statistics) can be. Therefore, graphical explorations are necessary. In the
Anscombe’s quartet (fig 5.9) the first plot shows the correlation as it should
be, the second shows a curvilinear relationship, the third shows the effect of an
outlier, and the fourth shows the effect of an extreme outlier. If the data is
biased and the outliers are a necessary part of the data (not random errors),
the Pearson correlation should not be used. In such cases, Spearman’s rank
2 Anscombe, F. J. (1973). Graphs in statistical analysis. The American Statistician, 27(1),

17-21. Retrieved from http://www.sjsu.edu/faculty/gerstman/StatPrimer/anscombe1973.pdf

CHAPTER 5. ANALYSIS 34

Figure 5.9: Anscombe’s quartet. Correlation coefficients, means and standard

deviations of X’s and Y’s and regression lines are almost identical although the
data sets are considerably different.
CHAPTER 5. ANALYSIS 35

correlation can be used, which is a similar measure to the Pearson correlation,

but is not influenced by the effect of outliers.

5.3.1 Linear regression

Once we have familiarized ourselves with the basic idea of correlation, it is time
to move on to linear regression, which is a technique for modeling the relation-
ship between two or more variables. The scatterplot in the figure 5.10 depicts
the regression line that passes through the dots, representing the observations.
This regression line is similar to a linear polynomial function that can be repre-
sented as a linear equation having a constant "a" (the intersection with Y-axis),
the slope "b" (coefficient for defining how steep the line is), and the variables
x and y, whose relationship the function describes. This type of equation for
a linear line takes the following form: y = a + bx, and can be thought of as
what goes in as x comes out as y. It is important to note that theoretically
speaking we are not dealing with associations any more, but making a causal
claim that from x follows y. When x changes, y changes in a way specified by
the right side of the equation. In the language of regression analysis, x is called
the independent variable (the one that is not influenced; also “predictor”) and y
is called the dependent variable (the one that is dependent on x ). The symbol
for constant is β0 (“beta”) and for coefficient(s) βx .
If you take a look at the plot in figure 5.10, age and salary are depicted as a
linear line, and the equation describing the position and slope of the line is also
labeled. In the equation y stands for “salary”, and x for “age”. The numbers
2340 and 10.23 are estimates for β0 and β1 . The R2 in the top left of the plot is
the coefficient of determination, which describes how well the line – the linear
model – fits the data, i.e. how well the line accounts for the variation of the
observations. R2 can be interpreted as a percentage. In this instance the model
captures no more than 1.5% of the variation.
The equation can be written in the following way: ysalary = 2340 + 10.23 ×
xage . The constant β0 is 2340 and the coefficient β1 is 10.23. How to interpret
the model? First, the coefficient is positive, so age has a positive effect on salary.
Secondly, the constant (or intercept) is 2340. So if age is 0, the salary would be
2340 when β1 cancels out. The interesting part, the effect of the age on salary
(read: the effect of x on y) is described by the coefficient β1 . It is interpreted in
the following way: when x increases by 1 unit (which is 1 year in this case), y
will increase by 10.23. So each time we get one year older, our salary increases
CHAPTER 5. ANALYSIS 36

Figure 5.10: A plot showing the regression line and equation for age and salary.
CHAPTER 5. ANALYSIS 37

by 10.23 euros according to the model.

The idea of a regression model is to predict the values of y on the basis of the
values of one or more x variables. Let’s try to predict the salary for a 40-year-old
person. We start by inserting 40 in place of x so that the right of the equation
will be 2340+10.23*40, which yields 2749. This a very simple linear model.
Considering the low coefficient of determination (98.5% was not explained), the
next task of a researcher is to find better predictors, i.e. x -variables that can
predict changes in y. The motive for using regression analysis is that it allows
us to add as many x variables into the equation as needed. The regression
analysis will account for them all together, considering their simultaneous ef-
fect on y. The motive for adding new predictors to the equation comes both
from theoretical and statistical reasons, i.e. we know that x influences y under
certain conditions and R2 increases accordingly. Linear regression analysis also
has some constraints, however. Most importantly, the dependent variable need
to be continuous. Independent variable(s) need to be discrete/continuous or
dichotomous. All variables need to be more or less normally distributed, and
independent variables should not correlate strongly together (a condition called
multicollinearity).
Let’s add two new variables into the regression equation. We introduce a new
variable called "education", which is dichotomous and hence can take two values:
“1” standing for “person has a university/college degree” and “0” for “person does
not have a university/college degree” (i.e. they have a degree lower than that or
no degree at all). We are interested in testing the effect of a university/college
degree on the predicted salary. The second new variable is “gender”, testing the
effect of “being female” (=1) on the predicted salary. The regression equation
is now ysalary = β0 + β1 x1(age) + β2 x2(education) + β3 x3(gender) . The output of
the regression analysis is in the table 5.3.
From the table 5.3 we can see that the effect of age is 16.787, meaning that
each year yields this much extra to the salary. The constant is now 2117.362,
meaning that this is the value when both age, education, and gender are zero.
The effect of education is that if a person has a college/university degree, 766.969
will be multiplied by 1, and hence added to the salary. The same goes if a
person is female, but this time we subtract 592.928 from the salary. Let’s make
the same prediction as above for a 40-year-old female with a college/university
degree. The salary is now 2117.362 + 16.787*40 + 766.969*1 - 592.928*1 =
2962.883. Which one of the predictors is stronger? Beta values in the table are
standardized values for B’s, making it possible to compare their effects on y. It
CHAPTER 5. ANALYSIS 38

Table 5.3: Regression coefficients.

Table 5.4: Model summary table.

seems to be that education is the strongest one, having almost twice as big an
effect as age. t-tests at the right end of the table show whether the predictors
are statistically significant in the model (p-values should be <0.05).
From the other output table 5.4 we can also see that R2 increased from
1.5% to 24.8% ("adjusted R2 " accounts for the error caused by increasing the
number of predictors), so it is a big change in the goodness of fit of the model.
The standard error of estimate is 791.659, and it is a measure of error in the
model. It can be interpreted as the average deviation of the predicted values
(salaries given by the model) from the observed values (the observed salaries in
the data).
The difference between observed and predicted values is called residual vari-
ation in regression analysis and it can be marked with e for “error”: y =
β0 + β1 x1 + e. This means that y is equal to the model plus the rest of the
variation of y that is not captured by the model. If we modify the equation
slightly, we can say that e = y − (β0 + β1 x1 ). Technically speaking, the goal of
a regression analysis is to minimize the amount of errors. This happens usually
by the procedure called least square estimation, which is done by finding such
values for β0 and β1 that the sum of the squared differences between y and
(β0 + β1 x1 ) become as small as possible.
Finally, a researcher needs to assess how well the model actually fits to the
CHAPTER 5. ANALYSIS 39

Figure 5.11: A residual plot should show a homogenous cloud around the center
line.

data. This is typically done by examining the residual variance, i.e. the part of
the variation that was not explained by the model. For this purpose a residual
plot can be drawn (figure 5.11). If the residual variance is normally distributed
and varies in an unsystematic manner around the predicted values, we can state
that the model is quite a good predictor of the values for y.