Research methodology module-2

RESEARCH METHODOLOGY
MODULE-2
Prepared by
Satyajit Behera
ARYAN INSTITUTE OF ENGINEERING AND
TECHNOLOGY, BHUBANESWAR

Module-2
Data collection(1):
Data collection is a process of collecting information from all the relevant sources
to find answers to the research problem, test the hypothesis and evaluate the
outcomes. Data collection methods can be divided into two categories: secondary
methods of data collection and primary methods of data collection.
Data collection is an important social research. It is also known as field work. It
involves administrating the research tools to geather data, It connects link to the
reality of the work for the researchers. Data collection consists of taking ordered
information from reality and transferring to some recording systems so that social
behaviour can be understood and predicted, It is based on research design.
Data
Data are observations and evidence regarding some aspects of the
problems/issue under study. According to John Geltang: “A datum is what is
observed, in manifest or phonotypical". Data provide information for decision
making. Information reduces uncertainties in a decision making.
Types of data
Nearly endless varieties of data existence can be obtained but only few types are
relevant to each research study. They can be classified on the basis source,
quantification, function and others. By nature of data, there are two types of data:
facts and opinion.
Facts
Facts describes tangible things. They measure anything that actually exists or
can exist. Facts then described as things done or a piece of information having
objectives reality. Facts can be intangible as long as they can really be
determined.
Examples of facts:
 The distance between Bhubaneswar and Cuttack is 25 km.
 We have 8 planets.
The information presented in the above example give us the accurate picture of
the distance and the number of planets respectively.

Opinion
Opinions are how people perceive something. They are what people believe
about something and what whose beliefs signify. They are the results of people
attitudes, intensions, knowledge and motives. These all reflects people perception
about matter. It can be an attitude or image. Attitudes are mental sets or
predispositions to some manner. An image is what something is like.
Examples of opinion: I believe there is life on Mars.
Importance of data collection
 Data collection completely fulfills the data requirements of a research
project. It is the connecting link for the researchers to the world of reality.
 It provides the sources of comparative data by which data can be
interpreted and evaluated against each other. Based on the data collection.
data are presented and analyzed.
 It suggests the type and method of data for meeting the information
needed. Several data collection methods are used to collect several types
of data.
 It serves as a source of future reference and evidence because they are used
to prepare written records. They can now provide lots of material for the
subsequent research.
 It helps to takes ordered information from reality and transferring into
some recording system so that can be later examined and analyzed. It is
from that pattern that social behaviour can be predicted.
Primary data
Primary Data Collection Methods
 Primary data collection methods can be divided into two groups:
quantitative and qualitative.
 Quantitative data collection methods are based in mathematical
calculations in various formats. Methods of quantitative data collection
and analysis include questionnaires with closed-ended questions, methods
of correlation and regression, mean, mode and median and others.
 Quantitative methods are cheaper to apply and they can be applied within
shorter duration of time compared to qualitative methods. Moreover, due
to a high level of standardisation of quantitative methods, it is easy to make
comparisons of findings.

 Qualitative research methods, on the contrary, do not involve numbers or
mathematical calculations. Qualitative research is closely associated with
words, sounds, feeling, emotions, colours and other elements that are non-
quantifiable.
 Qualitative studies aim to ensure greater level of depth of understanding
and qualitative data collection methods include interviews, questionnaires
with open-ended questions, focus groups, observation, game or role-
playing, case studies etc.
 Your choice between quantitative or qualitative methods of data collection
depends on the area of your research and the nature of research aims and
objectives.
Sources of data collection
Data may be collected from several sources. It is not easy to list them in
details. Researchers use these sources according to their data needs. However, the
general classification of data collection sources can be presented under two
groups:
Primary sources:
It provides primary data. Primary data are first hand, original data collected. They
are attained directly. The researcher obtained them directly. They are collected
by researcher and they have not been processed once. They are collected from
families, representatives, organization, etc. Interviews, questionnaire,
observation are the major tools for collecting data from primary sources.
Data collection methods/techniques
No matter what the basic design of the research, it is necessary to collect accurate
data to achieve useful results Researchers use a number of methods to collect
data. They are as follows:
Survey
The term survey has two constituents “sur" which means over and “view" which
means to see. Thus the word survey means to oversee, that is, to look over
something from high place. A survey is a data collection method based on the
study of a given population. It is a systematic gathering of information from the
people for the purpose of understanding or predicting some aspect of their
behaviour.

The survey method gathers data from a relatively large number of cases at a
particular time. It is not concerned with character of individuals. It is concerned
with generalized statics that result when data are abstract.
Types of surveys
 Census survey:
It covers the survey of population. It is very expensive and time and effort
consuming. Bu t provides diverse range of data.
 Sample survey:
It covers the study of a sample group only. A part of the population or unit.
It is less expensive and less time and effort consuming.
 Regular survey:
It is conducted after regular intervals. Generally, the government uses it to
obtain data about economics problems etc.
 Ad Hoc survey:
It is conducted for certain purpose and is undertaken once for all. Mostly,
is conducted for testing hypothesis, getting missing or new information.
 Primary survey:
It is conducted in order to acquire directly the relevant facts and
information. It is more reliable than secondary survey.
 Secondary survey:
It is conducted after the primary survey has been completed.
 Official survey:
It is conducted by government to serve general or specific information for
formulating plans and policies.
 Non official survey:
It is conducted by non-government persons or agency.
 First survey:
It is conducted subsequent to first survey. It is made for second or third
time and so on.
 Open survey:
It is also called public survey. The repetitive survey is publicly available.
It is of general importance.
 Confidential survey:
The result of the survey is not made public. Information is not revealed to
the common people.

 Public opinion:
It is conduced to know the views of the people in any kind of legalized
abortion, open prostitution, homosexual activities etc.
Interview
It is a technique of primary data collection. It is an oral method in which one
person asks another person questions designed to obtain answer pertinent to the
research problem. It is the most commonly used direct method in the study of
human behaviour. The interview is a face to face interpersonal role situation in
which one person asks another person being interviewed. The respondent
question designed to obtain answer pertinent to the purpose of the research
problem. The interview may be regarded as a systematic method by which one
person enters more or less imaginatively with the inner life of another who is
generally a comparative stranger to him. The purpose of interview is to find out
what’s in or on someone else’s mind.
Research interview schedule
Interviewing itself is an art. But planning and writing an interview schedule is
even more or so. The research interview schedule is a guideline which the
interviewer follows. As interviewer asks the question, s/he records the response.
A good interviewer will let the respondent do the most talking.
Formats or research interview schedule
There are three formats of interview schedule providing 3 kinds of information.
They are:
Fixed-alternative item interview
It offers the respondents a choice between two or more alternatives. The
responses are limited to stated alternatives. These items are called or full
questions. The commonest kind of fixed-alternative items is dichotomous. If asks
for yes-no, agree-disagree, and other two alternative items often a third alternate
I don’t know or undecided is added.
Open-end item
Open-end item is an interview schedule that lists only the main question. It
permits the respondent to answer the question in the way s/he likes. The contents
of the schedule are dictated by the research problems. They impose no other
restriction on the contents and the manner of respondents answer. Open-end
questions are more flexible. They have possibilities of depth.

Example: Do you have any contacts with the faculty outside of class?
If yes, how often is that?
Scale item
A scale is a set of verbal items to which an individual responds by expressing
degree of agreement or disagreement or some other mode or response. Scale items
have fixed alternatives and place the responding individual on some point on
Observation Method:(2)
Observation (watching what people do) would seem to be an obvious method of
carrying out research in psychology. However, there are different types of
observational methods and distinctions need to be made between:
1. Controlled Observations
2. Natural Observations
3. Participant Observations
In addition to the above categories observations can also be either overt/disclosed
(the participants know they are being studied) or covert/undisclosed (the research
keeps their real identity a secret from the research subjects, acting as a genuine
member of the group).
In general observations, are relatively cheap to carry out and few resources are
needed by the researcher. However, they can often be very time consuming and
longitudinal.
Controlled Observation
Controlled observations (usually a structured observation) are likely to be carried
out in a psychology laboratory. The researcher decides where the observation will
take place, at what time, with which participants, in what circumstances and uses
a standardised procedure. Participants are randomly allocated to each
independent variable group.
Rather than writing a detailed description of all behaviour observed, it is often
easier to code behaviour according to a previously agreed scale using a behaviour
schedule (i.e. conducting a structured observation).
The researcher systematically classifies the behaviour they observe into distinct
categories. Coding might involve numbers or letters to describe a characteristics,
or use of a scale to measure behaviour intensity. The categories on the schedule
are coded so that the data collected can be easily counted and turned into statistics.

For example, Mary Ainsworth used a behaviour schedule to study how infants
responded to brief periods of separation from their mothers. During the Strange
Situation procedure infant's interaction behaviours directed toward the mother
were measured, e.g.
1. Proximity and contacting seeking
2. Contact maintaining
3. Avoidance of proximity and contact
4. Resistance to contact and comforting
The observer noted down the behaviour displayed during 15 second intervals and
scored the behaviour for intensity on a scale of 1 to 7.
Sometimes the behaviour of participants is observed through a two-way mirror
or they are secretly filmed. This method was used by Albert Bandura to study
aggression in children (the Bobo doll studies).
A lot of research has been carried out in sleep laboratories as well. Here electrodes
are attached to the scalp of participants and what is observed are the changes in
electrical activity in the brain during sleep (the machine is called an
electroencephalogram – an EEG).
Controlled observations are usually overt as the researcher explains the research
aim to the group, so the participants know they are being observed. Controlled
observations are also usually non-participant as the researcher avoids any direct
contact with the group, keeping a distance (e.g. observing behind a two-way
mirror).

Strengths
1. Controlled observations can be easily replicated by other researchers by
using the same observation schedule. This means it is easy to test
for reliability.
2. The data obtained from structured observations is easier and quicker to
analyze as it is quantitative (i.e. numerical) - making this a less time
consuming method compared to naturalistic observations.
3. Controlled observations are fairly quick to conduct which means that
many observations can take place within a short amount of time. This
means a large sample can be obtained resulting in the findings being
representative and having the ability to be generalized to a large
population.
Limitations
1. Controlled observations can lack validity due to the Hawthorne
effect/demand characteristics. When participants know they are being
watched they may act differently.
Naturalistic Observation
Naturalistic observation (i.e. unstructured observation) involves studying the
spontaneous behaviour of participants in natural surroundings. The researcher
simply records what they see in whatever way they can.
Compared with controlled/structured methods it is like the difference between
studying wild animals in a zoo and studying them in their natural habitat.
With regard to human subjects Margaret Mead used this method to research the
way of life of different tribes living on islands in the South Pacific. Kathy Sylva
used it to study children at play by observing their behaviour in a playgroup in
Oxfordshire.
Strengths
1. By being able to observe the flow of behaviour in its own setting studies
have greater ecological validity.
2. Like case studies naturalistic observation is often used to generate new
ideas. Because it gives the researcher the opportunity to study the total
situation it often suggests avenues of enquiry not thought of before.

Limitations
1. These observations are often conducted on a micro (small) scale and may
lack a representative sample (biased in relation to age, gender, social class
or ethnicity). This may result in the findings lacking the ability to be
generalized to wider society.
2. Natural observations are less reliable as other variables cannot be
controlled. This makes it difficult for another researcher to repeat the study
in exactly the same way.
3. A further disadvantage is that the researcher needs to be trained to be
able to recognise aspects of a situation that are psychologically significant
and worth further attention.
4. With observations we do not have manipulations of variables(or control
over extraneous variables) which means cause and effect relationships
cannot be established.
Participant Observation
Participant observation is a variant of the above (natural observations) but here
the researcher joins in and becomes part of the group they are studying to get a
deeper insight into their lives. If it were research on animals we would now not
only be studying them in their natural habitat but be living alongside them as
well!
This approach was used by Leon Festinger in a famous study into a religious cult
who believed that the end of the world was about to occur. He joined the cult and
studied how they reacted when the prophecy did not come true.
Participant observations can be either cover or overt. Covert is where the study is
carried out 'under cover'. The researcher's real identity and purpose are kept
concealed from the group being studied.
The researcher takes a false identity and role, usually posing as a genuine member
of the group. On the other hand, overt is where the researcher reveals his or her
true identity and purpose to the group and asks permission to observe.
Limitations
1. It can be difficult to get time / privacy for recording. For example, with
covert observations researchers can’t take notes openly as this would blow
their cover. This means they have to wait until they are alone and reply on
their memory. This is a problem as they may forget details and are unlikely
to remember direct quotations.

2. If the researcher becomes too involved they may lose objectivity and
become bias. There is always the danger that we will “see” what we expect
(or want) to see. This is a problem as they could selectively report
information instead of noting everything they observe. Thus reducing
the validity of their data.
Recording of Data
With all observation studies an important decision the researcher has to make is
how to classify and record the data. Usually this will involve a method of
sampling. The three main sampling methods are:
1. Event sampling. The observer decides in advance what types of behaviour
(events) she is interested in and records all occurrences. All other types of
behaviour are ignored.
2. Time sampling. The observer decides in advance that observation will
take place only during specified time periods (e.g. 10 minutes every hour,
1 hour per day) and records the occurrence of the specified behaviour
during that period only.
3. Instantaneous (target time) sampling. The observer decides in advance
the pre-selected moments when observation will take place and records
what is happening at that instant. Everything happening before or after is
ignored.
Personal Interview(3)
The face-to-face contact between researcher and respondent is not equal in terms
of the potential quality of data that can be obtained. In the face-to-face interview
it is possible to record more than the verbal responses of the interviewee, which
are often superficial. When human beings communicate directly with each other
much more information is communicated between them. When two people face
one another, the dialogue is conducted on several levels. It goes beyond verbal
expression. The nature of words used, facial expressions and body language all
communicate what the other party means.
Objective
 Become aware of the different forms which personal interviews can take
 Learn how to structure both individual and group interviews
 Recognise the main difficulties encountered when conducting interviews,
and
 Understand the role of the moderator in focus group sessions.

Types of personal interview
The two main types of interviews conducted in marketing research are
structured and unstructured.
Unstructured informal interview
The unstructured informal interview is normally conducted as a preliminary step
in the research process to generate ideas/hypotheses about the subject being
investigated so that these might be tested later in the survey proper. Such
interviews are entirely informal and are not controlled by a specific set of detailed
questions. Rather the interviewer is guided by a pre-defined list of issues. These
interviews amount to an informal conversation about the subject.
Informal interviewing is not concerned with discovering 'how many' respondents
think in a particular way on an issue (this is what the final survey itself will
discover). The aim is to find out how people think and how they react to issues,
so that the ultimate survey questionnaire can be framed along the lines of thought
that will be most natural to respondents.
The respondent is encouraged to talk freely about the subject, but is kept to the
point on issues of interest to the researcher. The respondent is encouraged to
reveal everything that he/she feels and thinks about these points. The interviewer
must note (or tape-record) all remarks that may be relevant and pursue them until
he/she is satisfied that there is no more to be gained by further probing. Properly
conducted, informal interviews can give the researcher an accurate feel for the
subject to be surveyed. Focus groups, discussed later in this chapter, make use of
relatively unstructured interviews.
Structured standardised interview
With structured standardised interviews, the format is entirely different. A
structured interview follows a specific questionnaire and this research instrument
is usually used as the basis for most quantitative surveys. A standardised
structured questionnaire is administered where specific questions are asked in a
set order and in a set manner to ensure no variation between interviews.
Respondents' answers are recorded on a questionnaire form (usually with pre-
specified response formats) during the interview process, and the completed
questionnaires are most often analysed quantitatively. The structured interview
usually denies the interviewer the opportunity to either add or remove questions,
change their sequence or alter the wording of questions.

Depth interviews
Depth interviews are one-to-one encounters in which the interviewer makes use
of an unstructured or semi-structured set of issues/topics to guide the discussion.
The object of the exercises is to explore and uncover deep-seated emotions,
motivations and attitudes. They are most often employed when dealing with
sensitive matters and respondents are likely to give evasive or even misleading
answers when directly questioned. Most of the techniques used in the conduct of
depth interviews have been borrowed from the field of psychoanalysis. Depth
interview are usually only successful when conducted by a well trained and
highly skilled interviewer.
Other instances when depth interviewers can be particularly effective are: where
the study involves an investigation of complex behaviour or decision-making
processes; when the target respondents are difficult to gather together for group
interviewers (e.g. farmers, veterinary surgeons, haulage contractors, government
officials); and where the interviewee is prepared to become an informant only if
he/she is able to preserve his/her anonymity.
Dillon etal. believe that to be effective, the interviewer must adhere to six
fundamental rules. These are:
 he/she must avoid appearing superior or condescending and make use of
only familiar words
 he/she must put question indirectly and informatively
 he/she must remain detached and objective
 he/she must avoid questions and questions structure that encourage 'yes'
or 'no' answers
 he/she must probe until all relevant details, emotions and attitudes are
revealed
 he/she must provide an atmosphere that encourages the respondent to
speak freely, yet keeping the conservation focused on the issue(s) being
researched
Depth interviews involve a heavy time commitment, especially on the part of the
marketing researcher. Interview transcripts have to be painstakingly recovered, if
they are to be accurate, either from terse interview notes or from tape-recordings
of the interviews. This can take many hours of often laborious work. The
transcripts then have to be read and re-read, possibly several times, before the
researcher is able to begin the taxing process of analysing and interpreting the
data.

Types of personal interview
Conducting the interviews
It is essential, for both types of interview format, that the interviewer has a good
grasp of the study's objectives, and of the information that is to be collected. This
will enable 'probing' to elicit the right data required, and ensure all relevant issues
are covered. Furthermore, some respondents may ask why a particular question
was included in an interview, and it may be necessary for the interviewer to be
able to 'justify' particular questions.
In rural areas it is customary before embarking on a formal interviewing survey
to notify the relevant public authorities, e.g. village head, district union, etc. to
ensure co-operation from respondents. Sometimes individuals may refuse to co-
operate unless they are convinced that the interviewer has permission and
approval to conduct the survey from the recognised local authorities.
Before commencing on interviews it is as well for the interviewer to prepare what
he/she is going to say when he/she first meets a respondent. Decisions need to be
made as to whether the respondent is to be told who is sponsoring the study, the
purpose of the study, or how the data is to be used, and so on. These points need
to be decided beforehand to ensure that a 'standardised' approach is used for each
interview. Variations in approach style may lead to different types of response
from respondents and therefore variations in results. If suitable introductions are
prepared in advance, no time will be lost during the interview in lengthy
explanations, and a good impression can be created from the start.
Interview approach in the field: It is important that the interviewer keeps as
low a profile as possible in the rural setting. Interviewers should walk as much as
possible and in small numbers - two in a team is often best. If the research team
is large, it is advisable to divide the study area into a number of zones to avoid
duplicating efforts or interviewing the same respondents.
Once an individual who appears to be worth interviewing is spotted in the field,
it is best not to wander around indecisively creating suspicion. He/she should be
approached directly. However, one should avoid startling potential respondents
by running up to them and pulling out the questionnaire for interview. Blending
into the local context as much as possible is obviously the best strategy. One
should always be sensitive to the fact that most people may be suspicious of
outsiders.
The timing of the interview can be very important. One should be aware of the
daily schedule, seasonal activities, and work habits of potential respondents. For

example, if a farmer is irrigating and receives water only once a week for an hour,
he/she may not be interested in participating in an interview at that time.
Interview introduction:
The introduction to an interview is crucial. A good introduction can effectively
gain the respondent's co-operation and a good interview, but a bad introduction
could result in refusal to co-operate or biased responses.
Greeting: This should be made according to local custom.
Small
talk:
Being approached by a stranger will make the potential respondent
feel uncomfortable. It is necessary to help him/her feel at ease by
starting with polite small talk about the weather or crop conditions,
(in the case of a farmer) or about the health of the family and the
general economic climate in the case of non-farmers.
Overcoming apprehension: The approach of an interviewer is still an unfamiliar
experience to most people. Many people are suspicious of outsiders and
particularly interviewers. Some may think the interviewer is an 'official' who has
come to check up on them for taxes. Certainly many potential respondents will
fear that the information they give will be used against them at a later date, or that
the interviewer is trying to probe family secrets. To ensure cooperation it is
important to:
 Keep the atmosphere relaxed and informal.
It can be helpful if the interviewer plays down the fact that he/she wishes to
conduct a 'formal' interview. Respondents can be encouraged to think that the
interviewer is interested in conversation rather than interrogation.
 Explain why the interview is necessary.
The respondent should be given a brief background as to the nature and purpose
of the study. This will bring him/her into the interviewer's confidence.
 Stress the value/benefit of the study to the respondent
Respondents are more likely to co-operate if they think they will ultimately
benefit from the study. If one can indicate that as a result of the study it will be
possible to develop better and cheaper products for the respondent, then they
should be encouraged to co-operate.
 Appeal to the instincts of pride and vanity of the respondent

The respondent needs to be made to feel important. He/she needs to be made to
feel that the interviewer is particularly interested in his/her opinion because
he/she is the 'expert' and 'informed'.
Additional points that may help to put the respondent at ease could include:
Language: It is advisable that marketing researchers should adopt the language
of those from whom they hope to obtain information.
"... using local names for socio-economic characteristics, bio-physical
characteristics, lands, customs, time, intervals and measures".
Length of interview: The respondent can be assured that the interview will be
brief. It is unwise to be deceitful here, otherwise there is a danger that the
interview may be stopped mid-way by an angry respondent.
Confidentiality: The respondent can be assured that the interviewer will not
reveal the respondent's identity (and will use the data only in aggregate form) or
give the results to official organisations.
Closing interview: After all relevant topics have been covered or the
respondent's time exhausted, the conversation should be brought to an end. If the
weather is unfavourable (too hot or too wet) or the respondent seems pressed for
time it is best to prematurely stop the interview. The departure is best done
gracefully, naturally and not too abruptly. The business-like 'Got to go' departure
should be avoided. The respondent should be thanked for his/her time and given
the appropriate customary farewell.
Interview recording
All the best interviewing is useless if it has not been adequately recorded, so it is
important to ensure good recording conditions. In an open-ended interview it is
difficult to make notes on everything during the interview. The best approach in
team-work is to appoint a scribe, i.e. a person whose job it is to write everything
down. How long one waits before writing up full field-notes depends on the
setting, and the interviewer's personal style but it should be borne in mind that an
interviewer's memory is limited. It is surprising how facts, ideas and important
observations that one thinks one will never forget quickly slip away. Half of the
details from an interview can be forgotten within 24 hours, three-quarters can be
lost within 2 days and after this only skeletal notes can be salvaged. Jotted notes
will help prompt memory later, but it is best to write up interview notes while
they are still fresh in the interviewer's mind after the interview or at the end of the
interviewing day.

Use of tape-recorders: A tape recorder can often be useful. It enables the
interviewer to give THE respondent his/her full attention during the interview and
avoid the need to be constantly scribbling notes. It can also enable data to be left
until such time as analysis can be applied more rigorously and in a more leisurely
way. It should be borne in mind, however, that not everyone likes to be tape-
recorded. If taping is contemplated the respondents' permission should be sought
first.
Sources of error and bias
In personal interviews there are many ways in which 'errors' can be made by both
the respondent and the interviewer, and this can lead to 'bias' in the results. The
objective of the interviewer should be to minimise the likelihood of such bias
arising.
Respondent induced bias
Faulty memory: Some respondents may answer a question incorrectly simply
because they have a poor memory. The key to avoiding this problem is to steer
clear of questions requiring feats of memory. For example, questions such as,
"Can you tell me what your crop yield was four years ago?" should be avoided.
Other aspects of faulty memory that were mentioned in the previous chapter were
telescoping and creation.
Exaggeration and dishonesty: There can be a tendency on the part of some
respondents to exaggerate claims about their conditions and problems if they
think it will further their cause and lead to improvement in their well-being. The
interviewer must be alert to, and note any, inconsistencies arising. This is best
achieved by checking key pieces of information with a variety of sources.
Failure to answer questions correctly: If rapport is not developed sufficiently,
the respondent may be unwilling to respond or fail to give sufficient attention or
consideration to the questions asked, and if the respondent does not understand a
question properly he may give inappropriate answers. The interviewer needs to
ensure that the respondent fully understands the questions being asked and is
responding in the appropriate context.
Misunderstanding purpose of interview: Some respondents may perceive the
purpose of the survey to be a long-winded form of 'selling', particularly if the
interviewer is asking them what they think about a new product. Their comments,
therefore, about such issues as 'propensity to purchase' need to be looked at within
a context where they may be expecting to have to buy the product at some stage
and are trying to strike a hard bargain. To avoid such problems arising it is
important to carefully explain the objectives of the survey, the identity of the

interviewer and sponsor, and what is required of the respondent, prior to the
interview proper.
Influence of groups at interview: During interviews the presence of other
individuals is almost inevitable. Most of the time other family members or
neighbours will wish to join in the discussion. Such a situation has can have
important implications for the type of data obtained. The respondent may be
tempted to answer in a way that gives him/her credibility in the eyes of onlookers,
rather than giving a truthful reply. In circumstances where the presence of third
parties cannot be avoided, the interviewer must ensure as far as possible that the
answers being given are the honest opinions of the individual being interviewed.
The interviewer must again be alert to inconsistencies and closely observe and
monitor the way in which the respondent is reacting and interacting with those
around him.
Courtesy bias: In interview situations it is quite possible that one will come
across the problem of courtesy bias, i.e. the tendency for respondents to give
answers that they think the interviewer wants to hear, rather than what they really
feel. The respondents may not wish to be impolite or to offend the interviewer,
and may therefore endeavour to give 'polite' answers. Courtesy bias can be an
obstacle to obtaining useful and reliable data and therefore needs to be minimised.
Generally, however, the creation of a good interview environment and an
appropriate relationship between the interviewer and the respondent can help
avoid too much courtesy bias arising:
Bias induced by interviewer
It is also possible for the interviewer him or herself to introduce bias into an
interview, and this must be avoided at all costs.
Desire to help the respondent: The interviewer may become too sympathetic to
the problems and conditions of the respondent, and this can affect the conduct of,
and results obtained from, the interview. Objectivity must be retained at all times.
Failure to follow instructions in administering the questions: It is often
tempting for the interviewer to change the wording of a question or introduce
inflections in questions. This can affect the respondent's understanding and can
bias his/her replies. Particular problems may arise if the respondent does not
understand the question as stated and the interviewer tries to simplify the
question. The altered wording may constitute a different question. When
questions are open-ended, this can involve the interviewer in formulating probing
questions that go beyond the printed words. Unless the probes follow instructions
faithfully the potential for bias is great.

Reactions to responses: When respondents give answers, the interviewer must
be careful not to 'react.' A note of 'surprise' or 'disbelief may easily bias the
respondent's subsequent answers. Interviewers must respond with a uniform
polite interest only.
Focus group interviews
Focus group interviews are a survey research instrument which can be used in
addition to, or instead of, a personal interview approach. It has particular
advantages for use in qualitative research applications. The central feature of this
method of obtaining information from groups of people is that the interviewer
strives to keep the discussion led by a moderator focused upon the issue of
concern. The moderator behaves almost like a psycho-therapist who directs the
group towards the focus of the researcher. In doing so, the moderator speaks very
little, and encourages the group to generate the information required by
stimulating discussion through terse provocative statements.
Characteristics of focus group interviews
The groups of individuals (e.g. housewives, farmers, manufacturers, etc.) are
invited to attend an informal discussion. Usually between 6 and 8 participants are
involved and the discussion would last between 1 and 2 hours. Small groups tend
to lose the mutual stimulation among participants, whilst large groups can be
difficult to manage and may prevent some participants having the opportunity to
get fully involved in the discussion.
The researcher raises issues for discussion, following a 'guide list of topics' rather
than a structured questionnaire. The participants are encouraged to discuss the
issues amongst themselves and with the researcher in an informal and relaxed
environment. The researcher records comments made by the participants (usually
utilising a tape or video recorder). Figure 5.2 shows how this list of topics is
arrived at.
The process of developing a topic list for focus groups
In contrast to a personal interview survey, the number of interviews in a typical
group interview survey is very small, usually between 3 and 4 would be sufficient
for each type of respondent-sector (e.g. farmers or manufacturers). Generally
from the first interview on an unfamiliar subject the researcher will learn a great
deal. The second and third interviews will produce more information, but not all
of it

will not be new. By the fourth interview most of what is revealed will have been
covered before, and the diminishing returns involved would generally not justify
the cost of further groups.
The participants within a focus group are selected in such a way that they exhibit
a high degree of homogeneity with respect to either background, behaviour or
both. Consider, for example, a study carried out by a small African nation that is
looking for a niche market for a new range of sparkling wines. It is decided that,
as a first step, a series of focus groups be conducted. The researchers are keen to
ensure that each group comprises people who are similar in age and behaviour
with respect to wine consumption. Figure 5.3 depicts the kind of screening
questionnaire that the marketing researcher would use.
The first two questions will eliminate those who are likely to be too aware
of the focus group process and distracted from the research topic. Questions 3 and
4 prevent those whose experience of wine consumption is not sufficiently recent
from taking part. Question 5 would enable the researcher to allocate prospective
participants to homogeneous groups. Thus, for example, there may be a group
comprised entirely of people whose favourite wine is one of the sparkling wines.
Other groups would be made up of people who have never tried sparkling wine
and another may involve those who have tried and rejected sparkling wine.
Clearly, the line of questioning would be different in emphasis for each of these
groups. Question 6 also helps the researcher balance groups in terms of age
distribution or he/she can make sure that only people within a narrow age range
participate in a particular group. The seventh question allows the researchers to
keep to whatever male/female ratios are appropriate given the research topic.
One has a choice of three different types of venue for group interviews,
each having particular advantages and problems. Firstly, one could hold
interviews at or near farmers' or manufacturers' residences. Such a venue has the
advantage that the participants would feel they are on safe ground and may
therefore feel more secure to express candid opinions, and also the advantage that
the participants do not incur expense in attending the interview. However, such a
venue can be problematic to organise, costly for transportation if equipment is to
be demonstrated, and it can be difficult for the researcher to retain control over
the interviewing environment.

Figure 5.3 An example of a screening questionnaire
WINE CONSUMPTION FOCUS GROUP SCREENER
Hello, I am from Marketing Research Centre and we are conducting research among people who
enjoy drinking wine and 1 would like to ask you a few questions.
1. Do you or does anyone in your household work in any of the following professions: marketing
research, advertising, public relations, or in the production or distribution of wine?
Yes terminate and tally
No continue
2. Have you participated in a group discussion, survey, or been asked to test any products for market
research purposes in the past 6 months?
No continue
3. Have you purchased and/or consumed any wine during the past 3 months?
No continue
4. Are you currently under medical treatment which prevents you from drinking wine at the present
time?
No continue
5. Next I am going to read you a list of statements about drinking wine. Please tell me if any of the
following statements apply to yourself. (Circle the letters that appear alongside the statements that
apply to you).
a. I prefer sparkling wines to any other type.
b. I often drink sparkling wines although it is not my preferred type of wine.
c. I only occasionally drink sparkling wine.
d. I have tried sparkling wine and did not like it so I never drink it.
e. I have never tried sparkling wine.
6. Which of the following groups include your age?
under 18 terminate
18-24
25-29
30-39
40-49
50-59
60 and older terminate
7. Sex (by observation)
Male check quotas
Female check quotas

Secondly, one could select a 'neutral' location such as a government
agricultural research centre or a hotel. Again, here, one might avoid respondents'
fears of attending, but there are still the problems associated with organisation,
transportation of equipment, and the deterring cost involved for those participants
who have to travel to the venue.
Group discussions can be invaluable research instruments for investigating
why individuals behave in a particular way. They can be used to uncover motives,
attitudes, and opinions through observing and recording the way the individuals
interact in a group environment. Group discussions are used primarily to generate
in-depth qualitative information rather than quantitative data, and are generally
applied in the context of evaluating individuals' reactions to existing products or
new product/concept ideas.
Structuring a focus group session
This example assumes that the problem to hand involves a concept (or idea) for
a new product.
Group discussions are also useful as a cost-effective means of generating
background information and hypotheses on a particular subject prior to the launch
of a quantitative survey. In this respect group interviews can have advantages
over personal interviews in a number of ways:
Synergism: The combined effort of the group will produce a wider range of
information, insight, and ideas than will the accumulation of responses of a
number of individuals when these replies are secured in personal interviews.
Snowballing: A bandwagon effect often operates in that a comment by one
person triggers a chain of responses from other participants.
Stimulation: Usually after a brief introductory period the participants become
enthusiastic to express their ideas and feelings as the group begins to interact. In
a personal interview, the respondent may not be willing to expose his/her views
for fear of having to defend his/her view or fear of appearing 'unconcerned' or
'radical'. Like most animals, the human being feels safer psychologically - as well
physically - when he/she is in a group.
Spontaneity: Since no individual is required to answer any given question in a
group interview, the individual's responses can be more spontaneous, less
conventional, and should provide a more accurate picture of his position on some

issues. In short, respondents are able to speak when they have definite feelings
about a subject and not because a question requires an answer.
Serendipity: It is more often the case in a group interview than a personal
interview that unexpected responses or ideas are put forward by participants. The
group dynamics encourages ideas to develop more fully.
Specialisation: The group interview allows the use of a more highly trained, but
more expensive, interviewer since a number of individuals are being 'interviewed'
simultaneously.
Scientific scrutiny: It allows closer scrutiny in several ways: the session can be
observed by several observers. This allows some check on the consistency of the
interpretations. The session can be taped or even video-taped. Later detailed
examination of the recorded session allows the opportunity of additional insight
and also can help clear up points of disagreement among analysts with regard to
exactly what happened.
Figure 5.5 lists some of the main applications of focus groups in marketing
research.
Figure 5.5 Applications of focus groups
APPLICATIONS OF FOCUS
GROUPS
 New product development
 Positioning studies
 Usage studies
 Assessment of packaging
 Attitude and language studies
 Advertising/copy evaluations
 Promotion evaluations
 Idea generation
 Concept tests....

Problems with group interviews
While group interviews have many advantages as a research instrument for
market research it should be borne in mind that they also have inherent problems.
Careful planning and management is required to obtain the most value from
group-based surveys.
Qualitative data: The researcher cannot produce hard quantitative data or
conduct elaborate statistical analysis because of the usually small number of
participants involved in group surveys. It is unlikely that one will be able to
include a statistically representative sample of respondents from the population
being studied.
Analysis: Analysis of the dialogue produced by group interviews can be a
difficult and time- consuming process. This point was made earlier where the
time taken to create transcripts from brief notes or tape recordings can take many
tedious hours. Thereafter the researcher has to analyse and interpret these
transcripts.
Potential biases
There are many potential opportunities for bias to creep into the results of group
discussions:
 Some participants may feel they cannot give their true opinions due to the
psychological pressure on them arising from their concern as to what other
members of the group may think. Some may feel tempted to give opinions
that they feel will be respected by the group.
 The presence of one or two 'dominant' participants may repress the
opinions of others. Some may not feel confident about expressing an
opinion. Some may prefer to submit to the opinions of others rather than
cause conflict/argument to develop.
Comparisons across groups: When a number of group interviews are being
conducted, comparisons of the results between groups can be hampered if the
setting, mix of participants, and/or interviewer is varied. Different interviewers
may vary the way they ask questions and vary the order of questions in response
to the answers being given. Differences in the settings of different groups may
produce variability in the quality of results.
These potential problems should not be taken as reasons for avoiding using group
discussions. The advantages far outweigh the problems, and careful planning and
management will avoid many difficulties arising in the first place.

Role of the researcher/moderator in discussion group
The researcher organising the group discussion acts as a 'moderator' not an
interviewer. The purpose of the interview technique is to get others talking and
interacting amongst themselves, and does not involve an interviewer asking them
a pre-set series of questions. The role of the researcher is thus to moderate the
discussion, encouraging participants to talk, prompting the discussion in
appropriate directions to ensure all issues are covered, and changing the direction
of the discussion when a point is felt to have been sufficiently covered. The
moderator is also required to 'control' the group interaction to ensure that the
viewpoints of all participants are allowed to be expressed.
In every interview situation one will find three types of participant who will need
to be controlled:
The
Monopolist:
the participant who wants to do all the talking. The moderator
must allow him/her a say, but ensure that he/she is quietened
when others wish to express an opinion.
The Silent Shy: The participant who cannot bring himself to participate. Direct
questioning of such individuals is often necessary to produce full
co-operation and contribution.
The Silent
Aggressive:
The participant that has plenty to say, but believes he is no good
at articulating it. The moderator needs to probe his feelings and
have these discussed by the others in the group.
The moderator has to identify and minimise the effect of these types of
participant. By anticipating the likely behaviour of individuals, the moderator can
be in a better position to maintain continuity and an easy exchange of opinions
and thoughts between individuals.
Questions and prompts must be completely free of bias. The discussion must
consist of genuine opinions of the group participants and not 'assisted answers'.
The neutrality of the moderator must be maintained at all times. It is also
important to ensure that the interview atmosphere is not too artificial. In group
interviews which aim to uncover attitudes towards products, it is always helpful
to have the product concerned available (and, if possible, demonstrated or tried
by respondents) to elicit realistic and valid opinions.
It is important in the group interview situation that the moderator is not so
involved in writing/recording participants' comments that he cannot listen or react
to the discussion which ensues. For this reason it is recommended that group
interviews are tape-recorded (audio or visual, where possible). Subsequent

analysis can then be more comprehensive, more rigorous and can be conducted
at a more leisurely pace.
Due to the nature of group discussions and the number of participants involved,
the data obtained can only be qualitative. Analysis is problematic (particularly in
deciphering which participant said what) but appropriate qualitative techniques
are available and should always be used. Tape recordings of discussions should
be fully transcribed, reduced and processed, and their content analysed.
Constructing the interview schedule
The interview schedule has at least four distinct sections: the warm-up,
exploration of discussion points, the core discussion section and a summary.
Structuring an interview Schedule
The warm-up: This section has the objective of creating an atmosphere
conducive to an open an free-flowing discussion. One technique that can be used
to break down the initial bashfulness among group members who, in most
instances, are strangers to one another is to divide them into pairs and exchange
simple facts about themselves (e.g. their names, details of the families, place of
work, interests etc.). Each group member is then asked to introduce their
neighbour to the rest of the group.
The warm-up phase of the session then moves on to encourage the group
members to engage in a free-ranging discussion around the topic upon which the
discussion will eventually focus. For example, a municipal authority considering
establishing a new fruit and vegetable wholesale market positioned outside a
congested city centre would ultimately wish to determine what innovative
facilities might attract traders to use the new market which is less convenient to
them in terms of location. During the warm-up phase the moderator will direct
the discussion in such a way as to obtain general information on how participants
currently behave with respect to the topic, issue or phenomenon under
investigation. The emphasis is upon a description of current behaviour and
attitudes. For instance, the traders would be asked to describe their own modes of
operation within the wholesale market as well as those of fellow traders.
Exploration of discussion points: In this phase the discussion moves on to the
participants' attitudes, opinions and experiences of existing products, services (or
in this case facilities) and on to what they like and dislike about those
products/services. With reference to the wholesale markets example, at this stage
traders would be invited to comment on the advantages and disadvantages of the
facilities within which they currently operate.

Core discussion: This part of the group discussion focuses directly upon the
principal purpose of the research. The flow of the session moves on to the
participants' perceptions of new concepts, possible developments or innovations.
The wholesale traders, for instance, would be guided towards discussing peri-
urban wholesale markets and the kinds of facilities which might attract traders
like themselves. A common approach is to follow a sequence of first exploring
the ideas which participants generate themselves and then to solicit participants'
reactions to ideas preconceived by researchers, or their clients, about possible
future developments.
Summary:
The final phase of the focus groups session allows participants to reflect upon
the foregoing discussion and to add any views or information on the topic that
they may have previously forgotten or otherwise have omitted. A common
tactic is to conclude the session by inviting the group, as well as its individual
members, to "advise the manufacturer" (or whoever) on the issue at hand.
Telephone interviews: Synchronous communication of time
asynchronous communication of place(4)
Due to the asynchronous communication of place, one of the advantages of
telephone interviewing is the extended access to participants, compared to FtF
interviews. MANN and STEWART (2000) make a distinction in the following
categories:
 Wide geographical access. People from all over the globe can be interviewed—
of course if they have access to telephone or computer. FtF interviewing can be
very expensive and takes too much time.
 Hard to reach populations. It enables researchers to contact populations that might
be difficult to work with on an FtF basis for example mothers at home with small
children, shift workers, computer addicts and people with disabilities.
 Closed site access. It is a possible means of access to people on sites, which have
closed or limited access (such as hospitals religious communities, prisons, the
military, and cults).
 Sensitive accounts. Some personal issues are so sensitive that participants might
be reluctant to discuss them FtF with an interviewer.

 Access to dangerous or politically sensitive sites. With telephone, interviewers
can interview people living or working in war zones, or sites where diseases are
rife, without needing to grapple with the danger—and the bureaucracy—of
visiting the area.
Although the interviewer can interview people that are not easy to access, one of
the disadvantages of asynchronous communication of place by telephone is the
reduction of social cues. The interviewer does not see the interviewee, so body
language etc. can not be used as a source of extra information. But social cues as
voice and intonation are still available. Although social cues are reduced, enough
social cues remain for terminating a telephone interview without a problem.
Another disadvantage of asynchronous communication of place is that the
interviewer has no view on the situation in which the interviewee is situated.
Because of this the interviewer has lesser possibilities to create a good interview
ambience. FtF interviews can make more use of a standardisation of the situation.
Due to this lessened possibility to create a standardisation of the situation with
telephone an extra disadvantage is that the interviewee can stay "visible" for other
employees and managers in the organisation. As I experienced for example the
interviewee was called away by his manager, so the interview had to be stopped
abruptly.
As in FtF interviews synchronous communication of time implies that interviewer
and interviewee can directly react to what the other says. This also leads to the
advantage that the interviewee is more spontaneous in his response and does not
deliberate too long. But on the other hand, the interviewer has to concentrate
much more on the questions that need to be asked and the answers given.
Another advantage of synchronous communication of time concerning telephone
interviews is, as in FtF interviews, the interview can be tape recorded. Tape
recording a telephone interview depends on the equipment. A speakerphone is
recommended (BURKE & MILLER, 2001). As with FtF interview the telephone
interview is also time consuming due to the fact that the tape has to be transcribed.
Mail Survey
Mail survey is a data collection method in which questionnaires are mailed to
potential respondents who in turn fill and return them at their convenience. This
method has the following advantages:
 Less cost of data collection

 Less time of data collection .
 Wider coverage of population
 Better accuracy of data
 Absence of interviewer's bias.
But it has the following drawbacks.
 The identity of the respondents is not known to the interviewers.
 The questionnaire may be filled in by the assistant family members of the
respondent.
 Some respondents may not return filled-in questionnaires.
 There may be delay from the part of the respondents in returning the filled-
in questionnaires.
In mail survey, the researcher selects the required number of potential
respondents of the study from mailing list mailing panel provided by some
organizations. Then a carefully designed questionnaire is despatched to each of
the potential respondents. The mailing of questionnaires involves the following
tasks:
 Selecting the type of envelope
 Determining the mode of postage.
 Designing covering letter
 Deciding questionnaire length, content, layout and format
 Notification and follow-up details
 Type of incentive, if any, to be given to potential respondents.
In some cases, even before mailing the questionnaires, a notification is sent to the
respondents which makes them aware of the purpose of the study before hand.
After mailing the questionnaires, reminders are to be sent to the respondents at
predetermined intervals to speedup the matter.

If the response rate is very poor, then more reminders should be mailed to the
respondents. Inspite of this effort, if the response rate is low, to cope up with the
deficit number of respondents, either the personal interview or the telephone
interview may be used. As an alternative, a kind of extrapolation can be done
based on the responses of the respondents who replied very late. Here, these
delayed respondents are equated to non-response type respondents for the
purpose of extrapolation.
As an alternative to mail survey, with the emergence of communication facilities,
one can use either fax survey or Web survey. In a fax survey, the questionnaire is
sent to respondents through fax. In turn, the respondents are expected to send the
filled-in questionnaires through fax or mail. This method quickens the entire
process. The success of this method relies on the infrastructural facilities
available with the respondents. Hence, this is used only for the organizational
respondents for whom such facilities are available. In the sprit of the study,
organizations support its employees to use such fax facility. In a Web-based
survey, the questionnaire is posted on a secure Web site. The invitation to
participate in the Web survey would be posted on a company Web site which
receives high traffic from target customers. If a customer visits that particular
Web site and clicks that particular invitation banner, then the customer will be
connected to the secure Web site. where a detailed interview is conducted. After
finishing the interview, the customer is placed back to his/ her original point of
entry into this system. The moment, a customer finishes the interview in a secure
Web site, the researcher can view the results of the interview instantaneously at
his/her station. When compared to other methods, this has the provision of even
filtering nonsensical responses. It is considered to be the fastest method of data
collection. But, the units of the sampling frae are limited to the customers who
own computers or visits Web sites very often on rental machines.
Secondary sources
It provides secondary data. Secondary data are already gathered by others.
They are attained indirectly. The researcher doesn’t obtain them directly. They
are collected by some other researchers before and have been processed at least
once.

Types of secondary data
Internal secondary data
They are available from in-house source. The source like within the organization.
Sources of such data include representative's report, shipment records, accounting
data etc.
External secondary data
They are available from the sources outside the organization. Such sources
include published report, library, computer, data books. Etc.
Sampling(5)
The Basic Idea of Sampling
Survey sampling is really quite remarkable. In research we often want to know
certain characteristics of a large population, but we are almost never able to do a
complete census of it. So we draw a sample—a subset of the population—and
conduct research on that relatively small subset. Then we generalize the results,
with an allowance for sampling error, to the entire population from which the
sample was selected. How can this be justified?
The capacity to generalize sample results to an entire population is not inherent
in just any sample. If we interview people in a “convenience” sample—those
passing by on the street, for example—we cannot be confident that a census of
the population would yield similar results. To have confidence in generalizing
sample results to the whole population requires a “probability sample” of the
population. This chapter presents a relatively non-technical explanation of how
to draw a probability sample.
Key Principles of Probability Sampling
When planning to draw a sample, we must do several basic things:
1. Define carefully the population to be surveyed. Do we want to
generalize the sample result to a particular city? Or to an entire nation? Or
to members of a professional group or some other organization? It is
important to be clear about our intentions. Often it may not be realistic to
attempt to select a survey sample from the whole population we ideally
would like to study. In that case it is useful to distinguish between the
entire population of interest (e.g., all adults in the U.S.) and the population
we will actually attempt to survey (e.g., adults living in households in the
continental U.S., with a landline telephone in the home). The entire

population of interest is often referred to as the “target population,” and
the 2 more limited population actually to be surveyed is often referred to
as the “survey population.”1
2. Determine how to access the survey population (the sampling frame).
A welldefined population is only the starting point. To draw a sample from
it, we need to define a “sampling frame” that makes that population
concrete. Without a good frame, we cannot select a good sample. If some
persons or organizations in the survey population are not in the frame, they
cannot be selected. Assembling a sampling frame is often the most
difficult part of sampling. For example, the survey population may be
physicians in a certain state. This may seem welldefined, but how will we
reach them? Is there a list or directory available to us, perhaps from some
medical association? How complete is it?
3. Draw a sample by some random process. We must use a random
sampling method, in order to obtain results that represent the survey
population within a calculable margin of error. Selecting a few convenient
persons or organizations can be useful in qualitative research like focus
groups, in-depth interviews, or preliminary studies for pre-testing
questionnaires, but it cannot serve as the basis for estimating
characteristics of the population. Only random sampling allows
generalization of sample results to the whole population and construction
of confidence intervals around each result.
4. Know the probability (at least in relative terms) of selecting each
element of the population into the sample. Some random sampling
schemes include certain population elements (e.g., persons or
organizations) at a higher rate than others. For example, we might select
5% of the population in one region but only 1% in other regions. Knowing
the relative probabilities of selection for different elements allows the
construction of weights that enable us to analyze all parts of a sample
together.
The remainder of this chapter elaborates on and illustrates these
principles of probability sampling. The next two sections cover basic methods for
sampling at random from a sampling frame. We proceed to more complicated
designs in the sections that follow.

Assemble or identify the list from which the sample will be drawn
Once we have defined the survey population – that is, the persons or organizations
we want to survey—how do we find them? Is there a good list? Or one that is
“goodenough”? Lists are rarely perfect: common problems are omissions,
duplications, and inclusion of ineligible elements.
Sometimes information on population elements is found in more than one
file, and we must construct a comprehensive list before we can proceed. In
drawing a sample of schools, for instance, information on the geographic location
of the schools might be in one file, and that on academic performance scores in
another. In principle, a sampling frame would simply merge the two files. In
practice this may be complicated, if for example the two files use different school
identification codes, requiring a “crosswalk” file linking the corresponding codes
for a given school in the different files.
Dealing with incomplete lists
An incomplete list leads to non-coverage error – that is, a sample that does
not cover the whole survey population. If the proportion of population elements
missing from the list is small, perhaps 5% or less, we might not worry. Sampling
from such a list could bias results only slightly. Problems arise when the
proportion missing is quite large.
If an available list is incomplete, it is sometimes possible to improve it by
obtaining more information. Perhaps a second list can be combined with the
initial one. If resources to improve the list are not available, and if it is our only
practical alternative, we might redefine the survey population to fit the available
list. Suppose we initially hoped to draw a sample of all physicians in a state, but
only have access to a list of those in the medical association. That frame omits
those physicians who are not members of the association. If we cannot add non-
members to that frame, we should make it clear that our survey population
includes only those physicians who are members of the medical association. We
might justify making inferences from such a sample to the entire population of
physicians (the target population) by arguing that non-member physicians are not
very different from those on the list in regard to the variables to be measured. But
unless we have data to back that up, such arguments are conjectures resting on
substantive grounds – not statistical ones.

Duplicates on lists
Ideally a list includes every member of the survey population – but only
once. Some elements on a list may be duplicates, especially if a list was compiled
from different sources. If persons or organizations appear on a list more than
once, they could be selected more than once. Of course, if we select the same
element twice, we will eventually notice and adjust for that. The more serious
problem arises if we do not realize that an element selected only once had
duplicate entries on the frame. An element that appears twice on a list has double
the chance of being sampled compared to an element appearing only once, so
unrecognized duplication could bias the results. Such differences in selection
probabilities should be either eliminated or somehow taken into account (usually
by weighting) when calculating statistics that will be generalized to the survey
population.
The most straightforward approach is to eliminate duplicate listings from
a frame before drawing a sample. Lists available as computer files can be sorted
on any field that uniquely identifies elements—such as a person’s or
organization’s name, address, telephone number, or identification code.
Duplicate records should sort together, making it easier to identify and eliminate
them. Some duplicates will not be so easily isolated and eliminated, though,
possibly because of differences in spelling, or recordkeeping errors.
Alternately, we can check for duplicates after elements are selected. A
simple rule is to accept an element into the sample only when its first listing on
the frame is selected. This requires that we verify that every selected element is a
first listing, by examining the elements that precede the position of that selection
on the list. Selections of second or later listings are treated as ineligible entries
(discussed next). This procedure can be extended to cover multiple lists. We
predefine a certain ordering of the lists, and after selecting an element we check
to see that it was not listed earlier on the current list or on the list(s) preceding the
one from which the selection was made. This procedure requires that we check
only the selected elements for duplication (rather than all elements on the frame),
and that we check only the part of the list(s) preceding each selection.

Ineligible elements
Ineligible elements on a list present problems opposite to those posed by
an incomplete list. Ineligible entries are elements that are outside the defined
survey population. For example, a list of schools may contain both grade schools
and high schools, but the survey population may consist only of high schools.
Lists are often out of date, so they can contain ineligible elements—like schools
that have closed, or persons who have died.
It is best to delete ineligible elements that do not fit study criteria, if they
are easily identified. Nevertheless, ineligible records remaining on the frame do
not pose major problems. If a selected record is determined to be ineligible, we
simply discard it. One should not compensate by, for example, selecting the
element on the frame that follows an ineligible element. Such a rule could bias
the sample results, because elements immediately following ineligible ones
would have higher selection probabilities – their own probability plus that of the
immediately preceding ineligible element(s).
When a list includes ineligible entries, we must ensure that the sample
includes enough usable selections by anticipating the ineligibility rate and
sampling additional elements. If the target sample size is 500, for example, and
we expect that 20% of the elements on the frame are ineligible, selecting 500
elements would leave only 400 usable selections. To end up with 500, we should
select 500/(1-0.20)=625. If we anticipate further that only 70% of the eligible
selected elements (persons or organizations) will agree to participate in the
survey, we should increase the sample size even further to 625/0.70 = 893.
Indeed, once we decide on a certain target number of completed
interviews, it is usually necessary to make many more than that number of
selections, to compensate for anticipated losses due to ineligibles, duplicates,
refusals, language problems, and other issues. Such adjustments in sample
selection plans are an important part of sampling work.

Figure 5.1
Simple Random Sampling From a List
Want to select 2 out of 10 elements
Generate a few random numbers between 1 and 10:
8
4
7
6
6
List of elements Selected?
Element 1
Element 2
Element 3
Element 4 Yes
Element 5
Element 6
Element 7
Element 8 Yes
Element 9
Element 10
Formula (in Excel) for generating a random
number between 1 and 10:
=INT(RAND()*(10-1) + 1)

Basic Methods for Random Sampling from Lists
Selecting persons, organizations or other elements from a list is the simplest
and most straightforward sampling method. It illustrates the main points in
sampling and provides groundwork for more complex methods. Variations on the
basic theme exist, however, even for this simplest sample selection method.
Once the frame has been assembled, we can draw one or more samples.
Three commonly used sampling methods are simple random sampling,
systematic sampling, and selection with probability proportional to size.
Simple Random Sampling
Simple random sampling (SRS) is the standard basic method of sampling.
With SRS, each element on the list has the same selection probability, and
selections are made independently of one another. SRS serves as a baseline
against which other methods are evaluated.
Selection can be carried out either “with replacement” or “without
replacement.” To understand the terminology, think of selecting little numbered
balls from a big jar. If we put a ball back in the jar after selecting it, we could
select the same ball more than once. If we do not replace selected balls, we cannot
select the same ball more than once. A valid random sample can be drawn either
way. The statistical theory of random sampling is a little simpler if sampling is
done with replacement. In practice, however, we almost always prefer not to
select the same person or organization more than once, and therefore we usually
sample without replacement.
Figure 5.1 illustrates a very simple procedure for drawing simple random
samples. Suppose we would like to select 2 of the 10 elements in Figure 5.1 at
random. We could generate some independent random numbers between 1 and
10 using a spreadsheet, a computer program, or a table of random numbers. In
this example we generated (in order) 8, 4, 7, 6, and 6. The first random number
selects element #8 on the list, and the second selects element #4.
(Figure 5.1 about here) The element numbers could refer to the sequential
position of elements on the list, or to another unique identifier for each element,
so that each random number refers to no more than one element. If the element
numbering system has gaps, some random numbers might not correspond to any
element. In that case, we simply discard such a random number and move on to
the next one. In Figure 5.1, we generated more than two random numbers even
though we wanted only two selections, because we planned to select elements
without replacement. Since random numbers are usually generated

independently, some could be duplicates. (Indeed, the fourth and the fifth random
numbers are both 6.) If a random number is the same as an earlier one, we discard
it and move on to the next unique one.
Many lists used as sampling frames are available as computer files. In such
cases we can use a spreadsheet or a statistical program such as SPSS, SAS, or
Stata to select a simple random sample.
Systematic Random Sampling
Systematic sampling selects elements from a list by using a fixed selection
interval, calculated by dividing the number of elements on the list by the desired
number of selections. Randomness is introduced by choosing a random number
within the first interval to make the first selection. To make subsequent selections,
the interval is added successively to the preceding selection number.
For example, to select 20 elements from a list of 100, we use an interval of
100/20 = 5, and we select every 5th
element. To begin, we would take a random
number between 1 and 5, say 3. Then we would select elements 3, 8, 13, 18, and
so on up to 98. The random number should be obtained from a table of random
numbers or generated by a computer program, not a number we happened to
think of “at random.” Notice in this example that there are only five distinct
samples of elements that can be selected, corresponding to the five possible
random starts between 1 and 5. This simplicity makes the method easy to use, but
it has to be used with some care.
Systematic selection is used for many kinds of lists, but it is especially
convenient for sampling from lists that are not computerized and when records
are not numbered sequentially. One only has to estimate the number of entries on
the list, calculate the interval that will produce the desired sample size, generate
a random start, and then just count off the selections.
Systematic selection never draws the same element more than once (unless
a list has duplicates or occasionally when sampling is done with probability
proportional to size, to be discussed below). Moreover, a systematic sample is
always spread out over all parts of a list. For example, if our list is ordered
chronologically by the dates of transactions or records, such a sample will cover
the whole time period represented in the frame.
Systematic selection is relatively simple, and commonly used. At least two
potential complications can arise– the ordering of elements on the list, and dealing
with fractional intervals.

Order of the List
The ordering of elements within the list can pose the most important risk in
systematic sampling. The size of the fixed selection interval should not
correspond with any periodicity on the list. Suppose we are studying the
prevalence of different types of recreational activities, and we sample records by
systematic selection from a list that sequentially orders consecutive dates. If we
use an interval of 7 (or some multiple of 7), all dates in the sample would fall on
the same day of the week as the first selection. Since activity patterns vary across
days (Monday and Saturday activities are quite different for many), we would not
want a sample of dates consisting of only one day of the week. Any interval other
than a multiple of 7 would yield a good mix of days and provide a more
representative picture.
Periodicity is a particularly obvious example, but other, more subtle, issues
of ordering can also arise. Consider a list of persons ordered from youngest to
oldest. Depending on the size of the list and the interval size, different random
starts could produce samples with noticeably different age distributions. If the
interval spans multiple ages, the random start will make a difference: a low
random start will result in a younger sample, and a high one will produce an older
sample. On the other hand, if the interval is smaller than the number of persons
in the frame with any given age, the age distribution will not depend noticeably
on the random start. If the highest and lowest possible random starts would fall
on persons in substantively different age groups at the beginning and the end of
the frame, it would probably be best to order the frame by some other variable.
If the frame cannot be reordered and the order of the list is of concern, a simple
and effective approach is to change the random start as selection proceeds. With
an interval of 10 and a random start of 2, for example, our first selections would
be elements 2, 12, 22, 32, and so on. After reaching element 100, we could select
a new random start, say 8, selecting elements 108, 118, 128, 138, and so on, until
we change the random start again. This involves little more work than using a
single random start.
This point anticipates a subsequent discussion of “implicit stratification.” Often
a frame is deliberately sorted in a certain order to ensure that samples include all
parts of a distribution. Ordering persons by age and selecting systematically
ensures that we sample our “fair share” of older, middle-aged, and younger
persons without creating explicit strata. Samplers like to take advantage of
opportunities to stratify frames in such a simple manner. We must remain
sensitive to the possible impact of the random start on a systematic sample,
however, even when a list is ordered deliberately

Fractional Intervals
Fractional intervals are the other complication in systematic sampling. If
systematic selection is done by hand, it is easier to use a whole-number interval.
Suppose a list contains 9,560 elements and we want to select 200, so that the
interval is 9,560/200 = 47.8. A simple approach is to round fractional intervals.
Rounding up lowers the sample size and rounding down raises it. The calculated
interval of 47.8 in this example could be rounded up to 48, yielding 9,560/48 =
199 selections (for most random starts), or down to 47, leading to 9,560/47 = 203
or 204 selections (depending on the random start). Usually it does not matter if
the sample is a little larger or smaller, especially if we have to allow for losses
due to ineligibility and non-response.
If we really need to select a specific number of elements, Figure 5.2 illustrates
a procedure to do so, using a fractional interval. The procedure is as follows:
 Calculate the fractional interval. To select exactly 4 elements from a list of
10, use the interval 10/4 = 2.5.
 The random start should be a fractional number greater than 0 and less than
or equal to the interval. In Figure 5.2 the random start is 1.5. To obtain a
fractional random start between 0.1 and 2.5, one could pick a random
integer between 1 and 25 (10 times the interval), and divide by 10. For
example, the random integer 15 would yield 1.5.
 Add the interval repeatedly to the random start to generate a series of
selection numbers, retaining the decimal fractions, until a selection number
is beyond the end of the list. In the example, the series is 1.5, 4.0, 6.5, 9.0,
and 11.5.
 Truncate each selection number to a whole number by dropping its decimal
portion. The truncated selection numbers in the example are 1, 4, 6, 9, and
11. Numbers that truncate to 0 and those beyond the end of the list (like the
last number, 11) are discarded. Truncation is simple to do, and it yields the
correct probability of selection for all elements on the list (Kish, 1965, p.
116).
In the example, the interval between selections alternates between 2 and 3.
It is 3 between 1 and 4 and between 6 and 9, but it is 2 between 4 and 6. The
procedure yields exactly the desired number of selections.
Simple random sampling and systematic sampling are most commonly
used to select samples in which each element in the frame has the same selection
probability. Both techniques can also be applied to select elements with unequal

probabilities. We next cover the most common such situation, selection with
probability proportional to size
Figure 5.2
Systematic Random Sampling
with a Fractional Selection Interval
Number on the list: 10
Number to select: 4
Selection interval 2.5
Random start: 1.5
Selection series: With
fractions Truncated
1.5 1
4.0 4
6.5 6
9.0 9
(beyond end of list:) 11.5 11
Sampling with Probability Proportional to Size
Sampling with probability proportional to size (PPS) gives “larger” elements
on a list a greater chance of selection than “smaller” ones. Specifically, the
probability of selecting an element is directly proportional to its size. If one
element is twice as large as another, it will have double the chance of being
sampled.
Selecting with PPS is common in two-stage (or multi-stage) cluster samples
(discussed below), in which first-stage selections are areas or other clusters that
contain varying numbers of last-stage units (e.g. persons or households). First-
stage units (clusters) are selected with PPS, while last-stage units are usually
drawn with probability inversely proportional to size. PPS selection also is used
for single-stage samples of units that vary in size, such as schools or businesses.
In such cases, for a fixed number of selections, a PPS sample usually generates
more information than a sample selected with equal probability. The PPS sample
will tend to include more of the larger units than an equal probability sample in
which small and large units have the same chance of selection.

Preparing the frame
In order to select a PPS sample, each element in the frame must have an associated
“measure of size” (MOS). The size measure provides the basis for selecting some
elements with greater probability than others. Very often the MOS is a measure
of estimated size, so this procedure is sometimes called selection with probability
proportional to estimated size (PPES). However, we ignore that distinction and
refer to the method simply as PPS.
Figure 5.3 illustrates PPS selection. The bottom part of that figure lists 10
elements. The second column gives the measure of size associated with each
element, which ranges from 1 to 7. The MOS can be in any appropriate units –
population totals, sales figures, square footage, number of students, or whatever,
provided that the units are the same for all elements on the frame. The scale of
the units is less important than the relative size of the measure for different
elements.
(Figure 5.3 about here)
The third column in the figure shows the cumulative running total of the
MOS as we go down the list. The total of the MOSs for the 10 elements in the
frame is 40 units. We calculate a selection interval using this total if we draw a
PPS sample using systematic sampling. The fourth column in the figure shows
the selection range for each element—how the total range of 40 MOS units is
divided among the 10 elements in the frame. The width of each element’s
selection range corresponds to its MOS, larger elements having wider ranges than
smaller ones.
Methods of PPS selection
With selection ranges determined for the elements, we can select a sample.
Because samplers usually want to minimize the chance of selecting the same
element more than once, they often select PPS samples using systematic
selection. However, as for an equal probability sample, we can use either simple
random or systematic selection.
Simple random selection with PPS works in the same way as for equal
probability samples, except that random numbers refer to the selection range of
each element instead of its position on the list or some other identifier. The MOS
of an element determines the width of its selection interval and in turn its chances
of being selected. In Figure 5.3, selection ranges for all the elements together
extend from 1 to 40, so the generated random numbers should lie within that
range. Suppose we generate or look up the random number 5. That random
number selects the element with a selection range that includes 5: element #1,

with a selection range of 1 to 5. Because element #1’s selection range is five times
larger than element #3’s (of width 1), a randomly generated number will, on
average, select element #1 five times as often as element #3. Using MOSs to
determine selection ranges makes the probabilities of selection proportional to
the size of each element.
Systematic selection of a PPS sample works the same way as SRS
selection, except that the numbers for selections are generated systematically by
adding the selection interval to a random start, instead of independently. It is
important to understand that the selection interval must be based on the total
MOS. In the example shown in Figure 5.3 we want to select three elements, so
the interval is 40/3 = 13.3. We then generate a random start between 0.1 and 13.3,
say 5.5. Using the method for fractional intervals with truncation, we generate
three selection numbers, 5.5, 18.8, and 32.1, which are then truncated to 5, 18,
and 32, respectively. These numbers fall within the selection intervals of elements
#1, #5, and #9, so those three elements are selected. Once again, letting selection
intervals differ according to the MOS makes probabilities of selection
proportional to size.
If an element’s MOS exceeds the magnitude of the selection interval, it is
certain to be selected once and might even be selected more than once. Rather
than leaving such elements on a list for PPS selection, we often include them in
the sample automatically as “certainty selections” and remove them from the list
before sampling. In single-stage PPS samples, weights adjust for differences in
selection probabilities for certainty selections. For multi-stage samples, certainty
selections are treated as distinct strata, and subsamples of other units are drawn
from them.
It is also possible to leave large elements on a list for PPS selection when
drawing multi-stage samples, even though they must be selected at least once.
This may be the most convenient approach with long lists. If a large first-stage
element is selected twice, then the size of the second-stage subsample from it is
doubled.
Problems can also arise if some first-stage elements are too small to yield
sufficiently large second-stage samples. In such cases, groups of two or more
first-stage elements can be formed. Grouped units will be selected (or not)
together, with an MOS based on their combined MOSs. Kish (1965, pp. 244-245)
describes a clever objective method of linking small units after selection,
especially if they are too numerous to link by hand in advance.

We have described and illustrated the basic methods of random sampling
from a single list. The next sections discuss topics involving sample design rather
than the mechanics of drawing samples: these topics are stratification and
clustering.
Stratification
Stratification is a procedure whereby we divide the sampling frame for a
population into separate subpopulation frames, in order to draw a separate sample
from each subpopulation. In practice, stratification usually entails dividing a big
computer file up into smaller files, so that we can sample separately from each.
There are several good reasons for dividing the overall frame into subpopulation
frames. Unlike sample selection, however, this division is not based on some
random process. We first review some reasons for stratifying, and then we show
how to apply the random sampling methods of previous sections to the strata.
Reasons to stratify
Both theoretical and practical reasons underlie the technique of
stratification. The practical considerations are usually the more decisive. The two
most common reasons behind stratification are to facilitate making estimates3 for
subgroups and to increase sample precision (that is, to reduce the size of standard
errors and confidence intervals).
Separate reporting areas – proportionate sampling
Research studies often seek to obtain separate estimates for parts of the
population. For example, a sample of schools might need to produce results
separately for different geographic regions. A reasonably large simple random
sample would probably include some schools in all major regions, but it might
not (because of the random selection process) contain enough schools to make
adequately precise estimates for some of the smaller regions. Stratifying the
frame by region and drawing separate samples would allocate a proportionate
share of the total sample to each region.
Figure 5.4 illustrates stratification. There, a frame including 1800 schools
is divided into subpopulation frames for three regions. Then a separate sample is
drawn from each regional frame. Following the design in the second column, we
select the same proportion of schools from each region, with a sampling fraction,
f, of 0.10 or 10%. This is known as a “proportionate stratified sample.”

(Figure 5.4 about here)
Figure 5.4
Stratification
Proportionate Sampling Disproportionate
Sampling
STRATIFIEDFRAME
Region 1 (large)
School 1
School 2
School 3 f = 10% f = 5%
…
School 1000
Region 2 (small)
School 1
School 2
School 3 f = 10% f = 15%
…
School 300
Region 3 (medium)
School 1
School 2
School 3 f = 10% f = 10%
…
School 500
A proportionate stratified sample design ensures that each stratum (here, region)
will be represented in the sample in proportion to its size in the population–
including, in this case, exactly 10% of the schools in each region. A simple
random sample from the entire frame should yield approximately 10% of the
schools in each region, but the actual percentage in each region will vary from
sample to sample. We may not want to risk ending up with a smaller than
expected sample from a small stratum (like Region #2 in Figure 5.4). Stratifying
guarantees that we will have a certain number of cases in each stratum. If we must
report survey results separately for values of some variable, stratifying by that
variable is a good idea.
Stratifying requires that information on every element’s stratum be in the
frame before the sample is selected. We cannot stratify on variables that will only
be measured during the survey itself. Geography is often used for stratification
since geographic variables are usually known ahead of time for all elements in a
frame.

Oversampling some strata – disproportionate sampling
Stratifying by some variable such as region and selecting proportionately will
ensure that the sample includes a certain fraction of cases from each stratum, but
that may not be enough for some smaller strata. If we want good estimates for
certain subgroups (strata) of the population, we may need to allocate more than a
proportionate share of the sample to those strata. Having larger samples in those
strata will allow us to calculate results for those strata with more precision. This
type of sample is called a “disproportionate stratified sample.”
The design in the third column of Figure 5.4 illustrates disproportionate
stratification. The sampling fraction, f, differs across strata. In the figure, large
Region #1 (with 1,000 schools) is sampled at a low rate (5%), small Region #2
(300 schools) is sampled at a high rate (15%), while medium-sized Region #3
(500 schools) is sampled at an intermediate rate (10%). This increases the sample
size in the smaller strata, to provide enough cases to make reasonably good
within-stratum estimates of the variables of interest. Limited budgets may often
require reducing the sampling fraction in the bigger strata to compensate for
larger samples in smaller strata.
Although disproportionate sampling improves the precision of estimates
within the smaller strata, it generally reduces the precision of estimates for the
overall sample, compared to a proportionate sample of the same size. Because the
sample is no longer spread over all strata (regions) in proportion to the population,
we need to use weights when calculating statistics describing the whole sample.
These compensate for disproportionate selection, which results in having “too
many” cases from smaller strata and “not enough” cases from larger strata in the
sample. The consequence of having to use such weights is a reduction in precision
for the overall sample.4 Disproportionate selection involves a tradeoff between
overall precision and precision in smaller strata. This tradeoff is the price we pay
to have a single survey do multiple jobs. If we want reasonably good estimates
for small subgroups, and if we can sacrifice some precision in the estimates for
the population as a whole, then disproportionate sampling can be a good strategy.
Disproportionate sampling based on screening
Suppose we want to oversample certain ethnic groups in a population. If
our frame (e.g. a list of students or hospital patients) includes a race or ethnicity
code, we can create strata for the ethnic groups and sample some groups with
higher sampling fractions than others. However, if we must use another frame
(e.g., a list of telephone numbers or addresses) that lacks ethnicity data, we cannot
stratify ahead of time. Instead we must begin the interview with “screening”
questions, to ascertain the ethnicity of those selected, and then oversample by
continuing with the full interview at different rates for different groups. For
instance, we might interview all African Americans and Latinos in a sample, but
only half of those in other groups.

Fieldwork planning and supervision must control the implementation of
screening procedures like this “continue half of the time” rule. Our preference is
to control such selection rules by dividing the sample into random parts
(“replicates”) and then assigning a different selection rule to each part. For the
example in the preceding paragraph, we would divide the sample at random into
two halves. In one half, interviewers would attempt to complete the interview
with everyone. In the other half, they would attempt to interview only African
Americans and Latinos. African Americans and Latinos would then have double
the probability of selection into the overall sample, compared with the other
groups.
Reducing sampling error – “optimal allocation”
Often a major reason for stratifying is to attempt to increase the precision
of statistics by creating strata based on one or more variables that are correlated
with the primary variable we are trying to estimate. If the variation of our primary
variable within strata is less than its variation overall, proportionate stratification
will increase the precision of the estimate of our primary variable (see Groves et
al., 2009: pp. 114-120; Kalton, 1983: pp. 20-24).
Disproportionate stratification can sometimes be used to increase
precision even more, by using a strategy called “optimal allocation” (see the
Frankel and the Land and Zheng chapters in this volume). Optimal allocation is
a strategy for allocating more (than proportionate) cases to those strata with
relatively high variability in the primary variable of interest. Specifically, if data
collection costs are the same in all strata, the sampling fractions in the strata
should be proportional to the primary variable’s standard deviation in each
stratum. For instance, if the primary variable’s standard deviation is twice as large
in stratum #1 as in stratum #2, the sampling fraction in stratum #1 should be
double the sampling fraction in stratum #2.
If data collection costs differ across strata, optimal allocation also calls for
increasing the sampling fraction in low-cost strata, and decreasing it in more
expensive strata. More specifically, sampling fractions should be inversely
proportional to the square root of the cost per case in a stratum. For example, if
costs per case are four times greater in one stratum compared to a second, the
more expensive stratum should be sampled at half the rate.
The combined criteria of variability and cost can be summarized as:
fh = k * Sh / √Ch
where fh is the sampling fraction in stratum h, Sh is the standard deviation
in stratum h of the primary variable to be estimated, Ch is cost per element in that
stratum, and k is a constant used to scale the sampling fractions to produce the
target sample size.
When these criteria can be applied, sampling theory shows that confidence
intervals for means, percentages, and totals based on the overall sample will be
as small as possible for a given budget (Kish 1965, pp. 92-98; Kalton 1983, pp.
24-26).

Unfortunately we often lack the information necessary for applying those
optimization criteria. Unless estimates are available from prior studies, we may
not know the details of the primary variable’s distribution in advance, and will
not be able to estimate its standard deviation in various strata. Moreover, costs
per case are often little different for different parts of the frame.
And finally, one rarely conducts a whole survey just to obtain estimates
for a single variable. Surveys are almost always multi-purpose, and the optimal
sample allocation for one variable may not be optimal for some other variable of
equal importance. Proportionate stratified sampling, with the same sampling
fraction for all strata, is usually best – unless we have a good reason to oversample
a particular subgroup.
Nevertheless, optimal allocation is a very helpful heuristic for designing a
sample. Stratification is not simply a matter of convenience or a way of producing
reports for separate parts of the sample. The goal of good sample design is to
generate samples that produce results that are as precise as possible, and
stratification helps to do that. It is among the most useful tools available for
designing samples.
Methods of stratification
Stratification may be achieved explicitly by creating sub-frames, or implicitly
by exploiting the order of elements in a single frame. Some sample designs
combine the two.
Explicit stratification
In introducing stratification, we tacitly assumed that strata are created
explicitly, by physically dividing the overall frame into separate sub-frames or
files. Then a separate sample is drawn from each. This is the basic method of
stratification. No formulas dictate how many strata to create. From a practical
point of view, the number of strata required depends on the number of separate
subgroups for which results must be presented and on whether we can subdivide
the population based on a variable that is correlated with the variable(s) of
primary interest.
If we plan to use disproportionate stratified sampling, we must keep track
of the relative sampling fractions for strata, so that the strata can be weighted
appropriately to reflect the population. Then we will be able to use those weights
to combine the data from different strata when calculating results for the overall
sample, If, on the other hand, we do not plan to apply different sampling fractions
to different parts of the frame, we do not always need to stratify explicitly. A
simpler method, implicit stratification, is often sufficient.
Implicit stratification
Stratifying a frame before sample selection ensures that the sample is
distributed over the various segments of the population. “Implicit stratification”
accomplishes this without creating explicit strata for the various segments.

With implicit stratification, we sort the frame by some variable and then
select a systematic random sample. For example, to ensure that a sample of
addresses is spread over all regions of a state, we could first sort the address list
by zip code, and then select addresses with systematic sampling (not with SRS,
which would defeat the purpose of sorting). By selecting the sample in this
manner, we can be sure that the sample will include addresses from all of the
major geographic areas included in the frame.
Spreading the sample over the distribution of a variable may also improve
the precision of the statistics we are estimating. In a study of health variables, for
instance, sorting a frame of persons by their age will usually be helpful, since age
is highly correlated with health status. Controlling the age distribution in the
sample should therefore reduce the sampling error of estimated health statistics.
Stratifying implicitly is often more practical than stratifying explicitly.
Creating explicit strata for zip code groups, for example, could require a fair
amount of work: examining the distribution of elements in the frame by different
series of zip codes, deciding how many strata to create, and finally dividing the
frame into separate files. Sorting by zip code is much easier than going through
all those steps.
Another reason to stratify implicitly on a variable is that we might prefer
to base explicit strata on other variables. For example, we might need to stratify
a list of schools by type of school (public, private, charter) and by grade level.
Creating explicit strata for groups of zip codes would reduce our opportunity to
stratify on these other important variables. It may be preferable to sort on zip code
within explicit strata defined by the other variables. We comment further below
on this very useful combination of explicit and implicit stratification.
Implicit stratification is very useful and common, but it cannot achieve all
the goals of stratification. In particular, using disproportionate stratification to
oversample certain subgroups requires the creation of explicit strata so that a
larger sampling fraction can be applied in certain strata. Also, implicit
stratification cannot guarantee a specific number of selections in any particular
segment of the frame. Explicit strata should be created if this is important for
reporting results. Finally we should check for ordering effects in any systematic
sample. If the selection interval is large compared to the number of elements in
each category of the variable we are sorting on, high or low random starts could
produce samples that differ in non-random ways.
Combining explicit and implicit stratification
Stratification imposes some control on the sample selection process by
ensuring that a sample is spread over the distributions of certain variables in a
predictable way. In general, more strata yield better control. Consequently,
samplers tend to stratify the sampling frame as much as they can.
It is often desirable to stratify by more than one variable at the same time
(for instance, by creating a stratum for each school type within each region).

Explicit stratification offers the most control over sample selection, but a frame
can be divided into only so many categories at once. A solution is to create
explicit strata based on some variables, and then sort the frame on other variables
within each explicit stratum, to gain the benefit of some additional implicit
stratification. This combination of explicit and implicit stratification is common.
Explicit stratification is often used for major geographic areas such as
regions or states, especially if we know in advance that separate results will be
required for those segments of the population. If information for further
stratification is available in the frame, the simple device of sorting on one or more
variables and then selecting systematically within each explicit stratum takes
advantage of additional opportunities to attain the goals of stratification.
Cluster Sampling
When we sample, our eventual goal is to collect data on a specific type of
“element” (e.g., students). An “element sample” selects elements directly, as from
a list. So far, everything in this chapter has been about “element sampling.” Often,
however, we plan to sample elements only though groups of elements known as
“clusters,” usually to reduce costs. Such circumstances require “cluster
sampling.”
Figure 5.5 presents an example of a cluster design for sampling students in
a state. Often we cannot sample students (the elements) directly, because listing
them would be too costly, or because we wish to concentrate the sample in a
limited number of schools to reduce costs during data collection. So instead of
selecting students directly, we might select students within a sample of schools
(clusters). Within each selected school we will select some (or all) of the students.
In the figure, School #1 and School #3 are selected as clusters for further sampling
of students, but School #2 and School #4 are not.
Because the same groups of elements (like schools) could be used either as
strata or as clusters, the distinction between stratification and clustering can be
confusing. Strata and clusters differ in an important way. After dividing the
elements in a frame into strata, we subsequently sample elements from every
stratum. The point of grouping elements into clusters, however, is that we select
elements only from some of the clusters.

Figure 5.5
Cluster Sampling
Selected?
ELEMENTS WITHIN CLUSTERS
School 1 Yes
Student 1
Student 2
Student 3
…
Student 190
School 2 No
Student 1
Student 2
Student 3
…
Student 215
School 3 Yes
Student 1
Student 2
Student 3
…
Student 350
School 4 No
Student 1
Student 2
Student 3
…
Student 220
Effect of cluster sampling on precision
Cluster sampling usually increases the size of standard errors and confidence
intervals of the statistics we calculate from the sample results. Notice in Figure
5.5 that we will not sample any students in schools #2 and #4. Nevertheless, we
certainly will want to generalize results to all students in the state– not only to
students in those schools that happen to have been selected as clusters. Since
clusters are selected at random, the results can be generalized to the whole
population, but the sampling of clusters introduces a new level of uncertainty into
our results.
What if we had selected, by chance, other clusters into the sample – how
different would the study results be? How different are the clusters (schools) of

students from one another, in regard to the variables we want to study? If the
sampled schools are not very different, we can reasonably infer that our results
would have been similar had we sampled other schools instead. If, on the other
hand, the sampled schools turn out to be quite different from one another, our
uncertainty due to the sampling of clusters increases, which correspondingly
increases the width of confidence intervals for statistics based on the sample.
Campbell and Berbaum (this volume) cover methods of computing these
confidence intervals for cluster samples; here we try to provide an intuitive
understanding of the issues.
Comparing two extreme cases is informative. Consider a sample of 2,000
students within 100 schools, an average of 20 students in each. Suppose that some
characteristic (a certain test result, for instance) of all students within each school
is exactly the same, but the results for all sampled schools differ from one another.
In this case, all the information about test results in a school could have been
obtained from a single student in each school. Instead of sampling 2,000 different
students, we could have learned just as much from only 100 students, with one
student per school. So our cluster sample of 2,000 students is the equivalent of a
simple random sample of only 100 students. Calculating a confidence interval by
assuming that we have a simple random sample of 2,000 independent selections
overstates sample precision, because of the high (here, perfect) correlation
between elements within clusters. When elements within clusters are
homogeneous, sampling additional elements within clusters provides less
information than one might expect.
Now consider the other extreme case. Consider the same sample of 2,000
students within 100 schools. What if the average of some characteristic (e.g., a
certain test result) was exactly the same for all schools, though students within
schools differed from one another on that characteristic? Then there would be no
“cluster effect” on the results; it would have made no difference if we had
sampled 2,000 students from 100 schools, or 40 schools, or even 2 schools (if
they were large enough). In this ideal case, the cluster sample of 20 students
within each of 100 schools is equivalent to a simple random sample of 2,000
students from a statewide list. Both samples would have the same confidence
intervals. This is ideal: we conserve resources by dealing with only 100 schools,
but we obtain results as precise as those from a sample of 2,000 students spread
around the state.
In reality, of course, the effect of clustering almost always lies somewhere
between these two extremes. Results usually differ between clusters, and rarely
are all elements within clusters exactly the same. The more the variability
between clusters and the less variability among elements within clusters, the
lower the precision of sample statistics in a cluster sample.

Cluster effect and design effect
Quantifying the “cluster effect” can help us resolve this tradeoff.5
Sampling theory calls this effect the “coefficient of intraclass correlation” and
represents it by roh or the Greek letter ρ. Kish (1965, p. 161) clarifies by calling
it a “rate of homogeneity.” Like the familiar Pearson correlation coefficient, roh
is scaled to range between zero and one.
We can calculate roh only after a study is completed and standard errors
have been computed (as discussed in by Campbell and Berbaum, this volume).
When designing a cluster sample, however, it is useful to have a guess about the
probable size of roh, perhaps based on results of other studies that used similar
samples. Most research reports do not present values of roh itself, but they
sometimes report the “design effect,” from which we can calculate roh.
The design effect, deff, is the ratio of the variance of a statistic calculated
from a cluster sample (or any complex sample) to that of the same statistic
calculated from a simple random sample of the same size. For example, if the
variance of a statistic in a cluster sample is twice as large as its variance under
SRS, the design effect is 2.
The following important formula (Kish 1965, pp.161-164; Groves et al.
2009, pp. 109-112) gives the relationship between roh and deff, where b is the
average number of elements per cluster:
deff = 1 + roh(b -1 )
As the formula makes clear, we can reduce the design effect, and improve
precision, either by using clusters that have a low roh (low homogeneity), or by
designing a cluster sample with a low cluster size b. If roh is zero, the design
effect will be 1 regardless of the cluster size b. But if roh is high, even a relatively
small cluster size will result in a high deff.
Solving for roh in terms of deff and b yields:
roh = (deff – 1) / (b – 1)
If a study similar to ours reports design effects and provides the information
needed to calculate average cluster size (the total number of elements and the
number of clusters), we can calculate roh and use that information to design our
cluster sample. Or, if we have access to the data file of a prior study, we can
calculate deff and roh for ourselves, using newer versions of statistical packages
like Stata or SAS that calculate the correct variances and standard errors for
cluster samples.
In any case, to optimize the design of a cluster sample we must make
some guess about the value of roh that we expect to encounter. In some studies
roh is relatively small, like 0.05. A moderate roh is 0.10, and a high one is 0.20.
Notice that even a moderate roh of 0.10 will produce a deff of 2 if the average
cluster size is 11, so that the confidence intervals for the cluster sample will be

40% wider than those for a simple random sample of the same size. (If the
variance is two times larger, standard errors are larger by the factor √2 = 1.4)
Optimal cluster size
With an estimate of roh for the primary variable of interest in a sample
that uses a specific type of cluster design, we can begin to resolve the precision-
cost tradeoff described above. We also require information on the relative cost of
adding a new cluster versus collecting data from one more case in an already
selected cluster. An easy-to- apply formula gives the optimal cluster size, b, for a
given roh and relative cost (Kish, 1965, p. 269):
optimal b = √ ( relative cost * (1-roh)/roh )
For example, with a roh of 0.05 and a relative cost of 10, the optimal b is
√(10*19)=14 (rounded). This means that we should plan to sample about 14
elements per cluster. That degree of clustering should produce the narrowest
confidence intervals possible for a given budget, for those variables having a roh
of 0.05. Precision will be lower for variables with a higher roh, and greater for
those with a lower roh. Table 5.1 gives the optimal cluster size for several
combinations of relative cost and roh. Notice that only when relative cost is very
high or roh is very low do larger cluster sizes give the optimal result.
(Table 5.1 about here)
Different variables can and do have different values of roh, and therefore
different optimal cluster sizes. Moreover, we are often guessing about the size of
roh. In practice, then, the cluster size is often set using a compromise figure.
Nevertheless, the exercise of calculating optimum cluster size has heuristic value
for designing good samples, by requiring us to think systematically about the
tradeoffs. Reducing costs is not the sole object of cluster sampling. For any given
budget, we want a sample design that provides the most precise results possible.
Selecting clusters
Selecting clusters requires a frame of clusters, and uses the techniques
already described above for selecting individual elements from a frame. As a first
step, it can be advantageous to stratify clusters, to ensure that the selected clusters
are spread over the whole population. We may also plan to oversample certain
strata (types of clusters). Stratification of clusters could also reduce sampling
error, if the clusters can be grouped into strata likely to differ on the variables of
interest, since the standard errors for statistics will be computed based on
differences between clusters within the same stratum. Through such stratification,
we might mitigate some of the loss of precision that usually results from cluster
sampling.
Cluster sampling can be carried out either as a one-stage sample or as part
of a two-stage (or multi-stage) sample. An example of a one-stage cluster sample
is a sample of students within schools, in which we collect data on all students
within the selected schools. One-stage samples have large clusters, and usually
large design effects as well, so confidence intervals for most statistics will be
wider than one might expect for the number of students sampled.

Nevertheless, the type of data involved, and the cost structure for collecting
them, may justify sampling complete clusters. Suppose that the main cost of a
survey of students is the initial cost of contacting a school and getting access to
its records. After that, the marginal cost of data on additional students within that
school may be negligible, especially if the data are computerized. That is, the
relative cost of selecting an extra cluster (school), compared to that of collecting
data on an individual element (student), may be so high that it justifies large
clusters even with a high expected roh.
Two-stage cluster sampling
Often, however, we want to sample only some of the elements in the
selected clusters. Then we need a two-stage sample. A certain number of clusters
are selected in the first stage, and then elements are selected only within the
selected clusters in the second stage. Clusters are stepping stones providing
access to the elements within each cluster. Large-scale area probability samples
(Harter et al., this volume) are an important application of such designs. We
briefly discuss their use in smaller scale studies here.
In two-stage cluster sampling, one should decide on the selection method
for the two stages jointly. The simplest method is to select clusters with equal
probability at the first stage, and then to select elements, also with equal
probability, within the selected clusters. This method produces an equal-
probability sample that would not require sampling weights to be used in
analyses. For example, we could select 1% of the schools in a state and then
subselect 10% of the students in each selected school. The overall probability of
selection would be 1/100 x 1/10 = 1/1000 and would be the same for all students
in the state. However, this method yields little control over the total sample size.
If the selected clusters happen to be larger schools, the 10% subsamples will also
be large; if they happen to be small, the 10% subsamples will be correspondingly
small. Stratifying the schools by size could control the sample size to some extent,
but then we give up the opportunity to stratify using some other, perhaps more
interesting, variable(s).
A more efficient way of maintaining control over sample size is to sample
clusters with probability proportional to size (PPS), and then to subsample
elements within the selected clusters with probability inversely proportional to
size. Suppose we plan to select 5 elements per cluster. If the first-stage PPS
sample selects a cluster with a measure of size (MOS) of 100, we would
subsample elements within it with the fraction 5/100: either sampling elements at
random at the rate of 5%, or systematically sampling them using an interval of 20
and a random start between 1 and 20. Element samples within each of the other
selected clusters would be drawn using a fraction based on its respective MOS –
that is, 5 / MOSi . This procedure can be summarized with the following equation:
Probability = (a * MOSi / Total_MOS) * (5 / MOSi)
where MOSi is the measure of size for cluster i, and a is the number of clusters
selected.

Sampling with PPS at the first stage and inverse PPS at the second stage
produces an equal-probability sample. Notice that the MOSi in the equation
above then cancels out: the overall sampling fraction (or probability of selection)
is the same (i.e., 5a/Total_MOS) for all elements in all clusters. Therefore it is
not necessary to use sampling weights in analyses. The advantage of this method
is that total sample size is quite predictable, provided that the actual cluster sizes
found later during fieldwork are not very different from the MOSs for the clusters.
To ensure that the overall sample remains equal-probability, subsampling from
each selected cluster must be based on its MOS, not its actual number of elements
found later during fieldwork (otherwise the MOSi in the equation above will not
cancel out).
If we decide to select exactly 5 units in a cluster (instead of applying the
secondstage fraction 5/MOSi), our second-stage sampling fraction will be 5/Ni
where Ni is the actual number of units in the cluster found during fieldwork. Then
the overall probability of selection would be:
Probability = (a * MOSi / Total_MOS) * (5 / Ni).
Notice that MOSi and Ni do not cancel each other out of this equation, unless
they are exactly the same in every cluster (which is unlikely). The units selected
in cluster i would therefore be selected with probability proportional to the ratio
MOSi / Ni which could be different for every cluster. We should compensate for
such a departure from equal-probability sampling by using weights, a topic we
turn to next.
Weighting(5)
Several features of samples, even for small-scale studies, may require that
weights be used in data analysis. This section provides a brief summary of the
principles of weighting.
Weights give some cases more influence (weight) than others when
calculating statistics. Their basic purpose is to correct for biases in the data,
resulting from either the sample design or data collection procedures, that end up
producing “too many” sample elements from one population segment, and “not
enough” from some other segments. The sample designer should provide
instructions for creating basic sampling weights for any sample design other
than an equal-probability sample.
Relative weights versus expansion weights
One distinction cuts across all types of weights: that between relative
weights and expansion weights. This difference is simply a matter of scale.
Expansion weights scale the total weighted number of cases up to the size
of the population that the sample represents. For example, if we sampled 1% of

students from some list, each student would be given a weight of 100 (on
average). If that 1% sample yielded 500 students, the expansion weights would
project sample results up to the 50,000 students in the population. Expansion
weights are especially useful when presenting results to policymakers or other
publics interested in knowing not only what percentage of people have some
characteristic but also how many.
Relative weights scale the weighted number of cases to the actual size of
the sample, and they usually have a mean of 1. Some cases have relative weights
greater than 1, and others have relative weights less than 1, but the total weighted
number of cases is the same as the actual sample size. Data analyses and
presentations of results often use relative weights, to convey an approximate
sense of the precision of sample statistics. Using expansion weights could give
the misleading impression that statistics are based on tens of thousands of cases,
when in fact the sample may only include a few hundred.
Expansion and relative weights for different cases in a given sample
should have the same proportionality to one another. For example, one case might
have a relative weight of 1.5, and another a relative weight of 0.75. The
corresponding expansion weights might be 1,000 and 500 – in the same ratio of
2:1. When calculating descriptive statistics other than totals, using either type of
weight should give the same results. All weighting adjustments discussed below
can be used to construct both expansion weights and relative weights. Expansion
weights can readily be converted into relative weights by dividing them by the
mean of the expansion weights. To convert a relative weight into an expansion
weight, we must know the total population size or the sampling fraction.
Adjusting for selection probabilities
Section 5.4 introduced disproportionate stratified sampling, in which we
divide a sampling frame into several strata and sample the strata at different rates.
For instance, with a sampling frame divided into geographic regions, we might
sample smaller regions at higher rates than larger ones, to increase the sample
size and thus the precision of estimates in smaller regions.
It is crucial to keep track of the sampling rate used in each stratum. When
we combine results from different strata into estimates for the full population,
data from different strata must receive different weights to take into account the
oversampling of some strata and the undersampling of others. This first weighting
adjustment factor, applied to every case in the data file, is based on the inverse of
the sampling fraction in each case’s stratum:
Weight factor #1 = 1/fh

where fh is the sampling fraction for stratum h. If we sample elements in stratum
1 with the fraction 1/100, and those in stratum 2 with the fraction 5/100, the first
weight factor for the cases in stratum 1 will be 100, and the factor for stratum 2
will be 100/5 = 20.
Sometimes the information needed to adjust for different probabilities of
selection is only available after the fieldwork has been completed. For example,
in household samples of adults, usually only one adult is selected at random to be
interviewed within each sampled household. An adult who lives alone will always
be selected if we select her or his household. In comparison, the chance of
selecting an adult who lives with one other adult is only half as large. However,
we do not know the number of adults in the household until after it is selected
and contacted.
Differences in selection probabilities for households due to multiple
telephone numbers in random-digit-dialed telephone samples are another
common example. A household with two separate telephone numbers (regularly
answered and not used exclusively for a fax machine or a computer modem) has
twice the chance of selection as one with a single telephone number. Likewise, if
cell phone numbers as well as landline numbers are in the sampling frame, they
also affect the probability of selecting individuals. Someone who receives calls
via a cell phone has one chance to be called on the cell phone, and another to be
selected through the household’s landline. Whenever the elements in the survey
population are selected at different rates, we must compensate by using another
weighting factor. This adjustment requires that the survey obtain data on the
source of differences in selection probabilities (e.g. the number of adults in a
household, and the number of telephone numbers). This second weighting
adjustment factor is
Weight factor #2 = 1/pi ,
where pi reflects influences on selection probabilities for case i.
This weight factor can combine more than one factor affecting differential
selection probabilities. If, for example, a household has two telephone lines and
three eligible adults, the value of the combined value of pi for an adult in that
household is 2/3, the product of the telephone factor of 2 and the adults factor of
1/3. Since weight factor #2 is the inverse of pi , the second weighting adjustment
for such an adult would be 1/(2/3) = 3/2 = 1.5.

Non-response adjustments
Survey response rates are rarely 100%. Not adjusting for differential
nonresponse tacitly assumes that all non-respondents are similar to the average
respondent with respect to the variables measured. If non-response is
concentrated in certain subgroups, statistics for the sample will under-represent
those groups. Weighting adjustments for non-response compensate for this. Such
adjustments assume that nonrespondents in a subgroup are more like the
respondents in that subgroup than the average respondent. If the subgroup
classification is related to the variables we are estimating, a non-response
adjustment may improve our estimates.
To make a weighting adjustment for non-response, we must calculate a
separate response rate for each subgroup. In order to do that, we must know the
subgroup membership for all elements in the sample – non-respondents as well
as respondents. We cannot use a subgroup classification to adjust for non-
response if it becomes known only after fieldwork. For example, we usually do
not know the race or ethnicity of sampled persons before interviewing them, so
we cannot usually calculate separate response rates for race/ethnicity subgroups.
Sampling strata, therefore, are commonly used subgroup classifications for
purposes of non-response adjustment, since we know the stratum membership for
every sampled element.
Weighting adjustment factors for non-response are the inverse of a
subgroup’s response rate:
Non-response factor = 1/rrg
where rrg is the response rate for group g, expressed as a proportion, like
0.50 or 0.45.
If response rates are similar in all subgroups, this non-response adjustment
factor will also be similar for all subgroups, and it will have little or no impact on
the relative size of weights. It will, however, increase the weighted number of
cases. That can be important when creating expansion weights, to estimate the
number of elements in the population having a certain characteristic.
Putting the factors together
After calculating the factors that adjust for differences in probabilities of
selection and non-response, a weight variable is constructed by multiplying them
together. The value of the weight variable for case i in stratum h and subgroup g
in the sample is the product of the factors described above:

weightghi = (1/fh) * (1/pi) * (1/rrg)
where:
fh is the sampling fraction for elements in stratum h, and
pi is the probability factor for selecting element i, as learned during fieldwork,
and
rrg is the response rate for elements in group g.
This weight will be an expansion weight if the sampling fractions have
been expressed in absolute terms (like 1 / 10,000) instead of relative terms (for
example, that stratum 1 was sampled at double the rate of stratum 2). Relative
weights that yield the same number of weighted cases as the actual number of
completed cases in the data file (n) can be calculated by dividing the above-
calculated weightghi for each case by the mean of the weights:
relative weightghi = weightghi /(Σ(weightghi)/n)
This weight (either expansion or relative), adjusting for selection probabilities
and response rates, is sufficient for many studies. Sometimes, however, we want
to go further and adjust the sample distributions to match some criterion
distribution. We turn to that topic next.
Post-stratification weights
After the weighting adjustments for selection probabilities and response rates
have been made, noticeable differences between the distributions of certain
variables in the sample and in the population may still exist. One common
difference is for the percentage of women in the sample to exceed that in the
population. The response rate is generally a little higher among women than
among men, but we usually cannot adjust for differential non-response by gender
because the gender of respondents becomes known only during fieldwork.
Another reason that a sample distribution may differ from a criterion
distribution like the U.S. Census is that the sampling frame may not cover some
groups as well as others. Race and ethnic distributions could diverge from Census
figures because the sampling frame is less apt to include very low income
households (because they are less likely to have telephones, for instance), and
those missing households might be concentrated in particular ethnic groups.
Post-stratification weighting adjustments make the distributions of key
variables in the sample match Census figures or some other criterion distribution.
Matching the distributions of several different variables at once (e.g. gender, age,
education, race, and income) can be quite complicated. But post-stratification on

one or two variables, each with only a few categories, is not difficult. Simply
follow these steps:
A. Calculate the percentage of cases in the sample within the categories
you want to adjust. For example, we could use the percentage of
respondents in each cell of the cross-tabulation of race by gender.
The percentages must add up to 100%. Be sure to use the weight for
differential selection probabilities and non-response when
generating those percentages8 , and use at least a few decimal
places. Also, you should have at least about 20 cases in each cell;
otherwise, use fewer categories.
B. Find the corresponding percentages of the population in those same
categories, from Census data or some other criterion source. These
too must add up to 100%.
C. For each category in the distribution, divide its population
percentage (B) by its sample percentage (A). This ratio is the post-
stratification adjustment factor that applies to all cases in that
category. For example, making the gender distribution for the
sample match the Census distribution could require adjustment
factors like 1.1234 for males and 0.8902 for females. This would
have the effect of increasing the weighted number of males in the
sample, and decreasing the weighted number of females.
D. Finally, produce a new weight for each case, i, by multiplying the
previous weight variable by the post-stratification adjustment
appropriate to that case:
post-stratification weightghi = post-stratification adjustmenti *
weightghi
Since the post-stratification weight includes all the adjustments
incorporated into the previous weight variable, it would usually be used as the
primary weight variable when analyzing the data.

Table 5.1
Optimum Cluster Size
Roh
0.01 0.02 0.05 0.10 0.15 0.20
Relative Cost
1 10 7 4 3 2 2
2 14 10 6 4 3 3
3 17 12 8 5 4 3
4 20 14 9 6 5 4
5 22 16 10 7 5 4
6 24 17 11 7 6 5
7 26 19 12 8 6 5 8
28 20 12 8 7 6 9
30 21 13 9 7 6
10 31 22 14 9 8 6
11 33 23 14 10 8 7
12 34 24 15 10 8 7
13 36 25 16 11 9 7
14 37 26 16 11 9 7
15 39 27 17 12 9 8
20 44 31 19 13 11 9
50 70 49 31 21 17 14
100 99 70 44 30 24 20
500 222 157 97 67 53 45
1000 315 221 138 95 75 63
1500 385 271 169 116 92 77
For example: If roh is .05 and the relative cost is 10, the optimal cluster size is 14.
Simple cost model: Total Cost = a * (cost per cluster) + n * (cost per case) where a = number
of clusters, and n = number of interviews or cases
Relative cost = (cost per cluster) / (cost per case)
Optimal cluster size = sqrt( (relative cost) * (1 - roh)/roh)

Measurement & scaling Technique(6)
Measurement scales in Research Methodology are used to categorize and/or
quantify variables. From what has been stated above, we can write that scales of
measurement can be considered in terms of their mathematical properties. The
most widely used classification of measurement scales are:
 nominal scale
 ordinal scale
 interval scale and
 ratio scale.
Nominal scale: Nominal scale is simply a system of assigning number symbols
to events in order to label them. The usual example of this is the assignment of
numbers of basketball players in order to identify them. Such numbers cannot be
considered to be associated with an ordered scale for their order is of no
consequence; the numbers are just convenient labels for the particular class of
events and as such have no quantitative value. Nominal scales provide
convenient ways of keeping track of people, objects and events. One cannot do
much with the numbers involved. For example, one cannot usefully average the
numbers on the back of a group of football players and come up with a
meaningful value. Neither can one usefully compare the numbers assigned to one
group with the numbers assigned to another. The counting of members in each
group is the only possible arithmetic operation when a nominal scale is
employed. Accordingly, we are restricted to use mode as the measure of central
tendency. There is no generally used measure of dispersion for nominal scales.
Chi-square test is the most common test of statistical significance that can be
utilized, and for the measures of correlation, the contingency coefficient can be
worked out.
Nominal scale is the least powerful level of measurement. It indicates no
order or distance relationship and has no arithmetic origin. A nominal scale
simply describes differences between things by assigning them to categories.
Nominal data are, thus, counted data. The scale wastes any information that we
may have about varying degrees of attitude, skills, understandings, etc. In spite
of all this, nominal scales are still very useful and are widely used in surveys and
other ex-post-facto research when data are being classified by major sub-groups
of the population.

Ordinal scale: The lowest level of the ordered scale that is commonly used is
the ordinal scale. The ordinal scale places events in order, but there is no attempt
to make the intervals of the scale equal in terms of some rule. Rank orders
represent ordinal scales and are frequently used in research relating to qualitative
phenomena. A student’s rank in his graduation class involves the use of an
ordinal scale. One has to be very careful in making statement about scores based
on ordinal scales. For instance, if Ram’s position in his class is 10 and Mohan’s
position is 40, it cannot be said that Ram’s position is four times as good as that
of Mohan. The statement would make no sense at all. Ordinal scales only permit
the ranking of items from highest to lowest. Ordinal measures have no absolute
values, and the real differences between adjacent ranks may not be equal. All
that can be said is that one person is higher or lower on the scale than another,
but more precise comparisons cannot be made.
Thus, the use of an ordinal scale implies a statement of ‘greater than’ or
‘less than’ (an equality statement is also acceptable) without our being able to
state how much greater or less. The real difference between ranks 1 and 2 may
be more or less than the difference between ranks 5 and 6. Since the numbers of
this scale have only a rank meaning, the appropriate measure of central tendency
is the median. A percentile or quartile measure is used for measuring dispersion.
Correlations are restricted to various rank order methods. Measures of statistical
significance are restricted to the non-parametric methods.
Interval scale: In the case of interval scale, the intervals are adjusted in terms of
some rule that has been established as a basis for making the units equal. The
units are equal only in so far as one accepts the assumptions on which the rule is
based. Interval scales can have an arbitrary zero, but it is not possible to
determine for them what may be called an absolute zero or the unique origin.
The primary limitation of the interval scale is the lack of a true zero; it does not
have the capacity to measure the complete absence of a trait or characteristic.
The Fahrenheit scale is an example of an interval scale and shows similarities in
what one can and cannot do with it. One can say that an increase in temperature
from 30° to 40° involves the same increase in temperature as an increase from
60° to 70°, but one cannot say that the temperature of 60° is twice as warm as
the temperature of 30° because both numbers are dependent on the fact that the
zero on the scale is set arbitrarily at the temperature of the freezing point of water.
The ratio of the two temperatures, 30° and 60°, means nothing because zero is
an arbitrary point.
Interval scales provide more powerful measurement than ordinal scales for
interval scale also incorporates the concept of equality of interval. As such more

powerful statistical measures can be used with interval scales. Mean is the
appropriate measure of central tendency, while standard deviation is the most
widely used measure of dispersion. Product moment correlation techniques are
appropriate and the generally used tests for statistical significance are the ‘t’ test
and ‘F’ test.
Ratio scale: Ratio scales have an absolute or true zero of measurement. The term
‘absolute zero’ is not as precise as it was once believed to be. We can conceive
of an absolute zero of length and similarly we can conceive of an absolute zero
of time. For example, the zero point on a centimeter scale indicates the complete
absence of length or height. But an absolute zero of temperature is theoretically
unobtainable and it remains a concept existing only in the scientist’s mind. The
number of minor traffic-rule violations and the number of incorrect letters in a
page of type script represent scores on ratio scales. Both these scales have
absolute zeros and as such all minor traffic violations and all typing errors can
be assumed to be equal in significance. With ratio scales involved one can make
statements like “Jyoti’s” typing performance was twice as good as that of
“Reetu.” The ratio involved does have significance and facilitates a kind of
comparison which is not possible in case of an interval scale.
Ratio scale represents the actual amounts of variables. Measures of
physical dimensions such as weight, height, distance, etc. are examples.
Generally, all statistical techniques are usable with ratio scales and all
manipulations that one can carry out with real numbers can also be carried out
with ratio scale values. Multiplication and division can be used with this scale
but not with other scales mentioned above. Geometric and harmonic means can
be used as measures of central tendency and coefficients of variation may also
be calculated.
Thus, proceeding from the nominal scale (the least precise type of scale) to
ratio scale (the most precise), relevant information is obtained increasingly. If
the nature of the variables permits, the researcher should use the scale that
provides the most precise description. Researchers in physical sciences have the
advantage to describe variables in ratio scale form but the behavioural sciences
are generally limited to describe variables in interval scale form, a less precise
type of measurement.
Methods of Data Processing in Research(7)
Data processing is concerned with editing, coding, classifying, tabulating and
charting and diagramming research data. The essence of data processing in
research is data reduction. Data reduction involves winnowing out the irrelevant

from the relevant data and establishing order from chaos and giving shape to a
mass of data. Data processing in research consists of five important steps. They
are:
1. Editing of Data
Editing is the first step in data processing. Editing is the process of examining
the data collected in questionnaires/schedules to detect errors and omissions and
to see that they are corrected and the schedules are ready for tabulation. When the
whole data collection is over a final and a thorough check up is made. Mildred B.
Parten in his book points out that the editor is responsible for seeing that the data
are;
1. Accurate as possible,
2. Consistent with other facts secured,
3. Uniformly entered,
4. As complete as possible,
5. Acceptable for tabulation and arranged to facilitate coding tabulation.
There are different types of editing. They are:
1. Editing for quality asks the following questions: are the data forms
complete, are the data free of bias, are the recordings free of errors, are the
inconsistencies in responses within limits, are there evidences to show
dishonesty of enumerators or interviewers and are there any wanton
manipulation of data.
2. Editing for tabulation does certain accepted modification to data or even
rejecting certain pieces of data in order to facilitate tabulation. or instance,
extremely high or low value data item may be ignored or bracketed with
suitable class interval.
3. Field Editing is done by the enumerator. The schedule filled up by the
enumerator or the respondent might have some abbreviated writings,
illegible writings and the like. These are rectified by the enumerator. This
should be done soon after the enumeration or interview before the loss of
memory. The field editing should not extend to giving some guess data to
fill up omissions.
4. Central Editing is done by the researcher after getting all schedules or
questionnaires or forms from the enumerators or respondents. Obvious
errors can be corrected. For missed data or information, the editor may

substitute data or information by reviewing information provided by likely
placed other respondents. A definite inappropriate answer is removed and
“no answer” is entered when reasonable attempts to get the appropriate
answer fail to produce results.
Editors must keep in view the following points while performing their work:
1. They should be familiar with instructions given to the interviewers and
coders as well as with the editing instructions supplied to them for the
purpose,
2. While crossing out an original entry for one reason or another, they should
just draw a single line on it so that the same may remain legible,
3. They must make entries (if any) on the form in some distinctive color and
that too in a standardized form,
4. They should initial all answers which they change or supply,
5. Editor’s initials and the data of editing should be placed on each completed
form or schedule.
2. Coding of Data
Coding is necessary for efficient analysis and through it the several replies may
be reduced to a small number of classes which contain the critical information
required for analysis. Coding decisions should usually be taken at the designing
stage of the questionnaire. This makes it possible to pre-code the questionnaire
choices and which in turn is helpful for computer tabulation as one can straight
forward key punch from the original questionnaires. But in case of hand coding
some standard method may be used. One such standard method is to code in the
margin with a colored pencil. The other method can be to transcribe the data from
the questionnaire to a coding sheet. Whatever method is adopted, one should see
that coding errors are altogether eliminated or reduced to the minimum level.
Coding is the process/operation by which data/responses are organized into
classes/categories and numerals or other symbols are given to each item
according to the class in which it falls. In other words, coding involves two
important operations; (a) deciding the categories to be used and (b) allocating
individual answers to them. These categories should be appropriate to the
research problem, exhaustive of the data, mutually exclusive and uni – directional
Since the coding eliminates much of information in the raw data, it is important
that researchers design category sets carefully in order to utilize the available data
more fully.

The study of the responses is the first step in coding. In the case of pressing –
coded questions, coding begins at the preparation of interview schedules.
Secondly, coding frame is developed by listing the possible answers to each
question and assigning code numbers or symbols to each of them which are the
indicators used for coding. The coding frame is an outline of what is coded and
how it is to be coded. That is, a coding frame is an outline of what is coded and
how it is to be coded. That is, coding frame is a set of explicit rules and
conventions that are used to base classification of observations variable into
values which are which are transformed into numbers. Thirdly, after preparing
the sample frame the gradual process of fitting the answers to the questions must
be begun. Lastly, transcription is undertaken i.e., transferring of the information
from the schedules to a separate sheet called transcription sheet. Transcription
sheet is a large summary sheet which contain the answer/codes of all the
respondents. Transcription may not be necessary when only simple tables are
required and the number of respondents are few.
3. Classification of Data
Classification or categorization is the process of grouping the statistical data
under various understandable homogeneous groups for the purpose of convenient
interpretation. A uniformity of attributes is the basic criterion for classification;
and the grouping of data is made according to similarity. Classification becomes
necessary when there is a diversity in the data collected for meaningless for
meaningful presentation and analysis. However, it is meaningless in respect of
homogeneous data. A good classification should have the characteristics of
clarity, homogeneity, equality of scale, purposefulness and accuracy.
Objectives of Classification are below:
1. The complex scattered and haphazard data is organized into concise,
logical and intelligible form.
2. It is possible to make the characteristics of similarities and dis – similarities
clear.
3. Comparative studies is possible.
4. Understanding of the significance is made easier and thereby good deal of
human energy is saved.
5. Underlying unity amongst different items is made clear and expressed.
6. Data is so arranged that analysis and generalization becomes possible.

Classification is of two types, viz., quantitative classification, which is on the
basis of variables or quantity and qualitative classification, in which classification
according to attributes. The former is the way of, grouping the variables, say,
quantifying the variables in cohesive groups, while the latter groups the data on
the basis of attributes or qualities. Again, it may be multiple classification or
dichotomous classification. The former is the way of making many (more than
two) groups on the basis of some quality or attributes while the latter is the
classification into two groups on the basis of presence or absence of a certain
quality. Grouping the workers of a factory under various income (class intervals)
groups come under the multiple classification; and making two groups into skilled
workers and unskilled workers is the dichotomous classification. The tabular
form of such classification is known as statistical series, which may be inclusive
or exclusive.
4. Tabulation of Data
Tabulation is the process of summarizing raw data and displaying it in compact
form for further analysis. Therefore, preparing tables is a very important step.
Tabulation may be by hand, mechanical, or electronic. The choice is made largely
on the basis of the size and type of study, alternative costs, time pressures, and
the availability of computers, and computer programmes. If the number of
questionnaire is small, and their length short, hand tabulation is quite satisfactory.
Table may be divided into: (i) Frequency tables, (ii) Response tables, (iii)
Contingency tables, (iv) Uni-variate tables, (v) Bi-variate tables, (vi) Statistical
table and (vii) Time series tables.
Generally a research table has the following parts: (a) table number, (b) title of
the table, (c) caption (d) stub (row heading), (e) body, (f) head note, (g) foot note.
As a general rule the following steps are necessary in the preparation of table:
1. Title of table: The table should be first given a brief, simple and clear title
which may express the basis of classification.
2. Columns and rows: Each table should be prepared in just adequate
number of columns and rows.
3. Captions and stubs: The columns and rows should be given simple and
clear captions and stubs.
4. Ruling: Columns and rows should be divided by means of thin or thick
rulings.

5. Arrangement of items; Comparable figures should be arranged side by
side.
6. Deviations: These should be arranged in the column near the original data
so that their presence may easily be noted.
7. Size of columns: This should be according to the requirement.
8. Arrangements of items: This should be according to the problem.
9. Special emphasis: This can be done by writing important data in bold or
special letters.
10.Unit of measurement: The unit should be noted below the lines.
11.Approximation: This should also be noted below the title.
12.Foot – notes: These may be given below the table.
13.Total: Totals of each column and grand total should be in one line.
14.Source : Source of data must be given. For primary data, write primary
data.
It is always necessary to present facts in tabular form if they can be presented
more simply in the body of the text. Tabular presentation enables the reader to
follow quickly than textual presentation. A table should not merely repeat
information covered in the text. The same information should not, of course be
presented in tabular form and graphical form. Smaller and simpler tables may be
presented in the text while the large and complex table may be placed at the end
of the chapter or report.
5. Data Diagrams
Diagrams are charts and graphs used to present data. These facilitate getting the
attention of the reader more. These help presenting data more effectively.
Creative presentation of data is possible. The data diagrams classified into:
1. Charts: A chart is a diagrammatic form of data presentation. Bar charts,
rectangles, squares and circles can be used to present data. Bar charts are
uni-dimensional, while rectangular, squares and circles are two-
dimensional.
2. Graphs: The method of presenting numerical data in visual form is called
graph, A graph gives relationship between two variables by means of either
a curve or a straight line. Graphs may be divided into two categories. (1)
Graphs of Time Series and (2) Graphs of Frequency Distribution. In graphs

of time series one of the factors is time and other or others is / are the study
factors. Graphs on frequency show the distribution of by income, age, etc.
of executives and so on.
REFERENCES
1. Data Collection Methods [Internet]. Research-Methodology. [cited 2018 Oct
28]. Available from: https://research-methodology.net/research-
methods/data-collection/
2. Observation Methods in Research | Simply Psychology [Internet]. [cited
2018 Oct 29]. Available from:
https://www.simplypsychology.org/observation.html#part
3. Chapter 5: Personal Interviews [Internet]. [cited 2018 Oct 29]. Available
from: http://www.fao.org/docrep/W3241E/w3241e06.htm
4. Opdenakker R. Advantages and Disadvantages of Four Interview
Techniques in Qualitative Research. Forum Qual Sozialforschung Forum
Qual Soc Res [Internet]. 2006 Sep 30 [cited 2018 Oct 30];7(4). Available
from: http://www.qualitative-research.net/index.php/fqs/article/view/175
5. Piazza T. Fundamentals of Applied Sampling. :42.
6. Measurement Scales in Research Methodology - Measurement Scales in
Research Methodology (11478) [Internet]. Wisdom Jobs. [cited 2018 Nov
2]. Available from: https://www.wisdomjobs.com//e-university/research-
methodology-tutorial-355/measurement-scales-11478.html
7. Francis A. Methods of Data Processing in Research [Internet]. MBA
Knowledge Base. 2012 [cited 2018 Nov 2]. Available from:
https://www.mbaknol.com/research-methodology/methods-of-data-
processing-in-research/

Research methodology module-2

More Related Content

What's hot (20)

Similar to Research methodology module-2 (20)

More from Satyajit Behera (13)

Recently uploaded (20)

Research methodology module-2