Vous êtes sur la page 1sur 22


qxd 7/28/03 6:48 PM Page 435

Sampling, representativeness
and generalizability

Giampietro Gobo

Sampling in qualitative research has had a hard 4 only findings from a representative sample
time. On the one hand, it has been long neglected are automatically generalizable to the
by many qualitative researchers as a mere posi- population;
tivistic worry; on the other hand, it has been 5 (therefore) representativeness leads to
undervalued by survey researchers because of generalizability;
the use of non-probability methods. Qualitative 6 if research is not carried out on a representative
researchers often maintain that qualitative sample, its findings are not generalizable;
research does not need to sample or to consider 7 findings of qualitative researchers are not
seriously sampling issues, arguing that the most generalizable.
theoretically significant and important studies in
field research (accomplished by Gouldner, These sentences have become such common-
Dalton, Becker, Goffman, Garfinkel, Cicourel, places that they form an undisputed part of most
Sudnow and so on) were based on opportunistic researchers background assumptions. However,
samples. This argument may lead to the idea that survey researchers do not realize that in social
thinking about issues of sampling, representa- science (as I shall show later) probability
tiveness and generalizability is a waste of time. samples are achieved very rarely because of a
However, defining sampling units clearly before number of theoretical and practical reasons
choosing cases is essential in order to avoid (such as the nature of social or cultural objects,
messy and empirically shallow research. As a the lack of a population list for most variables,
matter of fact, in contemporary organizational the phenomenon of non-response, and so on).
research the problem of representativeness is a This implies that it is in survey research as well
constant and growing concern of many that issues of sampling, representativeness and
researchers. In addition, traditional qualitative generalizability need to be reframed in a new
researchers often forget that sampling is an perspective.
unavoidable consideration because it is, first of For decades, field and survey researchers
all, an everyday life activity deeply rooted in avoided each other and failed to respect the
thought, language and practice. scientific credentials of their respective works.
On the other hand, survey researchers have However, recently a conciliatory offer has been
often disqualified qualitative research because it made by some qualitative researchers, legitimat-
is based on non-probability samples. Their rea- ing both ways of doing research by stating that
soning, expressed in numerous research methods there are two kinds of generalizations: a general-
texts, lies in the following logic: ization about a specific group or population
(which aims at estimating the distribution in a
1 only surveys and polls use representative population) and a generalization about the nature
samples; of a process. Sampling requirements are com-
2 to be representative a sample needs to be pletely different in the two cases. The former
drawn up using probability methods; generalization, which is implemented in a survey
3 as qualitative research is based on non- or a poll, is based on statistical logic; the latter,
probability samples, its samples are not applied in field research, is based on the notion of
representative; theoretical sampling (Glaser and Strauss, 1967).
3110-Ch-28.qxd 7/28/03 6:48 PM Page 436


To make clearer the difference between these between these two activities. In the social
two kinds of generalization, some authors process that starts with the creation of a repre-
have called the second one transferability sentative sample and ends with the generaliza-
(Guba, 1981), analytical generalization (Yin, tion of findings, the researchers activity is
1984), extrapolation (Alasuutari, 1995) and constantly driven by biases and organizational
moderate generalization (Williams, 2002). obstacles.
While this offer was welcome and undoubtedly Issues of sampling, representativeness and
wise from a political point of view, it seems not generalizability need to be faced in a practical
to solve completely the theoretical and practical way, quite differently from the abstract way
problems involved in notions of representative- usually suggested in textbooks where methodo-
ness and generalizability. logical principles and rules stand on their own
In this chapter I shall try to follow an alterna- with only a weak relation to practice. On the
tive path, in search of more adequate answers to contrary, it is necessary to approach the whole
the two main linked questions usually raised in question of sampling sequentially because it is
social research: misleading to plan the whole strategy before
starting. In order to obtain representativeness,
1 How do we know to which extent our cases the sampling plan needs to exist in dialogue
(sample) are representative of all members of with field incidents, contingencies and dis-
the population from which the cases were coveries.
selected? Some of these issues will be discussed through
2 Can we generalize from a few cases (a sam- an (invented) dialogue between a teacher and a
ple) to a population without following a class. The content of this conversation is a col-
purely statistical logic? lection of questions, doubts, observations and
objections that I encountered several times when
To understand this complex matter, several pro- teaching research methods over a long period of
gressive analytical steps will be followed. First, time.
some (misleading) commonplaces on which sta-
tistical sampling is grounded will be decon-
structed, showing how probability samples are Setting: a professor (P) with students (S) in a class
rarely achievable in survey research. Second, the
P: Today well talk about how to sample.
theoretical legitimacy of non-probability samples
will be described. As a matter of fact, experi- S1: Are we going to taste some wine?
ments (reckoned by many scientists as the best Class: ha ha ha (laughing)
possible examples of scientific procedure) are P: What? Oh yeah, I get it. If I think it
not based on probability samples. Third, the over in fact its not a silly joke it
research practice of disciplines (such as palaeon- has some truth.
tology, ethology, biology, astronomy, anthropol- I told you in earlier weeks that
ogy, cognitive science, linguistics, and so on) there are many links between ordinary
whose scientific work is based on few cases, will language and scientific language, or
be considered to see if we can learn something between common sense and scientific
useful for qualitative social research. knowledge.This is just another case.
In order to answer the first question, we need
to go back to the core problem of representative-
ness: the variance of the phenomenon under
study. In social studies representativeness is SAMPLING IN EVERYDAY LIFE
often a practical matter, hardly ever an outcome
of automatic (statistical) procedures, which are
often useless (as well as difficult to implement) P: Before becoming a scientific procedure
because in social research we look at the social sampling probably was, and still is, a prac-
significance of samples instead of a statistical tical activity of daily life: the cook takes
logic. one macaroni out of the pot to know if
In order to answer the second question as
the pasta is ready to be served; the buyer
listed above, we need to distinguish two analyt-
ically different problems that are usually con- tastes to choose a wine or a cheese; the
fused with one another: the representativeness teacher asks a student some questions to
of samples and the generalizability of findings. assess his or her knowledge on the whole
Offering these issues as two sides of the syllabus. In everyday life social actors sam-
same coin neglects the existing social space ple constantly. If we refer to traditional
3110-Ch-28.qxd 7/28/03 6:48 PM Page 437


classifications of knowledge (see Frege,

1892; Peirce, 1902; Gomperz, 1905; Ogden
thought language
and Richards, 1923; Morris, 1938; Whorf, (world of concepts) (world of words)
1956, Popper, 1972; Marradi, 1994), which
divide it in three worlds (see Figure 28.1),
we notice that there are sampling activi-
ties in all three worlds.
With regard to thought, cognitive psy-
chologists Kahneman and Tversky showed
that human beings use particular ways of
reasoning called heuristics. One of these
is the heuristic of representativeness action
(world of events)
(Kahneman and Tversky, 1972; Tversky
and Kahneman, 1974), that is, peoples ten-
dency to generalize based upon a few
observed characteristics or events. With Figure 28.1 The three worlds of knowledge.
regard to the world of language, the same
function is performed, as Becker (1998)
has stated, by synecdoche, a rhetorical
figure in which we use a part for some- or grave goods specimens, and so on. As
thing to refer the listener or reader to the Becker states:
whole it belongs to (1998: 67). Finally, in
Archeologists and paleontologists have this
the world of action the representative problem to solve when they uncover the remnants
function is performed by the sample. The of a now-vanished society. They find some bones,
seller shows a sample of cloth to the cus- but not a whole skeleton; they find some cooking
tomer; in a paint shop the buyer skims equipment, but not the whole kitchen; they find
through the catalogue of shading colours some garbage, but not the stuff of which the
in order to choose a paint. garbage is the remains. They know that they are
lucky to have found the little they have, because
the world is not organized to make life easy for
archeologists. So they dont complain about having
FROM EVERYDAY lousy data. (Becker, 1998: 701)
LIFE TO SCIENCE S2: Yes, but still they do not know if these
samples are representative
P: Yes, please (seeing a student with a raised P: This concerns the nature of the out-
hand). come (the sample) and the associated
S2: But we dont know if they are represen- process for obtaining it, not the outcome
tative it could be a fake or damaged itself. Well face this issue in a few min-
piece of cloth] utes. For now the only thing you have to
P: Wait, wait you are going too fast. remember is that the process or procedure
Before arriving at this point we first employed to select cases (e.g. a random
need to define a sample. one) has nothing to do with the concept of
After having specified a population, the sampling itself. They are two separate
researcher decides if he or she will col- issues.
lect information on all its individuals
(complete enumeration) or on a sub-set
only (sample). So the sample is just a sub-
set of cases. Nevertheless, it might also be
one case only.Various scientific disciplines
use one case as sample: geologists with a P: One important question is that of why
fossil or a fragment of stone, palaeontol- we sample.
ogists with fragments of skull, bones or a Sampling was not always dominant in
skeleton, archaeologists with ornaments social sciences. Once it was common to
3110-Ch-28.qxd 7/28/03 6:48 PM Page 438


select the whole population of P: It means that a sample always has a

residents (a census), car owners, drawback.The idea of a sample was born
magazine subscribers, people in the from the material impossibility of exam-
phone directory, and so on. The main ining the whole population (see Pinto,
problem of this method was not the 1964: 623; Galtung, 1967: 51; Perrone,
difficulty of reaching the whole popula- 1977: 71). Also, Poppers refutability and
tion, but the difficulty of obtaining feed- falsification principles are based on such
back because many of them did not a recognition of the impossibility of the
send back mailed questionnaires.1 ideal of checking and controlling all
However, at that time researchers members of a specified population. We
trusted in the law of large numbers, that wish we could do it but we cannot.Also,
is, if you have collected a great number of when this is needed, as in a census,
cases (it doesnt matter how big is the some residents are always missing. For
related population) you are all right and example, the 1990 US census failed
do not have to worry. However, today to include about 10 million residents
we can state that those studies were not, and double counted about 6 million
in a strict sense, studies of populations others for a total net undercount of 4
but of samples, that is, sub-sets million people.2 So for the 2000 census
formed by people who participated the US Census Bureau decided to use
through their answers. scientifically proven statistical methods
S1: So whats new about the proposal about to account for non-respondents. A
sampling from the statisticians George paradox!
Gallup and Elmo Roper, the two bad S3: How can we overcome this problem?
guys who forecast correctly the 1936 P: There is no way to get out of it even
US presidential election? Samples were, though natural scientists believe that it is
even though unconsciously and against possible through the idea of the repre-
researchers intentions, used also before sentativeness of a sample.
1936! S1: Thank God! Finally we arrive at the point
P: Yes and no. You are right in pointing out of representativeness
that the change was not from the popula- P: Yes, but you will have many surprises.
tion method to the sampling method. The sampling procedures in use in
However, the change was in the nature of social sciences come from biology. This
samples. In the population studies, sam- was made possible by the positivistic cli-
ples were always an output of the failure of mate of the beginning of the twentieth
recruiting participants; in the subsequent century when social disciplines, in order
studies, samples were a rational and delib- to be scientific, imported language and
erate input. Researchers wanted to rule procedures from natural sciences. It was
the game (that is, to decide who to con- a bad period and its consequences
tact and how many) instead of waiting for unfortunately are still present. Anyway
participants to decide whether to answer the first problem, long and welldebated,
a mailed questionnaire. arises from this emulation of natural
S1: Thats it? science. Many authors have observed
P: No, of course. There were other major that biology or physics differ from social
advantages to this new method of science in the nature of their units of
recruiting participants, such as lower analysis (Goode and Hatt, 1952: 327): a
costs, saved time, and fewer human Russian chemist has no doubt that his or
resources needed to contact them. her atom of nitrogen acts in the same
However, in the whole business there way as the atom of nitrogen of his or her
is one enormous disadvantage, which Japanese or French colleague; however,
troubled statisticians for long time: no human being is similar to another
a sample will always be a sub-set of a one. So the presupposition of homo-
population. geneity, on which sampling in natural
S2: What does that mean? science is based, does not work in social
3110-Ch-28.qxd 7/28/03 6:48 PM Page 439


science because individuals are not But before describing those I need to
interchangeable in the same way as make a digression
atoms of nitrogen or little balls in a box. S1: Oh boy! How long it is taking to see the
S2: Is that all? end of this tunnel!
S3: It is an old criticism based on an abstract P: Hey hey hey Its not my fault if in the
argument and has been solved anyway by past methodologists have complicated
survey reseachers! things instead of solving them.
P: I will show later on that these problems My digression relates to the concept
have never been solved but only swept of probability. In many textbooks of
under the carpet. For the next few min- methodology it has been written that for
utes I want to stay on the abstract level, a sample to be representative it needs to
as you said, because statistical principles be based on probability methods. On this
are abstract and there are many other basis they have disqualified qualitative
theoretical pitfalls in treating social research.
sciences as if these were a branch of Usually such textbooks distinguish
mechanics. between probability samples (simple ran-
dom, systematic, proportional stratified,
non-proportional stratified, multistage
cluster) and non-probability samples
(haphazard or convenience, purposive
sampling, snowball) used by qualitative
research. As for the latter, Bailey (1978:
P: We previously made the important 115) says that the evident drawback of
distinction between a representative non-probability sampling is that, as the
sample and a sample. Let us go back to researcher does not know each subjects
the cook, the buyer and the teacher. probability of being selected, usually she
They focus on a part to infer informa- or he cannot state that his or her sample
tion about the whole because they think is representative of the population. This
that the part (the sample) is representa- reduces the possibility of generalizing the
tive of the whole (the population).While findings beyond the sample in use; non-
one macaroni drawn from the boiling pot probability sampling can only be ade-
may look like the others, answering quate if the researcher does not aim at
some questions correctly does not nec- generalizing his or her findings beyond
essarily mean the student knows the the sample.
whole syllabus. As a matter of fact it is S3: It seems to me quite reasonable what
crucial to differentiate between a repre- Bailey said.
sentative sample (a sub-set that is a P: Yes, it could be but only if it is possi-
miniature of the population) and a sample ble in practice to work also with a prob-
(a sub-set of cases the researcher ability sample. Lets see an example. You
focuses on during any research). While did this calculation several times in
one macaroni is a representative sample, maths or statistics classes: when rolling a
the students answers might be only a dice, what is the probability of the
sample. number two appearing in one throw
Strictly speaking, using representative only?
samples is plausible if there is no doubt S2: One in six.
that they mirror the characteristics of P: Right! To make this calculation you need
the population. to know in advance the whole range of
S3: Let us say that researchers aim at obtain- alternatives, that is, the population of the
ing generalizable findings: how can they number of the die. In social science it is
be sure that the sample they are focusing the same. Firstly, in a simple random
on is also representative? sample, to know the probability if each
P: There are two different ways (or criteria) person really has the same probability
to check this: inductive and deductive. of being included in the sample the
3110-Ch-28.qxd 7/28/03 6:48 PM Page 440


researcher has to know this probability It consists of choosing an individual (even

(no matter whether each individual will without a population list) at pre-set k intervals,
be chosen or discarded). Secondly, to obtaining random samples. There is only one
make sure someone is not missed during condition: the whole population must be gath-
the selection, the researcher needs a ered in one place only, so that the researcher can
choose at regular intervals the individuals who
complete list of the population, that is, to
will be included in the sample. This procedure is
know exactly the number of cases within in widespread use in areas such as the quality
the population. control of handmade products, exit polls, market
research, and surveys of visitors to museums,
cinemas and theatres. Even this is quite difficult
The first drawback: the list of population as in social sciences it is not always possible to
gather all the individuals with the characteristic
According to statistical criteria, we now realize (or characteristics) observed in one place to
we ca0nnot go on without the complete popula- select them. Therefore, when there is no popula-
tion list. It means that in order to know if the sam- tion list (or an equivalent that allows the
ple actually mirrors the characteristics of the researcher to reach the whole population theoreti-
population, you need to know empirically (for cally), we have to give up using probability sam-
these characteristics) both the sample and the pling procedures. As a matter of fact, each
population (Marradi, 1989: 53). As Bailey individual of the population cannot have the
(1978: 10814) says, random selection is possible same probability of being selected.
if we have accurate information about the distrib- The significance of this argument is even
ution of the statuses among the characteristics of greater when we think of other more dynamic
the population, which we try to re-create propor- units: as we do not know how characteristics
tionally in the sample using probability criteria. concerning emotions, attitudes, opinions and
As a matter of fact, finding a complete popula- behaviour are distributed in the population, aim-
tion list is often difficult and sometimes impossi- ing at statistical representativeness of samples is
ble because the list is not available, so that simple technically groundless. Is there a population list
random sampling is seldom applied in surveys. At on authoritative or unselfish behaviour?
registry offices, at town halls or in electoral regis- The argument in favour of the implementation
ters, complete lists of the population are available, of probability sampling in social sciences is the-
but they record only a small number of character- oretically weak because it works only when the
istics, mainly socio-demographic. For most of the population is known.4 But what do we do when
sub-sets of population or the characteristics soci- the population is infinite or unknown?
ology is interested in (emotions, attitudes or
behaviours), these lists are not available. As
Corbetta points out: how can we obtain a random Three problems in statistical sampling theory
sample of unemployed peopleif we do not
have the whole list of unemployed people? Statisticians state that probability sampling is
(1999: 332). Many unemployed people register applied only when the size of the population is
for work at employment agencies, yet not all of unknown. The method helps to find quantitative
them have registered. Thus: estimates (parameters) that are unknown in the
a population list is not available for most of studies of
population and to determine the sample size,
special groups of people. Let us think about studies of
even though we do not know the magnitude of
blue collar workers, the unemployed, people who work the population. It estimates the variance of a vari-
at home, artists, immigrants, housewives, pensioners, able (for example churchgoers) in the population.
sports team supporters, churchgoers, members of polit- This method comes from classical statistics and
ical movements, volunteers who do social work, elderly is based on sampling theory, involving the theo-
people who live alone, cohabitants and so on. (Corbetta, rem of the central limit. However, in social
1999: 333)3 sciences the application of this method involves


SYSTEMATIC SAMPLING VARIABLES First, such a sample size is valid only

when doing a univariate analysis (an analysis
focused on one variable only). If we want to
Survey researchers say that the alternative to carry out a bivariate (two variables) or multivariate
simple random sampling is systematic sampling. (three or more variables) analysis, to determine
3110-Ch-28.qxd 7/28/03 6:48 PM Page 441


the sample size we should make a calculation SOCIETY IS NOT RANDOM Another assumption of
that takes account of each variable. The sample statistics is that cases are distributed at random in
is therefore bound to increase, otherwise when the population. However, as Gilli (1971: 230ff.)
analysing the data some cells might lack cases. reminds us, society is not random. There is no
However, according to Sudman (1983: 157), this evidence that the sampling assumptions under-
occurs often: while it is rare that published stud- lying natural sciences (i.e. that cases are inter-
ies have an insufficient sample if considered in changeable because they are equal and
their totality, in multivariate analysis samples distributed at random in the population) works
have to be subdivided and in most of such subdi- well in social sciences. On the contrary, in
visions samples are inadequate (see also society almost nothing is at random: for exam-
Capecchi, 1972: 501). ple, there are social inequalities affecting
peoples position in the population.
THE PROBLEM OF THE NON-TRANSFERABILITY OF Probability sampling equalizes the chance of
THE LOGIC OF REPRESENTATIVENESS Secondly, every case, including the odd ones, turning up.
following the previous argument, representa- This seems, sociologically speaking, nonsense
tiveness has to be achieved on each variable of because not every person has the same relevance in
the research. Capecchi (1972: 501) exempli- society. We need a probability sample if we study
fied clearly that in an interview with 80 ques- a topic such as the political vote, where everybody
tions (consequently with 80 variables), counts: one head, one vote. However, when we
representativeness has to be achieved for each study political behaviours not everybody counts in
of them. However, if we follow the statistical the same way: leaders, some politicians and their
method recommended in sampling theory we advisers, lobbies, minorities, financial supporters
would need an enormous sample. But nobody count more than voters. If we study black minori-
does it. What survey researchers really do (in ties, we do not consider Martin Luther King or
order to construct a sample that claims to be a Malcolm X to be on the same level as others; how-
miniature of the population) is to estimate the ever, random sampling does this. More than thirty
variance of a few variables only, usually socio- years ago Capecchi said: we can state without any
demographic ones (such as gender, age, educa- doubt that no national or regional sociological
tion, political vote, residency and a few others) research based on interviews has ever used a rep-
of which we know truly the population parame- resentative sample that follows the rigorous conse-
ters from previous censuses. But what about the quences of statistical logic (1972: 53).
remaining 70 or so variables regarding atti-
tudes, behaviours, feelings where we do not
know their distribution in the population? As THE PHENOMENON OF
stated by Marradi (1989: 60), representative- NON-RESPONSE
ness cannot be transferred from one variable to
Let us look at a trivial example: if survey Following the logic of statistical theory it would
researchers aim to study sexual behaviour (a seem that, at least, polls on a few socio-
variable whose distribution in the population is demographic variables might be done with repre-
unknown), they usually construct a (claimed) sentative samples. If this is true in theory,
representative sample based upon the gender practice shows that another drawback always lies
variable, that is, they choose some men and some in wait: non-response. As Marradi points out:
women. They calculate the variance of sexual
the concept of random selection is theoretically very
behaviour by looking at the two values (male and
simple and, thanks to the ideal-typical image of the box,
female) of the gender variable. But this is not
quite clear to public opinion. However this clarity is
behaviour; it is gender! In other words, we can- misleadinghuman beings differ from balls in the box
not study behaviour indirectly through a repre- on two features: they are not at the researchers hand
sentative sample constructed on the basis of the and they are free to decide not to answer. (1989: 78)
gender variable. Indeed, the category sexual
behaviour has at least six sub-categories: We have to take into account the gap (which
heterosexual, homosexual, bisexual, abstinence, varies according to the research project) between
mostly heterosexual and mostly homosexual. the initial sample (all the individuals about
The idea according to which a sample that is whom we want to collect information) and the
representative of some socio-demographic charac- final sample (the cases we managed to get infor-
teristics is automatically representative of psy- mation on): the two sets may correspond but
chological features (or behaviours or opinions) usually a part of objects of the first sample is not
is highly problematic. gathered (Gasperoni and Marradi, 1996: 628).
3110-Ch-28.qxd 7/28/03 6:48 PM Page 442


The phenomenon of non-response is composed to other research methods: the statistical inference
of three different aspects: from sample to population. If the sample strays
from the probability model, nothing can be said
1 Lack of contact with the person who had been about its general representativeness, that is about
selected (because she or he has moved, is the fact that it truly reproduces all the character-
unknown at that address, not there at the istics of the population.
moment, ill, a prisoner, dead, or for other
2 Refusal to be interviewed. REPRESENTATIVENESS IN
3 Refusal to fill in the whole questionnaire if QUALITATIVE RESEARCH
there are questions considered too sensitive
by the respondent. In Italy these missing
answers may vary from 5 per cent up to S3: This detailed argument leads to a sort of
60 per cent, especially on questions concerning nihilism because it seems impossible to do
income (Gobo, 1997: 177). research with representative samples
P: Dont be too pessimistic. The aim was
Usually there is about a 30 per cent non-response
only to show that the criticisms by
with peaks of 5060 per cent. This would not be a
serious problem if the 70 per cent taking part in the survey methodologists of qualitative
poll were similar to the people who refused; in this research are exaggerated because in
case the representativeness of the sample would not most of their studies not even probabil-
be jeopardized. But in practice quite the reverse ity samples are what they are claimed to
often occurs: those who refuse to take part in the be. In addition, I rejected the forced
poll are not a random sub-set of the sample choice, as some scholars portray it,
(Marradi, 1989: 736). As Kish (1965: 558) states, between an (approximately) random
replacing non-responses is often a mistake because sample and a totally subjective one, or
substitutes are like those who accepted to be inter- between a partial probability sample and
viewed rather than those who refused to answer. a sample whose representativeness we
Non-responses may have systematically different
cannot be sure of. This false dilemma
statistical distributions compared to interviewed
people concerning their main socio-demographic does not take into account an alternative
characteristics (Castellano and Herzel, 1971, 302; path, which can be developed by thinking
Marradi, 1989; Brehm, 1993: 17). Non-response radically differently about the problem of
makes non-random a sample chosen at-random. representativeness in social science
This brings about serious distortions of the uni- S2: In which way?
variate characteristic values such as mean and vari- P: First, admitting that the most important
ance and of bivariate and multivariate coefficients. studies in qualitative research, which pro-
However, survey researchers have found a duced significant theories, were based on
remedy: weighting. As a matter of fact, in order non-probability samples: Gouldner (1954)
to obtain a representative sample, some criteria
observed one small gypsum extraction
have been developed to weight the answers of
the interviewed people who, on some socio- and refining factory located close to his
demographic characteristics, are similar to those university, so this was a haphazard sample;
who have not been reached by interviewers or Dalton (1959) did a study at Milo and
have refused the interview. However, this process Fruhling, two companies in a highly indus-
seems too artificial and quite arbitrary (Pitrone, trialized area in the US; Becker (1951)
1984: 14950; Marradi, 1989: 6878). Indeed, studied many dance musicians; anthropol-
statisticians avoid it and suggest doing the analy- ogist De Martino (1961) observed 21
sis on actually collected data only. people suffering from tarantism disease;
If according to survey methodologists ethnomethodologists Sudnow (1967) and
Converse and Schuman (1974: 40) the 20 Cicourel (1968) observed two hospitals
per cent is a reasonable amount of missing data
and two police districts respectively;
and does not jeopardize the representativeness of
the sample, it has to be stressed that we are far Goffman (1963) analysed various inter-
from working on representative samples. As actions between various people.
survey methodologists Groves and Lyberg S3: Does this mean that thinking about
(1988: 191) pointed out, non-response error threat- issues of sampling, representativeness
ens the surveys unique characteristic compared and generalizability is a waste of time?
3110-Ch-28.qxd 7/28/03 6:48 PM Page 443


P: That is a good point! Many qualitative In sociology and political science the
researchers think as you. main trend is choosing clearly defined
However, defining the sampling unit and easily detectable individual or
clearly (which comes before choosing collective units: persons, households,
the cases and thus picking the sample) is groups, associations, movements, parties,
important in order to avoid messy and institutions, organizations, regions and
empirically shallow research. During states. The consistency of these collec-
their analysis of some Finnish studies of tive subjects is not very real. In practice,
artists, Mitchell and Karttunen (1991) members of these groups are inter-
noticed different findings according to viewed individually: the head of the
the definition of artist employed by the family, the human resources manager, the
researchers. Some studies included in statistical department manager, and so
the category artist only those who on. This means that the sampling unit is
defined themselves as artists; in other different from the observational unit (i.e.
cases, only those who created durable the respondent). Only a focus group can
works of art were included; in other (at least in part) preserve the integrity of
cases, only those who were considered the collective subject. On the other
artists by the whole community; and in hand, as Galtung (1967: 37) stated,
yet other studies, those who were regis- choosing individuals implies an atomistic
tered by artists associations. A compari- view of society, whose structural ele-
son is therefore nonsense. ments are taken for granted or reckoned
S1: So we need to sample with accuracy to be mirrored in the individual (which
P: Yes, and fortunately the problem of means neglecting the sociological tradi-
representativeness is a constant and tion that gives priority to relations
growing concern of many young qualita- instead of individuals).
tive researchers. S1: What is the alternative?
S4: But how do we know the extent to P: Other, more dynamic units are wrongly
which our cases (the sample) are repre- neglected:
sentative of all members of the popula-
beliefs, attitudes, stereotypes,
tion from which the cases are selected?
S5: Can we generalize from cases-study
emotions, motivations;
findings to a population without following
behaviours, social relations, meetings,
a purely statistical logic?
interactions, ceremonies, rituals,
P: You have raised two important questions
that until now have not been solved by
cultural products (such as pictures,
qualitative methodologists.
paintings, movies, theatre plays, tele-
First of all, in order to produce
vision programmes);
persuasive answers, we need to keep
rules and social conventions;
these two questions separate. Putting
documents and texts (historical, lit-
them together belongs to an old way of
erary, journalistic);
framing the problem. Methodologists
situations and events (wars, elections).
too often equate representativeness
and generalizability, forgetting that the The researcher should focus her or
former characterizes samples only and his investigation on these kinds of units,
the latter concerns findings.We shall see not only because social processes are
this clearly later. more easily detectable and observable,
but also because these units allow a
more direct and deeper analysis of
the observed characteristics. When
Corsaro, for instance, tried to analyse
systematically childrens behaviour in
P: Let us start with sampling units and then kindergartens, the unit he chose was the
look at their representativeness. interactive episode. He defined this as
3110-Ch-28.qxd 7/28/03 6:48 PM Page 444


sequences of behavior which begin with S1: I can scarcely believe my ears
two actors in an ecological area and P: The conversation analyst Harvey Sacks
their try/tries to reach a shared meaning (1992, vol. 1: 485, quoted in Silverman,
of an ongoing or starting activity. The 2000: 109) reminds us of anthropologist
episodes end with a physical movement and linguist Benjamin Lee Whorf, who
of the actors who leave the area where managed to reconstruct Navajo gram-
the behavior had started (Corsaro, mar by interviewing extensively only one
1985: 24), disappearing from the stage. native Indian speaker. Usually grammars
Corsaro emphasizes that the most have low variance. It would probably
important thing is not his definition have been different if he wanted to study
(which may be questionable) but the how Navajos bring up children or how
need to create and employ explicitly they have fun. In his famous study,
units based on the researchers theoreti- anthropologist Clifford Geertz (1972)
cal and practical aims. attended 57 cockfights and ethnologist
Ernesto De Martino (1961) observed 21
people (socially labelled as) suffering
SOCIAL VERSUS from tarantism disease. They selected a
sample with so many or so few cases
probably because this reflected the vari-
ance.This is the only very important rule
P: After having chosen the sampling unit we for selecting the sample, and it is a pity
turn to the representativeness of samples. that many qualitative and quantitative
In order to evaluate this, it is crucial to researchers seem to forget it.
distinguish between logical versus social S2: Can you give us another example?
universes. P: Let us take my recent study of call
S1: What is this? centres. In Italy it has been estimated
P: The former belongs to statistics, the (this means that nobody knows the real
latter to social sciences. The conse- number) that there were 1020 call centres
quence is that the former aims at statisti- in 2000. I was interested in customer
cal significance, the latter aims at the relationship management (the sampling
social significance or the sociological unit) done by call takers in order to solve
relevance of the population.As Goode and clients requests. What is the variance of
Hatt (1952: 339) pointed out, if in a study this sampling unit? If I followed the tradi-
of marriage you extract a sample from tional way of sampling, focusing on the
the universe of marriages that happened call centre as organization, I would have
in New York in the last ten years, from a selected three types of call centres: pri-
statistical point of view this procedure is vate or marketing oriented; public, as for
correct because for statistics the actual example medical emergency dispatch
reality of such a universe is irrelevant. centres; and non-profit. However, if I am
However, from a sociological view this is interested in customer relationship
nonsense because such a universe simply management practices, to make a repre-
doesnt exist, it doesnt have any social sentative sample one has to consider
substance, it isnt a collective subject as a their variance in the population instead
social movement or an organization. It is of the variance of call centres. In other
a researchers invention only. words, the reference is not the variance
S2: So how can we proceed? of the call centre but the variance of
P: First, you look at the variance of the phe- customer relationship management
nomenon under the study. If it is high, practices within call centres. Reviewing
you need many cases in order to include literature, talking with experts and doing
in your sample each category or class of ethnographic research, I found that there
your phenomenon. If its variance is low are mainly four types of such practices
you need few cases; sometimes one case based on the nature and structure of tasks:
could be enough counselling, marketing, interviewing and
3110-Ch-28.qxd 7/28/03 6:48 PM Page 445


advertising. In addition, looking closely at the universe,is better when the population
each of them, while the three latter from which it is drawn is homogeneous, and
seem to have low variance, counselling it is lower when it is more heterogeneous.
has an higher variance due to the nature This is the key issue for legitimating quali-
of counselling itself, which ranges over tative research sampling.
different kinds of assistance: technical, We already know that an inductive
medical, psychotherapeutic, and so on. procedure used to check the correspon-
S1: This is quite complicated. dence between sample and population is
P: Yes, but it is a good way to obtian repre- almost impossible for practical reasons.
sentativeness. I did not have enough S4: So we have to change procedure and
resources to select four samples; I could follow the deductive one.
study only one type of practice using this P: This implies thinking through whether
approach.This meant that when attempt- the social phenomenon under study is
ing to generalize the findings, one will be pervasive. In other words, we may expect
conscious about the extent to which the there are not significant differences
sample is representative. In addition, this is between the population and the sample.
a way of getting away from many sociologi- For this reason a few cases, mirroring a
cal stereotypes. We sample leftists versus pervasive phenomenon in society, may be
rightists to know what they think about enough if its population is quite homoge-
immigrants. Why not do the reverse? neous. Following this reasoning is per-
Sample attitudes about immigrants and fectly acceptable in many qualitative
then look at where leftists and rightists studies. For example, anthropologist
are located. This procedure avoids the vincent Crapanzano (1980) studied
effects of conventional theories. Moroccan social relations through the
S3: Maybe I am too stupid with statistics but experience, of Tuhami, a furnace-workman.
I still cannot believe that a few cases, even Anthropologist Marcel Griaule (1948)
selected with the accuracy you describe, reconstructed the cosmology of the
can make a representative sample Dogon, a population from Mali, question-
P: There is some evidence for it. At the end ing a small group of informants only;
of the 1980s I did a study of interview- Bourdieus book (1993) on professions
ing, particularly on rituals and rhetorical is based on 50 interviews with police-
strategies used by one interviewer in his women, temporary workers, attorneys,
telephone contacts in order to obtain blue-collar workers, civil servants and
the consent for the interview, with ten the unemployed. In general, biographical
Italian respondents (Gobo, 1990, 2001). method relies on this methodological
I tape-recorded the verbal exchanges assumption (Javeau, 1987: 176ff.) Referring
and analysed them. Some years later I to conversation analysis, Harvey Sacks
read a similar study conducted in the (1984: 22, quoted in Silverman, 2000: 109)
US by Maynard and Schaeffer (1999) and has provocatively argued:tap into whom-
I found in it more or less the same soever, wheresoever and we get much
patterns. This happened not because I the same things.
was lucky but because interviewers
receive a similar training and use in their
work similar cognitive artefacts (scripts
and questionnaires). STUDYING SIMILARITY

PERVASIVENESS OF SOME Because probability sampling has many theoreti-

SOCIAL PHENOMENA cal problems and is also almost impossible to
apply in social research, the only practical and
sensible way is to work with non-probability
P: As I have already said, the precision of a samples. We have two broad criteria in building
sample, its being an accurate miniature of a non-probability sample depending on a
3110-Ch-28.qxd 7/28/03 6:48 PM Page 446


researchers interest: looking at similarities or psychiatric clinics; Cicourel and Boese (1972) on
differences among cases. The former criterion deaf childrens interpersonal communications;
maximizes the probability of extracting odd Garfinkel (1967) on achievement of sex status in
cases, but focuses on their similarities. The latter an intersexed person; Pollner and McDonald
selects only odd (deviant) cases. Wikler (1985) on interactions in a family with a
In the 1700s in biology there was a great con- mentally retarded child] or to explore subculture
troversy about this, associated with the two fun- or emergent or avant-garde phenomena that
damental aims of natural history: disposition and could become dominant or significant in the
denomination. This controversy has been mar- future [e.g. Festinger, et al. (1956) on Jehovahs
vellously described by Foucault (1966: 15079). Witnesses; Becker (1963) on marijuana smokers;
To accomplish these two tasks, naturalists of that Hebdige (1979) on style groups such as mods,
time used one of two different and opposite tech- punks and skinheads; Fielding (1981) on right-
niques: the System (Linnaeus) and the Method wing political movements]. The deviant case can
(Buffon, Adanson, Bonnet). On the one hand, be used also as a proof of refutability and falsifi-
Carolus Linnaeuss taxonomic technique was cation of a well-known and received theory, as
concerned with finding similarities among Rosenhan (1973) did against the psychiatric
animals and plants; on the other hand Buffon, theory that locates mental disease in a persons
criticizing Linnaeus, maintained that our general head, or De Martino (1961) did against the
ideas medical theory that considered the tarantism syn-
drome as being caused by the sting of a spider.
are relative to a continuous scale of objects of which we
The latter procedure (widespread in palaeontol-
can clearly perceive only the middle rungs and the
ogy) does not consider to what extent the studys
extremities increasingly flee from and escape our con-
finding is distributed in the population, but only
siderations. The more we increase the number of
divisions in the productions of nature, the closer we
notes that the phenomenon exists and must there-
shall approach to the true, since nothing really exists in fore be reviewed and understood theoretically.
nature except individuals, and since genera, orders, and
classes exist only in our imagination. (quoted in
Foucault, 1966, 1467) THEORETICAL SAMPLING
Bonnet seems no less resolute:
Theoretical sampling concerns both. The expres-
There are no leaps in nature: everything in it is gradu-
sion was coined by Glaser and Strauss (1967),
ated, shaded. If there were an empty space between any
and refined by Schatzman and Strauss (1973),
two beings, what reason would there be for proceeding
from the one to another? Above and below every human
Strauss (1987) and Strauss and Corbin (1990). In
being we find other beings that are close to him by some the following refinements the notion has been
characters and differentiated from him by others. expanded, producing different versions and loos-
ening it from its initial meaning. Theoretical
It is therefore always possible to discover inter- sampling means sampling on the basis of con-
mediate organisms, such as the polyp between cepts that have proven theoretical relevance to
the animal and the vegetable, the flying squirrel the evolving theory. [Relevance means] that con-
between the bird and the quadruped, the monkey cepts are deemed to be significant because they
between the quadruped and the man. Conse- are repeatedly present or notably absent when
quently, our divisions into species and classes comparing incident after incident, and are of suf-
are purely nominal; they represent no more ficient importance to be given the status of cate-
than means relative to our needs and to the gories (Strauss and Corbin, 1990: 176). In
limitations of our knowledge (1764, quoted in addition, as Mason writes, theoretical sampling
Foucault, 1966: 147). is concerned with constructing a sample which is
This historical digression is epistemologically meaningful theoretically because it builds in cer-
important, as in social research these two meth- tain characteristics or criteria which help to
ods have been largely used. Most studies and develop and test your theory and explanation
associated samples follow the former criterion in (1996: 94).
search of dominant characteristics of pervasive Theoretical sampling was and still is a tremen-
phenomena: biographical method, conversation dously important way of reasoning in order to
and discourse analysis, organizational ethno- select cases for your sample. However, at least in
graphy, and so on. However, there are some its first versions, it does not explicitly and clearly
important studies that have focused on deviant take account of the problem of representative-
cases in order to understand standard behaviour ness, but simply avoids it. In such a useful con-
[e.g. Goffman (1961) on ceremonies and rituals in cept there is nevertheless no suggestion on how
3110-Ch-28.qxd 7/28/03 6:48 PM Page 447


to consider the representativeness of selected Why did grievers and managers form cross-cliques?
categories. The selection of cases is driven only Why were staff personnel ambivalent toward line offi-
by their relevance for the phenomena under study cers? Why was there disruptive conflict between
and not by the need for representativeness. How, Maintenance and Operation? If people were awarded
then, should we address Denzins concern that posts because of specific fitness, why the disparity
the researcher has to show that these units, com- between their given and exercised influence? Why
among executives on the same formal level, were some
pared to all other similar phenomena, are repre-
distressed and some not? And why were there such
sentative (1971: 232), recently pointed out also
sharp differences in viewpoint and moral concern about
by Becker (1998: 67)? For past generations of
given events? What was the meaning of double talk
qualitative researchers the notion of theoretical about success as dependent on knowing people rather
sampling (beyond being extremely useful) was than on possessing administrative skills? Why and how
also pedagogically harmful because (wrongly) it were control staffs and official guardians variously
was received as a slogan and an implicit invita- compromised? What was behind the contradictory
tion not to care about representativeness. policy and practices associated with the use of company
materials and services? Thus the guiding question
embracing all others was: what orders the schism and
PRACTICAL ADVICE ties between official and unofficial action? (Dalton,
1959: 274)
Research questions contain concepts and cate-
gories (behaviours, attitudes and so on) the
What are the criteria for selecting cases in order researcher is interested in studying.
to construct a representative sample in social (2) Then she or he carries out primary (or
research? Some criteria have been already provisional and open,5 following Strauss and
described, but I shall summarize them together Corbin, 1990: 193) sampling in order to collect
here even though there is not a precise logical cases in accordance with concepts.
itinerary because methodological principles and (3) Because not every concept can be directly
rules do not have to stand on their own, having studied, in crafting the provisional sample she or
only a weak relation to practice. On the contrary, he reasons about some aspects:
it is necessary to approach the whole question of
sampling sequentially and it would be mislead- (a) specificity (focusing on precise social
ing to plan the whole strategy before starting. In activities, with clear contours as a ritual or
order to obtain representativeness, the sampling a ceremony);
plan needs to be set in dialogue with field inci- (b) degree of openness of the field (open or
dents, contingencies and discoveries. Here is an closed places);
example of changing or adding to the sampling (c) intrusiveness (the will to reduce the
plan on the basis of something the researcher researchers visibility);
learns in the field: (d) institutional accessibility (free-entry
versus limited-entry situations within the
Blanche Geer and I were studying college students. At organizations);
a certain point, we became interested in student lead-
(e) significance (frequent and high organiza-
ers, students who were heads of major organizations at
tional relevance of social activities).
the university (there were several hundred of them). We
wanted to know how they became leaders and how they (4) It is recommended to sample types of
exercised their powers. So we made a list of the major actions or events, incidents and not persons per
organizations (which we could do because we had been se! (Strauss and Corbin, 1990: 177, bold in the
there for a year and knew what those were, which we original text), in contrast with the common habit
would not have known when we began) and inter-
of sampling bodies and of seeking information
viewed twenty each of men and women student leaders.
from these bodies about behaviour and events
And got a great result it turned out that the men
got their positions through enterprise and hustling,
that are never observed directly (Cicourel, 1996).
while the women were typically appointed by someone
This important recommendation has two reasons:
from the university! (Becker, 13 July 2002, personal first, we do not want to replicate the survey
communication) sampling mistake about the transferability of
ideas about representativeness second, the same
Afterwards consistency must be reached in the person can engage in several activities. For
sampling reasoning, not through just applying pro- example, Dalton (1959), exploring power strug-
cedural steps. The reasoning could be as follows: gles in companies, found five types of cliques:
(1) The researcher usually starts from his or vertical (symbiotic and parasitic), horizontal
her research questions. Daltons ones were: (defensive and aggressive) and random. If we
3110-Ch-28.qxd 7/28/03 6:48 PM Page 448


sample individuals we will notice that they stay only if they are quite recurrent and significant in
in more than one clique according to the situa- the organization under study. In addition, our
tion, intentions and so on. If we look at activities, concern is with representativeness of concepts
everything is simpler. (Strauss and Corbin, 1990: 190).
(5) Until now in social research four main (9) The researcher can sample new incidents
types of sampling have been invented: purposive, or she or he can review incidents already col-
quota, emblematic and snowball. We will see lected: Theoretical sampling is cumulative. This
more of these later. In selecting cases you need to is because concepts and their relationships also
pay attention to the variance of concept in order accumulate through the interplay of data collec-
to include in the sample different voices or cases. tion and analysisuntil theoretical saturation of
(6) During the study the researcher will cer- each category is reached (Strauss and Corbin,
tainly refine his or her ideas, categories and con- 1990: 178, 188).
cepts, or come up with new ones. The important (10) This interplay between sampling and test-
thing is to make connections among them, thus ing hypotheses is needed because
formulating working hypotheses. Even if some
qualitative researchers believe that field research (a) representative samples are not predicted in
is carried out without hypotheses, making advance but found, constructed and discov-
hypotheses is as much an everyday life activity ered gradually in the field;
as comparing, sampling, making inference, find- (b) it reflects the researchers experience, pre-
ing causes, and so on. A hypothesis is a conjec- vious studies and literature regarding the
tural statement or assertion about the relation topic. In other words, the researcher will
between two or more properties. Not every know the variance of a phenomenon cumu-
hypothesis is testable; indeed the most interest- latively, study by study;
ing ones often arent. However, if we want to (c) representative samples are used to justify
persuade the reader we need to formulate them in the researchers statements.
a testable way.
(7) When the researcher has set up some
hypotheses, then she or he restarts sampling in STRATEGIES OF SAMPLING
order to collect systematically cases that relate
each hypothesis, trying to make his or her analysis
consistent. Strauss and Corbin call this second There are four important sampling procedures in
sampling relational and variational: it is associ- qualitative research.
ated with axial coding. It aims to maximize the
finding of differences at the dimensional level
(1990: 176). The authors frame the research Purposive sampling
process as a funnel-shape: through three better
and better focused steps (open, axial and selec- Purposive sampling consists of detecting cases
tive) the researcher clarifies his or her statements within extreme situations as for certain charac-
because consistency here means gathering data teristics or cases within a wide range of situa-
systematically on each category (Strauss and tions in order to maximize variation, that is, to
Corbin, 1990: 178). When the researcher finds an have all the possible situations. We can choose
interesting aspect she or he would always have two different elementary schools where, thanks
to control for it if this finding happens in other to the press, to previous studies or interviews or
samples (Perrone, 1977: 27). personal experiences, we know we can find two
(8) To guarantee representativeness she or he extreme situations: in the first school there are
needs to collect cases on recurrent behaviour at enormous integration difficulties among natives
different moments. As the researcher cannot and immigrants, while in the latter there are vir-
observe the population 24 hours a day, she or he tually none. We can also pick three schools: the
has to decide at what time and where she or he first with huge integration difficulties, the second
will observe the population (Schatzman and with average difficulties, and the third with rare
Strauss, 1973: 3941; Corsaro, 1985: 2832). cases. In the 1930s and 1940s the American
Social practices always recur in certain places anthropologist W. Lloyd Warner (18981970)
and at certain times of the day. If the researcher and his team of colleagues and students carried
knows all the different rituals of the organization out some studies on various communities in the
observed she or he can draw a representative United States. When he had to choose the sam-
sample. We do not aim to know the distribution ples, he decided to select different communities
of such behaviours (how many times), a purpose whose social structure mirrored some important
that surveys cannot succeed in, as we saw, but features of American society. Four communities
3110-Ch-28.qxd 7/28/03 6:48 PM Page 449


(called by assumed names) were chosen: a city in by Robert S. and Helen M. Lynd (1929, 1937), is
Massachusetts (Yankee City) ruled by traditions a pioneer attempt to study a typical American
on which he wrote five volumes; a lone community applying social anthropology meth-
Mississippi county (Deep South, 1941);6 a ods. The households chosen were typically
Chicago black district (Bronzetown, 1945),7 and American and not statistically representative
a city in the Midwest (Jonesville, 1949).8 ones. Other examples are two studies that
Gouldner (1954) carried out between 1948 and
1951 in a small gypsum extraction and refining
Quota sampling factory, Daltons research (1959) at Milo and
Fruhling, two companies in a highly industrial-
Quota sampling is employed for objects that ized area in the US, and, as a covert observer,
contain a wide range of statuses. The population is Kanters (1977) observations of a company with
divided up into as many sub-sets as the charac- high technological density for five years. As par-
teristics we want to observe and the proportion of ticular elements are more easily detectable
each sub-set in the sample is the same as in the through comparison, during each study it is
population. This sampling method is in wide- advisable to observe at least two cases.
spread use in market research and polls and is
usually associated with the use of questionnaires.
Curiously, even though it is not a probability Snowball sampling
sample, it is largely employed in quantitative
research because it helps obtain good findings Snowball sampling means picking some subjects
while cutting down on costs. Even though statis- who feature the necessary characteristics and,
ticians do not consider it to be scientific, because through their recommendations, finding other
cases are not selected at random but by a human subjects with the same characteristics. The
being (always the bearer of some bias), in the last research of Whyte (1943) may be an example:
few decades quota samples have been shown to thanks to Doc, a young unemployed man who
produce findings as precise (and sometimes even attended the social service, the author gradually
more precise) as probability samples.9 A good managed to get in touch with the network of his
example of this sampling method (with non- acquaintances, the people and the groups belong-
proportional quotas) applied to ethnographic ing to the street corner society. If snowball
research is Jankowskis (1991) study of criminal sampling respects some criteria, it may also be
gangs. He observed for ten years 37 different included in probability sampling (see TenHouten
gangs selected according to their ethnicity, size et al., 1971).
and members age in Los Angeles, New York As we can see, most samples used in qualita-
and Boston. tive research are still linked to a traditional style
Another example is Gouldners (1954) of constructing samples, i.e., sampling individuals
research. At that time the gypsum company instead of concepts.
employed approximately 225 people. In his
methodological appendix he reported that his
team did 174 interviews, that is, almost all of the REPRESENTATIVENESS
population (precisely 77 per cent of it): 132 of
174 interviews were done with a representative
sample of blue-collar workers in the company.
Gouldner used quota sampling stratified by age, P: We have arrived at the last issue: the
rank and tasks. Then he did another representa- relationship between representativeness
tive sample of 92 blue-collar workers who were
and generalizability. Even here there
given a questionnaire.
are some commonplaces left to be
The emblematic case 1 only findings from a representative
sample are automatically generaliz-
The emblematic case may have up to three
able to the population;
features: average (the typical provincial hospital,
the organization of a typical mountain-village 2 (therefore) representativeness leads
town hall), excellence (a well-known car- to generalizability;
manufacturing firm) and emerging (or avant- 3 if a study is not carried out on a
garde, such as recent juvenile phenomena). representative sample, its findings
Middletown, the two famous studies carried out are not generalizable;
3110-Ch-28.qxd 7/28/03 6:48 PM Page 450


4 findings of qualitative studies are not intentions: the research is carried out on
generalizable. a representative sample and then, thanks
In the methodological literature, to the findings of the ballot (a census of
generalizability is usually considered a votes), we can check the generalizability
direct and automatic consequence of of the data drawn from the sample.
representativeness. Nevertheless, as we saw before, success
S1: Why is it not? or failure of forecasts could not depend
P: Let us take another example.Why, some- only on the representativeness of the
times, do pre-electoral polls fail? As their sample. There are other factors, which
samples are statistically representative, have nothing to do with the method, that
generalization should be automatic and might have an impact on the success of
predict correctly the (probability of) the forecasts: chance, luck, artificial
results. manipulations during data processing
S4: If polls fail it means that their samples (weights),
werent representative! S1: Lack of new political parties.
P: This is a trick. Do you mean that repre- P: Right! Indeed, pre-electoral polls usually
sentativeness is evaluated only after the show a difference when a new party
result? It seems to me quite tautological, enters the electoral competition for the
a kind of statistical Darwinism whereby first time. This failure happens because
success reframes the past. What about sampling is always based on the previous
two different polling agencies working census (the past election), not on the
with the same statistically representa- present situation. So forecasts may be
tive sample and producing different unsuccessful even though statistically
results? representative samples are employed.
Too often we forget that between the This means that, if findings deriving from
representativeness of a sample and the samples may be eliminated by the find-
generalizability of its findings there is a ings of the elections, it is always advisable
number of activities that depend on at to check the population: theoretically,
least eight different domains: the trust- without the confirmation given by the
worthiness of operational definition, the elections, all findings might be valid.
trustworthiness of operationalization, Without this control we would never be
the reliability of method, the suitability of able to know which different data pub-
conceptualization, the researchers accu- lished by polling agencies mirror the
racy, his or her degree of success with population. Unfortunately, there is no
field relations (people can lie), and data similar census on most of the topics
and interpretation validity (data analysis sociological research is interested in.
error see Groves, 1989).These aspects
and their associated mistakes (called
measurement error) may jeopardize the GENERALIZABILITY
equivalence between representativeness IN STATISTICS
and generalizability. And it happens quite
frequently in any research study.
S4: So we are not certain P: Let us see briefly how generalization in
P: The only way is to check. However, in statistics works.This is called inference. It
social research the generalizability of is an academic issue since we have
findings is quite problematic because it is already seen that in most surveys there
objectively very difficult to check to see arent statistically representative samples.
if, at the end of the study, the findings To estimate the probability that the
from a sample actually mirror the rest of finding (e.g. an existing relation between
the population. To do such a thing we variable A and variable B) drawn from
would have to check the population your sample is also in the population,
anew by a census. This seldom occurs; you are helped by some statistical tests.
an example is given by polls on voting There are many statistical tests for
3110-Ch-28.qxd 7/28/03 6:48 PM Page 451


controlling hypotheses. The best known these tests and many field researchers
is chi-square. However, if you do a study are still afraid.
of a whole population (e.g. all of the 300
students in the Research Methods class
instead of just some), you do not need ARE QUALITATIVE RESEARCHERS
such a test and an inspection cross-tab FINDINGS NOT GENERALIZABLE?
by cross-tab is needed. In other words,
statistical tests of significance are used
only when you want to infer from a S1: So methodologists who say that qualita-
sample to the whole population. In the tive studies are not generalizable
past there was a long controversy about should be more careful because survey
the appropriateness of applying statistical studies also have many troubles with
significance tests in social sciences (see generalization?
Morrison and Henkel, 1971). The main P: You got the point: statistical inference in
criticisms, well known by survey social studies is quite problematic.
methodologists, are: Unfortunately this hasnt been pointed
out by most qualitative methodologists,
1 the arbitrariness of the probability who (on this issue) accepted the received
threshold of 0.05 below which the view. For example, Strauss and Corbin
null hypothesis is rejected; state:
2 the associated crudeness of this
dichotomous decision, without any in terms of making generalization to a larger
gradation, in order to accept or population, we are not attempting to generalize
as such but to specify the condition under
refuse the probability of the
which our phenomena exist, the action/interaction
null hypothesis being valid for the that pertains to them, and the associated out-
population; comes or consequences. This means that our
3 the statistical test assumes a rigor- theoretical formulation applies to these situation
ously random sample, otherwise the or circumstances but to no others. (1990: 191,
procedure is nonsense. However, as bold in the original text)
we have already seen, this is quite
S3: They refer to a kind of internal
difficult to obtain in social research;
4 the chi-square is heavily affected by
P: Yes. And a similar position is held by
sample size: it increases in signifi-
Lincoln and Guba (1985: 2021),
cance when sample size increases!
Hammersley (1992: 186ff.) and many
With very large samples even small
others. The underlying idea is that there
trends will appear significant.
are two kinds of generalizations: a gener-
So statistical tests are of modest use- alization about a specific group or popu-
fulness in social research. They can be lation (applied in surveys and polls) and a
helpfully used only as orienting criteria generalization about the nature of a
for the researchers decision-making. process (applied in field research).
S3: In addition, chi-square is a significance Consequently sampling requirements are
test only completely different in the two cases.
P: Right. It tells if a relation is likely to exist The latter is based on the notion of
in the population only. It doesnt say any- theoretical sampling. To make clearer
thing about the strength of this relation- the difference between these two kinds
ship.You need other tests, associated to of generalization, some authors have
the nature of your variables (nominal, called the second one transferability
ordinal, interval and so on), such as phi, (Guba, 1981; Lincoln and Guba, 1985: 77,
Cramers V, Pearsons C, and so on. 217),10 naturalist generalization (Stake,
However, they are rarely used. Generally 1983: 282), analytical generalization (Yin,
speaking, all these issues have been 1984: 31), extrapolation (Alasuutari,
neglected with the consequence that 1995: 157) and moderate generalization
many survey researchers still venerate (Williams, 2002). These proposals are
3110-Ch-28.qxd 7/28/03 6:48 PM Page 452


politically wise because they try to practices forms of marriage, or work,

give qualitative research scientific legiti- or habitation which dont change much,
mation; however, they do not solve even though the people who perform
adequately the theoretical and practical them are continually replaced through
problems involved in the notions of the ordinary demographic process of
representativeness and generalizability. birth, death, immigration, and emigration
We might ask: if a qualitative (2000: 6).
researchers findings are not generaliz- S4: Yes, but we do not know the distribution
able, why does she or he carry out a in the population.
study? What is the point, for instance, for P: But nobody can know this; neither can
an ethnographer spending months plan- statistics help you. You can quantify or
ning a study, weeks negotiating access to correctly estimate the distribution only
the field, months observing social actors for characteristics already counted in a
behaviour, days organizing and analysing census. How many of them are there in
ethnographic notes and, finally, writing a social research? So the only method (as
research report if then she or he is told far as I know) is by evaluating the social
that the findings are only applicable to pervasiveness of your concepts under
his or her community/organization/ study.
group (the sample) and not to many/ As Alberoni and his colleagues wrote
all cases of the population? This is in their Introduction about a study car-
masochism! ried out on 108 political activists of the
Besides, this was not the way in which Italian Communist Party and the
the studies done by Goffman, Gouldner, Christian Democratic Party:
Whyte and so on were received. No one
if we want to know, for instance, how many activists
would remember them for the findings there are of both parties in the whole country
of one setting only. coming from families of the Catholic or Communist
S2: What is the trick then? tradition, [this] study is useless; quite the reverse,
P: The logical mistake is in confusing the if we want to show that family background is
representativeness of the case with the important in determining if a citizen will be an
representativeness of its characteristics activist in the Communist rather than in the
Christian Democratic party, this research can give
observed by the ethnographer. Even
the right answer. If we want to find out what are
though the case may be a firm, we and what have been the percentages of the differ-
should take into account the relation- ent types of activistsin both parties, the study is
ship among employees, the psychological useless, while if we want to show these types exist
and relational effects of a new work the study gives a certain answer. The study does
organization, the power relationships not aim at giving a quantitative objective description
between chiefs and subordinates, and so of Italian activism but can help in understanding
some of its essential aspects, basic motivations,
on. So the sample is a compound of such crucial experiences and typical situations which
characteristics, and is not just the firm. gave birth to Italian activism and help to keep it
For this reason I consider it misleading alive. (Alberoni et al., 1967: 13)
to use the widespread expression case
study and also damaging to the image of Alberoni was too afraid of the pre-
qualitative research. As a matter of fact, dictable quantitative criticism.
to use Masons words you may have
sampled people but what you really
want to compare are their experiences A NEW CONCEPT OF
(1996: 96). It is what Gouldner (1954) GENERALIZABILITY FOR
did: he sampled blue-collar workers QUALITATIVE RESEARCH?
to discover their values, motivations and
so on. In this chapter I have documented the need for a
And Becker, writing about the cogni- new, bottom-up, socially informed and practi-
tive process of generalization, stated: cally driven theory of sampling, representative-
in every city there is a body of social ness and generalization. This theory goes back
3110-Ch-28.qxd 7/28/03 6:48 PM Page 453


and remains faithful to the original central are two different ways of thinking, and most
problem of representativeness as theorized by social sciences aim at finding such patterns.
statisticians: the variance of the phenomenon While the verbal expressions of an interactive
under study. The variance is the only worry the exchange may vary, exchange based on the
researcher needs to take into consideration. questionanswer pattern features a formal trans-
The concept of generalizability now intro- institutional (even though not universal) structure.
duced is based on the idea of social representa- While maybe to lay a page of a newspaper on the
tiveness, which goes beyond the limits of floor and declare ones sovereignty over it
statistical representativeness. The aim is to (Goffman, 1961) is a behaviour observed in one
observe extensively the relations between vari- psychiatric clinic only, the need to have a private
ables, not only to assess (which is always a quite space and control of territory has been reported
problematic task) the number of persons who fea- many times, though in different forms. However,
ture one characteristic. Therefore, generalizabil- the extension of this structure must be well orga-
ity is mainly a practically and contingent outcome nized. On the other hand, as Rositi states, we
related to the variance of the research topic; in may reasonably doubt the generalizability of
other words it is a function of the invariance (reg- findings of studies of 10002000 cases which
ularities) of the phenomenon, not a standard or claim to sample the whole population. We have
automatic algorithm of a statistical rule. to wonder if we should prefer such samples with
This is a reason why the findings of the stud- such aims. Studies with samples of 100200
ies of Goffman on social embarrassment, defer- conversational interviews, structured to
ence and demeanour, Sacks and his colleagues describe variables rather than a population, are
on telephone conversation machinery, Whyte definitely more suitable for a new model of
(1943) on social organization and leadership in a studying society (1993: 198).
group, Gouldner (1954) on bureaucracy in
medium enterprises, Cicourel (1968) on the
description of a typical juvenile offender, van ACNOWLEDGEMENT
Dijk (1983) on the cognitive processes of racial
prejudice, and Norman (1988) on accidents in I wish to thank Howard Becker, Aaron Cicourel,
the use of technology in everyday life, and so on, two anonymous referees and the editors for help-
have always been considered generalizable. ful comments and suggestions.
In qualitative research, generalizability con-
cerns general structures rather than single social
practices, which are only an example of this NOTES
structure. The ethnographer does not generalize
one case or event that, as Max Weber pointed 1 Incidentally I emphasize that this remains a problem
out, cannot recur but its main structural aspects even for studies based on samples.
that can be noticed in other cases or events of the 2 See http://www.census.gov/dmd/www/samfaq.htm
same kind or class. For example, in the conclu- 3 The author also writes: the situation is slightly better
sions of his study of the relationship between a where the unit is a group because an aggregate of indi-
psychotherapist and a patient suffering from viduals is usually institutionalized (schools, universi-
AIDS, Perkyl says: ties, companies, hospitals, commercial activities);
however, the situation becomes more complicated for
The results were not generalizable as descriptions of unofficial structures: private language or computer
what other counselors or other professionals do with courses, sport, leisure and cultural associations, char-
their clients; but they were generalizable as descriptions ities, amateur theatre groups, bowling clubs, pensioners
of what any counselor or other professional, with his or recreation groups and so on (Corbetta, 1999: 333).
her clients, can do, given that he or she has the same 4 Even among those who seem to accept the legitimacy
array of interactional competencies as the participants of inferential procedure, there are authors who have
of the AIDS counseling session have. (Perkyl, 1997: doubts about it, such as Isidor Chain (1963: 512), who
216, quoted in Silverman, 2000: 109) states that we are never sure that findings on sample
can be applied to population unless we also simulta-
Something similar Perkyl, happens in film and neously do a census.
radio production with noise sampling. The 5 As Strauss and Corbin (1990: 176) explain: open
squeak of the door (which gives us the shivers sampling is associated with open coding. Openness
when we watch a thriller or a horror) does not rather than specificity guides the sampling choices.
represent all squeaks of doors but we associate it Open sampling can be done purposively (e.g.
with them. We do not think about the differences pp. 1834) or systematically (e.g. p. 184) or occurs
between that squeak and the one made by our fortuitously (e.g. pp. 1823). It includes on-site sampling.
front door; we notice the similarities only. They 6 Cf. Davis et al. (1941).
3110-Ch-28.qxd 7/28/03 6:48 PM Page 454


7 Cf. Drake and Cayton (1945). Anthropological Approach and Psychological

8 Cf. Warner (1949). Approach. New York: Teachers College Press.
9 As Barisione and Mannheimer write: The experi- Converse, Jean Marie and Schuman, Howard (1974)
ence in public opinion surveys has showed that quota Conversations at Random. New York: Wiley.
samples are on the average reliable; as a matter of Corbetta, Piergiorgio (1999) Metodologia e tecniche della
fact their results do not differ from results obtained ricerca sociale. Bologna: Il Mulino. Translated as
with random samples of the same size more than two Social Research. London: Sage, 2001.
random samples differ between them (1999: 54). Corsaro, William A. (1985) Friendship and Peer Culture
This is another case when experience and practice in the Early Years, Norwood, NJ: Ablex.
predominate over abstract statistical principles Crapanzano, Vincent (1980) Tuhami, Portrait of a
(1999: 54). Moroccan. Chicago: University of Chicago Press.
10 According to the authors the expression transferabil- Dalton, Melvin (1959) Men Who Manage. New York: Wiley.
ity does not stand for a researchers inference Davis, Allison, Gardner, Burleigh B. and Gardner, Mary R.
process but a choice of the reader, who may transfer (1941) Deep South. Chicago: University of Chicago
this information to other situations on the base of a Press.
solid thick description provided by the researcher De Martino, Ernesto (1961) La terra del rimorso. Milan:
(Lincoln and Guba, 1985: 362). Il Saggiatore.
Denzin, Norman K. (1971) Symbolic interactionism and
ethomethodology, in Jack D. Douglas (ed.),
REFERENCES Understanding Everyday Life. London: Routledge &
Kegan Paul, pp. 000.
Alasuutari, Pertti (1995) Researching Culture, London: Drake, St Clair and Cayton, Roscoe Horace (1945) Black
Sage. Metropolis. New York, Harcourt Brace.
Alberoni, Francesco, et al. (1967) Lattivista di partito. Festinger, L., Henry, H.W. and Schachter, S. (1956) When
Bologna: Il Mulino. Prophecy Fails. Minneapolis: University of Minnesota
Bailey, Kenneth D. (1978) Methods in Social Research. Press.
New York: Free Press. Fielding, Nigel (1981) The National Front. London:
Barisione, M. and Mannheimer, R. (1999) I sondaggi. Routledge & Kegan Paul.
Bologna: Il Mulino. Foucault M. (1966) Les mots et les choses. Paris:
Becker, Howard (1951) The professional dance musician Gallimard. Translated as The Order of Things.
and his audience, American Journal of Sociology, New York: Vintage Books, 1973.
57: 000. Frege, G.F.L. (1892) ber Sinn und Bedeutung,
Becker, Howard (1963), Ousiders. Glencoe, IL: Free Press. Zeitschrift fr Philosophie und philosophische Kritik,
Becker, Howard (1998), Tricks of the Trade. Chicago: 000.
University of Chicago Press. Galtung, John (1967) Theory and Methods of Social
Becker, Howard (2000) Italo Calvino as urbanologist, Research. Oslo: Universitets Forlaget.
LAnne Sociologique, 000 Garfinkel, Harold (1967) Studies in Ethnometodology.
Bourdieu, Pierre, et al. (1993) La misre du monde. Paris: Englewood Cliffs, NJ: Prentice-Hall.
Editions du Seuil. Gasperoni, Giancarlo and Marradi, Alberto (1996)
Brehm, John (1993) The Phantom Respondents: Opinion Metodo e tecniche delle scienze sociali, in
Surveys and Political Representation. Ann Arbor: Enciclopedia delle scienze sociali. Rome: istituto della
University of Michigan Press. Enciclopedia Italiana, vol. 5, pp. 62443.
Capecchi, Vittorio (1972) Struttura e tecniche della Geertz, Clifford (1972) Deep play: notes on the Balinese
ricerca, in Pietro Rossi (ed.), Ricerca sociologica e cockfight, Dedalus, 101: 137.
ruolo del sociologo. Bologna: Il Mulino. Gilli, Gian Antonio (1971) Come si fa ricerca. Milan:
Castellano, V. and Herzel, A. (1971) Elementi di teoria Mondadori.
dei campioni. Rome: Ilardi. Glaser, Barney G. and Strauss, Anselm L. (1967) The
Chain, Isidor (1963) An introduction to sampling, in Discovery of Grounded Theory. Chicago: Aldine.
C. Selltiz and M. Jahoda (eds), Research Methods Gobo, Giampietro (1990), The first call: rituals and
in Social Relations. New York: Holt & Rinehart, rhetorical strategies in the first telephone call with
pp. 50945. Italian respondents. Paper presented at Annual
Cicourel, Aaron Victor (1968) The Social Organization of Meeting of the ASA, Washington, DC, 1115 August.
Juvenile Justice. New York: Wiley. Gobo, Giampietro (1997) Le risposte e il loro contesto.
Cicourel, Aaron Victor (1996) ecological validity and Milan: Angeli.
white room effects , Pragmatics and Cognition, Gobo, Giampietro (2001) Best practices: rituals and
4(2): 22164. rhetorical strategies in the initial telephone contact ,
Cicourel, Aaron Victor and Boese, R. (1972) Sign lan- Forum: Qualitative Social Research, 2(1), http://www.
guage acquisition and the teaching of deaf children, in qualitative-research.net/fqs-texte/101/101gobo-e.
Dell Hymes et al. (ed.), The Functions of Language: An htm
3110-Ch-28.qxd 7/28/03 6:48 PM Page 455


Goffman, Erving (1961) Asylums. New York: Doubleday. Morrison, D.D. and Henkel, R.E. (1971) The Significance
Goffman, Erving (1963) Behavior in Public Places. Tests Controversy: A Reader. Chicago: Aldine.
Glencoe, IL: Free Press. Norman, Donald A. (1988) The Psychology of Everyday
Gomperz, H. (1905) Weltanschauungslehre, vol. I: Things. New York: Basic Books.
Methodologie, Jena: Diederichs. Ogden, C.K. and Richards, I.A. (1923) The Meaning of
Goode, William, and Hatt, Paul K. (1952) Methods in Meaning. London: Routledge & Kegan Paul.
Social Research. New York: McGraw-Hill. Peirce, Charles S. (1902) Sign, in J.M. Baldwin (ed.),
Gouldner, Alvin G. (1954) Patterns of Industrial Dictionary of Philosophy and Psychology, vol. II.
Bureaucracy. New York: Free Press. New York: Macmillan.
Griaule, Marcel (1948) Dieu deau: entretiens avec Perkyl, Anssi (1997) Reliability and validity in
Ogotemmli. Paris: Editions du Chne. research based upon transcripts, in David Silverman
Groves, Robert M. (1989) Survey Errors and Survey (ed.), Qualitative Research. London: Sage, pp. 20119.
Costs. New York: Wiley. Perrone, Luca (1977) Metodi quantitativi della ricerca
Groves, Robert M. and Lyberg, L.E. (1988) An overview sociale. Milan: Feltrinelli.
of nonresponse issues in telephone surveys, in Robert Pinto, Roger (1964) Mthodes des sciences sociales.
M. Groves et al. (eds), Telephone Survey Methodology. Paris: Dalloz.
New York: Wiley. Pitrone, Maria Concetta (1984) Il sondaggio. Milan:
Guba, Egon G. (1981) Criteria for assessing the trust- Angeli.
worthiness of naturalistic enquires, Educational Pollner, Melvin and McDonald-Wikler, L. (1985) The
Communication and Technology Journal, 29: 7592. social construction of unreality, Family Process, 24:
Hammersley, Martyn (1992) Whats Wrong with 24154.
Ethnography? London: Routledge. Popper, Karl R. (1972) Objective Knowledge. Oxford:
Hebdige, Dick (1979) Subculture: The meaning of Style. Oxford University Press.
London and New York: Routledge. Rosenhan, David L. (1973) On being sane in insane
Jankowski, Martn Snchez (1991) Islands in the Street. places, Science, 179: 2508.
Berkeley: University of California Press. Rositi, Franco (1993) Strutture di senso e strutture di
Javeau, C. (1987) Analisi dei singolare e sociologia, in dati, Rassegna Italiana di Sociologia, 2: 177200.
Roberto Cipriani (ed.), La metodologia delle storie di Sacks, Harvey (1984) On doing being ordinary , in
vita. Rome: Euroma-La Goliardica. J.M. Atkinson and J. Heritage (eds), Structures of
Kahneman, Danny and Tversky, Amos (1972) Subjective Social Action. Cambridge: Cambridge University Press,
probability: a judgment of representativeness, pp. 51329.
Cognitive Psychology, 3: 43054. Sacks, Harvey (1992) Lectures on Conversation, vol. 1.
Kanter, Rosabeth Moss (1977) Men and Women of the Oxford: Blackwell.
Corporation. New York: Basic Books. Schatzman, Leonard and Strauss, Anselm L. (1973) Field
Kish, L. (1965) Survey Sampling. New York: Wiley. Research. Englewood Cliffs, NJ: Prentice-Hall.
Lincoln, Yvonna S. and Guba, Egon G. (1985) Naturalist Silverman, David (2000) Doing Qualitative Research.
Inquiry. Beverly Hills, CA: Sage. London: Sage.
Lynd, Robert S. and Lynd, Helen M. (1929) Middletown. Stake, Robert (1983) The case study method in social
New York: Harcourt, Brace. enquiry, in G. Madaus, M. Scriven and D. Stufflebean
Lynd, Robert S. and Lynd, Helen M.(1937) Middletown in (eds), Evaluation Models. Boston: Kluwer-Nijhoff,
Transition. New York: Harcourt, Brace. pp. 27986.
Marradi, Alberto (1989) Casualit e rappresentativit di Strauss, Anselm (1987) Qualitative Analysis for Social
un campione: contributo a una sociologia del linguag- Scientists. Cambridge: Cambridge University Press.
gio scientifico, in Renato Mannheimer (ed.), I Strauss, Anselm and Corbin, Juliet (1990) Basics of
sondaggi elettorali e le scienze politiche. Milan: Qualitative Research. London: Sage.
Angeli, pp. 51133. Sudman, Seymour (1983) Applied sampling, in
Marradi, Alberto (1994) Referenti, pensiero e linguaggio: P.H. Rossi, G.D. Wright and A.B. Anderson (eds),
una questione rilevante per gli indicatori, Sociologia e Handbook of Survey Research. San Diego: Academic
Ricerca Sociale, 43: 137207. Press, pp. 000.
Mason, Jennifer (1996) Qualitative Researching. Sudnow, David (1967) Passing On, Englewood Cliffs,
Newbury Park, CA: Sage. NJ: Prentice-Hall.
Maynard, Douglas W. and Schaeffer, Nora Cate (1999) TenHouten, Warren, Stern, John, and TenHouten, Diana
Keeping the gate, Sociological Methods and (1971) Political leardership in poor communities:
Research, 1: 3479. application of two sampling methodologies, in
Mitchell, Ritva and Karttunen, Sari (1991) Perch e come P. Orleans and W.R. Ellis Jr (eds), Race, Change and
definire un artista?, Rassegna italiana di Sociologia, Urban Society, vol. V. Beverly Hills, CA: Sage.
32(3): 34964. Tversky, Amos and Kahneman, Danny (1974) Judgment
Morris, Charles W. (1938) Foundations of the Theory of under uncertainty: heuristics and biases, Science, 185:
Signs. Chicago: University of Chicago Press. 112331.
3110-Ch-28.qxd 7/28/03 6:48 PM Page 456


van Dijk, Teun Adrian (1983), Cognitive and Whyte, William Foote (1943). Street Corner Society.
conversational strategies in the ethnic prejudice, Text, Chicago: University of Chicago Press.
3(4): 375404. Williams, Malcolm (2002) Generalization in interpretive
Warner, L. William (1949) Democracy in Jonesville. research, in Tim May (ed.), Qualitative Research in
New York: Harper & Row. Action. London: Sage, pp. 000.
Whorf, Benjamin Lee (1956) Language, Thought, and Yin, Robert K. (1984) Case Study Research. Thousand
Reality. Cambridge, MA: MIT Press. Oaks, CA: Sage.