Andrew Abbott (2004), Methods of Discovery: Heuristics for the Social Sciences. New York: Norton.

Chapter One
EXPLANATION
I. EXPLANATION
II. METHODS
A. ETHNOGRAPHY
B. HISTORICAL NARRATION
C. STANDARD CAUSAL ANALYSIS
D. SMALL-N COMPARISON
E. FORMALIZATION
III. EXPLANATORY PROGRAMS

SCIENCE IS A CONVERSATION between rigor and imagination. What one proposes,

the other evaluates. Every evaluation leads to new proposals, and so it goes, on
and on.
Many people think of social science less as a conversation than as a
monologue. For them, it is a long speech that ends with a formal question, to
which reality meekly answers yes or no like the plastic heroine of a Victorian
novel. Yet no good researcher believes in such monologues. Researchers know
all about the continual interchange between intuition and method, just as they
know about the endless teasing of reality as it evades them. Social science in
practice is less old-style romance than modern soap opera.
The monologue version of social science is of course easier to describe.
There are many excellent books about its machinery: how to propose a
question, how to design a study, how to acquire and analyze data, how to draw
inferences. Indeed, many books are organized around particular ways of doing
these things, the various “methods,” as we call them: ethnography, surveys,
secondary data analysis, historical and comparative methods, and so on. All
that is fine and good.
But such books forget the other voice, the imaginative voice of whimsy,
surprise, and novelty. This discovery side of social science is more systematic
than we think. Social scientists use gambits of imagination, mental moves they
employ to hasten discovery. Like gambits in chess, these mental moves are
formulas for the opening, developing, and realizing of possibilities. Some are
general gambits implicit in the nature of argument and description, while
others arise in conceptual issues that pervade the disciplines. All of these
gambits work within any kind of method. They make up the heuristic of social
science, the means by which social science discovers new ideas.
We need heuristic because, as I said, social reality often resists the charms of
methodology. As social scientists, we aim to say something interesting—
perhaps even true—about social life. Yet social reality often makes a stingy
reply to even the best of our methodological monologues, returning tiny
correlations even though challenged by the best of questionnaires, returning
simpleminded truisms even though watched by months of earnest ethnography,
returning boring stories even though questioned by years of painstaking
archival research. Social reality wants a subtler wooing; it wants rigor and
imagination.1
2 Abbott (2004), Methods of Discovery

So this is a book about heuristic, a book of aids to the social scientific

imagination. Because I am a sociologist, many of the examples I use in the book
come from sociology. But because the social sciences are all mixed up together,
not all of the examples will be sociological. The social sciences share subject
matters, theories, and a surprising amount of methodology. They are not
organized into a clearly defined system but take their orientations from various
historical accidents. Loosely speaking, economics is organized by a theoretical
concept (the idea of choice under constraint), political science by an aspect of
social organization (power), anthropology by a method (ethnography), history
by an aspect of temporality (the past), and sociology by a list of subject matters
(inequality, the city, the family, and so on). Thus, there is no single criterion for
the distinctions among disciplines. As a result, when one or another discipline
becomes too much of a bore, the others make fun of it and steal its best ideas
to put them to better use elsewhere. All of this flux means that a heuristics
book can range widely, as this one will.

THE FIRST TWO CHAPTERS introduce the aims, means, and assumptions of social
science research. I begin with explanation because explanation is the purpose
of social science. I then introduce some types of methods—some of the various
ways in which social scientists have tried to be rigorous. I treat these methods
as concrete realizations of “explanatory programs,” programs that carry out the
different concepts of explanation introduced earlier in the chapter.
Chapter Two turns to a more customary approach. I characterize methods in
terms of a set of conceptual issues—nine of them, in fact. I first introduce these
conceptual issues, then give the customary account of methods (I skipped it in
Chapter One), which says that methods are best defined in terms of these nine
issues. Then I leave the beaten path. I discuss the critiques that each method
poses to the others and show that these critiques lead us into an endless
cycling through the methods (both in theory and in practice). Moreover, the
conceptual issues themselves turn out not to be fixed things; they have an
unstable, fractal character. Not only do they differentiate one method from
another, they also differentiate internal strands within each method—and
internal strands within the internal strands. And so on.
Chapters One and Two are the heavy lifting before the fun part begins. While
the main aim of the book is to stimulate imagination, it needs to present a clear
sense of rigor as well. Otherwise, we won’t be able to tell the difference
between imagination and foolishness. Recognizing that difference means
getting a secure sense of what explanation is, of why we seek explanations, and
of what different kinds of explanations and programs of explanation exist in
social science. It also means having a solid grasp of more traditional ways of
thinking about rigor, which are presented in Chapter Two, with its litany of the
classic methodological debates in social science and its endless isms.
(Ultimately, I will turn these isms from dead methodological debates into live
heuristics.)
Chapter One: Explanation 3

Having set forth the basics of rigor in Chapters One and Two, I then turn to
imagination. Chapter Three discusses the general concept of heuristic and sets
forth the two simplest heuristic strategies: the additive heuristic of normal
science and the use of commonplace lists to generate new ideas. Chapter Four
considers in detail the general heuristic gambits that search for importable
novelty elsewhere and produce it by transforming our existing arguments.
Chapter Five looks at the heuristics of time and space, the heuristics that
change ways of describing or envisioning social reality so as to produce new
ideas. Chapter Six examines the gambits that arise out of the basic debates and
methodological concerns of Chapter Two—making a positivist move within an
interpretive tradition, for example. Finally, Chapter Seven discusses the
problem of evaluating the ideas produced by heuristics. It asks how we know a
good idea when we see one.
I have drawn examples from as far back as the 1920s and as recently as
1999. Old work is not necessarily bad work. Newton himself is a good example.
Newton became the greatest name in modern science by giving up on the
medieval question of the nature and origins of motion. He solved the problem
of motion by simply assuming that (a) motion exists and (b) it tends to persist.
By means of these assumptions (really a matter of declaring victory, as we
would now put it), he was able to develop and systematize a general account of
the regularities of motion in the physical world. That is, by giving up on the
why question, he almost completely answered the what question. So following
his example, we learn that switching questions is a powerful heuristic move.
The very same move has occurred in social science. One of the great
difficulties in the work of Talcott Parsons, the dominant American sociologist of
the mid-twentieth century, was in explaining social change. Parsons held that
social behavior was governed by norms, which were themselves governed by
values, which were themselves governed by yet more general values. In such a
system, change could be conceived only as local breakdown, a problem event
that had somehow escaped the supervising norms. Later writers handled the
same problem—explaining change—by simply assuming that social change was
not unusual at all; rather, it was the normal state of affairs. With this
assumption, the various historical sociologists who challenged Parsons were
able to develop much more effective accounts of social movements, of
revolutions, and, indeed, of the rise of modernity in general. This was exactly
the Newtonian move: historical sociologists gave up on explaining change and
simply assumed it was happening all the time. Then all they had to do was
figure out what is regular about the way it happens. (They should have gone on
to explain stability, of course, but they pretty much forgot about that!)
Thus, old work provides useful examples of heuristics just as new work does.
This means that as I introduce the reader to the basic tool kit of heuristics in
social science, I can simultaneously introduce some of the great heritage that
that tool kit has produced. Let’s begin, then, at the beginning—with
explanation.

I. EXPLANATION
4 Abbott (2004), Methods of Discovery

Social science aims to explain social life. There are three things that make a
social scientist say that a particular argument is an explanation. First, we say
something is an explanation when it allows us to intervene in whatever it is we
are explaining. For example, we have explained the economy when we can
manage it. We have explained poverty when we know how to eradicate it.
Second, we say an account explains something when we stop looking for
further accounts of that something. An explanation is an account that suffices.
It frees us to go on to the next problem by bringing our current problem into a
commonsense world where it becomes immediately comprehensible. So socio-
biologists say they have explained altruistic behavior when they show it to be
merely an accidental result of selfish behavior. They go no further because they
think selfish behavior is self-evident; it needs no explanation.
Third, we often say we have an explanation of something when we have made
a certain kind of argument about it: an argument that is simple, exclusive,
perhaps elegant or even counterintuitive. Thus, we may think Freudian
psychology is better than folk psychology because it is better worked out, more
complex, and more surprising. In this third sense, an account is an explanation
because it takes a certain pleasing form, because it somehow marries simplicity
and complexity.
The first of these views—the pragmatic view that an explanation is an
account that enables us to intervene—is the most familiar. Consider the
explanation of germ-based disease. We think discovering a germ is explaining a
disease because by discovering the germ, we have discovered something that
enables us to stop the disease. Note that this pragmatic approach to ex-
planation works best for phenomena that have somewhere a narrow neck of
necessary causality: something absolutely necessary to the phenomenon yet
clearly defined and subject to outside action. It is this narrow neck—the
necessity of a particular organism—that makes the germ-based diseases easier
to fight than diseases “caused” by the interaction of millions of small random
events—cancer, heart disease, and arthritis. The move to the microcellular level
in studying these diseases aims precisely to find a new realm where there may
be a narrow neck— the necessary presence of a certain gene or enzyme, for
example. In social science, however, relatively few phenomena seem to have
this narrow-neck pattern. So, as we shall see, the pragmatic approach to
explanation in social science has taken a different path.
In the second view of explanation, where an explanation is an account that
enables us to stop looking for further accounts, things are different. This kind
of explanation works by transposing the thing we want to explain from a world
that is less comprehensible to one that is more comprehensible. The attempt to
explain all human activities without any reference to group phenomena is a
good example. The utilitarian philosophers tried to show that systematic
pursuit of self-interest by everyone (an individual phenomenon repeated many
times) would, when aggregated, result in the social world that was best for all.
Social reality was just an additive total of individual realities. Apparent social
phenomena, like the (to them unbelievable) phenomenon of people getting
Chapter One: Explanation 5

along without obvious coordination, must be explained as the result of some

ensemble of individual behaviors.
This second view of explanation—in which we think explanation is a move
from one conceptual world to another—is not a pragmatic but rather a semantic
view. It defines explanation as translating a phenomenon from one sphere of
analysis to another until a final realm is reached with which we are intuitively
satisfied. So the utilitarians “explain” prosocial behavior as an outcome of
individual selfishness because they feel the latter realm—that of individual
selfish activity—is more real, more intuitive, than any other. It doesn’t need to
be explained any further. It is a “final realm” for explanation.
Of course, different schools of thought have different final realms for
explanation. Utilitarians and their followers, the economists, aren’t happy until
they have translated a phenomenon into something recognizable on their
familiar turf of individuals with preferences and constraints. But anthropologists
are equally unhappy until they have translated those very same preferences into
what is for them the familiar realm of culture. This difference makes it awkward
to refer to the semantic view of explanation as reduction, which is the usual
name for it in the philosophy of science. The word reduction seems to imply a
hierarchy of explanation, in which “emergent” phenomena are “reduced” to
“lower-level” ones. Such a view may make sense for the natural sciences, where
it is common to think about reducing chemistry to physical chemistry and
ultimately to physics. But it isn’t very helpful in social science, where the final
realms of the various disciplines and research traditions are not shared or
ordered in any way.
The third view of explanation, as I noted, derives from the characteristics of
explanation itself. Often we think an explanation is satisfactory simply because
it is logically beautiful and compelling. Indeed, sometimes we find an
explanation beautiful and satisfying without believing it at all. This is the reac-
tion most people have to Freud on a first reading. It may or may not work, but
how elegant it is! How simple yet comprehensive! Many have the same reaction
to Jean Piaget’s early work on the origins of intelligence in children. From such
tiny postulates, he managed to produce so many insights! Reflective life creates
in us a desire for pretty argument. We may not like its premises, its content, or
its results, but we all appreciate its enticing mixture of complexity and clarity.2
Formal writing about explanation has usually taken this third view, that
explanation has to do with the properties of an argument—specifically, its
logical structure. In the most famous article on explanation in the twentieth
century, the philosopher Carl Hempel argued that to explain is to demonstrate
that the starting conditions in the case that we want to explain fit the
hypothesis conditions of some general “covering law” (1942). For example, we
might have the covering law that when a political party has a substantial
majority in a parliament, it will be able to have a large effect on the country.
Then we demonstrate in a particular case (say, Great Britain in 1997, after the
labour landslide) that one party had such a substantial majority. We can then
say we have explained why the Labour Party has had a strong effect on British
6 Abbott (2004), Methods of Discovery

policies in the years after 1997: the conjunction of our covering law—“whenever
a party has a strong majority, it has a big effect”—with our empirical premise—
“Labour in 1997 got a strong majority”—logically entails the empirical conclu-
sion that “Labour had a large effect on the country.” By combining the general
law with a demonstration that our particular case fits the condition of that law,
we can use the conclusion of the law to explain the particular outcome in our
particular case.
Hempel’s view of explanation focused on the logical pattern of an account,
on the way its parts are put together. His is a syntactic view of explanation, for
it emphasizes the syntax of an account rather than its ability to help us act (the
pragmatic view) or its ability to translate a phenomenon into a realm we think
we understand intuitively (the semantic view).
Now the goal of social science, as I have said, is explanation of social life in
whichever of these three senses we choose. A century or so of experience has
taught social scientists some standard ways to go about this.3

II. METHODS
Social scientists have a number of methods, stylized ways of conducting their
research that comprise routine and accepted procedures for doing the rigorous
side of science. Each method is loosely attached to a community of social
scientists for whom it is the right way to do things. But no method is the
exclusive property of any one of the social sciences, nor is any social science,
with the possible exception of anthropology, principally organized around the
use of one particular method.4
One might expect that the various social science methods would be versions
of a single explanatory enterprise or that they would be logical parts of some
general scheme, but in practice they don’t work that way. Far from being parts
of a general scheme, they are somewhat separated from one another and often
mutually hostile. In fact, many social scientists use methods that take for
granted that other methods—used by other social scientists—are useless. But
nobody cares much. The various methodological traditions roll along, happily
ignoring one another most of the time.
It is therefore not at all obvious how best to classify methods. If we recall the
basic questions of method—how to propose a question, how to design a study,
how to draw inferences, how to acquire and analyze data—we can see that any
one of these questions might be used to categorize methods. If we categorize
by type of data gathering, there are four basic social science methods:
1. ethnography: gathering data by personal interaction
2. surveys: gathering data by submitting questionnaires to respondents or
formally interviewing them
3. record-based analysis: gathering data from formal organizational records
(censuses, accounts, publications, and so on)
4. history: using old records, surveys, and even ethnographies
If, by contrast, we begin with how one analyzes data, we might have three
methods:
Chapter One: Explanation 7

1. direct interpretation: analysis by an individual’s reflection and synthesis

(for example, narration)
2. quantitative analysis: analysis using one of the standard methods of
statistics to reason about causes
3. formal modeling: analysis by creating a formal system mimicking the world
and then using it to simulate reality
If we begin with how one poses a question, we might note the important
issue of how many cases we consider. This would give us three kinds of
methods:
1. case-study analysis: studying a unique example in great detail
2. small-N analysis: seeking similarities and contrasts in a small number of
cases
3. large-N analysis: emphasizing generalizability by studying large numbers
of cases, usually randomly selected
Any one of these categorizations could be used to classify methods.
Moreover, putting these three category systems together gives one 4 x 3 x 3 =
36 possible subtypes. And in fact, the majority of these subtypes have been
tried by someone at some point or other.
Because there is no obvious list or categorization of methods, I will simply
give five examples of conspicuously successful methodological traditions:
ethnography, historical narration, standard causal analysis, small-N
comparison, and formalization. Most of these have been hybridized in various
ways, but we can look at the hybrids later if we need to. (Actually, small-N
comparison will serve as an example of hybrid methods throughout.) Note that
these five examples do not make up an exhaustive list. Indeed, they come out
of different ways of categorizing methods. Ethnography is a way of gathering
data, narration is a way of writing it up, small-N comparison is a choice of data
size, standard causal analysis is a general analytic approach, and formalization
is a specific analytic approach using purely abstract data. Let me reiterate.
There is no one basic way to categorize methods, nor is there any simple set of
dimensions for arraying them. Methodological traditions are like any other
social phenomena. They are made by people working together, criticizing one
another, and borrowing from other traditions. They are living social things, not
abstract categories in a single system. Each of the five methods that follow is a
living mode of inquiry with a long and distinguished lineage.

A. Ethnography
Ethnography means living inside the social situation one is studying and
becoming to some extent a participant in it. One’s participation can range from
mere observation to going native, from occasional afternoons to round-the-
clock immersion. One can augment this participation with interviews, guidance
from key informants, and review of official records.
An ethnographer’s questions are often not very detailed before the field
research begins, although the researcher will have a general puzzle or problem.
As an ethnographer proceeds, he or she generates a mass of field notes:
8 Abbott (2004), Methods of Discovery

records of events, interviews, observations, and reflections about personal

reactions, as well as endless verbatim records of conversations and interac-
tions. The ethnographer floats into and out of the field situation, trying to keep
an outsider’s view even while developing an insider’s one as well. Continually
reading and rereading field notes, the ethnographer thinks up new questions to
ask and new avenues to explore. This constant reflection is difficult, and as a
result the field experience is disorienting, as is evident in the famous field
diaries of the anthropologist Bronislaw Malinowski (1989).
When the fieldwork is done, the ethnographer returns home and
contemplates these hundreds of pages of notes. Questions become clearer.
Connections and themes begin to surface as the inchoate data are classified
and reclassified, thought and rethought. The result is most often a monograph
of some sort, with chapters that pose the now clear question, set the ethno-
graphic scene, present extensive data from the field, and in the end provide a
theoretical insight.
As an example, consider Witchcraft, Oracles, and Magic among the Azande
by E. E. Evans-Pritchard. Evans-Pritchard made several extended sojourns
among the Azande between 1926 and 1930. Interestingly, he did not go to the
field to study what he eventually wrote about: “I had no interest in witchcraft
when I went to Zandeland, but the Azande had; so I had to let myself be guided
by them” (1976:242). As a result of that guidance, Evans-Pritchard wrote a
monumental book that explores not only withcraft but all the “metaphysical”
ideas of the remarkable Azande. The central question eventually became one of
why the Azande held the beliefs they held about the supernatural and the
nonobservable. Evans-Pritchard gave a functional answer to this question;
beliefs in witchcraft, oracles, and magic served mainly to reinforce the social
and cultural status quo. But this simplistic summary of the book belies its ex-
traordinary richness. One comes away from it having questioned not only
Azande beliefs but also one’s own.

B. Historical Narration
Historical narration is another methodological tradition. Much of historical
work is descriptive, examining the question of what really was the state of
affairs in a particular place and time. But historians often pose a specific
narrative question: most commonly, why did such and such an event take
place? Historians apply many methods to such questions. Much of historical
work consists of amassing published or archival materials from the time and
place studied, so-called primary materials. Strange as it may seem, historical
data are often embarrassingly rich; we often know too much about the details
of the past. As a result, historical method often takes the form of trolling these
seas of old data for important materials.
The heart of historical method is the reading of documents themselves. An
informed historical reading of primary materials presupposes extensive—
indeed overwhelming—knowledge of the time, and place that produced them.
Often this includes not only knowing the environing historical record but also
Chapter One: Explanation 9

knowing foreign languages (or old usages in one’s own language) and indeed
recognizing the historical and regional varieties both of languages and of the
many forces behind the survival of the documents read. The historian (or any
social scientist employing historical methods) walks a thin line between
overinterpreting and underinterpreting sources. No source should be read out
of context, but the art of historical discovery often lies in figuring out how
previous conceptions of that context were wrong. Thus, reading documents
seems easy but is difficult.
Like the ethnographer, the historian carries out many tasks simultaneously,
now seeking documents, now reading them, now looking for more, now
assembling preliminary arguments and recasting earlier interpretations. As with
ethnography, there is a long and painstaking process by which a researcher
assembles a synthetic view of something that is first perceived only through a
welter of particular detail. But it has long been a custom of historians to hide
their arduous research process under an elegant mantle of prose. Without
question, history is the best written of the social sciences, perhaps the only
social science that is read widely for pleasure by nonspecialists. As a result,
history and in particular historical narrative seem at their best to be simple and
effortless. That simplicity, however, is deceptive.
A classic example of historical work is A. J. P. Taylor’s celebrated and
contentious Origins of the Second World War. Taylor set himself the task of
showing why the European war of 1939 broke out. One of the revolutionary
aspects of Taylor’s book was that it asked this question at all; previous writers
had seen Hitler’s war as requiring no explanation. Taylor’s materials included
thousands of documents, memoirs, and published works in all the languages of
Europe. As with most first-rate history, the methodological efforts that
produced the book—the reading of this enormous mass of material, the
interpretations tried and rejected, the sources sought but missed—disappear
behind Taylor’s smooth, ironic prose. His basic interpretation—that German
foreign policy in the interwar period was brilliantly (and successfully)
opportunistic and that Hitler’s ingenuity deserted him only when he
gratuitously invaded the Soviet Union and declared war on the United States—
caused a furor for decades after its publication.

C. Standard Causal Analysis

Standard causal analysis (SCA) takes large numbers of cases, measures
various aspects of them, and employs statistical models to draw inferences
about the relationships among those measurements. It then uses the inferences
to consider the causal factors that might have produced the correlational
patterns that are observed in the data.
Causal analysis starts by defining a universe of cases in which it is interested.
These can be anything: people, organizations, families, nations, cities. The
cases are then measured by some common yardsticks. These variables can be
unordered categories, like race, gender, graduate degree, occupation, or color
of eyes. They can be ordered categories, like the familiar five-point attitude
10 Abbott (2004), Methods of Discovery

scale from “strongly disagree to disagree,” “don’t care,” “agree,” and “strongly
agree.” Or they can be continuous scales, like income, wealth, age, and level of
education. Much of the hard work in standard causal analysis takes the form of
finding, measuring, and assessing the distributions of these variables. As in
ethnography and historical research, this apparently simple task of data
gathering is easy to do badly if one is not careful.
One of the variables is taken, in each particular study, to be the dependent
variable. That is, the analyst will seek to know the effects of all the other
(independent) variables on this dependent one. Mathematically, the analyst tries
to replace the dependent variable with a weighted sum of the independent
variables. So if the dependent variable is income, for example, one takes so
many parts education and so many parts occupation and so many parts gender,
and so on, and sees how well one can predict income. There are many
mathematical complexities to this approach, and there are several different
ways of estimating the results, but the basic approach is always to vary the
weights in order to find the weighted sum of the independent variables that
best predicts the dependent variable. Note, however, that what is independent
in one study can be dependent in another, and vice versa.
Analysts choose their variables by trying to think up causal stories that would
imply that some variable has a powerful effect on another. Someone predicting
individual racial attitudes will probably use region of birth as a predictor, for
example. Note, too, that the mathematics does its best to control the in-
terdependencies of the variables. Either education or occupation does pretty
well predicting income by itself, but when the two are together, they aren’t
twice as good, because they are highly correlated with each other.
A classic example of this type of study is The American Occupational
Structure by Peter Blau and Otis Dudley Duncan. In this great work, Blau and
Duncan wanted to understand the forces that determine the kinds of
occupations people end up in. They were particularly concerned with the degree
to which parents’ occupations influenced their children’s occupations. Twenty
thousand male respondents filled out a questionnaire on many topics, among
them their race, their occupation and education, and their parents’ occupation,
education, and employment. The occupations were not treated as categories
(doctor, lawyer, and so on) but were converted to a single continuous prestige
scale. Thus, the actual dependent variable was the prestige of the occupation
held by the respondent at the time of the survey (1962). In their basic model,
Blau and Duncan showed that the most important factors in determining a re-
spondent’s current job status were his educational level and the status of his
first job (since the men were of widely varying ages, some had had many jobs).
Nearly all the effects of respondent’s father’s education and job came through
these two “intervening” variables. (That is, father’s education and father’s
occupation affected respondent’s education and first job, which in turn affected
the respondent’s job as of 1962.) The Blau and Duncan study, which of course
had dozens of other findings, helped inaugurate two decades of research on
this process of “occupational status attainment.”
Chapter One: Explanation 11

D. Small-N Comparison
Partway between the detailed analysis of the historical or current reality of a
single case and the statistical analysis of many cases lies a method we can call
small-N comparison. Typically, small-N comparison investigates a handful of
cases, from three to perhaps a dozen. The cases can be many different kinds of
things—bureaucracies, nations, social service agencies, communities, or any
other form of social organization.
The particular form of data gathering employed in small-N analysis can vary.
There are ethnographies comparing several different field sites as well as
histories comparing several different trajectories of nations or classes. Small-N
analysis typically emerges within ethnographic and historical traditions and is
usually seen as a way of improving generalizations by invoking more (and
different) cases. It occasionally arises from the reverse process, in which a
quantitative analyst focuses on a small number of cases to improve his or her
“reading” of the variables.5
Small-N comparison attempts to combine the advantages of single-case
analysis with those of multicase analysis, at the same time trying to avoid the
disadvantages of each. On the one hand, it retains much information about
each case. On the other, it compares the different cases to test arguments in
ways that are impossible with a single case. By making these detailed
comparisons, it tries to avoid the standard criticism of single-case analysis—
that one can’t generalize from a single case—as well as the standard criticism
of multicase analysis—that it oversimplifies and changes the meaning of
variables by removing them from their context.
Small-N analysis has been characteristic of a number of areas in social
science. The field of comparative politics has been built on small-N comparison,
as has historical sociology. In both cases, there is heavy reliance on secondary
literatures concerning the individual cases. Most anthropologists, by contrast,
have gone directly from single-case analysis to abstract generalizations based
on categorization of dozens of cases (for example, in studies of kinship,
totemism, or folklore), although anthropological linguists have often used
comparisons of relatively small numbers of cases.
A classic example of small-N analysis is Barrington Moore’s Social Origins of
Dictatorship and Democracy. This book compares routes to modernity in
England, France, the United States, China, Japan, and India. Germany and Russia
are also considered, but not in depth. Moore’s sources included hundreds of
histories of this or that aspect of each country. After endless reading,
comparison, and reflection, Moore theorized three basic routes to modernity,
all of them depending on how the traditional agricultural classes—lords and
peasants—dealt with the coming of commercial agriculture and the rise of the
bourgeoisie. In the first route, that of England, France, and the United States, a
powerful commercial middle class overthrew the landed classes or forced them
to accept middle-class terms. The result was democracy. In Germany and Japan,
the bourgeois revolution failed, and the landed classes determined the shape
12 Abbott (2004), Methods of Discovery

and dynamics of capitalism as it emerged, leading to fascism. In China and

Russia, an enormous peasant class provided the main force behind revolution,
thus undercutting the drive to capitalism and leading to a standoff between the
revolutionaries in the advanced capitalist sector (the Communists) and the
peasants. Moore’s book provided the stimulus for much of comparative politics
and historical sociology in the 1970s, 1980s, and 1990s.

E. Formalization
There are methods in social science that work without much data at all. Or
rather they work with what are called stylized facts. These methods are not
methods in the usual sense but rather modes of reasoning about social reality
that require some “quasi-factual” input. They are thus halfway between theories
and methods.
A good example of this kind of formalization is analysis of the life table. A
life table is a description of what happens to a cohort (traditionally, 100,000
individuals) after n years of life: how many are still living, what number and
percentage died that year, what the expectation of life is for those remaining,
and so on. By combining life tables with birth-rate information, we can work out
age distributions for a population, investigate the structure of generations,
predict future family structure, and make many other useful demographic
projections. We haven’t gathered new information but have simply worked out
the details implied by the information we already have.
Formalization has gone furthest, of course, in economics, where it has
sometimes lost contact with social reality altogether. But formal thinking is
important throughout social science. The great anthropologist Claude Levi-
Strauss attempted a largely formal analysis of myths, breaking myths up into a
linear, narrative dimension on the one hand and a timeless, structural
dimension on the other (1967). The sociologist Harrison White treated job
markets (like those for clergymen and college presidents) as if they were
electron-hole systems, in which vacancies rather than moving people had the
initiative (1970). Mathematical geographers treat arrangements of political
boundaries as if they were the product of universal mathematical relationships
(Haggett, Cliff, and Frey 1977).
More than any other methodological tradition, formalization lives by
borrowing. By nature, formalization is portable, and many a formal analyst has
made a reputation by borrowing. Economists borrowed much of their formalism
from thermodynamics. Sociologists have borrowed formalisms from physics
and biology.
A good example of formalization is Thomas Schelling’s famous model of
segregation, originally published in 1971 and republished in his remarkable
Micromotives and Macrobehavior. The Schelling model presumes two kinds of
people, one much more numerous than the other, and a neighborhood that
people of both kinds would like to live in. Both groups have a similar “tolerance
distribution,” which describes how willing they are to live in communities of
varying mixes of the two populations. The most tolerant within each group will
Chapter One: Explanation 13

live in a neighborhood as a one-third minority, while the least tolerant will live
only in a totally segregated neighborhood, all of their own kind. Under these
conditions, Schelling shows, the only two stable equilibriums for the particular
neighborhood considered are the fully segregated ones. He goes on to
demonstrate that if the two groups were of equal size and if the most tolerant
of each group were a little more tolerant, there would be a stable fifty-fifty
equilibrium. He also shows that if the larger group included more intolerant
people, there would be a stable integrated equilibrium (because people from
the larger group wouldn’t keep moving into the neighborhood, frightening out
the less tolerant members of the smaller group).
The Schelling models require no real data, only stylized data. But they tell us
something important and counterintuitive. They tell us that even somewhat
tolerant populations have a hard time producing integrated neighborhoods
when the populations vastly differ in size and indeed that sometimes more
tolerance leads to more segregation.6

ETHNOGRAPHY, historical narration, standard causal analysis, small-N analysis,

and formalization are thus five examples of reasonably successful
methodological traditions. Each has its style and its proponents. Each has been
combined with these and other methods in a bewildering variety of ways. I want
to reiterate that these methodological traditions are not associated absolutely
with any discipline, although ethnography and narration are somewhat
associated with anthropology and history, respectively. I also want to reiterate
that these methods do not follow from a single mode of categorization of
methods. As I noted, some are methods of analysis, some are ways of gathering
data, and so on. They are, if anything, best thought of as practices, as ways of
doing social science. As such, they are produced by communities of researchers
who practice them, teach them, and develop them. They are living traditions,
not abstract recipes.

III. EXPLANATORY PROGRAMS

You may be wondering when you would use one of these methods as
opposed to another. Are there hypotheses or empirical problems particularly
well suited to particular methods? The usual answer to this question is yes, and
the usual procedure would be to present here a list of what method is good for
what kind of problem. But my answer to the question of suitability is no. I don’t
think there are methods that are particularly good for particular questions. So I
have no such list. Rather, I will show that the different methods are in fact
aiming to do different things; they envision different kinds of explanations.
That argument takes up the rest of this chapter. Chapter Two then shows how
the standard idea of “well-suited methods” rests on false assumptions about the
methods, and as a result suitability falls apart as a concept. The good news is
that that falling apart creates important openings for heuristics, which are, after
all, what we are looking for.
We begin by seeing how different methods are in fact trying to accomplish
different things. We do this by putting sections 1 and II of the chapter together,
14 Abbott (2004), Methods of Discovery

relating the methods just discussed to the three broad senses of explanation
introduced earlier.
Each of the three senses of explanation defines an explanatory program, a
general style of thinking about questions of explanation. And each explanatory
program has some versions that are more concrete and some versions that are
more abstract. With three explanatory programs, each having concrete and
abstract versions, there are six total possibilities. To give the whole analysis in
simple form ahead of time:
1. Ethnography is a concrete version of the semantic explanatory program.
2. Historical narration is a concrete version of the syntactic explanatory
program.
3. Formalization is an abstract version of the syntactic explanatory program.
4. SCA is an abstract version of the pragmatic explanatory program.
Note that there are two missing possibilities. I shall say very little about one
of them: the concrete version of the pragmatic program. Think of this as simple
experimentation, something we don’t do much of in social science unless you
think of psychology—which involves a lot of experiments—as a social science. I
shall say more about the other missing cell: the abstract version of the
semantic program. Although it has no single name, this is probably the most
rapidly evolving area of methods in the social sciences.
This analysis can be seen visually in the figure. The three dimensions are the
three types of explanations. For
each of these, the origin stands for
explanations focused on everyday
particulars, on commonsense
events. These are an anchor for
each explanatory program, rooting
it in the everyday world. From this
base, “universalizing” moves reach
from the origin toward abstraction
along each of the principal axes of
explanation. The syntactic program
explains the social world by more
and more abstractly modeling its
particular action and in-
terrelationships. The semantic
program explains the world of social
particulars by assimilating it to
more and more general patterns,
searching for regularities over time or across social space. Finally, the purely
pragmatic program tries to separate more and more clearly the effects of
different potential interventions or causes from one another.
The reader should not read this little exercise as a definitive classification of
methods but rather as a way to see that the various methods are in many ways
trying to do different kinds of things. In particular, I am not assuming, as much
Chapter One: Explanation 15

of empirical social science does, that all explanation involves thinking about
causality. We should separate the concept of explanation from that of
understanding the causes of something. Our notion of understanding the
causes of things has become very narrow in social science, in contrast to the
much more general idea of causality that obtains, for example, in the law.
Let me now show in more detail how this argument works. We start with the
programs relating to particulars: concrete, real events rather than abstract
ones. Ethnography exemplifies semantic explanation of particular events, while
historical narration exemplifies syntactic explanation of particular events. Both
are found near the origin of the figure above, but they lie on different
dimensions. This is not because of their difference in temporality but because
of their difference in general explanatory style: translation-semantic type on the
one hand, narrative-syntactic type on the other.
A brief aside about temporality. Temporality is a particularly important issue
in explanation. Some explanations are focused on processes, on the embedding
of social life in moving time. Others devote most of their attention to complex
interrelationships in a static “present”; they think social life takes place within a
given structure, which they treat as fixed for the time being.7 It is important to
recognize that all explanatory programs have temporal and atemporal versions.
For example, there are temporal versions of history (narrative histories like
Thucydides’ History of the Peloponnesian War) and atemporal ones
(descriptions of a moment, like Sir Lewis Namier’s Structure of Politics at the
Accession of George III). Temporality is another dimension I could have used to
classify methods, but I prefer to leave it for later chapters because of the
importance of time in heuristics. What must be emphasized here is that
temporality is not one of the dimensions that differentiates types of explana-
tions or explanatory programs more broadly. All explanations have to think
about time in one way or another.
Returning then to the main argument. In ethnography, the act of explanation
is chiefly semantic. When we say that Malinowski, in his great Argonauts of the
Western Pacific, has explained why the Trobrianders paddle around the islands
giving and receiving shells, what we mean is that he has told us enough about
their culture and their social life that we can understand why they would do
this. We can envision what it is that they see themselves doing, and we can see
what they are doing as reasonable, as something we would do if we were in
their place. The field-worker has translated, however imperfectly, their world
into one that we find comprehensible. Typically, ethnography accomplishes this
by providing detail, by showing ramifications, and by embedding the strange
habits of unfamiliar people in the everyday habits of those same people and
then connecting their everyday world with our own. The ethnographer may have
other professional aims, of course. To return to an earlier example, Evans-
Pritchard takes pains, in Witchcraft, Oracles, and Magic, to explain to us that
the idea of witchcraft serves the epistemological and social function of
explaining unfortunate events, an argument by which he sets forth his
functional theory of culture. But the explanation of witchcraft lies less in the
16 Abbott (2004), Methods of Discovery

syntax of functionalist explanation than in Evans-Pritchard’s ability to translate

the activities of the Azande into something thinkable by Western minds. Evans-
Pritchard does this semantic translation, for example, in his offhand remark
about using the Azande poison oracles to run his everyday life. The Azande
make daily decisions by posing a yes-or-no question (for example, should I do
ethnography today or not?) while feeding young chickens a small dose of
poison. A chicken then makes the decision by living (yes) or dying (no):
I always kept a supply of poison for the use of my household and neighbours and
we regulated our affairs in accordance with the oracles’ decisions. I may remark that I
found this as satisfactory a way of running my home and affairs as any other I know of.
(1976:126)
It is not Evans-Pritchard’s functional theory that persuades, but this homey
detail. Witchcraft, Oracles, and Magic is an explanatory success because of its
semantic virtues, not its syntactic ones.
Of course, ethnography can have pragmatic and syntactic virtues as well.
Ethnography of the drug culture is probably our only effective means to
pragmatic intervention in that culture. And Levi-Strauss’s structural
anthropology had as its chief virtue an extraordinary syntactic elegance that
sometimes amounted to a kind of monomania. But the deep virtue of
ethnography as an explanatory program lies in translation. It is principally a
semantic program.
By contrast, the great virtue of narrative explanation lies in syntax. The
longstanding literature on the philosophy of history is clear on this point. When
Alexis de Tocqueville tells us, in The Old Régime and the French Revolution,
why that revolution came about, he may here and there employ general laws
about social life. But the reason we think his book explains the revolution is
that he tells a followable, reasonable story in which a particular sequence of
events under those general laws leads in some inevitable way to the revolution.
We don’t notice his assumptions of general causal laws (for example, “people
with large amounts of power don’t give it away”). What we notice is the
sweeping story that draws us along with France into the maelstrom of
revolution.
This syntactic strength is, of course, by no means an abstract one. Narration
seems persuasive precisely because telling stories is how we explain most
things in daily life. To be sure, there are some quite abstract narrative concepts:
evolution (in Herbert Spencer’s sociology), habituation (in Max Weber’s
sociology and throughout psychology), dialectical conflict (in Marxian social
analysis), and the like. But these are for scholars.
The real reason we feel that historical narration explains is that narration is
the syntax of commonsense explanation, the one we use all the time ourselves.
So there is no need to justify it. Indeed, the analytical philosophers of history
never could really demonstrate how narration explains; they just said over and
over that it does.
Like ethnography, narration has other explanatory virtues. Narration often
moves us toward a simpler semantic plane. The narrative ideals of followability
Chapter One: Explanation 17

(Gallie 1968) and reenactment (Collingwood 1946) follow the same semantic
principles as ethnography. They measure a narrative’s ability to locate us as
reasonable persons within itself, as people who would have done what was
done had we been the actors of whom we read. And narration can also have
pragmatic virtues. Often, the first step to undertaking action in any particular
situation is developing a narrative of how it got to be the way that it is. But
again, neither of these is a basic virtue. Serious narration explains things for us
because we use unserious narration all day every day. Narration is the syntax of
everyday understanding.
The explanatory programs illustrated by both ethnography and narration thus
appeal to the commonsense world; the first appeal to the commonsense
content of everyday experience, the second to the basic explanatory syntax of
everyday life. Two major streams of explanatory practice in social science grow
out of moves to make these two programs more abstract and formal. (This
means moving away from the origin in the figure on page 12) On the one hand,
we have the attempt to formalize explanatory syntax in modeling and
simulation, which embody what I will here call the syntactic explanatory
program. This is the explanatory practice that is the abstract version of what
narration is at the concrete level. On the other hand, we have the equivalent
effort to formalize semantics, embodied in the family of techniques loosely
known as data reduction and pattern search. This strand is the abstract version
of what ethnography is at the concrete level; I shall call it the semantic
explanatory program. (It is the important omitted cell mentioned a few pages
back, pattern search in its most general version.)
Formal modeling and simulation embody the attempt (atemporal in formal
modeling and temporal in simulation) to improve syntactic explanation by
making it more abstract. The crucial quality sought in the syntactic explanatory
program is elegance. In it, a set of statements “explain” some phenomenon if
they offer a rigorous, complex, yet simple formal representation of it. On the
atemporal side, there are many embodiments of this program: game theory,
classical microeconomics, the Markovian tradition in social mobility analysis,
the group theoretic version of network theory. The temporal side—expressed
most clearly in simulation—has had fewer adherents in social science, although
Jay Forrester gave it a very public demonstration in his studies of industrial,
urban, and world dynamics in the 1960s, and it has returned in the guise of
simulation games. These various methods are astonishingly elegant, some in
their mathematics, some in their simplicity, some in their ability to produce
unexpected results, some in their extraordinary coherence. All are clear,
parsimonious, and in a deep way intellectually pleasing to the abstract mind.
At the same time, these methods share a breathtaking disattention to
semantics, to the reference from model to reality. This is well shown by the
diversity of some models’ application. Microeconomics was systematized by
Irving Fisher (in the early twentieth century) by borrowing whole cloth the
methods of statistical thermodynamics, as if gases and people behaved in the
same way. Group theory (a particular branch of modern algebra) saw major
18 Abbott (2004), Methods of Discovery

application in crystallography and in pure mathematics as well as in sociology’s

network theory and even anthropology’s kinship analysis. Game theory has
journeyed from psychological experiments to explaining the stock market and
modeling family-planning decisions. Of course, proponents of the syntactic
program argue that semantics in fact doesn’t matter. These empirical realities
all have the same general semantic form, they say, and so one can write
abstract syntax for them.
But most readers find the semantic assumptions of the syntactic program
quite worrisome. What is the point of game-theory models if we can write ten
different models for any given social situation? We must choose between those
models on semantic grounds, and about those semantic grounds the syntactic
program tells us nothing. What is the point of admiring the elegance of
microeconomics if microeconomics frankly admits that preferences cannot be
generated from inside the system without undercutting the assumptions of the
whole edifice? Essentially, microeconomics is telling us that if we can explain
what people want to do, it can then explain that they do it. So what?
In summary, the syntactic program buys elegance and breadth at the price of
semantic indeterminacy and limitation. By contrast with this syntactic
explanation via elegant and highly general arguments, the semantic program
seeks to explain social reality by a different kind of abstraction. It directly
simplifies the complexity of the social world, turning it into a reduced
description that a reasonable reader can grasp with the syntax of everyday
explanation. Thus, techniques like cluster analysis and multidimensional scaling
take data of enormous detail and turn it into simple categories and pictures.
Pierre Bourdieu, for example, “explained” consumption patterns in France (in
his book Distinction) by showing that those patterns constitute a language of
class distinctions. From the reader’s point of view, the explanation is a matter
of common sense once Bourdieu has visually presented the “geometry” of the
consumption patterns by using a scaling technique that turns raw data on
people’s preferences for cultural materials into a picture locating types of
goods and types of people on the same map.
The semantic program has been strong in psychology and particularly strong
in market research; marketers routinely use cluster analysis to reduce the
American consumer market to one hundred or so basic types of consumers. In
that sense, the semantic program has shown considerable pragmatic strength
as well. (These are the techniques that are used to figure out your consumption
preferences from your Internet use, for example.) On the syntactic side,
however, the semantic program has been weak. Its overwhelming focus on one-
time analysis makes it static. It can abstractly describe a state of affairs but
cannot account for how it changes. Network analysis is one of the glories of
abstract semantic explanation, but there is still no real conceptualization for
the temporal development of networks. Only when some researchers recently
began to think about applying pattern search techniques to over-time data did
any kind of syntactic development arrive in the semantic program. In short, as
Chapter One: Explanation 19

with the syntactic program, power of one type was bought at the price of
indeterminacy of the other.
I have so far described concrete and abstract versions of the syntatic program
(history and formal modeling, respectively) and concrete and abstract versions
of the semantic program (ethnography and pattern search, respectively). There
is a third abstracting move in social scientific explanation, the one that moves
out from the origin along the pragmatic dimension of the figure on page 12.
Oddly enough, this program has become so successful that social scientists
have forgotten that pragmatics is its origin. This is the program carried out by
the standard forms of causal analysis in social science, both analysis of the
cross-sectional type (as in structural equations models or path analysis) and of
the temporal type (as in durational models). Because the SCA program is so
dominant in empirical social science, we need to look at it in some detail.
The SCA paradigm arose out of a rationalization of the methods it uses,
methods that were originally used to interpret practical experiments. As we saw
earlier, these methods work by taking apart the complex particulars in the data
(the cases) and treating them as intersections of abstract, universal properties
(the variables). Analysis then isolates one of those variables—an arbitrarily
chosen dependent variable—and searches out the effects of the other, so-called
independent variables on it. Interaction effects—that is, effects arising from two
or more variables “working together—are treated as secondary.
The great explanatory virtue of this method, as originally conceived, was
pragmatic. Sir Ronald Fisher and his followers devised these statistical
techniques in the 1920s and 1930s to test the effects of experimental
manipulations. Should one add fertilizer or not? Was soil A better than soil B?
They put the fertilizer on some fields but not others, measured the effects, and
figured out a probability theory for the resulting numbers. They had no
particular concern for causes, for why or how growth happened. The point was
to decide whether to take some action, not to understand mechanisms. Since
the original applications were experimental, these statistical techniques were in
fact explanatorily quite persuasive for the pragmatic purpose they served. Used
in an experimental context—as they still often are in psychology—they remain
so.
Later in the century, however, this approach was applied to nonexperimental
data and combined with new ideas about causality. This led to the hybrid
explanatory program that is now general throughout the empirical social
sciences, the standard causal analysis program. The SCA program still has
some pragmatic relevance; the methods are still used in evaluation research, for
example. But its main uses are not now pragmatic. Rather, they pretend to be
syntactic. So we say (using the weighted-sums approach mentioned earlier) that
differences in wages in civil service systems are “caused by” gender, bureau-
cracy, unionization, and so on. Semantically, of course, this whole language of
variables is a mirage. The words gender and bureaucracy do not refer to real
entities. Gender and bureaucracy do not exist as independent things; they exist
only as properties of real things (in this case, of civil service systems). So this
20 Abbott (2004), Methods of Discovery

“properties” syntax has to be justified by further semantic reference. We have to

have some way to give empirical meaning to statements about relationships
between abstract things like gender and bureaucracy. In economics, this
semantic reference is made to formal and simplified models of action. So
typical economics articles in the SCA tradition justify their SCA with a mass of
formalizing and calculus that typically begins each article. In sociology and
political science, this external reference is made to a set of simplified
narratives. So sociology and political science articles of the SCA type begin not
with the calculus of the economists but with commonsense historical narratives
of the form “such and such people are likely to do such and such things under
such and such conditions.” These stories try to justify the “variables-level
syntax” by reaching toward the semantic world of everyday reasonable
understanding. Thus, in order to be explanatory, the SCA program has to com-
bine its variables-level causal syntax with unrelated semantic references to
other, more credible syntactic approaches to reality: stylized action in the
economics case, followable narratives in the sociology one.
All of this complexity happens because in reality the SCA program has no
causal foundation at all; it was originally designed to help us make decisions, to
be pragmatic. Dressed up as a syntactic program, it is ungracious and silly. (It is
also surprisingly difficult to learn, since its rationale—as this long discussion
shows—is quite tortured.) Its strongest point remains its ability to tell us about
the comparative size of variables’ pragmatic effects on other variables, given
the implicit assumption that we have a quasi-experimental situation (which we
almost never do). But it can’t even tell us in which direction the causal forces
work nor how causes work together. All of those judgments must be imported
from elsewhere.8
In summary, there is no free lunch. Strongly developing any one aspect of
explanation ends up losing much of the rest. In particular, the present moment
in social science is probably one in which the syntactic and semantic programs
are about to turn the tables on the pragmatic one, which has dominated social
science for about sixty years. The latter remains the best program when we
think about social policy. But if we are trying to understand why and how things
happen, it has little to recommend it.

Notes:
1 As the great anthropologist Evans-Pritchard once remarked,
Anyone who is not a complete idiot can do fieldwork, and if the people he is working among have not been
studied before he cannot help making an original contribution to knowledge. . . . Anyone can produce a
new fact; the thing is to produce a new idea.” (1976:243)
In the more theoretical phraseology of Imre Lakatos (1970:132ff.), the most important
quality of research programs is their “heuristic power,” their ability to keep producing new
ideas and point the way to new findings.
2. Among many writers who have made the case for beauty” in scientific argument, see
Chandrasekhar (1979).
3. Syntactics, semantics, and pragmatics are the three fundamental aspects of all systems of
signs, of which explanation is an example. See Morris (1938),
4. The words for denoting methods are changing. Properly speaking, a method is a set of
routine procedures for rigorous inquiry. Methodology is (literally) discussion of methods.
Chapter One: Explanation 21

Ethnography or standard causal analysis (SCA), then, is a method, while to write about
ethnography or SCA is to write methodology. In practice, people are now often using
methodology to mean “method,” as in the familiar seminar question, “What’s your method-
ology.” Note that people using these terms do not customarily use methodical as the adjective
form of method; they use methodological, which is thus the adjective form used for both
method and methodology. I have tried to maintain the traditional distinction between method
and methodology throughout.
5. Sometimes quantitative analysts do undertake detailed study of several cases. For an
example, see Paige (1975).
6. Of course, when we look at the facts, the situation is much more drastic. Black and white
tolerances are by no means as auspicious as the Schelling models presume. It is, then, hardly
surprising that American neighborhoods stably integrated at any ratio beyond 20 percent
black are extremely rare.
7. One should not necessarily think that one or the other of these has priority as a mode of
thinking even about causality. If we consider the literature on causality, there have been
distinguished exponents both of the idea that causality must involve passage of time and of
the idea that it cannot involve passage of time. See Abbott (2001b:c. 3).
8. To save space, I do not comment in depth on the temporal versions of the SCA program. But
in fact, the same discussion governs them. To be sure, they are embedded in time and
because of that acquire a semantic verisimilitude that cross-sectional studies lack. But they
still function on the semantic level of variables, far removed from narrative understandings of
the unfolding of events. Durational methods can predict “particular” events, like the passing of
a law or the founding of a newspaper, but they do so with the same kinds of disembodied
variables (rather than complex particulars) that are used by cross-sectional methods. So they
remain at a considerable semantic distance from immediately familiar worlds.