Paradigms and

Sand Castles
Theory Building and Research Design
in Comparative Politics
FLACSO- Biblioteca
BARBARA GEDDES

The University of Michigan Press

Ann Arbor

f?nn2QB24

wrs.
CU T.

r'T.' 'OT' CN - RACSO

Copyright by the U niversity of M ichigan 2003
AH rights reserved
Published in the United States o f A m erica by
T h e U niversity o f M ichigan Press
M anufactured in the United States o f A m crica
Printed on acid-free paper
2006

2005

2004

2003

4 3 2 1

N o p ari o f this publication m ay be reproduced, stored in a retrieval system , or

transm itted in any form or by any m eans, electronic, m echanical, or otherw ise,
without the written perm ission o f the publisher.

A C1P catalog record fo r this book

is available from

the British Library.

Library o f Congress Cataloging-in-Publication D ata

G e d d e s. B arbara.
Paradigm s and sand castles : theory building and research design
in com parative politics / B arb ara G ed d es.
p.

cm. (A nalytical perspectives on politics)

In d u d cs bibliographical references and index.

isbn

0-472-09835-7 (cloth . alk. paper) isbn 0-472-06835-0

(pbk. : alk. paper)

1. Com parative governm ent.
I. T itle.

2. Political science R esearch .

IT. Series.

JA 8 6 .G 3 5

2003

320.3 d c2i

2003002160

1
'* - A .
I

1 r : v ' ... :
r
i

Analytical Perspectives on Politics

A D V I S O R Y EDI TORS:

Jo h n A ld rich , D uke U n iversity

B ru c c B u en o de M esqu ita, H oover Institution and N ew Y ork U niversity
Robert Jack m a n , U niversity o f C alifo rn ia , D av is
D avid R ohde. M ich igan State U niversity
Political S c ien c e is d evelop in g rapidly and ch an gin g m arked ly K e e p in g in
touch with new ideas across the d iscip lin e is a ch allen ge for p olitical scientists
and fo r their students.
T o help m eet this ch allen ge, the series A n a ly tical P ersp ectives on Politics
presents creativ e and sophisticated syn theses o f m ajo r areas o f research in the
field s o f p olitical science. In each book, a high -caliber author p rovid es a clear
and d iscrim inating description o f the current state o f the art and a strong m inded
prescription and structure fo r future w ork in the field.
T h ese distin ctive books p rovid e a com pact rev iew fo r p olitical scien tists, a
helpful introduction for graduate students, and central reading fo r advanced
undergraduate courses.
Robert W. Jack m a n , Pow er without F o rce : The Political Capacity o f

Nation-States
L in d a L. F o w ler, Candidates, Congress, and the Am erican Dem ocracy
O le R. H olsti, Public Opinion and American Foreign Policy
Scott G ates and B rian D. H u m es, Gam es, Information, and Politics:

Applying Gam e Theoretic M odels to Political Science

Law ren ce B au m , The Puzzle o f Ju d icia l Behavior
B arb ara G ed d es, Paradigm s and Sand Castles: Theory Building and Research

D esign in Com parative Politics

To John

26

Paradigm s and San d C astles

may not have had much experience with it or who may have
misconceptions about it. The chapter also reviews the many sub
stantive areas in which rational choice arguments have been
used, so that those interested will know where to find out more. I
do not seek to proselytize for the rational choice approach. In
stead, I use it as an example of an approach that incorporates
many of the features that contribute to theoretical fruitfulness.
We should keep these features in mind as we search for ap
proaches that suit particular subjects better than rational choice
does.
The choice of approach logically precedes the proposal of
hypotheses and testing, though one may, of course, test hypothe
ses drawn from multiple approaches. If the order o f chapters in
this book followed this logic, the chapter on approaches would
come near the beginning. I put it at the end, however, because I
do not want rational choice to frame the whole book. I believe
the methodological suggestions made here are relevant and use
ful to anyone who seeks to explain political outcomes, regardless
o f approach.
The book covers a narrow and perhaps idiosyncratic range of
topics. Certainly there are other aspects of research design that
deserve attention. I do not claim that if we follow the advice in
this book, scholars in other disciplines will stop responding to the
term political science with amusement. We have a long way to go
before the wishful thinking embodied in that label approaches
reality, and some believe that the aspiration displayed in that
name is not only unrealistic but undesirable. For those who find
science as a vocation 1* a compelling goal, however, the advice
in this book aims to prevent casual, uninformed, or unintended
ventures off the long path leading to that goal. It aims not to
provide a set of clear and mechanical rules of research design,
but rather to foment thoughtful and innovative ways of using the
inadequate and fuzzy evidence we actually have in order to build
theories.

14. Science as a Vocation is the title o f M ax W ebers inspiring essay on the

pursuit o f knowledge (W eber 1958, 129).

CH APTER 2

Big Questions, Little Answers

How the Questions You Choose Affect
the Answers You G et

For the scholar who wants to contribute to the accumulation of

knowledge, the first step in the process is choosing a question to
investigate. This chapter makes some suggestions about how to
shape research questions to increase the likelihood that they will
yield compelling and robust theories. The early part of'the chap
ter is an attempt to articulate some of the values and emotions
that I believe motivate good scholars. These values and emotions
undergird much research, but they are rarely expressed. On the
contrary, the advice given to beginning scholars often implies the
opposite values. Here I discuss the role of curiosity, indignation,
and passion in the selection and framing of research topics.
In the second and much longer part of the chapter, I suggest a
change in the way we usually think about the kinds of big, worldtransforming subjects that comparativists often choose to study.
Large-scale phenomena such as democratic breakdown, eco
nomic development, democratization, economic liberalization,
and revolution result from the convergence of a number of differ
ent processes, some of which occur independently from others.
No simple theory is likely to explain such compound outcomes.
In principle, a complex, multifaceted theory might successfully
do so, but in practice the task of constructing such theories has
daunted most analysts. I propose changing the way we approach
these questions. Instead of trying to explain such compound
outcomes as wholes, I suggest a focus on the various processes
that contribute to the final outcome, with the goal of theorizing
these processes individually and generating testable propositions
about them. In contrast to much of the methodological advice
given in this book, the suggestions in this chapter do not derive
from the logic of quantitative research. I cannot make any claim
that this research strategy is more correct than any other. My
27

28

Paradigm s and San d C astle s

argument rests, rather, on the judgment that it is a more effective

route to an accumulation of theoretical knowledge. The proof of
the pudding is in the eating, however, and until we have some
pudding, we cannot taste it. The last part of the chapter is an
extended example of how to break up a big question into multi
ple processes, theorize one of them, and then test one of the
implications o f the theory thus devised.
"Science as a Vocation," Not Just a Job: Choosing a
Research Topic

Students are often advised to choose research topics by looking

for holes in the literature or by reading the ads in the Personnel
Newsletter of the American Political Science Association to see
what topics are hot. This advice conveys the impression that the
search for research topics can and perhaps should be methodi
cal and instrumental. This impression is false, and the advice,
if followed, leads to a number of perversities; taking The Lit
erature seriously whether it merits it or not; the selection of a
topic that will become outdated before the dissertation is done;
boredom.
Curiosity, fascination, and indignation should guide the choice
o f research topic. Emotion has been banned from most of the
research enterprise, and properly so. But one place it should
remain is in the choosing of research topics. The standard advice
on how to choose a topic leaves out the role of such emotions as
commitment, irritation, and obsession.
An especially thoughtful version of the standard advice is ar
ticulated by G ary King, Robert Keohane, and Sidney Verba
(1994, 15 - 17 ) , who advise students to pick topics that are impor
tant in the world and that contribute to an identifiable scholarly
literature. Beginning scholars are advised to:
1. Choose a hypothesis seen as important by scholars in
the literature but for which no one has completed a
systematic study. . . .
2. Choose a hypothesis in the literature that we suspect is
false (or one we believe has not been adequately con
firmed) and investigate whether it is indeed false. . . .
3. Attempt to resolve or provide further evidence of one
side of a controversy in the literature perhaps demon-

B ig Questions, Little Answers

29

strate that the controversy was unfounded from the

start.
4. Design research to illuminate or evaluate unques
tioned assumptions in the literature.
5. Argue that an important topic has been overlooked in
the literature and then proceed to contribute a system
atic study to the area. ( 16 - 17 )
It would be difficult to disagree with any of this advice, and it is
probably very useful to students in the sciencelike parts o f social
science. It assumes, however, that the relevant literature really
does contain a considerable accumulation of theory and stylized
facts. It thus fails to take into account the real state of a good deal
of the literature in comparative politics. The literature on some
subjects contains only a few arguments generally accepted as true;
many controversies in which the hypotheses on both sides lack
both clarity and strong empirical support; and large amounts of
opinion and conjecture, unsupported by systematic evidence but
nevertheless often referred to as theory. Such a literature creates a
fuzzy research frontier. The reader finds not well-defined holes in
the literature but swampy quagmires. Students who wade into
these literatures often find themselves sinking into the quicksand
of contested definitions and chasing after nebulous dependent
variables that flit around like will-o'-the-wisps.
Consequently, good research in the field is more often moti
vated by curiosity about the world and intuition about cause-andeffect relationships than might be true in a field with more accu
mulated knowledge and a more clearly defined research frontier.
Much of what is eventually judged to be exciting research in the
comparative field either addresses subjects not covered in the
literature or addresses old subjects in very novel ways, rather
than extending the existing literature.
Contrary to the advice about looking for holes in the litera
ture. good research in the comparative field often begins either
with an intense but unfocused curiosity about why some event or
process has happened or with a sense of sputtering indignation at
the patent idiocy of some particular argument advanced in the
literature. Sometimes political commitments or an aroused sense
o f injustice drive this curiosity or indignation. Potential research
ers who feel little curiosity, intuition, or indignation in response
to the social world and the arguments published about it should
consider the possibility that they have chosen the wrong job.

30

Paradigm s and San d C astles

The literature does play a role in the choice of a research

topic, but not in the mechanical way suggested by the standard
advice.
It stimulates the indignation, annoyance, and irritation that
often fuel good research. When one reads an argument, finds it
utterly implausible, and believes that one can find evidence to
demonstrate that it is wrong beyond the possibility of refutation,
one has indeed found a hole in the literature. Generally, how
ever, these holes are not found through coolheaded searches.
Instead, our gut-level response of irritation causes us to pause
and notice them while we are reading for some other purpose.
Moreover, such holes cannot be found unless the reader has
sufficient background knowledge of facts to notice that the argu
ment seems inconsistent with reality.
Arguments in the literature also create expectations about
how events will play out in as yet unexamined cases. When we
have some information about such cases that leads us to believe
they may not meet expectations, our curiosity is aroused. Cases
and outcomes may capture our interest because they differ from
other cases or from what theory has led us to expect. Such out
comes call for explanation because they are anomalous when
compared with other known or apparently understood instances.
A t this stage, the comparison may be entirely implicit, and the
analyst may focus on the anomalous case; but without the im
plicit comparison, there would be no basis for considering the
case interesting or puzzling.
I emphasize the emotional aspects of choosing research topics
because these emotions contribute to the intense commitment to
finding out what really causes things to happen that leads to good
research. A s Max Weber asserted, Without this strange intoxica
tion, ridiculed by every outsider; without this passion . . . you
have no calling for science and you should do something else
(19 5 8 ,13 5 ).
Fostering Creativity

In the same essay, Weber also stressed the importance of the

nonmethodical aspects of thought intuition and inspiration. He
emphasized the importance of having ideas, that is, of creativity:
Certainly enthusiasm is a prerequisite of the inspiration
which is decisive. Nowadays . . . there is a widespread no-

B ig Questions, Little A nsw ers

3i

tion that science has become a problem in calculation, fabri

cated in laboratories or statistical filing systems just as in a
factory, a calculation involving only the cool intellect and
not ones heart and soul . . . . [but] some idea has to oc
cur to someone's mind . . . if one is to accomplish anything
worthwhile. And such intuition cannot be forced. It has
nothing to do with any cold calculation. (1958, 135)
Creativity is distributed as unequally among us as everything
else, and very little is understood about it. Nevertheless, I be
lieve that the way we train ourselves in graduate school and for
the rest of our lives determines how much combustible material
our creative sparks will find to ignite. Weber stressed that the soil
in which ideas grow is normally prepared by very hard work. I
will go further and suggest that some kinds of hard work are
more likely to bear fruit than others.
Original ideas grow out of having individual and autonomous
reactions to the world. We can have such reactions only on the
basis of our own inner sense o f how the world works. The task of
the apprentice scholar, therefore, is to develop this inner sense.
This process can be helped along in a somewhat conscious and
systematic manner. Good scholarship arises from the interaction
of observation and conjecture. We can intentionally increase the
amount o f observation we have to draw upon and thus deepen
our ability to speculate fruitfully; we do this by exposing our
selves to large amounts of information, whether by wide reading
about many countries and over long historical periods or via the
scrutiny o f masses of quantitative data. However it is done, the
scholar is filling his or her stores with information within which
to hunt for patterns and with which to probe the plausibility of
hypotheses. I would urge all students to get into the habit of
creating formal or informal data sets, that is, collecting and
storing in some place other than their own fallible brains large
quantities of factual information. (For some kinds o f informa
tion, Excel spreadsheets are the perfect storage medium, but in
other situations there may be no substitute for old-fashioned
index cards.)
The kind of information that should be collected, of course,
depends on the scholars interests. Whatever the topic, however,
it is always useful to find out about it in countries and times
outside ones primary area o f expertise. For example, if the stu
dent's interests center on how oil wealth has affected government

32

Parad igm s and S an d C astles

in Middle Eastern countries, he should also stockpile some infor

mation on forms o f government and uses of oil revenues in coun
tries of other regions. The student interested in the effects of
political institutions on the development of party systems in new
democracies should resist the temptation to base her speculations
on the experience of the countries most thoroughly covered in
the literature for example, Chile, Argentina, and Brazil and
should make sure she knows basic facts about electoral institu
tions and party systems in the small, less studied countries of
Latin Am erica and, if at all possible, in the new democracies of
the rest of the world.
Having this kind of factual knowledge base helps the scholar to
avoid making unfounded claims about the uniqueness of particu
lar events, processes, or countries and also to avoid mistaking the
simplified portraits of events often found in the literature for realis
tic descriptions. Much of the literature on many subjects in the
comparative field is dominated by descriptions of events in a few
much-studied countries. Much of the transitions literature on L a
tin Am erica, for example, focuses on Brazil, Argentina, and Chile
(along with Spain). Scholars working on transitions in other parts
of the world assume that this literature accurately describes the
general transition experience of Latin Am erica, but it does not.
Transitions have also occurred in Bolivia, Ecuador, Paraguay,
Peru, and most of Central Am erica. In some ways, the transitions
in most of these latter countries more closely resemble those of
African countries than those of the more often studied and more
industrialized Southern Cone countries. If ones knowledge of
Latin American transitions comes solely from the best known
transitions literature, conclusions about differences between L a
tin American transitions and those in other regions will be inaccu
rate. Increases in factual information, however, improve the
chance of finding the patterns that really exist.
A second aid to creativity arises from becoming fluent in the
use of various kinds of models. When models even such simple
ones as the prisoner's dilemma enter our imaginative reper
toires, they make possible interpretations of information that
simply would not have occurred to us otherwise.
A model is a simplified representation of a process. Its pur
pose is to illuminate a basic logic underlying the process that
might not be perceptible from observation of the entire compli
cated reality overlaid, as all reality is. with multitudinous irrele
vant details. A good model one that is useful, fruitful, or

FL A C S O

Pihiinteca
Big Questions, L iule Answers

33

exciting shows both its creator and those who are exposed to it
something about the process that they had not perceived before.1
When a model seems to fit the essential features of a situation, it
enables the analyst to understand that situation more clearly and
deeply than before. It also aids in communicating this under
standing to others.
The collective action problem, usually expressed in purely
verbal terms, is probably the best-known example of a model
that simply changed the way we understand the world. Prior to
the dissemination of the idea that individuals will not find it
rational to expend their own resources in order to secure public
goods for groups of which they are members, the failure of vari
ous disadvantaged groups to organize politically was considered
puzzling. Much ink was spilled explaining false consciousness.
Since Mancur Olson's very striking articulation (1965) of the
collective action problem, our baseline expectations about politi
cal mobilization have been inverted. We now find it puzzling,
and hence worthy of explanation, when large groups do manage
to organize in order to press for some public good.
Another widely used model is the idea of evolutionary selec
tion. The central idea here is that outcomes may occur in the
absence of intentional decision making because the actors, orga
nizations, states, parties, or other entities that fail to behave in
certain ways will lose office or go out o f existence. Thus the only
ones that remain will be those that did behave as required, even
though they may not have understood their situation or made
conscious decisions about it. Probably the most famous example
o f the use of this logic comes from Richard Nelson and Sidney
Winter (1982), who found that managers of firms do not really
think much about maximizing profit. Nevertheless, they argue,
firms behave as though their managers sought to maximize
profit, because the firms of those managers who deviate greatly
from what they would do if they were maximizing profit go bank
rupt. The same logic can be used to explain the prevalence of
contiguous territorial states as the main large-scale form of gover
nance in the world today. A t one time, many rulers laid claims to
noncontiguous pieces of territory, and they did not decide to give
up outlying bits in order to concentrate on consolidating their
rule in the contiguous areas. Instead, wars, uprisings, and the
i.
F o r an extensive and wonderfully useful discussion o f m odels in the social
sciences, see L a ve and M arch (19 75).

34

Paradigm s and S an d C astles

spread of nationalism led to the consolidation o f contiguous

states that were relatively large (compared to what preceded
them) at the expense of smaller and more scattered ones. In
other words, even though rulers may not have consciously sought
to limit their domains to contiguous areas, competition among
them eliminated noncontiguous areas, which were much more
difficult to defend militarily, and allowed the rulers of large con
tiguous areas to consolidate their territorial claims at the expense
of others.
The two models described here can be applied using only
words. There are many others that usually need to be expressed
mathematically or graphically because the processes they exam
ine are too complicated to be captured easily by words alone.
Widely used models include divide-the-dollar games, which illu
minate how different rules and time horizons affect the outcome
of bargaining over distribution; signaling models, which describe
the effect o f costly symbolic actions on the perceptions of others;
information cascades, which describe the diffusion o f changes in
information or perception of risk; spatial models o f preferences,
used to think about voting behavior, policy choice by legislators,
and lots of other issues; prospect theory, which models the effect
of prior gains and losses on risk aversion; and contagion models,
which can be used to think about anything from the diffusion of
technological innovation to the spread of religious fundamental
ism. The internal logic of these models is too complicated to be
fully and simply articulated in words. In such arguments, equa
tions and graphical representations are used in addition to verbal
descriptions as a way of making all aspects of the logic precise
and clear.
Even if the student has no interest in becoming proficient in
the use of such models, exposure to them enriches the theoretical
imagination. It improves the quality of our speculations, which
are, in the words of Charles Lave and Jam es March (1975, 2),
the soul of social science.
A form of hard work that seems to me much less likely to
fertilize the soil in which the imagination may grow is the kind of
reading that is often considered preparation for qualifying ex
ams. Being able to read the introductions and conclusions of
great books in order to summarize the main argument in a few
sentences is a skill in its own right, a skill often rewarded in
graduate school. But it does not seem to be correlated with the
ability to do imaginative research.

Big Questions, Little A n sw ers

35

When one reads, whether in preparation for qualifying exams

or not, one should ponder and even brood over the discussions of
why one thing causes another. This, not the simple identification
o f cause and effect, is the crux o f a work o f scholarship, and the
reader needs to think about whether it rings true, whether it fits
with what he thinks he already knows about the world. If individuals are not the unit of analysis in the argument, it is very
useful to think through which individuals would have to be motivated, in what way, and to do what in order for the argument to
hold. If individuals are the main actors in the argument, it is
useful to ask oneself whether the motivations implied by the
argument seem to be plausible accounts of how, on average,
people behave.
Readers should also scrutinize the evidence the author offers
to support the argument. They should never accept an authors
assertion that evidence supports the claims made without looking
at the evidence and thinking about it. One cannot assess the
evidence supporting arguments without reading the middle parts
of books. II one does not have time to read everything one
should and one never does it is better to read carefully what
one can of the evidence than to read only the introduction and
conclusion for a summary of the argument.
Although it is all-important to absorb both information about
the world and models o f how information can be organized and
interpreted, this is not enough. Scholars must also constantly,
though often implicitly, ask themselves the question, What do I
think? Do I believe this? Students cannot develop an autono
mous reaction to the world by constantly worrying about what
others think. They must worry about what they think them
selves, and make sure they think something. The vocation o f
science is not for the other-directed. The gradual accretion of
thoughts entertained in response to information and models will
be the basis of ones own creative ideas and scientifically impor
tant discoveries.
The Mentor's Role

I turn now to the delicate subject of mentors. Having an appren

tice relationship with an experienced scholar can be a very useful
training experience. The student can learn how a seasoned scholar
approaches intellectual puzzles and how to make practical use of
the statistical and modeling tools acquired in classes. Typically, an

36

P arad igm s and San d C astles

experienced researcher has figured out, stumbled on, and bor

rowed lots of tricks and efficiencies over the years that can be
passed on to students. The student also gains professional social
ization and sometimes a leg up in the job market via coauthored
publications. These are the advantages of a close mentoring rela
tionship, and they are very substantial.
The mentoring relationship can be intellectually seductive to
the student, however. Graduate students, like the peasants de
scribed by James Scott (1976), feel powerless in an unpredictable
world. Among other survival strategies, students often attempt
to cultivate patron-client relationships with faculty members,
who they hope can protect them from the various hazardous
forces of nature they face.2 In this environment of situationally
induced dependence, students may become so imbued with the
mentor's worldview and research project that they dismiss evi^
dence that conflicts with the mentors arguments. Students may
even experience something akin to hostage syndrome, in which
they come to identify completely with the mentor's point of view,
feeling that all the adviser's opponents and all other ways of
thinking are wrongheaded or even contemptible. Such narrow
minded partisanship is a rather common but perverse result of
the mentoring relationship. Students should guard against it, and
mentors should make all possible efforts to limit students' natu
ral impulse toward partisanship. When students rely so heavily
on the mentor, they may be unable to conceive of research proj
ects other than subsets of the mentors research.
Advisers may, through inertia or inattention, seem to want
students to defer to all their ideas, but what good advisers really
want is for their students to be unafraid to challenge them in
sophisticated and well-informed ways. The best scholars are not
the best research assistants in graduate school, but rather those
who challenge, extend, and go beyond their teachers, and good
advisers know that.
B y the dissertation stage, a student should be perched on the
mentors shoulder, having absorbed what the mentor has to teach
and poised to take off in independent flight. He should not be
huddled under the mentor's protective wing. It is part of the men
tor's job to push reluctant fledglings out of the nest if necessary.
Just as young people in the West do not allow their parents to
2.
Students also m ake use o f the weapons o f the weak*' noted by Scott gossip,
slander, ostracism , and shirking to punish the village notables who fail to perform
their allotted roles in the departm ental moral economy.

Dig Questions, Little Answers

37

choose mates for them, students should not allow advisers to

choose their dissertation topics. They should listen carefully to
the advisers advice, as one listens to parental advice, but ulti
mately the scholar must feel an intense fascination in order to
sustain the commitment needed for such a massive research en
deavor. The average comparativist lives with a dissertation topic
for between eight and twenty years, from starting to think about
it to publication of the dissertation-book and possible spin-offs
and extensions. That is as long as many marriages last. Many
comparativists continue working on their dissertation subject for
the rest of their careers. No one but the person who will be
putting in this massive amount of time and effort is really quali
fied to choose the subject.
Romantic Questions, Reliable Answers

Having allowed passion, fascination, or indignation to influence

the choice of topic, the researcher then faces a very different
kind of task: devising a research strategy. Many of the classic
works in the comparative field focus on big, romantic questions,
and the same kinds of questions draw many into the field. The
choice of a strategy for investigating such topics requires methodi
cal thought as well as romantic attraction. Outcomes such as
democratization, the collapse of empires, and revolution result
from the convergence of a number of different processes, some
o f which may occur independently of others. Insufficient atten
tion to research strategy when approaching such big questions
accounts for quite a few sand castles.
Because the complex outcomes are rare and undertheorized,
inductive research strategies prevail. Either researchers immerse
themselves in the history and social structure of a few cases that
have experienced the outcome of interest and come up with a list
of events and characteristics that predate the outcome, or they
cull indicators of potential causes from large public data sets and
plop them into statistical models. Thus, the implicit or explicit
model of explanation, even for those who reject quantitative
research, turns out to be a kitchen-sink regression. But correla
tion is not causation, even in nonquantitative research.
A t its best, this unstructured inductive approach to investigat
ing complex social outcomes is analogous to that of mcdical re
searchers who try to understand the onset of cancer by amassing
data on all the dietary and environmental factors that correlate

38

Paradigm s and S an d C astles

with an increased incidence of the disease. These studies are

useful. They lead to the accumulation of hypotheses, some of
which are ultimately confirmed and some not. But though this
sort of fact-collecting has been essential to the origin of many
significant sciences, Thomas Kuhn notes that anyone who exam
ines famous instances of pretheoretic work will discover that it
produces a morass (1970, 16). It does not by itself lead to an
understanding of the process through which cancer develops. For
that, researchers have had to step back from the aggregate out
come, the diseased person, and focus instead on basic mecha
nismsfor example, the factors that regulate cell division and
death. They must concentrate on the units within which the pro
cess occurs (the cell and the gene) rather than on the outcome
(the diseased organism).
In a similar manner, students of comparative politics need to
seek to understand underlying political processes rather than to
explain, in the sense of identifying the correlates of, complex
outcomes. What I am proposing here bears a resemblance to the
research strategy that Robert Bates et al. have called analytic
narratives. I concur with their belief that we need to
seek to locate and explore particular mechanisms that shape
the interplay between strategic actors and that thereby gen
erate outcomes. [We need to] focus on the mechanisms that
translate such macrohistorical forces into specific political
outcomes. B y isolating and unpacking such mechanisms,
analytic narratives thus contribute to structural accounts.
( 1998, 12 - 13 )
In order to unpack these mechanisms, we need to focus on the
fundamental unit of politics, in most cases individuals. We need
to break up the traditional big questions into more precisely
defined questions about what individuals do in specific situations
that recur often enough to support generalizations about them. I
depart from Bates et al., however, in that I see analytic narra
tives as an essential part of the research enterprise, but not its
end product. A carefully constructed explanatory argument built
up from fundamentals usually has multiple implications, at least
some of which are testable. The research effort is not complete
until empirical tests have shown that implications drawn from the
argument are consistent with reality.
Figuring out the implications of an argument involves repeat

B ig Questions , Little A nsw ers

39

edly asking, If this argument were true, what would I see in the
real w orld? Some scholars seem impelled by intuition to engage
in this kind of reasoning, but anyone can train himself to do it as
part of a regular routine. To demonstrate deriving implications
from an argument, let us use Barrington M oores famous apho
rism no bourgeois, no democracy 3 (1966, 418) as a simple
example. Since there are no contemporary societies that are liter
ally without a bourgeoisie, and since the aphorism is stated in
absolutes but the world is probabilistic, it can be restated in
social sciencese as: The likelihood of democracy increases once
the size of the bourgeoisie has passed a certain threshold. If
bourgeois is taken to refer to the commercial and industrial bour
geoisie but not government bureaucrats, the implications of this
argument include the following:
Democracies would not be expected to occur before the
industrial and commercial revolutions.
The establishment of democracies would be expected
first in the countries that industrialized first.
In the contemporary world, democracy would be more
likely in more industrialized countries.
Democracy would be less likely in countries in which
wealth comes mainly from the export of mineral re
sources (because comparative advantage might be ex
pected to reduce industrial investment).
The likelihood of democracy would decline as state own
ership of economic resources rose.
Democracy would be less likely in countries in which
foreigners or pariah capitalists excluded from the politi
cal community own most enterprises.
The point of this rather simpleminded exercise is that to test
the famous aphorism, one need not count the members of the
industrial and commercial bourgeoisie in each country and then
correlate the count with the Freedom House democracy scale.
Instead of, or in addition to, a direct test of an argument, one can
figure out some of its observable implications and test them.
Some of the implications of any argument will be consistent with
more than one theory, but if enough implications can be drawn,
3.
This academ ic sound bite is M oores summ ary o f M arx, not o f his own argu
ment. It is useful in the current context because it is so sim ple, not because it captures
M o o res argum ent.

40

Paradigm s and S an d C astles

not all will be consistent with both the proposed argument and
the same rival hypothesis. Although one cannot test all argu
ments and cannot always reject alternative interpretations for
given sets of findings, one can, through tests of multiple implica
tions, build support for a particular causal explanation one brick
at a time.
If instead of the aphorism which is itself an assertion about a
correlation I had used an argument that, like those advocated
by Bates et al., showed the moving parts in the causal mecha
nism. the number of implications would have been multiplied.
Implications can be drawn from every link in the logical chain,
not just from the hypothesized relationship between initial cause
and final effect. Big, romantic, untestable ideas can be made
amenable to rigorous investigation by first breaking them up into
their component processes and then theorizing these processes
one at a time. In the example below, I demonstrate drawing
implications from causal mechanisms.
Breaking up the traditional big questions of comparative poli
tics into the processes that contribute to them would make pos
sible the construction and testing of theories. I would not label
this shift in the focus of analysis as a move from grand to mid
range theory. A persuasive theory, backed by solid evidence,
about one of the several processes that combine to lead to a
transformational outcome strikes me as very grand indeed.
An Example of Breaking Up a Big Question
into Processes

Abstract methodological prescriptions are rarely compelling or

even fully intelligible. In an effort to move from the abstract to the
concrete and thus make a more persuasive argument for a change
in research strategy, the rest of this chapter focuses on transitions
from authoritarianism as an extended illustration of both the prob
lems associated with big questions and the usefulness of disaggre
gation into multiple processes as a research strategy. It will also
demonstrate the leverage that very simple models can bring to
bear on a question and show the usefulness of collecting a large
mass of information about a subject.
When we read research results in books and journals, we usu
ally sec only the finished product reporting the encounter be
tween argument and evidence. Often, however, the most diffi
cult part of research comes before any evidence is collected,

B ig Questions, Little Answers

41

during the stage when the analyst has to figure out how to think
about the problem in a fruitful way. This example goes through
those initial steps in considerable detail.
I chose transitions as an example because of its normative and
academic importance. During recent decades, the last authoritar
ian holdouts in capitalist Europe, nearly all countries in Latin
Am erica and Eastern Europe, and some countries in Asia and
Africa have democratized. At the beginning of 1974, the year
identified by Samuel Huntington (1991) as the start of the third
wave of democratization, dictatorships of one kind or another
governed 80 countries.4 Only 15 of these dictatorships still sur
vived at the end of 2000. During these years, 93 authoritarian
regimes collapsed (some countries endured more than one dicta
torship during the period). These transitions had resulted in 40
democracies that survived at the end of 2000, some quite flawed
but many stable and broadly competitive; 9 democracies that
lasted only a short time before being overthrown in their turn;
and 35 new authoritarian regimes, 15 of which lasted into the new
millennium.5 No one knows if these will be the last transitions for
these countries, but so far, contrary to initial expectations, new
democracies have proved fairly resilient. The study of these transi
tions has become a m ajor focus of scholarly attention.
Some of the finest minds in comparative politics have worked
on this subject. The body of literature on transitions now includes
hundreds, perhaps thousands, of case studies of particular transi
tions, dozens of comparisons among small numbers of cases, and
at least half a dozen important efforts at theoretically informed
generalizations. A number of descriptive generalizations have
become rather widely accepted. One example is the observation
that there is no transition whose beginning is not the conse
quence direct or indirect of important divisions within the au
thoritarian regime itself" (O Donnell and Schmitter 1986, 19); a
second is that pacts between competing elites facilitate the success
ful transition to democracy (Karl 1986, 1990; Higley and Gunther
1992).
4. Figures here and elsewhere in this chapter are drawn from a data set I have
collected that includes all authoritarian regim es (except m onarchies) lasting three
years or m ore, in existence at any time since 1946, in countries with a million or
m ore inhabitants. If m onarchies and countries with less than a million inhabitants
were included, the number o f authoritarian regim es would be larger. See G eddes
(1999a) for more details about the data set.
5. O utcom e numbers exclude regim es in countries created as a result o f border
changes during transitions, thus they do not sum to 93.

42

Paradigm s and Sand C astles

These and similar inductive generalizations emerging from

studies of particular groups of countries have added to our fac
tual knowledge, and they have forced the abandonment o f some
dearly held preconceptions. These are important advances. Nev
ertheless, despite the passage of more than twenty-five years
since the current wave of democratization began and the sacrifice
of whole forests to the production of literature on the subject,
few new theories of democratization have been created. When
fine scholarsseveral o f whom have in the past constructed theo
ries of great elegance and plausibility seem to have backed
away from theorizing about this topic, it behooves us to think
about why.
A part of the difficulty, I believe, stems from certain common
choices about research design. O f the fifty-six volumes on transi
tions reviewed in the American Political Science Review between
1985 and 1995, thirty-one were studies of single countries, and
many of the others were edited volumes made up of individual
case studies of several countries but lacking a theoretical synthesis
of the different experiences. In nearly all these books, the cases
were selected on the dependent variable; that is, authors sought
to explain one or more cases of political liberalization or democra
tization without comparing them to cases in which change had
failed to occur. Many of these studies supply readers with valu
able factual information, but the research design chosen prevents
their authors from testing their theoretical claims.
Furthermore, in the majority of the studies, the outcome of
interest (liberalization, transition, or consolidation) had not yet
finished happening when the study was written. The desire of
authors to write about the most important political events of the
time, and of publishers to publish things at the peak o f interest in
them, is understandable. This rush to publish, however, has dev
astating effects on the accumulation of theoretical knowledge.
There is no way to test causal arguments if the outcome being
explained has not yet happened at the time the study is done.
Becoming embroiled in controversies over the causes of some
thing that has not happened is like arguing about what the angels
dancing on the head of a pin look like without first having made
sure that at least one angel really performs there.
These would be short-term problems, with theories emerging
over time, if analysts continued working on the same problems
after the outcomes had become clear and if readers treated very
tentative conclusions with appropriate skepticism, but most do

B ig Questions, Little A n sw ers

43

not. Around the time it became clear that transitions to democ

racy really had occurred in a large number o f countries, many
scholars shifted their attention to trying to explain the consolida
tion o f democracy, which o f course had not yet happened. Inter
est in transitions declined at precisely the time when enough
experience had accumulated to make theory building possible.
The rush to publish is not unique to the study of regime
change, of course, and thus cannot carry all the blame for its
modest generation of theory. A further cause, I suggest, arises
from the choice of a compound outcome that is, an outcome
that results from the confluence of multiple causal processes as
the object o f study, while maintaining an approach more suited to
simple outcomes.
To show exactly what I mean, in the pages that follow I de
velop a concrete research strategy that begins with the disaggre
gation of the big question why democratization occurs into a
series o f more researchable questions about mechanisms. The
second step is a theorization of the specific process chosen for
study in this case, the internal authoritarian politics that some
times lead to transition. The third step is the articulation of
testable implications derived from the theorization. Decisions
about the domains of different testable implications constitute
the fourth step. The fifth is the actual discovery or collection of
evidence on which to test the implications; the sixth is the testing
itself; and the seventh is the interpretation of and response to
test results.
What I am aiming for here, and in other examples in this
book, is the self-conscious articulation of steps in the research
process that, like the values discussed above, occur in the prac
tice of good scholarship but are rarely described in detail. At
various points, I shall step back from the description of the steps
involved in setting up the research question to comment on why I
made certain decisions, to mention where ideas came from, or to
reiterate methodological points. The example in this chapter em
phasizes steps one through three as outlined in the previous para
graph, leaving detailed discussion of issues involved in testing to
later chapters.
Theory-Based Disaggregation

The first issue that confronts the researcher attempting to follow

the research strategy suggested here is figuring out how to

44

P arad igm s and S a n d C astle s

disaggregate the processes leading to the compound outcome.

There will always be multiple ways to do this, some more fruitful
than others. The only general advice that can be given is that the
disaggregation should be based on theoretical intuition and that
more than one should be tried. The paragraphs below sketch an
example of the process involved based on my theoretical intu
itions and fairly wide reading about transitions. Another ob
servers intuitions might be different and at least equally useful.
The Intuition
A regime transition is a change in the basic institutions that
determine who will rule, how rulers will be chosen, and how
basic distributive decisions will be made. When such a change in
institutions occurs as a result of revolution or violent seizure of
power, a standard way of simplifying reality for the purpose of
theory building is to focus attention on the winning and losing
groups in the power struggle, assuming implicitly that institutions
chosen will reflect the interests of the winners and that any bar
gaining that occurs over institutions is bargaining over details
among winners. Then, to explain such regime changes, we try to
understand why groups concluded that the old regime had be
come intolerable and how they developed the organizational
strength and popular support needed to overthrow it.
Our intuitions about regime change in general seem to derive
from observing such forcible seizures of power, but these arc not,
as it happens, very useful for understanding most transitions to
democracy. The breakdown o f an authoritarian regime need not
lead to democratization, but when it does, the transition involves
bargaining and negotiation. Unlike revolutionary victories and
authoritarian seizures of power, transitions to more participatory
forms of government cannot be accomplished entirely by force,
and the institutions that emerge during such transitions reflect
compromises among groups, not domination by a single group.
Even when the authoritarian regime is overthrown by the mili
tary. bargaining is necessary in order to complete the transition
to democracy. No single group wins and imposes its institutional
choices on all others. Furthermore, this bargaining occurs over a
period of time, during which the identity of particular negotia
tors can change. Institutional changes may be accomplished in
increments. It is only at the end of the process that the observer
can look at the set of institutional changes and make a judgment

Dig Questions, Little Answers

45

about whether democratization has occurred. In short, bargain

ing over institutions is a central feature of regime change.
Several different processes can affect this bargaining. Political
competition and rivalry within the authoritarian elite can cause
splits that may increase the willingness of factions to bargain, as
Guillermo O Donnell and Philippe Schmitter (1986) have noted.
Members of the upper class who had initially benefited from re
gime policies may become critical of later policies or performance
and may withdraw their support and their investments, thereby
destabilizing the economy and the regime. Economic crisis or
some other disaster may push ordinary citizens into clamorous
opposition, despite its risks. Such societal changes can strengthen
opposition bargainers and weaken elites. Changes in the interna
tional economy or the influence of powerful neighbors may alter
the eost-benefit calculations of both leaders and led about the
feasibility of regime change. Not all these processes will be salient
in every transition, but often several of them are. They may inter
act with each other, but they may also be independent.
The theoretical disaggregation that begins the research strat
egy should focus on such possibly independent processes identi
fied by the researcher. The disaggregation I suggest places the
bargaining over institutions at the center of analysis and seeks to
explain how these processes affected bargaining among different
actors at different times during the transition.
The Topics
With these ideas in mind, a possible set of topics would include
the following:
i. The politics within authoritarian governments, that is,
how political rivalries, policy disagreements, and bar
gaining within different kinds of authoritarian regimes
affect the incentives of authoritarian rulers to liberalize6
6.
E xccp t for discussions o f hard-liners and soft-liners who cannot he identified a
priori (e .g .. Przew'orski T992). this is a topic that received little attention in the early
analyses o f regim e change. Przeworski ( 19 9 1) has even asserted that characteristics of
the old regim e do not affect outcom es in the new one. Rem m er (1989) and Bratton
and van dc Walle (1994, 1997). however, have argued that different kinds o f authori
tarian regim es dissolve in characteristically different w ays, which has consequences
both for the likelihood o f transition and for the kind o f regim e likely to em erge as a
result. For a review o f som e o f these issues, see Sn yder and M ahoney (1999).

46

Paradigm s and Sand C astles

2. The determinants of upper-class support for authoritar

ian rule and the effects of loss of such support on bar
gaining between government and opposition, and
hence on regime maintenance7
3. The causes and risks of mass expressions of discontent
and the influence o f mass mobilization on bargaining
between government and opposition8
4. The effect of the relationship between opposition
elites and masses on bargaining between government
and opposition9
5. The relationship among (a) the timing of institutional
choices, (b ) the interests of the bargainers at particular
times, and (c) extent of democratization10
6. The relationship between economic modernization and
citizen influence on regime choice11
7. The effect of international economic and geopolitical
shocks on the decisions and actions of regime leaders,
regime supporters, and ordinary citizens
7. M any case studies note the fickleness and ingratitude o f bourgeois and other
upper-class supporters o f authoritarian regim es, along with the role these groups
have played in opposition to authoritarian governm ents. C ard oso s study (1986) of
the Brazilian bourgeoisie during dem ocratization is one o f the earliest and most
insightful.
8. M any case studies describe the effect o f dem onstrations and other mass ac
tions on the decisions of authoritarian rulers. In addition, several authors have em
phasized the im portance o f popular opposition in bringing about transitions (e.g.,
B ratton and van de Walle 1997; C asper and Taylor 1996; C ollier 1999; C ollier and
M ahoney 1997; B erm eo 1997). T hese studies are largely descriptive, however.
Though initial theoretical steps have been taken from several different directions to
account for why large numbers o f people, after having suffered oppression and
poverty for long periods o f tim e, suddenly rise up to voice their indignation
(Przeworski 1986; G ed d es and Z aller 1989; Lohm ann 1994), much m ore work re
mains to be done. Furtherm ore, to my know ledge, no one has offered a com pelling
explanation o f why authoritarian regim es som etim es respond with coercion to mass
protests and at other times hasten to com prom ise.
9. It should be possible to extend work on nested gam es (Tsebelis 1990) to deal
with this subject, though adaptation will be required to accom m odate the institu
tional fluidity characteristic o f transitions.
10. Much of the work on this subject has focused on pacts (e .g ., H igley and
G unther 1992; K a rl 1986, 1990). This topic has only begun to be m ore fully and
system atically explored (e .g ., Przew orski 19 9 1; G ed d es 1995, 1996; M ainwaring
1994).
1 1 . T he correlation between econom ic developm ent and dem ocracy is one o f the
best established in com parative politics (B ollen 1979; Bollen and Jackm an 1985;
Przeworski et al. 2000; B urkhart and Lew is-Beck 1994; B arro 1999). T he causes o f
this relationship, however, continue to be debated.

B ig Questions, Little Answers

47

Though different researchers would break up the big question in

different ways, any disaggregation into constituent processes
should have some of the characteristics of the topics on this list.
Each topic is posed as a general comparative question. We would
hesitate to propose an argument about any of these topics based
on the experience of only one country. Research on these topics
seems, on the contrary, to demand comparison across cases.
None of these topics imply selection bias, that is, none imply
limiting studies to those countries that have completed transi
tions. A ll governments face opposition, and the absence o f bar
gaining in particular times and places requires explanation; it is
not a reason to exclude cases from examination. I will return to
issues related to appropriate selection o f cases in later chapters.
For now, the important thing to note is that each topic listed here
is worthy of a project in itself. When processes are described
separately in this way, it becomes clear why it might be difficult
to theorize transitions as a whole.
Some of these topics, especially 2 and 3, have received consider
able attention in the case study literature. The next step in develop
ing research strategies to investigate them would be to build theo
ries that subsume and explain the observations made in the case
studies. In the extended example below, 1examine the first topic,
to which somewhat less attention has been paid. I propose an
argument about the incentives facing leaders in different kinds of
authoritarianism that helps to explain, first, why some authoritar
ian governments initiate liberalization when they face little soci
etal pressure to do so; and, second, why and when the factions that
always exist within dictatorships may contribute to democratiza
tion. This argument thus offers an explanation for two elements in
the process o f regime change that a number of studies note with
out explaining, but it does not try to account for the final outcome
of democratization itself.
A Theorization of One Process: Politics in
Authoritarian Regimes

O Donnell and Schmitters observation (1986) about the im

portance of splits within authoritarian governments, noted
above, alerts us to the importance of individuals near the center
of power during the transition process. Although political fac
tions and disagreements can be found within any authoritarian

48

P arad igm s and San d C astles

governm ent,12every transition is not actually a consequence o f im

portant divisions within the authoritarian regime itself (O Don
nell and Schmitter 1986, 19). The Salazar-Caetano dictatorship in
Portugal did not fall as a result o f internal splits (though such
splits existed, of course), unless that regime is defined as includ
ing the midlevcl military officers who had spent most of their
careers in Africa and who forcibly overthrew the dictatorship.
Arguably, a number of other dictators (such as Somoza in Nicara
gua and the Shah in Iran) fell not because of divisions within the
regime itself meaning splits among those with decision-making
power but because of desertions from the societal and mili
tary coalition originally supporting them. And, though O Don
nell and Schmitter had no way of knowing it at the time they
wrote, the collapses of communist regimes in countries such as
Bulgaria and East Germ any were not caused by splits within the
regime, unless those regimes are defined as including not only
high officials of the Bulgarian and East German communist par
ties but also their Soviet allies. Nor, according to Michael Bratton
and Nicolas van de Walle (1992, 1997), has the initiation of liberal
ization in many African countries been a consequence of splits
internal to regimes.
It is nevertheless true that in a large number of the recent
transitions from authoritarianism, the initial steps toward what
became democratization were taken by those in power, for rea
sons internal to the ruling elite rather than in response to pres
sure from cither supporters or opponents in the larger society.
The observation of this pattern in a number of cases surprised
observers who were accustomed to thinking of institutional
changes as consequences of power shifts, not as the causes of
them.
One of the reasons that regime transitions have proven so
theoretically intractable is that different kinds of authoritarian
ism break down in different ways. The beginnings of some can be
traced to splits within the regime, but others begin in other ways.
Dictatorships can differ from each other as much as they differ
from democracy, and these differences affect the way they col
lapse. They draw on different groups to staff government offices
and on different segments of society for support. They have
different procedures for making decisions, different characteris
12.
Num erous descriptions exist of factionalism within authoritarian regim es in
evcrv region of the world; sec. for exam ple, Stepan (11/71)*. Fontana (1987);
Sandbrook ( ig86); and W atcrbury (19 73).

B ig Questions, Little Answers

49

tic forms o f intraelite factionalism and competition, different

ways of choosing leaders and handling succession, and different
ways of responding to societal interests and opposition. Because
analysts have not studied these differences systematically, what
theorizing exists about authoritarian regimes is posed at a highly
abstract level, and few authors have considered how characteris
tics of dictatorships affect transitions. These differences, how
ever, cause authoritarian regimes to break down in systemati
cally different ways, as 1 show below.
To explain the first incremental institutional changes that set
some countries on the path toward democratization, we need a
theory of politics within authoritarian regimes. Where do we get
one? Standard theories o f politics in democratic regimes begin
with two simplifying assumptions: first, that officials want to
remain in office; second, that the best strategy for doing so is to
give constituents what they want. Much of the literature on demo
cratic politics concerns how different political institutions affect
the survival strategies o f politicians. The analysis of transitions
requires an analogous investigation o f the effects of differences
among various kinds o f authoritarian institutions.
To begin the task of investigating the effects of authoritarian
institutions, wc need first to assess the plausibility of the standard
assumptions, and then, possibly, to revise them. Most obviously,
in the absence of routine ways for citizens to remove authoritar
ian leaders from office, empirical investigation is needed to an
swer questions about who exactly the constituents of dictators
are, how satisfied they have to be, and what factors besides satis
faction with regime performance affect constituents' acquies
cence. These questions cannot be answ'cred in the abstract, nor
can answers be assumed, as in the study of democratic politics.
Topics 2, 3, 6, and 7 as outlined above deal with these issues.
Less obviously, it should not be assumed that the officers, par
ties, and cliques supporting authoritarian leaders always want to
remain in power. Military officers, in contrast to cadres in single
party and personalist regimes, may not want to. If there are cir
cumstances in which they can achieve their ends better while out
of power, as I will argue there are, then we can expect them to
return voluntarily to the barracks. Furthermore, the costs of
leaving office vary for different kinds of authoritarian leaders.
Military officers can return to the profession that called them in
the first placc, usually without suffering punishment for actions
while in office. Cadres in single parties lose their monopoly on the

50

Paradigm s and San d Castles

advantages of office, but they also usually remain free to compete

for office after a transition and thus to continue their chosen
profession. The allies of a personalist leader, however, generally
find it hard to continue the life to which they have become accus
tomed. Compared with other kinds o f authoritarians, they are
more likely to lose the opportunity for future office, and possibly
also their property and lives, in the wake of a transition.
To begin building an understanding o f authoritarian politics, I
focus on rivalries and relationships within the entity from which
authoritarian governments are drawn: the officer corps, the
single party, the clique surrounding the ruler, or some combina
tion of these. Most of the time, the greatest threat to the survival
of the leader in office though not necessarily to the survival of
the regime comes from within this ruling group, not from out
side opposition. In normal times, most of what we would call
politics, namely, the struggle over office, spoils, and policy deci
sions, takes place within this ruling group.
Politics within the ruling group tells only part of the story of
regime change, but it is a part about which we understand little.
Opposition from outside the ruling coalition and exogenous
shocks, such as the Soviet collapse, the international economic
crisis of the 1980s, and the economic reforms induced by that
crisis, have affected regime survival, sometimes decisively. By
focusing on the political dynamics within different kinds of au
thoritarian regimes, however, I aim to show why some forms of
authoritarianism are more vulnerable than others to exogenous
shocks and popular opposition.

The Classification of Authoritarian Regimes

Before we can use differences among authoritarian regimes as
the basis for elaborating theoretical arguments about the conse
quences of these differences, we need to develop a simplifying
classificatory scheme of regime types and clear criteria for assign
ing cases to categories. Without this kind of simplification of
reality, we would be inundated by complexity and unable to see
the patterns underlying it. The aim here is to carve nature at
its joints, that is, to find the places in the complicated whole
at which elements seem to divide naturally. A s with carving a
chicken, we must know a fair amount about the basic structure of
the beast in order to find the right places to hack. Because I
consider the most important differences among authoritarian re

B ig Questions, Little A nsw ers

51

gimes to be qualitative, I create a typology for measuring re

gimes rather than a scale or index. Typologies are theoretical
constructs used when variables can only be measured nominally.
Like other theoretical constructs, they are useful or not useful
rather than true or false. To be useful, they have to capture
differences that are essential to the argument being m ade.13
In this section, I discuss the bases for assigning regimes to one
category or another.14 I initially classified regimes as personalist,
military, or single-party. In military regimes, a group of officers
decides who will rule and exercises some influence on policy. In
single-party regimes, one party dominates access to political o f
fice and control over policy, though other parties may exist and
compete as minor players in elections. Personalist regimes differ
from both military and single-party in that access to office and
the fruits of office depend much more on the discretion o f an
individual leader. The leader may be an officer and may have
created a party to support himself, but neither the military nor
the party exercises independent decision-making power insu
lated from the whims of the ruler (cf. Bratton and van de Walle
1997, 61-9 6; Chehabi and Linz 1998, 4-45; and Snyder 1998). I
had to add intermediate categories to this classification scheme
after discovering how many of the cases simply resisted being
crammed into one or another o f the original categories.
My initial guess about what kind of classification would best
capture the important differences among authoritarian regimes
grew out of reading about many such regimes. Let me again
emphasize the importance of collecting information about a wide
range of cases. Although I had to hunt for information in a much
more systematic way further on in the research process, the ini
tial ideas that motivated this study came from reading military
sociology and descriptions of events in many countries during
transitions just because I was curious.
In this classification scheme, a military regime, in contrast to a
personalist regime led by a military officer, is one in which a group
of officers determines who will lead the country and has some
13. Typologies have been much and ju stly maligned in the com parative field
because their creation was at one lime seen by som e as an end in itself, and scholars
used lo waste their time com paring them and arguing about them . T h ey have a useful
role, however, as a way o f categorizing causes and effects that cannot be m easured
using numbers.
14. T he classification o f individual eases is a measurement'* issue. M easurem ent
will be discussed later in this chapter and in chapter 4. H ere I am concerned with
creating an overall 'coding schem e.

52

Parad igm s and S an d C astles

influence on policy. In an institutionalized military regime, senior

officers have agreed upon some formula for sharing or rotating
power, and consultation is somewhat routinized. Exam ples of
military regimes include that of Brazil (1964-85), in which senior
officers, in consultation with a small number of civilian allies,
picked each successive president in keeping with rules specified
by the institutions o f the authoritarian regime; and that of Argen
tina (1976-8 3). in which, despite intense factional struggle and
the efforts of some military presidents to renege on precoup agree
ments establishing an elaborate arrangement for consultation and
predictable rotation in office, senior officers did not permanently
lose control of succession and policy.
Many regimes headed by a military officer are not, however,
really controlled by a group of senior officers. It is common for mili
tary interventions to lead to short periods of collegial military rule
followed by the consolidation of power by a single officer and the
political marginalization of much of the rest of the officer corps.
These are personal dictatorships, even though the leader wears a
uniform. Regimes such as that of Rafael Trujillo in the Dominican
Republic (19 3 0 -6 1) and Idi Amin in Uganda (19 7 1-7 9 ) are some
what extreme instances of the transformation of a military inter
vention into personal tyranny. Other regimes, such as that of Augusto Pinochet in Chile (1973-89) and Sani Abacha in Nigeria
(1993-99), are harder to classify; the military institution retained
some autonomy and influence, but the concentration of power in
the hands of a single man prevents them from being categorized
simply as military.15 1 classify regimes on the margin between the
two categories as military-personalist hybrids.
Since most dictators form parties to support themselves, distin
guishing between real and nominal single-party regimes in
volves the same difficulties as distinguishing between military
regimes and personalist regimes led by military officers. In real
single-party regimes, a party organization exercises some power
over the leader at least part of the time, controls the selection of
officials, organizes the distribution of benefits to supporters, and
mobilizes citizens to vote and show support for party leaders in
other ways. Exam ples of single-party regimes include that of the
Institutional Revolutionary Party (PR I) in M exico, the Revolu
tionary Party of Tanzania (C CM ), and the Leninist parties in
15.
lhis classification o f Pinochet is supported by R em m er's analysis (1989) and
by Huntington (19 9 1) . The classification o f A bacha is supported by O basanjo (1998).

Big Questions, Little Answers

53

various East European countries. Regimes in which the leader

himself maintains a near monopoly over policy and personnel
decisions despite the existence o f a support party such as those
led by Manuel Odria in Peru and Etienne Eyadem a in Togo are
personalist.
Personalist dictators range from vicious psychopaths to benevo
lent populists. Institutionally, what they have in common is that
although they are often supported by parties and militaries, these
organizations have not become sufficiently developed or autono
mous to prevent the leader from taking personal control of policy
decisions and the selection of regime personnel. The fear of poten
tial rivals leads such rulers to undermine these and other institu
tions that might serve as power bases for potential challenges.
They rely instead on informal, and sometimes quite unstable,
personal networks sometimes based on kinship, ethnicity, or
region within which particularistic favors are exchanged for sup
port. Typically, regime personnel are rotated frequently to pre
vent them from developing autonomous bases of support, and
erstwhile supporters who become rivals or dissidents arc quickly
and unceremoniously deprived o f office, influence, and some
times their lives (cf. Bratton and van de Walle 1994).

Leaders Interests and Intraregime Politics

In order to build a theory about particular actors, one must first
have some knowledge about their goals. I have argued that the
goals of leaders in different kinds of authoritarian regimes typi
cally differ from each other. In this section, I discuss their differ
ent interests and the evidence supporting my assessment of these
interests.
The dictator who leads a personalist regime after having
clawed his way to the top in intense and often d' :idly struggles
among regime insiders can reasonably be assumed to have a
strong and abiding determination to remain in office. No similar
assumption can be made, however, about most of the officials of
military regimes. Some individual leaders, especially those who
have managed to scramble to the very top during the early chaos
of military takeovers, undoubtedly feel as intense a desire to
remain there as any other leader, but many officers do not. The
discussion below describes the interests of members of the pri
mary supporting institution or informal group in each type of
regime, starting with military.

54

Paradigm s and S an d C astles

Research on the attitudes and preferences of military officers

in many different societies finds that officers in different coun
tries come from different socioeconomic, ethnic, and educational
backgrounds. They have different ideologies and feel sympa
thetic toward different societal interests. No generalizations can
be made about the societal interests or policies they are likely to
support. According to the scholarly consensus, however, most
professional soldiers place a higher value on the survival and
efficacy of the military itself than on anything else (Janowitz
i960, 1977; Finer 1975; Bienen 1978; Decalo 1976; Kennedy
1974; Van Doom 1968, 1969).
This corporate interest implies a concern with the mainte
nance of hierarchy, discipline, and cohesiveness within the mili
tary; autonomy from civilian intervention in postings and promo
tions; and budgets sufficient to attract high-quality recruits and
buy state-of-the-art weapons. Officers also value their nations
territorial integrity and internal order, but the effective pursuit of
these goals requires unity, discipline, and adequate supplies
(Stepan 19 71; Nordlinger 1977; Barros 1978). Such preferences
might result from socialization in military schools (Stepan 1971;
Barros 1978) or from a rational calculation of the effect of the
health of the military institution on the officers own career pros
pects. For the purposes of this study, the source of these prefer
ences does not matter.
In countries in which joining the military has become a stan
dard path to personal enrichment (as, for example, during some
time periods in Bolivia, Nicaragua, Nigeria, Thailand, Indone
sia, and the Congo), acquisitive motives can be assumed to rank
high in most officers preferences. Such motives will occupy first
place for some officers and rank second or third for others, if
only because the continued existence of lucrative opportunities
for officers may depend on the survival of the military as an
effective organization. Where acquisitive motives have swamped
concern for corporate survival and effectiveness, however, the
professionalism of the military deteriorates, and the officer corps
is less likely to serve as a successful counterweight to ambitious
political leaders.
Where corporate interests prevail, most officers agree to join
coup conspiracies only when they believe that the civilian govern
ment prevents the achievement of their main goals. Many offi
cers, in fact, will join only if they believe that the military institu
tion itself is threatened. These preferences are consistent with

B ig Questions, Little A nsw ers

55

the observations by Alfred Stepan ( 1971) and Eric Nordlinger

(1977) about the importance o f threats to the military as an
institution in the decisions of officers to join coup conspiracies.
They are also consistent with the observation that coups do not
usually occur in fully professionalized armies until a consensus
exists among senior officers (Stepan 19 7 1; Valenzuela 1978),
since the worst possible outcome for the military as an institution
is a civil war in which part o f the military fights on each side.
Consequently, the most important concern for many officers
in deciding whether to join a coup conspiracy is their assessment
of how many other officers will join. What Nordlinger, Stepan,
and others are describing resembles a classic battle-of-the-sexes
game. The insight behind this game comes from the following
scenario: One member of a couple would prefer to go to a movie
and the other would prefer the symphony, but each would prefer
doing something together to doing something alone. Going to
either event together is a potential equilibrium, but no dominant
strategy exists, since the best outcome for either player always
depends on what the other chooses.
The logic of decisions about seizing power or returning to the
barracks is the same. Some officers always want to intervene,
others have legalist values that preclude intervention except in
the most extreme circumstances, and most are located som e
where in between but almost all care most about the survival
and efficacy of the military and thus want the military to move
either in or out of power as a cohesive whole. Figure 2.1 depicts
this set o f preferences as a game.
In the figure, the two numbers in each cell represent the re
spective payoffs to the two factions, the first number being the
payoff for the majority faction and the second number the payoff
for the minority faction.'6 In the game depicted, the majority
prefers that a united military remain in the barracks. The payoffs
to both factions for remaining in the barracks are shown in the
lower right cell. The upper left cell shows the payoffs for a suc
cessful intervention carried out by a united military. The minor
ity is better off than it was in the barracks, but the majority is
slightly worse off since it would have preferred not to intervene.
The minority faction prefers to intervene, but it would be far
worse off if it initiated an unsuccessful coup without support
16.
I have used numbers in this and other matrices because I think they are easier
to understand. T he specific numbers used here, however, have no m eaning. The logic
of the gam e would be the sam e for any numbers that m aintained the sam e order.

56

Paradigm s and S an d C astles

Minority Faction
intervene

barracks

intervene

4, 5

2, -10

barracks

3, -10

5, 4

M a jo rity
Faction

Fig. 2.1.

Game between military factions

from the majority than if they remained unhappily in the bar

racks (payoffs for this outcomc are shown in the lower left cell).17
Participants in an unsuccessful coup attempt face possible demo
tion, discharge, court-martial, or execution for treason, so their
payoff is shown as negative. The majority faction that opposed
the coup is also damaged by the attempt, since the armed forces
are weakened and the government is likely to respond with
greater oversight, reorganization, and interference with promo
tions and postings to try to ensure greater future loyalty, all of
which reduce military autonomy.
The final possible, though unlikely, outcome is a successful
coup carried out despite minority opposition (payoffs are shown
in the upper right cell).'8 In this event, the minority that remains
loyal to the ousted civilian government is likely to face the same
costs as unsuccessful conspirators: demotion, discharge, exile,
prison, death. The winners achieve power, but a weakened mili
tary institution reduces their chances of keeping it. Future con
spiracies supported by those demoted or discharged after the
coup become more likely. Once factions of the military take up
arms against each other, it takes years or decades to restore unity
and trust.
This is a coordination game: once the military is either in
power (upper left cell) or out of power (lower right cell), neither
17. The use o f m ajority and m inority here is not meant to imply that the success of
coup attempts is determ ined by which side has the most support. Support affects the
likelihood o f success but is by no m eans decisive. F or ease o f exposition, however, I
describe the m ajority faction as successful if it attempts a coup and the m inority
faction as unsuccessful. A more realistic game would introduce uncertainty about the
likelihood o f a succcssful coup, but the p ayoffs in the off-diagonal cells would remain
lower for both actors than those in the upper left and low er right cells regardless of
the outcome of the coup attempt.
18. Since the m ajority prefers not to intervene, it is hard to imagine anything
other than profound misinformation that would lead to this outcom e. Even if the
m ajority preferred intervention, however, their payoff for intervening without full
support would be lower than foi rem aining in the barracks.

B ig Questions, Little Answers

57

faction can improve its position unilaterally. Each faction must

have the others cooperation in order to secure its preferred
option. When the military is out of power, even if the majority
comes to believe it should intervene, it cannot shift equilibria
without cooperation from the minority.
There are two ways to solve coordination problems: one is to
negotiate until consensus is reached, and the other is to make a
credible first move that confronts the second mover with the
choice between joining the first mover or receiving the payoff
associated with a divided choice. Some military decisions to seize
power have been carefully negotiated over a period of months
until rules for sharing power have been hammered out and the
last legalist holdout has either given in or retired. Such negoti
ated interventions occurred in Argentina in 1976, Brazil in 1964,
and Chile in 1973 (Fontana 1987; Stepan 19 71; Valenzuela 1978).
Since extended negotiations carry considerable potential for dis
covery, however, most military interventions have employed a
first-mover strategy in which a small group of conspirators seizes
the presidential palace, the airport, television and radio stations,
military installations in the capital, and perhaps a few other key
buildings. It then announces that it has taken power and counts
on the rest of the armed forces to go along (Nordlinger 1977).
They usually do but not always.
For the first-mover strategy to work, the first move has to be
credible, meaning that other officers have to be convinced that
the seizure of power is irreversible. The attempted coup of 1981
in Spain is an example of a failed first-mover strategy. The coup
plotters seized the requisite number of installations in Madrid
and had reason to believe that garrison commanders in the rest
of the country would go along with them. King Juan Carlos,
however, immediately began telephoning the garrison command
ers, telling them that he opposed the intervention and that if they
joined it, they would be guilty of treason. He also went on televi
sion to rally citizens against the coup. Once the king had taken
such a strong stand, the first move lost its credibility, and most of
the military refused to go along. Josep Colomer reports that one
of the coup conspirators, when interviewed later, said, Next
time, cut the king's phone line1' (1995, 12 1). Colomer suggests
that had the king not been able to use television and the phone to
rally support, the first-mover strategy might well have worked,
because many in the officer corps sympathized with the goals of
the conspirators (110 - 2 3 ).

58

Paradigm s and San d C astles

When the military controls the government, the logic remains

the same. Most officers will go along with a credible move by one
faction to return to the barracks. Strong disagreements among
leading officers over how to respond to economic difficulties or
who among them will next occupy the presidency lead to intense
factionalization. When this happens, one group is likely to prefer
returning to the barracks as a way of avoiding institution-damag
ing conflict. Observers see splits in the officer corps at the time of
the first moves toward democracy because the concern over divi
sions within the military causes some factions to prefer a return
to the barracks. Both hard-line and soft-line factions can use
the first-mover strategy, however. Military presidents can make
quite credible first moves heading back toward the barracks, and
most officers will go along. Hard-liners can also chance firstmover strategies, ousting more moderate military presidents.
Again, if the move is credible, most of the officer corps will go
along.
For the officer who ends up as paramount leader of the post
coup junta, the game may change after a successful seizure of
power, as it did for Pinochet and those like him in other countries
who sought to concentrate power in their own hands; but other
officers usually see their situation as resembling a battle-of-thesexes game, even in the most politicized and factionalized militar
ies. Repeated coups by different factions, as occurred in Syria
prior to 1970 and Benin (then called Dahomey) before 1972,
would not be possible if most of the army did not go along with
the first mover, either in seizing power or in handing it back to
civilians.
This analysis demonstrates the usefulness of having some
simple models in ones theoretical toolkit. The militarys concern
about professional unity has been described verbally by a num
ber of scholars, but using the game to show the logic of the
situation demonstrates the consequences of this concern in a very
clear and stark manner. The comparison between this game and
the ones to be developed below will show the effects of differ
ences in the interests of cadres in different kinds of regimes more
clearly than could a verbal comparison alone. These models are
not, of course, the endpoint of the analysis. A fter the models
have been described, they will need empirical confirmation.
The preferences of party cadres are much simpler than those
of officers. Like democratic politicians, party cadres simply want
to hold office. Some value office because they want to control

B ig Questions, Little A nsw ers

59

Majority
(Leader's Faction)
in office

out of office

in office

8, 10

5, 1

out of office

3, 9

0, 0

Rival
Faction

Fig. 2.2.

Game between factions in single-party regimes

policy, some for the pure enjoyment of influence and power, and
some for the illicit material gains that can come with office. The
game between single-party factions is shown in figure 2.2. The
insight behind this game is that everyones cooperation is needed
in order to achieve a desired end, and no one can achieve it
alone. In this game, no one ever has an incentive to do anything
but cooperate to remain in office.
In this game the best outcome for everyone is for both the
majority faction and the rival faction to hold office (payoffs are
shown in the upper left cell). The worst outcome occurs when
both are out o f power (shown in the lower right cell). The upper
right cell shows the payoffs when the party has lost control of
government but the minority faction still fills some seats in the
legislature or holds other offices as an opposition to the new
government. The minoritys payoff when in opposition is lower
than when its party holds power because the opposition has
fewer opportunities to exercise influence or line pockets. In the
lower left cell, the minority faction is excluded from office, but
the dominant faction of the party still rules. In this case, the
minority continues to receive some benefits, since its policy pref
erences are pursued and party connections are likely to bring
various opportunities, but members of the excluded minority
receive none of the specific perquisites of office. The majority is
also worse off, because exclusion gives the minority an incentive
to try to unseat the majority. Combatting the minority is both
risky and costly for the majority.
Factions form in single-party regimes around policy differences
and competition for leadership positions, as they do in other kinds
of regimes, but everyone is better off if all factions remain united
and in office. This is why co-optation rather than exclusion charac
terizes established single-party regimes. Neither faction would be
better off ruling alone, and neither would voluntarily withdraw

6o

Parad igm s and S an d C astles

from office unless exogenous events changed the costs and bene
fits of cooperating with each other (and hencc changed the game
itself).19
In contrast to what happens in military and single-party re
gimes, the political fate of the close allies of a personalist dicta
tor is tied to the fate of the dictator himself. [IJnsiders in a
patrimonial ruling coalition are unlikely to promote reform. . . .
Recruited and sustained with material inducements, lacking an
independent political base, and thoroughly compromised by cor
ruption, they are dependent on the survival of the incumbent
(Bratton and van de Walle 1997, 86). Personalist dictatorships
rarely survive for long after the death or ouster of the dictator,
perhaps bccause dictators, in their efforts to defend themselves
from potential rivals, so assiduously eliminate followers who
demonstrate high levels of ability and ambition.
In personalist regimes, one leader dominates the military, the
state apparatus, and the ruling party, if there is one. Because so
much power is concentrated in the hands of this one individual, he
generally controls the coalition-building agenda. Consequently,
the game between factions in a personalist regime must be de
picted as a game tree instead of a two-by-two matrix in order to
capture the leader's control over first m oves.20 As shown in fig
ure 2.3, the leaders faction has the initiative, choosing to share
the spoils and perks with the rival faction or not. The choice I have
labeled hoard can be interpreted either as limiting the opportu
nities and rents available to the rival faction or as excluding some
of its members altogether. In the example shown in this figure, the
amount of hoarding is small (the payoff to members of the rival
faction for continued cooperation despite hoarding by the rulers
faction is 6); perhaps members of the rival faction are not offered
the choicest opportunities, or perhaps a few' of its members are
jailed but the rest continue to prosper. If the whole rival faction
were excluded from all benefits, their payoff for continued co
19. The econom ic shocks o f the 1980s and 1990s changed these costs and benefits
in many countries, reducing the incentive o f potential rivals to cooperate with rulingparty leaders and thus destabilizing regim es. T h e gam e used here shows the incen
tives to cooperate during good times. A different gam e would be needed to capture
(he choiccs facing single-party cadres after serious exogenous shocks.
20. Two-by-two m atrices, often used to dcpict simple prisoners dilem m as, battleof-the-sexes gam es, chicken gam es, and so on. assume sim ultaneous decisions by the
players or a lack o f information about how the other has chosen. M ore com plicated
gam es, including those in which one player chooses first and the second chooses
know ing how the first has chosen, have to be depicted using a gam e tree or equations.

F L A C S G -B ih lio te c a

B ig Questions, Little Answers

61

Majority faction

share

hoard

Minority faction

cooperate /

Nature

10
8
success
(probability = p),

-2 p
12 p
Fig. 2.3.

1 0 ( 1 - P)
-2(1 - p)

-2 p
12p

12(1 -p )
-2(1 -p )

Game between factions within personalist clique

operation would be much lower, but rarely lower than the payoff
for refusing to cooperate.
After the leader's faction has chosen its strategy, the rival
faction must decide whether to continue supporting the regime
or not. During normal times, it has strong reasons to continue.
Because its members face the prospect of losing all visible
means of support in a political transition, they have little option
but to cling to the regime, to sink or swim with it (Bratton and
van de Walle 1997, 86).
Unlike in single-party regimes, the leaders faction in a per
sonalist regime may actually increase benefits to itself by exclud
ing the rival faction from participation. Where the main benefits
of participation in the government come from access to rents
and illicit profit opportunities, the payoff to individual mem
bers of the ruling group may be higher if these benefits need not
be shared too widely. It may also be easier to keep damage to
the economy below the meltdown threshold, and thus increase

62

Paradigm s and Sand C astles

the likelihood of regime survival, if the predatory group is rela

tively small. Hoarding by the leaders faction is thus likely. If the
hoarding is not too extreme, as in figure 2.3, the rival faction is
better off continuing to cooperate, and most of the time that is
what they do.
If the rival faction withdraws its support and begins to plot the
leaders overthrow, its members risk life, liberty, and property.
The rewards of a successful overthrow are high, but so are the
costs of detection, betrayal, or defeat. In the game, the uncer
tainty over the outcome of plots is shown as a play by Nature.
The plot succeeds with probability p , usually a low number, and
fails with probability 1 p. The rival faction decides whether to
continue its support for the leaders faction by comparing its
payoff for support with its expected payoff from a plot. Two
considerations thus affect the choice: the benefits being derived
from the status quo and the potential plotters assessment of the
risk o f plotting. A s long as the personalist ruler seems powerful
enough to detect plots and defeat coup attempts, the rival faction
will continue to cooperate if it gets some benefits from the re
gime. The leaders faction has an incentive to reduce the benefits
to the rival faction to a level just above that needed to prevent
plotting. This system is very stable as long as the ruler can distrib
ute the minimum level of benefits needed to deter plotting and
can maintain control over an effective security apparatus and
loyal military. The situations in which these conditions become
less likely are discussed below.
B y drawing on some rudimentary game theory, I have begun
to develop insights into how the interests of cadres in different
kinds of authoritarian regimes might play out in different con
texts and how resilient to stress the cadres loyalties to regime
leaders might be. These insights motivate the analysis in the next
section.

The Consequences of Differences in Interests

The interests described above provide a starting point for figur
ing out whether the splits and rivalries that exist within all kinds
of governments will lead to regime breakdown. Because most
military officers view their interests as following a logic similar to
that o f a battle-of-the-sexes game, they acquiesce in continued
intervention regardless of whether military rule becomes institu
tionalized, the leader concentrates power in his own hands, or a

B ig Questions, Little Answ ers

63

rival ousts the original leader. The officer corps will not, how
ever, go along with disintegration of the military into openly
competing factions. If elite splits threaten military unity and effi
cacy, some factions will opt for a return to the barracks. If the
soft-line faction can make a credible first move in that direction,
most other officers will go along.
Military regimes thus contain the seeds of their own destruc
tion. When elite rivalries or policy differences intensify and these
factional splits become threatening, a return to the barracks be
comes attractive to most officers. For officers, there is life after
democracy, because all but the highest regime officials can usu
ally return to the barracks with their status and careers untar
nished and their salaries and budgets often increased by nervous
transitional governments (Nordlinger 1977; Huntington 19 91).
Leaders of single-party regimes also face competition from
rivals, but most of the time, as in personalist regimes, the bene
fits of cooperation are sufficient to ensure continued support
from all factions. Leadership struggles and policy disagreements
occur, but they do not affect the desire of most cadres to remain
in office. For them, life after democracy would require some
unpleasant changes in lifestyle. They would have to compete for
the benefits they have become accustomed to monopolizing. Dur
ing leadership struggles, most ordinary cadres just keep their
heads down and wait to see who wins. Thus, leadership struggles
within single-party regimes usually do not result in transitions.
The close allies of personalist dictators have even less reason
to desert the ship in normal times. If the ship goes down, they are
likely to go with it. A s long as the dictatorship is able to supply
some benefits and has a sufficiently competent repressive appara
tus to keep the probability of successful plotting reasonably low,
they will remain loyal.
These differences explain why the early transitions literature,
drawing insights primarily from the transitions from military rule
in Latin Am erica, emphasized splits within the regime as causes
for the initiation of democratization though later studies did not.
In other parts of the world, where rule by the military as an
institution is less common, factions and splits could be identified
within authoritarian regimes, but they did not result in transi
tions. Instead, observers emphasized the importance of other
factors in bringing down long-standing dictatorships: economic
crisis (Haggard and Kaufman 1995), foreign pressure (Hunting
ton 19 91), and popular protest (Bratton and van de Walle 1992,

64

Paradigm s and San d C astles

1997; Casper and Taylor 1996). In short, the theorization of

intra-authoritarian politics makes it possible to subsume the find
ings of a number of studies with differing regional foci. In regions
where the military led most of the authoritarian regimes that
broke down, the first steps toward democratization could be
traced to splits within the military leadership, but where single
parties or personalist autocrats tended to rule, pressures of vari
ous kinds external to the ruling party or clique played larger
roles.
The many studies of transitions, most of which draw essential
insights primarily from one part of the world, bear some similar
ity to the parable about five blind men encountering an elephant.
Each offers a useful and insightful description of the part of the
elephant he touches, but cannot describe the whole. The early
studies could not do so because they were trying to explain a
process that had barely gotten under way, though of course they
had no way of knowing how many countries democratization
would eventually affect. Later studies either have made no at
tempt to survey all cases or, in their attempt to set their own
region in the broader context, have misinterpreted studies of
some of the most frequently examined cases in particular regions
as being representative of the general experience of that region.
To repeat two of the most basic pieces of advice in this book: lots
of factual information is always good; and it is hard to explain an
outcome that has not yet finished coming out.
Drawing Testable Implications from the Argument

In this section, I detail the derivation of testable implications from

the analytic argument above. A s is often the case in comparative
politics, it is not feasible to test in a rigorous way the argument
about cadre interests proposed here. To gather the necessary de*
tailed information about the internal politics of a large number of
authoritarian regimes would require learning many languages and
traveling to many places. Although numerous books and articles
have been written about authoritarian governments in the larger,
more developed, and for other reasons more interesting coun
tries, it is difficult to find even detailed descriptions of events in
smaller, less developed countries such as Burkina Faso, Niger,
and Laos, especially those in which democratization has not taken
place. In situations like this, one must rely on tests of the itriplica
tions of the argument, which can sometimes be done with less

Big Questions, Little Answers

65

detailed information than would be needed to test the argument

itself. Testing implications in this kind of situation can make it
possible to avoid the selection bias that would almost inevitably
arise in an attempt to test the argument itself.
The argument sketched above claims that because officers
see their interests in terms similar to a battle-of-the-sexes game,
military regimes break down more readily than do other types
of authoritarianism in response to internal splits, regardless of
the cause of the splits. If that is true, we should expect mili
tary regimes to last less long, on average, than other forms of
authoriiarianism.
We should also expect economic crisis, which weakens support
for all governments, to have a stronger disintegrating effect on
military governments because of their underlying fragility. This
suggestion might at first seem surprising, since most military
governments hold no elections and tend to be more insulated
from societal interests than other types of dictatorship. Thus, we
might suppose them less vulnerable to pressures emanating from
citizens unhappy with the regimes economic performance.
The cadre-intcrests argument, however, implies that officers
may decide to step down even without the inducement of overt
public pressure. Officers and cadres are aware of their govern
ments economic performance, and they are linked to society via
their families and friends. Typically, when officers perceive their
government's performance as unsuccessful, some of them advo
cate intensifying the economic strategy being pursued while oth
ers advocate changing it. The backers of each policy prescription
support the presidential aspirations of a different officer, and com
petition between them intensifies, sometimes leading to coups
and countercoups. A split over economic strategy has the same
effect as any other kind of split: if it threatens to get out of hand,
most officers prefer to return to the barracks.
Observers such as Bratton and van de Walle (1997) note the
importance of material inducements to loyalty in personalist re
gimes. We might suspect that where loyalty depends on the
leaders ability to deliver individual benefits, economic crisis
would cause regime breakdown, but that would be an insuffi
ciently cynical view. Run-of-the-mill poor economic performance
hurts ordinary citizens but does not preclude rewarding support
ers. It takes a true economic disaster to do that. We should thus
expect personalist regimes to be destabilized by economic cata
strophe but, in comparison to military regimes, less affected by

66

Paradigm s and San d CastJes

ordinary poor economic performance. Recent African experience

suggests that reforms reducing state intervention in the economy,
and hence the rents and corruption opportunities often used to
reward supporters, may be as destabilizing to personalist regimes
as economic crisis itself.
Because officers tend to decide to return to the barracks for
reasons relating to internal military concerns rather than being
forced out of office by popular protest or external events, we
should expect them to negotiate their extrication. When officers
decide to withdraw from power, they enter into negotiations with
civilian political leaders to arrange an orderly transition and to
safeguard, if possible, their own interests after the transition. We
should thus expect that military regimes will be more likely than
other kinds of authoritarianism to end in negotiation.
Because of the internal sources of fragility in military regimes,
we should expect them to be overthrown by armed insurgents or
ousted by popular uprisings only rarely. Demonstrations against
them occur, but most of the time such demonstrations persuade
factions o f the military to initiate a transition before popular
opposition develops into rebellion. Coups are common in mili
tary regimes, but they rarely end the regime. They are usually
leadership changes, the analogue of votes of no confidence in
parliamentary systems. Coups that bring a liberalizing military
president to power often precede transitions in military regimes;
such coups can be interpreted as first-mover strategies. They
demonstrate that a shift in officer opinion has occurred and that
a substantial faction prefers to return to the barracks.
In strong contrast to military officers, the leaders of personal
ist regimes generally fight tooth and claw to hang on to power. In
Bratton and van de Walles words, They resist political open
ings for as long as possible and seek to manage the process of
transition only after it has been forced on them (1997, 83). If
they are forced by foreign pressure, for example to negotiate
with opponents, they renege on agreements at the first opportu
nity.21 Military governments rarely renege on the agreements
they make, not because they cannot, but because agreements are
made at a time when most officers want to return to the barracks.
The cadre-interests argument claims that in normal times, the
2 1.
N ote, for exam ple, the way M obutu o f Z aire (now C on go), Eyad em a of
Togo, and various other long-ruling A frican leaders m anipulated electoral rules and
intimidated opponents after agreeing, under pressure from intem aiional aid donors,
to initiate m ultiparty elections.

B ig Questions, Little A nsw ers

67

members of a ruling personalist clique have little reason to desert

their leader or oppose the regime. We should expect to see elite
desertions of the regime only if rents and opportunities can no
longer be distributed to supporters or if the leader loses control
over the security apparatus and armed forces, thus reducing the
risk of plotting his overthrow. Loss of control of the security
apparatus can happen for various reasons, but one obvious and
usually insurmountable reason is the death or physical incapacity
o f the leader. Dead or incapacitated leaders are replaced in all
political systems, but the demise o f the leader does not usually
end other forms of authoritarianism. Because control of the
armed and security forces is usually concentrated in the dictators
hands in personalistic regimes, however, his death or serious
illness often reduces the risks of opposition. A testable implica
tion of this argument is that the death o f the leader is more likely
to lead to regime breakdown in personalist than in other types of
authoritarian regimes.
According to the cadre-interests argument, most of the mili
tary prefers to return to the barracks in some circumstances.
Even for those officers who would prefer to remain in govern
ment, the cost for most of resuming a more ordinary military
career is low. The cost of losing office is higher for cadres in a
dominant party, but not, on average, devastating. M any promi
nent politicians in post-transition democratic regimes were once
cadres of the formerly dominant party. Although the cadres of a
single-party regime cannot be expected to desert when times are
good, if it looks as though the partys hegemony will soon end,
those cadres who think they possess the skills to make a success
of democratic politics and whose ambitions are frustrated within
the ruling party can be expected to form or join opposition par
ties. Even those who remain in the ruling party to the bitter end
need not despair of life after democratization. Many previously
dominant parties continue to function as effective political actors
after democratization (cf. van de Walle and Butler 1999). In fact,
in a number of ex-communist and African countries, such parties
have achieved executive office in the second free and fair elec
tion after democratization.
The members of personalist cliques, however, have fewer op
tions. Joining the opposition prior to a transition can have very
high costs, and many who desert the regime must go into exile in
order to protect their lives and liberty. From exile, they may plot
and organize, but few who remain at home are willing to risk

68

Parad igm s and S an d C astles

public opposition. Those who stick with the regime to the bitter
end are much less likely to find a rcspected place in the post
transition political world than are the close supporters of single
party and military regimes. For these reasons, the end of a personalist regime is more likely to be violent in one way or another
than is the end of a single-party or military regime. Thus, an
other testable implication of the cadre-interests argument is that
personalist regimes should be more likely than other forms of
authoritarianism to end in the assassination of the leader, popu
lar uprising, armed insurgency, civil war, revolution, or armed
invasion (cf. Skocpol and Goodwin 1994).
Violence and upheaval do not segue naturally into democratic
elections; consequently, transitions from personalist rule should
be more likely to end in renewed authoritarianism than are transi
tions from other forms of authoritarianism. Transitions accom
plished by uprisings, invasions, or assassinations often allow the
consolidation of power by those who overthrow the old regime.
In contrast, negotiations during transitions usually set a time for
elections and hammer out rules for how they will be conducted.
Thus, competitive regimes are more likely to succeed military
regimes than other forms of authoritarianism.
Like members of personalist cliques, cadres in single-party
regimes have few reasons to desert in normal circumstances.
Furthermore, because power is less concentrated in single-party
regimes, they are less vulnerable to the death or illness of lead
ers. Thus, we should expect single-party regimes to last longer
than either military or personalist regimes.
Because the dominant strategy of the ruling coalition in singleparty regimes is to co-opt potential opposition, such regimes
tend to respond to crisis by granting modest increases in political
participation, increasing opposition representation in the legisla
ture, and granting some opposition demands for institutional
changes. They attempt to give the opposition enough to deter
them from risky plots and uprisings while continuing to hang on
to power.
In the most common kind of regime crisis one caused by
poor economic performance leading to antiregime demonstra
tionsthe ruling elite in any kind of authoritarian regime tends
to divide into intransigents and moderates as they struggle to
respond. In military regimes, that division itself tends to per
suade many officers that the time has come for a return to the
barracks. In personalist regimes, the ruling coalition narrows as

Big Questions, Little Answers

69

the intransigents circle the wagons and exclude moderates from

access to increasingly scarce spoils. Former regime moderates
may then join the opposition bccause they have been excluded
from the distribution of spoils (cf. Bratton and van de Walle
1997). Ruling parties, however, attempt to distract citizens from
their economic grievances by granting them modest political
rights. This strategy works only sometimes, but it works often
enough to extend the average life span of single-party regimes.

Hypotheses Derived from Implications of

the Argument
To summarize, we can list a number of expectations about what
we would see in the real world if the basic logic of how elite
politics works in different kinds of authoritarian regimes were
correct. Compared to other kinds of authoritarianism,
military regimes survive less long;
military regimes are more quickly destabilized by poor
economic performance;
military regimes are more likely to end in negotiation;
military regimes are more likely to be followed by com
petitive forms of government;
personalist regimes are more likely to end when the dic
tator dies;
personalist regimes are more likely to end in popular
uprising, rebellion, armed insurgency, invasion, or other
kinds of violence;
personalist regimes are more likely to be followed by
new forms of authoritarianism;
single-party regimes last longest, on average.
Testing the Implications: "Measurem ent"

Since this is a book on research methods, I include here many

details about case selection and classification that might ordinarily
appear in the appendix of an article or book. To test the implica
tions discussed above, I have collected basic information about all
authoritarian regimes (except monarchies) lasting three or more
years that existed or began between 1946 and 1996, in countries
with a population of more than a million that became independent
before 1990. Authoritarian regimes already in existence in 1946,

70

Paradigm s and S an d C astles

such as those in the Soviet Union, M exico, and Turkey, are

included, with their length of time in office calculated from the
time they actually took power. Countries that became indepen
dent after 1945 enter the data set at the time of independence
(if authoritarian). Countries that have achieved independence
since 1990 because of the breakup o f the Soviet Union and
other communist states (and that remain authoritarian) have
not been included, because the inclusion of a fairly large num
ber of countries with severely truncated regimes might have
biased conclusions.
The purpose of the three-year threshold is to distinguish re
gimes from temporary authoritarian interventions and periods of
chaos. Regimes are defined as sets of formal and informal rules
and procedures for selecting national leaders and policies. Under
this definition, periods of instability and temporary moderating
military interventions (Stepan 19 7)) are considered interreg
nums, not regimes. That is, they are periods of holding customary
rules in temporary abeyance, struggle over rules, or transition
from one set of rules to the next. The three-year threshold is
simply a way of excluding such periods from the data set. This
cutoff was chosen after considerable empirical investigation of
very short-lived authoritarian interludes because it introduced
the least misclassification into the data. A lower threshold would
lead to the inclusion of a few moderating interventions and
interventions that never managed to establish a new set of rules.
The military governed during most of these interregnums. If they
were included in the data set, the findings I report below would be
stronger. A higher threshold would lead to the exclusion of some
authoritarian governments that have been included in other litera
ture on transitions.
I use a dichotomous measure of regime type (authoritarian
versus not authoritarian) to identify cases for inclusion in the
study, because the hypotheses I want to test require being able to
identify the endpoints of regimes. Zachary Elkins (2000) argues
that continuous measures of regime type are better, and for some
purposes they are. In this study, however, they would add noth
ing. The argument makes no predictions about whether regimes
are moving incrementally toward somewhat more press freedom
or allowing minority parties a few seats in the legislature. It does
make predictions about the conditions that cause regimes to end.
To test those, I need to be able to identify unambiguously when
an authoritarian regime has ended. I could have dichotomized an

B ig Questions, Little A n sw ers

71

available continuous indicator such as the Freedom House index

rather than measuring the cases myself, but Freedom House
indices are made up of measures of citizens political and civil
rights. Since I define regimes as the sets of rules for choosing
leaders and policies, Freedom House and other commonly used
measures of democracy do not seek to measure the concept I
have in mind.
Measurement decisions should derive from definitions of con
cepts in the theory being investigated and from the needs of
particular tests of hypotheses. The usefulness of different indi
cators depends on their purpose and cannot be judged in the
abstract.
In this data set, most decisions about whether governments
were sufficiently authoritarian to deserve inclusion were easy,
but a few were not. A significant complication was that norms for
defining countries as democratic vary by region. Few Latin
Americanists would classify Mexico as democratic before 1997,
and some would not do so until the P R I finally lost the presi
dency in 2000. Am ong Africanists, however, Botswana, in which
the ruling party has never lost control of the executive and at
least two-thirds of the seats in parliament, is always called demo
cratic. Needing a single standard to apply across regions, I clas
sify regimes as authoritarian if opposition parties have been
banned or subjected to serious harassment or institutional disad
vantage, or if the ruling party has never lost control of the execu
tive and has controlled at least two-thirds of legislative seats in all
elections before 1985. Once a regime is labeled authoritarian, I
do not consider it fully democratized until one turnover of execu
tive power has occurred. Where it appears that conclusions
might be affected by the stringency of these criteria, I also show
results using less demanding rules.22
A basic point to be made about using concepts with contested
definitions (such as authoritarianism) in research is that the con
crete criteria used for classifying cases or observations need to be
clear. The researcher must take care to apply the same criteria to
all cases. Where these coding decisions are complicated or
22.
T hese regim e type classifications are sim ilar to those o f Huntington (19 9 1) ,
and my "cod ing judgm ents are very close to his. M y decision rule for determ ining
whether a political system had crossed the threshold to dem ocracy is essentially the
sam e as that o f Przeworski and Lim ongi (19 9 7). T he biggest difference between my
classification scheme and that o f Linz and Stepan (1996) is that I collapse what they
call * sultanistic and 'civilianized regimes into one categ o ry personalist.

72

Paradigm s and San d C astles

require careful judgm ent, I suggest using a written coding

scheme and reporting the classification of all cases along with the
research findings.2-1 If one suspects that the concrete criteria used
to assign cases to categories will be controversial or that deci
sions about classification drive results, one should also show re
sults using alternative classificatory criteria.
The rationale for the stringent classification rule used here is
that a party (or clique) that has concentrated great power in its
hands over the years can, like the current Malaysian govern
ment, very quickly and easily reinstate strict limits on opposition
when threatened. Such a regime contains few institutionalized
limitations on the power of rulers, even if the rulers have not
previously felt the need for repressive measures and hence have
not relied on them. The consequence of this rule is that a few
cases that are sometimes considered democratic notably, B o
tswana, Tanzania, M alaysia, and Taiwan (before the election in
2000) are classified as single-party regimes here. Classifying
these countries as democracies would reduce the average life
span of single-party regimes by about a year.
To classify authoritarian regimes as military* single-party, per
sonalist, or hybrids of these categories, I relied on the following
criteria. Military regimes were defined as those governed by an
officer or retired officer, with the support of the military estab
lishment and some routine mechanism by which high-level offi
cers could influence policy choice and appointments. Single
party regimes were defined as those in which the party had some
influence over policy, controlled most access to political power
and government jobs, and had functioning local-level organiza
tions. Regim es were considered personalist if the leader, who
usually came to power as an officer in a military coup or as the
leader of a single-party government, had consolidated control
over policy and recruitment in his own hands, in the process
marginalizing other officers influence and/or reducing the influ
ence and functions of the party. In the real world, many regimes
have characteristics of more than one regime type. When re
gimes had important characteristics of more than one pure re
gime type, especially when the area specialist literature con
tained disagreements about the importance of military and party
institutions, I put them in hybrid categories.
In all cases, I attempted to rely on a regime's actual rules for
23. D evising a coding scheme is discussed in more detail in chapter 4.

B ig Questions, Little Answers

73

selecting leaders and making allocative decisions rather than for

mal designations of regime type. In practice, many regimes have
characteristics o f more than one of these classifications, and
many move from one category to another over time even though
the same person holds the highest office.
Dictators sometimes succeed in transforming the regimes
they lead from one kind to another. A s noted above, the transi
tion from military to personalist occurs frequently. I did not
count these transformations as regime changes, since that would
artificially reduce the length of what we in everyday language
call regimes, and one of the implications I want to test involves
length of survival. If an early period of uncertainty or transition
was followed by consolidation of a different regime type, I as
signed the regime to the category in which it seemed to stabi
lize. Some cases, however, had to be assigned to intermediate
categories.
In deciding whether a regime led by the single leader of a
single party should be classified as personalist or single-party, I
gave more weight to the party if it existed prior to the leaders
accession to power, especially if it had organized the fight for
independence, a revolution, or some equivalent mass movement,
rather than being formed by the leader after his accession; the
heir apparent or the successor to the first leader already held a
high position in the party and was not a relative or a member of
the same tribe or clan as the leader; the party had functioning
local-level organizations that did something important, such as
distributing agricultural credit or organizing local elections; the
party either faced competition from other parties or held intra
party competitive elections for some offices; and party member
ship was more or less required for government jobs. I gave the
party less weight if its membership seemed to be almost all urban
(with little or no grassroots organization); its politburo (or equiva
lent) served as a rubber stamp for the leader; all members of the
politburo and assembly were in effect selected by the leader; its
membership was dominated by one region, tribe, clan, or religion
(in heterogeneous societies); and the dictator's relatives occupied
high offices.
To classify a regime led by an officer as either military or
personal, I leaned toward military if relationships within the
junta or military council seemed relatively collegial; the ruler
held the rank of general or its equivalent; the regime had some
kind of institutions for deciding succession questions and for

74

Paradigms and Sand Castles

routinizing consultation between the leader and the rest o f the

officer corps; the military hierarchy remained intact; the security
apparatus remained under military control rather than being
taken over by the leader himself; succession in the event o f the
leaders death was in hierarchical order; the officer corps in
cluded representatives of more than one ethnic, religious, or
tribal group (in heterogeneous countries); and the rule of law
was maintained (perhaps after rewriting the laws). I treated the
following as evidence of greater personalism: seizure of execu
tive office by an officer who was not a retired or active duty
general (or the air force or navy equivalent); disintegration of
military hierarchy; dissolution of military councils and other mili
tary consultative institutions; the forced retirement or murder of
officers within the leaders cohort or from tribes or clans other
than the leader's; the murder or imprisonment of dissenting offi
cers or of soldiers loyal to them; the formation of a party led by
the leader as an alternative base of support for himself; and the
holding of plebiscites to legitimize the leaders role. See appen
dix A for a summary of the regime classification criteria.
Most of the time it was not hard to distinguish between military
and single-party regimes, though a few cases, especially in the
Middle East, were problematic. Probably the most difficult deci
sions in this data set involved the current Egyptian regime and
post-1963 Syria. Egypt posed a problem because the regime that
took power in 1952 has gone through a series of changes. In my
judgment, it began as a military regime under Naguib and the
Free Officers but was transformed when Nasser consolidated his
personal power beginning in 1954. Though the military continued
to support the regime, Nasserand Sadat to an even greater
extentincreasingly marginalized it (Springborg 1989). Begin
ning under Nasser, efforts were made to create a single party; this
party achieved some real importance in the mid-1960s but was
then undermined by Nasser (Waterbury 1983; Richards and Waterbury 1990). The Nasser period thus seems primarily personalist. Under Sadat, the party became more important, though his
government also retained large personalist elements (Hinnebusch
1985). The dominant party has played a more important role as
the regime has gone through a modest liberalization. In the Syrian
case, some experts refer to the period after 1963 as a B a athist
regime (Ben-Dor 1975; Perlmutter 1969; Richards and Waterbury
1990), while others emphasize the personal power of Hafez alAsad until his death (Hopwood 1988; M aoz 1986, 1988; Rabino

B ig Questions, Little A n sw ers

75

vich 1972). A s in the Egyptian case, the military is an important

supporter of the regime but seems to have been excluded from
most decision making. The best way to deal with these difficult
cases seemed to be to put them, along with the regimes of Suharto
in Indonesia, Stroessner in Paraguay, and Ne Win in Burm a (or
Myanmar), into a triple hybrid military/personalist/single-party
category. The second section of appendix A lists all the regimes
used in the data analysis and their classifications.
How long an authoritarian regime lasts is not always obvious.
The beginning is usually clear, because dictatorships start either
with an illegal seizure o f power or with a change in rules such
as the banning of opposition parties that in effect eliminates
meaningful competition for the top national office, though oppo
sition parties may be allowed minority representation. But the
end of an authoritarian regime may be less clearly demarcated. I
counted an authoritarian regime as defunct if either the dictator
and his supporters had been ousted from office or a negotiated
transition resulted in reasonably fair, competitive elections and a
change in the party or individual occupying executive office.
Where ousters occurred, I used that date as the endpoint. Where
elections occurred, I used the date of the election, but I did not
include the case unless the winner of the election was allowed to
take office. Elections did not have to be direct, but the body
electing the executive had to be made up mainly of elected mem
bers. Cases in which elections deemed free and fair by outside
observers have been held but have not led to a turnover in
personnel are not treated as transitions because, until they actu
ally step down, we do not know if long-ruling parties such as the
United Malay National Organization (UM NO) or the R evolu
tionary Party of Tanzania (CCM ) really will relinquish power
if defeated.24 The 1992 Angolan elections were deemed free
and fair by outside observers, but few would have called Angola
a democracy in subsequent years. Several of the countries in
which long-ruling parties have won officially free and fair elec
tions, however, probably have taken irreversible steps toward
democracy. Since observers disagree about the classification of
these free and fair countries, tests should be done classify
ing them first as continuing authoritarian regimes and then as
24.
In a study o f transitions in A frica, van de Walle and Butler (1999) show that a
strong relationship exists between executive turnover and scoring at the dem ocratic
end of the Freedom H ouse scale, which suggests an additional reason for not treating
dem ocratization as com plete until a turnover in power has occurred.

76

Parad igm s and S an d C astle s

authoritarian regimes that ended at the time of the free and

fair election. In this data set, these reclassifications make no
substantive difference in the results.
Some of the most difficult classification decisions involved
judgments about whether successive authoritarian governments
should be considered one regime (defined as a set of formal and
informal rules and procedures for choosing leaders and policies)
or not. Authoritarian regimes often follow one another, as, for
exam ple, the Sandinista regime followed the Somozas in Nicara
gua. Data sets that simply identify regimes as authoritarian or
democratic create the impression that amhoritarian regimes are
more stable and longer-lived than they really are, because they
fail to note that one has broken down and another taken its
place. This problem may undermine some of the findings in a
series of studies by Adam Przeworski and coauthors on the rela
tionship between regime type and growth (e.g., Przeworski and
Limongi 1993; Przeworski et al. 2000). In putting together their
data set, they simply coded each country as democratic or not in
December of each year. If a country was coded authoritarian two
years in a row. the regime was considered to have survived,
regardless of whether one authoritarianism had been replaced by
another or a democracy had been formed and then overthrown
during the intervening year.25
I relied on a num ber o f decision rules to avoid this problem .
W here a period o f dem ocracy intervened betw een two periods o f
authoritarianism , I counted the authoritarianism s as separate en
tities. W here one kind o f authoritarian regim e succeeded an
other, as with Som oza-Sandinista, I counted them as separate.
Som e o f these decisions were much m ore difficult. In a num ber
of cases, periods of collegial military rule were succeeded by one
officer's consolidation of his personal power. These I classified as
single regim es undergoing consolidation, unless there was persua
sive evidence that the support base o f the regim e had changed.
W here a c o u p especially if accom panied by a change in clan or
tribal dom inance 01 a substantial m ove down the m ilitary hierar
chy (e .g ., a coup by sergeants against a governm ent led by the
25.
This coding decision does not affect their main finding about the robustness o f
dem ocracy at high levels o f developm ent, but it does undermine conclusions about
the effect o f economic perform ance on authoritarian stability, since ouster.s o f dicta
torships followed by renewed authoritarian rule within (he year arc coded as on-going
regimes.

B ig Questions, Little Answers

77

high command) led to the change in most of the leadership, I

counted it as a regime change. Where one individual who was
already part of a governing junta overthrew another but most of
the rest continued, I counted it as a single regime.
In any study, but especially when the project involves compli
cated or contested decisions about how to classify cases, it is
important to carry out the analogue of sensitivity analysis in
statistics. That is, one should reclassify the cases and see if it
affects conclusions. This might involve including or excluding
cases from the data set, as in the decision about how many years
a dictatorship has to survive in order to be classified as a regime.
If I had followed the usual practice of including every period of
authoritarianism that lasted a year or more, military regimes
would appear even more fragile than they do in the results below,
because most of these very short interventions are military. The
three-year threshold seems to me theoretically correct in that it
derives from the definition of a regime as a set of rules, but it is
also a methodologically conservative decision. If the empirical
investigation turns out to support the argument even though the
most short-lived military interventions have been excluded from
the data set, then we can have greater confidence in the argu
ment. because changing that decision rule would only strengthen
the findings.
The reclassification of cases could, alternatively, involve mov
ing them from one category to another on one of the variables.
For this project, I classified a number of rulers who are often
described as military for example, Barrientos in Bolivia and
Ershad in Bangladesh as personalist because, although they
were officers and camc to power in coups, the military was not
their primary constituency; they organized civilian support and
held popular elections to legitimate their rule. If I were to elimi
nate this criterion for discriminating between personalist and
military rule, a certain number of cases would move from the
personalist to the military category. The changes would not affect
conclusions about the length of military rule, because most of the
cases that would be affected were quite short-lived. The reclassi
fication would, however, increase the number of regimes clas
sified as military that ended in violence, and this could affect
conclusions about another of the implications of the cadre interests argument not tested here: that personalist regimes are
more likely than others to be violently overthrown.

78

Paradigm s and S an d C astles

TABLE 2.1. Durability of Different Types of Authoritarian Regime

R egim e T ype

A verage
L ength of
R ule (years)*

M ilitary
M ilitary/personalist
Personalist'
Single-party hybrids '1
Single-party (stringent
transition criteria)'
Single-party (less stringent
transition criteria)
Triple hybrid

9.5
(33)
11.3
( 12 )
15.5
(46)
19.6
(14)
29.0
(2 1 )
27,9
(24)
33.0
(3)

A verage
A ge o f
Surviving
R egim es '1

P ercent
R egim es
Surviving
in 2000

10.0
(2)

5.7

12.7
(3)
18.0
( 12 )
25.2
(6 )
34.0
(13)
35.4
(9)
43.5
(2 )

20.0

20.7
30.0
38.2
27.3
40.0

Note: Regimes maintained by foreign occupation or military threat are excluded. Number of
observations on which averages are based is shown in parentheses.
^Includes only regimes that had ended by December 2000.
includes regimes in existence in 1946, or that have come to power since then, that still
survived at the end of 2000.
The Rawlings government in Ghana held elections deemed free and fair by international
observers in 1996 (and elections boycotted by the opposition in 1992), and voters reelected
Rawlings. Many then considered Ghana democratic, but by the criteria used for this study its
transition was completed in 2000. If Ghana were classified as having made a transition in 1996,
this change would have no effect on the average length of personalist regimes.
dCategory includes both military/single-party and personalist/single-party regimes.
Six countries in this category have held elections deemed free and fair but nevertheless
returned the ruling party to power. The results if these countries are classified as having
democratized ai the time of the first free and fair elections are shown in the next row.

Testing One of the Implications

Here I describe a test of the first implication above: that mili

tary regimes should be expected to survive less long than other
kinds of authoritarianism.25 Preliminary evidence bears out the
expectation that sources of fragility endogenous to military re
gimes cause life spans shorter on average than those of other
forms of dictatorial rule. Table 2 .1 shows the average life spans
of both the pure and hybrid regime types. Am ong regimes that
had ended by December 2000, military regimes lasted on aver
26 .
Tests of sonic o f the o th er im plications o f the cad re-in terests argum ent are
rep o rted in G ed d es ( 1999 a).

B ig Questions, Little A n sw ers

79

age 9-5 years,27 personalist regimes 15 .5 years, and single-party

regimes (excluding those maintained by foreign occupation or
threat o f intervention) 29 years.28
Another way to assess the durability of different regime types
is to compare their current survival rates. A s shown in column 3
of table 2. 1 , the proportion of surviving military regimes is
quite low. Only 5.7 percent of those that once existed still sur
vive. In contrast, 20.7 percent of personalist regimes remain in
existence, and 38.2 percent of single-party regimes still sur
vived in 2000 if stringent transition criteria are used to deter
mine regime endpoints (27.3 percent if less stringent criteria are
used).29 Military regimes that had come to power by 1997 and
still survived had lasted an average of to years by 2000. Single
party regimes that remained in power, on the other hand, had
lasted an average of 34 years (35.4 if less stringent transition
criteria are used).
Although these differences in the average length of different
types of regime are quite large, we cannot be sure that they are
really caused by regime type. Military regimes are more common
in Latin Am erica, where levels of economic development are
relatively high, and personalist regimes are most common in
A frica, where countries tend to be poorer. It might be that the
stronger demand for democracy by citizens of more developed
countries accounts for the shorter duration of military regimes.
Alternatively, it might be that military regimes last less long
because they are responsible for worse economic performance

27. R em ind er: authoritarian interludes lasting less than three years have been
excluded from the data set. T h e military ruled during most o f these interludes. If they
were included, the average length o f m ilitary rule would be reduced. Nordlinger, who
did not exclude them from his calculations, found that m ilitary regimes last five years
on average (19 77, 139).
28. Regim es maintained in power by direct foreign occupation or the threat of
m ilitary intervention have been excluded from the calculation of average life span
here and from the statistical analysis below because their longevity depends on exter
nal events. The excluded regim es are those in A fghanistan, 19 79 -9 2; B u lgaria, 19 4 7 go; C am bod ia, 19 79 -9 0 ; C zech oslovakia, 19 48-9 0; G erm an D em ocratic R epublic,
19 4 5 -9 0 ; Hungary, 1949-90; and Poland, 1947-89. The average length of these
regim es is 34 years.
29. The stringent criteria fo r determ ining the end o f an authoritarian regim e re
quire not only that com petitive elections be held but also that the executive change
hands. The less stringent criteria count authoritarian regim es as ended if com petitive
elections are held and are considered free and fair by outside observers, regardless of
w ho wins.

8o

Paradigm s and San d C astle s

than other kinds of authoritarianism. To test for these possibili

ties, I have carried out statistical tests of the effect of regime type
on the probability of regime breakdown, controlling for level of
development, growth rate, and region.
I
use a hazard model to assess these rival arguments. Hazard
models are used in medical research and other areas to predict
the survival of individuals with certain conditions, given various
treatments. This type of model also seems appropriate for ex
plaining the survival of a different kind o f entity. A logit model
produces the same substantive results, so they do not depend on
the particular specification.
To rule out the possibility that the apparent relationship be
tween regime type and length of time in office might really be
caused by level of development, the statistical analysis includes
an indicator of development, the natural log of G D P per capita.
Since a number of studies have found that current economic
performance affects the likelihood of regime breakdown, an indi
cator for growth is also included in the models as a control
variable. The measure of growth used is change in G D P per
capita in the prior year. I use the prior year bccause credit or
blame for the prior year's economic performance is unambigu
ous. In years in which a transition takes place, the outgoing
regime might be responsible for only part of the years perfor
mance. Furthermore, economic performance is often erratic in
transition years. It can plummet in response to government insta
bility. but it can also improve rapidly during the euphoria that
sometimes accompanies a transition. Thus, the previous years
growth seems a better indicator of the regime's recent economic
performance. Adam Przeworski and Fernando Liniongi (1997)
also found, after trying a number of possibilities and lags, that
growth during the prior year was the best predictor of regime
change.
Economic data are from the Penn World Tables, the longest
time series for the largest number of countries I have been able
to find. For most countries, it covers 1950-92, which means that
regime years prior to 1951 and after 1992 are excluded from the
statistical analysis. In addition, no economic data are available
from Albania, Cambodia. Cuba. North Korea, Libya, Vietnam,
and South Yemen, and there are some years missing from a few
other countries. Since the period covered is quite long and I
cannot think of any reason to believe that transitions during the
years covered would be different from those in the years immedi

B ig Questions, Little Answers

81

ately before and after, I do not think the years excluded intro
duce bias into the results.
The countries left out of the data set, however, differ from
those included. Most had or have single-party or personalist/
single-party regimes. Their regimes have lasted an unusually
long time (excluding Cambodia, 32 .1 years on average). Dicta
tors still rule in five o f seven, and nearly all the countries are very
poor. If they were included in the data analysis, they would
probably further strengthen the coefficient for the effect of
single-party regime and reduce the effect of level of development
on the probability of regime stability. Since the data set is large,
however, and not very many cases have to be left out, I do not
think their exclusion has much effect on conclusions.
Region is used as a quasi-fixed effects estimator.30 Fixed ef
fects estimators are used to hold constant aspects of history and
culture that might affect the outcome of interest but that cannot
be directly measured. I have used region to hold constant some
of the possible effects of colonial history and cultural heritage.
Because regime types are nominal categories, they are en
tered into the model as dummy variables: if the regime is, for
example, military, it is coded on e ; otherwise, it is coded
zero. The lef(-out regime type is personalist, the middle cate
gory in terms o f longevity. Thus, the hazard ratios reported
should be interpreted as referring to differences between the
effect o f the type of regime associated with a particular ratio and
the effect of personalist regimes.
Hazard ratios have a simple intuitive interpretation. Ratios
above one mean that the variable associated with them increases
the probability of regime collapse. In the first column of table
2.2, the hazard ratio for military regime is 2 .8 1, which means
30.
Usually, country dumm y variables are used as fixed effects estim ators, but
they could not be used to analyze this data set because they cause countries with only
one regime to be dropped from the analysis. In this data set, half the countries have
had only one authoritarian regim e, either because one stable regim e remained in
power for several decades or because the country is usually dem ocratic and had only
one postwar authoritarian interlude. A more serious problem than the loss of cases
per se is that regim es in the cases with only one regim e, are. on average, unusually
long-lived, and they are especially likely to be single-party regim es. T he use of
country fixed effects estim ators eliminates 60 percent o f the single-party regimes
from the analysis. When the analysis was done using country fixed effects estim ators,
the coefficient for the effect o f m ilitary regim e was artificially strengthened (since the
longest m ilitary regim es w ere elim inated), and th e effect of single-party regim e was
greatly w eakened (since most of the single-party regim es w ere elim inated, leaving an
unrepresentative set of m ostly A frican cases).

82

Paradigm s and Sand C astles

that, all else being equal, military regimes are nearly three times
as likely to break down as personalist regimes. Hazard ratios
between zero and one mean that the variable reduces the proba
bility of breakdown. In column i, the hazard ratio for single
party regimes, .39, means that, all else being equal, single-party
regimes have about 40 percent the chance of collapsing that per
sonalist regimes do.
As can be seen in column 1, military regimes break down
more readily than all other types. The hazard ratio for the mili
tary regime variable is substantively large and statistically signifi
cant. The two intermediate regime types, military/personalist
and single-party hybrid (in which personalist/single-party re
gimes predominate, since there are very few military/single-party
regimes), are, not surprisingly, not very different from personal
ist regimes. Single-party regimes, however, are more resilient
than personalist regimes to about the same extent that military
regimes are less resilient, and this difference is also statistically
significant. Finally, the triple hybrid regimes, which combine
characteristics of single-party, personalist, and military regimes,

TABLE 2.2. Effect of Regime Type on Authoritarian Survival (Weibull

regression, log reiative-hazard form)
D ependent V a riab le: Regim e C ollapse (hazard ratios)

1
2 .81 **
1.31
1.24
0 .39 **
0 .04 **
0 .53 *

M odel
M ilitary
M ilitary/personalist
Single-party hybrids
Single-party
Triple hybrid
Log G D P per capita
Grow th G D P per capita
A sia
C entral A m erica, Caribbean
Eastern E u ro p ea
M iddle East
North A frica
South A m erica
Sub-Saharan A frica
Percent Muslim
D ependence on oil
Dependence on minerals
N o f observations

0.02*
1.2 2

0.99
0 . 18 *
7.46
0 .42 *
3.44
0.61

2
2 .83 **
1.15
1.47
0 .38 *

M odel

0.00

0 .54 *
0.02*

0.97
0.98
0.16
3.29
0.15
3.95
0.41

1 .0 1 *

1,694

1,627

Note: Lefi-oul regime category is personalist; left-oul region is Southern Europe.

Excludes regimes maintained by foreign intervention.
* Statistically significant at .05 to .01; **stalisiically significant at .01 or better.

3
10. 26 **
2.07
3.44 *
0.59

M odel

0.00

0.40
0 .004 **
0.19
0.23
0.00

0.40

0 .0 1* *

0.35
0 .05 **
1.0 2**
1.02

1.00

861

IKLALSO - tttoiioteca
B ig Questions, Little A n sw ers

83

are the strongest of all. An alternative to the substantive interpre

tation of this category is that it simply serves the purpose of
controlling for five very long-lived and unusual regimes that
might otherwise inflate the apparent longevity of single-party
regimes.
The control variables used in the regression also show some
interesting effects. A s the level of development rises, authoritar
ian regimes, like democratic ones, become more stable. This
finding is consistent with that of John Londregan and Keith
Poole (1990, 1996), who found that the best predictor of coups,
in both democratic and authoritarian regimes, was poverty. It
raises some questions, however, about traditional demandcentered explanations for the relationship between increased de
velopment and democracy. It is inconsistent with the idea that
the citizens of more affluent countries are more likely to demand
democratization. Rather, it suggests that when authoritarian gov
ernments manage the economy well over the long term, regime
allies remain loyal and citizens remain supportive, or at least
acquiescent. That interpretation is reinforced by the very strong
negative effect of short-term economic growth on the probability
of regime breakdown. In other words, both long- and short-term
economic performance affects authoritarian stability.
In light of various arguments about the effects of religion,
culture, and colonial heritage on the development of democratic
values, it is somewhat surprising that most of the region variables
show little effect. The left-out region here is southern Europe
(Portugal, Spain, and Greece), and we might have expected the
regions most culturally distinct from Europe to exhibit differ
ences in the likelihood o f regime transition. The only two regions
with statistically significant hazard ratios in model 1, however,
are Eastern Europe and North Africa. Since the governments
kept in place by the threat of Soviet intervention have been
excluded from these tests, the East European region contains
only the Soviet Union, Rom ania, and Yugoslavia (Albania had
to be excluded because of missing economic data). Controlling
for level of development, growth, and regime type, regimes in
these countries were unusually resilient. There are internal rea
sons for this resilience, but in the context of the full data set, this
region dummy variable in effect controls for the very unusual
longevity of the Soviet regime and prevents it from inflating the
apparent effect of the single-party regime type.
The unusual resilience of authoritarian regimes in North

84

Paradigm s and S an d C astles

A frica leads to speculations about other possible causes of au

thoritarian stability that might have been left out of these tests.
Several studies have shown that countries with large Islamic
populations are less likely to be democratic, and Michael Ross
(2001) shows that oil wealth is associated with authoritarianism.
North African exceptionalism might be caused by Islam or by
oil wealth, though it should be noted that the hazard ratio for
the Middle East is not only insignificant but suggests the oppo
site effect.
The results of tests of these possibilities are shown in columns
2 and 3 o f table 2.2. For the hazard model used in column 2, the
percentage of the population that is Islamic was added as an
additional control variable. The first thing to notice is that the
inclusion of this control variable has virtually no effect on the
relationship between regime type and breakdown. Those rela
tionships look as strong as ever, and the hazard ratios for growth
and level of development are also unaffected. The absence of
change increases our faith in the importance of the variables of
interest.
The effect of percent Muslim population on the probability of
breakdown would be quite interesting in its own right if we be
lieved it. It is statistically significant, though the effect is the
opposite of that expected. A s Muslim population increases, the
probability of regime collapse becomes more likely. This finding
does not, of course, mean that democratization is more likely in
Islamic countries. Authoritarian regimes collapse and are fol
lowed by other authoritarian governments. If this finding were to
be replicated in additional tests, it would disconfirm the idea that
authoritarianism is more stable in Islamic countries because of an
affinity between Muslim culture and authoritarian values. The
results so far have to be considered quite tentative, however,
because of the exclusion of monarchies from the data set. Since
most of the extant monarchies are both long-lived and in pre
dominantly Muslim countries, their inclusion might well cause
the disappearance of this finding.
Adding dependence 011 oil and minerals to the model (col
umn 3) reduces the number of observations by about half and
causcs the hazard ratios to bounce around quite a bit. Statistical
significance is harder to achieve in the smaller data set. Because
of peculiarities in the set of cases for which data on oil exports
are available, findings from the third model should probably

B ig Questions, Little Answers

85

not be taken too s e r i o u s l y .^ 1 For what they are worth, however,

the hazard ratio for military regimes is still statistically signifi
cant and remarkably large, growth still reduces the likelihood of
breakdown, and percent Muslim has about the same effect as
before. Single-party regime and level of development still re
duce the probability of breakdown but have lost statistical sig
nificance. Dependence on oil apparently has no effect on au
thoritarian stability, contrary to much that has been written in
the literature on rentier states, though here also the exclusion
of monarchies renders the conclusion suspect. In short, these
findings are mostly consistent with expectations drawn from the
cadre-interests argument, though if we had faith in the quality
of the data on which the findings are based, the loss of statisti
cal significance for the single-party variable would be cause for
concern.
To summarize the findings, the hypotheses about the average
duration o f different types of authoritarian regime have been
mostly confirmed by statistical analysis, holding constant the most
obvious challenges to the apparent relationship. Growth was
found to have the expected effect of reducing the probability of
regime breakdown. Higher levels of development also probably
reduced the likelihood of authoritarian breakdown. In short, em
pirical investigation of the first implication of the cadre-interests
argument about authoritarian breakdown has failed to disconfirm
expectations.
A series of methodological observations can be made about the
empirical test described above. The first observation, though
obvious, may need restating: this was a test of an implication of
the argument, not of the argument itself. The implication is quite
simple and, once the data had been collected, easy to test. Never
theless, its confirmation adds to the persuasiveness of the argu
ment, and if the other implications listed above also proved con
sistent with reality, we would be pretty much convinced that the
argument captures a key aspect of the explanation of the break
down of authoritarian regimes.
31.
F or nearly all A frican countries, most years after 1983 are m issing, which
means that most A frican transitions are missing. In other w ords, the missing cases are
almost all from the poorer half o f the data set. which is probably the reason that level
of developm ent loses significance. T he later years of a considerable nuniher o f longlived A frican single-party regim es are also left out.

86

Paradigm s and San d C astles

Second, in order to test the implication, evidence had to be

gathered about a large number of cases, although the kind of
evidence needed was not complicated. It is not always necessary
or feasible to include the whole universe of cases, as I did here. If
examining the whole universe is infeasible, however, the whole
universe nevertheless needs to be identified so that cases from
within the universe can be selected at random or in some other
way that does not bias conclusions. (Figuring out the domain of
an argument is discussed in chapters 3 and 4.) Cases should be
selected to ensure that the outcome of interest varies across
them. (Case selection is treated in detail in chapter 3.) Cases that
have been studied repeatedly by other scholars are, on average,
larger, more developed, and more geopolitically important than
the cases that have not been studied, and conclusions based on
experience in such cases are therefore unlikely to be representa
tive of the whole group.
The number of cases used to test an argument needs to be
reasonably large, since it is very hard to be sure that a result has
not been caused by chance events when only a few cases are
examined. The more arguments one can think of other than the
one of interest that could also explain an apparent relationship,
and the more factors one thinks need to be held constant in order
to exclude the effects of irrelevant forces on the relationship of
interest, the more cases need to be examined. The possibility of
spurious correlation that is, the possibility that what appears to
be a relationship between some cause and effect really results
from some outside factor that causes both of them can rarely
be dismissed without using statistics, and not always even then.
All data sets, whether gathered by the researcher or taken off
the shelf, contain some missing data and some mistakes. The
researcher should always think carefully about how the missing
cases differ from the cases included in the study and how their
inclusion, if it could be managed, would affect conclusions.
When using an off-the-shelf data set known or rumored to con
tain mistakes, the researcher should try to figure out whether the
mistakes are likely to affect conclusions. The Freedom House
democracy indicator, for example, is rumored to contain a proWashington bias, especially during the early years. For some
purposes, this bias would not matter, but for others it might
seriously undermine the credibility of findings. One would want
to use a different indicator o f democracy in the latter situation.

B ig Questions, Little A nsw ers

87

Conclusion

This chapter began with the claim that one of the practices hinder
ing the accumulation o f theory in comparative politics is the way
we usually go about trying to explain compound outcomes such as
democratization. I argued that greater progress could be made
toward actually understanding how such outcomes occur by exam
ining the mechanisms and processes that contribute to them,
rather than through inductive searches for the correlates of the
undifferentiated whole. Coherent deductive arguments can be
devised to explain constituent processes, and hypotheses derived
from the arguments can be tested.
I attempted to demonstrate the usefulness of this approach
with an extended example. After identifying seven constituent
processes of the large, complicated phenomenon of democratiza
tion, 1 proposed a deductive argument based on the individual
interests of regime insiders to explain why elite splits play a
larger role in some instances of authoritarian breakdown than
others, and why some authoritarian regimes initiate political lib
eralization in the absence of societal pressure to do so. Although
this argument as a whole is not testable, it was a simple matter to
derive implications of the argument that could potentially be
falsified.
The only impediment to testing these hypotheses was the need
to gather an appropriate data set. Data gathering was a major
and time-consuming effort, but once the data had been gathered,
it was possible to show not only that predicted differences ex
isted, but that they were quite large and statistically significant.
I make no grand claims for the cadre-interests argument itself.
It may not be true. It is possible that when other variables are
taken into consideration, the relationship that seems apparent
now between regime type and longevity will disappear. Even if
true, the argument explains only one element of the compound
process of regime transformation. I do claim, however, that an
argument from which an implication has been tested on evidence
from a large number of cases is more likely to prove of lasting
value than untested arguments induced from a handful of cases.
And once data have been gathered, more implications can be
tested. If those tests also conform to expectations generated by
the argument, our confidence that the argument is true will in
crease. I also claim that to tack down an explanation of one

88

Paradigm s and S an d C astles

process that contributes to a compound outcome of great theo

retical and real-world importance such as democratization would
constitute serious intellectual progress.
While inductive explorations of instances of transition may
have been the only possible research strategy at the beginning of
the current wave of democratization, we now have enough basic
information to move on to theory building. I have tried to show
here that the theoretical edifice can best be built one deductive
brick at a time, testing as many of them as possible using evi
dence from a large number of cases.

CH APTER 3

How the Cases You Choose Affect

the Answers You Get
Selection Bias and Related Issues

Comparative politics, like other subfields in political science, has

norms and conventions about what constitutes an appropriate
research strategy and what kind of evidence makes an argument
persuasive. Although the norm has begun to change, for many
years one of our most durable conventions was the selection of
cases for study from one end of the outcome continuum we
wished to explain.1 That is, if we want to understand something
for example, revolution we select one or more occurrences and
subject them to scrutiny to see if we can identify antecedent
events or characteristics as causes.
Most graduate students learn in the statistics courses forced
upon them that such selection on the dependent variable often
leads to wrong answers, but few remember why, or what the
implications of violating this rule might be for their own work.
And so, comparativists often ignore or forget about it when un
dertaking or assessing nonquantitative research.
This chapter demonstrates the consequences of violating the
rule. It does so by comparing the conclusions reached in several
influential studies based on cases selected on the dependent vari
able with retests of the same arguments using samples not corre
lated with the outcome. A ll the studies discussed in this chapter
are intelligent, plausible, insightful, and possibly correct in their
knowledge claims. All have been advanced by highly respected
social scientists. The effort here is not to discredit arguments or
belittle authors who are, after all, working within accepted
conventions but to demonstrate the deficiencies of the conven
tions themselves.
i.
Com parative politics is not the only field bedeviled by problem s with selection
bias (see A chcn and Snidal 1989).

90

Paradigm s and Sand C astles

These conventions affect not only authors but readers of com

parative politics. Authors, including some of those discussed
below, are frequently aware of the tentativeness of the evi
dence supporting their arguments and indicate their awareness
in the caveats they attach to them. Readers, however, tend to
ignore the caveats and give greater weight to unsystematic evi
dence than it deserves. Many studies in which authors have care
fully hedged their explanatory claims are discussed in seminars,
cited in literature reviews, and summarized in qualifying exams
as though the tentative arguments advanced were supported by
solid evidence. The purpose of this chapter is as much to de
crease the credulity of readers as to increase the sophistication of
researchers.
The message of the chapter is not that the examination of
cases selected because they have experienced a particular out
come is never warranted, but rather that the analyst should un
derstand what can and cannot be accomplished with cases se
lected for this reason. Some kinds of tests of conditions proposed
as necessary or sufficient for explaining outcomes can be carried
out using only cases that have experienced an outcome, although
assessment of what Braumoeller and Goertz (2000) refer to as
trivialness requires at least some information about the rest of
the universe of cases.2
The close examination of an anomalous case with a particular
outcome can also serve a useful role in either generating a pro
posed revision of current theory or suggesting domain conditions
not previously understood. A test of the proposed revision or
domain condition would require examining a wider range of
cases, however. Although the proposal of a revision is a useful
2.
A s D ion (1998) has pointed out, selection on the dependent variable does not
undermine tests of necessary but not sufficient or necessary and sufficient argu
ments. B rau m o eller and G oertz (2000) propose a series o f tests that, taken together,
would increase confidence in a necessary or sufficient argum ent. C arryin g out these
tests requires: ( 1 ) being able to estim ate the error in the m easurem ent o f both
proposed causes and effects; (2) including enough cases selected to have the outcom e
so that an appropriate statistical test can reject the null hypothesis (with no m easure
ment error, the minimum number is seven; as m easurem ent error increases, so does
the required number of cases); (3) collecting enough inform ation about the full
universe o f cases to assure on eself that there is enough variation in both purported
cause and outcom e to avoid trivialness. T he issue o f trivialness is discussed below. It
refers to proposed necessary conditions that are theoretically m eaningless because
they vary little if at all. B rau m o eller and G oertz note, for exam ple, that the argument
that dem ocratic dyads are necessary for peace is trivial before 1800 because there
were no dem ocratic dyads then.

H ow the Cases You Choose A ffect the A nsw ers You G et

91

contribution to knowledge building, the revision should not be

accepted until it has been tested and confirmed on a representa
tive sample o f cases.
The Nature of the Problem

The adverse effects of selecting cases for study on the dependent

variable stem from the logic of inference. When one sets out to
explain why countries A and B have, say, developed more rapidly
than countries C through I, one is implicitly looking for some
antecedent factors X through Z that countries A and B possess in
greater degree than do countries C through I. The crux o f the
difficulty that arises when cases are selected on the dependent
variable is that if one studies only countries A and B , one can
collect only part of the information needed, namely, the extent of
factors X through Z in countries A and B . Unless one also studies
countries C through I (or a sample of them) to make sure they
have less of X through Z, one cannot know whether the factors
identified really vary with the outcome under investigation.
The problem becomes more obvious when shown in graphs
rather than expressed in words. Suppose a universe of developing
countries A through I, where A and B are among the fastest grow
ing. On the basis of an intensive study of A and B , one concludes
that factor X is the cause of their success. In concluding this, one
implicitly assumes that if countries C through I were examined,
they would turn out to have less of factor X than do A and B , and
that one would observe the relationship shown in figure 3 .1.
Yet if one examines only countries A and B , it is possible that
the full range of cases would look more like one of the scatterplots
in figure 3.2. That is, it is possible that there is no relationship

*H

Faclor X

Fig. 3.1.

Assumed relationship between factor X and growth

92

Paradigm s and S an d C astles

Growth

Fad or X

Fig. 3.2.

Alternative possible relationships between factor X and growth

between X and the rate of development. The only things that can
actually be explained using a sample selected on the dependent
variable are differences among the selected cases.
When one looks only at the cases above the broken line in
figure 3 .1. two kinds of mistaken inference can occur. The first
involves jumping to the conclusion that any characteristic that
the selected cases share is a cause. The other involves inferring
that relationships (or absence of relationships) between variables
within the selected set of cases reflect relationships in the entire
population of cases.
In the statistical literature, attention has focused on the sec
ond kind of faulty inference (Achen 1986; King 1989). If the true
relationship between factor X and the dependent variable is that

H ow the Cases You Choose Affcct the Answers You Get

93

shown in figure 3.1 but one selects cases in a manner that results
in the examination only of cases located above the broken line,
statistical procedures carried out on the selected cases may indi
cate that no relationship exists or even that the relationship is the
opposite o f the true one. Selection on the dependent variable
biases statistical results toward finding no relationship even when
one does, in fact, exist.
In nonquantitative work, however, the first kind of faulty
inference is at least as common as the second. If the main causes
of the dependent variable are factors R through T, not including
X, and one selects cases from one end of the dependent variable,
X may appear to be important in the selected sample either
because of random variation or because it explains some of the
differences among cases still remaining in the data set even after
the selection has limited it (01 because it is correlated with some
other factor that explains the remaining differences). In the
former situation, the true relationship might look like one of
the panels in figure 3.2, but the analyst on the basis of bits and
pieces of information rather than a systematic check assumes
that cases C through I are located in the lower left quadrant and
concludes that factor X causes the outcome of interest even
though, in fact, no relationship exists. In the latter situation,
factor X makes a minor contribution to the outcome, but the
analyst overestimates its importance. An example should help to
make these points clearer.
A Straightforward Case of Selection on the
Dependent Variable

Analysts trying to explain why some developing countries have

grown so much more rapidly than others regularly select a few
successful new industrializing countries (NICs) for study. Prior to
the debt crisis, which began in 1982, the cases most often exam
ined were Taiwan, South K orea, Singapore, Brazil, and Mexico
(see, e.g., Haggard 1990). In all these countries, during the peri
ods o f most rapid growth, governments exerted extensive control
over labor and prevented most expressions of worker discontent.
Having noted this similarity, analysts argue that the repression,
co-optation, discipline, or weakness of labor contributes to high
growth. Chalmers Johnson (1987, 149). for example, asserts that
weak unions and federations of unions devoid of all but token
political power are real comparative advantages in international

94

Paradigms and Sand Castles

economic competition. Frederic D eyo (1984, 1987) argues that

an export-led growth strategy depends on cheap skilled labor
and, consequently, a disciplined and quiescent labor force. H a
gen Koo (1987) claims that labor control is needed in order to
attract foreign investment.3
These claims draw additional plausibility from their conver
gence with arguments made in studies aimed not at explaining
growth but at understanding authoritarian interventions in the
more developed countries o f Latin Am erica. Am ong the best
known of these is Guillerm o O Donnells argument (1973) that
the transition from the easy stage of import-substitution industri
alization to a more capital-intensive stage creates a need for
reduced consumption and, hence, a demand for the repression of
labor.4 In the same vein, Fernando Henrique Cardoso (1973a)
and Peter Evans (1979) argue that labor repression helps attract
foreign investment.
Whatever the details of the argument, many scholars who have
studied the N ICs seem to agree that repression or co-optation of
the labor force contributes to growth. Taiwan, South K orea (espe
cially between 1961 and 1986), Singapore (after 1968), Brazil
(19 6 7 -8 ]), and M exico (before 1982) all had repressed and/or
co-opted labor forces and relatively high growth rates. In other
words, all have the outcome o f interest and all exhibit another
common trait labor repression so analysts conclude that la
bor repression has caused the outcome.
But that conclusion is unwarranted. Perhaps there are other
countries in which labor suffers at least as much repression as in
the high-growth countries examined but that have failed to pros
per. In order to establish the plausibility o f the claim that labor
repression contributes to development, it would be necessary to
select a sample of cases without reference to their position on the
dependent variable (growth), rate each case on its level of labor
repression, and show that, on average, countries with higher
levels of repression grow faster.
To be persuasive, theories must be tested on at least a few
cases other than those examined in the initial development o f the
3. H aggard (1986, 3 5 4 -5 6 ) provides a careful and nuanced review o f several of
these argum ents.
4. T h e dependent variable in O D on nell s study is regim e typ e, not grow th, and
its research design is exemp/ary. O D onneJJ com pared the two countries that had
experienced m ilitary intervention with a set o f other Latin A m erican countries that,
at the time he w rote, rem ained dem ocratic.

H ow the Cases You Choose A ffect the Answers You Get

95

idea. A t the stage of theory development, it is virtually im pos

sible to avoid overfitting, that is, tailoring arguments to fit the
circumstances found in particular cases. Testing arguments on
other cases allows the analyst to discover which factors proposed
as possible causes during the discovery stage o f theory building
really do have general causal influence and which should, in the
context o f a general argument, be thought of as part of the error
term. The error term contains all those serendipitous, conjunctural, and other kinds o f factors that contribute to particular
outcomes in particular cases but that do not systematically influ
ence outcomes.
Domain of the Argum ent

To test this or any other hypothesis, one must first identify the
universe of cases to which the hypothesis should apply and then
find or develop measures of the hypothesized causes and effects.
The theory or hypothesis being tested determines the appropri
ate unit of analysis and the universe of potential observations.
If a theory suggests a relationship between some cause and
individual behavior, the test of hypotheses derived from that
theory should be based on observations of individuals. Where
the unit of analysis is the individual, valid inferences can often be
made in studies of single countries or even single towns, because,
unless the town has been chosen precisely because the particular
kind of individual behavior to be explained prevails within it,
observing a range of individuals within a town does not entail
selection on the dependent variable. The full range of individual
variation may well occur within a town. Thus, for example, the
research design used in William Sheridan Allens The Nazi Sei
zure o f Power (1973) avoids selection bias by including both
individuals who embraced Nazism and those who resisted, and
also by including change in individual attitudes over time.5
If, however, the hypothesis predicts country-level outcomes,
as those linking labor repression and growth usually do, one
should test it on a set of countries that reflects a reasonable range
5.
In his critique o f K in g, K eoh an e, and Verba (19 9 4 ), R ogow ski (1995) has
noted that A lle n s thoughtful study o f one town in which Nazism enjoyed an early
and substantial success deepens and enriches our understanding of the rise o f Nazism.
In the com parative field, we are inclined to equate cases autom atically with territorial
entities, but the unit o f analysis used by A llen is clearly the individual, not the town,
and thus he did not select on the dependent variable.

96

P arad igm s and San d C astle s

of variation on the country-level outcome. In short, the cases on

which an argument is tested should reflect the level of analysis at
which the argument is posed.
In everyday language, a case is a single entity, most often a
country, but possibly a city, region, agency, administration, social
movement, party, revolution, election, policy decision, or virtu
ally anything else that involves interacting human beings. The
more technical definition of a case is a unit within which each
variable measured takes on only one value or is classified in only
one category (Eckstein 1975). Many everyday language case stud
ies include multiple technical cases, otherwise known as observa
tions. Much of the disagreement in the literature over the useful
ness of case studies has arisen from a confusion between, on the
one hand, the everyday usage of the word case to mean (usually)
a country; and, on the other hand, the more technical usage of
case to mean an observation the sense intended by those who
give methodological advice, such as King, Keohane, and Verba
(1994).
The appropriate universe of observations on which to test a
hypothesis depends 011 the domain implied by the hypothesis. In
other words, the domain depends on the substantive content of
the theory or hypothesis itself, not necessarily on the author's
statements about where the argument should apply. If an analyst
proposes a theory about the effects of industrialization on late
developing democratic countries, then tests of the theory can and
should be carried out on a sample of countries drawn from the
universe of all late developing democratic countries. Theories
can contain substantive elements that limit their domain to par
ticular regions of the world or time periods, and, if so, those
limitations should be kept in mind during testing. Theories are
not, however, automatically limited to the domain within which
they were first proposed. Authors sometimes fail to realize that
their arguments might apply to countries with which they are
unfamiliar.
Well-intentioned scholars can disagree about what constitutes
the appropriate domain of a theory, but their disagreements
should derive from different interpretations of the implications
of the theory. Tests of hypotheses in controversial domains can
be useful in establishing clearer limits to the domain, extending
it. and suggesting new hypotheses about why the domain has the
limits it does. It is also legitimate to test arguments in domains
outside those implied by theories to see whether the theories

H ow the Coses You Choose A ffect the Answers You Get

97

have greater generality than their creators realized, though nega

tive results in such tests fail to disconfirm the argument within its
original domain.
If the whole universe of cases is too large to study, examina
tion of a random sample is usually recommended as a means of
ensuring that the criteria of selection do not correlate with the
dependent variable. One can, however, make valid inferences
from any sample selected in a way that does not inadvertently
result in a set of cases clustered at one end of the outcome
continuum. Moreover, randomization does not guarantee the
absence of correlation. If, at a particular time, the universe con
tains only cases that have passed a certain threshold of success
because nature" has in some fashion weeded out the others,
then even random or total samples will, in effect, have been
selected on the dependent variable. If, for example, potential
states that failed to adopt a given military innovation in the
fifteenth century were later defeated and incorporated into other
states, one would not be able to find evidence of the importance
of this innovation by examining a random sample of the states
that existed in the eighteenth century. All surviving states would
have the innovation.6
Some theories have implications that apply to only one end
of the dependent variable. To test hypotheses based on these
implications, the analyst must, of course, choose cases from the
relevant part of the outcome continuum. This may appear at
first glance to entail selection on the dependent variable, but it
does not. The outcome relevant for the test of a particular
implication is the outcome predicted by this hypothesis about
that implication, not the outcome explained by the theory. The
full range of variation in the outcome predicted by the hypothe
sis may be contained at one end of the outcome predicted by
the theory. For example, one of the implications of the cadreinterests argument described in chapter 2 is that military govern
ments are more likely to negotiate their extrication from power
than are personalist regimes. One way to test this implication is
to compare the incidence of negotiation by different kinds of
dictatorship during the years in which breakdown occurs. In
other words, only regimes that had experienced breakdown
would be included in the test (one end of the breakdown versus
6.
A n extensive and thought-provoking discussion o f selection by nature can he
found in Przeworski and Lim ongi (19 9 3).

98

Paradigm s and Sand C astles

persistence outcome continuum), but the hypothesis actually

being tested is about the incidence of negotiation during transi
tions, not about the causes of breakdown. The outcome contin
uum relevant for testing this hypothesis is the negotiation versus
no negotiation continuum, not the breakdown versus persis
tence continuum.
For the hypothesis that labor repression contributes to growth,
different arguments about the specific reasons a weak labor force
might have this effect imply different domains for the argument.
One possibility is that the domain should simply include all devel
oping countries. In one of the tests of the argument below, I have
included all developing countries for which the Penn World
Tables collected data between 1970 and 1982, except those with
communist governments, those embroiled in civil war for more
than a third o f the period covered, and those that are extremely
small (fewer than 500,000 inhabitants).7 Communist countries
are excluded because the various theories apply only to countries
with capitalist or mixed economies. The other exclusions involve
countries with characteristics not related to labor repression that
could be expected to affect greatly their growth rates and thus
might distort the apparent relationship between labor repression
and growth. In the second test, I narrow the domain to conform
to arguments associated with O'Donnell and others who expect
labor repression to contribute to growth once a certain threshold
of development has been reached.
Measurement

The outcome to be explained, growth rate, presents no measure

ment problems; various measures are readily available. For this
test, I used the Penn World Tables to calculate growth in G D P
per capita between 1970 and 1982, since most of the studies of
development strategies focus on the period before the debt crisis.
A further test of the hypothesis that included economic perfor
mance in the far more adverse post-1982 international economic
environment would also be interesting and useful.
The hypothesized cause labor repression, co-optation, or
quiescence is more difficult to measure. Standard indicators
are not available, and labor repression can take different forms
7.
D eveloping countries are defined as those with per capita incom e below $4,200
in 1979. T his cut-off point excludes wealthy oil exporters (per capita income above
$4,200), Saudi A ra b ia , K u w ait, L ib y a, and the United A ra b E m irates.

H ow the Cases You Choose Affect the Answers You G et

99

in different contexts, for example, state co-optation in one coun

try and private violence against workers in another. To deal with
this difficulty, I developed criteria for ranking each country on
labor repression, using the Country Reports on Human Rights
Practices prepared for congressional committees on foreign rela
tions (U.S. Department o f State, 1979-83), Amnesty Interna
tional Annual Reports (19 7 3-8 3), and many studies of labor in
particular countries.
Eighty-four developing countries were given scores between
zero and one for every year between 1970 and 1981 on five
factors expected to contribute to the ability of workers to defend
their interests:
The extent to which unions are legal and free to
function
The autonomy o f unions from government or rulingparty control or manipulation
The right to bargain collectively and to strike
The degree o f political participation allowed to workers
and the organizations that represent them
Freedom from violence, arbitrary arrest, and other
forms of repression
When these factors are combined, possible scores range from
zero to five, with high scores indicating extreme control and
repression and low scores reflecting freedom to organize, inde
pendence from ruling parties, legal protection of the right to
bargain and strike, freedom to participate in politics, and protec
tion from violence and repression. Countries with very low
scores include F iji, Mauritius, and Jamaica. The highest scorers
are Uganda, Haiti, and Iraq. The countries included and their
average labor repression scores are shown in appendix B .
In countries that experienced regime changes, policies toward
labor usually changed along with the government. Yearly scor
ing o f each country allowed those changes to be tracked. The
coding sheet that was used to keep track of information while
consulting multiple sources is shown in appendix B , along with
the coding scheme. The coding scheme gives careful rules for
translating the information gathered into numbers.
The purpose of coding sheets and coding rules, discussed at
greater length in chapter 4, is to help make sure that the same
factors are assessed in every case and that they are all judged

io o

Paradigm s and San d C astles

using the same criteria. While it is obvious that some such aid to
memory is required for dealing with eighty-four cases, explicit
coding rules also increase the precision of studies that focus on
only a few cases. I would urge getting into the habit of writing
down explicit coding rules, no matter what the number of obser
vations. It helps the analyst stick to the same rules across coun
tries and time, and it also helps readers understand exactly what
the analyst means when she makes assessments of key causal
factors. The phrase labor repression no doubt has somewhat
different connotations for scholars with different areas of exper
tise. but the person who has read the coding scheme in appen
dix B will have a very clear idea of what is meant by the term
here.
Although the indicator of labor repression created in this way
is an imperfect measure of a complex set of phenomena, and
experts might have small disagreements about the placement of a
few cases, this measure is at least as precise as the verbal descrip
tions available in the literature. It seems, therefore, adequate to
the present task of demonstrating a methodological point.
Tests of the hypothesis linking labor repression to growth us
ing these data are shown in figures 3.3. 3.4, 3.5, and 3.6. Figure
3.3 shows the relationship between average labor repression and
average growth from 1970 to 1981 for the sample of NICs most
frequently studied (Taiwan, South K orea, Singapore, Brazil, and
M exico). This scatterplot reflects the most commonly chosen
research strategy for studying the NICs in the 1970s and 1980s. It
shows that repression and growth were both relatively high in all
five countries. Analysts assumed, without checking carefully,
that most of the cases they had not examined would lie in the
lower left quadrant of the figure. From data like thesebut in
verbal form researchers have concluded that labor repression
contributes to growth. The plot shown here actually lends some
plausibility to the argument because, using the quantitative mea
sure of labor repression I created, it is possible to show small
differences in labor repression that are not discernible in the
verbal descriptions. Original statements o f the argument did not
distinguish levels of repressiveness among these cases of rela
tively high repression.
Note that the faulty inference expressed in the literature on
the NICs is the opposite of the one that a thoughtless analyst
using statistical methods would have drawn. A number cruncher
might have concluded, on the basis of these data points, that

roi

H o w the Cases You C h o o se A ffe c t the A n sw e rs You G e l

Singapore

* T iM t n
S o uth K orea

Brazil
M co #

Growth In
G O P ptr capita

2.0

2.5

3.0

Labor R*pm*lon

Fig. 3.3. Labor repression and growth in the most frequently studied
cases, 1970-81. (GDP per capita from Penn World Tables.]

repression did nol cause growth, because the variance in repres

sion explained little of the variance in growth rate within this
high-growth sample: on the other hand, the nonquantitative comparativist would conclude that since all cases are high on both
growth and repression, repression must be a cause o f growth.
But. in fact, no conclusion can be drawn from figure 3.3. It
simply contains too little information.
Scholars working on East A sia, where the fastest-growing
NICs have historically been located, played an important role in
developing the argument linking labor repression to growth. If,
rather than selecting the five industrializing countries most fre
quently described in the literature, we examine the cases most
familiar to East Asia specialists, it appears that repression does
indeed contribute to growth, as shown in figure 3.4.
Based on an image of the world drawn from a few countries in
one part of the world, some analysts advanced general argu
ments about the role of labor repression in growth, implying that
the relationship that seemed apparent in A sia would also charac
terize the entire developing world. Such an inference cannot be
justified, because the selection of cases by virtue of their location
in East Asia biases the sample just as surely as would selection
explicitly based on growth rates. This is so because, on average,
growth rates in East Asia are unusually high. (See table 3.1.)
Geographical area is correlated with growth, and consequently

10 2

Paradigm s and San d C astles

H b c r R*pr*lon
Slope c oe ffic ien t (S) 2 73. R 1* 36

Fig. 3.4. Labor repression and growth in the Asian cases, 1970-81.
(GDP per capita from Penn World Tables; for Thailand, from World Bank
1984.1

the selection of cases by geographical location amounts to selec

tion on the dependent variable.
When one looks at the relationship between average labor
repression and average growth for a larger sample o f countries
that includes slow-growing as well as fast-growing ones, the ap
parent relationship shown in figure 3.4 disappears. A s figure 3.5
shows, the slope is approximately flat, and the R 2 is near zero. In
other words, level of labor repression has no discernible effect on
growth in the larger sample.
It might be objected that several o f the arguments linking
labor repression to growth were never intended to apply to the
entire Third World. Rather, their togic depends on tensions that
develop only after industrialization has progressed to a certain

TABLE 3.1. Average Country Growth Rates by Region

E a st A sia
South A sia
A fric a
Latin A m erica
M iddle E a st and N orth A frica

1960-82

1965-86

(% per capita)

(% per capita)

5.2
1.4

5.1
1.5
0.5

1.0
2.2

4.7

Source: Calculated from data in World Bank (1984. 1988).

1.2

3.6

H ow the Cases You Choose A ffect the Answers You G el

S lop* coefflcw u ( S ) -.07,

R >c 0Q1

10 3

Labor FUpr**ion

Fig. 3.5. Labor repression and growth in the full universe of developing
countries, 1970-81. The countries included, and their labor repression
scores, appear in appendix B. (GDP per capita from Penn World Tables.)

stage. Since the literature is unclear about exactly what level of

industrialization countries would need to achieve before labor
repression would be expected to contribute to growth, I had to
decide on a reasonable cutoff point. I used the level of develop
ment in South Korea at the beginning of the 1970s as the thresh
old, since South Korea was the least developed of the countries
often discussed as successful examples of labor repression and
growth. Figure 3.6 shows the relationship between average labor
repression and average growth in the subset of countries that
were at least as developed as South Korea in 1970. A s figure 3.6
shows, there is no linear relationship between labor repression
and growth, even in this subset of cases. The slope is only slightly
positive, and the R 2 remains near zero.
In this set of cases, the country with the lowest average growth
is Iran, which also scores very high on labor repression. Since
Irans growth rate was depressed toward the end of this period by
the revolution, it could be argued that it should be removed from
the data set, even though its civil war did not last very long. If this
is done, the slope coefficient rises to 0.27, but it remains far from
statistical significance, and the R 2 remains near zero. Thus, even
with Iran removed, the analysis fails to support the claim that
labor repression contributes to growth.
It has been suggested that not labor repression per se but the
repression of a previously well organized and mobilized working

(0 4

Paradigm s and S a n d C astle s

Qrottth In

O DP p*r c*ptt

00
Slop#

* .09

05

1.0
003

2.0

2.5

3.0

3.5

Labor Repr*t*lofi

Fig. 3.6. Labor repression and growth in higher-income developing

countries, 1970-81. The countries included are those from appendix B
w hose GDP per capita in 1970 w as greater than that of South Korea.
(GDP per capita from Penn World Tables.)

class would improve economic performance (O'Donnell 1973;

Collier 1979). To test whether increasing repression increases
growth, I have estimated time-series models of the effect of
yearly labor repression on growth in the following year. In one of
the models, two factors that might also be expected to affect
growth oil exports and level of development at the beginning
of the period are controlled for. In the other model, instead of
trying to identify the various things that might be expected to
affect growth, country fixed effects estimators are used to hold
constant all the various country-specific factors that could affect
growth rates. When country fixed effects estimators are used,
coefficients, can be interpreted as reflecting the effect of changes
in the variable of interest here, labor repressionwithin each
country, rather than cross-country differences.
Table 3.2 shows the results o f these two regressions. In the
model with control variables, the effect o f labor repression on
growth is both minuscule and statistically insignificant. In the
model using fixed effects, the coefficient for labor repression is
positive, but not statistically significant. If the coefficient were
reliable, it W'ould indicate that for each unit o f increase in the
labor repression score, a little under a third o f a percentage point
of extra growth could be expected. T h e low R 2 for the model

How the Cases You Choose Affect the Answers You Get

105

TABLE 3.2. The Effect of Changes in Labor Repression on Growth

D ependent V aria b le: A nnual G row th in G D P P er Capita
O L S with
C ontrol Variables'1
Coefficient
L abor repression (range
Oil exports
D evelopm ent level
R2

0 - 5)

.018
.008
- .000
.000

P>\Z\

.917
.850
.751

O L S with
Fixed Effects"
Coefficient

.288

P > \Z\

.286

.099

'I anel corrected standard errors.

shows that even with the inclusion of seventy-nine country fixed

effects,8 the regression explains almost none of the variance in
growth.9
The point of this exercise is not to demonstrate that the hy
pothesis that labor repression contributes to growth is false. It
may have a small positive effect. It might be that the addition of
appropriate control variables or an elaborate lag structure would
make clear a relationship that does not show in the simple tests
done here. These tests do show, however, that the strong relation
ship that seems to exist when the analyst examines only the most
rapidly growing countries is hard to find when a more representa
tive sample of cases is examined. If analysts interested in the
success of the N ICs had examined a more representative sample,
they would probably have reached different conclusions about
the relationship between labor repression and growth. A s figures
3.5 and 3.6 show, labor is as often repressed in slow-growing
Third World countries as in fast-growing ones.
To sum up, the first example above (fig. 3.3) demonstrates
selection bias in its simplest form: the cases are selected precisely
because they share the trait one wants to explain. In the second
example (fig. 3.4), cases are selected on the basis of a character
istic-geographical region that is correlated with the dependent
variable. In both instances, the hypothesized relationship was a
simple, direct one: a higher level of X (labor repression) seemed
to result in a higher level of Y (growth).
Not all causal arguments are so simple. Researchers some
times posit arguments with complicated structures of prior and
8. F o u r countries had to be excluded because of missing data.
9. O ther m odels were tried using different error specifications and corrections for
autocorrelation, even though the regressions reported in table 3.2 disclosed no auto
correlation. In none did the coefficient for repression reach statistical significance.

io 6

Paradigm s and San d C astles

intervening variables that are more difficult to test rigorously.

The consequences of selection on the dependent variable, how
ever, are the same no matter how complicated the argument.
The next section will consider another frequently encountered
variation on this theme: selection on the dependent variable in a
complicated, contingent historical argument.
Selection on the Dependent Variable in a Complicated
Historical Argument

Theda Skocpols stimulating and thoughtful book States and S o

cial Revolutions ( 1979) combines selection on the dependent vari
able with a complex historical argument. She wants to explain
why revolutions occur, so she picks the three most well known
instances the French, Russian, and Chinese revolutions to
examine in detail. She also examines a few cases in which revolu
tion failed to occur, using them as contrasts at strategic points in
her chain o f argument. The use of cases selected from both ends
of the dependent variable makes this a more sophisticated design
than the studies of the NICs.
Skocpol argues that external military threats cause state offi
cials to initiate reforms opposed by the dominant class. If the
dominant class has an independent economic base and a share of
political power, its opposition will be effective and will cause a
split in the elite. If, in addition, peasants live in solidary commu
nities autonomous from day-to-day landlord supervision, they
will take advantage of the elite split and rebel, which will lead to
revolution. (This argument is schematized in fig. 3.7.) This expla
nation, according to Skocpol, mirrors the historical record in

REVOLUTION

Peasant
Rebellion
Dominant Class Has Independent
Economic Base and Shares Power,
Either through Representative
Institutions or Decentralization
Village
Autonomy,
Solidarity

Fig. 3.7.

Schematization of Skocpol's argument

H ow the Cases You Choose A ffect the A nsw ers You G et

10 7

France and in the parts of China controlled by the Communists

during Japanese occupation. The Russian case differs from the
other two in that the upper class lacked the independent eco
nomic base necessary to impede state-sponsored reforms. Conse
quently, the elite remained unified, and revolution failed to oc
cur after the Crimean War. Nevertheless, defeat in World War I
caused elite disintegration, which opened the way for revolution
in 1917.
A t two points in the chain of argument, Skocpol introduces
contrasting cases to strengthen her contention that structural fea
tures identified as causes in these three cases have general signifi
cance. In an examination of Prussia during the late eighteenth to
early nineteenth century and Japan during the late nineteenth
century, she finds that dominant classes lacked the independent
economic base necessary to obstruct state reforms. Both Prussia
and Japan faced military threats at least as severe as that facing
France, but elites remained unified, and revolution failed to oc
cur. She also looks at Britain during the Civil War and Germ any
in 1848 and finds levels of village autonomy low. In both cases,
elites split, but peasants failed to take advantage of the situation;
as a result, revolutions did not occur. These comparisons are
summarized in figures 3.8 and 3.9. A s the figures show, the cases
she examines appear to provide strong support for the argument.
There is no question that the examination of contrasting
cases makes the argument more persuasive than it would other
wise be, though an assessment of the argument based on a few
cases selected from the other end of the dependent variable
carries less weight than would a test based on more cases se
lected without reference to the dependent variable. Neverthe
less, examination of contrasting cases is a solid step in the right

E lite S p lit
D o m in a n t C lass E co n o m ic ally
In d e p e n d e n t, S hares P ow er

France
C hina, a fte r T a ip in g R ebellion

D o m in a n t Class D e p e n d e n t,
E xcluded fro m P o w er

Russia, W o rld W a r 1

E lite C ohesive

Prussia
J ap a n
C h in a , b e fo re T a ip in g
R e b e llio n
Russia, b e fo re W o rld W a r 1

Fig. 3.8. Given external threat, the effect of dominant-class power on

the likelihood of an elite split

io 8

Parad igm s and San d C astles

Revolution

V illage A utonom y

No Revolution

Russia
France
China, in area controlled by
Communists

1640-60
1848

V illage Dependence

Britain,
G erm any,
China, before Communists

Fig. 3.9. Given an elite split, the effect of village autonomy on the
likelihood of revolution

direction and one of the reasons that Skocpol's study has been
considered so persuasive.
Skocpol makes no effort, however, to test other links in the
chain of argument with comparable care. In particular, she offers
no contrasting cases to strengthen her claim that
developments within the international states system as
such especially defeats in wars or threats of invasion and
struggles over colonial controls have directly contributed
to virtually all outbreaks of revolutionary crises. (2 3)10
This claim, which looms large in the overall thesis, seems espe
cially problematic if we accept her implicit definition of threat
ened," that is, as threatened as late-eightcenth-century France.
Fiance arguably the most powerful country in the world at the
time was certainly less threatened than its neighbors.
Most countries in the world have suffered foreign pressures as
great as those suffered by prerevolutionary France, and yet revo
lutions occur infrequently. This raises the question. A re revolu
tions infrequent because of the absence of appropriate structural
conditions, as Skocpol's argument implies, or because foreign
threats have less causal impact than Skocpol believes? To distin
guish between these two possibilities, one would need to choose
a set of cases in which the structural conditions identified by
Skocpol did in fact exist (in effect, holding the structural condi
10.
Note that "contributed to virtually all" is a probabilistic statem ent, not a
statem ent that foreign threat is necessary but not sufficient to explain revolution.
O ther statem ents o f this argum ent, however, can be interpreted as m eaning that
external threats are necessary but not sufficient causes o f revolution (D ion 1998).

H ow the Cases You Choose A ffect the Answers You Get

109

tions constant). Within this set of countries, one would then need
to assess the relationship between level of threat and revolution
ary outcome. If threat and occurrence of revolution tended to go
together in this set of cases, we would have greater faith in the
correctness of Skocpols argument. If, however, high levels of
threat did not seem to increase the likelihood of revolution within
this set of eases, we would feel more skeptical about it.
To carry out this test, as with the prior one, we first need to
establish the appropriate domain of the argument. The question
of what would constitute an appropriate domain for testing Skocpols argument is controversial. Skocpol herself is extremely
modest about the domain for her argument, stating at one point:
Can [the arguments presented in this book] be applied beyond
the French, Russian, and Chinese cases? In a sense, the answer
is unequivocally 4no'. . . . [T]he causes of revolutions . . . neces
sarily vary according to the historical and international cir
cumstances of the countries involved (288). Skocpol does not
eschew generalizability entirely, however, since she evidently
considers scventeenth-century England, eighteenth- and nine
teenth-century Prussia, and mid-nineteenth-century Germany
and Japan within the domain of her argument. But she does
explicitly limit her argument to agrarian states/ which I take as
including countries in the early stages o f industrialization (since
all the cases included in her study had begun to industrialize) but
excluding fully industrialized countries and preagrarian primitive
societies. She also limits the argument to countries that have
never been colonized; wealthy, historically autonomous and
well-established imperial states" (288); and countries whose
state and class structures had not been recently created" (40).
In the face o f such modesty, the rest of the scholarly community
has two options in assessing the study. One is to accept the selfimposed limitations suggested by the author and try to test the
argument on the set of cases implied by them. The broadest inter
pretation of these limiting criteria suggests that the appropriate
universe thus defined would include, besides some (but not all) of
those actually used by Skocpol, only the larger and wealthier p reWorld War 1 states of Europe: Belgium , the Netherlands, Spain,
Portugal, Sweden, Lithuania before 1795, Poland before parti
tion, Austria, the Austro-Hungarian Em pire, and the Ottoman
Empire. This universe includes a fair number of nonrevolutions,
so it would be quite possible to retest the argument on this set of
cases.

i io

Paradigm s and San d C astles

Limiting the argument to this domain does, however, restrict

its interest, since twentieth-century revolutions, with the excep
tion of the Russian revolution, have occurred in poor countries
that had been al least partly colonized and would thus be out
side the domain of Skocpols argument. Moreover, Skocpols
own selection of cases casts some doubt on the appropriateness
of the domain she describes. Japan was not a well-established
imperial state1 in the mid-nineteenth century. Nor was China in
the twentieth. Germ anys state structure, though not affected
by colonialism, had been recently created in 1848. China, Ja
pan, and arguably Russia were poor. The time period and geo
graphical location identified by Skocpol as those in which the
Chinese peasantry had the autonomy to rebel were precisely the
time period and area of Japanese colonization. In short, many
of Skocpols cases violate her own criteria for limiting the do
main of her argument.
The alternative approach is to derive the domain of the argu
ment directly from the substantive claims of the argument itself.
If we do this, the appropriate domain would seem to include all
independent, not fully industrialized states (and possibly em
pires). These restrictions are necessary because the argument
seems to require (1) the existence of an indigenous state elite and
dominant class; and (2) a peasantry. Skocpol herself is most ada
mant about excluding colonized nations from the domain (28890). This seems a reasonable exclusion during the period of colo
nization (when the state elite and often the dominant class as
well are not indigenous) and perhaps for some limited time a
decade or two after independence.11 The claim that any coun
try that has ever been colonized should be forever excluded does
not seem to flow from anything in the argument itself, however,
and also seems to ignore the role of conquest in the development
of the states included in the original argument. A fter all, E n
gland was once colonized by the Normans, large parts of Russia
by the Mongols and Tatars, and China by the Mongols. In all
three, aspects of subsequent state organization and development
are commonly traced to the effects of these conquests.
11.
If we think o f the dom ain as derived from the argum ent itself rather than from
Sk ocp ols somewhat ad hoc comments about it, then her inclusion o f China during
the time that much o f the country was colonized by Jap an seem s less puzzling.
Throughout the Japan ese occupation, an indigenous C hinese state elite and dominant
class continued to exist in southern C hin a, and it was they whom the Chinese Com m u
nists eventually d efeated, not the Jap an ese.

H o w the C ases You C h o o se A ffe c t the A n sw e rs You G et

Skocpol also argues that small countries should be excluded

because revolutions in them may be caused or prevented by
outside intervention (289). This is a legitimate concern. It should
not lead to the blanket exclusion of all small countries, however,
since we know that outside intervention has failed to prevent
revolution in a number of them. Yet it might be reasonable to
eliminate some cases in which a persuasive argument can be
made about the decisiveness of intervention.
Ideally, a test of Skocpors hypothesis about the effects of
military competition would examine all independent, not fully
industrialized states characterized by the structural features
village autonomy and a dominant class with an independent eco
nomic base and access to political power that she identifies as
necessary to complete the sequence from military threat to revo
lution. Then one could determine whether revolutions occur
more frequently in countries that have faced military threats.
In practice, identifying the universe of cases that meet these
structural criteria is probably im possible. It would require exten
sive know ledge about every country in the world from the E n
glish C ivil War to the present. N onetheless, m oderately serious
tests o f her argum ent are possible, and one is shown below.

A s it happens, several Latin American countries (M exico, G u a

temala, El Salvador, Honduras, Nicaragua, Ecuador, Peru, B o
livia, and Paraguay) have the structural characteristics she identi
fies and so can be used as a set of cases on which to test the
hypothesis linking military threat to revolution. These cases are
obviously not selected at random, but since their geographical
location is not correlated with revolution, geography does not
serve as a proxy for the dependent variable (as occurred in the test
of the relationship between labor repression and growth among
the East Asian N ICs).
In all these countries, dominant classes had an independent
economic base in land and/or mining from the nineteenth cen
tury until well into the twentieth. They also shared political
power. Thus, they had the economic and political resources that
Skocpol identifies as necessary to oppose state-sponsored re
forms successfully and so pave the way for revolution.
These countries also all contained (and most still contain)
large, severely exploited indigenous and mestizo populations,
many of whom lived in autonomous, solidary villages. Spanish
colonial policy reinforced, and in some areas imposed, corporate
village structure. A fter independence, changes in property rights

Paradigm s and San d C astle s

reduced village control over land, but this reduction in the func
tions that had contributed to building village autonomy and soli
darity was at least partially offset by the increase in absentee
landlordism that accompanied increasing commercialization.
Much of the land in these countries was held in large tracts.
Some peasants lived on the haciendas, but many lived in tradi
tional villages, owned tiny parcels of land or had use rights to
communal land, and worked seasonally on the haciendas. These
villages often had long histories of conflict with large landowners
over land ownership, water rights, and grazing rights. Villages
governed themselves in traditional ways. Landlords have rarely
lived in villages in these countries. In short, the rural areas of
these Latin American countries approximate Skocpol's descrip
tion of the autonomous, solidary village structure that makes
possible peasants' participation in revolution. Differences of
opinion are, of course, possible about whether peasants in these
countries were really autonomous enough from day-to-day land
lord control to enable them to play the role Skocpol allots to
peasants in bringing about social revolutions. Perhaps the best
evidence that they were is that revolutions have in fact occurred
in several of these countries, and peasant rebellions have oc
curred in most of them.
With these structural features on which the outcome is contin
gent held constant, it becomes possible to test the relationship
between external threat and revolution. In the test below. I have
used a higher level o f threat than that experienced by France in
the late eighteenth century. I wanted to choose a criterion for
assessing threat that would eliminate arguments about whether a
country was really threatened enough, and I found it hard to
establish an unambiguous criterion that corresponded to the
France threshold." Consequently, the criterion used here is loss
of a war, accompanied by invasion and/or loss of territory to the
opponent. With such a high threat threshold, finding cases of
revolution in the absence of threat will not disconfirm Skocpols
argument, since the countries may have experienced external
pressures sufficient to meet her criteria even though they did not
lose wars. If, however, several countries did lose wars (and the
structural conditions identified as necessary by Skocpol are pres
ent) but have not had revolutions, this test will cast doubt on her
argument.
Figure 3.10 shows eight instances of extreme military threat
that failed to lead to revolution, two revolutions (if the Cuban

H o w the C ases You C h o o se A ffe c l the A n sw e rs You G e l

Revolution

N o Revolution

D efeated and Invaded or Lost

Territory

Bolivia (1 9 3 5 ), revolution 1952

Peru (1 8 3 9 )
Bolivia (1 8 39 )
M exico (1 8 4 8 )
M exico (1 8 6 2 -6 6 )
P araguay (1 8 69 )
Peru (1 8 83 )
Bolivia (1 8 8 3 )
Colom bia (1 9 03 )

N ot D e fe ated within 2 0 Y ears

Mexico, revolution 1 9 1 0 -1 7
Nicaragua, revolution 197 9

All O thers

113

Note: T h e C uban Revolution is not, in Skocpol's term s, a social revolution because it did
not entail m assive uprisings of the low er classes.

Fig. 3 10 Relationship between military defeat and revolution in Latin

America (with Skocpol's structural variables held constant)

revolution of 1959 is not counted, because it does not fit Skocpol*s definition of a social revolution as entailing massive upris
ings of the lower classes) that w'ere not preceded by any unusual
degree of external competition or threat, and one revolution, the
Bolivian, that fits Skocpol's argument. I do not think any foreign
power deserves credit or blame for any of the revolutions that
have occurred, and thus the finding that two revolutions oc
curred without unusual foreign threat is not undermined by for
eign influences on revolutionary outcome. The United States
may deserve credit or blame for the nonoccurrence of revolu
tion in El Salvador and Guatem ala, but if these revolutions had
been successful, they would have increased the number of cases
in which revolutions occurred in the absence of unusual foreign
threat and thus added to the evidence undermining Skocpol's
argument. In short, among these cases there is little support for
the claim that foreign threat increases the likelihood of revolu
tion. If we accept the idea that the domain depends on the argu
ment itself, then these findings suggest that if Skocpol had se
lected a broader range of cases to examine, rather than selecting
on the dependent variable, she would have reached different
conclusions.
This test does not constitute a definitive disconfirmation of
Skocpols argument. Competing interpretations of all the con
cepts used in the argumentvillage autonomy, dominant-class
independence, military pressure exist, and different opera
tionalizations might lead to different results. In particular, my

114

Paradigm s and S an d Castles

operationalization of threat fails to capture the complexity of

Skocpols idea, and a different operationalization might put Nica
ragua and Mexico in the threat-revolution cell. A ny indicator of
threat that identified Nicaragua in 1979 and Mexico in 1910 as
threatened, however, would add hundreds of other countryyears to the threat-no revolution cell, because the amount of
U.S. pressure experienced by these countries at these times was
not at all unusual in the region. In short, despite some deficien
cies in operationalization, this cursory examination of cases not
selected on the dependent variable does cast doubt on the origi
nal argument.
Arguments about Necessary Causes

Some have interpreted Skocpols statements as meaning that she

sees external threat as a necessary but not sufficient cause of
revolution. A s Douglas Dion (1998) and others have noted, the
logic underlying tests of arguments about necessary causes dif
fers from that described above. Methods for testing arguments
about necessary causes have only begun to be developed, but
Dion suggests a Bayesian approach.12 Bayesian analysis provides
a way of assessing the impact of new information on ones prior
beliefs about the likelihood that a particular theory is true. If, in
order to keep things simple, we set aside the possibility of mea
surement error and think of only one rival hypothesis lo the one
being tested, Bayes rule can be expressed as:
P ..* ( W H | D )=

P( D,WH)
P,il,(W H)P(D|W H) + Pllo,(R H )P (D |R H )'

where
Posterior (WH|D) is the probability that the working hypothesis
(the one being tested) is true, in light of the new evidence
collected in a study.
Prior(WH) is the analysts belief about whether the working
hypothesis is true before conducting the study.
P(D|WH) is the probability that the data uncovered in this
study would turn up i f the working hypothesis were true.
Prior(RH ) is the analysts belief about the likelihood that the
rival hypothesis (the most likely alternative to the working
hypothesis) is true prior to conducting the study.
12. See B au m oeller and G oertz (2000) for a careful non-Bayesian approach.

How the Cases You Choose A ffect the Answ ers You G et

115

P(D|RH) is the probability that the data uncovered in this

study would have emerged i/the rival hypothesis were true.
If we interpret the Skocpol argument as one about the neces
sity of external threat, then the appropriate initial research strat
egy is to choose cases that have experienced revolutions and then
check to see if an external threat preceded the revolution. The
information about these external threats is the new data that will
be used to update the assessment of the likelihood that the w ork
ing hypothesis is true. Note, however, that the only way to assess
the likelihood of observing the new data given that the rival
hypothesis is true is to know enough about the whole relevant
universe of cases (not just those that experienced revolution) to
be able to estimate the probability of observing these events (in
this case, external threats) if the rival hypothesis better describes
reality. In other words, we need to know something about the
frequency of the hypothesized preceding event in the universe as
a whole.
In order to use B ayes rule, it is also necessary to state a level
of prior belief that the working hypothesis is true. These prior
beliefs come from prior research on a subject. When little prior
research has been done on a subject, it has become conventional
to treat prior beliefs as neutral between the two competing hy
potheses, that is, to set P(WH) = P(RH ) = 0.5.
To return to the Skocpol example, if we use as threatened as
France as the appropriate threat threshold, then I would esti
mate that 95 percent of all countries that would otherwise fit
within the domain of the theory have experienced such a threat
at some time, many of them repeatedly. With this estimate, the
probability o f observing the data (external threat) in any particu
lar country between 1600 and the present, given the rival hy
pothesis that external threat does not cause revolution, can be
calculated.
Skocpol examined three cases and found external threats in
all three. The probability of seeing these data, i f the working
hypothesis is true (and there is no measurement error) equals
one. If the rival hypothesis is true and 95 percent of the coun
tries in the domain of the argument have experienced similar
levels of threat at some time, then the probability of observing
three instances of threat if the rival hypothesis is true equals
0.95 X 0.95 X 0.95 = 0.857. Plugging these numbers into Bayes
rule, we get:

Paradigm s and Sand C astles

Posterior

(WH)D) = [0.5(1)]

[0.5(1)

0.5(0.857)]

O.539

In other words, when the hypothesized necessary cause is very

common in the world, the increase in ones level of belief in the
argument is increased only very modestly (from 0.5 to 0.539)
when a few new cases are examined. If the hypothesized neces
sary cause only occurred in ten percent of the cases in the appro
priate universe, then examining three cases and finding the data
expected would increase our prior belief in the argument to
above 99 percent, Thus the number of observations needed to
affect posterior beliefs about a hypothesis depends very dramati
cally on the general distribution of the hypothesized necessary
cau se.'3
From the point of view of research design, this discussion of
Bayesian inference leads to two conclusions that have not been
much emphasized in the literature on testing arguments about
necessary causes. First, the Bayesian approach requires that the
data used to assess the likelihood that the theory is true be newly
observed. It must come from cases observed for the purpose of
testing the argument, not from the cases from which the hypothe
sis was induced. The original cases, along with other research
and general knowledge about the world, influence the observers
prior beliefs about whether the argument is true. B ayes rule
provides a way of judging how much more convinced by an
argument we should be after seeing new data, not how much
faith we should put in a plausible but untested argument.
Second, although arguments about necessary conditions can
be tested using only cases selected on the dependent variable,
the use of Bavesian logic to assess how much has been learned
from the test requires gathering enough information about how
often the hypothesized necessary cause occurs in the world more
generally in order to estimate the probability that the data that
were actually observed would have been observed if the rival
hypothesis were true. In the non-Bayesian approach suggested
bv Braumoeller and Goertz (2000), this issue has been ad
dressed under the label trivialness. Braum oeller and Goertz ar
gue that to be non-trivially necessary, the hypothesized neces
sary cause must be shown to vary more than a little in the full
) ?. D ion ( 1 9 9 J*) provides a chart show ing how m any cases w ould be needed to
reach gs percent confidence th at an argum ent is tru e , given different prior levels of
belief and different estim ates of the likelihood o f observing, the d ata if the rival
hypothesis is true.

H ow the Coses You Choose A ffect the Answ ers You Get

i!7

relevant universe. W hatever am ount o f variation is deem ed suffi

cient for non-trivialness, we cannot discover it without exam in
ing at least some cases not selected on the dependent variable.
If the hypothesized necessary cause occurs infrequently, then
examining only a few additional cases would suffice to meet
their condition.
Time Series, Case Studies, and Selection Bias

Case studies, perhaps the most common form of research in the

comparative field, can often be thought of as nonquantitative
time-series research designs. They usually examine a single coun
try over a period of time, often for the purpose of explaining
some outcome at the end or showing the effects of some change
that occurred during the time examined. Case studies are often
criticized as single data points and hence incapable of reveal
ing anything about cause-and-effect relationships, but most can
be more reasonably thought of as a series of observations of the
same case at different times. In fact, most of what are called case
studies actually include unsystematic observations at multiple
levels of analysis (for example, individuals, government adminis
trations, and parties) and observations of multiple entities at the
same level of analysis (for example, several parties in one coun
try) as well as observations over time. For now, however, let us
focus on the simplest kind of case study say, a study of the
evolution of one party over time.
Such case studies are subject to several methodological pit
falls, solutions to which are discussed at greater length in chapter
4. Here I want to note the methodological issues related to selec
tion bias that can arise in the context of case studies and single
case quantitative time series. In the typical study of a single case,
a country, organization, or group is chosen for examination be
cause it has experienced something unusual, sometimes because
it is considered typical of a group of cases that have experienced
the unusual. The variation 011 the dependent variable is supplied
by observations of the same case at other times (when it was not
experiencing the unusual). Whether such a research design in
volves actual selection bias depends on whether the variation
over time within the case reflects the full range of outcomes in
the relevant universe. Often it does not. When the selection of
multiple observations of a single case results in a truncated sam
ple relative to the appropriate universe, the result is inadvertent

118

Paradigm s and San d C astles

selection on the dependent variable, and selection bias can be

expected to have the same results noted above.
The key concern for the researcher, then, is identifying the
universe relevant to the question being asked or hypothesis being
tested. Only when that has been done can he assess whether
outcomes vary widely enough within the single case to avoid
selection bias. H ere, as elsewhere, the question under examina
tion determines the appropriate universe. Sometimes one wants
to understand the effect of a particular policy change in a particu
lar setting. In this situation, one is not asking What caused
outcome y ? but rather "What was the effect of cause A ? If
cause X occurred only in one setting, then a one-country time
series or case study is the appropriate research design (Campbell
and Ross 1968). If X occurred in multiple places, the analyst
would be wise to examine its effects in all of those places or a
sample of them. Otherwise, he risks the possibility of attributing
to X anything that might have happened in the chosen country
during the time following X . When the analyst wants to know
what caused outcome Y, it is always risky to examine only one
entity in which Y occurred, even if he cares only about why it
happened in a particular place and time. A case study of the
particular place and time o f interest may not provide the answer.
The reason it might not is that it is quite possible that selecting
multiple observations of the same case will have the effect of
holding constant or near constant some of the true causes of the
outcome of interest, even if the dependent variable spans a con
siderable range. A t the same time, whatever potential causal
factors do vary within the single case over time will seem to
explain differences in the outcome. These causes o f the withincase variation can be less important causal variables that belong
in a complete explanation, or they can be idiosyncratic factors
that affect this case but not others and therefore do not belong in
a general explanation. The analyst has no way of knowing. Ei
ther way, he will be tricked into focusing on these factors while
giving short shrift to causa! factors that may be changing slowly
and not very noticeably during the time under study but that
nevertheless explain the general trend in the outcome.
This problem is caused by inadvertent selection on one or
more causal variables. Case studies are highly vulnerable to inad
vertent selection from one end of the continuum of potential
causes because so many factors remain constant or change slowly
over time in a single entity. In statistical work, selection on the

H ow the Cases You Choose A ffect the Answ ers You Get

\ 19

dependent variable leads to biased estim ates o f the effects o f

causal factors, but the practical result is usually a failu re to dem
onstrate a causal relationship that actually exists, because within
the truncated sam ple there is little variation in the outcom e for
differences in the causal factors to explain. Selection on the inde
pendent variable, as often happens in case studies, does not lead
to biased estim ates in statistical w o rk . N everth eless, in practice it
is hard to show that a relationship betw een a cause and effect
exists if the cause varies little within the sam ple. W hether o b se r
vations are quantitative does not affect the logic o f the research
design. In either case, it is quite possible to overlook factors o f
real causal im portance, because they do not vary much o ver time
or follow an increm ental trend that country observers take for
granted.
If one knows quite a bit about the underlying causal m odel, a
single-case tim e-series design can be a good w ay to assess the
effect o f one potential cause while holding many other things
constant (because they do not vary within the single case), but it
will be less useful in the more typical situation where the analyst
does not know the underlying m odel. T he analyst will then fail to
identify any causal factor that varies little within the case and will
tend to overem phasize serendipitous contributors to the outcom e.
A s an exam ple, let us contem plate A lb ert H irschm ans careful
and insightful study o f inflation in C hile (19 73)- In this essay,
Hirschman review s C h iles m ajor bouts of inflation betw een the
nineteenth century and 1961. H e reconsiders the role o f foreign
experts in Chilean policy form ulation and shows the im portance
of dogmatic econom ic ideologies and policy m istakes in causing
inflationary episodes. Hirschm an argues in this study, and in the
book o f which it is a part, that policym akers gradually learn to
resolve persistent problem s, that the search for solutions has
positive externalities in that it brings hitherto unnoticed issues to
policymakers attention, and that reform ism , though messy and
often em otionally unsatisfying, leads o ver time to significant im
provements. Hirschm an describes first the intermittent difficul
ties with inflation that Chile experienced between 1870 and 1939,
and then the persistent and worsening inflation o f 1940 to 1959.
At each point in the story, Hirschm an, with his custom ary flair
and sensitivity to detail and context, discusses the policy m is
takes and other factors that increased inflation. D uring the early
period, inflationary episodes were caused by wars and civil w ars,
serious policy m istakes, and business expansions, all o f which

120

Paradigm s and San d C astles

1973-76

Y*r
Fig. 3.11. Inflation in Chile, 1930-96. (Data for 1930-61 from Hirschman 1973; 1962-63, Corbo Loi 1974; 1964-96, IMF 1997.)

seem tohave been largely self-correcting. Beginning in 1940, how

ever. inflation became more persistent and more serious. It no
longer returned to normal between episodes, and the trend line,
though masked by zigzagging, began a determined upward slope
(see fig. 3 .1 1 ) . Between 1940 and 1959, inflation averaged 28
percent per year (Hirschman 1973, 160). Chilean inflation rates
are shown in table 3.3.
Hirschman attributes this worsening mostly to the effects of
specific policies, especially the failure to restrict credit to the pri
vate sector and the routinization of wage adjustments. The empha
sis throughout the essay is on the details of policy and the political
context that influenced them. In explaining the control of infla
tion, which appeared to have been achieved in Chile in 1960-61,
Hirschman stresses the intense political struggle over ending auto
matic wage adjustments accomplished during an inflationary
peak in 1956 (203-5) and strengthening the system of credit
control under President Jorge Alessandri in 1959 (219).
Although Hirschman mentions general economic factors such
as fiscal deficits and exchange rates, the reader is left with the
impression that Chilean inflation was caused by some fairly dis
crete policy mistakes. This impression is strengthened by the low
inflation rates of i960 and 1961, caused by specific policy changcs
introduced by the Alessandri administration. The reader never
sees the bigger and more general picture: that the policy strategy

blow the Cases You Choose Affect the Answers You Get

12 1

of state-sponsored import-substitution industrialization (ISI),

which was initiated in Chile in 1939 and varied little during the
next thirty-five years, caused increasingly serious inflation.
Although the general policy strategy remained stable and was
not subjected to the kind of intense political debate that accompa
nied the policy changes emphasized by Hirschman, its implemen
tation over time entailed increasingly distortionary tariff and ex
change rate policies. These policies led to the same problems with
balance of payments crises and inflation that afflicted so many
other developing countries. Inflation plagued all developing

TABLE 3.3. Chilean Inflation, 1930-96

Y ear

Inflation R ate (% )

Y ear

Inflation R ate (% )

1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963

5
4
26
5
9

1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
l ()82
1983
1984
1985
1986
1987
1988
19S9
1990
1<)91
1992
1993
1994
1995
1996

46
29
23
19
26
30
32

-1

12
10
2

10

23
26
8

15
8

30
23
17
21

17
23
12

56
71
84
38
17
33
33
5
in

14
44

Source: 1930-61, Hirschman (1973); 1962-63.

(2002).

20

75
361
505
375
2 12

92
40
33
35
20
10

27
20

31
19
20

15
17
26
22

15
13
11
8

Corbo Loi (1974): 1964-96. World Bank

12 2

Paradigm s and San d C astles

countries that followed state-led import-substituting development

strategies. The average inflation rate for low- and middle-income
countries between 1970 and 1978, a time when virtually all but the
most backward were following state-led ISI strategies, was 18
percent per year, compared to about 9 percent per year for indus
trialized countries (World Bank 1980, 1 1 0 - 1 1 ) . Hirschman is, of
course, entirely correct in noting that discrete policy mistakes
worsen inflation; they account for some of the zigzags so apparent
in figure 3 .1 1 . Nevertheless, the reader interested in under
standing why Chile suffered from recurrent and worsening bouts
of inflation for nearly four decades after 1939 will not have recog
nized the main underlying cause after reading this essay.14
In fairness, let me note that Hirschman did not aim to explain
inflation in this essay. Rather, he sought to show that inflation,
like other seemingly intractable problems, could be gradually
conquered as policymakers learned to understand it and took
advantage of occasionally propitious political circumstances to
initiate reforms. When the primary underlying cause of some
thing has not been identified, however, identifying and fixing
less important and less systematic causes may not result in long
term improvement. The last measure of inflation in Hirschmans
study is for 1961, when it appeared that policymakers had at last
brought inflation under control. The apparent cure, however,
turned out to be a very brief remission.
A s is apparent in table 3.3 and figure 3 .1 1 , Chilean inflation did
not begin its long-term downward trend until the abandonment of
state-sponsored import-substitution development policies during
the Pinochet administration. The extremely high inflation rates of
the Allende years and their immediate aftermath cannot be
blamed on development strategy, but if those years are excluded,
it is still clear that the conquest of inflation in Chile began in the
late 1970s. Current Chilean economic policy-making and low
inflation demonstrates that Hirschman was correct in believing
that human beings, including policymakers, learn. But in Chile,
inflation was not finally conquered by reformist muddling through
and discrete policy changes, as Hirschman had hoped. It was con
quered by traumatic policy changes that reversed four decades of
basic economic policy strategy.
14.
In another essay, written a few years before the one on Chile and drawing on the
experiences o f several Latin A m erican countries rather than only one, Hirschman (1968)
was one of the first to identify a number o f the systematic ill effects o f the importsubstitution strategy of industrialization, including its tendency to cause inflation.

How the Cases You Choose A ffect the Answ ers You Get

123

The methodological point is that even if one cares only about

what caused inflation in Chile, the best research strategy for
discovering these causes may require examining other cases. Im
portant causes, such as the basic thrust of development strategy,
may not change very much within a few decades in a single
country. Consequently, analysts may overlook their importance
and instead concentrate attention on less important causes or on
conjunctural factors that turn out to have no general causal e f
fect. Case studies generally help to explain zigzags in the trend
line, but they sometimes offer little leverage for explaining the
trend itself.
Case Studies, Time Series, and Regression to the Mean

The remainder of this chapter focuses on some less obvious pitfalls

that face the researcher who must choose not only which cases to
examine but also the beginning and end points of the study. If
either the starting or ending dates of a case study or time series are
chosen because of their extreme scores, the analyst must be con
cerned about the effects of regression to the mean, in addition to
the other possibilities for mistaken inference associated with
selection bias. Because extreme outcomes typically result from a
combination of extremes in their systematic causes and extreme
unsystematic influences (what would be called the error term in
quantitative work), terrible conditions at the initiation of a study
are likely to improve with the passage of time, and wonderful
situations are likely to deteriorate even if there has been no
change at all in the systematic factors causing them. Such changes
in the unsystematic influences on outcomes lead unwary analysts
to attribute improvement or deterioration to their favorite hero or
villain among intervening events, even though the only real
change that has occurred is in the random factors that influence
everything in social science and the rest of the world.
Regression to the mean is the name given to the tendency of
any extreme situation, score, outcome, or event to be followed
by one that is less extreme simply because fewer extreme random
factors happened to influence things the second time. Regression
to the mean causes the mismeasure of systematic change in out
comes over time. A fter having misunderstood the amount of
change that has actually occurred, the analyst often then com
pounds the mistake by building an argument to explain the
changes that never occurred.

124

Parad igm s and Sand C astles

Regression to the mean has been most fully analyzed in the

context of educational research. The classic example involves
researchers trying to assess the usefulness of a new technique for
teaching remedial reading. Students reading ability is tested.
Those who score below some threshold are selected to receive
special help, using the new technique. After some time has
elapsed, they are tested again, and the rest of the class is also
retested as a control, since all students are expected to be increas
ing their skill over time. The students who received special help
always make greater gains than the group that did not receive
help, no matter what technique is tried. Illiteracy has not disap
peared, however, because at least some and perhaps all of the
gain demonstrated by these students is an artifact caused by
regression to the mean, not a genuine effect of the remedial
reading techniques. The students who scored lowest on the first
test did so for two reasons: they read less well than others; and,
for unsystematic reasons such as being sick or tired, they did
especially badly on the first test. The second test, like the first,
measures both the systematic component of reading ability and
also random factors such as sickness and tiredness. Since it is
unlikely that the same children would be sick or tired during both
tests, on average the scores of the remedial group would not
include extreme unsystematic elements the second time, and
thus they would score higher even if their reading ability had not
improved.
Regression to the mean has two sources, one conceptually triv
ial but practically very important in the social sciences, and the
other both conceptually and practically important. The first is
that every measurement contains an element of error. For simple
physical measurements of things, such as temperature and length,
the small element of error in the measure is usually of no practical
consequence, but in social science most of the things we want to
explain can be measured" only very inadequately. One of the
unsystematic contributors to every outcome is thus measurement
error simply the inaccuracy of all measures, whether quantita
tive or not. Every outcome is also affected by happenstance, by
events that will never occur again and have no theoretical impor
tance, by luck, by the particular skills, failings, and longevity of
certain individuals, and so on. These unsystematic contributors to
outcomes are also, if the research goal is systematic explanation,
part of the error term . That is, both real measurement error
and serendipitous factors contribute to every assessment of an

H ow the Coses You Choose Affect the Answers You G et

12 5

outcome. They are the causes of the statistical artifact, regression

to the mean.
A n y time that cases are selected for study or treatment" on
the basis of high (or low) scores on some variable, the analyst
unintentionally selects a sample with unusually positive (or nega
tive) error terms, in both senses discussed above. When the
selected cases are measured a second time, the inaccuracy in
their measurement is no more likely to be positive than negative,
and the serendipitous events that affect the outcome then are
also no more likely to be positive than negative. In consequence,
cases with especially positive outcomes in a first measurement
will look as though they are doing less well in a second one, and
cases with especially negative outcomes in a first assessment will
seem to improve over time even if nothing systematic has
changed.
Because most of the work on regression to the mean has oc
curred in the context of educational and psychological testing,
students of comparative politics are sometimes unaware of its
implications for their own work. Consider the following hypo
thetical study. The analyst wants to know what effect structural
adjustment loans from international financial institutions have
on the economic performance of developing countries. To an
swer this question, she compares the growth rate per capita in
the countries that have received such loans with the growth rates
of developing countries that have not. She needs to compare the
countries of interest with another set of countries during the
same time period as a means of controlling for international
factors that affect growth all over the world. Since, however,
structural adjustment loans went to countries with economies in
crisis, and since these crises are caused by bad luck, bad weather,
and unrepeatable events as well as by more systematic factors,
their economic performance can be expected to improve, on
average, whether or not the loans help. Thus, the unwary re
searcher who simply compares changes in the performance of
countries receiving loans with changes in those that did not may
be misled about the effect of the loans, since a certain amount of
improvement in the countries with the worst performance could
have been expected in any case.
Whenever research focuses on comparison of growth rates
over time, the analyst needs to be attentive to the possibility of
regression to the mean. The underlying causes of economic per
formance, such as resource endowment, human capital, savings

12 6

Paradigm s and Sand C astles

10

0
Tim e 2
Growth

-10
/

/
20

-20

/
-10

Tim e 1
Growth

10

Fig. 3.12. Regression of growth in GDP per capita for 1991 on growth
in GDP per capita for 1990 for developing countries. (GDP per capita
from Penn World Tables.)

rate, competence o f policymakers, and basic thrust o f economic

policy, do not change much from year to year in most countries.
Consequently, we would expect a high correlation between rates
of growth from one year to the next, and our casual observation
that most Asian countries grow rapidly almost every year and
most African countries do not supports that expectation. At the
same time, unsystematic factors and measurement error contrib
ute to the observed growth rate in every country every year. This
unsystematic component of measured growth always leads to the
appearance of faster growth in countries with the worst perfor
mance in an earlier time period and slower growth in countries
with the best performance.
Figure 3 .12 shows the relationship between growth in 1990
and growth in 1991 for developing countries.5 The solid line is
the regression line, which shows the estimated growth rate in
1991 given any particular growth rate in 1990. The dashed diago
nal line is the hypothetical relationship we would expect to exist
15.
T he data set used to construct this scatterplot includes all low- and middleincome countries with more than a million inhabitants for which data w ere available
from the Penn World Tables. Countries with few er than a million inhabitants were
excluded because their econom ies tend to be unusually volatile, and I did not want
the results shown here to depend on unusual cases. T he countries for which data are
not available include most o f those engaged in civil war during these years.

How the Cases You Choose A ffect the Answ ers You Get

1 27

if all the causes of economic performance remained stable from

year to year and therefore growth, on average, remained the
same from year to year. The part of the regression line that
reflects the performance of the countries with the highest growth
rates in 1990 lies below the diagonal, showing that they tended to
grow less rapidly in 1991 than they had in 1990. Meanwhile, the
part o f the regression line for countries with the lowest growth
rates in 1990 lies above the diagonal, indicating that they grew
more rapidly in 1991. Countries growing at above 5 percent per
capita in 1990 grew, on average, only 3.7 percent in 1991. A t the
other extreme, countries with growth declining at 5 percent or
more in 1990 were declining at only 2.8 percent per capita, on
average, in 1991.
These tendencies were not caused by some vicissitude in the
international economy that for once advantaged the poor and
disadvantaged the rich. (The reader who suspects that this might
be the case is urged to try this regression on other years. In every
single pair, the fast-growing countries will do a little less well in
the second year, and the slow-growing ones a little better. This
result does not mean that growth rates are gradually evening out
among countries.)
These tendencies are not caused by anything systematic, but
rather by changes in the error term . The countries with the
highest scores at any time are those with not only good system
atic economic performance but also, on average, those with posi
tive error terms either real measurement errors or serendipi
tous events and luck that cannot be expected again the following
year. In the subsequent measurement, economic performance is,
on average, still good, but the unsystematic component of the
outcome is, on average, neither positive nor negative, and thus
the overall score is lower. A s a consequence, any time one selects
cases for study because they are doing especially well, one can
expect that their subsequent performance will decline a bit, and
the inverse is true for cases selected because they are doing
especially badly.
Regression to the mean is especially likely to interfere with
reaching correct conclusions when one is trying to assess the
effect o f some treatment, such as structural adjustment loans
or aid programs aimed at meeting basic needs. This is because
treatments are often provided only for those who donors think
need them, usually those experiencing some sort of crisis. In such
situations, the problem is not that the analyst selects cases from

128

P aradigm s and Sand C astles

one end of the continuum, but that the agencies supplying the
treatment" do.
Regression to the mean can also affect one's ability to assess the
effect of spontaneous treatments such as military interventions.
If democratic breakdown usually occurs during economic crises,
then a research design that compares economic performance be
fore and after military interventions is likely to overestimate the
beneficial effects o f military rule on the economy, for exactly the
same reasons that educational researchers might be tempted to
overestimate the beneficial effects of a remedial reading tech
nique. On average, the poor economic performance of the prebreakdown period was caused by both systematic and unfortunate
serendipitous factors, but the serendipitous factors that affect per
formance during the later period under military rule will, on aver
age, be average. If, for example, one compares the growth rate in
Argentina, Brazil. Chile, and Uruguay during the year prior to the
most recent breakdown of democracy with the average during the
first five years of military rule as a number of authors attempting
to assess the effects of bureaucratic authoritarianism did, though
usually not quantitatively one is tempted to conclude that mili
tary rulers handle the economy belter than do elected politicians.
On average, per capita income declined by 1.5 percent during the
year prior to breakdown in these countries, but it grew 0.8 percent
per year, on average, during the first five years under military rule
(not including the breakdown year itself).16
One cannot, however, conclude from these figures that mili
tary regimes perform better. To assess that question, one would
need to model the regression to the mean that would be expccted
in the relevant years and then compare economic performance
under the military with that predicted by the model. Alterna
tively, one might compare growth during military rule with long
term growth in the same countries, since the ups and downs
in the error term would be evened out by averaging over many
years. Average growth in these four countries from T95t to the
year before military intervention ranged from 0.9 percent for
Uruguay to 3.2 percent for Brazil, all higher than the average
during the first five years of military rule.17 A more careful test
could certainly be done, but this simple one is sufficient to sug
16 . T hese p ercen tages, as well as those in the follow ing p arag rap h , w ere calcu
lated from the Penn W orld Tables.
17 . Years included for A rg entin a arc 19 5 1 to 1965 . because the m ilitary ruled for
m ost o f the tim e after that.

H ow the Cases You Choose Affect the Answers You G et

12 9

gest that these military governments were not especially success

ful at delivering rapid growth.
C o n clu sio n

The examples in this chapter have shown that choosing cases for
study from among those that cluster at one end of the outcome to
be explained can lead to the wrong answers. Apparent causes that
all the selected cases have in common may turn out to occur just as
frequently among cases in which the effect they arc supposed to
have causcd has not occurred. Relationships that seem to exist
between causes and effects in a small sample selected on the depen
dent variable may disappear or be reversed when cases that span
the full range of the dependent variable are examined. Arguments
that seem plausible if a historical study or time series begins or
ends at a particular date may seem less persuasive if the dates of
the study are changed. Regression to the mean can lead the un
wary researcher into explaining changes that did not occur. In
short, selecting cases without giving careful thought to the logical
implications of the selection entails a serious risk of reaching false
conclusions.
This is not to say that studies of cases selected on the depen
dent variable have no place at all in comparative politics. They
are useful for digging into the details of how phenomena come
about and for developing insights. They identify plausible causal
variables. They bring to light anomalies that current theories
cannot accommodatc. In so doing, they contribute to lhe cre
ation and revision o f proposed theories. By themselves, how
ever, the} cannot test the theories they propose (cf. Achen and
Snidal 1989). To test theories, one must select cases in a way that
docs not undermine the logic of inference.
If wr want to begin accumulating a body of theoretical knowl
edge in comparative politics, we need to change the conventions
governing the kinds of evidence we regard as theoretically rele
vant. Conjectures based on eases selected on the dependent vari
able have a long and distinguished history in the subfield, and
they will continue to be important as generators of insights and
hypotheses. Regardless of how plausible such conjectures are,
however, they retain probationary status as accumulated knowl
edge until they have been tested, and testing them usually re
quires the thoughtful selection of cases from the full range of
possible outcomes.