How the Cases You Choose Affect the
Answers You Get: Selection Bias in
Comparative Politics
Barbara Geddes
Abstract
This article demonstrates how the selection of cases for study on the basis of
outcomes on the dependent variable biases conclusions. It first lays out the
logic of explanation and shows how it is violated when only cases that have
achieved the outcome of interest are studied. It then examines three wellknown
and highly regarded studies in the field of comparative politics, comparing
the conclusions reached in the original work with a test of the arguments
on cases selected without regard for their position on the dependent
variable. In each instance, conclusions based on the uncorrelated sample
differ from the original conclusions.
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. One of our most durable
conventions is the selection of cases for study on the dependent variable.1
That is, if we want to understand something, for example, revolution, we
select one or more occurrences and subject them to scrutiny.
Most graduate students learn in the statistics courses forced upon them
that selection on the dependent variable is forbidden, but few remember why,
or what the implications of violating this taboo are for their own work. And
so, comparativists often ignore or forget about it when carrying out or assessing
nonquantitative comparative research.
I am grateful to Chris Achen, Jim Caporaso, David Collier, Jeff Frieden, Miriam Golden,
Jonathan Hartlyn, Robert Jackman, and John Zaller for comments on earlier versions of this
article.
1. The subfield of comparative politics is not the only one bedeviled by problems with
selection bias (compare Achen and Snidal 1989).
131
132 Political Analysis
This article will show the consequences of violating the taboo. It will do
so by comparing the conclusions reached in several influential studies that
selected cases on the dependent variable with conclusions reached in tests of
the same arguments using samples not correlated with the outcome on the
dependent variable.
All the studies discussed in this paper are intelligent, plausible, insightful,
and possibly true. 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 conventions.
The Nature of the Problem
The problem with selecting cases for study on the dependent variable stems
from the logic of explanation. When one sets out to explain why countries A
and B have, say, developed more rapidly than countries C through G, one is
implicitly looking for some antecedent factors X through Z that countries A
and B possess, but that countries C through G do not. The crux of the
difficulty that arises when cases are selected on the dependent variable is that
if one studies only countries A andfi, one can collect only half the information
needed, namely what A and B have in common. Unless one also studies
countries C through G (or a sample of them) to make sure they lack factors X
through Z, one cannot know whether or not the factors identified are crucial
antecedents of the outcome under investigation. Countries A and B may be the
only countries that have X through Z, in which case the hypothesis seems
plausible. But many other countries may also have them, in which case the
hypothesis would seem dubious.
This point can also be made graphically. Suppose a universe of developing
countries A through G, where A and B are among the fastest growing. 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 G 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 1.
Yet, if one examines only countries A and B, it is possible that the full
range of cases would look more like figure 2. That is, it is possible that there
is no relationship 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 dotted line in figures 1 and 2,
two kinds of mistaken inferences can easily occur. The first, as shown in the
figures, involves jumping to the conclusion that any characteristic that the
Selection Bias in Comparative Politics 133
High
Dependent
Variable
(e.g.,
Growth)
Low
Low
Independent Variable
(Factor X)
High
Fig. 1. Assumed relationship between factor X and the dependent
variable
selected cases share is a cause. The other involves assuming that a relationship
(or the absence of a relationship) 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 second kind of
faulty inference (Achen 1986; King 1989). If the true relationship between
factor X and the dependent variable is that shown in figure 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
indicate that no relationship exists. Thus, selection on the dependent variable
often biases statistical results toward finding no relationship even when a
relationship, in fact, exists.
In nonquantitative work, however, the first kind of faulty inference is at
least as common as the second. This occurs when figure 2 shows the true
relationship, but the analyst—based on bits and pieces of information—
assumes that cases C through G are located in the lower left quadrant, and
concludes that factor X causes the outcome of interest even though, in fact, no
relationship exists. An example may make these points clearer.
High
Dependent
Variable
(e.g..
Growth)
Low
•A-
•B
• CD
•
Low High
Independent Variable
(Factor X)
Fig. 2. An alternative possibility for the relationship between factor X and
the dependent variable
134 Political Analysis
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, most often Taiwan, South Korea, Singapore,
Brazil, and Mexico. In all these countries, during the periods of most
rapid growth, governments exerted extensive controls over labor and prevented
most expressions of worker discontent. Having noted this similarity,
analysts argue that the repression, cooptation, discipline, or quiescence of
labor contributes to high growth (compare Haggard 1986, 354-56).
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 economic competition." Different
reasons for this advantage have been advanced, some more plausible than
others. Guillermo O'Donnell (1973) argues that the transition from the easy
stage of import substitution industrialization to a more capital-intensive stage
creates a need for reduced consumption and, hence, a demand for the repression
of labor. Both Fernando Henrique Cardoso (1973) and Hagen Koo (1987)
assert that labor control is necessary in order to attract foreign investment.
Frederic Deyo (1984 and 1987) maintains that an export-led growth strategy
requires cheap skilled labor and, consequently, a disciplined and quiescent
labor force.
Whatever the details of the argument, many scholars who have studied
the NICs seem to agree that repression or cooptation of the labor force contributes
to growth. Taiwan, South Korea (especially after 1961), Singapore
(after 1968), Brazil (1964-81), and Mexico (before 1982) all had repressed
and/or coopted labor forces and very high growth rates. In other words, all
have the outcome of interest and all exhibit another common trait—labor
repression—so analysts conclude that labor repression has caused the
outcome.
But the conclusion does not follow. Perhaps there are other countries that
suppress labor to the same or a greater degree and that have failed to prosper
economically. In order to establish the plausibility of the claim that labor
repression contributes to development, it is necessary to select a sample of
cases without reference to their position on the dependent variable, rate each
on its level of labor repression, and show that, on average, countries with
higher levels of repression grow faster.
The two tasks crucial to testing any hypothesis are to identify the universe
of cases to which the hypothesis should apply, and to find or develop
measures of the variables. A sample of cases to examine then needs to be
Selection Bias in Comparative Politics 135
selected from the universe in such a way as to insure that the criteria for
selecting cases are uncorrelated with the placement of cases on the dependent
variable.
If the whole universe is too large to study, examination of a random
sample is usually recommended as a means of insuring that the criteria of
selection do not correlate with the dependent variable. One can, however,
make valid inferences from any sample selected by some rule not correlated
with the dependent variable. Moreover, randomization does not guarantee the
absence of correlation. If, at a particular time, the universe itself only contains
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 nations that failed to adopt a given military innovation in the sixteenth
century were 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.
For the hypothesis that labor repression contributes to growth, the universe
is easy to identify: all developing countries. In the test to follow, I have
included Taiwan and all developing countries for which the World Bank
collects data except high income oil exporters, those with Communist governments,
those embroiled in civil war for more than a third of the period
covered, and those that are extremely small (less than a million inhabitants).
Communists countries are excluded because the theory only applies to countries
with capitalist or mixed economies. The other exclusions involve countries
with characteristics not related to labor repression which could be expected
to affect greatly their growth rates.
The dependent variable, growth rate, also presents no problems. Various
measures are readily available. For the test, I used World Bank (1984, 218—
19) calculations of GNP per capita between 1960 and 1982, since most of the
studies of development strategies focus on the period before the debt crisis.
Labor repression/cooptation/quiescence is more difficult to measure.
Standard indicators are not available, and labor repression can take different
forms in different contexts, for example, state cooptation in one country 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 the congressional
committees on Foreign Relations and Foreign Affairs (U.S. Department of
State 1981 and 1989).
Countries received a score of one if unions are free to organize and
choose their own leaders; labor organizations are not controlled by the gov-
136 Political Analysis
ernment or dominant party; strikes are legal, not constrained to any significant
degree by government regulations, and occur reasonably often; and labor can
participate in politics. India, Jamaica, and Venezuela are examples.
Countries received a score of two if unions are free to organize and
choose their own leaders, labor organizations are not controlled by the government
or dominant party, and strikes are legal but constrained by government
regulations or simply occur infrequently in practice; or if unions and
strikes are legal, etc., but violence against workers curtails the exercise of
workers' rights, though without eliminating strikes and demonstrations. Colombia,
Zimbabwe, and Malaysia are examples.
Countries received a score of three if union organizations are constrained
by links to the government or dominant party; strikes are legal in some cases,
but subject to considerable government regulation; and government or private
violence against workers is no more than moderate. South Korea, Brazil, and
Pakistan are examples.
Countries received a score of four if unions are illegal or completely
controlled by the government or dominant party; the right to strike is severely
constrained or strikes simply never occur; or if violence against workers is
very severe. Chile (1973-79), Syria, and Benin are examples. Where radical
changes in levels of repression had occurred between 1960 and 1982,1 scored
countries in accordance with the more recent regime unless it had been in
power less than five years in 1982.
Although this is an imperfect measure of a complex set of phenomena,
and experts might disagree about the placement of cases between adjacent
categories, this measure is at least as precise as the verbal descriptions 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 using these
data are shown in figures 3, 4, and 5. Figure 3 shows the relationship for the
sample of NICs most frequently studied. This scatter plot reflects the most
commonly chosen research strategy for studying the NICs. It shows that
repression is moderately high in all five countries. Analysts assume, without
checking rigorously, that most of the cases they have not examined lie in the
lower left quadrant of the figure. From data like these—but in nonquantitative
form—researchers have concluded that labor repression contributes to economic
growth. Yet, from inspection of these data, that inference is obviously
unfounded. There is no relationship at all between labor repression and
growth among the cases shown in figure 3.
Note that the faulty inference here is the opposite of the one a thoughtless
analyst using statistical methods would have drawn. Where a number
cruncher would have jumped to the conclusion that repression made no contribution
to growth because the variance in repression explained none of the
Selection Bias in Comparative Politics 137
Growth in
GNP per capita
(1960-82)
6-
2o
Singapore
0 South Korea
Taiwan
o Brazil
o Mexico
1 2 3 4
Level of Labor Repression
Fig. 3. Growth and labor repression in selected NICs. (Repression scores
based on U.S. Department of State 1981 and 1989. Growth rates taken
from World Bank 1984.)
variance in growth rate within this high-growth sample, the comparativist
jumps to the conclusion that since all cases are high on both growth and
repression, repression must be a cause of growth. But, in fact, no conclusion
can be drawn from figure 3. It simply contains too little information.
Scholars working on East Asia, where the fastest growing NICs are
located, have played an important role in developing the argument linking
labor repression to growth. If, rather than selecting the five industrializing
8 -I
6 Growth
in
GNP per capita
(1960-82) 4
2 Singapore
>^
o S
South Korea - y^
Taiwan /r
Thailand
8/
Malaysia
/ Philippines
Indonesia
0
R'. .56
2 3
Level of Labor Repression
Fig. 4. Growth and labor repression in East Asia. (Repression scores
based on U.S. Department of State 1981 and 1989. Growth rates taken
from World Bank 1984.)
138 Political Analysis
Growth in
GNP per capita
(1960-82)
Argentina
o Chile
1 2 3 4
Level of Labor Repression
Fig. 5. Growth and labor repression in the Third World. (Repression
scores based on U.S. Department of State 1981 and 1989. Growth rates
taken from World Bank 1984.)
countries most frequently described in the literature, one examines the cases
most familiar to East Asia specialists, it appears that repression does contribute
to growth, as shown in figure 4.2
Based on findings of this sort, analysts make general arguments about the
role of labor repression in growth which imply that the relationship will also
characterize the entire Third 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 selection explicitly based on growth rates. This is
so because, on average, growth rates in East Asia are unusually high. (See
table 1.) Geographical area is correlated with growth, and consequently the
selection of cases by geographical location amounts, in effect, to selection on
the dependent variable.
Up to this point, I have reminded the reader that one should not make
inferences based on samples selected on the dependent variable. This is not to
2. A significance test for a difference of means between the two repression categories
shows that the difference is significant at .02. The regression line is shown in fig. 4 only to make
the relationship more immediately intelligible.
Selection Bias in Comparative Politics 139
TABLE 1. Average Country
Growth Rates
East Asia
South Asia
Africa
Latin America
Middle East and
North Africa
1960-82
5.2
1.4
1.0
2.2
4.7
1965-86
5.1
1.5
0.5
1.2
3.6
Source: Calculated from data in World Bank 1984
and 1988.
say that a relationship found within the sample will never characterize the
population, but only that one cannot assume that it does. In the example used
here, as it happens, it does not.
This point is made apparent in figure 5. When one looks at the relationship
between labor repression and growth for a larger sample of cases that
includes slow-growing countries as well as fast, the apparent relationship
between labor repression and growth disappears. The slope coefficient is
slightly negative, and the R2
is .07. In other words, level of labor repression
has no effect at all on growth, as shown in figure 5.
It might be objected that several of the arguments linking labor repression
to growth were never intended to apply to the entire Third World. Rather,
their logic depends on tensions that only develop after industrialization has
progressed to a certain stage. Figure 6 shows the relationship between labor
repression and growth in the subset of countries at least as advanced as South
Korea. Since some confusion exists in the literature about exactly how advanced
countries are when the hypothesized difficulty with labor participation
begins, I have, in order to be conservative, used the country from among the
most discussed cases that was least advanced in the 1970s as the cutoff point.
As •figure 6 shows, there is no linear relationship between labor repression
and growth, even in this subset of cases. Two possible interpretations of
the scatterplot are possible. The first is that the relationship between repression
and growth within the more advanced Third World countries is curvilinear.
That is, moderate amounts of repression or cooptation contribute to
growth, but extreme repression and violence against workers do not. A second
possible interpretation is that there is something unusual about Singapore,
Taiwan, and South Korea that causes very high growth rates, and that their
presence in any small sample will give the appearance of a curvilinear relationship,
whether one exists or not. Some support for the latter interpretation
can be found in the fact that if the cutoff point for inclusion in the sample of
140 Political
Growth in
GNP per capita
(1960-82)
Analysis
8 -
6-
4-
2-
0 -
Panama
8
o
o
Jamaica
1
Tunisia
0
0
8
o
_o
Peru
Singapore
o
g Korea
Taiwan
o
0
o
Algeria
2 3
Level of Labor Repression
Syna
o
O j
o
o Chile
o Argentina
4
Fig. 6. Growth and Labor repression in the more advanced nations of the
Third World. (Repression scores based on U.S. Department of State 1981
and 1989. Growth rates taken from World Bank 1984.)
more advanced countries is lowered to include a few additional cases, the
curvilinearity declines markedly.
Whichever interpretation is correct, the point here is not to demonstrate
that the hypothesis that labor repression contributes to growth is false. This
simple bivariate test cannot disconfirm the hypothesis. It may be that the
addition of appropriate control variables would make clear a relationship that
does not show in the bivariate test. This test does show, however, that the
simple relationship that seems to exist when the analyst examines only the
most rapidly growing countries disappears when a more representative sample
is examined. If analysts who try to explain the success of the NICs had
examined a more representative sample, they would probably have reached
different conclusions about the relationship between the repression of labor
and growth. As figure 5 shows, labor is just as frequently repressed in slowgrowing
Third World countries as in fast.
The first example above (fig. 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. 4), cases are selected on a
variable—geographical region—that is correlated with the dependent variable.
In both cases, the hypothesized relationship was a simple, direct one: the
higher the level of X (labor repression), the higher the expected level of Y
(growth).
Selection Bias in Comparative Politics 141
Not all causal arguments are so simple. Researchers sometimes posit
arguments with complicated structures of prior and intervening variables that
are more difficult to test rigorously. The consequences of selection on the
dependent variable, however, are the same no matter what the form of argument.
Subsequent sections will consider two frequently encountered variations
on this theme: selection on the dependent variable in a complicated,
contingent historical or path-dependent argument; and selection of the end
points of a time-series or historical case study on the dependent variable.
Selection on the Dependent Variable
in a Path-Dependent Argument
Theda Skocpol's stimulating and thoughtful book States and Social Revolutions
(1979) combines selection on the dependent variable with a complex,
path-dependent argument. She wants to explain why revolutions occur, so she
picks the three most well-known instances—the French, Russian, and
Chinese—to examine. She also examines a few cases in which revolution
failed to occur as contrasting cases at strategic points in her chain of 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.
The central argument in States and Social Revolutions is schematized in
figure 7. Skocpol's argument is that external military threats will cause state
officials 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,
peasant villages are solidary and autonomous from day-to-day landlord super-
Revolution
Peasant
Dominant Class Has Independent Rebellion
Economic Base and Shares Power, i,
Either through Representative
Institutions or Decentralization
Village
Autonomy,
Solidarity
Fig. 7. Schematization of Skocpol's argument
External
Military
Threat
State
Officials
Initiate
Reforms
Opposition
by ~~*"
Dominant
Classes ^
Disintegration
of Old Regime
142 Political Analysis
Dominant Class
Economically
Independent,
Shares Power
Dominant Class
Dependent,
Excluded from
Power
Elite Splits
France
China, after
Taiping
Rebellion
Russia,
World War 1
Elite Cohesive
Prussia
Japan
China, before
Taiping
Russia, before
World War 1
Fig. 8. Effect of external military threat
vision, peasants will take advantage of the elite split and rebel, which will
lead to revolution. This explanation, according to Skocpol, mirrors the historical
record in France and in the parts of China controlled by the Communists.
The Russian case differs from the other two in that the upper class lacked the
independent economic base necessary to impede state-sponsored reforms,
and, consequently, the elite remained unified and revolution failed to occur
after the Crimean War. Nevertheless, defeat in World War I caused elite
disintegration that opened the way for revolution in 1917.
At two points in the chain of argument, Skocpol introduces contrasting
cases to strengthen her contention that structural features identified as important
in these three cases are generally important. 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 faced military threats
at least as severe as that facing France, but did not experience the disintegration
of the elite or, in consequence, revolution. She also looks at Britain
during the Civil War and Germany in 1848 and finds levels of village autonomy
low. In both, elites fragmented, but peasants were not in a position to
take advantage of the situation and, as a result, revolutions did not occur. The
results of these comparisons are summarized in figures 8 and 9.
There is no question but that the examination of a few contrasting cases
makes the argument more persuasive than it would otherwise be, though a test
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
selected without reference to the dependent variable. Nevertheless, it is a step
in the right direction. A rigorous test would be nearly impossible since the
Selection Bias in Comparative Politics 143
Revolution No Revolution
Village
Autonomy
Village
Dependent
Russia
France
China, in area
controlled by
Communists
Britain, 1640-60
Germany, 1848
China, before
Communists
Fig. 9. Effect of elite split
amount of research needed to assess the situation of the dominant class and
the degree of autonomy in peasant villages precludes, for all practical purposes,
the examination of many randomly selected cases.
Skocpol makes no effort, however, to test other links in the chain
of argument. 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. (1979, 23)
This claim seems especially problematic if we accept her implicit definition of
"threatened," that is, as threatened as late eighteenth-century France.
Prance—arguably the most powerful country in the world at the time—was
certainly less threatened than France's neighbors.
Many countries in the world have suffered foreign pressures as great as
those suffered by France and yet revolutions occur infrequently. This raises the
question: are revolutions infrequent because of the absence of appropriate
structural conditions, as Skocpol's argument implies, or because foreign
threats only occasionally set off the sequence of events leading to revolution?
The cases Skocpol selects for examination confirm her argument, but
would a differently selected set of cases do so? Ideally, a test of this link in
Skocpol's argument would examine all nations characterized by the structural
features—that is, village autonomy and a dominant class with an independent
economic base and access to political power—she identifies as necessary to
complete the sequence from military threat to revolution. Then one could
144 Political Analysis
determine whether revolutions occur more frequently in countries that have
faced military threats or not.
In practice, identifying the universe of cases that meet the structural
criteria is probably an impossible task. It would require extensive knowledge
about every country in the world from the French Revolution to the present.
Nonetheless, a serious test of her argument is possible. As it happens, several
Spanish-American countries (Mexico, Guatemala, El Salvador, Honduras,
Nicaragua, Ecuador, Peru, Bolivia, and Paraguay) have the structural characteristics
she identifies 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 NICs).
In all these countries, dominant classes had an independent economic
base in land and/or mines from the nineteenth century until well into the
twentieth. They also shared political power. Thus, they had the economic and
political resources that Skocpol argues are needed in order to successfully
oppose state-sponsored reforms and thus 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. After independence, changes in
property rights reduced the village control over land, but this reduction in the
functions which had contributed to building village autonomy and solidarity
was at least partially offset by the increase in absentee landlordism that accompanied
increasing commercialization.
Most of the land in these countries was held in large tracts. Some peasants
lived on the haciendas, but many lived in traditional villages, owned tiny
parcels of land, and worked seasonally on the haciendas. These villages often
had long histories of conflict with large landowners over land ownership,
water rights, grazing, etc. 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 description of
the autonomous, solidary village structure that makes possible peasants' participation
in revolution.3
With these structural features on which the outcome
is contingent held constant, it becomes possible to test the relationship between
external threat and revolution.
3. Differences of opinion are, of course, possible about whether peasants in these countries
were really autonomous enough from day-to-day landlord control to enable them to play the role
Skocpol allots to them 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 that widespread,
sustained peasantrevolutionarymovements have occurred in most of them.
Selection Bias in Comparative Politics 145
In the following test, I have used a higher level of 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 Skocpol's
argument since the countries may have experienced external pressures
sufficient to meet her criteria even though they have not lost wars. If several
countries have lost wars (and the structural conditions identified as necessary
by Skocpol are present) but have not had revolutions, however, this test will
cast doubt on her argument.
Figure 10 shows the relationship between external threats and revolutions
in these Latin American cases since independence. Skocpol's definition of
revolution is used: rapid political and social structural change accompanied
and, in part, caused by massive uprisings of the lower class (1979, 4-5). It is
not clear how long the sequence from external threat to revolution should
take, but I thought 20 years a reasonable amount of time for the effects of
military threat to be felt.
Figure 10 shows seven instances of extreme military threat that failed to
lead to revolution, two revolutions not preceded by any unusual degree of
external competition or threat, and one revolution, the Bolivian, that fits
Skocpol's argument. These findings suggest that if Skocpol had selected a
broader range of cases to examine, rather than selecting three cases because of
their placement on the dependent variable, she would have come to different
conclusions.
This test does not constitute a definitive disconfirmation of Skocpol's
argument. Competing interpretations of all the variables—village autonomy,
dominant-class independence, military pressure, etc.—exist, and different
operationalizations might lead to different results. In particular, my operationalization
of threat fails to capture the complexity of Skocpol's argument,
and a different operationalization might put Nicaragua and Mexico in the
threat/revolution cell. 1 would argue, however, that any indicator of threat that
identified Nicaragua in 1979 and Mexico in 1910 as threatened would add
dozens of other country-years to the threat/no revolution cell. In short, despite
some deficiencies in operationalization, this cursory examination of cases not
selected on the dependent variable does cast doubt on the original argument.
Selection of the Endpoints of a Time-Series
The final example of how case selection can bias one's conclusions involves
time-series analyses, that is, research designs based on the observation of coun-
146 Political Analysis
Defeated and
Invaded or
Lost Territory
Not Defeated
within 20
Years
Revolution
Bolivia,
Defeated 1935,
Revolution 1952
Mexico, 1910
Nicaragua, 1979
No Revolution
Peru, 1839
Bolivia, 1839
Mexico, 1848
Paraguay, 1869
Peru, 1883
Bolivia, 1883
Bolivia, 1903
All Others
Cuba, 1959*
[El Salvador]**
[Peru]**
[Guatemala]**
i
'The Cuban Revolution is an intermediate case in Skocpol's terms because it
did not entail massive uprisings by tire lower classes.
"These countries have experienced long, peasant-based insurgencies that
have not so far resulted in revolution but may yet.
Fig. 10. Relationship between defeat in war and revolution in Latin
America
tries (or other units of analysis) over a period of time. Here the problem facing
the analyst centers on choosing the endpoints of the study in a neutral way.
Perhaps the most widely known demonstration of the effect of the selection
of endpoints on conclusions is the work by Raul Prgbisch, which apparently
demonstrates a secular decline in the terms of trade for primary products
(1950). The Prdbisch thesis influenced academics and policymakers all over
the developing world, but subsequent work has demonstrated that his findings
hold only for the particular period of time included in the original study.
Studies using different endpoints have failed to replicate Pr£bisch's results
(e.g., Haberler, 1961; Gongalves and Barros 1982).
The selection problem in this instance is that the analyst chooses as the
endpoint for the study a year in which the variable to be explained has attained
some high or low point. The analyst may feel that he or she has no choice in
selecting the endpoint; it may be the last year for which information is available.
Nevertheless, if one selects a case because its value on some variable at
Selection Bias in Comparative Politics 147
the end of a time series seems particularly in need of explanation, one, in
effect, selects on the dependent variable. If the conclusions drawn depend
heavily on the last few data points, they may be proven wrong within a short
space of time as more information becomes available.
An easily examined—because quantitative—example of this comes
from Albert Hirschman's Journeys Toward Progress (1973). In the essay on
inflation, Hirschman advances two interlocked arguments. He suggests that
observers should take a more benign view of inflation, first, because it may
serve as a peaceful alternative to political violence or even civil war. Inflation
may give all groups battling over shares of the national pie the illusion that
they are gaining. Second, he contends that inflation will be brought under
control in time as competing groups realize the futility of their competition
and politicians come to understand the problem better. He concludes the
essay:
Inflation then offers an almost miraculous way of temporizing in a situation
in which two or more parties who are psychologically not ready for
peaceable compromise appear to be set on a collision course. It permits
them . . . to maintain a militant and hostile stance while playing an
elaborate, largely non-violent game in which everybody wins sham victories.
. . . [A]fter having played the game a few times, the parties will
realize its futility. . . . (Hirschman 1973, 223)
Hirschman bases his argument on a case study of inflation in Chile that
ends in 1961 when the conservative administration of Jorge Alessandri seemed
to have succeeded in bringing inflation under control. Chile's yearly inflation
rates from 1930 to 1961 are shown in table 2.
Alessandri was elected in 1958 and began an all-out stabilization effort in
mid-1959 that succeeded in lowering inflation. Does this show, as Hirschman
suggests, that inflation was at long last being conquered after serving the
purpose of averting violent confrontation between classes? Figure 11 shows
Chile's inflation rate from 1930-72. Even with the hyperinflation of 1973-76
excluded, the figure shows no evidence that groups had learned the futility of
pressing inflationary demands or that political leaders had learned to solve the
problem. Rather, 1960 and 1961 appear to be unusual years, best explained by
the orthodox stabilization policies of Chile's last conservative administration
before the military seized power in 1973.
Hirschman would probably have reached different conclusions if he had
waited a few years to write his analysis. Even writing at the time Hirschman
did, a more cautious observer would have been hesitant to rest an argument on
two data points markedly below the trend line. Those two points might turn
out to be harbingers of a new phase of declining inflation and thus evidence of
148 Political Analysis
TABLE 2. Chilean Inflation, 1930-1961
Year
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
Rate
-5%
-4
26
5
9
-1
12
10
2
7
Year
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
Rate
10%
23
26
8
15
8
30
23
17
21
Year
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
I960
1961
Rate
17%
23
12
56
71
84
38
17
33
33
5
10
Source: Hirschman 1973. 160.
"political learning" as Hirschman thought, but they might not. And an alternative
explanation for them—the never to be repeated election by a slim plurality
(31 percent of the vote) of a conservative president—was readily
available.
Conclusion
The reexamination of these three well-known arguments has shown that
choosing cases for study on the basis of their scores on the dependent variable
Inflation
Rate
Endpoinl of
Hirschman's study
70
Fig. 11. Inflation in Chile, 1930-72. (Data for 1930-61 from Hirschman
1973,160. Data for 1962-70 from Valenzuela 1978, 19. Data for 1971-72
from Ramos 1986, 14.)
Selection Bias in Comparative Politics 149
may bias the conclusions one reaches. Apparent causes that all the selected
cases have in common may turn out to be just as common among cases in
which the effect they were supposed to have caused has not occurred. Relationships
that seem to exist between causes and effects in a small sample
selected on the dependent variable may disappear or be reversed in a sample
uncorrelated with position on the dependent variable. Arguments that seem
plausible if a historical study or time-series ends at a particular date may seem
ludicrous if the years included in the study are changed. In short, selecting
cases on the dependent variable entails a high probability of getting the wrong
answer.
This is not to say that studies of cases selected on the dependent variable
have no place in comparative politics. They are ideal 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 accommodate. In so doing, they contribute to building and revising
theories. By themselves, however, they cannot test the theories they propose
and, hence, cannot contribute to the accumulation of theoretical knowledge
(compare Achen and Snidal 1989). To develop and test theories, one must select
cases in a way that does not undermine the logic of explanation.
If we want to begin accumulating a body of theoretical knowledge in
comparative politics, we need to change the conventions governing the kinds
of evidence we regard as theoretically relevant. Speculative arguments based
on cases selected on the dependent variable have a long and distinguished
history in the subfteld, and they will continue to be important as generators of
insights and hypotheses. For arguments with knowledge-building pretensions,
however, more rigorous standards of evidence are essential.
REFERENCES
Achen, Christopher. 1986. The Statistical Analysis of Quasi-Experiments. Berkeley:
University of California Press.
Achen, Christopher, and Duncan Snidal. 1989. "Rational Deterrence Theory and
Comparative Case Studies." World Politics 41:143-69.
Cardoso, Fernando Henrique. 1973. "Associated Dependent Development: Theoretical
and Practical Implications." In Authoritarian Brazil: Origins, Policies, and
Future, ed. Alfred Stepan. New Haven: Yale University Press.
Deyo, Frederic. 1984. "Export Manufacturing and Labor: The Asian Case." In Labor
in the Capitalist Economy, ed. Charles Bergquist. Beverly Hills: Sage.
Deyo, Frederic. 1987. "State and Labor: Modes of Political Exclusion in East Asian
Development." In The Political Economy of the New Asian Industrialism, ed.
Frederic Deyo. Ithaca, N.Y.: Cornell University Press.
Gongalves, Reinaldo, and Amir Coelho Barros. 1982. 'Tendencias do termos de troca:
150 Political Analysis
A tese de PrSbisch e a economia brasileira—1850/1979" [Tendencies in the
Terms of Trade: The Pr£bisch Thesis and the Brazilian Economy, 1850-1979].
Pesquisa e Planejamenw Economico 12(1): 109-31.
Haberler, Gottfried. 1961. 'Terms of Trade and Economic Development." In Economic
Development for Latin America, ed. Howard S. Ellis and Henry C. Wallich.
London: MacmiMan.
Haggard, Stephan. 1986. "The Newly Industrializing Countries in the International
System." World Politics 38:343-70.
Hirschman, Albeit O. 1973. Journeys Toward Progress: Studies of Economic PolicyMaking
in Latin America. New York: Norton.
Johnson, Chalmers. 1987. "Political Institutions and Economic Performance: The
Government-Business Relationship in Japan, South Korea, and Taiwan." In The
Political Economy of the New Asian Industrialism, ed. Frederic Deyo. Ithaca,
N.Y.: Cornell University Press.
King, Gary. 1989. Unifying Political Methodology: The Likelihood Theory of Statistical
Inference. New York: Cambridge University Press.
Koo, Hagen. 1987. "The Interplay of State, Social Class, and World System in East
Asian Development: The Cases of South Korea and Taiwan." In The Political
Economy of the New Asian Industrialism, ed. Frederic Deyo. Ithaca, N.Y.: Cornell
University Press.
O'Donnell, Guillermo. 1973. Modernization and Bureaucratic-Authoritarianism:
Studies in South American Politics. Berkeley: University of California Institute of
International Studies, Politics of Modernization Series, no. 9.
Prdbisch, Raul. 1950. The Economic Development of Latin America and Its Principal
Problems. New York: United Nations.
Ramos, Joseph. 1986. Neo-Conservative Economics in the Southern Cone of Latin
America, 1973-1983. Baltimore: Johns Hopkins University Press.
Skocpol, Theda. 1979. States and Social Revolutions: A Comparative Analysis of
France, Russia, and China. Cambridge: Cambridge University Press.
U.S. Department of State. 1981. Country Reports on Human Rights Practices for
1980. Washington: U.S. Government Printing Office.
U.S. Department of State. 1989. Country Reports on Human Rights Practices for
1988. Washington: U.S. Government Printing Office.
Valenzuela, Arturo. 1978. The Breakdown of Democratic Regimes: Chile. Baltimore:
Johns Hopkins University Press.
World Bank. 1984. World Development Report 1984. New York: Oxford University
Press.
World Bank. 1988. World Development Report 1988. New York: Oxford University
Press.
Controversy