Академический Документы
Профессиональный Документы
Культура Документы
Process-Tracing Methods
Derek Beach and Rasmus Brun Pedersen have completely reworked and
updated the first book-length guide for using process-tracing in social
science research. They begin by introducing a more refined definition of
process-tracing, differentiating it into four distinct variants, and explaining
the applications for and limitations of each. The authors develop the un-
derlying logic of process-tracing, including how one can understand causal
mechanisms. They walk the researcher through each stage of the research
process, starting with issues of developing mechanistic theoretical explana-
tions, formulated in either minimalist terms or as systems that detail how
a process could work in a given case. Building on recent advances in case-
selection techniques, the book elaborates guidelines for which cases are
appropriate for the different variants of process-tracing. The book then de-
velops an innovative framework that enables generalizations about mecha-
nistic claims to be made. The authors then develop guidelines for making
causal inferences using Bayesian logic in an informal, nontechnical fashion
and for translating empirical material into mechanistic evidence that may
or may not confirm how causal mechanisms operate in particular cases.
The book concludes with three chapters that provide a step-by-step guide
to using the different variants of process-tracing.
This updated book makes three major contributions to the method-
ological literature on case studies. First, it develops the underlying logic
of process-tracing methods in a level of detail that has not been presented
previously and thus establishes a new standard for the use of these methods
in tracing causal mechanisms in real-world cases. Second, by explaining
the application of Bayesian logic to process-tracing methods in a non-
technical manner, the volume provides a coherent framework for drawing
strong causal inferences within cases using mechanistic evidence. Finally,
the book develops guidelines for case selection and generalizing that are
appropriate for tracing mechanisms, whereas existing guidelines are blind
to the potential risk of mechanistic heterogeneity.
Widely acclaimed instructors, the authors draw on their extensive ex-
perience at the graduate level in university classrooms and professional
workshops across the globe.
Process-Tracing Methods
Foundations and Guidelines
Second Edition
A CIP catalog record for this book is available from the British Library.
Contents
Acknowledgmentsix
Chapter 1. What Is Process-Tracing? 1
1.1. Introduction: The Three Components of Process-Tracing
as a Social Science Method 1
1.2. What Are We Tracing? 2
1.3. What Are Traces? 4
1.4. Case Selection and Generalization 6
1.5. Four Variants of Process-Tracing 9
1.6. A Primer on the Foundations of Case-Based Methods 12
Chapter 2. What Are We Tracing? 29
2.1. Introduction 29
2.2. The Nature of Causal Mechanisms 30
2.3. Five Common Questions Relating to Mechanistic Approaches 41
Chapter 3. Theorizing Concepts and Causal Mechanisms 53
3.1. Introduction 53
3.2. Defining Concepts to Be Compatible with
Mechanistic Explanations 54
3.3. Theorizing Mechanisms—Minimalist and Systems
Understandings64
3.4. Level of Abstraction of Mechanistic Explanations 73
3.5. The Importance of Context and the Dangers of Mechanistic
Heterogeneity77
3.6. The Temporal Dimension and Levels of Analysis 81
3.7. The Building Blocks of Mechanism-Based
Theoretical Explanations 84
3.8. How to Theorize about Mechanisms in Practice 86
vi Contents
Contents vii
Acknowledgments
x Acknowledgments
dards for case-based designs in their own research. We are greatly in debt
to the participants in these courses on process-tracing methods, including
but by no means limited to the participants at the ECPR Summer Schools
(2011–17) and ECPR Winter Schools (2012–18), the IPSA Summer Schools
in São Paulo (2012–18), the Berlin Graduate School for Transnational Stud-
ies (2012–14, 2016–18), APSA short courses (2013–17), the Concordia Work-
shops for Social Science Research (2013–17), and the short courses at the
ICPSR Summer School (2016–17). We particularly thank our teaching as-
sistants at the ECPR and IPSA schools, who served as vital interlocutors and
sparring partners in developing our ideas for this book. We thank Terra Bu-
dini, Natália Nahas Carneiro Maia Calfat, Peter Marton, Hilde van Meeg-
denburg, Kim Sass Mikkelsen, Yf Reykers, and Camila Rocha. The picture
used on the cover is part of a visual depiction of a theorized causal mecha-
nism linking economic development and democratization developed by two
participants (Camilla Rocha and Betina Sarue) in a PhD course at the IPSA
Summer School in São Paulo, Brazil.
C h a p ter 1
What Is Process-Tracing?
You know a conjurer gets no credit when once he has explained his
trick; and if I show you too much of my method of working, you will
come to the conclusion that I am a very ordinary individual after all.
—Sherlock Holmes, A Study in Scarlet
2 Process-Tracing Methods
What Is Process-Tracing? 3
2010; Hedström and Yklikowski 2010; Waldner 2012). However, the essence
of mechanistic explanations is that we shift the analytical focus from causes
and outcomes to the hypothesized causal process in between them. That is,
mechanisms are not causes but are causal processes that are triggered by causes
and that link them with outcomes in a productive relationship. But beyond
this core point, there is considerable disagreement about the nature of mecha-
nisms. We discuss these issues in chapters 2 and 3. In chapter 2, we first
discuss two understandings of causal mechanisms that are not compatible
with the goal of tracing mechanisms using within-case studies: descriptive
narratives and intervening variables. Descriptive narratives are by definition
not causal because the causal linkage between events is not developed. While
viewing mechanisms as intervening variables can be relevant when engaging
in large-n cross-case analyses, it results in the black-boxing of precisely the
causal links that we wanted to trace because relevant evidence for the causal
effect of the intervening variable (i.e., mechanism) is the difference that pres-
ence/absence of it makes across cases that are otherwise similar in all other
factors. Instead of tracing the process as it played out within a case, we would
investigate cross-case variation. This means that we lose focus on the process
that links together causes and outcomes, which was our original objective.
The chapter then develops two different understandings of mechanisms
that are compatible with process-tracing and that are widely used in the
social and natural sciences, a minimalist understanding and a systems un-
derstanding. In a minimalist understanding, the causal mechanisms—the
arrow between a cause and outcome—is not unpacked in much detail, either
empirically or theoretically. This understanding is appropriate to use early in
a research process, where many different plausible mechanisms might link a
cause and an outcome and we want to explore whether there is any evidence
for particular mechanisms, enabling us to focus our analytical attention on
one or a small handful of mechanisms. It is also appropriate as a follow-up to
in-depth process-tracing as an analytically less taxing tool to explore whether
similar mechanisms operate in other cases.
In a systems understanding, the core elements of a causal mechanism
are unpacked theoretically and studied empirically in the form of the traces
left by the activities associated with each part of the process. Mechanisms
are here viewed in a more holistic fashion, where the effects of a mechanism
are more than the sum of their parts. Each of the parts of the mechanism
can be described in terms of entities that engage in activities (Machamer
2004; Machamer, Darden, and Craver 2000). Entities are the factors (ac-
tors, organizations, or structures) engaging in activities, whereas the activi-
4 Process-Tracing Methods
ties are the producers of change or what transmits causal forces or powers
through a mechanism. When a causal mechanism is unpacked theoretically
as a system, the goal becomes understanding how a process actually works
by tracing the operation of each part (or at least the most critical parts) in
one or more cases. This requires considerable analytical resources, meaning
that in-depth process-tracing is often done only after we have some idea that
there is something to look at. In-depth process-tracing is like bringing out
an electron microscope: we only do it after we have an idea that there might
be something to look at because we have engaged in case research using
minimalist process-tracing.
Chapter 3 discusses how we can theorize causal mechanisms in process-
tracing. We start by developing how defining concepts (causes and out-
comes) differs in process-tracing from other methods, including case-based
comparisons and variance-based approaches. We then explore how theorized
mechanisms can look when we raise or lower the analytical level of abstrac-
tion from detailed, case-specific mechanisms to abstract, midrange theorized
mechanisms that may be present in a large number of cases.
What Is Process-Tracing? 5
6 Process-Tracing Methods
What Is Process-Tracing? 7
causal mechanism in the studied case (Illari 2011). Generalizing these find-
ings to other cases requires the ability to document that other cases are caus-
ally similar to the studied case, meaning that we can then expect similar
processes to operate there also. However, given the contextual sensitivity of
mechanistic explanations, we argue that it is not enough to assume mecha-
nistic homogeneity in other cases based on cross-case comparisons because
of the risk of mechanistic heterogeneity produced by known/unknown con-
textual differences. In chapter 4, we uncover a range of different sources
of mechanistic heterogeneity that can result in flawed generalizations if we
merely assume that a similar process should operate in other cases that look
similar in cross-case terms. Instead, we suggest a snowballing-outward strat-
egy for exploring a bounded population to test whether or not mechanistic
heterogeneity is present, thereby reducing the risk of flawed generalizations.
A critical reader might view these arguments as a recipe for piecemeal,
noncumulative research, with process-tracing resulting only in idiographic
studies of mechanisms that operate in only one or a small number of cases.
This is a common argument put forward by neopositivist variance-based
scholars. Gerring writes that “social science gives preference to broad infer-
ences over narrow inferences. First, the scope of an inference usually cor-
relates directly with its theoretical significance. . . . Second, broad empirical
propositions usually have greater policy relevance, particularly if they ex-
tend to the future. They help us to design effective institutions. Finally, the
broader the inference, the greater its falsifiability” (2017: 234).
However, the alternative to taking mechanistic heterogeneity seriously
by appreciating the complexity of how processes play out in real-world cases
and the limited bounds of generalization of mechanisms because of contex-
tual sensitivity is to lift the level of abstraction about mechanisms to such
a high level that our theorized mechanisms are in essence nothingburgers
that tell us precious little, if anything, about how a process works in the real
world (for examples of nothingburger mechanisms, see chapters 3 and 4).
Despite Gerring’s claim about policy relevance, the field of policy evaluation
is increasingly interested in using process-tracing and the tracing of mecha-
nisms as an analytical tool to study how interventions actually work in par-
ticular contexts instead of working with broad propositions that tell us little
about how things work in the real world (Bamanyaki and Holvoet 2016;
Wauters and Beach 2018; Befani and Stedman-Bryce 2016; Cartwright 2011;
Cartwright and Hardie 2012; Clarke et al. 2014; Schmitt and Beach 2015).
The claim about breadth of claims is also increasingly under fire in the
natural sciences, where the appreciation of complexity and the contextual
8 Process-Tracing Methods
What Is Process-Tracing? 9
Theory-Testing Process-Tracing
10 Process-Tracing Methods
with a mechanism and its parts. The predictions about mechanistic evidence
should be as clear as possible, making it easier to determine whether or not
they are then actually found in the subsequent case study. The researcher
then collects and assesses the available empirical record to determine whether
mechanistic evidence suggests that the mechanism was present and worked
as theorized or whether the theory needs to be modified. If the predicted evi-
dence is found, we can then infer that the hypothesized causal mechanism is
present in the case and worked as we theorized. If evidence is not found for a
given part (or for the overall mechanism, if the minimalist understanding is
used), the researcher should engage in a round of theory-building using the
insights gained from the empirical analysis of what went wrong as inspira-
tion for building theories of new parts of the mechanism.
What Is Process-Tracing? 11
the empirical record, with material collected without knowing what it is evi-
dence of yet. Here it can also be helpful to develop a descriptive narrative of
what happened in the case to shed light on potential mechanisms. The next
step involves inferring that the found observable empirical material is actual
evidence that reflects the empirical fingerprints left by the operation of a
plausible causal mechanism in the case. Tentative hunches about potential
mechanisms (and their parts in the systems understanding) are made based
on the first round of empirical probing, after which the researcher evaluates
whether any of the collected material is actually evidence of the tentative
mechanism (or parts of the mechanism). This evaluation of evidence pro-
ceeds in a slightly different fashion than in theory-testing process-tracing,
given that discussing the certainty of evidence is not relevant because it has
already been found; instead, one evaluates only the uniqueness in relation to
the tentative hypothesized mechanism or its parts.
Evidence does not speak for itself. Theory-building often has a strong
testing element in that scholars seek inspiration about what to look for from
existing theoretical work and previous observations. Here, existing theory
can be thought of as a form of grid to detect systematic patterns in empirical
material, enabling inferences about predicted evidence. In addition, one can
look to research on mechanisms on similar research topics for inspiration
about how parts of the mechanism may look. In other situations, the search
for mechanisms is based on hunches drawn from puzzles that existing work
cannot explain.
Theoretical-revision process-tracing involves a combination of tracing a
mechanism in a deviant case, where a mechanism should have been opera-
tive but where it broke down, and using the information garnered about
where in the process the mechanism broke down to engage in a focused
comparison of what differs between a typical case where the mechanism
worked and the studied deviant case. The goal is to uncover unknown omit-
ted conditions that must be present for the mechanism to function properly.
12 Process-Tracing Methods
What Is Process-Tracing? 13
relationship plays out. The temporal and spatial scope and the level of analy-
sis (micro, mezzo, or macro) of a case are defined by the causal theory with
which we are working. The temporal bounds are defined as starting with
the occurrence of a cause and ending with the manifestation of the out-
come. The spatial bounds of a case are determined by the scope in which
a theory operates, which could be a meeting room for social psychologi-
cal theories like groupthink that deal with decision-making within small
groups (Janis 1983) or the international system as a whole over many years
for international relations theories like neorealism (Waltz 1979). The level of
analysis also depends on the theory, ranging from individual-level analyses
to macrolevel theories dealing with collective actors (social groups, states,
international organizations, and so forth). Therefore, the bounds of a “case”
are determined by the causal theory with which we are working.
In any given case, a hypothesized causal relationship either has taken
place or not, which logically means there are no within-case likelihoods or
propensities of a relationship taking place. Based on our knowledge of con-
textual conditions, we might estimate that a given case has a higher or lower
propensity for the hypothesized causal relation to exist, but when we are
engaging in an actual study of that case, the relationship is either present or
not. In case-based research, because individual cases are the analytical point
of departure, we are working with a form of bottom-up analysis of causal
relationships, which then logically implies that we must adopt a determin-
istic and asymmetric understanding of causation if within-case analysis is
to make any sense. We now turn to the reasons underlying this argument.
14 Process-Tracing Methods
What Is Process-Tracing? 15
16 Process-Tracing Methods
a reason (or more realistically, a set of reasons) (Adcock 2007; Bhaskar 1979:
70–71; Mahoney 2008). This is a form of bottom-up claim about causation
(Russo and Williamson 2011) in which outcomes have occurred for a set of
reasons within a particular case, but moving upward from the individual
to a population can only realistically be done when we translate our claims
into probabilistic ones except if the population exhibits a “law-like” causal
homogeneity (Cartwright 1999: 154–59)—that is, if the same cause produces
the same outcome through the same process in all of the cases in the popula-
tion. This does not mean that we empirically are always able to figure out
why something happened in a case. However, just because we cannot figure
out empirically why something occurred does not mean that the outcome
was the product of randomness. Things do not “just happen”; things happen
for a reason.
The distinction between ontological probabilism and determinism is one
of the most misunderstood distinctions in the social sciences. One reason is
that many scholars conflate ontological and epistemological issues. Here we
are discussing the nature of causal claims (ontology), not how we can learn
about these causal relationships using empirical research (epistemology).
Building on the work of social science methodologists such as Goertz and
Mahoney and philosophers of science such as Illari, Russo, and William-
son, we argue that only a deterministic ontology is compatible with study-
ing mechanisms in single cases using process-tracing (Goertz and Mahoney
2012; Illari and Russo 2014; Mahoney 2008; Russo and Williamson 2011);
anything else makes the study of individual cases a relatively superfluous
exercise that merely illustrates that a found cross-case mean causal effect ac-
tually makes sense at the case level. The evidential heavy lifting of assessing a
probabilistic causal claim about a mean causal effect is, however, done by the
cross-case analysis, preferably in the form of a controlled experiment (e.g.,
Gerring 2011). At the same time, process-tracing should adopt a probabilistic
epistemology based on Bayesian logic (see chapter 5), which tells us that our
knowledge of the world will never be certain but instead that our degree of
confidence in a causal relationship being valid depends on the quality of
the evidence produced. Epistemological determinism would mean that we
can gain knowledge about causal relationships that is 100 percent certain,
which is naturally an impossibility outside of the realm of trivial facts and
conspiracy theories.
Surprisingly, the distinctive combination of ontological determinism and
epistemological probabilism has not been clearly developed in social science
methodology, although it has long existed in the philosophical literature
What Is Process-Tracing? 17
(Chatterjee 2011). Given the lack of development of this position, many ex-
isting accounts of case-based research methods adopt a weaker position as a
means of protecting themselves from critiques of epistemological determin-
ism by softening their deterministic ontological claims. For example, many
case-based scholars use such terms as an “almost deterministic” relationship
to refer to a pattern where C might be necessary in 19 out of 20 cases (e.g.,
Ragin 2000). In practice, this loosening has quite serious implications for
the building and testing of theories both within and across cases, given that
we never really know whether we have found an exception that proves the
rule (the 20th case where a case-specific condition meant that the relation-
ship did not work there) or whether there is a much more complicated
causal pattern (e.g., the cause has a positive impact in some of the cases but
a negative or no impact in other cases, depending on the context), or even
no relationship at all (spurious causation).
Another common misunderstanding is that ontological determinism is
the same as the theoretical determinism of structural, macrolevel theories.
Dunning (2017) conflates the two when he claims that deterministic theo-
ries make claims about outcomes being “inevitable.” If this argument holds,
there would no place for human free will (i.e., agency) in deterministic theo-
ries. While this is true for strong structural macrolevel theories, ontological
determinism is also compatible for theories that incorporate agency. Indeed,
the claim that it is always possible for one or more conditions to interfere
with the workings of a theorized mechanism, leading to the process break-
ing down in a given case, is completely compatible with an ontologically
deterministic causal logic.
We can illustrate this conflation of structural deterministic theories
and ontologically deterministic understandings of causation itself by com-
paring a structural theory with a theory that incorporates agency. Waltz’s
(1979) structural realist theory of the international system leaves little if
any space for agency or contingency by either states in their foreign poli-
cies or individual leaders within states. The distribution of relative power
determines the type of international system that exists, acting as a struc-
tural straightjacket on states that produces the broad contours of their
foreign policies. Here, there is little wiggle room for state agency, especially
when viewed over the longer term. But this is a theoretical position on the
agent-structure/micro-macro debate (see section 2.4). Taking a position
about the causal priority of structure is not the same thing as understand-
ing causation in a deterministic fashion.
Indeed, many theories hypothesize that actor-level causes such as agency
18 Process-Tracing Methods
What Is Process-Tracing? 19
the actual producing large deviations between the predicted and the actual
over time. We thus must be very careful in generalizing from a studied case
to other, nonstudied cases, but we do not necessarily need to adopt ontologi-
cal probabilistic models. Instead, when studying past cases, we should try to
understand these complexities and model them theoretically.
1. The “randomness” in the quantum world deals in particular with the indeterminacy of
properties, the nonlocalizability of quantum objects, and the nonseparability of quantum
states as a result of entanglement. For a good introduction, see Kuhlmann and Glennan 2014.
20 Process-Tracing Methods
world than importing lessons from quantum behavior that cannot be ex-
pected to hold at our level of existence is that when we are making causal
claims across a number of cases, we can at best detect trends (e.g., mean
causal effects) because of the causal complexity of the social world.2 This
relates to the question of causal homogeneity/heterogeneity. A causally ho-
mogenous set of cases is where the same cause(s) produce the same out-
comes, whereas a set of causally heterogeneous cases is one where the same
outcome might be produced by different causes, or the same cause produces
different outcomes.3 For example, when explaining what causes democratic
transitions, we might find a cross-case pattern that suggests that higher lev-
els of inequality and distributional conflict in authoritarian societies make
democratic transition more likely (Haggard and Kaufmann 2016). However,
the population of all cases of democratic transitions since 1980 is very het-
erogeneous because of the contextual conditions that differentiate the South
American cases in the 1980s from the collapse of authoritarianism in Central
and Eastern Europe in the 1990s. Therefore, Haggard and Kaufmann find
that the cross-case trend of inequality and distributional conflict with transi-
tions does not hold in particular cases of democratic transition: for example,
in Czechoslovakia, a combination of the withdrawal of Soviet hegemony
and a desire to “return to Europe,” not the level of distributional conflict,
created conditions under which a transition became possible.
At the end of the day, whether the stochastic elements of the social world
result from inherent randomness or are the product of causal complexity
across cases is irrelevant—the methodological implications of ontological
probabilism are the same (Gerring 2005, 2011; King, Keohane, and Verba
1994: 211; Marini and Singer 1988). Variance-based scholars approach the
study of mechanisms and process using probabilistic terms, treating causal
mechanisms as analogous to intervening variables that make the occurrence
of outcome O more likely within a population of cases, other things equal.
The process is translated into a simple pathway “variable” whose causal effect
within a population of cases can then be assessed statistically (Weller and
Barnes 2014). The mean causal effect can then be assessed by investigating
cross-case trends.
Many scholars would argue that studying single cases is meaningless
2. As the rest of this book explores, complexity does not necessarily mean that we must
adopt probabilistic ontological assumptions; it can also mean that we just have to be more
cautious in the scope of generalizations being made.
3. Within variance-based approaches, the term typically used to refer to causal homogene-
ity is stable unit treatment value assumption (SUTVA). See Morgan and Winship 2007: 37–40.
What Is Process-Tracing? 21
when we are dealing with mean causal effects. Indeed, King, Keohane, and
Verba (1994: 129) go so far as to say that an investigation of one case tells
us nothing: “Nothing whatsoever can be learned about the causes of the
dependent variable without taking into account other instances when the
dependent variable takes on other values.” We could make the probabilis-
tic claim that when the star player on our favorite team (cause) plays, the
probability of our team winning increases. A single game in which the star
player was absent and the team lost would provide no information about the
relationship because there is no variation in the outcome of the game. The
probabilistic claim about the star player increasing the probability that our
team wins should instead be assessed by measuring the difference that the
star player makes across a number of cases, holding other factors constant.
If we still insisted on engaging in a case study in this variance-based
probabilistic world, King, Keohane, and Verba’s (1994) methodological
guidance is to recommend that we transform one case into many to assess
the mean causal effects that variation in the values of X have for values of
Y across cases. In their opinion, a case can be disaggregated geographically
(e.g., moving from the federal (n = 1) to state level (n = 50)), temporally (e.g.,
before/after), or by splitting a negotiation into distinct issue areas. We might
disaggregate a game by treating quarters or halves as cases. We could then
investigate whether our team was winning when the star player was present
and losing when he was absent. However, as chapter 5 explores, we cannot
treat these cases as independent cases, which is required to assess difference-
making variation. We might have been winning in the first half of the game
when the star player was on the field. However, at halftime, the other team
rallies in the dressing room, and they return to the field so highly motivated
that they cause an injury to our star player. Our team then loses, but not
because of his absence but because the “outcome” of the first half (our team
winning) acted as a “cause” for the different outcome (loss) in the second
half. In other words, the value of the outcome in t0 influences the value of
a cause in t1.
Other scholars suggest that it can make sense to study single cases if we
are able to estimate the likelihood of finding a causal variable being op-
erative (here a mechanism) in a given case (Gerring 2007: 115; Macartan
Humphreys and Jacobs 2015; Levy 2008: 12).4 This would take the form of
a propensity score, which is a function of the impact of the other variables
4. In the second edition of his book, Gerring (2017) suggests that single-case studies merely
serve an “illustrative function.”
22 Process-Tracing Methods
present in a given case along with the level of the mechanism present in the
case if it is measured as an ordinal-or interval-scale variable. “A ‘least-likely’
case is one that, on all dimensions except the dimension of theoretical inter-
est, is predicted not to achieve a certain outcome, and yet does so” (Gerring
and Seawright 2007: 115). In contrast, “a ‘most-likely’ case is one that, on
all dimensions except the dimension of theoretical interest, is predicted to
achieve a certain outcome, and yet does not” (115). The logic, then, is that
if we find that the mechanism operates in the least-likely case, we increase
our cross-case estimate of the strength of the causal effect in the population,
whereas if we do not find the mechanism in a most-likely case, we down-
grade our estimate of the strength of the causal effect in the population.
However, this updating of cross-case propensities based on the mismatch
between predicted and actual propensity of a single case requires the very
strong assumption of cross-case causal homogeneity within a population of
cases (Cartwright 1999). This means that we must be able to assume that
all variables have a continuous and linear relationship with the outcome.
For example, what if a particular causal condition matters in a subset of the
cases (e.g., a geographic region) but not in the other cases? We might have
selected a “least-likely” case from this region, where we had expected that
the relationship would not hold because of adverse contextual conditions,
but nevertheless found that it did matter. Based on the logic of least-likely
cases, we would then make a generalization with more confidence that the
same relationship holds across all cases, also in other regions. However, this
inference would be a flawed unless we can demonstrate that the strong as-
sumption of perfect causal homogeneity within the population holds. In
the real world, context matters for how causal relationships work. A causal
condition might work in a small set of cases but have precisely the opposite
effect in another context (see chapter 4).
What Is Process-Tracing? 23
expect it (a least-likely case), we should revise the theory about the context in
which the relationship holds rather than make stronger generalizations. For
example, when studying elite decision-making, we might have selected what
looked like a least-likely case for the relationship between efficient presi-
dential leadership (C) and rational decision-making (O), where in a severe
crisis situation, conditions such as the extreme stakes involved and the short
time frame for decisions would be expected to prevent efficient presidential
leadership (C), thereby resulting in poor decision-making processes (~O).
But if we then found in a case study that the outcome was actually rational
decision-making (O), we would not make a stronger cross-case inference
that C matters across the population of more typical cases because it mat-
tered when we least expected it. Instead, we would want to delve into the
case to understand why our theoretical expectations about the impact of the
adverse context were confounded, probing which factors would enable us to
more accurately delineate the contextual conditions required for presidential
leadership to matter. One plausible result would be that we revise our theory
to claim that the combination of extreme stakes and short time frame makes
actors more vigilant in their decision-making, as occurred in the Cuban
Missile Crisis (see Janis 1983).
In a case-based understanding of process-tracing, the tables are flipped.
Here the analytical focus is on understanding how the mechanism works
within single cases. Cases are the natural focus because causes and the mech-
anisms linking them with outcomes are causally operative within cases, not
across cases (Russo and Williamson 2011: 61). This means that the traces of
their operation are found within cases, making individual cases the analyti-
cal point of departure. This implies a bottom-up form of research in which
we first assess whether mechanism A was present and functioned as hypothe-
sized in a single case. The next question then is whether the same mechanism
is present in other cases, and if so, under what conditions mechanism A links
together the cause and outcome. But for this bottom-up approach to study-
ing mechanisms to make sense, an ontological deterministic understanding
of causation must be adopted.
For process-tracing in cases to make sense, therefore, we logically have
to adopt an ontological deterministic position. This does not mean that the
mechanism will always occur; whether or not it does is a function of the
presence or absence of the contextual conditions required to trigger it. Nor
does it mean that the outcome will always occur when the cause and mecha-
nism are present; this is dependent on whether other causal and contextual
conditions are present that together are sufficient to produce the outcome.
24 Process-Tracing Methods
What Is Process-Tracing? 25
Our claims in process-tracing are asymmetric: that is, we make claims about
the mechanisms triggered by a cause (or set of causes) but make no claims
about other mechanisms triggered in other circumstances or what happens
when the cause is not present. Whether we make symmetric or asymmetric
claims has major implications for research design, notably with regard to
how we conceptualize concepts and mechanisms (chapters 2–3), how we
make causal inferences (chapters 5–7), and proper case selection (chapter 4).
Symmetric causation means that we hypothesize that there are causal
effects across different values of both causes and outcomes. For example, a
deterministic symmetric claim would be that low values of X result in low
values of Y and high values of X result in high values of Y. A probabilistic
symmetric claim would be that as values of X increase, so does the probabil-
ity that higher values of Y will occur. Symmetric claims are typically made by
variance-based scholars who build on constant conjunction and/or counter-
factual understandings of causation (e.g., Gerring 2011; J. Woodward 2004).
26 Process-Tracing Methods
What Is Process-Tracing? 27
C h a p ter 2
2.1. Introduction
Even though the term process is in the title, there is widespread disagreement
about what we are actually tracing when engaging in process-tracing. Most
scholars acknowledge that process- tracing is tracing causal mechanisms
(Beach and Pedersen 2016a; Bennett 2008a, b; Bennett and Checkel 2014;
Checkel 2008; Hall 2008; Rohlfing 2012), but there is considerable confu-
sion and discord about what causal mechanisms actually are. These disagree-
ments about what we are tracing have resulted in considerable methodologi-
cal debate about what good process-tracing is.
This chapter cuts through this confusion and discord by developing two
distinct positions relating to the nature of causal mechanisms that are com-
patible with the goal of tracing processes empirically within cases. Both posi-
tions are relevant for process-tracing, but the appropriate choice depends on
the research situation. Early in a project, it can make sense to operate with a
minimalist understanding, where mechanisms remain relatively black-boxed
as mere “sketches” in which the parts and the causal logics linking them
are not specified. When there is considerable uncertainty about whether
a mechanism links a cause (or set of causes) and an outcome or multiple
mechanisms might plausibly link them together, it makes sense to engage in
a form of plausibility probe to see whether there is any evidence of a link be-
fore engaging in a detailed tracing of a full-fledged mechanism. In contrast,
tracing a full-fledged mechanism makes sense when we have a strong hunch
29
30 Process-Tracing Methods
that a particular mechanism might be operative and want to test more rig-
orously whether it operates as we had theorized, or when there is strong
evidence of a causal relationship and we want to find what mechanism links
together the causes and outcome.
Drawing on recent advances in the philosophy of science related to
mechanisms and case-based methodologies, this chapter delineates the dif-
ferent positions in the debates about the nature of mechanisms, explain-
ing which positions are compatible with a case-based approach to study-
ing causal mechanisms using process-tracing. We first explain why tracing
events in a case is not the same thing as tracing a causal mechanism and
why mechanisms should be understood in case-based approaches as some-
thing more than mere intervening variables.1 After discussing what causal
mechanisms are not, we then explore two different conceptions of what
causal mechanisms are (minimalist and systems) that are compatible with
a case-based understanding of process-tracing. We discuss the methodologi-
cal consequences and trade-offs of adopting either a minimalist or systems
understanding of mechanisms. The chapter concludes with a discussion that
addresses five common questions regarding mechanisms.
research designed to make causal inferences. The second two positions in-
volve a minimalist and a maximalist (systems) understanding of the nature
of mechanisms. Causal mechanisms should not be understood as series of
events, given that a descriptive narrative is not the same thing as explaining
a process that links together causes and outcomes. An intervening-variable
understanding is not compatible with the goal of tracing the process using
in-depth, within-case analysis. Therefore, these two positions are depicted in
gray in table 2.1.
This leaves us with the minimalist and systems-understanding positions.
Both are compatible with a case-based understanding of process-tracing; the
choice of which one is more appropriate depends on the question in a par-
ticular research project.
32 Process-Tracing Methods
Mahoney 2012: 571; Roberts 1996: 16; Suganami 1996: 164–68). This type of
research takes the form of an empirical narrative in the form of “actor A did
X to actor B, who then changed its position on issue Y,” etc.
Yet causal mechanisms are more than just a series of events. A sequence of
events tells us who did what but does not tell us why or how the events were
linked together in a causal sense. Describing a series of events can provide
a plausible descriptive narrative about what happened but does not shed
light on the causal question of why things happened. Causal explanation
therefore involves more than just tracing temporal sequences (Gryzmala-
Busse 2012: 1268–71). Tracing a series of events between C and O is in reality
merely crafting a “just so” story of case-specific happenings between C and
O, with the causal logic underlying the linkages between events completely
black-boxed. Therefore, sequences of events do not qualify as mechanistic
explanations (Craver and Darden 2013). The result of viewing mechanisms
as series of events is that the causal mechanism linking causes and outcomes
is completely black-boxed (Bunge 1997). This is depicted in figure 2.1.
Abell’s account of analytical narratives goes a bit further than suggest-
ing we should just trace events; rather, he contends that we need to de-
velop narrative structures that include action linkages between events that
build on subjective counterfactuals, where we ask actors who participated
in a process whether things could have been different at critical junctures
(2004: 295–96). However, this position has two problems. First, logically it
collapses back down onto the counterfactual understanding of causation,
which means that we shift our attention from how a process worked within
a case to the investigation of the difference that variations in action linkages
make for the outcome across cases (i.e., comparing the actual with the hypo-
thetical scenario imagined by a participant). Second, it significantly reduces
the scope of research questions to only those that can be assessed by asking
actors whether things could have been different.
Some scholars view mechanisms as an intervening or mediating variable
between C and O (Gerring 2007; King, Keohane, and Verba 1994; Weller
and Barnes 2014). For example, Mahoney (2015: 206) writes, “I use the term
mechanism to refer to a factor that intervenes between a cause and outcome.
I treat mechanisms in the same way as causes and outcomes; they are par-
ticular events or specific values on variables. Mechanisms are different from
causes and outcomes because of their temporal position: they stand between
a cause and outcome in time. Thus, in the expression X → M → Y, the
letters refer to events or specific values on variables, with X being treated
as the cause, M as the mechanism, and Y as the outcome.” In this section
we discuss the theoretical problems with treating mechanisms as variables,
whereas in chapter 5 we discuss the empirical challenges of claiming that we
are tracing mechanisms within cases by using evidence of difference-making
across cases.
At the theoretical level, treating a mechanism as a variable means logi-
cally that it collapses down onto a counterfactual understanding of causa-
tion (Woodward 2003: 350–58). Counterfactual causation is defined as the
claim that a cause (or mechanism) produced an outcome because its absence
would result in the absence of the outcome, all other things being held equal
(Lewis 1986: 160; Woodward 2003). Treated as variables, mechanisms have
an independent causal impact that must be assessed by measuring the im-
pact of the mechanism’s absence on values of the outcome across cases, with
everything else held constant—that is, using evidence of difference-making.
Yet treating mechanisms as intervening variables (and thereby as
counterfactual-based claims) results in the theoretical black-boxing of the
causal links in our theory (Groff 2011). A “mechanism explanation for some
happening that perplexes us is explanatory precisely in virtue of its capacity
to enable us to understand how the parts of some system actually conspire to
produce that happening” (Waskan 2011: 393). In the words of Bogen (2005:
415), “How can it make any difference to any of this whether certain things
that did not happen would have or might have resulted if other things that
did not actually happen had happened?” Groff (2011: 309) claims that mech-
anisms are real processes that involve the exercise of causal powers in the real
world, not in logically possible counterfactual worlds. Machamer (2004: 31)
writes, “In the case of counterfactuals, the modality of the subjunctive con-
34 Process-Tracing Methods
trast is somehow supposed to warrant the necessity of the actual causal case.
But even if this were true, and I am not sure it is, there still would be no
explanation, for they would still have forsaken the process of production by
which these certain entities and activities produced such and such changes,
which was what was to be explained.” The distinction between understand-
ing parts of mechanisms as lower-level counterfactuals and the mechanistic
account can be understood by using an analogy to an integrated electronic
circuit. In a counterfactual-based analysis, we would assess whether each
part turned on or not as current passed through. The evidence of difference-
making would be the difference made to the part by turning the current on
or not. In contrast, in a mechanistic account, while there is a counterfactual
“difference” overall between current on and off states, we are interested in
tracing how the current moved through each part of the system. Here, the
circuit would be assessed as a holistic system, using the circuit schematic to
understand the step-by-step transference of current from input to output.
Engineers “tracing” this system in detail might then look for traces in the
form of heat or photons to determine whether current leakage occurs in
particular parts, helping them design more efficient and robust circuits or
to understand where the circuit failed. In plain English, learning that some-
thing could have been different—that is, identifying a causal effect—is not
the same thing as learning about how a process works in an actual case.
The counterfactual-based account on which an intervening-variable un-
derstanding of mechanisms builds has the result of transforming the within-
case tracing of causal mechanisms into a cross-case analysis of patterns of
variation, just at a lower level of aggregation. But by studying difference-
making by assessing the effect of variation of intervening variables between
a cause and outcome, we lose focus on the process itself that links the cause
and outcome (Bogen 2005: 415; Dowe 2011; Illari 2011; Mayntz 2004: 244–
45; Waldner 2012). This means that if we treat mechanisms as intervening
variables, we are no longer studying the causal links in the theory—that is,
how the process worked within a case. Yet this was precisely the reason we
wanted to trace the causal mechanism using process-tracing!
But our arguments here do not mean that mechanisms cannot be re-
searched using an intervening-variable approach. Indeed, there are excellent
methodological guidelines for designing research on mechanisms defined
as intervening variables (e.g., Weller and Barnes 2014). But understand-
ing mechanisms as intervening variables is not compatible with the goal of
learning how processes actually work within cases because studying interven-
ing variables by definition means that we need to explore the difference that
36 Process-Tracing Methods
mechanism scheme would include more parts (e.g., Cause → Part 1 → Part 2
→ Outcome), but the causal logics linking parts together are not described,
but instead are merely depicted as arrows that are assumed to link parts
together in a relationship of conditional dependence. In both instances, the
theorized minimalist mechanism does not provide enough information to
answer fully the “how does it work” question (Craver and Darden 2013:
90–91).
An example of a theorized “minimalist” mechanism can be found in Nina
Tannenwald’s 1999 article on the impact of norms on U.S. decision-making.2
She theorizes that norms against the nonuse of A-weapons (cause) (a nu-
clear “taboo”) contributed to U.S. decision-makers’ avoidance of using those
weapons (outcome), but the mechanism remains firmly within a theoretical
gray-box because no mechanism is detailed to link the cause to the outcome.
The closest she gets to unwrapping causal mechanisms is in the conclusion,
where she mentions three plausible links between norms and nonuse in the
form of minimalist “one-liners”: constraints imposed by individual decision-
makers’ personal moral convictions or domestic or world opinion (Tannen-
wald 1999: 462; see also Tannenwald 2007: 47–51). Yet these brief descrip-
tions do not explain the causal process that links norms with nonuse; that
is, how does the existence of the taboo actually produce behavioral changes?
How does the process work to link individual moral convictions to nonuse
in a situation where nuclear weapons might otherwise have been used. Do
these individuals have to deploy normative speech acts to shame other actors?
Given that we cannot answer these questions, we can conclude that the actual
causal process remains minimalist. This is depicted in figure 2.2, where the
process is gray-boxed because we are not told how the process actually works.
However, Tannenwald’s research situation warranted not unpacking mecha-
nisms in any detail because there was a low prior confidence in the existence
of a causal relationship between norms and nonuse (1999: 438). But after she
found within-case evidence of a relationship, the natural follow-up would
have been to probe mechanisms in more detail.
While a minimalist mechanism is by definition an inadequate mechanis-
tic explanation because key parts of the process are not unpacked and there-
fore are not subject to theoretical and empirical scrutiny, this is a deliberate
choice determined by the research situation. Of course, the analytical result
of not unpacking the mechanism in any detail theoretically means that the
2. For a more in-depth discussion of Tannenwald’s research design, including the theorized
mechanism and the evidence presented in her research, see chapter 7.
At its core, the systems understanding means that the core elements of the
causal mechanisms are unpacked theoretically and studied empirically in
38 Process-Tracing Methods
the form of the traces that the activities associated with parts of the process
leave within cases. Mechanisms in this understanding are defined as systems
of interlocking parts that transmit powers or forces between a cause (or a
set of causes) to an outcome (see, e.g., Beach and Pedersen 2016a; Bhaskar
1978; Bunge 1997, 2004; Glennan 1996, 2002; Machamer 2004; Machamer,
Darden, and Craver 2000; Mayntz 2004; Rohlfing 2012; Waldner 2012). In
the systems understanding, we are operating at a lower level of analytical
abstraction because we are trying to capture how actual causal processes play
out within cases. The level of abstraction can range from very detailed, case-
specific models of process to more abstract processes that can in theory be
present within a bounded population of cases.
The ambition is to unpack explicitly the causal process that occurs be-
tween a cause (or set of causes) and an outcome and trace each of its con-
stituent parts empirically. Here the goal is to dig deeper into how things
work by tracing each part of the mechanism empirically using mechanistic
evidence. In particular, observing the empirical fingerprints left by the activ-
ities of entities in each part of the process should enable us to make stronger
inferences about how causal processes actually worked in real-world cases
(Illari 2011; Russo and Williamson 2007). In comparison, in the minimalist
understanding, we have less direct mechanistic evidence because we have
not explicitly detailed the process, resulting in weaker inferences about the
operation of a causal process.
When mechanistic explanations are viewed as systems, they are under-
stood in a holistic way, with the effects of the mechanism more than the
sum of its parts (Sawyer 2004). Parts have no independent existence (i.e.,
they are not variables) in relation to producing an outcome; instead, they
are integral elements of a system that transmits causal forces to the outcome.
In a system, a complex interrelationship often exists between the parts of the
mechanisms, with the effects of the individual parts only manifesting them-
selves fully in conjunction with the effects of other parts. In the words of
Cartwright (2007: 239), “There are any number of systems whose principals
cannot be changed one at a time without either destroying the system or
changing it into a system of a different kind.” This means that a mechanism-
as-system explanation cannot be reduced to counterfactual dependencies.
Each of the parts of the mechanism can be described in terms of entities
that engage in activities (Machamer 2004; Machamer, Darden, and Craver
2000). Entities are the factors (actors, organizations, or structures) that en-
gage in activities, whereas the activities are the manifestations of the causal
powers of the entities in each part (Kaiser 2017; Piccinini 2017)—in other
40 Process-Tracing Methods
42 Process-Tracing Methods
44 Process-Tracing Methods
Singular causation (also termed token causation) means that claims about
a causal relationship in a particular case can be made, whereas the regularity
position holds that causation by definition requires that the individual event
is subsumed under a more general relationship between cause and effects.
Regular associations can also be termed type causation.
In many respects, positioning in this debate is determined by founda-
tional assumptions regarding the nature of science itself (Groff 2011; Jackson
2016). Whereas neopositivism is typically associated with a neo-Humean
regularity understanding (e.g., Hume 1975), pragmatists and analyticists fa-
vor singular causation because they believe that real-world events are too
complex and multifaceted to compare—as in, the causes of the start of
World War I. The critical or philosophical realist school has adherents to
both singular and regularity understandings of causation, although regu-
larity tends to come in the form of contingent generalization within very
bounded populations (e.g., midrange theories).
In the regularity understanding, for a relationship to be causal it must
occur more than once. In the philosophy of science, this position is asso-
ciated with David Hume. The neo-Humean understanding of causality as
nothing more than patterns of regular empirical association has traditionally
prevailed in social science (Brady 2008; Jackson 2016; Kurki 2008: 33–59).
Hume’s claim about causation as regularity can be understood by using the
example of a pen falling to the ground. We can observe that the pen falls
to the ground, but we cannot observe the gravitational forces that caused
the object to fall.3 Given this inability to empirically observe the forces that
link the cause with the outcome for most phenomena in the natural world
(at the time), Hume argued that we should merely define causes in terms
of constant conjunction (correlations) between factors: any theorization of
“undetectable” mechanisms would quickly, in his opinion, degenerate into
metaphysics. Causation is taken to involve the regular association between
C and O, controlled for other relevant possible causes (Chalmers 1999: 214;
Holland 1986; Marini and Singer 1988). Adherents of regularity deny that
causal claims can be made about individual outcomes. This is can be seen
in King, Keohane, and Verba’s (1994: 81–82) definition of mean causal ef-
fects, where causation is defined in regularity terms when they write, “The
difference between the systematic component of observations made when
the explanatory variable takes one value and the systematic component of
comparable observations when the explanatory variables takes on another
3. At least until very recently. See Castelvecchi and Witze 2016.
46 Process-Tracing Methods
nomena in the philosophical debate (e.g., Andersen 2012; Hedström and Yk-
likowski 2010) (i.e., they are midrange theories that exist within a bounded
population of cases) or whether singular causal mechanisms can apply only
in a particular case (Bogen 2005: 399; Glennan 2010, 2011; Russo and Wil-
liamson 2011; Waskan 2011).
However, the debate about whether singular mechanistic claims can be
made is actually more of an epistemic question than an ontological one. This
can be clearly seen in the arguments used for/against singular causal claims.
Arguing for regularity, Anderson (2012: 426) states that
a key task is essentially winnowing out, from all of the many entities
engaging in activities that are candidates for causally contributing to
the phenomenon, those that actually did so. Were we to provide “an
explanation” by citing all of the causal interactions that took place in
the spatiotemporal vicinity in question, we would be swamped by in-
formation, most of which would be irrelevant . . . Multiple instances
of some phenomenon allow us to judge the entities and activities that
were genuinely causally involved in the mechanism that produced
the phenomenon. Comparing different occurrences thus provides
the grounds to identify the stages that genuinely contribute to the
mechanism.
48 Process-Tracing Methods
4. For positions in this debate, see Archer 2000; Coleman 1990; Emirbayer and Mische
1998; Giddens 1984; Hedström and Swedberg 1998.
50 Process-Tracing Methods
52 Process-Tracing Methods
5. We prefer the term empirical fingerprint, but it has the same meaning as observable mani-
festation.
C h a p ter 3
3.1. Introduction
1. We use both the terms influenced and produced. Produced implies sufficiency, which
of course depends on whether the cause (or set of causes) that triggered the mechanism is
sufficient to produce the outcome. We also reserve a place for tracing mechanisms trig-
gered by nonsufficient but contributing causes; influence denotes a nonsufficient but causal
contribution.
53
54 Process-Tracing Methods
imalist process-tracing case study in a broader research design. But it has the
consequence that the mechanistic explanation remains within a gray box.
In contrast, the goal of the systems understanding is to formulate a full-
fledged mechanism-based explanation, where the theorized mechanism is
unpacked into its constituent parts in enough detail to explain the work-
ings of the process between the causes and the outcomes (e.g., Cartwright
1999; Craver and Darden 2013; Illari and Russo 2014; Machamer, Craver,
and Darden 2000; Russo and Williamson 2007).
The chapter then turns to the issue of the level of abstraction of our theo-
ries of causal mechanisms, which is an issue when operating with unpacked
mechanisms (systems understanding). Theorized mechanisms can range
from detailed, case-specific mechanisms to abstract, midrange mechanisms
that are formulated in such a way that they can in theory be present in many
different cases within a bounded population. We discuss issues relating to
the temporal bounds of the operation of mechanisms and differences be-
tween macro-and microlevel mechanisms.
Next, we turn to the challenges of context sensitivity and mechanistic
heterogeneity. Mechanistic claims are often very sensitive to the surround-
ing context, meaning that the same causes can trigger different mechanisms
in different contexts, a phenomenon that we term mechanistic heterogeneity.
Because mechanistic heterogeneity can be produced by contextual differ-
ences, when theorizing causal mechanisms it is also important to specify
the context in which the mechanism is expected to operate as theorized. Of
course, this information might not be known before engaging in research,
which is why chapter 4 suggests engaging in extensive probing of cases across
a bounded population to develop a better understanding of the context in
which the particular mechanism can be expected to operate.
The chapter closes with a discussion of modularity and building blocks of
mechanistic explanations in the social sciences before offering a set of practi-
cal suggestions for translating cause → outcome theories into minimalist or
full-fledged mechanistic theories.
Theoretical concepts such as war or democracy are very abstract, with many
different potential meanings. Therefore, before we can move to theorizing
about mechanisms, we first must define our causes and outcomes in more
concrete terms (Adcock and Collier 2001: 533).
2. Research situations can exist where we want to study mechanisms that should have
worked but broke down during the process. In such situations, we use theoretical-revision
process-tracing, which involves studying deviant cases where the cause(s) are present but the
expected outcome did not occur to learn about omitted conditions that also must be present.
See chapter 9.
56 Process-Tracing Methods
3. For more on the distinction between homogeneity at the cause/outcome level and the
mechanistic level, see chapter 4.
58 Process-Tracing Methods
Include Only Attributes That Are Relevant for the Mechanistic Claim
4. This attribute of democracy could be measured by using the “deliberative component
index” or the more specific “respect for counterarguments” measures in the V-DEM project’s
dataset of democracy. See Coppedge et al. 2017: 53–54, 203.
5. This attribute could be measured using the “electoral democracy index” of the V-DEM
project. See Coppedge et al. 2017: 49–50.
60 Process-Tracing Methods
62 Process-Tracing Methods
Defining the positive pole also involves setting a qualitative threshold—a cat-
egorical difference in kind—between cases that are in and out of the set of the
concept. A difference in kind demarcates a change in causal properties. Cases
that are members of the positive pole of a concept are expected to have the
causal properties that our theory hypothesizes, whereas cases outside the set
have either different or no causal properties. When working with mechanistic
claims, thresholds can be understood as the point at which a given mechanism
should be expected to become operative. In other words, the threshold is a
representation of an underlying pattern of asymmetric causation.
64 Process-Tracing Methods
Mechanistic explanations are causal claims about the processes that link
together causes and outcomes. This section provides guidelines for what
“good” theorized causal mechanisms can look like in both minimalist and
systems understandings. When we are theorizing about causal mechanisms,
we are turning our attention away from causes and outcomes to focus ana-
lytically on what is going on between them.
While theorizing causal relationships in cause → outcome terms is (rel-
atively) straightforward, it is much more difficult to conceptualize causal
mechanisms, a challenge made even greater by the existence of two distinct
understandings of mechanisms in the literature (minimalist versus systems).
In the minimalist understanding, mechanisms are not unpacked in any de-
tail. In the systems understanding, the causal process is unpacked in greater
detail in an attempt to theorize a mechanism that ideally achieves what Ma-
chamer, Craver, and Darden (2000) term “productive continuity,” where
the mechanism composed of entities engaging in activities is theorized to
transmit causal forces in an uninterrupted process that links cause(s) and
outcome.
66 Process-Tracing Methods
nation effects (145). But how would learning really operate in terms of who
does what and why? What activities should we expect to see if coordination
is taking place? Because we are told little about how the process works, we
are left uncertain about the logics that provide the causal links in the pro-
cess. We are therefore unable to answer the how does it work question that is
the baseline for an adequate mechanistic explanation (Craver and Darden
2013: 90–91). The analytical result can be seen in the subsequent case studies
(Pierson 1996), where some mechanistic evidence makes plausible that fac-
tors like sunk costs and policy constraints are present, but we are left in the
dark about how the process actually works.
An incomplete mechanism scheme unpacks mechanisms slightly more by
including constituent parts of mechanisms, but the causal logics linking parts
together are still depicted as arrows that are assumed to link parts together in
a relationship of conditional dependence. However, the actual causal logic
linking parts together remains black-boxed. One good example of this type of
minimalist understanding of mechanisms is Waldner’s proposed use of causal
graphs as a technique for depicting causal mechanisms. Waldner (2014: 134,
135) depicts a causal graph as X → M1→ M2 → Y, writing,
Waldner does not offer methodological guidance for how to do this (e.g.,
Waldner 2014, 2015), while the guidelines for tracing mechanisms as systems
in this book provide a set of tools for doing so. However, Waldner’s frame-
work can be a good tool for probing the plausibility of a mechanism before it
is traced more rigorously and can also be used after in-depth process-tracing
to test the bounds of valid mechanistic generalizations.
An example of a minimalist mechanism can be seen in Ziblatt (2009).
He develops and tests a theory linking landholding inequality (C) to elec-
toral fraud (O) in Germany between 1871 and 1914, during which time “old
patterns of cooptation are recast as landed elites ‘capture’ local institutions,
thereby equipping landed elites with the institutional, coercive, and mate-
rial resources to subvert free and fair elections, even as their traditional social
power erodes” (3). The core of the article is a cross-case statistical analysis that
demonstrates a strong correlation between C and O (3–12). The cross-case
analysis is followed by a case study that aims to make the correlation more
plausible by attempting to answer the question “What was the causal process
through which this relatively abstract and distant concept of landholding
inequality actually operated on the ground leading to election fraud?” (14).
Given that the case study acts as a plausibility probe, the causal mecha-
nism is not unpacked in any detail. The closest we get to a theoretical de-
scription of the causal mechanism is where Ziblatt (2009: 14) writes that
the landed elites “exert influence indirectly via the capture of rural public
officials such as mayors, county commissioners, police officials, and election
officials, who in turn are the actors that interfere with free and fair elec-
tions.” This can be depicted as a causal graph: landholding inequality (cause)
→ landed elites capture rural public officials (part 1) → officials interfere in
elections (part 2) → electoral fraud (outcome).We are told earlier that this
capture occurs through deployment of institutional, coercive, and material
resources as well as that social resources, while on the decline, might also
matter. But the causal logics binding the cause with part 1, part 1 with part 2,
and part 2 with the outcome remain black-boxed because we are not able to
answer questions such as What types of power resources do landed elites deploy
to capture officials? Does capture occur through the use of material resources such
as the power to control revenue or through control of appointment processes? Or
perhaps by deploying more institutional power resources? Do landed elites have to
actively intervene to capture officials, or do officials anticipate what local officials
want? And once captured, what is the process whereby local officials actually en-
gage in electoral fraud? What types of actions do they use? Removal of voters from
electoral rolls, pressuring poll officers, and so on?
68 Process-Tracing Methods
6. In Aviles and Reed’s (2017: 717–18) terms for mechanisms in sociology, we are operat-
ing with a “substantial standard” of what constitutes a mechanistic explanation, where the
70 Process-Tracing Methods
mechanism works and the time span of its operation. Entities should be
formulated as nouns, but they should be described as things that can have
properties, structures, or orientations that enable them to engage in activi-
ties. They can be identified by their properties and spatiotemporal location.
For example, in social science mechanisms, common entities would be in-
dividual voters, lobbyists, or politicians or macroentities such as political
parties, social movements, or governments.
The activities in which entities engage move the mechanism from an
initial or start condition through different parts to an outcome. Activities
should be formulated as active verbs that describe what entities do in terms
of something that can transfer causal forces to the next part of the theorized
process. Examples in social science of activities could include voting, mobi-
lizing, lobbying, cajoling, demonstrating, attacking, and so on. For example,
a group of voters can vote for a politician, which might then lead the elected
politician to try to represent their interests, at least in theory.
Figure 3.2 depicts a very simple, linear mechanism. But there is noth-
ing in the logic of theorizing mechanisms as systems that tells us that they
cannot include more complicated forms. For example, we might expect a
mechanism to split into two parallel mechanisms for part of the causal pro-
cess, or feedback loops and other forms of nonlinear relationships might ex-
ist. Multiple causes might affect the process at different times. Cause 1 might
trigger the process, but cause 2 might be required to take the process further.
For example, the initial parts of a theorized mechanism of democratization
72 Process-Tracing Methods
Fig. 3.3. Causal mechanism linking policy dispute among states with policy
adoption
Source: Based on O’Mahoney 2017.
74 Process-Tracing Methods
can range from case-specific causal mechanisms, which are highly specified
mechanisms that describe how a causal process works in a particular case;
over contingent mechanisms, which can be present in at least two cases; to
midrange mechanisms, which are processes that are described in quite ab-
stract terms but that still identify interlocking parts of the process in terms
of entities engaging in activities. Entities and activities in a case-specific
mechanism will typically be described using proper nouns (e.g., particular
people or institutions) and the specific activities in which they engage. In
contrast, in a midrange mechanism, the parts are quite abstract, focusing
only on the most causally critical elements shared across a range of similar
cases. Logically, lifting the level of abstraction by dropping specifics expands
the potential scope of generalizations about the mechanism, whereas low-
ering the level of abstraction involves further specification of parts of the
process, thereby narrowing the scope of potential generalization.
But in all instances, to qualify as a mechanism-based explanation in a sys-
tems understanding, we still have to answer the How does it work? question,
which requires that the causal arrow between causes and outcome be eluci-
dated in enough detail that the critical parts of the process are made clear in
terms of entities engaging in activities that link one part to the next (Craver
and Darden 2013: 31; Hedström and Ylikoski 2010: 53). We suggest thinking
about raising the level of abstraction using an analogy of reduction of sauces
in cooking, where the goal is to reduce the amount of liquid while preserv-
ing and intensifying the flavor. When theorizing mechanisms, we can reduce
a case-specific mechanism by cooking out the liquid (case-specific parts) into
the essence that still captures the key working parts of the process.
To exemplify what a mechanism can look like as we raise/lower lev-
els of analytical abstraction, we utilize a case-specific mechanism de-
scribed in Haggard and Kaufman (2016). The authors provide a set of
case studies that can be interpreted as descriptions of very case-specific
mechanisms linking repressive autocratic regimes and economic griev-
ances (causes) with democratic transition (outcome). We ignore for the
moment the fact that the case studies sometimes veer toward narratives
of events where the causal logics linking parts together become foggy.7
7. Haggard and Kaufman put forward no real evidence to substantiate most of the key
causal links in their case studies. For example, in the Argentine case, they write, “But the
decision to invade was itself a major gamble that was directly precipitated by growing pres-
sure from below and the increasing internal rifts within the regime itself that resulted from
the surge in mass mobilization” (2016: 112–13). But they do not provide the evidence to
We also ignore the fact that the mechanisms described in different cases
differ widely (see chapter 4).
Instead, we utilize the analytical narrative in the Argentine case to flesh
out a case-specific mechanism and then demonstrate what it could then look
like if we raise the level of abstraction to both a less detailed but still contin-
gent mechanism and an abstract midrange mechanism that could travel to a
range of contextually similar cases.
Table 3.1 describes a fourteen-part collective-action mechanism based on
the narrative provided by Haggard and Kaufman (2016: 111–13). If we were
to lift the level of abstraction to a midrange mechanism, we would want
to ask ourselves what elements of this process are case-specific and what in
theory could be present in other democratic transitions in a similar context
of a labor-repressive regime with few channels of representation, a history of
union involvement in politics, and repeated economic crises.
Here it is obvious that the military invasion of the Falkland/Malvinas
Islands and subsequent lost war is a case-specific factor. This detail of the
Argentine case could be abstracted away, putting it down into a category of
severe grievances with the regime that undermines its support with critical
constituencies. In addition, there is a repeated cycle of mass mobilization
followed by repression that then triggers further protests. While these partic-
ular events in the Argentine case and the postulated causal linkages in them
are case-specific, we could theorize that a midrange mechanism might have
to go through multiple iterations of the protest-repression cycle before the
threshold is reached to push toward democratic transition. This can be mod-
eled as a feedback loop that repeats several times. Which particular groups
eventually support the unions beyond human rights organizations is prob-
ably case-specific, but the critical point is that unions act as the spearhead
for broader middle-class mobilization. In other cases, this feedback cycle will
have a different face, but the key testable feature would be that we find an
iterated cycle where union protests and demonstrations are eventually sup-
ported by a broader middle-class coalition, resulting in a level of protest at
which further repression is no longer feasible, leading to the regime to accept
a transition to democracy.
Figure 3.4 depicts a more abstract process in which the essential elements
of the Argentine case-specific mechanism are reduced into three critical
back up this postulated causal link between the invasion and the pressures created by mass
mobilization.
Beach_2d-edn.indd 76
Cause
Autocratic Unions initially Military repres- More combative In midst of deep Unions respond Strikes deepen Military com-
repression AND divided over sion hardens factions organize recession in by increasing divisions be- mand replaces
economic whether to resolve and wave of strikes 1981, military pressure on re- tween hard-line Viola with
grievances negotiate or arrests of older beginning in high command gime, July 1981 officers and hard-line officer
oppose conservatives 1977 puts in power strikes more moderate (Galtieri)
clears way for more moderate president
more combat- leader (Viola)
ive younger
militants
Outcome
Militant opposi- Government Protests surge Army and Question of Formal proposal Military govern- Transition to de-
tion to regime launches inva- as information navy withdraw amnesty for for legal amnesty ment abandons mocracy in 1983
increases, March sion of British- about losses representatives military crimes withdrawn in attempts, allows
1982 largest held Falkland spreads and from ruling sticking point in face of massive transition to
demonstrations Islands to divert economy dete- junta, army negotiations opposition rally competitive
yet, now includ- popular pres- riorates command ap- elections to go
ing the support sures and rally points caretaker forward without
of human rights military and government preconditions in
organizations society around (Bignone) to or- October 1983
and political nationalism ganize elections
parties in negotiation
with opposition
coalition
Source: Haggard and Kaufman 2016: 111–13.
9/28/2018 1:08:26 PM
Revised Pages
parts: the mass mobilization (part 1), and the iterative sequence of repression
and protest between ruling elites and union opposition (parts 2 and 3) that
finally results in democratic transition. In Haggard and Kaufman (2016),
this mechanism is then applicable in the case of Bolivia in 1982, where there
is a similar context although in the Bolivian case it includes elements such
as church support of protests that are not theorized to matter in the Argen-
tine case (113–16). If we find evidence suggesting that we cannot understand
this process in the Bolivian case without including church support, and we
cannot categorize the church as “other societal groups,” we would conclude
that the midrange mechanism developed in the Argentine case does not even
travel to the Bolivian case. This collective-action mechanism does not travel
to other cases in the book, like Niger in 1985, where the role of unions and
economic crisis is replaced by another dynamic involving a weaker regime
and unions of state workers mobilizing within the state instead of the soci-
etal mass mobilization described in Argentina and Bolivia (see chapter 4).
78 Process-Tracing Methods
The literature on causal mechanisms frequently points out that the ways
mechanisms unfold in a specific case are sensitive to the surrounding con-
text (Bunge 1997; Falleti and Lynch 2009; Gerring 2010; Goertz and Ma-
honey 2009; Steel 2004, 2008; Tilly 1995: 1601). Contextual conditions are
sometimes termed “scope” conditions in the literature, but the terms have
the same meaning. We understand contextual conditions as all “relevant as-
pects of a setting (analytical, temporal, spatial, or institutional)” in which
the analysis is embedded and that might have an impact on the constitutive
parts of a mechanism (Falleti and Lynch 2009: 1152; similarly, Mahoney and
Goertz 2004, 660–61).
Contextual conditions can sometimes be difficult to distinguish from
causal conditions. It is relatively easy to distinguish the two when we are
operating with causal theories where C and O are linked together with a
causal mechanism. When operating with mechanism-based understandings
of causation, a cause is defined as something that triggers a mechanism,
meaning it is in a productive relationship with the outcome. Here a cause
does something, whereas a contextual condition is merely an enabler or in-
hibitor; it does not do anything active. For example, if we conceptualize a car
as being a mechanism that transfers causal forces from a cause (the burning
of fuel) to the outcome (forward movement), we might theorize that the
contextual conditions in which the car mechanism can be expected to oper-
ate include the presence of oxygen and relatively level ground. If we throw
the car mechanism into a lake, even though the mechanism might be in
perfect shape, it will not work, as it is outside of the contextual conditions in
which it will run. But the presence of oxygen or ground does not actually do
anything in a causal sense; only the absence of these contextual conditions
can prevent C from producing O.
Even when the same causes and outcome are present, different contex-
tual conditions can trigger differences in the mechanisms linking them to-
gether (Falleti and Lynch 2009; Steel 2008). Context sensitivity of mecha-
nisms means that two cases that look causally homogeneous at the level of
conditions (same causes, same outcome) might be heterogeneous when we
move to the mechanism level. Mechanistic heterogeneity can both mean
that in two or more cases, the same cause triggers different processes, thereby
resulting in different outcomes, or it can mean that the same cause is linked
to the same outcome through different processes. Different processes can be
either a whole mechanism that is completely different or only one or more
parts that diverge between two cases.
Depicted in abstract terms, mechanistic heterogeneity between the same
cause and same outcome can be illustrated as in table 3.2, where there are
two scenarios: either completely different mechanisms link the same cause
and outcome in two different cases, or the mechanisms diverge at one part.
The parts that diverge are shaded gray. In both situations, a contextual differ-
ence exists between the two cases that produces mechanistic heterogeneity.
In situation 1, the same cause and outcome are linked through completely
different processes (parts 1a and 2a), whereas in situation 2, the two cases
diverge only in part 2. But inferring that the same process operated in both
cases based on studying only one case would be a faulty inference because of
mechanistic heterogeneity.
A real-world example of mechanistic heterogeneity found by policy evalua-
tion produced by differences in context can be found in Howard White (2009).
He describes a mechanism that links a cause (World Bank funding of educa-
tion of mothers in nutrition) with an outcome (improved nutritional outcomes
for children) that was found to have worked in a case (the Tamil Nadu Inte-
grated Nutrition Project in India). The unpacked mechanism can be described
as: Cause (mother participates in program) → (1) mother receives nutritional
counseling → (2) exposure results in knowledge acquisition by mother → (3)
knowledge used by mother to change child nutrition → Outcome (improved
nutritional outcomes for children) (4–5). Based on the program’s success in
Tamil Nadu, the World Bank tried it in Bangladesh. However, the mechanism
did not function as expected in the different context and instead broke down.
The reason was a key contextual difference. In Bangladesh, mothers were not the
key decision-makers in households, with men doing the shopping and mothers-
in-law in joint households (a sizable minority) acting as decision-makers about
what food went onto the table. The mechanism therefore worked until part 3,
but a contextual difference caused it to break down in the Bangladesh case.
The problem of mechanistic heterogeneity cannot be resolved merely
Scenario 2
Case 1 Cause → part 1 (e1 * a1) → part 2 (e2 * a2) → Outcome
→ part 2a (e2a * a2a)
Case 2 Cause → part 1 (e1 * a1) → Outcome
80 Process-Tracing Methods
Causal mechanisms vary according to the time horizon of the causal forces
that produce an outcome and the time horizon of the manifestation of the
8. Randomized controlled experiments across a number of subsets of people would also
shed light on contextual differences, although such a design would be a very onerous and
expensive because we would have to engage in experiments on a large number of different
subsets. We would need a large number because we would have no knowledge about what
contextual differences might matter. And even if we could do this, we would still not know
why these contextual differences mattered because we did not trace the operation or break-
down of the process.
82 Process-Tracing Methods
change (outcome) and contends that institutional change is not always self-
evident. Different forms of institutional change are theorized to vary de-
pending on their time span, including evolutionary change (composed of
small incremental steps along a single path), punctuated equilibrium (where
nothing happens for long periods followed by a period of relatively rapid and
profound institutional change), and punctuated evolution (which features
long periods of evolutionary change followed by a rapid period) (33–35).
It is therefore important to theorize explicitly about the time dimension
involved in the workings of a mechanism along with how an outcome mani-
fests itself. Activities in an incremental mechanism will not be dramatic inter-
ventions but instead will be small nudges whose effects accumulate over time.
A longer-term mechanism will therefore look very different empirically from
a short-term mechanism. In particular, these differences manifest themselves
in the types of empirical fingerprints that an incremental, long-term mecha-
nism will be expected to have in comparison to a short-term mechanism
(see chapter 5). Here, we should expect small, almost unnoticeable empirical
traces that will only be found if we know what we are looking for.
Incremental mechanisms are also challenging to study empirically. This
type of mechanism produces very little observable evidence until the mecha-
nism has reached a cusp, after which a very sudden development (outcome)
occurs. This type of mechanism could be mistaken for a short-term mecha-
nism, but incorrectly theorizing that the outcome resulted from a short-term
mechanism would cause us to miss the longer-term incremental process that
actually produced the outcome.
84 Process-Tracing Methods
86 Process-Tracing Methods
Theoretical Brainstorm
88 Process-Tracing Methods
C h a p ter 4
Derived from the sloppy search for causes and their location in
large events is the tendency to slight the importance of conditions
and circumstances under which the outcome occurred. When a
person constructs an account of events that stresses the overriding
importance of a few variables and simple connections between
them, he will learn a set of rigid rules that will not be a good guide
to a changing world. . . . The person will see the world as more
unchanging than it is and will learn overgeneralized lessons.
—Robert Jervis, Perceptions and Misperceptions
in International Politics (1976: 230)
4.1. Introduction
This chapter develops guidelines for selecting appropriate cases for process-
tracing and for generalizing findings about mechanisms. In-depth case stud-
ies using methods like process-tracing only enable within-case inferences
to be made about how causal mechanisms work within the studied case.
To make insights travel to other cases, process-tracing case studies need
to be nested into broader cross-case comparisons, where the studied case
is compared with other cases to enable the generalizing inference that the
mechanism(s) found in the examined case(s) should also be present in simi-
lar targeted cases.
The first task is to select cases that are appropriate for engaging in process-
tracing. If we are interested in tracing a causal mechanism linking a cause or
89
90 Process-Tracing Methods
set of causes and an outcome, we want to trace it in cases where it could have
been present, at least in theory. Tracing a nonexistent mechanism in a case
where we a priori knew it was not present tells us nothing about how the
mechanism works. Using case selection guidelines that are appropriate for
variance-based designs when tracing mechanisms using in-depth case studies
therefore creates the risk that we select analytically irrelevant cases or that we
make flawed generalizing inferences to other cases.
Selecting appropriate cases for process-tracing requires that we first map
a population of cases by scoring them according to whether the cause (or
set of causes) that is theorized to trigger a mechanism and the outcome
are present, along with contextual conditions that can plausibly affect how
the process works. Once we have done this mapping, we can distinguish
between four types of cases: (1) typical cases where the hypothesized cause,
outcome, and contextual conditions are all present; (2) deviant consistency
cases, where the known cause and contextual conditions are present but the
outcome is not present; (3) deviant coverage cases, where the cause(s) is not
present but the outcome is; and (4) irrelevant cases, where neither the cause
or outcome are present.1 In process-tracing, typical cases are used for build-
ing and testing theories about mechanisms, whereas deviant consistency
cases shed light on why a process did not work as expected—for example, an
omitted contextual condition must be present for the mechanism to work.
While relevant for other methods, deviant coverage and irrelevant cases tell
us nothing about the mechanisms linking causes and outcomes and there-
fore have limited use for process-tracing.
We then turn to the issue of generalizing about mechanisms based on
successful process-tracing of mechanisms. If we have found a mechanism in
a case, how can we generalize these findings to other similar typical cases?
Unfortunately, most of the existing literature has focused on establishing
cross-case similarity as the main criterion for generalization, thereby assum-
ing mechanistic homogeneity (different mechanisms operating in cases that
share similar causes and outcomes). There are two forms of mechanistic het-
erogeneity: (1) where the same causes trigger different processes in two or
more cases, thereby resulting in different outcomes, or (2) where the same
cause is linked to the same outcome through different processes. This chap-
ter focuses on the second form because the risks of the first variant should
be reduced through careful mapping of the population by scoring cases on
1. The terms deviant coverage and consistency were developed in the literature on QCA. See,
e.g., Schneider and Rohlfing 2013.
their values of the cause, outcome, and contextual conditions. The contex-
tual sensitivity of mechanistic explanations—in particular, when unpacked
as systems—means that mechanistic heterogeneity might be lurking under
what might look like a homogeneous set of cases at the level of causes. After
having reviewed the deficiencies in the state of the art regarding generalizing
after process-tracing, we explore potential sources of mechanistic heteroge-
neity that can result from mapping using either variance-or case-based com-
parisons. This is followed by a snowballing-outward strategy for generalizing
about mechanisms within a bounded population of cases that can reduce
the risk of flawed generalizations about mechanisms. As discussed in the
introduction, while the risk of mechanistic heterogeneity produced by the
contextual sensitivity of mechanisms reduces the scope of valid generaliza-
tions about mechanisms, the goal of cumulative research about mechanisms
should be to delineate the proper bounds within which specific mechanisms
work, enabling us to more confidently claim that the mechanism will work
as hypothesized in a particular context. The alternative of lifting the level of
abstraction of our mechanistic claims to one-liners that can be generalized
across many cases is not an attractive alternative because we learn next to
nothing about how the process works in real-world cases.
The chapter ends with two examples (appendixes 1 and 2) illustrating
that the problems of mechanistic heterogeneity discussed in a hypothetical
example in the chapter also are present in published research.
When the research goal is to say something beyond a single case, after theo-
retical concepts are defined, the next step is to map a population of cases that
are relevant for tracing mechanisms. The overall mapping of a population
should include both positive and negative cases on both the potential cause
(or set of causes) and outcome, although in process-tracing, only positive
cases on the cause(s) are relevant for in-depth tracing of mechanisms. The
mapping should enable cases to be categorized based on similarities and
differences on key conditions and the outcome. The initial mapping can
include cases that score both positive and negative on the outcome.
Mapping a population involves scoring cases on values of the cause(s),
outcome, and relevant contextual conditions. Contextual conditions can be
defined as any factor that could impact how a process works. For example,
the degree of ethnic heterogeneity in a country might impact how societal
92 Process-Tracing Methods
conflicts play out. Permissive contextual conditions are factors that must be
present for process to work, whereas inhibiting conditions prevent a given
mechanism from working. Modifying conditions are factors that impact how
a process works—for example, by shifting it from one pathway to another.
Mapping cases results in a simple dataset as depicted in table 4.1. Here,
there are four cases: two positive cases of the outcome (cases 1, 2) and two
negative cases of the outcome (cases 3, 4). If our theory was that C1 and C2
are causal conditions that together are sufficient to produce the outcome,
cases 1 and 2 would be cases in which it would be relevant to trace the
mechanism(s) linking C1 and C2 with O (positive, typical cases). Cases 1
and 2 differ on a contextual condition (C3), which potentially can impact
what mechanism links C1 and C2 with O.
There are many existing large-n datasets that may provide case scores on one
or more conditions being utilized. However, three modifications are typically
required before they can be used as the basis for mapping a population for
selecting cases for process-tracing. First, most existing datasets have already
defined concepts, which means that the attributes they include might not be
relevant for the mechanism(s) we are interested in studying.2 For example, if
we are interested in studying the mechanisms that link democratic systems
with more peaceful resolution of domestic conflicts, it would be problematic
to use an off-the-shelf dataset like the Polity IV aggregate measure of democ-
racy because it does not include the attribute civil liberties in its definition
of democracy (Marshall, Gurr, and Jaggers 2017: 14). Therefore, before exist-
94 Process-Tracing Methods
3. Considerable uncertainty can of course exist about where the exact threshold should be
set, and it can vary across contexts. For this reason, case-based scholars tend to favor selecting
cases that are obviously in—what can be termed ideal-typical cases. An example would be
Sweden in 1985 as a member of the set of countries that have welfare states. This case would be
included in the set of welfare states irrespective of how the concept is defined and how high
the threshold is set.
96 Process-Tracing Methods
Types of Cases
The result of a mapping of the population is that cases can be divided into
four types depending on whether they are members of the cause(s) and req-
uisite contextual conditions and the outcome: typical, deviant coverage,
deviant consistency, and irrelevant (Beach and Pedersen 2016a, b; Schnei-
der and Rohlfing 2013). In table 4.2, cases are divided into four quadrants
based on these types. Cases in quadrant I are “typical” cases, understood
as the cases where a priori we can expect the theorized C → O relation-
ship through the theorized mechanism to be operative. Quadrants II and IV
represent two different types of deviant cases: quadrant II is deviant cover-
age cases, where the cause(s) and/or the required contextual conditions are
not present but the outcome is; quadrant IV is deviant consistency cases,
where the cause(s) and contextual conditions are present but the outcome is
not. Cases in quadrant III are irrelevant and thus tell us nothing about the
mechanisms operative.
Disregarding cases in quadrant III as analytically irrelevant clashes with
most existing case-selection guidelines (but see Mahoney and Goertz 2004).
According to Lieberman (2005: 444), we should select cases for in-depth
study based on the size of their residuals in relation to a regression-based
analysis of X:Y correlations. Gerring and Seawright (2007: 89) make simi-
lar claims, defining typical cases of a given X:Y relationship as regression
on-liers that could include cases in quadrant III. Yet these guidelines are
relevant only when investigating the magnitude of causal effects, where
variation between cases where the cause is present and absent is key to our
ability to make inferences based on evidence of difference-making (e.g.,
Gerring 2007; King, Keohane, and Verba 1994: 142–46; Lieberman 2005;
Tarrow 2010: 249). When making asymmetric claims, as we do when study-
ing mechanisms using process-tracing, positive typical cases are the most
interesting. For example, if we are studying mechanisms that link mutual
democracy (cause) to avoidance of war (outcome), tracing a mechanism in a
negative case where the cause and outcome are not present (two nondemoc-
racies went to war) would shed light on mechanisms that lead to war but not
the mechanism (if any) between democracy and peaceful conflict resolution.
While negative (aka deviant consistency) cases can provide important infor-
mation about contextual conditions, they tell us nothing about the actual
mechanisms and how they work in positive, typical cases.
Typical Cases—Quadrant I
98 Process-Tracing Methods
4. While this guidance might seem similar to what Tarrow (2010: 251) terms “progressively
testing scope conditions,” here we are talking about within-case analyses that trace mecha-
nisms instead of using paired comparisons as a research strategy.
100 Process-Tracing Methods
different processes that lead to different outcomes, meaning that when en-
gaging in process-tracing we have to choose which process(es) to focus on
and ignore others. For example, budget support to developing countries can
produce (at least hypothetically) a range of outcomes, among them an in-
crease in public goods and services, strengthened public sector institutions,
and increased corruption if mismanaged (see Schmitt and Beach 2015).
Given this, we would not be able to select a real-world case in which only
one of these mechanisms was triggered.
Cases within quadrant II are deviant coverage cases where existing causal
theories are unable to explain the outcome—that is, they are not covered
by existing causal theories. This type of deviant case can be used to find
new causes of the outcome, although we are skeptical about whether tracing
mechanisms is the most efficient methodological tool for these purposes, in
contrast to the arguments found in the existing literature (e.g., Lieberman
2005: 443; Rohlfing 2008: 1510; Schneider and Rohlfing 2013).
The basic idea is that one traces backward from the occurrence of the
outcome to find a new cause. But determining what one is actually tracing
under these circumstances is far from straightforward. Existing suggestions
that we engage in backward-tracing build on an understanding of mecha-
nisms that sees them as mere events. Yet tracing events is not the same thing
as tracing causal mechanisms that can be understood as theoretical systems
found in multiple cases (see chapter 2). And if we have no idea about the
cause, we also have no clue about the mechanism(s) linking the mystery
cause with the outcome, meaning that we are in effect blindly groping in
the dark. We therefore suggest that a comparative design would be a much
more efficient analytical first step in detecting potential new causes (Beach
and Pedersen 2016a: 241–45). For example, we might systematically compare
two cases that are similar in all aspects except the occurrence of the out-
come (O, ~O) (i.e., a most-similar-systems design). We would then want
to know what new, undiscovered cause is different between the two cases.
After having found a potential cause using a paired comparison, it would
then be possible to engage in an exploratory case study that focuses on this
new candidate cause, attempting to discern whether the new condition is
actually causally related to the outcome. Alternatively, one could use Mill’s
method of agreement, where we look to see what condition is shared across
102 Process-Tracing Methods
a range of positive cases of the outcome to find a plausible cause (see Beach
and Pedersen 2016a).
Given that both theoretical and empirical uncertainty about the exact
threshold of membership can exist, we recommend a degree of caution
about selecting cases close to the qualitative threshold that demarcates caus-
ally relevant differences of kind. There can be theoretical and empirical rea-
sons why we are not exactly sure where the appropriate threshold is (Beach
and Pedersen 2016a; Ragin 2000: 224–25; 2008; Schneider and Wagemann
5. We denote guidelines as being “variance-based” when the comparative mapping of cases
is based on a statistical, large-n analysis or when it is based on probabilistic assumptions
about the propensity of individual cases to exhibit a particular causal relationship (most/least
likely cases). For example, Lieberman’s (2005) nested analysis approach is categorized as being
variance-based, given that case selection and generalization are based on large-n statistical
analysis.
104 Process-Tracing Methods
6. We do not discuss Weller and Barnes’s pathway-analysis approach, given that it builds
on an understanding of mechanisms as intervening variables (2014: 14) and that they explicitly
state that we cannot learn anything about mechanisms by studying a single case (42).
106 Process-Tracing Methods
regression line? Given that we are not told about the threshold at which a
mechanism is theorized to kick in, we might expect that similar mechanisms
also operated in cases 1 and 2, just to a lesser degree. Yet as discussed in chap-
ter 2, mechanism-based explanations are intrinsically asymmetric, meaning
that we should not expect the mechanism to operate in cases below the
threshold. In addition, tracing mechanisms in two cases that merely exhibit
maximum variation on values of X and Y does not inoculate us from the risk
of mechanistic heterogeneity produced by contextual differences (Zs). In a
standard regression model, these factors would be held constant to isolate
the difference that X makes on values of Y. However, this focuses our atten-
tion solely on the cross-case cause/outcome level, ignoring the potential that
the mechanisms linking X and Y might vary significantly in cases with dif-
fering Zs. Therefore, based on Lieberman’s guidelines, we might generalize
about a particular mechanism found in one high X, high Y case with small
residuals to all other cases with small residuals without exploring whether
different processes actually operate in cases with different values on the con-
trol variables. This would mean generalizing based on one or two cases to the
several hundred or thousand used in the large-n analysis. Given that Lieber-
man’s recommendations completely ignore the risk of mechanistic heteroge-
neity, we strongly suggest not using Lieberman’s guidelines when engaging
in mechanism-focused research.
7. Macartan Humphreys and Jacobs (2015: 671) recognize the potential for mechanisms to
differ across cases but provide no guidance for dealing with this problem.
8. It is not a simple proportional updating in terms of number of studied cases/size of total
population; instead, it is a more complicated multilevel Bayesian model. But the gist of the
generalization procedure is proportional.
108 Process-Tracing Methods
TABLE 4.4. Different Definitions of Most Likely/Least Likely Cases in the Literature
Eckstein 1975: 118–20 • Cases categorized based on scores on variables, but if marked change
in X has occurred and other causes remain constant, the case can also
be crucial
• Most likely = X1 strongly predicts Y but X2 found to matter (discon-
firmatory for X1)
• Least likely = X2 strongly predicts Y but X1 found to matter (confir-
matory for X1)
King, Keohane, and • Cases categorized based on case scores on pertinent variables
Verba 1994: 209 • Most likely = “if predictions of what appear to be an implausible
theory confirm with observations of a most likely observation, the
theory will not have passed a rigorous test but will have survived a
‘plausibility probe’ and may be worthy of further scrutiny”
• Least likely = case that “seems on a priori grounds unlikely to accord
with theoretical predictions—a ‘least likely’ observation—but the
theory turns out to be correct regardless”
George and Bennett • Cases categorized based on scores on variables
2005: 109–25 • Most likely = single variable X at such an extreme value that its
underlying causal mechanism, even when considered alone, should
strongly determine Y; if the predicted Y does not occur, then hypoth-
esized causal mechanism strongly impugned
• Least likely = case least likely for causal mechanism and alternative
hypotheses offer different predictions, but causal mechanism correctly
predicts Y
Gerring 2007: 115 • Cases categorized based on scores on variables, in particular the
outcome
• Most likely = “on all dimensions except the dimension of theoretical
interest, is predicted to achieve a certain outcome, and yet does not”
• Least likely = “on all dimensions except the dimension of theoretical
interest, is predicted not to achieve a certain outcome, and yet does”
Levy 2008: 12 • Cases categorized based on values of key variables or theory’s assump-
tions and scope conditions satisfied or not satisfied
• Most likely = case is likely to fit a theory but data from case does not
fit, strongest if case is least likely for alternative theory
• Least likely = case is not likely to fit a theory but data supports theory,
strongest if case is most likely for alternative theory
110 Process-Tracing Methods
and asymmetric causal terms (see chapter 2), the relevant distinction is
possible and not possible. If a mechanism is present when we did not
theoretically expect it (a least likely case), we should revise the theory
about the contextual conditions under which the mechanism is pres-
ent instead of making a strong cross-case inference that the relationship
should also work in other more-typical cases.
Second, case studies produce within-case mechanistic evidence, updat-
ing our posterior confidence in a mechanism operating within a single case.
Inferring beyond the single case to the rest of the population requires that
it is causally similar to the studied case. However, given the large differences
in the contextual conditions present in least likely and most-likely cases and
more typical cases and given the sensitivity of causal processes to contextual
conditions that are typically assumed in case-based research, we should ex-
pect that cases with different likelihoods will exhibit high degrees of mecha-
nistic heterogeneity. This means that we cannot just infer that because we
found confirming evidence of a causal mechanism in a least likely case, it
should also be present in potentially causally dissimilar cases throughout the
population (e.g., in most likely cases).
Therefore, the Sinatra inference does not hold for process-tracing of
mechanisms. Irrespective of the musical virtues of the song, a case-based
scholar would rebut that Sinatra was basically wrong in claiming that some-
one who can “make it” as a crooner in a major venue in New York would
“make it” everywhere. He ignores the importance of context for causal rela-
tionships. For example, the style of music might matter, leading us to expect
that Sinatra probably would not “make it” in a bluegrass club in Nashville
or in a Chinese opera in Beijing. The type of audience would probably also
matter, meaning that just because Sinatra made it in a New York nightclub,
we should not infer that he would also rock an audience that is hard of hear-
ing in a nursing home in Albuquerque. Case-based scholars recognize that
context matters, but when extended to the level of mechanisms, this means
that we should be very cautious in generalizing about mechanisms across
causally heterogeneous sets of cases.
We therefore recommend the use of the term typical case without any
qualifications about likelihood in relation to cases. The critical distinction
is not the propensity of a causal relationship in a case in terms of most/least
likely but instead the dichotomous question of possible/not possible—that
is, whether a causal mechanism being operative is possible or not a given
context.
112 Process-Tracing Methods
and Rohlfing 2016: 541). Schneider and Rohlfing (2016: 535–39) suggest that
for tracing mechanisms, we should start by selecting a typical case that is
not a member of other conjunctions (unique membership). In other words,
we look for typical cases that are (1) either full members in both the final
configuration and the outcome (Ci = 1; O = 1) or as close as possible to that
and (2) exclusively belong to one term. Unique membership is arguably not
required when engaging in process-tracing because we can distinguish em-
pirically between different mechanisms in operation at the evidential level.
After we have found a mechanism in the typical case, existing case-based
guidelines suggest that we can generalize to all other typical cases that belong
to the same conjunction (Schneider and Rohlfing 2013, 2016). The validity
of making such strong claims about the generalizability of process-tracing
findings is based on the assumption of mechanism homogeneity, according
to which, in the absence of model misspecification, “the same sufficient term
cannot give rise to different mechanisms” (Schneider and Rohlfing 2016:
555). Assumptions about homogeneity with respect to the causal effects on
the cross-case level are thereby extended into the sphere of within-case anal-
ysis—in other words: mechanism homogeneity is assumed to follow from
homogeneity of the cause-effect relation. This means in the context of QCA
that once causal heterogeneity across cases is delineated via identifying differ-
ent sufficient configurations, “cases of the same type are qualitatively identi-
cal and cases belonging to different types are qualitatively different,” which
is why “findings from the study of one, say, typical case, travel to all other
typical cases of the same term, but not beyond this term” (Schneider and
Rohlfing 2016: 556, see also Schneider and Rohlfing 2013: 30; Williams and
Gemperle 2017: 127).
However, given the risks of mechanistic heterogeneity, assuming homo-
geneity without exploring whether the assumption holds creates significant
risks for faulty generalizations about mechanisms.
While the assumption of causal homogeneity across cases that share similar
scores on causes and outcomes makes sense in the context of cause-effect
relations (see, e.g., Collier, Brady, and Seawright 2010: 41–43; Rohlfing 2012:
44–47), it becomes more problematic to assume mechanistic homogeneity a
priori based on homogeneity found at the cross-case level via either statisti-
cal or QCA-based analysis (B. Clarke et al. 2014: 349). In other words, how
do we know whether mechanistic heterogeneity is actually present across
cases that look similar at the level of causes and outcomes?
In many respects, it can seem strange that existing guidelines ignore the
risks of flawed generalizations created by hidden mechanistic heterogeneity.
In policy interventions, we would not blindly recommend the introduction
of a new policy based solely on evidence that it worked in one case without
exploring first whether the conditions required for the policy to work were
present in the target case. Why, then, are social scientists so willing to blindly
generalize about mechanisms without looking more closely at how the op-
eration of the process can be impacted by contextual factors?
Five potential sources of mechanistic heterogeneity can lurk in what
looks at the level of conditions/outcomes like a causally homogenous set of
cases. The first three sources are shared by variance-based (e.g., regression
results) and case-based (e.g., using QCA) mappings of populations, whereas
the final two are relevant only for case-based mapping using QCA. The five
potential sources of mechanistic heterogeneity are the product of
114 Process-Tracing Methods
It is obvious that the same cause in a different context can trigger different
mechanisms or even no mechanism (see chapter 3). Unknown conditions
are the product of exclusion of conditions before the cross-case analysis.11
Known conditions are ones that are ignored when selecting cases because the
cross-case analysis found they did not matter for the identified causal effect/
invariant association.
Variance-based approaches typically ignore the problem of omitted con-
textual conditions. In the Lieberman (2005) guidelines, cases are selected
solely on values of X and Y, holding other conditions constant. The crucial
case logic (most/least likely) deliberately works against contextual condi-
tions by arguing that if we find a process working in a context in which we
least expect it to occur, we would be much more confident that the process
is also present in more likely cases. However, this logic is flawed because it
ignores the impact of context on how mechanisms work. If a mechanism
is present when we did not theoretically expect it (a least likely case), this
should result in revision of the theory about the contextual conditions un-
der which the mechanism is present instead of enabling us to make a strong
cross-case inference that the relationship should also work in more typical
cases. For example, when studying elite decision-making, we might have
selected a least likely case for finding a process linking presidential leadership
(C) and rational decision-making (O). In a severe crisis situation, contextual
conditions like the extreme stakes involved and the short time frame for
decisions would be expected to hinder C’s production of Y. But if a least
likely case study then found that the outcome was actually rational decision-
making (O) and that a process linked the two (M1), we would not make
a stronger cross-case inference that C is linked through M1 with O across
the population of more typical cases because we found M1 where we least
expected it to function. Instead, the proper generalization would be that our
knowledge about the requisite contextual conditions needs to be updated.
To do this, we would want to delve into the studied case to understand why
our theoretical expectations about the impact of the adverse context were
confounded, probing the factors that enabled the process linking X and Y to
function and thereby enabling us to delineate more accurately the contex-
tual conditions required for M1 to function.
11. Schneider and Rohlfing (2016: 555) acknowledge this in a brief discussion of model mis-
specification as a source of potential mechanistic heterogeneity in QCA research.
12. We do not seek to criticize these works, since their primary purpose was not to in-
tegrate their analysis with further in-depth case studies or to hypothesize about the causal
mechanisms at play.
116 Process-Tracing Methods
tions are redundant across cases, we cannot automatically infer that these
conditions do not matter for the mechanisms being operative within cases.
While this notion of difference-making makes sense for the cross-case level
(e.g., Baumgartner 2009; Ragin 2008; Rohlfing and Schneider 2016), it can
easily produce mechanistic heterogeneity because minimized conditions
might be important contextual factors that matter for how the causal pro-
cesses plays out within cases. As an example, in the study on military (non-)
participation in the Iraq war, Mello (2012) finds a solution term linked to
troop deployment that involves the absence of parliamentary veto rights
and constitutional restrictions together with a right executive. According
to the QCA results, whether the center of gravity in the legislature tilts to
the right or the left of the ideological spectrum is logically redundant across
the set of cases. However, it is likely that mechanistic heterogeneity could
be present here, as it is easy to see why the actual causal processes leading to
military participation might differ in cases where a right executive is backed
by a right parliament—as in Australia or the UK—from cases where it had
to face a legislature with a more polarized or even ideologically juxtaposed
center—like in Spain.
We can illustrate the dangers of (un)known omitted conditions to mech-
anistic generalizations by turning to our hypothetical example. Table 4.5
depicts a subset of the full set of cases analyzed using the QCA (see table
4.11). Applying existing guidelines on combining process-tracing after QCA
(Schneider and Rohlfing 2013, 2016; T. Williams and Gemperle 2017) on
the cases included in table 4.11, we could have chosen case 4 as typical and
uniquely covered by the solution term C1*C2.13
TABLE 4.5. Mechanistic Heterogeneity Hidden behind Known and Unknown Omitted
Conditions
Case O C1 C2 C3 C4 UC Mechanism Operative Comment
4 1 1 1 0 0 0 C1*C2 → part 1a → studied with initial
part 2a → part 3a → O process-tracing study
5 1 1 1 0 0 1 C1*C2*UC → part 1 mechanism differs at
→ part 2 → part 3a → the final stage
part 3b → O
3 1 1 1 0 1 0 C1*C2*C4 → part 1c mechanism differs at
→ part 2a*2c → part every stage
3a*3c → O
13. In the absence of further guidelines, we could have equally selected any other most typi-
cal case—i.e., cases 1–3 in table 4.11—for in-depth process-tracing, which, of course, would
lead to the same problems.
118 Process-Tracing Methods
presidential public approval ratings (C1*C2) constitute the initial causal fac-
tors that trigger a three-part mechanism. The majority party first searches
for a legislative proposal that matches the popular president’s agenda (part
1) and then introduces a congressional bill that closely tracks the president’s
priorities (part 2). Part 3 deals with the legislative bargaining, where opposi-
tion within the majority party is whipped into supporting the party line
both through the deployment of material resources (e.g., carrots in the form
of earmarked spending in districts),14 and through anticipation of the util-
ity of a popular president for representatives’ reelection bids. Condition C2
is also a direct cause in part 3 since it ensures the silence of the opposition
party. Once the majority has its troops in line through whipping and silenc-
ing, legislation favored by the president will pass (O). The theory qualifies as
a mechanistic explanation because the causal logics underlying key parts of
the story are made explicit—for example, that material power resources have
to be deployed to whip opposing party members into line.
In table 4.5, we label unknown omitted conditions UC, while known
omitted conditions are C3 and C4, both of which were identified to be
redundant in the QCA analysis in relation to the conjunction C1*C1. C3 is
a strong level of partisan conflict, and C4 is intensive presidential involve-
ment. Case 5, for example, shares the same configuration across all known
conditions with case 4 (C1 to C4), yet the actual causal process might look
different as a result of conditions that are not included in the explanatory
model. One reasonable scenario for a UC in case 5 could be that the bill’s
substance was contested between the House and the Senate, leading to the
two chambers passing divergent policy proposals. Because bills must pass
each house with the exact same wording, resolving differences between the
House and the Senate is crucial; however, it also changes the way the mecha-
nism plays out at the last stage of the bargaining process, where part 3a
(legislative bargaining) found in case 4 would include a second step in case
5 and would involve a conference committee of members of both chambers
(part 3b).
But even if we constitute a very homogeneous population, thus keeping
the effects of potential causally important omitted conditions at bay, the
challenge of mechanistic heterogeneity might persist. As table 4.5 shows,
case 3 is an example of known omitted conditions, differing from case 4 on
condition C4. For example, it would be reasonable to expect a very differ-
14. Both parties stopped the practice of explicit earmarking in 2011. After 2011, this part of
the process worked differently—an example of mechanistic heterogeneity.
120 Process-Tracing Methods
their degree of fit to theoretical concepts (Ragin 2000; Schneider and Wage-
mann 2012). Moreover, combining several attributes into one set of a con-
cept when engaging in QCA is explicitly advocated to reduce the number
of conditions in an analysis with the goal of restricting the property space of
the truth table and consequently reducing the potential for limited diversity
(see, e.g., Ragin 2000, 321–28; Schneider and Wagemann 2012).
Yet whenever we define concepts and calibrate sets using more than one
attribute, we need to be attentive to the fact that these attributes can carry
different causal properties at the level of mechanisms. This becomes appar-
ent in a simple two-way OR combination of attributes but is even more
problematic if more complex concept-formation strategies are employed—
for example, radial concepts or indexes (Beach and Pedersen 2016a; Collier
and Levitsky 1997). The underlying logic with OR combinations is that two
attributes capture two different dimensions of the same concept but are sub-
stitutable in a causal sense (e.g., Goertz 2006, 39–46).
Multiple attributes in concepts can result in mechanistic heterogeneity
when the attributes that are equivalent from a cause-effects perspective at
the cross-case level trigger different mechanisms at the within-case level. For
example, strong left parties or strong unions can be perceived as functional
equivalent causes of the creation of an established welfare state regime (Ra-
gin 2008; Schneider and Wagemann 2012), yet it is reasonable to expect
that the processes through which the two conditions produce a welfare state
would look very different—for example, through introducing policies inside
parliament versus mobilization of public pressure and strikes outside the
parliamentary arena.
In our example, let us assume that the president’s legislative engagement
(condition C4) is defined using two attributes that tap into two different
types of involvement: the degree of direct negotiations with members of
Congress (C41) or a high degree of presidential public activity (C42), which
were combined via the logical OR, treating the two dimensions as substitut-
able legislative instruments in the president’s toolkit. In case 3, let us assume
that a positive score on C4 is achieved because the president directly negoti-
ated with members of Congress (C41). We might have found the mechanism
operative in case 3 in table 4.6.
In contrast, in case 1, a positive score on C4 is achieved through another
attribute in the OR concept (C42): a high degree of presidential public activ-
ity. Here, we should expect that the process would look quite different from
case 3. For example, instead of part 1c, we could expect the president to hold
public speeches and rallies, to using the bully pulpit to convince wavering
122 Process-Tracing Methods
states,” which might imply that there are three categorical anchor points
establishing four different types of cases—those that are fully in, partially in,
partially out, and fully out. Yet from Schneider and Rohlfing’s (2016: 555) ar-
gument that inferences about mechanisms identified in one typical case can
travel to all other typical cases irrespective of their partial set membership,
we can infer that only the categorical difference established by the 0.5 thresh-
old matters for the presence of the respective causal mechanism. Mikkelsen
(2017), conversely, introduces the notion of partial mechanisms in his treat-
ment of fuzzy-set case studies, which mimics the logic of the cross-case level.
A cautious middle position regarding the causal role of fuzzy sets can also be
found in Schneider and Wagemann (2012: 30), which states that although
the categorical threshold trumps quantitative nuances, differences of degree
nevertheless constitute real gradual differences with respect to the cases.
While using information about differences of degree in fuzzy sets can
make sense in cross-case analyses using QCA, what this means for mecha-
nisms at the within-case level remains ambiguous. From a mechanistic per-
spective, we agree that categorical thresholds in concepts are crucial in trig-
gering different mechanisms because of the change of status in the concept’s
causal properties. However, we often do not pay enough attention to the fact
that the way fuzzy sets are constructed might include more than one categor-
ical difference at the level of mechanisms. When conducting process-tracing
case studies utilizing fuzzy sets, researchers therefore must be alert to hidden
categorical differences that may change the underlying causal mechanisms
connecting a configuration and an outcome.
We can clarify the underlying challenge by revisiting our hypothetical ex-
ample as shown in table 4.7. If only the categorical difference established by
the 0.5 anchor point matters for producing the causal mechanism, we should
be able to generalize the mechanism found in case 4 to all other typical cases
belonging to the configuration C1*C2, independent of levels of fuzzy-set
scores. However, if we translate the fuzzy-set values back into what they
actually mean at a conceptual level, we might doubt whether the processes
124 Process-Tracing Methods
leading to presidential legislative success are really the same in cases that are
similar in kind but differ in degree.
In case 4, for example, a very popular president (C2 = 1) can rely on
support from an overwhelming partisan majority in both the House and
the Senate, including the sixty-vote supermajority in the latter necessary to
break a filibuster (C1 = 1), producing the mechanism described in figure 4.2.
In this process, the minority party has only limited leverage and avenues to
influence the legislative process. In the end, the president’s legislative pro-
posal passes largely unchanged.
Following existing guidelines, we would expect to find the same mecha-
nism at work in case 12, which is also a typical and unique member of con-
junctions C1*C2. But there is a difference: here, the popular president has
only a slim congressional majority (A = 0.6). It seems reasonable to expect
that this difference of degree might affect how the mechanism unfolds. For
example, we can expect the negotiations in Congress to follow a different
script, with vital consequences for the dynamics in part 3 of the mechanism.
Since the president’s party lacks a sixty-vote supermajority, interparty bar-
gaining becomes crucial to convince enough senators from the other side of
the aisle not to block the legislation (part 3e).
The problem of categorical differences lurking behind differences of de-
gree is far from hypothetical. We know, for example, from the literature in
psychometrics that categorical differences are often masked by degree differ-
ences in concepts (Mitchell 2011). In the QCA context, Hasebrouck (2015)
shows that fuzzy sets can mask categorical differences that become visible
when using crisp or multivalue sets. Going one step further, this means that
the possibility exists that other mechanisms stemming from categorical dif-
ferences link together Ci and O across various set levels. One real-world
example can be found in Samford’s (2010) study on trade liberalization in
Latin America. His case narratives based on the QCA show strikingly dif-
ferent processes at work in Peru (1990–95) and Uruguay (1972–85) although
both reflect the conjunction of hyperinflation and unconstrained executive
power. In particular, inflation plays a different causal role in the two cases, as
might be expected given that the inflation rate in Peru was at 7,481 percent
while Uruguay’s rate was only (!) 58 percent.
Against this backdrop, it becomes clear that assuming mechanistic ho-
mogeneity across cases displaying differences of degree is not the best choice.
Instead, we need to pay closer attention to the possibility that the sets em-
ployed at the cross-case level might conceal categorical differences in the
underlying concept and empirically probe whether these differences might
result in different processes at the within-case level.
15. This holds true irrespective of whether we look at configurations at complex stages—-
i.e., single-truth table rows—or at configurations at later stages after the minimization pro-
cess. Elaborate designs such as temporal QCA (Caren and Panofsky 2005; Ragin and Strand
2008), which allow for including temporal ordering among conditions in the analysis, or
coincidence analysis (Baumgartner 2009, 2013), which concentrates on causal dependencies
among conditions, cannot uncover interplay between the conditions at the level of mecha-
nisms and hence do not account for explanations at the level of causal processes.
126 Process-Tracing Methods
128 Process-Tracing Methods
Conclusion
1a 2a 3a
• president’s party • president’s party • internal bargain-
leadership searches puts forward ing by party
for proposal that legislation leadership
matches popular • silencing due to
president’s agenda strong public
support
4 1f 2a 3a Configurative dynamics
• president’s party C2 does not play role
introduces bill in part 1
as an item of a
shared partisan
agenda
(continues)
TABLE 4.10.—Continued
Differences created
Problem Part 1 Part 2 Part 3 by . . .
5 1g 2a 3g Configurative dynamics
• President’s party • internal bargain- (weakest link)
introduces bill ing by party C1 has high score
that includes leadership (large congressional
some but not all majority); C2 has low
of president’s leg- score (0.6) (president
islative priorities more popular than
not)
132 Process-Tracing Methods
identify any mechanism in other cases, which would mean that the case we
selected for process-tracing is largely idiosyncratic and that we abstain from
making generalizing inferences beyond the single case.
Empirically testing the assumption of mechanistic homogeneity signifi-
cantly increases the analytic burdens for researchers to generalize based on
their single process-tracing case studies. Yet it is better to be right about a
little than wrong about a lot. To lighten the analytical load, we suggest three
tools that can be employed—either alone or in combination—to make the
snowballing-outward strategy of tracing mechanisms in multiple cases more
feasible in research practice.
One way to reduce the analytical costs of conducting multiple process-
tracing case studies is to focus only on critical parts of the identified mech-
anism. The critical stages of a mechanism can be determined by asking
whether certain parts of the process are particularly crucial from a causal
perspective and where we have theoretical reasons to expect that the pro-
cesses might most plausibly differ across cases (Steel 2008: 88–92). If we have
theoretical and/or empirical justifications for expecting that a given mecha-
nism will always have to travel through a particularly critical stage, we can
focus our analytical attention only on this stage in additional case studies,
engaging in more focused process-tracing case studies.
An alternative means of lightening the burden of engaging in a
snowballing-outward case-selection strategy is to develop signatures of a
particular process that can be tested across a number of cases. If our initial
process-tracing case study has unpacked a causal process linking a condition
(or set of conditions) with an outcome, we can look for empirical observ-
ables in critical stages of our theorized mechanism that (1) are highly theo-
retically unique for this critical part and (2) can be expected to have similar
evidential weight in other cases. In this way, we can utilize such observables
as a form of signature of the operative mechanism as a whole in other cases
instead of empirically tracing each part of the process, using a minimalist
version of theory-testing process-tracing (see chapter 8). This can take the
form of a single observable, or a cluster of observables, i.e. engage in mini-
malist process-tracing.
Finally, researchers can explore multiple sources of mechanistic hetero-
geneity simultaneously in one analytic step to figure out what exactly causes
divergent processes across cases. In addition, it seems unlikely that a QCA
analysis or large-n statistical analysis will run into all problems simultane-
ously. Our review of existing applications of QCA and case studies in the
literature found that they experienced only some of the pitfalls. Researchers
can make the informed choice to skip certain steps if doing so can be sub-
stantiated based on theoretical and/or empirical grounds.
Table 4.12 summarizes our snowballing-outward approach and its ba-
sic logic regarding the five potential sources of mechanistic heterogeneity:
(1) known/unknown omitted contextual conditions, (2) multiple attributes
within conditions, (3) differences of degree, (4) configurational dynamics
related to the sequencing of conditions in a conjunction, and (5) different
rankings of conditions that are part of the configuration producing the out-
come. We start the stepwise exploration of mechanism homogeneity by ex-
amining the effects of potential omitted contextual conditions based on the
expectations that the risks of lurking mechanistic heterogeneity are greatest
in this scenario. Moreover, we elaborate the first step in detail and present
the procedure for the subsequent steps in an abbreviated manner since they
follow the same underlying logic. If we are operating with a mapping of
cases using a variance-based dataset,16 the data first must be translated into
set-theoretical terms by demarcating cases that are in and out of the sets of
all of the theoretical concepts.
Table 4.11 presents the set of cases whose fuzzy-set scores on C1, C2, and
O demarcate them as members of the conjunction C1*C2. In our example,
our initial case is case 4. Let us assume that our in-depth study of case 4 has
found confirming evidence for the operation of mechanism C1*C2 → part
1a → part 2a → part 3a → O (see figure 4.2). The question then becomes
how to generalize this causal inference about this mechanism to additional
cases. We should not simply assume mechanism homogeneity but instead
should test it empirically.
Given that we cannot repeat the process-tracing if there are more than
a handful of potential cases that are members of the conjunction and out-
come (table 4.11 has 31 cases), we need to strategically select cases to explore
the bounds of valid generalizations about the found mechanism. Based on
16. Statistical clustering analysis offers a particularly useful tool for mapping cases and clus-
tering them into sets that are relatively homogeneous at the level of causes. See, e.g., Ahlquist
and Breunig 2012.
2) Test for importance of configuration dynamics (e.g. temporal order of conditions) (problem 4)
using focused e.g., divide popula- subsets
comparison C1*C2*C3 → tion into subsets e.g.,
• Divide M1 → O along last found C1*C2 → M1
population into C1*C2*~C3 → condition →O
subsets based M2 → O • If most different, C2*C1 → M2
on evidence go backward to →O
of mechanistic select more simi-
1) Test for importance of known/unknown omitted conditions (problem 1)
similar case to
C1*C22 → M2 figure out where
→O mechanistic
heterogeneity
kicks in
↑ ↑ ↑
Test for impact Continue sam-
Test for OR
of difference of pling until all
relationship in
degree in most conditions tested,
conditions (con-
similar case (one or most different
tinue until all OR
difference of de- case selected
relationships in
gree in fuzzy set
conditions tested)
or composition
of condition)
↓ ↓ ↓
• Conclude that
all differences
of degree in
• Conclude that conditions do
• Conclude that difference of not matter
OR relationships degree in con- (weaker claim if
in concepts does dition does not most different
not matter matter strategy used)
If find that similar mechanism = mechanistic homogeneity
• Theory- • Theory-building
building process-tracing to
process-tracing find new mecha-
to find new nism (whole or
mechanism part)
(whole or part) • If all previous
• Divide conditions tested,
population into divide population
subsets accord- into subsets along
ing to which last difference
mechanism found
operative
5) Testing for ranking differences within conjunction (weakest-link) (problem 5)
e.g.,
C1 = 1*C2
= 0.6 → M1
→O
C1 = 0.6*C2 =
1 → M2 → O
↑ ↑
Test for impact Continue sampling
of ranking dif- until all differences
ferences within tested
conjunction in
most similar case
(smallest differ-
ence in ordering)
↓ ↓
• Conclude that
ranking differ-
ences do not
matter within • Conclude that all
conjunction ranking differenc-
• Ability to es do not matter
recode tested • Ability to recode
conditions as all conditions as
crisp set crisp set
If find that similar mechanism =
mechanistic homogeneity
138 Process-Tracing Methods
Once we have figured out what different process operates in case 5, we need
to identify the omitted contextual condition by comparing cases 4 and 5. If
the mechanisms in the two cases initially run in parallel but then diverge,
the information about divergence can be used to inform a focused compari-
son of the two cases at this point. Once the omitted condition is identified
(UC) and all cases are scored on it, the population of cases can be split into
two subsets: one where UC is not present and M1 works, and another where
M2 operates because the UC contextual condition is present. While further
division of a population of cases might pose problems for engaging in a
subsequent QCA analysis—that is, increasing problems related to limited
diversity, skewed ratio between cases displaying the outcome and nonout-
come, and so forth—this is not problematic here because we are merely
using QCA as a tool to explore the bounds within which we can make valid
generalizations about mechanisms.
If we find evidence for a similar mechanism in case 5, we can conclude
that it does not look like unknown conditions matter, although we can by
no means be 100 percent confident about mechanism homogeneity, in par-
ticular if numerous plausible contextual conditions are omitted from the
dataset. In Pennings (2003), the set of cases spanned a number of different
contextual factors such as region and time (e.g., during or after the Cold
War, recently decolonized countries, and so on) that might be expected to
affect which processes are operative. In this type of situation, testing for
potential omitted conditions should proceed in a more strategic fashion,
with multiple cases sampled across plausible differences. If we still do not
find differences, then we can conclude that there is a relatively low risk of
lurking mechanistic heterogeneity as a consequence of unknown omitted
conditions.
Testing for known omitted conditions involves selecting a case in which
there are the fewest differences (case 2 or 3). Ideally, the impact of each
known omitted condition should be tested—for example, by starting with
case 2 (testing for whether the presence of C3 matters) and then following
with case 3 (testing for whether C4 matters). If we find that there is a similar
process, we can conclude that the known omitted condition does not appear
to matter, thereby enabling mechanistic generalizations to all similar cases.
In table 4.11, if we find M1 also present in case 2, we could then generalize
with more confidence that M1 should be present in all cases similar to case 1
and case 2—that is, irrespective of whether or not C3 is present. This analysis
would then be followed by an analysis of case 3 to determine whether the
140 Process-Tracing Methods
When working with conjunctions that are found in a QCA for sufficiency,
the QCA analysis tells us nothing about whether the temporal ordering of
the conditions matters (Beach and Rohlfing 2018). Testing for temporal im-
portance can in principle be done in parallel with step 1. If C1 occurred
before C2 in case 4 and we found a different mechanism operative in case 2,
we would want to investigate whether evidence suggests that the temporal
ordering of C1 and C2 mattered. If we find either that C1 and C2 occur si-
multaneously or that C2 occurs before C1 in case 2, this could be a plausible
source of mechanistic heterogeneity that should be explored in more detail
using theory-building process-tracing together with a focused comparison
of the two cases. If evidence suggests that temporal ordering mattered for
the process triggered, we should divide the population into a subset of cases
where C1 occurs before C2 and a subset where C2 occurs before C1.
142 Process-Tracing Methods
of 1 (case 4)? If the answer is yes, we could then generalize that quantitative
differences in contextual condition also do not matter.
If we do not find a similar mechanism in the two cases, we should engage
in theory-building process-tracing to figure out what mechanism is present
in the second case. When we figure out what mechanism is operative, we
should recode the condition to incorporate this causal threshold relating to
mechanisms, reintroducing categorical distinctions into the measure. One
tool could be to recode cases using a recent methodological innovation—
what Skaaning, Gerring, and Bartusevičius (2015) term a “lexical scale.” A
lexical scale is a categorical ordinal scale where membership at each level is
determined by the presence of necessary and sufficient attributes in an ad-
ditive fashion. With this approach, each category in the scale demarcates a
causally relevant difference in kind. For example, if we found that similar
mechanisms operated in cases 4 and 9–12 but that a different mechanism
(M2) was present in case 13, we could recode condition C2 into two distinct
concepts, where cases with full membership in C2 are members of C2a and
cases with fuzzy-set scores of 0.8 or below are members of C2b, where an-
other mechanism operates. In the example of Samford (2010) , the two very
different processes in Uruguay and Peru could be accounted for by a causal
threshold relating to mechanisms lurking within the degree differences in
the condition hyperinflation. In the case of Peru, extreme hyperinflation
produced a very strong imperative for reforms. This finding could lead to
the recoding of cases into two distinct categories of hyperinflation: moderate
(e.g., Uruguay) and extreme (Peru).
144 Process-Tracing Methods
case scores on C1, C2, or O lowest matters. This could be done by testing
whether M1 is present in cases 19 (C2 and O lowest) and 23 (C1 and O low-
est). If we find that M1 is present in both, we have empirically validated the
homogeneity assumption regarding this set of cases, enabling us to general-
ize cautiously to other relatively similar cases (cases 20–22, 23–26). Ideally,
we would also explore whether differences of degree in this relative ranking
matters, but when a bounded population is medium-n, this would in most
instances be practically infeasible given the large number of possibilities.
4.6. Conclusions
9/28/2018 1:08:29 PM
Revised Pages
148 Process-Tracing Methods
Fig. A.1a. A highly abstract one-liner mechanism in the Venezuela (Chávez 2) case
Source: Based on Kuehn and Trinkunas 2017.
tives, although the police rather than the military saw its prerogatives threat-
ened. Indeed, in the period before the “coup,” Correa had increased the
military budget by 91 percent between 2006 and 2009 (Rittinger and Cleary
2013: 413). In both Bolivia and Ecuador, there is little evidence suggesting
that oil rents played a causal role in triggering opposition among military
officials, as in Venezuela.
If we lower the level of abstraction in our theorization about the causal
mechanism from a minimalist mechanism to a more detailed description of
the key steps in the process, comparing Venezuela with other typical cases
suggests an even greater degree of heterogeneity at the mechanistic level. In
the Venezuela case, there was a period of growing popular protest against
the Chávez regime. The events leading up to the 2002 coup featured a cycle
of escalating protests and counterprotests, with the trigger occurring when
Chávez ordered the military to stop the protests (Kuehn and Trinkunas 2017:
871) in April 2002. Instead of obeying, military leaders rebelled by detaining
Chávez, fearing the legal and reputational costs of suppressing protests (871).
After three days, the coup collapsed because of popular opposition and mili-
tary infighting (871). The “mechanism” Kuehn and Trinkunas describe in the
Chávez 2 case is depicted in figure A.1b.
In contrast, as figure A.1c shows, in the 2009 Bolivian “coup,” the lower
level of abstraction mechanism would be one in which the military had
few grievances with the president (Eaton 2014: 1143) but in which the core
conflict was between regional elites and the indigenous groups that Morales
mobilized in support of his conflicts with the regional elites (Cyr 2015: 299–
300; Eaton 2014: 1142–45; Madrid 2013: 255; Torre 2013). The “coup” attempt
took place in this context; during a visit by Morales to a province that was
the bastion of opposition to his statist and ethnopopulist policies. Moreover,
the 2008 constitutional settlement had actually calmed the regional conflict
(Cyr 2015: 299; Eaton 2014: 1142), although the push for autonomy of gas-
rich areas was perceived as a threat to regional elites. In the run-up to the
“coup,” Bolivia did not see the same type of popular protests as occurred
in Venezuela, nor did the military actually take part in the “coup” attempt.
In the case of Ecuador (Correa 2), statism and radical foreign policies had
produced considerable upset among elites. In 2009 a decline in oil revenue
resulted in widely unpopular austerity policies. However, the actual “coup”
event resulted from protests and strikes by police officers in reaction to a
proposed new civil service law that cut wages and advancement opportuni-
ties. Correa confronted protesting policemen at a barracks, provoking an an-
gry response in which tear gas was used and a police officer allegedly kicked
the president’s knee. Correa was taken to a hospital and was rescued by
military units several hours later. Figure A.1d depicts this causal mechanism.
This example illustrates that what look like causally homogeneous cases
via a QCA-based comparison are not when we shift to the level of causal
mechanisms that actually played out. This example shows why we cannot as-
sume mechanistic homogeneity based on homogeneity at the cross-case level
150 Process-Tracing Methods
found using QCA; instead, we must test whether the assumption actually
holds at the level of mechanisms to avoid making incorrect generalizations.
Based on our preliminary process-tracing in the other cases, we suggest that
the population of positive cases should be split into different subsets—for
example, by recoding the outcome into real coups (Chávez 2) and coup-like
events (Correa 2 and Morales 1).
9/28/2018 1:08:30 PM
Revised Pages
152 Process-Tracing Methods
important factors in the case narratives. They engage in seven different case
studies out of a total population of roughly forty cases,17 providing analytical
narratives of events occurring from the start of serious disturbances to the
final democratic transition.
As discussed in chapter 3, these narratives are not theorized mechanisms
because they do not flesh out the underlying causal mechanisms18 in the
seven cases, nor do they provide evidence of causal links. Unfortunately,
Haggard and Kaufman’s (2016: 128) descriptions of the mechanisms remain
at the level of very abstract one-liners: the pathway is “credible and sustained
mass mobilization” (although they also note that the processes played out
in many different ways: “Violence and direct displacement of incumbents
did occur in some of the distributive conflict cases, but the dominant pat-
tern of exit was credible and sustained mobilization that raised the costs of
continued authoritarian rule. Elections did play a focal point in some cases,
as did political parties, but even where they did, it was typically in conjunc-
tion with the mobilization of both protest and grievance on the part of civil
society organizations”).
Using our snowballing-outward strategy and unpacking mechanisms in
more detail than one-liner “stylized arcs,” based on the Argentina case, they
might have theorized a midrange abstract mechanism as depicted in fig-
ure A.2. When we look at the empirical narrative for the Bolivian case (the
most similar case), we find a relatively similar mechanism in operation, even
though in the Bolivian case there is a potential omitted contextual condition
(church support of protests) that did not matter in the Argentine case but is
important in Bolivia.
The next step in exploring the bounds of generalization of the union
collective-action mechanism would be to probe whether it holds in the Pol-
ish case, which differs on two known conditions (as well as on unknown
omitted conditions—e.g., the impending collapse of the Soviet empire in
Central and Eastern Europe). When we look at the analytical narrative in
the Polish case, we find mass mobilization organized around the Solidarity
unions in the early 1980s that experienced repression in October 1981 and
a resurgence of this phenomenon in the wake of the easing of Soviet pres-
17. Either 36 or 42 depending on the dataset used. See Haggard and Kaufman 2016: 45.
18. Haggard and Kaufman (2016: 128) claim that they “used these cases to trace causal
mechanisms more closely, and also to take note of several other causal factors that operated.”
As discussed in chapter 3, the narratives are not tracing mechanisms but can be translated into
mechanisms by cooking out the essence of the causal links.
154 Process-Tracing Methods
similar mechanisms operate across cases in different contexts. Here the ad-
age of “the more one knows, the less one knows” is very applicable. The goal
in process-tracing is to get to know our cases, but the more we learn about
how processes play out, the harder it becomes to make sweeping generaliza-
tions about mechanisms across sets of cases. Instead, serious process-tracing
is compatible only with more bounded generalizations about mechanisms.
C h a p ter 5
5.1. Introduction
155
156 Process-Tracing Methods
158 Process-Tracing Methods
Based on this logic, we suggest that we need to ask three critical questions
when trying to figure out what evidence in theory can tell us about the op-
eration of a mechanism: (1) What is our prior confidence in the mechanism
(overall and for particular parts)? (2) Do we have to find the posited observ-
ables (theoretical certainty)? (3) If we find them, are there alternative expla-
nations for finding the observables other than the mechanism or part was
operating as theorized (theoretical uniqueness)? Ideally, if we have clearly
theorized activities that enable us to formulate theoretically unique propo-
sitions about observables, we can collect relatively direct evidence of the
operation of a mechanism (or a part). In contrast, if we can formulate only
theoretically certain but not unique propositions, we only have more cir-
cumstantial evidence. If we do not find the evidence, we would disconfirm
the theorized mechanism (or part) to some degree; if we do find it, we would
not be much more confident that the mechanism actually was operating.
Our framework uses Bayesian reasoning in an informal fashion, as the
logical foundation for the questions we need to ask of empirical material to
transform it into evidence of causal relationships. The informal use of Bayes-
ian logic enables scholars to focus on what matters most in case studies—
learning about how a causal relationship works (or does not work) by under-
standing what particular pieces of empirical material mean in the context of
a particular case. In place of unnecessarily complicated analytical techniques
that divert attention from this contextual interpretation, we develop a set
of guidelines that analysts can use to ask the right questions when assessing
what their collected empirical material can demonstrate in a given research
context, thus enabling the interrogation of empirical material in a more ro-
bust and transparent fashion and consequently resulting in the production
of stronger confirming or disconfirming causal inferences about the opera-
tion of mechanisms in particular cases.
160 Process-Tracing Methods
nism (i.e., a natural experiment). If we find that the absence of the mecha-
nism (or a part of it) results in the absence of the outcome, we can infer that
the presence of the mechanism produces a difference (Runhardt 2015: 1304).
The comparison can be enabled by finding two cases that are similar in all
respects other than the presence/absence of a mechanism (or of a part). If
we unpack a mechanism into multiple parts (in-depth process-tracing), we
would try to find cases in which all other factors are present except the part
of the mechanism being assessed, requiring strong theoretical assumptions
that no parts of the mechanism are redundant (i.e., substitutable) and about
the lack of impact of contextual conditions. We would then have to repeat
this natural experiment for each part of the mechanism!
Alternatively, we can disaggregate a single case into multiple subcases,
engaging in a “one-into-many” strategy (King, Keohane, and Verba 1994:
217–28). King, Keohane, and Verba suggest several different options2 for
disaggregating cases,3 including splitting a case up into its subunits or tem-
porally. For example, a case can be split temporally by studying a single
negotiation as a series of phases (t0 = prenegotiation, t1 = agenda-setting, t2 =
bargaining phase, and t3 = negotiation end game), or the single negotiation
can be split into subissues. In both, we transform the single case into a set
of cases that can be compared to detect patterns of cross-case variation in
an intervening variable (i.e., mechanism). The most fundamental problem
with the recommendation to transform one case into many is that we move
away from the level at which causal relationships play out. When studying
mechanisms using within-case studies, we use mechanistic evidence that is
produced by tracing how a given causal process plays out in a particular case.
However, if we transformed the case from one to many, we would no longer
have evidence at the unit level at which the causal relationship is theorized
to operate. For example, if we are studying group dynamics, disaggregation
into individuals would move the focus of the research away from the very
level at which the causal relationship is theorized to play out, producing
“cross-case” evidence of difference-making among individuals that does not
match the level of the causal relationship we intended to study.
But irrespective of whether we compare cases that were found or we have
transformed a single case into many cases, using observational evidence of
2. King, Keohane, and Verba offer two additional suggestions, but both involve shifting
the causal claim being assessed (e.g., by changing the outcome being explained). We believe
that this advice basically boils down to creating variation by comparing apples and oranges.
3. King, Keohane, and Verba use the term observation in the same way that we use the term
case in this chapter, defined as one measure of one dependent variable on one unit.
4. If we have many cases to compare, the unit homogeneity assumption logically becomes
less necessary if we are making claims about mean causal effects because potential differences
in the causal effect in individual cases “wash out,” as long as cases are independent of each
other (ensured, for example, through randomization in experiments) (Brady 2008: 261–66)
because the case-related differences will wash out across many cases if they are not systemic.
However, this is a second-best strategy. For an excellent discussion, see Gerring 2011: 246–55.
162 Process-Tracing Methods
methodological tools that are appropriate for case-based research than to im-
port variance-based ideas of assessing difference-making that result in a set
of guidelines in which case studies are always inferior to large-n comparative
research or experiments.
Similar problems occur when we try to identify two or more cases that
can be compared. As Runhardt (2015: 1306) admits, “A similarity compari-
son in areas like political science is, however, difficult to defend.” Because of
the complexity of the social world, it is difficult to find cases in which the
“all other things equal” assumption required in a natural experiment actually
holds (Ragin 1987: 48). Levy (2015: 390) writes that “Controlled comparison
and matching face the difficulty of finding real-world cases that are identi-
cal in all respects but one.” But unless we can substantiate that all other
things are equal except for the presence/absence of a cause or mechanism,
we cannot make a causal inference that its absence made a difference for the
outcome.
Another route is to utilize logical counterfactuals. Single-case counter-
factual comparisons involve comparing an existing, real-world case with a
hypothetical counterfactual case, where the logical argument is then made
that if a particular cause or mechanism had not been present, the outcome
would not have occurred (Fearon 1991; Goertz and Levy 2007; Lebow 2000–
2001; Levy 2015; Mahoney and Barrenechea 2017; Tetlock and Belkin 1996).
In effect, a single-case counterfactual comparison attempts to approximate
a most-similar-systems comparison between a real and hypothetical possible
other world. To achieve an analysis as close as possible to a most-similar-
systems test, the most important criteria is the minimal-rewrite rule, which
states that our hypothetical changes should be as minor as possible to see
whether the change can logically produce a major change in an outcome
(Tetlock and Belkin 1996). For example, gaming through the consequences
if Archduke Franz Ferdinand had not been assassinated in 1914 is an example
of a minimal rewrite that could have prevented the outcome (World War I)
from occurring (see Lebow 2007 and Schroeder 2007). But despite many
attempts to build a methodology for logical counterfactuals, there are no
objective empirical truth conditions for assessing a nonexistent “possible”
alternative world.
In addition, a minimal rewrite is a nonachievable ideal in most circum-
stances, as the absence of all but trivial conditions would have significant
knock-on effects for other causal conditions, meaning that everything else
would not be the same. In the words of one critic, “Though it is logically
defensible to think up counterfactual questions with which to confront the
historical record, the exercise seems pointless or at best of limited value from
a practical standpoint because even so-called ‘easily imagined variations’ in-
troduced into the complex matrix of historical developments can change so
many variables [conditions] in so many unpredictable or incalculable ways,
leading to so many varied and indeterminate consequences, that the proce-
dure quickly becomes useless for helping us deduce or predict an alternative
outcome” (Schroeder 2007: 149).
While this critique partially reflects a historian’s skepticism of theory-
driven social science research in general, it is almost impossible to minimize
the scope of the rewrite. However, certain types of research questions may
be more amenable to counterfactual comparisons than others. Rewrites that
involve changes in political leaders are good examples of what are claimed
to be minimal rewrites. For example, Harvey (2012) games through whether
the Iraq War would have occurred if Al Gore rather than George W. Bush
had been elected U.S. president in 2000, concluding that Gore likely would
also have gone to war. But even this minimal rewrite can be critiqued as
involving more changes than just leader. If Gore had been elected, the ar-
gument can be made that the terrorist attacks of 9/11 might have been pre-
vented because of the amount of attention the Clinton/Gore administration
was paying to the threat from al-Qaeda in the late 1990s (see, e.g., Clarke
2004). The incoming Bush administration chose instead to focus on more
traditional state-based threats such as China and Russia, with the result that
despite strong signals that an attack was imminent, the Bush administration
ignored the threat (9/11 Commission Report 2004; B. Woodward 2004). And
if 9/11 had not occurred, it is highly unlikely that a Gore administration
would have elected to go to war against Iraq. Harvey does not explore this
knock-on effect, merely assuming that 9/11 would also have occurred if Gore
had been elected president.
The challenges grow even greater when we attempt to apply the hypo-
thetical counterfactual at the level of parts of mechanisms with in-depth
process-tracing, where we would have to assess hypothetically the difference
that the absence of each part of a mechanism made for the outcome, holding
everything else equal.
At a more fundamental level, even if we could engage in meaningful
assessment of the difference-making of mechanisms through comparisons
using either real or hypothetical cases, assessing the difference that an indi-
vidual part of a mechanism makes assumes that the mechanism has no em-
bedded redundancy. Biological mechanisms have redundancy of key parts,
meaning that if we remove one part to see what happens, the mechanism
164 Process-Tracing Methods
166 Process-Tracing Methods
evidence. In plain English, how can we claim that our hoop test is evidence
when we are never told what is jumping through the hoop?
The key reason for this ambiguity in much of the literature on CPOs is
that the process linking a cause and outcome together remains in a theoreti-
cal black box. Instead, CPOs, also sometimes called “diagnostic evidence”
(Bennett and Checkel 2014), are merely produced by asking If causal mecha-
nism M exists, what observables would it leave in a case? However, the mecha-
nism is not unpacked theoretically in any detail, meaning that it is difficult
to claim that empirical material is evidence of a process since we are not told
what process it is evidence of.
For example, in Fairfield and Charman’s (2017) discussion of evidence in
process-tracing, basically any empirical material that tells us about a mecha-
nism can act as evidence. However, they provide precious little guidance for
scholars because they never tell us what the process is. The analytical result
of this lack of clarity is that, in a reconstruction of Fairfield’s (2013) analysis
that presents a set of CPOs, they claim that the CPOs beyond reasonable
doubt confirm a postulated link between legitimacy appeals (cause) and tax
reforms (outcome) that are otherwise opposed by business interests (Fair-
field and Charman 2017: appendix A).5
Fairfield and Charman (2017: 16) claim, for example, that a confessional
from a right-wing party deputy who states that “Our candidate made a com-
mitment, and it was also a difficult moment for him. Therefore the political
decision was made to support what the candidate said.” is particularly strong
evidence. In their view, this statement “captures the causal mechanism un-
derlying HEA more completely than any of the pieces of evidence previously
analyzed” (16). However, even if the confessional can be trusted because it
goes against the assumed interests of the respondent, this information at
best provides us with evidence of a causal effect (C → O) and tells us next
to nothing about the actual process that linked the two. In this example, the
confessional cannot be strong evidence of a causal process because there are
multiple potential abstract causal processes that could be compatible with
the causal hypothesis of C (appeal to legitimacy norm) → O (tax reform).
One process might be a bottom-up reaction to pressure from voters, where
the process could be appeal to norm → appeal resonates with voters →
voters contact their representatives → these representatives inform party
leadership about a surge in public opposition → party leadership decides
5. Fairfield and Charman astonishingly claim that the evidence enables them to reach a
posterior of 99.4 percent, which approaches verbally “no question about” level of confidence!
168 Process-Tracing Methods
to support reform. This appears to be what Fairfield (2013: 45) had in mind
when discussing the need to mobilize public opinion in support of a norm
that puts opponents publicly on the defensive. However, another potential
process could be more elite-driven, with party leaders anticipating a poten-
tial public reaction before it happens, meaning that the mobilization does
not need to actually take place. Indeed, the confessional appears to be more
compatible with parts of this type of anticipatory process.
Other scholars have produced narrower definitions of CPOs but have
not solved the problem of what we are observing with CPOs. Mahoney
(2012: 571), for example, describes CPOs as “diagnostic pieces of evidence—
usually understood as part of a temporal sequence of events—that have pro-
bative value in supporting or overturning conclusions.” He then shifts from
discussing CPOs as within-case evidence to discussing them as cross-case
counterfactuals that compare the existing with an “alternative universe” case.
He describes a smoking-gun test as “identifying a mechanism that is neces-
sary for the outcome. The analyst then determines whether the cause itself
is necessary for this mechanism” (581). But empirical material from the case
itself is not doing the inferential heavy lifting in enabling a claim that a cause
is necessary for the mechanism in Mahoney’s framework; rather, that role is
played by cross-case evidence that enables us to either eliminate or substanti-
ate the claim of a necessary relationship. The lack of any noticeable role for
actual within-case empirical evidence in his framework is best illustrated in
the appendix, where he discusses how three classic works use empirical tests
to make inferences. Mahoney claims that Downing’s test involves claim-
ing that a mechanism (M) that leads to the outcome (liberal democracy)
is found to be necessary “using cross-case evidence” (592). This is followed
by the claim that three causes (A, B, C) are necessary for M, substantiated
by “the fact that A, B, and C are temporally proximate to M, such that it
is easier to show that the counterfactual absence of any one of them would
have eliminated M. Once [Downing] has persuaded us that A, B, and C are
necessary for M, he then logically reasons that the three factors must also
be necessary for Y, given that M is established to be necessary for Y” (592).
Beyond the temporal sequence of occurrence of A, B, C, and M, within-case
evidence is not used. Therefore, while logically clear, Mahoney’s test types
provide us only with a counterfactual-based framework based on evidence
of difference-making, which we contend is not compatible with mechanism-
based claims as developed in chapters 2 and 3.
Blatter and Haverland (2012: 23) define CPOs as “a cluster of infor-
mation that is used (a) to determine the temporal order in which causal
170 Process-Tracing Methods
gists have not satisfactorily answered Beck’s critical questions: What exactly
does a CPO look like? And, more important, how do CPOs enable causal
inferences to be made in case studies? We now turn to answering precisely
those questions.
172 Process-Tracing Methods
Bayesian reasoning can act as the logical underpinning for using mechanistic
evidence to trace causal mechanisms in operation. At the core of Bayesian
reasoning is the idea that science is about using new evidence to update our
confidence in causal theories, either within a single case or across a popu-
lation.6 Bayesians are pragmatic about what type of empirical material can
be used to make inferences, making the approach particularly well suited
as a logical epistemological foundation for process-tracing research using
mechanistic evidence. Further, Bayesian empirical updating goes in both
directions—confirmation and disconfirmation—which is a significant im-
provement on much existing Popper-inspired social science methodology,
where falsification (disconfirmation) is claimed to be the only type of infer-
ence possible. In many respects, Bayesian reasoning makes explicit the back-
and-forth relationship between empirics and theories found in abductive
analytical tools (e.g., Tavory and Timmermans 2014).
Before we present and discuss the core elements of Bayesian reasoning,
we must address a few issues about how it can be applied to process-tracing
research to avoid misunderstandings. When working with mechanistic evi-
dence, we are asking questions such as Did we observe that a politician actu-
ally attempted to mobilize voters as the part of the mechanism would predict?
This is a piece of evidence that contains no variation within a case because it
is either present or not. However, this one piece of evidence (match or not
match) still can enable inferences about a causal link because it is an observa-
tion of an activity linking one part of a process with another.
Another common point of misunderstanding arises when we use Bayes-
ian probabilistic language in case studies when making inferences based on
empirical material while also claiming that case-based methods build on a
deterministic ontological understanding of causation (see chapter 1). Con-
fusion typically arises when we fail to distinguish between ontology and
epistemology. It is more analytically fruitful to talk about deterministic on-
tology as regards causal relationships at the case level. By putting forward
deterministic claims that can be wrong in a case or small set of cases, we
progressively refine our theories through repeated meetings with empirical re-
ality. But at the same time, our evidence-based knowledge about relationships
in the real world will always be imperfect. Determinism at the ontological
level does not logically imply that we can gain 100 percent certain empiri-
cal knowledge about why things occur. Bayesian reasoning is epistemologi-
cally probabilistic—we can never absolutely confirm or disconfirm a theory
because of the inherent complexity and ambiguity of the empirical world;
instead, we attach varying degrees of confidence to our theories based on the
6. For good popular science introductions to the Bayesian approach, see McGrayne 2011;
Silver 2013. For an exposition of how the Bayesian approach underlies medical reasoning, see
Gill, Sabin, and Schmid 2005.
174 Process-Tracing Methods
7. This can also be interpreted as the “old evidence problem.” For a more technical discus-
sion of the topic, see Wagner 2001.
Bayesian reasoning can provide us with a set of logical tools for evaluat-
ing what finding particular pieces of evidence tells us about our theories of
causal mechanisms. As used here, empirical material acts as evidence that
can either increase or decrease our confidence in the validity of a hypothesis.
Evidence in the form of mechanistic evidence enables within-case inferences
about causal mechanisms (Illari 2011; Russo and Williamson 2007).
Updating our confidence in a hypothesis about the validity of a causal
mechanism (or a part) operating in a case is a function of (1) our prior con-
fidence in the mechanism (or a part thereof ), (2) the confirming or discon-
firming power of evidence in theory (theoretical evaluation), and (3) the em-
pirical evaluation of a particular observation (on this point, see chapter 6).
After we have collected new evidence, we update our degree of confidence in
the validity of the mechanism (or a part thereof )—“posterior probability.”
We recommend the use of Bayesianism as an informal rather than for-
mal logic used to estimate quantitative values for probabilities, as others
have begun advocating (Bennett 2014; Humphrey and Jacobs 2015). There
are several reasons to avoid trying to quantify theoretical certainty and
uniqueness when evaluating the probative value of mechanistic evidence
in process-tracing, in contrast to Bayesian statistical applications. First, if
we want to quantify priors in relation to whether we should expect a causal
mechanism to hold in a particular case, our prior knowledge in a quanti-
fied form would typically be drawn from estimates about the propensity
of individual cases to exhibit a causal relationship based on population-
level research (e.g., Humphrey and Jacobs 2015: 661). Even if we ignore
problems such as conflicting ontological foundations of the causal claims
being made in case-based and variance-based designs that result in evi-
dence of different things (probabilistic versus deterministic, symmetric
versus asymmetric, and mechanisms operating within cases versus mean
176 Process-Tracing Methods
178 Process-Tracing Methods
Bayes’s Theorem
theory tell us are posterior probability, likelihood, and the prior. The full
theorem is expressed as (Howson and Urbach 2006: 21):
p(h)
p(h|e) =
p(e|~h)
p(h) + p(~h)
p(e|h)
180 Process-Tracing Methods
100 percent and disconfirmation never reaches 0, both of which are logical
impossibilities in Bayesian reasoning (Howson and Urbach 2006: 103–5).
The term p(h) is the prior, which is the researcher’s degree of confidence
in the validity of a hypothesis prior to gathering evidence, based on existing
theorization, empirical studies, and other forms of expert knowledge.
The likelihood ratio is composed of two elements: theoretical certainty
(p(e│h)) and theoretical uniqueness (p(e│~h)) of evidence. Theoretical cer-
tainty should be read here as the probability that the piece of evidence will
be found if the hypothesis is true, whereas theoretical uniqueness describes
the expected probability of finding the predicted evidence with alternative
explanations of the evidence.
Theoretical certainty relates to the disconfirming power of evidence. If
it was theoretically very certain that the evidence-generating process associ-
ated with an activity should have left an empirical fingerprint and we do
not find it in the case despite a thorough search of the empirical record, we
can significantly downgrade our confidence in the hypothesis (Sober 2009).
However, if we are engaging in an empirics-first study (i.e., we have already
found empirical material that we think might be evidence), we would not
evaluate certainty of evidence in relation to a given proposition because we
have already found the evidence! Theoretical certainty can also be thought
of as the rate of false negatives: if it is highly unlikely that we do not find
the proposed evidence, not finding it would suggest that the mechanism (or
part) is not present. In table 5.1, as the value of p(e│h) rises, the proposition
about mechanistic evidence becomes more certain, meaning more disconfir-
mation if not found.
Theoretical uniqueness describes the expected probability of finding the
observable empirical fingerprints if the mechanism (or part thereof ) does
not exist, telling us about the confirmatory power of evidence. In plain Eng-
lish, if the predicted evidence is found, can we plausibly account for it with
any alternative plausible explanation, depicted as ~h? We must evaluate the
plausibility of all alternative explanations for finding the evidence (Sober
2009). If we find the predicted evidence and it was theoretically unique, we
can make a strong confirming inference. Theoretical uniqueness can also
be understood as the rate of false positives. If it is highly unlikely that we
would find the evidence without the mechanism (or part) being operative,
then finding it is confirmatory to some degree. In table 5.1, as the value of
p(e│~h) declines, the proposition regarding mechanistic evidence becomes
more unique and confirmation increases if it is found.
This equation deals with what updating is possible if we find the posited
evidence. If we do not find the posited evidence (i.e., we find ~e), we use
another variant of the theorem. Given that p(e│~h) = 1—p(~e│~h), and
p(e│h) = 1—p(~e│h), and using the same values for p(h) and p(~h), the
formula becomes:
p(h)
p(h|~e) =
p(~e|~h)
p(h) + p(~h)
p(~e|h)
For example, let us quantify the updating possible when using Tannen-
wald’s (1999) proposition about “taboo talk” evidence. The minimalist mecha-
nism that she assesses using the proposition lies between norms of nuclear
nonuse and the actual nonuse in the Korean War, where the personal convic-
tions informed by beliefs about American values and conceptions of the ap-
propriate behavior constrain self-interested decision-makers (462). The propo-
sition about the empirical fingerprint of the process (“taboo talk”) is defined
as “non-cost-benefit-type reasoning along the lines of ‘this is simply wrong’
in and of itself (because of who we are, what our values are, ‘we just don’t do
things like this,’ ‘because it isn’t done by anyone,’ and so on)” (440).
We can quantify the prior as being relatively low for a norm-based ex-
planation (0.3—i.e., norms probably do not matter), based on the idea that
“dominant explanations are materialist” (Tannenwald 1999: 438). The prop-
osition “taboo talk” has a relatively low theoretical certainty to be found in
later cases, because when the norm is very strong, it “might become a shared
but ‘unspoken’ assumption of decision-makers” (440). However, given that
the norm was still developing in the Korean War case, we should expect to
find “taboo talk” if the norm matters because it is not yet dominant, mean-
ing that the theoretical certainty in the case might be taken to be 0.7 (“prob-
ably” present evidence). Finally, she claims that “taboo talk” is theoretically
unique: the predicted evidence “is not just ‘cheap talk,’ as realists might
imagine” (440). We might set theoretical uniqueness at 0.2 (“very probably”
we will not find the evidence with alternatives). If we then actually find
sources that suggest that the observable “taboo talk” was present in the case,
this mechanistic evidence would strengthen our confidence in the opera-
tion of the mechanism from the prior of 0.3 to the posterior of 0.6. In plain
English, we would have gone from believing that “norms probably do not
matter” to believing that it is “somewhat more likely than not that norms
matter.” If we did not find the posited evidence, we would have downgraded
our confidence from 0.3 to 0.14 (norms almost certainly did not matter).
182 Process-Tracing Methods
184 Process-Tracing Methods
case we select to study. Overall, this low prior confidence means that even if
our process-tracing case study can produce only relatively weak confirming
evidence, it would still enable some updating to take place if we found the
predicted evidence for each part of the mechanism.
The priors for each individual part determine the type and strength of
evidence necessary to further update our confidence for each part—what
can be thought of as miniature case studies embedded within the overall
study of the mechanism in the case. This means that the prior for each part
of the mechanism should be described. For example, the prior for part 2
(p(h2)) is set relatively high, given what we know about modern represen-
tative governments in general. That is, it is almost self-evident that they
would be monitoring public opinion in some fashion on important issues.
Therefore, only very strong confirming evidence would further increase our
confidence. In contrast, the low priors for parts 1 and 3 reflect our overall low
confidence in the mechanism given that these are the real “working parts” of
the causal relationship if it exists.
New empirical evidence for each individual part then enables us to in-
186 Process-Tracing Methods
cogency in the light of group dynamics” (33–34). One can infer from his
description of the state of the art of the theory and knowledge of the case
that he believes that the prior confidence in the group dynamic (group-
think) hypothesis is relatively low. Therefore, it makes sense when he then
claims that plausibility probes with relatively weak evidence are sufficient
to update prior low confidence in the hypothesis. He writes, “For purposes
of hypothesis construction—which is the stage of inquiry with which this
book is concerned—we must be willing to make some inferential leaps from
whatever historical clues we can pick up” (ix).
Scholars should be explicit about which forms of existing knowledge
have contributed to their derivation of an estimation of prior confidence
by answering three questions. First, at the most abstract level, is the postu-
lated causal relationship theoretically plausible? Do we find plausible causal
mechanisms linking C with O in the literature, or is it possible to derive
these mechanisms from existing research? Second, what does existing re-
search tell us about the relationship at the cross-case level? Given that this
evidence typically shows correlations across cases, it provides relatively weak
support for a causal relationship, especially when applied to single cases. In
addition, why should we expect that the cross-case relationship might actu-
ally be present in the selected case? That is, how can we justify the assump-
tion of causal homogeneity in the population?
Third, what does the evidence from existing research on the case itself
tell us about the plausibility of the mechanism? This answer often takes the
form of historical work on the case, which can shed light on the plausibil-
ity of the relationship in the particular case. Finally, what is the plausibility
that the theorized mechanism and each of its individual parts are present in
a particular case?
Before we can get an idea about possible empirical fingerprints that can be
left by the evidence-generating processes of the activities of a mechanism
(or its parts), we need to have some ideas about the evidential setting within
which our theorized mechanism is expected to function. The predicted evi-
dence that a given mechanism will leave depends on the mechanism and
the case being studied. A banal point is that different mechanisms will leave
different empirical fingerprints. More interesting is the claim that the same
mechanism can leave different predicted evidence in different cases. Theories
and mechanisms often have different empirical fingerprints in different cases,
despite being theorized to be a similar process. Operationalizing the empirical
fingerprints of mechanisms that are sensitive to the particulars of individual
cases therefore requires considerable case-specific knowledge and expertise. If
we are investigating a mechanism related to lobbying, while the causal process
might be the same, the fingerprints it might leave in the United States may
differ substantially from the fingerprints it would leave in an Italian context.
Gaining an overview of the evidential settings is therefore an important step
before operationalizing the theoretical expectations about the empirical fin-
gerprints, since the operationalization of expected empirical manifestations
of the underlying causal mechanism will always be case-specific.
Pedersen and Reykers (2017) utilize a process-tracing research design to
test a bandwagon-for-status hypothesis, where they aim to study whether
small states joined U.S. interventions to improve their relative status within
their geographical peer groups by gaining recognition and prestige. The au-
thors conceptualize a general status-seeking bandwagoning mechanism that
188 Process-Tracing Methods
nins ate (Cerling et al. 2013). These authors found that as vegetation patterns
shifted between three million and two million years ago toward more C4
grasses, one hominin species (Paranthropus bosei) ate a narrow, grass-based
diet, whereas early humans (Homo) ate a more varied diet. This suggests that
what differentiated us from our close ancestors was the ability to exploit and
manipulate our relationship with the environment. Creatively exploiting
different forms of empirical fingerprints as evidence has therefore resulted
in new knowledge about early human evolution. We strongly recommend
trying to be just as creative about developing propositions about empirical
evidence in the social sciences to strengthen the conclusions we can make
based on case study research.
When designing theory-first propositions, we need to state clearly which
types of evidence the causal relationship should leave in the empirical record
and what that evidence may tell us regarding a given theory (theoretical cer-
tainty and uniqueness). The clearer we are in describing what we expect to
find, the easier it is to determine whether we have actually found it. When
evaluating what evidence might tell us, it is also important to provide con-
textual information that enables us to assess the evidence’s probative value
because of the importance of case-specific context (Kreuzer 2014).
Evidence might also be in the form of what is not mentioned in the text
or in an archive. This type of evidence can be termed e silentio—based on
silence or the absence of an expected statement or message in a text or in
an archive. If we expect an author to have a strong interest in presenting an
event in a certain light or we would expect the author to take credit for a
given decision and this is not the case, this omission might have inferential
weight in our analysis. When a certain event is not mentioned in the text,
one possible explanation might be that the event did not take place. Con-
versely, if we are very confident based on other sources that an event took
place, its omission from a source where it should be is a highly unique obser-
vation that would have strong evidential weight, other things equal.
The activities associated with different types of theoretical explanations
can be expected to leave different types of empirical fingerprints. Activities
in explanations that deal with the role that ideas play in decision-making
might be expected to leave particular discursive fingerprints—for example,
a key phrase might be found in an initial document and then make its
way through successive drafts until the final text is agreed (Jacobs 2014;
Parsons 2016). Here, we would need to use methods that enable us to
observe and evaluate language, such as discourse analysis (e.g., Fairclough
1995; Schreier 2012; Schwartz-Shea and Yanow 2011). Theories about how
190 Process-Tracing Methods
192 Process-Tracing Methods
that cause C1 matters does not mean that other causes are not also pres-
ent. Further, the empirical fingerprints left by various theories often differ
greatly, meaning that evidence suggesting that a process associated with C1
matters does not necessarily imply that other causes and/or processes were
not present. For example, Tannenwald’s (1999; 2007), finding of evidence
of “taboo talk” suggests that a process associated with norms was present
(constructivist explanation of the outcome) but does not mean that processes
associated with cost-benefit calculations (rationalist explanation) played no
role. Indeed, Tannenwald (2007: 53–54) explicitly declares that that norms
are necessary but not sufficient and that cost-benefit calculations also mat-
ter. Therefore, finding confirming mechanistic evidence that suggests that
norms matter does not disconfirm a cost-benefit rationalist explanation.
When we are making inferences about parts of mechanisms using in-
depth process-tracing, the logic of rival theories completely breaks down
because it is difficult to imagine a situation in which two or more different
causes would trigger two parallel processes in which each part is mutually
exclusive of the other theoretically and empirically. Given the complex na-
ture of many of the types of causes we are studying in social science (e.g.,
authoritarian repression, electoral democracy, popular pressure from voters),
it is very likely that a given cause might trigger any number of potential
processes (see chapter 2; see also Steel 2008). Therefore, finding confirming
evidence of the activities associated with a particular mechanism triggered
by authoritarian repression does not in most circumstances enable us to dis-
confirm the existence of other mechanisms triggered by the same cause. This
means that when we are tracing mechanisms using in-depth process-tracing,
~h should be understood simply as any plausible explanation for finding the
evidence in a universe in which the hypothesized part of the mechanism is
not working as theorized.
We can illustrate the logic of developing predictions of relatively unique
empirical fingerprints for the activities associated with different competing
mechanisms in a situation that variance-based approaches would categorize
as overdetermination. The classic example from the philosophy of science in-
volves a person found dead after wandering in the desert, with poison in the
person’s canteen as well as a hole in the canteen. Therefore, both poison (C1)
and thirst (C2) are potentially sufficient causes, but only one can actually
have had a causal relationship with the outcome—that is, they are mutually
exclusive because the person died either of poisoning or of dehydration.
Overdetermination—that is, multiple sufficient causes—wreaks havoc in a
variance-based design that uses evidence of difference-making because we
cannot isolate the differences that C1 and C2 make for the outcome be-
cause both are present. But this is not a problem when engaging in in-depth
process-tracing using mechanistic evidence. Here, overdetermination at the
cause/outcome level does not mean that we cannot figure out which mecha-
nism operated because we can distinguish at the empirical level between the
fingerprints that the activities associated with mechanisms C1 and C2 would
leave. For example, we would expect to find relatively theoretically unique
mechanistic evidence in the form of poor skin turgor, tinting of skin, sunken
eyes, and/or dry galea and dry organ surfaces if the mechanisms associated
with dehydration had proceeded from start to finish (Madea and Lachen-
meier 2005), whereas the mechanisms associated with poisoning could be
detected if we find increased levels of poison in the blood in an autopsy
conducted relatively quickly after death (Musshoff et al. 2002). If we found
the mechanistic evidence predicted by the mechanism associated with poi-
soning, we could then conclude that the person died of poison and not
dehydration.
Evaluating theoretical uniqueness can be done by brainstorming about
other activities that might have left the empirical fingerprint. Here, it can be
helpful to act as one’s own devil’s advocate, drawing on other theories and
case-specific knowledge to explore additional potential explanations for the
presence of the piece of mechanistic evidence. It can also be helpful to ask
case experts about alternative interpretations of evidence.
We can also attempt to improve the uniqueness of empirical fingerprints
by increasing the specificity of the prediction. Instead of putting forward a
prediction that “taboo talk” would be found (Tannenwald 1999), the unique-
ness of this prediction could be increased by making it even more specific—
that is, by stating that central actors would be expected to be making this
type of speech act during critical phases of the decision-making process. This
more specific prediction would have a higher degree of uniqueness, other
things equal. The downside of making predictions more unique is that do-
ing so often decreases the theoretical certainty of the test, meaning that we
become less likely to find the evidence, other things equal. And not finding
unique but not certain evidence tells us little if anything about the operation
of the mechanism (or its parts).
C h a p ter 6
6.1. Introduction
195
196 Process-Tracing Methods
Fig. 6.1. The relationship between observations and propositions about evidence
Source: Based on Sober 2009: 68.
have found the posited mechanistic evidence but it could in reality be a false
positive finding (upper right corner). In this situation, the correct inference
would be to state that while we have found what looks like evidence, no
inferences should actually be made.
Empirical certainty deals with whether we can conclude that a not found
observable is absence of the posited mechanistic evidence. If we do not find
observables, we had extensive access to the empirical record, and/or we can
trust that sources did not cover anything up, we can then disconfirm to a con-
siderable degree the existence of the posited mechanistic evidence (evidence
of absence—lower left corner). In contrast, if we either lack good access or
cannot trust sources, we should not claim that we have not found the posited
evidence (lower right corner). Instead, no inferences should be made.
We should not conflate the empirical assessment of what found/not
found observables mean and whether we can trust the sources with the the-
oretical evaluation of what found/not found mechanistic evidence tells us
about the existence of the hypothesized activities associated with a mecha-
nism. In theory, a piece of mechanistic evidence might be certain, but if we
are unable to gain access to the empirical record that would enable us to
assess whether it was actually present, no inferences would be possible even
if we do not observe the predicted evidence. Here, absence of evidence is not
evidence of absence.
We start by discussing the challenges associated with the collection of
observations. We then discuss the different questions to ask depending on
whether we find or do not find observables. We split the discussion into ques-
tions relating to what finding/not finding observables can mean in a given
context. After we have evaluated whether the found observation reflects the
proposed fingerprint or whether not found observations occurred because we
lacked access to relevant parts of the empirical record, we must assess the sec-
ond question: Can we trust the source? Here we are assessing the accuracy
of our observations. We discuss a variety of different challenges relating to
accuracy and develop a set of questions that assist in evaluating whether we
can trust a given source. We then turn to different challenges associated with
different types of sources, including interviews and archival sources.
198 Process-Tracing Methods
200 Process-Tracing Methods
we could conclude that the theoretically certain proposition was not present
empirically, leading us to disconfirm the operation of the part of the mecha-
nism associated with the proposition.
202 Process-Tracing Methods
tion to the stream of other communications and activities within the policy-
making process? It is also important to note the circumstances surrounding
the document: Why has it been declassified or released? For what purpose?
What has not been released?
Sometimes assessing what an observable means can be relatively straight-
forward. If we had a proposition about mechanistic evidence that stated that
we would expect a leader to use bellicose rhetoric toward an opponent in
public speeches, and we find obviously bellicose language in a speech, we
can conclude with reasonable confidence that the observation matches the
proposed empirical fingerprint. However, we would still want to investigate
whether the particular speech was representative of all of the speeches or
whether extenuating circumstances can lead us to conclude that the speech
does not represent the overall pattern of discourse in the leader’s public
speeches. A “seemingly explosive document” might be followed by a “note
shortly thereafter counteracting the previous verdict” (Darnton 2017–18:
116). Even more insidious is the possibility that a published speech does not
exactly match what an actor actually said in a public speech.
Almost all sources are ambiguous to some extent. According to Munson
(1976: 74), “An expression is ambiguous when it has more than one meaning
and it is used in a situation or context in which it can be understood in at
least two different ways.” Unfortunately, many social science sources can be
deliberately ambiguous—for example, when parties in a negotiation agree
to use equivocal language that can mask disagreement on issues (Iklé 1964:
15–16; Smeets 2015).
And when we think that language in documents or interviews is unam-
biguous, it is typically because we know too little about the context. Lan-
guage and its usage are terribly complicated. Building on Austin (1970) and
Wittgenstein (1973), Schaffer (2016: 34–39) has suggested a set of helpful
questions for interpreting what language can potentially mean. For example,
are there differences in how close synonyms are used in a document and how
opposites and negations are used? This can involve looking for closely related
keywords in United Nations resolutions to see whether they are used differ-
ently, thereby shedding light on what they might mean in the context. This
investigation can be aided by asking diplomats or lawyers what particular
phrases or words are intended to mean and confronting them with subtle
differences in the usage of terms in particular situations (45–47).
In some situations, one observation can be so definitive (high empirical
uniqueness) that it alone is enough to confirm the presence of the proposi-
tion about evidence. If we are investigating a proposition dealing with lob-
204 Process-Tracing Methods
206 Process-Tracing Methods
370).1 This idea conflates the empirical evaluation of particular sources with
the theoretical evaluation of what predicted mechanistic evidence can tell us
about a hypothesized process (or its parts). In particular, reliability is usually
not a function of a causal hypothesis but instead relates to factors linked
to the source itself. There are, for example, natural cognitive limitations to
human memory, where actors do not remember that something took place
even though it did or incorrectly recall events. But even validity concerns are
not always correlated with our hypotheses about mechanisms.
An additional reason to separate theoretical and empirical evaluations
relates to the different procedures by which we can increase the probative
value of evidence by increasing either theoretical certainty/uniqueness or
empirical certainty/uniqueness. Theoretical uniqueness or certainty can be
increased only by changing predictions about which evidence will be found
or not found (e.g., uniqueness can be increased by making an expectation
about evidence more specific, such as precisely described phrases in particu-
lar documents). Empirical uniqueness can be improved not only by search-
ing for better sources that help to develop a more representative picture
of the empirical record but also by engaging in extensive source criticism,
asking critical questions in relation to particular sources (e.g., assessing po-
tential motivations for distorting accounts and corroborating with other in-
dependent sources).
Unless we explicitly discuss our confidence a given source using ques-
1. Fairfield and Charman also claim that this cannot be expressed mathematically. How-
ever, accuracy can be expressed in terms of probability (p(a)), where an unreliable measure
has a low probability of accuracy, and vice versa. Entered into Bayes’s theorem, an unreliable
measure reduces the ability of evidence to update our confidence in whether or not a hypoth-
esis is true. Howson and Urbach (2006:111) provide Bayesian reasoning for this statement.
(The probability of accuracy (p(a)) enters the calculation of the posterior probability of a hy-
pothesis through the likelihood function. In the nominator of the likelihood ratio is p(e|~h),
which logically equals p(e|~h & a)p(a) + p(e|~h & ~a)p(~a). The latter means that the prob-
ability of e being found if the hypothesis is not true when the measure is accurate is multiplied
by the probability of the accuracy of the instrument. This is then added to the product of the
probability of e being found if h is not true and the instrument is not accurate multiplied by
the probability that the instrument is not accurate. In the denominator, p(e|h) = p(e|h & ~a)
* p(~a) means that the probability of e being found if the hypothesis is true equals the prob-
ability of e being found if h is true when the measure is inaccurate times the probability of
the measure being inaccurate. This shows that a very unreliable measure increases the size of
the denominator and decreases the numerator. In plain English, if we find e but p(a) is low,
finding e does little to update our confidence in the veracity of the hypothesis, irrespective of
the theoretical weight of e (based on the likelihood ratio p(e|~h)/p(e|h)), meaning that the
posterior probability is not updated substantially.
tions that historians typically ask to evaluate veracity, we should assume that
the found/not found observable has little if any probative value. We can ask
five questions of our sources to evaluate whether they are trustworthy.
Can the source have known something firsthand about the events he or she
describes? If careful assessment of the position that a particular person held
at the time of an event demonstrates that he or she would not have been
privy to key discussions, we would conclude that the source is not an eye-
witness but instead heard the information from some other source. Further-
more, given that the source misled us about his or her role, should we really
trust anything else this source says? In both instances, our assessment of the
source’s measurement accuracy would fall considerably.
Question 2: How Many Steps Removed from the Events Is the Source?
208 Process-Tracing Methods
to which his policies are responsible for the outcome. . . . When the other’s
behavior is undesired, the actor is likely to see it as derived from internal
sources rather than as being a response to his own actions” (Jervis 1976: 343).
But it is by no means certain that the original participants produce a
record of the event. Instead, they might discuss what took place with col-
leagues or with an official record-maker after events have occurred, meaning
that if we interview the colleagues rather than the participant, the interview
would be a secondary source, two steps removed from the original activity. It
is therefore vital to discuss with interviewees the sources of the information
they provide to determine whether they are a primary source who took part
in the event or are referring to official minutes of the meeting or discussions
with actual participants, in which case the interviewee is a secondary source.
Finally, accessibility issues are endemic in the social sciences, especially
for research questions involving corruption or political violence in nondem-
ocratic systems. Therefore, the record of events that is accessible is often
even further removed from the original activities that generated empirical
fingerprints. We may not even know about primary sources closer to events
because they are classified or otherwise inaccessible to us, thereby forcing us
to rely on insiders or the information that authorities decide to make public.
We can use tools such as the dating of documents or text analysis to de-
termine which sources are primary and secondary as well as the number of
steps that our source is removed from events (Milligan 1979). For example,
if source A has the same linguistic phrasing as source B, and we can show
that the production of source A preceded the production of source B, this
suggests that source B should be treated as secondary material that builds on
source A, as it is highly unlikely that they would have used the same phras-
ing otherwise. The sources are therefore dependent on each other.
Third, what are the reasons to trust a source? Was the source competent to
observe what took place and provide an accurate account of it? For example,
can interviewees accurately recall complicated events that might have taken
210 Process-Tracing Methods
place several years earlier? Was the source able to comprehend events? For
example, in a complicated legal negotiation in a subcommittee, did a partic-
ular politician possess the requisite knowledge to understand what was going
on? If we are working with a document, we would ask ourselves what is nor-
mally included in the type of source (e.g., minutes from a cabinet meeting)?
One way to test for reliability is to use our knowledge of the case to
evaluate a particular source in relation to what we know from other sources
(in effect a form of triangulation). If we are dealing with trade statistics,
in which multiple different indicators are produced by different authori-
ties, do some sources appear to produce more reliable indicators than other
sources? In interviews, we can use questioning techniques used in criminal
interrogations, where reliability is probed by asking similar but differently
phrased questions at different times to assess whether we hear similar things.
If a respondent has an unusually clear recollection of events, we can probe
which factor has enabled the respondent to remember events so well. If, for
example, we find that the events took place on the respondent’s birthday or
on the same day that the low-level civil servant met the U.S. president, the
respondent might be more likely to actually recall events.
Can we provide justifications for the claim that the source had no significant
motives for distorting content? If bias exists (and it always does), the found/
not found observable can measure something other than the predicted ev-
idence (low empirical uniqueness or certainty), resulting in flawed infer-
ences. We must use all available knowledge to assess whether the source has
motivations for distorting content.
Bias can express itself in many different forms, depending on the type of
source. For example, particular documents may be declassified while oth-
ers remain classified, with actors wanting to release only documents that fit
their interpretation of events (Trachtenberg 2006: 157). If we study only the
available documents without taking into consideration what might appear
in other documents, the result could be a skewed picture. Further, bias can
manifest itself in what the source tells us. If we go to a prison in a well-
functioning legal system and uncritically ask convicted criminals whether
they committed the crimes for which they are incarcerated, we might be
amazed at the number of “innocent” people in prison.
Each collected observable should be evaluated relative to what is known
about the actors, their possible intentions, their interactions, and the situa-
tion in which they found themselves (Thies 2002: 357). How does the docu-
ment fit into the political system, and what is the document’s relation to
the stream of other communications and activities within the policymaking
process? Such assessments require considerable background knowledge. For
example, how does the political system work? Is anything amiss in the events
that have been uncovered? What is normally included in the type of source
(e.g., minutes of a cabinet meeting)?
Unfortunately, for most of social science research questions, actors of-
ten have strong incentives to distort content. Even seemingly innocuous
sources of data such as unemployment figures can be politically motivated
to distort information—for example, by excluding parts of the labor force to
produce lower unemployment figures. More insidious are the validity prob-
lems created by public authorities deliberately cooking the books, as seen in
the systematic underreporting of levels of political violence in authoritarian
systems.
When our sources have strong interests in distorting content, it can be
nearly impossible to evaluate with any degree of confidence the level of bias
without extensive corroboration with other independent sources. Other
things equal, we should therefore assume a much lower empirical certainty
or uniqueness when the political stakes are high unless we can provide strong
justifications for trusting a particular source.
In general, open sources such as public statements contain more poten-
tial bias, given that there can be many reasons for strategic communication
in public. This means that we should place greater trust in confidential re-
cords that will be made public only long after events have taken place, other
things equal (Khong 1992; Trachtenberg 2006: 151–53). Further, we should
trust accounts that go against the motives we would expect a given actor to
have, although this can be very difficult to establish. Historians tend to trust
interviews much less than archival documents, given that respondents “have
a real interest in getting you to see things in a certain light” (Trachtenberg
2006: 154). However, our default position regarding all sources should be
one of skepticism, and any ranking is quite arbitrary. Wohlforth (1997: 229)
warns, “Documentary records are ambiguous in their intrinsic meaning for
many reasons, not least because the historical actors who created them were
deliberately deceptive in their efforts to bring about desired results. Theories
and interpretive debates focus on precisely the information that statesmen
and diplomats face incentives to obscure.”
When assessing actors’ potential incentives actors to distort, we can ask
212 Process-Tracing Methods
ourselves whether, on the basis of what we know about the case, the par-
ticular source might have institutional or personal interests in skewing the
account. For example, if we are studying decision-making before the 2003
Iraq War, we might be very pleased with ourselves for securing an interview
with a central participant such as former secretary of state Colin Powell. Yet
if we assessed motives even superficially, we could seriously doubt Powell’s
accuracy as a witness to events, since he has strong institutional and personal
interests in producing an unflattering account of the Bush administration
and particularly former vice president Cheney and former secretary of de-
fense Rumsfeld. Powell “lost” the internal debates and consequently was
publicly humiliated when he went to the United Nations Security Council
before the invasion with dubious intelligence. His position finally became
so untenable that he left office, meaning that his revenge motive is quite
strong. Of course, if Powell were to give an account that ran against what we
expected based on all available information regarding his potential motives,
we would tend to trust it even more. Here, the Bayesian reasoning is this:
How probable is it that Powell would tell us an account that manifestly goes
against everything we know about his potential motives unless it is actually
a true account? Conversely, actors might have motives of which we are not
aware, meaning that there is a limit to our confidence in accounts that ap-
pear to go against actors’ interests.
as area experts often have a horse in the race and thus cannot be treated as
neutral, nonbiased sources.
Interviews
214 Process-Tracing Methods
Archival Material
216 Process-Tracing Methods
ridden analysis of civil rights leader Martin Luther King Jr.’s activities, “The
number one thing I’ve learned in 40 years of doing this, is just because you
see it in a top-secret document, just because someone had said it to the FBI,
doesn’t mean it’s all accurate” (cited in Phillips 2017). We therefore need to
be very careful before we assume that all documentary materials are hard
primary sources.
Archival material can provide all four types of evidence. Pattern evidence
could, for example, entail counting the number and length of documents.
Sequence evidence could be in the form of documents describing what
meetings took place over a period of time (e.g., agendas for meetings). Trace
evidence could take the form of meeting minutes, proving that a meeting
actually took place. Finally, meeting minutes could also be used as account
evidence for what took place in a meeting.
The first step in evaluating the accuracy of archival material is examining
the authenticity of a document. If something seems amiss in the author-
ship, time period, style, genre, or origin of a document, we must uncover
its past to evaluate whether the source can tell us anything. Relevant ques-
tions are: (1) Was the document produced at the time and place when an
event occurred, or was it produced later and/or away from where the events
took place? (2) Is the document what it claims to be? (3) Where and under
what circumstances was it produced? (4) Why was the document created? (5)
What would such a document be expected to tell us? Naturally, if the docu-
ment is not authentic, it has no inferential value for us.
The rest of this section deals primarily with challenges related to deal-
ing with official archives, such as foreign ministerial archives. However, for
many social science research questions, such archives may be either unin-
formative or irrelevant. The personal archives of participants can also be
relevant but can raise daunting challenges of access. And for many research
questions, there are no relevant archives, meaning that the historical record
must be traced using other types of sources.
The historical record is usually quite ambiguous (Wohlforth 1997), and
we do not suggest that social scientists enter the archives before they have
operationalized some tests of the presence/absence of each part of the hy-
pothesized causal mechanism. Finding answers in archival records is often
akin to finding a needle in a haystack. Without a clear set of theory tests to
guide the search, it can be more akin to a mission impossible—finding some
small unspecified object in the haystack.
Before we can admit observations gathered in archival material as evi-
218 Process-Tracing Methods
220 Process-Tracing Methods
Historical Scholarship
selects work produced by historians who share this bias, resulting in the
mistaken confirmation of the social scientist’s preferred theory (Lustick
1996: 606).
For example, 1950s scholarship on the start of the Cold War tended to at-
tribute the conflict to either the innate expansionistic tendencies inherent in
communist doctrine (the “neurotic bear”) or Soviet behavior as a traditional
great power (T. J. White 2000). Revisionist accounts of the 1960s then con-
tended that U.S. expansion in keeping with capitalist interests triggered the
Cold War (e.g., W. Williams 1962). During the 1970s, the “postrevisionist”
school saw the Cold War primarily as the product of misunderstandings that
could have been avoided at least to some degree (e.g., Gaddis 1972). If we are
interested in testing a liberal theory dealing with the impact of perceptions
and misperceptions, we would find supporting evidence in the postrevision-
ist school and disconfirming evidence in the previous two schools. In Bayes-
ian terms, such bias implies a form of rigging of results that undermines our
ability to update our confidence in the accuracy of a hypothesis.
Lustick’s (1996) solution to this bias problem is to first know the historio-
graphic schools and then triangulate across them. We should select histori-
cal works that are best suited to provide a critical test. If we are testing an
ideational mechanism dealing with actor perceptions, we should not choose
works from the postrevisionist school as sources of evidence.
Newspaper Sources
222 Process-Tracing Methods
C h a p ter 7
7.1. Introduction
Once we have evaluated our sources for what they tell us and whether we
can trust them and evaluated theoretically what a particular fingerprint can
tell us, we possess individual pieces of mechanistic evidence that enable us
to confirm or disconfirm to some degree the existence/ nonexistence of a
mechanism or a part thereof. But when we are working with mechanistic
evidence in process-tracing, we typically operate with many different pieces
of evidence for the hypothesized causal process (or for each part of the causal
mechanism when using in-depth process-tracing). This means that we need
to sum together the collective picture painted by multiple pieces of mecha-
nistic evidence. This chapter develops a procedure for the aggregation of
mechanistic evidence using argument road maps that delineate the two-
stage evidence-evaluation framework (from theory to propositions and from
propositions to actual sources of observations). The chapter concludes with
an extended reconstruction of the two-stage evidence-evaluation framework
as applied to Tannenwald’s (1999) case study of the nuclear taboo in the
Korean War.
The first step in aggregating mechanistic evidence in support of infer-
ences about a causal relationship is to map the structure of the arguments
about the empirical fingerprints that a causal relationship might have left
223
224 Process-Tracing Methods
(propositions) and the actual evidence found (or not found) for each propo-
sition. Some propositions might have the character of a smoking gun in rela-
tion to the overall causal relationship being studied because they are very
theoretically unique (van Evera 1997), where finding evidence of the propo-
sition is highly confirming but not finding it would do little to disconfirm
the causal relationship. Other propositions can be more like hoops, with
high theoretical certainty—that is, not finding supporting evidence would
be highly disconfirming, but finding it tells us little (van Evera 1997). We
also have to position each found/not found observation in relation to each
of these propositions about mechanistic evidence.
In most process-tracing research situations, there is a complex, multilayer
structure in the collective body of evidence.1 Multiple propositions about evi-
dence may be related to each part of the mechanism. Even more complex is
the fact that some propositions can be logically composed of multiple support-
ing propositions. For example, the famous smoking gun is actually a cluster
of supporting propositions about evidence in the form of “the gun is the same
as used in the crime” and “the suspect’s fingerprints are on the gun.” If, for
example, the gun is the same one used in the crime, but no evidence links the
suspect to the gun, the found gun is not the proverbial smoking gun.
Propositions about mechanistic evidence need to be decomposed into
supporting propositions because each of them might have different logi-
cal relationships to the proposition: some might be theoretically unique,
whereas others might be theoretically certain. In the case of the cluster of
propositions related to the smoking gun, both are theoretically certain indi-
vidually but only theoretically unique when combined together.
Levels of theoretical certainty and uniqueness are developed for the logical
links between propositions and supporting propositions about empirical fin-
gerprints. Empirical uniqueness or certainty is assessed for the links between
found/not found observations and propositions and supporting propositions.
individually and then evaluating what it can tell us about the existence of the
proposition before we move onto evaluating how the supporting proposi-
tions relate to overall propositions, and so on.
In the argument road map depicted in figure 7.1, we are testing a causal
theory about a suspect who is hypothesized to have killed someone. Proposi-
tion 1 states that there should be evidence that the victim was actually killed
by a gun. This proposition is highly theoretically certain, for if the person
was killed by a gun, it would leave clearly observable signs (fingerprints).
If our source of the observation is an autopsy performed by a credible and
competent authority, we should also expect that the empirical uniqueness
and certainty will be relatively high. This means that if the autopsy tells
us that the person was killed by a gun, we can beyond reasonable doubt
confirm that the person was killed by a gun. However, just because we find
strong confirming evidence that the victim was killed by a gun does not
mean that the suspect committed the crime: the proposition has low theo-
retical uniqueness in relation to the causal hypothesis.
Proposition 2 states, “There is a smoking gun that links the suspect to the
crime.” In relation to the overall causal hypothesis that the suspect is guilty
of murdering the victim, the smoking gun proposition is highly confirm-
ing (high theoretical uniqueness), but the proposition tells us little if we do
not find the smoking gun (low theoretical certainty). Each of the support-
ing propositions is theoretically certain but not unique individually; only
together are they relatively theoretically unique. Sources for proposition 2a
could be the suspect’s fingerprints on the gun or a witness who has seen the
suspect holding the gun and then dropping it on the ground as he or she
fled. The fingerprints, if found, would have relatively low empirical unique-
ness because they confirm only that the suspect held the gun at some point,
not that he actually used it.
Each of these propositions or supporting propositions has different mecha-
nistic evidence that might be relevant; even more important, the inferential
relationship in terms of empirical certainty or uniqueness between found and
not found evidence and supporting propositions is not necessarily the same
as the theoretical uniqueness and certainty between supporting propositions
and higher-level propositions and causal hypotheses. For example, relevant
confirming evidence for proposition 2 could be a forensic report that shows
a ballistic match between the smoking gun and the crime. However, finding
strong confirming evidence of P2a alone does not enable us to confirm the
smoking gun proposition 2; in the absence of other evidence, it could be just
as probable that someone else used the gun to kill the victim. In contrast, if we
find a credible witness who testifies that she saw the suspect fire the gun at the
victim (evidence of P2b), given that (1) the piece of found evidence is highly
empirically unique, (2) proposition 2b is highly theoretically unique in rela-
tion to proposition 2, and (3) proposition 2 is theoretically unique in relation
to hypothesis 1, we could, based on this one piece of evidence, confirm the
overall hypothesis 1 with a reasonable degree of confidence.
An additional consideration is whether the pieces of evidence support-
ing a particular proposition are independent of each other, in which case
they can be summed together to enable more updating, or whether they
have varying degrees of dependence, which lowers our ability to sum them
together.
The argument road map can be presented in table form, as shown in table
7.1. A table is an efficient form of presentation when dealing with multilevel
propositions and/or more pieces of evidence. Presentation in table form also
enables us to provide transparent justifications for why we believe, for ex-
ample, that particular sources of evidence is highly empirically certain and/
or unique in relation to a given proposition. The “lowest” level is the actual
pieces of empirical evidence on which inferences are based, followed by the
logical links from empirical evidence to propositions and/or supporting prop-
ositions. We could also present our case study in an analytical narrative form,
but key parts of the narrative should be explicitly linked to the underlying set
of propositions and evidence and the justifications for why empirical material
is evidence (theoretical and empirical certainty or uniqueness). The table can
then be presented as an appendix (either printed or online). However, aggre-
gating pieces of evidence is not just a simple additive process in most research
situations. In particular, we often face situations where we find both pieces of
confirmatory and disconfirmatory evidence (a mixed evidential picture) in re-
lation to propositions, or we find evidence only for some propositions but not
for others in relation to the overall causal mechanism being assessed.
When aggregating, we therefore first need to map the structure of the
propositions about empirical fingerprints left by activities and how they re-
late logically to the underlying hypothesized causal mechanism. This takes
the form of an assessment of theoretical certainty and uniqueness, where
justifications are provided for why the proposed empirical fingerprints could
be evidence of the activities of the mechanism or a part thereof. This should
be followed by the evaluation of the found (or not found) observations for
each proposition or supporting proposition in terms of what it can tell us
about the proposition and whether we can trust it (empirical certainty or
uniqueness). Using the road map and the logical links between evidence
and propositions, we then start the aggregation process at the lowest level
proposition, assessing in which direction the weight of evidence for each
proposition or supporting proposition points. For example, if we have found
multiple independent pieces of confirmatory evidence that we can trust for
a given proposition, we would add together their probative value, enabling
a strong confirmatory inference. However, the picture is often murkier, with
different pieces of evidence pointing in different directions. When we face
this common situation, the aggregation process must be as transparent as
possible, with clear justifications for why we conclude that the collective
body of evidence points more in one direction than the other.
We develop two sets of rules for aggregating evidence, enabling the system-
atic and transparent evaluation of what causal inferences and the degree
of confidence in them that are warranted based on the collective body of
empirical evidence. The first set of rules relates to adding together evidence
at one level; we then discuss how to move from lower to higher levels of
proposition using the second rule.
The basic rule for aggregating evidence is that two pieces of evidence that are
independent of each other have an additive effect in terms of the amount of
updating they enable (Fitelson 2001: S125; Good 1991: 89–90). This results
in rule 1a:
2. For the debate on this issue of “old” evidence and sequential updating, see, e.g., Eels and
Fitelson 2000; Gallow 2014; Weisberg 2009.
230 Process-Tracing Methods
is taken in single bites, where the first bite updates our confidence and then
informs the prior for the next bite of evidence. We use quantified terms in
this example, but we believe that numbers cannot meaningfully express the
probative value of evidence in process-tracing.
In simplified terms, if finding E1 updates our prior confidence by 5 per-
cent and E2 also updates it by 5 percent, then finding both would update
our confidence by 10 percent, contingent on E1 and E2 being independent
of each other (rule 1b).3 In principle, this means that adding more pieces of
“weak” evidence will provide stronger evidence confirming or disconfirming
a given proposition, contingent on rules 1b and 1d holding.4 However, there
is also a pragmatic limit to adding more pieces of weak evidence. Rule 403
in US evidence law states, “The court may exclude relevant evidence if its
probative value is substantially outweighed by a danger of one or more of
the following: unfair prejudice, confusing the issues, misleading the jury,
undue delay, wasting time, or needlessly presenting cumulative evidence” (em-
phasis added, Michigan Legal Publishing, 2018). When adding more pieces
of evidence would provide only a small amount of additional updating and
would provoke boredom, no more evidence should be produced. In other
words, when adding new pieces of evidence does not tell us anything new,
we should stop.
If the evidence points in different directions in relation to a particu-
lar proposition, we use individual pieces’ direction and relative probative
value to determine the overall inference they enable. If we find two pieces
of confirming evidence that each enable 1 percent updating but do not find
a highly certain piece of evidence (disconfirming evidence) that reduces our
confidence by 10 percent, the overall result might be a disconfirming infer-
ence where we are 8 percent less confident in a proposition holding based on
the evidence. According to Rule 1b,
3. This is a simplification, given that the subsequent test would actually provide a bit less
updating, which is also contingent on how close our prior confidence approaches either 0
percent or 100 percent (see rule 1d on the declining marginal effects of adding more pieces
of evidence).
4. Bennett (2014) talks about adding multiple straw-in-the-wind pieces of evidence to
paint a more compelling overall picture. We agree with this, though with the caveat that we
should avoid a more-is-always-better logic in which we attempt to mask the poor quality of
the evidence (i.e., low probative value) with a high quantity.
232 Process-Tracing Methods
table 50–50 deal where no one “won” or “lost.” The winner might want to
avoid gloating about victory because she is engaged in a long-term, iterated
game of repeated negotiations with the loser, whereas the loser might want
to avoid being portrayed as such for domestic political reasons. Therefore,
we cannot assume that two actors on opposing sides are independent: we
must empirically verify their independence. And we will never be 100 per-
cent confident—a degree of uncertainty will always exist about the indepen-
dence of our sources.
Rule 1c states,
tion and find that their accounts are broadly similar, and if we can verify
that the accounts are independent of each other, we would increase our
confidence in the accuracy of the evidence (more empirically unique),
given that it would be highly unlikely to find similar accounts unless the
observed material is a true measure of what happened. However, finding
similar accounts could also mean that the participants met afterward to
agree on a common account of the events or were just retelling the publicly
available account of the event. And finding that the accounts are too simi-
lar should actually decrease our assessment of the empirical uniqueness of
the observations quite dramatically, as it is highly unlikely that we would
find close-to-perfect correspondence down to particular phrases. Here, the
collected evidence is too good to be true, meaning that such a close fit with
our expectations is not plausible and therefore should significantly down-
grade our confidence in the empirical uniqueness of the observations. If
we are using a secondary source and the collected evidence matches too
closely—more than seems reasonable—this would suggest that we have
selected a biased historical source whose implicit theories run in the same
direction as our own (Lustick 1996: 608, 615). In this situation, we should
significantly downgrade our assessment of the source’s accuracy, with the
result that the collected evidence would not enable us to make inferences
about the veracity of our hypothesis.
Corroboration does not increase confidence in empirical uniqueness un-
less we can substantiate that sources are actually independent of each other.
Doing three interviews and postulating that we have triangulated sources
is not enough—we need to provide proof that the interviews are actually
independent of each other. And so rule 1d holds,
234 Process-Tracing Methods
Claims about propositions can never be stronger than the actual pieces of
evidence that support them directly or through supporting propositions.
Figure 7.4 illustrates a situation with a mixed evidential picture but in
which we could reasonably conclude that there is support for the overall
hypothesis because the confirming evidence for proposition 1 is stronger
than the disconfirming evidence for proposition 2. Here, the contrast in
236 Process-Tracing Methods
Tannenwald (1999: 443–51) uses four parallel case studies to assess whether
norms against the use of atomic weapons (C) affected the U.S. decision not
to use them after 1945 (the outcome, O). She uses within-case evidence to
assess whether C is causally related to O in the Korean War case.
While Tannenwald claims to put forward only one proposition about ev-
idence in relation to the causal hypothesis norms → non-use (“taboo talk”),
the Korean War case study actually includes (at least) four propositions
about mechanistic evidence. When we then aggregate the found evidence
in support of the core propositions and supporting propositions, only very
weak confirming evidence appears to support her core proposition (1—the
operation of norms leaves empirical fingerprints) (table 7.2), mostly because
we are unable to assess what finding “taboo talk” by a particular actor means
in the context of the case (low empirical uniqueness). Similarly, while she
claims that norms against using nuclear weapons may have resulted in a
lack of planning and readiness (proposition 3—inhibitions about nuclear
weapons may have operated in more indirect ways as well—e.g., by influ-
encing perceptions about suitable targets and the state of readiness for tacti-
3 “Inhibitions about nuclear weapons may have operated in more indirect ways as well, for ex-
ample, by influencing perceptions about suitable targets and the state of readiness for tacti-
cal nuclear warfare” (446) (no information provided on theoretical certainty or uniqueness)
• Theoretical certainty and uniqueness not discussed explicitly, meaning that we are not told
how this claim relates to the underlying causal hypothesis.
• observation P3(i) “In March 1951 a Johns Hopkins University research group working
with the Far East Command informed MacArthur that many ‘large
targets of opportunity’ existed for nuclear attack. But the group
found U.S. forces ill-prepared for tactical nuclear warfare” (446).
• Lu: No information given for whether this finding is linked to
moral inhibitions or merely states that the United States was not
ready (low uniqueness). We also have no information about the
context of the remarks, preventing us from further evaluating why
it could be evidence of P3.
• No confirmation of proposition 3
• observation P3(ii) “Truman’s general reluctance to consider nuclear weapons as like any
other weapon . . . must be taken into account. Because of this, as
Rosenberg and others have documented, U.S. planning for nuclear
warfare lagged in the years before Korea” (447).
• Lu: Citing three historical works without discussion means that
we cannot evaluate accuracy. The two other works could build on
Rosenberg’s, but we have no information that would enable us to
assess whether or not this is the case. Some historiographic source
criticism is warranted.
• Weak confirmation of proposition 3
• Aggregation of • Tannenwald’s conclusions: “In short, inhibitions about using
evidence for nuclear weapons in general may have delayed readiness and plan-
proposition 3 ning for tactical nuclear use” (447).
• No information about independence of sources given (P3ii might
have built on P3i).
• Overall confirmation not warranted, given that P3i is not really
evidence of the link between planning and readiness and norms,
and P3ii is of questionable accuracy.
4 “US leaders perceived a taboo developing and sought to challenge it” (448) (no information
provided on theoretical certainty and uniqueness of empirical proposition)
• Theoretical certainty and uniqueness not discussed explicitly, although it is relatively obvi-
ous that if found it would be difficult to account for leaders challenging a taboo unless a
taboo actually existed.
• observation P4(i) “General Nicholas . . . expressed his disappointment over the failure
to use nuclear weapons in Korea. . . . ‘I knew that many individuals
in the United States opposed such thinking for idealistic, moral, or
other reasons’” (447).
(continues)
In sum, during the Korean War, an emerging taboo shaped how U.S.
leaders defined their interests [no evidence is put forward that warrants
concluding that the norm defined interests]. In contrast to the moral
opprobrium Truman personally felt [weak confirming evidence of
proposition 1 supports this], the taboo operated mostly instrumentally
for Eisenhower and Dulles, constraining a casual resort to tactical
nuclear weapons [not warranted—only evidence for their attempt to
counter the emerging norm (proposition 4) is produced]. The burden of
proof for a decision to use such weapons had already begun to shift.
For those who wanted to challenge the emerging taboo, the best way
to do so would have been to actually use such weapons, but the politi-
cal costs of doing so were already high [confirmation of proposition 2
supports the last claim about political costs, but the first parts of the claim
are speculative]. . . . [T]he political debate over the categorization of
nuclear weapons as “unordinary” weapons suggests the early develop-
ment of constitutive effects [evidence for proposition 1 suggests that this
might be warranted, although evidence very weak—therefore using the
term suggest is warranted]. (Tannenwald 1999: 448; bracketed text
added)
244 Process-Tracing Methods
C h a p ter 8
Theory-Testing Process-Tracing
8.1. Introduction
245
246 Process-Tracing Methods
Theory-Testing Process-Tracing 247
of what types of mechanism link a given cause and outcome and under
which conditions a particular mechanism provides the link. It makes sense
first to engage in a form of a plausibility probe, where mechanisms are not
unpacked in any detail. In this situation, we first want to know whether a
causal link exists; if it does, we then want to determine what mechanism
links a cause (or set of causes) and an outcome. Only then do we turn to
learning about the inner workings of that mechanism. And after we have
engaged in more in-depth process-tracing of one or more cases, we can use
minimalist process-tracing to determine whether what we found in the stud-
ied case(s) also holds in other cases within the population of causally similar
cases (e.g., cases with similar contextual conditions).
The key distinctions in minimalist theory-testing relate to how “un-
packed” the mechanism is in theoretical terms (superficial or incomplete)
(see chapter 2) and to how many different empirical fingerprints are op-
erationalized for the mechanism, ranging from a singular proposition to a
cluster of predictions. Figure 8.1 depicts in abstract terms the research pro-
cess of minimalist process-tracing in a single case. The mechanism shown
is a superficial one that is not disaggregated into parts. In a singular test,
a proposition about potential evidence is assessed multiple times during a
temporal process or across space (e.g., issues in a negotiation). In the cluster
test, multiple nonoverlapping propositions about evidence are assessed em-
pirically. Operationalizing both singular empirical tests and clusters of tests
becomes considerably easier when potential activities associated with the
process are made as explicit as possible.
Singular tests of superficial mechanisms can be thought of as simple
plausibility probes that can be deployed early in a research process or as
follow-ups aimed at generalizing insights about mechanisms across addi-
tional cases. In contrast, a stronger version of minimalist process-tracing
involves clusters of empirical tests and/or unpacking the process to some
extent (incomplete mechanistic explanations). In the following, we first dis-
cuss the differences between singular and clusters of tests, followed by a
discussion of when to use plausibility probes and more robust versions of
minimalist process-tracing.
248 Process-Tracing Methods
Van Evera 1997). With a singular proposition, the more (relatively indepen-
dent) sources we find that suggest the presence of the proposed empirical
fingerprint is present, the more we can be confident that we have found the
mechanistic evidence. Finding the proposition multiple times also increases
the empirical uniqueness of the evidence, because we can rule out the alter-
native empirical explanations of the evidence—for example, that the source
does not represent the full empirical record. The exception to this “more is
better” logic would be if the first found observation itself was highly empiri-
cally unique—for example, if we can trust it and it documents the activities
of a particularly central actor at a particularly critical juncture.
A cluster, in contrast, involves operationalizing a battery of nonoverlap-
ping propositions to assess the causal process. These propositions about evi-
dence often are theoretically certain but not very unique (also termed “hoop
tests” in the literature; Van Evera 1997). These are disconfirming when taken
individually, but if we find independent evidence for each of the proposi-
tions, the small probative value of each individual proposition is summed,
enabling a relatively strong overall confirmation of the theory.
For example, our propositions might make predictions about evidence
in the form of the timing and nature of decisions that we should find if
Theory-Testing Process-Tracing 249
250 Process-Tracing Methods
process when we have developed a singular test that can act as a signature of
a process, enabling us to engage in a broader exploration of the bounds of
valid generalization about a mechanism (see chapter 4).
Even though we are employing weak tests in a plausibility probe, the
analysis offers important insights that enable us to refine theories of mecha-
nisms and/or choose to deploy stronger empirical tests. If we do not find
the predicted evidence of our propositions, we would want to engage in a
more comprehensive assessment of why. Was the absence of the predicted
evidence the result of deploying a theoretically low certainty test? To assess
whether the negative finding resulted from the weak test, we should repeat
the plausibility probe with one or more different tests to conclude more con-
fidently that there is or is not a causal relationship. We would then want to
expand our probing outward by brainstorming other potential causal links
that could then be tested empirically. If we still do not find any confirming
evidence, we can then cautiously conclude that—at least for the selected
case—there is no evidence of a link. If we find confirming evidence, we can
then proceed to a stronger case study test of the theorized mechanism.
If we have higher prior confidence in the presence of a theoretical rela-
tionship because of either existing research or our own plausibility probe, we
can then deploy stronger empirical tests in the form of a cluster of nonover-
lapping propositions. If the propositions are independent of each other, their
probative value can be summed together (see chapter 7). The individual tests
in a cluster often are not very theoretically unique, but taken together they
might enable relatively strong inferences. If they are theoretically certain
(either by themselves or together), not finding them would enable relatively
strong disconfirming inferences to be made.
Theory-Testing Process-Tracing 251
252 Process-Tracing Methods
ber of cases within the population studied: the more cases we study within
a population, the better we can infer that what was found in the chosen
cases should also be present in other typical cases throughout the popula-
tion (Rohlfing 2012). Although more rigorous studies of individual cases
enable additional updating about causal relations, other things equal, they
also require significant analytical resources, meaning that in most research
situations we can never conduct more than two or three rigorous case stud-
ies. An additional challenge is more practical, in that presenting the results
of rigorous case studies in a transparent fashion requires significant space,
ruling out more than one or two case studies in article-length work. Yet if
we study only one or two cases in a bounded population of a dozen or more
cases, our cross-case inferences are very weak at best. The trade-off involves
increased confidence about a few cases versus information about a greater
number of cases. One cannot have one’s cake and eat it too.
Brooks and Wohlforth (2000–2001) assess whether a causal link exists be-
tween Soviet economic decline and Soviet leaders’ decisions to engage in a
major strategic shift of policy that resulted in the end of the Cold War. The
causal mechanism linking them together is not unpacked, meaning that it
remains firmly in a theoretical gray box, as figure 8.2 depicts.
The authors test whether material constraints mattered by deploying a
cluster test of three distinct predictions of evidence of a relationship be-
tween the occurrence of the cause and the outcome. The description of their
propositions, in the form of the evidence that the authors “predict” that they
will find and an explanation of why its presence would constitute confirm-
ing evidence, is not as explicitly operationalized as it could be. In particular,
the theoretical certainty and uniqueness could be better developed. At least
three predictions of evidence can be understood as nonoverlapping theoreti-
cally certain propositions (see figure 8.2). We do not describe the empirical
sources used to evaluate whether or not the predicted evidence was actually
present, although multiple sources are used for each proposition.
Theory-Testing Process-Tracing 253
254 Process-Tracing Methods
cause to the outcome, along with the contextual conditions that can be ex-
pected to affect the functioning of the mechanism.
The theoretical conceptualization of the entities includes nouns, while
the activities include verbs that transmit causality through the mechanism.
In social science terms, social entities have causal powers, which can be un-
derstood as “a capacity to produce a certain kind of outcome in the presence
of appropriate antecedent conditions” (Little 1996: 37). A good theorized
mechanism as a system should clearly describe what links together each of
the parts, ideally exhibiting productive continuity between cause and out-
come in a more or less seamless causal story. In practice, our theorized mech-
anisms will often be more mechanistic sketches, but the critical difference
between minimalist and systems understandings is whether our mechanistic
explanation provides a good enough account of how the cause is linked to
the outcome through a process.
The amount of theoretical work necessary to flesh out a causal mecha-
Theory-Testing Process-Tracing 255
While large differences exist in how the natural and social sciences conduct
within-case analysis based on mechanistic evidence, in some social science
research situations, it is appropriate to focus our analytical attention on the
256 Process-Tracing Methods
Theory-Testing Process-Tracing 257
Conceptualization
Before a mechanism can be theorized, we must define the condition (C) and
the outcome (O), attempting to conceptualize what about C can trigger a
mechanism or mechanisms. What are the spatial and temporal bounds of
the operation of the link theorized to be operative (see chapter 3)? In simple
terms, we should be able to clearly state what a given case is a “case of ”—
for example, the democratic peace, where two democratic countries resolve
their conflicts in a peaceful fashion in a situation in which nondemocracies
might have gone to war. If C and/or O is not known, a theory-building
design should be adopted to allow their identification. This could involve
258 Process-Tracing Methods
Theory-Testing Process-Tracing 259
Operationalization
After a typical case is selected, we estimate our prior confidence for the
mechanism (and each part when using in-depth theory-testing) to deter-
mine the relative strength of test required (see chapter 5). Predicted empiri-
cal fingerprints of the traces that activities associated with a mechanism (or
parts thereof ) should then be developed and evaluated for whether they
are theoretically certain and/or unique (see chapter 5). Predictions about
what evidence we should find translate the theoretical concepts of the causal
mechanism into case-specific tests. We must evaluate whether the activity
would have to leave a particular fingerprint (theoretical certainty) and if that
fingerprint is found, whether any plausible alternative explanations exist for
it (theoretical uniqueness). Theoretically certain but not unique tests can
be thought of as hoop tests that provide relatively circumstantial evidence
that enables only disconfirming inferences. In contrast, theoretically unique
evidence is more direct evidence of the activities of a causal process because
it enables confirming inferences to be made.
The next step involves the collection and assessment of empirical observa-
tions (see chapter 6). Before our sources of evidence become mechanistic
evidence on which we can make inferences, we need to assess the content
and accuracy of found/not found empirical observations. Particularly im-
portant here is assessing what found/not found observations mean in a spe-
cific case context. Here, we must be as cautious as historians, not overstating
what found/not found observations actually mean and whether we can trust
them. As the adage holds, “The more you know, the less you know.” Unless
we can demonstrate that a found observation actually means what we think
it means and that it is trustworthy, we should not admit it as confirming
evidence.
After the probative value of each individual piece of confirming or dis-
confirming evidence is found, the final part of the assessment involves aggre-
gating evidence for the mechanism or its parts (see chapter 7). If the evidence
points in the direction of confirmation for each part of a mechanism, we can
infer that the mechanism worked as theorized. If disconfirming evidence is
found for a part (or overall link when minimalist theory-testing), theoretical
building should then be used (see chapter 9).
260 Process-Tracing Methods
Generalization
Theory-Testing Process-Tracing 261
the outcome is the adaptation of government policies in line with the prefer-
ence of the epistemic community. Haas’s (1992) theory is not explicit about
the cause that triggers the mechanisms, but the theory seems to generally
assume that decision-makers face complexity and lack of information about
possible solutions in the specific situation. To establish a possible mecha-
nism between C and O, Löblová uses Haas’s suggestion regarding how such
a mechanism might look:
Theory-Testing Process-Tracing 263
Prior Confidence
Löblová does not utilize very explicit Bayesian language, but she indicates
that she has a relatively high prior confidence in the presence of the causal
Evidence of P2 (i) “The first step, though, was capacity building in the topic. Neither Poland
nor the Czech Republic had any experience with HTA and had only limited
expertise in pharmacoeconomics in general. EBM was also a relatively new
concept in both countries in the early 2000s. In fact, HTA activities in
Poland and in the Czech Republic both evolved from an interest in quality
issues in health care, and related notions of EBM and clinical guidelines,
introduced to the domestic environment by prominent clinicians. Nowhere
is this logical link more visible than in Poland, where first HTA reports were
produced by a department of a newly established National Center for Qual-
ity Assessment in Health Care” (12). Source: Interviews: I-110, I-185, I-27,
I-73; Nizankowski and Wilk 2009.
Lu: Unclear how this piece of evidence actually relates to capacity-building
and dissemination. We must assume that there is a high degree of accuracy
in the evidence provided, but the evidence could be make more transparent.
Weak confirmation of proposition 2
Evidence of P2 (ii) “In the absence of formal university programs or trainings on HTA or
health economics, HTA enthusiasts in both countries learned from text-
books and international conferences” (12). Source: Interviews I-10, I-110,
I-169, I-185, I-27, I-4, I-94.
Lu: Difficult to evaluate whether this was a unique evidence, since it is not
clear whether the absence of a program actually induced enthusiasts to learn
from textbooks, thereby weakening our ability to make confirming infer-
ences based on evidence found. Instead, it would be relevant to document
that members took active measures to enhance and spread knowledge. The
interview data might provide more direct proof of this.
Very weak confirmation of proposition 2
Evidence of P2 (iii) “CMJ employees in Poland received trainings in the early 2000s from inter-
national HTA academics” (12). Source: Nizankowski and Wilk 2009.
Lu: Not clear whether the training was part of a general capacity-building
process. We also lack a clear presentation of the source.
Weak confirmation of proposition 2
266 Process-Tracing Methods
TABLE 8.1.—Continued
Evidence P2 (viii) “A nonprofit branch of a pharmacoeconomic consultancy (iHETA) gave
a number of presentations on HTA and pharmacoeconomics and, similar
to CEESTAHC activities in Poland, has offered health economics train-
ings since 2011, and ran a ‘capacity-building’ project cofinanced by Swiss
development funds” (13). Source: iHETA, 2013
Lu. No direct reference to what CEESTAHC actually did in Poland or any
direct sources on this.
Weak confirmation of proposition 2
Aggregation of General weak discussion of the accuracy of the evidence provided. Each
evidence for proposi- test has a very low degree of uniqueness. The low empirical uniqueness
tion 2 of all of the found sources means that it is possible only to provide weak
confirmation.
Inferences made:
The evidence suggests that the expected part of the causal mechanism was present in the case of
Poland.
Discussion of whether inferences warranted:
• More discussion needed of why empirical material is evidence, first by developing more clearly
the logical relationship between the propositions about observables and the parts of the causal
mechanism and then via discussions of why actual empirical material is evidence for each of
the propositions. Little or no source criticism, although the breadth of sources and multiple
pieces of evidence used increase our confidence in measurement accuracy.
relationship in Poland and the Czech Republic based on the existing litera-
ture, but also notes that the literature has not actually assessed Haas’s pos-
tulated causal mechanism. Her reading of the existing literature thus sup-
ports having a relatively high prior confidence but also leaves us somewhat
in the dark regarding the functioning of the causal mechanism itself. This
suggests that Löblová believes that the prior confidence in the mechanism
itself is relatively low, meaning that she is doing a form of plausibility probe,
where any evidence of a mechanism updates our confidence. Therefore, not
developing stronger reasoned propositions and testing them using stronger
empirical evidence is justified. When we have low prior confidence, even
relatively weak confirming evidence can be enough to update our confi-
dence. Löblová (2017: 7) therefore argues for the need to test the workings
of the causal mechanism “by verifying the mechanism with the help of a
case of an epistemic community’s success”; doing so would enable us to “see
if all the individual parts Haas proposes are necessary to how an epistemic
community influences policy.”
Theory-Testing Process-Tracing 267
As a whole, the empirical evidence for each individual part should enable us
to increase or decrease our confidence in the presence of the entire mecha-
nism. The level of our posterior confidence in the whole mechanism reflects
the lowest posterior level of confidence for each of the parts after our empiri-
cal research. Despite its innovations regarding theoretical clarity about the
causal mechanism, Löblová’s analysis only slightly increases our confidence
in the causal mechanism she theorizes. However, this use of relatively weak
evidence is warranted in her research situation, with her study framed as
a plausibility probe where even weak evidence of the disaggregated causal
mechanism is better than no evidence. Löblová’s article could however ben-
efit from additional reflection on how the individual pieces of evidence are
aggregated and how the collected body of evidence increases or decreases
our confidence in the presence of each part and for the whole mechanism.
C h a p ter 9
Theory-Building Process-Tracing
9.1. Introduction
269
270 Process-Tracing Methods
and the use of multiple cases increases our confidence that a candidate cause
is not case-specific (Beach and Pedersen 2016a: 241–45; Rohlfing 2008).
Theoretical-revision process-tracing is used when we have engaged in
process-tracing in a number of typical cases and want to explore cases where
a cause (or causes) is present and the outcome should have occurred but did
not. In this type of deviant (consistency) case, process-tracing seeks to trace
the mechanism until it breaks down, thereby shedding light on omitted con-
textual and/or causal conditions that potentially must be present. The actual
tracing of mechanisms until failure is, however, an adjunct method used to
focus a subsequent comparative analysis of similar typical and deviant cases
and figure out what is missing in the deviant case.
Theory-Building Process-Tracing 271
272 Process-Tracing Methods
Theory-Building Process-Tracing 273
from the scene and then would ask whether the clue could link the suspect’s
car to the crime, whether there is any plausible theoretical link between the
clue and the crime, and whether this proposition about evidence matches
what has been found empirically. By itself, the clue would be only a weak
piece of evidence in support of a proposition that a person who might be
a suspect was present, but if combined with other pieces of evidence, like
footprints at the crime scene that roughly match those of the owner of the
car, the evidence would be stronger confirmation of the theory.
Empirical probing is an iterative process in which hunches about poten-
tial clues are explored to determine what they can tell us about the newly
found candidate part of the process. If our initial hunch holds empirically
and theoretically, we could then engage in more focused testing, exploring
whether any other independent empirical fingerprints of the theorized part
of the mechanism exist. Returning to the crime scene, we could then inves-
tigate whether the suspect’s footprints or other forms of forensic evidence are
present. If we find the evidence and if it is relatively unique both theoreti-
cally and empirically (e.g., there is no evidence of other people’s presence at
the scene), we would be more confident that our theory holds empirically.
Once we are relatively confident about the newly identified part of a process,
we can then engage in more systematic and robust case studies.
Step 2 in the figure involves inferring from the found observations that
there is actual mechanistic evidence that an underlying plausible part of a
causal mechanism was present. But mechanistic evidence does not speak
for itself. Theory-building often in practice has a testing element, in that
scholars seek inspiration from existing theoretical work and previous obser-
vations for what to look for. For example, an analyst investigating socializa-
tion of administrative officials within international organizations could seek
inspiration in theories of domestic public administration or in psychological
theories of small group dynamics while reading more descriptive accounts
of the workings of international organizations. Existing theory can be con-
ceived as a form of grid that aids in the detection of systematic patterns
in empirical material, enabling inferences about predicted evidence to be
made. In other situations, the search for mechanisms is based on hunches
drawn from puzzles for which existing work cannot account.
In step 3, the secondary inferential leap is made from found evidence
to infer that it reflects an underlying causal mechanism. In reality, theory-
building process-tracing is usually an iterative and creative process, with
the results of each search forming the background for further searches. This
means that steps 1 and 2 are often repeated before step 3 is reached.
274 Process-Tracing Methods
Theory-Building Process-Tracing 275
276 Process-Tracing Methods
Conceptualization
Before a mechanism can be found, we must define the condition (C) and the
outcome (O) and then conceptualize what about C might trigger a mecha-
nism or mechanisms. While we do not know what mechanism links the
two, the definitions of C and O should give us some idea of the spatial and
temporal bounds of the link theorized as operating (see chapter 3). If C
and/or O are not known, a theory-building design should be adopted that
would allow their identification. Theory-building process-tracing should be
virtually the last methodological tool used, given the amount of analytical
resources that would be required to trace an unknown mechanism triggered
by an unknown cause. More appropriate methods would include probing
case studies or comparisons such as a most-similar-system design or Mill’s
method of agreement (Beach and Pedersen 2016a: 241–45).
After C and O are defined, we should devote some thought to potential
contextual conditions. However, given that we do not know what mecha-
nism (if any) links together C and O, we often have little information about
the required contextual conditions.
Theory-Building Process-Tracing 277
Generalization
278 Process-Tracing Methods
Theory-Building Process-Tracing 279
of underlying causal mechanisms. Inferences are then made that part of the
mechanism existed (step 2), resulting in the secondary inference that an un-
derlying mechanism was present in step 3. According to Janis (1983: ix), “For
purposes of hypothesis construction—which is the stage of inquiry with which
this book is concerned—we must be willing to make some inferential leaps
from whatever historical clues we can pick up. But I have tried to start off
on solid ground by selecting the best available historical writings and to use
as my springboard those specific observations that appear to be solid facts in
the light of what is now known about the deliberations of the policy-making
groups.” Further, “What I try to do is to show how the evidence at hand can
be viewed as forming a consistent psychological pattern, in the light of what
is known about group dynamics” (viii).
The presentation of the empirical evidence is not in the form of an ana-
lytical narrative describing events or causal steps between C and O. Instead,
Janis writes, “Since my purpose is to describe and explain the psychological
processes at work, rather than to establish historical continuities, I do not
present the case studies in chronological order. The sequence I use was cho-
sen to convey step-by-step the implications of group dynamics hypotheses”
(1983: viii–ix). He describes four different “symptoms” that can be under-
stood as evidence of a groupthink mechanism, including the illusion of in-
vulnerability, the illusion of unanimity, the suppression of personal doubts,
and the presence of self-appointed mind guards. For example, the shared
illusions of invulnerability and unanimity helped members of the group
maintain a sense of group solidarity, resulting in a lack of critical appraisal
and debate that produced a dangerous level of complacent overconfidence.
Janis (1983: 47) concludes, “The failure of Kennedy’s inner circle to de-
tect any of the false assumptions behind the Bay of Pigs invasion plan can
be at least partially accounted for by the group’s tendency to seek concur-
rence at the expense of seeking information, critical appraisal, and debate.
The concurrence-seeking tendency was manifested by shared illusions and
other symptoms, which helped the members to maintain a sense of group
solidarity. Most crucial were the symptoms that contributed to complacent
overconfidence in the face of vague uncertainties and explicit warnings that
should have alerted the members to the risks of the clandestine military
operation—an operation so ill-conceived that among literate people all over
the world the name of the invasion site has become the very symbol of per-
fect failure.”
C h a p ter 1 0
Explaining-Outcome Process-Tracing
10.1. Introduction
1. The neopositivist version of theory-testing process-tracing focuses on the empirical trac-
281
282 Process-Tracing Methods
ing of minimalist mechanisms that are generalizable to a larger number of cases. In contrast,
critical realists take mechanisms more seriously and consequently believe that generalizations
are possible only across small sets of cases. Critical realists believe that mechanisms are un-
observable but are nevertheless useful analytical constructs whose fingerprints can be traced
empirically.
Explaining-Outcome Process-Tracing 283
284 Process-Tracing Methods
Explaining-Outcome Process-Tracing 285
research designs are open to idiographic explanation in a way that case study
research is not. But single-outcome researchers should not assume, ex ante,
that the truth about their case is contained in factors that are specific to that
case” (Gerring 2006: 717).
Second, given that the inferential ambition is case-centric, seeking a min-
imally sufficient explanation of a particular outcome, more case-sensitive
parts must usually be included in the causal mechanism. The explanation
cannot be detached from the particular case. Theorized mechanisms are
therefore seen as heuristic instruments whose function is to help build the
best possible explanation of a particular outcome (A. Humphreys 2010;
Jackson 2016).
Case selection in explaining-outcome process-tracing is driven by a
strong interest in accounting for a particular interesting and/or historically
important outcome. The outcome is viewed not as a case of something
but instead as a particular event expressed as a proper name. Explaining-
outcome process-tracing can therefore be thought of as a single-outcome
study, seeking the causes of a specific outcome in a single case (Gerring
2006). Examples of this type of study in the literature include Layne’s (1996)
study of U.S. grand strategy toward Western Europe after World War II and
Schiff’s (2008) analysis of the creation of the International Criminal Court.
While case selection in explaining-outcome process-tracing can resemble
the selection of extreme cases (Gerring and Seawright 2007), a case such as
the Cuban Missile Crisis, when understood in a more holistic fashion, is
not just a case of a theoretical concept like crisis bargaining. In explaining-
outcome process-tracing, the goal is to craft a minimally sufficient explana-
tion that captures the unique character of a specific historical event. We
choose the case because it is the Cuban Missile Crisis—it is, in and of itself,
historically important to understand.
The findings of explaining-outcome process-tracing cannot be general-
ized to other cases for two reasons. First, the case itself is unique, given our
broader conceptualization of outcomes (the Cuban Missile Crisis instead
of a narrower theoretical phenomenon such as a case of deterrence bargain-
ing). Second, given the inclusion of nonsystematic parts and case-specific
combinations of mechanisms in our explanations, the actual explanation is
also case-specific. However, it is possible to point outwards when presenting
findings. This occurs when there are elements of the explanation that appear
to have a more “general” character, meaning they potentially can travel to
other cases because they appear to be less contextually specific to the par-
ticular case.
286 Process-Tracing Methods
Explaining-Outcome Process-Tracing 287
288 Process-Tracing Methods
The Sherlock Holmes story “Silver Blaze” (Doyle 1975) is a frequently used
example in teaching process-tracing methods. In contrast to other accounts
though, there are no “variables” in the story since there is no variation within
the case (Collier 2011). The story can better be understood when it is un-
packed as a causal mechanism linking a cause (motive) and outcome (death
of the trainer), with attached propositions about empirical fingerprints and
the actual sources of evidence. We depict this as an argument road map in
table A, which details a causal mechanism that Sherlock Holmes might have
theorized and then tested empirically. We reconstruct the reasoning about
the probative value of evidence for each part based on what we are told in
the story.
The cause that triggers the process is the incurrence of debts by the
trainer, Straker, in connection with his affair with Madame Derbyshire. By
all accounts, the theory that the trainer had nefarious intentions is not very
plausible beforehand, because the trainer had served the owner of the horse
as a jockey for five years and as a trainer for seven years, “always [showing]
himself to be a zealous and honest servant” (Doyle 1975: 14).
While the first part of the process is not tested empirically—although it
logically has to be present—the second part is expected to leave some em-
pirical fingerprints. Evidence that the trainer practiced is not theoretically
certain, because he could have practiced well before the crime occurred and
the injuries might have healed. However, depending on the animal chosen
for practice, the number of mistakes he made in practicing, and the number
of animals hurt, the fingerprint would be relatively unique because most
farmyard animals are robust. The actual evidence found is testimony from
the maid who tends the sheep that three sheep have gone lame. Holmes
289
Beach_2d-edn.indd 290
Part 1 Part 2 Part 3 Part 4 Part 5 Part 6
Theorized mechanism
Trainer (Straker) Straker develops Straker prac- Straker drugs Straker takes horse Straker leads horse Straker gets ready Horse gets
in debt because plans for betting tices procedure for stable boy in from stable in to hollow in moor to hurt horse spooked and kicks
of expensive affair on injured horse as hurting horse evening by putting middle of night (needs light and Straker (death)
with Madame way to cover debts poison in curried ability to engage in
Derbyshire mutton delicate procedure)
Propositions about fingerprints
Bills for expenses No fingerprints ef2a—Animals ef3a—stable boy ef4a—dog not ef5a—horse prints ef6a—some form Fatal wound that
related to Madame discussed in the that have been (Hunter) bears barking at night at hollow of lighting needed could come from
Derbyshire story used for practice signs of poisoning (Hc, Hu) (Hc, Hu) (Hc, Lu) horse
(Lc, Mu) (Mc, Mu) ef6b—jacket taken
ef2b -instrument off and laid on
for procedure bush
(Hc, Lu) (Hc, Mu)
Actual sources
1—bill for dress e2a—lame sheep e3a—other stable e4a—neither Mrs. e5a—“abundant e6a1—wax vesta “His head had
2—photograph of (Hu) boys find Hunter Straker nor other proofs in the at scene of crime been shattered
Straker identified by e2b—small cata- groggy (“no sense stable boys heard mud” (Hu) (Mu) by a savage blow
dressmaker ract knife found could be got out dog at night (Mu) e6a1—candle from some heavy
on Straker (Mu) of him”) (Mu) found in Straker’s weapon”
pocket (Lu)
e6b—“overcoat
flapping from a
furze-bush” (Mu)
9/28/2018 1:08:39 PM
Revised Pages
assesses the maid as a credible and independent source, and it is highly un-
likely that three sheep would go lame coincidentally because they are very
robust animals. Holmes evaluates this evidence as highly empirically and
theoretically unique, stating to the inspector, “Gregory, let me recommend
to your attention this singular epidemic among the sheep” (23).
Another fingerprint expected to be found was the actual instrument in-
tended to injure the horse, which, given Holmes’s mechanistic explanation,
would have to be found at the scene of the crime. It did have a lower theo-
retical uniqueness, as the police believed that the trainer had grabbed the
knife as a weapon to stop the horse thief. The actual source of evidence was
the found cataract knife, which Watson describes as relatively empirically
unique because it is a “very delicate blade devised for very delicate work. A
strange thing for a man to carry with him upon a rough expedition” (15).
The next part relating to drugging the stable boy is expected to leave
fingerprints in the form of the stable boy bearing some signs of having been
poisoned. Aside from the maid, only the trainer and other stable boys had
access to the food and the knowledge that they were having curried mutton,
which could hide the smell of the drug. The actual evidence was testimony
from the other stable boys that they found Hunter groggy in the morning.
Given that Hunter probably only drank water that night, it is unlikely that
he was hungover after drinking alcohol.
Holmes theorized that after drugging the stable boy, Straker took the
horse out of the stable in the middle of the night. The stable dog was a very
vigilant guard dog who always barked when strangers were around but not
when known people were present. Therefore, the e silentio fingerprint of the
dog not barking was expected to be found in theory. However, turning to
evaluate the observation process, while it is implausible that sources would
have lied about hearing/not hearing the dog, there are plausible reasons that
the dog might not have been heard, such as people sleeping soundly (only
medium empirical uniqueness).
The next parts in the causal story relate to the trainer taking the horse to
a hollow on the moor and getting ready to injure the horse. The predicted
empirical fingerprints for part 5 were the hoofprints in the hollow, which
were obviously both highly theoretically certain and unique. In contrast,
the fingerprints for part 6 (preparing) included expectations about finding
traces of plans to light a candle and taking off his jacket to engage in the
delicate procedure. The need for light made traces of a candle very theoreti-
cally certain, but there could be innocent reasons why Straker would have
had a candle or matches (relatively low theoretical uniqueness). Similarly,
taking his jacket off was certain because of the delicate nature of the proce-
dure and because his death prevented him from putting it back on again.
However, he could have taken his jacket off to fight Simpson, as the police
theorized, although hanging it up carefully on a bush before a fight is a less
plausible explanation of the evidence. The actual sources of evidence in-
cluded Holmes’s discovery of a wax vesta (match) at the scene of the crime,
the candle in Straker’s pocket, and the jacket hung carefully on a bush at the
crime scene. The presence of the actual wax vesta at the scene of the crime
was difficult to explain as random chance and even more difficult to explain
with the police’s theory.
Finally, if Holmes’s theorized mechanism was present, we would expect
that the killing blow could have come from the horse. Indeed, the blow was
consistent with a kick from a horse, although it also could have come from
a heavy stick such as what Simpson possessed.
What makes the story intriguing to read is that Holmes’s argument road
map is not put forward at the start of the story. Instead, the reader is left
wondering what theory of the crime (cause and causal mechanism) Holmes
is testing. In the story, it is plausible that Holmes arrived at Dartmoor with
a sketch of a mechanistic explanation along with propositions about poten-
tial empirical fingerprints based on his prior study of published accounts of
the puzzling events. His preliminary theory-building is described as, “For a
whole day my companion had rambled about the room with his chin upon
his chest and his brows knitted, charging and recharging his pipe with the
strongest black tobacco, and absolutely deaf to any of my questions or re-
marks. Fresh editions of every paper had been sent up by our news agent,
only to be glanced over and tossed down into a corner. Yet, silent as he
was, I knew perfectly well what it was over which he was brooding” (1). On
the train ride, Holmes admits that he has a preliminary mechanistic sketch.
Watson asks whether he has formed a theory, and Holmes responds, “At least
I have a grip of the essential facts of the case,” although he then describes
only events, not his mechanistic explanation (3).
On arriving at Dartmoor, Holmes states that he first wants to “go into
one or two questions of detail” before moving on to the crime scene (14).
When he looks through the trainer’s possessions, he finds a cataract knife: as
a surgeon, Watson can assess the knife’s meaning (“A strange thing for a man
to carry” [15]). Other pieces of potential evidence are also found, including
a bill for an expensive dress. We can view Holmes’s initial exploration as a
form of plausibility probe: he is seeking potential evidence for his theorized
mechanism to update his confidence in the theory before expending the ef-
fort to walk to the crime scene. It is obvious here that he is not sure exactly
what fingerprints his theory could have left in the empirical record of the
crime; instead, he is learning about the empirical record to strengthen his
propositions about fingerprints before testing them.
Later in the story, it becomes evident that Holmes is explicitly testing
propositions based on his theorized mechanism, particularly when he arrives
at the crime scene:
Holmes took the bag, and descending into the hollow he pushed the
matting into a more central position. Then stretching himself upon
his face and leaning his chin upon his hands he made a careful study
of the trampled mud in front of him.
“Halloa!” said he, suddenly, “what’s this?’
It was a wax vesta, half burned, which was so coated with mud
that it looked at first like a little chip of wood.
“I cannot think how I came to overlook it,” said the Inspector
with an expression of annoyance.
“It was invisible, buried in the mud. I only saw it because I was
looking for it.” (17)
The story has an auxiliary outcome: what happened with the prize race-
horse, Silver Blaze. Holmes’s theory expected that the horse had wandered
off, and given that Holmes is told that another racing stable is located nearby,
it was a natural corollary to expect that a prize racehorse wandering around
the moor would be found by the staff at the other stable.
We believe that an argument road map can be a helpful tool for mak-
ing explicit a theorized mechanism and systematizing the presentation of
evidence confirming/disconfirming the mechanism. We do not suggest
keeping the argument road map hidden—as Sherlock Holmes does—until
the end of a theory-testing analysis. Instead, it can be included explicitly
at the start, with the subsequent analysis presenting and discussing only
the pieces of mechanistic evidence with highest probative value—in par-
ticular, providing justifications for claims about theoretical and empirical
certainty or uniqueness.
References
295
296 References
References 297
cess Tracing with Bayesian Analysis. In Process Tracing: From Metaphor to Analytic
Tool, ed. Andrew Bennett and Jeffrey Checkel, 276–98. Cambridge: Cambridge
University Press.
Bennett, Andrew, and Jeffrey Checkel. 2014. Process Tracing: From Metaphor to Analytic
Tool. Cambridge: Cambridge University Press.
Berg-Schlosser, Dirk. 2012. Mixed Methods in Comparative Politics: Principles and
Applications. Houndmills: Palgrave Macmillan.
Berg-Schlosser, Dirk, and Gisèle De Meur. 2009. Comparative Research Design: Case
and Variable Selection. In Configurational Comparative Methods: Qualitative Com-
parative Analysis, ed. Benoit Rihoux and Charles Ragin, 19–32. Thousand Oaks,
CA: Sage.
Bhaskar, Roy. 1978. A Realist Theory of Science. Brighton: Harvester.
Blatter, Joachim, and Till Blume. 2008. In Search of Co-Variance, Causal Mecha-
nisms, or Congruence? Towards a Plural Understanding of Case Studies. Swiss
Political Science Review 14 (2): 315–56.
Blatter, Joachim, and Markus Haverland. 2012. Designing Case Studies: Explanatory
Approaches in Small-N Research. Houndmills: Palgrave Macmillan.
Bogaards, Matthijs. 2012. Where to Draw the Line? From Degree to Dichotomy in
Measures of Democracy. Democratization 19 (4): 690–712.
Bogen, Jim. 2005. Regularities and Causality: Generalizations and Causal Explana-
tions. Studies in History and Philosophy of Biological and Biomedical Sciences 36 (2):
397–420.
Bowman, Kirk, Fabrice Lehoucq, and James Mahoney. 2005. Measuring Political
Democracy: Case Expertise, Data Adequacy, and Central America. Comparative
Political Studies 38 (8): 939–70.
Brady, Henry E. 2008. Causation and Explanation in Social Science. In The Oxford
Handbook of Political Methodology, ed. Janet M. Box-Steffensmeier, Henry E.
Brady, and David Collier, 217–70. Oxford: Oxford University Press.
Brady, Henry E., and David Collier, eds. 2011. Rethinking Social Inquiry: Diverse Tools,
Shared Standards. 2nd ed. Lanham, MD: Rowman and Littlefield.
Bretthauer, J. M. 2015. Conditions for peace and conflict: applying a fuzzy-set Qualita-
tive Comparative Analysis to cases of resource scarcity. Journal of Conflict Resolu-
tion 59 (4): 593–616.
Brilmayer, Lea. 1986. The Role of Evidential Weight in Criminal Proof. Boston Univer-
sity Law Review 66 (4): 673–91.
Brooks, Stephen G., and William C. Wohlforth. 2000–2001. Power Globalization and
the End of the Cold War. International Security 25 (3): 5–53.
Bunge, Mario. 1997. Mechanism and Explanation. Philosophy of the Social Sciences 27
(4): 410–65.
Bunge, Mario. 2004. How Does It Work? The Search for Explanatory Mechanisms.
Philosophy of the Social Sciences 34 (2): 182–210.
Campbell, John L. 2005. Institutional Change and Globalization. Princeton: Princeton
University Press.
Caren, Neal, and Aaron Panofsky. 2005. TQCA: A Technique for Adding Temporality to
Qualitative Comparative Analysis. Sociological Methods and Research 34 (2): 147–72.
298 References
Cartwright, Nancy. 1999. The Dappled World: A Study of the Boundaries of Science.
Cambridge: Cambridge University Press.
Cartwright, Nancy. 2007. Hunting Causes and Using Them: Approaches in Philosophy
and Economics. Cambridge: Cambridge University Press.
Cartwright, Nancy. 2011. Predicting “It Will Work for Us”: (Way) beyond Statistics.
In Causality in the Sciences, ed. Phyllis McKay Illari, Federica Russo, and Jon Wil-
liamson, 750–68. Oxford: Oxford University Press.
Cartwright, Nancy, and Jeremy Hardie. 2012. Evidence-Based Policy: A Practical Guide
to Doing It Better. Oxford: Oxford University Press.
Cerling, Thure E., Fredrick Kyalo Manthi, Emma N. Mbua, Louise N. Leakey, Meave
G. Leakey, Richard E. Leakey, Francis H. Brown, Frederick E. Grine, John A.
Hart, Prince Kaleme, Hélène Roche, Kevin T. Uno, and Bernard A. Wood. 2013.
Stable Isotope-Based Diet Reconstructions of Turkana Basin Hominins. Proceed-
ings of the National Academy of Sciences of the United States of America 110 (26):
10501–6.
Chalmers, A. F. 1999. What Is This Thing Called Science? Buckingham: Open Univer-
sity Press.
Charman, Andrew, and Tasha Fairfield. 2015. Applying Formal Bayesian Analysis to
Qualitative Case Research: An Empirical Example, Implications, and Caveats.
Unpublished paper.
Chatterjee, Partha. 2011. Lineages of Political Society: Studies in Postcolonial Democracy.
New York: Columbia University Press.
Checkel, Jeffrey T. 2008. Tracing Causal Mechanisms. International Studies Review 8
(2): 362–70.
Christiansen, Thomas, and Knud Erik Jørgensen. 1999. The Amsterdam Process: A
Structurationist Perspective on EU Treaty Reform. http://eiop.or.at/eiop/texte/1999-
01a.htm
Christiansen, Thomas, and Christine Reh. 2009. Constitutionalizing the European
Union. Basingstoke: Palgrave Macmillan.
CIA. 1968. Intelligence Report: Bayes’ Theorem in the Korean War. No. 0605/68, July.
https://www.cia.gov/library/readingroom/docs/DOC_0001205738.pdf
Clarke, B., D. Gillies, Phyllis Illari, Federica Russo, and Jon Williamson. 2014. Mecha-
nisms and the Evidence Hierarchy. Topoi 33 (2): 339–60.
Clarke, Richard A. 2004. Against All Enemies: Inside America’s War on Terror. New
York: Free Press.
Coleman, James. 1990. Foundations of Social Theory. Cambridge: Harvard University
Press.
Collier, David, Henry E. Brady, and Jason Seawright. 2010. Sources of Leverage in
Causal Inference: Toward an Alternative View of Methodology. In Rethinking
Social Inquiry: Diverse Tools, Shared Standards, 2nd ed., ed. Henry E. Brady and
David Collier, 161–200. Lanham, MD: Rowman and Littlefield.
Collier, David, and Steven Levitsky. 1997. Democracy with Adjectives: Conceptual
Innovation in Comparative Research. World Politics 49 (3): 430–51.
Collier, David, and James Mahoney. 1996. Research Note: Insights and Pitfalls: Selec-
tion Bias in Qualitative Research. World Politics 49 (1): 56–91.
References 299
Coppedge, Michael. 1999. Thickening Thin Concepts and Theories: Combining Large
N and Small in Comparative Politics. Comparative Politics 31 (4): 465–76.
Coppedge, Michael. 2012. Democratization and Research Methods. Cambridge: Cam-
bridge University Press.
Coppedge, Michael, John Gerring, Staffan I. Lindberg, Svend-Erik Skaaning, Jan Teo-
rell, David Altman, Frida Andersson, Michael Bernhard, M. Steven Fish, Adam
Glynn, Allen Hicken, Carl Henrik Knutsen, Kyle L. Marquardt, Kelly McMann,
Valeriya Mechkova, Pamela Paxton, Daniel Pemstein, Laura Saxer, Brigitte
Seim, Rachel Sigman, and Jeffrey Staton. 2017. V-Dem Codebook v7.1. Varieties
of Democracy (V- Dem) Project. https://www.v-dem.net/en/reference/version-
7-1-july-2017/
Craver, Carl F., and Lindley Darden. 2013. In Search of Mechanisms. Chicago: Univer-
sity of Chicago Press.
Cyr, Jennifer. 2015. Making or Breaking Politics: Social Conflicts and Party-System
Change in Democratic Bolivia. Studies in Comparative International Development
50:283–303.
Darnton, Christopher. 2017–18. Archives and Inference. International Security 42 (3):
84–126.
Day, Timothy, and Harold Kincaid. 1994. Putting Inference to the Best Explanation in
Its Place. Synthese 98 (2): 271–95.
Diani, Mario, and Doug McAdam. 2003. Social Movements and Networks. Oxford:
Oxford University Press.
Dion, Douglas. 2003. Evidence and Inference in the Comparative Case Study. In Nec-
essary Conditions: Theory, Methodology, and Applications, ed. Gary Goertz and H.
Starr, 95–112. Oxford: Rowman and Littlefield.
Dowe, Phil. The Causal-Process-Model Theory of Mechanisms. In Causality in the
Sciences, ed. Phyllis McKay Illari, Federica Russo, and Jon Williamson, 865–79.
Oxford: Oxford University Press.
Doyle, A. Conan. 1892. The Adventures of Sherlock Holmes. London: George Newnes.
Doyle, A. Conan. 1975. The Memoirs of Sherlock Holmes. London: George Newnes.
Dunning, Thad. 2017. Contingency and Determinism in Research on Critical Junc-
tures: Avoiding the “Inevitability Framework.” Qualitative and Multi- Method
Research 15 (1): 41–47.
Dyson, Kenneth, and Kevin Featherstone. 1999. The Road to Maastricht: Negotiating
Economic and Monetary Union. Oxford: Oxford University Press.
Eaton, Kent. 2014. Recentralization and the Left Turn in Latin America: Diverg-
ing Outcomes in Bolivia, Ecuador, and Venezuela. Comparative Political Studies
47:1130–57.
Eckstein, Harry. 1975. Case Study and Theory in Political Science. In Handbook of
Political Science, vol. 7, Strategies of Inquiry, ed. Fred I. Greenstein and Nelson W.
Polsby, 79–138. Reading, MA: Addison-Wesley.
Economist. 2010. A Strike against Democracy. October 9, 64–65.
Eells, Ellery, and Brandon Fitelson. 2000. Measuring Confirmation and Evidence.
Journal of Philosophy 97 (12): 663–72.
Elman, Colin, and Miriam Fendius Elman. 2001. Introduction: Negotiating Interna-
300 References
tional History and Politics. In Bridges and Boundaries: Historians Political Scientists
and the Study of International Relations, ed. Colin Elman and Miriam Fendius
Elman, 1–36. Cambridge: MIT Press.
Elster, Jon. 1998. A Plea for Mechanisms. In Social Mechanisms, ed. P. Hedström and
R. Swedberg, 45–73. Cambridge: Cambridge University Press.
Emirbayer, Mustafa, and Ann Mische. 1998. What Is Agency? American Journal of
Sociology 103 (4): 962–1023.
English, Robert D. 1997. Sources, Methods, and Competing Perspectives on the End
of the Cold War. Diplomatic History 21 (2): 283–94.
Erslev, K. R. 1963. Historisk Teknik: Den Historiske Undersøgelse Fremstillet I sine Grund
lier. Copenhagen: Gyldendalske Boghandel.
Evangelista, Mathew. 2014. Explaining the Cold War’s End: Process Tracing All the
Way Down? In Process Tracing: From Metaphor to Analytic Tool, ed. Andrew Ben-
nett and Jeff Checkel, 153–85. Cambridge: Cambridge University Press.
Evans, Peter 1995. The Role of Theory in Comparative Politics. World Politics 48 (1):
3–10.
Fairclough, Norman. 1995. Critical Discourse Analysis: The Critical Study of Language.
London: Routledge.
Fairfield, Tasha. 2013. Going Where the Money Is: Strategies for Taxing Economic
Elites in Unequal Democracies. World Development 47 (1): 42–57.
Fairfield, Tasha, and Andrew E. Charman. 2017. Explicit Bayesian Analysis for Process
Tracing: Guidelines, Opportunities, and Caveats. Political Analysis 25 (3): 363–80.
Falleti, Tulia G., and Julia F. Lynch. 2009. Context and Causal Mechanisms in Politi-
cal Analysis. Comparative Political Studies 42 (9): 1143–66.
Falleti, Tulia G., and James Mahoney. 2015. The Comparative Sequential Method.
In Advances in Comparative-Historical Analysis, ed. James Mahoney and Kathleen
Thelen, 211–39. Cambridge: Cambridge University Press.
Fearon, James. 1991. Counterfactuals and Hypothesis Testing in Political Science.
World Politics 43 (2): 169–95.
Federoff, Howard J., and Lawrence O. Gostin. 2009. Evolving from Reductionism to
Holism: Is There a Future for Systems Medicine? Journal of the American Medical
Association 302 (9): 994–96.
Finke, Daniel. 2009. Domestic Politics and European Treaty Reform: Understanding
the Dynamics of Governmental Position Taking. European Union Politics 10 (4):
482–506.
Fitelson, B. 2001. A Bayesian Account of Independent Evidence with Applications.
Philosophy of Science 68:S123–40.
Ford, Harold P. 1998. CIA and the Vietnam Policymakers: Three Episodes, 1962–1968.
Washington, DC: Center for the Study of Intelligence.
Freedman, David A. 1991. Statistical Models and Shoe Leather. Sociological Methodol-
ogy 21:291–313.
Frieden, Richard D. 1986a. A Close Look at Probative Value. Boston University Law
Review 66 (4): 733–59.
Frieden, Richard D. 1986b. A Diagrammatic Approach to Evidence. Boston University
Law Review 66 (4): 571–620.
References 301
Friedrichs, Jörg, and Friedrich Kratochwill. 2009. On Acting and Knowing: How
Pragmatism Can Advance International Relations Research and Methodology.
International Organization 63 (4): 701–31.
Gaddis, John Lewis. 1972. The United States and the Origins of the Cold War. Ithaca:
Cornell University Press.
Gaddis, John Lewis. 1992–93. International Relations Theory and the End of the Cold
War. International Security 17 (3): 5–58.
Gallow, J. Dimitri. 2014. How to Learn from Theory-Dependent Evidence; or, Com-
mutativity and Holism: A Solution for Conditionalizers. British Journal of the Phi-
losophy of Science 65: 493–519.
George, Alexander L., and Andrew Bennett. 2005. Case Studies and Theory Develop-
ment in the Social Sciences. Cambridge: MIT Press.
Gerring, John. 1999. What Makes a Concept Good? A Criterial Framework for Under-
standing Concept Formation in the Social Sciences. Polity 31 (3): 357–93.
Gerring, John. 2005. Causation: A Unified Framework for the Social Sciences. Journal
of Theoretical Politics 17 (2): 163–98.
Gerring, John. 2007. Case Study Research: Principles and Practices. Cambridge: Cam-
bridge University Press.
Gerring, John. 2010. Causal Mechanisms: Yes but . . . Comparative Political Studies 43
(11): 1499–526.
Gerring, John. 2011. Social Science Methodology: A Unified Framework. Cambridge:
Cambridge University Press.
Gerring, John. 2017. Case Study Research. 2nd ed. Cambridge: Cambridge University Press.
Gerring, John, and Jason Seawright. 2007. Techniques for Choosing Cases. In Ger-
ring, Case Study Research: Principles and Practices, 86–150. Cambridge: Cambridge
University Press.
Giddens, Anthony. 1984. The Constitution of Society: Outline of the Theory of Structura-
tion. Cambridge: Polity Press.
Gill, Christopher, Lora Sabin, and Christopher Schmid. 2005. Why Clinicians Are
Natural Bayesians. British Medical Journal 330 (7499): 1080–83.
Glennan, Stuart S. 1996. Mechanisms and the Nature of Causation. Erkenntnis 44 (1):
49–71.
Glennan, Stuart S. 2002. Rethinking Mechanistic Explanation. Philosophy of Science
69 (S3): 342–53.
Glennan, Stuart S. 2010. Ephemeral Mechanisms and Historical Explanation. Erken-
ntnis 72 (2): 251–66.
Glennan, Stuart S. 2011. Singular and General Causal Relations: A Mechanist Perspec-
tive. In Causality in the Sciences, ed. Phyllis McKay Illari, Federica Russo, and Jon
Williamson, 789–817. Oxford: Oxford University Press.
Goertz, Gary. 2006. Social Science Concepts: A User’s Guide. Princeton: Princeton Uni-
versity Press.
Goertz, Gary. 2017. Multimethod Research, Causal Mechanisms, and Case Studies: An
Integrated Approach. Princeton: Princeton University Press.
Goertz, Gary, and Jack S. Levy, eds. 2007. Explaining War and Peace: Case Studies and
Necessary Condition Counterfactuals. London: Routledge.
302 References
Goertz, Gary, and James Mahoney. 2009. “Scope in Case-Study Research.” In The Sage
Handbook of Case-Based Methods, ed. David Byrne and Charles C. Ragin, 307–17.
Thousand Oaks, CA: Sage.
Goertz, Gary, and James Mahoney. 2012. A Tale of Two Cultures: Qualitative and
Quantitative Research in the Social Sciences. Princeton: Princeton University Press.
Good, L. J. 1968. Corroboration, Explanation, Evolving Probability, Simplicity, and a
Sharpened Razor. British Journal of Philosophy 19 (2): 123–43.
Good, L. J. 1991. Weight of Evidence and the Bayesian Likelihood Ratio. In Use of Statistics
in Forensic Science, ed. C. G. Aitken and D. A. Stoney, 85–106. London: CRC Press.
Groff, Ruth. 2011. Getting Past Hume in the Philosophy of Social Science. In Causal-
ity in the Sciences, ed. Phyllis McKay Illari, Federica Russo, and Jon Williamson,
296–316. Oxford: Oxford University Press.
Grzymala-Busse, Anna. 2011. Time Will Tell? Temporality and the Analysis of Causal
Mechanisms and Processes. Comparative Political Studies 44 (9): 1267–97.
Haas, Peter M. 1992. “Introduction: Epistemic Communities and International Policy
Coordination.” International Organization 46 (1): 1–35.
Haesebrouck, Tim. 2016. The Added Value of Multi-Value Qualitative Comparative
Analysis. Forum: Qualitative Social Research 17 (1). http://dx.doi.org/10.17169/
fqs-17.1.2307
Haggard, Stephan, and Robert R. Kaufman. 2016. Dictators and Democrats: Masses,
Elites, and Regime Change. Princeton: Princeton University Press.
Hall, Peter A. 2008. Systematic Process Analysis: When and How to Use It. European
Political Science 7 (3): 304–17.
Harish, S. P., and Andrew T. Little. 2017. The Political Violence Cycle. American Politi-
cal Science Review 111 (2): 237–55.
Harvey. Frank P. 2012. Explaining the Iraq War: Counterfactual Theory, Logic and Evi-
dence. Cambridge: Cambridge University Press.
Hedström, Peter, and Richard Swedberg, eds. 1998. Social Mechanisms: An Analytical
Approach to Social Theory. Cambridge: Cambridge University Press.
Hedström, Peter, and Petri Ylikoski. 2010. Causal Mechanisms in the Social Sciences.
Annual Review of Sociology 36:49–67.
Hempel, Carl G. 1965. The Function of General Laws in History. In Aspects of Scientific
Explanation and Other Essays, 231–44. New York: Free Press.
Holland, Paul W. 1986. Statistics and Causal Inference. Journal of the American Statisti-
cal Association 81 (396): 945–60.
Howson, Colin, and Peter Urbach. 2006. Scientific Reasoning: The Bayesian Approach.
3rd ed. La Salle, IL: Open Court.
Hug, Simon, and Thomas König. 2002. In View of Ratification: Governmental Pref-
erences and Domestic Constraints at the Amsterdam Intergovernmental Confer-
ence. International Organization 56 (4): 447–76.
Hug, Simon, and Thomas Schultz. 2007. Referendums in the EU’s Constitution
Building Process. Review of International Organizations 2 (2): 177–218.
Hume, David. 1975. Enquiries Concerning Human Understanding and Concerning the
Principles of Morals. Ed. P. H. Nidditch. 3rd ed. Oxford: Oxford University Press.
Humphreys, Adam R. C. 2010. The Heuristic Explanation of Explanatory Theories in
International Relations. European Journal of International Relations 17 (2): 257–77.
References 303
Humphreys, Macartan, and Alan Jacobs. 2015. Mixing Methods: A Bayesian Approach.
American Political Science Review 109 (4): 653–73.
Iklé, Fred Charles. 1964. How Nations Negotiate. New York: Praeger.
Illari, Phyllis McKay. 2011. Mechanistic Evidence: Disambiguating the Russo-
Williamson Thesis. International Studies in the Philosophy of Science 25 (2): 139–57.
Illari, Phyllis McKay, and Federica Russo. 2014. Causality: Philosophical Theory Meets
Scientific Practice. Oxford: Oxford University Press.
Illari, Phyllis McKay, and Jon Williamson. 2011. Mechanisms Are Real and Local. In
Causality in the Sciences, ed. Phyllis McKay Illari, Federica Russo, and Jon Wil-
liamson, 818–44. Oxford: Oxford University Press.
Illari, Phyllis McKay, and Jon Williamson. 2013. In Defense of Activities. Journal for
General Philosophy of Science 44 (1): 69–83.
Jackson, Patrick T. 2016. The Conduct of Inquiry in International Relations. 2nd ed.
London: Routledge.
Janis, Irving L. 1983. Groupthink: Psychological Studies of Policy Decisions and Fiascoes.
Boston: Houghton Mifflin.
Jervis, Robert. 1976. Perceptions and Misperceptions in International Politics. Princeton:
Princeton University Press.
Jervis, Robert. 2010. Why Intelligence Fails: Lessons from the Iranian Revolution and the
Iraq War. Ithaca: Cornell University Press.
Joos, Erich, H. Dieter Zeh, Claus Kiefer, Domenico J. W. Giulini, Joachim Kupsch,
and Ion-Olimpiu Stamatescu. 2003. Decoherence and the Appearance of a Classical
World in Quantum Theory. Berlin: Springer-Verlag.
Kaiser, Marie I. 2017. The Components and Boundaries of Mechanisms. In The Rout-
ledge Handbook of Mechanisms and Mechanistic Philosophy, ed. Stuart Glennan and
Phyllis Illari, 116–30. London: Routledge.
Kaye, David H. 1986. Comment: Quantifying Probative Value. Boston University Law
Review 66 (4): 761–66.
Khong, Yuen Foong. 1992. Analogies at War: Korea, Munich, Dien Bien Phu, and the
Vietnam Decisions of 1965. Princeton: Princeton University Press.
King, Gary, Robert O. Keohane, and Sidney Verba. 1994. Designing Social Inquiry:
Scientific Inference in Qualitative Research. Princeton: Princeton University Press.
Kitano, Hiroaki. 2002. Systems Biology: A Brief Overview. Science 295 (5560): 1662–
64.
Klein, Richard G. 2013. Comments: Stable Isotope-Based Diet Reconstructions of
Turkana Basin Hominins. Proceedings of the National Academy of Sciences in the
United States of America 110 (26): 10470–72.
Kramer, Mark. 1990. Remembering the Cuban Missile Crisis: Should We Swallow
Oral History? International Security 15 (1): 212–16.
Krebs, R. R., and P. T. Jackson. 2007. Twisting Tongues and Twisting Arms: The Power
of Political Rhetoric. European Journal of International Relations 13 (1): 35–66.
Kreuzer, Markus. 2014. The Structure of Description: Elements of and Criteria for
Evaluating Historical Analysis. Paper presented at the annual meeting of the
American Political Science Association.
Kuehn, David, and Harold Trinkunas. 2017. Conditions of Military Contestation in
Populist Latin America. Democratization 24 (5): 859–80.
304 References
Kuhlmann, Meinard, and Stuart Glennan. 2014. On the Relation between Quantum
Mechanical and Neo-Mechanistic Ontologies and Explanatory Strategies. Euro-
pean Journal of the Philosophy of Science 4 (3): 337–59.
Kurki, Milja. 2008. Causation in International Relations: Reclaiming Causal Analysis.
Cambridge: Cambridge University Press.
Larson, Deborah Welch. 2001. Sources and Methods in Cold War History: The Need
for a New Theory-Based Archival Approach. In Bridges and Boundaries: Historians,
Political Scientists, and the Study of International Relations, ed. Colin Elman and
Miriam Fendius Elman, 327–50. Cambridge: MIT Press.
Layne, Christopher. 2006. The Peace of Illusions: American Grand Strategy from 1940 to
the Present. Ithaca: Cornell University Press.
Leamer, Edward E. 2010. Tantalus on the Road to Asymptopia. Journal of Economic
Perspectives 24 (2): 31–46.
Lebow, Richard Ned. 2000–2001. Contingency, Catalysts, and International System
Change. Political Science Quarterly 115 (4): 591–616.
Leifeld, Philip, and Volker Schneider. 2012. Information Exchange in Policy Net-
works. American Journal of Political Science 53 (3): 731–44.
Levi-Montalcini R., and P. Calissano. 2006. The Scientific Challenge of the 21st Cen-
tury: From a Reductionist to a Holistic Approach via Systems Biology. BMC Neu-
roscience 7 (Suppl 1): S1. https://doi.org/10.1186/1471-2202-7-S1-S1
Levy, Jack. 2008. Case Studies: Types, Designs, and Logics of Inference. Conflict Man-
agement and Peace Science 25 (1): 1–18.
Levy, Jack. 2015. Counterfactuals, Causal Inference, and Historical Analysis. Security
Studies 24 (3): 378–402.
Lewis, David K. 1986. Postscripts to “Causation.” In Philosophical Papers, 2:172–213.
Oxford: Oxford University Press.
Lieberman, Evan S. 2005. Nested Analysis as a Mixed-Method Strategy for Compara-
tive Research. American Political Science Review 99 (3): 435–51.
Little, Daniel. 1996. Causal Explanation in the Social Sciences. Southern Journal of
Philosophy 34 (S1): 31–56.
Löblová, Olga. 2017. When Epistemic Communities Fail: Exploring the Mechanism
of Policy Influence. Policy Studies Journal. https://doi.org/10.1111/psj.12213
Lustick, Ian S. 1996. History, Historiography, and Political Science: Multiple Histori-
cal Records and the Problem of Selection Bias. American Political Science Review
90 (3): 605–18.
Machamer, Peter. 2004. Activities and Causation: The Metaphysics and Epistemology
of Mechanisms. International Studies in the Philosophy of Science 18 (1): 27–39.
Machamer, Peter, Lindley Darden, and Carl F. Craver. 2000. Thinking about Mecha-
nisms. Philosophy of Science 67 (1): 1–25.
Mackie, J. L. 1965. Causes and Conditions. American Philosophical Quarterly 2 (2):
245–64.
Madea, Burkhard, and Dirk W. Lachenmeier. 2005. Postmortem Diagnosis of Hyper-
tonic Dehydration. Forensic Science International 155 (1): 1–6.
Madrid, Raúl. 2013. Bolivia: Origins and Politics of the Movimiento al Socialismo. In
The Resurgence of the Latin American Left, ed. Stephen Levitsky and Kenneth M.
Roberts, 239–59. Baltimore: Johns Hopkins University Press.
References 305
306 References
Milligan, John D. 1979. The Treatment of a Historical Source. History and Theory 18
(2): 177–96.
Moore, Barrington. 1991 [1966]. Social Origins of Dictatorship and Democracy: Lord and
Peasant in the Making of the Modern World. London: Penguin.
Moravcsik, Andrew. 1998. The Choice for Europe. Ithaca: Cornell University Press.
Moravcsik, Andrew. 2010. Active Citation: A Precondition for Replicable Qualitative
Research. PS: Political Science and Politics 43 (1): 29–35.
Morgan, Stephen L., and Christopher Winship. 2007. Counterfactuals and Causal
Inference: Methods and Principles for Social Research. Cambridge: Cambridge Uni-
versity Press.
Munck, Gerardo L., and Jay Verkuilen. 2002. Conceptualizing and Measuring Democ-
racy: Evaluating Alternative Indices. Comparative Political Studies 35 (1): 5–34.
Munson, Ronald. 1976. The Way of Words. Boston: Houghton Mifflin.
Musshoff, Frank, Peter M. D. Schmidt, Thomas Daldrup, and Burkhard Madea. 2002.
Cyanide Fatalities: Case Studies of Four Suicides and One Homicide. American
Journal of Forensic Medicine and Pathology 23 (4): 315–20.
New York Times. 2009. Bolivian President Says Plot on His Life Was Tied to Coup
Attempt. April 18.
O’Mahoney, Joseph O. 2017. Making the Real: Rhetorical Adduction and the Bangla-
desh Liberation War. International Organization 71 (2): 317–48.
Owen, John M. 1994. How Liberalism Produces Democratic Peace. International Secu-
rity 19 (2): 87–125.
Owen, John M. 1997. Liberal Peace, Liberal War: American Politics and International
Security. Ithaca: Cornell University Press.
Parsons, Craig. 2007. How to Map Arguments in Political Science. Oxford: Oxford
University Press.
Parsons, Craig. 2016. Ideas and Power: Four Intersections and How to Show Them.
Journal of European Public Policy 23 (3): 446–63.
Pedersen, Rasmus Brun, and Y. Reykers. 2017. Small States Bandwagon for Status?
A Comparative Study of Denmark and Belgium. Paper presented at The XVIII
Nordic Political Science Congress, University of Denmark (Odense), August 8–11.
Peirce, C. S. 1955. Philosophical Writings of Peirce. Ed. J. Buchler. New York: Dover.
Pennings, Paul. 2003. Beyond Dichotomous Explanations: Explaining Constitutional
Control of the Executive with Fuzzy-Sets. European Journal of Political Research 42
(4): 541–67.
Piccinini, Gualtiero. 2017. Activities Are Manifestations of Causal Powers. In Eppur
Si Muove: Doing History and Philosophy of Science with Peter Machamer, ed. M. P.
Adams, Z. Biener, U. Feest, and J. A. Sullivan, 171–82. Cham: Springer.
Pierson, Paul. 1996. The Path to European Integration: A Historical Institutionalist
Perspective. Comparative Political Studies 29 (1): 123–63.
Pierson, Paul. 2003. Big, Slow-Moving, and . . . Invisible: Macrosocial Processes in
the Study of Comparative Politics. In Comparative Historical Analysis in the Social
Sciences, ed. James Mahoney and Dietrich Rueschemayer, 177–207. Cambridge:
Cambridge University Press.
Pierson, Paul. 2004. Politics in Time: History Institutions and Social Analysis. Princeton:
Princeton University Press.
References 307
Phillips, Kristine. 2017. In the Latest JFK Files: The FBI’s Ugly Analysis on Martin
Luther King Jr., Filled with Falsehoods. Washington Post, November 4.
Ragin, Charles C. 1987. The Comparative Method: Moving beyond Qualitative and
Quantitative Strategies. Berkeley: University of California Press.
Ragin, Charles C. 2000. Fuzzy-Set Social Science. Chicago: University of Chicago
Press.
Ragin, Charles. 2004. Turning the Tables: How Case-Oriented Methods Challenge
Variable-Oriented Methods. In Rethinking Social Inquiry: Diverse Tools, Shared
Standards, ed. Henry E. Brady and David Collier, 125–41. Lanham, MD: Rowman
and Littlefield.
Ragin, Charles C. 2008. Redesigning Social Inquiry: Fuzzy Sets and Beyond. Chicago:
University of Chicago Press.
Ragin, Charles C., and Sarah Ilene Strand. 2008. Using Qualitative Comparative
Analysis to Study Causal Order: Comment on Caren and Panofsky (2005). Socio-
logical Methods and Research 36 (4): 431–41.
Reskin, Barbara F. 2003. Including Mechanisms in Our Models of Ascriptive Inequal-
ity. American Sociological Review 68 (1): 1–21.
Rickles, Dean. 2009. Causality in Complex Interventions. Medical Health Care and
Philosophy 12:77–90.
Rittinger, Eric R., and Mathew R. Cleary. 2013. Confronting Coup Risk in the Latin
American Left. Studies in Comparative International Development 48 (4): 403–31.
Roberts, Clayton. 1996. The Logic of Historical Explanation. University Park: Pennsyl-
vania State University Press.
Rohlfing, Ingo. 2008. What You See and What You Get: Pitfalls and Principles of
Nested Analysis in Comparative Research. Comparative Political Studies 41 (11):
1492–514.
Rohlfing, Ingo. 2012. Case Studies and Causal Inference: An Integrative Framework.
Houndmills: Palgrave Macmillan.
Rohlfing, Ingo. 2014. Comparative Hypothesis Testing via Process Tracing. Sociological
Methods and Research 43 (4): 606–42.
Rohlfing, Ingo, and Carsten Q. Schneider. 2016. A Unifying Framework for Caus-
al Analysis in Set- Theoretic Multimethod Research. Sociological Methods and
Research. https://doi.org/10.1177/0049124115626170
Rudalevige, Andrew. 2002. Managing the President’s Program: Presidential Leadership
and Legislative Policy Formulation. Princeton: Princeton University Press.
Rueschemeyer, Dietrich. 2003. Can One or a Few Cases Yield Theoretical Gains? In
Comparative Historical Analysis in the Social Sciences, ed. James Mahoney and Diet-
rich Rueschemayer, 305–37. Cambridge: Cambridge University Press.
Runhardt, Rosa W. 2015. Evidence for Causal Mechanisms in Social Science: Recom-
mendations from Woodward’s Manipulability Theory of Causation. Philosophy of
Science 82 (5): 1296–307.
Russo, Federica, and Jon Williamson. 2007. Interpreting Causality in the Health Sci-
ences. International Studies in the Philosophy of Science 21 (2): 157–70.
Russo, Federica, and Jon Williamson. 2011. Generic versus Single-Case Causality: The
Case of Autopsy. European Journal of the Philosophy of Science 1 (1): 47–69.
Salmon, Wesley. 1998. Causality and Explanation. Oxford: Oxford University Press.
308 References
Samford, Steven. 2010. Averting “Disruption and Reversal”: Reassessing the Logic of
Rapid Trade Reform in Latin America. Politics and Society 38 (3): 373–407.
Sarkees, Meredith Reid. 2010. Defining and Categorizing Wars. In Resort to War: A
Data Guide to Inter-State, Extra-State, Intra-State, and Non-State Wars, 1816–2007,
ed. Meredith Reid Sarkees and Frank Whelon Wayman, 39–73. Washington, DC:
CQ Press.
Sartori, Giovanni. 1970. Concept Misformation in Comparative Politics. American
Political Science Review 64 (4): 1033–53.
Sartori, Giovanni. 1984. Guidelines for Concept Analysis. In Social Science Concepts: A
Systematic Analysis, ed. Giovanni Sartori, 15–85. Beverly Hills, CA: Sage.
Sawyer, R. Keith. 2004. The Mechanisms of Emergence. Philosophy of the Social Sci-
ences 34 (2): 260–82.
Schaffer, Frederic Charles. 2016. Elucidating Social Science Concepts: An Interpretivist
Guide. London: Routledge.
Scheier, Margrit. 2012. Qualitative Content Analysis in Practice. London: Sage.
Schimmelfennig, Frank. 2001. The Community Trap: Liberal Norms, Rhetorical
Action, and the Eastern Enlargement of the European Union. International Orga-
nization 55 (1): 47–80.
Schimmelfennig, Frank. 2015. Efficient Process Tracing: Analyzing the Causal Mecha-
nisms of European Integration. In Process Tracing: From Metaphor to Analytical
Tool, ed. Andrew Bennett and Jeffrey T. Checkel, 98–125. Cambridge: Cambridge
University Press.
Schmitt, Johannes, and Derek Beach. 2015. The Contribution of Process Tracing to
Theory-Based Evaluations of Complex Aid Instruments. Evaluation 21 (4): 429–47.
Schmitter, Philippe C. 1979. Still the Century of Corporatism? In Trends towards Cor-
poratist Intermediation, ed. P. Schmitter and G. Lehmbruch, 7–49. London: Sage.
Schneider, Carsten Q., and Ingo Rohlfing. 2013. Combining QCA and Process Trac-
ing in Set-Theoretical Multi-Method Research. Sociological Methods and Research
42 (4): 559–97.
Schneider, Carsten Q., and Ingo Rohlfing. 2016. Case Studies Nested in Fuzzy-Set
QCA on Sufficiency: Formalizing Case Selection and Causal Inference. Sociologi-
cal Methods and Research 45 (3): 526–68.
Schneider, Carsten Q., and Claudius Wagemann. 2012. Set-Theoretic Methods for the
Social Sciences: A Guide to Qualitative Comparative Analysis. Cambridge: Cam-
bridge University Press.
Schwartz-Shea, Peregrine, and Dvora Yanow. 2012. Interpretive Research Design: Con-
cepts and Processes. New York: Routledge.
Scriven, Michael. 2011. Evaluation, Bias, and Its Control. Journal of Multidisciplinary
Evaluation 7 (15): 79–98.
Siewert, Markus B. 2017. Qualitative Comparative Analysis. In Weiterentwicklungen
in den politikwissenschaftlichen Methoden. Innovative Techniken für qualitative und
quantitative Forschung, ed. Sebastian Jäckle. Wiesbaden: VS Springer.
Sil, Rudra, and Peter J. Katzenstein. 2010. Beyond Paradigms: Analytical Eclecticism in
the Study of World Politics. Basingstoke: Palgrave Macmillan.
Silver, Nate. 2013. The Signal and the Noise. New York: Penguin.
References 309
310 References
Verkuilen, Jay. 2005. Assigning Membership in a Fuzzy Set Analysis. Sociological Meth-
ods and Research 33 (4): 462–96.
Wagemann, Claudius. 2017. Qualitative Comparative Analysis (QCA) and Set
Theory. Oxford Research Encyclopedia of Politics. https://doi.org/10.1093/acre-
fore/9780190228637.013.247
Wagner, Carl. G. 2001. Old Evidence and New Explanation III. Philosophy of Science
68 (3): S165–75.
Waldner, David. 2012. Process Tracing and Causal Mechanisms. In Oxford Handbook
of the Philosophy of Social Science, ed. H. Kincaid, 65–84. Oxford: Oxford Univer-
sity Press.
Waldner, David. 2014. What Makes Process Tracing Good? Causal Mechanisms,
Causal Inference, and the Completeness Standard in Comparative Politics. In
Process Tracing: From Metaphor to Analytic Tool, ed. Andrew Bennett and Jeffrey
Checkel, 126–52. Cambridge: Cambridge University Press.
Waldner, David. 2015. Process Tracing and Qualitative Causal Inference. Security Stud-
ies 24 (2): 239–50. https://doi.org/10.1080/09636412.2015.1036624
Walker, Vern. 2007. Discovering the Logic of Legal Reasoning. Hofstra Law Review 35
(4): 1687–1708.
Waltz, Kenneth N. 1979. Theory of International Politics. New York: McGraw-Hill.
Waskan, Jonathan. 2008. Knowledge of Counterfactual Interventions through Cogni-
tive Models of Mechanisms. International Studies in the Philosophy of Science 22
(3): 259–75.
Waskan, Jonathan. 2011. Mechanistic Explanation at the Limit. Synthese 183 (3): 389–
408.
Wasserman, Stanley, and Katherine Faust. 1994. Social Network Analysis. Cambridge:
Cambridge University Press.
Wauters, Benedict, and Derek Beach. 2018. Process-Tracing and Congruence Analysis
to Support Theory-Based Impact Evaluation. Evaluation 24 (3): 284–305.
Weisberg, Jonathan. 2009. Commutativity or Holism? A Dilemma for Conditional-
izers. British Journal of the Philosophy of Science 60 (4): 793–812.
Weller, Nicholas, and Jeb Barnes. 2014. Finding Pathways: Mixed-Method Research for
Studying Causal Mechanisms. Cambridge: Cambridge University Press.
Wendt, Alexander. 1999. Social Theory of International Politics. Cambridge: Cambridge
University Press.
White, Howard. 2009. Theory-Based Impact Evaluation: Principles and Practice. Inter-
national Initiative for Impact Evaluation Working Paper 3. http://www.3ieimpact.
org/media/filer_public/2012/05/07/Working_Paper_3.pdf
White, Timothy J. 2000. Cold War Historiography: New Evidence behind Traditional
Typologies. International Social Science Review 75 (3–4): 35–46.
Wight, Colin. 2004. Theorizing the Mechanisms of Conceptual and Semiotic Space.
Philosophy of the Social Sciences 34 (2): 283–99.
Williams, Malcolm and Wendy Dyer. 2009. Single Case Probabilities. In The SAGE
Handbook of Case-Based Methods, ed. David Byrne and Charles C. Ragin, 85–100.
London: Sage.
Williams, Timothy, and Sergio M. Gemperle. 2017. Sequence Will Tell! Integrating
References 311
Index
313
314 Index
Index 315
316 Index
Empirical predictions (continued) mechanistic, 4–6, 14, 35, 40, 47, 55–
theoretical uniqueness of predic- 56, 164
tions, 44, 100, 138, 142, 155–56, 158, pattern, 172, 198, 216
175–76, 178, 180–81, 190–93, 205–6, probative value, 157, 166, 168, 175–77,
224–36, 244, 259, 263 204, 223–24
Empirical test types sequence, 172, 213, 216, 219
hoop tests, 166, 167, 248, 259 trace, 172, 216, 265
smoking gun, 166, 168, 169, 205, 224, Experiments, 8, 13–16, 44, 55, 56, 81, 157,
225, 228, 234, 247, 250, 251 159–62, 182
straw-in-the-wind, 230 Explaining-outcome process-tracing, 2,
Empirical uniqueness, 155–56, 196, 200– 6, 9, 11–12, 281
206, 210, 211, 213, 224–36 Abductive (see Abduction)
Endogeneity, 161 composite / conglomerate mecha-
Entities. See Operationalization of causal nisms, 284
mechanisms empirics first, 286–87
Epistemological probabilism, 16, 174 historical outcomes, 2, 9, 11, 281, 283,
Epistemology, 16, 19, 173 284, 287
Equifinality. See Causal mechanisms theory first, 286–87
Errors Extrapolation, 255–57
nonsystematic, 103, 204, 284, 285
systematic, 204, 217 Falsification, 173, 174, 178
Essentialist position, 57 Findings
Evaluating evidence negative, 59, 250
accuracy of observations, 195, 197, positive, 250–51
204–7, 212–13, 214, 216, 233, 259
bias / systematic errors (see Bias) Group-think, 13, 86, 185–86, 198, 278–
content of observations, 5, 172, 195, 79
201, 205, 210, 211, 259
reliability / nonsystematic errors (see Historical scholarship, 199–207, 214, 221,
Bias) 231, 259, 282, 284
Events /empirical narratives. See Nar- Homogeneity. See under Causal homo-
rative geneity
Evidence Homogeneous populations. See Popula-
account, 157, 172, 198, 213, 215, 216, tion
219, 220 Hume, David: neo-Humean, 45–46
archival observations (see Data) Hypothesis, 93, 167, 169, 172, 175, 178,
confirming, 110, 179, 182, 184, 192, 179, 180, 182, 186
225, 230, 235, 249
difference making, 13, 21, 26, 33, 34, Idiosyncratic, 132, 138, 140, 251, 255
46, 55, 56, 97, 100, 116, 157, 159–65, Independence of evidence / observa-
182 tions, 205, 215, 230–32
disconfirming, 172, 174, 221, 235 Independent variable, 61
e silentio evidence, 189, 291 In-depth theory-testing process-tracing.
evidentiary weight, 178, 229, 234 See Process-tracing
Index 317
Large-N, 3, 92, 93, 105, 106, 162 Narrative, 124, 147, 152, 156
Least likely case. See Case selection descriptive, 2, 3, 11, 31, 32, 87, 88, 159,
Likelihood ratio. See Bayesian logic 277
events, 45, 46, 51, 87, 95, 101, 148, 156,
Manipulation, 164 161
Measurement Necessary condition. See Causal relation
accuracy, 199, 205, 207, 212 Nested analysis, 103, 105–6
error, 195, 204, 214, 215, 217 Neo-positivist, 7, 45, 46, 47, 52, 96,
Mechanismic evidence. See Evidence 281
Mechanistic heterogeneity, 6, 7, 41, 54, Non-homogeneous populations. See
77–81, 90, 91, 99, 104, 106, 112–32, Population
135, 136
Mechanistic homogeneity, 41, 90, 96, Observations
112, 119, 124, 125, 129, 131, 132, 135, independence of sources, 222, 231
136, 140, 144, 149, 176, 260, 277 raw data material, 5, 166, 177
Mechanisms as systems. See Causal Observable implications, 166, 177, 195
mechanisms Omitted causal conditions, 6, 11, 55, 102,
Membership. See Comparative methods, 116, 118, 135, 138–40, 144, 152, 274
fuzzy set; Crisp set concepts; Set Ontological determinism, 15–19
theory Ontology, 16, 19, 61, 173
Metaphysic, 45 Operationalization of causal mechanisms
Method of agreement. See Comparative activities, 3–5, 18, 37–40, 64, 66, 68–
methods 72, 74, 80, 83
Meteor, 170 entities, 3, 9, 34, 38, 47, 52, 64, 69–73
Methodological alignment, 14 Operationalization of empirical tests,
Micro / macro debate, 12, 17. See also 187, 245, 258–60
Actor / micro level Outcome, 13, 14, 18, 21, 22–25
Midrange theory, 77, 153
Mill’s methods of agreement. See Com- Parsimony, 96
parative methods Parts of a mechanism
Mill’s methods of disagreement. See nonsystematic, 103, 284, 285
Comparative methods systematic, 45–46, 103
318 Index
Index 319