This article was downloaded by: [UQ Library]

On: 16 July 2015, At: 13:31

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: 5 Howick Place, London, SW1P 1WG

Security Studies
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/fsst20

Process Tracing and Qualitative Causal

Inference
David Waldner
Published online: 22 Jun 2015.

Click for updates

To cite this article: David Waldner (2015) Process Tracing and Qualitative Causal Inference, Security
Studies, 24:2, 239-250, DOI: 10.1080/09636412.2015.1036624

To link to this article: http://dx.doi.org/10.1080/09636412.2015.1036624

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
 Security Studies, 24:239–250, 2015
Copyright © Taylor & Francis Group, LLC
ISSN: 0963-6412 print / 1556-1852 online
DOI: 10.1080/09636412.2015.1036624

Process Tracing and Qualitative Causal

Inference

DAVID WALDNER

This essay critically reviews this symposium’s essays on process trac-

ing and security studies by James Mahoney, Andrew Bennett, and
Nina Tannenwald. It covers three major issues that have not been
Downloaded by [UQ Library] at 13:31 16 July 2015

adequately addressed by previous writings on process tracing: the

relationship of single case studies to more general causal claims,
the conceptualization of causation, and the criteria of valid causal
inference. It introduces the “completeness standard,” which com-
bines causal graphs, event history maps, and invariant causal
mechanisms. The completeness standard, it argues, bridges unit-
level causal inferences and average treatment effects, invokes an
epistemologically warranted conceptualization of causation, and
better satisfies existing standards of causal inference by making
unit homogeneity assumptions more credible.

The articles by Andrew Bennett, James Mahoney, and Nina Tannenwald in

this symposium are powerful statements of the value of process tracing for
scholars of security studies. Yet from the perspective of causal inference,
three critical issues remain under-developed: the relationship of single case
studies to more general causal claims, the conceptualization of causation,
and the criteria of valid causal inference. I survey each issue and suggest
some new ways of thinking about process tracing and causal inference. The
“completeness standard” I propose below is a procedure for making unit-
level causal inferences given the challenge of the fundamental problem of
causal inference.

David Waldner is Associate Professor of Politics at the University of Virginia. He is the

author of State Building and Late Development (Ithaca, NY: Cornell University Press, 1999).

239
 240 D. Waldner

AVERAGE TREATMENT EFFECTS

All three authors express strong preferences for using process tracing to study
particular historical cases. Mahoney identifies process tracing as a procedure
for developing explanations of individual historical cases, averring that “the
notion of an average causal effect . . . is not particularly useful when applied
to this kind of historical question.”1 Bennett privileges the particular as well,
observing that “policymakers are not very interested in knowing how most
countries usually respond to a policy instrument; rather, they want to know
how country X will respond this time to this instrument.”2 Tannenwald draws
connections to diplomatic history and urges process tracers to emphasize the
habit of “good storytelling” along with methodological rigor.3
Yet each author also gives reason, at least tacitly, to reconsider this
devaluation of average treatment effects. Tannenwald observes that an im-
portant contribution of process tracing is to identify relevant causal mecha-
Downloaded by [UQ Library] at 13:31 16 July 2015

nisms; we might reasonably assume that these mechanisms will have rele-
vance beyond the specific case being studied.4 Mahoney’s reconstruction of
procedures for explaining individual historical cases echoes this observation,
as it rests at crucial moments on “relevant preexisting theories and gener-
alizations.”5 Bennett excoriates George W. Bush administration officials for
ignoring “well-developed academic theories” that would have predicted a
high likelihood of civil conflict and a low likelihood of successful democ-
ratization in Iraq.6 Each of these remarks points to the value of knowledge
about average treatment effects.
The relevant question is not whether to privilege case-specific expla-
nations or average treatment effects over the other; the question is how
to advance our knowledge of the general and the specific simultaneously.
The standard I develop below is a procedure to merge the general and the
specific.

CAUSATION

Mahoney’s state-of-the-art essay identifies INUS conditions as the relevant

causal concept for explaining individual cases.7 John L. Mackie’s account of

1
James Mahoney, “Process Tracing and Historical Explanation,” Security Studies 24, no. 2 (April–June
2015): 202–203.
2 Andrew Bennett, “Using Process Tracing to Improve Policy Making: The (Negative) Case of the

2003 Intervention in Iraq,” Security Studies 24, no. 2 (April–June 2015): 229.
3 Nina Tannenwald, “Process Tracing and Security Studies,” Security Studies 24, no. 2 (April–June

2015): 227.
4 Ibid.
5 Mahoney, “Process Tracing and Historical Explanation,” 202.
6 Bennett, “Using Process Tracing to Improve Policy Making,” 239 (quotation, 9).
7 Mahoney, “Process Tracing and Historical Explanation.”
 Process Tracing and Qualitative Causal Inference 241

INUS conditions is an important amendment to more common understand-

ings of causation based on necessary, sufficient, and contributing conditions.8
But it also retains a well-known defect stemming from its derivation from
a Humean account of causation as constant conjunction. Consider Mackie’s
reconstruction of a particular event, a house fire. Suppose experts determine
that a short circuit in the home’s wiring caused the fire. The short circuit
cannot be a necessary condition because in its absence, other events, such
as a lightning strike or a turned-over lantern, could have caused the fire; nor
can it be a sufficient condition, because had there been no flammable mate-
rial nearby, the short circuit could not have caused a fire. The short circuit,
Mackie concludes, was an INUS condition, or “an insufficient but necessary
part of a condition which is itself unnecessary but sufficient for the result.”
But as Mackie subsequently emphasized, his account of INUS causation
is fully consistent with a regularity or Humean theory of causation, in which
one event can be said to cause another if the two events exhibit constant
Downloaded by [UQ Library] at 13:31 16 July 2015

conjunction.9 Any regularity theory of causation is subject to two substantial

criticisms. First, a regularity theory cannot distinguish accidental from causal
associations; sometimes one event (the administration of a vaccination) pre-
cedes a second event (the development of a serious medical complication)
without any causal connection between them. Second, a regularity theory of
causation leaves unopened the “black box” of causality. Consider, after all,
that Mackie’s accounts of individual fires are a particular instantiation of an
underlying causal model known as the “fire tetrahedron”:

Fires start when a flammable and/or a combustible material, in combi-

nation with a sufficient quantity of an oxidizer such as oxygen gas or
another oxygen-rich compound (though non-oxygen oxidizers exist that
can replace oxygen), is exposed to a source of heat or ambient tempera-
ture above the flash point for the fuel/oxidizer mix, and is able to sustain
a rate of rapid oxidation that produces a chain reaction.10

It should strike most readers as odd to talk about the causes of a fire with-
out any understanding of the underlying chemical processes that produce
fire, but this is precisely what Mackie’s INUS account does. This is not to
deny the importance of finding that a particular fire was caused by a short
circuit; rather it is to insist that a complete causal account of a fire requires
two types of features, the underlying causal model and the specific circum-
stances of the particular fire. Note, then, that insofar as we combine the
general causal model with the reconstruction of the particular fire, we satisfy

8 John L. Mackie, The Cement of the Universe: A Study of Causation (Oxford: Oxford University Press,
1974).
9 John L. Mackie, “Causes and Conditions,” in Causation, ed. Ernest Sosa and Michael Tooley

(Oxford: Oxford University Press, 1993), 52.

10 “Fire,” Wikipedia.com, http://en.wikipedia.org/wiki/Fire, accessed 24 September 2013.
 242 D. Waldner

the desideratum of merging the general (average treatment effects) and the
particular.
Mahoney complements his discussion of causation as INUS conditions
with discussion of sequence analysis and mechanisms. For many scholars,
the defects of the regularity or Humean account of causality prompts the
requirement that knowledge of causation requires knowledge of the un-
derlying mechanisms; this argument was introduced to the field of security
studies in David Dessler’s plea for a mechanism-based causal theory of war.11
Thus, Mahoney counsels process tracers to identify mechanisms such that X
→ M → Y. This is an excellent suggestion: producing causal chains marks
progress over establishing simple associations between X and Y , because
causal chains yield more fine-grained knowledge and they yield additional
opportunities for theory testing. The value of this suggestion is restricted,
however, because Mahoney defines mechanisms as “a factor that intervenes
between a cause and outcome. I treat mechanisms in the same way as causes
Downloaded by [UQ Library] at 13:31 16 July 2015

and outcomes; they are particular events or specific values on variables.”12

Defining mechanisms as variables or events self-evidently reproduces a reg-
ularity theory of causation, with all its attendant defects.
An alternative position, acknowledged but not endorsed by Mahoney, is
that mechanisms, to perform any useful epistemic function, must be under-
stood as invariant causal principles that generate the links between events.13
Mechanisms-as-events can be referred to as “weak” mechanisms, because in
this view mechanisms passively occupy a specific location in a causal chain
but they do not do any work; in contrast, mechanisms-as-invariant-causal-
principles are “strong mechanisms,” because mechanisms in this account per-
form work—they generate and propagate causal effects and hence actively
connect elements of the causal chain. Here is an example that illustrates
strong mechanisms. The fire tetrahedron, discussed above, is generated by
the underlying mechanism of combustion:

Combustion is the sequence of exothermic chemical reactions between

a fuel and an oxidant accompanied by the production of heat . . . The
release of heat can produce light in the form of either glowing or a
flame.14

11 David Dessler, “Beyond Correlations: Toward a Causal Theory of War,” International Studies

Quarterly 35, no. 3 (September 1991): 337–55.

12 Mahoney, “Process Tracing and Historical Explanation,” 206.
13 For discussion, see David Waldner, “What Are Mechanisms and What Are They Good For?”

Qualitative and Multi-Method Research 8, no. 2 (Fall 2010): 30–34; David Waldner, “Process Tracing
and Causal Mechanisms,” in The Oxford Handbook of Philosophy of Science, ed. Harold Kincaid (Oxford:
Oxford University Press, 2012), 65–84. Alexander L. George and Andrew Bennett, Case Studies and Theory
Development in the Social Sciences (Cambridge: MIT Press, 2005), 131–45, provides an early statement of
a mechanistic position, albeit without the critical property of invariance.
14 “Combustion,” Wikipedia.com, http://en.wikipedia.org/wiki/Combustion, accessed 24 September

2013.
 Process Tracing and Qualitative Causal Inference 243

What does it mean to refer to combustion as an invariant causal principle?15

The chain reaction that constitutes combustion in oxygen is dependent upon
the unusual structure of the dioxygen molecule, specifically the spin and
angular momentum of its electrons. The subatomic structure, in other words,
is a constitutive element of its action. Thus, given the conditions described by
the underlying causal model, the structure and derived action of the dioxygen
molecule ensures that combustion invariably occurs, generating the heat and
light that we experience as fire. A causal account of fire, then, requires
more than intervening events, like short circuits or lightning strikes, to use
Mackie’s examples; it requires bringing together the underlying causal model,
its invariant causal mechanisms, and understanding the concrete instantiation
of the causal model in the specific circumstances.16
Mahoney and others who define mechanisms as intervening
events—weak mechanisms, in other words—might raise three concerns
about this understanding of strong mechanisms: that the account is philo-
Downloaded by [UQ Library] at 13:31 16 July 2015

sophically unsatisfactory, that knowledge of strong mechanisms is simply not

available to political scientists, or that the inordinate efforts needed to obtain
knowledge of strong mechanisms is not epistemically warranted. There is no
space here to address the philosophical critiques of strong mechanisms, and
I will briefly address the pragmatic concerns at the conclusion of this essay.
The remainder of this essay will thus focus on the epistemic virtues of strong
mechanisms. This seemingly abstruse discussion of the nature of mechanisms
matters because it establishes justification for using process tracing for causal
inference. My proposal, elaborated below, is that we can combine the idea
of causation as strong mechanisms with the potential outcomes framework
to resolve the fundamental problem of causal inference.17

INFERENTIAL VALIDITY

What does it mean to make a causal inference based on process tracing?

Borrowing from Van Evera’s initial formulation, Mahoney equates causal
inference to the passage of two main types of hypothesis tests: “hoop tests,”
whose passage is necessary but not sufficient for the truth of a hypothesis that

15 See Peter Machamer, Lindley Darden, and Carl F. Craver, “Thinking about Mechanisms,” Philosophy

of Science 67, no. 1 (March 2000): 1–25; Stuart Glennan, “Rethinking Mechanistic Explanation,” Philosophy
of Science 69, no. 3 (September 2002): S342–S35, for important philosophical statements of mechanisms
as invariant causal processes.
16 Of course the underlying causal model and mechanisms may be suppressed when specific causal

accounts are given; that the model is often taken as background knowledge is an indicator of our
confidence in the account, not an indicator of its insignificance.
17 The fundamental problem of causal inference is discussed by Paul Holland, “Statistics and Causal

Inference,” Journal of the American Statistical Association 81, no. 396 (December 1986): 945–60; Gary
King, Robert L. Keohane, and Sidney Verba, Designing Social Inquiry: Scientific Inference in Qualitative
Research (Princeton, N.J.: Princeton University Press, 1994), 79.
 244 D. Waldner

explain some event, and “smoking-gun tests,” whose passage is sufficient but
not necessary for the truth of a hypothesis.18 Furthermore, our beliefs about
the truth status of a hypothesis and its rivals conditional on these tests can be
expressed in terms of Bayesian probabilities.19 As Tannenwald summarizes,
“Not all evidence is equal.”20 Some evidence might have such high probative
value that it allows us to support (provisionally and fallibly, to be sure) one
hypothesis and reject its rivals.
There is much to be said about this well-articulated and highly valuable
framework of hypothesis testing. Let me confine my remarks to three points:
interpretive debates about the status of a hypothesis test; ambiguity about
how much of a process must be tested before a hypothesis alleged to explain
it “passes” the test; and lingering questions about whether hoop and smoking
gun tests, even with the Bayesian armature, are sufficient to satisfy the criteria
of causal inference.
As we know, hoop and smoking gun tests are, respectively, necessary
Downloaded by [UQ Library] at 13:31 16 July 2015

but not sufficient and sufficient but not necessary for the acceptance of a
hypothesis. But necessity and sufficiency are derived characteristics; Van Ev-
era originally wrote of degrees of uniqueness and certitude. The change is
subtle but highly consequential, because, as Van Evera notes, “interpretive
disputes also arise from quarrels over the uniqueness and certitude of predic-
tions.”21 Uniqueness and certitude, it must be emphasized, are not objective
features of a hypothesis. Perhaps in many cases there will be ready assent
to the degree of uniqueness or certainty of a set of hypotheses; but it is
equally reasonable to note, as Van Evera did, that many hypotheses lack
either uniqueness or certainty, making them “straw-in-the-wind tests” with
limited inferential value. Tannenwald’s discussion of rival process tracing
accounts of the end of the Cold War illustrates the difficulty of arbitrating
these interpretive disputes.22 At minimum, we should ask for clear guide-
lines for measuring the degree of uniqueness and certainty; these features of
a test must be demonstrated, not asserted. And as David Collier has already

18 Stephen Van Evera, Guide to Methods for Students of Political Science (Ithaca, NY: Cornell Uni-

versity Press, 1994), 30–34; Mahoney, “Process Tracing and Historical Explanation.” For his efforts to
distinguish easy and hard tests, see James Mahoney, “The Logic of Process Tracing Tests in the Social
Sciences,” Sociological Methods & Research 41, no. 4 (November 2012): 570–97.
19 Andrew Bennett, “Process Tracing: A Bayesian Perspective,” in The Oxford Handbook of Political

Methodology, ed. Janet M. Box-Steffensmeier, Henry E. Brady, and David Collier (Oxford: Oxford Uni-
versity Press, 2008), 702–21. More recently, Bennett has suggested modifications of the basic framework
that relax the unnecessarily restrictive categorical logic of ncessity and sufficiency. See Andrew Ben-
nett, “Process Tracing with Bayes: Moving beyond the Criteria of Necessity and Sufficiency,” Qualitative
and Multi-Method Research 12, no. 1 (Spring 2014): 46–51, which draws on Macartan Humphreys and
Alan Jacobs, “Mixing Methods: A Bayesian Unification of Qualitative and Quantitative Approaches,” pre-
sented at the Annual Meeting of the American Political Science Association, Chicago, August 2013. These
developments complicate but do not fully obviate the critical points raised below.
20 Tannenwald, “Process Tracing and Security Studies,” 226.
21 Van Evera, Guide to Methods, 33.
22 Tannenwald, “Process Tracing and Security Studies.”
 Process Tracing and Qualitative Causal Inference 245

observed, it will often be difficult to establish distinctions between types of

tests, a state-of-the-world that must surely give pause to advocates of process
tracing.23
Next, consider this question: how many steps of a process must be traced
before we consider a hypothesis to be (provisionally) true? In the Bayesian
approach, after all, discovery of even small amounts of highly surprising
evidence about one link in a causal chain could in principle be sufficient
to raise the posterior probability to a sufficiently high level to confirm a
hypothesis and refute its rivals. Yet this point clashes with the nominal
meaning of process tracing—to trace a process must surely mean something
more than finding highly surprising evidence about one or a few links in the
argument. Indeed, in an early and still influential statement of process tracing,
Alexander L. George and Andrew Bennett propose that “all the intervening
steps in a case must be predicted by a hypothesis, or else that hypothesis
must be amended—perhaps trivially or perhaps fundamentally—to explain
Downloaded by [UQ Library] at 13:31 16 July 2015

that case.”24 Yet the Bayesian logic favored by Mahoney and Bennett does
not necessarily imply this continuity criterion, which is conspicuously absent
from virtually all recent writings on process tracing.
Suppose, after all, that an analyst is confident that there are only five
plausible hypotheses to explain outcome Y. Each of the first four hypotheses
fails a hoop test on its independent variable and so all four are eliminated;
the fifth survives its hoop test. Must we continue tracing the process linking
X to Y, or shall we be satisfied by eliminative induction on the independent
variables? From a Bayesian perspective, the answer depends upon the prior
and posterior likelihoods. Yet from the perspective of categorical logic, all
rivals have been eliminated and hence the sole survivor must be confirmed,
even without tracing the entire causal chain. Or suppose the first four of
these hypotheses each passes a hoop test on their independent variable and
so are not eliminated; but the fifth hypothesis passes both a hoop test and
a smoking gun test on its independent variable. By the categorical logic
of smoking gun tests, this passage should be sufficient for the truth of the
hypothesis. Must we continue tracing the process linking X to Y? It would
be highly valuable—perhaps even obligatory—for process tracers to address
these questions, for they seem to imply that the systematic tracing of an
entire process is not necessary for making inferences.
Finally, let us return to the question of whether the current approach
to process tracing, when successful, is equivalent to causal inference. I have
suggested that critical components of current statements of process tracing
rest on Humean or regularity notions of causation that are insufficient to
warrant causal inference. Explicating the Bayesian logic of process tracing

23 David Collier, “Understanding Process Tracing,” PS: Political Science and Politics 44, no. 3 (October

2011): 825.
24 George and Bennett, Case Studies and Theory Development, 207.
 246 D. Waldner

does not alter this proposition. Bayesianism is an “all-purpose” statement

about how to express degrees of belief in order to compare hypotheses;
nothing about Bayesian probabilities or Bayesian statistics implies that a
hypothesis is causal, as opposed to a descriptive hypothesis. Indeed, Bennett
explicitly links process tracing to descriptive inference, not causal inference.25
Current understandings of causal inference, in contrast, place heavier
burdens on drawing valid causal inferences from observational data.26 A
powerful challenge to process tracers is the fundamental problem of causal
inference. Define a causal effect as the difference in an outcome under treat-
ment and under control—the presence or absence of the explanatory cause.
The problem is that we cannot observe a unit under treatment and under
control simultaneously. Because of this problem of missing data, current un-
derstandings of causal inference imply that unit-level causal inference—what
process tracers call “within-case analysis”—is simply impossible.
Process tracers must take this fundamental challenge seriously. The
Downloaded by [UQ Library] at 13:31 16 July 2015

claim I advance here is that the combination of strong mechanisms with

the potential outcomes framework can support an assumption of unit homo-
geneity, permitting unit-level causal inference.27

WHAT NEXT FOR PROCESS TRACING?

I have suggested three challenges for process tracers: to resolve the ten-
sion between the particular and the general (between unit-level and average
treatment effects); to articulate and operationalize a non-Humean notion of
causation; and to justify using process tracing as a means of valid causal
inference, given the constraints of the fundamental problem of causal infer-
ence. Here I show how to address all three concerns using what I have called
elsewhere the “completeness standard.”28 I illustrate the proposal using John
Owen’s work on the liberal origins of the democratic peace.29

25 Bennett, “Using Process Tracing to Improve Policy Making.”

26 For the emergent literature on causal inference, see Judea Pearl, Causality: Models, Reasoning, and
Inference (Cambridge: Cambridge University Press, 2000); Stephen L. Morgan and Christopher Winship,
Counterfactuals and Causal Inference: Methods and Principles for Social Research (Cambridge: Cambridge
University Press, 2007); Thad Dunning, Natural Experiments in the Social Sciences: A Design-Based
Approach (Cambridge: Cambridge University Press, 2012); John Gerring, Social Science Methodology: A
Unified Framework, 2nd ed. (Cambridge: Cambridge University Press, 2012).
27 For formal treatment of unit-level qualitative causal inference, see David Waldner, “Aspirin,

Aeschylus, and the Foundations of Qualitative Causal Inference,” unpublished manuscript, University
of Virginia, 2015.
28 For further discussion and illustrations of the completeness standard, see David Waldner, “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 T. Checkel (Cambridge: Cambridge University Press, 2014), 126–52.
29 John M. Owen IV, Liberal Peace, Liberal War: American Politics and International Security (Ithaca,

NY: Cornell University Press, 1997); Owen, “How Liberalism Produces Democratic Peace,” International
Security 19, no. 2 (Autumn 1994): 87–125.
 Process Tracing and Qualitative Causal Inference 247

FIGURE 1 Causal Graph for Owen, Liberalism and Democratic Peace.

In order to effect a merger between the general (average treatment ef-

fects) and the particular (unit-level treatment effects), I propose first that
process tracers articulate a causal graph. Causal graphs, or directed acyclic
Downloaded by [UQ Library] at 13:31 16 July 2015

graphs, are composed of nodes or vertices representing random variables

(features of a causal system that can undergo change) and directed edges
that connect nodes and thus represent relations of probabilistic causal de-
pendence due to the “strong” causal mechanism left tacit in the graph and
represented by the arrow. Two nodes connected by an arrow have this pre-
cise meaning: the parent node is a direct cause of its descendant such that an
exogenous intervention on X will change the distribution of the descendant
variable holding all other variables in the model fixed.30 The causal graph is
thus a complete statement of the causal relations (the average treatment ef-
fects) connecting X and Y. Figure 1 is an example of a causal graph, copied
directly from Owen’s 1994 International Security article.31 Liberalism mat-
ters, in Owen’s account, because it initiates a chain of direct causal relations
that terminates in the democratic peace. Intervening on any non-terminal
node to set that random variable to a new value would therefore alter the
probability distribution of the democratic peace.
My second proposal is that alongside of a causal graph, process tracers
articulate an event-history map for each particular case they study. Whereas
causal graphs represent average treatment effects, event-history maps repre-
sent unit-level causal effects. A causal graph is an abstract representation of
an historical narrative. A scholar does not trace the process depicted by the
graph, but rather the process constituted by events that instantiate the graph,
in all their contextual specificity. Although the syntax of the event-history
map resembles that of the causal graph, the former depicts singular events

30 James Woodward, Making Things Happen: A Theory of Causal Explanation (Oxford: Oxford

University Press, 2003).

31 Owen, “How Liberalism Produces Democratic Peace,” 102. Owen refers to “causal pathways” and

does not use the concepts of a causal graph, random variables, or joint probability distributions.
 248 D. Waldner

FIGURE 2 Event-History Map for Franco-American Crisis, 1796–98

and thus does not represent a joint probability distribution. The combina-
tion of a causal graph with a set of event-history maps thus bridges the gap
Downloaded by [UQ Library] at 13:31 16 July 2015

between the search for generality and the privileging of specific historical
outcomes and explanations. Figure 2 is a possible depiction of the events
in one of Owen’s early case studies, Franco-American relations from 1796
to 1798.32 By my proposal, Owen would construct a separate event-history
map for each of his ten historical case studies.
The third step in the procedure I am advocating is descriptive inference,
or checking the correspondence between the event-history map and the
causal graph. This procedure uses the standard tools of measurement theory:
construct validity, measurement reliability, and measurement validity. Do the
events in the map represent the conceptual connotation of the corresponding
node? Is the author’s evidence sufficient to confirm the descriptive inference?
For example, corresponding with the node “refusal to fight democracies =
true” in the causal graph is the event-history “republicans agitate against war
with France.” Here we ask whether the evidence about the events support the
event-history as a concrete historical representation of the abstract node in
the causal graph. Owen provides a detailed and well-documented historical
narrative of Republican members of Congress introducing legislation against
the war and refusing to vote for a war declaration.33 Owen gives substantial
empirical support to his inference that although President John Adams clearly
wanted war with France, going so far as to draft a war message to Congress,
he did not present the message to Congress because he knew that Republican
opposition would block its passage. Notice that Owen provides three direct
pieces of evidence: Republican actions in Congress, such as introducing

32 Owen does not formally and graphically represent his narrative, and the historical narrative does

not precisely correspond to the template set by the causal graph. I thank Owen for helping me with the
reconstruction of this event-history map.
33 Owen, “How Liberalism Produces Democratic Peace,” 86–87; Owen, Liberal Peace, Liberal War,

107.
 Process Tracing and Qualitative Causal Inference 249

anti-war resolutions; Adams’s drafting of a war message; and Adams’s non-

presentation of the message to Congress. These three “primitive” facts can be
directly observed. The reason that Adams did not present the war message,
on the other hand, is an inference from the primitives, an inference based on
a tacit causal mechanism—a decision to not present a request for war that
had no chance of being passed. This inference would only be made stronger
by the inclusion of formal discussion of the underlying decision-theoretic
model.
We can make two points about this procedure of making descriptive
inferences from event-history to causal graph. First, the entire procedure
should be as transparent and replicable as possible. Providing both the
causal graph and the event-history map and explicitly denoting moments
of inference should aid transparency and replicability.34 Second, the existing
apparatus of hypothesis testing—hoop tests, smoking gun tests, and Bayesian
probability—is centrally important to descriptive inference from event-history
Downloaded by [UQ Library] at 13:31 16 July 2015

map to causal graph. For example, we might interpret Republican anti-war

agitation as a hoop test, while Adams’s drafting of a message to Congress but
its non-presentation to Congress could fruitfully be interpreted as a smoking
gun test.
The final step in the completeness standard is causal inference: demon-
strating that the causal links represented in the causal graph are instantiated
as causal connections between subsets of events in the event-history map.
Causal inference relies on causal mechanisms with invariant generative ca-
pacities; these capacities make intervention possible. Ideally, process tracers
will identify causal mechanisms for each arrow connecting two nodes in the
causal graph. It is the identification of these causal mechanisms with their
critical property of invariance that allows us to justify the assumption of unit
homogeneity and hence make unit-level causal inferences.
Owen does not explicitly identify these strong mechanisms; for the most
part, he too adopts a perspective of weak mechanisms and implicit causal
claims. Sometimes implicit mechanisms can easily be made explicit; for ex-
ample, institutional rules may simply prevent a president from going to war
once Congress refuses to pass a declaration of war. At other times, the causal
mechanism will not likely be so easily recovered. Indeed, close scrutiny of
causal claims might undermine their initial plausibility. For example, some of
the nodes in Owen’s causal graph, such as liberal ideas and liberal ideology,
are not conceptually distinct from one another and so we might claim that
they cannot exist in a relationship of causal dependence on one another.
My overall claim, then, is that process tracing that satisfies the complete-
ness standard has four components: a causal graph, a set of event-history
maps, a set of descriptive inferences, and a set of causal inferences via strong

34 An additional key tool of transparency is active citation, as advocated by Andrew Moravcsik,

“Trust, but Verify: The Transparency Revolution and Qualitative International Relations,” Security Studies
23, no. 4 (October 2014): 663–88.
 250 D. Waldner

mechanisms that support a unit-homogeneity assumption and thus address

the fundamental problem of causal inference. This proposal clearly addresses
the tension between the general and the specific; it is based on a clear and
defensible notion of causation (generative causal mechanisms that make in-
terventions possible); and it sets relatively clear standards of valid causal
inference. Indeed, the proposal has important heuristic features; it forces
process tracers to articulate chains of direct causation, separate events from
variables, and identify causal mechanisms. Each of these procedures can help
stimulate theoretical insight. The proposal has important evaluative features
as well; the entire procedure is methodologically more transparent, allowing
us to distinguish good process tracing from bad. Finally, it incorporates the
existing methodological practices of conducting hoop and smoking gun tests
and interpreting their implications using Bayesian probabilistic logic.
The completeness standard places a heavy methodological burden on
process tracers; the payoff is the epistemic virtue of providing more complete
Downloaded by [UQ Library] at 13:31 16 July 2015

explanations that can potentially satisfy the demands of unit-level causal in-
ference. The framework can accommodate other understandings of process
tracing as well. Some qualitative scholars prefer to immerse themselves in
context and sequence; they look to process tracing as disciplined “soaking
and poking,” to use Richard Fenno’s time-honored phrase. Think of this
inductive narrative approach as the construction of an event-history map
without the corresponding construction of a causal graph. Mahoney ad-
vocates using process tracing for theory discovery, extracting causal factors
from fine-grained analysis of a case.35 Think of this approach as the inductive
construction of a causal graph from the details of an event-history map, con-
verting events into random variables. The completeness standard is also fully
consistent with Bayesian hypothesis testing, which would be used to make
inferences from event-history maps to causal graphs. The additional struc-
ture of the completeness standard, however, promises substantial epistemic
value-added. These various uses of process tracing all have considerable
value, but scholars using them should carefully tailor the claims that they
make to the standards that they satisfy. Scholars using process tracing that
does not identify invariant causal mechanisms and that does not identify an
alternative instrument for overcoming the fundamental problem of causal
inference should adjust their claims of causal inference accordingly.

ACKNOWLEDGMENTS

The author thanks Andrew Bennett, Colin Elman, James Mahoney, John
Owen, and two anonymous reviewers for Security Studies for their very
helpful comments on earlier drafts of this essay.

35 Mahoney, “Process Tracing and Historical Explanation.”