7 Process tracing, causal inference,
and civil war
Jason Lyall
Introduction
Process tracing is an invaluable tool in the civil war scholar’s toolkit. Or, rather,
it should be, for it provides the ability to move beyond statistical association
toward causal inference about why (and how) outcomes are produced in civil
war settings. Yet, scholars have too often neglected its use. Instead, great pains
have been taken to construct research designs that (at best) are able to identify
suggestive correlations between variables, but lack the ability to test the
mechanism(s) at work. Qualitative research is not immune to this criticism,
either, for process tracing, when properly conducted, establishes a standard for
rigor that often goes unmet even in detailed historical cases (see also
Evangelista, this volume, Chapter 6). This is an unfortunate state of aﬀairs;
without understanding the causal processes that underpin associations, we
foreclose opportunities to advance our theories of civil war and contribute to
debates about the eﬃcacy of diﬀerent policies in violent settings.1
This chapter emphasizes the practicalities of marrying design-based inference
with the strengths of process tracing to improve our ability to build and
(especially) test theories about civil war onset and dynamics. Bennett and
Checkel’s ten best practices for process tracing (this volume, pp. 20–31)
provide a springboard for a discussion of how to identify and conduct
rigorous process tracing in settings marked by poor (or no) data, security
concerns, and ﬂuid events. The chapter also introduces ideas from the nowburgeoning
literature on causal inference to help guide decisions about case
I thank Jeﬀ Checkel and Andrew Bennett for helpful comments on an earlier draft, and Helinna Ayalew,
Nicole Pﬂug, Andrey Semenov, and Boloroo Uuganbayar for excellent research assistance. Support from the
Air Force Oﬃce of Scientiﬁc Research (Grant FA#9550-09-1-0314) is gratefully acknowledged. All errors
are mine.
1
Civil war is deﬁned here as an armed confrontation resulting in at least 1,000 battle deaths between two or
more combatants that were subject to the same political and legal system prior to the war.
18:06:25
selection and evidentiary standards. In particular, the approach advocated
here draws on a potential outcomes framework that hinges on the use of
counterfactual observations, “elaborate” theory, and qualitative evidence on
treatment assignment to facilitate drawing causal inferences about why wars
break out and how they are fought (see also Dunning, this volume, Chapter 8,
on using process tracing to assess assignment to treatment).
I proceed as follows. The ﬁrst section details the near absence of process
tracing as a methodological approach in journal articles published since 1994
on civil war onset and dynamics. The second section draws on Elisabeth
Wood’s Insurgent Collective Action and Civil War in El Salvador (2003) as
an illustration of Bennett and Checkel’s ten “best practices” of process tracing.
The third section discusses four additional “best practices” that arise from the
causal inference literature and that are especially likely to be useful in civil war
settings. Next, I detail potential research designs and the utility of process
tracing for two literatures: the cross-national study of why civil wars break out,
and the micro-level (for example, subnational) study of civilian victimization
and its eﬀects on subsequent participation in an insurgency. A ﬁfth section
brieﬂy details the ethical and practical challenges faced by researchers in these
environments. I conclude with thoughts about the use of process tracing to
further our theoretical and practical understandings of civil war.
Process tracing and civil war
The meteoric rise of research on civil war has largely centered around two
questions. One research agenda, heavily dominated by cross-national statistical
analyses of the post-1945 era, has sought to explain civil war onset. These
studies seek to draw an association between structural factors – state capacity,
lootable resources, and ethnic exclusion from political power, to name three –
and the outbreak of civil war. A second research program has drawn on a
“micro-level” framework that explores the dynamics of violence – including
its location, nature, and timing, especially toward civilians – at the subnational
level. Unlike cross-national studies, these micro-level studies typically pay
close attention to identifying the causal relationship between independent
variables and outcomes using disaggregated time-series data and a host of
sophisticated approaches, including quasi- and natural experiments, matching,
and instrumental variable regression.
What role has process tracing played in these two research programs? Very
little, it turns out. Figure 7.1 plots the sharp increase in the number of articles
187 Process tracing, causal inference, and civil war
18:06:25
published annually in ﬁfteen political science journals on the topic of civil war
from 1995 to 2012.2
Of these 448 articles, only 12 explicitly claim to be
employing process tracing (all since 2004). While other work may be drawing
implicitly on process-tracing insights (Checkel 2013b: 6), these articles reﬂect
a more general trend away from the use of case studies, process tracing’s
natural habitat. Indeed, the share of articles with evidence from at least one
case study has fallen from 80 percent in 1995 to about 50 percent in 2012. Over
the past ﬁve years, an average of 44 percent of articles have had some form of
case study, generously deﬁned as a systematic discussion of a particular
historical case at least four paragraphs in length.
The curious under-utilization of process tracing in civil war studies to date
likely has several causes. Detailed process tracing can be diﬃcult to execute
within journal word limits. Perhaps relatedly, both cross-national and microlevel
studies have increasingly adopted research designs built to measure the
PublishedArticles
0
1994 1996 1998 2000 2002 2004 2006 2008 2010 2012
10
20
30
40
50
Figure 7.1 Number of articles published on civil war onset or dynamics in ﬁfteen political science journals,
1995–2012
Note: The sample consists of 448 articles. Review articles and those in related ﬁelds (for example,
genocide studies) were not included.
2
The journals surveyed include: American Political Science Review, American Journal of Political Science,
Perspectives on Politics, Journal of Conﬂict Resolution, International Organization, International Security,
Journal of Peace Research, Security Studies, Journal of Politics, World Politics, Comparative Politics,
Comparative Political Studies, Civil Wars, Terrorism and Political Violence, and International Studies
Quarterly.
188 Jason Lyall
18:06:25
direction and magnitude of the relationship between independent variables and
outcomes rather than the mechanisms that underpin this relationship. This is a
pragmatic move for research programs in their early stages. It can be diﬃcult
enough simply to identify the existence of a relationship given the multiple
threats to inference, poor or absent data, and noisy proxy measures that often
characterize research in conﬂict settings. Moreover, research designs that are
tasked with establishing associations between variables may not be suitable for
testing mechanisms. Yet, without moving beyond correlation, we are left blind
about the processes and dynamics that drive these relationships, impoverishing
both our theories and our ability to contribute to policy debates.
Process tracing in action: an example
The apparent neglect of process tracing in journal articles notwithstanding,
there are still exemplars of the craft within political science and civil war
studies. I use Elisabeth Wood’s (2003) book, Insurgent Collective Action and
Civil War in El Salvador, as an illustration of the ten “best practices” of process
tracing outlined by Bennett and Checkel (this volume, pp. 20–31).3
Insurgent
Collective Action tackles the twin questions of why peasants supported (and
joined) an armed insurrection against El Salvador’s government during the
1970s and 1980s and how that participation evolved over time. Wood’s argument,
developed inductively and deductively in equal measure, is a nuanced
one. Individuals supported the armed opposition, she argues, through a series
of emotional mechanisms, including a belief in the moral purpose of acting,
deﬁance in the face of state repression, and “pleasure in changing unjust social
structures through intentional action” (Wood 2003: 235). More simply, pride
in the “authorship” of their wartime actions (ibid.: 231) led some individuals to
eschew the relative safety of fence-sitting in favor of risky acts that carried no
credible promise of immediate (or future) material pay-oﬀ.
This interpretation of high-risk collective action is pitted against alternative
explanations that emphasize the need for material incentives (Olson 1965;
Popkin 1979), protection from state violence (Mason and Krane 1989;
Goodwin 2001), or strong horizontal networks among peasants (Moore
1978; Scott 1976) to induce participation. In the language of this volume’s
best practices, Wood clearly “casts her net widely” for alternative explanations
3
Waldner, this volume, Chapter 5, assesses Wood’s use of process tracing in a diﬀerent book – and comes
to similar conclusions on its quality.
189 Process tracing, causal inference, and civil war
18:06:25
(criterion 1). She is also equally tough on these alternative explanations
(criterion 2), marshaling an impressive array of ethnographic evidence from
prolonged ﬁeldwork to build her case.
To test these claims about the connection between emotions and participation,
Wood initially engaged in eighteen months of ﬁeldwork in four diﬀerent
sites in Usulután, a wealthy but conﬂicted department of El Salvador, and one
site in Tenancingo in the northern department of Cuscatán.4
Interviews with
200 campesinos, all but 24 of whom participated in the insurgency in some
fashion, and mid-level Farabundo Martí National Liberation Front (Frente
Farabundo Martí para la Liberación Nacional, or FMLN) commanders comprise
the bulk of her evidence. In a particularly innovative (and non-intrusive)
practice, twelve campesino teams engaged in collective map-making during
three workshops in 1992 to provide a window into how peasant culture,
especially pride in collective achievements, manifested itself. Wood is alert
to the potential biases of her sources (criterion 3), particularly the problems
associated with memory and (selective) recall of wartime activities. She also
notes that her interviewees were not randomly selected, but instead chosen
through campesino organizations, skewing her sample toward individuals
who participated in the insurgency.
These materials, and the process of gathering them, enable Wood to generate
inductively a wealth of insights (criterion 8). Yet, Wood’s empirical
claims do not rest solely on induction, for she also outlined the argument a
priori using a formal model of individual decision-making (Wood 2003:
267–274). The micro-level motives for individual actions are also supported
by insights from laboratory experiments developed by social psychologists. As
a result, the book’s argument draws on both inductive and deductive
approaches to discipline its data gathering and to identify the speciﬁc processes
that lead to campesino participation (criterion 9).
Wood selected her ﬁve ﬁeld sites according to a fourfold criterion: their
accessibility to an outside researcher; the presence of both supporters and
non-supporters (for example, the regions had to be “contested”); variation in
agrarian economies (to examine multiple pathways that peasants could take into
the insurgency); and the presence of only one or two guerrilla factions (Wood
2003: 52–54). Taken together, it appears that these regions do oﬀer representative
examples of broader patterns of participation and violence in El Salvador’s
contested areas. What remains unclear, however, is whether these cases represent
a “most likely” or “least likely” test for alternative explanations (criterion 4).
4
The book draws on additional research and visits over the following twelve years (Wood 2003: xiii).
190 Jason Lyall
18:06:25
By truncating variation on the degree of state control or rebel presence, we may
be working outside scope conditions where material incentives or desire for
protection from state violence are most operative, for example.
Moreover, while Wood’s “starting rule” (criterion 5) is clearly justiﬁed –
sometimes researchers must simply take advantage of opportunities to start
work that are created exogenously by lulls in ﬁghting – her “stopping rule” is
less clear (criterion 6). It appears that repetition in the campesino’s own stories
for why they participated was the decision rule for ceasing data collection;
once the researcher has heard the same stories repeated across diﬀerent
respondents, data collection stops.
In this instance, however, the process tracing is not necessarily conclusive
(criterion 10). The decision to over-sample participants, for example, even
though two-thirds of the population did not participate meaningfully in the
insurgency (Wood 2003: 242), could overestimate the importance of emotive
mechanisms. Wood herself notes how past patterns of state violence and
proximity to insurgent forces (ibid.: 237–238) conditioned whether these
emotions could be acted upon. Sorting out the relative causal weight between
emotions and mechanisms of control or prior exposure to violence would
require additional interviews among non-participants both within and outside
of these ﬁve areas. Not all process tracing is deﬁnitive – indeed, the best
examples typically raise more questions that could be tackled by adjusting the
research design or sample frame to provide additional empirical leverage on
the original process under study.
Avoiding “just-so” stories: additional best practices
In the spirit of this volume’s emphasis on practicality, I oﬀer four additional
process-tracing best practices that can help researchers avoid “just-so” stories
when exploring civil war dynamics. These include: (1) identifying counterfactual
(“control”) observations to help isolate causal processes and eﬀects;
(2) creating “elaborate” theories where congruence across multiple primary
indicators and auxiliary measures (“clues”) is used to assess the relative
performance of competing explanations; (3) using process tracing to understand
the nature of treatment assignment and possible threats to causal
inference; and (4) out-of-sample testing. The emphasis here is on situations
where researchers wish to test empirical claims, but cannot randomize the
“treatment” (for example, state violence, rough terrain, etc.) due to practical
limitations or ethical concerns.
191 Process tracing, causal inference, and civil war
18:06:25
First – and taking the Rubin Causal Model (RCM) as a point of departure –
I emphasize the need for counterfactual reasoning to measure causal eﬀects
(Rubin 2006; Rosenbaum 2010; see also Evangelista, this volume, Chapter 6).
The intuition here is a simple one: every unit – be it a village, province, or
state – has a diﬀerent potential outcome depending on its assignment to a
particular treatment. Since we cannot by deﬁnition observe all outcomes in
the same unit, we must engage in counterfactual reasoning to supply the
“match” (or “control”) for the unit where an outcome was unobserved. The
more similar the control and treated observations along the values of their
independent variables, the greater the conﬁdence we have in our estimates of
the treatment’s causal eﬀects.
The comparative nature of the RCM framework strengthens inferences
from process tracing in several ways. By matching treated and control observations,
the number of possible alternative explanations is reduced, simplifying
the task of process tracing since some (ideally all but one, but hopefully
many or even most) mechanisms are being held constant by a research design
pairing cases that have similar values on independent variables. Process
tracing can then be used to assess whether the treatment variable and the
variables that could not be properly controlled for might account for observed
outcomes. More generally, without the counterfactual, we cannot rule out the
possibility that the same causal process is present in both the treated and
control cases. To be conﬁdent about one’s inferences, within-case process
tracing should thus be paired with cross-case process tracing in a control
observation where the presumed relationship between treatment and outcomes
is not present.
The RCM framework also provides a natural bridge to emerging Bayesian
approaches to process tracing (Bennett, this volume, Appendix; see also Beach
and Pedersen 2013a: 83–88).5
At its core, the Bayesian principle of “updating”
one’s prior beliefs in light of new evidence hinges on counterfactual reasoning.
Bayesian updating is guided by the prior probability of a theory’s validity and
the likelihood ratio between “true positives” (instances where the evidence
suggests a theory is true and the theory is in fact true) and “false positives”
(instances where the evidence is consistent with a theory, but the theory itself
is in fact false). The likelihood ratio itself relies, often implicitly, on control
observations to provide both aﬃrmative evidence for the preferred theory and
eliminative induction that rules out alternative explanations and the possibility
that a theory’s claims are false. As Bayesian reasoning underscores, ruling
5
See also Humphreys and Jacobs 2013: 20–22.
192 Jason Lyall
18:06:25
out alternative explanations can sometimes generate greater discriminatory
power for a test between hypotheses than discovering evidence that (further)
conﬁrms a preferred theory’s validity.
Second, scholars should craft elaborate theories (Rosenbaum 2010: 329) that
articulate multiple measures for the mechanism(s) at work (see also Jacobs, this
volume, Chapter 2; Schimmelfennig, this volume, Chapter 4). If multiple
mechanisms are thought to be present, then the sequence by which a process
or eﬀect is created should also be mapped out.6
These hypotheses and measures
should be speciﬁed before moving to empirical testing. Backward induction
from a known outcome to the mechanisms that produced the outcome should
be avoided, especially if counterfactuals are not used to eliminate the possibility
that these mechanisms are also present in control cases.
Specifying multiple measures a priori enables the researcher to test for the
congruence between these observations, helping to diﬀerentiate competing
explanations that might rely on the same mechanism to explain an outcome.
Put diﬀerently, the comparative strength of a particular argument may be
decided not on the strength of evidence linking a variable to a mechanism, but
instead on its ability to account for auxiliary observations as well as the
sequence producing the outcome itself. From a Bayesian perspective, these
auxiliary observations are “clues’’ that can shift beliefs about a theory’s validity
since their presence denotes that a speciﬁed process – and only that process –
is responsible for the observed outcome.7
Third, treating potential outcomes explicitly also focuses one’s attention on
the key question of treatment assignment. The non-random nature of most
“treatments” that interest civil war scholars means dealing with a host of
methodological issues that can frustrate causal inference. Process tracing can
help here, too. Qualitative data can be used to trace how the treatment was
assigned to treated and control units, for example, a procedure Thad Dunning
in Chapter 8 refers to as a treatment-assignment causal-process observation
(see also Dunning 2012: 209). Understanding how the treatment was assigned,
and whether it was truly assigned “as-if” random across units, is pivotal for
micro-level studies that rely on natural or quasi-experiments to ﬁnd starting
points in the dynamics of civil war violence. Tracing the logic of assignment is
especially important when evidence for these conditioning variables is private
information among combatants, making it diﬃcult to match across cases.
6
In Chapter 5, Waldner formalizes this insight through the use of “causal graphs.”
7
It is worth emphasizing that the probative value of these clues hinges on whether they are uncovered in a
treated, but not a control, case.
193 Process tracing, causal inference, and civil war
18:06:25
More broadly, process tracing can be used to explore whether the proposed
causal pathway between an independent variable (or treatment) and the suggested
mechanism is even plausible. This task is especially relevant for crossnational
studies, where the language of mechanisms is often invoked in fairly
coarse terms – “state capture,” for example, or “opportunity costs” – which
obscures rather than reveals the causal processes unfolding at diﬀerent subnational
levels (Sambanis 2004; see also Checkel 2013b: chapter 1).8
Similarly,
cross-national studies that rely on exogenous events such as price commodity
shocks to explore changes in conﬂict incidence across diﬀerent states could be
strengthened by using process tracing to clarify the channel(s) through which a
shock aﬀects state capacity or rebel recruitment at the subnational level. In this
setting, since numerous mechanisms are plausible, process tracing the link
between the shock and the mechanism would also be an important step in
reducing the problem of equiﬁnality that plagues cross-national studies.
Fourth, the distinction between process tracing for theory building versus
theory testing is an important one (Bennett and Checkel, this volume,
Chapter 1, pp. 7–8; see also Beach and Pedersen 2013a). While comparative
observations (say, villages) within a particular case (say, a region within a
country) are useful for theory building, out-of-sample tests are generally preferred
for empirical testing to avoid “ﬁtting” one’s argument to the cases used to
develop it. Lubkemann (2008) provides a neat illustration of this principle at
work. Seeking to explain forced migration as a function of war, he began his
empirical investigation in the Machaze district of Mozambique, which witnessed
a high degree of violence and refugee outﬂow. He then followed the
trail of internally displaced persons to new ﬁeld sites, treating “dispersion as a
ﬁeld site” (ibid.: 25), including dispersion to the capital of the neighboring
district and to the area across the border in South Africa. While his ﬁeldwork in
Machaze was formative in establishing propositions about refugee ﬂows, it is the
testing of these insights in locations not originally envisaged by the research
design – process tracing out-of-sample, as it were – that provides greater
conﬁdence in his claims about the nature of wartime forced migration.
Working examples
I draw on two empirical examples to demonstrate the importance of process
tracing to civil war studies. I ﬁrst concentrate on the (mostly) cross-national
8
See also Bazzi and Blattman 2011; and Berman and Couttenier 2013.
194 Jason Lyall
18:06:25
debate about the determinants of civil war onset. I then turn to emerging
micro-level debates about the eﬀects of civilian victimization on subsequent
insurgent violence. In each case, I suggest possible research designs that use
process tracing within a potential outcomes framework to adjudicate between
proposed mechanisms linking independent variables to outcomes.
Working example 1: civil war onset
Why do civil wars break out? To date, scholars have sought answers to this
question by predominantly utilizing cross-national regressions that link
national level characteristics to the probability of civil war onset. In one
notable example, James Fearon and David Laitin draw on data from 127
conﬂicts in the 1945 to 1999 era to argue that war is driven by opportunities
for rebellion, not percolating grievances within the population. Instead, weak
state capacity, as proxied by per capita income, and mountainous terrain are
key drivers of insurgency; the weaker and more mountainous the state, the
more likely we are to witness war (Fearon and Laitin 2003).
A recent spate of work has taken exception to this state capacity claim,
however, and has instead argued that the exclusion of ethnic groups from
executive political oﬃce better captures the origins of civil war onset. The larger
the size of the excluded ethnic group, the greater the likelihood of civil war,
especially if the now-excluded group once held the reins of political power
(Cederman and Girardin 2007; Buhaug et al. 2008; Cederman et al. 2010).
This is an important and productive debate, but one subject to diminishing
returns if the underlying processes that produce these outcomes continue to
be left unexamined or measured with crude national-level proxy indicators.
Absent new cross-national data, the greatest returns to investment appear to
lie in the testing of proposed mechanisms at the subnational level.9
Take the argument by Cederman et al. (2010). These authors identify
124 ethnic civil wars (1946 to 2005) and employ a new data set (Ethnic
Political Relations, or EPR) that measures the annual level of political exclusion
from executive power for relevant ethnic groups within a given state.
Using multivariate regression and several measures of political exclusion, they
conclude that “we are able to establish an unequivocal relationship between
the degree of access to state power and the likelihood of armed rebellion”
(Cederman et al. 2010: 114).
9
For examples of the use of qualitative case studies to reﬁne cross-national models, see Sambanis 2004;
Collier and Sambanis 2005.
195 Process tracing, causal inference, and civil war
18:06:25
The authors cite ﬁve possible mechanisms that could undergird the relationship
between rising ethnic exclusion and a greater likelihood of ethnic civil
war. First, political exclusion can generate a fear of domination and resentment
among excluded individuals, leading to a desire for (armed) revenge.
Such motives are especially likely if the ethnic group was only recently
excluded from political oﬃce. Second, the larger the excluded group, the
greater its mobilizational capacity, and the greater its likelihood of leading
an armed challenge against the state. Third, a history of prior conﬂict between
ethnic groups can heighten the risk of war via three channels: (1) ethnonationalist
activists glorify their group’s history through one-sided narratives
that stress their own victories and attribute blame for military losses to
traitors, weak-spirited leaders, or a ruthless enemy; (2) past experiences of
violence may become part of oral tradition or oﬃcial narratives, nourishing
calls for revenge; and (3) prior exposure to combat means that violence is no
longer unthinkable, but constitutes part of the accepted repertoire of action.
These hypothesized mechanisms are summarized in Table 7.1. Mechanisms
suggested by other theories are also listed, although these are illustrative rather
than comprehensive. While the mechanisms oﬀered by Cederman et al. (2010)
are plausible, the evidence marshaled to support their presence is thin, consisting
typically of a few short sentences (see, for example, ibid.: 110–111).
How could we go beyond statistical associations to examine the causal
processes at work? One possible approach uses a potential outcomes framework
to identify a series of comparative cases that isolate the mechanisms and
their role in producing war onset. Political exclusion would be recast as a
“treatment,” while countries without ethnic group-based discrimination
would represent the pool of available control observations. Matching could
then be used to identify pairs of cases that have similar values across a range of
theoretically important independent variables (or “covariates”), including
Table 7.1 Mechanisms and measures as proposed by Cederman et al. (2010)
Proposed mechanisms Possible measures
Status reversal Fear of domination; desire for revenge
Mobilization capacity % of population (collective action)
Prior exposure to violence Nationalist histories; violence as “thinkable”
State capacity Force structure; deployment; bureaucracy; police
Spoils Center-seeking behavior; spoil-seeking
Note: Below the dotted line are alternative mechanisms and proposed measures.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
196 Jason Lyall
18:06:25
level of state capacity, ruggedness of terrain, and size of standing army.
Assuming the statistical relationship identiﬁed in the full data set survives
the matching procedure, we could then identify matched pairs of cases that are
dissimilar only in their treatment status and the outcome (war onset/no war
onset). Since the proposed argument rests on at least ﬁve mechanisms, no one
matched pair will be able to test all possible mechanisms and their relationship
to war onset. Instead, the matching procedure creates a pool of available
paired comparisons that could be used to isolate individual mechanisms
through a series of cascading comparisons.
For example, Comparison A could involve process tracing within and
across a pair of similar cases where civil war onset was observed in the treated
case (for example, the politically exclusionary state), but not in the control
case. Each state could also have been subjected to an external shock – ideally,
the same shock, such as a sharp decrease in commodity prices – that impacts
each in a similarly negative fashion. This type of design would allow for
separation of the eﬀects of political exclusion from those of state capacity, as
the price shock should aﬀect each state in equal measure, yet civil war is only
observed in the politically exclusionary state. Similarly, matching on additional
(new) measures of state capacity such as bureaucratic penetration or the
nature of infrastructure would enable the sifting out of the eﬀects of status
reversal or mobilizational capacity from the potentially confounding eﬀects of
(weak) state capacity (Comparison B).
Disaggregating an ethnic group’s experience with political exclusion can
provide additional causal leverage. Comparison C could involve two states
that have similar characteristics, including presence of political exclusion, but
where one group has experienced a sudden and recent reversal, while the other
excluded group has not. A related set-up could examine a matched pair where
the size of the excluded group varies (one large group, one small group) to test
the link between mobilizational capacity and war onset (Comparison D).
Another matched pair could examine two similar states with equivalent levels
of political exclusion, but where one marginalized ethnic group has experienced
prior violence at the hands of the state, while the “control” group has
not suﬀered prior victimization (Comparison E). More ambitious designs
could use matched pairs that control for several mechanisms across cases –
say, status reversal and mobilizational capacity – and vary a third mechanism
such as prior exposure to state violence (Comparison F).
Once the relevant comparisons have been established via matching, the
actual process tracing can begin. To establish the credibility of ties between
ethnic exclusion and war onset, we might consider qualitative evidence from
197 Process tracing, causal inference, and civil war
18:06:25
the recruitment drives of insurgent organizations. What types of appeals do
they use to mobilize individuals? Are insurgents organized along ethnic lines?
We should also observe that proportionately larger ethnic groups more readily
overcome collective action problems when attempting to mobilize recruits.
Ideally, evidence from both public and private claims about the nature of
(ethnic) grievances would be uncovered and would dominate more tactical
considerations such as perceptions of state weakness (“now is the time to
strike because the state is weak”) or a desire for spoils.
Process tracing is also essential for articulating the sequence of events
leading up to the war. Cederman et al. (2010) suggest that rebels, not the
state, should initiate the conﬂict. Did fear of ethnic domination precede the
conﬂict, or were such concerns actually a product of the ﬁghting? Were
nationalist histories and memories of prior violence widespread, or did such
myths emerge as a post-hoc rationalization for the war? And, perhaps most
importantly, are these myths only actionable in political systems that exclude
along ethnic lines, or can would-be rebels craft such narratives even in the
absence of prior ethnic exclusion?
Finally, process tracing can play a crucial role in sifting out the indirect eﬀects
that state capacity might have on the mechanisms proposed by the ethnic
exclusion argument. Although these arguments are typically pitted against
one another, it is possible, indeed likely, that state actions can condition the
eﬀects of ethnic exclusion (and vice versa). Political exclusion may be a
response to state weakness, for example, as an embattled elite seeks to “harden”
its regime against potentially disloyal populations (Roessler 2011). More subtly,
fear of ethnic domination may be a reﬂection of the military’s ethnic composition,
while opportunities for group mobilization may be conditioned by the size
and deployment patterns of a state’s armed forces. Cederman et al. (2010: 95,
106) also note that rapid and sudden ethnic reversal is especially likely in weak
states, suggesting a more complicated relationship between state power (and
violence) and grievance-based mechanisms (see also Wood 2003).
In short, adopting a potential outcomes framework involves the use of
multiple comparisons (“cascades”) to screen out competing theories and
their mechanisms. It also enables a closer examination of the sequence by
which ethnic exclusion translates into a heightened risk of conﬂict onset,
helping to guard against reverse causation. Articulating an elaborate theory
with numerous measures for each mechanism also strengthens our inferences
about these processes by permitting congruence tests across multiple
indicators, increasing our conﬁdence that we have correctly identiﬁed the
process(es) at work.
198 Jason Lyall
18:06:25
Working example 2: civilian casualties and insurgent violence
Civilian victimization and its eﬀects on subsequent insurgent violence represents
one of the fastest growing research areas in the study of civil war
dynamics. Despite divergent methods, it has become a near article of faith
that indiscriminate victimization of civilians facilitates the recruitment of
newly abused individuals by insurgents, contributing to bloody spirals of
escalatory violence between counterinsurgent and rebel forces (for example,
Kalyvas 2006; US Army 2007; Jaeger and Paserman 2008; Kocher et al. 2011;
Condra and Shapiro 2012; Schneider and Bussmann 2013). While this view is
not uncontested (Lyall 2009), much of the debate now centers around the
causal processes linking victimization to subsequent patterns of insurgent
violence. To date, however, our research designs have not kept pace with the
profusion of mechanisms cited by scholars as facilitating insurgent recruitment
or producing escalatory spirals.
Setting aside for the moment the inherent diﬃculties in process tracing
such a sensitive issue, the abundance of possible mechanisms, operating
singularly or jointly, can frustrate eﬀorts to establish defensible causal claims.
Consider the following example from a January 2013 drone strike in Yemen,
which killed at least one, and possibly ﬁve, innocent civilians:
As the ﬁve men stood arguing by a cluster of palm trees, a volley of remotely operated
American missiles shot down from the night sky and incinerated them all, along with
a camel that was tied up nearby.
In the days afterward, the people of the village vented their fury at the Americans
with protests and brieﬂy blocked a road. It is diﬃcult to know what the long-term
eﬀects of the deaths will be, though some in the town – as in other areas where drones
have killed civilians – say there was an upwelling of support for Al Qaeda, because
such a move is seen as the only way to retaliate against the United States.
Innocents aside, even members of Al Qaeda invariably belong to a tribe, and when
they are killed in drone strikes, their relatives – whatever their feelings about Al
Qaeda – often swear to exact revenge on America.
“Al Qaeda always gives money to the family,” said Hussein Ahmed Othman al
Arwali, a tribal sheik from an area south of the capital called Mudhia, where Qaeda
militants fought pitched battles with Yemeni soldiers last year. “Al Qaeda’s leaders
may be killed by drones, but the group still has its money, and people are still joining.
For young men who are poor, the incentives are very strong: they oﬀer you marriage,
or money, and the ideological part works for some people.”10
10
“Drone Strikes Risks to Get Rare Moment in the Public Eye,” New York Times, February 6, 2013, A1.
199 Process tracing, causal inference, and civil war
18:06:25
This brief example usefully highlights at least ﬁve of the mechanisms that
scholars typically invoke to explain the process from victimization to participation
in an insurgency. A desire for revenge, tribal (group) ties, selective
incentives in the form of money and marriage, and ideology all intermingle as
plausible mechanisms in just this one instance. We might also add property
damage, which leads to economic hardship and shifting reservation values for
joining an insurgency (Abadie 2006),11
and the belief that greater risk is
associated with non-participation in an insurgency (Kalyvas and Kocher
2007), as two additional mechanisms not captured by this example.
The example also illustrates a second, less appreciated, issue: without prior
baseline levels for these mechanisms, and without a similar control village that
was not struck, we cannot assess the relative importance of these mechanisms
or the causal eﬀects of the air strike on subsequent behavior. Once again, a
potential outcomes framework that emphasizes counterfactual observations
provides insights not possible with a singular focus on within-case observations.
Without a control observation, for example, we cannot establish either
the direction or the magnitude of the air strike’s eﬀect on support for Al
Qaeda. Similarly, without a before-and-after comparison of civilian attitudes
and behavior across cases, we cannot determine whether the air strike
increased, decreased, or had no eﬀect on subsequent insurgent recruitment
and violence.
Given the number of plausible mechanisms and the possibility that they
might interact, how could process tracing be used to explore the links between
victimization, recruitment, and subsequent participation in an insurgency?
Table 7.2 outlines one possible research design.12
The basic idea is again one of maximizing comparisons by exploiting
variation in the nature of the victimization and how it was administered.
More speciﬁcally, we can create additional comparisons by decomposing the
“treatment” – here, experiencing a drone strike – into diﬀerent types of
victimization, while including individuals in the sample who were present
(i.e. in the same village) at the time of the strike, but who were not hurt, as
counterfactual observations.
Variation in civilian victimization, for example, can be used to create
comparisons that enable process tracing to link state violence to insurgent
behavior. To separate the “revenge” mechanism from an economic hardship
11
See also Lyall 2013.
12
This design draws on the author’s experiences with USAID’s Afghan Civilian Assistance Program II,
administered by International Relief and Development (IRD) in Afghanistan during 2012 to 2013.
200 Jason Lyall
18:06:25
one, we could compare individuals who are victimized but do not experience
property damage (Type A) with those who only have property damage
(Type B). We could then compare individuals A and B to individual C, who
was present but unharmed by the drone strike. These individuals could be
chosen via random selection (for example, from a list of victimized individuals
and locations). A screening question could be used to insure that these
individuals share similar socioeconomic characteristics. This procedure creates
a two-control group comparison (Rosenbaum 2010) between individuals
A and B, and between A and C, permitting in-depth process tracing to sort out
the role played by diﬀerent mechanisms in shaping an individual’s attitudes.
We can also draw on process tracing inductively to explore the nature of the
sample and the context in which the civilian victimization occurred. In
particular, we should stratify our sample by levels of key covariates to account
for victimization’s conditional eﬀects. In Table 7.2, I use the example of prior
violence in a village by the counterinsurgent as one key conditioning factor
with varying levels (here, high/medium/low). We might imagine that diﬀerent
mechanisms operate under diﬀerent circumstances; a one-time event may
have a diﬀerent meaning from repeated violence, and so revenge motives or
nationalism may have more purchase when heavy oppression is used rather
than a one-time, possibly accidental, event. Stratifying our sample along these
important covariates before process tracing also aids in illustrating gaps in our
coverage. It may be impossible, for example, to access high violence areas,
placing an important limit on the generalizability of our ﬁndings.
Table 7.2 Sample research design for assessing effects of civilian
victimization using process tracing
Assignment
Context Random Targeted
(violence) (“as-if”) (selective)
Low A, B, C A, B, C
Medium A, B, C A, B, C
High A, B, C A, B, C
Note: A represents personal victimization; B represents property damage;
C is a control individual in the selected location, but who was not
victimized. A, B, and C are in the same village in this design. Violence is
used as an important example of context. Assignment refers to the manner
in which individual(s) were targeted, i.e. plausible claim to “as-if” random
or selected according to some criteria. Context is by village.
201 Process tracing, causal inference, and civil war
18:06:25
The credibility of our estimates about the eﬀects of violence is also enhanced
if we can demonstrate that this victimization occurred “as if” randomly. For
most micro-level studies, the problem of selection bias looms large. That is, the
individuals victimized diﬀer in some important fashion from non-victims, since
they were selected by the state for victimization. Some studies (for example,
Condra and Shapiro 2012), however, contend that we can assume casualties are
inﬂicted more or less randomly – unlucky individuals are in the “wrong place
and time” – and so we can treat these casualties as unconnected (“plausibly
exogenous”) to broader patterns of war. The beneﬁt, of course, is clear. If civilian
casualties are not intimately tied to broader patterns of violence, then we are
able to estimate cleanly the eﬀects of these casualties on subsequent violence,
without worrying about selection eﬀects that might confound our study.
Whether this claim is plausible given the possibility of substantial heterogeneity
in how civilians are victimized, variation in the meaning of victimization
depending on the perpetrator’s identity, and the prospect that civilians
are often targeted strategically, is a central question for inductive process
tracing. Determining whether (and when) the “as-if” random assumption
holds also helps determine to which populations we can generalize when
making claims about the eﬀects of violence.
What form does the process tracing actually take? Given the observational
equivalence of these mechanisms, it makes sense to shift the debate to examine
how victimization aﬀects attitudes, not behavior. Once again, we witness the
virtues of elaborate theories, which force us (in this case) to create attitudinal
measures for each mechanism that enable us to distinguish among causal
pathways to insurgency. Table 7.3 oﬀers an initial cut at measures for ﬁve
Table 7.3 Possible mechanisms linking civilian victimization to insurgent recruitment and violence
Proposed mechanisms Possible measures
Revenge View of government/counterinsurgent, sense of loss
Economic hardship Changes in livelihood, beliefs about (future) well-being
Group identity Perception of status; magnitude of co-ethnic bias
Risk Willingness to consider risky actions
Selective incentives Receipt and views of rebel provision of goods/services
Note: Proposed measures (not exhaustive) are designed to be consistent with multiple
methodologies, including survey and behavioral experiments, focus groups, interviews, and
ethnographic approaches that remain open to post-positivist notions of causation. Measured
relative to control observations (individuals with no or diﬀerent exposure to civilian victimization).
202 Jason Lyall
18:06:25
mechanisms that link victimization to increased participaion in an insurgency
via changes in attitudes.
Creating multiple measures for each mechanism also creates more space to
adopt diﬀerent methodologies when process tracing (see also Checkel and
Bennett, this volume, Chapter 10). Interviews with rebels, for example, have
become a standard tool in the civil war scholar’s methodological toolkit
(Wood 2003; Weinstein 2007; Ladbury 2009), although care should be taken
to insure that non-rebels are also interviewed. Survey experiments could also
tap into these concepts using indirect measurement techniques that mitigate
incentives for interview subjects to dissemble due to social desirability bias or
concerns about reprisals (Humphreys and Weinstein 2008).13
Focus groups
provide an opportunity to explore not just individual level dynamics, but also
the construction of narratives about civilian victimization and, in particular,
how blame for these events is assigned. Behavioral “lab-in-the-ﬁeld”
experiments provide an additional means of measuring how violence aﬀects
attitudes, including preferences over risk, time horizons, and decision-making
(Voors et al. 2012). Finally, ethnography may oﬀer a window into how
these dynamics shift over time. These processes are diﬃcult to capture
with surveys or one-oﬀ interviews, especially if the process between victimization
and subsequent behavior has more of a “slow burn” than a “quick fuse”
logic.
Each of these methods has its own particular strengths and weaknesses.
Moreover, the environment after a civilian casualty event is among the most
sensitive a researcher can experience. These factors combine to make
“smoking-gun” evidence elusive in such settings; it is unlikely that evidence
will be found to support one mechanism while trumping all others. Good
process tracing may still not yield wholly conclusive evidence, as emphasized
by Bennett and Checkel (this volume, Chapter 1). Instead, it may be more
productive to explore the scope conditions that make certain pathways more
or less likely to lead to insurgency. A potential outcomes framework that
stresses the role of counterfactuals (i.e. non-victims), the need for multiple
measures for each mechanism (i.e. “elaborate theory”), and a clear understanding
of the selection mechanisms (was victimization deliberate or by
chance?) oﬀers one means for harnessing process tracing to the task of
producing generalizable claims.14
13
See also Lyall et al. 2013.
14
The relation of process tracing to theory type (mid-range, typological, general) remains a key challenge
for future work. See also Checkel (this volume, Chapter 3); and Checkel and Bennett (this volume,
Chapter 10).
203 Process tracing, causal inference, and civil war
18:06:25
Practicalities
My arguments thus far have tacitly assumed that ﬁeldwork is necessary to
gather most, if not all, of the data required for process tracing. Indeed, many of
the methodologies best suited for process tracing – including lab-in-the-ﬁeld
and survey experiments, in-depth interviews, and ethnography – mandate an
often-substantial investment in ﬁeld research.
Yet, ﬁeldwork in (post-)conﬂict settings presents a host of methodological,
logistical, and ethical challenges (Wood 2006). A short list of such issues
includes: the threat of physical harm to the researcher, his or her team, and
local respondents; variable (and unpredictable) access to ﬁeld sites due to
changing battleﬁeld conditions; the twin dangers of social desirability bias and
faulty memories that may creep into interview and survey responses, especially
in areas contested between combatants; the often-poor quality of data
for key measures; the changing nature of causal relationships, where eﬀects of
a particular intervention may be large in the initial conﬂict period, but
diminish over time as the conﬂict churns on; and reliance on outside actors
and organizations for access and logistics that might shape perceptions of the
researcher’s work among potential respondents.
Context typically trumps generalization in these environments, so solutions
to these problems are necessarily local in nature. That said, there are three
issues that all researchers are likely to face when gathering data for process
tracing in conﬂict zones.
First, researchers must obtain the voluntary consent of would-be interviewees
and respondents. Though this is a common injunction for Institutional
Review Board (IRB) approval at American universities, the requirement takes
on a special cast in conﬂict settings, where individuals may run risks for
simply meeting with (foreign) researchers or survey teams. Informed consent
in these settings requires that participants understand the nature of the study
(at least broadly), its funding source, and plans for dissemination, so that they
can properly judge the risk associated with participating. It also requires that
individuals recognize that they will receive no material beneﬁts – for example,
new disbursements of economic assistance – from participation.
Moreover, in many settings, such as Afghanistan, obtaining consent is a
two-step process: ﬁrst, with the stakeholders who control access to a given
village and, second, with the prospective participant(s). Obtaining consent
from these gatekeepers, whether government oﬃcials, local authorities, or
rebel commanders, can mean the diﬀerence between accessing or being
204 Jason Lyall
18:06:25
excluded from certain locations. In addition, obtaining permission from local
authorities can lower individuals’ concerns about participating, potentially
also reducing the bias in their responses to interview or survey questions.
Consent from local authorities and individuals becomes especially important
if one’s process tracing hinges on gathering longitudinal data.
Second, maintaining the anonymity of interviewees and survey respondents
is essential in wartime settings. Researchers must work to secure data and to
insure that if compromised, it does not allow third parties to identify their
sources. The simplest expedient is not to record an individual’s name and
instead use a randomized identiﬁcation number. The advent of computers,
cell phones, and portable data storage devices in the ﬁeld has changed the
calculus, however, making it possible to reconstruct an individual’s identify
even if his or her name was not recorded. Survey ﬁrms routinely use respondents’
telephone numbers to call back for quality control purposes, while
enumerators use GPS devices and maps to track their “random walks” in
selected villages when creating samples.
Conﬁdentiality and guarantees of anonymity must extend to these personal
data, not simply an individual’s identity, especially given the prospects for
rapid dissemination if these electronic storage devices are compromised. In
areas with good cell phone coverage – an increasing share of once remote
locations – data from interviews, surveys, or maps should be stored remotely
(for example, on a “cloud” storage site) and local copies deleted to mitigate the
risks of unwanted data capture. Researchers should also maintain robust
networks for returning surveys, interview notes, or other sensitive materials
to a central safe location if electronic means are not available. In Afghanistan,
for example, trucks carrying market wares to Kabul can be enlisted to deliver
sealed packages of completed surveys back to Kabul, where they are then
scanned and destroyed. This system avoids having dozens of enumerators risk
exposure while carrying materials through potentially hundreds of checkpoints
between their ﬁeld sites and Kabul. Similarly, quality control call-backs
can be completed by a manager at the ﬁeld site; the phone’s log is then deleted,
thereby avoiding transporting these data across checkpoints.
Third, researchers must work to safeguard both themselves and members of
their team. Surprisingly, existing IRB guidelines do not address the issue of
researcher safety nor that of the enumerators, translators, ﬁxers, and others
who might work under the researcher’s direction and who also assume risks
by participating in the research. Establishing a baseline of risk before conducting
research – How violent? Which actors are present? What types of
movement restrictions exist? – can be useful in detecting sudden changes that
205 Process tracing, causal inference, and civil war
18:06:25
suggest increased risk for one’s team. This baseline is also useful in selecting
potential ﬁeld sites as well as replacements, often via matching, which enables
researchers to switch sites quickly without compromising their research
design. Locals, who often have a far better sense of security risks than outsiders,
should also be consulted when establishing notions of baseline risk.
Finally, it is useful to construct a “kill-switch” protocol that can be activated if
team members have been threatened (or worse). Activating the “kill-switch”
(often via SMS) would signal to team members to wipe their data and withdraw
to central points to avoid a credible threat, such as speciﬁc targeting of
the team by rebel or government forces.
Conclusion
The explosion of research on the origins and dynamics of civil wars has not
(yet) been accompanied by a turn to process tracing to identify and test the
causal mechanisms that underpin our theories. This state of aﬀairs is unfortunate,
not least because political scientists have developed an increasingly
sophisticated and eclectic methodological toolkit that could be applied toward
process tracing in violent settings. Certainly, feasibility and safety concerns are
paramount in these environments. Yet, as this chapter has sought to demonstrate,
there are research designs and strategies that can be adopted to
heighten our ability to make casual inferences despite these challenges.
The advantages of incorporating process tracing into conﬂict research also
spill over to the policy realm. Process tracing oﬀers an excellent means of
uncovering the contextual “support factors” (Cartwright and Hardie 2012:
50–53) that help produce a causal eﬀect. Without exploring these contextual
factors, as well as the nature of the link between treatment and its mechanisms,
we are left on shaky ground when trying to determine whether a
particular eﬀect or process generalizes to other settings. Moreover, process
tracing is ideally suited to investigating possible interactions between multiple
mechanisms. Policymakers, not to mention scholars conducting impact evaluations,
are likely operating in settings marked by multiple mechanisms that
interact in complex ways to produce a given eﬀect. Pre-specifying the possible
causal pathways and identifying several measures for these mechanisms, as
called for by elaborate theorizing, will also help to avoid ﬁshing for the
“correct” mechanism via backward induction. The result of these eﬀorts is
likely to be a better understanding of how these processes unfold, thus
contributing to our theories of civil wars as well.
206 Jason Lyall
18:06:25
Process tracing does have its limits, however. Without explicitly incorporating
counterfactuals to facilitate cross- and within-case comparisons, theorytesting
process tracing can lead to mistaken causal inferences about the
robustness of a presumed relationship between an independent variable and
outcomes. Moreover, crafting research designs that are capable of both identifying
a statistical association and then competitively testing the mechanisms
responsible for it may be a bridge too far. What may be required is a shift
toward designs that take a particular relationship as a given and instead
explicitly engage in process tracing to detail why this pattern is present.
Danger lies in this type of strategy, though: the more micro-level the process
tracing, the more contextual factors trump abstraction. The result may be a
wonderfully nuanced account of a speciﬁc process that doesn’t generalize to
other settings even within the same case. Finally, a too-speciﬁc focus on
mechanisms and process tracing might lead to neglecting the importance of
structural factors that might condition which mechanisms are present and the
magnitude of their eﬀects (Checkel 2013b: 19).
Of course, process tracing is not unique in having drawbacks; no methodological
approach is without its shortcomings. And the pay-oﬀs, measured in
terms of theoretical progress and policy insights, are considerable. By seeking
to move beyond statistical associations to understanding why these relationships
are present, scholars can open new avenues for exciting research into
substantively important questions about the onset and battleﬁeld dynamics of
civil wars.
207 Process tracing, causal inference, and civil war
18:06:25