 

mm
c)

J‘5I

59aaﬁfﬁq

This is to certify that the
dissertation entitled

GLOBALIZATION, POLITICAL COSTS, AND STRATEGIC
DECISIONS

presented by
WONJAE HWANG

has been accepted towards fulfillment
of the requirements for the

Ph.D. degree in POLITICAL SCIENCE

 

.assiggésﬁ:;*””“

 

Maia? Professor’s Signature

JUNE/1 5/2004

 

Date

MSU is an Affirmative Action/Equal Opportunity Institution

 

LIBRARY

Micmgan State
University

 

 

PLACE IN REFURN Box to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJClFiCIDateDuepss-p. 1 5

GLOBALIZATION, POLITICAL COSTS,
AND STRATEGIC DECISIONS

By

Wonjae Hwang

A DISSERTATION

Submitted to
Michigan State University
in partial fulﬁllment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Political Science

2004

ABSTRACT
GLOBALIZATION, POLITICAL COSTS, AND STRATEGIC DECISIONS
By

Wonjae Hwang

The “third wave of democratization” along with the ongoing economic
globalization in the last two decades has made domestic politics important in the modern
society. Characterized by the advent of new actors and new global issues, globalization or
economic interdependence mandates that scholars in international relations look carefully
at the process of decision-making in international politics. With the growing importance
of domestic politics and domestic actors, a structural proposition that international
structural factors dominate states’ behavior has been weakened. In fact, international
political events are basically determined by the interaction between these multiple levels:
the state level, the dyadic state level, and the international system level. In this regard,
this dissertation assesses multilevel structures in decisions on international political
events.

Investigating political leaders’ strategic consideration on political cost, it attempts
to explain the varying effect of economic liberalization on interstate conﬂict. Contrary to
the liberal peace, the effect of economic interdependence on the onset of international
conﬂict is not constant but heterogeneous, depending on domestic political situation. The
effect of economic globalization on national welfare should be examined not only by

overall economic growth but also by inequality within countries. If economic

globalization is associated with the increase of inequality within countries, the pacifying
effect of trade on interstate conflict can be offset.

In sum, this study explores a theoretically and empirically proper way of
analyzing the effect of economic globalization on international relations and political

leaders’ strategic decisions.

Copyright by
Wonjae Hwan g

2004

For My Beloved Family

ACKNOWLEDGMENTS

By its very nature, a dissertation is a collective enterprise. Although only my
name appears as author, I could not have completed this immense undertaking without
the help and guidance of many people. While I cannot thank all of them by name without
doubling the size of this volume, I want to give special thanks to a few special people.

The ﬁrst and most important person in the long list is my dear wife, Soon-Joo Jo.
Without Soon-Joo’s love, patience, and innumerable sacrifices, this dissertation would
not have been completed. I am also grateful to Chan-Hee Hwang, my son, and Se-Hee
Hwang, my daughter, for their many years of acceptance that dad had to spend time
mostly in my office. My family and Soon-Joo’s family have also played a major role
encouraging me as I worked. I especially want to acknowledge my parents who raised me
to cherish education and pursue this ﬁnal degree.

I have also had many mentors at Michigan State who have skillfully guided me
along the way. I would like to thank Scott Gates for his academic guidance and
inspiration with faith and confidence in me. He has provided me with unquestioned
support and counsel and has spent countless hours helping me to hone my professional
skills. Mark Jones always stood by me through this challenging process with unflagging
support and encouragement. He also has been a great mentor throughout my tenure as a
graduate student. Without his invaluable help in various aspects, this dissertation would
be impossible. Bil] Reed also encouraged my professional development and sparked my
interest in international relations. I was privileged to learn invaluable knowledge of

international relations from him. I must extend thanks to Burt Monroe and Susan Chun

VI

Zhu for their constructive advices and genuine support. I am grateful for their balanced
and thoughtful critiques.

I have also beneﬁted enormously from discussion with and assistance of other
scholars and friends. Especially, Andy Muwais has spent his invaluable hours to review
and edit my chapters. The department of Political Science support staff, Kelly Fenn,
Rhonda Burns, Char Rogers, Karen Battin, Shirley Gordon, and Steve Andre, also helped
me in many ways.

I also received quite a bit of assistance beyond East Lansing’s borders. Sara
McLaughlin Mitchell offered her data to me, despite the fact that it would be used to
question the methods of her own work. While I attended the EITM program at University
of Michigan, I received assistance from other participants, staffs, and instructors.

While many names are not included in this list, their help is greatly appreciated

and not forgotten.

VII

TABLE OF CONTENTS

LIST OF TABLES

LIST OF FIGURES

INTRODUCTION

CHAPTER I
A MULTILEVEL ANALYSIS OF INTERNATIONAL POLITICS:
A SYSTEMIC APPROACH TO DEMOCRATIC PEACE
Multilevel Interaction And International Relations Studies
Problems Caused By Disregarding Multilevel Structure
Theoretical Issues
Statistical Issues
Why Multilevel Models?
Application of the Multilevel Model Method
Democratic Norms And Third Party Conflict Resolution
Empirical Analysis
Result

Conclusion

CHAPTER II
TRACING MONADIC EFFECTS ON DYADIC BEHAVIOR:
A STRATEGIC APPROACH TO THE LIBERAL PEACE
Multilevel Data Structure in International Relations
Monadic Effects vs. Dyadic Effects
Why a Multilevel Model Analysis?
The Liberal Peace and Heterogeneous Domestic Politics

Propositions, Modeling, and Estimation

VIII

XI

11
11
15
17
21
21
24
30

46
51
51
54
56
62

Result 73

Conclusion 83
CHAPTER HI
GLABAIZATION, INEQUALITY WITHIN COUNTRIES, AND INTERSTATE
CONFLICT 85
Diversionary Use of Interstate Conﬂict 90
Trade and Inequality Within Countries 94
Hypotheses 97
Speciﬁcation 98
Result 104
Conclusion 107
CONCLUSION 1 10
APPENDICES 115
Appendix A - Classification of the size of winning coalition 116
Appendix B - The Onset of Militarized Interstate Disputes 118
Appendix C — Graphical inspection of convergence 119
Appendix D — Global Inequality of Individual Incomes 120
BIBLIOGRAPHY 122

LIST OF TABLES

Table 1: Third-Party Attempts to Settle Territorial Claims

Table 2: ANOVA (Analysis of variance)

Table 3: Variable Descriptions

Table 4: Analysis of variance with Random Effects

Table 5: The Onset of Militarized Interstate Disputes (Politically relevant dyads)

Table 6: Classification of the size of winning coalition based on Bueno de
Mesquita et a1. (2000).

Table 7: Classiﬁcation of the size of winning coalition based on DPI 2000.

Table 8: The Onset of Militarized Interstate Disputes (All dyads)

Table 9: Variable Descriptions

Table 10: Initiation of Militarized Interstate Disputes

32

37

70

74

76

116

117

118

103

105

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Figure 6:

Figure 7:

Figure 8:

LIST OF FIGURES

Theoretical Logic of the Regression Models

Probability of Third Party Settlement Attempts

Third Party Settlement Attempts

Third Party Settlement Attempts

The Conditional Effect of Interdependence on MIDS

The Conditional Effect of Interdependence on MIDS

Global Inequality of Individual Incomes, 1820-1992

Global Inequality of Individual Incomes, 1820-1992

XI

16

35

39

42

80

82

120

121

INTRODUCTION

Over the last two centuries, proponents and critics of an open global economy
have debated whether the free ﬂows of goods, capital, and services make the world more
or less peaceful.1 One view is that as nations increase their commercial, ﬁnancial, and
cultural ties, they are less likely to go to war. The opposing view holds that world
economic liberalization worsens inequality between and/or within countries conﬂict
(Agenor 2003; Bourguignon and Morrisson 2002; Lindert and Williamson 2001),
increasing interstate or intrastate conﬂict.

Scholarly efforts of clarifying the relationship between economic liberalization
and peace have recently intensified in international relations studies. A host of scholars
advocating the liberal peace argue that free trade or economic openness is beneﬁcial to
countries. Since trade is assumed to bring beneﬁts to all trading partners, the concern of
losing benefits from potentially lost trade weaken political leaders’ likelihood of
choosing military actions when they face international crisis. However, this treatment of
the effect of economic liberalization on interstate conﬂict is oversimpliﬁed because it
considers the effect of trade on domestic politics as homogeneous. Depending on the
domestic political situation, however, economic liberalization may increase or at least
may not decrease the probability of interstate conﬂict onset. This happens where policy
costs do not necessarily affect political leaders’ political costs. It implies that we Should

look at domestic politics to have a comprehensive understanding of decision-making.

 

l A host of scholars in international economics view that globalization began from the decade of 18208. For
example, O’Rourke and Williamson (2000) Show that international commodity price convergence did not
start until the decade, and that a powerful move toward liberal policy appeared during that decade as well.
Maddison (1995) and Bourguignon and Morrisson (2001) also monitor the world economy from the year
of1820.

The “third wave of democratization” (Huntington 1991) along with the ongoing
economic globalization in the last two decades has made domestic politics important in
the modern society. The advent of new actors and new global issues, which soften
borders between nation states, mandates that scholars in international relations look
carefully at the process of decision-making. With the growing importance of domestic
politics and domestic actors, a structural proposition that international structural factors
dominate states’ behavior has been weakened. Decisions on international political events

should be analyzed at multiple levels.

Multilevel Structure and Strategic Decisions in International Politics

This dissertation calls attention to multilevel structures in international politics.
International politics is often intertwined with domestic politics. Domestic institutions
affect decisions on interstate relations. The existing international structure or
international norms constrain or limit the interaction between nation states. Therefore,
international political events are basically determined by the interaction between multiple
levels. On the other hand, in the era of economic globalization or democratization, the
role of agents — such as individual states in the international system, political leaders in a
state, nongovernmental organizations, or multinational companies — is more important
than ever before. The growing importance of agents in international politics requires us to
revisit and reexamine carefully the agent-structure problem. That is, the theoretical
relationship and the interaction between agents and structure should be clarified (Desseler

1989; Waltz 1999).

The emphasis on the role of agents makes the strategic approach seem reasonable
in explaining international political phenomena. Political leaders’ strategic consideration
of political costs is the primary source of decision-making. It is a reasonable assumption
that political leaders are strongly concerned with the implications of foreign policy
decisions for their own political survival.2 It implies that political leaders should “take
domestic interests as well as foreign realities carefully into account because they want to
survive to enjoy the benefits of office” (Bueno de Mesquita 2000, p.111), especially
when political leaders do not have strong political support at home. Considering
“political” risk is essential to understanding political decisions (Lambom 1991; Baldwin
1995; Bueno de Mesquita, Morrow, Siverson, and Smith 1999). As a consequence,
theoretical models of strategic politics must address these multi-level game structures,
clearly thereby providing crucial answers regarding the relationship between opportunity
(structure) and willingness (agents) simultaneously. This type of rigorous specification
and subsequent empirical testing is demanding, but necessary.

The strategic approach in international political studies is important especially in
that it emphasizes the substantial importance of domestic institutions in international
politics. By investigating the relationship between agents and multilevel structures, this
approach makes it possible to integrate domestic and international games in a
theoretically coherent way. This argument also implies that both willingness
(preferences) and decision contexts (opportunity) should properly be addressed and
empirically tested in rational choice models. I develop an analytical model and

empirically test hypotheses from a multi-level hierarchical model on the liberal peace.

 

2 Political survival indicates that a political leader is not deposed from his ofﬁce in the year in question.

That is, this research addresses questions on the relationship between economic

interdependence and rnilitarized conﬂict using a multi-level analysis.

Three Essays

This dissertation consists of three different but inter-connected essays. The first
essay highlights a multilevel structure in the analyses of international politics. Potential
theoretical and statistical problems caused by failing to account for a multilevel structure
in an analysis are discussed. After reviewing the agent-structure problem, this study
shows how the theoretical relationship between agent-centric factors and structure-centric
factors is set up with logical consistency. As a way of specifying and testing the
relationship between multiple levels, a multilevel model is introduced. The potential of
this model is discussed along with the comparison with other statistical models.

Speciﬁcally, this study explores if changes in the structure of the international
system or international norms affect the interaction between countries, and also if
countries respond to the structural constraints in different manners depending on their
domestic institutions. A systemic approach to democratic peace is examined as a
substantive example. This essay shows that international democratic norms have a
positive effect on third party settlement attempts to territorial disputes, and that
depending on the regime type of dyadic states, the behavior of dyadic pairs of States
varies. Finally, it argues that as the proportion of democracies in the international system
increases, nondemocratic states are more likely to rely on such peaceful measures as third

party intervention to settle their territorial disputes.

The second essay attempts to trace monadic effects on dyadic behavior. Paying
attention to two types of heterogeneity inherent in a dyadic approach to the liberal peace,
this essay aggregates dyadic observations into individual states. Political leaders’
strategic positions in domestic politics as a main source of heterogeneity at the sate level
are analyzed. To maintain theoretical consistency in explaining the relationship between
domestic politics and international politics, this analysis also uses a multilevel
hierarchical model. As a result, this research provides a way of examining the
heterogeneous effect of economic interdependence on interstate conﬂict, which depends
on domestic institutions. Using a Bayesian simulation method and various statistical tests,
the second essay shows that trade or economic liberalization has mitigating effects on
interstate conﬂict more likely in democratic systems or systems where political leaders’
behavior is signiﬁcantly checked by people than other systems. In other words, this essay
argues that domestic political ‘institutions’ do matter in understanding economic ‘policy’
effects on interstate conﬂict.

The third and ﬁnal essay focuses on the effect of trade on inequality within
countries. There is a controversy about whether or not trade has increased or decreased
the level of inequality within countries. There are also empirical difﬁculties in measuring
the level of inequality within countries. 3 Nevertheless, as the world economy has
continued to be integrated, evaluating the effect of trade on inequality within countries

becomes more important and necessary. The diversionary theory posits that political

 

3 For example, a recent article in a journal notices the discrepancy between the ‘national-account estimates’
(income poverty) and the ‘household-survey estimates’ (consumption poverty) in measuring the level of
inequality. It argues that to the evaluate changes in inequality, we should compare growth in average
consumption with the growth in average income, highlighting the danger of focusing too much on the
dollar-a-day threshold. (For details, refer to The Economist (2004), using the following link:
http://www.economist.com/printedition/displayStory.cfm?Story ID=2498851.)

F

leaders have an incentive to take military actions to remain in power by diverting
attention from internal problems such as a weak economy to external issues (DeRouen
1995; Leeds and Davis 1997; Levy 1989; Morgan and Bickers 1992). It implies that a
high level of inequality within countries can intensify political leaders’ willingness of
diversionary use of force. As a result, if economic liberalization has something to do with
a high level of inequality within countries, economic liberalization or trade also has an
indirect effect on interstate conﬂict. Finally, this essay tests if trade has increased
inequality within countries, controlling other theoretically important variables including
the size of population in a country. This essay ascertains the extent to which inequality
within countries is relevant factor to political leaders faced with the decision to use force.
It tests a model of the diversionary theory that treats the inequality within countries
variable as an endogenous variable. Finally, the third essay demonstrates that, as the level
of inequality within countries grows, political leaders are more likely to engage in
international conﬂict as a way of diverting attention from domestic issues, and that the
pacifying effect of trade on interstate conﬂict is offset by its negative effect on equality
within countries.

Overall, this dissertation highlights the importance of employing multilevel
analyses in explaining international political events. A theoretically and empirically
proper way of analyzing the effect of economic globalization on international relations,
the interaction between the international system and nation state, and the effect of trade

on domestic politics, is explored through this study.

CHAPT ER I

A Multilevel Analysis of International Politics:

A Systemic Approach To Democratic Peace

Abstract

Analysis of the dynamics of the interactions between multiple levels is demanding,
because it requires us to answer theoretical puzzles and statistical challenges inherent in
multilevel data. Multilevel data has been widely used in studies of international relations.
Nevertheless, it is hard to find a well-defined statistical model that accounts for the
dynamic interplay across levels of analysis with theoretical consistency, properly
determining the relationship between levels. This paper argues that the failure to account
for multilevel structure in data of international relations may cause serious theoretical and
statistical problems, and explores the potential of multilevel models for studies of
international politics. A systemic approach to the democratic peace is examined as a
substantive example. International democratic norms have a positive effect on third party
settlement attempts to territorial disputes. Depending on the regime type of dyadic states,
the behavior of dyadic pairs of states varies. That is, as the proportion of democracies in
the international system increases, nondemocratic states are more likely to rely on such

peaceful measures as third party intervention to settle their territorial disputes.

Multilevel Interaction And International Relations Studies

One of the persistent themes in research of international politics involves
analyzing the impact of international structural factors on the behavior of individual
states. The relationship between polarity and international stability (Bueno de Mesquita
1975; Deutsch and Singer 1964; Waltz 1979), the impact of international norms on third
party intervention (Dixon 1993; Mitchell 2002; Raymond 1993), and the inﬂuence of
international regimes on capital control policy (Kastner and Rector 2003) are such
examples. In these studies, the primary theoretical emphasis is placed on the structure of
the international system, the role of international norms, ideas, or international
governance (Finnemore and Sikkink 1998; Katzenstein 1996; Krasner 1983). Here, we
observe the existence of a hierarchical structure between multiple levels, where
international environments condition interactions between states.

Analysis of the dynamics of the interactions between multiple levels is demanding,
because it requires us to answer theoretical puzzles and statistical challenges inherent in
multilevel data. Multilevel data consists of multiple units of analysis, where, based on
theoretical expectations, one unit should be grouped or clustered with others: for instance,
states are grouped within the structure of the international system.

First, to be a powerful tool for analyzing the relationship between multiple levels,
multilevel analyses must explore how the whole system works, how multiple levels are
structured in the system, and how agents are connected with structures, properly handling
the issue of the relationship between agents and structures (Braumoeller 2003a; Buzan,

Jones, and Little 1993; Goertz 1994; Wendt 1987). By failing to explain a theoretical

relationship between multiple levels, our arguments suffer from a lack of theoretical
clarity and logical consistency.

Despite causal complexity4 in theoretical models (Braumoeller 2003b), it is
difﬁcult to find a well-deﬁned statistical model that accounts for the dynamic interplay
across levels of analysis with theoretical consistency. Statistically, the failure to account
for the hierarchical structure of the data may result in “inconsistent” parameter estimates
as well as “underestimation of the variance” of the estimated coefficients (Beck 2001;
Rodriguez and Goldman 1995). This problem is also related to unmeasured
heterogeneity. In this regard, this paper underscores the need for an empirical evaluation
of theoretical models in multilevel analyses.

Discussing theoretical and statistical issues in regard to modeling multilevel data
structures, this paper explores the potential of multilevel (hierarchical) modeling for
improving scholarship in international relations. Multilevel (hierarchical) modeling is
now established as the appropriate tool for modeling data with hierarchical structures
(Goldstein 1995; Hox 2002; Raudenbush and Bryk 2002; Steenbergen and Jones 2002).
Comparing multilevel modeling with other methods of modeling, this paper demonstrates
how multilevel modeling can signiﬁcantly improve the study of international politics.

In the following, ﬁrst, I discuss theoretical and Statistical issues related to
multilevel structure in the data commonly used in international relations research. The
theoretical and statistical potential of multilevel modeling in dealing with multilevel data

is examined in the second section. As a substantive example, the argument of the

 

4 If the effect of an explanatory variable on the dependent variable is in some ways affected by the other
explanatory variables in a model, the causal path of impact is complex.

10

systemic approach to democratic peace is examined and empirically tested in the third

section.

Problems Caused By Disregarding Multilevel Structure

Theoretical Issues
Agent-Structure Problem

A basic tenet of social science research is that scholars make basic assumptions
about the main actor and its goal. Based upon these assumptions, the level of analysis is
determined. In multilevel analyses of international politics, however, it is often hard to
define these basic assumptions required for projecting a theoretical argument.

A systems—level theory in general deﬁnes a state as a unitary actor (agent), and
power, security, or wealth5 as its main goals. Structural (contextual) factors at the
international system level, which changes over time, are believed to constrain and set
boundaries on activities of individual states, a type of agents. For example, structural
features of the international system such as the number of major powers (whether or not
it is bipolar, multi-polar, etc.), the format of alliances, or the existence of a hegemonic
state, are assumed to have constraining effects on the behavior of states in the
international political arena, determining the nature of interactions between states
(Deutsch and Singer 1964; Krasner 1983; Keohane and Nye 1977; Waltz 1979). Since
structural factors external to individual states give shape to foreign policy, and thus
induce a common trend in international politics, the role of agents is negligible in this

approach. As a result, the main level of analysis is the international system.

 

5 This assumption is often used as one of the main criteria of differentiating realism, neo-realism, and neo-
liberalism in international politics.

11

Meanwhile, it is also argued that agents are autonomous in determining
international political outcomes. State-centric properties or dyadic state relational factors
are believed to affect international political events. From this perspective, theoretical
emphasis is given to the role of agents and their interactions rather than international
structural conditions in explaining international political consequences. Structure
explains only the possibilities of agents’ action, and even structure can be the outcome
caused by agents’ interactions (Dessler 1989)6. Agents are the main unit of analysis in
this approach.

Students of international politics, therefore, face the longstanding question about
the “agent-structure problem.” The agent-structure problem indicates the difficulty of
identifying a clear deterministic cause at the level of structure or agents (Giddens 1979;
Waltz 1979; Wendt 1987, 1999). Since “human agency is the only moving force behind
the actions, events, and outcomes of the social world” and at the same time “human
agency can be realized only in concrete historical circumstances that condition the
possibilities for action and inﬂuence its course” (Dessler 1989, p.443), it is a hard task to
explain both agents’ power and the causal relevance of structural factors in a logically
and ontologically consistent way. In this regard, this is also a unit of analysis problem
(Raudenbush and Bryk 2002).

Accordingly, any attempts to analyze multilevel data should clearly discuss the
ontological and epistemological standpoints about the agent-Structure problem in
explaining political phenomena (Braumoeller 2003a; Wendt 1987). Since multilevel data

consists of multiple actors playing at possibly multiple structures, solving the agent-

 

6 Dessler (1989) deﬁnes this agent-centric view as “the transformational model”, while he deﬁnes the
structure-centric view as “the positional model.”

12

structure problem becomes complicated. That is, in multilevel analyses such as the two-
level game analyses (Fearon 1994; Mo 1995; Morrow 1991; Putnam 1988) or studies of
the strategic approach, multi-units of actors at multi-levels are compounded and thus
detecting the main actor and its basic motivation becomes difficult. Here, an important
theoretical question is, “if two or more units and sources of explanations are operating
together, how are their different analyses to be assembled into a whole understanding?”
(Buzan 1995) Failing to properly answer this question is very likely to cause logical
inconsistency in an argument.
Structures as a Conditional Factor

As a way of dealing with the theoretical and empirical difficulty in connecting
agents with structure, studies of international politics often assume that agent-centric
factors and structural factors are independent from each other. These studies thus assume
that the effect of agent-centric variables on international political events is not inﬂuenced
by international structural variables. Even in this case, however, a relative weight of
theoretical importance should be assigned to each level. Nevertheless, not all scholars
clearly state how this weighting works. Methodologically it is not clear how it can be
done, either. More importantly, as the strategic approach explains, domestic and
international factors should often be interdependent. If this is the case, a theory should
tell the story about the interactions between structure and agents (Dessler 1989;
Lieberman 2002; Wendt 1987, 1999), and an empirical test Should distinguish agent-
centric variables from structure-centric variables, illustrating how the causal or

conditional relation between these variables can be examined.

13

When a group of explanatory variables in a model are not independent from each
other, the relationship between these variables can be causal, correlational, or conditional.
Causal relation means that one of the explanatory variables is endogenous to at least one
of the other explanatory variables. Since the existence of endogeneity in a model causes
biased inferences, researchers try to cope with this problem using various remedies
(King, Keohane, and Verba 1994).

Nevertheless, when it is theoretically difficult to identify a causal relationship
between two explanatory variables, although these variables are associated with each
other, we often refer to the relationship as correlational. In this case, since we do not
make a theoretical assumption about causality, the direction of impact between these
variables is bilateral. In order to solve the inefficiency problem caused by high degree of
collinearity, several remedial measures exist (Wooldridge 2000).

In models of international politics with multilevel data, the relationship between
explanatory variables is often conditional. Since there is no causality between variables,
the relationship is not causal. Since the interaction is not bilateral, the relationship is not
correlational, either. Conditional relationships indicate that, given a theoretical
framework, an explanatory variable constrains the causal path of impact of another
explanatory variable, although these variables are by nature independent from each other.
It is the case where “the effect of some explanatory variables on the dependent variable is
contingent on structural features that vary from country to country” (Beck 2001).

Conditional relationships between explanatory variables are typical in multilevel data,

14

where units at a low level can be grouped into a high level. Figure 17 describes the
conditional relationship between multiple levels.

As Figure 1 indicates, international systemic structural factors can have a
constraining impact on the behavior of dyadic states. Structure (or context) is not simply
another variable in a list of explanatory factors, because its inﬂuence is pervasive
(Western 1998). In a single level of analysis, international affairs between states are
mostly explained based on dyadic relational variables such as the relative size of power.
In multilevel analyses, however, international systemic structures, as constant features,
can limit states’ choices and their behavior. Structures of the international system vary
across different time points in the latter case, while domestic institutions vary across

states in the former case.

Statistical Issues

Heterogeneity And Type I Error

When we use pooled data, it is assumed the world is completely homogeneous.
Based on this assumption, one can obtain pooled estimates of a model. Unfortunately, the
complete homogeneity assumption is very unrealistic in the ‘real world,’ especially when
we deal with almost all nation states in our analyses over a long time period. Individual
states operate under heterogeneous structural environments. Since states share common
experiences to some degree given structural environments and their experience varies due

to changes of international systemic structure, norms, or culture, the complete

 

7 Goertz (1994) explains context (structure) in three different modes of relationship: context as cause,
context as barrier, and context as changing meaning. In the conditional relationship, context (structure) is
understood as changing meaning in his terminology. The relationship between cause and effect varies
according to the surroundings in which agents’ behavior occurs.

15

Figure 1. Theoretical Logic of the Regression Models

 

 

 

 

 

Dependent Variable 1 Levell: Explanatory Variable
(Dyadic Level Data)
Militarized Conﬂict or Alliance. Contiguity.
Third Party Intervention T Capability Ratio,
Economic Interdependence

 

 

 

 

 

Level 2: Constraining Variable

(Domestic Level Data) or
(International System Level Data)

 

 

 

l6

homogeneity assumption can hardly be maintained in data with a relatively long time
period.

To address heterogeneous properties, we need to cluster dyadic relations in
accordance with varying structures of the international system. By disregarding common
properties of states within a cluster, we fail to explain the correlation between states
within a cluster. Since states within a cluster show similar patterns of behavior, the
direction of correlation between states is often positive. This means that we are likely to
underestimate the Size of Standard errors in a statistical test, if we disregard
heterogeneous structure. As a result, our statistical test results are not credible. We will
commit a Type 1 error by accepting false hypotheses. In the worst cases, even wrongly
signed coefficients will be obtained (Raudenbush and Bryk 2002; Steenbergen and Jones
2002; Western 1998).

This problem is related to unmeasured heterogeneity. Unmeasured heterogeneity
is a type of omitted variable bias. Thus, if the unobservable structure speciﬁc effects are
not constant across time points, biased estimators will be obtained.8 If the underlying
relationship between the outcome variable and the explanatory variables is nonlinear, the
failure to account for the hierarchical structure of the data may result in “inconsistent”
parameter estimates as well as “underestimation of the variance” of the estimated

coefficients (Beck 2001; Rodriguez and Goldman 1995).

Why Multilevel Models?

Causal Heterogeneity, Uncertainty, and Conditional Eﬁ‘ect

 

8 See Greene (2000, pp.590-651).

17

In Figure l, the reality of the multilevel structure of international relations data is
empirically modeled in several ways: dummy variable models, interactive models, and
multilevel models. Multilevel models are also referred to as hierarchical models
(Goldstein 1995; Rasbash et al. 2002; Rodriguez and Goldman 1995; Western 1998),
random-coeﬁ‘icient models 9 in the econometrics literature (Rosenberg 1973), or
covariance components models in the statistical literature (Dempster, Rubin, and
Tsutakawa 1981).

Dummy variable models employ a series of dummies to measure institutional
difference. For example, when we believe that each state has its own speciﬁc domestic
institutions, and that the impact of state specific institutions on the relationship between
states is theoretically meaningful, we can introduce unique dummies for all of the
individual states (save one). ‘Temporal’ dummies can be used for capturing heterogeneity
or structural differences in the international system. Statistically, this way of modeling is
advantageous in that it accounts for the differences between heterogeneous states. On the
other hand, theoretically, dummy variable modeling provides no substantive explanation
about the origin of heterogeneity at the subgroup level because it does not introduce any
predictors from that level. Moreover, since this modeling exhausts the degrees of freedom
by including a number of dummy variables equivalent to the number of cases minus one,
this methodology often represents a problem, especially when the total number of
observations is small.

An interactive model includes an interaction term between an explanatory

variable at the subgroup level and another explanatory variable measuring heterogeneity

 

9 According to Raudenbush and Bryk’s (2002) terminology, the multilevel model is the same as the random
coefﬁcients model (RCM) when RCM allows both the intercept and the slope(s) in the model to be random,
introducing predictors of structural features.

18

at the group level. By including a predictor at the group level, interactive models provide
a substantive explanation on the origin of heterogeneity. This modeling method is
possible if we assume the partial effect of an explanatory variable on a dependent
variable depends upon the magnitude of another explanatory variable (Wooldridge 2000).
The decision to introduce an interaction term is generally based on theoretical reasoning
that there is correlation between two variables.

Unfortunately, this way of model specification also entails costs. First, by using
an interaction term, we allow the partial effects of both variables in the interaction term to
be mutually dependent on each other. In other words, the direction of correlation in the
interaction term is bilateral. However, we cannot keep the bilateral dependence
assumption in all cases.

In multilevel analyses, this is significant problem. This is because we should
clarify the ontological relationship between agents and structure (or between multiple
levels) and also because the ontology of a model varies depending on a theoretical belief.
For instance, we can assume structure has dominant inﬂuence over an outcome. Thus,
structural features at the system level (for instance, the international system level) are
considered importantly in a model, while agent-centric variables (for instance, dyadic
state level variables) are believed to have only a minor impact on the outcome. Agents
play virtually little role in this case (Waltz 1979).

Another ontological position assumes a reverse relationship between structure and
agents. Agent—centric variables are mainly used to explain international political
consequences. Moreover, agents’ interaction may bring changes in the structure of the

international system. Finally, we can assume that both structure and agents are equally

l9

important and that the relationship between agents and structure (or between multiple
levels) is reciprocal (Dessler 1989; Wendt 1999). While structure constrains agents’
choice and behavior, agents’ interaction may also change the structure of the international
system. Likewise, models with a multilevel data structure should clarify the ontological
relationship between multiple levels. In interactive models, however, we cannot easily
specify the ontology of the model.

In addition, interactive models must make the unrealistic assumption that there is

I.10 In other words, by

no remaining heterogeneity at the group (international system) leve
relying on the unrealistic assumption that one or a few predictors fully account for the
variation at the group level (the zero error term assumption), modeling method becomes
statistically incomplete (Steenbergen and Jones 2002).

Compared to previous ways of modeling, multilevel models have several
advantages. First, by allowing the intercept in a model to be random, they can identify
group (structure) effects. A statistical test will show if there are significant group effects.
Second, monitoring if the effect of a subgroup level variable on a dependent variable
plays out differently as institutional settings change, multilevel models can explain causal
heterogeneity. Theoretically, we can clearly make a distinction between causal effects of
subgroup level variables and constraining effects of group level variables. Third, by
tracing dyadic states’ behavior trajectory over structure, we can identify the effect of
dyadic interaction. That is, if dyadic states react to the structural changes in a different

way, it shows that agents can behave autonomously, independent from structural

constraint. Finally, although multilevel modeling is similar to interactive modeling in that

 

‘0 Steenbergen and Jones (2002) provide a clear explanation about the differences between these three ways
of modeling. Refer to their article for details.

20

it introduces predictors for the existing heterogeneity into the model specification,
multilevel modeling does not require holding the zero error term assumption. In this
sense, multilevel (hierarchical) models supply more accurate forecasts and estimates,
providing “a more realistic account of all the sources of uncertainty associated with the
comparative research design than the model with interaction terms” (Western. 1998,
p.1238). In addition, compared to complete pooling or un- pooling methods, multilevel
models as a partial pooling method provide compromising estimates between the two
extreme methods. As a result, estimates of multilevel models are more realistic than
others.

In order to see how multilevel modeling works and how significantly different
results are brought about depending on which of the above-mentioned methodologies is
employed, in the next section I test a model of a systemic theory of the democratic peace

using a multilevel model.

Application of the Multilevel Model Method
Democratic Norms And Third Party Conﬂict Resolution

In this section, we examine a systemic approach to the democratic peace to
demonstrate the application of the multilevel model method to international relations
studies. According to normative dyadic approaches to the democratic peace, democratic
dyads should be more likely to resolve disputes peacefully and also turn to third parties in
the dispute resolution process. This is because democratic states share an affinity with
international legal processes and institutions (Simmons 1999), favor the use of

compromise and nonviolence (Dixon 1993; Russett 1993), or create a norm of trust in

21

legal procedures (Raymond 1994). The means of third party conﬂict resolution can
include mediation, arbitration, and so on. A systemic approach to the democratic peace,
meanwhile, sheds light on the interactions between democratic states and the behavior of
nondemocratic states. While most of the democratic peace literature focuses on a
tendency of democracies, this line of research pays attention to the idea that democratic
norms can become international norms affecting a tendency of nondemocracies (Mitchell
2002). That is, as the number of democracies and interactions between democracies
increase, democratic norms can be externalized into international norms and therefore
even nondemocratic states can adopt democratic norms.

Democratic norms can affect the behavior of nondemocratic dyads in several
ways. The increasing number of peaceful territorial dispute settlements through third
party intervention between democracies can provide useful lessons to nondemocratic
dyads. In addition, the increasing proportion of democracies in the international system
indicates the increase of the number of third parties that maintain a good reputation as
mediators. In other words, as the number of democracies increases, the opportunity for
solving disputes through democratic third parties also increases for nondemocratic dyads.
When one of the member states in a nondemocratic dyad11 is a democracy, the dyad is
more likely to solve territorial dispute through third party intervention than a
nondemocratic dyad where neither of the members in the dyad is a democracy. This is
because one of the dyadic partners holds democratic norms. The existence of a
democratic partner in a nondemocratic dyad can be a good signal to the other partner in
terms of succeeding in dispute resolution through third party intervention. This signaling

is intensified as the proportion of democratic states increase.

 

” A dyad is deﬁned to be nondemocratic when one (or both states) in a dyad is nondemocratic.

22

Hypothesis

Any test of a systemic theory of the democratic peace that fails to distinguish the
structural effect of the international system on states from the effect of dyadic relational
factors will provide an incomplete, and quite possibly inaccurate, evaluation of the theory
(Waltz 1979)”. In other words, the effect of international norms on states must be
distinguished from the effect of dyadic relational factors. To sort out the systemic effect
of international norms on states from the effect of dyadic relational variables, ﬁrst the
theoretical relationship between the international system unit and the dyadic unit (i.e., the
relationship between structure and agents) should be clariﬁed, and then the processes of
interactions between the two units should be explained in a consistent manner.

If structural factors and dyad-centric factors are assumed to interact with each
other, this seemingly complex relationship can be understood time serially. First, given a
time point, states’ behavior is inﬂuenced by structural factors. Prevailing international
norms will drive all states including nondemocracies to behave in a similar way. In this
case, international norms or Structure has conditional effects on states’ behavior.
However, the relationship between structure and agents is not unilateral in that interaction
between individual states can cause changes in the implication of structural effect of the
international system with respect to third party intervention over time. In other words, if
interaction between democratic states has an impact on the behavior of nondemocratic
states, and thus causes changes in the effect of international norms on nondemocratic
states’ behavior over time, these two types of dyadic states will exhibit different trends in

their behavior regarding third party dispute resolution. That is, as interaction between

 

'2 Waltz argues that without showing how the systems level, or structure, is distinct from the level of
interacting units, “one does not have a systems approach or a systems theory at all” (1979, p.40).

23

democratic states increases and accordingly the effect of democratic international norms
on the behavior of nondemocracies is intensified, nondemocracies will adopt third party
intervention at a higher rate than democracies do. Therefore, if nondemocracies show
significantly different trends over time from democracies with respect to the likelihood of
third-party dispute resolution, we can identify the agents’ effect. The main hypothesis of

the systemic theory of democratic peace is provided below.

Hypothesis 1.1: As the proportion of democracies in the international system
increases, the likelihood of third-party dispute resolution

in nondemocratic dyadic disputes increases.

In this hypothesis”, the main actors are nondemocratic (dyadic) states and the structure
factor is the proportion of democracies in the international system. Under this hypothesis
we expect that as the proportion of democracies increases, so to will the extent of third
party intervention. In addition, nondemocratic states are more likely to adopt democratic
norms as the proportion of democracies increases, thus increasing third party intervention.
To test this hypothesis, a multilevel model, rather than an interactive model, is the ideal
vehicle to employ given its ability to specify the hierarchical structure present in the data,

not to mention its above-mentioned statistical advantages discussed.

Empirical Analysis

 

‘3 This hypothesis is actually tested in Mitchell’s article (2002). By showing the signiﬁcance of this
hypothesis in a series of tests, her work makes an invaluable contribution to the development of a systemic
theory of the democratic peace.

24

Speciﬁcation and OperationalizationM

In the hypothesis, the dependent variable is a dummy variable coded one for third-
party attempts to settle a territorial claims and zero for bilateral settlement attempts. Six
hundred and ninety different settlement attempts, for a total of 114 dyadic territorial
claims in the Americas (North, Central, and South America) from 1816 to 1997 are
analyzed. The data were collected by the Issue Correlates of War (ICOW) Project
(Hensel 1999, 2001). The Polity 111 data set (J aggers and Curt 1995) is used to measure a
nation’s democratic status. Any nation with a score of six or higher on the scale (ranging
from 0 to 10) of the democracy variable in the data set is considered to be democratic. To
create a dyadic measure of regime type, if one or both nations in a dyad are
nondemocratic, it is coded as one, and zero otherwise. The same data set is used to
measure the annual proportion of democracies in the international system.

Multilevel models are used when clustering observations is possible and
meaningful. Therefore, the decision regarding how to cluster (group) dyadic observations
is very important. If there is any systematic change over time, time also plays a role as a
high level in a hierarchy (Goldstein 1995; Raudenbush et al. 2002). Since the structure of
the international system has group effects on dyadic behavior in a time period and the
effects change over time by assumption, time consists of a hierarchical structure in this
model. Here, dyadic observations are clustered within time periods in two ways. There
are 690 observations over a 182 year of time period in the data. Therefore, the average
number of cases per year is very small, which might bring incorrect estimates

(Raudenbush and Bryk 2002). In addition, it is hard to say that the structure of

 

‘4 The data used in this paper come primarily from Mitchell’s research (2002). I am very grateful to her for
sharing this data.

25

international system changes yearly. To solve these problems, 1 clustered all observations
by 5 year time periods, resulting in 37 clusters. Due to the concern that this way of
clustering may have a signiﬁcant effect on outcomes, as a diagnostics I also conducted
analysis where I grouped observations into 19 clusters by 10 year time period.
Comparison of these two multilevel models will show how different ways of clustering
affect the results.

There are three control variables. If one side is a dominant power, it should be
able to impose its preferred settlement (Hansel 2001). Thus, power asymmetry is
expected to have negative effect on third party intervention. Power asymmetry is
measured by the capability ratio of the challenger to the target. The Correlates of War
composite capability index is used for capability, which is created with the EUGene
program (Bennett and Stam 1998). If a dyad has experienced prior militarized disputes, it
will attract the attention of third parties and this will therefore increase the likelihood of
third party settlement attempts (Dixon 1994; Hansel 2001). This second control variable
is coded as one if a dyad fought a militarized dispute in the ten years prior to the start of
settlement attempt, and zero otherwise. The ﬁnal control variable indicates whether there
have been any prior attempts to manage the territorial claim. Prior management activity is
expected to increase the likelihood of third party intervention (Dixon 1993).

The Multilevel Logit Model
To perform a multilevel analysis with the binary dependent variable, the multilevel
logit model is used. The multilevel (in our case, two level) logit model can be written in

the general form

ﬁ.—,-=f(X.~,ﬂ,~), (1.1)

26

where 71",,- is the probability that the i-th dyadic territorial claim in the j—th structure of the

international system is associated with third party intervention. f is a nonlinear function

(logit link function) of the linear predictor X,,,6,. We need to specify a distribution for the
observed response y,,|7r,,. The y,, are the observed (0,1) responses with the standard

assumption that they are binomially distributed y,, ~ Binomial(n 7r,,) , where n,, is the

q"
denominator for the probability 7r,,. We also have var(y,,|7r,,) = 7r,,(l — 7r,,)/n,, . The
model including the dyadic level variation can now be written as

Y.) : ”11+ 6,,x0 r

where x0 is used to specify the variation at the first level. With the assumptions that the

n,, are all equal to 1 and x0 is also equal to 1,15 the variance of the dyadic level random

term 6,, is rr,,(l—rr,,). Full details can be found in Goldstein’s (1995), Rodriguez and

Goldman’s (1995), and Rasbash et al.’s (2002).
The usual interactive model with binary response is specified in the following form:
logit(7r,) = ,30 + ,6, * Asymmetry, + ,8, * priorMID, + ,63 * priorManagement, +
,6, * Nondemocracy, + ,35 * Nondemocracy * proportion, + ,66 * proportion, + 8, ,
where It, is the probability that the i-th dyadic territorial claim is resolved by third party

intervention. In this interactive model, we have to assume that the proportion of

democracies perfectly account for the effect of international system on third party

 

'5 This assumption means that “the binomial model is assumed to hold precisely” (Hox 1995, 77). To test if
a Binomial assumption holds or if the model exhibits ‘extra Binomial’ variation (overdispersion), I obtained
a dyadic level variance based on the extra Binomial distributional assumption and reported it in Table 1.
The dyadic level variance multiplier is only slightly different from one. And the other parameters obtained
with the extra Binomial distributional assumption have hardly changes (these results are not reported here).
As a result, there is little evidence that the response is not Binomially distributed.

27

intervention, which could be incorrect. In other words, the interactive model assumes no
error term at the international system level. But, there could be other international
systemic factors such as the number of major power (poles), which affect the behavior of
nondemocratic dyads.

With the two level logit model, we define the probability a, as a function of the
intercept and linear predictors as follows:
logit(7r,,) = ,60, + ,6, * Asymmetry, + ,6, * priorMIDU. + ﬂ, * priorManagementU +
,64, * Nondemocracy, + 8,,- . (1.2)
Additionally, the intercept and the coefficient of nondemocracy variable are functions of
the international system level predictor, proportion variable:
,60, = To + y, * proportion, +210, , (1.3)
ﬂu =a0+a, *proportion, +,u4,, (1.4)
where 7:, is the probability that the i-th dyadic territorial claim in the j-th structure of the
international system is resolved by third party intervention. This speciﬁcation allows the

coefﬁcient of dyadic type variable, ,64, , and the intercept term, 190,, to be a function of

the international system variable. That is, it assumes that the proportion of democracies in
the international system affects the behavior of nondemocratic states in regard to third
party intervention (the slope) as well as the probability of third party intervention for all
states (the intercept term). The subscript j denotes different clusters in terms of the
proportion of democracies in the international system. To simplify the model, we
substitute Equations (1.3) and (1.4) into Equation (1.2):

logit(n',,) = (}/0 + y, * proportion, + #0,) + ,6, * Asymmetry, + 52 * priorMID,,

28

+ﬂ3 * priorManagementU + (a0 + a, * proportion, + #4,) * Nondemocracy, + 6,,
= yo + ,6, * Asymmetry, + ,8, * priorMID,, + ,63 * priorManagement, + y, * proportion,
+a'0 *Abndemocracy, +0!l *proportion, *Nonderrrracy, +,u4, * Nondemocracy, + #0, + 60.. (1.5)
In this Equation (1.5), yo denotes the intercept and y, is the effect of the international
system level predictor. ,B,,,62,,63 and are are the effect of the dyadic level predictors. a,

is the effect of the cross-level interaction between the dyadic and the international system
levels. The multilevel disturbance term is composed of the three different

components, porn“, and 6,. ,uo, is the international system level disturbance in the
equation (1.3) of the dyadic level intercept. #4, is the international system level
disturbance in the equation (1.4) of the dyadic level slope. 8,, is the dyadic level

disturbance capturing any possibly omitted predictors at the dyadic level. As a result, the
multilevel disturbance term does not have constant variance. The multilevel disturbance

variance is a function of the dyadic type variable unless #4, is zero.

Estimation

To estimate parameters illustrated above, we need to specify a few assumptions
about the three components of the disturbance term. First, like in OLS analysis, we make
the zero mean assumption for the three components of the multilevel disturbance

term: Elaoj]=EIp,,]=EIe,,]=O. Each component originated from the international system
level has a constant variance and a normal distribution: Variao,]=a)2, Varip4,]=r2. The

component from the dyadic level has a constant variance: Var{£,,]=0'2. The variance of
the 6,, is given as 1. This means the model is fitted to have a Binomial error distribution.

Since the international system level disturbances on the intercept and the slope may be

29

correlated, we need to estimate the covariance term between these disturbances (Snijders

and Bosker 1999): Covlﬂoj.#4,-]=It/. To identify the multilevel model, we also need to

assume zero covariance between the system level disturbances in both the intercept and

the slope equations and the dyadic level disturbance: Cov[,u0,,e,,] = Cov[,u4,,e,,] =0.

In order to estimate the parameters in equation (1.1), we need to linearize the
exponential function and then to apply quasi-likelihood estimation using the binomial
distribution assumption to deﬁne the dyadic level variation (Goldstein and Rasbash 1996).
I use the iterative generalized least squares (IGLS) algorithm for implementing
estimation. Since the IGLS procedure produces biased estimates in general by taking no
account of the sampling variation of the fixed parameters especially when the sample size
is small, I use the restricted version of iterative generalized least squares (RIGLS)
algorithm that leads to unbiased estimates of random parameters (Goldstein 1995).'6 I
use a stringent convergence criterion for the parameter estimates. The relative change
from one iteration to the next is at most 0.001. As a specific method for linearization and
estimation, I use the second-order penalized quasi-likelihood (PQL) procedure to obtain
estimates. As the previous studies conﬁrm (Goldstein 1995; Goldstein and Rasbash 1996;
Raudenbush and Bryk 2002), the second-order PQL procedure produces almost unbiased
estimates for the fixed and the random parameters. MLwiN statistical package is used for

obtaining these estimates (Rasbash et a]. 2002).

Results

 

'6 For a detailed discussion and proofs, refer to Goldstein (1995) or Hox (1995).

30

Table 1 provides the estimates in all three different models: one interactive model
and two multilevel models.‘7 The nondemocracy variable is statistically insignificant in
all three models, while the negative sign of the coefficient in the interactive model
changed into a positive Sign in the second multilevel model. While the interaction term’s
coefﬁcient in the interactive model is statistically Significant, it is not in the two
multilevel models. This difference is generally due to the increase of its standard error.
Therefore, this result shows that we are likely to underestimate standard errors and thus
make erroneous conclusions about statistical significance by ignoring the multilevel data
structure. As expected, a history of costly conﬂict contributes to a higher likelihood of
third party intervention, while power asymmetry has negative effect on it. Prior
management activity is not statistically signiﬁcant in explaining third party intervention.

In regard to interpretation of the test results, however, special care is required.
The effect of the nondemocracy variable should be jointly evaluated with its interaction
term. It means we should not separately look at t-statistics of nondemocracy variable and
its interaction term (Wooldridge 2000). Although either the nondemocracy variable or its
interaction term, or both, is statistically insignificant, we cannot conclude that the
nondemocracy variable is statistically insigniﬁcant. The p-value for the Likelihood Ratio

(LR) test of the joint signiﬁcance in the interactive model is .0023, which means that both

 

'7 The estimates of the interactive model were obtained by replicating the model in Mitchell’s (2002) paper.
The estimates of the multilevel model and all other estimates in this paper were obtained using MLwiN
V1.1 (Rasbash et al. 2002).

31

Table 1. Third-Party Attempts to Settle Territorial Claims

 

Variable Interactive Model

 

Multilevel Model
Clustered by Clustered by
10-year term 5-year term

 

Proportion of Democracies 4.927** (1.933)

Nondemocratic Dyad — 0.421 (0.665)
Proportion*Nondem 4.737* (2.338)
Asymmetry — 0.0158* (0.008)
Prior MID 1.1869** (0.219)
Prior Management 0.2268 (0.361)
Constant — 3.77** (0.648)
Variance Components

System Level

Constant ( a?)

Nondemocracy (22)

Constant, Nondemocracy

(Covariance: I/l)

Dyadic Level (02)
12 (1) = 22.33**

4.062 (4.968)
——0.280 (1.376)
4.409 (5.012)
— 0.018**(0.007)
l.154**(0.333)
0.189 (0.452)

— 3.696**(1.525)

0.950 (0.523)
0.016 (0.251)
—0.304 (0.354)

a1.034=I=*(0.110)

6.592 (5.270)
0.106 (1.344)
3.084 (5.242)
- 0.017**(0.005)
1.122**(0.304)
0.201 (0.433)

— 4.327**(1.518)

1.199 (0.837)
0.827 (0.714)
—0.899 (0.749)

al.087**(0.144)

 

Note: Multilevel table entries are maximum likelihood (RIGLS) estimates with robust standard

errors in parentheses.
*=p<0.05,**=p<0.01

a By a Binomial distributional assumption, the random coefﬁcient at the ﬁrst (dyadic) level 5,,- was

deﬁned with a zero mean and a variance which was constrained to equal one. However, to test
whether or not a Binomial assumption holds or whether the model exhibits ‘extra Binomial’ variation,
I obtained a dyadic level variance based on the extra Binomial distributional assumption

and reported it here. As we see, the dyadic level variance multiplier is only slightly different from
one. And although it is not reported here, the other parameters have hardly changed. As a result,

there is little evidence that the response is not Binomially distributed.

32

variables are jointly signiﬁcant at the conventional significance level.18 The LR test
results in the two multilevel models give the same answer. The p-value for the LR test is
almost zero19 in each model.

If we simply look at the coefﬁcient of nondemocracy variable in the interactive
model or in the first multilevel model, we will incompletely conclude that nondemocracy
dyads have a negative effect on third party settlement attempts. Its positive Sign in the
second multilevel model is also conﬂicting with the one in the interactive model. But, the
coefﬁcient of the nondemocracy variable supposedly measures the effect when the
proportion of democracies is zero, which is not interesting and practically impossible (in
the sample, the smallest proportion of democracies is .01). To obtain the substantively
meaningful partial effect of the nondemocracy variable on third party settlement attempts,
we must plug-in interesting values of the proportion of democracies into both additive
term and interaction term.

Moreover, since the model is nonlinear, the effect of the nondemocracy variable
on the dependent variable is not constant. Accordingly, looking at the effect of the
nondemocracy variable on the changes in the predicted probability of third party
settlement attempts is preferable to estimating the partial effect. For this purpose, I
simulate the interactive model using “Clarify” program20 (King, Tomz, and Wittenberg

2000) in order to obtain detailed information regarding the predicted probability of third

 

'8 The LR test is based on the difference in the log-likelihood functions for the unrestricted and the
restricted models. In the restricted model, nondemocracy variable and its interaction term are dropped. The
likelihood ratio Statistic is twice the difference in the log-likelihoods: LR = 2[ln(Lu,) — ln(L,)]. The LR
statistic has an approximate chi-square distribution with two degrees of freedom.

'9 The LR test for joint signiﬁcance of the nondemocracy variable and its interaction term was performed
for each multilevel model. There are few differences between the results in the two models. The two
variables are jointly signiﬁcant at the 1% signiﬁcance level.

20 This program uses Monte Carlo simulation to convert the raw output of statistical procedures into
interesting results. For more details, refer to their article (King, Tomz, and Wittenberg 2000) or the program
website: httpzllgking.harvard.edu/clarify/docs/clarify.html.

33

party settlement attempts. To capture the effect of increasing the proportion of
democracies on the predicted probability and its 95% confidence intervals for both
democratic dyads and nondemocratic dyads, I vary the proportion of the democracy
variable from its minimum to maximum, holding all other variables at their mean values.
All binary variables are set equal to their modal value.

Figure 2 reveals that both democratic dyads and nondemocratic dyads are more
likely to rely on third party resolution as the proportion of democracies increases. It
signiﬁes that regardless of the regime type, all dyads are constrained by the structural
effect of the international system to some extent. To identify the systemic effect of
democratic norms on nondemocracies, we should observe that the increasing rate of third
party intervention in nondemocracies should be higher than the rate from democracies. In
fact, Figure 2 Shows that nondemocratic dyads have different trends in their behavior
from democratic dyads over time.” As a result, this is consistent with the hypothesis that
the democratic norm of third party intervention on territorial dispute settlement becomes
an international norm. One of the reasons why nondemocratic states are more sensitive to
the systemic effect than democratic states is that nondemocratic states have few
alternative methods for resolving territorial disputes. Democratic states have relatively
many options with respect to dispute settlement. Democratic states, which share
democratic norms, can peacefully resolve disputes bilaterally without relying on third
party intervention. The large variation in the probability of third party intervention for

democratic dyads in Figure 2 supports this reasoning.

 

2‘ Table 3 in Mitchell’s article (2002) shows a similar result. She explains this result in three ways. “The
percentage of democratic dyads in the Americans that have disputed over territory is quite small." “The
United states has been able to resolve the vast majority of its disputes bilaterally.” And, “Democracies are
initially able to resolve disputes peacefully without external aid.”

34

Probability of Third Party Intervention

1.0

0.8

0.6

0.4

0.2

0.0

* The solid lines indicate the mean and its 95% conﬁdence intervals of probability
of third party settlement attempts for democratic dyads, while the dotted lines
indicate the same terms for nondemocratic dyads. Before the proportion of
democracies reaches its 21% point, the behavior of democratic dyads and
nondemocratic dyads are not meaningfully distinguishable each other. However,
after the proportion of democracies exceeds its 21% point, the behavior of

nondemocratic dyads is signiﬁcantly different from that of democratic dyads.

Figure 2. Probability of Third Party Settlement Attempts

 

 

 

 

/
/
/ /
/ / /
Nondemocratic Dyads / / / /
/ / /
/ /
? é /
/
/ /
é
/ ﬁ / /
/ / Democratic
:://
0.0 0.1 0.2 0.3 0.4 0

Proportion of Democracies

Dyads

.5

Nondemocratic dyads are more likely to rely on third party resolution.

35

Of great interest, the ﬁgure shows that the nondemocracy variable (and thus
democratic norms) is substantively meaningful when the proportion of democracies in the
international system becomes about 21% or higher. Before the proportion of democracies
reaches 21%, the effect of the systemic variable on the behavior of nondemocratic dyads
is not distinguishable from that on the behavior of democratic dyads. This contributes to
the statistical insignificance of the nondemocracy variable. However, after the proportion
of democracies exceeds 21%, the behavior of nondemocratic dyads is signiﬁcantly
different from that of democratic dyads.22 This finding illustrates when the systemic
effect of democratic norms on nondemocratic dyads can be effective in the international
system.

With the signiﬁcant theoretical implication of the interactive model, this way of
modeling entails cost. It is based on the mutual dependence assumption between two
components in the interaction term, and it also requires the zero error term assumption at
the international system level, which are inappropriate and unrealistic. We might have the
underestimation problem. In multilevel models, however, we don’t have these problems
anymore, and even we get additional information: the necessity of introducing variables
at the international system level into the model specification, and statistical test results of
the multiple error components.

Table 2 shows the results of analysis of variance (ANOVA). This test is important
because it shows whether there is significant variation in third party intervention at both
the dyadic level and the international system level. All of the variance components are

statistically signiﬁcant in the two multilevel models. This means that both dyadic and

 

22 Gleditsch and Hegre (1997) make a similar argument. They report that with increasing democratization
over time, the frequency of war starts to decline only after 36% of states are democratized.

36

Table 2. ANOVA (Analysis of variance)

 

 

 

 

Parameter 10- Year Term(19 clusters) 5 - Year Term( 3 7 clusters)

Fixed Effects 0.268** (0.045) 0.269** (0.034)

Constant

Variance

Components

System Level 0.033** (0.013) 0.032** (0.009)
16.3% 15.9%

Dyad Level 0.170** (0.009) 0.169** (0.014)
83.7% 84.1%

** = p < .01

Note: Table entries are maximum likelihood (IGLS) estimates with estimated standard
errors in parentheses and the percentages of the variance in third party intervention. The
ratio of each variance component to the total variance in third party intervention can be
used to indicate the importance of each level in the model. Since this comparison is
difﬁcult to be done for a binary model, I fitted the model as a Normal response model as
a crude way of performing this test (Rasbash et al. 2002). The result suggests that by
ignoring the multilevel character of third party intervention, we will miss an important
source of variance at the system level. It also shows that the method of clustering has a

negligible effect on the ratio of each variance component to the total variance.

37

international system levels are important for third party intervention. The ratio of each
variance component to the total variance in third party intervention can be used to
indicate the substantive importance of each level in the model (Rasbash et al. 2002;
Steenbergen and Jones 2002). Since this comparison is difﬁcult to carry out for a binary
model, I ﬁt the model as a Normal response model as a crude way of performing this test
as recommended by Rasbash et a1. (2002). The international system level accounts for
about 16% of the total variance in third party intervention, while the dyadic level explains
about 84% of the total variance. The result suggests that by ignoring the multilevel
character of third party intervention, we will miss an important source of variance in third
party intervention and make incomplete conclusions about third party intervention. It
underscores the validity of fitting this model as a multilevel model. Table 2 also
demonstrates that the method of clustering has a negligible effect on the size of variance
in each level.

Figure 3, obtained from the ﬁrst multilevel model with 19 clusters, Shows changes
in the means of predicted probability of third party intervention and their 90 percent
conﬁdence intervals. For this purpose, I employ Markov Chain Monte Carlo (MCMC)
estimation methods with the Metropolis-Hastings (MH) algorithm (see J ackman 2000 for
details).23 The output of MLES is used to initialize the MH sampler. A standard Wishart

prior is assumed for the international system level variance parameter.24

 

23 Since I use MCMC methods only to obtain this ﬁgure, detailed explanation on this method is skipped in
this paper to save space.

2‘ To perform convergence diagnostics, I utilize two distinct convergence diagnostics (Geweke,
Heidelberger and Welch) (Cowles and Carlin 1996; Smith 2003). These diagnostics indicate that the sample
output perfectly converges to the posterior distribution.

38

Figure 3. Third Party Settlement Attempts

1.0

 

0.8 - /

Nondemocratic Dyads / / /

 

Probability of Third Party Intervention

/ /
M \Democratic Dyads
/

0.0 I j I I
0.0 0.1 0.2 0.3 0.4 0.5

 

 

 

 

 

Proportion of Democracies

* This ﬁgure comes from the first multilevel model with 19 clusters. The output is
obtained by the Markov Chain Monte Carlo (MCMC) simulation method. These lines
indicate the mean of the last 10,000 samples of the 150,000 sample run and its interval
from the 5‘h to the 95th percentiles of these samples. The solid lines indicate probability of
third party settlement attempts for democratic dyads, while the dotted lines indicate the
same term for nondemocratic dyads. As the proportion of democracies increases, the
behavioral difference between these two types of states gets bigger. Nondemocratic dyads
are more likely to rely on third party resolution. The slope of the mean line of
nondemocratic dyads is steeper than the one of the democratic dyads. The variation of the

nondemocratic dyads is smaller than the one of the democratic dyads.

39

This figure illustrates that the proportion of democracies has an increasing effect
on the probability of third party intervention in both nondemocratic dyads and democratic
dyads. Furthermore, it shows that the behavioral difference between nondemocratic dyads
and democratic dyads become larger, as the proportion of democracies increases. The
slope of the nondemocratic dyads’ line is steeper than that of the democratic dyads. Since
the multilevel model solves the underestimate problem, the confidence intervals are wider
than the ones in Figure 2. As Figure 2 informs, the nondemocracy variable is
substantively meaningful when the proportion of democracies reaches a certain level,
about 28%.25 As a result, it identiﬁes the necessity of clustering structure differences in
the international system when we look at structural effect on dyadic behavior.26 It also
implies that the effect of the nondemocracy variable (agent-centric variable) should
simultaneously be evaluated with the constraining effect of international systems or
international norms especially in nonlinear models.

The systemic variable, the proportion of democracies, is statistically significant in
the interactive model, while it is not statistically significance in the two multilevel models
at the conventional signiﬁcance level. However, to evaluate the effect of the proportion
of democracies on third party intervention, again, we have to look at the joint signiﬁcance
of this variable with its interaction term using the LR test. The p-values for the LR test of
the joint signiﬁcance in both the interactive model and the two multilevel models are

almost zero, which means that both variables are jointly signiﬁcant.

 

25 Gleditsch and Hegre (1997) make a similar argument. They report that with increasing democratization
over time, the frequency of war starts to decline only after 36% of states are democratized.

26 The White test also conﬁrmed the existence of heteroskedasticity problem. A graphical method of
checking the source of the heteroskedastic variance informed that the proportion of democracies
contributed to the existence of heteroskedasticity.

40

’11

Figure 4 shows how the proportion of third party intervention changes in each
type of nondemocratic dyad, as the proportion of democracies in the international system
increases. The proportion of third party intervention in each type of nondemocratic dyad
greatly increases as the proportion of democracies increases. This figure shows that the
systemic constraining power of democratic norms on nondemocratic states can be
effective both indirectly through a democratic partner in the dyad and directly through the
international system. Especially noteworthy, when the proportion of democracies exceeds
a certain point, approximately 21% of democracies in the international system,

nondemocratic dyads are more involved in third party intervention.

A hypothesis test for the variance (22) and covariance (to) of the system level

random term (6,) is interesting in multilevel models. This test shows if there is still

signiﬁcant effect of this random term on the unexplained variation at the system level in a
multilevel model after we controlled the effect of nondemocracy variable on the
heterogeneous variance in this model. I carried out the Wald test for this purpose. This
yielded a chi-squared value for the joint hypothesis that both variance and covariance
parameters are zero of 1.467 on 2 degrees of freedom. As a consequence, this test
provides little evidence for an effect at the 5% signiﬁcance level. The separate test for
each term in Table 1 conﬁrms this result. Neither the variance nor the covariance of the
system level disturbance is statistically signiﬁcant at the 5% significance level.

Discussion

41

Figure 4. Third Party Settlement Attempts

 

30

25-

Probability of Third Party Intervention

 

 

 

I I I r I I I I

2 3 4 5 6 7 8 9 10

Proportion of Democracies

 

+ One Democracy
--O-- No Democracy

 

 

 

* This figure shows the changes of proportion of third party intervention for
nondemocratic states due to the increase of proportion of democracies in the international
system. It shows that the structural changes of the international system affects the
behavior of nondemocratic states both directly and indirectly. Regardless of the specific
type of a nondemocratic dyad (whether it has one democracy), the proportion of third
party intervention increases as the proportion of democracies increases. Especially when
the proportion of democracies exceeds a certain level, a nondemocratic dyad with one

democracy comes to behave signiﬁcantly differently.

42

,ll

The multilevel model is very useful in testing the institutional (systemic) effect on
individual states. Multilevel modeling is possible when we are able to cluster structures
of the international system that vary over time. Unlike studies in comparative politics,
however, there is no general agreement (or at least understanding) on the way of
clustering units (in this case, the structure of the international system) like parties for
individuals, nation states for parties, or regions for states. Introducing year dummies into
a model is the conventional method of clustering structural differences in the
international system. However, in a model with year dummies, we are unable to explain
what factor actually accounts for the structural differences.

The nature of differences between clusters can be qualitative or categorical. For
example, if the unit of clustering in multilevel data is the school district for students or
the region for states, the systemic variable is qualitative and categorical. In this case, the
number of clusters carries theoretical importance. The “names” of districts or regions
convey theoretically important meaning. On the other hand, if the difference between
clusters is quantitative, how many clusters we have in multilevel data is not as important
as in the former case. In this case, theoretical emphasis will be given to the origin of the
difference between clusters, not to the number of clusters. Unlike schools or regions,
clustering observations within time periods is quantitative. Therefore, the number of
clusters is not so important, once we have enough observations in each cluster. The
results in the two multilevel models conﬁrm this argument, suggesting that the method of
clustering does not significantly bring changes in interpreting theoretibal implication.

Meanwhile, there could be various ways in which the international system affects

the behavior of individual states. The applicability of multilevel modeling into various

43

1"

multilevel analyses of international politics should be tested in further studies. Efforts to
develop and reﬁne appropriate methods of clustering structural differences in the

international system should be continued.

Conclusion

This paper highlights the importance of the empirical evaluation of theoretical
models. Modeling multilevel structure in the data of international politics is an issue of
speciﬁcation. Inappropriately specified models cannot properly account for a theory. On
the other hand, well-deﬁned statistical models can make a theoretical argument sharp and
distinct. In this regard, this paper contributes to theory testing and development in
international politics by introducing multilevel modeling into the literature of
structuralism.

Systemic theories stress the effect of the international systemic environment on
the behavior of states and states’ response to the system. In spite of the theoretical
importance of this perspective, scholarly effort to test these systemic theories has to date
been insufﬁcient. The multilevel model provides an adequate test of systemic theories
and shows how states interact with the structure of the international system. The paper’s
results, employing a multilevel model, support the systemic account for democratic
peace: nondemocratic nations are more likely to involve third parties in their dispute
resolution processes as the proportion of democracies increases. The spread of
democratic norms into nondemocratic states identiﬁes the effect of democratic norms in
the international system. Recognizing this trend in nondemocratic dyads, this paper’s

findings also bolster the arguments of constructivism. They demonstrate the manner in

which international norms become effective in changing the behavior of states in the
international political arena.

By ignoring multilevel structure in the data or incorrectly defining the relationship
between multiple levels, we are likely to have invalid and unreliable estimates in an
empirical model. As a strong analytical tool for dealing with multilevel data, a multilevel
model provides more accurate estimates than an interactive model, taking account of
uncertainty at the international systemic level. A multilevel model also clariﬁes the
theoretical relationship between agent-centric variables and structure-centric variables.
By ignoring constraining effects of structure on agents’ behavior and also by overlooking
agents’ response to structural constraints, we can fail in making correct prediction on the
trend of agents’ behavior. As all three models in this paper identify, the increase in the
proportion of democracies in the international system (the strength of structural effect)
has an impact on both democratic and nondemocratic dyads’ behavior. It demonstrates
that agents’ behavior is dependent upon the constraining power of structure. More
interestingly, after the proportion of democracies exceeds a certain threshold,
nondemocracies’ behavoir is significantly different from that of democracies. As a result,

this research illustrates how agents and structures interact with each other.

45

l

CHAPTER II

Tracing Monadic Effects on Dyadic Behavior:

A Strategic Approach to the Liberal Peace

46

'l

Abstract

Foreign policy decisions are often strategic in that international politics is
intertwined with domestic politics. By ignoring the constraining effect of domestic
politics on foreign policy decisions, studies in international politics suffer from
signiﬁcant theoretical and empirical deﬁciencies. Paying attention to two types of
heterogeneity inherent in a dyadic approach to the liberal peace, this research aggregates
dyadic observations into individual states. Political leaders’ strategic positions in
domestic politics as the main source of heterogeneity at the sate level are analyzed. To
maintain theoretical consistency in explaining the relationship between domestic politics
and international politics, this analysis uses a multilevel hierarchical model. This way of
modeling helps clarify the constraining effect of domestic institutions on international
politics. As a result, this research provides a way of examining the heterogeneous effect
of economic interdependence on interstate conﬂict, which depends on domestic
institutions. In short, by highlighting multi-level data structure in decision-making and
empirically testing multilevel models using a Bayesian simulation method, this research

traces monadic effects on dyadic behavior.

47

Introduction

A host of scholars have refined the liberal peace arguments over the last decade
(e. g. Maoz and Russett 1993; Oneal and Russet 1999, 2002; Barbieri and Schneider 1999;
Beck, Katz, and Tucker 1998; Hegre 2000). Arguing that an expansion of civil liberty
enhances the prospects for peace in the international system, this scholarship has
explored the link between domestic institutional arrangements or liberty norms and the
onset of interstate conﬂict. The central argument has regarded the pacifying effect of
economic interdependence or democratic regimes on interstate conﬂict. This scholarship
has made signiﬁcant theoretical and empirical developments in international relations
studies.27

Despite this progress, these studies still limited in their ability to respond to some
theoretical and empirical challenges. First, the literature on the liberal peace lacks a
strategic perspective, overlooking the role of political actors (agents) in explaining the
relationship between domestic politics and international politics. Political leaders are
strategic decision makers in that they are selective in choosing foreign policies, given
domestic and international political situations. For instance, depending on political
leaders’ personal stability in domestic politics and their strategic consideration on
political costs, decisions concerning engagement in international conﬂict can vary across
nation states and even in a single state over time. Since theoretical justiﬁcation of the
liberal peace is built on its emphasis on domestic politics, liberal peace studies should
explore the relationship between the political leadership and civil society and its impact

on interstate interactions.

 

27 Especially, this scholarship has contributed to the improvement of statistical analysis. It centers on the
unit of analysis, rare events, temporal dependence, endogeneity, unmeasured heterogeneity, etc.

48

f .

Second, in pursuit of the link between domestic institutions and international
politics, recent international relations studies accept comparative political methods (e.g.
Bueno de Mesquita, Morrow, Siverson, and Smith 2000, 1999; Hegre 2000). However,
comparative cross-country studies have been handicapped “by a lack of detailed data on
the political and institutional characteristics of countries, and their change over time”
(Beck, Clarke, Groff, Keefer, and Walsh 2001). Since international relations (IR) studies
have relied on pooled data or time-series-cross-sectional data (TSCS)28, the limitation in
obtaining functionally equivalent data in comparative studies is also a serious obstacle for
IR scholars in specifying domestic institutional conditions under which governments or
political leaders make decisions. To examine the nexus of domestic institutions to foreign
policies, we have to answer two fundamental questions: Which domestic political
institutions are relevant to decision-making? And under what conditions do such
institutions matter? Unfortunately, data providing a disaggregated picture of a particular
state’s political context and simultaneously maintaining a strong power of functional
equivalence across states have been rare.29

Third, to incorporate domestic covariates into the analysis, IR scholars often use
directed dyad year format of data. It implies that we take dyads as units of analysis and at
the same time treat individual states in a dyad as another units of analysis. Accordingly,

the use of directed dyad data requires us to answer theoretical and statistical questions.

Theoretically, we face the unit of analysis problem. To maintain logical consistency, we

 

28 In the TSCS data, cross-sectional observations (N) are not randomly selected (Beck 2001). Accordingly,
country-speciﬁc factors carry substantive meaning more in this data structure than in other data structure
like the panel data.

29 One of the popular concepts used in Studies of domestic politics is “the winning coalition”. However, the
previous used measures for this concept have been time-invariant in most cases, and thus could not fully
capture the changes of domestic institutional impact on the real politics (e.g. Bueno de Mesquita, et al.
1999, 2000)

49

should provide theoretical explanation and justification regarding the relationship
between multiple units from multiple levels in data. Despite the growing popularity of
using directed dyad data in IR studies, scholars in this area have paid little attention to the
mixture of units of analysis and its theoretical problem. Statistically, we have to deal
with two types of unexplained variance (error terms) originated from multiple levels of
analysis. Thus, the unobserved heterogeneity inherent in most of the models in the liberal
peace literature (Green, Kim, and Yoon 2001; King 2001) should be explained at the two
levels: the dyadic level and the state level.

Finally, the theoretical reasoning of the liberal peace is primarily based on
monadic factors. That is, the level of liberty within an individual state is the primary
interest in the liberal peace theory. Yet, the actors in a dyad are engaged in a strategic
interaction. Such strategic interactions go between two states or between political leaders
in two states, while it occurs between a political leader and the citizens in a monadic
approach. If the theoretical emphasis is given to monadic factors, its statistical test should
be performed in a theoretically consistent way by distinguishing monadic effects from
dyadic effects on decision-making.30 This final criticism raises an issue about hierarchical
data structure. The repeated observations at the dyadic level are nested within the state
level. In other words, if a state’s behavior is explained or determined by domestic
political factors, repeated measures at the dyadic level should be grouped within the state.
To sort out monadic effects from dyadic effects, this paper suggests using a multilevel

(hierarchical) analysis. Through this analysis, the hierarchical structure of data can

 

3° In this regard, Gleditsch and Hegre (1997) raise a similar issue about three levels of analysis. They
suggest that peace and regime type can be examined at the dyadic, nation, and system level.

50

properly be defined, and the constraining effect of domestic institutions on the dyadic
relationship can thoroughly be explored.

In the followings, first, I discuss multilevel data structure in studies of
international conﬂict. The theoretical and statistical necessity of distinguishing monadic
effects from dyadic effects in the strategic approach is argued. Second, based on a brief
review of the liberal peace argument and its development, a strategic speculation of the
liberal peace is introduced. Third, a series of statistical tests on the relationship between
domestic institutions and interstate conﬂict are conducted. To capture the constraining
monadic effects of domestic politics on international politics, a multilevel analysis is

performed.

Multilevel Data Structure in International Relations

Monadic Eﬂects vs. Dyadic Eﬁects

Including the onset of militarized interstate disputes, most international political
events have often been understood as outcomes caused mainly by the interaction between
pairs of dyadic states with the structural approach. For this reason, dyadic analyses with
dyad-year data, which emphasize the importance of dyad-centric factors, have recently
prevailed in international relations studies. In a dyadic analysis, however, we are likely to
face the unobserved heterogeneity problem or ‘misestimated precision’ (Raudenbush and
Bryk 2002). The unobserved heterogeneity problem or misestimated precision occurs
when we miss theoretically important dyadic characteristics in explaining outcomes. As

well known, unobserved heterogeneity causes biased estimates, incorrect standard errors,

51

and the problem of inconsistency.31 However, this problem also occurs at the State level.
Since the repeated measures convey information about a state’s behavioral trajectory over
time, grouping observations within a state is necessary. Speciﬁcally, we notice the
necessity of grouping observations within a state in several ways.

First, each state has its own geopolitical conditions. Since geopolitical conditions
often limit or direct a state’s interest and behavior, affecting decision—making, a state’s
decisions made at various time points can be explained by constant geopolitical factors.
Moreover, geopolitical conditions are often a source of conﬂict (Diehl 1991). Second,
each state has its own cultural background and military history, which have shaped the
preferences of its leaders. These factors are not easily changeable but gradually
cumulative. For instance, a political leader or the people in a state may be war-prone,
peace-prone, or neutral, regardless of the relationship with any dyadic opponent. A state’s
risk-relational type in many cases persists over time and often provides powerful
explanation on political decisions such as militarized disputes. Finally, reﬂecting on
history, we observe that a set of states frequently have engaged in militarized disputes
compared to other states. Whatever the reasons could be”, these states reveal the fact that
individual states’ (not dyadic) characteristics do matter with respect to the likelihood of
interstate disputes. It is not simply a higher level of economic interdependence, or
“longer histories of economic interdependence” between dyadic states, that mitigates

conﬂict (Box-Steffensmeier, Reiter, and Zorn 2003). The likelihood of militarized

 

3’ See Greene for details (2000: 590-651).

32 Geopolitical explanations such as Mackinder’s “heartland-rimland” analysis or Ratzel’s
“Anthropogeographie” have appealed in this ﬁeld. AS well known, geopolitics, which a group of German
geographers developed, was used to justify Nazi expansion (Dougherty and Pfaltzgraff 1997). Explanation
based upon economic factors such as imperial wars by major states against considerably weaker states is
another rationale.

52

1"

disputes is also dependent on individual states’ characteristics or domestic institutional
context.

The strategic approach to international conﬂict summarizes nicely the necessity of
distinguishing monadic effects from dyadic effects. In models of strategic politics,
political decisions on the initiation of (or participation in) interstate conﬂict are not
determined only by international structural factors such as the distribution of power,
alliances, contiguity, and economic interdependence between dyadic states. These
decisions, however, can also be inﬂuenced by the domestic institutional context. For
example, when political leaders have a strong incentive to divert people’s attention from
domestic problems to international issues, they can initiate interstate conﬂict strategically
to rally domestic support. Since strategic theories place the theoretical importance on the
interaction between domestic politics and international politics, models of strategic
politics must address the multi-level structure in an analysis, clearly thereby providing
crucial answers regarding the ontological relationship between domestic politics and
international politics. Statistically, the failure to account for multilevel data structures
may result in inconsistent parameter estimates and inefﬁcient variance of the estimated
coefficients (Beck and Katz 2001b; Rodriguez and Goldman 1995).

In a dyadic analysis, the problem of unobserved heterogeneity basically centers on
dyadic heterogeneity. In a dyadic analysis with individual state level accounts, on the
other hand, the unobserved heterogeneity problem can also occur at the individual state
level as well as at the dyadic level. By disregarding heterogeneity existing at the state
level, we face the same statistical problems that we might have at the dyadic level. A

theoretical explanation on the outcome of interest will be incomplete. As a result, we

53

have to deal with two types of unexplained variations originated from both levels of
analysis. If multiple units from multiple levels are introduced into an analysis, we should
provide a theoretical explanation on the relationship between multiple units to maintain
logical consistency. This type of rigorous specification, measurement, and subsequent
empirical testing is demanding, but necessary.

Why a Multilevel Model” Analysis?

Various remedies for unobserved heterogeneity have been suggested. To
determine a proper way of dealing with the unobserved heterogeneity problem in TSCS
data with binary response variables (BTSCS) (which is a typical data format in IR
studies), continuous scholarly effort has been made to testing validity of fixed effects
models (Green et al. 2001; Beck and Katz 2001b; King 2001), random effects (frailty)
models (Box-Steffensmeier et al. 2003), or the between-and within-cluster estimation
(Zorn 2001). Simply, ‘fixed effects’ models assume that each dyad differs in its intercept,
while each dyad differs in its error term in random effects models. The between- and
within-cluster estimation is for capturing both cross-sectional and temporal differences in
dyads. Despite the previous efforts having made signiﬁcant contributions to
methodological developments in the arena of international politics, much attention needs
to be paid to two types of heterogeneity in multilevel data and also to a theoretical
explanation of the origin of heterogeneity.

Multilevel hierarchical models (or random coefficient models) have rapidly
become established as the appropriate tool for modeling data with complex hierarchical

structures (Beck 2001; Beck and Katz 2001a; Congdon 2003; Martin 2001, Steenbergen

 

33 Multilevel models (Goldstein 1995; Hox 1995) are also called as hierarchical models (Raudenbush and
Bryk 2002; Western 1998) or random coefﬁcient models (Beck 2001).

54

and Jones 2002). Allowing the intercept and parameter coefﬁcients to be random,
multilevel hierarchical models can monitor how domestic institutional contexts affect
international political outcomes. This way of modeling also has advantages in providing
“more accurate forecasts and estimates, providing a more realistic account of all the
sources of uncertainty associated with the comparative research design” than other
models such as the interaction model (Western 1998, p.1238).

Although the heterogeneity problem has been viewed as a methodological issue,
the causes of heterogeneity are of substantive interest. That is, if a theory tells that
political leaders’ decisions are systematically affected by certain domestic institutional
conditions (Putnam 1988; Morrow 1991; Fearon 1994; Mo 1994, 1995), the constraining
effect of domestic institutions on the interaction between dyadic states should be
explored. In this case, ﬁrst of all, we need to clarify the theoretical relationship between
dyadic units and state units. In the strategic approach, domestic institutional context often
explains internal effects on policy decisions, while dyadic relational factors create
external structural conditions that affect decision-making. In this regard, by employing
multilevel hierarchical structure into the model, we are able to successfully separate
domestic political factors from dyad-centric factors, clarifying the relationship between
the two levels.

Despite the importance of illustrating causes of heterogeneity at the state level, it
has been a hard task for international relations scholars to properly define and empirically
test the effects of domestic institutions on international political events. Since each state
has its own domestic context, it seems difficult to explain domestic variations using a

single measure. However, since a multilevel model does not require holding the ‘zero

55

error term’ assumption, we do not have to assume that a single predictor explains all
sources of heterogeneity existing at the state level (Steenbergen and Jones 2002; Western
1998).

Practically, to trace monadic effects on dyadic behavior, we need to explore
information at the state level. When the sample size within a state is large enough to
bring us precise estimates, grouping is possible and necessary. Given a relatively long
time span in most of IR studies, for instance from the year 1816 to the year 2000 in this
paper, we observe that most of states have survived through this time span. Finally, with
theoretical rationales and empirical opportunity, IR scholars can sort out monadic effects

on international political events from dyad-centric factors in multilevel analyses.

The Liberal Peace and Heterogeneous Domestic Politics

This research reevaluates liberal peace arguments, which posit that economic
interdependence has a pacifying effect on militarized conﬂict. The logic of the liberal
peace is simply that increased costs of potentially lost trade diminish the probability of
conﬂict. Since trading states recognize that militarized disputes might bring serious
consequences to their economic relations, they become reluctant to get involved in
disputes against trading partners as the level of trade goes up. In the liberal peace studies,
therefore, a high level of economically important trade, as usually indicated by the
bilateral trade-to-GDP ratio, is associated with a low level of incidences of rrrilitarized
interstate disputes and war, even after controlling for theoretically interesting other
inﬂuences: geographic distance, the balance of power, alliance, and the existence of

major powers.

56

The liberal peace maintains a theoretical assumption that policy costs caused by
lost trade are significantly associated with decision makers’ political costs. Relaxing this
assumption, recent studies have focused on domestic politics, based on the idea that
domestic political factors condition the pacifying effect of economic interdependence on
conﬂict (Hegre 2000). For instance, indicating that the interdependence literature ignores
state-market relations, Garztke and Li analyze how “costly signaling through global
markets facilitates ex ante bargains by allowing observers to infer the value of strategic
variables” (Garztke and Li 2001, p.7).

The strategic approach views that political leaders’ foreign policy decision
making can vary in response to domestic political situation. If political leaders enjoy a
relatively low level of autonomy domestically, they are more likely to be concerned about
policy outcomes and the likelihood (opportunity) of policy failure. On the other hand, if
political leaders are relatively free from domestic constraints, they are less likely to
concern about policy failure and thus more likely to stick to their preferences
(willingness). This perspective is based on the assumption that policy decisions are
mainly determined by political leaders’ drive to survive. In other words, political leaders’
behavior is basically analyzed in relation to political leaders’ political survival. In order
to survive and enjoy the beneﬁts of ofﬁce, political leaders should carefully take
domestic interests as well as the international political reality into account, especially
when political leaders do not have strong political support at home. Accordingly,
considering political risk is essential to understanding political decisions (Lambom 1991;
Baldwin 1995; Bueno de Mesquita et al. 1999). Such a theoretical innovation makes it

possible to integrate domestic and international “games” (Putnam 1988) played by

57

various types of political leaders. Overall, this perspective improves scholarly knowledge
on the dynamic relationship between domestic politics and interstate conﬂict, providing a
more comprehensive understanding on political leaders’ strategic decision-making
(Reiter 2003).

Policy Costs and Political Costs

The liberal peace advocates contend that, Since trading states recognize that
militarized conﬂict might bring serious economic losses, they are more likely to be
reluctant to engage in interstate disputes as the level of trade goes up. This argument is
held on the basic assumption that policy costs are directly associated with political costs34
for political leaders. In other words, it assumes that political leaders lose (or beneﬁt)
equally from adversity (or prosperity). However, if this assumption cannot be held, that is,
if policy failure does not significantly risk political leaders’ survival”, we should explore
the structure of domestic politics. Domestic politics is an important factor in calculating
the political costs of potentially lost trade. By ignoring the role of domestic politics and
political costs in evaluating the liberal peace, we also fail to explain the behavior of
‘liberal autocracies’ and ‘illiberal democracies.’ Liberal autocracies may frequently
engage in interstate conﬂict, while illiberal democracies may keep a peaceful relationship
with other states. As a result, the pacifying effect of economic interdependence on
interstate conﬂict should be examined along with domestic institutional context.

This research attempts to explain heterogeneity at the state level by testing a
proposition that the pacifying effect of trade on conﬂict is contingent on the domestic

institutional context. Depending on the level of personal stability in domestic politics,

 

3" Political costs are associated with political leader’s power position in a country, while policy costs with
the national wealth.
35 Political survival indicates that a political leader is not deposed from his ofﬁce in the year in question.

58

political leaders’ decision making on conﬂict can change. In turn, the consequences of
conﬂict can bring changes in political leadership and their political survival. This implies
that the signiﬁcant economic loss resulted from interstate conﬂict (policy costs) does not
always entail political leaders’ political failure (political costs) in domestic politics.

Measuring Domestic Institutions

In the previous studies, the measure of domestic institutions is often another name
of the regime type, which is typically classiﬁed into two categories, democracies or
autocracies (non-democracies). Since this measure is time—constant for a relatively long
time period, it is hard to capture dynamic changes in the relationship between the state
and the civil society in a country over time. Especially, this measurement problem is
serious when we deal with autocracies.36 Even in democracies, since various domestic
sectors are acting in each country, we cannot sufficiently explain overall domestic
institutional contexts which are associated with political leaders’ political survival.
Undeveloped states seem even more problematic. Even more troubling is that data tends
to be systematically missing from such countries — the state is simply so underdeveloped
that it cannot gather statistical data.

To evaluate domestic institutional contexts, this study uses two concepts: the
winning coalition and the selectorate. A winning coalition is defined as: “those members
of the selectorate whose support is essential to keep the incumbent leadership in office.”
The selectorate is defined as: “all those people in a country who have an institutionally

granted right or norm that gives them a say in choosing the government” (Bueno de

 

3" It is hard to overcome this problem when we deal with autocracies (nondemocracies). Maybe, it is
because it is difﬁcult to get good information on autocracies, or because autocracies do not actually have
variations in their domestic institutional conditions. Due to this reason, many studies have often limited
their research objects within OECD countries or democratic countries.

59

Mesquita et al. 1999, p.148). The mechanism through which these political institutions
affect political leaders’ perception about their political survival is straightforward.

When the size of the winning coalition is large, political leaders are able to hold on
to political power in relying on policy success, rather than on ‘private goods allocations.’
In such cases, political leaders are more likely to perceive themselves to be constrained
by domestic interests and domestic institutions when they make decisions on

37 Therefore, political leaders are more likely to behave

international politics.
strategically. Their behavior can be strategic because political leaders’ desire for political
survival makes them place less weight on the importance of international structural
factors (opportunity) than they would otherwise do with respect to interstate conﬂict.
This is the case where high economic interdependence has a pacifying effect on interstate
conﬂict. Since policy costs are strongly associated with political costs, political leaders
should be very cautious in engaging in conﬂict against trading partners. As a result, as the
size of winning coalition increases, domestic institutional context is more likely to affect
decision-making on conﬂict.38 Alternatively, if the winning coalition is relatively small,
leaders are less likely to care for policy success because they can easily maintain support

from the winning coalition by supplying private goods (Bueno de Mesquita et al. 2000;

1999; Gelpi and Grieco 2001). There is less of an incentive for political leaders to behave

 

37 In general, there is a high level of political competition on the leadership in these countries with a large
winning coalition.

38 This paper assumes that, under strong domestic constraint (a relatively big size of winning coalition),
political leaders are less likely to decide to engage in interstate conﬂict. Some may view that strong
domestic constraint leads political leaders to interstate conﬂict in some cases, especially when a majority of
people in a country support armed conﬂict. Although the latter argument is useful in explaining the onset of
interstate conﬂict in some cases, it does not harm the theoretical maintenance of this paper. In this paper,
we focus on the indirect contextual effect of domestic politics on the probability of MIDS. That is, based on
the liberal peace arguments, we assume that citizens do not want to get disadvantages from the loss of trade.
Based on this assumption, this research argues that the pacifying effect of interdependence on conﬂict is
conditional on domestic institutional context. The more decision makers are constrained by domestic
politics, the more the pacifying effect of economic interdependence on conﬂict is intensiﬁed.

60

strategically for their political survival. They are less likely to be cautious in making
decisions. Political leaders have a bigger leeway for maintaining personal preferences. In
this case, domestic institutional context may not strongly affect decision-making on
conﬂict.

Meanwhile, holding the size of the winning coalitions constant, the size of the
selectorate is expected to have a negative effect on the chance that a defector in a winning
coalition will be accepted into the rival’s winning coalition. This means that members in
a wining coalition will be more loyal to their political leaders as the size of the selectorate
increases (Bueno de Mesquita et al. 2000). As a result, given a certain size of the winning
coalition, political leaders are less likely to perceive themselves to be constrained by
domestic political factors as the size of the selectorate increases.

Tsebelis’ work (2002) is related to this argument. He shows how the number of
“veto players” in a political system affect policy stability, impeding significant departures
from the status quo. According to his argument, we are able to analyze political systems
regardless of the level of their institutional complexity, observing the number of veto
players or their ideological distance from each other.

With respect to the effect of domestic politics on political leaders’ strategic
decisions, the diversionary theory provides a somewhat different perspective than the one
I presented above. The contribution of the diversionary theory to the studies of
international relations is signiﬁcant. It explains well the strategy that political leaders can
use international conﬂict to divert people’s attention from domestic matters to

international issues (Leeds and Davis 1997”; Smith 1996). According to the rationale of

 

39 They suggest that, in order to extemalize domestic discontent, states in economic crisis may initiate
militarized conﬂicts against states with high growth.

61

this theory, political leaders whose domestic political positions are not stable often have a
stronger incentive to be involved in interstate conﬂict than leaders with stable domestic
political positions. This perspective seems contradictory with the theoretical argument in
this research. But, a strategic view on the liberal peace in this study does not come into
conﬂict with the argument of the diversionary theory, because this study focuses only on
the constraining effect of domestic politics on the relationship between economic
interdependence and interstate conﬂict. In other words, it is not assumed that the issue of
political leaders’ political survival has a direct effect on the likelihood of interstate
conﬂict. 40 This paper examines why and how economic interdependence can have
heterogeneous effect on interstate conﬂict, and, as a conditional factor, political leaders’
strategic position in domestic politics is introduced in order to provide an explanation to
the puzzle. Due to the assumption that political survival is a political leaders’ main goal,
scholars of the diversionary theory cannot deny the existence of the constraining effect of
domestic politics on the relationship between economic interdependence and interstate
conﬂict. Under relatively strong domestic constraints, it is not an easy decision for
political leaders to cut off a strong economic tie with their trading partners by initiating
conﬂict against them. In a multilevel model structure, we can set up the relationship
between domestic politics and interstate conﬂict to be conditional and thus make a more

general statement on interstate conﬂict.

Propositions, Modeling, and Estimation

Propositions

 

4° In order to test the direct effect of domestic politics on interstate conﬂict, which is not the interest of this
paper, we need to control various speciﬁc domestic level variables.

62

The first proposition of this research tests the extent to which domestic politics
explains the onset of militarized interstate disputes. In other words, this proposition
examines if a multi-level hierarchical model is necessary in testing the liberal peace
arguments:

Proposition 2.1: The onset of militarized interstate disputes is signiﬁcantly

explained at the state level.
If domestic institutional context does matter, then the probability of militarized interstate
onset should partly be explained at the state level. Analysis of variance at each level in
the multilevel data is used as a preliminary step to test this proposition. If political leaders
are constrained by domestic institutions, they tend to be more careful with conﬂict
initiation or participation than they would otherwise be. Therefore, the next proposition
tests the strategic view of the liberal peace:

Proposition 2.2: As the size of the winning coalition increases, the pacifying
effect of economic interdependence on militarized dispute onset
will increase, holding all else constant.

If political leaders’ position in domestic politics is not stable, they should be concerned
about political costs from potentially lost trade. Accordingly, as the size of the winning
coalition increases, political leaders are more likely to be cautious of engaging in conﬂict
against strong economic partners. On the other hand, if political leaders perceive their
political positions to be very stable in regard to their political survival, they will behave
non-strategically. Political leaders are less likely to be cautious of having conﬂict against

trading partners, since to some extent they can endure the costs coming from lost trade.

63

Testing this proposition helps us understand why economic interdependence may have
heterogeneous effects on the onset of militarized disputes.

A Three-Level Hierarchical Model

In order to set up the relationship between trade and conﬂict in a three-level

framework, I use a “three-level hierarchical model.” In the ﬁrst level of the three-level
hierarchical model, we observe a set of binary outcomes Y,,-,- representing either the onset
of militarized interstate disputes or not:

Yu‘j “' Be mm”,- )r

n is the probability of the outcome. Of particular interest in this analysis are the

"'1'
modeling of heterogeneity at levels 2 and 3 (dyads and states), and resulting impacts on
ﬁxed regression coefficients. The responses are classified by state j, dyad i, and by dyadic
year t within dyads. The first level of this three level model with a logit link for the

binary outcome is speciﬁed in the following form:

. _ * * . .
(15,-,- * Distance,,-, + 016,-,- * Contiguity,,, + a7!) * Onepowern-j + 61/8,,- * Twopowers,,-,- +

:1:
619,-,- Peaceyear,

,, + 0:10,,- *Spline1 + a, 1,,- *Spline2,,-, + am,- *Spline3 + £,,-,. (2.1)

rij tij
By defining the first level intercept coefﬁcient as random, we can specify the second

level as:
aOij 3600} “I’ﬂorj , (2-2)
where A”, is the average intercept across the second level units (dyads), #0,, is the

unique increment to the intercept associated with the second level unit i. Since we are

mainly focusing on the effect of level-3 (states) on fixed coefficients, no predictors are

introduced in this model. We can also specify that the second level intercept vary

randomly:
where ym, is the average intercept across the third level units (states), v00, is the unique

increment to the intercept associated with the third level unit j. Equations (2.2) and (2.3)
allow us to estimate the variability in the regression coefﬁcients (intercepts) across both
level 2 and 3 units. The next step is to model this variability especially at the level-3.
Based on a strategic theory of the liberal peace, we can specify that the effect of
economic interdependence varies as a function of the sizes of the winning coalition and
the selectorate:

any =710j+711j*WCz0j+712j*5101+V10p (2.4)

where y,0,, 7”,, and ya, are level-3 coefficient (fixed effects), and V10) is a level-3 random

effect. Unlike in Equation (2.4), we don’t include the same level-3 predictors in Equation
(2.3). As illustrated above, this is because we assume that domestic institutions affect the
probability of nrilitarized interstate disputes through its relationship with economic
interdependence. Now, if we substitute Equation (2.2), (2.3), and (2.4) into Equation (2.1),
it yields:
logit(ﬂ,,ﬂ=i’qn +710) *Irtterzlqrarderroe,,, +711} *“(TOj *Wzy +712) *5101' ”"1an
+020 * Democracy,,-,+ ...... +€,,-,- + #0,,- + V00,- + V10,- * Interdependence",- . (2.5)

As a result, in Equation (2.5), the multilevel disturbance term is composed of the four

different components, 5",, p0,,,vm,, and 1110,.

For the a and yparameters, I assume a normal distribution with a ﬂat prior. And

65

1" .

for each level-2 and level-3 random effects, I assume for each a normal distribution and a

multivariate normal distribution: #0,- ~N(0,S2,,), vpo, ~MN(O,SZVP). The assumed prior for
level-l variance parameter is £,,-,- ~ N(O,1), while level-2 variance parameter is assumed to
have a Gamma distribution with a ﬂat prior: S2,”)~G(0(I)LO(I)1). For the level-3 variance
parameter, I assume a standard Wishart prior: 9;; ~W(&,,2). To identify the multilevel

model, we need to assume zero covariance between three level disturbances:

Comrqy]=Corivm,,e,,,]=(briv10,,e,,,]=CovLuﬂ,,I/m,]= Com,,v,0,]=0
All the priors are chosen not to signiﬁcantly dominate the iteration procedure. And all the
distributional assumptions are based on the information found out in many previous
studies (Congdon 2003).

This formulation has a three-level hierarchical structure. In this model framework,
we can separate domestic context (level-3) from dyadic factors (level-2). The second
level of this analysis relates dyadic relational factors to MIDS, and the third level
incorporates domestic context by explaining variation in the parameter estimates.

The ﬁrst step to test this argument is to check a multilevel structure in the data. A
one-way ANOVA (analysis of variance) with random effects can help to decompose the

variance in militarized interstate disputes:
Disputes“,- : yooo + ij + [110,-]- + 8“):
In this model, 7111) is the grand mean of militarized interstate disputes. v00,- contains cross-

state variation, while #0,,- contains cross-dyadic variation. Finally, a captures dyadic

tij
year temporal variation. In order to argue that all three levels of analysis are important for

militarized interstate disputes, all three-variation terms should be statistically signiﬁcant.

66

Estimation

Beck and Katz (2001a) report that parameters of multilevel models (random
coefficients models in their terminology) with TSCS data are poorly estimated in Stata or
Limdep statistical packages, which mainly use GLS algorithm for obtaining estimates.
Following their suggestion, I use a Bayesian simulation method to estimate the
parameters. Estimation and inference via Bayesian simulation method such as Markov
Chain Monte Carlo (MCMC) methods are advantageous in various aspects (Jackman
2001). First of all, instead of ﬁnding simple point estimates for the parameters of interest,
we can make a large number of simulated random draws from the joint posterior
distribution of all the parameters. And then we can use these random draws to form a
summary of the underlying distribution, which provides much reliable estimates. In
addition, when we have to deal with very complex models like multilevel models,
Bayesian simulation methods help to solve computational difficulties.

The parameters that I must estimate are as follows: 6 = {y,a,£2}. Given the data

set, Y = {Y, } , we can estimate the parameters from the posterior distribution:

i,
f(9lY)°<f(Yll9)f(6’) ,

where f (Y|6) is the likelihood. The likelihood function summarizes the information about

Bin Y. Thus, the function of His proportional to L(6| Y) oc f (YIB). f (6) is the prior, which

is constructed from the priors and the hyperparameters of the model. Finally, the
posterior density is proportional to the product of the prior distribution and the likelihood:
f(¢9lY)°c L(9|Y)f(0).
Due to the complex data structure in which the parameters y vary across domestic

contexts, it is very difficult to estimate this model using maximum likelihood estimation.

67

In order to estimate simultaneously all the parameters, I employ Markov Chain Monte

Carlo (MCMC) estimation methods. Speciﬁcally, the posterior distribution f (6|Y) is

investigated with the Metropolis Hastings sampler. Compared to the Gibbs sampling, the
Metropolis Hastings (MI-I) sampler is useful when the conditional posterior distributions
do not have simple forms. This is because the MH sampling method allows the proposal
distribution to have any form. Since, for binary response models, the conditional
posterior distributions for both ﬁxed effects and the residuals do not have standard forms,
the MH sampling is used for the parameters in this model. The Winbugs41 and MLwiN
(Rasbash et al. 2002) statistical packages are used for this purpose.

The binary dependent variable in this analysis has few events compared to the
large amount number of nonevents. Rare events can bring biased logit coefﬁcients in
small samples, causing underestimation of event probabilities (King and Zeng 2001). To
see if the data in this research reveals the ‘rare events’ problem, I employ a new, robust
estimation procedure that King and Zeng (2001) have developed.42 Comparison of the
estimates obtained through this suggested method with the estimates in the normal logit

model can clarify if the ‘rare events’ problem is serious in the data.

Operationalization

The dependent variable is militarized interstate disputes. This variable is coded as
one if a militarized interstate dispute occurred between the members of that dyad in that
year, and zero otherwise. There has been controversy about the way of measuring

rrrilitarized interstate disputes (MIDS) (Hegre 2000). Heightening the cutoff point will

 

4' This soft ware is available at <http://www.mrc-bsu.cam.ac.uk/bugs/winbugs/contents.shtml>.
’2 This estimation procedure can be done using the software “ReLogit”, which is publicly available at
<http://gking.harvard.edu>.

68

aggravate the ‘rare events’ problem. Adversely, lowering the threshold for inclusion can
yield more ﬂexibility in dealing with data. That is, it will help to reduce both the ‘rare
events’ problem and the problem of unreasonably combining data obtained from
relatively long periods of time (Gleditsch, Strand, Eriksson, Sollenberg, and Wallensteen
2001).

As a result, I use Oneal and Russet’s (2002) MIDS data set compiled by the
Correlates of War (COW) Project. A militarized interstate dispute involves the threat,
display or use of force short of war by one member state, explicitly directed towards the
government, ofﬁcial representatives, official forces, property or territory of another state.
They incorporate Zeev Maoz’s (1999) revised data for more accurate dyadic analyses.
The data set covers the time period of 1816-2000. Directed dyads data format will be
used. Including many dyads that have almost zero chance of being involved in militarized
disputes is unrealistic. Thus, this research uses “politically relevant dyads.”43 A pair of
dyadic states is defined as a politically relevant dyad if it includes at least one major
power or it is geographically contiguous. By limiting the pool of cases to relevant dyads,
the rare events problem is reduced. The measures of variables are described in Table 3.

This research corrects for the temporal dependence problem by creating a spline
function for peace years (Beck, Katz, and Tucker 1998). To measure level-2 predictors,
the winning coalition and the selectorate, ﬁrst, I use Bueno de Mesquita et al.’s

suggestion (1999). Their measures have recently been popular in many international

 

’3 To see if ﬁndings remain signiﬁcant only when we restrict our analysis to politically relevant dyads,
analyses with all dyads are also performed and reported in Appendix 2. And, there are no substantively
signiﬁcant differences in the outcomes between the two data sets.

69

Table 3. Variable Descriptions

 

 

Variable Description

MIDS The onset of militarized dispute (1: onset). The Correlates of War (COW)
MID data. Available in EuGene.

Dependence The ratio of dyadic trades to the GDP of each partner in constant US
dollars. The trade dependence value for the least commercially engaged
member of the dyad is used for the level of economic interdependence.
Available in EuGene.

Democracy The regime score for the less democratic dyad member. This score come
from the Polity III “institutionalized democracy” score minus the
“institutionalized autocracy” score for each state (J aggers & Gurr 1995).
Available in EuGene.

Alliance Whether or not the dyad members are allied (l=allied). Available in
EuGene.

Parity A continuous variable bounded between zero and one, indicating the ratio
of the weaker dyad member’s capabilities to those of the stronger dyad
member. Capabilities are based on the COW project’s Composite
Capabilities Index. Available in EuGene.

Contiguity It is coded as one when the dyad members are geographically contiguous
by a land border. Available in EuGene.

Distance The natural log of distances between capital cities. Available in EUGene.

Major Powers Two dichotomous variables, one equal to one if only one dyad member is a

Winning Coalition

Selectorate

Checks

 

major power, the other equal to one only if both dyad members are major
powers. Available in EUGene.

A composite of four items: (1) “chomp” (the competitiveness of
executive recruitment), (2) “Xropen” (the openness of executive
recruitment), (3) “Parcomp” (the competitiveness of political
participation), and (4) “Regtype” (regime type) in Polity II and Polity HI
data (Jaggers and Gurr 1995). Available in EUGene and ICPSR (Inter-
university Consortium for Political and Social Research) data archive in
the university of Michigan.

“Legislative Selection” variable in Polity II data is used. Available from
ICPSR data archive.

The number of veto players measured between 1975 and 2000. Available
in the Database of Political Institutions (DPI) 2000 (Beck et al. 2001).

 

70

relations studies. “Legislative Selection” variable in Polity H data set is used to measure
the size of the selectorate, which varies only from 1 to 3. This variable measures the
selective breath of the members of each country’s legislature. If there is no legislature, it
is coded as 0. If the legislature is chosen by heredity, by ascription, or by the effective
executive, it is coded as 1. If members of the legislature are selectedby popular election,
it is coded as 2. Therefore, it is believed that the larger the value of this variable, the
larger the size of the selectorate.

To measure the size of the winning coalition, we construct a composite index
based on the variables “chomp”, “Xropen”, “Parcomp”, and “Regtype” in Polity 111 data
set44 (Jaggers and Gurr 1995). chomp variable measures the competitiveness of
executive recruitment. The value of this variable is 1, 2, or 3. A coded value of 1
indicates that the chief executive is selected by heredity or in unopposed elections. If this
variable is larger than or equal to 2, one point is assigned to the winning coalition
variable. Xropen variable measures the openness of executive recruitment. If the
executive is not recruited by heredity, another point is given to the winning coalition
variable. Parcomp variable measures competitiveness of political participation. If
political participation is highly competitive, another point is added to the winning
coalition variable. Regtype variable measures the regime type. If a political system is not
a military or military/civilian regime, one more point is assigned to the winning coalition
variable. With the base value of one, finally, this index varies one to ﬁve. The data for

this variable is from Polity II data set.

 

4" “Regtype” variable existed in only Polity II data set and was not updated in Polity III and beyond. A
revised version of the data, Polity IIID, was released in 2000.

71

Due to the small range of variations, these measures may not fully explain
variations in domestic institutions. As an alternative way of measuring domestic context,
the “Checks” variable in the Database of Political Institutions (DPI) 2000 (Beck et al.
2001) is used for the winning coalition. This variable is originally designed to measure
the number of veto players across different political systems. By counting the number of
veto players in a political system, we may be able to have a ﬂexible way of measuring the
effect of domestic institutions on policy outcomes (Tsebelis 2002). The Checks variable
is coded as 1 if legislatures are not competitively elected. This variable is incremented by
one in each following case: if there is a chief executive, if the chief executive is
competitively elected, or if the opposition controls the legislature. In addition, in
presidential systems, this variable is incremented by one both “for each chamber of the
legislature unless the president’s party has a majority in the lower house and a closed list
system is in effect” and “for each party coded as allied with president’s party and which
has an ideological orientation closer to that of the main opposition party than to that of
the president’s party.” In parliamentary systems, the Checks variable is incremented by
one both “for every party in the government coalition as long as the parties are needed to
maintain a majority” and “for every party in the government coalition that has a position
on economic issues closer to the largest opposition party than to the party of the
executive” (Keefer 2002). Finally, an increase of the number of veto players in a political
system indicates that political leaders are more likely to be checked by their opponents or
by the people. Therefore, this variable goes well with the concept of the winning coalition

as a proxy variable. Compared to Bueno de Mesquita et al.’s composition index (1999;

72

2000), the Checks variable that varies from one to eighteen is believed to efficiently

capture changes in domestic context45 (See Appendix A for details).

Result

The first step in multilevel models is to check if there is signiﬁcant variation in
the probability of militarized dispute occurrence at both state and dyad levels. For this
purpose, Table 4 reports the result in the one-way ANOVA model. This test provides us
the evidence that the multilevel character of the data should not be ignored. The ratio of
each variance component to the total variance in the onset of rrrilitarized disputes can be
used to indicate the importance of each level in the model (Rasbash et al. 2002;
Steenbergen and Jones 2002). Since this comparison is difﬁcult to carry out for a binary
model, I fit the model as a Normal response model as a crude way of performing this test
as recommended by Rasbash et al. (2002). Although we should not rely completely on
this test result in evaluating the necessity of introducing a multilevel model, it can be a
preliminary step to begin a multilevel analysis. To see if the applicability of multilevel
models is limited only to directed dyad year format of data, I also report the result from
the non-directed dyad year data.

All of the variance components are statistically significant in each multilevel
model. In the two-level model with non-directed dyad year data, the dyadic level explains
about 17% of the total variance in the onset of interstate conﬂict, while the dyadic year

level explains about 83% of the total variance. The two-level model with directed dyad

 

’5 One of the examples, which show the goodness of this measure over Bueno de Mesquita et al.’s
composition index, is Argentina. Between 1983 and 2000, the value of Bueno de Mesquita et al.’s
composition index for Argentina is ﬁxed at 3. On the other hand, DPI index indicates that the size of
winning coalition has frequently ﬂuctuated between the value of l and 5 during the same time period,
which is much closer to the trends in the history of Argentina.

73

Table 4. Analysis of variance with Random Effects

 

 

 

 

Parameter Non-Directed Dyad Year Directed-Dyad Year
Fixed Effects 0.004" 0.022“ 0.031" 0.028"
Constant (0.000) (0.001) (0.003) (0.003)
Variance
Components
State Level 0.002** 9% 0.001** 4.4%
(0.000) (0.000)
Dyad Level 0.001“ 16.7% 0.004” 17.4% 0.003“ 13.0%
(0.000) (0.000) (0.000)
Dyadic Year 0.005“ 83.3% 0.019** 82.6% 0.022" 91% 0.019“ 82.6%
Level (0.000) (0.000) (0.000) (0.000)
** = p < .01

Note: Estimated standard errors are in parentheses. The percentage of each level variance
of the onset of militarized interstate conﬂict is reported. The ratio of each variance
component to the total variance in the onset of conﬂict can be used to indicate the
importance of each level in the model. Since this comparison is difficult to be done for a
binary model, I fitted the model as a Normal response model as a crude way of
performing this test. The result suggests that by ignoring the multilevel character of the
onset of conﬂict, we will miss an important source of variance at the dyadic level of the

state level to some degree.

74

year data reports similar results. Meanwhile, when we compare variance at the dyadic
year level with the one at the state level, about 9% of the total variance is resulted from
the state level. Finally, in the three-level model, variance components to the total variance
take 82.6%, 13%, and 4.4% at each dyadic year, dyad, and state level. Overall, these
results imply that there are signiﬁcant variances in the onset of conﬂict at both dyadic
and state levels as well as at the dyadic year level. The result demonstrates that, by
ignoring the multilevel character of the onset of militarized conﬂict, we will miss
important sources of information on the total variance and make incomplete conclusions
about the onset of interstate conﬂict. It underscores the validity of ﬁtting this model as a
multilevel model.

In Table 5, estimates from four different models are reported. Using a MH
sampler, I generate a large number of samples from the joint posterior density of the
parameters and obtain the summary statistics used for inference in multilevel models.46 In
the logit with splines model, the corrected logistic estimates are obtained by controlling
the possible rare events problem. The normal logit estimates and corrected logit estimates
are almost identical in this analysis, implying that there might not be the ‘rare events’
problem in the data.

To check if parameter estimates converge to the posterior distribution, I use
various diagnostic methods. Graphical inspection of MCMC sequences is important in
assessing problems with convergence. According to trace plots of the iterative history of
MCMC sequences, it is hard to say that there is a problem with convergence (See

Appendix C). As an alternative, I use the Raftery-Lewis diagnostic (Raftery and Lewis

 

’6 I let the sampler run for more than 100,000 iterations, and saved the last 40,000 (or 60,000) iterations for
inference.

75

Table 5. The Onset of Militarized Interstate Disputes (Politically relevant dyads)

 

Variable

Democracy
Interdependence
Parity

Alliance

Distance
Contiguity
One Power
Two Powers

1nterdep*WC

Interdep*S
Interdep*Checks
Constant

Variance
Components
Sate Level
Constant
Interdependence
Covariance

Dyad Level
Dyadic Year Level

Deviance (—2LL)
DIC

Logit with splines

modela
-0.046** (0.004)

—19.591**(5.951)
1.0116**(0.092)
—0. 142** (0.059)

—0.265** (0.026)
0.468" (0.098)
—0.012 (0.101)

0.972** (0.123)

0.101 (0.210)

14045.028

ANCOVA with
Random Effiects

Model
— 0.030M (0.007)

—l6. 140**(7.501)

1.049**(0.180)
— 0.303** (0.086)

— 0.538“ (0.063)
0.402" (0.193)
— 0.377 (0.208)

0.231 (0.251)

0.893 (0.490)

0.930** (0.199)

1.238“ (0.141)
b0.996"""(0.004)

10829.40
11393.10

The Slopes-as-
outcomes
Modell
—0.025** (0.007)

89.73** (34.10)

1.056“ (0.181)
—0.279** (0.087)

-0.542** (0.064)
0.400** (0.206)
-0.383 (0.208)
0.210 (0.252)

—27.19**(11.16)

—4.122 (14.44)

0.924 (0.535)

0.935 ** (0.199)
884.61 (1227.5)
4.129 (12.881)

1.238" (0.140)
b0.923"""(0.000)

10811.01
11380.66

The Slopes-as-
outcomes
Mode12
—0.030** (0.012)

—44.54c (35.62)

1.372**(0.348)
—0.179 (0.181)

—0.505*"‘(0.l 10)
1.855” (0.433)
0.302 (0.472)

2.200**(0.710)

—9.058C (8.564)
—-2.432** (0.872)

1.013** (0.315)
4522.2 (3358.1)
2.894 (24.301)

2.683**(0.4l7)
b0.876"""(0.007)

3786.99
4150.70

 

Note: Multilevel table entries are Bayesian estimates obtained from the Metropolis-Hastings (MH)
sampling methods. Standard errors are in parentheses. Spline variables are not reported to save space.

(N=60,961).

*=p<001**=p<001

a Logit estimates are corrected for controlling the rare events problem. Despite the concern of rare
events problem, it seems that it is not the case in this analysis. The normal logit estimates and corrected
logit estimates (King and Zeng 2001) are almost identical.

By a Binomial distributional assumption, the random coefﬁcient at the ﬁrst level was deﬁned with a
zero mean and a variance which was constrained to equal one. However, to test whether or not a
Binomial assumption holds or whether the model exhibits ‘extra Binomial’ variation, I obtained a
dyadic level variance based on the extra Binomial distributional assumption.
c. LR test for joint Signiﬁcance of “Interdependence” and its interaction term is statistically signiﬁcant.

76

1992). This diagnostic tests for convergence to the stationary distribution and estimates
the length of Markov chain needed to accurately estimate quantiles of interest. The test
results show that the parameter estimates are significant with 95% of Monte Carlo
probability.

The logit with spline variable model in the first column supports the previous
liberal peace argument (Oneal and Russett 1999, 2002). Strong economic
interdependence mitigates the probability of militarized dispute onset. To see if the effect
of economic interdependence on MIDS is conditional on domestic politics, I obtained
three different multilevel models.

In the ANCOVA with random effects model in the second column, we let the
intercept vary across both dyads and states. In this model, the dyadic and the state effects
are conceived as random effects. Accordingly, our main interest is given to error variance
terms both at the dyad and the state levels. The two error variance terms at both the state
level and the dyad level are all statistically signiﬁcant, which suggests that we need to
incorporate both dyadic and state level predictors into the models in order to explain
between-variations in both dyads and states. Another important thing to be noticed in the
random coefﬁcients model is the standard error of the interdependence variable.
Compared to the one in the standard logit model, the standard error in the ANCOVA
model, as well as in the two other multilevel models increased. This is because we
grouped total observations within states.

In the two slopes-as-outcome models, we include state level predictor(s) into the
model so that we can explain the variability of the slope of interdependence variable. The

state level error variance term, which comes from the variability of the slope of

77

interdependence variable, and also its covariance term with the state level intercept error
variance, are statistically insignificant. It implies that we might not need further state
level variables to explain the slope variability. A hypothesis test for the joint signiﬁcance
of these two terms, the Wald test, also confirms this fact. The Wald test in the ﬁrst slope-
as-outcome model yields a chi-squared value of 0.554 on two degrees of freedom,
providing little evidence for an effect at the 5% signiﬁcance level. The Wald test in the
second slope-as-outcome model reports the same result. Meanwhile, as can be expected,
the dyadic level error variance and the state level intercept error variance are still
statistically significant. This is because we did not incorporate any predictors for the state
and dyad level intercepts.

In these models, the role of newly introduced state level variables is important.
First of all, we need to check if the introduction of state level variables into the models
have an impact on the interpretation of the interdependence variable. Statistically, the
effect of the interdependence variable should jointly be evaluated with its interaction
term(s). Although the interdependence variable may not be statistically insigniﬁcant (in
the last model), we should not look at separate t statistics of interdependence variable and
its interaction term(s) (Wooldridge 2000). The p-values for the Likelihood Ratio (LR)
test of the joint significance in the two slopes-as-outcomes models are almost zero, which
means that interdependence variable and its interaction term(s) are jointly signiﬁcant at
the conventional significance level. 47 Therefore, we cannot say that economic

interdependence does not affect the onset of militarized disputes, although both

 

47 The LR test is based on the difference in the log-likelihood functions for the unrestricted and the
restricted models. In the restricted model, nondemocracy variable and its interaction term are dropped. The

likelihood ratio statistic is twice the difference in the log-likelihoods: LR = 2[ln(Lu,) — ln(L,)]. The LR
statistic has an approximate chi-square distribution.

78

Interdependence variable and its interaction term with Checks variable in the second
slope-as-outcomes model are statistically insignificant.

The interaction term between interdependence and the winning coalition variables
in the third model is statistically signiﬁcant, which means that the pacifying effect of
economic interdependence on MIDS varies between states depending on the size of
winning coalition. As the size of the winning coalition increases, economic
interdependence is more likely to mitigate the probability of militarized dispute onset.
Interestingly, the effect of interdependence variable is not monotonic anymore. When the
size of the winning coalition is relatively small, which means that a political leader is
weakly inﬂuenced by domestic politics, there is no pacifying effect of trade on
rrrilitarized disputes. However, as the size of winning coalition is bigger, which means
that a political leader is more likely to be concerned with policy outcomes and political
costs, a pair of dyadic states with strong economic ties is less likely to involve in
militarized disputes. Figure 5 well illustrates this point. In Figure 5, we observe ﬁve
different lines, where each line indicates the effect of economic interdependence on the
probability of MIDS at a different level of the winning coalition variable. From the
highest line to the lowest one, the size of the winning coalition increases. For this ﬁgure,
all other control variables are ﬁxed at their means or modes (for dummy variables), while
the selectorate variable is ﬁxed at three (its maximum value).

As Figure 5 illustrates, the economic interdependence variable mitigates the
probability of militarized dispute onset only when the size of the winning coalition is big
enough (more than 3 in the scale of the variable). It implies that if political leaders do not

have to concern about political costs from lost trade, a strong economic tie with a

79

 

Figure 5. The conditional effect of interdependence on
MIDs

DIS-r

0.23» /
020* /

018~~ /’
015—
013~
010«~

008«

Probability of MIDS

005-

 

003-

 

 

 

 

 

 

 

0.00 i g :
0,00 0.01 0.02 0.03 0.04 005

Economic Interdependence

 

 

 

* These ﬁve lines show the changes of probability of militarized dispute onset according
to the changes of the size of winning coalition. From top to bottom line, the size of
winning coalition varies from one to ﬁve. All the control variables are ﬁxed at their
means or modes (for dummy variables). This ﬁgure indicates that the effect of economic
interdependence on the probability of MIDS varies across states depending upon domestic

institutional context.

8O

potential opponent does not decrease the probability of militarized dispute onset. In the
meantime, the interaction term between interdependence and the selectorate variables has
a negative sign. Although it is different from our theoretical expectation, the interaction
term is not statistically signiﬁcant.

Figure 6 also shows that decisions on international conﬂict are inﬂuenced by
domestic institutional context. Using Checks variable, which measures the number of
veto players in a state, this figure48 monitors if the probability of militarized dispute onset
significantly changes as the number of veto players increases. As expected, the pacifying
effect of trade on MIDS becomes more signiﬁcant as the value of Checks variable
increases. Checks variable is statistically significant in the model with all dyads. All
substantive implications in the multilevel models in Table 5 do not significantly change
in the multilevel models in Table 8 in Appendix B, where all dyads are used.

All four models report similar results for the control variables. Democratic dyads
tend not to ﬁght with each other. Power parity leads to interstate disputes. Alliance
mitigates the onset of conﬂict. The longer the distance between dyadic states, the less
they are involved in conﬂict. Contiguity is another important factor of contributing to the
onset of militarized interstate disputes. If dyadic members are both major powers, they
are more likely to have disputes. If only one dyadic member in a dyad is a major power,
the dyadic states are less likely to engage in interstate disputes, although this last variable
is not statistically signiﬁcant in any model.

Table 5 also reports deviance and the Deviance Information Criterion (DIC) for

each multilevel model as a way of model assessment (Spiegelhalter, Best, Carlin, and van

 

’8 This ﬁgure is obtained from the model with all dyads. Since Checks variable is not statistically
signiﬁcant in the model with politically relevant dyads, the constraining effect of Checks variable is tested
in the model with all dyads.

81

Probability of MIDs

Figure 6. The conditional effect of interdependence on MIDs

 

0.07

0.06 -

 

o'w TM-

 

 

 

 

 

 

o.04~ \_\
\\\ —— Checks=1
o.w_ \. ....... W3
\ \\ —— Checks=5
-. — Checks=7
0.024 \ \ \\ — W
.‘ \
\\_. \\\ ..............
0.01" \\-.\ \\\ .
0.“) r I \f ~——~|qi'__—_ulnm=-l
0.00 0.01 0.02 0.03 0.04 0.05

Econom'c Interdependence

* These five lines show the changes of probability of militarized dispute onset according

to the changes in “Checks” variable (the number of veto players). From top to bottom

line, Checks variable varies from one to 9. Since few observations in this variable take

the value 10 or more, these cases are not included in this figure. All the control variables

are ﬁxed at their means or modes (for dummy variables) except Contiguity variable

(ﬁxed at one). This ﬁgure indicates that the pacifying effect of economic interdependence

on the probability of MIDs is signiﬁcantly intensified as the number of veto players

increases in domestic politics. All dyads are used for this figure.

82

der Linde 2002). The DIC is a generalization of the Akaike Information Criterion (AIC).
The DIC diagnostic can be used to compare models, as it consists of the sum of two
terms: the measure of the ‘ﬁt’ and the measure of ‘complexity’ (taking account of the
number of parameters) of a particular model. Since a smaller value of the DIC indicates a
better model, the multilevel model in the third column seems slightly better than the

ANCOVA model with random effects model.

Conclusion

The theoretical reasoning of the liberal peace is based on an assumption that lost
trade always disadvantages political leaders. Relaxing the homogeneity assumption, this
paper attempts to connect domestic politics to the analysis of international politics. This
paper also addresses the question of whether political leaders’ strategic position in
domestic politics inﬂuences their decisions on whether to engage in militarized conﬂict.
The results of this analysis suggest that, depending on domestic institutional context and
the expected size of political costs, political leaders take different courses of action even
under the same dyadic relational conditions. That is, as the size of the winning coalition
increases, economic interdependence is more likely to mitigate the probability of
militarized dispute onset. Thus, political leaders’ strategic consideration on their political
survival may explain why economic interdependence has heterogeneous effect on
interstate conﬂict.

Interstate conﬂict is not solely explained by dyadic effects, but also by monadic
effects. In sorting out monadic effects from dyadic effects, we should be cautious of two

types of heterogeneity originated from both dyad and state levels in directed dyad (year)

83

data. In this case, dyadic observations should be grouped within states. By failing to
properly specify multilevel data structure in an analysis, we are likely to have inefﬁcient
and inconsistent estimates and thus to wrongly reject the null hypothesis. A response of
this paper to the existence of two types of heterogeneity and to the limitation in obtaining
functionally equivalent data at the state level is to use a multilevel model. Statistically, a
multilevel model is better than other models in that it does not require a strong
assumption that a single measure explains all existing heterogeneity at the group (the
state or dyads) level. Theoretically, a multilevel model can clarify the ontological
relationship between multiple units (actors) from multiple levels of analysis. Testing the
liberal peace using a multi-level hierarchical model with a Bayesian simulation method,
this paper shows how domestic institutions affect dyadic behavior and how the causal
heterogeneity can be explained without a strong statistical assumption.

Meanwhile, this research has some limitations in applying a strategic theory to the
liberal peace. The measures of domestic political institutions used in this paper are not
perfect. To investigate the linkage between domestic politics and international politics,
we need to develop a well-reﬁned measure applicable to various political systems for
comparative purpose. Decisions about interstate conﬂict can also be explained
(constrained) by international systemic factors such as the power of a hegemonic state or
prevailing international norms, not only by dyadic relational factors. Investigating the

possibility of much complex modeling will be a further step for future research.

84

CHAPTER III

Globalization, Inequality Within Countries,

and Interstate Conﬂict

85

Abstract

This study ascertains the extent to which inequality within countries is relevant
factor to political leaders faced with the decision to use force. It tests a diversionary
theory of the use of force that treats inequality within countries as an endogenous
variable. As important covariates for inequality within countries, political leaders’
winning coalition and trade dependence are introduced. This design allows the author to
assess the direct and indirect effects of trade on the probability of using force, supporting
for a revised diversionary theory. Finally, this study demonstrates that political leaders
are more likely to engage in international conﬂict as a way of diverting attention from
domestic problems, as the level of inequality within countries grows, and that trade
indirectly increases the probability of militarized interstate dispute onset by increasing

the level of inequality within countries.

86

Introduction

The world economy has continued to be integrated over the last two centuries.
The ongoing economic globalization in the 21St century has made scholars revisit a
traditional topic in international relations studies, interstate conﬂict, from an economic
point of view. A political economy of international relations emphasizes the linkage
between the economy and uses of force. A host of scholars have investigated the
pacifying effect of economic interdependence on interstate conﬂict (e.g. see Barbieri and
Schneider 1999; Beck, Katz, and Tucker 1998; Maoz and Russett 1993; Oneal and Russet
1999, 2002). The logic of the liberal peace is simply that increased costs of potentially
lost trade diminish the probability of interstate conﬂict. Since trading states recognize
that militarized disputes might bring serious consequences to their economic relations,
they become reluctant to engage in militarized disputes with their trade partners, as the
level of trade goes up. The basic assumption in this argument is that economic
liberalization beneﬁts trade partners, providing ‘absolute gains’49 to all trading states.
Therefore, relative gains or inequality between countries is not a big issue in the liberal
theories.

The effect of economic liberalization on world inequality and its political
implication should be viewed in two ways: inequality between countries and inequality
within countries.50 Although there is a controversy about an appropriate empirical method

to test the relationship between inequality and political violence, many studies report that

 

’9 The meaning of ‘absolute gains’, a contrary concept to ‘relative gains’, used in international relations
studies is different from the concept of ‘absolute advantage’, a contrary concept to ‘comparative
advantage’, used in international trade theories. ‘Absolute gains’ indicate the net gains that a trading partner
can get from trade, while ‘absolute advantage’ indicates the economic status where a country can produce a
goood or service using a smaller amount of resources than the one its trade partner uses.

For example, Bourguignon and Morrisson (2002) show the trend of global inequality in terms of
inequality between and within countries. Refer to Figure 7 and 8 in Appendix D for a graphical description
of their report.

87

economic inequality within countries induces political violence such as revolution,
rebellion, coups, or civil wars (e.g. Muller and Seligson 1987; Wang, Dixon, Muller, and
Seligson 1993). Economic inequality within countries can induce political violence
because ‘relative deprivation’ generates domestic discontent (Gurr 1970), or because
inequality is often concomitant with social cleavages between classes, religions, or
regions (Lichbach 1989). In this regard, high inequality within countries is a potential
threat to political leadership especially in democracies, causing domestic discontent and
political violence.

According to diversionary theory, political leaders have an incentive to take
international military actions to remain in power by diverting attention from internal
problems such as a weak economy to external issues (e.g. DeRouen 1995; Leeds and
Davis 1997; Levy 1989; Morgan and Bickers 1992). In many cases, diversionary tactics
are strongest vis-a-vis a shock to the economy such as a very low growth rate. Regarding
the relationship between a shock to the economy and inequality, previous research reports
that inequality within countries is negatively correlated with growth (Knack and Keeger
1997; Persson and Tabellini 1994). It implies that a high level of inequality within
countries can provide an incentive of diversionary use of force to political leaders.51 If
trade or economic liberalization aggravates inequality within countries, political leaders’
incentive of diversionary use of force against trading partner is intensiﬁed. As a result,
the effect of trade on interstate conﬂict should be evaluated in two ways: a direct

pacifying effect and an indirect precipitating effect through inequality within countries.

 

5' A high level of inequality between countries also could be a crucial threat to political leadership. Since
this research focuses on the linkage between trade and inequality within countries and the linkage between
inequality and interstate conﬂict, the effect of inequality between countries on interstate conﬂict is not
discussed here.

88

Directly, trade may mitigate the probability of interstate conﬂict onset, keeping trade
partners holding onto the beneﬁts from trade. Indirectly, trade may heighten the level of
inequality within countries, domestic discontent, and thus the probability of interstate
conﬂict onset. These two seemingly contradictory arguments indicate that the pacifying
effect of trade can be offset by its negative effect on inequality within countries.
Therefore, to account for the relationship between trade and interstate conﬂict, we should
look at the duel aspects of trade.

The key questions in this research are as follows: Does trade alleviate or
aggravate the level of inequality within countries? Does inequality within countries
increase the probability of militarized interstate disputes? What is the overall effect of
trade or economic interdependence on the probability of militarized disputes? Based on
these research questions, this study ascertains the extent to which inequality within
countries is a relevant factor for political leaders faced with the decision to use force. It
tests a diversionary theory of the use of force that treats inequality within countries as an
endogenous variable. That is, since the level of inequality within countries can be a
function of other explanatory variables in the model of interstate conﬂict onset, inequality
within countries should be treated as an endogenous variable. This design allows the
author to assess the direct and the indirect effects of trade on the probability of interstate
conﬂict onset. It will also help to solve theoretical and statistical problems in the previous
studies of diversionary theory.

In what follows, first, I review diversionary theory and discuss theoretical and
statistical problems in models of diversionary theory. Inequality within countries as a

source of diversionary use of force is discussed. In the second section, the relationship

89

between trade and inequality within countries is examined. I argue that political leaders’
winning coalition is important in accounting for inequality within countries. In the ﬁnal
section, a two-step estimation approach to the model of the diversionary theory is

introduced, and the following test results are reported.

Diversionary Use of Interstate Conﬂict

Scholars of international relations have argued that political leaders have an
incentive to use military force to divert attention away from domestic problems to foreign
affairs (e.g. James and Hristoulas 1994; Leeds and Davis 1997; Levy 1989; Morgan and
Bickers 1992; Smith 1996). This is the case when political leaders have serious political
or economic problems at home. The rationale of the diversionary (or scapegoat) theory is
based on ‘the rally-around-the ﬂag effect’. When the public face a threat from the
outside, they tend to become more cohesive and supportive of their leader (Stein 1976).
Since a political leader’s popularity generally increases during an international conﬂict
(Mueller 1970; MacKuen 1983), he or she faced with domestic discontent has an
incentive to draw attention to international conﬂict issues.

The sources of domestic problems or discontent with respect to political leaders’
diversionary use of force are various. A faltering economy, domestic turmoil, and a low
public approval rate of political leaders are those examples. If political leaders fail to
manage to make an economy grow, political leaders will face domestic criticism and
challenges. If an economy declines with a high unemployment rate, high inﬂation, or a
negative growth rate, few political leaders can maintain their positions in ofﬁce.

Domestic turmoil such as civil wars and riots also endangers political leaders’ power

90

position. As the approval rate of political leaders comes down, it is likely that political
leaders’ incentive of using external conﬂict for diversionary purpose increases (MacKuen
1983; Russett 2000). For example, Kaiser (1983) argues that “German foreign policy
after 1897 must be understood as a response to the internal threat of socialism and
democracy” (1983, p.443).

A high level inequality within countries is also a source of social discontent. This
is because psychologically the perception of ‘relative deprivation’ stirs social discontent.
People feel relative deprivation when there is “a perceived discrepancy between men’s
value expectations and their value capabilities” (Gurr 1970, p.13). Income is the value
that contributes directly to physical well—being and self-realization. Therefore, a high
level of inequality in income distribution within countries can cause a strong feeling of
relative deprivation in a society. Economic inequality within countries and relative
deprivation spur social discontent and provide motivation for political instability and
collective violence such as revolution, rebellion, coups, or civil wars. For example,
Muller and Seligson (1987) report that income inequality strongly affects mass
violence.52 Midlarsky (1988) also argues that there is a strong systemic relationship
between land inequality and deaths from political violence in Latin American and Eastern
European countries. If income distribution is systematically associated with social
cleavages between classes, religions, or regions, inequality within countries heightens the
probability of political violence (Lichbach 1989). As a result, inequality within countries
undermines political leadership. Accordingly, political leaders’ incentive of diversionary

use of force is intensified.

 

52 In their research, income inequality is measured as the share of income going to the top 20% of all
households.

91

To evaluate diversionary theory, we also need to consider the opportunity to use
diversionary force abroad. Since political leaders’ intention regarding the use of
diversionary force is not always materialized in real politics, the opportunity to engage in
international conﬂict should be controlled for in the model of the diversionary theory
(Meemik 1994). Examples include studies about the relationship between the
diversionary use of force and electoral cycles, rivalry relationships, and alliances. Since
the rally effect lasts shortly but strongly, the incentive of a diversionary use of force gets
bigger, as an election comes near. Incumbent political leaders often stimulate the
economy for electoral gains (James and Hristoulas 1994; Morgan and Bickers 1992).
When states have enduring rivals,53 political leaders can easily divert domestic attention
to rivalry conﬂict (Leeds and Davis 1997; Mitchell and Prins 2001). Also, the greater the
threat to a nation’s defense commitments to allies, political leaders can be led to use
greater levels of military force (Meemik 1994). There are other opportunity-related
variables. Since major powers are less constrained by international systemic factors, they
are expected to be more prone to diversionary use of force than are minor power (Russett
1990). Democratic governments are more likely to Show diversionary behavior because
the governments are sensitive to public opinion and electoral outcomes (Morgan and
Bickers 1992; Smith 1996).

Examining inequality within countries as a source of diversionary use of force
abroad is especially interesting in regard to the effect of trade on the onset of interstate

conﬂict. One of the limitations in the previous studies of the liberal peace is that they

 

53 Enduring rivals are deﬁned as “actors whose relations are characterized by disagreement or competition
over some stakes that are viewed as important, where each perceives that the other poses a signiﬁcant
security threat, and where this competition and threat perception last for substantial periods of time”
(Hensel 1998, p.163).

92

have only accounted for the effect of trade on inequality between countries. If a high level
of trade is strongly associated with a high level of inequality within countries, inequality
within countries as well as inequality between countries is an important factor that
political leaders should consider when they make a decision to initiate interstate conﬂict.
Therefore, by controlling the effect of trade on inequality within countries, we will be
able to clearly evaluate the linkage between trade and interstate conﬂict.

This argument is relevant to the endogeneity problem. That is, inequality within
countries can be a function of other explanatory variables such as the regime-type and
trade variables. Since trade has a direct effect on the income distribution in an economy,
trade is a theoretically important variable in explaining the level of inequality within
countries (Agenor 2003; Dollar and Kraay 2001). The regime type of a state is also
assumed to have a strong effect on inequality within countries (Rodrik. 1999; 2000). The
existence of endogeneity in the model brings us biased estimates (Greene 2000).
Therefore, when we investigate the effect of inequality on the onset of interstate conﬂict,
the endogeneity problem should be controlled in an adequate way.

In addition to the endogeneity problem, studies of the diversionary theory face
some criticisms (Levy 1989). Theoretically, the rationale of the diversionary theory is
based on the in-group/out-group hypothesis (Coser 1956; Stein 1976). Therefore,
extending this argument on small groups to relatively large organizations such as nation
states can be problematic (Morgan and Bickers 1992). It implies that political leaders
more seriously take domestic discontent into account when it comes from within their

winning coalition than when it comes from other domestic groups. To clearly specify the

93

conditions under which political leaders use diversionary tactics, we should control for
opportunity-related variables (Mitchell and Prins 2001).

Empirically, if we use pooled data to account for whether internal problems lead
to external conﬂict, we can have incorrect empirical results. This is because states have
different internal conﬂict thresholds for diversionary use of force (Morgan and Bickers
1992).54 More seriously, the causal relationship between inequality within countries and
the external use of force can be reversed (Blainey 1973; Levy 1989). That is, the onset of
interstate conﬂict may contribute to domestic problems such as inequality within
countries. To control this problem, we need time lags between domestic problems and
interstate conﬂict.

The following section explores in detail the relationship between inequality within
countries and economic liberalization. It also demonstrates how this study responds to the

theoretical and statistical criticisms mentioned in the previous paragraph.

Trade and Inequality Within Countries

The impact of trade or economic globalization on world inequality can be
examined in two ways: inequality between countries and inequality within countries. By
ignoring inequality within countries, the empirical growth literature gives “a biased view
of the evolution of world inequality over time, clearly underestimating it” (Bourguignon
and Morrisson. 2002). In this regard, recent Studies have focused on the effect of
globalization or economic liberalization on inequality within countries (Agenor 2003;

Dollar and Kraay 2001; Frieden 2001; Lindert and Williamson 2001).

 

5" Since this research uses time-series-cross-section data, it can avoid or at least reduce this type of problem
(Morgan and Bickers 1992).

94

Despite scholarly efforts to clarify the effect of trade or economic liberalization
on inequality within countries, it seems that there is still no unanimous agreement on the
relationship between globalization and inequality within countries. For example, Agenor
(2003) argues that “trade liberalization leads to higher poverty by reducing the demand
for unskilled labor and worsening income distribution.” In contrast, Dollar and Kraay
(2001) report that expanded trade has positive effect on the income distribution. The
effect of globalization on inequality within countries “exhibits no ubiquitous trend”
(Lindert and Williamson 2001), or the effect is “non-monotonic” (Agenor 2003). In effect,
it seems that some countries have improved the level of equality within countries through
trade, while others have not.

To explain why trade or economic liberalization has heterogeneous effects on
inequality within countries, scholars suggest various reasoning. For example, Agenor
(2003) argues that economic liberalization or globalization may hurt the poor not because
it went too far, but rather because it did not go far enough. In this argument, trade or
economic liberalization is expected to have a long-term positive effect on equality within
countries.

A host of scholars suggest that there is a strong relationship between inequality
within countries and domestic political factors (e.g. Gradstein, Milanovic, and Ying
2001; Knack and Keefer 1997; Milanovic 2000). According to these arguments, a certain
forms of domestic political institutions favor equality within countries, and serve to
intervene in the relationship between trade and inequality within countries. Speciﬁcally,
in democracies, it is assumed that a more egalitarian distribution of political rights is

accompanied by a more equal income distribution (Lipset 1959; Meltzer and Richard

95

1981). Their arguments imply that democratic leaders perform better in achieving policy
goals than nondemocratic leaders.

In contrast to this theoretical expectation, the existing empirical evidence on the
relationship between the regime type and inequality within countries often fail to detect a
robust relationship between these two factors (Bollen and Jackman 1985; Levy 1989; Li,
Squire and Zou 1998). This research suggests that this discrepancy occurs because we
disregard the existence of various groups in domestic politics and their competing
interests. Political leaders should take support from the winning coalition more seriously
than the one from other domestic groups. A winning coalition is deﬁned as: “those
members whose support is essential to keep the incumbent leadership in ofﬁce” (Bueno
de Mesquita et al. 1999, p.148). Therefore, as the size of the winning coalition decreases,
leaders are less likely to care about producing good public policies such as equal income
distribution.” This is because they can easily maintain support from the winning
coalition by supplying private goods (Bueno de Mesquita et al. 2000; 1999; Gelpi and
Grieco 2001).

Finally, to properly evaluate the effect of trade or economic liberalization on
inequality within countries, this study controls for a theoretically important variable, the
winning coalition. The regime type variable may not efficiently capture the effect of
domestic political institutions on inequality within countries. This is because the size of
the winning coalition can explain inequality within countries more efﬁciently than the
regime type variable. This is also because the size of the winning coalition can vary a lot

even across the same regime type countries like democracies. Thus, the effect of the

 

55 This paper assumes that members in the winning coalition mainly care for their own beneﬁts. Thus,
members in a relatively small size of the winning coalition are less likely to be concerned about policies for
public goods.

96

regime type variable on inequality within countries can be conditional on the size of the
winning coalition. This research uses an interaction term between the regime type
variable and the winning coalition variable.56 To resolve the potential simultaneous
problem in the relationship between inequality within countries and the diversionary use

of force, this study uses one year lagged variable for inequality within countries.

Hypotheses

The ﬁrst hypothesis of this research tests the diversionary theory by treating

inequality within countries as a source of a diversionary use of force.

Hypothesis 3.1: As the level of inequality within countries increases, the
probability of initiating militarized interstate disputes increases,
holding all else constant.

In the first hypothesis, we expect that there is a Signiﬁcant relationship between
inequality within countries and the decision on initiating militarized interstate disputes.
The second hypothesis is to test an indirect effect of trade on interstate conﬂict. Trade is
assumed to indirectly increase the probability of militarized interstate disputes by

increasing the level of inequality within countries.

Hypothesis 3.2: A marginal increase in the level of trade indirectly increases
the probability of conﬂict by increasing inequality within country,

holding all else constant.

 

5" Gradstein et al.’s study (2003) is interesting in this sense. They report that democratization reduces
inequality in societies where equality is not valued as much, while democratization has a negligible effect
on inequality in societies where equality is highly valued. State penetration can also be related to
redistribution.

97

Hypothesis 3.3: As the size of the winning coalition increases,

the level of inequality within countries is lowered.

The final hypothesis tests if the size of the winning coalition matters for political
leaders’ behavior in regard to producing good public policies such as equality within
countries. Due to payable political costs from a high level of inequality within countries,
political leaders are believed to put more efforts on producing public goods, as the size of

the winning coalition increases.

Speciﬁcation
We have a binary dependent variable and a continuous endogenous explanatory

variable in the model. Therefore, the model is specified in the following form:

Yizzrfsl+aIY2+urr (3-1)
Y2 = 21621172522 +V2 = Z52 +V2 , (3-2)
y.=1 lyl>01 . (3.3)

where (u,,v2) has a zero mean, bivariate normal distribution and is independent of z. y,‘
is a latent variable of y1 , and y1 represents either the initiation of militarized interstate

dispute or not. Equation (3.1), along with Equation (3.2), is the structural equation. And,

Equation (3.2) is a reduced form for y,. In this Speciﬁcation, y1 = 1 is observed if there is
onset of a militarized dispute (Jones, Bremer and Singer 1996); otherwise y1 = 0. z1 is a

matrix of covariates such as the regime type, trade, rivalry, major power status, power

parity, the existence of alliance, and contiguity variables. y2 indicates the continuous
endogenous variable, inequality within countries. We are assuming that y2 given 2 is

normal. Inequality within countries is explained by some of the explanatory variables, 2, ,

98

in Equation (3.1) such as the regime type and trade variables and also by new explanatory

variables, z, , such as growth rate and the winning coalition variables.

To estimate a probit model with an endogenous explanatory variable, we can use
a two-step approach suggested by Rivers and Vuong (1988). Under the assumption of
joint normality of (u1 , v2) with Var(u,) = 1, we can write

u1 = By, + e,, (3.4)
where 6, = 77, / 2'3, 77, = Cov(v2,u,), 1% = Var(vz). And, e, is independent of z and V2. el
is normally distributed with E(e,) = 0 and Var( eI ) = Var(u1)-rﬁ/ I": =1—pf , where
.0. =Corr(v2.ul).
Finally, we can write a new equation:

yf = ad, + my, + 6,v2 +e,, (3.5)
where eI is normally distributed with zero mean and ( 1— pf ) variance. A standard
calculation shows that

z,6,+a,y2+6,v2]
(l-pf)” '

 

P()’r : 1|Z.)’2,V2)=<I)[

Assuming for the moment that we observe v2, probit of y1 on zl , y2 , and v2 consistently

estimates 5,,

rad/(1-901”, up, “em/(117%)”, and6p1561/(1—pf)”2 (refer to Wooldridge

(2002) for details).

Variables and Measurement
The dependent variable is an initiation of militarized interstate dispute by the
challenger state. This is coded as one if the challenger state initiated a militarized

interstate dispute in that year, and zero otherwise. Zeev Maoz’s (1999) revised data on

99

militarized interstate disputes is used. This data is expected to minimize the problem that
individual states in a dyad may never have threatened or used force against particular
states on the opposing side.

The endogenous variable is the inequality within countries variable. Challenger’s
inequality score in a dispute is used. Data for inequality within countries is from
Galbraith and Kum’s (2003) new data set, the Estimated Household Income Inequality
(EHII). This data set is intended to correct the deﬁciencies in the World Bank’s -
Deininger and Squire — data set on household income inequality (Deininger and Squire
1996). According to Galbraith and Kum’s argument, the World Bank’s data set is
problematic in that it includes “sparse coverage, problematic measurements, and the
combination of diverse data types into a single data set” (Galbraith and Kum 2003, p.1).
The Deininger and Squire data also has another problem, biased missingness. Certain
types of countries (i.e. poor, authoritarian, and those with extremely high levels of
inequalities) are less likely to report their data than other types of countries. Attempting
to correct these problems in the World Bank’s data set, Galbraith and Kum combine the
information in the Deininger and Squire data with the information in the University of
Texas Inequality Project (UTIP) data, along with additional information such as the ratio
of manufacturing employment to population and the share of urban population. Finally,
they provide estimated measures of household income inequality (the EHII Gini
coefﬁcients), based on the relationship between inequality of household incomes
(Deininger and Squire’s Gini coefficients), inequality of industrial pay (UTIP’S Theil
measures), and other variables (see Galbraith and Kum 2003 for details). The data set has

approximately 3,200 observations for about 150 countries over 36 years (1963-1998).

100

The Trade variable is measured by the ratio of dyadic trade to the GDP of each
partner, in constant US dollars. Low trade dependence value in a dyad is used for dyadic
trade. Rivalry is deﬁned by the frequency of rrrilitarized interstate disputes between the
same pair of states. The existence of a militarized rivalry is indicated by the occurrence of
militarized disputes as deﬁned by the COW—MID data set. Disputes, which occur within
10—15 years of each other, are considered to be part of the same rivalry. A dispute is
considered part of the same rivalry if it involves the same two states and occurs within 11
years of the ﬁrst dispute of the sequence, 12 years after the second dispute, and up to 15
years after the ﬁfth dispute.

For the Democracy variable, challenger’s polity score is used (Polity IV).
“Institutionalized democracy” score minus the “institutionalized autocracy” score for
each state is used as a measure for the level of democracy (Jaggers and Curt 1995).
Following Oneal and Russet’s suggestion (Oneal and Russett 1999, 2001), we do not
include a measure of economic growth in the structural equation. This is because it is not
robustly related to the likelihood of dyadic disputes. However, the reduced form equation
includes the GDP growth rate variable, because previous studies argue for the
theoretically strong relationship between inequality and growth rate (Perotti. 1992; Dollar
and Kraay. 2002). Challenger’s GDP per capita growth rate is used as a measure.

To measure the size of the winning coalition, we construct a composite index
based on the variables “chomp”, “Xropen”, “Parcomp”, and “Regtype” in Polity 111 data
set 57 (Jaggers and Gurr 1995). chomp variable measures the competitiveness of

executive recruitment. The value of this variable is 1, 2, or 3. A coded value of 1

 

57 “Regtype” variable existed in only Polity II data set and was not updated in Polity III and beyond. A
revised version of the data, Polity IIID, which more accurately dated political changes, was released in
2000.

101

indicates that the chief executive is selected by heredity or in unopposed elections. If this
variable is larger than or equal to 2, one point is assigned to the winning coalition
variable. Xropen variable measures the openness of executive recruitment. If the
executive is not recruited by heredity, another point is given to the winning coalition
variable. Parcomp variable measures competitiveness of political participation. If
political participation is highly competitive, another point is added to the winning
coalition variable. Regtype variable measures the regime type. If a political system is not
a military or military/civilian regime, one more point is assigned to the winning coalition
variable. With the base value of one, ﬁnally, this index varies one to ﬁve. The data for
this variable is from Polity II data set. We have many control variables. Detailed
explanation on data and its source is described in Table 9.

I include a smoothed function of elapsed time based on peace years. Peace years
spline estimation is used to control time dependence. This method is recommended by
Beck, Katz, and Tucker (1998). Although several studies Show that electoral cycles also
affect a govemment’s decision of diversionary use of force, we do not include the
electoral cycle variable in this research. This is because, unlike previous studies, this
research deals with a lot of nondemocracies, where formal elections have not been held
or elections are politically meaningless. Since not having elections is systematically
related to the regime type variable, and observing elections in nondemocracies can bring
biased results, we do not include the electoral cycle variable in this research. Actually, to
investigate the relationship between electoral cycles and the external use of force,

previous studies have limited objects only within democratic states (Smith 1996),

102

Table 9. Variable Descriptions

 

 

Variable Description

MIDs The onset of militarized dispute (1: onset). Zeev Maoz’s (1999) revised
data on militarized interstate disputes is used. Available in EuGene.

Trade The ratio of dyadic trades to the GDP of each partner in constant US

Inequality within
Countries

Democracy

Rivalry

Alliance

Parity

Contiguity

Distance

Major Power

Winning
Coalition

GDP Growth

Population
Growth

 

dollars. The trade dependence value for the least commercially engaged
member of the dyad is used for the level of economic interdependence. The
trade dependence value for the challenger state is used for the challenger.
Available in EuGene.

Galbraith and Kum’s (2003) new data set, the Estimated Household Income
Inequality (El-HI). This data set is intended to correct the deﬁciencies in the
World Bank’s — Deininger and Squire — data set on household income
inequality. The time period covered by this data set is 1963-1998.

The regime score for the challenger state. This score come from the Polity
IV “institutionalized democracy” score minus the “institutionalized
autocracy” score for each state (J aggers & Gurr 1995). Available in
EuGene.

Data used in "War and Peace in International Rivalry" by Paul Diehl and
Gary Goertz (2000). All the data used in all the analyses are available from
this web site: httpzllwsi.cso.uiuc.edu/poliscilfaculty/diehl.html.

Whether or not the dyad members are allied (l=allied). Available in
EuGene.

A continuous variable bounded between zero and one, indicating the ratio
of the weaker dyad member’s capabilities to those of the stronger dyad
member. Capabilities are based on the COW project’s Composite
Capabilities Index. Available in EuGene.

It is coded as one when the dyad members are geographically contiguous
by a land border. Available in EuGene.

The natural log of distances between capital cities. Available in EUGene.

A dichotomous variable. Coded as one if the challenger is a major power.
Available in EUGene.

A composite of four items: (1) “chomp” (the competitiveness of executive
recruitment), (2) “Xropen” (the openness of executive recruitment), (3)
“Parcomp” (the competitiveness of political participation), and (4)
“Regtype” (regime type) in Polity H and Polity 1H data (J aggers and Gurr
1995). Available in EUGene and ICPSR (Inter-university Consortium for
Political and Social Research) data archive in the university of Michigan.

The growth rate of GDP per capita. The average score of GDP per capita in
the last two years is used for comparing GDP growth rate. Virtually all the
data are derived from The World Economy: A Millennial Perspective
(Maddison 2001).

The growth rate of population [Population(t)/Population(t-1)]. Virtually all
the data are derived from The World Economy: A Millennial Perspective
(Maddison 2001).

103

advanced industrialized democracies (Leeds and Davis 1997), or the US. (Morgan and

Bickers 1992).

Result

As discussed in the previous sections, the inequality within countries variable is
endogenous to trade and democracy variables. The ﬁrst column in Table 10 shows what
happens if we disregard the endogeneity problem when actually we have.

The normal probit model in the ﬁrst column assumes no endogeneity problem in
it. It seems that all test results go along with theoretical expectation. As the trade
relationship becomes strong, the likelihood of militarized disputes between trade partners
decreases. As the level of inequality within countries increases, the probability of
initiating militarized interstate disputes increases. Major powers are more prone to
diversionary use of force than minor power. States with rivals are more easily engaged in
militarized disputes than states without rivals.

In the mean time, contrary to the theoretical expectation, it reports that democratic
governments are less likely to show diversionary behavior, and that this variable is not
statistically signiﬁcant. As discussed above, however, the normal probit model is very
likely to have the endogeneity problem, providing biased estimates in the analysis.
Accordingly, the ﬁrst step in this model is to check the existence of the endogeneity
problem. We can do this test by following Rivers and Vuong’s (1988) suggestion (see

also Wooldridge 2002). After running the OLS regression y, on 2 and saving the
residuals 17,, we run the probit y, on 2,, y, , and ix, to get consistent estimators of

interest. The usual probit t statistic on fl, is a valid test of the endogeneity. The

104

Table 10. Initiation of Militarized Interstate Disputes

 

 

MID Initiation (t) Probit Probit (a 2-step estimation)
Challenger Variable
Inequality (t-l) 0.0057** (0.0025) 0.0752*** (0.0126)
Democracy (t) — 0.0050 (0.0031) 0.0203*** (0.0057)
Major power (t) 0.5442*** (0.0692) 0.8284*** (0.1067)
Dyadic Variable
Rivalry (t) 0.9081 *** (0.0534) 0.9725*** (0.0604)
Trade (t) — 12.7545” (5.9746) — 8.8498 (6.2004)
Parity (t) — 0.0468 (0.0780) — 0.0684 (0.1014)
Alliance (t) — 0.0471 (0.0571) — 0.2156*** (0.0776)
Distance (t) — 0.2301*** (0.0239) -— 0.3197*** (0.0339)
Contiguity (t) 0.7175*** (0.0720) 0.4353*** (0.0860)
Constant - 1.1686*** (0.2201) — 3.2562*** (0.4892)

 

Inequality (t-l)

 

Trade (t-2)
GDP Growth (t-2)

Democracy (t- 1)

0.8385" (0.4071)
1.8903 (2.4140)

 

Winning Coalition (t- 1) 0.5951*** (0.0082)
Dem (t-l)*Winning Coalition (t-l) — 2.1189*** (0.0220)
Population Growth (t- 1) - 0.2209*** (0.0024)
Constant 74.6835*** (7.8149)
— 27.5962*** (8.4761)
Log-Likelihood -2087.49 - 1173.43
Chi-squared (12) 2993.40
Observations 291,568 150,767

 

 

Note: Standard errors are in parentheses. To correct the potential temporal dependence problem, I use
peace year spline variables suggested by Beck et al. (1998). These variables are not reported here to save
space.

*=p<0.1, **=p<0.05, ***=p<0.01

105

coefﬁcient of i2, is about 0.01 with its standard error, 0.004 (t-value is 0.012). Finally,

according to this test, we cannot say that there is no endogeneity problem in the normal
probit model.

The second column reports the results from the probit model with a two-step
estimation procedure, which corrects the endogeneity problem. In the probit model with a
two-step approach, trade is perceived to have dual effects on interstate conﬂict. Trade has
a direct effect on militarized interstate disputes. Since trade is assumed to affect the level
of inequality within countries, it has also an indirect effect on militarized interstate
disputes through inequality within countries. Without controlling the endogeneity
problem in the probit model, the Trade variable is statistically signiﬁcant, showing a
pacifying effect on the initiation of interstate conﬂict. When we control the endogeneity
problem in the probit model with a two-step approach, however, trade is not statistically
signiﬁcant anymore. But, in the reduced form equation, trade appears to increase the level
of inequality within countries, and accordingly the likelihood of militarized interstate
dispute onset. This result implies that the pacifying effect of trade on militarized disputes
can be offset by it negative effect on equality within countries, controlling other
theoretically signiﬁcant variables.

Most of empirical evidence supports our theoretical expectation. As inequality
within countries increases, the probability of a diversionary use of force also increases.
And, the bigger the size of the winning coalition is, the less political leaders care for
equality within countries. By correcting the endogeneity problem, the Democracy
variable in the structural equation now suggests that democratic governments are more

likely to show diversionary behavior than nondemocratic governments. This variable is

106

also statistically significant. However, the Democracy variable in the reduced form
equation has a decreasing effect on inequality within countries, when the size of the
winning coalition is big enough (bigger than 2.7 in the scale between 1 to 5, in our
analysis). As a result, these results imply that the Democracy variable also has dual
effects on the likelihood of militarized interstate disputes.

The probit model with a two-step approach has fewer observations than the
normal probit model does, due to the limitation in obtaining data for some state level
variables. To see if differences in the two models are originated from the difference in the
total observations, I ran the same probit model with the same size of observations in the
probit model with endogenous variable. The statistical results are almost same in the new
test. It conﬁrms that the statistical differences in the two models in Table 10 are not

caused by the differences in the total number of observations.

Conclusion

Understanding the effect of trade or economic liberalization on interstate conﬂict
is not simple, because trade has an impact on inequality within countries as well as on
other various areas in an economy. According to the result in this study, trade aggravates
the level of inequality within countries, holding other important variables constant. And,
as the level of inequality within countries increases, political leaders are more likely to
engage in international conﬂict as a way of diverting attention from domestic problems.
Finally, these results imply that the pacifying effect of trade on interstate conﬂict can be
offset by an increase in the level of inequality within countries. Accordingly, it warns us

that we can overestimate the pacifying effect of trade on interstate conﬂict.

107

The diversionary theory assumes that democratic governments are more likely to
Show diversionary behavior than nondemocratic governments, because of their sensitivity
to public opinion and electoral outcomes. However, the inequality within countries
variable can be endogenous to the regime type variable. This study shows that we can get
biased estimates by overlooking the endogeneity problem. As we analyze the effect of
trade on interstate conﬂict, we should distinguish the direct effect of the regime type
variable on diversionary use of force from its indirect effect through inequality within
countries. Given domestic problems, democracies are more likely to have diversionary
use of force. On the other hand, democracies are less likely to have domestic problems
such as inequality within countries than nondemocracies.

There are some potential problems in this research. First, there could be an
endogenous relationship between growth rate and trade in the reduced form equation,
although there is no clear scholarly agreement on the relationship between trade and
growth rate. Second, trade is not a sufﬁcient and a necessary condition for inequality
within countries. To get better understanding on the conditions under which the level of
inequality within countries increases, we may need to include other interesting economic
indicators such as ﬁnancial integration, technology transfer, labor migration, etc. into the
analysis. Finally, the probit model with continuous endogenous explanatory variable can
also be estimated by conditional maximum likelihood. Maximum likelihood estimation
(MLE) has some advantages over two-step procedures, which is used in this paper.
Although MLE is computationally demanding, this method is efﬁcient in estimating the
parameters of interest for computing partial effects (Wooldridge 2002). Testing the probit

model with conditional MLE will be desirable for future research.

108

Meanwhile, inequality within countries and domestic discontent may be more
strongly relevant to the onset of civil wars than interstate conﬂict onsets. Arguing for the
pacifying effects of globalization, analysts note that both the number and intensity of
conﬂicts between countries declined after the Cold War. But, internal conﬂicts like civil
wars have been proliferated in the early 19908. Given a growing number of civil wars in
recent history, this research can be extended to investigate the effect of globalization on

inequality within countries and accordingly on civil wars.

109

CONCLUSION

110

Globalization, Political Costs, and Strategic Decisions

To make a generalized argument, theories in social science often have to make
homogeneity assumptions even though they are dealing with heterogeneous units like
nation-states. In a multilevel analysis such as a two-level game model (Putnam 1988), we
often assume that there are no interactions between levels, or simply fail to specify a
theoretical relationship between levels. By disregarding the heterogeneity problem, or by
nus-specifying (or not-specifying) the relationship between multiple levels in a
theoretical analysis, we are likely to obtain a misleading or incomplete conclusion. These
seemingly unrealistic assumptions and mistakes are made partly because of empirical
difﬁculties in specifying a multilevel structure.

In this regard, the ﬁrst essay in this dissertation highlights the importance of the
empirical evaluation of theoretical models. To properly test the systemic account of the
democratic peace, this study employed multilevel modeling, showing how states interact
with the structure of the international system. Modeling multilevel structure in the data of
international politics is an issue of speciﬁcation, not just another way of estimating a
statistical model. Inappropriately speciﬁed models cannot properly account for a theory.
Just as the results of an empirical test are baseless without support from a theory, a well-

’58 without well-

developed theoretical argument cannot be recognized to be ‘scientiﬁc
deﬁned empirical speciﬁcation and test. The failure to set a bridge between a theory and a
well-deﬁned empirical method can cause an incorrect evaluation of a theoretical

argument. Contrast, well-deﬁned statistical models can make a theoretical argument

sharp and distinct.

 

58 Of course, not all theories are necessarily scientiﬁc. Descriptive and qualitative theories are also
contributive to the improvement of our understanding of social phenomena. For detailed arguments of
scientiﬁc research refer to Chalmer (1999).

111

The second essay shows how misspecification can bring a hasty judgment of the
effect of trade on interstate conﬂict. The liberal peace is based on an assumption that
policy costs from lost trade affect political costs for political leaders, forcing political
leaders to stick to trade beneﬁts. Sorting out monadic effects from dyadic effects in the
strategic approach to the liberal peace, the second essay demonstrates that political
leaders can take different courses of action even under the same dyadic relational
conditions, depending on domestic institutional context and the expected amount of
payable political costs.

The third essay views the effect of globalization on the world economy from the
perspective of the poor. In countries with a high level of inequality, economic
liberalization may be related with corruption, discontent, or violence in domestic politics.
In non-democratic countries especially with a small size of the winning coalition,
economic globalization has worsened the level of inequality within countries. The third
essay shows that political leaders are more likely to engage in international conﬂict as a
way of diverting attention from domestic issues, as the level of inequality within
countries grows, and that the pacifying effect of trade on interstate conﬂict can be offset
by its negative effect on equality within countries.

This study has broad implication on international politics. As the issues and actors
in the arena of international politics are diversiﬁed and interconnected with each other,
analyses on the relationship between countries becomes hard and complex. A political
economy of international relations emphasizes the nexus between the economy and
interstate conﬂict. In this regard, this study suggests a way of solving puzzles of

economic theories by incorporating domestic political factors into analyses. For example,

112

according to the Stopler-Samuelson theorem, international economic integration raises
inequality in the North (developed countries) but lowers it in the South (underdeveloped
countries). The reasoning behind this theorem is that trade liberalization benefits all
factors in the export industry but hurts all factors in the import-competing industry.
Unfortunately, empirical evidence does not go with this theoretical expectation. The level
of inequality has increased in both the North and the South. By incorporating domestic
political factors into economic analyses and capturing a disaggregated picture of
domestic politics, we may be able to improve the explanatory power of a theory and
reconcile the discrepancy between theoretical reasoning and empirical evidence.

Answering various research questions in three essays, this dissertation shows the
limitations in analyzing the effect of economic liberalization on international relations.
Globalization or economic liberalization may have decreased the number and the
intensity of interstate conﬂict. The increasing democracies in the international system
may also contribute to peace in the world politics. But, to evaluate the net effect of
economic liberalization on international politics, we should consider multilevel factors by
incorporating domestic politics into our analyses on international politics. It warns us not
to overestimate the effect of economic liberalization on interstate conﬂict, by
disregarding the importance of inequality within countries, domestic institutions, and
political leaders’ strategic behavior. Political leaders’ strategic consideration on payable
political costs affects decision-making on interstate conﬂict.

The strategic approach is advantageous in that it accounts for the interaction
between domestic politics and international politics. Due to the development of empirical

methods, a theoretical model in strategic approach can properly be tested. And, new

113

theoretical refinement will follow. The issue of economic liberalization and inequality
within countries may also seriously appeal to other areas as well as interstate conﬂict.
The increasing number of civil wars and terrorist attacks are those examples. By relaxing
unrealistic assumptions that states are equally affected by economic liberalization, and
that international politics is isolated with domestic politics, we can make a Signiﬁcant
improvement in scholarly understanding on international political events with logical

consistency.

114

APPENDICES

115

APPENDIX A.

Table 6. Classification of the size of winning coalition based on Bueno

 

 

de Mesquita et al. (2000).
Winning Coalition Non-democracy Democracy“
1 16.26% 0.00%
2 34.64% 0.00%
3 39.66% 17.65%
4 9.01% 42.24%
5 0.43% 40.1 1%

 

 

* A country is deﬁned as a democracy if the country has more than or equal to 6 value in

“dem” variable in Polity 111 data set.

** In general, Table 6 shows that democracies have bigger sizes of winning coalitions
than non-democracies do. However, it seems hard to explain variations in domestic
context (especially in democracies) using this measure. With respect to the size of

winning coalition, democracies are classiﬁed into only three categories, which may not

be realistic.

116

Table 7. Classification of the size of winning coalition based on DPI 2000.

 

 

Winning Non-democracy Democracy All
Coalition
1 90.16% 15.89% 42.69%
2 3.92% 6.42% 5.52%
3 3.24% 17.53% 12.38%
4 1.54% 30.82% 20.26%
5 0.92% 20.74% 13.59%
6 0.22% 2.14% 1.45%
7 0.00% 1.78% 1.14%
8 0.00% 0.24% 0.15%
9 0.00% 2.22% 1.42%
10 0.00% 0.57% 0.36%
l 1 0.00% 0.58% 0.36%
12 0.00% 0.00% 0.00%
13 0.00% 0.00% 0.00%
14 0.00% 0.00% 0.00%
15 0.00% 0.07% 0.04%
16 0.00% 0.93% 0.60%
17 0.00% 0.04% 0.02%
18 0.00% 0.04% 0.02%

 

 

* “Checks” variable in DPI 2000 data set is used as a measure for winning coalition in
this classiﬁcation. This table shows that most of non-democracies have very small sizes
of winning coalition. Theoretically, non-democracies have less variation in the size of
winning coalition than democracies. Political leaders in non-democracies can usually
maintain their political positions based on support from a relatively small group of people.
On the other hand, the necessary size of supporting group for keeping political leaders’
political positions in democracies varies a lot depending on domestic political situations.
In this regard, Checks variable seems advantageous especially for explaining variations
between democracies. The size of winning coalition in democracies varies a lot,

speciﬁcally, between zero and eighteen.

117

APPENDIX B.

Table 8. The Onset of Militarized Interstate Disputes (All dyads)

 

Variable

Democracies
Interdependence
Parity

Alliance
Distance
Contiguity

One power

Two powers
Interdep*WC
Interdep*S
Interdep*Checks

Constant

Variance
Components
State Level
Constant

Interdependence

Covariance
Dyad Level

Deviance (-2LL)
DIC

Logit with
splines modela

—0.050** (0.004)
—51.46** (7.321)
1.26** (0.080)
—0.l38 (0.071)
—0.640**(0.026)
2.781**(0.071)

2.271**(0.057)

3.162**(0.117)

0.669** (0.209)

19724.224

ANCOVA with
Random Effects
Model

—0.035** (0.007)

-22.598** (8.059)

0.969** (0.151)
—0.222** (0.088)
—l.078** (0.059)

2.620“ (0.154)

2.309M (0.136)

2.567" (0.203)

1.069** (0.465)

1.254" (0.225)

b3.002** (0.184)

14382.62
15466.04

The Slopes-as-
outcomes
Model 1

—0.031** (0.007)

61.806 (38.019)

0.979** (0.154)

—-0.207** (0.087)

—l.073** (0.058)

2.618** (0.148)

2.306“ (0.138)

2.554" (0.204)

—15.454 (8.005)

— 9.725 (14.791)

1.025** (0.438)

1.289“ (0.236)
1.515 (3.881)
0.060 (0.962)

The Slopes-as-
outcomes
Mode12

—0.033** (0.011)
14.206 (19.362)
0.955“ (0.253)
—O.l67 (0.176)
—1.360** (0.082)
2.949“ (0.209)

3.086“ (0.219)

4.916“ (0.637)

—16.80** (6.021)

1.549" (0.582)

1.736” (0.379)
0.326 (0.546)
-0.091 (0.370)

b2.978"‘* (0.209) b2.909** (0.339)

14389.210
15466.71

5241.254
5758.64

 

Note: Multilevel table entries are Bayesian estimates obtained from the Metropolis-Hastings (MH)
sampling methods. Standard errors are in parentheses. Spline variables are not reported here to save

space. (N: 466,626).

*=p<0.05,**=p<0.01

a Logit estimates are corrected for controlling the rare events problem.

b By a Binomial distributional assumption, the random coefﬁcient at the ﬁrst level was deﬁned with a
zero mean and a variance which was constrained to equal one. However, to test whether or not a
Binomial assumption holds or whether the model exhibits ‘extra Binomial’ variation, I obtained a
dyadic level variance based on the extra Binomial distributional assumption.

118

APPENDIX C.

Graphical inspection of convergence (politically relevant dyads)

State level variance Dyad level variance

 

   

 

 

Interdependence Democracy
0.00
. I . ’ 1,1,, A ,' ~ 1‘ 1‘ ., l . , :1, ”will”: ,lf 00
-0.
-O.
-0.1
Parity Alliance
2.0, o
1.61, .
l I ‘ l
1 I ‘ ‘ l
0 . , 1 ll 1
04,1
00 Leo -i+¢*§> a—e—v tﬁ—ﬁHl—Hnﬁﬁ—i—l

 

 

 

 

Distance Contiguity
0.0 v . i v. . 2.
l.
1.
0
0.4
0.0
-0.
O.
Two Powers
One Power ,&
12i
09 , . , .
o.s“‘ 1,. “ ,j
g: . .111 l. 1 I‘m .11 l. .‘ Illl. 1 11.11.
»0. ‘ ‘ ‘ ‘

 

-o. ‘ ‘ ‘ ‘ .
-0.9
.12

119

APPENDIX D.

Figure 7. Global Inequality of Individual Incomes, 1820-1992

 

1.0

0.8 -

0.6 n

0.4 -

Coefficients of Inequality

0.2 -

 

/ .
0.0 I I I I I I I I
1820 1840 1 860 1 880 1 900 1 920 1 940 1 960 1 980 2000

 

 

 

Year

 

— Global Inequality (Theil)

------- Inequality within countries (Theil)
— — Inequality within countries (Gini)
— -- Inequality between countries (T heil)

 

 

 

* Source: Bourguignon and Morrisson (2002). The data consists of 15 single countries
with abundant data and large populations and plus 18 groups of countries. The 18 groups
of countries are aggregates of geographical neighbors having similar sizes of GDP per

capita, as estimated by Maddison (1995).

120

Figure 8. Global Inequality of Individual Incomes, 1820-1992

1.0

 

Coefficients of Inequality

 

 

0.0 I I I I I I I I

 

 

1 820 1 840 1 860 1 880 1 900 1 920 1 940 1 960 1 980 2000

Year

 

— Global Inequality (Mean logarithmic deviation)
------- Inequality within countries (Mean logarithmic deviation)
— — Inequality between countries (Mean logarithmic deviation)

 

 

 

* Source: Bourguignon and Morrisson (2002). The data consists of 15 single countries
with abundant data and large populations and plus 18 groups of countries. The 18 groups
of countries are aggregates of geographical neighbors having similar sizes of GDP per

capita, as estimated by Maddison (1995).

121

BIBLIOGRAPHY

122

BIBLIOGRAPHY

Agnor, Pierre-Richard. 2003. “Does Globalization Hurt the Poor?” Forthcoming, World
Bank PRE. Working paper.

Alt, James E. & Gary King & Curtis Signorio. 2000. “Aggregation Among Binary,
Count, and Duration Models: Estimating the Same Quantities from Different
Levels of Data.” Political Analysis 9:1

Ames, Barry. 1987. Political Survival: Politicians and Public Policy in Latin America.
Berkeley, California. University of California Press.

Baldwin, David. 1995. “Security Studies and the End of the Cold War.” World Politics
48: 117-34.

Barbieri, Katherine & Gerald Schneider. 1999. “Globalization and Peace: Assessing New
Directions in the Study of Trade and Conﬂict.” Journal of Peace Research. Vol.36,
No.4: 387-404.

Barro, Robert. 2000. “Democracy and The Rule of Law.” In Governing for Prosperity,
ed. Bruce Bueno de Mesquita and Hilton L. Root. New Haven, CT. Yale
University Press: 209-231.

Beck, Nathaniel. 2001. “Time-series-cross-section data.” Statistica Neerlandica. Vol.55,
No.2:111-133.

Beck, Nathaniel. 2003. “Modeling Dynamics in the Study of Conﬂict: A Comment on
Oneal & Russett.” In Gerald Schneider Katherine Barbieri, Nils Petter Gleditsch.
Eds. Globalization and Armed Conﬂict. Lanham: Rowman & Littleﬁeld.

Beck, Nathaniel & Jonathan N. Katz. 2001a. “Random Coefﬁcient Models for Time-
Series-Cross-Section Data: The 2001 Version.” Presented at the 2001 Annual
Meeting of the Society for Political Methodology, Emory University.

Beck, Nathaniel & Jonathan N. Katz. 2001b. “Throwing Out the Baby with the Bath
Water: A Comment on Green, Kim, and Yoon.” International Organization.
Vol.55, No.2:487-495.

Beck, Nathaniel & Jonathan N. Katz, & Richard Tucker. 1998. “Taking Time Seriously:

Time-Series-Cross-Section Analysis with a Binary Dependent Variable.” American
Journal of Political Science. Vol.42, N 0.4: 1260-1288.

123

Beck, Nathaniel, and Richard Tucker. 1997. Conﬂict in Time and Space. Center for
International Affairs Working Paper, No. 97-8. Harvard University. [On line]
Available http://wwwl.columbia.edu/sec/dlc/ciao/wpsfrm.html.

Beck, Thorsten, George Clarke, Alberto Groff, Philip Keefer, and Patrick Walsh, 2001.
"New tools in comparative political economy: The Database of Political
Institutions." World Bank Economic Review. 15 (1): 165-176.

Bennett, D. Scott, and Allan C. Stam, III. 1998. EUGene: Expected Utility Generation
and Data Management Program. [On line] Available
http://wizard.ucr.edu/cps/eugene/eugene.html.

Benson, Michelle and Jacek Kugler. 1998. “Power Parity, Democracy, and the Severity
of Internal Violence.” Journal of Conﬂict Resolution. Vol. 42, No. 2: 196-209.

Berry, William D., Michael B. Berkman. 2000. “Legislative Professionalism and
incumbent Reelection: The Development of Institutional Boundaries.” American
Political Science Review. Vol. 94, No. 4: 859-874.

Blainey, G. 1973. The Causes of War. New York: Free Press.

Bollen, K. and R.W. Jackman. 1985. “Political democracy and the size distribution of
income.” American Sociological Review 50: 438-457.

Bourguignon, Francois. 1979. “Decomposable Income Inequality Measures.”
Econometrica. Vol.47, No.4:901-920.

Bourguignon, Francois & Christian Morrisson. 2002. “Inequality Among World Citizens:
1820 - 1992.” The American Economic Review. Vol.92, No.4:727-744.

Box-Steffensmeier, Janet M., Dan Reiter, & Chistopher Zorn. 2003. “Temporal
Dynamics and Heterogeneity in the Quantitative Study of International Conﬂict.”
In Edward D. Mansfield and Brian C. Pollins, eds. New Perspectives on Economic
Exchange and Armed Conﬂict. Ann Arbor: University of Michigan Pressz274-288.

Braumoeller, Bear F. 2003a. “A Dynamic Solution to the Agent-Structure Debate in
International relations.” Paper presented at the CEEISA/ISA International
Convention, Budapest, Hungary, June 26-28.

Braumoeller, Bear F. 2003b. “Causal Complexity and the Study of Politics.” Political
Analysis. Vol.11, No.3: 209-233.

Bueno de Mesquita, Bruce. 1975. “Measuring Systemic Polarity.” Journal of Conﬂict
Resolution 19: 187-216.

124

Bueno de Mesquita, Bruce. 2000. Principles of International Politics. Washington, DC,
Congressional Quarterly Inc.

Bueno de Mesquita, Bruce and Hilton L. Root. 2000. “When bad economics is good
politics.” In Governing for Prosperity, ed. Bruce Bueno de Mesquita and Hilton
L.Root. New Haven, CT. Yale University Press.

Bueno de Mesquita, Bruce, James D. Morrow, Randolph M. Siverson & Alastair Smith.
1999. Policy Failure and Political Survival. Journal of Conﬂict Resolution. Vol.43,
No.2: 147-161.

Bueno de Mesquita, Bruce, James D. Morrow, Randolph M. Siverson, and Alastair
Smith. 2000. “Political Institutions, Political Survival, and Policy success.” In
Governing for Prosperity, ed. Bruce Bueno de Mesquita and Hilton L. Root. New
Haven,CT. Yale University Pressz59-84.

Bueno de Mesquita, Bruce and Hilton L. Root. 2000. “When bad economics is good
politics.” In Governing for Prosperity, ed. Bruce Bueno de Mesquita and Hilton
L.Root. New Haven, CT. Yale University Press.

Buzan, Barry. 1995. “The Level of Analysis Problem Reconsidered.” In International
Relations Theory Today. Eds. by Ken Booth & Steve Smith. The Pennsylvania
State University Press. Pennsylvania: 198 - 216.

Buzan, Barry., Charles Jones, and Richard Little. 1993. The Logic of Anarchy:
Neorealism to Structural Realism. New York: Columbia University Press.

Chalmers, A.F. 1999. What is this thing called Science? 3rd ed. Hackett Publishing
Company, Inc. Indianapolis, Cambridge.

Congdon, Peter. 2003. Applied Bayesian Modeling. John Wiley & Sons Ltd, UK.

Coser, L. 1956. The function of social conﬂict. New York: Free Press.

Deininger, K. and L. Squire. 1996. “A New Data Set Measuring Income Inequality.”
World Bank Economic Review 10: 565-591. [On line] Available

httpzllwww.worldbank.org/research/growth.

DeRouen, Karl R. 1995. “The Indirect Link: Politics, The Economy, And The Use of
Force.” Journal of Conﬂict Resolution 39 (4): 671-695.

Dempster, A.P., D.B. Rubin, & R.K. Tsutakawa. 1981. “Estimation in Covariance
components models.” Journal of the American Statistical Association, 76: 341-353.

Dessler, David. 1989. “What’s at Stake in the Agent-Structure Debate?” International
Organization, 43: 441-74.

125

Deutsch, Karl W. and J. David Singer. 1964. “Multipolar Power Systems and
International Stability.” World Politics 16(2): 390 — 406.

Diehl, Paul F. 1991. “Geography and War: A Review and Assessment of the Empirical
Literature.” International Interactions, 17(1): 16-23.

Dixon, William J. 1993. “Democracy and the Management of International Conﬂict.”
Journal of Conﬂict Resolutions 37(1):42-68.

Dixon, William J. 1994. “Democracy and the Peaceful Settlement of International
Conﬂict.” American Political Science Review 88(1): 14-32.

Dollar, David and Aart Kraay. 2002. “Growth Is Good for the Poor.” Journal of
Economic Growth 7(3): 195-225.

Dougherty, James E. Robert L. Pfaltzgraff, Jr. 1997. Contending Theories of
International Relations: A Comprehensive Survey. 4th ed. Addison-Wesley
Educational Publishers Inc. New York.

Fearon, James D. 1994. “Domestic Political Audiences and the Escalation of
International Disputes.” American Political Science Review 88:577-592.

Feng, Yi. 2000. “Political institutions, economic growth, and democratic evolution.” In
Governing for Prosperity, ed. Bruce Bueno de Mesquita and Hilton L. Root. New
Haven, CT. Yale University Press.

Feng, Yi. 2001. “Political Freedom, Political Instability, and Policy Uncertainty: A Study
of Political Institutions and Private Investment in Developing Countries.”
International Studies Quarterly 45: 271-294.

Feng, Yi & Paul J. Zak. 1999. ”The Determinants of Democratic Transition.” Journal of
Conﬂict Resolution 43(2): 162-177.

Finnerrnore, Martha, and Kathryn Sikkink.1998. “International Norm Dynamics and
Political Change.” International Organization 52 (4): 887-917.

Gartzke, Erik. 1998. “Kant We All Just Get Along? Opportunity, Willingness, and the
Origins of the Democratic Peace.” American Journal of Political Science 42:1-27.

Gartzke, Erik and Quan Li. 2001. “War, Peace, and the, Invisible Hand: Positive Political
Extemalities of Economic Globalization.” Paper presented at the 2001 APSA
meetings, August 30 — September 2, San Francisco, CA.

Gates, Scott. 2001. “Empirically assessing the causes of civil war.” Working paper.
Michigan State University.

126

Gates, Scott, Haavard Hegre, Mark P. Jones, and Haavard Strand. 2001. “A
Multidimensional Institutional Representation of Political Systems (MIRPS).”
Paper presented at the Uppsala Conference on Data Collection, Uppsala, June 8-9.

Gelpi, Christopher & Joseph H. Grieco. 2001. “Attracting Trouble: Democracy,
Leadership tenure, and the Targeting of militarized challenges, 1918-1992.”
Journal of Conﬂict Resolution, Vol 45, No.6:794-817.

Giddens, Anthony. 1979. Central Problems in Social Theory: Action, Structure, and the
Contradiction in Social Analysis. Berkeley, University of California Press.

Gleditsch, Nils Petter & Kristian Skrede. 2002. “Expanded trade and GDP data Version
1.” Journal of Conﬂict Resolution 46 (5): 712-724.

Gleditsch, Nils Petter, & Havard Hegre. 1997. “Peace and Democracy: Three Levels of
Analysis.” Journal of Conﬂict Resolution 41(2): 283-310.

Gleditsch, Nils Petter, Havard Strand, Mikael Eriksson, Margareta Sollenberg & Peter
Wallensteen. 2001. “Armed Conﬂict 1946-99: A New Dataset.” Paper presented at
the 42'1d Annual Convention of the ISA, 2001, Chicago.

Goertz, Gary. 1994. Contexts of International Politics. Cambridge University Press. New
York.

Goldstein, Harvey. 1995. Multilevel Statistical Models. 2“. The Americas by Halsted
Press. New York.

Goldstein, Harvey, and Jon Rasbash. 1996.”Improved Approximations for Multilevel
Models with Binary Responses.” Journal of Royal Statistical Society. A, 159(3):
505-513.

Gradstein, Mark, Branko Milanovic, & Yvonne Ying. 2003. “Democracy and Income
Inequality: An Empirical Analysis.” Working Paper.

Greene, William H. 2000. Econometric Analysis. 4th ed. Prentice-Hall, Inc. New Jersey.

Green, Donald P., Kim Soo Yeon, & Yoon David H. 2001. “Dirty Pool.” International
Organization. Vol.55, No.2: 441-468.

Gurr, TR. 1970. Why Men Rebel? Princeton, NJ. Princeton University Press.

Hegre, Haavard. 2000. "Development and the Liberal Peace: What Does it Take to be a
Trading State?" Journal of Peace Research, Vol.37, No.1:5-30.

Hegre, Haavard & Arvid Raknerud. 1997. “The Hazard of WarzReassessing the Evidence
for the Democratic Peace.” Journal of Peace Research, Vol.34, No.4:385—404.

127

Hensel, Paul R. 1999. “An Evolutional Approach to the Study of Interstate Rivalry.”
Conﬂict Management and Peace Science 17: 175-206.

Hensel, Paul R. 2001. “Contentious Issues and World Politics: The Management of
Territorial Claims in the Americas, 1816-1992.” International Studies Quarterly 45
(1):81-109.

Hox, J .J . 1995. Applied Multilevel Analysis. Amsterdam. TT-Publikaties.

Huntington, Samuel P. 1991. The Third Wave: Democratization in the Late Twentieth
Century. Norman, OK and London: University of Oklahoma Press. .

Jackman, Simon. 2000. “Estimation and Inference via Bayesian Simulation: An
introduction to Markov Chain Monte Carlo.” American Journal of Political
Science 44 (2): 369-98.

Jaggers, Keith, and Ted R. Gurr. 1995. “Tracking Democracy’s Third Wave with the
Polity III Data.” Journal of Peace Research 32 (4): 469-82.

James, Patrick and Athanasios Hristoulas. 1994. “Domestic Politics and Foreign Policy:
Evaluation a Model of Crisis Activity for the United States.” The Journal of
Politics 56: 327-348.

Jones, Daniel, Stuart A. Bremer and J. David Singer. 1996. “Militarized Interstate
Disputes, 1816-1992: Rationale, Coding Rules, and Empirical Patterns.” Conﬂict
Management and Peace Science 15: 163-213.

Kaiser, D. E. 1983. “Germany and the Origins of the First World War.” Journal of
Modern History 55:442-74.

Kastner, Scott L., and Chad Rector. 2003. “International Regimes, Domestic Veto-
Players, and Capital Control Policy Stability.” International Studies Quarterly
47:1-22.

Katzenstein, Peter J ., ed. 1996. The Culture of National Security: Norms and Identity in
World Politics. New Yorszolumbia University Press.

Keefer, Philip. 2002. “DP12000, Database of Political Institutions: Changes and Variable
Deﬁnitions” Development Research Group, The World Bank. [On line] Available
httpzl/www.worldbank.org/research/bios/pkeefer.htm.

Keefer, Philip & Stephen Knack. 1997. “Why don’t countries catch up? A cross-national
test of an institutional explanation.” Economic Inquiry. Vol. XXXV:590-602.

King, Gary. 2001. “Proper Nouns and Methodological Propriety: Pooling Dyads in
International Relations Data.” International Organization. Vol.55, No.2: 497-507.

128

King, Gary & Langche Zeng. 2001. “Explaining Rare Events in International Relations.”
International Organization Vol. 55, No.3, Summerz693-715.

King, Gary, Michael Tomz, and Jason Wittenberg. 2000. “Making the Most of Statistical
Analyses: Improving Interpretation and Presentation.” American Journal of
Political Science 44, No.2, Aprilz347-61.

Knack, Stephen & Philip Keefer. 1997. “Does Inequality Harm Growth Only in
Democracies? A Replication and Extension.” American Journal of Political
Science Vol. 41, No.1, Januaryz323-332.

Krasner, Stephen D. 1983. “Structual Causes and Regime Consequences: Regimes as
Intervening Variables.” In International Regimes, ed. Stephen D. Krasner. Ithaca,

N.Y.: Cornell University Press.

Kugler, Jacek and Yi Feng. 1999. “Explaining and Modeling Democratic Transitions.”
Journal of Conﬂict Resolution. Vol. 43, No. 2: 139-146.

Lambom, Alan C. 1991. The Price of Power. Boston: Unwin Hyman.

Leblang, David. 2002. “The Political Economy of Speculative Attacks in the Developing
World.” International Studies Quarterly 46: 69-91.

Leeds, B. Ashley. 1999. “Domestic Political Institutions, Credible Commitments, and
International Cooperation.” American Journal of Political Science 43(4):979-1002.

Leeds, Brett Ashley, & David R. Davis. 1997. “Domestic Political Vulnerability and
International Disputes.” Journal of Conﬂict Resolution. Vol.41, No.6: 814-834.

Levy, Jack. 1989. “The diversionary theory of war: A critique.” In Handbook of War
Studies, edited by Manus Midlarsky. Boston: Unwin Hyman: 259-288.

Li, Hongyi, Lyn. Squire, and Heng-fu Zou. 1998. “Explaining international and
intertemporal variations in income inequality.” The Economic Journal 108: 26-43.

Lieberman, Robert C. 2002. “Ideas, Institutions, and Political Order: Explaining Political
Change.” American Political ScienceReview. Vol.96, No.4, December: 697-712.

Lindert, Peter H. & Jeffrey G. Williamson. 2001. “Does Globalization Make the World
More Unequal?” Working Paper.

Lipset, SM. 1959. “Some social requisites of democracy: Economic development and
political development.” America Political Science Review 53: 69-105.

MacKuen, M. 1983. “Political drama, economic conditions, and the dynamics of
presidential popularity.” American Journal of Political Science 27: 165-192.

129

Maddison, Angus. 1995. Monitoring the World Economy, 1820-1992. Paris, OECD.
Maoz, Zeev. 1999. Dyadic Militarized Interstate Disputes Dataset version 1.1.

Maoz, Zeev & Bruce M. Russett. 1993. “Normative and Structural causes of democratic
Peace.” American Political Science Review. Vol.87, No.3, September: 624-638.

Marshall, Monty G. and Keith J aggers. 2000. POLITY I V PROJECT. Intergrated Network
for Societal Conﬂict Research (INSCR) Program Center for International
Development and Conﬂict Management (CIDCM). University of Maryland.
[Online] Available http://www.bsos.umd.edu/cidcm/inscr/polity.

Martin, Andrew D. 2002. “Congressional Decision Making and the Separation of
Powers.” American Political Science Review. Vol.95, No.2:361-378.

Meemik, James. 1994. “Presidential Decision Making and the Political Use of Force.”
International Studies Quarterly 38: 121-138.

Meltzer, A. and S. Richard. 1981. “A rational theory of the size of government.” Journal
of Political Economy 89:914-927.

Milanovic, Branko. 2000. “The median-voter hypothesis, income inequality, and income
redistribution: an empirical test with the required data.” European Journal of
Political Economy. Vol. 16:367-410.

Mitchell, Sara McLaughlin. 2002. “A Kantian System? Democracy and Third-Party
Conﬂict Resolution.” American Journal of Political Science. Vol.46, No.4,
October: 749-759.

Mitchell, Sara McLaughlin and Brandon C. Prins. 2001. “Rivalry and Diversionary Uses
of Force.” Presented at the Annual Meeting of the Peace Science Society
(International), Atlanta, Georgia, October 26-28, 2001.

Mo, Jongryn.1994.”Two-Level Games with Endogenous Domestic Coalitions.” Journal
of Conﬂict Resolution 42:397-426.

Mo, Jongryn.1995.”Domestic Institutions and International Bargaining: The Role of
Agent Veto in Two-Level Games.” American Political Science Review. Vol.89,
No.4:914-924.

Morgan, T. Clifton, and Kenneth N. Bickers. 1992. “Domestic Discontent and the
External Use of Force.” Journal of Conﬂict Resolution 36(1): 25-52.

Morrow, J.D.1991.”Electoral and Congressional Incentives and Arms control.” Journal
of Conﬂict Resolution 35:245-65.

130

Most, Benjamin A., & Harvey Starr. 1989. “Opportunity and Willingness.” Inquiry,
Logic and International Politics. Columbia, SC: USC Press.

Mueller, J .1970.“Presidential popularity from Truman to Johnson.” American Political
Science Review 64: 18-34.

Newey, W. 1987. “Simultaneous estimation of limited dependent variable models with
endogenous explanatory variables.” Journal of Econometrics 36: 231-250.

Oneal, John R. & Bruce M. Russett. 1997. “The Classical Liberals Were Right:
Democracy, Interdependence, and Conﬂict, 1959-1985.” International Studies
Quarterly. 41: 267-94.

Oneal, John R. & Bruce M. Russett. 1999. “Assessing the Liberal Peace with Alternative

Specifications: Trade Still Reduces Conﬂict.” Journal of Peace Research. Vol.36,
No.4: 423-442.

Oneal, John R. & Bruce M. Russett. 2002. “Modeling Conﬂict While Studying
Dynamics,” in Gerald Schneider, Katherine Barbieri, and Nils Petter Gleditsch,
eds., Globalization and Armed Conﬂict.

O’Rourke, Kevin H. and Jefferey G. Williamson. 2000. “When Did Globalization
Begin?” NBER Working Paper 7632.

Ostrom, C.W., and BL. Job. 1986. “The President and the Political Use of Force.”
American Political Science Review 80 (2): 541-566.

Perotti, Roberto. 1992. “Income Distribution, Politics, and Growth.” Political-Economic
Equilibrium. Vol.82,No.2:3l 1-316.

Polachek, Solomon W., John Robst, & Yuan-Ching Chang. 1999. “Liberalism and
Interdependence:Extending the Trade-Conflict Model.” Journal of Peace
Research. Vol.36, No.4: 405-422.

Prins, Brandon C. & Christopher Sprecher. 1999. “Institutional Constraints, Political
Opposition, and Interstate Dispute Escalation: Evidence from Parliamentary

Systems, 1946-89.” Journal of Peace Research 36(3):271-87.

Przeworski, Adam and Fernando Limongi. 1993. “Political Regimes and Economic
Growth.” Journal of Economic Perspectives 7 (3): 51—69.

Putnam, Robert. 1988. “Diplomacy and Domestic Politics: The Logic of Two-Level
games.” International Organization 42:427-460.

131

Raftery, A. E. & S. M. Lewis. 1992. “How many iterations in the Gibbs sampler?” In
Bayesian Statistics 4, eds. J.M. Bernardo, J.O. Berger, A.P. Dawid and A.F.M.
Smith. Oxford, Oxford University Press: 765-76.

Rasbash, Jon, William Browne, Harvey Goldstein, Min Yang, Ian Plewis, Michael Healy,
Geoff Woodhouse, David Draper, Ian Langford, and Toby Lewis. 2002. A User’s
guide to MLwiN. Centre for Multilevel Modelling, Institute of Education ,
University of London. London, England.

Raudenbush, Stephen W. and Anthony S. Bryk. 2002. Hierarchical Linear Models:
Applications and Data Analysis Methods. 2"d . Sage Publications, Inc. Thousand
Oaks, California.

Raymond, Gregory A. 1994. “Democracies, Disputes, and Third-Party Intermediaries.”
Journal of Conﬂict Resolution 38 (1): 24-42.

Reed, William & David R. Clark. 2000. “War Initiators and War Winners.” Journal of
Conﬂict Resolution 44(3):378-395.

Reither, Dan & Allan C. Stam.1998.“Democracy, War Initiation, & Victory.” American
Political Science Review 92(2):377-390.

Reither, Dan.2003.“Exploring the Bargaining Model of War.” Perspectives on Politics
1(1):27-43.

Rivers, D., and Q. H. Vuong. 1988. “Limited Information Estimators and Exogeneity
Tests for Simultaneous Probit Models.” Journal of Econometrics 39: 347-366.

Rodriguez, German, and Noreen Goldman. 1995. “An Assessment of Estimation
Procedures for Multilevel Models with Binary Responses.” Journal of Royal
Statistical Society A, 158 (1): 73-89.

Rodrik, Dani. 1999. “Democracies Pay Higher Wages.” The Quarterly Journal of
Economics. Vol. CXIV, Issue 3.

Rodrik, Dani. 2000. “Institutions for High-Quality Growth: What They Are and How to
Acquire Them.” Studies in Comparative International Development. Vol.35,
No.3:3-31.

Russett, B. 1990. “Economic decline, electoral pressure, and the initiation of international
conﬂict.” In Prisoners of war? Edited by C. Gochman and A. Sabrosky: 123-140.
Lexington, KY. Lexington Books.

Russett, Bruce. 1993. Grasping the Democratic Peace: Principles for a Post-Cold War
World. Princeton: Princeton University Press.

132

Simon, Beth. 1999. “See You in ‘Court’? The Appeal to Quasi-Judicial Legal Processes
in the Settlement of Territorial Disputes.” In A Road Map to War: Territorial
Dimensions of International Conﬂict, ed. Paul F. Diehl. Nashville: Vanderbilt
University Press.

Smith, Alastair. 1996. “Diversionary Foreign Policy in Democratic System.”
International Studies Quarterly 40(1): 133-153.

Spiegelhalter, D., N. Best, B. Carlin, and A. van der Linde. 2002. “Bayesian measures of
model complexity and ﬁt.” Journal of Royal Statistical Society, 64B: 1-34.

Steenbergen, Macro R., and Bradford S. Jones. 2002. “Modeling Multilevel Data
Structures.” American Journal of Political Science, Vol.46, No.1 January: 218-237.

Stein, Arthur A. 1976. “Conﬂict and cohesion: A review of the literature.” Journal of
Conﬂict Resolution 20:143-165.

Trumbore, Peter F. & Mark A. Boyer. 2000. “International Crisis Decision-making as a
Two-Level Process.” Journal of Peace Research, Vol.37, No.6: 679 — 697.

stelis, George. 2002. Veto Players: How Political Institutions Work. Princeton
University Press. Princeton, New Jersey.

Waltz, Kenneth N. 1979. Theory of International Politics. Addison-Wesley Publishing
Company. California.

Wendt, Alexander. 1987. “The Agent-Structure Problem in International Relations
Theory.” International Organization 41(3):336-370.

Wendt, Alexander. 1999. Social Theory of International Politics. Cambridge: Cambridge
University Press.

Western, Bruce. 1998. “Causal Heterogeneity in Comparative Research: A Bayesian
Hierarchical Modeling Approach.” American Journal of Political Science 41(4):
1233-1259.

Wooldridge, Jeffrey M. 2000. Introductory Econometrics: A Modern Approach. South-
Westem College Publishing, Thomson Learning.

Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross Section and Panel Data.
The MIT Press, Cambridge, Massachusetts.

Zorn, Christopher. 2001. “Estimating Between- and Within-Cluster Covariate Effects,
with an Application to Models of International Disputes.” International
Interactions 27(4): 433—445.

133

   

, 'n~rﬂrw\tu\.\..‘.'